content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
In recent years the scientific community renewed its interest in understanding the properties of
multi-orbital Hubbard models, and this has been intensified by
the discovery of the iron-based superconductors.\cite{JohnstonAP2010,StewartRMP2011,
DaiNaturePhys2012,vanRoekeghem} On a theoretical front, this is a
challenging problem due to a lack of non-perturbative methods for treating multi-orbital
Hubbard models at intermediate or strong couplings and on extended systems.
Nevertheless, considerable progress has been made using mean-field-based approaches,
\cite{vanRoekeghem,YuPRL2013, FanfarilloPRB2015,GeorgesReview,Zhang,Knecht,WernerPRL2007,
LiebschPRL2005,deMediciPRL2009,deMediciPRB2005,MukherjeePRB2016,SemonPreprint,Ferrero}
resulting in new concepts such as that of a Hund's metal \cite{Zhang,YinNatureMaterials2011,GeorgesReview,LanataPRB2013}
and the orbital-selective Mott phase (OSMP).\cite{Anisimov2002,GeorgesReview} These concepts are
central to our understanding the paradoxical appearance of both localized and
itinerant characteristics in many multi-orbital systems \cite{deMediciPRL2014,MannellaReview}
and bad metallic behavior in the presence of sizable electronic correlations.\cite{deMediciPRL2014}
The most widely used numerical approach in this context
is single-site multi-orbital dynamical mean-field theory (DMFT).\cite{Footnote1,Georges,vanRoekeghem}
Generally speaking, DMFT maps the full lattice problem onto an impurity problem embedded in an
effective medium, which approximates the electron dynamics on a larger length scale
as a local renormalization. \cite{Georges}
While this technique has had considerable success in addressing many aspects of the OSMP and
other physics related to the multi-orbital problem,\cite{Knecht,LiebschPRL2005,ChanPRB2009,
LiebschPRB2004,BiermannPRL2005,deMediciPRL2009,deMediciPRB2005,WernerPRL2007,LiebschPRB2010,GregerPRL2013} it is unable to
capture spatial fluctuations and non-local correlations encoded in the $k$-dependent
self-energy $\Sigma({\bf k},\omega)$. This is a potential short coming
as non-local correlations are known to have an impact in the case of the
single-band Hubbard model.\cite{GullEPL2008,ParkPRL2008}
It is therefore important to assess the
importance of such non-local effects on multi-orbital properties such as the OSMP.
To date, most non-perturbative studies have used
cluster DMFT or the dynamical cluster approximation (DCA); \cite{NomuraPRB2015,LeePRB2011,LeePRL2010,BeachPRB2011,LeoPRL2008,SemonPreprint}
however, these techniques are typically limited to a handful of sites when multiple orbitals
are included in the basis. This is due to technical issues related to each choice in impurity solver,
such as the Fermion sign problem in the case of quantum Monte Carlo or the exponential growth of the
Hilbert space in the case of exact diagonalization. As a result, these studies have only addressed
short-range spatial fluctuations. One study of the OSMP has been carried out on a larger two-dimensional cluster
using determinant quantum Monte Carlo (DQMC). In that case, however, the OSMP was imposed by the model by
assuming that electrons in a subset of orbitals were localized as Ising spins.\cite{BouadimPRL2009}
In light of these limitations it is desirable to find situations where multi-orbital physics can be
modeled on extended clusters that support long-range spatial fluctuations and where the properties
under study emerge from the underlying many-body physics of the model.
In this regard, one dimensional (1D) models are quite promising. For example, two recent density matrix
renormalization group (DMRG) studies have been
carried out for an effective 1D three-orbital model representative of the iron-based superconductors.
\cite{RinconPRL2014,RinconPRB2014} More recently, it was demonstrated that
DQMC simulations for a simplified version of the same model can also be carried out to low temperatures
due to a surprisingly mild Fermion sign problem.\cite{LiuPreprint2016} These observations
open the doorway to non-perturbative studies of this model on extended clusters,
thus granting access to the momentum-resolved self-energies and non-local correlations.
1D studies along these lines are also directly relevant for the recently-discovered quasi-1D
selenide Ba$_{1-x}$K$_x$Fe$_2$Se$_3$.\cite{CaronPRB2011,CaronPRB2012,Patel,Shaui,LuoPRB2013,LuoPRB2014}
In this context, it is important to note that DMFT becomes more accurate in higher dimensions and
therefore one expects its ability to describe multi-orbital Mott physics in 1D to be
diminished.
Motivated by these considerations, we examine the properties of a three-orbital Hubbard
Hamiltonian on an extended 1D cluster using DQMC and DMRG, with a particular focus
on its $k$-resolved self-energies and spectral properties.
We thus gain explicit access to non-local correlations occurring on longer length scales than
those addressed in previous non-perturbative studies. In general, we find that the
OSMP leads to a mixture of localized and itinerant bands, where the former are characterized by a
localized (momentum-independent)
self-energy while the latter exhibits significant non-local (momentum-dependent) correlations.
This leads to a band-dependent shift of the underlying band structures. We also
identify an insulating state driven by orbital ordering in a region of parameter space previously
associated with an OSMP.\cite{RinconPRL2014,RinconPRB2014}
\section{Methods}\label{Sec:Methods}
\subsection{Model Hamiltonian}
We study a simplified three-orbital model defined on a 1D chain as
introduced in Ref. \onlinecite{RinconPRL2014}.
This model displays a rich variety of phases including block ferromagnetism,
antiferromagnetism, Mott insulting phases, metallic and band insulating phases,
and several distinct OSMPs.\cite{RinconPRL2014,RinconPRB2014,LiuPreprint2016}
The Hamiltonian is $H = H_\mathrm{0} + H_\mathrm{int}$, where
\begin{equation}\label{Eq:H0}
H_\mathrm{0} = - \sum_{\substack{\langle i,j \rangle\\\sigma,\gamma,\gamma^\prime}}
t^{\phantom{\dagger}}_{\gamma\gamma^\prime} c^\dagger_{i,\gamma,\sigma}c^{\phantom{\dagger}}_{j,\gamma^\prime,\sigma}
+ \sum_{i,\sigma,\gamma} (\Delta_\gamma-\mu) \hat{n}_{{i},\gamma,\sigma}
\end{equation}
contains the non-interacting terms of $H$, and
\begin{equation}\nonumber
\begin{split}
H_\mathrm{int}&=U\sum_{i,\gamma} \hat{n}_{i,\gamma,\uparrow}\hat{n}_{i,\gamma,\downarrow} +
\left(U^\prime - \frac{J}{2}\right) \sum_{\substack{i,\sigma,\sigma^\prime\\
\gamma < \gamma^\prime}} \hat{n}_{i,\gamma,\sigma}\hat{n}_{i,\gamma,\sigma^\prime}\\
&+ J \sum_{i,\gamma<\gamma^\prime} S^\mathrm{z}_{i,\gamma} S^\mathrm{z}_{i,\gamma^\prime}
\end{split}
\end{equation}
contains the on-site Hubbard and Hund's interaction terms. Here, $\langle \dots \rangle$
denotes a sum over nearest-neighbors, $c^\dagger_{i,\gamma,\sigma}$
$(c^{\phantom{\dagger}}_{i,\gamma,\sigma})$ creates (annihilates) a spin $\sigma$ electron in orbital $\gamma = 1,2,3$
on site $i$, $\Delta_\gamma$ are the on-site energies for each orbital,
$S^\mathrm{z}_{i,\gamma}$ is the z-component of the spin operator ${\bf S}_{i,\gamma}$,
and $\hat{n}^{\phantom{\dagger}}_{i,\gamma,\sigma} = c^\dagger_{i,\gamma,\sigma}c^{\phantom{\dagger}}_{i,\gamma,\sigma}$ is the
particle number
operator. The pair-hopping and spin-flip terms of the interaction have been neglected in order
to manage the sign problem in the DQMC calculations.
Following Ref. \onlinecite{RinconPRL2014}, we set $t_{11} = t_{22} = -0.5$, $t_{33} = -0.15$,
$t_{13} = t_{23} = 0.1$, $t_{12} = 0$, $\Delta_1 = -0.1$, $\Delta_2 = 0$, and $\Delta_3 = 0.8$
in units of eV while the chemical potential $\mu$ is adjusted to obtain the desired filling.
These parameters produce a non-interacting band structure
analogous to the iron-based superconductors, with two hole-like bands
centered at $k = 0$ and an electron-like band centered at $k = \pi/a$,
as shown in Fig. \ref{Fig:Bands}. Due to the weak inter-orbital hopping, each of the bands
is primarily derived from a single orbital, as indicated by the
line thickness and colors in Fig. \ref{Fig:Bands}. One can therefore (loosely)
regard the orbital character as an indicator of the band in this model.
For example, the top most band is primarily composed of orbital $\gamma = 3$. The total bandwidth
of the non-interacting model is $W = 4.9|t_{11}| = 2.45$ eV. This will serve as our unit of energy.
We further set $a = 1$ as our unit of length. The interaction parameters are fixed
to $U^\prime = U - 2J$, $J = U/4$, while $U$ is varied. This parameter regime results in a
robust OSMP for intermediate values of $U$, which is our focus here.
\begin{figure}
\includegraphics[width=0.75\columnwidth]{./noninteractingbands.pdf}
\caption{\label{Fig:Bands} (color online) A fat band plot of the non-interacting band structure
at a total filling of $\langle \hat{n} \rangle = 4$, where the thickness of the
lines indicates the majority orbital content of the band.
The top most band has the narrowest bandwidth and is primarily of orbital 3 character.
The lower two bands disperse over a much larger energy range and are primarily composed
of orbitals 1 and 2, respectively. }
\end{figure}
\subsection{DQMC and DMRG Calculations}
The model is studied using non-perturbative DQMC and DMRG methods.
The details of these techniques can be found in Refs. \onlinecite{WhitePRB1989,DQMC_Review,RademakerPRB2013} (DQMC)
and Refs. \onlinecite{White1} and \onlinecite{White2} (DMRG).
These approaches are complementary to one another;
DMRG works in the canonical ensemble and provides access to the ground state properties of the system while
DQMC works in the grand canonical ensemble and provides access to finite temperatures and fluctuations in
particle number. Both methods are capable of treating large cluster sizes such that non-local
correlations can be captured without approximation for the specified Hamiltonian.
The primary drawback to DQMC is the Fermion sign problem,\cite{Sign1,Sign2}
which typically limits the range of accessible temperatures for many models. Indeed, when the
spin-flip and pair hopping terms of the Hund's interaction are included in the Hamiltonian, we find that the
model has a prohibitive sign problem. But when these terms are
neglected the corresponding sign problem becomes very mild,\cite{LiuPreprint2016}
even in comparison to similar simplified multi-orbital models in 2D.\cite{RademakerPRB2013,BouadimPRB2008}
Given that these terms do not qualitatively affect the
phase diagram \cite{LiuPreprint2016} for the current model, we have neglected them here.
This has allowed us to study clusters of up to $L = 24$ sites in length
($3L$ orbitals in total) down to temperatures as low as $\beta = 74/W$.\cite{LiuPreprint2016}
At this low of a temperature we begin to see the onset of magnetic correlations in our cluster,
however, as we will show, the OSMP forms at a much higher temperature. Since the latter phase is
our focus here, we primarily show DQMC results for $\beta \le 19.6/W$ throughout.
In all cases shown here,
the average value of the Fermion sign is greater than $0.87 \pm 0.01$. Unless otherwise stated, all of our
DQMC results were obtained on an $L =24$ site cluster with periodic boundary conditions and for
an average filling of $\langle n\rangle = 4$ electrons, which corresponds to $2/3$ filling.
DQMC provides direct access to various quantities defined in the
imaginary time $\tau$ or Matsubara frequency $\mathrm{i} \omega_n$ axes.
In Sec. \ref{Sec:Spectral} we will examine the spectral properties of our model,
which requires an analytic continuation to the real frequency axis.
This was accomplished using the method of Maximum Entropy,\cite{MaxEnt}
as implemented in Ref. \onlinecite{FuchsPRE2010}.
Our DMRG results were obtained on variable length chains with open boundary conditions.
The chemical potential term in Eq. (\ref{Eq:H0}) is dropped for these calculations.
In all of the DMRG calculations the truncation tolerance is between $10^{-5}$ -- $10^{-7}$.
We performed three to five full sweeps of finite DMRG algorithm and used 300 states
for calculating both the ground state and the spectral function.
Once the ground state is obtained using the standard DMRG algorithm, we
computed the spectral function using the correction vector targeting
in Krylov space
\cite{DMRG_Method1,DMRG_Method2}, with an broadening of $\eta = 0.001$ eV.
\section{Results}\label{Sec:Results}
\subsection{Self-energies in the OSMP}
We begin by examining some of the standard metrics for the formation of an OSMP,
namely the average filling per orbital and the quasiparticle residue $Z_\gamma(k,\mathrm{i}\omega_n)$.
DQMC results for $\langle n \rangle = 4$ and $U/W = 0.8$ are summarized in Fig. \ref{Fig:n_beta}. The temperature
dependence of the individual orbital occupations $\langle n_\gamma \rangle$,
plotted in Fig. \ref{Fig:n_beta}a, has the standard indications of the
formation of an OSMP: At high temperature (small $\beta$) we see noninteger fillings for all
three orbitals. As the temperature is lowered (large $\beta$), however, orbitals one and two
smoothly approach fillings of $\sim 1.53$ and $\sim 1.47$, respectively, while orbital three
locks into an integer value of exactly 1. In some studies these values for the
average occupation are often taken as an
indication of an OSMP,\cite{RinconPRL2014,RinconPRB2014} where orbital three has undergone a transition to a Mott
insulating state while orbitals one and two host itinerant electrons. However,
as we will show, this is not always the case. For $U/W =
0.8$ the two fractionally filled orbitals are in fact itinerant, but for larger
values of $U/W$ these same orbitals retain a fractional filling but are
driven into an insulating state by the onset of orbital ordering in these two orbitals.
\begin{figure}[t]
\includegraphics[width=0.66\columnwidth]{./n_beta.pdf}
\caption{\label{Fig:n_beta} (color online) Orbitally resolved electronic
properties for $U/W = 0.8$ ($W = 2.45$~eV) at
different temperatures. (a) The temperature dependence of orbital occupations.
(b) The orbital resolved quasiparticle residue $Z_\gamma(k,\mathrm{i}\pi/\beta)$ at an
inverse temperature $\beta = 19.6/W$. (c)
The normalized electron self energies Im$\Sigma_\gamma(k,\mathrm{i}\pi/\beta)$ at
$\omega_n=\pi/\beta$ as a function of momentum.
Each curve is normalized by its $k = 0$ value to highlight the overall momentum dependence.
The scale is determined by $\mathrm{Im}\Sigma_\gamma(0,\mathrm{i}\pi/\beta) = -0.53$, $-0.57$,
and $-2.53$ for $\gamma = 1,2,3$, respectively, and in units of the bandwidth $W$. The blue,
red, and green dash lines in (b) and (c) correspond to the bare Fermi
momentum of the non-interacting bands. Panel (d)
shows orbitally resolved quasiparticle residues $Z_\gamma(k_\mathrm{F}^0,\mathrm{i}\pi/\beta)$ and
self energies
Im$\Sigma_\gamma(k_\mathrm{F}^0,\mathrm{i}\pi/\beta)$ at Fermi momentum as a function of temperature.
In each panel, error bars smaller than the marker size have been
suppressed for clarity.
}
\end{figure}
The mixed itinerant/localized nature of the OSMP at $U/W = 0.8$ is reflected in the momentum
dependence of quasi-particle residue $Z_\gamma(k,\mathrm{i}\pi/\beta)$ and
the orbitally resolved normalized self-energies
$R(k) = {\mathrm{Im}}\Sigma_\gamma(k,\mathrm{i}\pi/\beta)/{\mathrm{Im}}\Sigma_\gamma(0,\mathrm{i}\pi/\beta)$, plotted in
Figs. \ref{Fig:n_beta}c and \ref{Fig:n_beta}d, respectively, for $\omega_n = \pi/\beta$.
The self-energy is extracted from the dressed Green's function using Dyson's equation
\begin{equation}
\hat{G}^{-1}(k,\mathrm{i}\omega_n) = \hat{G}^{-1}_0(k,\mathrm{i}\omega_n) - \hat{\Sigma}(k,\mathrm{i}\omega_n),
\end{equation}
where the $\hat{G}$ notation denotes a matrix in orbital space,
$\hat{G}_0(k,\mathrm{i}\omega_n) = [\mathrm{i}\omega_n\hat{I} - \hat{H}_0(k)]^{-1}$ is the non-interacting Green's function,
and $\hat{H}_0(k)$ is the Fourier transform of the non-interacting Hamiltonian defined in orbital space.
The quasi-particle residue is obtained from the diagonal part of the
self-energy using the identity
\begin{equation}
\hat{Z}(k,\mathrm{i}\pi/\beta) = \left(\hat{I} - \frac{\mathrm{Im}\hat{\Sigma}
(k,\mathrm{i}\pi/\beta)}{\pi/\beta}\right)^{-1},
\end{equation}
where $\hat{I}$ is a $3\times 3$ unit matrix.
As can be seen from Fig. \ref{Fig:n_beta}c, the self-energies for each orbital have a sizable $k$-dependence
at this temperature. (In this case we have normalized the self-energy by its value at $k = 0$ in
order to highlight the overall momentum dependence. The magnitude of $\mathrm{Im}\Sigma_\gamma(0,\mathrm{i}\pi/\beta)$
is given in the figure caption.) In the case of orbitals one and two, the magnitude of the self-energy
varies by nearly 50\% throughout the Brillouin zone. In contrast,
the momentum dependence of $\Sigma_3(k,\mathrm{i}\pi/\beta)$ for orbital three is
much weaker, varying by only 5-10\% and reflecting the localized nature
of the carriers in these
orbitals. Similarly, the quasi-particle residue for the orbital three is essentially
momentum independent, while it increases for the two itinerant orbitals as
$k$ tracks towards the zone boundary.
The $k$ dependence at the remaining Matsubara frequencies accessible to our simulations
(not shown) exhibits a similar trend, with orbitals one and two having a strong $k$-dependence
while orbital three is nearly momentum independent at each $\omega_n$.
The momentum dependence shown in Fig. \ref{Fig:n_beta} indicates that the local self-energy
approximation introduced by DMFT may miss quantitative aspects of the electronic correlations
in the OSMP with mixed itinerant and local characteristics.
It should be noted that our results have been obtained in 1D,
which is the worst case situation for DMFT. \cite{Akerlund_PRD}
It is expected that the local approximation will perform better in higher dimensions, since DMFT
becomes exact in the limit of infinite dimensions; however, it is unclear how well the method
will capture similar non-local correlations in two dimensions relevant for the Fe-based superconductors.
A recent study \cite{SemonPreprint} has argued that the local approximation is quite accurate
for parameters relevant to the iron-based superconductors, however, it remains to be seen if this
will remain true for all parameter regimes or when longer range fluctuations are included.
Our results further highlight the need for the
continued development of numerical methods capable of handling the strong
Hubbard and Hund's interactions in intermediate dimensions and on extended clusters.
Figure \ref{Fig:n_beta}d examines the temperature dependence of
$Z(k^0_\mathrm{F},\frac{\mathrm{i}\pi}{\beta})$ and Im$\Sigma(k^0_\mathrm{F},\frac{\mathrm{i}\pi}{\beta})$ at the
Fermi momenta $k^0_\mathrm{F}$ of the non-interacting system.
(These are indicated by the dashed lines in \ref{Fig:n_beta}b and \ref{Fig:n_beta}c.)
Here, we find indications of anomalous behavior for the itinerant electrons,
where the quasiparticle residues of all three orbitals decrease with temperature.
This is accompanied by an increase in Im$\Sigma(k_\mathrm{F},\frac{\mathrm{i}\pi}{\beta})$ as $T$
is lowered. This is perhaps expected for orbital three, as $Z$ (Im$\Sigma$) for
the localized orbitals should decrease (increase) as this orbital becomes more
localized. For the itinerant orbitals, however, one would naively expect the
self-energy to decrease as temperature is lowered, which is
opposite to what is observed. We believe that this is due to the Hund's interaction between
the itinerant electrons and the localized spins on
orbital three. At this temperature we find no evidence of a magnetic ordering in our
model,\cite{LiuPreprint2016} despite the fact that a local moment has clearly formed in
the OSMP.
This means that the orientation of the local moment is random and fluctuating
at these temperatures. This produces a fluctuating potential
acting on the itinerant electrons via the Hund's coupling,
thus generating a residual scattering mechanism at low temperatures that reduces
the quasiparticle residue and increases the self-energy.
\subsection{Momentum and Temperature Dependence of the Spectral Weight}
Next, we turn to the momentum dependence of the spectral weight for the three orbitals in the vicinity of the
Fermi level. This can be estimated directly from
the imaginary time Green's function, where the spectral weight at momentum $k$ is proportional to
$\beta G(k,\tau=\beta/2)$.\cite{spectral} Using this relationship we do not have to
perform the extra step of analytically continuing the data to the real frequency axis.
Figures \ref{Fig:Gbeta}a-\ref{Fig:Gbeta}c summarize
$\beta G(k,\beta/2)$ for $U/W=0.1$, $U/W=0.8$, and $U/W=2$, respectively.
The results in the weak coupling limit ($U/W = 0.1$, Fig. \ref{Fig:Gbeta}a) are consistent with that
of a fully itinerant system: all three orbitals have a maximal spectral weight at a
momentum point very close to the Fermi momenta of the non-interacting system (indicated by the dashed lines).
This is exactly the behavior one expects for a well-defined quasi-particle band dispersing
through $E_\mathrm{F}$, where the peak in the spectral weight occurs at $k_\mathrm{F}$. The proximity of the
peaks in $\beta G(k,\beta/2)$ to the non-interacting values of $k_\mathrm{F}$ indicates that the
Fermi surface is only weakly shifted for this value of the interaction parameters. However,
as we will show in Sec. \ref{Sec:BandShifts}, these shifts are band dependent.
\begin{figure}[t]
\includegraphics[width=0.7\columnwidth]{./Gbeta.pdf}
\caption{\label{Fig:Gbeta} (color online) The
momentum dependence of Green functions $G(k,\tau=\beta/2)$ for a) $U/W=0.1$,
b) $0.8$, and c) $2.0$. The inverse temperature in all three cases is $\beta=19.6/W$.
The blue, red, and green dash lines in each panel indicate the Fermi
momentum of the three non-interacting bands.
(d) $G(k_\mathrm{F},\tau=\beta/2)$ as a function of inverse temperatures $\beta$
for the OSMP $U/W=0.8$. Error bars smaller than the marker size
have been suppressed for clarity.
}
\end{figure}
In the intermediate coupling regime ($U/W = 0.8$, Fig. \ref{Fig:Gbeta}b), where the
OSMP has formed, we again see both localized and itinerant characteristics. The
spectral weight of the localized orbital is small and independent of momentum,
as expected for the formation of a localized Mott state. Conversely, the
spectral weight of the remaining orbitals still exhibits a momentum dependence
characteristic of dispersive bands. Despite this, the total spectral weight is
decreased, indicating that spectral weight has been transferred to higher binding
energies by the Hubbard and Hund's interactions.
This is also reflected in the position of the maximum spectral weight, which has
shifted to a slightly larger $k$ value due to a
renormalization of the Fermi surface by the interactions. We also observe
that the spectral weight at the zone boundary increases relative to the zone center,
consistent with a flattening of the bands and a broadening of the spectral
function with increasing $U$. (This will be confirmed shortly when we examine
the spectral functions directly.) A similar transfer of spectral weight was observed in a
two-dimensional cluster DMFT study.\cite{NomuraPRB2015}
The temperature evolution of spectral weight $\beta G(k_\mathrm{F},\beta/2)$ at the
Fermi momentum for the OSMP ($U/W = 0.8$) is shown in Figure \ref{Fig:Gbeta}d.
In a metallic system one generally expects the spectral weight at the Fermi
level to increase as the temperature is decreased.
Initially, this is what is observed for all three orbitals,
however, the spectral weight for orbital three reaches a maximum around
$\beta = 7.5/W$ before decreasing as the temperature is lowered further and the OSMP gap
forms on this orbital. Conversely, the spectral weight of the itinerant orbitals
continues to rise until saturating at $\beta/W \approx 15$. This saturation is
again due to the presence of a residual scattering channel, which we associate with
the fluctuating localized spins present on the localized orbital three.
The $U/W = 0.8$ results confirm the mixed itinerant/local character of the
model at intermediate coupling.
When the value of $U$ is further increased, we find that all three bands
become localized while maintaining partial occupancies for each band.
To demonstrate this, Fig. \ref{Fig:Gbeta}c shows results for $U/W = 2$.
In this case, the orbital occupations
for the three orbitals are $\langle n_1 \rangle = 1.55$, $\langle n_2 \rangle = 1.44$,
$\langle n_3 \rangle = 1$, which are similar to those obtained
at $U/W = 0.8$. At face value one might therefore conclude that the system is in an
OSMP,\cite{RinconPRL2014,LiuPreprint2016} however, an examination of the spectral
weight reveals that the system is in fact insulating.
As can be seen in Fig. \ref{Fig:Gbeta}c, at $U/W = 2$ and $\beta = 19.6/W$,
$\beta G(k,\beta/2)$ is nearly momentum independent and the total spectral weight of
all three orbitals has significantly decreased (note the change in scale of the y-axis). This
behavior is indicative of the formation of a charge gap throughout the
Brillouin zone. The ultimate origin of this insulating behavior is the formation
of a long-range orbital ordering, as we will show in Sec. \ref{Sec:Spectral}.
\subsection{Band-dependent Fermi surface renormalization}\label{Sec:BandShifts}
\begin{figure}[t]
\includegraphics[width=0.7\columnwidth]{./n_k.pdf}
\caption{\label{Fig:n_k} (color online)
The momentum dependence of the number operator
$n_\gamma(k) = \frac{1}{2}\sum_\sigma \langle c^\dagger_{{\bf k},\gamma,\sigma} c_{{\bf k},\gamma,\sigma}\rangle$ for each
band. Results are shown for the non-interacting case $U = 0$ (black dashed, $\square$),
$U/W = 0.1$ (blue solid, $\bigtriangleup$), $U/W = 0.8$ (red solid $\circ$), and
$U/W = 2$ (green solid $\diamond$) and at an inverse temperature of $\beta = 19.6/W$.}
\end{figure}
It is now well known that {\em ab initio} band structure calculations based on
density functional theory (DFT) do not describe
the electronic structure of the iron based superconductors as measured in ARPES
experiments. (For a recent review, see Ref. \onlinecite{vanRoekeghem}.) Generally speaking,
the calculated band structure usually needs to be rescaled by an overall factor, which
is attributed to reduction in bandwidth driven by electronic correlations. In addition,
the size of the Fermi surfaces is often overestimated by DFT in comparison to
measurements. A prominent example of this is LiFeAs,\cite{Chi} where the inner most hole pocket
realized in nature is
substantially smaller than the one predicted by DFT \cite{Klaus,Klaus2}. In order to correct this,
the electron- and hole-bands need to be shifted apart,\cite{vanRoekeghem} which
requires a momentum-dependent self-energy correction.
We examine this issue within our model in Fig. \ref{Fig:n_k}, which plots the
expectation value of the orbitally-resolved number operator in momentum space
$n^{\phantom{\dagger}}_\gamma(k) = \frac{1}{2}\sum_\sigma \langle c^\dagger_{{\bf k},\gamma,\sigma} c^{\phantom{\dagger}}_{{\bf k},\gamma,\sigma}\rangle$
for various values of the interaction strength.
In the non-interacting limit, and in a single-band case, this quantity is equal to the
Fermi-Dirac distribution and the location of the leading edge corresponds to $k_\mathrm{F}$.
In a multi-band system the mixing of the orbital character complicates this
picture; however, in our model the leading edge still corresponds to $k_\mathrm{F}$
due to the weak hybridization between orbitals. In the weak coupling case ($U/W = 0.1$)
we observe a small shift in the position of the leading edge. Within error bars, the
curve $n_1(k)$ and $n_2(k)$ shift to slightly larger momenta while
$n_3(k)$ shifts towards smaller momenta. This indicates that the size of the
Fermi surfaces are increasing and the electron-like and hole-like bands are
shifted towards one another by the interactions. This trend continues as $U/W$ is
increased to $0.8$; however, in this case the electron-like band is significantly
smeared out due to the formation of the OSMP.
We note that the direction of the band shifts is reversed from what is generally
required for the two-dimensional iron-based superconductors, where the calculated
hole-like Fermi surfaces generally need to be shrunk relative to the electron-like
Fermi surfaces. We attribute this to differences in the underlying tight-binding model
and differences in dimensionality. In this light, it would be interesting to compare the ARPES
observed band structures in the quasi-one-dimensional pnictides against the predictions
of our model and DFT calculations.\cite{Patel} Nevertheless, our results do show that non-local
correlations arising from a local interaction
can produce relative shifts of the electron-like and hole-like bands
in a multi-orbital system.
\subsection{Spectral Properties}\label{Sec:Spectral}
\subsubsection{Intermediate Coupling $U/W = 0.8$}
We now examine the spectral properties of the model, beginning with
the OSMP. Figure \ref{Fig:dos}a shows the temperature evolution of the total
density of states (DOS) at $U/W = 0.8$, which is obtained from the trace of the
orbital-resolved spectral function $N(\omega) = \sum_{k,\gamma}
-\frac{1}{\pi}{\mathrm{Im}}\hat{G}_{\gamma\gamma}(k,\omega+i\delta)$. In the
non-interacting limit (the long-dashed (blue) curve), the DOS has a double peak
structure, where the lower (upper) peak corresponds to the bands derived from
orbitals one and two (orbital three). The overall structure of the DOS in the
interacting case is similar at high temperatures, but some spectral weight is
transferred to a broad incoherent tail extending to lower energies. As the temperature is
decreased, the peak on the occupied side shifts towards the Fermi level and
sharpens. At the same time, a small amount of spectral weight is
transferred from the vicinity of the Fermi level into this peak. The appearance
of this apparent ``pseudogap" is a direct consequence of the OSMP
forming on orbital three, which is easily confirmed by examining the
orbital-resolved DOS $N_\gamma(\omega) = -\frac{1}{\pi}\sum_k
\mathrm{Im}\hat{G}_{\gamma,\gamma}(k,\omega)$ shown in Fig. \ref{Fig:dos}b.
As can be clearly seen, orbitals one and two have a finite DOS at $\omega = 0$,
while orbital three is fully gapped at low-temperature.
\begin{figure}[t]
\includegraphics[width=0.75\columnwidth]{./dos.pdf}
\caption{\label{Fig:dos} (color online) (a) The density of states at
different temperatures. (b) The orbitally-resolved density of states for each
orbital at an inverse temperature $\beta=19.6/W$. (c) The density of states at
the Fermi surface of the orbital 3 as a function of inverse temperatures
$\beta$. The Coulomb interaction strength is $U/W=0.8$ in all three graphs.
}
\end{figure}
We also begin
to see the formation of an additional peak near the Fermi level
at the lowest temperature we examined ($\beta = 19.6/W$).
This feature is more clearly seen in the orbital-resolved DOS (Fig. \ref{Fig:dos}b),
where it is found to originate from the itinerant orbitals.
This peak is due to a hybridization between the itinerant
and localized orbitals, which is observable in the $k$-resolved spectral functions
(see Fig. \ref{Fig:spectral}).
The relevant temperature scale for the formation of the OSMP can be estimated
by tracking $N_{3}(0)$ as a function of
temperature, as shown in Fig. \ref{Fig:dos}c. Here, a continuous suppression of
$N_3(0)$ is observed, with the value reaching zero at $\beta \approx 20/W$. The
rate at which $N_3(0)$ decreases also undergoes a distinct change at $\beta
\approx 10/W$, which coincides with the temperature at which the spectral
weight for this orbital at $k_\mathrm{F}$ is largest (see Fig.
\ref{Fig:Gbeta}d). We interpret this to mean that the Mott gap on orbital three
begins to form at $\beta W \approx 10$ (on the $L = 24$ site lattice), growing
continuously from zero as the temperature is lowered. In this case, the finite
spectral weight between $\beta W = 10$ -- $20$ is due to thermal
broadening across this gap. Since we have observed similar behavior on
smaller clusters with DQMC and at zero temperature using DMRG, we
believe that the transition to the OSMP will survive
in the thermodynamic limit, however, the gap magnitude has some finite size
dependence.
\begin{figure*}[t]
\includegraphics[width=0.8\textwidth]{spectral.pdf}
\caption{\label{Fig:spectral} (color online) (a) The spectral function
for $U/W=0.8$. (b), (c), and (e) are the orbital 1, 2, and 3 parts of the
spectral function in (a), respectively. (e) The spectral function for $U/W =2$.
(f), (g), and (h) are the orbital 1, 2, and 3 parts of the spectral function in
(e), respectively. The dash white line labels the Fermi surface. The inverse
temperature is set as $\beta=19.6/W$. Results where obtained with Maximum Entropy DQMC.
}
\end{figure*}
The extended length of our 1D cluster grants us access to the momentum
dependence of the spectral function, which is shown in Fig. \ref{Fig:spectral}.
The top row of Fig. \ref{Fig:spectral} shows the
results in the OSMP with $U/W = 0.8$ and $\beta = 19.6/W$, which is the same
parameter set used in Fig. \ref{Fig:dos}.
The total spectral function $A(k,\omega) =-\frac{1}{\pi}\mathrm{Tr}\left[\mathrm{Im}\hat{G}(k,\omega)\right]$
is shown in Fig. \ref{Fig:spectral}a and the orbital-resolved components
$A_\gamma(k,\omega)=-\frac{1}{\pi}\mathrm{Im}\hat{G}_{\gamma\gamma}(k,\omega)$
are shown in Figs. \ref{Fig:spectral}b-d, as indicated.
The lower row of Fig. \ref{Fig:spectral} shows similar results obtained for
$U/W = 2$ and $L = 8$. (In this case a smaller cluster is sufficient
due to the non-dispersing nature of the band dispersions.)
The results in the OSMP with $U/W = 0.8$ reveal localized and itinerant
characteristics that are consistent with the spectral weight analysis
presented earlier. The itinerant orbitals primarily contribute to
dispersing bands that track through the $E_\mathrm{F}$ ($\omega = 0$), while
orbital three has split into two relatively dispersionless upper and lower Hubbard bands
above and below $E_\mathrm{F}$. At first glance, these Hubbard bands appear to be
sharper than the corresponding Hubbard bands in the single-band Hubbard model; however,
an examination of the DOS (Fig. \ref{Fig:dos}b) reveals that they are spread out over
an energy interval that is larger than the non-interacting bandwidth of the
top most band ($W_3 \sim 0.3W \sim 0.735$ eV).
In addition to the formation of the Hubbard bands for orbital 3, we also observe two additional
effects. The first is an expected narrowing of the bandwidth of the
itinerant bands. For this parameter
set we obtain $W_1 \sim 1.7$ and $W_2 \sim 1.65$ eV for orbitals one and two,
respectively, which should be compared to the non-interacting values of 1.88 and 1.97 eV.
The second is the aforementioned hybridization and level repulsion between the itinerant and
localized orbitals. This is manifest in the spectral
function as a slight ``buckling" of orbital three's upper Hubbard band near $k = 0$,
and the tracking orbital one's spectral weight along $E_\mathrm{F}$ near $k = \pm\pi/2a$.
It is this trailing intensity that forms the peak observed in the DOS
just above the Fermi level at low temperatures.
\subsubsection{Strong Coupling $U/W = 2$}
The spectral properties of the model are very different when
the Hubbard interaction is increased to $U/W = 2$.
In this case, the total spectral function (Fig. \ref{Fig:spectral}e) and its
orbitally-resolved components (Fig. \ref{Fig:spectral}f-\ref{Fig:spectral}h)
all split into relatively flat Hubbard-like bands above and below $E_\mathrm{F}$.
(In the case of orbital three, the lower band below $E_\mathrm{F}$ has been pushed outside of
the energy range shown in the figure.) For this value of the interaction strength there is
no spectral weight at the Fermi level, and the system is insulating even though
orbitals one and two have on average 1.55 and 1.44 electrons/orbital,
respectively.
(These values are obtained both from the measured equal time orbital occupancies, and
from integrating the total spectral weight above and below $E_\mathrm{F}$.)
The imaginary axis spectral weight analysis (Fig. \ref{Fig:Gbeta}c) and the spectral function
analysis (Fig. \ref{Fig:spectral}) both indicate that for
$U/W = 2$ the model is an insulator. The origin of this behavior
is the combined action of the Hund's coupling and the onset of an orbital ordering
of the itinerant orbitals. All indications show that orbital three
has already undergone an orbital selective Mott phase transition (OSMT)
when $U/W = 2$. This has the effect of localizing one electron per site within this
subset of orbitals while leaving
three additional electrons to be distributed among the remaining two itinerant orbitals.
A sizable Hund's coupling will decouple the
individual orbitals when the crystal field splittings are smaller
than the bandwidth of the material.\cite{deMediciPRL2014} This is precisely the situation at hand,
and thus the remaining nominally itinerant orbitals are decoupled from the localized orbital
by the large $J = U/4$. This results in an effective nearly-degenerate two-band system
with (nearly) three-quarters filling. This is special case for the two-orbital
Hubbard model, which is prone to orbital ordering
in one and two-dimensions.\cite{HeldEPJB,ChanPRB2009,KuboPRB2002}
The situation is sketched in Fig. \ref{Fig:OrbitalOrdering}.
Assuming ferromagnetic nearest neighbor correlations for orbital three,
we have a low-energy ground state configuration as shown in the left side of \ref{Fig:OrbitalOrdering}a.
Here, orbitals one and two adopt alternating double occupations
in order to maximize their delocalization energy through virtual hopping
processes. This results in near-neighbor orbital correlations.
Subsequent charge fluctuations such as the one shown
in the right side of the Fig. \ref{Fig:OrbitalOrdering}a cost a potential energy
$PE \sim U^\prime - J = W/2$. This is compensated for by a kinetic energy gain
$KE \sim 4t_{11} \sim 4W/4.9$. The ratio between these competing energy
scales is $\sim 5/8$, suggesting that charge fluctuations are strongly
suppressed by the strong electronic correlations in this subsystem.
Note that the situation is worse for antiferromagnetic nearest neighbor correlations
in orbital three. The energy cost in this case
increases to $\sim U^\prime$, as shown in Fig. \ref{Fig:OrbitalOrdering}b.
Thus both ferro- and antiferromagnetic correlations in orbital three will
suppress charge fluctuations and promote orbital ordering.
Since the type of magnetic correlations does not matter, such orbital ordering tendencies
can be expected in the paramagnetic phases, provided the localized moments
have formed in orbital three. This picture is then consistent with
insulating behavior (and short-range orbital ordering tendencies, see below)
at high temperatures, where no magnetic correlations are observed.
\begin{figure}
\includegraphics[width=0.9\columnwidth]{./Orbital_Ordering.pdf}
\caption{\label{Fig:OrbitalOrdering}
A cartoon sketch of the relevant charge fluctuation processes leading to
the insulating state when $U/W = 2$ assuming (a) ferromagnetic and (b)
antiferromagnetic nearest neighbor correlations within the orbital that has
undergone the orbital selective Mott transition (orbital three).
}
\end{figure}
We verify this picture explicitly in Fig. \ref{Fig:Orbital_correlation}, which plots the
equal-time orbital correlation function $\langle \hat{\tau}_{i+d}\hat{\tau}_i\rangle$,
with $\hat{\tau}_{i} = (\hat{n}_{i,2}-\hat{n}_{i,1})$. Here, results are shown for finite
temperature DQMC calculations (Fig. \ref{Fig:Orbital_correlation}a) and zero
temperature DMRG calculations (Fig. \ref{Fig:Orbital_correlation}b) and with $U/W=2$ in
both cases. The ``long-range" (with respect to the cluster size)
anti-ferro-orbital correlation is clear in the
zero temperature results obtained on $L = 8$ and $L=16$ chains.
At finite temperatures ($\beta = 19.6/W$) we find that the
orbital correlations are suppressed at long distances, but local
anti-ferro-orbital correlation remains on shorter length scales. These
combined results demonstrate the
presence of short-range orbital correlations at higher temperatures, which
grow in length as the temperature is decreased.
The corresponding orbitally resolved DOS are plotted in Fig. \ref{Fig:Ueq2}
for both cases. Both methods predict that the system is insulating, with a charge
gap width on orbitals one and two of about 0.5 eV. The presence of a gap at
finite temperature also confirms that the short range orbital correlations
are sufficient to open a gap in the spectral function. Finally, we stress these
results will survive in the thermodynamic limit $L \rightarrow \infty$.
This is confirmed in the inset in Fig. \ref{Fig:Ueq2}b, which shows
the evolution of the $T = 0$ gap $\Delta$ as a function of $L$,
as obtained from DMRG. Here, the gap size decreases with
increasing chain lengths, however, it saturates to 0.2 eV for an infinite length
chain.
\begin{figure}
\includegraphics[width=0.8\columnwidth]{./Orbital_Correlation.pdf}
\caption{\label{Fig:Orbital_correlation}
Results for the orbital correlation function
for the system in the strong coupling case $U/W = 2$. Results are obtained
at (a) finite temperature using DQMC and (b) $T = 0$ ($\beta = \infty$)
using DMRG.
In both cases, results are shown on $L = 8$ (red dots) and $L = 16$
(blue triangles) chains. The DQMC results were obtained on a chain with
periodic boundary conditions. The DMRG results were obtained on a chain with
open boundary conditions.
}
\end{figure}
\section{Discussion and Summary}\label{Sec:Discussion}
\begin{figure}
\includegraphics[width=0.8\columnwidth]{./DOS_DQMCvsDMRG.pdf}
\caption{\label{Fig:Ueq2}
Results for the orbitally-resolved density of states for each orbital
obtained for $U/W = 2$ and on $L = 8$ site
chains. Panel (a) shows DQMC results at $\beta = 19.6/W$
and the inset zooms in to energy around Fermi
surface. Panel (b) shows DMRG results for the same conditons but at
zero temperature ($\beta = \infty$). The inset plots a finite size
scaling analysis of the charge gap obtained within DMRG (see text).
The dash line in both panels indicates the Fermi energy.
}
\end{figure}
We have performed a momentum-resolved study of a multi-orbital model defined on
extended 1D chains using non-perturbative DQMC and DMRG.
This has allowed us to compute the
several properties of an OSMP in a momentum resolved manner without resorting to
approximate methods. We find that several properties do indeed exhibit significant momentum
dependencies, not be captured by local approximations
introduced by DMFT; however, the 1D case we have considered
represents the worst case for DMFT. In that sense our results
complement existing DMFT efforts by providing analysis in a region where the method is expected to
perform badly.
Our results establish the hierarchy of charge and magnetic orderings in this model.
At low temperatures, our DMRG calculations (as well as those in Ref. \onlinecite{RinconPRL2014})
demonstrate that orbital three is ferromagnetically ordered at $T = 0$. Contrary to this,
our finite temperature DQMC calculations find no indications of any magnetic order
for $\beta < 19.6/W$; the magnetic structure factor $S(q)$ is completely featureless
as a function of $q$ at these temperatures. Despite this, our finite $T$ calculations find an orbital-selective
Mott phase, as well as a fully insulating phase arising due to short-range orbital ordering, depending on
the strength of the Hubbard interaction $U$.
We therefore conclude that the charge ordering occurs before any magnetic ordering in this model.
The results shown in Fig. \ref{Fig:n_beta}d and \ref{Fig:Gbeta}d
show that orbital three in our model, which has the narrowest band width, undergoes a transition to
a Mott phase at $\beta W \sim 10-15$. This in combination with the lack of magnetic signal
means that OSMP in this parameter regime is a true Mott phase as opposed to a Slater insulator where
the insulating behavior is driven by magnetism.
Our results also demonstrate that it is insufficient to identify an OSMP using
the orbital occupations only in some instances. One should be particularly
careful in regions of parameter space where the itinerant bands have average occupations
close to special cases known for one and two-orbital Hubbard models. In our case,
the average fillings of the itinerant orbitals are $\langle n_1\rangle \sim 1.53$
and $\langle n_2 \rangle \sim 1.47$, values very close to the special case of $3/4$ filling in
a degenerate two-band Hubbard model. At zero temperature, our DMRG results obtain
fillings of 1.5 for each orbital.
Finally, we discuss our results in the context of recent experimental work.
ARPES results for AFe$_2$As$_2$ have found evidence that the OSMP in these
materials disappears as the temperature is lowered.\cite{YiPRL2013} This
behavior was explained using a slave-boson approach and attributed to an increase in
entropy associated with the OSMP. Our results do not show this behavior, and the
OSMP is found at low temperature as one might naively expect.
This difference may be
related to the differences in the dimensionality (one vs. two) or number of orbitals
(three vs. five) between the models or the differences between our non-perturbative
approach and other mean-field methods. However, one would expect the entropy to be
more important in one dimension. This highlights the need for
continued application of non-perturbative methods to tractable multi-orbital Hubbard
models.
{\em Acknowledgements} ---
The authors thank G. Liu for useful discussions.
S. L., Y. W., and S. J. are supported by the
University of Tennessee's Science Alliance Joint Directed Research and
Development (JDRD) program, a collaboration with Oak Ridge National Laboratory.
N. K. and E. D. were supported by the National Science Foundation (NSF) under
Grant No. DMR-1404375.
Y. T. and T. A. M. acknowledge support by the Laboratory Directed Research
and Development Program of Oak Ridge National Laboratory, managed by
UT-Battelle, LLC, for the U.S. Department of Energy.
Part of this work was conducted at the Center for Nanophase Materials Sciences,
sponsored by the Scientic User Facilities Division (SUFD), BES, DOE,
under contract with UT-Battelle. A.N. and G.A. acknowledge support by
the Early Career Research program, SUFD, BES, DOE. CPU time was
provided in part by resources supported by the University of Tennessee and Oak
Ridge National Laboratory Joint Institute for Computational Sciences
(http://www.jics.utk.edu).
|
\section{Introduction\label{sec:intro}}
Observations beginning in the 1990s established galactic winds as ubiquitous phenomena associated with star-forming galaxies \citep{1993ASSL..188..455H,1995PASA...12..190B,1997PASP..109.1298D,2000ApJS..129..493H}.
These observations focused on optical emission lines images and spectroscopy \citep{1993ASSL..188..455H}.
Optical imagery established the physical morphology of galactic winds and spectroscopy provided the kinematics and warm plasma diagnostics.
While emission traced the interaction of the warm ISM with the hot wind, absorption lines probed the interaction between warm and cold gas and the hot wind \citep{2000ApJS..129..493H}.
X-ray emission, first observed in M82 \citep{1984ApJ...286..144W}, would also become important for identifying galactic outflows and measuring wind energetics \citep{1988ApJ...330..672F,1990ApJ...355..442F,1993ASSL..188..455H,1995ApJ...448...98H}. While some studies of galactic winds focused on X-ray emission \citep{StricklandStevens,2009ApJ...697.2030S}, \citet{1995PASA...12..190B} predicted that multi-band observations of galactic winds would become standard in characterizing galactic winds, and \citet{2005ARA&A..43..769V} have shown that subsequent multi-band studies are important in characterizing the galactic wind.
More recent observations \citep{2012ApJ...760..127M,2014A&A...568A..14A,2014ApJ...794..156R,2015ApJ...811..149C,2015ApJ...809..147H,2016ApJ...822....9H} show that galactic winds are ubiquitous for star forming galaxies.
Galactic winds are detected in 45\% \citep{2012ApJ...760..127M}, 74\% \citep{2015ApJ...811..149C}, 66\% (89\% for face on, 45\% for edge on galaxies) \citep{2014ApJ...794..156R}, and 90\% \citep{2015ApJ...809..147H,2016ApJ...822....9H} of star forming galaxies surveyed.
Outflow kinematics are typically measured using UV absorption lines such as: Mg II and Fe II \citep{2014ApJ...794..156R}, Si II, Si III, Si IV and O I \citep{2015ApJ...811..149C,2016MNRAS.457.1257H,2015ApJ...809..147H,2016ApJ...822....9H}, Na D \citep{2000ApJS..129..493H,2005ApJ...621..227M}, H$\alpha$ and O III \citep{2016A&A...588A..41C}, and C II and N II \citep{2015ApJ...809..147H,2016ApJ...822....9H}.
\citet{2000ApJS..129..493H} found that starburst galaxies whose Na D absorption line is dominated by the ISM, typically exhibited outflow velocities of $> 100$ \kms, with maximum velocities ranging from $300-700$ \kms.
They were able to map outflow gas up to 10 kpc from the galactic center.
They concluded that dense clouds in the ISM were being disrupted by the galactic wind, and that the ablated gas was being accelerated up to the terminal wind velocity.
\citet{2005ApJ...621..227M} investigated the relationship between outflow velocities, as measured by the Na D lines, and the SFR.
She found that the maximum wind velocity correlates as SFR$^{.35}$, and that stellar luminosity suffices to accelerate cool outflows to the terminal velocity.
Martin noted that the covering fraction of the cold gas is not complete, which indicates that it is not a continuous fluid but is broken into clouds or shells.
\citet{2014ApJ...794..156R} extended previous work using Mg II and Fe II absorption lines to find that outflows are detected for all $M_\star$, SFR and SFR density (\sfrd) studied.
Interestingly they found no evidence of a minimum threshold for \sfrd.
This indicates that galactic winds can still form in galaxies with extremely low \sfrd.
Although outflows were detected for all parameter ranges, a correlation was only found between outflow velocity and $M_\star$.
These findings are both consistent with and conflict with previous work \citep{2009ApJ...692..187W,2010AJ....140..445C,2011ApJ...730....5H,2012ApJ...760..127M}.
Conversely \citet{2015ApJ...811..149C} found correlations between $M_\star$ and SFR, but not with \sfrd, using Si II absorption lines.
They found a weak correlation between SFR and maximum velocity, but a slightly stronger correlation between SFR and the velocity as measured by the line center.
In agreement with \citet{2014ApJ...794..156R}, \citet{2015ApJ...811..149C} found that there is no minimum \sfrd at which outflows form.
\citet{2016A&A...588A..41C} studied a large sample of galaxies and found that the outflow velocity correlates with both the SFR and the specific SFR, but only for a SFR $> 1$~\msyr.
\citet{2015ApJ...809..147H} and \citet{2016ApJ...822....9H} found that the outflow velocity correlates strongly with SFR and \sfrd but weakly with galactic stellar mass.
They also also noted that at sufficiently high \sfrd the correlation flattens out.
In this paper we resolve these differences and explain why some surveys of starburst galaxies have found correlations between the outflow velocity of starburst-driven winds and both SFR and \sfrd, while others have not.
We also address why even among studies of star forming galaxies that have found a correlation, there is no agreement on the measured scaling relation.
We use 3D hydro models of nuclear starbursts to study how the outflow velocities of galactic winds scale with SFR and \sfrd.
In Section \ref{sec:setup} we explain our setup and model parameters.
In Section \ref{sec:lines} we describe how we generate synthetic absorption profiles for Si ions to use as tracers of outflow velocity at different gas temperatures.
In Section \ref{sec:res} we describe how the outflow velocity correlates with SFR and \sfrd, followed with a discussion in Section \ref{sec:dis} of how our results relate to observational results.
\section{Setup}\label{sec:setup}
Our models use the setup of \citet[][hereafter \citetalias{Tanner1}]{Tanner1} with a few modifications described below. As in \citetalias{Tanner1} we simulate a nuclear starburst inside a box 1000 pc on a side. The stellar gravitational potential and associated parameters are set to correspond to an M82 sized galaxy. All models are run for 1.5 Myr.
\subsection{Static Mesh Refinement}\label{sec:setup:smr}
In \citetalias{Tanner1} we employed a single grid, but in this paper we use the static mesh refinement (SMR) available in Athena \citep{Stone-Athena}.
We use two levels of refinement covering the starburst and wind region directly above the starburst extending to the +z boundary, as shown in Figure \ref{fig:test:smr}.
The base grid is divided into $64^3$ cells, the first level divided into $64\times64\times112$ cells, and the second level divided into $128\times128\times160$ cells.
This gives spatial resolution of $15.6$ pc on the base grid and $7.8$ and $3.9$ pc on each respective level. The resolution of the top level corresponds to our medium resolution models from \citetalias{Tanner1}.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.5\textwidth]{Figure1.pdf}
\protect\caption[SMR Levels]{XZ plane slice of initial gas density (particles cm$^{-3}$) scaled logarithmically. White lines indicate SMR levels of refinement.}
\label{fig:test:smr}
\end{figure}
\subsection{Analytic Wind Velocity}\label{sec:setup:va}
As in \citetalias{Tanner1} we model the starburst as a spheroidal region where we inject thermal energy ($\dot{E}$) and mass ($\dot{M}$) at each time step.
The analytical wind velocity depends on both $\dot{E}$ and $\dot{M}$ as,
\begin{equation}\label{eq:vA}
v_A = \sqrt{2\frac{\dot{E}}{\dot{M}}}.
\end{equation}
In this paper and in \citetalias{Tanner1} we refer to $v_A$ as the analytic wind velocity, whereas \citet{2005ARA&A..43..769V,2009ApJ...697.2030S} refer to it as the terminal wind velocity.
Both $\dot{E}$ and $\dot{M}$ are proportional to the total contribution from supernovae (SN) and stellar winds (SW) in the starburst.
The total thermal energy injected into the ISM depends on the thermalization efficiency ($\epsilon$), which gives the fraction of SN and SW energy ($\dot{E}_{SN+SW}$) retained by the ISM as shown in Equation \ref{eq:Edot}.
\begin{equation}\label{eq:Edot}
\dot{E} = \epsilon\dot{E}_{SN+SW}
\end{equation}
The total mass added to the ISM by the starburst includes contributions from SN and SW ($\dot{M}_{SN+SW}$) and cold gas ($\dot{M}_{cold}$) remaining from star formation that cannot be resolved by our simulations.
The additional mass from cold gas can be factored in as a scaling factor ($\beta$) for $\dot{M}_{SN+SW}$ and is referred to as the mass loading factor \citep{2005ARA&A..43..769V,2009ApJ...697.2030S}, as shown in Equation \ref{eq:Mdot}.
\begin{equation}\label{eq:Mdot}
\dot{M} = \dot{M}_{SN+SW} + \dot{M}_{cold} = \beta\dot{M}_{SN+SW}
\end{equation}
We calculate $\dot{E}_{SN+SW}$ and $\dot{M}_{SN+SW}$ using Starburst99 models \citep{Leitherer1999}. We assume continuous star formation (CSF) with a \citet{2001MNRAS.322..231K} IMF, and Geneva stellar evolutionary tracks with solar metallicity.
Both $\dot{E}_{SN+SW}$ and $\dot{M}_{SN+SW}$ scale with SFR \citep{2005ARA&A..43..769V} as,
\begin{eqnarray}
\dot{E}_{SN+SW} &=& 4.324e41~(\text{erg}~\text{s}^{-1})~\text{SFR}\label{eq:ESNSW}\\
\dot{M}_{SN+SW} &=& 0.1902~(\text{M}_\Sun~\text{yr}^{-1})~\text{SFR}.\label{eq:MSNSW}
\end{eqnarray}
With the above assumptions $\dot{E}_{SN+SW}$ and $\dot{M}_{SN+SW}$ are constant after $\sim5$ Myr.
Thus we assume the starburst begins CSF 5 Myr before the start of our simulations, and we run our simulations for 1.5 Myr by which time a steady state wind has formed.
Inserting Equations \ref{eq:Edot}, \ref{eq:Mdot}, \ref{eq:ESNSW}, and \ref{eq:MSNSW} into Equation \ref{eq:vA} and simplifying we get,
\begin{equation}\label{eq:vAsimp}
v_A = (1894~\text{km}~\text{s}^{-1})\sqrt{2\frac{\epsilon}{\beta}}.
\end{equation}
Thus $v_A$ does not depend on the SFR, but only $\epsilon$ and $\beta$.
The actual velocity of the hot gas will be some fraction of $v_A$, and will be discussed in Section \ref{sec:res}.
We use $v_A$ as a free parameter that we set in our models.
\subsection{Model Series}\label{sec:setup:mod}
We run two series of models, which we label series S and R, to test how the measured velocity of a galactic wind scales with the SFR and the SFR density (\sfrd) respectively.
For all models we set $\epsilon=1.0$, $v_A$ to 1000, 1500 or 2000 \kms, and use Equation \ref{eq:vAsimp} to calculate $\beta$.
The S series varies the SFR from 1 to 100 \msyr in steps of $0.1$ dex.
For the S series we range over all SFRs with $v_A$ set to 1000, 1500 or 2000 \kms ($\beta = 7.1745$, 3.1887, 1.7936 respectively) for a total of 63 models.
As in \citetalias{Tanner1}, our S series uses a starburst radius of 150 pc.
Model numbers denote first the $v_A$ then the SFR.
Thus model number S\_15\_79 has $v_A=1500$ \kms and SFR $=7.9$ \msyr.
The R series varies the radius (r) of the starburst from 50 to 500 pc in steps of $0.1$ dex.
Each model in the R series has a fixed SFR of 10, 50 or 100 \msyr for a total of 33 models.
The SFR density for this series is calculated using, \sfrd $= \text{SFR}/\pi r^2$.
All R series models use $v_A=2000$ \kms ($\beta = 1.7936$).
Model numbers denote first the SFR then the starburst radius.
Thus model number R\_50\_79 has SFR 50 \msyr and starburst radius $79$ pc.
\section{Absorption Profiles}\label{sec:lines}
To probe the outflow velocities of different temperature regimes we generate synthetic absorption profiles for various silicon ions.
We generate our absorption profiles by calculating the optical depth in each cell for a set of velocity channels directly above the starburst ($z>100$ pc) ranging from -2800 \kms to 200 \kms.
We then integrate along a column of cells in the +z direction to get a total optical depth for each velocity channel, and then average each channel over all columns.
We use a channel resolution of $\Delta v_{ch} = 0.25$ \kms.
The absorption coefficient for a single velocity channel ($v_{ch}$) is,
\begin{equation}\label{eq:lines:abs}
\kappa(v_{ch}) = N(v_{ch})a(v_{ch})
\end{equation}
where $N(v_{ch})$ is the column density and $a(v_{ch})$ is the absorption per atom.
The column density is calculated from the cell density and the ionization fraction \citep{1998A&AS..133..403M}.
For simplicity, and to compare with observations \citep{2015ApJ...811..149C,2016ApJ...822....9H}, we choose to use silicon ions for our analysis.
The ionization fractions for Si I-XIII in collisional ionization equilibrium and at various temperatures are shown in Figure \ref{fig:lines:ion}.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.5\textwidth]{Figure2.pdf}
\protect\caption[Ionization Fraction]{Ionization fractions for Silicon ions \citep{1998A&AS..133..403M} in collisional equilibrium at various temperatures.\label{fig:lines:ion}}
\end{figure}
Contributions from Doppler broadening and spontaneous radiative transitions $a(v_{ch})$ is given as,
\begin{equation}\label{eq:lines:abspat}
a(v_{ch}) = \frac{\pi e^2}{m_e c}\frac{1}{\sqrt{\pi}}\frac{1}{\Delta \nu_{1/2}}f H(v_{ch}).
\end{equation}
Here $m_e$ is the mass of an electron, $c$ is the speed of light, $\Delta \nu_{1/2}$ is the half width half maximum (HWHM) of the Gaussian component, $f$ is the oscillator strength, and $H(v_{ch})$ is a Voigt profile.
The Gaussian HWHM is calculated \citep{kwok2007physics} using,
\begin{equation}
\Delta \nu_{1/2} = \frac{2}{c}\sqrt{\frac{2k_BT}{m}\ln(2)}\nu_0
\end{equation}
with $k_B$ Boltzmann's constant, $T$ the gas temperature, $m$ the atomic mass of the ion, and $\nu_0$ the frequency of the line center from the NIST Atomic Spectra database \citep{NIST_ASD}.
We calculate the Voigt profile ($H(v_{ch})$) using Matlab code\footnote{\href{http://www.mathworks.com/matlabcentral/fileexchange/45058-deconvolution-mordenite-zeolite}{http://www.mathworks.com/matlabcentral/fileexchange/45058-deconvolution-mordenite-zeolite}} written by Dr. Nikolay Cherkasov which employs the method of \citet{Schreier20111010}.
We calculate a normalized Voigt profile for each cell and then Doppler shift the profile using the $z$ velocity of the gas in the cell.
The Doppler shift for each cell is calculated with respect to the systemic velocity of the galaxy, with negative velocities toward the observer and positive velocities away from the observer.
Using Equation \ref{eq:lines:abs} we calculate the optical depth of each cell
\begin{equation}
\tau_i(v_{ch}) = \kappa_i(v_{ch}) dz.
\end{equation}
The optical depth for each velocity channel is then summed along a column in the z direction. The intensity for each velocity channel is calculated using,
\begin{equation}
I(v_{ch}) = I_0(v_{ch})\textrm{e}^{-\tau(v_{ch})}.
\end{equation}
The resulting profile is then averaged over all columns directly over the starburst and then re-normalized so that we can compare the profiles of different ions.
Figure \ref{fig:lines:lineex} gives an example of a synthetic absorption profile for the Si IV line.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\textwidth]{Figure3.pdf}
\protect\caption{Synthetic absorption profile for a Si IV line. S\_20\_100 model with an analytic wind velocity of 2,000 \kms and a SFR of 10 \msyr. Vertical lines indicate $v_{cen}$ and $v_{90}$ velocities. \label{fig:lines:lineex}}
\end{figure}
Using the synthetic absorption lines, we measure outflow velocities of each ion in two ways.
$v_{cen}$ is the velocity at half of the full width at half maximum (FWHM).
This effectively measures the average outflow velocity for the temperature range of the ion.
$v_{90}$ is the velocity on the blueward side of the line where the absorption profile returns to 90\% of full intensity.
This measures the maximum velocity of the gas for a specific temperature range.
These velocity measures were adopted by \citep{2005ApJ...621..227M,2009ApJ...692..187W,2012ApJ...759...26E,2015ApJ...811..149C,2016MNRAS.457.1257H,2016ApJ...822....9H}, while others have used different, but comparable velocity measures \citep{2005ApJS..160..115R,2014ApJ...794..130B,2015ApJ...809..147H,2016A&A...588A..41C}.
We use these velocities to determine the relationship between the galactic wind and the properties of the starburst.
The vertical lines in Figure \ref{fig:lines:lineex} show $v_{cen}$ and $v_{90}$ for a synthetic absorption profile.
\section{Results}\label{sec:res}
In this section we investigate the relationships between $v_{cen}$ and $v_{90}$, and the analytic wind velocity ($v_A$ from Equation \ref{eq:vAsimp}), the SFR, and the SFR density ($\Sigma_{SFR}$). This allows us to determine how the outflow velocities of the multi-phase medium scale with different starburst properties.
\subsection{Outflow Velocities of Different Ions}\label{sec:res:ions}
In Figure \ref{fig:lines:Siex} we plot synthetic absorption lines for Si I, II, VII, and XIII for our S\_20\_1000 model.
These four lines probe gas temperature ranges corresponding to $<1e4$ K, $1e4-2.5e4$ K, $4.5e5-7e5$ K, and $2e6-1e7$ K respectively.
We note that hotter, highly ionized gas has a higher outflow velocity.
This is in accordance with our results in \citetalias{Tanner1}.
For the Si XIII line $v_{\rm cen} \approx 1700$ \kms, and $v_{90} \approx 1900$ \kms.
These values are slightly less than the $v_A = 2000$ \kms of the S\_20\_1000 model.
This difference is due to energy lost to cooling and kinetic energy transferred to the surrounding ISM.
The cold gas, traced by Si I, moves much slower with $v_{\rm cen} \approx 300$ \kms, and $v_{90} \approx 800$ \kms.
The $v_{\rm cen}$ and $v_{90}$ velocities for Si II are slightly higher than Si I, and the values for Si VII are midway between Si I and Si XIII.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.5\textwidth]{Figure4.pdf}
\protect\caption[Si Absorption Lines]{Synthetic absorption lines for Si I, II, VII, and XIII from my S\_20\_1000 model, which has a $v_A$ of 2000 \kms and a SFR of 100 \msyr. \label{fig:lines:Siex}}
\end{figure}
To further determine how the velocity of the gas changes with increasing ionization we plot in Figure \ref{fig:lines:SRions} (Subplots (a) and (b)) $v_{cen}$ and $v_{90}$ respectively for Si I-XIII from our S series models.
The plots include models with $v_A$ of $1000$, $1500$, and $2000$ \kms at SFR of 10, 50, and 100 \msyr.
We find three distinct velocity regimes corresponding to Si I-II, Si III-XI, and Si XII-XIII, which correspond to temperatures ranges $<2.5e4$ K, $2.5e4-2e6$ K, and $>2e6$ K, respectively.
The outflowing gas in the cold regime has much lower $v_{\rm cen}$ and $v_{90}$ than the warm regime, which in turn has lower velocities than the hot regime.
In the warm regime the velocity plateaus and does not increase for several ions.
For each ion, the measured velocity increases with increasing $v_A$, with a few notable exceptions.
Models with $v_A=1000$ \kms show little variation in velocities for different SFRs, and the increase in velocity for hotter gas is less pronounced and occurs at Si VIII instead of Si XII or Si XIII as with models with higher $v_A$.
The measured velocities for all ions from our model with a SFR 10 \msyr and $v_A=1500$ \kms are lower than the velocities from models with higher SFRs and $v_A=1500$ \kms.
Our models with $v_A=2000$ \kms show a similar trend.
In Figure \ref{fig:lines:SRions} (Subplots (c) and (d)) we plot the $v_{cen}$ and $v_{90}$ velocities respectively for Si I-XIII from our R series models. A similar three part grouping of velocities in cold, warm and hot regimes is evident.
For each ion there is a trend of increasing velocity for increasing \sfrd.
\begin{figure*}[h]
\centering
\includegraphics[width=1.0\textwidth]{Figure5.pdf}
\protect\caption{Outflow velocities of Si ions I-XIII for select S and R series models. Top, S series: Blue lines are for models with $v_A=1,000$ \kms, green for $v_A=1,500$, and red for $v_A=2,000$. Solid lines with 'x' indicate models with SFR of 10 \msyr, dashed lines with circles indicate a SFR of 50 \msyr, and dot dashed lines with pentagrams indicate models with SFR of 100 \msyr. (a) $v_{cen}$ velocity of all Si ions. (b) $v_{90}$ velocity of all Si ions. Bottom, R Series: Models have a SFR of 100 \msyr. Numbers in legend indicate $\log(\Sigma_{\rm SFR})$ \msyrpc of the models. (a) $v_{cen}$ velocity, (b) $v_{90}$ velocity.\label{fig:lines:SRions}}
\end{figure*}
\subsection{Scaling Relations}\label{sec:res:scale}
\subsubsection{Outflow Velocity vs. SFR}\label{sec:res:scale:S}
To see how the maximum and central velocities scale with SFR, in Figures \ref{fig:res:Sscale} and \ref{fig:res:Sscalecen} we plot $v_{90}$ and $v_{\rm cen}$ respectively, versus SFR for Si I, II, IV and XIII from all S series models.
The plots in Figure \ref{fig:res:Sscale} show that the maximum outflow velocity is correlated with the SFR, but only for some SFRs depending on the $v_A$ of the starburst.
The measured $v_{90}$ velocity resulting from cold or warm gas correlates with the SFR until the velocity reaches some fraction of $v_A$, and then the relationship flattens out.
This effect is most clearly seen in the plot for Si IV lines (Figure \ref{fig:res:Sscale} (c)).
In the case of Si IV, the relationship flattens out when the $v_{90}$ velocity is $\sim 80$\% of $v_A$.
A similar effect is observed for Si I and II $v_{90}$ velocities, with a turnover in the relationship at $\sim 50$\% and $\sim 65$\% of $v_A$ respectively.
The scaling relationship for Si XIII is flat, indicating that there is no correlation between SFR and outflow velocity for the hot gas.
This is in agreement with Equation \ref{eq:vAsimp}, which shows that the analytic velocity does not depend on the the SFR.
For hot gas, the $v_{90}$ velocity is roughly constant at $\sim 90$\% of $v_A$ for all SFRs.
\begin{figure*}[h]
\centering
\includegraphics[width=1.0\textwidth]{Figure6.pdf}
\protect\caption{The $v_{90}$ velocity vs SFR for all S series models for select Si ions: (a) Si I, (b) Si II, (c) Si IV, (d) Si XIII. Lines indicate linear fits for models with $v_A=2000$ \kms (yellow, dot-dash line), $v_A=1000$ \kms (blue, dashed line), and all models together (purple, solid line). \label{fig:res:Sscale}}
\end{figure*}
Figure \ref{fig:res:Sscalecen} shows that there are similar scaling relations between $v_{\rm cen}$ and SFR.
But the relationship flattens out at $\sim 20$\%, $\sim 30$\% and $\sim 50$\% of $v_A$ for Si I, II and IV respectively.
Interestingly, Si XIII now shows a correlation between SFR and the outflow velocity for low SFRs, with the relationship flattening out at $\sim 85$\% of $v_A$.
\begin{figure*}[h]
\centering
\includegraphics[width=1.0\textwidth]{Figure7.pdf}
\protect\caption{The $v_{\rm cen}$ velocity vs SFR for all S series models for select Si ions: (a) Si I, (b) Si II, (c) Si IV, (d) Si XIII. Lines indicate linear fits for models with $v_A=2000$ \kms (yellow, dot-dash line), $v_A=1000$ \kms (blue, dashed line), and all models together (purple, solid line). \label{fig:res:Sscalecen}}
\end{figure*}
We perform a fit to each set of models with the same $v_A$ and then fit all models combined.
The scaling relationship is of the form,
\begin{equation}
\log(v) = \delta\log(\text{SFR}) + \alpha
\end{equation}
For both $v_{90}$ and $v_{\rm cen}$ the scaling relation with SFR is steepest for models with a $v_A=2,000$ \kms and nearly flat for models with a $v_A=1,000$ \kms.
For example, using $v_{\rm cen}$ velocities from Si IV lines, $\delta=0.35\pm0.08$ for models with $v_A=2,000$ \kms, and $\delta=0.14\pm0.08$ for models with $v_A=1,000$ \kms.
For all data combined the fit falls between those two extremes with $\delta = 0.25\pm0.04$.
All uncertainties are reported at a 95\% confidence level.
The lower $\delta$ values are a result of including models with a SFR above the point where the relationship flattens out.
As can be seen in Figures \ref{fig:res:Sscale} and \ref{fig:res:Sscalecen}, the same scaling relation exists for all models below the turnover point independent of $v_A$.
We restrict our data set to models with SFRs below the turnover point and then fit the remaining data.
The resulting slopes and y intercepts for these fits for $v_{90}$ and $v_{\rm cen}$ are given Table \ref{tab:fits}.
As can be seen in Table \ref{tab:fits} for $v_{90}$ the slope is higher for low ionization and decreases with increasing ionization.
But for $v_{\rm cen}$ the slope is lower for low ionization and increases with increasing ionization.
Thus colder gas has a greater $\delta$ value than warmer gas for $v_{90}$, while the opposite is true for $v_{\rm cen}$.
\begin{table}
\begin{center}
\caption{\label{tab:fits}Fit data for outflow velocities vs. SFR from S series models. Only models below the turnover point are used.}
\begin{tabular}{lll} \toprule
Ion & Slope ($\delta$) & Intercept ($\alpha$) \\
\midrule
\multicolumn{3}{l}{$v_{90}$ Velocities}\\
\midrule
Si I & $0.38 \pm 0.08$ & $2.37 \pm 0.06$ \\
Si II & $0.28 \pm 0.06$ & $2.56 \pm 0.05$ \\
Si III & $0.26 \pm 0.03$ & $2.72 \pm 0.03$ \\
Si IV & $0.25 \pm 0.03$ & $2.76 \pm 0.03$ \\
\midrule
\multicolumn{3}{l}{$v_{\rm cen}$ Velocities}\\
\midrule
Si I & $0.28 \pm 0.03$ & $2.11 \pm 0.03$ \\
Si II & $0.32 \pm 0.05$ & $2.20 \pm 0.04$ \\
Si III & $0.34 \pm0.06$ & $2.38 \pm 0.05$ \\
Si IV & $0.37 \pm 0.08$ & $2.39 \pm 0.06$ \\
\bottomrule
\end{tabular}
\end{center}
\end{table}
\subsubsection{Outflow Velocity vs. SFR Density}\label{sec:res:scale:R}
Using our R series models we determine the relationship between the outflow velocity and the SFR density (\sfrd).
In Figure \ref{fig:res:Rscale} we plot the $v_{90}$ velocity versus \sfrd for all R series models.
For our R series we set $v_A = 2000$ \kms and tested three SFRs (10, 50 and 100 \msyr) while varying the size of the starburst to achieve a range of \sfrd.
As can be seen in Figure \ref{fig:res:Rscale}, much like with Figure \ref{fig:res:Sscale}, the outflow velocity is correlated with \sfrd, up to a point, and then the correlation flattens out.
For all ions, including Si XIII, the turnover point for $v_{90}$ is at $\log(\Sigma_{\rm SFR})\approx -3.5$ (M$_{\odot}$ yr$^{-1}$ pc$^{-2}$).
The turnover point for $v_{\rm cen}$ is slightly higher at $\log(\Sigma_{\rm SFR})\approx -3.0$ (M$_{\odot}$ yr$^{-1}$ pc$^{-2}$).
Because we set $v_A = 2000$ \kms for our R series these turnover points are higher than they would be for models with a lower $v_A$.
As shown in Section \ref{sec:res:scale:S}, models with a lower $v_A$ have a lower turnover point.
If we compare the velocities at the turnover point for each ion in our R series we find that they are similar to the velocities at the turnover point for our S series models with $v_A = 2000$ \kms.
\begin{figure*}[h]
\centering
\includegraphics[width=1.0\textwidth]{Figure8.pdf}
\protect\caption{The $v_{90}$ velocity vs \sfrd for all R series models for select Si ions: (a) Si I, (b) Si II, (c) Si IV, (d) Si XIII. \label{fig:res:Rscale}}
\end{figure*}
We perform a similar analysis as was done in Section \ref{sec:res:scale:S} on our R series and determine scaling relations for models below the turnover point.
The fit parameters are summed up in Table \ref{tab:fits:R}.
Similar to our results for our S series, the slopes calculated using $v_{90}$ velocities are steeper for colder gas and shallower for warmer gas.
But the opposite trend between cold and warm gas slopes calculated using $v_{\rm cen}$ velocities is not as clear, since Si I produces the steepest slope.
It is interesting to note that the SFR does not affect the outflow velocity, even below the turnover point. In Figure \ref{fig:res:Rscale} for each ion we plot three sets of models with different SFRs.
There is significant overlap among the three sets, and all have similar scaling relations, turnover points, and velocities above the turnover point.
\begin{table}
\begin{center}
\caption{\label{tab:fits:R}Fit data for outflow velocities vs. \sfrd from R series models. Only models below the turnover point are used.}
\begin{tabular}{lll} \toprule
Ion & Slope ($\delta$) & Intercept ($\alpha$) \\
\midrule
\multicolumn{3}{l}{$v_{90}$ Velocities}\\
\midrule
Si I & $0.32 \pm 0.07$ & $4.0 \pm 0.3$ \\
Si II & $0.28 \pm 0.05$ & $3.9 \pm 0.2$ \\
Si III & $0.25 \pm 0.05$ & $4.0 \pm 0.2$ \\
Si IV & $0.22 \pm 0.05$ & $3.9 \pm 0.2$ \\
\midrule
\multicolumn{3}{l}{$v_{\rm cen}$ Velocities}\\
\midrule
Si I & $0.35 \pm 0.08$ & $3.7 \pm 0.3$ \\
Si II & $0.26 \pm 0.04$ & $3.5 \pm 0.1$ \\
Si III & $0.32 \pm 0.07$ & $3.9 \pm 0.3$ \\
Si IV & $0.32 \pm 0.07$ & $4.0 \pm 0.3$ \\
\bottomrule
\end{tabular}
\end{center}
\end{table}
\section{Discussion}\label{sec:dis}
\subsection{Absorption Line Shape}\label{sec:dis:shape}
There are two important characteristics of our synthetic absorption profiles, the smoothness and the asymmetry of the lines.
In Figure \ref{fig:lines:Siex} the Si I line is not smooth, but has a large number of jagged spikes especially at lower velocities.
As explained in \citetalias{Tanner1}, the colder gas entrained in the wind is confined in filamentary structures with lower velocity than the surrounding hot gas.
As these filaments fragment in the wind due to Kelvin-Helmholtz instabilities they produce many dense cores \citep{CooperII}.
The contributions from these dense cores embedded in the filaments produce the jagged shape of the Si I line.
The smoothness of the absorption line, or lack thereof, reflects the distribution of vertical velocities of the dense cores and filaments.
We note that there is measurable absorption at positive velocities from dense cores that have been elevated above the galactic disk but whose vertical movement has stalled.
The Si II line is significantly smoother than the Si I line.
This indicate that the gas traced by the Si II line is significantly less clumpy than the colder gas traced by Si I.
As the dense cores embedded in the wind are disrupted, the cold gas is ablated and accelerated to a higher velocity while being heated by the wind.
This produces the asymmetric profiles of the Si I and II lines.
These asymmetries have been observed in absorption lines from several galaxies \citep{2009ApJS..181..272G,2012ApJ...751...51J,2013ApJ...765..118W,2015ApJ...809..147H,2015ApJ...810..104A,2015ApJ...811..149C,2016MNRAS.457.3133C}.
The asymmetric tail to higher velocities results from ablated gas being accelerated as it is entrained in the hot wind.
The Si IV (Figure \ref{fig:lines:lineex}) and Si VII (Figure \ref{fig:lines:Siex}) lines are not highly asymmetric due to contributions from both slower gas ablated off of the cooler dense cores, and warmer gas beginning to mix with the hot gas that fills the superbubble created by the starburst.
The Si VII line is not entirely smooth as it traces the large scale structure of filaments embedded in the wind.
The asymmetry of the Si XIII line skews in the opposite direction. The long tail to lower velocities results from the hot gas being accelerated as it moves off the galactic disk.
It reaches terminal velocity at $\sim 300$ pc above the galactic center to form the deep dip in absorption in Figure \ref{fig:lines:Siex}.
\subsection{Outflow Velocities}\label{sec:dis:vel}
As shown in Figures \ref{fig:res:Sscale} and \ref{fig:res:Sscalecen} the measured outflow velocity for neutral and slightly ionized gas increases with SFR.
Considering just Si II (Subplot (b) in both Figures \ref{fig:res:Sscale} and \ref{fig:res:Sscalecen}), at SFR $= 1$~\msyr, $v_{\rm cen}\approx 160$~\kms and $v_{90}\approx 400$~\kms.
At SFR $= 100$~\msyr there is significant divergence in the measured velocities depending on the $v_A$, but the highest measured velocities are $v_{\rm cen}\approx 560$~\kms and $v_{90}\approx 1070$~\kms.
\citet{2015ApJ...811..149C} use Si II lines to measure outflow velocities from their sample of star forming galaxies with SFRs ranging from 0.01 \msyr to 100 \msyr.
At SFR $= 1$~\msyr they measure $v_{\rm cen}\approx 50$~to 150 \kms and $v_{90}\approx 350$~\kms.
At higher SFRs they also have increased scatter in their measured velocities as we have in our data.
At SFR $\lesssim 100$~\msyr their highest velocities are $v_{\rm cen}\approx 500$~\kms and $v_{90}\approx 1100$~\kms.
Thus our measured velocities correspond to the outflow velocities found by \citet{2015ApJ...811..149C}.
We find similar agreement with the measured velocities from \citet{2015ApJ...809..147H} and \citet{2016ApJ...822....9H}, but only for our models with the highest $v_A$.
The sample of galaxies used by \citet{2015ApJ...809..147H} and \citet{2016ApJ...822....9H} overlap with the sample used by \citet{2015ApJ...811..149C}, as both use data from \citet{2011ApJ...730....5H} and \citet{2015ApJ...810..104A}, but the overlap only accounts for about a third to a half of the galaxies.
This results in increased scatter in the measured velocities at higher SFRs for \citet{2015ApJ...809..147H}.
As shown in Figure \ref{fig:lines:SRions} the velocities of neutral or slightly ionized gas are significantly lower than the velocity of the hot, highly ionized gas.
Higher outflow velocities for higher ionization have been found in some surveys of starburst galaxies \citep{2009ApJS..181..272G,2016A&A...588A..41C}.
Thus surveys which rely on absorption lines which trace neutral gas \citep{2000ApJS..129..493H,2005ApJ...621..227M}, or low ionized gas \citep{2014ApJ...794..156R,2015ApJ...811..149C,2015ApJ...809..147H,2016ApJ...822....9H,2016A&A...588A..41C} will underestimate outflow energetics.
The difference is more pronounced at low SFRs below the turnover point where there is degeneracy in the measured velocities for models with differing $v_A$.
This difference is greatest for models with high $v_A$.
The degeneracy below the turnover point means the degree to which the energy of the outflow is underestimated cannot be known.
At higher SFRs above the turnover point, the velocity of neutral and low ionized gas can more reliably be used as a proxy for the outflow velocity of the hot gas.
\subsection{Scaling Relations}\label{sec:dis:scale}
A scaling relation between outflow velocity and SFR has been found in numerous surveys of starburst galaxies \citep{2000ApJS..129..493H,2005ApJ...621..227M,2005ApJS..160..115R,2009ApJ...692..187W,2012ApJ...760..127M,2012ApJ...759...26E,2012ApJ...758..135K,2014ApJ...794..130B,2014ApJ...794..156R,2015ApJ...811..149C,2016MNRAS.457.3133C,2015ApJ...809..147H,2016ApJ...822....9H,2016A&A...588A..41C}.
A similar relationship has been found between the outflow velocity and \sfrd \citep{2010AJ....140..445C,2012ApJ...758..135K,2014ApJ...794..156R,2015ApJ...811..149C,2016MNRAS.457.3133C,2015ApJ...809..147H,2016ApJ...822....9H}.
All of these surveys rely on either optical or UV absorption lines, which trace warm, not hot gas.
Our results in Section \ref{sec:res:scale:S} show that only the hot gas follows the relation given in Equation \ref{eq:vAsimp}, where the outflow velocity does not depend on the SFR.
The outflow velocity of the warm gas does depend on the SFR, but only if the velocity is below some fraction of $v_A$.
The turnover point where the relationship flattens out, also called the saturation point \citep{2015ApJ...809..147H,2016ApJ...822....9H}, depends on the $v_A$ associated with the starburst and the ion being used.
Below the turnover point the scaling relation for each ion is uniform and does not depend on $v_A$.
Many surveys find a correlation between outflow velocity and SFR, but there is no agreement about the slope of the scaling relation ($\delta$).
For example, using $v_{90}$ velocities \citet{2005ApJS..160..115R} find $\delta = 0.21$, \citet{2009ApJ...692..187W} find $\delta = 0.38$, \citet{2015ApJ...811..149C} find $\delta = 0.081$, and \citet{2016ApJ...822....9H} find $\delta = 0.32$.
Yet using $v_{\rm cen}$ velocities \citet{2005ApJ...621..227M} find $\delta = 0.35$, and \citet{2015ApJ...811..149C} find $\delta = 0.22$.
Our results show that the lower $\delta$ values result from samples comprising a combination of galaxies with high and low $v_A$ velocities, and therefore a mix of galaxies above and below the turnover point.
Including a mix of galaxies in the sample would produce increased scatter in the data at higher SFRs or \sfrd because it would include galaxies above the turnover point in the relationship.
In contrast, the higher $\delta$ values would come from samples of galaxies with either relatively uniform, and high, $v_A$, or a sample restricted to the lower range of SFRs below the turnover point.
For example, \citet{2015ApJ...809..147H} calculate $\beta\approx 2$ for all of the galaxies in their sample.
This would imply that the galaxies in their sample have relatively uniform $v_A$ velocities between 1500 and 2000 \kms, putting the turnover point in the velocity vs. SFR relation high enough that it would not significantly lower their $\delta$.
When we compare the galaxy samples used we find that high $\delta$ values come from samples consisting of ULIGs \citep{2005ApJ...621..227M}, high red-shift, luminous galaxies \citep{2009ApJ...692..187W}, or UV-bright galaxies \citet{2016ApJ...822....9H}.
Lower $\delta$ values come from IR-bright galaxies \citep{2005ApJS..160..115R} or samples of local starburst galaxies with a variety of morphologies \citep{2015ApJ...811..149C}.
While \citet{2015ApJ...809..147H} do not find a turnover in their relation between velocity and SFR, they note a turnover in the relation between velocity and \sfrd at $\sim 10^{-5}$ \msyrpc.
In \citet{2016ApJ...822....9H} they find the turnover point to be $\sim 10^{-4}$ \msyrpc.
The latter value is similar to our turnover point of $\sim 10^{-3.5}$ \msyrpc from our R series, which has a $v_A=2000$ \kms.
We would expect their turnover point to be slightly lower because the mass loading factor of their galaxies would imply a lower $v_A$ than our R series models.
Also, we set the radius of our starburst, and from that calculate \sfrd, while they find the starburst radius using the half light radius in the UV.
So we would expect affect the measured turnover point in the relationship.
\section{Conclusions}\label{sec:end}
Using synthetic absorption lines generated from our series of 3D starburst models, we have investigated how the velocity of starburst-driven galactic winds scales with both SFR and \sfrd.
As implied by Equation \ref{eq:vAsimp} the terminal wind velocity should not depend on SFR, and by extension \sfrd.
Our simulations, both in this paper and in \citetalias{Tanner1}, have shown that the gas entrained in a starburst-driven wind does not move as a single coherent structure.
Cold gas moves much slower than warm gas, which in turn moves slower than hot, highly ionized gas.
But the relative velocities of these three gas phases do not relate to each other in a simple straight forward manner.
The measured velocity of the hot, highly ionized, gas is independent of SFR, and instead depends on $v_A$, and by extension, the thermalization efficiency and mass loading factor associated with the starburst.
Meanwhile both the warm and cold gas scale with SFR, but only up to a point.
The point where the scaling relation flattens out and the velocity becomes independent of SFR is determined again by $v_A$.
Below this point the scaling relation is independent of $v_A$, and depends entirely on what temperature regime is being probed.
Because of this, there is degeneracy in the velocity for different values of $v_A$, which means at low SFRs neutral or warm gas tracers cannot be used as proxies for the velocity of the hot, highly ionized gas.
Using neutral or low ionized gas to determine outflow energetics will underestimate the total energy by an indeterminate amount.
Our models show that at low SFRs it is possible to have a high velocity ($\sim 2000$~\kms) outflow while the neutral gas embedded in the wind would have a much lower velocity ($\sim 300$~\kms).
The relationship between the outflow velocity and \sfrd behaves in a similar way.
There is a point, determined by $v_A$, where the relationship flattens out, but below that point the velocity scales with \sfrd.
The turnover has been found \citep{2015ApJ...809..147H}, and is in agreement with our results.
Even though we did not address how outflow velocity scales with other galaxy characteristics such as total stellar mass and the specific SFR, similar effects in the scaling relations may be present.
Our results indicate that the differences in the measured scaling relations stem from sample selection criteria that group galaxies with either a mix of $v_A$ velocities or uniform velocities.
This would explain why \citet{2014ApJ...794..156R} found no correlation between the outflow velocity and either SFR and \sfrd, while \citet{2015ApJ...811..149C} found a correlation for both, but with a modest slope.
But \citet{2015ApJ...809..147H} found a steeper slope to the correlation for both SFR and \sfrd.
If we only use models with a high $v_A$ we reproduce the high slopes found by \citet{2015ApJ...809..147H}, and when we include models a mix of both high and low $v_A$ we reproduce the slopes found by \citet{2015ApJ...811..149C}.
If we were to extend our models to include higher mass loading factors, and therefore lower $v_A$, our models would show no correlation, with a high scatter, as found by \citet{2014ApJ...794..156R}.
Thus when considering the scaling relations between outflow velocity, SFR and \sfrd we cannot treat a sample of star forming galaxies as a single coherent group.
\acknowledgments{}
NASA Herschel grant NHSC-OT-1-1436036 and NC Space Grant supported this work.
\medskip
|
\section{Introduction}
\label{sec introduction}
\paragraph*{Framework.}
Let us first describe our framework and state the main results of this article (see Section~\ref{sec random real algebraic submanifolds} for more details). Let $\mathcal{X}$ be a smooth complex projective manifold of positive complex dimension $n$. Let $\L$ be an ample holomorphic line bundle over $\mathcal{X}$ and let $\mathcal{E}$ be a rank $r$ holomorphic vector bundle over $\mathcal{X}$, with $r \in \{1,\dots,n\}$. We assume that $\mathcal{X}$, $\mathcal{E}$ and $\L$ are endowed with compatible real structures and that the real locus $M$ of $\mathcal{X}$ is not empty. Let $h_\mathcal{E}$ and $h_\L$ denote Hermitian metrics on $\mathcal{E}$ and $\L$ respectively that are compatible with the real structures. We assume that $h_\L$ has positive curvature $\omega$. Then $\omega$ is a Kähler form on $\mathcal{X}$ and it induces a Riemannian metric $g$ on $M$.
For any $d \in \mathbb{N}$, the Kähler form $\omega$, $h_\mathcal{E}$ and $h_\L$ induce a $L^2$-inner product on the space $\mathbb{R} \H$ of real holomorphic sections of $\mathcal{E} \otimes \L^d \to \mathcal{X}$ (see \eqref{eq definition inner product}). Let $d \in \mathbb{N}$ and $s \in \mathbb{R} \H$, we denote by $Z_s$ the real zero set $s^{-1}(0)\cap M$ of $s$. For $d$ large enough, for almost every $s \in \mathbb{R} \H$, $Z_s$ is a codimension $r$ smooth submanifold of $M$ and we denote by $\rmes{s}$ the Riemannian measure on $Z_s$ induced by $g$ (see Sect.~\ref{subsec random submanifolds}). In the sequel, we will consider $\rmes{s}$ as a positive Radon measure on $M$. Let us also denote by $\rmes{M}$ the Riemannian measure on $M$.
Let $s_d$ be a standard Gaussian vector in $\mathbb{R} \H$. Then $\rmes{s_d}$ is a random positive Radon measure on $M$. We set $Z_{d}=Z_{s_d}$ and $\rmes{d}=\rmes{s_d}$ to avoid too many subscripts. In a previous paper \cite[thm.~1.3]{Let2016}, we computed the asymptotic of the expected Riemannian volume of $Z_d$ as $d \to +\infty$. Namely, we proved that:
\begin{equation}
\label{eq reminder expectation volume}
\esp{\vol{Z_d}} = d^\frac{r}{2}\vol{M} \frac{\vol{\S^{n-r}}}{\vol{\S^n}} + O\!\left(d^{\frac{r}{2}-1}\right),
\end{equation}
where $\vol{M}$ is the volume of $M$ for $\rmes{M}$ and the volumes of spheres are Euclidean volumes. Here and in throughout this paper, $\esp{\cdot}$ denotes the expectation of the random variable between the brackets, and $\S^m$ stands for the unit Euclidean sphere of dimension $m$.
Let $\phi \in \mathcal{C}^0(M)$, we denote by $\Norm{\phi}_\infty = \max_{x \in M} \norm{\phi(x)}$ its norm sup. Besides, we denote by $\prsc{\cdot}{\cdot}$ the duality pairing between $\left(\mathcal{C}^0(M),\Norm{\cdot}_\infty\right)$ and its topological dual. Then, \eqref{eq reminder expectation volume} can be restated as:
\begin{equation*}
\esp{\prsc{\rmes{d}}{\mathbf{1}}} = d^\frac{r}{2}\vol{M} \frac{\vol{\S^{n-r}}}{\vol{\S^n}} + O\!\left(d^{\frac{r}{2}-1}\right),
\end{equation*}
where $\mathbf{1} \in \mathcal{C}^0(M)$ stands for the unit constant function on $M$. The same proof gives similar asymptotics for $\esp{\prsc{\rmes{d}}{\phi}}$ for any continuous $\phi : M \to \mathbb{R}$ (see \cite[section 5.3]{Let2016}).
\begin{thm}
\label{thm reminder expectation test function}
Let $\mathcal{X}$ be a complex projective manifold of positive dimension $n$ defined over the reals, we assume that its real locus $M$ is non-empty. Let $\mathcal{E} \to \mathcal{X}$ be a rank $r \in \{1,\dots,n\}$ Hermitian vector bundle and let $\L \to \mathcal{X}$ be a positive Hermitian line bundle, both equipped with compatible real structures. For every $d \in \mathbb{N}$, let $s_d$ be a standard Gaussian vector in $\mathbb{R} \H$. Then the following holds as $d \to +\infty$:
\begin{equation}
\label{eq reminder expectation test function}
\forall \phi \in \mathcal{C}^0(M), \qquad \esp{\prsc{\rmes{d}}{\phi}} = d^\frac{r}{2} \left(\int_M \phi \rmes{M}\right) \frac{\vol{\S^{n-r}}}{\vol{\S^n}} + \Norm{\phi}_\infty O\!\left(d^{\frac{r}{2}-1}\right).
\end{equation}
Moreover the error term $O\!\left(d^{\frac{r}{2}-1}\right)$ does not depend on $\phi$.
\end{thm}
In particular, we can define a sequence of Radon measures $\left(\esp{\rmes{d}}\right)_{d\geq d_0}$ on $M$ by: for every $d \geq d_0$ and every $\phi \in \mathcal{C}^0(M)$, $\prsc{\esp{\rmes{d}}}{\phi} = \esp{\prsc{\rmes{d}}{\phi}}$. Then Thm.~\ref{thm reminder expectation test function} implies that:
\begin{equation}
\left(d^{-\frac{r}{2}}\right)\esp{\rmes{d}} \xrightarrow[d \to + \infty]{} \frac{\vol{\S^{n-r}}}{\vol{\S^n}} \rmes{M},
\end{equation}
as continuous linear functionals on $\left(\mathcal{C}^0(M),\Norm{\cdot}_\infty\right)$.
\paragraph*{Statement of the results.}
The main result of this paper is an asymptotic for the covariances of the linear statistics $\left\{\prsc{\rmes{d}}{\phi}\mathrel{}\middle|\mathrel{} \phi \in \mathcal{C}^0(M)\right\}$. Before we can state our theorem, we need to introduce some additional notations.
As usual, we denote by $\var{X} = \esp{\left(X-\esp{X}\right)^2}$ the variance of the real random variable $X$, and by $\cov{X}{Y} = \esp{\left(X-\esp{X}\right)\left(Y-\esp{Y}\right)}$ the covariance of the real random variables $X$ and $Y$. We call \emph{variance} of $\rmes{d}$ and we denote by $\var{\rmes{d}}$ the symmetric bilinear form on $\mathcal{C}^0(M)$ defined by:
\begin{equation}
\label{def cov dVd}
\forall \phi_1,\phi_2 \in \mathcal{C}^0(M), \qquad \var{\rmes{d}}\left(\phi_1,\phi_2\right) = \cov{\prsc{\rmes{d}}{\phi_1}}{\prsc{\rmes{d}}{\phi_2}}.
\end{equation}
\begin{dfn}
\label{def continuity modulus}
Let $\phi \in \mathcal{C}^0(M)$, we denote by $\varpi_\phi$ its \emph{continuity modulus}, which is defined by:
\begin{equation*}
\begin{array}{cccc}
\varpi_\phi: & (0,+\infty) & \longrightarrow & [0,+\infty) \\
& \epsilon & \longmapsto & \sup \left\{ \norm{\phi(x)-\phi(y)} \mathrel{}\middle|\mathrel{} (x,y) \in M^2, \rho_g(x,y)\leq \epsilon \right\},
\end{array}
\end{equation*}
where $\rho_g(\cdot,\cdot)$ stands for the geodesic distance on $(M,g)$.
\end{dfn}
Since $M$ is compact, $\varpi_\phi$ is well-defined for every $\phi \in \mathcal{C}^0(M)$. Moreover every $\phi \in \mathcal{C}^0(M)$ is uniformly continuous and we have:
\begin{equation*}
\forall \phi \in \mathcal{C}^0(M), \qquad \varpi_\phi (\epsilon) \xrightarrow[\epsilon \to 0]{} 0.
\end{equation*}
Note that, if $\phi: M \to \mathbb{R}$ is Lipschitz continuous, then $\varpi_\phi(\epsilon) = O\!\left(\epsilon\right)$ as $\epsilon \to 0$.
\begin{dfn}
\label{def Jacobian}
Let $L : V \to V'$ be a linear map between two Euclidean spaces, we denote by $\odet{L}$ the \emph{Jacobian} of $L$:
\begin{equation*}
\label{eq def odet}
\odet{L} = \sqrt{\det\left(LL^*\right)},
\end{equation*}
where $L^*:V' \to V$ is the adjoint operator of $L$.
\end{dfn}
See Section~\ref{subsec Kac--Rice formula} for a quick discussion of the properties of this Jacobian. If $A$ is an element of $\mathcal{M}_{rn}(\mathbb{R})$, the space of matrices of size $r\times n$ with real coefficients, we denote by $\odet{A}$ the Jacobian of the linear map from $\mathbb{R}^n$ to $\mathbb{R}^r$ associated with $A$ in the canonical bases of~$\mathbb{R}^n$ and $\mathbb{R}^r$.
\begin{dfn}
\label{def XYt}
For every $t>0$, we define $\left(X(t),Y(t)\right)$ to be a centered Gaussian vector in $\mathcal{M}_{rn}(\mathbb{R})\times \mathcal{M}_{rn}(\mathbb{R})$ with variance matrix:
\begin{equation*}
\left(\begin{array}{ccccc|ccccc}
1-\frac{te^{-t}}{1-e^{-t}} & 0 & \cdots & \cdots & 0 & e^{-\frac{t}{2}}-\frac{te^{-\frac{t}{2}}}{1-e^{-t}} & 0 & \cdots & \cdots & 0\\
0 & 1 & \ddots & & \vdots & 0 & e^{-\frac{t}{2}} & \ddots & & \vdots\\
\vdots & \ddots & \ddots & \ddots & \vdots & \vdots & \ddots & \ddots & \ddots & \vdots \\
\vdots & & \ddots & 1 & 0 & \vdots & & \ddots & e^{-\frac{t}{2}} & 0\\
0 & \cdots & \cdots & 0 & 1 & 0 & \cdots & \cdots & 0 & e^{-\frac{t}{2}}\\
\hline
e^{-\frac{t}{2}}-\frac{te^{-\frac{t}{2}}}{1-e^{-t}} & 0 & \cdots & \cdots & 0 & 1-\frac{te^{-t}}{1-e^{-t}} & 0 & \cdots & \cdots & 0\\
0 & e^{-\frac{t}{2}} & \ddots & & \vdots & 0 & 1 & \ddots & & \vdots\\
\vdots & \ddots & \ddots & \ddots & \vdots & \vdots & \ddots & \ddots & \ddots & \vdots \\
\vdots & & \ddots & e^{-\frac{t}{2}} & 0 & \vdots & & \ddots & 1 & 0\\
0 & \cdots & \cdots & 0 & e^{-\frac{t}{2}} & 0 & \cdots & \cdots & 0 & 1
\end{array}\right) \otimes I_r,
\end{equation*}
where $I_r$ is the identity matrix of size $r$. That is, if we denote by $X_{ij}(t)$ (resp.~$Y_{ij}(t)$) the coefficients of $X(t)$ (resp.~$Y(t)$), the couples $\left(X_{ij}(t),Y_{ij}(t)\right)$ with $1 \leq i \leq r$ and $1 \leq j \leq n$ are independent from one another and the variance matrix of $\left(X_{ij}(t),Y_{ij}(t)\right)$ is:
\begin{align*}
&\begin{pmatrix}
1-\frac{te^{-t}}{1-e^{-t}}& e^{-\frac{t}{2}}\left(1-\frac{t}{1-e^{-t}}\right)\\
e^{-\frac{t}{2}}\left(1-\frac{t}{1-e^{-t}}\right) & 1-\frac{te^{-t}}{1-e^{-t}}
\end{pmatrix} & &\text{if } j=1, \text{ and} & &\begin{pmatrix}
1& e^{-\frac{t}{2}}\\ e^{-\frac{t}{2}} & 1
\end{pmatrix} & &\text{otherwise.}
\end{align*}
\end{dfn}
\begin{ntn}
\label{ntn alpha0}
We set $\alpha_0 = \dfrac{n-r}{2(2r+1)(2n+1)}$.
\end{ntn}
We can now state our main result.
\begin{thm}
\label{thm asymptotics variance}
Let $\mathcal{X}$ be a complex projective manifold of dimension $n\geq 2$ defined over the reals, we assume that its real locus $M$ is non-empty. Let $\mathcal{E} \to \mathcal{X}$ be a rank $r \in \{1,\dots,n-1\}$ Hermitian vector bundle and let $\L \to \mathcal{X}$ be a positive Hermitian line bundle, both equipped with compatible real structures. For every $d \in \mathbb{N}$, let $s_d$ be a standard Gaussian vector in $\mathbb{R} \H$.
Let $\beta \in (0,\frac{1}{2})$, then there exists $C_\beta >0$ such that, for all $\alpha \in (0,\alpha_0)$, for all $\phi_1$ and $\phi_2 \in \mathcal{C}^0(M)$, the following holds as $d \to +\infty$:
\begin{multline}
\label{eq asymptotics variance}
\var{\rmes{d}}\left(\phi_1,\phi_2\right) = d^{r-\frac{n}{2}} \left(\int_M \phi_1 \phi_2 \rmes{M}\right) \frac{\vol{\S^{n-1}}}{(2\pi)^r} \mathcal{I}_{n,r}\\
+ \Norm{\phi_1}_{\infty}\Norm{\phi_2}_{\infty} O\!\left(d^{r-\frac{n}{2}-\alpha}\right) + \Norm{\phi_1}_{\infty}\varpi_{\phi_2}\left(C_\beta d^{-\beta}\right) O\!\left(d^{r-\frac{n}{2}}\right),
\end{multline}
where
\begin{equation}
\label{eq def Inr}
\mathcal{I}_{n,r} = \frac{1}{2} \int_0^{+\infty} \left(\frac{\esp{\odet{X(t)}\odet{Y(t)}}}{\left(1-e^{-t}\right)^\frac{r}{2}} -(2\pi)^r\left(\frac{\vol{\S^{n-r}}}{\vol{\S^n}}\right)^2\right)t^{\frac{n-2}{2}}\dmesure\! t< + \infty.
\end{equation}
Moreover the error terms $O\!\left(d^{r-\frac{n}{2}-\alpha}\right)$ and $O\!\left(d^{r-\frac{n}{2}}\right)$ in~\eqref{eq asymptotics variance} do not depend on $\left(\phi_1,\phi_2\right)$.
\end{thm}
We obtain the variance of the volume of $Z_d$ by applying Thm.~\ref{thm asymptotics variance} to $\phi_1=\phi_2= \mathbf{1}$. When $\phi_1=\phi_2= \phi$ we get the following.
\begin{cor}[Variance of the linear statistics]
\label{cor variance of linear statistics}
In the same setting as Thm.~\ref{thm asymptotics variance}, let $\beta \in (0,\frac{1}{2})$, then there exists $C_\beta >0$ such that, for all $\alpha \in (0,\alpha_0)$ and all $\phi \in \mathcal{C}^0(M)$, the following holds as $d \to +\infty$:
\begin{multline}
\label{eq variance of linear statistics}
\var{\prsc{\rmes{d}}{\phi}} = d^{r-\frac{n}{2}} \left(\int_M \phi^2 \rmes{M}\right) \frac{\vol{\S^{n-1}}}{(2\pi)^r} \mathcal{I}_{n,r}\\
+ \Norm{\phi}_{\infty}^2 O\!\left(d^{r-\frac{n}{2}-\alpha}\right) + \Norm{\phi}_{\infty}\varpi_{\phi}\left(C_\beta d^{-\beta}\right) O\!\left(d^{r-\frac{n}{2}}\right).
\end{multline}
Moreover, the error terms do not depend on $\phi$.
\end{cor}
\begin{rems}
Some remarks are in order.
\begin{itemize}
\item The value of the constant $\alpha_0$ should not be taken too seriously. This constant appears for technical reasons and it is probably far from optimal.
\item If $\phi_2$ is Lipschitz continuous with Lipschitz constant $K$, then the error term in eq.~\eqref{eq asymptotics variance} can be replaced by:
\begin{equation*}
\Norm{\phi_1}_\infty \left(\Norm{\phi_2}_\infty + K\right) O\!\left(d^{r-\frac{n}{2}-\alpha}\right)
\end{equation*}
by fixing $\beta > \alpha_0$, which is possible since $\frac{1}{2} > \alpha_0$.
\item Thm.~\ref{thm asymptotics variance} shows that $\var{\rmes{d}}$ is a continuous bilinear form on $\left(\mathcal{C}^0(M),\Norm{\cdot}_\infty\right)$ for $d$ large enough. Moreover, denoting by $\prsc{\cdot}{\cdot}_M$ the $L^2$-inner product on $\mathcal{C}^0(M)$ defined by $\prsc{\phi_1}{\phi_2}_M = \displaystyle\int_M \phi_1\phi_2 \rmes{M}$, we have:
\begin{equation*}
d^{\frac{n}{2}-r} \var{\rmes{d}} \xrightarrow[d \to + \infty]{} \frac{\vol{\S^{n-1}}}{(2\pi)^r} \mathcal{I}_{n,r} \prsc{\cdot}{\cdot}_M
\end{equation*}
in the weak sense. A priori, there is no such convergence as continuous bilinear forms on $\left(\mathcal{C}^0(M),\Norm{\cdot}_\infty\right)$ since the estimate~\eqref{eq asymptotics variance} involves the continuity modulus of $\phi_2$.
\item The fact that the constant $\mathcal{I}_{n,r}$ is finite is part of the statement and is proved below (Lemma~\ref{lem integrability Dnr}). This constant is necessarily non-negative. Numerical evidence suggests that it is positive but we do not know how to prove it at this point.
\item Thm.~\ref{thm asymptotics variance} does not apply in the case of maximal codimension ($r=n$). This case presents an additional singularity which causes our proof to fail. However, we believe a similar result to be true for $r=n$, at least in the case of the Kostlan--Shub--Smale polynomials described below (compare \cite{Dal2015,Wsc2005}).
\end{itemize}
\end{rems}
\begin{cor}[Concentration in probability]
\label{cor concentration}
In the same setting as Thm.~\ref{thm asymptotics variance}, let $\alpha > \frac{r}{2}- \frac{n}{4}$ and let $\phi \in \mathcal{C}^0(M)$. Then, for every $\epsilon >0$, we have:
\begin{equation*}
\P\left( \norm{\prsc{\rmes{d}}{\phi}-\rule{0pt}{4mm}\esp{\prsc{\rmes{d}}{\phi}}} > d^{\alpha} \epsilon \right) = \frac{1}{\epsilon^2} O\!\left( d^{r-\frac{n}{2}-2\alpha} \right),
\end{equation*}
where the error term is independent of $\epsilon$, but depends on $\phi$.
\end{cor}
\begin{cor}
\label{cor connected components}
In the same setting as Thm.~\ref{thm asymptotics variance}, let $U \subset M$ be an open subset, then as $d\to + \infty$ we have:
\begin{equation*}
\P\left( Z_d \cap U = \emptyset \right) = O\!\left( d^{-\frac{n}{2}} \right).
\end{equation*}
\end{cor}
Our last corollary is concerned with the convergence of a random sequence of sections of increasing degree. Let us denote by $\dmesure\! \nu_d$ the standard Gaussian measure on $\mathbb{R}\H$ (see~\eqref{eq Gaussian density}). Let $\dmesure\! \nu$ denote the product measure $\bigotimes_{d \in \mathbb{N}} \dmesure\! \nu_d$ on $\prod_{d \in \mathbb{N}} \mathbb{R} \H$. Then we have the following.
\begin{cor}[Almost sure convergence]
\label{cor as convergence}
In the same setting as Thm.~\ref{thm asymptotics variance}, let us assume that $n\geq 3$. Let $(s_d)_{d \in \mathbb{N}} \in \prod_{d \in \mathbb{N}} \mathbb{R} \H$ be a random sequence of sections. Then, $\dmesure\! \nu$-almost surely, we have:
\begin{equation*}
\forall \phi \in \mathcal{C}^0(M), \qquad d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\phi} \xrightarrow[d \to +\infty]{} \frac{\vol{\S^{n-r}}}{\vol{\S^n}} \left(\int_M \phi \rmes{M}\right).
\end{equation*}
That is, $\dmesure\! \nu$-almost surely,
\begin{equation*}
d^{-\frac{r}{2}}\rmes{s_d} \xrightarrow[d \to +\infty]{} \frac{\vol{\S^{n-r}}}{\vol{\S^n}}\rmes{M},
\end{equation*}
in the sense of the weak convergence of measures.
\end{cor}
\begin{rem}
We expect this result to hold for $n=2$ as well, but our proof fails in this case.
\end{rem}
\paragraph*{The Kostlan--Shub--Smale polynomials}
Let us consider the simplest example of our framework. We choose $\mathcal{X}$ to be the complex projective space $\mathbb{C}\P^n$, with the real structure defined by the usual conjugation in $\mathbb{C}^{n+1}$. Then $M$ is the real projective space $\mathbb{R} \P^n$. Let $\L = \mathcal{O}(1)$ be the hyperplane line bundle, equipped with its natural real structure and the metric dual to the standard metric on the tautological line bundle over $\mathbb{C}\P^n$. Then the curvature form of $\L$ is the Fubini--Study form $\omega_{FS}$, normalized so that the induced Riemannian metric is the quotient of the Euclidean metric on the unit sphere of $\mathbb{C}^{n+1}$. Let $\mathcal{E} = \mathbb{C}^r \times \mathbb{C}\P^n \to \mathbb{C}\P^n$ be the rank $r$ trivial bundle with the trivial real structure and the trivial metric.
In this setting, the global holomorphic sections of $\L^d$ are the complex homogeneous polynomials of degree $d$ in $n+1$ variables and those of $\mathcal{E} \otimes \L^d$ are $r$-tuples of such polynomials, since $\mathcal{E}$ is trivial. Finally, the real structures being just the usual conjugations, we have:
\begin{equation*}
\mathbb{R} \H = \mathbb{R}_{\text{hom}}^d[X_0,\dots,X_n]^r,
\end{equation*}
where $\mathbb{R}_{\text{hom}}^d[X_0,\dots,X_n]$ is the space of real homogeneous polynomials of degree $d$ in $n+1$ variables. The $r$ copies of this space in $\mathbb{R} \H$ are pairwise orthogonal for the inner product~\eqref{eq definition inner product}. Hence a standard Gaussian in $\mathbb{R}\H$ is a $r$-tuple of independent standard Gaussian in $\mathbb{R}_{\text{hom}}^d[X_0,\dots,X_n] = \mathbb{R} H^0\left(\mathcal{X},\L^d\right)$.
It is well-known (cf.~\cite{BSZ2000a,BBL1996,Kos1993}) that the monomials are pairwise orthogonal for the $L^2$-inner product~\eqref{eq definition inner product}, but not orthonormal. Let $\alpha =\left(\alpha_0,\dots,\alpha_n\right) \in \mathbb{N}^{n+1}$, we denote its length by $\norm{\alpha} = \alpha_0 + \dots + \alpha_n$. We also define $X^{\alpha} = X_0^{\alpha_0} \cdots X_n^{\alpha_n}$ and $\alpha ! = (\alpha_0!) \cdots (\alpha_n!)$. Finally, if $\norm{\alpha}=d$, we denote by $\binom{d}{\alpha}$ the multinomial coefficient $\frac{d!}{\alpha!}$. Then, an orthonormal basis of $\mathbb{R}_{\text{hom}}^d[X_0,\dots,X_n]$ for the inner product~\eqref{eq definition inner product} is given by the family:
\begin{equation*}
\left(\sqrt{\frac{(d+n)!}{\pi^n d!}} \sqrt{\binom{d}{\alpha}}X^\alpha\right)_{\norm{\alpha} = d}.
\end{equation*}
Thus a standard Gaussian vector in $\mathbb{R}_{\text{hom}}^d[X_0,\dots,X_n]$ is a random polynomial:
\begin{equation*}
\sqrt{\frac{(d+n)!}{\pi^n d!}} \sum_{\norm{\alpha} =d} a_\alpha \sqrt{\binom{d}{\alpha}}X^\alpha,
\end{equation*}
where the coefficients $(a_\alpha)_{\norm{\alpha}=d}$ are independent real standard Gaussian variables. Since we are only concerned with the zero set of this random polynomial, we can drop the factor $\sqrt{\frac{(d+n)!}{\pi^n d!}}$.
Finally, in this setting, $\rmes{d}$ is the common zero set of $r$ independent random polynomials in $\mathbb{R}_{\text{hom}}^d[X_0,\dots,X_n]$ of the form:
\begin{equation}
\label{KSS polynomials}
\sum_{\norm{\alpha} =d} a_\alpha \sqrt{\binom{d}{\alpha}}X^\alpha,
\end{equation}
with independent coefficients $(a_\alpha)_{\norm{\alpha}=d}$ distributed according to the real standard Gaussian distribution. Such polynomials are known as the Kostlan--Shub--Smale polynomials. They were introduced in \cite{Kos1993,SS1993} and were actively studied since (cf.~\cite{AW2005,Buer2007,Dal2015,Pod2001,Wsc2005}).
\paragraph*{Related works.}
As we just said, zero sets of systems of independent random polynomials distributed as~\eqref{KSS polynomials} were studied by Kostlan \cite{Kos1993} and Shub and Smale \cite{SS1993}. The expected volume of these random algebraic manifolds was computed by Kostlan \cite{Kos1993} and their expected Euler characteristic was computed by Podkorytov \cite{Pod2001} in codimension $1$, and by Bürgisser \cite{Buer2007} in higher codimension. Both these results were extended to the setting of the present paper in \cite{Let2016}.
In \cite{Wsc2005}, Wschebor obtained an asymptotic bound, as the dimension $n$ goes to infinity, for the variance of number of real roots of a system of $n$ independent Kostlan--Shub--Smale polynomials. Recently, Dalmao \cite{Dal2015} computed an asymptotic of order $\sqrt{d}$ for the variance of the number of real roots of one Kostlan--Shub--Smale polynomial in dimension $n=1$. His result is very similar to~\eqref{eq asymptotics variance}, which leads us to think that such a result should hold for $r=n$. He also proved a central limit theorem for this number of real roots, using Wiener chaos methods.
In \cite[thm.~3]{KL2001}, Kratz and Le\`{o}n considered the level curves of a centered stationary Gaussian field with unit variance on the plane $\mathbb{R}^2$. More precisely, they considered the length of a level curve intersected with some large square $[-T,T]\times [-T,T]$. As $T \to + \infty$, they proved asymptotics of order $T^2$ for both the expectation and the variance of this length. They also proved that it satisfies a central limit theorem as $T \to +\infty$. In particular, their result applies to the centered Gaussian field on $\mathbb{R}^2$ with correlation function $(x,y)\mapsto \exp\left(-\frac{1}{2}\Norm{x-y}^2\right)$. This field can be seen as the scaling limit, in the sense of \cite{NS2015}, of the centered Gaussian field $\left(s_d(x)\right)_{x \in M}$ defined by our random sections, when $n=2$ and $r=1$.
The study of more general random algebraic submanifolds, obtained as the zero sets of random sections, was pioneered by Shiffman and Zelditch \cite{SZ1999,SZ2008,SZ2010}. They considered the integration current over the common complex zero set $Z_d$ of $r$ independent random sections in $H^0(\mathcal{X},\L^d)$, distributed as standard complex Gaussians. In \cite{SZ1999}, they computed the asymptotic, as $d$ goes to infinity, of the expectation of the associated smooth statistics when $r=1$. They also provided an upper bound for the variance of these quantities and proved the equivalent of Cor.~\ref{cor as convergence} in this complex algebraic setting. In \cite{SZ2008}, they gave an asymptotic of order $d^{2r-n-\frac{1}{2}}$ for the variance of the volume of $Z_d \cap U$, where $U \subset \mathcal{X}$ is a domain satisfying some regularity conditions. In \cite{SZ2010}, they proved a similar asymptotic for the variance of the smooth statistics associated with $Z_d$. When $r=1$, they deduced a central limit theorem from these estimates and an asymptotic normality result of Sodin and Tsirelson \cite{ST2004}. Finally, in \cite[thm.~1.4]{SZZ2008}, Shiffman, Zelditch and Zrebiec proved that the probability that $Z_d \cap U = \emptyset$, where $U$ is any open subset of $\mathcal{X}$, decreases exponentially fast as $d$ goes to infinity.
Coming back to our real algebraic setting, one should be able to deduce from the general result of Nazarov and Sodin \cite[thm.~3]{NS2015} that, given an open set $U \subset M$, the probability that $Z_d \cap U = \emptyset$ goes to $0$ as $d$ goes to infinity. Corollary~\ref{cor connected components} gives an upper bound for the convergence rate. In particular, this bounds the probability for $Z_d$ to be empty. In the same spirit, Gayet and Welschinger \cite{GW2015} proved the following result. Let $\Sigma$ be a fixed diffeomorphism type of codimension $r$ submanifold of $\mathbb{R}^n$, let $x \in M$ and let $B_d(x)$ denote the geodesic ball of center $x$ and radius $\frac{1}{\sqrt{d}}$. Then, the probability that $Z_d \cap B_d(x)$ contains a submanifold diffeomorphic to $\Sigma$ is bounded from below. On the other hand, when $n=2$ and $r=1$, the Harnack--Klein inequality shows that the number of connected components of $Z_d$ is bounded by a polynomial in $d$. In \cite{GW2011}, Gayet and Welschinger proved that the probability for $Z_d$ to have the maximal number of connected components decreases exponentially fast with $d$.
Another well-studied model of random submanifolds is that of Riemannian random waves, i.e.~zero sets of random eigenfunctions of the Laplacian associated with some eigenvalue $\lambda$. In this setting, Rudnick and Wigman \cite{RW2008} computed an asymptotic bound, as $\lambda \to +\infty$, for the variance of the volume of a random hypersurface on the flat $n$-dimensional torus $\mathbb{T}^n$. On $\mathbb{T}^2$, this result was improved by Krishnapur, Kurlberg and Wigman \cite{KKW2013} who computed the precise asymptotic of the variance of the length of a random curve. In \cite{Wig2010}, Wigman computed the asymptotic variance of the linear statistics associated with a random curve on the Euclidean sphere $\S^2$. His result holds for a large class of test-function that contains the characteristic functions of open sets satisfying some regularity assumption. In relation with Cor.~\ref{cor connected components}, Nazarov and Sodin \cite{NS2009} proved that, on the Euclidean sphere $\S^2$, the number of connected components of a random curve times $\frac{1}{\lambda}$ converges exponentially fast in probability to a deterministic constant as $\lambda \to + \infty$.
\paragraph*{About the proof.}
The idea of the proof is the following. The random section $s_d$ defines a centered Gaussian field $(s_d(x))_{x \in \mathcal{X}}$. The correlation kernel of this field equals the Bergman kernel, that is the kernel of the orthogonal projection onto $\H$ for the inner product~\eqref{eq definition inner product} (compare \cite{BSZ2000a,Let2016,SZ1999,SZ2008,SZ2010}).
In order to compute the covariance of the smooth statistics $\prsc{\rmes{s}}{\phi_1}$ and $\prsc{\rmes{s}}{\phi_2}$, we apply a Kac--Rice formula (cf.~\cite{AW2009,BSZ2000a,Dal2015,TA2007,Wig2010}). This allows us to write $\var{\rmes{d}}(\phi_1,\phi_2)$ as the integral over $M\times M$ of some function $\mathcal{D}_d(x,y)$, defined by~\eqref{eq def density}. This density $\mathcal{D}_d(x,y)$ is the difference of two terms, coming respectively from
\begin{align*}
\esp{\prsc{\rmes{d}}{\phi_1}\prsc{\rmes{d}}{\phi_2}}& & &\text{and} & \esp{\prsc{\rmes{d}}{\phi_1}} \esp{\prsc{\rmes{d}}{\phi_2}}.
\end{align*}
Since the Bergman kernel decreases exponentially fast outside of the diagonal $\Delta$ in $M^2$ (see Section~\ref{subsec far off diagonal estimates}), the values of $s_d(x)$ and $s_d(y)$ are almost uncorrelated for $(x,y)$ far from~$\Delta$. As a consequence, when the distance between $x$ and $y$ is much larger than $\frac{1}{\sqrt{d}}$, the above two terms in the expression of $\mathcal{D}_d(x,y)$ are equal, up to a small error (see Sect.~\ref{subsubsec far off diagonal correlated} for a precise statement). Thus, $\mathcal{D}_d(x,y)$ is small far from $\Delta$, and its integral over this domain only contributes a remainder term to $\var{\rmes{d}}(\phi_1,\phi_2)$.
The main contribution to the value of $\var{\rmes{d}}(\phi_1,\phi_2)$ comes from the integration of $\mathcal{D}_d(x,y)$ over a neighborhood of $\Delta$ of size about~$\frac{1}{\sqrt{d}}$. We perform a change of variable in order to express this term as an integral over a domain of fixed size. This rescaling by $\frac{1}{\sqrt{d}}$ explains the factor $d^{-\frac{n}{2}}$ in~\eqref{eq asymptotics variance}. Besides, the order of growth of $\mathcal{D}_d(x,y)$ close to $\Delta$ is $d^r$, that is the order of growth of the square of $\esp{\rmes{d}}$ (see Thm.~\ref{thm reminder expectation test function}). Finally, we get an order of growth of $d^{r-\frac{n}{2}}$ for $\var{\rmes{d}}(\phi_1,\phi_2)$. The constant in~\eqref{eq asymptotics variance} appears as the scaling limit of the integral of $\mathcal{D}_d(x,y)$ over a neighborhood of $\Delta$ of typical size $\frac{1}{\sqrt{d}}$.
The difficulty in making this sketch of proof rigorous comes from the combination of two facts. First, we do not know exactly the value of the Bergman kernel (our correlation function) and its derivatives, but only asymptotics. In addition, the conditioning in the Kac--Rice formula is singular along $\Delta$, and so is $\mathcal{D}_d$. Because of this, we lose all uniformity in the control of the error terms close to the diagonal. Nonetheless, by careful bookkeeping of the error terms, we can make the above heuristic precise.
\paragraph*{Outline of the paper.}
In Section~\ref{sec random real algebraic submanifolds} we describe precisely our framework and the construction of the random measures $\rmes{s_d}$. We also introduce the Bergman kernel and explain how it is related to our random submanifolds.
In Section~\ref{sec estimates for the bergman kernel}, we recall various estimates for the Bergman kernel that we use in the proof of our main theorem. These estimates were established by Dai, Liu and Ma \cite{DLM2006}, and Ma and Marinescu \cite{MM2007,MM2013,MM2015} in a complex setting. Our main contribution in this section consists in checking that the preferred trivialization used by Ma and Marinescu to state their near-diagonal estimates is well-behaved with respect to the real structures on $\mathcal{X}$, $\mathcal{E}$ and $\L$ (see Section~\ref{subsec real normal trivialization}).
Section~\ref{sec proof of the main theorem} is concerned with the proof of Thm.~\ref{thm asymptotics variance}. In Sect.~\ref{subsec Kac--Rice formula}, we prove a Kac--Rice formula adapted to our problem, using Federer's coarea formula and Kodaira's embedding theorem. In Sect.~\ref{subsec an integral formula for the variance} we prove an integral formula for the variance, using the Kac--Rice formula (Thm.~\ref{thm Kac-Rice var}). The core of the proof is contained in Sect.~\ref{subsec asymptotic for the variance}.
Finally, we prove Corollaries~\ref{cor concentration}, \ref{cor connected components} and~\ref{cor as convergence} in Section~\ref{sec proofs of the corollaries}.
\paragraph*{Acknowledgments.}
I am thankful to Damien Gayet for his guidance in the course of this work and for countless mathematical discussions, on this topic and others.
\tableofcontents
\section{Random real algebraic submanifolds}
\label{sec random real algebraic submanifolds}
\subsection{General setting}
\label{subsec general setting}
In this section, we introduce our framework. It is the same as the algebraic setting of \cite{Let2016}, see also~\cite{GW2014b,GW2015}. Classical references for the material of this section are \cite[chap.~0]{GH1994} and \cite[chap.~1]{Sil1989}.
Let $\mathcal{X}$ be a smooth complex projective manifold of complex dimension $n\geq 2$. We assume that $\mathcal{X}$ is defined over the reals, that is $\mathcal{X}$ is equipped with an anti-holomorphic involution $c_\mathcal{X}$. The real locus of $(\mathcal{X},c_\mathcal{X})$ is the set of fixed points of $c_\mathcal{X}$. In the sequel, we assume that it is non-empty and we denote it by $M$. It is a classical fact that $M$ is a smooth closed (i.e.~compact without boundary) submanifold of $\mathcal{X}$ of real dimension $n$ (see \cite[chap.~1]{Sil1989}).
Let $\mathcal{E} \to \mathcal{X}$ be a holomorphic vector bundle of rank $r \in \{1,\dots,n-1\}$. Let $c_\mathcal{E}$ be a real structure on $\mathcal{E}$, compatible with $c_\mathcal{X}$ in the sense that the projection $\pi_\mathcal{E} : \mathcal{E} \to \mathcal{X}$ satisfies $c_\mathcal{X} \circ \pi_\mathcal{E} = \pi_\mathcal{E} \circ c_\mathcal{E}$ and $c_\mathcal{E}$ is fiberwise $\mathbb{C}$-anti-linear. Let $h_\mathcal{E}$ be a real Hermitian metric on $\mathcal{E}$, that is $c_\mathcal{E}^\star(h_\mathcal{E})=\overline{h_\mathcal{E}}$.
Similarly, let $\L \to \mathcal{X}$ be an ample holomorphic line bundle equipped with a compatible real structure $c_\L$ and a real Hermitian metric $h_\L$. Moreover, we assume that the curvature form $\omega$ of $h_\L$ is a Kähler form. Recall that if $\zeta$ is any non-vanishing holomorphic section on the open set $\Omega \subset \mathcal{X}$, then the restriction of $\omega$ to $\Omega$ is given by:
\begin{equation*}
\omega_{/\Omega} = \frac{1}{2i}\partial \bar{\partial} \ln\left(h_\L(\zeta,\zeta)\right).
\end{equation*}
This Kähler form is associated with a Hermitian metric $g_\mathbb{C}$ on $\mathcal{X}$. The real part of $g_\mathbb{C}$ defines a Riemannian metric $g = \omega(\cdot,i\cdot)$ on $\mathcal{X}$, compatible with the complex structure. Note that, since $h_\L$ is compatible with the real structures on $\mathcal{X}$ and $\L$, we have $c_\L^\star(h_\L)=\overline{h_\L}$ and $c_\mathcal{X}^\star\omega = - \omega$. Then we have $c_\mathcal{X}^\star g_\mathbb{C} = \overline{g_\mathbb{C}}$, hence $c_\mathcal{X}^\star g = g$ and $c_\mathcal{X}$ is an isometry of $(\mathcal{X},g)$.
Then $g$ induces a Riemannian measure on every smooth submanifold of $\mathcal{X}$. In the case of $\mathcal{X}$, this measure is given by the volume form $\dmesure\! V_\mathcal{X}=\frac{\omega^n}{n!}$. We denote by $\rmes{M}$ the Riemannian measure on $(M,g)$.
Let $d \in \mathbb{N}$, then the rank $r$ holomorphic vector bundle $\mathcal{E} \otimes \L^d$ can be endowed with a real structure $c_d = c_\mathcal{E} \otimes c_\L^d$, compatible with $c_\mathcal{X}$, and a real Hermitian metric $h_d = h_\mathcal{E} \otimes h_\L^d$. If $x \in M$, then $c_d$ induces a $\mathbb{C}$-anti-linear involution of the fiber $(\mathcal{E} \otimes \L^d)_x$. We denote by $\mathbb{R}(\mathcal{E} \otimes \L^d)_x$ the fixed points set of this involution, which is a dimension $r$ real vector space.
Let $\Gamma(\mathcal{E} \otimes \L^d)$ denote the space of smooth sections of $\mathcal{E} \otimes \L^d$. We can define a Hermitian inner product on $\Gamma(\mathcal{E} \otimes \L^d)$ by:
\begin{equation}
\label{eq definition inner product}
\forall s_1, s_2 \in \Gamma(\mathcal{E} \otimes \L^d), \qquad \prsc{s_1}{s_2} = \int_\mathcal{X} h_d(s_1(x),s_2(x)) \dmesure\! V_\mathcal{X}.
\end{equation}
We say that a section $s \in \Gamma(\mathcal{E} \otimes \L^d)$ is real if it is equivariant for the real structures, that is: $c_d \circ s = s \circ c_\mathcal{X}$. Let $\mathbb{R} \Gamma(\mathcal{E} \otimes \L^d)$ denote the real vector space of real smooth sections of $\mathcal{E} \otimes \L^d$. The restriction of $\prsc{\cdot}{\cdot}$ to $\mathbb{R} \Gamma(\mathcal{E} \otimes \L^d)$ is a Euclidean inner product.
\begin{ntn}
\label{ntn bracket}
In this paper, $\prsc{\cdot}{\cdot}$ will always denote either the inner product on the concerned Euclidean (or Hermitian) space or the duality pairing between a space and its topological dual. Which one will be clear from the context.
\end{ntn}
Let $\H$ denote the space of global holomorphic sections of $\mathcal{E} \otimes \L^d$. This space has finite complex dimension $N_d$ by Hodge's theory (compare \cite[thm.~1.4.1]{MM2007}). We denote by $\mathbb{R} \H$ the space of global real holomorphic sections of $\mathcal{E} \otimes\L^d$:
\begin{equation}
\label{eq dfn RH0}
\mathbb{R} \H = \left\{ s \in \H \mathrel{}\middle|\mathrel{} c_d \circ s = s \circ c_\mathcal{X} \right\}.
\end{equation}
The restriction of the inner product \eqref{eq definition inner product} to $\mathbb{R} \H$ makes it into a Euclidean space of real dimension $N_d$.
\begin{rem}
\label{rem inner product}
Note that, even when we consider real sections restricted to $M$, the inner product is defined by integrating on the whole complex manifold $\mathcal{X}$.
\end{rem}
\subsection{Random submanifolds}
\label{subsec random submanifolds}
This section is concerned with the definition of the random submanifolds we consider and the related random variables.
Let $d \in \mathbb{N}$ and $s \in \mathbb{R} \H$, we denote the real zero set of $s$ by $Z_s=s^{-1}(0) \cap M$. If the restriction of $s$ to $M$ vanishes transversally, then $Z_s$ is a smooth submanifold of codimension $r$ of $M$. In this case, we denote by $\rmes{s}$ the Riemannian measure on $Z_s$ induced by $g$, seen as a Radon measure on $M$. Note that this includes the case where $Z_s$ is empty.
Recall the following facts, that we already discussed in \cite{Let2016}.
\begin{dfn}[see \cite{Nic2015}]
\label{dfn 0 ample}
We say that $\mathbb{R} \H$ is \emph{$0$-ample} if, for any $x \in M$, the evaluation map
\begin{equation}
\label{eq dfn 0 ample}
\begin{array}{rccc}
\ev_x^d:& \mathbb{R} \H & \longrightarrow &\mathbb{R} \left(\mathcal{E} \otimes \L^d\right)_x\\
& s & \longmapsto & s(x)
\end{array}
\end{equation}
is surjective.
\end{dfn}
\begin{lem}[see \cite{Let2016}, cor.~3.10]
\label{lem dfn d1}
There exists $d_1 \in \mathbb{N}$, depending only on $\mathcal{X}$, $\mathcal{E}$ and $\L$, such that for all $d \geq d_1$, $\mathbb{R} \H$ is $0$-ample.
\end{lem}
\begin{lem}[see \cite{Let2016}, section~2.6]
\label{lem 0 ample implies nice}
If $\mathbb{R} \H$ is $0$-ample, then for almost every section $s \in \mathbb{R} \H$ (for the Lebesgue measure), the restriction of $s$ to $M$ vanishes transversally.
\end{lem}
From now on, we only consider the case $d \geq d_1$, so that $\rmes{s}$ is a well-defined measure on $M$ for almost every $s \in \mathbb{R} \H$. Let $s_d$ be a standard Gaussian vector in $\mathbb{R} \H$, that is $s_d$ is a random vector whose distribution admits the density:
\begin{equation}
\label{eq Gaussian density}
s \mapsto \frac{1}{\sqrt{2\pi}^{N_d}} \exp\left(-\frac{1}{2}\Norm{s}^2\right)
\end{equation}
with respect to the normalized Lebesgue measure on $\mathbb{R} \H$. Here $\Norm{\cdot}$ is the norm associated with the Euclidean inner product \eqref{eq definition inner product}. Then $Z_{s_d}$ is almost surely a submanifold of codimension $r$ of $M$ and $\rmes{s_d}$ is almost surely a random positive Radon measure on $M$. To simplify notations, we set $Z_d = Z_{s_d}$ and $\rmes{d}=\rmes{s_d}$. For more details concerning Gaussian vectors, we refer to \cite[appendix~A]{Let2016} and the references therein.
Let $\phi \in \mathcal{C}^0(M)$, for every $s \in \mathbb{R} \H$ vanishing transversally, we set
\begin{equation}
\label{eq def linear stats}
\prsc{\rmes{s}}{\phi} = \int_{x \in Z_s} \phi(x) \rmes{s}.
\end{equation}
Such a $\phi$ will be refered to as a \emph{test-function}. Following \cite{SZ2010}, we call \emph{linear statistic} of degree $d$ associated with $\phi$ the real random variable $\prsc{\rmes{d}}{\phi}$.
\subsection{The correlation kernel}
\label{subsec correlation kernel}
Let $d \in \mathbb{N}$, then $(s_d(x))_{x \in \mathcal{X}}$ is a smooth centered Gaussian field on $\mathcal{X}$. As such, it is characterized by its correlation kernel. In this section, we recall that the correlation kernel of $s_d$ equals the Bergman kernel of $\mathcal{E} \otimes \L^d$. This is now a well-known fact (see \cite{BSZ2000a,GW2014b,SZ1999,SZ2010}) and was already used by the author in \cite{Let2016}.
Let us first recall some facts about random vectors (see for example \cite[appendix~A]{Let2016}). In this paper, we only consider centered random vectors (that is their expectation vanishes), so we give the following definitions in this restricted setting. Let $X_1$ and $X_2$ be centered random vectors taking values in Euclidean (or Hermitian) vector spaces $V_1$ and $V_2$ respectively, then we define their \emph{covariance operator} as:
\begin{equation}
\label{eq def covariance operator}
\cov{X_1}{X_2} : v \longmapsto \esp{X_1 \prsc{v}{X_2}}
\end{equation}
from $V_2$ to $V_1$. For every $v \in V_2$, we set $v^*=\prsc{\cdot}{v} \in V_2^*$. Then $\cov{X_1}{X_2} = \esp{X_1 \otimes X_2^*}$ is an element of $V_1 \otimes V_2^*$. The \emph{variance operator} of a centered random vector $X \in V$ is defined as $\var{X} = \cov{X}{X} = \esp{X \otimes X^*} \in V \otimes V^*$. We denote by $X \sim \mathcal{N}(\Lambda)$ the fact that $X$ is a centered Gaussian vector with variance operator $\Lambda$. Finally, we say that $X \in V$ is a \emph{standard} Gaussian vector if $X \sim \mathcal{N}(\Id)$, where $\Id$ is the identity operator on $V$. A standard Gaussian vector admits the density~\eqref{eq Gaussian density} with respect to the normalized Lebesgue measure on $V$.
Recall that $(\mathcal{E}\otimes \mathcal{L}^d) \boxtimes (\mathcal{E}\otimes \mathcal{L}^d)^*$ stands for the bundle $P_1^\star\left(\mathcal{E}\otimes \mathcal{L}^d\right) \otimes P_2^\star \left(\left(\mathcal{E}\otimes \mathcal{L}^d\right)^*\right)$ over $\mathcal{X}\times \mathcal{X}$, where $P_1$ (resp.~$P_2$) denotes the projection from $\mathcal{X} \times \mathcal{X}$ onto the first (resp.~second) factor. The covariance kernel of $(s_d(x))_{x \in \mathcal{X}}$ is the section of $(\mathcal{E}\otimes \mathcal{L}^d) \boxtimes (\mathcal{E}\otimes \mathcal{L}^d)^*$ defined by:
\begin{equation}
\label{eq def covariance kernel}
(x,y) \mapsto \cov{s_d(x)}{s_d(y)} = \esp{s_d(x) \otimes s_d(y)^*}.
\end{equation}
The orthogonal projection from $\mathbb{R} \Gamma(\mathcal{E} \otimes \L^d)$ onto $\mathbb{R} \H$ admits a Schwartz kernel (see \cite[thm.~B.2.7]{MM2007}). That is, there exists a unique section $E_d$ of $(\mathcal{E}\otimes \mathcal{L}^d) \boxtimes (\mathcal{E}\otimes \mathcal{L}^d)^*$ such that, for any $s \in \mathbb{R} \Gamma(\mathcal{E} \otimes \L^d)$, the projection of $s$ onto $\mathbb{R} \H$ is given by:
\begin{equation}
\label{eq def Schwartz kernel}
x \longmapsto \int_{y \in \mathcal{X}} E_d(x,y)\left(s(y)\right) \dmesure\! V_\mathcal{X}.
\end{equation}
This section is called the \emph{Bergman kernel} of $\mathcal{E} \otimes \L^d$. Note that $E_d$ is also the Schwartz kernel of the orthogonal projection from $\Gamma(\mathcal{E}\otimes\L^d)$ onto $\H$, for the Hermitian inner product \eqref{eq definition inner product}.
\begin{prop}
\label{prop covariance equals Bergman}
Let $d \in \mathbb{N}$ and let $s_d$ be a standard Gaussian vector in $\mathbb{R} \H$. Then, for all $x$ and $y \in \mathcal{X}$, we have:
\begin{equation}
\label{eq covariance equals Bergman}
\cov{s_d(x)}{s_d(y)} = \esp{s_d(x) \otimes s_d(y)^*} = E_d(x,y).
\end{equation}
\end{prop}
\begin{proof}
We will prove that $(x,y)\mapsto \esp{s_d(x)\otimes s_d(y)^*}$ is the kernel of the orthogonal projection onto $\mathbb{R} \H$, i.e.~satisfies \eqref{eq def Schwartz kernel}. Let $s \in \mathbb{R} \Gamma(\mathcal{E} \otimes \L^d)$, then
\begin{equation*}
\int_{y \in \mathcal{X}} \esp{s_d(x)\otimes s_d(y)^*}\left(s(y)\right) \dmesure\! V_\mathcal{X} = \esp{s_d(x) \int_{y\in \mathcal{X}} s_d(y)^*(s(y))\dmesure\! V_\mathcal{X}} = \esp{s_d(x)\prsc{s}{s_d}}.
\end{equation*}
If $s$ is orthogonal to $\mathbb{R} \H$ this quantity equals $0$. If $s \in \mathbb{R} \H$ then
\begin{equation*}
\esp{s_d(x) \prsc{s}{s_d}} = \esp{\ev_x^d(s_d) s_d^*(s)} = \ev_x^d \left(\esp{s_d \otimes s_d^*}(s)\right) = \ev_x^d (\var{s_d} s) = \ev_x^d(s) = s(x)
\end{equation*}
since $\var{s_d} = \Id$. Thus, for any $s \in \mathbb{R} \Gamma(\mathcal{E} \otimes \L^d)$, $\esp{s_d(x)\prsc{s}{s_d}}$ is the value at $x$ of the orthogonal projection of $s$ on $\mathbb{R} \H$. Finally, the correlation kernel of $(s_d(x))_{x \in \mathcal{X}}$ satisfies \eqref{eq def Schwartz kernel} and equals $E_d$.
\end{proof}
\begin{rem}
\label{rem Bergman kernel basis}
If $(s_{1,d},\dots,s_{N_d,d})$ is any orthonormal basis of $\mathbb{R} \H$, we have:
\begin{equation}
\label{eq Bergman in coordinates}
E_d : (x,y) \longmapsto \sum_{i=1}^{N_d} s_{i,d}(x) \otimes s_{i,d}(y)^*.
\end{equation}
\end{rem}
\begin{rem}
\label{rem Bergman kernel splitting}
If $\mathcal{E}$ is the trivial bundle $\mathcal{X} \times \mathbb{C}^r \to \mathcal{X}$ then $E_d$ splits as $E_d = \Id \otimes\, e_d$, where $\Id$ is the identity of $\mathbb{C}^r$ and $e_d$ is the Bergman kernel of $\L^d$. There is no such splitting in general.
\end{rem}
\begin{rem}
\label{rem Bergman kernel cov complex}
In a complex setting, $E_d$ is also the covariance kernel of the centered Gaussian field associated with a standard complex Gaussian vector in $\H$.
\end{rem}
The Bergman kernel also describes the distribution of the derivatives of $s_d$. Let $\nabla^d$ denote any connection on $\mathcal{E} \otimes \L^d\to \mathcal{X}$. Then $\nabla^d$ induces a connection $(\nabla^d)^*$ on $(\mathcal{E} \otimes \L^d)^*\to \mathcal{X}$, which is defined for all $\eta \in \Gamma\left(\left(\mathcal{E} \otimes \L^d\right)^*\right)$ by:
\begin{equation}
\label{eq def dual connection}
\forall s \in \Gamma\left(\mathcal{E} \otimes \L^d\right),\ \forall x \in \mathcal{X}, \qquad d_x\prsc{s}{\eta} = \prsc{\nabla^d_x s}{\eta(x)}+\prsc{s(x)}{(\nabla^d)_x^*\eta},
\end{equation}
where $\prsc{\cdot}{\cdot}$ is the duality pairing. Let $s \in \Gamma\left(\mathcal{E} \otimes \L^d\right)$, then $s^\diamond: x \mapsto s(x)^* = \prsc{\cdot}{s(x)}$ defines a smooth section of $\left(\mathcal{E} \otimes \L^d\right)^*$. Note that we use the notation $s^\diamond$ because $s^*$ already denotes $\prsc{\cdot}{s}$ which is a linear form on $\Gamma\left(\mathcal{E} \otimes \L^d\right)$. We want to understand the relation between $(\nabla^d)_x^*s^\diamond : T_x\mathcal{X} \to \left(\mathcal{E} \otimes \L^d\right)_x^*$ and $\left(\nabla^d_xs\right)^*$. Recall that $\left(\nabla^d_xs\right)^* = \prsc{\cdot}{\nabla^d_xs}$, where the inner product is the one on $\left(\mathcal{E} \otimes \L^d\right)_x \otimes T^*_x\mathcal{X}$ induced by $h_d$ and $g_\mathbb{C}$. That is, $\left(\nabla^d_xs\right)^*$ is the adjoint operator of $\nabla^d_xs : T_x \mathcal{X} \to \left(\mathcal{E} \otimes \L^d\right)_x$. In order to get a nice relation, we have to assume that $\nabla^d$ is a metric connection, i.e.~that:
\begin{equation}
\label{eq def metric connection}
\forall s,t \in \Gamma\left(\mathcal{E} \otimes \L^d\right),\ \forall x \in \mathcal{X}, \qquad d_x\prsc{s}{t} = \prsc{\nabla^d_xs}{t(x)} + \prsc{s(x)}{\nabla^d_xt}.
\end{equation}
\begin{lem}
\label{lem dual connection}
Let $\nabla^d$ be a metric connection on $\mathcal{E} \otimes \L^d$, let $s \in \Gamma\left(\mathcal{E} \otimes \L^d\right)$ and let $x \in \mathcal{X}$. Then for all $v \in T_x\mathcal{X}$,
\begin{equation}
\label{eq lem dual connection}
(\nabla^d)^*_x s^\diamond \cdot v = \left(\nabla_x^ds\cdot v\right)^* = v^* \circ \left(\nabla_x^ds\right)^*.
\end{equation}
\end{lem}
\begin{proof}
First, for all $s,t \in \Gamma\left(\mathcal{E} \otimes \L^d\right)$ and all $x \in \mathcal{X}$,
\begin{equation}
\label{eq truc1}
\prsc{t(x)}{s(x)} = \prsc{t(x)}{s(x)^*} = \prsc{t(x)}{s^\diamond(x)}.
\end{equation}
Then, by taking the derivative of~\eqref{eq truc1}, we get that for all $s,t \in \Gamma\left(\mathcal{E} \otimes \L^d\right)$, for all $x \in \mathcal{X}$ and $v \in T_x\mathcal{X}$:
\begin{equation*}
\prsc{t(x)}{\nabla^d_xs\cdot v}+\prsc{\nabla^d_xt\cdot v}{s(x)} = d_x\left(\prsc{t}{s}\right)\cdot v = \prsc{t(x)}{(\nabla^d)^*_xs^\diamond\cdot v}+ \prsc{\nabla^d_xt\cdot v}{s^\diamond(x)}.
\end{equation*}
The first equality comes from the fact that $\nabla^d$ is metric (see~\eqref{eq def metric connection}) and the second from the definition of the dual connection~\eqref{eq def dual connection}. Besides $\prsc{\nabla^d_xt\cdot v}{s^\diamond(x)} =\prsc{\nabla^d_xt\cdot v}{s(x)}$, hence for all $s \in \Gamma\left(\mathcal{E} \otimes \L^d\right)$ and all $x \in \mathcal{X}$ we have:
\begin{equation*}
\forall v \in T_x\mathcal{X}, \qquad (\nabla^d)^*_xs^\diamond \cdot v = \left(\nabla^d_xs\cdot v\right)^*.
\end{equation*}
Recall that $\left(\nabla_x^ds\right)^*$ is the adjoint of $\nabla_x^ds$. Hence for all $v \in T_x\mathcal{X}$ and all $\zeta \in \left(\mathcal{E} \otimes \L^d\right)_x$,
\begin{equation*}
\prsc{\zeta}{\nabla^d_xs\cdot v} = \prsc{\left(\nabla^d_xs\right)^*\zeta}{v} = v^* \circ \left(\nabla^d_xs\right)^* (\zeta),
\end{equation*}
which proves the second equality in~\eqref{eq lem dual connection}.
\end{proof}
\begin{rem}
\label{rem metric connection}
Conversely, one can show that a connection satisfying the first equality in eq.~\eqref{eq lem dual connection} for every $s,x$ and $v$ is metric.
\end{rem}
From now on, we assume that $\nabla^d$ is metric. Then $\nabla^d$ induces a natural connection $\nabla_1^d$ on $P_1^\star(\mathcal{E} \otimes \L^d) \to \mathcal{X} \times \mathcal{X}$ whose partial derivatives are: $\nabla^d$ with respect to the first variable, and the trivial connection with respect to the second. Similarly, $(\nabla^d)^*$ induces a connection $\nabla_2^d$ on $P_2^\star\left((\mathcal{E} \otimes \L^d)^*\right)$ and $\nabla_1^d \otimes \Id + \Id \otimes \nabla_2^d$ is a connection on $(\mathcal{E} \otimes \L^d) \boxtimes (\mathcal{E} \otimes \L^d)^*$. We denote by $\partial_x$ (resp. $\partial_y$) its partial derivative with respect to the first (resp. second) variable. By taking partial derivatives in \eqref{eq covariance equals Bergman}, we get the following.
\begin{cor}
\label{cor variance of the 1-jet}
Let $d \in \mathbb{N}$, let $\nabla^d$ be a metric connection on $\mathcal{E} \otimes \L^d$ and let $s_d$ be a standard Gaussian vector in $\mathbb{R} \H$. Then, for all $x$ and $y \in \mathcal{X}$, for all $(v,w) \in T_x\mathcal{X} \times T_y\mathcal{X}$, we have:
\begin{align}
\label{cov bergman 10}
\cov{\nabla^d_xs \cdot v}{s(y)} &= \esp{\left(\nabla^d_xs\cdot v\right) \otimes s(y)^*} = \partial_x E_d(x,y)\cdot v,\\
\label{cov bergman 01}
\cov{s(x)}{\nabla^d_ys\cdot w} &= \esp{s(x) \otimes \left(\nabla^d_ys\cdot w\right)^*} = \partial_y E_d(x,y)\cdot w,\\
\label{cov bergman 11}
\cov{\nabla^d_xs\cdot v}{\nabla^d_ys\cdot w} &= \esp{\left(\nabla^d_xs\cdot v\right) \otimes \left(\nabla^d_ys\cdot w\right)^*} = \partial_x\partial_y E_d(x,y)\cdot (v,w).
\end{align}
\end{cor}
\begin{proof}
The first equality of each line is simply the definition of the covariance operator. By applying $\partial_x$ to~\eqref{eq covariance equals Bergman} we get:
\begin{equation*}
\esp{\left(\nabla^d_xs\right) \otimes s(y)^*} = \partial_x E_d(x,y),
\end{equation*}
which proves~\eqref{cov bergman 10}. We can rewrite~\eqref{eq covariance equals Bergman} as: $\forall x,y \in \mathcal{X}$, $E_d(x,y) = \esp{s(x) \otimes s^\diamond(y)}$. By applying $\partial_y$ to this equality, we get:
\begin{equation*}
\esp{s(x) \otimes \left(\nabla^d\right)^*_y s^\diamond} = \partial_y E_d(x,y).
\end{equation*}
Then we apply this operator to $w \in T_y\mathcal{X}$, and we obtain~\eqref{cov bergman 01} by Lemma~\ref{lem dual connection}. The proof of~\eqref{cov bergman 11} is similar.
\end{proof}
We would like to write that $\partial_y E_d(x,y)$ is $\cov{s(x)}{\nabla^d_ys}=\esp{s(x) \otimes \left(\nabla^d_ys\right)^*}$. Unfortunately, this can not be true since
\begin{align*}
\partial_y E_d(x,y) &\in T^*_y\mathcal{X} \otimes \left(\mathcal{E} \otimes \L^d\right)_x \otimes \left(\mathcal{E} \otimes \L^d\right)_y^*\\
\intertext{while} \esp{s(x) \otimes \left(\nabla^d_ys\right)^*} &\in T_y\mathcal{X} \otimes \left(\mathcal{E} \otimes \L^d\right)_x \otimes \left(\mathcal{E} \otimes \L^d\right)_y^*.
\end{align*}
Let $\partial_y^\sharp E_d(x,y) \in T_y\mathcal{X} \otimes \left(\mathcal{E} \otimes \L^d\right)_x \otimes \left(\mathcal{E} \otimes \L^d\right)_y^*$ be defined by:
\begin{equation}
\label{eq def sharp}
\forall w \in T_y\mathcal{X}, \qquad \partial_y^\sharp E_d(x,y) \cdot w^* = \partial_y E_d(x,y) \cdot w.
\end{equation}
Similarly, let $\partial_x\partial_y^\sharp E_d(x,y) \in T^*_x\mathcal{X} \otimes T_y\mathcal{X} \otimes \left(\mathcal{E} \otimes \L^d\right)_x \otimes \left(\mathcal{E} \otimes \L^d\right)_y^*$ be defined by:
\begin{equation}
\label{eq def sharp 2}
\forall (v,w) \in T_x\mathcal{X} \times T_y\mathcal{X}, \qquad \partial_x\partial_y^\sharp E_d(x,y) \cdot (v,w^*) = \partial_x\partial_y E_d(x,y)\cdot (v,w).
\end{equation}
Then by Lemma~\ref{lem dual connection} and Corollary~\ref{cor variance of the 1-jet}, we have the following.
\begin{cor}
\label{cor variance of the 1-jet sharp}
Let $d \in \mathbb{N}$, let $\nabla^d$ be a metric connection on $\mathcal{E} \otimes \L^d$ and let $s_d$ be a standard Gaussian vector in $\mathbb{R} \H$. Then, for all $x$ and $y \in \mathcal{X}$, we have:
\begin{align}
\label{cov bergman 10 sharp}
\cov{\nabla^d_xs}{s(y)} &= \esp{\nabla^d_xs \otimes s(y)^*} = \partial_x E_d(x,y),\\
\label{cov bergman 01 sharp}
\cov{s(x)}{\nabla^d_ys} &= \esp{s(x) \otimes \left(\nabla^d_ys\right)^*} = \partial_y^\sharp E_d(x,y),\\
\label{cov bergman 11 sharp}
\cov{\nabla^d_xs}{\nabla^d_ys} &= \esp{\nabla^d_xs \otimes \left(\nabla^d_ys\right)^*} = \partial_x\partial_y^\sharp E_d(x,y).
\end{align}
\end{cor}
\section{Estimates for the Bergman kernel}
\label{sec estimates for the bergman kernel}
The goal of this section is to recall the estimates we need for the Bergman kernel. Most of what follows can be found in \cite{MM2007}, with small additions from \cite{MM2013} and \cite{MM2015}. The first to use this kind of estimates in a random geometry context were Shiffman and Zelditch \cite{SZ1999}. They used the estimates from \cite{Zel1998} for the related Szegö kernel (see also \cite{BSZ2000a,SZ2008}). Catlin \cite{Cat1999} proved similar estimates for the Bergman kernel independently.
In order to state the near-diagonal estimates for the Bergman kernel, we first need to choose preferred charts on $\mathcal{X}$, $\mathcal{E}$ and $\L$ around any point in $M$. This is done in Section~\ref{subsec real normal trivialization}. Unlike our main reference \cite{MM2007}, we are only concerned with a neighborhood of the real locus of $\mathcal{X}$, but we need to check that these charts are well-behaved with respect to the real structures. Sections~\ref{subsec near diagonal estimates}, \ref{subsec diagonal estimates} and \ref{subsec far off diagonal estimates} state respectively near-diagonal, diagonal and far off-diagonal estimates for $E_d$.
\subsection{Real normal trivialization}
\label{subsec real normal trivialization}
In this section, we define preferred local trivializations for $\mathcal{E}$ and $\L$ around any point in $M$. We also prove that these trivializations are compatible with the real and metric structures.
Let $R >0$ be such that the injectivity radius of $\mathcal{X}$ is larger than $2 R$. Let $x_0 \in M$, then the exponential map $\exp_{x_0} : T_{x_0}\mathcal{X} \to \mathcal{X}$ at $x_0$ is a diffeomorphism from the ball $B_{T_{x_0}\mathcal{X}}(0,2R) \subset T_{x_0}\mathcal{X}$ to the geodesic ball $B_\mathcal{X}(x_0,2R) \subset \mathcal{X}$. Note that this diffeomorphism is not biholomorphic in general.
\begin{ntn}
\label{ntn open ball}
Here and in the sequel, we always denote by $B_A(a,R)$ the open ball of center $a$ and radius $R>0$ in the metric space $A$.
\end{ntn}
Since $c_\mathcal{X}$ is an isometry (see Sect.~\ref{subsec general setting}), we have that $c_\mathcal{X} \circ \exp_{x_0} = \exp_{x_0} \circ \, d_{x_0}c_\mathcal{X}$. Then $\exp_{x_0}$ sends $T_{x_0}M = \ker\left(d_{x_0}c_\mathcal{X} - \Id \right)$ to $M$ and agrees on $T_{x_0}M$ with the exponential map at $x_0$ in $(M,g)$. By restriction, we get a diffeomorphism from $B_{T_{x_0}M}(0,2R) \subset T_{x_0}M$ to the geodesic ball $B_M(x_0,2R) \subset M$. Moreover, on $B_{T_{x_0}\mathcal{X}}(0,2R)$ we have:
\begin{equation}
\label{eq real equivariance exp}
d_{x_0}c_\mathcal{X} = (\exp_{x_0})^{-1} \circ c_\mathcal{X} \circ \exp_{x_0}.
\end{equation}
We say that $\exp_{x_0}$ defines a \emph{real normal chart} about $x_0$.
Since $i\cdot T_{x_0}M = \ker\left(d_{x_0}c_\mathcal{X} + \Id \right)$, we have $T_{x_0}\mathcal{X} = T_{x_0}M \oplus i\cdot T_{x_0}M$. Note that $T_{x_0}M$ and $i\cdot T_{x_0}M$ are orthogonal for $g_{x_0}$, since these are distinct eigenspaces of an isometric involution. Moreover, we know from Sect.~\ref{subsec general setting} that $c_\mathcal{X}^\star g_\mathbb{C} = \overline{g_\mathbb{C}}$. This implies that $(g_\mathbb{C})_{x_0}$ takes real values on $T_{x_0}M\times T_{x_0}M$, i.e.~the restrictions to $T_{x_0}M$ of $(g_\mathbb{C})_{x_0}$ and $g_{x_0}$ are equal. Thus, $(g_\mathbb{C})_{x_0}$ is the sesquilinear extension of $g_{x_0}$ restricted to $T_{x_0}M$. Let $\mathcal{I}$ be an isometry from $T_{x_0}M$ to $\mathbb{R}^n$ with its standard Euclidean structure, $\mathcal{I}$ extends as a $\mathbb{C}$-linear isometry $\mathcal{I}_\mathbb{C} : T_{x_0}\mathcal{X} \to \mathbb{C}^n$, such that $\mathcal{I}_\mathbb{C} \circ d_{x_0}c_\mathcal{X} \circ \mathcal{I}_\mathbb{C}^{-1}$ is the complex conjugation in $\mathbb{C}^n$. Thus, $\exp_{x_0} \circ\, \mathcal{I}_\mathbb{C}^{-1}:B_{\mathbb{C}^n}(0,2R) \to B_{\mathcal{X}}(x_0,2R)$ defines normal coordinates that induce normal coordinates $B_{\mathbb{R}^n}(0,2R) \to B_M(x_0,2R)$ and such that $\mathcal{I}_\mathbb{C} \circ (\exp_{x_0})^{-1}\circ c_\mathcal{X} \circ \exp_{x_0} \circ\, \mathcal{I}_\mathbb{C}^{-1}$ is the complex conjugation in $\mathbb{C}^n$. Such coordinates are called \emph{real normal coordinates} about $x_0$.
We can now trivialize $\mathcal{E}$ over $B_\mathcal{X}(x_0,2R)$. Let $\nabla^\mathcal{E}$ denote the Chern connection of $\mathcal{E}$. We identify the fiber at $\exp_{x_0}(z) \in B_{\mathcal{X}}(x_0,2R)$ with $\mathcal{E}_{x_0}$, by parallel transport with respect to $\nabla^\mathcal{E}$ along the geodesic from $x_0$ to $\exp_{x_0}(z)$, defined by $t \mapsto \exp_{x_0}(tz)$ from $[0,1]$ to $\mathcal{X}$ (cf.~\cite[sect.~1.6]{MM2007} and~\cite{MM2013}). This defines a bundle map $\varphi_{x_0} : B_{T_{x_0}\mathcal{X}}(0,2R) \times \mathcal{E}_{x_0} \to \mathcal{E}_{/ B_\mathcal{X}(x_0,2R)}$ that covers $\exp_{x_0}$. We say that $\varphi_{x_0}$ is the \emph{real normal trivialization} of $\mathcal{E}$ over $B_\mathcal{X}(x_0,2R)$.
Since $x_0 \in M$, $c_\mathcal{E}(\mathcal{E}_{x_0})=\mathcal{E}_{x_0}$ and we denote by $c_{\mathcal{E},x_0}$ the restriction of $c_\mathcal{E}$ to $\mathcal{E}_{x_0}$. Then $(d_{x_0}c_\mathcal{X},c_{\mathcal{E},x_0})$ is a real structure on $B_{T_{x_0}\mathcal{X}}(0,2R) \times \mathcal{E}_{x_0}$ compatible with the real structure on $B_{T_{x_0}\mathcal{X}}(0,2R)$. We want to check that $\varphi_{x_0}$ is well-behaved with respect to the real structures, i.e.~that for all $z \in B_{T_{x_0}\mathcal{X}}(0,2R)$ and $\zeta^0 \in \mathcal{E}_{x_0}$,
\begin{equation}
\label{eq equivariance trivialization real structures}
c_\mathcal{E}(\varphi_{x_0}(z,\zeta^0))= \varphi_{x_0}\left(d_{x_0}c_\mathcal{X} \cdot z, c_{\mathcal{E},x_0}(\zeta^0)\right).
\end{equation}
This will be a consequence of Lemma~\ref{lem Chern connection is real} below.
\begin{dfn}
\label{dfn real connection}
Let $\mathcal{E} \to \mathcal{X}$ be a holomorphic vector bundle equipped with compatible real structures $c_\mathcal{E}$ and $c_\mathcal{X}$ and let $\nabla$ be a connection on $\mathcal{E}$, we say that $\nabla$ is a \emph{real connection} if for every section $s \in \Gamma(\mathcal{E})$ we have:
\begin{equation*}
\forall x \in \mathcal{X}, \qquad \nabla_x \left( c_\mathcal{E} \circ s \circ c_\mathcal{X}\right) = c_\mathcal{E} \circ \nabla_{c_\mathcal{X}(x)} s \circ d_xc_\mathcal{X}.
\end{equation*}
\end{dfn}
\begin{rem}
\label{rem real connection in a real direction}
Let $x \in M$, $v \in T_xM$ and $s \in \mathbb{R} \Gamma(\mathcal{E})$. If $\nabla$ is a real connection on $\mathcal{E}$, then $\nabla_xs\cdot v \in \mathbb{R} \mathcal{E}_x$. Indeed,
\begin{equation*}
\nabla_xs\cdot v = \nabla_{c_\mathcal{X}(x)}s \circ d_xc_\mathcal{X}\cdot v = c_\mathcal{E} \left(\nabla_x \left( c_\mathcal{E} \circ s \circ c_\mathcal{X}\right)\cdot v\right)= c_\mathcal{E}\left(\nabla_xs\cdot v\right).
\end{equation*}
\end{rem}
\begin{lem}
\label{lem Chern connection is real}
Let $\mathcal{E} \to \mathcal{X}$ be a holomorphic vector bundle equipped with compatible real structures $c_\mathcal{E}$ and $c_\mathcal{X}$ and a real Hermitian metric $h_\mathcal{E}$. Then, the Chern connection $\nabla^\mathcal{E}$ of $\mathcal{E}$ is real.
\end{lem}
\begin{proof}
Since $c_\mathcal{E}$ and $c_\mathcal{X}$ are involutions and $(d_xc_\mathcal{X})^{-1} = d_{c_\mathcal{X}(x)}c_\mathcal{X}$, we need to check that
\begin{equation}
\label{eq Chern connection is real}
\forall s \in \Gamma(\mathcal{E}), \forall x \in \mathcal{X} \qquad \nabla^\mathcal{E}_x s = c_\mathcal{E} \circ \nabla^\mathcal{E}_{c_\mathcal{X}(x)}\left( c_\mathcal{E} \circ s \circ c_\mathcal{X} \right) \circ d_xc_\mathcal{X}.
\end{equation}
Let $\tilde{\nabla}$ be defined by $\tilde{\nabla}_x s = c_\mathcal{E} \circ \nabla^\mathcal{E}_{c_\mathcal{X}(x)}\left( c_\mathcal{E} \circ s \circ c_\mathcal{X} \right) \circ d_xc_\mathcal{X}$, for all $s \in \Gamma(\mathcal{E})$ and $x \in \mathcal{X}$. Then $\tilde{\nabla}$ is a connection on $\mathcal{E}$ and it is enough to check that it is compatible with both the metric and the complex structure. Indeed, in this case $\tilde{\nabla} = \nabla^\mathcal{E}$ by unicity of the Chern connection, which proves \eqref{eq Chern connection is real}.
Let us check that $\tilde{\nabla}$ satisfies Leibniz' rule. Let $s \in \Gamma(\mathcal{E})$ and $f:\mathcal{X} \to \mathbb{C}$. We have:
\begin{align*}
\tilde{\nabla}_x(fs)&= c_\mathcal{E} \circ \nabla^\mathcal{E}_{c_\mathcal{X}(x)} \left( (\overline{f\circ c_\mathcal{X}}) (c_\mathcal{E} \circ s \circ c_\mathcal{X}) \right) \circ d_xc_\mathcal{X}\\
&= c_\mathcal{E} \circ \left(\overline{f(x)}\nabla^\mathcal{E}_{c_\mathcal{X}(x)}\left( c_\mathcal{E} \circ s \circ c_\mathcal{X} \right) + d_{c_\mathcal{X}(x)}(\overline{f\circ c_\mathcal{X}})\otimes c_\mathcal{E}(s(x)) \right) \circ d_xc_\mathcal{X}\\
&= f(x) \tilde{\nabla}_xs + d_xf \otimes s(x).
\end{align*}
Since $\nabla^\mathcal{E}$ is the Chern connection, its anti-holomorphic part is $\overline{\partial}^\mathcal{E}$. Then, $d_xc_\mathcal{X}$ and $c_\mathcal{E}$ being anti-linear (resp. fiberwise), the anti-linear part of $\tilde{\nabla}_xs$ equals $c_\mathcal{E} \circ \overline{\partial}^\mathcal{E}_{c_\mathcal{X}(x)}\left(c_\mathcal{E} \circ s \circ c_\mathcal{X} \right) \circ d_xc_\mathcal{X}$. By computing in a local holomorphic frame, one can check that:
\begin{equation*}
\forall s \in \Gamma(\mathcal{E}), \forall x \in \mathcal{X}, \qquad c_\mathcal{E} \circ \overline{\partial}^\mathcal{E}_{c_\mathcal{X}(x)}\left(c_\mathcal{E} \circ s \circ c_\mathcal{X} \right) \circ d_xc_\mathcal{X}=\overline{\partial}^\mathcal{E}_xs.
\end{equation*}
Thus, $\tilde{\nabla}$ is compatible with the complex structure. Finally, we check the compatibility with the metric structure. Let $s,t \in \Gamma(\mathcal{E})$ and $x \in \mathcal{X}$, since $h_\mathcal{E}=c_\mathcal{E}^\star(\overline{h_\mathcal{E}})$ we have:
\begin{align*}
d_x(h_\mathcal{E}(s,t)) &= d_x\left(\overline{h_\mathcal{E}}(c_\mathcal{E} \circ s, c_\mathcal{E} \circ t)\right) = d_{c_\mathcal{X}(x)}\left(\overline{h_\mathcal{E}}(c_\mathcal{E} \circ s \circ c_\mathcal{X}, c_\mathcal{E} \circ t \circ c_\mathcal{X})\right) \circ d_x c_\mathcal{X}\\
&\begin{aligned}=\overline{h_\mathcal{E}}\left(\nabla^\mathcal{E}_{c_\mathcal{X}(x)}(c_\mathcal{E} \circ s \circ c_\mathcal{X}),c_\mathcal{E}(t(x))\right)&\circ d_xc_\mathcal{X}\\+&\overline{h_\mathcal{E}}\left(c_\mathcal{E}(s(x)),\nabla^\mathcal{E}_{c_\mathcal{X}(x)}(c_\mathcal{E} \circ t \circ c_\mathcal{X})\right)\circ d_x c_\mathcal{X}\end{aligned}\\
&\begin{aligned}= h_\mathcal{E}\left(c_\mathcal{E} \circ \nabla^\mathcal{E}_{c_\mathcal{X}(x)}(c_\mathcal{E} \circ s \circ c_\mathcal{X}),t(x)\right) &\circ d_xc_\mathcal{X}\\ + &h_\mathcal{E}\left(s(x),c_\mathcal{E} \circ\nabla^\mathcal{E}_{c_\mathcal{X}(x)}(c_\mathcal{E} \circ s \circ c_\mathcal{X})\right)\circ d_x c_\mathcal{X}\end{aligned}\\
&= h_\mathcal{E}\left(\tilde{\nabla}_xs,t(x)\right)+ h_\mathcal{E}\left(s(x),\tilde{\nabla}_xt\right).\qedhere
\end{align*}
\end{proof}
Let us now prove \eqref{eq equivariance trivialization real structures}. Let $z \in B_{T_{x_0}\mathcal{X}}(0,2R)$, let $\zeta^0 \in \mathcal{E}_{x_0}$ and let $\zeta : B_\mathcal{X}(x_0,2R) \to \mathcal{E}$ be the section defined by $\zeta : x\mapsto \varphi_{x_0}\left((\exp_{x_0})^{-1}(x),\zeta^0\right)$. We denote by $\gamma : [0,1] \mapsto \mathcal{X}$ the geodesic $t \mapsto \exp_{x_0}(tz)$. We have for all $t \in [0,1]$, $\zeta(\gamma(t))=\varphi_{x_0}(tz,\zeta^0)$ and, by the definition of $\varphi_{x_0}$, we have:
\begin{equation}
\label{eq parallel transport}
\forall t \in [0,1],\qquad \nabla^\mathcal{E}_{\gamma(t)}\zeta\cdot \gamma'(t)= 0.
\end{equation}
Let us denote $\tilde{\zeta}=c_\mathcal{E} \circ \zeta \circ c_\mathcal{X}$ and $\overline{\gamma}=c_\mathcal{X} \circ \gamma$. Since $\nabla^\mathcal{E}$ is real, \eqref{eq parallel transport} implies that for all $t \in [0,1]$,
\begin{equation}
\label{eq parallel transport 2}
\nabla^\mathcal{E}_{\overline{\gamma}(t)}\tilde{\zeta} \cdot \overline{\gamma}'(t) = \nabla^\mathcal{E}_{c_\mathcal{X}(\gamma(t))}\tilde{\zeta} \circ d_{\gamma(t)}(c_\mathcal{X}) \cdot \gamma'(t) = c_\mathcal{E} \circ \nabla^\mathcal{E}_{\gamma(t)} \zeta \cdot \gamma'(t)= 0.
\end{equation}
Since $c_\mathcal{X}$ is an isometry, $\overline{\gamma}$ is a geodesic. More precisely, $\overline{\gamma}:t \mapsto \exp_{x_0}(td_{x_0}c_\mathcal{X} \cdot z)$. Besides, $\tilde{\zeta}(x_0)=c_\mathcal{E}(\zeta(x_0))=c_{\mathcal{E},x_0}(\zeta^0)$. Then by \eqref{eq parallel transport 2}, for all $t\in [0,1]$,
\begin{equation*}
\varphi_{x_0}\left(td_{x_0}c_\mathcal{X} \cdot z,c_{\mathcal{E},x_0}(\zeta^0)\right)=\varphi_{x_0}\left(td_{x_0}c_\mathcal{X} \cdot z,\tilde{\zeta}(x_0)\right)=\tilde{\zeta}(\overline{\gamma}(t)).
\end{equation*}
Finally, we get \eqref{eq equivariance trivialization real structures} for $t=1$:
\begin{equation*}
\varphi_{x_0}\left(d_{x_0}c_\mathcal{X} \cdot z,c_{\mathcal{E},x_0}(\zeta^0)\right)= \tilde{\zeta}(\overline{\gamma}(1))=c_\mathcal{E}(\zeta(\gamma(1)))=c_\mathcal{E}\left(\varphi_{x_0}(z,\zeta^0)\right).
\end{equation*}
Recall that $\mathbb{R} \mathcal{E}$ is the set of fixed points of $c_\mathcal{E}$. Then $\mathbb{R} \mathcal{E}$ is naturally a rank $r$ real vector bundle over $M$, as a subbundle of $\mathcal{E}_{/M}$. Let $\zeta^0 \in \mathbb{R} \mathcal{E}_{x_0}$, and $\zeta : x\mapsto \varphi_{x_0}\left((\exp_{x_0})^{-1}(x),\zeta^0\right)$ then, for all $x \in B_\mathcal{X}(x_0,2R)$,
\begin{align*}
c_\mathcal{E} \circ \zeta \circ c_\mathcal{X} (x) &= c_\mathcal{E} \circ \varphi_{x_0}\left((\exp_{x_0})^{-1}(c_\mathcal{X}(x)),\zeta^0\right)\\
&= c_\mathcal{E} \circ \varphi_{x_0}\left(d_{x_0}c_\mathcal{X} \circ(\exp_{x_0})^{-1}(x),\zeta^0\right)\\
&= \varphi_{x_0}\left((\exp_{x_0})^{-1}(x),c_{\mathcal{E},x_0}(\zeta^0)\right)\\
&= \zeta(x).
\end{align*}
Hence, $\zeta$ is a real local section of $\mathcal{E}$ and in particular, $\forall x \in M$, $\zeta(x) \in \mathbb{R}\mathcal{E}_x$. This shows that $\varphi_{x_0}$ induces, by restriction, a bundle map $B_{T_{x_0}M}(0,2R) \times \mathbb{R} \mathcal{E}_{x_0} \to \mathbb{R}\mathcal{E}_{/B_M(x_0,2R)}$ that covers the restriction of $\exp_{x_0}$ to $B_{T_{x_0}M}(0,2R)$.
Let $(\zeta_1^0,\dots,\zeta_r^0)$ be an orthonormal basis of $\mathbb{R} \mathcal{E}_{x_0}$. Since $\mathbb{R} \mathcal{E}_{x_0} = \ker\left(c_{\mathcal{E},x_0}-\Id\right)$ and $c_{\mathcal{E},x_0}$ is $\mathbb{C}$-anti-linear, we have $\mathcal{E}_{x_0} = \mathbb{R} \mathcal{E}_{x_0}\oplus i \cdot \mathbb{R} \mathcal{E}_{x_0}$. Moreover, since $h_{\mathcal{E},x_0}$ and $c_{\mathcal{E},x_0}$ are compatible, $(\zeta_1^0,\dots,\zeta_r^0)$ is also an orthonormal basis of $\mathcal{E}_{x_0}$. Let $i\in \left\{1,\dots,r\right\}$, we denote by $\zeta_i : B_\mathcal{X}(x_0,2R) \to \mathcal{E}$ the real local section defined by:
\begin{equation*}
\label{eq zeta i}
\zeta_i : x \mapsto \varphi_{x_0}\left((\exp_{x_0})^{-1}(x),\zeta_i^0\right).
\end{equation*}
Then, for every $x \in B_\mathcal{X}(0,2R)$, $(\zeta_1(x),\dots,\zeta_r(x))$ is an orthonormal basis of $\mathcal{E}_x$. Indeed, the sections $\zeta_i$ are obtained by parallel transport for $\nabla^\mathcal{E}$ along geodesics starting at $x_0$, and $\nabla^\mathcal{E}$ is compatible with $h_\mathcal{E}$. Hence, for all $i$ and $j\in \left\{1,\dots,r\right\}$, for all $z \in B_\mathcal{X}(x_0,2R)$,
\begin{multline*}
\label{eq constant along geodesic}
\frac{\dmesure\!}{\dmesure\! t}\left(h_\mathcal{E}(\zeta_i(\exp_{x_0}(tz)),\zeta_j(\exp_{x_0}(tz)))\right) = h_\mathcal{E}\left(\nabla^\mathcal{E}_{\exp_{x_0}(tz)}\zeta_i\circ d_{tz}\exp_{x_0}\cdot z,\zeta_j(\exp_{x_0}(tz)))\right)\\
+h_\mathcal{E}\left(\zeta_i(\exp_{x_0}(tz))),\nabla^\mathcal{E}_{\exp_{x_0}(tz)}\zeta_j\circ d_{tz}\exp_{x_0}\cdot z\right) = 0.
\end{multline*}
The function $x \mapsto h_\mathcal{E}(\zeta_i(x),\zeta_j(x))$ is then constant along geodesics starting at $x_0$, hence on $B_\mathcal{X}(x_0,2R)$. Since $\left(h_\mathcal{E}(\zeta_i(x),\zeta_j(x))\right)_{1\leq i,j\leq r}$ is the identity matrix of size $r$ at $x_0$, $(\zeta_1,\dots,\zeta_r)$ is a smooth unitary frame for $\mathcal{E}$ over $B_\mathcal{X}(0,2R)$. In particular, this shows that the real normal trivialization $\varphi_{x_0}$ is unitary. Since the $\zeta_i$ are real, $(\zeta_1(x),\dots,\zeta_r(x))$ is an orthonormal basis of $\mathbb{R} \mathcal{E}_x$ for all $x \in M$. Hence $(\zeta_1,\dots,\zeta_r)$ is also a smooth orthogonal frame for $\mathbb{R} \mathcal{E}$ over $B_M(0,2R)$. We say that $(\zeta_1,\dots,\zeta_r)$ is a local \emph{real unitary frame}.
Similarly, let $\varphi'_{x_0}$ denote the real normal trivialization of $\L$ over $B_\mathcal{X}(x_0,2R)$. Then any unit vector $\zeta^0_0\in\mathbb{R} \L_{x_0}$ defines a local real unitary frame $\zeta_0$ for $\L$:
\begin{equation*}
\zeta_0 : x \mapsto \varphi'_{x_0}\left((\exp_{x_0})^{-1}(x),\zeta^0_0\right).
\end{equation*}
Then, for any $d \in \mathbb{N}$, $\varphi_{x_0}$ and $\varphi'_{x_0}$ induce a trivialization $\varphi_{x_0} \otimes (\varphi'_{x_0})^d$ of $\mathcal{E} \otimes \L^d$. This trivialization is the real normal trivialization of $\mathcal{E} \otimes \L^d$ over $B_\mathcal{X}(x_0,2R)$, i.e.~it is obtained by parallel transport along geodesics starting at $x_0$ for the Chern connection of $\mathcal{E} \otimes \L^d$. Moreover, a local real unitary frame for $\mathcal{E} \otimes \L^d$ is given by $(\zeta_1 \otimes \zeta_0^d,\dots,\zeta_r \otimes \zeta_0^d)$.
\subsection{Near-diagonal estimates}
\label{subsec near diagonal estimates}
We can now state the near-diagonal estimates of Ma and Marinescu for the Bergman kernel. In the sequel, we fix some $R >0$ such that $2R$ is smaller than the injectivity radius of $\mathcal{X}$. Let $x \in M$, we have a natural real normal chart
\begin{equation*}
\label{eq normal chart 2}
\exp_x\times \exp_x : B_{T_x\mathcal{X}}(0,2R)\times B_{T_x\mathcal{X}}(0,2R) \to B_\mathcal{X}(x,2R) \times B_\mathcal{X}(x,2R).
\end{equation*}
Moreover, the real normal trivialization of $\mathcal{E} \otimes \L^d$ over $B_\mathcal{X}(x,2R)$ (see Section~\ref{subsec real normal trivialization}) induces a trivialization
\begin{equation*}
\label{eq normal trivialization 2}
B_{T_x\mathcal{X}}(0,2R)\times B_{T_x\mathcal{X}}(0,2R) \times \End\left(\left(\mathcal{E} \otimes \L^d\right)_x\right) \simeq \left(\mathcal{E} \otimes \L^d\right) \boxtimes \left(\mathcal{E} \otimes\L^d\right)^*_{/B_\mathcal{X}(x,2R) \times B_\mathcal{X}(x,2R)}
\end{equation*}
that covers $\exp_x\times \exp_x$. This trivialization coincides with the real normal trivialization of $\left(\mathcal{E} \otimes \L^d\right) \boxtimes \left(\mathcal{E} \otimes\L^d\right)^*$ over $B_\mathcal{X}(x,2R) \times B_\mathcal{X}(x,2R)$.
Recall that $\dmesure\! V_\mathcal{X}$ denotes the Riemannian measure on $\mathcal{X}$. When we read this measure in the real normal chart $\exp_x$, it admits a density $\kappa : B_{T_x\mathcal{X}}(0,2R)\to \mathbb{R}_+$ with respect to the normalized Lebesgue measure of $(T_x\mathcal{X},g_x)$. More precisely, we have $\kappa(z) = \sqrt{\det(g_{ij}(z))}$ where $(g_{ij}(z))$ is the matrix of $\left((\exp_x)^\star g\right)_z$, read in any real orthonormal basis of $(T_x\mathcal{X},g_x)$. Since we use normal coordinates and $\mathcal{X}$ is compact, we have
\begin{equation}
\label{eq estimate kappa}
\kappa(z) = 1 + O\!\left(\Norm{z}^2\right)
\end{equation}
where $\Norm{\cdot}$ is induced by $g_x$ and the estimate $O\!\left(\Norm{z}^2\right)$ does not depend on $x$.
Similarly, on the real locus $(M,g)$, $\rmes{M}$ admits a density, in the real normal chart $\exp_x$, with respect to the normalized Lebesgue measure on $(T_xM,g_x)$. This density is:
\begin{equation}
\label{eq density real}
z \longmapsto \det \left(\left(\left(\exp_x^\star g\right)_z\right)_{/T_xM}\right)^\frac{1}{2},
\end{equation}
from $B_{T_xM}(0,2R)$ to $\mathbb{R}_+$. As we already explained in Sect.~\ref{subsec real normal trivialization}, on the real locus, $g_\mathbb{C}$ is the sesquilinear extension of the restriction of $g$ to $TM$. Hence, for all $z \in B_{T_xM}(0,2R)$ we have:
\begin{equation*}
\det \left(\left(\left(\exp_x^\star g\right)_z\right)_{/T_xM}\right)^2 = \det \left(\left(\exp_x^\star g\right)_z\right),
\end{equation*}
which means that the density of $\rmes{M}$ in the chart $\exp_x$ is $\sqrt{\kappa}:B_{T_xM}(0,2R) \to \mathbb{R}_+$.
The following result gives the asymptotic of the Bergman kernel $E_d$ (see Sect.~\ref{subsec correlation kernel}) and its derivatives, read in the real normal trivialization about $x$ of $\left(\mathcal{E} \otimes \L^d\right) \boxtimes \left(\mathcal{E} \otimes\L^d\right)^*$. It was first established by Dai, Liu and Ma in \cite[thm.~4.18']{DLM2006}.
\begin{thm}[Dai--Liu--Ma]
\label{thm Ma Marinescu}
There exists $C'>0$ such that, for any $p \in \mathbb{N}$, there exists $C_p$ such that $\forall k \in \{0,\dots,p\}$, $\forall d \in \mathbb{N}^*$, $\forall z,w \in B_{T_x\mathcal{X}}(0,R)$,
\begin{multline*}
\Norm{D^k_{(z,w)}\left(E_d(z,w) - \left(\frac{d}{\pi}\right)^{\hspace{-1mm}n}\frac{\exp\left(-\frac{d}{2}\left(\Norm{z}^2+\Norm{w}^2-2 \prsc{z}{w}\right)\right)}{\sqrt{\kappa(z)}\sqrt{\kappa(w)}}\Id_{(\mathcal{E}\otimes\L^d)_x}\right)}\\
\leq C_p d^{n+\frac{p}{2}-1}\left(1+\sqrt{d}(\Norm{z}+\Norm{w})\right)^{2n+6+p} \exp\left(-C'\sqrt{d}\Norm{z-w}\right)+O\!\left(d^{-\infty}\right),
\end{multline*}
where:
\begin{itemize}
\item $D^k_{(z,w)}$ is the $k$-th differential at $(z,w)$ for a map $T_x\mathcal{X} \times T_x\mathcal{X} \to\End\left(\left(\mathcal{E} \otimes \L^d\right)_x\right)$,
\item the Hermitian inner product $\prsc{\cdot}{\cdot}$ comes from the Hermitian metric $\left(g_\mathbb{C}\right)_x$,
\item the norm $\Norm{\cdot}$ on $T_x\mathcal{X}$ is induced by $g_x$ (or equivalently $\prsc{\cdot}{\cdot}$),
\item the norm $\Norm{\cdot}$ on $\left(T^*_x\mathcal{X}\right)^{\otimes q}\otimes \End\left(\left(\mathcal{E} \otimes \L^d\right)_x\right)$ is induced by $g_x$ and $(h_d)_x$.
\end{itemize}
Moreover, the constants $C_p$ and $C'$ do not depend on $x$. The notation $O\!\left(d^{-\infty}\right)$ means that, for any $l \in \mathbb{N}$, this term is $O\!\left(d^{-l}\right)$ with a constant that does not depend on $x$, $z$, $w$ or $d$.
\end{thm}
\begin{proof}
This is a weak version of \cite[thm.~4.2.1]{MM2007}, with $k=1$ and $m'=0$ in the notations of~\cite{MM2007}. We used the fact that $\mathcal{F}_0$ in \cite{MM2007} is given by:
\begin{equation*}
\mathcal{F}_0(z,w) = \frac{1}{\pi^n}\exp\left(-\frac{1}{2}\left(\Norm{z}^2+\Norm{w}^2-2 \prsc{z}{w}\right)\right)\Id_{(\mathcal{E}\otimes\L^d)_x},
\end{equation*}
(compare (4.1.84), (4.1.85) et (4.1.92) pp.~191--192 and (5.1.18) p.~46 in \cite{MM2007}) and $\mathcal{F}_1 = 0$. See~\cite[Rem.~1.4.26]{MM2007} and~\cite{MM2013}.
\end{proof}
\begin{rem}
\label{rem normalization Ma Marinescu}
Note that our formula differs from the one in \cite{MM2007,MM2013} by a factor $\pi$ in the exponential. This comes from different normalizations of the Kähler form $\omega$.
\end{rem}
We are only interested in the behavior of $E_d$ at points of the real locus, hence we restrict our focus to points in $M$ and derivatives in real directions. Similarly, for $x,y \in M$, $E_d(x,y)$ restricts to an element of $\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \otimes \mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)^*_y$, still denoted by $E_d(x,y)$. Note that we can recover the original $E_d(x,y) : \left(\mathcal{E} \otimes \L^d\right)_y \to \left(\mathcal{E} \otimes \L^d\right)_x$ from its restriction by $\mathbb{C}$-linear extension.
First, we need to know the behavior of $E_d$ and its derivatives up to order $1$ in each variable in a neighborhood of the diagonal in $M \times M$.
\begin{cor}
\label{cor near diag estimates}
There exist $C$ and $C' >0$, not depending on $x$, such that $\forall k \in \{0,1,2\}$, $\forall d \in \mathbb{N}^*$, $\forall z,w \in B_{T_xM}(0,R)$,
\begin{multline*}
\Norm{D^k_{(z,w)}\left(E_d(z,w) - \left(\frac{d}{\pi}\right)^n\frac{\exp\left(-\frac{d}{2}\Norm{z-w}^2\right)}{\sqrt{\kappa(z)}\sqrt{\kappa(w)}}\Id_{\mathbb{R}(\mathcal{E}\otimes\L^d)_x}\right)}\\
\leq C d^{n+\frac{k}{2}-1}\left(1+\sqrt{d}(\Norm{z}+\Norm{w})\right)^{2n+8} \exp\left(-C'\sqrt{d}\Norm{z-w}\right)+O\!\left(d^{-\infty}\right),
\end{multline*}
where $D^k$ is the $k$-th differential for a map from $T_xM \times T_xM$ to $\End\left(\mathbb{R}\left(\mathcal{E} \otimes \L^d\right)_x\right)$, the norm on $T_xM$ is induced by $g_x$ and the norm on $\left(T^*_xM\right)^{\otimes q}\otimes \End\left(\left(\mathcal{E} \otimes \L^d\right)_x\right)$ is induced by $g_x$ and $(h_d)_x$.
\end{cor}
\begin{proof}
We apply Theorem.~\ref{thm Ma Marinescu} for $p=k \in \{0,1,2\}$ and set $C=\max(C_0,C_1,C_2)$. Then we restrict everything to the real locus.
\end{proof}
\subsection{Diagonal estimates}
\label{subsec diagonal estimates}
In this section, we deduce diagonal estimates for $E_d$ and its derivatives from Thm.~\ref{thm Ma Marinescu}. Let $x \in M$, then the usual differential for maps from $T_x\mathcal{X}$ to $(\mathcal{E} \otimes \L^d)_x$ defines a local trivial connection $\tilde{\nabla}^d$ on $(\mathcal{E} \otimes \L^d)_{/B_\mathcal{X}(0,2R)}$, via the real normal trivialization. Since this trivialization is well-behaved with respect to both the metric and the real structure (cf.~Sect.~\ref{subsec real normal trivialization}), $\tilde{\nabla}^d$ is metric and real. By a partition of unity argument, there exists a real metric connection $\nabla^d$ on $\mathcal{E} \otimes \L^d$ such that $\nabla^d$ agrees with $\tilde{\nabla}^d$ on $B_\mathcal{X}(0,R)$. In the remainder of this section, we use this connection~$\nabla^d$, and the induced connection on $(\mathcal{E} \otimes \L^d) \boxtimes (\mathcal{E} \otimes \L^d)^*$, so that the connection is trivial in the real normal trivialization about $x$.
Recall that $\partial_y^\sharp E_d$ and $\partial_x\partial_y^\sharp E_d$ are defined by~\eqref{eq def sharp} and~\eqref{eq def sharp 2} respectively.
\begin{cor}
\label{cor diag estimates}
Let $x \in M$, let $\nabla^d$ be a real metric connection that is trivial over $B_{T_x\mathcal{X}}(0,R)$ in the real normal trivialization about $x$. Let $\partial_y^\sharp$ and $\partial_x$ denote the associated partial derivatives for sections of $(\mathcal{E} \otimes \L^d) \boxtimes (\mathcal{E} \otimes \L^d)^*$, then we have the following estimates as $d \to +\infty$.
\begin{align}
\label{eq value Bergman diag 00}
E_d(x,x) &= \frac{d^n}{\pi^n} \Id_{\mathbb{R}(\mathcal{E}\otimes\L^d)_x} + O\!\left(d^{n-1}\right),\\
\label{eq value Bergman diag 10}
\partial_xE_d(x,x) &= O\!\left(d^{n-\frac{1}{2}}\right),\\
\label{eq value Bergman diag 01}
\partial_y^\sharp E_d(x,x) &= O\!\left(d^{n-\frac{1}{2}}\right),\\
\label{eq value Bergman diag 11}
\partial_x\partial_y^\sharp E_d(x,x) &= \frac{d^{n+1}}{\pi^n} \Id_{\mathbb{R}(\mathcal{E}\otimes\L^d)_x} \otimes \Id_{T^*_xM} + O\!\left(d^n\right).
\end{align}
Moreover the error terms do not depend on $x$.
\end{cor}
\begin{proof}
Let $x \in M$ and let us choose an orthonormal basis of $T_xM$. We denote the corresponding coordinates on $T_xM \times T_xM$ by $(z_1,\dots,z_n,w_1,\dots,w_n)$ and by $\partial_{z_i}$ and $\partial_{w_j}$ the associated partial derivatives. Let us compute the partial derivatives of $E_d$ read in the real normal trivialization of $\left(\mathcal{E} \otimes \L^d\right) \boxtimes \left(\mathcal{E} \otimes\L^d\right)^*$ about $(x,x)$. By Cor.~\ref{cor near diag estimates}, we only need to compute the partial derivatives at $(0,0)$ of
\begin{equation}
\label{eq def xi d}
\xi_d : (z,w) \mapsto \frac{\exp\left(-\frac{d}{2}\Norm{z-w}^2\right)}{\sqrt{\kappa(z)}\sqrt{\kappa(w)}}
\end{equation}
for any $d \in \mathbb{N}$. For all $i$ and $j \in \{1,\dots,n\}$ and all $(z,w) \in B_{T_xM}(0,R)$ we have:
\begin{align}
\label{eq derivative xi d z}
\partial_{z_i}\xi_d(z,w) &= \left(-d(z_i-w_i)-\frac{1}{2}\frac{\partial_{z_i}\kappa(z)}{\kappa(z)}\right)\frac{\exp\left(-\frac{d}{2}\Norm{z-w}^2\right)}{\sqrt{\kappa(z)}\sqrt{\kappa(w)}},\\
\label{eq derivative xi d w}
\partial_{w_j}\xi_d(z,w) &= \left(d(z_j-w_j)-\frac{1}{2}\frac{\partial_{w_j}\kappa(w)}{\kappa(w)}\right)\frac{\exp\left(-\frac{d}{2}\Norm{z-w}^2\right)}{\sqrt{\kappa(z)}\sqrt{\kappa(w)}}
\end{align}
and
\begin{multline}
\label{eq derivative xi d zw}
\partial_{z_i}\partial_{w_j}\xi_d(z,w) = \frac{\exp\left(-\frac{d}{2}\Norm{z-w}^2\right)}{\sqrt{\kappa(z)}\sqrt{\kappa(w)}} \times\\ \left(d\delta_{ij}-d^2(z_i-w_i)(z_j-w_j)-\frac{d(z_j-w_j)}{2}\frac{\partial_{z_i}\kappa(z)}{\kappa(z)}+\frac{d(z_i-w_i)}{2}\frac{\partial_{w_j}\kappa(w)}{\kappa(w)}\right),
\end{multline}
where $\delta_{ij}$ equals $1$ if $i=j$ and $0$ otherwise. Recall that, by \eqref{eq estimate kappa}, $\kappa(0)=1$ and the partial derivatives of $\kappa$ vanish at the origin. Then evaluating the above expressions at $(0,0)$ gives:
\begin{align*}
\xi_d(0,0) &= 1, & \partial_{z_i}\xi_d(0,0)&=0=\partial_{w_j}\xi_d(0,0) & &\text{and} & \partial_{z_i}\partial_{w_j}\xi_d(0,0)&=\delta_{ij}d.
\end{align*}
By Cor.~\ref{cor near diag estimates}, we have the following estimates for the partial derivatives of $E_d$ read in the real normal trivialization about $x$: for all $i,j \in \{1,\dots,n\}$,
\begin{equation}
\label{eq diag values in chart}
\begin{aligned}
E_d(0,0) &= \frac{d^n}{\pi^n} \Id_{\mathbb{R}(\mathcal{E}\otimes\L^d)_x} + O\!\left(d^{n-1}\right), & \partial_{w_j}E_d(0,0) &= O\!\left(d^{n-\frac{1}{2}}\right),\\
\partial_{z_i}\partial_{w_j}E_d(0,0) &= \delta_{ij}\frac{d^{n+1}}{\pi^n} \Id_{\mathbb{R}(\mathcal{E}\otimes\L^d)_x} + O\!\left(d^{n}\right), & \partial_{z_i}E_d(0,0) &= O\!\left(d^{n-\frac{1}{2}}\right).
\end{aligned}
\end{equation}
Moreover these estimates are uniform in $x \in M$. Equations~\eqref{eq value Bergman diag 00}, \eqref{eq value Bergman diag 10}, \eqref{eq value Bergman diag 01} and \eqref{eq value Bergman diag 11} are coordinate-free versions of these statements.
\end{proof}
\subsection{Far off-diagonal estimates}
\label{subsec far off diagonal estimates}
Finally, we will use the fact that the Bergman kernel and its derivatives decrease fast enough outside of the diagonal. In this section we recall the far off-diagonal estimates of \cite[thm.~5]{MM2015}, see also \cite[prop.~4.1.5]{MM2007}.
Let $d \in \mathbb{N}$ and let $S$ be a smooth section of $\mathbb{R}\left(\mathcal{E} \otimes \L^d\right) \boxtimes \mathbb{R}\left(\mathcal{E} \otimes \L^d\right)^*$. Let $x,y \in M$, we denote by $\Norm{S(x,y)}_{\mathcal{C}^k}$ the maximum of the norms of $S$ and its derivatives of order at most $k$ at the point $(x,y)$. The derivatives of $S$ are computed with respect to the connection induced by the Chern connection of $\mathcal{E}\otimes\L^d$ and the Levi--Civita connection on $M$. The norms of the derivatives are the one induced by $h_d$ and $g$.
\begin{thm}[Ma--Marinescu]
\label{thm off diag estimates}
There exist $C'>0$ and $d_0 \in \mathbb{N}^*$ such that, for all $k \in \mathbb{N}$, there exists $C_k>0$ such that $\forall d \geq d_0$, $\forall x,y \in M$
\begin{equation*}
\Norm{E_d(x,y)}_{\mathcal{C}^k} \leq C_k d^{n+\frac{k}{2}} \exp \left(-C' \sqrt{d}\, \rho_g(x,y)\right),
\end{equation*}
where $\rho_g(\cdot,\cdot)$ denotes the geodesic distance in $(M,g)$.
\end{thm}
\begin{proof}
This is the first part of \cite[thm.~5]{MM2015}, where we only considered the restriction of $E_d$ and its derivatives to $M$. Note that the Levi--Civita connection on $M$ is the restriction of the Levi--Civita connection on $\mathcal{X}$. Hence the norm $\Norm{\cdot}_{\mathcal{C}^k}$, such as we defined it, is smaller than the one used in \cite{MM2015}.
\end{proof}
\section{Proof of Theorem~\ref{thm asymptotics variance}}
\label{sec proof of the main theorem}
In this section, we prove Theorem~\ref{thm asymptotics variance}. Recall that $\mathcal{X}$ is a compact Kähler manifold of dimension $n\geq 2$ defined over the reals and that $M$ denotes its real locus, assumed to be non-empty. Let $\mathcal{E}\to \mathcal{X}$ be a rank $r \in\{1,\dots,n-1\}$ real Hermitian vector bundle and $\L\to \mathcal{X}$ be a real Hermitian line bundle whose curvature form is $\omega$, the Kähler form of~$\mathcal{X}$. We assume that $\mathcal{E}$ and $\L$ are endowed with compatible real structures. For all $d \in \mathbb{N}$, $E_d$ denotes the Bergman kernel of $\mathcal{E} \otimes \L^d$. Finally, $s_d$ denotes a standard Gaussian vector in $\mathbb{R} \H$, whose real zero set is denoted by $Z_d$, and $\rmes{d}$ is the measure of integration over $Z_d$.
\subsection{The Kac--Rice formula}
\label{subsec Kac--Rice formula}
The first step in our proof of Thm.~\ref{thm asymptotics variance} is to prove a version of the Kac--Rice formula adapted to our problem. This is the goal of this section. First, we recall the Kac--Rice formula we used in \cite{Let2016} to compute the expectation of $\vol{Z_d}$ (Thm.~\ref{thm Kac-Rice exp}). Then we prove a Kac--Rice formula adapted to the computation of the covariance (Thm.~\ref{thm Kac-Rice var}), compare \cite[thm.~6.3]{AW2009} and \cite[chap.~11.5]{TA2007}.
Let $L : V \to V'$ be a linear map between two Euclidean spaces, recall that we denote by $\odet{L}$ its Jacobian (cf.~Def.~\ref{def Jacobian}). Since $LL^*$ is a semi-positive symmetric endomorphism of $V'$, $\det(LL^*) \geq 0$ and $\odet{L}$ is well-defined. The range of $L^*$ is $\ker(L)^\perp$, hence $\ker(LL^*)=\ker(L^*)=L(V)^\perp$. Thus $\odet{L} >0$ if and only if $LL^*$ is injective, that is if and only if $L$ is surjective. In fact, if $L$ is surjective, let $A$ be the matrix of the restriction of $L$ to $\ker(L)^\perp$ in any orthonormal basis of $\ker(L)^\perp$ and $V'$, then we have:
\begin{equation*}
\odet{L}=\sqrt{\det\left(A A^{\text{t}}\right)}=\norm{\det(A)}.
\end{equation*}
\begin{thm}[Kac--Rice formula]
\label{thm Kac-Rice exp}
Let $d \geq d_1$, where $d_1$ is defined by Lem.~\ref{lem dfn d1} and let $\nabla^d$ be any real connection on $\mathcal{E} \otimes \L^d$. Let $s_d$ be a standard Gaussian vector in $\mathbb{R} \H$. Then for any Borel measurable function $\phi : M \to \mathbb{R}$ we have:
\begin{equation}
\label{eq thm Kac-Rice exp}
\esp{\int_{x \in Z_d} \phi(x) \rmes{d}}= (2\pi)^{-\frac{r}{2}} \int_{x \in M} \frac{\phi(x)}{\odet{\ev_x^d}}\espcond{\odet{\nabla^d_{x}s_d}}{s_d(x)=0} \rmes{M}
\end{equation}
whenever one of these integrals is well-defined.
\end{thm}
The expectation on the right-hand side of~\eqref{eq thm Kac-Rice exp} is to be understood as the conditional expectation of $\odet{\nabla^d_{x}s_d}$ given that $s_d(x)=0$. This result is a consequence of~\cite[thm.~5.3]{Let2016}. See also Section~5.3 of \cite{Let2016}, where we applied this result with $\phi = \mathbf{1}$, in order to compute the expected volume of $Z_d$.
Let us denote by $\Delta = \{(x,y) \in M^2 \mid x = y\}$ the diagonal in $M^2$. Let $d \in \mathbb{N}$ and let $(x,y) \in M^2 \setminus \Delta$ we denote by $\ev^d_{x,y}$ the evaluation map:
\begin{equation}
\label{eq dfn ev map}
\begin{array}{rccc}
\ev_{x,y}^d:& \mathbb{R} \H & \longrightarrow & \mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \oplus \mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_y.\\
& s & \longmapsto & \left(s(x),s(y)\right)
\end{array}
\end{equation}
The following proposition is the equivalent of Lemma~\ref{lem dfn d1} for two points $(x,y) \notin \Delta$. One could prove this result using only the estimates of Section~\ref{sec estimates for the bergman kernel}. We give instead a less technical proof, using the Kodaira embedding theorem. See \cite[sect.~5.1]{MM2007} for a discussion of the relations between these approaches.
\begin{prop}
\label{prop amplitude 2 points}
There exists $d_2 \in \mathbb{N}$, depending only on $\mathcal{X}$, $\mathcal{E}$ and $\L$, such that for every $d \geq d_2$ and every $(x,y) \in M^2 \setminus \Delta$, the evaluation map $\ev^d_{x,y}$ is surjective.
\end{prop}
\begin{proof}
Recall that there exists $d_1 \in \mathbb{N}$ such that, for all $d\geq d_1$, the map $\ev_x^d$ defined by~\eqref{eq dfn 0 ample} is surjective for any $x \in M$ (see Lem.~\ref{lem dfn d1}). Then, for all $d \geq d_1$ and all $x \in M$, the complexified map $\tilde{\ev}_x^d:\H \to \left(\mathcal{E} \otimes \L^d\right)_x$ defined by $\tilde{\ev}_x^d(s)=s(x)$ is also surjective.
For any $l \in \mathbb{N}$, we denote by $\Psi_l: \mathcal{X} \to \P\left(H^0(\mathcal{X},\L^l)^*\right)$ the Kodaira map, defined by $\Psi_l(x) = \left\{s \in H^0(\mathcal{X},\L^l) \mathrel{}\middle|\mathrel{} s(x) = 0\right\}$. By the Kodaira embedding theorem (see \cite[chap.~1.4]{GH1994}), there exists $l_0 \in \mathbb{N}$ such that $\Psi_{l_0}$ is well-defined and is an embedding.
We set $d_2 = l_0+d_1$. Let $d \geq d_2$ and let $(x,y) \in M^2 \setminus \Delta$. Since $\Psi_{l_0}(x)$ and $\Psi_{l_0}(y)$ are distinct hyperplanes in $H^0(\mathcal{X},\L^{l_0})$, there exist $\sigma_x$ and $\sigma_y \in H^0(\mathcal{X},\L^{l_0})$ such that:
\begin{equation*}
\label{eq Kodaira}
\left\{\begin{aligned}
\sigma_x(x) &\neq 0,\\
\sigma_x(y) & = 0
\end{aligned}\right.
\qquad \text{and} \qquad
\left\{\begin{aligned}
\sigma_y(x) & = 0,\\
\sigma_y(y) & \neq 0.
\end{aligned}\right.
\end{equation*}
Since $d-l_0 \geq d_1$, $\tilde{\ev}_x^d$ is onto and there exist $\sigma_{1,x},\dots,\sigma_{r,x} \in H^0(\mathcal{X},\mathcal{E} \otimes \L^{d-l_0})$ such that $\left(\sigma_{1,x}(x),\dots,\sigma_{r,x}(x)\right)$ is a basis of $\left(\mathcal{E} \otimes \L^{d-l_0}\right)_x$. Similarly there exist $\sigma_{1,y},\dots,\sigma_{r,y}$ such that $\left(\sigma_{1,y}(y),\dots,\sigma_{r,y}(y)\right)$ is a basis of $\left(\mathcal{E} \otimes \L^{d-l_0}\right)_y$. We define global holomorphic sections of $\mathcal{E} \otimes \L^d$ by $s_{k,x} = \sigma_{k,x}\otimes \sigma_x$ and $s_{k,y} = \sigma_{k,y}\otimes \sigma_y$ for all $k \in \{1,\dots,r\}$. These sections are such that $(s_{k,x}(x))_{1\leq k \leq r}$ is a basis of $\left(\mathcal{E} \otimes \L^d\right)_x$, $(s_{k,y}(y))_{1\leq k \leq r}$ is a basis of $\left(\mathcal{E} \otimes \L^d\right)_y$ and for all $k \in \{1,\dots,r\}$, $s_{k,x}(y) = 0 = s_{k,y}(x)$. This proves that the map
\begin{equation*}
\label{eq dfn ev map complex}
\begin{array}{rccc}
\tilde{\ev}_{x,y}^d:& \H & \longrightarrow & \left(\mathcal{E} \otimes \L^d\right)_x \oplus \left(\mathcal{E} \otimes \L^d\right)_y.\\
& s & \longmapsto & \left(s(x),s(y)\right)
\end{array}
\end{equation*}
has rank at least $2r$ (as a $\mathbb{C}$-linear map). Since $\tilde{\ev}_{x,y}^d$ is the complexified map of $\ev_{x,y}^d$, the latter must have rank at least $2r$ (as a $\mathbb{R}$-linear map), hence it is onto.
\end{proof}
\begin{rem}
\label{rem non degeneracy and surjectivity}
In Prop.~\ref{prop amplitude 2 points}, $\ev_{x,y}^d$ is surjective if and only if $\odet{\ev_{x,y}^d}>0$, that is if and only if $\ev_{x,y}^d \left(\ev_{x,y}^d\right)^*$ is non-singular. Since the latter is the variance operator of $\ev_{x,y}^d(s_d)$, where $s_d \sim \mathcal{N}(\Id)$ in $\mathbb{R} \H$, we see that the surjectivity of $\ev_{x,y}^d$ is equivalent to the non-degeneracy of the distribution of $(s_d(x),s_d(y))$.
\end{rem}
We can now deduce a Kac--Rice type formula from Prop.~\ref{prop amplitude 2 points}. For any $d \in \mathbb{N}$, we define $F_d$ to be the following bundle map over $M^2$:
\begin{equation*}
\label{eq def bundle map Fd}
\begin{array}{rccc}
F_d:& \mathbb{R} \H \times M^2 & \longrightarrow & \displaystyle \mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)\times \mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right).\\
& (s,x,y) & \longmapsto & \left(s(x),s(y)\right)
\end{array}
\end{equation*}
Let $\nabla^d$ be any real connection on $\mathcal{E} \otimes \L^d\to \mathcal{X}$ (see Def.~\ref{dfn real connection}). Then by Rem.~\ref{rem real connection in a real direction}, the restriction of $\nabla^d$ defines a connection on $\mathbb{R} (\mathcal{E} \otimes \L^d) \to M$. Let $\nabla^d F_d$ denote the vertical component of the diffential of $F_d$. Then, for all $(s_0,x,y) \in \mathbb{R}\H \times M^2$, we have:
\begin{equation*}
\label{eq differential Fdp}
\begin{array}{rccc}
\nabla^d_{(s_0,x,y)}F_d:& \mathbb{R} \H \times T_xM \times T_yM & \longrightarrow & \mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \oplus \mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_y.\\
& (s,v,w) & \longmapsto & \left(s(x)+\nabla^d_xs_0\cdot v,s(y)+\nabla^d_ys_0\cdot w\right)
\end{array}
\end{equation*}
We denote by $\partial_1^dF_d$ the partial derivative of $F_d$ with respect to the first variable (meaning~$s$), and by $\partial_2^dF_d$ its partial derivative with respect to the second variable (meaning $(x,y)$). Then for all $(s_0,x,y) \in \mathbb{R}\H \times M^2$ we have:
\begin{align}
\label{eq partial derivatives Fdp}
\partial_1^dF_d(s_0,x,y) &= \ev_{x,y}^d & &\text{and} & \partial_2^dF_d(s_0,x,y): (v,w) \mapsto \left(\nabla_x^ds_0 \cdot v,\nabla_y^ds_0 \cdot w\right).
\end{align}
From now on, we assume that $d\geq d_2$, where $d_2$ is given by Prop.~\ref{prop amplitude 2 points}. We define an incidence manifold $\Sigma_d$ by:
\begin{equation*}
\label{eq def incidence manifold}
\Sigma_d = \left(F_d\right)^{-1}(0) \cap \left(\mathbb{R} \H \times \left(M^2 \setminus \Delta\right)\right).
\end{equation*}
By Prop.~\ref{prop amplitude 2 points} and eq.~\ref{eq partial derivatives Fdp}, for all $(s,x,y)\in \mathbb{R} \H \times \left(M^2 \setminus \Delta\right)$, $\partial_1^dF_{d,p}(s,x,y)$ is surjective. Thus, the restriction of $F_d$ to $\mathbb{R} \H \times \left(M^2 \setminus \Delta\right)$ is a submersion and $\Sigma_d$ is a submanifold of $\mathbb{R}\H \times M^2$ of codimension $2r$. Note that we are only concerned with the zero set of $F_d$, hence none of this depends on the choice of $\nabla^d$. We can now state the Kac--Rice formula in this context.
\begin{thm}[Kac--Rice formula]
\label{thm Kac-Rice var}
Let $d \geq d_2$, where $d_2$ is given by Prop.~\ref{prop amplitude 2 points}, and let $\nabla^d$ be any real connection on $\mathcal{E} \otimes \L^d$. Let $s_d$ be a standard Gaussian vector in $\mathbb{R} \H$. Then for any Borel measurable function $\Phi : \Sigma_d \to \mathbb{R}$ we have:
\begin{multline}
\label{eq thm Kac-Rice var}
\esp{\int_{(x,y) \in (Z_d)^2 \setminus \Delta} \Phi(s_d,x,y) \rmes{d}^2}= \frac{1}{(2\pi)^r} \int_{(x,y) \in M^2 \setminus \Delta} \frac{1}{\odet{\ev_{x,y}^d}}\times\\
\espcond{\Phi(s_d,x,y) \odet{\nabla^d_{x}s_d}\odet{\nabla^d_{y}s_d}}{s_d(x)=0=s_d(y)} \rmes{M}^2
\end{multline}
whenever one of these integrals is well-defined. Here, $\rmes{M}^2$ stands for the product measure on $M^2$ induced by $\rmes{M}$. Similarly, $\rmes{d}^2$ is the product measure on $(Z_d)^2$.
\end{thm}
The expectation on the right-hand side of~\eqref{eq thm Kac-Rice var} is to be understood as the conditional expectation of $\Phi(s_d,x,y) \odet{\nabla^d_{x}s_d}\odet{\nabla^d_{y}s_d}$ given that $s_d(x)=0=s_d(y)$.
\begin{proof}
The proof of Thm.~\ref{thm Kac-Rice var} uses the double fibration trick, that is apply Federer's coarea formula twice. See for example \cite[App.~C]{Let2016} and the reference therein.
The Euclidean inner product on $\mathbb{R} \H$ defined by eq.~\eqref{eq definition inner product} and the Riemannian metric $g$ induce a Riemannian metric on $\mathbb{R} \H \times M^2$, and on $\Sigma_d$ by restriction. Let $\pi_1:\Sigma_d\to \mathbb{R} \H$ and $\pi_2:\Sigma_d\to M^2 \setminus \Delta$ denote the projections from $\Sigma_d$ to the first and second factor, respectively. For all $s \in \mathbb{R} \H$, $\pi_1^{-1}(s)$ is isometric to $Z_s$ and we identify these spaces. Similarly, for all $(x,y) \in M^2 \setminus \Delta$ we identify $\pi_2^{-1}(x,y)$ with the isometric space $\ker(\ev_{x,y}^d)$.
We denote by $\dmesure\! s$ the Lebesgue measure on $\mathbb{R} \H$ or any of its subspaces, normalized so that a unit cube has volume $1$. Let $\Phi:\Sigma_d \to \mathbb{R}$ be a Borel measurable function. Then
\begin{equation*}
\esp{\int_{(Z_d)^2 \setminus \Delta} \Phi \rmes{d}^2}= \int_{s \in \mathbb{R} \H} \left(\int_{(x,y) \in \pi_1^{-1}(s)} \Phi(s,x,y)\frac{e^{-\frac{1}{2}\Norm{s}^2}}{(2\pi)^{\frac{N_d}{2}}} \rmes{d}^2\right) \dmesure\! s,
\end{equation*}
where $N_d$ is the dimension of $\mathbb{R} \H$. Then, by the double fibration trick \cite[Prop.~C.3]{Let2016} this quantity equals:
\begin{equation}
\label{eq double fibration}
\int_{(x,y)\in M^2 \setminus \Delta}\left( \int_{s \in \ker(\ev_{x,y}^d)} \Phi(s,x,y)\frac{e^{-\frac{1}{2}\Norm{s}^2}}{(2\pi)^{\frac{N_d}{2}}} \frac{\odet{\partial_2^dF_d(s,x,y)}}{\odet{\partial_1^dF_d(s,x,y)}} \dmesure\! s\right) \rmes{M}^2.
\end{equation}
Then eq.~\eqref{eq partial derivatives Fdp} shows that $\partial_2^dF_{d,p}(s,x,y)= \nabla^d_xs \oplus \nabla^d_ys$. Moreover, by definition of the metrics, $T_xM$ is orthogonal to $T_yM$ and $\mathbb{R}(\mathcal{E}\otimes\L^d)_x$ is orthogonal to $\mathbb{R}(\mathcal{E}\otimes\L^d)_y$. Thus
\begin{align*}
\odet{\partial_2^dF_d(s,x,y)} &= \det\left(\partial_2^dF_d(s,x,y)\left(\partial_2^dF_d(s,x,y)\right)^*\right)^\frac{1}{2}\\
&= \det \left(\begin{pmatrix}
\nabla_x^ds & 0 \\ 0 & \nabla^d_ys
\end{pmatrix}
\begin{pmatrix}
(\nabla^d_xs)^* & 0 \\ 0 & (\nabla^d_ys)^*
\end{pmatrix}\right)^\frac{1}{2}\\
&= \det\begin{pmatrix}
\nabla_x^ds (\nabla_x^ds)^* & 0 \\ 0 & \nabla_y^ds (\nabla_y^ds)^*
\end{pmatrix}^\frac{1}{2}\\
&= \odet{\nabla^d_xs}\odet{\nabla^d_ys}.
\end{align*}
Besides, eq.\eqref{eq partial derivatives Fdp} also shows that $\odet{\partial_1^dF_d(s,x,y)}= \odet{\ev_{x,y}^d}$, which does not depend on $s$, so that \eqref{eq double fibration} equals:
\begin{equation*}
\int_{(x,y) \in M^2 \setminus \Delta} \frac{1}{\odet{\ev_{x,y}^d}}\left(\int_{s \in \ker\left(\ev_{x,y}^d\right)}\Phi \odet{\nabla^d_xs}\odet{\nabla^d_ys}\frac{e^{-\frac{1}{2}\Norm{s}^2}}{(2\pi)^{\frac{N_d}{2}}} \dmesure\! s \right) \rmes{M}^2.
\end{equation*}
Finally, by Prop.~\ref{prop amplitude 2 points}, $\ker(\ev_x^d)$ is a subspace of codimension $2r$ of $\mathbb{R} \H$. Hence, the inner integral in~\eqref{eq double fibration} can be expressed as a conditional expectation given that $\ev_{x,y}^d(s_d)=0$, up to a factor $(2\pi)^r$. This concludes the proof of Thm.~\ref{thm Kac-Rice var}.
\end{proof}
\subsection{An integral formula for the variance}
\label{subsec an integral formula for the variance}
In this section, we fix some $d \geq \max(d_0,d_1,d_2)$ where $d_0$, $d_1$ and $d_2$ are defined by Thm.~\ref{thm off diag estimates}, Lem.~\ref{lem dfn d1} and Prop.~\ref{prop amplitude 2 points} respectively. We denote by $\nabla^d$ a real connection on $\mathcal{E} \otimes \L^d$. Let $\phi_1, \phi_2 \in \mathcal{C}^0(M)$, we want to compute:
\begin{equation}
\label{eq variance}
\begin{aligned}
\var{\rmes{d}}\left(\phi_1,\phi_2\right) &= \cov{\prsc{\rmes{d}}{\phi_1}}{\prsc{\rmes{d}}{\phi_2}}\\
&= \esp{\prsc{\rmes{d}}{\phi_1}\prsc{\rmes{d}}{\phi_2}} - \esp{\prsc{\rmes{d}}{\phi_1}}\esp{\prsc{\rmes{d}}{\phi_2}}.
\end{aligned}
\end{equation}
First, by Thm.~\ref{thm Kac-Rice exp}, we have:
\begin{multline}
\label{eq exp squared}
\esp{\prsc{\rmes{d}}{\phi_1}}\esp{\prsc{\rmes{d}}{\phi_2}}=\frac{1}{(2\pi)^r} \times\\
\int_{M^2}\phi_1(x)\phi_2(y)\frac{\espcond{\odet{\nabla^d_{x}s_d}}{s_d(x)=0}}{\odet{\ev_x^d}}\frac{\espcond{\odet{\nabla^d_{y}s_d}}{s_d(y)=0}}{\odet{\ev_y^d}}\rmes{M}^2.
\end{multline}
On the other hand,
\begin{align*}
\esp{\prsc{\rmes{d}}{\phi_1}\prsc{\rmes{d}}{\phi_2}} &=\esp{\left(\int_{x \in Z_d}\phi_1(x) \rmes{d}\right)\left(\int_{y \in Z_d}\phi_2(y) \rmes{d}\right)}\\
&= \esp{\int_{(x,y) \in (Z_d)^2\setminus \Delta}\phi_1(x)\phi_2(y) \rmes{d}^2}.
\end{align*}
Indeed, $Z_d$ is almost surely of dimension $n-r > 0$, so that $(Z_d)^2 \cap \Delta$ (that is the diagonal in $(Z_d)^2$) has measure~$0$ for $\rmes{d}^2$. We compute this integral by Thm.~\ref{thm Kac-Rice var}:
\begin{multline}
\label{eq case n>r}
\esp{\int_{(x,y) \in (Z_d)^2\setminus \Delta}\phi_1(x)\phi_2(y) \rmes{d}^2} = \frac{1}{(2\pi)^r}\int_{(x,y) \in M^2 \setminus \Delta} \frac{\phi_1(x)\phi_2(y)}{\odet{\ev_{x,y}^d}}\times\\
\espcond{\odet{\nabla^d_{x}s_d}\odet{\nabla^d_{y}s_d}}{s_d(x)=0=s_d(y)} \rmes{M}^2.
\end{multline}
Let $\mathcal{D}_d$ be the function defined by: $\forall (x,y)\in M^2 \setminus \Delta$,
\begin{multline}
\label{eq def density}
\mathcal{D}_d(x,y) = \left(\frac{\espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0}}{\odet{\ev_{x,y}^d}}\right.\\
\left.-\frac{\espcond{\odet{\nabla^d_{x}s_d}}{s_d(x)=0}\espcond{\odet{\nabla^d_{y}s_d}}{s_d(y)=0}}{\odet{\ev_x^d}\odet{\ev_y^d}}\right)
\end{multline}
Since $\dim M =n >0$, $\Delta$ has measure $0$ in $M^2$ for $\rmes{M}^2$. Thus, by~\eqref{eq variance}, \eqref{eq exp squared}, \eqref{eq case n>r} and~\eqref{eq def density}, we have:
\begin{equation}
\label{eq variance2}
\var{\rmes{d}}\left(\phi_1,\phi_2\right) = \frac{1}{(2\pi)^r} \int_{M^2}\phi_1(x)\phi_2(y) \mathcal{D}_d(x,y) \rmes{M}^2.
\end{equation}
\begin{rem}
\label{rem choice of nablad}
At this stage, it is worth noticing that the values of the conditional expectations appearing in the definition of $\mathcal{D}_d$ (see eq.~\eqref{eq def density}) do not depend on the choice of $\nabla^d$. In fact, the whole conditional distribution of $\nabla^d_xs_d$ given that $s_d(x)=0$ (resp.~of $\nabla^d_ys_d$ given that $s_d(y)=0$, resp.~of $(\nabla^d_xs_d,\nabla^d_ys_d)$ given that $s_d(x)=0=s_d(y)$) is independent of the choice of $\nabla^d$. Indeed, if $s_d(x)=0$ then $\nabla^d_x s_d$ does not depend on $\nabla^d$, and we conditioned on the vanishing of $s_d(x)$ (resp.~$s_d(y)$, resp.~$s_d(x)$ and $s_d(y)$). Thus, in the sequel, we can use any real connection we like, even one that depends on $(x,y) \in M^2 \setminus \Delta$.
\end{rem}
\subsection{Asymptotic for the variance}
\label{subsec asymptotic for the variance}
In this section we compute the asymptotic of the integral in eq.~\eqref{eq variance2}. The main point is to write $M^2$ as the disjoint union of a neighborhood of $\Delta$, of size about $\frac{\ln d}{\sqrt{d}}$, and its complement. In~\eqref{eq variance2}, the set of points that are far from the diagonal will contribute a term of smaller order than the neighborhood of $\Delta$. This is a consequence of the fast decrease of the Bergman kernel outside of the diagonal. The values of $s_d$ at $x$ and $y$ are not correlated, up to some small error, outside of a neighborhood of $\Delta$.
We still assume that $d \geq \max(d_0,d_1,d_2)$ and we denote by $s_d$ a standard Gaussian vector in $\mathbb{R} \H$.
\subsubsection{Asymptotics for the uncorrelated terms}
\label{subsubsec uncorrelated terms}
Let us first compute asymptotics for the terms in the expression of $\mathcal{D}_d$ (see eq.~\eqref{eq def density}) that only depend on one point, say $x \in M$. For all $x \in M$, $\ev_x^d$ is linear. Hence $s_d(x)=\ev_x^d(s_d)$ is a centered Gaussian vector in $\mathbb{R} \left(\mathcal{E} \otimes \L^d\right)_x$ with variance operator:
\begin{equation}
\label{eq variance evx}
\ev_x^d \left(\ev_x^d\right)^* = \esp{s_d(x) \otimes \left(s_d(x)\right)^*} = E_d(x,x),
\end{equation}
where $E_d$ is the Bergman kernel of $\mathcal{E}\otimes \L^d$ and the last equality is given by Prop.~\ref{prop covariance equals Bergman}.
\begin{lem}
\label{lem estimates odet evx}
For every $x \in M$, we have:
\begin{equation*}
\left(\frac{\pi}{d}\right)^\frac{nr}{2}\odet{\ev_x^d} = 1 + O\!\left(d^{-1}\right),
\end{equation*}
where the error term $O\!\left(d^{-1}\right)$ does not depend on $x$.
\end{lem}
\begin{proof}
Let $x \in M$, then $\odet{\ev_x^d}^2 = \det E_d(x,x)$ by~\eqref{eq variance evx}. By~\eqref{eq value Bergman diag 00}, we have:
\begin{equation*}
\left(\frac{\pi}{d}\right)^{nr}\odet{\ev_x^d}^2 = \det\left(\Id_{\mathbb{R}\left(\mathcal{E}\otimes \L^d\right)_x}+O\!\left(d^{-1}\right)\right) = 1 + O\!\left(d^{-1}\right).
\end{equation*}
The error term in~\eqref{eq value Bergman diag 00} is independent of $x$, therefore the same is true here.
\end{proof}
Let $\nabla^d$ be a real connection on $\mathcal{E} \otimes \L^d$. We assume that $\nabla^d$ is a metric connection, so that Lem.~\ref{lem dual connection} and Cor.~\ref{cor variance of the 1-jet} are valid in this context. Recall that the Chern connection is an example of real metric connection.
For all $x \in M$, let $j_x^d : s \mapsto \left(s(x),\nabla_x^ds\right)$ denote the evaluation of the $1$-jet at $x$, from $\mathbb{R} \H$ to $\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \otimes \left(\mathbb{R} \oplus T^*_xM\right)$. Since $j_x^d$ is linear, $\left(s_d(x),\nabla_x^ds_d\right)$ is a centered Gaussian vector with variance operator $j_x^d \left(j_x^d\right)^*$. This operator splits according to the direct sum $\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x\oplus \mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \otimes T^*_xM$:
\begin{equation}
\label{eq variance jx}
\begin{aligned}
j_x^d \left(j_x^d\right)^* &= \esp{j_x^d(s_d) \otimes \left(j_x^d(s_d)\right)^*}\\
&= \begin{pmatrix}
\esp{s_d(x)\otimes s_d(x)^*} & \esp{s_d(x)\otimes (\nabla^d_xs_d)^*}\\
\esp{\nabla^d_xs_d\otimes s_d(x)^*} & \esp{\nabla^d_xs_d\otimes (\nabla^d_xs_d)^*}
\end{pmatrix}\\
&= \begin{pmatrix}
E_d(x,x) & \partial_y^\sharp E_d(x,x)\\
\partial_xE_d(x,x) & \partial_x\partial_y^\sharp E_d(x,x)
\end{pmatrix},
\end{aligned}
\end{equation}
where the last equality comes from Cor.~\ref{cor variance of the 1-jet sharp}. We chose $d \geq d_1$, so that $\ev_x^d$ is surjective (see Lem.~\ref{lem dfn d1}), i.e.~$\det\left(\ev_x^d\left(\ev_x^d\right)^*\right)>0$. Hence, the distribution of $s_d(x)$ is non-degenerate. Then (see~\cite[prop.~1.2]{AW2009}), the distribution of $\nabla^d_xs_d$ given that $s_d(x)=0$ is a centered Gaussian in $\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \otimes T^*_xM$ with variance operator:
\begin{equation}
\label{eq conditional exp}
\partial_x\partial_y^\sharp E_d(x,x)- \partial_xE_d(x,x) \left(E_d(x,x)\right)^{-1}\partial_y^\sharp E_d(x,x).
\end{equation}
\begin{lem}
\label{lem estimates cond exp x}
For every $x \in M$, we have:
\begin{equation*}
\left(\frac{\pi^n}{d^{n+1}}\right)^\frac{r}{2} \espcond{\odet{\nabla^d_xs_d}}{s_d(x)=0} = (2\pi)^\frac{r}{2} \frac{\vol{\S^{n-r}}}{\vol{\S^n}} \left(1 + O\!\left(d^{-1}\right)\right),
\end{equation*}
where the error term is independent of $x$.
\end{lem}
\begin{proof}
Let $x \in M$, and let $L_d(x)$ be a centered Gaussian vector in $\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \otimes T^*_xM$ with variance operator:
\begin{equation}
\label{eq def Lambda d x}
\Lambda_d(x) = \frac{\pi^n}{d^{n+1}}\left(\partial_x\partial_y^\sharp E_d(x,x)- \partial_xE_d(x,x) \left(E_d(x,x)\right)^{-1}\partial_y^\sharp E_d(x,x)\right).
\end{equation}
By~\eqref{eq conditional exp} and the above discussion, the distribution of $\nabla^d_xs_d$ given that $s_d(x)=0$ equals that of $\left(\frac{d^{n+1}}{\pi^n}\right)^\frac{1}{2}L_d(x)$. Then,
\begin{equation}
\label{eq cond exp equals}
\espcond{\odet{\nabla^d_xs_d}}{s_d(x)=0} = \esp{\odet{\hspace{-1mm}\left(\frac{d^{n+1}}{\pi^n}\right)^{\hspace{-1mm}\frac{1}{2}} \hspace{-1mm} L_d(x)}}=\hspace{-1mm} \left(\frac{d^{n+1}}{\pi^n}\right)^{\hspace{-1mm}\frac{r}{2}} \esp{\odet{L_d(x)}}.
\end{equation}
Recall that the distribution of $\nabla^d_xs_d$ given that $s_d(x)=0$ does not depend on the choice of $\nabla^d$ (Rem.~\ref{rem choice of nablad}). Hence $\Lambda_d(x)$ does not depend on the choice of $\nabla^d$. For the following computation, we choose $\nabla^d$ to be trivial over $B_{T_xM}(0,R)$ in the real normal trivialization about $x$. Then we can use the diagonal estimates of Cor.~\ref{cor diag estimates} for the Bergman kernel and its derivatives. We have: $\Lambda_d(x) = \Id_{\mathbb{R}(\mathcal{E}\otimes\L^d)_x} \otimes \Id_{T^*_xM} + O\!\left(d^{-1}\right)$, where the error does not depend on $x$. Hence,
\begin{equation}
\label{eq det Lambda dx}
\det\left(\Lambda_d(x)\right) = 1 + O\!\left(d^{-1}\right).
\end{equation}
Besides, there exists some $K>0$ such that $\Norm{\Lambda_d(x)^{-1}-\Id} \leq Kd^{-1}$ for all $d$ large enough. Then, by the mean value inequality, for all $L \in \mathbb{R}(\mathcal{E}\otimes\L^d)_x \otimes T^*_xM$
\begin{equation*}
\norm{\exp\left(-\frac{1}{2}\prsc{\left(\Lambda_d(x)^{-1}- \Id\right) L}{L}\right)-1} \leq \frac{K}{2d}\Norm{L}^2\exp\left(\frac{K}{2d}\Norm{L}^2\right).
\end{equation*}
Let $L^0_d(x) \sim \mathcal{N}(\Id)$ in $\mathbb{R}(\mathcal{E}\otimes\L^d)_x \otimes T^*_xM$ and let $\dmesure\! L$ denote the normalized Lebesgue measure on $\mathbb{R}(\mathcal{E}\otimes\L^d)_x \otimes T^*_xM$. Then we have:
\begin{multline*}
(2\pi)^{\frac{nr}{2}}\norm{\det\left(\Lambda_d(x)\right)^\frac{1}{2}\esp{\odet{L_d(x)}}-\esp{\odet{L^0_d(x)}}}\\
\begin{aligned}
&\leq\int \odet{L} e^{-\frac{1}{2}\Norm{L}^2}\norm{\exp\left(-\frac{1}{2}\prsc{\left(\Lambda_d(x)^{-1}- \Id\right) L}{L}\right)-1} \dmesure\! L\\
&\leq \frac{K}{2d} \int \odet{L} \exp\left(-\frac{1}{2}\left(1-\frac{K}{d}\right)\Norm{L}^2\right) \dmesure\! L.
\end{aligned}
\end{multline*}
The integral on the last line converges to some finite limit as $d \to +\infty$. Thus, by~\eqref{eq det Lambda dx},
\begin{equation}
\label{eq last lemma}
\begin{aligned}
\esp{\odet{L_d(x)}}&= \det\left(\Lambda_d(x)\right)^{-\frac{1}{2}} \left(\esp{\odet{L^0_d(x)}}+ O\!\left(d^{-1}\right)\right)\\
&=\esp{\odet{L^0_d(x)}}+ O\!\left(d^{-1}\right),
\end{aligned}
\end{equation}
uniformly in $x \in M$. Lemma~\ref{lem estimates cond exp x} follows from~\eqref{eq cond exp equals}, \eqref{eq last lemma} and the following equality, that was proved in~\cite[lem.~A.14]{Let2016}:
\begin{equation}
\label{eq value exp odet lim}
\esp{\odet{L^0_d(x)}} = (2\pi)^\frac{r}{2}\frac{\vol{\S^{n-r}}}{\vol{\S^n}}.\qedhere
\end{equation}
\end{proof}
\subsubsection{Far off-diagonal asymptotics for the correlated terms}
\label{subsubsec far off diagonal correlated}
We can now focus on computing terms in the expression of $\mathcal{D}_d$ that depend on both $x$ and~$y$. For all $(x,y)\in M^2 \setminus \Delta$, $\ev^d_{x,y}(s_d)=\left(s_d(x),s_d(y)\right)$ is a centered Gaussian vector with variance operator:
\begin{equation}
\label{eq variance evxy}
\begin{aligned}
\ev^d_{x,y}(\ev^d_{x,y})^* &= \esp{\ev^d_{x,y}(s_d) \otimes \ev^d_{x,y}(s_d)^*}\\
&= \begin{pmatrix}
\esp{s_d(x) \otimes s_d(x)^*} & \esp{s_d(x) \otimes s_d(y)^*}\\ \esp{s_d(y) \otimes s_d(x)^*} & \esp{s_d(y) \otimes s_d(y)^*}
\end{pmatrix}\\
&= \begin{pmatrix}
E_d(x,x) & E_d(x,y)\\ E_d(y,x) & E_d(y,y)
\end{pmatrix},
\end{aligned}
\end{equation}
where we decomposed this operator according to the direct sum $\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \oplus \mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_y$. Since we assumed $d \geq d_2$, $\odet{\ev_{x,y}^d}>0$ (see Prop.~\ref{prop amplitude 2 points}) and the distribution of $\left(s_d(x),s_d(y)\right)$ is non-degenerate.
We denote by $j_{x,y}^d:s \mapsto \left(s(x),s(y),\nabla^d_xs,\nabla^d_ys\right)$ the evaluation of the $1$-jet at $(x,y)$. Then $j_{x,y}^d$ is a linear map from $\mathbb{R} \H$ to
\begin{equation}
\label{eq target space 1jet}
\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \oplus \mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_y \oplus \left(\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \otimes T^*_xM\right) \oplus \left(\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_y \otimes T^*_yM\right),
\end{equation}
and $j_{x,y}^d(s_d)$ is a centered Gaussian vector, with variance operator $j_{x,y}^d \left(j_{x,y}^d\right)^*$. We can split this variance operator according to the direct sum~\eqref{eq target space 1jet}. Then by Cor.~\ref{cor variance of the 1-jet sharp}, we have:
\begin{equation}
\label{eq variance jxy}
\begin{aligned}
&j_{x,y}^d \left(j_{x,y}^d\right)^* = \esp{j_{x,y}^d(s_d) \otimes \left(j_{x,y}^d(s_d)\right)^*}\\
&=\begin{pmatrix}
\esp{s_d(x)\otimes s_d(x)^*} & \esp{s_d(x)\otimes s_d(y)^*} & \esp{s_d(x)\otimes (\nabla^d_xs_d)^*} & \esp{s_d(x)\otimes (\nabla^d_ys_d)^*}\\
\esp{s_d(y)\otimes s_d(x)^*} & \esp{s_d(y)\otimes s_d(y)^*} & \esp{s_d(y)\otimes (\nabla^d_xs_d)^*} & \esp{s_d(y)\otimes (\nabla^d_ys_d)^*}\\
\esp{\nabla^d_xs_d\otimes s_d(x)^*} & \esp{\nabla^d_xs_d\otimes s_d(y)^*} & \esp{\nabla^d_xs_d\otimes (\nabla^d_xs_d)^*} & \esp{\nabla^d_xs_d\otimes (\nabla^d_ys_d)^*}\\
\esp{\nabla^d_ys_d\otimes s_d(x)^*} & \esp{\nabla^d_ys_d\otimes s_d(y)^*} & \esp{\nabla^d_ys_d\otimes (\nabla^d_xs_d)^*} & \esp{\nabla^d_ys_d\otimes (\nabla^d_ys_d)^*}
\end{pmatrix}\\
&= \begin{pmatrix}
E_d(x,x) & E_d(x,y) & \partial_y^\sharp E_d(x,x) & \partial_y^\sharp E_d(x,y)\\
E_d(y,x) & E_d(y,y) & \partial_y^\sharp E_d(y,x) & \partial_y^\sharp E_d(y,y)\\
\partial_xE_d(x,x) & \partial_xE_d(x,y) & \partial_x\partial_y^\sharp E_d(x,x) & \partial_x\partial_y^\sharp E_d(x,y)\\
\partial_xE_d(y,x) & \partial_xE_d(y,y) & \partial_x\partial_y^\sharp E_d(y,x) & \partial_x\partial_y^\sharp E_d(y,y)
\end{pmatrix}.
\end{aligned}
\end{equation}
Since the distribution of $\left(s_d(x),s_d(y)\right)$ is non-degenerate, the distribution of $\left(\nabla_x^ds,\nabla_y^ds\right)$ given that $\ev_{x,y}(s_d)=0$ is a centered Gaussian with variance operator:
\begin{equation}
\label{eq conditional variance full}
\begin{aligned}
\begin{pmatrix}
\partial_x\partial_y^\sharp E_d(x,x) & \partial_x\partial_y^\sharp E_d(x,y)\\
\partial_x\partial_y^\sharp E_d(y,x) & \partial_x\partial_y^\sharp E_d(y,y)
\end{pmatrix}&-\\
\begin{pmatrix}
\partial_xE_d(x,x) & \partial_xE_d(x,y)\\
\partial_xE_d(y,x) & \partial_xE_d(y,y)
\end{pmatrix}&
\begin{pmatrix}
E_d(x,x) & E_d(x,y)\\
E_d(y,x) & E_d(y,y)
\end{pmatrix}^{-1}
\begin{pmatrix}
\partial_y^\sharp E_d(x,x) & \partial_y^\sharp E_d(x,y)\\
\partial_y^\sharp E_d(y,x) & \partial_y^\sharp E_d(y,y)
\end{pmatrix}.
\end{aligned}
\end{equation}
\begin{dfn}
\label{def Lambda dxy}
For every $(x,y) \in M^2 \setminus \Delta$ and every $d$ large enough, we define $\Lambda_d(x,y)$ to be the operator such that $\frac{d^{n+1}}{\pi^n}\Lambda_d(x,y)$ equals~\eqref{eq conditional variance full}. That is, $\Lambda_d(x,y)$ is the conditional variance of $\left(\frac{\pi^n}{d^{n+1}}\right)^\frac{1}{2}\left(\nabla_x^ds,\nabla_y^ds\right)$ given that $\ev_{x,y}(s_d)=0$.
\end{dfn}
Let $C'>0$ be the constant appearing in the exponential in Thm.~\ref{thm off diag estimates}. We denote
\begin{align}
\label{eq def bn}
&b_n = \frac{1}{C'}\left(\frac{n}{2}+1\right)\\
\intertext{and}
\label{eq def Delta d}
\Delta_d = &\left\{(x,y) \in M^2 \mathrel{}\middle|\mathrel{} \rho_g(x,y) < b_n\frac{\ln d}{\sqrt{d}} \right\},
\end{align}
where, as before, $\rho_g$ is the geodesic distance in $(M,g)$.
\begin{lem}
\label{lem off diagonal evxy}
For every $(x,y) \in M^2 \setminus \Delta_d$, we have:
\begin{equation*}
\label{eq lem off diagonal evxy}
\odet{\ev^d_{x,y}} = \odet{\ev_x^d}\odet{\ev_y^d} \left(1 + O\!\left(d^{-\frac{n}{2}-1}\right)\right),
\end{equation*}
where the error term is uniform in $(x,y)\in M^2 \setminus \Delta_d$
\end{lem}
\begin{proof}
For all $(x,y) \in M^2 \setminus \Delta_d$, we have $\rho_g(x,y) \geq b_n \frac{\ln d}{\sqrt{d}}$. Then, by Thm.~\ref{thm off diag estimates},
\begin{equation*}
\Norm{E_d(x,y)} \leq C_0 d^n \exp\left(-C' b_n \ln d \right) \leq C_0 d^{\frac{n}{2}-1}.
\end{equation*}
Then, by~\eqref{eq variance evxy} we have:
\begin{align*}
\ev^d_{x,y}\left(\ev_{x,y}^d\right)^* &= \begin{pmatrix}
E_d(x,x) & E_d(x,y) \\ E_d(y,x) & E_d(y,y)
\end{pmatrix}\\
&= \begin{pmatrix}
E_d(x,x) & 0 \\ 0 & E_d(y,y)
\end{pmatrix} + O\!\left(d^{\frac{n}{2}-1}\right).
\end{align*}
Besides, by~\eqref{eq value Bergman diag 00},
\begin{equation}
\label{eq estimate condition inverse}
\begin{pmatrix}
E_d(x,x) & 0 \\ 0 & E_d(y,y)
\end{pmatrix}^{-1} = \left(\frac{\pi}{d}\right)^n \left(\Id + O\!\left(d^{-1}\right)\right) = O\!\left(d^{-n}\right),
\end{equation}
so that
\begin{equation}
\label{eq easier condition inverse}
\begin{pmatrix}
E_d(x,x) & E_d(x,y) \\ E_d(y,x) & E_d(y,y)
\end{pmatrix} = \begin{pmatrix}
E_d(x,x) & 0 \\ 0 & E_d(y,y)
\end{pmatrix} \left(\Id + O\!\left(d^{-\frac{n}{2}-1}\right)\right).
\end{equation}
We conclude the proof by taking the square root of the determinant of this last equality (recall~\eqref{eq variance evx}).
\end{proof}
\begin{lem}
\label{lem off diagonal cond exp}
For every $(x,y) \in M^2 \setminus \Delta_d$, we have:
\begin{multline*}
\label{eq lem off cond exp}
\espcond{\odet{\nabla^d_{x}s_d}\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0} =\\ \espcond{\odet{\nabla^d_{x}s_d}}{s_d(x)=0}\espcond{\odet{\nabla^d_{y}s_d}}{s_d(y)=0}\left(1 + O\!\left(d^{-\frac{n}{2}-1}\right)\right),
\end{multline*}
where the error term is uniform in $(x,y)\in M^2 \setminus \Delta_d$
\end{lem}
This lemma is a consequence of the following technical result.
\begin{lem}
\label{lem almost uncorrelated}
For every $(x,y) \in M^2 \setminus \Delta_d$, we have:
\begin{equation*}
\Lambda_d(x,y) = \begin{pmatrix}
\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{pmatrix} \left(\Id+ O\!\left(d^{-\frac{n}{2}-1}\right)\right),
\end{equation*}
uniformly in $(x,y) \in M^2 \setminus \Delta_d$.
\end{lem}
\begin{proof}[Proof of Lemma~\ref{lem off diagonal cond exp}]
Let $\left(L_d(x),L_d(y)\right)$ and $\left(L'_d(x),L'_d(y)\right)$ be centered Gaussian vectors in
\begin{equation*}
\left(\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \otimes T^*_xM\right) \oplus \left(\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_y \otimes T^*_yM\right)
\end{equation*}
such that: the variance of $\left(L'_d(x),L'_d(y)\right)$ is $\Lambda_d(x,y)$ (recall Def.~\ref{def Lambda dxy}), and $L_d(x)$ and $L_d(y)$ are independent with variances $\Lambda_d(x)$ and $\Lambda_d(y)$ respectively (see~\eqref{eq def Lambda d x}). Then, the distribution of $\left(L_d(x),L_d(y)\right)$ is a centered Gaussian with variance $\left(\begin{smallmatrix} \Lambda_d(x) & 0 \\ 0 & \Lambda_d(y) \end{smallmatrix}\right)$. From the definitions of $\Lambda_d(x)$, $\Lambda_d(y)$ and $\Lambda_d(x,y)$, we have:
\begin{align*}
\espcond{\odet{\nabla^d_{x}s_d}}{s_d(x)=0} &= \left(\frac{d^{n+1}}{\pi^n}\right)^{\hspace{-1mm}\frac{r}{2}}\esp{\odet{L_d(x)}},\\
\espcond{\odet{\nabla^d_{y}s_d}}{s_d(y)=0} &= \left(\frac{d^{n+1}}{\pi^n}\right)^{\hspace{-1mm}\frac{r}{2}}\esp{\odet{L_d(y)}},\\
\espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0} &= \left(\frac{d^{n+1}}{\pi^n}\right)^{\hspace{-1mm}r} \esp{\odet{L'_d(x)}\!\odet{L'_d(y)}}.
\end{align*}
Since $L_d(x)$ and $L_d(y)$ are independent, we only need to prove that:
\begin{equation}
\label{eq easier}
\esp{\odet{L'_d(x)}\odet{L'_d(y)}} =\esp{\odet{L_d(x)}\odet{L_d(y)}}\left(1 + O\!\left(d^{-\frac{n}{2}-1}\right)\right).
\end{equation}
By Lemma~\ref{lem almost uncorrelated},
\begin{equation}
\label{eq estimates det Lambda dxy}
\begin{aligned}
\det\left(\Lambda_d(x,y)\right) &= \det\left(\begin{pmatrix}
\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{pmatrix}\left(\Id+ O\!\left(d^{-\frac{n}{2}-1}\right)\right) \right)\\
&= \det\left(\Lambda_d(x)\right)\det\left(\Lambda_d(x)\right) \left(1+ O\!\left(d^{-\frac{n}{2}-1}\right)\right).
\end{aligned}
\end{equation}
Besides Lem.~\ref{lem almost uncorrelated} shows that:
\begin{equation*}
\label{eq inverse Lambda dxy}
\Lambda_d(x,y)^{-1} = \begin{pmatrix}
\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{pmatrix}^{-1} \left(\Id+ O\!\left(d^{-\frac{n}{2}-1}\right)\right).
\end{equation*}
By Cor.~\ref{cor diag estimates} and eq.~\eqref{eq def Lambda d x}, we have: $\left(\begin{smallmatrix}\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{smallmatrix}\right) = \Id + O\!\left(d^{-1}\right)$. Hence,
\begin{equation}
\label{eq Lambda d x Lambda d y -1}
\begin{pmatrix}\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{pmatrix}^{-1} = \Id + O\!\left(d^{-1}\right)
\end{equation}
uniformly in $(x,y)$, and
\begin{equation*}
\Lambda_d(x,y)^{-1} - \begin{pmatrix}
\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{pmatrix}^{-1} = O\!\left(d^{-\frac{n}{2}-1}\right).
\end{equation*}
Thus there exists $K>0$ such that, for all $d$ large enough,
\begin{equation*}
\Norm{\Lambda_d(x,y)^{-1} - \begin{pmatrix}
\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{pmatrix}^{-1}} \leq \frac{K}{d^{\frac{n}{2}+1}}.
\end{equation*}
Then, for every $L=(L_1,L_2) \in \left(\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x \otimes T^*_xM\right) \oplus \left(\mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_y \otimes T^*_yM\right)$, we have:
\begin{equation*}
\norm{\exp\left(-\frac{1}{2}\prsc{\left(\Lambda_d(x,y)^{-1}-\left(\begin{smallmatrix}
\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{smallmatrix}\right)^{-1}\right)L}{L}\right)-1} \leq \frac{K\Norm{L}^2}{2d^{\frac{n}{2}+1}} \exp\left(\frac{K\Norm{L}^2}{2d^{\frac{n}{2}+1}}\right).
\end{equation*}
Let $\dmesure\! L$ denote the normalized Lebesgue measure on this vector space. We have:
\begin{align*}
&\begin{aligned}
(2\pi)^{nr} \left\lvert \det\left(\Lambda_d(x,y)\right)^\frac{1}{2}\right. &\esp{\odet{L'_d(x)}\odet{L'_d(y)}}\\
&-\left. \det\left(\Lambda_d(x)\right)^\frac{1}{2}\det\left(\Lambda_d(y)\right)^\frac{1}{2} \esp{\odet{L_d(x)}\odet{L_d(y)}}\right\rvert
\end{aligned}\\
&\begin{aligned}
\leq \int \odet{L_1} & \odet{L_2} \exp\left(-\frac{1}{2}\prsc{\left(\begin{smallmatrix}
\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{smallmatrix}\right)^{-1}L}{L}\right)\times\\
&\norm{\exp\left(-\frac{1}{2}\prsc{\left(\Lambda_d(x,y)^{-1}-\left(\begin{smallmatrix}
\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{smallmatrix}\right)^{-1}\right)L}{L}\right)-1} \dmesure\! L
\end{aligned}\\
&\begin{aligned}
\leq \frac{K}{2d^{\frac{n}{2}+1}}\int \odet{L_1}& \odet{L_2} \Norm{L}^2\times\\
&\exp\left(-\frac{1}{2}\prsc{\left(\left(\begin{smallmatrix}
\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{smallmatrix}\right)^{-1}-\frac{K}{2d^{\frac{n}{2}+1}}\Id\right)L}{L}\right)\dmesure\! L
\end{aligned}\\
& = O\!\left(d^{-\frac{n}{2}-1}\right).
\end{align*}
Let us prove the last equality. By eq.~\eqref{eq Lambda d x Lambda d y -1}, for every $d$ large enough (uniform in $(x,y)$),
\begin{equation*}
\Norm{\begin{pmatrix}
\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{pmatrix}^{-1} -\left(1+\frac{K}{2d^{\frac{n}{2}+1}}\right)\Id} \leq \frac{1}{2},
\end{equation*}
so that:
\begin{multline*}
\int \odet{L_1} \odet{L_2} \Norm{L}^2 \exp\left(-\frac{1}{2}\prsc{\left(\left(\begin{smallmatrix}
\Lambda_d(x) & 0 \\ 0 & \Lambda_d(y)
\end{smallmatrix}\right)^{-1}-\frac{K}{2d^{\frac{n}{2}+1}}\Id\right)L}{L}\right)\dmesure\! L\\
\leq \int \odet{L_1} \odet{L_2} \Norm{L}^2 \exp\left(-\frac{1}{4}\Norm{L}^2\right) \dmesure\! L.
\end{multline*}
And the last integral is finite since $\odet{L_1} \odet{L_2} \Norm{L}^2$ is the norm of a polynomial in $L$.
Eq.~\eqref{eq det Lambda dx} and~\eqref{eq estimates det Lambda dxy} show that $\det(\Lambda_d(x,y)) = 1 +O\!\left(d^{-1}\right)$. Then, by the previous computations and eq.~\eqref{eq estimates det Lambda dxy}, we have:
\begin{multline*}
\esp{\odet{L'_d(x)}\odet{L'_d(y)}}\\
\begin{aligned}
&= \left(\frac{\det\left(\Lambda_d(x)\right)\det\left(\Lambda_d(y)\right)}{\det\left(\Lambda_d(x,y)\right)}\right)^\frac{1}{2} \esp{\odet{L_d(x)}\odet{L_d(y)}} + O\!\left(d^{-\frac{n}{2}-1}\right)\\
&= \esp{\odet{L_d(x)}\odet{L_d(y)}}\left(1 + O\!\left(d^{-\frac{n}{2}-1}\right)\right)+ O\!\left(d^{-\frac{n}{2}-1}\right).
\end{aligned}
\end{multline*}
Equations~\eqref{eq last lemma} and \eqref{eq value exp odet lim} proves that
\begin{equation*}
\esp{\odet{L_d(x)}\odet{L_d(y)}} = \esp{\odet{L_d(x)}}\esp{\odet{L_d(y)}}
\end{equation*}
converges to some positive constant. This proves~\eqref{eq easier} and establishes Lemma~\ref{lem off diagonal cond exp}.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{lem almost uncorrelated}]
First, recall that $\Lambda_d(x,y)$, $\Lambda_d(x)$ and $\Lambda_d(y)$ do not depend on the choice of $\nabla^d$ (see Rem.~\ref{rem choice of nablad}). In this proof, we use the Chern connection which is both real and metric. Let $(x,y)\in M^2 \setminus \Delta_d$, then $\rho_g(x,y) \geq b_n\frac{\ln d}{\sqrt{d}}$. By Thm.~\ref{thm off diag estimates}, we have:
\begin{equation*}
\Norm{\partial_xE_d(x,y)} \leq C_1 d^{n+\frac{1}{2}} \exp\left(-C' b_n \ln d \right) \leq C_1 d^{\frac{n}{2}-\frac{1}{2}}.
\end{equation*}
Similarly, $\Norm{\partial_xE_d(y,x)}$, $\Norm{\partial_y^\sharp E_d(x,y)}$ and $\Norm{\partial_y^\sharp E_d(y,x)}$ are smaller than $C_1 d^\frac{n-1}{2}$. Then
\begin{align}
\begin{pmatrix}
\partial_xE_d(x,x) & \partial_xE_d(x,y)\\
\partial_xE_d(y,x) & \partial_xE_d(y,y)
\end{pmatrix} &= \begin{pmatrix}
\partial_xE_d(x,x) & 0\\
0 & \partial_xE_d(y,y)
\end{pmatrix} + O\!\left(d^\frac{n-1}{2}\right)\\
\begin{pmatrix}
\partial_y^\sharp E_d(x,x) & \partial_y^\sharp E_d(x,y)\\
\partial_y^\sharp E_d(y,x) & \partial_y^\sharp E_d(y,y)
\end{pmatrix} &= \begin{pmatrix}
\partial_y^\sharp E_d(x,x) & 0\\
0 & \partial_y^\sharp E_d(y,y)
\end{pmatrix}+ O\!\left(d^\frac{n-1}{2}\right)
\intertext{and, by eq.~\eqref{eq easier condition inverse},}
\begin{pmatrix}
E_d(x,x) & E_d(x,y) \\ E_d(y,x) & E_d(y,y)
\end{pmatrix}^{-1} &= \begin{pmatrix}
E_d(x,x) & 0 \\ 0 & E_d(y,y)
\end{pmatrix}^{-1} \left(\Id + O\!\left(d^{-\frac{n}{2}-1}\right)\right).
\end{align}
Using eq.~\eqref{eq value Bergman diag 10}, \eqref{eq value Bergman diag 01} and~\eqref{eq estimate condition inverse}, we get:
\begin{multline}
\label{eq conditioning easier}
\begin{pmatrix}
\partial_xE_d(x,x) & \partial_xE_d(x,y)\\
\partial_xE_d(y,x) & \partial_xE_d(y,y)
\end{pmatrix}
\begin{pmatrix}
E_d(x,x) & E_d(x,y) \\ E_d(y,x) & E_d(y,y)
\end{pmatrix}^{-1}
\begin{pmatrix}
\partial_y^\sharp E_d(x,x) & \partial_y^\sharp E_d(x,y)\\
\partial_y^\sharp E_d(y,x) & \partial_y^\sharp E_d(y,y)
\end{pmatrix}\\
=\begin{pmatrix}
\partial_xE_d(x,x) & 0\\
0 & \partial_xE_d(y,y)
\end{pmatrix}
\begin{pmatrix}
E_d(x,x) & 0 \\ 0 & E_d(y,y)
\end{pmatrix}^{-1}
\begin{pmatrix}
\partial_y^\sharp E_d(x,x) & 0\\
0 & \partial_y^\sharp E_d(y,y)
\end{pmatrix}\\ +O\!\left(d^{\frac{n}{2}-1}\right)
\end{multline}
Using Thm.~\ref{thm off diag estimates} once more, we know that $\Norm{\partial_x\partial_y^\sharp E_d(x,y)}$ and $\Norm{\partial_x\partial_y^\sharp E_d(y,x)}$ are smaller than $C_2 d^\frac{n}{2}$. Then we have:
\begin{equation}
\label{eq conditioning easier 2}
\begin{pmatrix}
\partial_x\partial_y^\sharp E_d(x,x) & \partial_x\partial_y^\sharp E_d(x,y)\\
\partial_x\partial_y^\sharp E_d(y,x) & \partial_x\partial_y^\sharp E_d(y,y)
\end{pmatrix} = \begin{pmatrix}
\partial_x\partial_y^\sharp E_d(x,x) & 0\\
0 & \partial_x\partial_y^\sharp E_d(y,y)
\end{pmatrix} + O\!\left(d^\frac{n}{2}\right).
\end{equation}
We substract eq.~\eqref{eq conditioning easier} to eq.~\eqref{eq conditioning easier 2} and divide by $\frac{d^{n+1}}{\pi^n}$. By definition of $\Lambda_d(x,y)$, $\Lambda_d(x)$ and $\Lambda_d(y)$ (see Def.~\ref{def Lambda dxy} and eq.~\eqref{eq def Lambda d x}),
\begin{equation*}
\Lambda_d(x,y) = \begin{pmatrix}
\Lambda_d(x) & 0\\ 0 & \Lambda_d(y)
\end{pmatrix} + O\!\left(d^{-\frac{n}{2}-1}\right) = \begin{pmatrix}
\Lambda_d(x) & 0\\ 0 & \Lambda_d(y)
\end{pmatrix} \left(\Id + O\!\left(d^{-\frac{n}{2}-1}\right)\right),
\end{equation*}
where we used the fact that $\Lambda_d(x) = \Id + O\!\left(d^{-1}\right) = \Lambda_d(y)$ to obtain the last equality.
\end{proof}
\begin{prop}
\label{prop off diagonal is small}
Let $\phi_1,\phi_2 \in \mathcal{C}^0(M)$, then we have the following as $d \to +\infty$:
\begin{equation*}
\int_{M^2 \setminus \Delta_d} \phi_1(x)\phi_2(y) \mathcal{D}_d(x,y) \rmes{M}^2 = \Norm{\phi_1}_\infty\Norm{\phi_2}_\infty O\!\left(d^{r-\frac{n}{2}-1}\right),
\end{equation*}
where the error term is independent of $(\phi_1,\phi_2)$.
\end{prop}
\begin{proof}
We combine Lemmas~\ref{lem off diagonal evxy} and~\ref{lem off diagonal cond exp}, which gives:
\begin{multline*}
\frac{\espcond{\odet{\nabla^d_{x}s_d}\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0}}{\odet{\ev_{x,y}^d}}=\\
\frac{\espcond{\odet{\nabla^d_{x}s_d}}{s_d(x)=0}\espcond{\odet{\nabla^d_{y}s_d}}{s_d(y)=0}}{\odet{\ev_x^d}\odet{\ev_y^d}} \left(1+ O\!\left(d^{-\frac{n}{2}-1}\right)\right).
\end{multline*}
for all $(x,y) \in M^2 \setminus \Delta_d$. Besides, by Lemmas~\ref{lem estimates odet evx} and~\ref{lem estimates cond exp x},
\begin{equation*}
\frac{\espcond{\odet{\nabla^d_{x}s_d}}{s_d(x)=0}}{\odet{\ev_x^d}} = O\!\left(d^\frac{r}{2}\right) = \frac{\espcond{\odet{\nabla^d_{y}s_d}}{s_d(y)=0}}{\odet{\ev_y^d}}.
\end{equation*}
Recalling the definition of $\mathcal{D}_d$ (eq.~\eqref{eq def density}), we obtain that:
\begin{equation*}
\forall (x,y) \in M^2 \setminus \Delta_d, \qquad \mathcal{D}_d(x,y) = O\!\left(d^{r-\frac{n}{2}-1}\right),
\end{equation*}
uniformly in $(x,y) \notin \Delta_d$. Then, for any continuous $\phi_1$ and $\phi_2 \in \mathcal{C}^0(M)$, we have:
\begin{align*}
\norm{\int_{M^2 \setminus \Delta_d} \phi_1(x)\phi_2(y) \mathcal{D}_d(x,y) \rmes{M}^2} &\leq \Norm{\phi_1}_\infty\Norm{\phi_2}_\infty \vol{M^2} \left(\sup_{M^2 \setminus \Delta_d} \norm{\mathcal{D}_d}\right)\\
&= \Norm{\phi_1}_\infty\Norm{\phi_2}_\infty O\!\left(d^{r-\frac{n}{2}-1}\right).
\end{align*}
and the error term does not depend on $(\phi_1,\phi_2)$.
\end{proof}
\subsubsection{Properties of the limit distribution}
\label{subsubsec properties of the limit distribution}
Before we tackle the computation of the dominant term in~\eqref{eq variance2}, that is the integral over $\Delta_d$, we introduce the random variables that will turn out to be the scaling limits of $\left(\nabla^d_x s_d, \nabla^d_y s_d\right)$ given that $\ev_{x,y}^d(s_d) =0$. We also establish some of their properties.
\begin{ntn}
\label{ntn z* tens z}
Let $x \in M$ and $z \in T_xM$, we denote by $z^* \otimes z \in T^*_xM \otimes T_xM$ the linear map:
\begin{equation*}
\begin{array}{rccc}
z^* \otimes z : & T^*_xM & \longrightarrow & T^*_xM.\\
& \eta & \longmapsto & \eta(z)z^*
\end{array}
\end{equation*}
Let $\left(\deron{}{x_1},\dots,\deron{}{x_n}\right)$ be an orthonormal basis of $T_xM$ and let $\left( dx_1,\dots,dx_n\right)$ denote its dual basis. If $z = \sum z_i \deron{}{x_i}$ then $z^* \otimes z = \sum z_i z_j dx_i \otimes \deron{}{x_j}$, i.e.~the matrix of $z^* \otimes z$ in $\left(dx_1,\dots,dx_n\right)$ is $\left(z_i z_j\right)_{1\leq i,j \leq n}$.
\end{ntn}
\begin{dfn}
\label{def Lambda z inf}
For all $x \in M$ and $z \in T_xM \setminus\{0\}$, we define
\begin{equation*}
\Lambda_x(z) \in \End\left(\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x \otimes T^*_xM \otimes \mathbb{R}^2\right)
\end{equation*}
by:
\begin{align*}
\Lambda_x(z) = \begin{pmatrix}
\Id_{T^*_xM} - \frac{e^{-\Norm{z}^2}}{1-e^{-\Norm{z}^2}} z^* \otimes z & e^{-\frac{1}{2}\Norm{z}^2}\left(\Id_{T^*_xM} - \frac{z^*\otimes z}{1-e^{-\Norm{z}^2}}\right)\\
e^{-\frac{1}{2}\Norm{z}^2}\left(\Id_{T^*_xM} - \frac{z^*\otimes z}{1-e^{-\Norm{z}^2}}\right) & \Id_{T^*_xM} - \frac{e^{-\Norm{z}^2}}{1-e^{-\Norm{z}^2}} z^* \otimes z
\end{pmatrix}\otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x}.
\end{align*}
\end{dfn}
We need information about $\Lambda_x(z)$, especially concerning the vanishing of its eigenvalues. This will be useful in the estimates involving $\Lambda_x(z)$ below.
\begin{lem}
\label{lem eigenvalues Lambda z}
For all $x \in M$ and $z \in T_xM \setminus\{0\}$, the eigenvalues of $\Lambda_x(z)$ are:
\begin{itemize}
\item $1 - e^{-\frac{1}{2}\Norm{z}^2}$ and $1 + e^{-\frac{1}{2}\Norm{z}^2}$, each with multiplicity $(n-1)r$,
\item $\dfrac{1-e^{-\Norm{z}^2}+\Norm{z}^2e^{-\frac{1}{2}\Norm{z}^2}}{1+e^{-\frac{1}{2}\Norm{z}^2}}$ and $\dfrac{1-e^{-\Norm{z}^2}-\Norm{z}^2e^{-\frac{1}{2}\Norm{z}^2}}{1-e^{-\frac{1}{2}\Norm{z}^2}}$, each with multiplicity $r$.
\end{itemize}
\end{lem}
\begin{proof}
Let $x \in M$ and $z \in B_{T_xM}(0,b_n \ln d)\setminus\{0\}$. By definition of $\Lambda_x(z)$, its eigenvalues are the same as that of
\begin{equation}
\label{eq big operator}
\begin{pmatrix}
\Id_{T^*_xM} - \frac{e^{-\Norm{z}^2}}{1-e^{-\Norm{z}^2}} z^* \otimes z & e^{-\frac{1}{2}\Norm{z}^2}\left(\Id_{T^*_xM} - \frac{z^*\otimes z}{1-e^{-\Norm{z}^2}}\right)\\
e^{-\frac{1}{2}\Norm{z}^2}\left(\Id_{T^*_xM} - \frac{z^*\otimes z}{1-e^{-\Norm{z}^2}}\right) & \Id_{T^*_xM} - \frac{e^{-\Norm{z}^2}}{1-e^{-\Norm{z}^2}} z^* \otimes z
\end{pmatrix},
\end{equation}
with multiplicities multiplied by $r$. Hence, it is enough to compute the eigenvalues of the operator~\eqref{eq big operator}.
Let us choose an orthonormal basis $\left(\deron{}{x_1},\dots, \deron{}{x_n}\right)$ of $T_xM$ such that $z = \Norm{z}\deron{}{x_1}$, and let us denote by $\left(dx_1,\dots,dx_n\right)$ the dual basis. Then, $z^* \otimes z = \Norm{z}^2 dx_1 \otimes \deron{}{x_1}$. Let $(e_1,e_2)$ denote the canonical basis of $\mathbb{R}^2$, then the matrix of the operator \eqref{eq big operator} in the orthonormal basis $\left(e_1 \otimes dx_1,e_2 \otimes dx_1,\dots,e_1 \otimes dx_n,\dots,e_2 \otimes dx_n \right)$ is:
\begin{equation}
\label{eq big matrix}
\left(\begin{array}{cc|c}
1 - \frac{\Norm{z}^2 e^{-\Norm{z}^2}}{1-e^{-\Norm{z}^2}} & e^{-\frac{1}{2}\Norm{z}^2}\left(1 - \frac{\Norm{z}^2}{1-e^{-\Norm{z}^2}}\right) & 0 \\
e^{-\frac{1}{2}\Norm{z}^2}\left(1 - \frac{\Norm{z}^2}{1-e^{-\Norm{z}^2}}\right) & 1 - \frac{\Norm{z}^2 e^{-\Norm{z}^2}}{1-e^{-\Norm{z}^2}} & 0 \\
\hline
0 & 0 & \begin{pmatrix}
1 & e^{-\frac{1}{2}\Norm{z}^2} \\ e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix} \otimes
I_{n-1}\\
\end{array}\right),
\end{equation}
where $I_{n-1}$ stands for the identity matrix of size $n-1$.
The bottom-right block has eigenvalues $1-e^{-\frac{1}{2}\Norm{z}^2}$ and $1+e^{-\frac{1}{2}\Norm{z}^2}$, each with multiplicity $n-1$. To conclude the proof of Lemma~\ref{lem eigenvalues Lambda z}, we only need to observe that, for all $t>0$, the eigenvalues of
\begin{equation*}
\begin{pmatrix}
1 - \frac{t e^{-t}}{1-e^{-t}} & e^{-\frac{1}{2}t}\left(1 - \frac{t}{1-e^{-t}}\right)\\
e^{-\frac{1}{2}t}\left(1 - \frac{t}{1-e^{-t}}\right) & 1 - \frac{t e^{-t}}{1-e^{-t}}
\end{pmatrix}
\end{equation*}
are:
\begin{align*}
1 - \frac{t e^{-t}}{1-e^{-t}} + e^{-\frac{1}{2}t}\left(1 - \frac{t}{1-e^{-t}}\right) &= \frac{1-e^{-t}-te^{-\frac{1}{2}t}}{1-e^{-\frac{1}{2}t}}\\
\intertext{and}
1 - \frac{t e^{-t}}{1-e^{-t}} - e^{-\frac{1}{2}t}\left(1 - \frac{t}{1-e^{-t}}\right) &= \frac{1-e^{-t}+te^{-\frac{1}{2}t}}{1+e^{-\frac{1}{2}t}}.
\end{align*}
Note that the latter one is the largest.
\end{proof}
\begin{dfn}
\label{def f}
We define the function $f : (0,+\infty) \to \mathbb{R}$ by:
\begin{equation*}
\forall t >0, \qquad f(t) = \frac{1-e^{-\frac{1}{2}t}}{1-e^{-t}-te^{-\frac{1}{2}t}}.
\end{equation*}
\end{dfn}
\begin{cor}
\label{cor Lambda z det and inverse}
Let $x \in M$ and $z \in T_xM\setminus\{0\}$, then we have:
\begin{multline}
\label{eq Lambda z det}
\det\left(\Lambda_x(z)\right) =\\
\left(1-e^{-\Norm{z}^2}\right)^{r(n-2)}\left(1-e^{-\Norm{z}^2}+\Norm{z}^2 e^{-\frac{1}{2}\Norm{z}^2} \right)^r\left(1-e^{-\Norm{z}^2}-\Norm{z}^2 e^{-\frac{1}{2}\Norm{z}^2}\right)^r>0.
\end{multline}
Moreover,
\begin{align}
\label{eq Lambda z inverse}
\Norm{\Lambda_x(z)} &< 2 & &\text{and} & \Norm{\Lambda_x(z)^{-1}} &= f\left(\Norm{z}^2\right),
\end{align}
where $\Norm{\cdot}$ denote the operator norm on $\End\left(\mathbb{R}^2 \otimes \mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x \otimes T^*_xM\right)$.
\end{cor}
\begin{proof}
First, the formula for $\det\left(\Lambda_x(z)\right)$ is a direct consequence of Lem.~\ref{lem eigenvalues Lambda z}, and we only need to check that the eigenvalues of $\Lambda_x(z)$ are positive. Clearly, $1\pm e^{-\frac{1}{2}t} >0$ when $t>0$.
Then, for all positive $t$, we have:
\begin{equation*}
\frac{1-e^{-t}-te^{-\frac{1}{2}t}}{1-e^{-\frac{1}{2}t}} = \frac{e^{-\frac{1}{2}t}}{1-e^{-\frac{1}{2}t}}\left(e^{\frac{1}{2}t}-e^{-\frac{1}{2}t}-t\right) = \frac{e^{-\frac{1}{2}t}}{1-e^{-\frac{1}{2}t}}\left( 2 \sinh\left(\frac{t}{2}\right) - t\right),
\end{equation*}
and $2 \sinh\left(\frac{t}{2}\right) > t$. Besides,
\begin{equation*}
\frac{1-e^{-t}+te^{-\frac{1}{2}t}}{1+e^{-\frac{1}{2}t}} = \frac{e^{-\frac{1}{2}t}}{1+e^{-\frac{1}{2}t}}\left( 2 \sinh\left(\frac{t}{2}\right) + t\right) > 0.
\end{equation*}
Recall that $\Norm{\Lambda_x(z)}$ is the larger eigenvalue of $\Lambda_x(z)$, and $\Norm{\Lambda_x(z)^{-1}}$ is the inverse of the smallest eigenvalue of $\Lambda_x(z)$. For all $t>0$ we have
\begin{equation*}
0 < 1-e^{-\frac{t}{2}} < 1 + e^{-\frac{t}{2}} < 2.
\end{equation*}
Besides,
\begin{equation*}
\frac{1-e^{-t}-te^{-\frac{1}{2}t}}{1-e^{-\frac{1}{2}t}} + \frac{1-e^{-t}+te^{-\frac{1}{2}t}}{1+e^{-\frac{1}{2}t}} = 2\left(1 - \frac{t e^{-t}}{1-e^{-t}}\right) < 2,
\end{equation*}
and we just proved that both these terms are positive. Hence, each of them is smaller than~$2$. Thus, all the eigenvalues of $\Lambda_x(z)$ are smaller than $2$ and $\Norm{\Lambda_x(z)}< 2$.
For all $t>0$,
\begin{equation*}
\frac{1-e^{-t}-te^{-\frac{1}{2}t}}{1-e^{-\frac{1}{2}t}} < 1-e^{-\frac{t}{2}} \iff 1-e^{-t}-te^{-\frac{1}{2}t} < 1 - 2e^{-\frac{t}{2}} + e^{-t} \iff 1 - \frac{t}{2} < e^{-\frac{t}{2}},
\end{equation*}
and this is always true by convexity of the exponential. Thus, the smallest eigenvalue of $\Lambda_x(z)$ is $\dfrac{1-e^{-\Norm{z}^2}-\Norm{z}^2e^{-\frac{1}{2}\Norm{z}^2}}{1-e^{-\frac{1}{2}\Norm{z}^2}}=\dfrac{1}{f(\Norm{z}^2)}$, which proves our last claim.
\end{proof}
\begin{rem}
\label{rem norm Lambda z inverse}
To better understand the estimate~\eqref{eq Lambda z inverse}, note that $f$ is a decreasing function on $(0,+\infty)$. Moreover,
\begin{align*}
f(t) & \xrightarrow[t \to + \infty]{} 1, & &\text{and} & f(t) & \sim \frac{12}{t^2}
\end{align*}
when $t$ goes to $0$.
\end{rem}
\begin{dfn}
\label{def Lx 0 z}
For every $x \in M$, and $z \in T_xM \setminus \{0\}$, let $\left(L_x(0),L_x(z)\right)$ be a centered Gaussian vector in $\mathbb{R}^2 \otimes \mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x \otimes T^*_xM$ with variance operator $\Lambda_x(z)$.
\end{dfn}
Recall that we defined the random vector $\left(X(t),Y(t)\right)$ for all $t>0$ in the introduction (see Def.~\ref{def XYt}). Then $\left(X(t),Y(t)\right)$ and $\left(L_x(0),L_x(z)\right)$ are related as follows.
\begin{lem}
\label{lem same distribution}
Let $x \in M$ and $z \in T_xM \setminus \{0\}$, then there exists an orthonormal basis of $T_xM$ such that, for every orthonormal basis of $\mathbb{R}\left(\mathcal{E} \otimes \L^d\right)_x$, the couple of $r\times n$ matrices associated with $\left(L_x(0),L_x(z)\right)$ in these bases is distributed as $(X(\Norm{z}^2),Y(\Norm{z}^2))$.
\end{lem}
\begin{proof}
As in the proof of Lem.~\ref{lem eigenvalues Lambda z}, let us choose an orthonormal basis $\left(\deron{}{x_1},\dots,\deron{}{x_n}\right)$ of $T_xM$ such that $z = \Norm{z}\deron{}{x_1}$. Let $\left(dx_1,\dots,dx_n\right)$ denote its dual basis. Let $\left(\zeta_1,\dots,\zeta_r\right)$ be any orthonormal basis of $\mathbb{R} \left(\mathcal{E} \otimes \L^d \right)_x$, and let $(e_1,e_2)$ denote the canonical basis of $\mathbb{R}^2$.
Then $z^* \otimes z = \Norm{z}^2 dx_1 \otimes \deron{}{x_1}$ and the matrix of the operator~\eqref{eq big operator} in the orthonormal basis $\left(e_1\otimes dx_1,\dots, e_1 \otimes dx_n,e_2 \otimes dx_1,\dots, e_2 \otimes dx_n\right)$ is:
\begin{equation}
\label{eq big matrix 0}
\left(\begin{array}{cc|cc}
1 - \frac{\Norm{z}^2 e^{-\Norm{z}^2}}{1-e^{-\Norm{z}^2}} & 0 & e^{-\frac{1}{2}\Norm{z}^2}\left(1 - \frac{\Norm{z}^2}{1-e^{-\Norm{z}^2}}\right) & 0 \\
0 & I_{n-1} & 0 & e^{-\frac{1}{2}\Norm{z}^2} I_{n-1}\\
\hline
e^{-\frac{1}{2}\Norm{z}^2}\left(1 - \frac{\Norm{z}^2}{1-e^{-\Norm{z}^2}}\right) & 0 & 1 - \frac{\Norm{z}^2 e^{-\Norm{z}^2}}{1-e^{-\Norm{z}^2}} & 0 \\
0 & e^{-\frac{1}{2}\Norm{z}^2} I_{n-1} & 0 & I_{n-1}
\end{array}\right),
\end{equation}
where $I_{n-1}$ stands for the identity matrix of size $n-1$. Since $\Lambda_x(z)$ equals this operator tensored by $\Id_{\mathbb{R} \left(\mathcal{E} \otimes \L^d \right)_x}$, the matrix of $\Lambda_x(z)$ in the orthonormal basis:
\begin{multline*}
\left(e_1\otimes dx_1\otimes \zeta_1,\dots, e_1 \otimes dx_n\otimes \zeta_1,e_2 \otimes dx_1\otimes \zeta_1,\dots, e_2 \otimes dx_n\otimes \zeta_1,\right. \\
\left. e_1\otimes dx_1\otimes \zeta_2,\dots,e_2\otimes dx_n\otimes \zeta_2, \dots e_1\otimes dx_1\otimes \zeta_r,\dots,e_2\otimes dx_n\otimes \zeta_r\right)
\end{multline*}
is exactly the variance matrix of $(X(\Norm{z}^2),Y(\Norm{z}^2))$ (cf.~Def.~\ref{def XYt}).
Let us denote by $M_x(0)$ and $M_x(z)$ the matrices of $L_x(0)$ and $L_x(d)$ in the bases $\left(\deron{}{x_1},\dots,\deron{}{x_n}\right)$ and $\left(\zeta_1,\dots,\zeta_r\right)$. Then $\left(M_x(0),M_x(z)\right)$ is a centered Gaussian vector in $\mathcal{M}_{rn}(\mathbb{R})^2$. Moreover, we have just seen that the variance matrix of this random vector is the same as that of $(X(\Norm{z}^2),Y(\Norm{z}^2))$. This concludes the proof.
\end{proof}
\begin{cor}
\label{cor same distribution}
Let $x \in M$ and $z \in T_xM \setminus \{0\}$, then we have:
\begin{equation*}
\esp{\odet{L_x(0)}\odet{L_x(z)}} = \esp{\odet{X(\Norm{z}^2)}\odet{Y(\Norm{z}^2)}}.
\end{equation*}
\end{cor}
\begin{proof}
With the same notations as in the proof of Lemma~\ref{lem same distribution} above, we have:
\begin{equation*}
\esp{\odet{M_x(0)}\odet{M_x(z)}} = \esp{\odet{X(\Norm{z}^2)}\odet{Y(\Norm{z}^2)}},
\end{equation*}
since $\left(M_x(0),M_x(z)\right)$ and $(X(\Norm{z}^2),Y(\Norm{z}^2))$ have the same distribution. Besides,
\begin{align*}
\odet{L_x(0)}& =\odet{M_x(0)} & &\text{and} & \odet{L_x(z)}&=\odet{M_x(0)}.\qedhere
\end{align*}
\end{proof}
Let us now establish some facts about the distribution of $\left(X(t),Y(t)\right)$ for $t>0$.
\begin{lem}
\label{lem cond exp bounded}
For all $t>0$, we have:
\begin{equation*}
\esp{\odet{X(t)}\odet{Y(t)}} \leq n^r.
\end{equation*}
\end{lem}
\begin{proof}
First, by the Cauchy--Schwarz inequality, we have:
\begin{equation*}
\esp{\odet{X(t)}\odet{Y(t)}} \leq \esp{\odet{X(t)}^2}^\frac{1}{2} \esp{\odet{Y(t)}^2}^\frac{1}{2}.
\end{equation*}
Then, the definition of $\left(X(t),Y(t)\right)$ (Def.~\ref{def XYt}) shows that both $X(t)$ and $Y(t)$ are centered Gaussian vectors in $\mathcal{M}_{rn}(\mathbb{R})$ with variance matrix:
\begin{equation}
\label{eq variance L}
\begin{pmatrix}
1 - \frac{t e^{-t}}{1-e^{-t}} & 0 \\ 0 & I_{n-1}
\end{pmatrix}\otimes I_r.
\end{equation}
in the canonical bases of $\mathbb{R}^n$ and $\mathbb{R}^r$. Here $I_r$ and $I_{n-1}$ stand for the identity matrices of size $r$ and $n-1$ respectively. Hence,
\begin{equation*}
\esp{\odet{X(t)}\odet{Y(t)}} \leq \esp{\odet{X(t)}^2} = \esp{\det\left(X(t)X(t)^\text{t}\right)}.
\end{equation*}
We denote by $X_1(t),\dots,X_r(t)$ the rows of $X(t)$. Then
\begin{equation*}
X(t)X(t)^\text{t} = \left(\prsc{X_i(t)}{X_j(t)}\right)_{1\leq i,j\leq r},
\end{equation*}
where we see $X_i(t)$ as an element of $\mathbb{R}^n$ and $\prsc{\cdot}{\cdot}$ is the usual inner product on $\mathbb{R}^n$. Hence, $\det\left(X(t)X(t)^\text{t}\right)$ is the Gram determinant of the family $\left(X_1(t),\dots,X_r(t)\right)$, which is known to be the square of the $r$-dimensional volume of the parallelepiped spanned by these vectors. In particular,
\begin{equation*}
\det\left(X(t)X(t)^\text{t}\right) \leq \Norm{X_1(t)}^2 \cdots \Norm{X_r(t)}^2.
\end{equation*}
By~\eqref{eq variance L}, the $X_i(t)$ are independent identically distributed centered Gaussian vectors with variance matrix:
\begin{equation*}
\begin{pmatrix}
1 - \frac{t e^{-t}}{1-e^{-t}} & 0 \\ 0 & I_{n-1}
\end{pmatrix},
\end{equation*}
so that:
\begin{equation*}
\det\left(X(t)X(t)^\text{t}\right)\leq \esp{\Norm{X_1(t)}^2 \cdots \Norm{X_r(t)}^2} \leq \esp{\Norm{X_1(t)}^2}^r = \left(n - \frac{t e^{-t}}{1-e^{-t}}\right)^r\leq n^r.\qedhere
\end{equation*}
\end{proof}
\begin{lem}
\label{lem integrability D infinity}
We have the following estimate as $t \to + \infty$:
\begin{equation*}
\esp{\odet{X\left(t\right)}\odet{Y\left(t\right)}} = (2\pi)^r \left(\frac{\vol{\S^{n-r}}}{\vol{\S^n}}\right)^2 + O\!\left(te^{-\frac{t}{2}}\right).
\end{equation*}
\end{lem}
\begin{proof}
Let $\left(X(\infty),Y(\infty)\right)$ be a standard Gaussian vector in $\mathcal{M}_{rn}(\mathbb{R})^2 \simeq \mathbb{R}^{2nr}$, i.e.~$X(\infty)$ and $Y(\infty)$ are independent standard Gaussian vectors in $\mathcal{M}_{rn}(\mathbb{R})$. Then,
\begin{align*}
\esp{\odet{X(\infty)}\odet{Y(\infty)}} &= \esp{\odet{X(\infty)}}\esp{\odet{Y(\infty)}}\\
&=\esp{\odet{X(\infty)}}^2\\
&= (2\pi)^r \left(\frac{\vol{\S^{n-r}}}{\vol{\S^n}}\right)^2,
\end{align*}
where we used~\eqref{eq value exp odet lim} to get the last equality.
Then the proof is basically the same as that of Lemma~\ref{lem estimates cond exp x}. From Definition~\ref{def XYt}, we see that the variance operator $\Lambda(t)$ of $\left(X(t),Y(t)\right)$ equals $\Id + O\!\left(te^{-\frac{t}{2}}\right)$ as $t\to +\infty$. Hence:
\begin{align*}
\det\left( \Lambda(t)\right) &= 1 + O\!\left(te^{-\frac{t}{2}}\right) & &\text{and} & \Lambda(t)^{-1} =& \Id + O\!\left(te^{-\frac{t}{2}}\right).
\end{align*}
Let $C>0$ be such that $\Norm{\Lambda(t)^{-1}- \Id} \leq C te^{-\frac{t}{2}}$. We denote by $L=\left(L_1,L_2\right)$ a generic element of $\mathcal{M}_{rn}(\mathbb{R})^2$ and by $\dmesure\! L$ the normalized Lebesgue measure on this space. Then,
\begin{multline*}
\left(2\pi\right)^{rn} \norm{\det\left(\Lambda(t)\right)^\frac{1}{2} \esp{\odet{X(t)}\odet{Y(t)}} - \esp{\odet{X(\infty)}\odet{Y(\infty)}}}\\
\begin{aligned}
&\leq \int \odet{L_1}\odet{L_2} \norm{\exp\left(-\frac{1}{2}\prsc{\left(\Lambda(t)^{-1}-\Id\right)L}{L}\right)-1}e^{-\frac{1}{2}\Norm{L}^2} \dmesure\! L\\
&\leq \frac{C}{2}te^{-\frac{t}{2}} \int \odet{L_1}\odet{L_2}\Norm{L}^2\exp\left(-\frac{1}{2}\left(1-\frac{C}{2}te^{-\frac{t}{2}}\right)\Norm{L}^2\right) \dmesure\! L\\
&= O\!\left(te^{-\frac{t}{2}}\right).
\end{aligned}
\end{multline*}
Thus
\begin{align*}
\esp{\odet{X(t)}\odet{Y(t)}}&\\
=\det&\left( \Lambda(t)\right)^{-\frac{1}{2}}\left(\esp{\odet{X(\infty)}\odet{Y(\infty)}}+ O\!\left(te^{-\frac{t}{2}}\right)\right)\\
= (2\pi&)^r \left(\frac{\vol{\S^{n-r}}}{\vol{\S^n}}\right)^2 + O\!\left(te^{-\frac{t}{2}}\right).\qedhere
\end{align*}
\end{proof}
\begin{dfn}
\label{def D}
Let $D_{n,r} : (0,+\infty) \to \mathbb{R}$ be the function defined by:
\begin{equation*}
\forall t \in (0,+\infty), \qquad D_{n,r}(t) = \frac{\esp{\odet{X(t)}\odet{Y(t)}}}{\left(1-e^{-t}\right)^\frac{r}{2}} - (2\pi)^r \left(\frac{\vol{\S^{n-r}}}{\vol{\S^n}}\right)^2.
\end{equation*}
\end{dfn}
\begin{lem}
\label{lem integrability Dnr}
We have:
\begin{equation*}
\int_0^{+\infty} \norm{D_{n,r}(t)}t^{\frac{n-2}{2}} \dmesure\! t < +\infty .
\end{equation*}
\end{lem}
\begin{proof}
We first check the integrability of $\norm{D_{n,r}(t)}t^{\frac{n-2}{2}}$ at $t=0$. By Lemma~\ref{lem cond exp bounded}, about $t=0$ we have:
\begin{align*}
\norm{D_{n,r}(t)} t^{\frac{n-2}{2}} &\leq t^{\frac{n-2}{2}}\frac{\esp{\odet{X(t)}\odet{Y(t)}}}{\left(1-e^{-t}\right)^\frac{r}{2}} + t^{\frac{n-2}{2}}(2\pi)^r \left(\frac{\vol{\S^{n-r}}}{\vol{\S^n}}\right)^2\\
&\leq t^{\frac{n-2}{2}}\frac{n^r}{\left(1-e^{-t}\right)^\frac{r}{2}} + O(t^\frac{n-2}{2}) = O\!\left(t^\frac{n-2-r}{2}\right).
\end{align*}
And this is integrable at $t=0$ since $n-r \geq 1$.
Then, by Lemma~\ref{lem integrability D infinity}, we have: $\norm{D_{n,r}(t)}t^{\frac{n-2}{2}} = O\!\left(t^\frac{n}{2}e^{-\frac{t}{2}}\right)$ when $t$ goes to infinity. This proves the integrability at infinity.
\end{proof}
\subsubsection{Near-diagonal asymptotics for the correlated terms}
\label{subsub near diag correlated}
The next step of the proof is to compute the contribution of the integral~\eqref{eq variance2} on $\Delta_d$. Let $R >0$ be such that $2R$ is smaller than the injectivity radius of $\mathcal{X}$, as in Section~\ref{sec estimates for the bergman kernel}. Let $d_3 \in \mathbb{N}$ be such that $\forall d \geq d_3$, $b_n\frac{\ln d}{\sqrt{d}}\leq R$. In the sequel we consider $d \geq \max(d_0,d_1,d_2,d_3)$.
Since we chose $d$ large enough that $b_n \frac{\ln d}{\sqrt{d}} \leq R$ we can compute everything in the exponential chart about $x$. Let $\phi_1,\phi_2 \in \mathcal{C}^0(M)$, we have:
\begin{multline}
\label{eq double integral}
\int_{\Delta_d} \phi_1(x)\phi_2(y) \mathcal{D}_d(x,y) \rmes{M}^2\\
\begin{aligned}
&= \int_{x \in M} \left(\int_{y \in B_M\left(x,b_n \frac{\ln d}{\sqrt{d}}\right)} \phi_1(x)\phi_2(y) \mathcal{D}_d(x,y) \rmes{M}\right) \rmes{M}\\
&= \int_{x \in M} \left(\int_{z \in B_{T_xM}\left(0,b_n \frac{\ln d}{\sqrt{d}}\right)} \phi_1(x)\phi_2(\exp_x(z)) \mathcal{D}_d(x,\exp_x(z)) \sqrt{\kappa(z)}\dmesure\! z \right) \rmes{M},
\end{aligned}
\end{multline}
where $\sqrt{\kappa}$ is the density of $(\exp_x)^\star\rmes{M}$ with respect to the normalized Lebesgue measure on $T_xM$ (see Sect.~\ref{subsec near diagonal estimates}). Let $x \in M$, for all $z \in B_{T_xM}\left(0,b_n \ln d\right)$ we define
\begin{equation}
\label{def Dd xz}
D_d(x,z) = \mathcal{D}_d\left(x, \exp_x\left(\frac{z}{\sqrt{d}}\right)\right),
\end{equation}
where $\mathcal{D}_d$ is defined by~\eqref{eq def density}. Then, by a change of variable in~\eqref{eq double integral},
\begin{multline}
\label{eq double int rescaled}
\int_{\Delta_d} \phi_1(x)\phi_2(y) \mathcal{D}_d(x,y) \rmes{M}^2 =\\
d^{-\frac{n}{2}}\int_{x \in M} \left(\int_{z \in B_{T_xM}(0,b_n \ln d)} \phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) D_d(x,z) \left(\kappa\left(\frac{z}{\sqrt{d}}\right)\right)^\frac{1}{2}\dmesure\! z \right) \rmes{M},
\end{multline}
and we need to compute the asymptotic of $D_d(x,z)$ as $d$ goes to infinity. We start by computing $\odet{\ev_{x,y}^d}$ when $(x,y) \in \Delta_d$.
\begin{prop}
\label{prop ultim estimate evxy}
Let $\alpha \in \left(0,\frac{1}{2r+1}\right)$, let $x \in M$ and $z \in B_{T_xM}(0,b_n\ln d)$. We denote $y=\exp_x\left(\frac{z}{\sqrt{d}}\right)$. Then we have:
\begin{equation}
\label{eq prop ultim estimate evxy}
\left(\frac{\pi}{d}\right)^{2nr} \det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)^*\right) = \left(1-e^{-\Norm{z}^2}\right)^r \left(1+ O\!\left(d^{-\alpha}\right)\right),
\end{equation}
where the error term does not depend on $(x,z)$.
\end{prop}
We will deduce Proposition~\ref{prop ultim estimate evxy} from the following two lemmas.
\begin{lem}
\label{lem near diag estimate evxy}
Let $\beta \in (0,1)$ and $d \geq d_3$, then for every $x \in M$ and $z \in B_{T_xM}(0,b_n\ln d)$, we have:
\begin{equation*}
\left(\frac{\pi}{d}\right)^{2nr} \det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)\right) = \left(1-e^{-\Norm{z}^2}\right)^r + O\!\left(d^{\beta-1}\right),
\end{equation*}
where $y$ stands for $\exp_x\left(\frac{z}{\sqrt{d}}\right)$. Moreover the error term depends on $\beta$ but not on $(x,z)$.
\end{lem}
\begin{lem}
\label{lem near diag estimate evxy integrable}
There exists $\tilde{C}>0$ such that, for all $\beta \in [0,1)$, there exists $d_\beta \in \mathbb{N}$ such that: $\forall d \geq d_\beta$, $\forall x \in M$, $\forall z \in B_{T_xM}\left(0,d^{\beta-1}\right)\setminus \{0\}$,
\begin{equation*}
\norm{\left(\frac{\pi}{d}\right)^{2nr} \det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)^*\right) \left(1-e^{-\Norm{z}^2}\right)^{-r} - 1} \leq \tilde{C} d^{\beta-1},
\end{equation*}
where $y$ stands for $\exp_x\left(\frac{z}{\sqrt{d}}\right)$.
\end{lem}
Let us assume Lemmas~\ref{lem near diag estimate evxy} and~\ref{lem near diag estimate evxy integrable} for now, and prove Prop.~\ref{prop ultim estimate evxy}.
\begin{proof}[Proof of Proposition~\ref{prop ultim estimate evxy}]
First, note that if~\eqref{eq prop ultim estimate evxy} holds for $z \in B_{T_xM}(0,b_n\ln d)\setminus \{0\}$, then the same estimate holds for $z\in B_{T_xM}(0,b_n\ln d)$ since both sides of the equality vanish when $z=0$. In the sequel we assume that $z \neq 0$.
Let $\alpha \in \left(0,\frac{1}{2r+1}\right)$, let $d \geq d_3$ and let $x\in M$. Then for any $z \in T_xM$ such that $\Norm{z} \geq d^{-\alpha}$, we have:
\begin{equation}
\label{eq norm inverse}
\left(1-e^{-\Norm{z}^2}\right)^{-r} \leq \left(1-\exp\left(-d^{-2\alpha}\right)\right)^{-r}.
\end{equation}
Since $1-e^{-t} = t\left(1-\frac{t}{2}+ O(t^2)\right)$ as $t \to 0$, there exists $\tilde{C}_0$ such that for all $t \in (0,1)$,
\begin{equation}
\label{eq def C0 tilde}
\norm{\left(1-e^{-t}\right)^{-r} - t^{-r}} \leq \tilde{C}_0 t^{1-r}.
\end{equation}
Hence, by~\eqref{eq norm inverse}, for any $d \geq d_3$, for any $x \in M$ and any $z \in T_xM$ such that $\Norm{z} \geq d^{-\alpha}$, we have:
\begin{equation}
\label{eq norm inverse 2}
\left(1-e^{-\Norm{z}^2}\right)^{-r} \leq \left(d^{2r\alpha}+\tilde{C}_0d^{(2r-2)\alpha}\right) \leq d^{2r \alpha} \left(1+\tilde{C}_0\right).
\end{equation}
Let $\beta = 1 -(2r+1)\alpha$ and $\beta' = 1 - \alpha$, then $\beta$ and $\beta' \in (0,1)$. By Lemma~\ref{lem near diag estimate evxy}, there exists $\tilde{K}_\beta >0$ such that: for all $d \geq d_3$, $\forall x \in M$, $\forall z \in B_{T_xM}(0,b_n \ln d)$,
\begin{equation*}
\norm{\left(\frac{\pi}{d}\right)^{2nr} \det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)\right) - \left(1-e^{-\Norm{z}^2}\right)^r} \leq \tilde{K}_\beta d^{\beta-1} = \tilde{K}_\beta d^{-(2r+1)\alpha},
\end{equation*}
where $y = \exp_x\left(\frac{z}{\sqrt{d}}\right)$. Then, by~\eqref{eq norm inverse 2}, we have: $\forall d \geq d_3$, $\forall x \in M$, $\forall z \in B_{T_xM}(0,b_n \ln d)$ such that $\Norm{z}\geq d^{-\alpha} = d^{\beta'-1}$,
\begin{equation*}
\norm{\left(\frac{\pi}{d}\right)^{2nr} \det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)\right)\left(1-e^{-\Norm{z}^2}\right)^{-r}-1} \leq \tilde{K}_\beta d^{-\alpha}\left(1+\tilde{C}_0\right),
\end{equation*}
Besides, let $d \geq d_{\beta'}$ and $x \in M$, then for all $z \in B_{T_xM}\left(0,d^{-\alpha}\right)\setminus \{0\}$ we have:
\begin{equation*}
\norm{\left(\frac{\pi}{d}\right)^{2nr} \det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)\right)\left(1-e^{-\Norm{z}^2}\right)^{-r}-1} \leq \tilde{C}d^{-\alpha},
\end{equation*}
by Lemma~\ref{lem near diag estimate evxy integrable}. Finally, for all $d \geq \max(d_{\beta'},d_3)$, $\forall x \in M$, $\forall z \in B_{T_xM}\left(0,b_n \ln d\right)\setminus \{0\}$, we have:
\begin{equation*}
\norm{\left(\frac{\pi}{d}\right)^{2nr} \det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)\right)\left(1-e^{-\Norm{z}^2}\right)^{-r}-1} \leq d^{-\alpha} \max\left(\tilde{C},2\tilde{K}_\beta \left(1+\tilde{C}_0\right)\right).\qedhere
\end{equation*}
\end{proof}
\begin{proof}[Proof of Lemma~\ref{lem near diag estimate evxy}]
Let $d \geq d_3$, let $x \in M$ and let $z \in B_{T_xM}(0,b_n \ln d)$. We denote $y=\exp_x\left(\frac{z}{\sqrt{d}}\right)$. Since $\frac{\Norm{z}}{\sqrt{d}}< R$, let us write eq.~\eqref{eq variance evxy} in the real normal trivialization of $\mathcal{E} \otimes \L^d$ about~$x$ (see Sect.~\ref{subsec real normal trivialization}). We have:
\begin{equation*}
\left(\frac{\pi}{d}\right)^n \ev_{x,y}^d\left(\ev_{x,y}^d\right)^* = \left(\frac{\pi}{d}\right)^n\begin{pmatrix}
E_d(0,0) & E_d\left(0,\frac{z}{\sqrt{d}}\right) \\ E_d\left(\frac{z}{\sqrt{d}},0\right) & E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix}.
\end{equation*}
Then, by the near-diagonal estimates of Cor.~\ref{cor near diag estimates}, we have:
\begin{multline*}
\left(\frac{\pi}{d}\right)^n \begin{pmatrix}
E_d(0,0) & E_d\left(0,\frac{z}{\sqrt{d}}\right) \\ E_d\left(\frac{z}{\sqrt{d}},0\right) & E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix} =\\
\begin{pmatrix}
\Id_{(\mathcal{E}\otimes\L^d)_x} & e^{-\frac{1}{2}\Norm{z}^2}\left(\kappa\left(\frac{z}{\sqrt{d}}\right)\right)^{-\frac{1}{2}}\Id_{(\mathcal{E}\otimes\L^d)_x} \\
e^{-\frac{1}{2}\Norm{z}^2}\left(\kappa\left(\frac{z}{\sqrt{d}}\right)\right)^{-\frac{1}{2}}\Id_{(\mathcal{E}\otimes\L^d)_x} & \left(\kappa\left(\frac{z}{\sqrt{d}}\right)\right)^{-1}\Id_{(\mathcal{E}\otimes\L^d)_x}
\end{pmatrix} + O\left(\frac{(\ln d)^{2n+8}}{d}\right),
\end{multline*}
where the error term does not depend on $(x,z)$. Recall that $\kappa$ satisfies~\eqref{eq estimate kappa}. Hence for all $z \in B(0,b_n \ln d)$,
\begin{equation*}
\kappa\left(\frac{z}{\sqrt{d}}\right) = 1 + O\left(\frac{(\ln d)^2}{d}\right),
\end{equation*}
uniformly in $x$ and $z$. Let $\beta \in (0,1)$, then we have:
\begin{equation}
\label{eq odet near diag}
\left(\frac{\pi}{d}\right)^n \ev_{x,y}^d\left(\ev_{x,y}^d\right)^* = \begin{pmatrix}
1 & e^{-\frac{1}{2}\Norm{z}^2} \\ e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix} \otimes \Id_{(\mathcal{E}\otimes\L^d)_x} + O\left(d^{\beta-1}\right),
\end{equation}
and the constant in the term $O\!\left(d^{\beta-1}\right)$ does not depend on $(x,z)$. Since the dominant term on the right-hand side of~\eqref{eq odet near diag} has bounded coefficients, we get the result by taking the determinant of~\eqref{eq odet near diag}.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{lem near diag estimate evxy integrable}]
Let $d\geq \max(d_0,d_3)$ and let $x \in M$. Recall that $D^k_{(z,w)}$ denotes the $k$-th differential at $(z,w)$ of a map from $T_x\mathcal{X} \times T_x\mathcal{X}$ to $\End\left(\left(\mathcal{E} \otimes \L^d\right)_x\right)$.
The Chern connection reads $D + \mu^d_x$ in the real normal trivialization about $x$, where $\mu^d_x$ is a $1$-form on $B_{T_x\mathcal{X}}(0,2R)$. By definition of the real normal trivialization, $\mu^d_x(0)=0$. Besides $\mu^d_x(z)$ is a smooth function of $(x,z)$ and grows at most linearly in $d$. By compactness of $M$, there exist $A$ and $B>0$ such that $\Norm{\mu^d_x\left(\frac{z}{\sqrt{d}}\right)}\leq A + B\sqrt{d}$ for all $x \in M$ and all $z \in B_{T_x\mathcal{X}}(0,R)$. Hence, there exists $K_1 >0$ independent of $x$ such that, for any smooth section $S$ of $\mathbb{R}\left(\mathcal{E} \otimes \L^d\right) \boxtimes \mathbb{R}\left(\mathcal{E} \otimes \L^d\right)^*$ over $B_{T_xM}(0,R)\times B_{T_xM}(0,R)$, we have:
\begin{equation*}
\Norm{D_{(z,w)}\left(S\left(\frac{z}{\sqrt{d}},\frac{w}{\sqrt{d}}\right)\right)} \leq K_1 \Norm{S\left(\exp_x\left(\frac{z}{\sqrt{d}}\right),\exp_x\left(\frac{w}{\sqrt{d}}\right)\right)}_{\mathcal{C}^1},
\end{equation*}
where $\Norm{\cdot}_{\mathcal{C}^1}$ was defined in Section~\ref{subsec far off diagonal estimates}. Since we use the exponential chart, we can argue similarly for the Levi--Civita connection. This gives a similar result for the higher derivatives of $S$. For all $k \in \mathbb{N}$, there exists $K_k>0$ independent of $x$ such that, for any smooth section $S$ of $\mathbb{R}\left(\mathcal{E} \otimes \L^d\right) \boxtimes \mathbb{R}\left(\mathcal{E} \otimes \L^d\right)^*$ over $B_{T_xM}(0,R)\times B_{T_xM}(0,R)$, we have:
\begin{equation}
\label{eq def Kk}
\Norm{D^k_{(z,w)}\left(S\left(\frac{z}{\sqrt{d}},\frac{w}{\sqrt{d}}\right)\right)} \leq K_k \Norm{S\left(\exp_x\left(\frac{z}{\sqrt{d}}\right),\exp_x\left(\frac{w}{\sqrt{d}}\right)\right)}_{\mathcal{C}^k}.
\end{equation}
Since $d \geq d_0$, by eq.~\eqref{eq def Kk} and Thm.~\ref{thm off diag estimates} we have: $
\forall z,w \in B_{T_xM}(0,R)$,
\begin{equation}
\label{eq estimate D2Ed}
\Norm{D^2_{(z,w)} \left(E_d\left(\frac{z}{\sqrt{d}},\frac{w}{\sqrt{d}}\right)\right)} \leq K_2 \Norm{E_d\left(\exp_x\left(\frac{z}{\sqrt{d}}\right),\exp_x\left(\frac{w}{\sqrt{d}}\right)\right)}_{\mathcal{C}^2} \leq C_2 K_2 d^{n+1}.
\end{equation}
Let $x \in M$ and $z \in B_{T_xM}(0,b_n \ln d)\setminus\{0\}$. We denote $y=\exp_x\left(\frac{z}{\sqrt{d}}\right)$. Let us write eq.~\eqref{eq variance evxy}, in the real normal trivialization of $\mathcal{E} \otimes \L^d$ about~$x$, as in the proof of Lem.~\ref{lem near diag estimate evxy}. We have:
\begin{equation*}
\ev_{x,y}^d\left(\ev_{x,y}^d\right)^* = \begin{pmatrix}
E_d(0,0) & E_d\left(0,\frac{z}{\sqrt{d}}\right) \\ E_d\left(\frac{z}{\sqrt{d}},0\right) & E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix}.
\end{equation*}
Then, by elementary operations on rows and columns,
\begin{multline}
\label{eq elementery operations}
\frac{1}{\Norm{z}^{2r}}\det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)^*\right) = \frac{1}{\Norm{z}^{2r}}\det \begin{pmatrix}
E_d(0,0) & E_d\left(0,\frac{z}{\sqrt{d}}\right) \\ E_d\left(\frac{z}{\sqrt{d}},0\right) & E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix}\\
=\det \begin{pmatrix}
E_d(0,0) & \frac{1}{\Norm{z}}\left(
E_d\left(0,\frac{z}{\sqrt{d}}\right)-E_d(0,0)\right) \\ \frac{1}{\Norm{z}}\left(
E_d\left(\frac{z}{\sqrt{d}},0\right)-E_d(0,0)\right) & \frac{1}{\Norm{z}^2}\left(\begin{aligned}
E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)&-E_d\left(\frac{z}{\sqrt{d}},0\right)\\-E_d&\left(0,\frac{z}{\sqrt{d}}\right)+E_d(0,0)
\end{aligned}\right)
\end{pmatrix}.
\end{multline}
By Taylor's formula, for all $z \in B_{T_xM}(0,b_n \ln d)\setminus \{0\}$ we have:
\begin{multline}
\label{eq Taylor 1}
\Norm{E_d\left(0,\frac{z}{\sqrt{d}}\right)-E_d(0,0)-D_{(0,0)}E_d\cdot \left(0,\frac{z}{\sqrt{d}}\right)}\leq\\
\frac{\Norm{z}^2}{2d} \left(\sup_{w \in [0,z]}\Norm{D^2_{(0,w)} E_d\left(\exp_x\left(\frac{\cdot}{\sqrt{d}}\right),\exp_x\left(\frac{\cdot}{\sqrt{d}}\right)\right)}\right).
\end{multline}
Then, by~\eqref{eq estimate D2Ed}, we have:
\begin{equation}
\label{eq Taylor 2}
\left(\frac{\pi}{d}\right)^n \frac{1}{\Norm{z}}\Norm{E_d\left(0,\frac{z}{\sqrt{d}}\right)-E_d(0,0)-D_{(0,0)}E_d\cdot \left(0,\frac{z}{\sqrt{d}}\right)} \leq \Norm{z} C_2 K_2 \pi^n.
\end{equation}
Similarly, for all $z \in B_{T_xM}(0,b_n \ln d)\setminus \{0\}$ we have:
\begin{equation}
\label{eq Taylor 3}
\left(\frac{\pi}{d}\right)^n \frac{1}{\Norm{z}}\Norm{E_d\left(\frac{z}{\sqrt{d}},0\right)-E_d(0,0)-D_{(0,0)}E_d\cdot \left(\frac{z}{\sqrt{d}},0\right)} \leq \Norm{z} C_2 K_2 \pi^n.
\end{equation}
A second order Taylor's formula gives:
\begin{multline*}
\Norm{\left(E_d(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}})-E_d(\frac{z}{\sqrt{d}},0)-E_d(0,\frac{z}{\sqrt{d}})+E_d(0,0)\right) - D^2_{(0,0)}E_d \left((0,\frac{z}{\sqrt{d}}),(\frac{z}{\sqrt{d}},0)\right)}\\
\leq \left(\frac{\Norm{z}}{\sqrt{d}}\right)^3 \left(\sup_{w \in [0,z]}\Norm{D^3_{(0,w)} E_d\left(\exp_x\left(\frac{\cdot}{\sqrt{d}}\right),\exp_x\left(\frac{\cdot}{\sqrt{d}}\right)\right)}\right),
\end{multline*}
and since $d \geq d_0$, by Thm.~\ref{thm off diag estimates} and eq.~\eqref{eq def Kk} we have:
\begin{multline}
\label{eq Taylor 4}
\left(\frac{\pi}{d}\right)^n \frac{1}{\Norm{z}^2}\Norm{\left(E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)-E_d\left(\frac{z}{\sqrt{d}},0\right)-E_d\left(0,\frac{z}{\sqrt{d}}\right)+E_d(0,0)\right) -\right.\\
\left.D^2_{(0,0)}E_d \left(\left(0,\frac{z}{\sqrt{d}}\right)\left(\frac{z}{\sqrt{d}},0\right)\right)}\leq \Norm{z} C_3 K_3 \pi^n.
\end{multline}
Finally, by equations~\eqref{eq Taylor 2}, \eqref{eq Taylor 3} and~\eqref{eq Taylor 4},
\begin{equation}
\label{matrix distance 1}
\begin{aligned}
\left(\frac{\pi}{d}\right)^n & \begin{pmatrix}
E_d(0,0) & \frac{1}{\Norm{z}}\left(
E_d\left(0,\frac{z}{\sqrt{d}}\right)-E_d(0,0)\right) \\ \frac{1}{\Norm{z}}\left(
E_d\left(\frac{z}{\sqrt{d}},0\right)-E_d(0,0)\right) & \frac{1}{\Norm{z}^2}\left(\begin{aligned}
E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)&-E_d\left(\frac{z}{\sqrt{d}},0\right)\\-E_d&\left(0,\frac{z}{\sqrt{d}}\right)+E_d(0,0)
\end{aligned}\right)
\end{pmatrix}\\
&=\left(\frac{\pi}{d}\right)^n \begin{pmatrix}
E_d(0,0) & \frac{1}{\Norm{z}}D_{(0,0)}E_d\cdot \left(0,\frac{z}{\sqrt{d}}\right) \\ \frac{1}{\Norm{z}}D_{(0,0)}E_d\cdot \left(\frac{z}{\sqrt{d}},0\right) & \frac{1}{\Norm{z}^2} D^2_{(0,0)}E_d \left(\left(0,\frac{z}{\sqrt{d}}\right)\left(\frac{z}{\sqrt{d}},0\right)\right)
\end{pmatrix} + O\!\left(\Norm{z}\right),
\end{aligned}
\end{equation}
where the error term is uniform in $x$ and $d$.
On the other hand, for every $x \in M$ and every $z \in T_xM\setminus \{0\}$, the diagonal estimates of Sect.~\ref{subsec diagonal estimates} give (see~\eqref{eq diag values in chart}):
\begin{align*}
\left(\frac{\pi}{d}\right)^n \frac{1}{\Norm{z}^2} D^2_{(0,0)}E_d \left(\left(0,\frac{z}{\sqrt{d}}\right)\left(\frac{z}{\sqrt{d}},0\right)\right) &= \frac{\pi^n}{d^{n+1}} D^2_{(0,0)}E_d \left(\left(0,\frac{z}{\Norm{z}}\right)\left(\frac{z}{\Norm{z}},0\right)\right)\\
&= \Id_{\mathbb{R}(\mathcal{E}\otimes\L^d)_x} + O(d^{-1}),
\end{align*}
where the error term is independent of $x$ and $z$. Similarly,
\begin{align*}
\left(\frac{\pi}{d}\right)^n \frac{1}{\Norm{z}}D_{(0,0)}E_d\cdot \left(0,\frac{z}{\sqrt{d}}\right) &=\left(\frac{\pi}{d}\right)^n \frac{1}{\sqrt{d}}D_{(0,0)}E_d\cdot \left(0,\frac{z}{\Norm{z}}\right) = O(d^{-1}),\\
\left(\frac{\pi}{d}\right)^n \frac{1}{\Norm{z}}D_{(0,0)}E_d\cdot \left(\frac{z}{\sqrt{d}},0\right) &=\left(\frac{\pi}{d}\right)^n \frac{1}{\sqrt{d}}D_{(0,0)}E_d\cdot \left(\frac{z}{\Norm{z}},0\right) = O(d^{-1}),\\
\intertext{and}
\left(\frac{\pi}{d}\right)^n E_d(0,0) &= \Id_{\mathbb{R}(\mathcal{E}\otimes\L^d)_x} + O(d^{-1}).
\end{align*}
Thus
\begin{equation}
\label{matrix distance 2}
\left(\frac{\pi}{d}\right)^n \begin{pmatrix}
E_d(0,0) & \frac{1}{\Norm{z}}D_{(0,0)}E_d\cdot \left(0,\frac{z}{\sqrt{d}}\right) \\ \frac{1}{\Norm{z}}D_{(0,0)}E_d\cdot \left(\frac{z}{\sqrt{d}},0\right) & \frac{1}{\Norm{z}^2} D^2_{(0,0)}E_d \left(\left(0,\frac{z}{\sqrt{d}}\right)\left(\frac{z}{\sqrt{d}},0\right)\right)
\end{pmatrix} = \Id + O(d^{-1}),
\end{equation}
where the error term is uniform in $(x,z)$. By~\eqref{matrix distance 1} and~\eqref{matrix distance 2}, there exist $\tilde{C}_1$ and $\tilde{C}_2 >0$ such that we have: $\forall d \geq \max(d_0,d_3)$, $\forall x \in M$, $\forall z \in B_{T_xM}(0,b_n \ln d) \setminus \{0\}$,
\begin{multline}
\label{matrix distance 3}
\Norm{\left(\frac{\pi}{d}\right)^n \begin{pmatrix}
E_d(0,0) & \frac{1}{\Norm{z}}\left(
E_d\left(0,\frac{z}{\sqrt{d}}\right)-E_d(0,0)\right) \\ \frac{1}{\Norm{z}}\left(
E_d\left(\frac{z}{\sqrt{d}},0\right)-E_d(0,0)\right) & \frac{1}{\Norm{z}^2}\left(\begin{aligned}
E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)-E_d\left(\frac{z}{\sqrt{d}},0\right)&\\-E_d\left(0,\frac{z}{\sqrt{d}}\right)+E_d(0,0)&
\end{aligned}\right)
\end{pmatrix}-\Id\ }\\
\leq \tilde{C}_1 \Norm{z} + \tilde{C}_2 \frac{1}{d}.
\end{multline}
Let $\beta \in [0,1)$, then for all $d \geq \max(d_0,d_3)$, for all $x\in M$ and all $z \in B_{T_xM}\left(0,d^{\beta-1}\right)$, we have: $\tilde{C}_1 \Norm{z} + \tilde{C}_2 d^{-1} \leq d^{\beta-1}\left(\tilde{C}_1 + \tilde{C}_2\right)$. Let $d_\beta \in \mathbb{N}$ be such that $(d_\beta)^{\beta-1}\left(\tilde{C}_1 + \tilde{C}_2\right) \leq \frac{1}{2}$. Since the determinant is a smooth function, there exists $\tilde{C}_3>0$ such that, for every operator $\Lambda$, if $\Norm{\Lambda} \leq \frac{1}{2}$, then $\norm{\det\left(\Id + \Lambda\right) - 1} \leq \tilde{C}_3 \Norm{\Lambda}$. Hence, by eq.~\eqref{eq elementery operations} and~\eqref{matrix distance 3}, we have: for all $d\geq d_\beta$ , for all $x \in M$, for all $z \in B_{T_xM}\left(0,d^{\beta-1}\right)\setminus\{0\}$,
\begin{equation}
\norm{\frac{1}{\Norm{z}^{2r}}\left(\frac{\pi}{d}\right)^{2rn}\det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)^*\right) - 1} \leq \left(\tilde{C}_1+\tilde{C}_2\right) \tilde{C}_3 d^{\beta-1}.
\end{equation}
Recall that $\tilde{C}_0$ was defined in the proof of Prop.~\ref{prop ultim estimate evxy} (see eq.~\eqref{eq def C0 tilde}) and that, for all $x \in M$, for all $z \in B_{T_xM}(0,1)\setminus \{0\}$, we have:
\begin{equation*}
\norm{\frac{\Norm{z}^{2r}}{\left(1-e^{-\Norm{z}^2}\right)^{r}} -1} \leq \tilde{C}_0 \Norm{z}^2.
\end{equation*}
Then we have: $\forall d \geq d_\beta$, $\forall x \in M$, $\forall z \in B_{T_xM}\left(0,d^{\beta-1}\right)\setminus\{0\}$,
\begin{multline*}
\norm{\left(\frac{\pi}{d}\right)^{2nr} \det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)^*\right) \left(1-e^{-\Norm{z}^2}\right)^{-r} - 1}\\
\begin{aligned}
&= \norm{\left(\frac{\pi}{d}\right)^{2nr} \frac{1}{\Norm{z}^{2r}}\det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)^*\right) \frac{\Norm{z}^{2r}}{\left(1-e^{-\Norm{z}^2}\right)^{r}} - 1}\\
&\leq \frac{\Norm{z}^{2r}}{\left(1-e^{-\Norm{z}^2}\right)^{r}} \norm{\frac{1}{\Norm{z}^{2r}}\left(\frac{\pi}{d}\right)^{2rn}\det\left(\ev^d_{x,y}\left(\ev^d_{x,y}\right)^*\right) - 1} + \norm{\frac{\Norm{z}^{2r}}{\left(1-e^{-\Norm{z}^2}\right)^{r}} -1}\\
&\leq \left(1 + \tilde{C}_0 d^{2\beta-2}\right)\left(\tilde{C}_1+\tilde{C}_2\right) \tilde{C}_3 d^{\beta-1} + \tilde{C}_0 d^{2\beta-2}\\
&\leq d^{\beta-1} \left(\left(\tilde{C}_1+\tilde{C}_2\right) \tilde{C}_3\left(1+\tilde{C}_0\right) + \tilde{C}_0\right) = d^{\beta-1} \tilde{C},
\end{aligned}
\end{multline*}
where we define $\tilde{C}>0$ by the equality on the last line.
\end{proof}
We now want to compute the limit of the conditional distribution of $\frac{\pi^n}{d^{n+1}}\left(\nabla^d_x s_d, \nabla^d_y s_d\right)$ given that $s_d(x) = 0 = s_d(y)$ for $(x,y) \in \Delta_d$. It is enough to compute the limit of $\Lambda_d(x,y)$ as $d \to +\infty$. Recall that $\Lambda_d$ is defined by Def.~\ref{def Lambda dxy}. Since we work near the diagonal, we can write everything in the real normal trivialization centered at $x$ (see Sect.~\ref{subsec real normal trivialization}).
\begin{lem}
\label{lem near diagonal Lambda dxy 1}
Let $x \in M$ and let $\nabla^d$ be a real metric connection which is trivial over $B_{T_xM}(0,R)$ in the real normal trivialization about $x$. Let $\beta \in (0,1)$, then, in the real normal trivialization about $x$, we have: $\forall z \in B_{T_xM}(0,b_n \ln d)$,
\begin{multline*}
\frac{\pi^n}{d^{n+1}}\begin{pmatrix}
\partial_x\partial_y^\sharp E_d(0,0) & \partial_x\partial_y^\sharp E_d\left(0,\frac{z}{\sqrt{d}}\right)\\
\partial_x\partial_y^\sharp E_d\left(\frac{z}{\sqrt{d}},0\right) & \partial_x\partial_y^\sharp E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix} =\\
\begin{pmatrix}
\Id_{T^*_xM} & e^{-\frac{1}{2}\Norm{z}^2}\left(\Id_{T^*_xM}-z^* \otimes z\right) \\ e^{-\frac{1}{2}\Norm{z}^2}\left(\Id_{T^*_xM}-z^* \otimes z\right) & \Id_{T^*_xM}
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} + O\!\left(d^{\beta-1}\right),
\end{multline*}
where the error term does not depend on $(x,z)$.
\end{lem}
\begin{proof}
Let $x \in M$ and let us choose an orthonormal basis $\left(\deron{}{x_1},\dots,\deron{}{x_n}\right)$ of $T_xM$. We denote the corresponding coordinates on $T_xM \times T_xM$ by $(z_1,\dots,z_n,w_1,\dots,w_n)$ and by $\partial_{z_i}$ and $\partial_{w_j}$ the associated partial derivatives. Let $\left(dx_1,\dots,dx_n\right)$ denote the dual basis of $\left(\deron{}{x_1},\dots,\deron{}{x_n}\right)$. By definition of $\nabla^d$ and $\partial_x\partial_y^\sharp E_d$ (see eq.~\eqref{eq def sharp 2}), for all $z,w \in B_{T_xM}(0,R)$, the matrix of $\partial_x\partial_y^\sharp E_d(z,w)$ in the orthonormal basis $(dx_1,\dots,dx_n)$ is:
\begin{equation*}
\begin{pmatrix}
\partial_{z_i}\partial_{w_j}E_d(z,w)
\end{pmatrix}_{1\leq i,j \leq n}.
\end{equation*}
Note that this is a matrix with values in $\End\left(\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x\right)$. Recall that we defined the function $\xi_d$ by~\eqref{eq def xi d}. Then, by Cor.~\ref{cor near diag estimates}, for all $z,w \in B_{T_xM}(0,b_n \ln d)$, we have:
\begin{equation*}
\partial_{z_i}\partial_{w_j}E_d\left(\frac{z}{\sqrt{d}},\frac{w}{\sqrt{d}}\right) = \left(\frac{d}{\pi}\right)^n \partial_{z_i}\partial_{w_j}\xi_d\left(\frac{z}{\sqrt{d}},\frac{w}{\sqrt{d}}\right) \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} + O\!\left((\ln d)^{2n+8}\right).
\end{equation*}
Then, eq.~\eqref{eq derivative xi d zw} shows that:
\begin{align*}
\partial_{z_i}\partial_{w_j}&\xi_d\left(\frac{z}{\sqrt{d}},\frac{w}{\sqrt{d}}\right) = \exp\left(-\frac{1}{2}\Norm{z-w}^2\right)\kappa\left(\frac{z}{\sqrt{d}}\right)^{-\frac{1}{2}}\kappa\left(\frac{w}{\sqrt{d}}\right)^{-\frac{1}{2}} \times\\
&\left(d\delta_{ij}-d(z_i-w_i)(z_j-w_j)-\frac{\sqrt{d}(z_j-w_j)\partial_{z_i}\kappa\left(\frac{z}{\sqrt{d}}\right)}{2\kappa\left(\frac{z}{\sqrt{d}}\right)}+\frac{\sqrt{d}(z_i-w_i)\partial_{w_j}\kappa\left(\frac{w}{\sqrt{d}}\right)}{2\kappa\left(\frac{w}{\sqrt{d}}\right)}\right)\\
&= d\exp\left(-\frac{1}{2}\Norm{z-w}^2\right)\left(\delta_{ij}-(z_i-w_i)(z_j-w_j)\right)+ O\!\left((\ln d)^4\right),
\end{align*}
where we used the fact that, uniformly in $z \in B_{T_xM}(0,b_n \ln d)$, we have:
\begin{align*}
&\kappa\left(\frac{z}{\sqrt{d}}\right) = 1 + O\!\left(\frac{(\ln d)^2}{d}\right)\\
\text{and} \ \forall i \in \{1,\dots,n\}, \qquad &\partial_{z_i}\kappa\left(\frac{z}{\sqrt{d}}\right) = O\!\left(\frac{\ln d}{\sqrt{d}}\right).
\end{align*}
Hence, for all $z,w \in B_{T_xM}(0,b_n \ln d)$, we have:
\begin{multline*}
\frac{\pi^n}{d^{n+1}}\partial_{z_i}\partial_{w_j}E_d\left(\frac{z}{\sqrt{d}},\frac{w}{\sqrt{d}}\right) =\\
\exp\left(-\frac{1}{2}\Norm{z-w}^2\right)\left(\delta_{ij}-(z_i-w_i)(z_j-w_j)\right) \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} + O\!\left(\frac{(\ln d)^{2n+8}}{d}\right),
\end{multline*}
where the error term is independent of $x,z$ and $w$. Furthermore, for any $\beta \in (0,1)$, the term $O\!\left(\frac{(\ln d)^{2n+8}}{d}\right)$ can be replaced by $O\left(d^{\beta-1}\right)$. Finally, for all $z,w \in B_{T_xM}(0,b_n \ln d)$, we have:
\begin{multline*}
\frac{\pi^n}{d^{n+1}}\partial_x\partial_y^\sharp E_d(z,w) =\\
\exp\left(-\frac{1}{2}\Norm{z-w}^2\right)\left(\Id_{T^*_xM} - (z-w)^* \otimes (z-w)\right)\otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x}+ O\left(d^{\beta-1}\right),
\end{multline*}
which yields the result.
\end{proof}
A similar proof, using Cor.~\ref{cor near diag estimates} and the expressions~\eqref{eq derivative xi d z} and~\eqref{eq derivative xi d w} for the partial derivatives of $\xi_d$ yields the following.
\begin{lem}
\label{lem near diagonal Lambda dxy 2}
Let $x \in M$ and let $\nabla^d$ be a real metric connection which is trivial over $B_{T_xM}(0,R)$ in the real normal trivialization about $x$. Let $\beta \in (0,1)$, then, in the real normal trivialization about $x$, we have: $\forall z \in B_{T_xM}(0,b_n \ln d)$,
\begin{align*}
\frac{\pi^n}{d^{n+\frac{1}{2}}}\begin{pmatrix}
\partial_x E_d(0,0) & \partial_x E_d\left(0,\frac{z}{\sqrt{d}}\right)\\
\partial_x E_d\left(\frac{z}{\sqrt{d}},0\right) & \partial_x E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix} &=
e^{-\frac{1}{2}\Norm{z}^2}\begin{pmatrix}
0 & z^* \\ - z^* & 0
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} + O\!\left(d^{\beta-1}\right)\\
\frac{\pi^n}{d^{n+\frac{1}{2}}}\begin{pmatrix}
\partial_y^\sharp E_d(0,0) & \partial_y^\sharp E_d\left(0,\frac{z}{\sqrt{d}}\right)\\
\partial_y^\sharp E_d\left(\frac{z}{\sqrt{d}},0\right) & \partial_y^\sharp E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix} &=
e^{-\frac{1}{2}\Norm{z}^2}\begin{pmatrix}
0 & -z \\ z & 0
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} + O\!\left(d^{\beta-1}\right),
\end{align*}
where $z^* \in T^*_xM$ is to be understood as the constant map $t \mapsto z^*$ from $\mathbb{R}$ to $T^*_xM$ and $z \in T_xM$ is to be understood as the evaluation on $z$ from $T^*_xM$ to $\mathbb{R}$. Moreover, the error terms does not depend on $(x,z)$.
\end{lem}
We would like to get a similar asymptotic for the last term in the conditional variance operator~\eqref{eq conditional variance full}, namely:
\begin{equation*}
\begin{pmatrix}
E_d(0,0) & E_d\left(0,\frac{z}{\sqrt{d}}\right)\\
E_d\left(\frac{z}{\sqrt{d}},0\right) & E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix}^{-1}.
\end{equation*}
Unfortunately, this term is singular on $\Delta$, and this kills all hope to get a uniform estimate on $B_{T_xM}(0,b_n\ln d)\setminus \{0\}$. Instead, we obtain a uniform estimate on $B_{T_xM}(0,b_n\ln d)\setminus B_{T_xM}(0,\rho)$ for some $\rho>0$. We need to carefully check how this estimate depends on $\rho$.
\begin{lem}
\label{lem near diag ev inverse}
Let $\beta \in (0,1)$ and $\rho \in (0,1)$. Let $x \in M$ and $z \in B_{T_xM}(0,b_n \ln d)$ such that $\Norm{z} \geq \rho$. Then, in the real normal trivialization about $x$, we have:
\begin{multline*}
\left(\frac{d}{\pi}\right)^{n} \begin{pmatrix}
E_d(0,0) & E_d\left(0,\frac{z}{\sqrt{d}}\right)\\
E_d\left(\frac{z}{\sqrt{d}},0\right) & E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix}^{-1} =\\
\frac{1}{1-e^{-\Norm{z}^2}} \begin{pmatrix}
1 & -e^{-\frac{1}{2}\Norm{z}^2} \\ -e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x}\left(\Id + O\!\left(\frac{d^{\beta-1}}{1-e^{-\frac{1}{2}\rho^2}}\right)\right).
\end{multline*}
Here, the notation $O\!\left(\frac{d^{\beta-1}}{1-e^{-\frac{1}{2}\rho^2}}\right)$ means a quantity such that there exists $C>0$ and $\epsilon >0$, independent of $x,z,d$ and $\rho$, such that whenever $\frac{d^{\beta-1}}{1-e^{-\frac{1}{2}\rho^2}}\leq \epsilon$, the norm of this quantity is smaller than $C\frac{d^{\beta-1}}{1-e^{-\frac{1}{2}\rho^2}}$.
\end{lem}
\begin{proof}
By eq.~\eqref{eq variance evxy} and~\eqref{eq odet near diag}, we have:
\begin{equation*}
\left(\frac{\pi}{d}\right)^{n} \begin{pmatrix}
E_d(0,0) & E_d\left(0,\frac{z}{\sqrt{d}}\right)\\
E_d\left(\frac{z}{\sqrt{d}},0\right) & E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix} = \begin{pmatrix}
1 & e^{-\frac{1}{2}\Norm{z}^2} \\ e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} + O\!\left(d^{\beta-1}\right),
\end{equation*}
where the error term is independent of $(x,z)$. Besides,
\begin{equation}
\label{eq inverse of the limit}
\begin{aligned}
\left(\begin{pmatrix}
1 & e^{-\frac{1}{2}\Norm{z}^2} \\ e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} \right)^{-1} =&\\
\frac{1}{1-e^{-\Norm{z}^2}} &\begin{pmatrix}
1 & -e^{-\frac{1}{2}\Norm{z}^2} \\ -e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x},
\end{aligned}
\end{equation}
and the eigenvalues of
\begin{equation*}
\begin{pmatrix}
1 & e^{-\frac{1}{2}\Norm{z}^2} \\ e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x}
\end{equation*}
are $1-e^{-\frac{1}{2}\Norm{z}^2}$ and $1+e^{-\frac{1}{2}\Norm{z}^2}$, which shows that:
\begin{equation*}
\Norm{\left(\begin{pmatrix}
1 & e^{-\frac{1}{2}\Norm{z}^2} \\ e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} \right)^{-1}} \leq \frac{1}{1-e^{-\frac{1}{2}\Norm{z}^2}},
\end{equation*}
where $\Norm{\cdot}$ is the operator norm on $\End\left(\mathbb{R}^2 \otimes \mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x\right)$. Then, if $\Norm{z} \geq \rho$, we have:
\begin{equation*}
\frac{1}{1-e^{-\frac{1}{2}\Norm{z}^2}} \leq \frac{1}{1-e^{-\frac{1}{2}\rho^2}}.
\end{equation*}
Thus,
\begin{multline}
\label{eq interm 1}
\left(\frac{\pi}{d}\right)^{n} \begin{pmatrix}
E_d(0,0) & E_d\left(0,\frac{z}{\sqrt{d}}\right)\\
E_d\left(\frac{z}{\sqrt{d}},0\right) & E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix} \\
=\begin{pmatrix}
1 & e^{-\frac{1}{2}\Norm{z}^2} \\ e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} \left(\Id + O\!\left(\frac{d^{\beta-1}}{1-e^{-\frac{1}{2}\rho^2}}\right)\right).
\end{multline}
Taking the inverse of eq.~\eqref{eq interm 1}, we get:
\begin{multline*}
\left(\frac{d}{\pi}\right)^{n} \begin{pmatrix}
E_d(0,0) & E_d\left(0,\frac{z}{\sqrt{d}}\right)\\
E_d\left(\frac{z}{\sqrt{d}},0\right) & E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix}^{-1}=\\
\left(\begin{pmatrix}
1 & e^{-\frac{1}{2}\Norm{z}^2} \\ e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} \right)^{-1}\left(\Id + O\!\left(\frac{d^{\beta-1}}{1-e^{-\frac{1}{2}\rho^2}}\right)\right),
\end{multline*}
where we used the mean value inequality and the fact that the differential of $\Lambda \mapsto \Lambda^{-1}$ is bounded from above on the closed ball of center $\Id$ and radius $\frac{1}{2}$. Finally, eq.~\eqref{eq inverse of the limit} gives the result.
\end{proof}
Recall that $\Lambda_x(z)$ is defined for $x \in M$ and $z \in T_xM\setminus \{0\}$ by Def.~\ref{def Lambda z inf}. Recall also that $\Lambda_d\left(x,y\right)$ is defined by Def.~\ref{def Lambda dxy}.
\begin{lem}
\label{lem near diag asymptotic Lambda d xy}
Let $\beta \in (0,1)$ and $\rho \in (0,1)$. Let $x \in M$ and $z \in B_{T_xM}(0,b_n \ln d)$ such that $\Norm{z} \geq \rho$. We denote $y = \exp_x\left(\frac{z}{\sqrt{d}}\right)$. Let $\nabla^d$ be any real metric connection. Then, in the real normal trivialization about $x$, we have:
\begin{equation*}
\Lambda_d\left(x,y\right) = \Lambda_x(z) + O\!\left(\frac{d^{\beta-1}}{(1-e^{-\frac{1}{2}\rho^2})^2}\right),
\end{equation*}
where the constant in the error term does not depend on $(x,z)$, $d$ or $\rho$.
\end{lem}
\begin{proof}
We know that $\Lambda_d(x,y)$ does not depend on the choice of $\nabla^d$ (see Rem.\ref{rem choice of nablad}). Hence, we can compute $\Lambda_d\left(x,y\right)$ with $\nabla^d$ trivial over $B_{T_xM}(0,R)$ in the real normal trivialization of $\mathcal{E} \otimes \L^d$ about $x$.
Let $\beta \in (0,1)$ and $\rho \in (0,1)$, we apply Lemmas~\ref{lem near diagonal Lambda dxy 2} and \ref{lem near diag ev inverse} for $\frac{\beta}{2}$. Then, in the real normal trivialization about $x$, we have:
\begin{multline}
\label{eq interm 2}
\frac{\pi^n}{d^{n+1}} \begin{pmatrix}
\partial_xE_d(x,x) & \partial_xE_d(x,y)\\
\partial_xE_d(y,x) & \partial_xE_d(y,y)
\end{pmatrix}
\begin{pmatrix}
E_d(x,x) & E_d(x,y)\\
E_d(y,x) & E_d(y,y)
\end{pmatrix}^{-1}
\begin{pmatrix}
\partial_y^\sharp E_d(x,x) & \partial_y^\sharp E_d(x,y)\\
\partial_y^\sharp E_d(y,x) & \partial_y^\sharp E_d(y,y)
\end{pmatrix}\\
= \left( \begin{pmatrix}
0 & z^* \\ - z^* & 0
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} + O\!\left(d^{\frac{\beta}{2}-1}\right)\right) \left(\begin{pmatrix}
1 & e^{-\frac{1}{2}\Norm{z}^2} \\ e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix}\otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x}\right)^{-1}\times \\
\left(\Id + O\!\left(\frac{d^{\frac{\beta}{2}-1}}{1-e^{-\frac{1}{2}\rho^2}}\right)\right)\left(\begin{pmatrix}
0 & -z \\ z & 0
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} + O\!\left(d^{\frac{\beta}{2}-1}\right)\right).
\end{multline}
Since, $\rho \leq \norm{z} < b_n \ln d$, the norm of
\begin{equation*}
\left(\begin{pmatrix}
1 & e^{-\frac{1}{2}\Norm{z}^2} \\ e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix}\otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x}\right)^{-1}
\end{equation*}
is smaller than $\left(1-e^{-\frac{1}{2}\Norm{z}^2}\right)^{-1} \leq \left(1-e^{-\frac{1}{2}\rho^2}\right)^{-1}$, and the norms of the other matrices appearing in~\eqref{eq interm 2} are $O\!\left(\ln d\right)$. Hence, the expression~\eqref{eq interm 2} equals:
\begin{equation}
\label{eq conditioning uniform}
\begin{aligned}
\frac{e^{-\Norm{z}^2}}{1-e^{-\Norm{z}^2}} \begin{pmatrix}
0 & z^* \\ - z^* & 0
\end{pmatrix}\hspace{-1mm}\begin{pmatrix}
1 & -e^{-\frac{1}{2}\Norm{z}^2} \\ -e^{-\frac{1}{2}\Norm{z}^2} & 1
\end{pmatrix}\hspace{-1mm}
\begin{pmatrix}
0 & -z \\ z & 0
\end{pmatrix}\! \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x}\! + O\!\left(\frac{d^{\beta-1}}{(1-e^{-\frac{1}{2}\rho^2})^2}\right)&\\
=\frac{e^{-\Norm{z}^2}}{1-e^{-\Norm{z}^2}}
\begin{pmatrix}
z^*\otimes z & e^{-\frac{1}{2}\Norm{z}^2}z^*\otimes z \\ e^{-\frac{1}{2}\Norm{z}^2}z^*\otimes z & z^*\otimes z
\end{pmatrix} \otimes \Id_{\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x} + O\!\left(\frac{d^{\beta-1}}{(1-e^{-\frac{1}{2}\rho^2})^2}\right),&
\end{aligned}
\end{equation}
where the error term is independent of $(x,z)$. Finally, eq.~\eqref{eq conditioning uniform} and Lemma~\ref{lem near diagonal Lambda dxy 1} yield the result.
\end{proof}
\begin{lem}
\label{lem near diagonal cond exp}
Let $\beta \in (0,1)$ and $\rho \in (0,1)$. Let $x \in M$ and $z \in B_{T_xM}(0,b_n \ln d)$ such that $\Norm{z} \geq \rho$. We denote $y = \exp_x\left(\frac{z}{\sqrt{d}}\right)$. Let $\nabla^d$ be any real metric connection. Then,
\begin{multline*}
\left(\frac{\pi^n}{d^{n+1}}\right)^r \espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0} =\\
\esp{\odet{X(\Norm{z}^2)}\odet{Y(\Norm{z}^2)}} + O\!\left(f(\rho^2)^{\frac{r(n+1)}{2}+4}d^{\beta-1}\right),
\end{multline*}
where the constant in the error term does not depend on $(x,z)$, $d$ or $\rho$.
\end{lem}
\begin{proof}
Let $x \in M$ and $z \in B_{T_xM}(0,b_n \ln d)\setminus\{0\}$, let $y = \exp_x\left(\frac{z}{\sqrt{d}}\right)$ then we have:
\begin{multline*}
\left(\frac{\pi^n}{d^{n+1}}\right)^r \espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0}\\
\begin{aligned}
&=\espcond{\odet{\left(\frac{\pi^n}{d^{n+1}}\right)^\frac{1}{2}\nabla^d_{x}s_d}\!\odet{\left(\frac{\pi^n}{d^{n+1}}\right)^\frac{1}{2}\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0}\\
&=\esp{\odet{L'_d(x)}\odet{L'_d(y)}},
\end{aligned}
\end{multline*}
where $\left(L'_d(x),L'_d(y)\right)$ is a centered Gaussian vector in
\begin{equation*}
\mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x \otimes T^*_xM \oplus \mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_y \otimes T^*_yM
\end{equation*}
with variance operator $\Lambda_d(x,y)$. We can consider $\left(L'_d(x),L'_d(y)\right)$ as a random vector in $\mathbb{R}^2 \otimes \mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x \otimes T^*_xM$, via the real normal trivialization about $x$. From now on, we work in this trivialization. Let $\rho \in (0,1)$ and $\beta \in (0,1)$, we assume that $\rho \leq \Norm{z} < b_n \ln d$. Then, by Lemma~\ref{lem near diag asymptotic Lambda d xy}, we have:
\begin{equation*}
\Lambda_d\left(x,y\right) = \Lambda_x(z) + O\!\left(\frac{d^{\beta-1}}{(1-e^{-\frac{1}{2}\rho^2})^2}\right).
\end{equation*}
Moreover, by Cor.~\ref{cor Lambda z det and inverse} and Rem.~\ref{rem norm Lambda z inverse}, $\Norm{\Lambda_x(z)^{-1}} \leq f\left(\Norm{z}^2\right)\leq f(\rho^2)$. Hence, we have:
\begin{equation*}
\Lambda_d\left(x,y\right) = \Lambda_x(z) \left(\Id + O\!\left(f(\rho^2)\frac{d^{\beta-1}}{(1-e^{-\frac{1}{2}\rho^2})^2}\right)\right) = \Lambda_x(z) \left(\Id + O\!\left(f(\rho^2)^3d^{\beta-1}\right)\right),
\end{equation*}
where we used the fact that $\frac{1}{1-e^{-\frac{1}{2}\rho^2}} \leq f(\rho^2)$ (see the proof of Cor.~\ref{cor Lambda z det and inverse}). Then, we get:
\begin{equation}
\label{eq det near diag cond exp}
\det\left(\Lambda_d\left(x,y\right)\right) = \det\left(\Lambda_x(z)\right) \left(1+ O\!\left(f(\rho^2)^3 d^{\beta-1}\right)\right)
\end{equation}
and
\begin{equation*}
\Lambda_d\left(x,y\right)^{-1} = \Lambda_x(z)^{-1} \left(\Id + O\!\left(f(\rho^2)^3 d^{\beta-1}\right)\right) = \Lambda_x(z)^{-1} + O\!\left(f(\rho^2)^4 d^{\beta-1}\right).
\end{equation*}
Thus there exists $K>0$ and $\epsilon >0$ such that, whenever $f(\rho^2)^4 d^{\beta-1}\leq \epsilon$,
\begin{equation*}
\Norm{\Lambda_d(x,y)^{-1} - \Lambda_x(z)^{-1}} \leq K f(\rho^2)^4 d^{\beta-1}.
\end{equation*}
By the mean value inequality, for every $L=(L_1,L_2) \in \mathbb{R}^2 \otimes T^*_xM \otimes \mathbb{R}\!\left(\mathcal{E} \otimes \L^d\right)_x$ we have:
\begin{multline*}
\norm{\exp\left(-\frac{1}{2}\prsc{\left(\Lambda_d(x,y)^{-1}-\Lambda_x(z)^{-1}\right)L}{L}\right)-1}\\
\leq \frac{K}{2}\Norm{L}^2 f(\rho^2)^4 d^{\beta-1} \exp\left(\frac{K}{2}\Norm{L}^2f(\rho^2)^4 d^{\beta-1}\right),
\end{multline*}
whenever $f(\rho^2)^4 d^{\beta-1} \leq \epsilon$. Let $\dmesure\! L$ denote the normalized Lebesgue measure on this vector space, and recall that we defined $\left(L_x(0),L_x(z)\right)$ above (Def.~\ref{def Lx 0 z}). Then, we have:
\begin{align*}
&\begin{aligned}
(2\pi)^{nr} \left\lvert \det\left(\Lambda_d(x,y)\right)^\frac{1}{2}\right. &\esp{\odet{L'_d(x)}\odet{L'_d(y)}}\\
&-\left. \det\left(\Lambda_x(z)\right)^\frac{1}{2}\esp{\odet{L_x(0)}\odet{L_x(z)}}\right\rvert
\end{aligned}\\
&\begin{aligned}
\leq \int \odet{L_1} & \odet{L_2} \exp\left(-\frac{1}{2}\prsc{\Lambda_x(z)^{-1}L}{L}\right)\times\\
&\norm{\exp\left(-\frac{1}{2}\prsc{\left(\Lambda_d(x,y)^{-1}-\Lambda_x(z)^{-1}\right)L}{L}\right)-1} \dmesure\! L
\end{aligned}\\
&\begin{aligned}
\leq \frac{K}{2} f(\rho^2)^4 d^{\beta-1}\int \odet{L_1}& \odet{L_2} \Norm{L}^2\times\\
&\exp\left(-\frac{1}{2}\prsc{\left(\Lambda_x(z)^{-1}-\frac{K}{2} f(\rho^2)^4 d^{\beta-1}\Id\right)L}{L}\right)\dmesure\! L,
\end{aligned}
\end{align*}
whenever $f(\rho^2)^4 d^{\beta-1}\leq \epsilon$. Since $\Norm{\Lambda_x(d)} <2$ by Cor.~\ref{cor Lambda z det and inverse}, the smallest eigenvalue of $\Lambda_x(z)^{-1}$ is larger than $\frac{1}{2}$. Thus, if $f(\rho^2)^4 d^{\beta-1} \leq \frac{1}{2K}$, for every $L$ we have:
\begin{equation*}
\prsc{\left(\Lambda_x(z)^{-1}-\frac{K}{2} f(\rho^2)^4 d^{\beta-1}\Id\right)L}{L} \geq \frac{1}{4}\Norm{L}^2.
\end{equation*}
Hence, the last integral above is bounded by:
\begin{equation*}
\int \odet{L_1} \odet{L_2} \Norm{L}^2 \exp\left(-\frac{1}{8}\Norm{L}^2\right)\dmesure\! L < +\infty.
\end{equation*}
Then, we have:
\begin{multline*}
\det\left(\Lambda_d(x,y)\right)^\frac{1}{2}\esp{\odet{L'_d(x)}\odet{L'_d(y)}} =\\
\det\left(\Lambda_x(z)\right)^\frac{1}{2}\esp{\odet{L_x(0)}\odet{L_x(z)}} + O\!\left(f(\rho^2)^4 d^{\beta-1}\right),
\end{multline*}
and by~\eqref{eq det near diag cond exp}, we obtain:
\begin{multline*}
\esp{\odet{L'_d(x)}\odet{L'_d(y)}} =\\
\esp{\odet{L_x(0)}\odet{L_x(z)}} \left(1 + O\!\left(f(\rho^2)^3 d^{\beta-1}\right)\right)\\
+ \det\left(\Lambda_x(z)\right)^{-\frac{1}{2}}O\!\left(f(\rho^2)^4 d^{\beta-1}\right)\left(1 + O\!\left(f(\rho^2)^3 d^{\beta-1}\right)\right).
\end{multline*}
Since, for all $t>0$ we have (see Lem.~\ref{lem eigenvalues Lambda z}):
\begin{align*}
\frac{1}{1+e^{-\frac{1}{2}t}} & \leq 1, & \frac{1}{1-e^{-\frac{1}{2}t}} &\leq f(t) & &\text{and} & \frac{1+e^{-\frac{1}{2}t}}{1-e^{-t}+te^{-\frac{1}{2}t}} &\leq f(t),
\end{align*}
by Cor.~\ref{cor Lambda z det and inverse} we have: $\det\left(\Lambda_x(z)\right)^{-\frac{1}{2}} \leq f(\rho^2)^{\frac{r(n+1)}{2}}$. Besides, by Cor.~\ref{cor same distribution}, we have:
\begin{equation*}
\esp{\odet{L_x(0)}\odet{L_x(z)}}= \esp{\odet{X(\Norm{z}^2}\odet{Y(\Norm{z}^2}},
\end{equation*}
and by Lemma~\ref{lem cond exp bounded} this quantity is bounded from above by $n^r$. Finally, we have:
\begin{align*}
\esp{\odet{L'_d(x)}\odet{L'_d(y)}} =& \esp{\odet{X(\Norm{z}^2}\odet{Y(\Norm{z}^2}}\\
&+ O\!\left(f(\rho^2)^{4+\frac{r(n+1)}{2}} d^{\beta-1}\right).\qedhere
\end{align*}
\end{proof}
The following corollary is not necessary to the proof of Thm.~\ref{thm asymptotics variance} but is worth mentioning.
\begin{cor}
\label{cor near diagonal cond exp}
Let $\beta \in (0,1)$. Let $x \in M$ and $z \in B_{T_xM}(0,b_n \ln d)\setminus \{0\}$. We denote $y = \exp_x\left(\frac{z}{\sqrt{d}}\right)$. Let $\nabla^d$ be any real metric connection. Then, we have:
\begin{multline*}
\left(\frac{\pi^n}{d^{n+1}}\right)^r \espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0} =\\
\esp{\odet{X(\Norm{z}^2}\odet{Y(\Norm{z}^2}} + O\!\left(d^{\beta-1}\right),
\end{multline*}
where the error term depends on $z$ but not on $x$.
\end{cor}
\begin{proof}
Let us fix, $\beta$, $x$ and $z$, then we set $\rho = \Norm{z}$ and we apply Lemma~\ref{lem near diagonal cond exp}.
\end{proof}
Before we can conclude the proof of Thm.~\ref{thm asymptotics variance}, we need one last lemma.
\begin{lem}
\label{lem bounded cond exp Lambda d}
Let $x \in M$ and $z \in B_{T_xM}(0,b_n \ln d)\setminus \{0\}$. We denote $y = \exp_x\left(\frac{z}{\sqrt{d}}\right)$. Let $\beta \in (0,1)$ and let $\nabla^d$ be any real metric connection. Then, we have:
\begin{equation*}
\left(\frac{\pi^n}{d^{n+1}}\right)^r \espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0} \leq \frac{(2r)!}{r!}n^r + O\!\left(d^{\beta-1}\right),
\end{equation*}
where the error term is independent of $(x,z)$.
\end{lem}
\begin{proof}
Let $x \in M$, let $z \in B_{T_xM}(0,b_n \ln d)\setminus \{0\}$ and let $y = \exp_x\left(\frac{z}{\sqrt{d}}\right)$. As in the proof of Lem.~\ref{lem near diagonal cond exp}, let $\left(L'_d(x),L'_d(y)\right)$ be a centered Gaussian vector in $\mathbb{R}^2 \otimes \mathbb{R}\left(\mathcal{E} \otimes \L^d \right)_x \otimes T^*_xM$ whose variance operator is $\Lambda_d(x,y)$, read in the real normal trivialization about $x$. In the sequel, we work in this trivialization. We have:
\begin{multline*}
\left(\frac{\pi^n}{d^{n+1}}\right)^r \espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0}=\\ \esp{\odet{L'_d(x)}\odet{L'_d(y)}}.
\end{multline*}
The proof follows the same lines as that of Lem.~\ref{lem cond exp bounded}, the main difference being that the variance operator is not explicit. An additional difficulty comes from the fact that the estimate for $\Lambda_d(x,y)$ given by Lemma~\ref{lem near diag asymptotic Lambda d xy} is not uniform in $z \in B_{T_xM}(0,b_n \ln d)\setminus \{0\}$, hence it is useless here. Fortunately, we only need to bound its trace, which is bounded from above by that of the unconditional variance operator:
\begin{equation*}
\frac{\pi^n}{d^{n+1}}\begin{pmatrix}
\partial_x\partial_y^\sharp E_d(0,0) & \partial_x\partial_y^\sharp E_d\left(0,\frac{z}{\sqrt{d}}\right)\\
\partial_x\partial_y^\sharp E_d\left(\frac{z}{\sqrt{d}},0\right) & \partial_x\partial_y^\sharp E_d\left(\frac{z}{\sqrt{d}},\frac{z}{\sqrt{d}}\right)
\end{pmatrix},
\end{equation*}
and Lemma~\ref{lem near diagonal Lambda dxy 1} allows us to bound the latter.
By the Cauchy-Schwarz inequality,
\begin{equation}
\label{step 0}
\esp{\odet{L'_d(x)}\odet{L'_d(y)}} \leq \esp{\odet{L'_d(x)}^2}^\frac{1}{2} \esp{\odet{L'_d(y)}^2}^\frac{1}{2}.
\end{equation}
Let $\Lambda_{d,1}(x,y)$ and $\Lambda_{d,2}(x,y)$ denote the variance operators of $L'_d(x)$ and $L'_d(y)$ respectively, so that:
\begin{equation}
\label{eq def Lambda 1 and 2}
\Lambda_d(x,y) = \begin{pmatrix}
\Lambda_{d,1}(x,y) & * \\ * & \Lambda_{d,2}(x,y)
\end{pmatrix}.
\end{equation}
Let us choose orthonormal bases of $T_xM$ and $\mathbb{R}\left(\mathcal{E} \otimes \L^d\right)_x$. We denote by $\left(L'_d(x)_{ij}\right)_{\substack{1\leq i \leq r\\1\leq j\leq n}}$ the coefficients of the matrix of $L'_d(x)$ in these bases, and by $\left(L'_d(x)_{i}\right)_{1\leq i \leq r}$ its rows. As in the proof of Lem.~\ref{lem cond exp bounded}, we have:
\begin{equation}
\label{step 1}
\begin{aligned}
\odet{L'_d(x)}^2 &= \det\left(L'_d(x) \left(L'_d(x)\right)^*\right) = \det\left(\prsc{L'_d(x)_i}{L'_d(x)_j}\right)\\
& \leq \Norm{L'_d(x)_1}^2 \cdots \Norm{L'_d(x)_r}^2.
\end{aligned}
\end{equation}
Then, we have:
\begin{equation}
\label{step 2}
\begin{aligned}
\esp{\Norm{L'_d(x)_1}^2 \cdots \Norm{L'_d(x)_r}^2} &= \esp{\prod_{i=1}^r \left(\sum_{j=1}^n \left(L'_d(x)_{ij}\right)^2\right)}\\
&= \sum_{1 \leq j_1,\dots,j_r\leq n} \esp{\prod_{i=1}^r \left(L'_d(x)_{i(j_i)}\right)^2}.
\end{aligned}
\end{equation}
Let $j_1,\dots,j_r \in \{1,\dots,r\}$, we denote $X_i = L'_d(x)_{i(j_i)}$. Then, by Wick's formula (see \cite[lem.~11.6.1]{TA2007}), we have:
\begin{equation*}
\esp{\prod_{i=1}^r \left(L'_d(x)_{i(j_i)}\right)^2} = \esp{\prod_{i=1}^r (X_i)^2} = \sum_{\left(\{a_i,b_i\}\right)} \prod_{i=1}^r \esp{X_{\lfloor \frac{a_i}{2} \rfloor}X_{\lfloor \frac{b_i}{2} \rfloor}},
\end{equation*}
where we sum over all the partitions into pairs $\left(\{a_i,b_i\}\right)_{1 \leq i \leq r}$ of $\{1,\dots,2r\}$. Hence, by Cauchy-Schwarz inequality again, we get:
\begin{align*}
\esp{\prod_{i=1}^r \left(L'_d(x)_{i(j_i)}\right)^2} &\leq \sum_{\left(\{a_i,b_i\}\right)} \prod_{i=1}^r \esp{\left(X_{\lfloor \frac{a_i}{2}\rfloor}\right)^2}^\frac{1}{2} \esp{\left(X_{\lfloor \frac{b_i}{2}\rfloor}\right)^2}^\frac{1}{2}\\
& \leq \sum_{\left(\{a_i,b_i\}\right)} \prod_{k=1}^{2r} \esp{\left(X_{\lfloor \frac{k}{2}\rfloor}\right)^2}^\frac{1}{2}\\
& \leq \sum_{\left(\{a_i,b_i\}\right)} \prod_{l=1}^{r} \esp{\left(X_l\right)^2}\\
&\leq \frac{(2r)!}{2^r r!} \prod_{i=1}^{r} \esp{\left(L'_d(x)_{i(j_i)}\right)^2}.
\end{align*}
Thus, we have:
\begin{equation}
\label{step 3}
\begin{aligned}
\sum_{1 \leq j_1,\dots,j_r\leq n} \esp{\prod_{i=1}^r \left(L'_d(x)_{i(j_i)}\right)^2} &\leq \frac{(2r)!}{2^r r!} \sum_{1 \leq j_1,\dots,j_r\leq n} \prod_{i=1}^{r} \esp{\left(L'_d(x)_{i(j_i)}\right)^2}\\
&\leq \frac{(2r)!}{2^r r!}\prod_{i=1}^{r} \left(\sum_{j=1}^n \esp{\left(L'_d(x)_{ij}\right)^2}\right)\\
&\leq \frac{(2r)!}{2^r r!} \left(\sum_{i=1}^r \sum_{j=1}^n \esp{\left(L'_d(x)_{ij}\right)^2}\right)^r\\
&\leq \frac{(2r)!}{2^r r!} \tr\left(\Lambda_{d,1}(x,y)\right)^r,
\end{aligned}
\end{equation}
where $\tr$ stands for the trace operator. Finally, by~\eqref{step 1}, \eqref{step 2} and~\eqref{step 3}, we have:
\begin{equation*}
\esp{\odet{L'_d(x)}^2} \leq \frac{(2r)!}{2^r r!} \tr\left(\Lambda_{d,1}(x,y)\right)^r,
\end{equation*}
and similarly,
\begin{equation*}
\esp{\odet{L'_d(y)}^2} \leq \frac{(2r)!}{2^r r!} \tr\left(\Lambda_{d,2}(x,y)\right)^r.
\end{equation*}
Thus, by~\eqref{step 0}, we get:
\begin{equation}
\label{step 4}
\begin{aligned}
\esp{\odet{L'_d(x)}\odet{L'_d(y)}} &\leq \frac{(2r)!}{2^r r!}\tr\left(\Lambda_{d,1}(x,y)\right)^\frac{r}{2} \tr\left(\Lambda_{d,2}(x,y)\right)^\frac{r}{2}\\
&\leq \frac{(2r)!}{2^r r!} \tr\left(\Lambda_d(x,y)\right)^r.
\end{aligned}
\end{equation}
Let $\beta \in (0,1)$, by eq.~\eqref{step 4}, we only need to prove that $\tr\left(\Lambda_d(x,y)\right) \leq 2n + O\!\left(d^{\beta-1}\right)$ to complete the proof. By eq.~\eqref{eq variance evxy},
\begin{equation*}
\begin{pmatrix}
E_d(x,x) & E_d(x,y)\\ E_d(y,x) & E_d(y,y)
\end{pmatrix}
\end{equation*}
is a variance operator. Hence it is a positive symmetric operator and so is its inverse. Besides, by~\eqref{eq variance jxy}, we know that:
\begin{equation*}
\begin{pmatrix}
\partial_y^\sharp E_d(x,x) & \partial_y^\sharp E_d(x,y)\\
\partial_y^\sharp E_d(y,x) & \partial_y^\sharp E_d(y,y)
\end{pmatrix} =
\begin{pmatrix}
\partial_xE_d(x,x) & \partial_xE_d(x,y)\\
\partial_xE_d(y,x) & \partial_xE_d(y,y)
\end{pmatrix}^*.
\end{equation*}
Then, the diagonal coefficients of:
\begin{equation*}
\begin{pmatrix}
\partial_xE_d(x,x) & \partial_xE_d(x,y)\\
\partial_xE_d(y,x) & \partial_xE_d(y,y)
\end{pmatrix}
\begin{pmatrix}
E_d(x,x) & E_d(x,y)\\ E_d(y,x) & E_d(y,y)
\end{pmatrix}^{-1}
\begin{pmatrix}
\partial_y^\sharp E_d(x,x) & \partial_y^\sharp E_d(x,y)\\
\partial_y^\sharp E_d(y,x) & \partial_y^\sharp E_d(y,y)
\end{pmatrix}
\end{equation*}
are non-negative, and so is its trace. Finally, by the definition of $\Lambda_d(x,y)$ (Def.~\ref{def Lambda dxy}), we have:
\begin{equation}
\label{trace}
\tr\left(\Lambda_d(x,y)\right) \leq \frac{\pi^n}{d^{n+1}}\tr \begin{pmatrix}
\partial_x\partial_y^\sharp E_d(x,x) & \partial_x\partial_y^\sharp E_d(x,y)\\
\partial_x\partial_y^\sharp E_d(y,x) & \partial_x\partial_y^\sharp E_d(y,y)
\end{pmatrix}.
\end{equation}
Note that what we have done so far works for any choice of connection since $\Lambda_d(x,y)$ is independent of this choice. However, the right-hand side of eq.~\eqref{trace} depends on the choice $\nabla^d$. We use a real metric connection that is trivial on $B_{T_xM}(0,R)$ in the real normal trivialization about $x$. Then, by Lemma~\ref{lem near diagonal Lambda dxy 1}, we have:
\begin{equation*}
\tr\left(\Lambda_d(x,y)\right) \leq 2n + O\!\left(d^{\beta-1}\right).\qedhere
\end{equation*}
\end{proof}
\subsubsection{Conclusion of the proof}
\label{subsubsec conclusion of the proof}
We can now prove Theorem~\ref{thm asymptotics variance}.
\begin{lem}
\label{lem very short distance}
Let $\alpha >0$, let $\phi \in \mathcal{C}^0(M)$ and let $x \in M$, then we have:
\begin{multline*}
\norm{\int_{B_{T_xM}\left(0,d^{-\alpha}\right)} \phi\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) \kappa\left(\frac{z}{\sqrt{d}}\right)^\frac{1}{2} \left(\frac{1}{d^r} D_d(x,z) - D_{n,r}(\Norm{z}^2)\right) \dmesure\! z}\\
= \Norm{\phi}_\infty O\!\left(d^{(r-n)\alpha}\right),
\end{multline*}
where the error term does not depend on $x$ or $\phi$.
\end{lem}
\begin{proof}
We have:
\begin{multline*}
\norm{\int_{B_{T_xM}\left(0,d^{-\alpha}\right)} \phi\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) \kappa\left(\frac{z}{\sqrt{d}}\right)^\frac{1}{2} \left(\frac{1}{d^r} D_d(x,z) - D_{n,r}(\Norm{z}^2)\right) \dmesure\! z}\\
\leq \Norm{\phi}_\infty \left(\sup_{B_{T_xM}\left(0,b_n\frac{\ln d}{\sqrt{d}}\right)} \norm{\kappa}^\frac{1}{2} \right) \int_{B_{T_xM}\left(0,d^{-\alpha}\right)} \left(\frac{1}{d^r} \norm{D_d(x,z)} + \norm{D_{n,r}(\Norm{z}^2)}\right) \dmesure\! z.
\end{multline*}
Since $\kappa(z) = 1 + O\left(\Norm{z}^2\right)$ uniformly in $x$ (see~\eqref{eq estimate kappa}), we have:
\begin{equation*}
\sup_{B_{T_xM}\left(0,b_n \frac{\ln d}{\sqrt{d}}\right)}\norm{\kappa}^\frac{1}{2} = 1 + O\!\left(\frac{(\ln d)^2}{d}\right),
\end{equation*}
and this term is bounded. Thus, we only need to consider the integrals of $\frac{1}{d^r} \norm{D_d(x,z)}$ and $\norm{D_{n,r}(\Norm{z}^2)}$. By Lemma~\ref{lem cond exp bounded}, we have:
\begin{equation}
\label{integrability}
\begin{aligned}
\int_{B\left(0,d^{-\alpha}\right)} \norm{D_{n,r}(\Norm{z}^2)}\dmesure\! z \leq& \vol{\S^{n-1}}\int_{\rho=0}^{d^{-\alpha}} \frac{\esp{\odet{X(\rho^2)}\odet{Y(\rho^2}}}{\left(1-e^{-\rho^2}\right)^\frac{r}{2}} \rho^{n-1}\dmesure\! \rho\\
&+ (2\pi)^r \left(\frac{\vol{\S^{n-r}}}{\vol{\S^n}}\right)^2 \vol{B_{T_xM}\left(0,d^{-\alpha}\right)}\\
\leq& \frac{n^r}{2}\vol{\S^{n-1}} \int_{t=0}^{d^{-2\alpha}} \frac{t^\frac{n-2}{2}}{\left(1-e^{-t}\right)^\frac{r}{2}}\dmesure\! t + O\!\left(d^{-n\alpha}\right).
\end{aligned}
\end{equation}
Then, since there exists $C>0$ such that $\frac{t}{1-e^{-t}}\leq C$ for all $t \in (0,1]$, we get:
\begin{equation}
\label{eq integrability easy}
\int_{t=0}^{d^{-2\alpha}} \frac{t^\frac{n-2}{2}}{\left(1-e^{-t}\right)^\frac{r}{2}}\dmesure\! t \leq C \int_{t=0}^{d^{-2\alpha}} t^\frac{n-2-r}{2} \dmesure\! t = O\!\left(d^{(r-n)\alpha}\right).
\end{equation}
Hence, $\displaystyle\int_{B\left(0,d^{-\alpha}\right)} \norm{D_{n,r}(\Norm{z}^2)}\dmesure\! z = O\!\left(d^{(r-n)\alpha}\right)$. Let us denote $y = \exp_x\left(\frac{z}{\sqrt{d}}\right)$. By the definition of $D_d(x,z)$ (cf.~\eqref{def Dd xz}), we have:
\begin{align*}
\frac{1}{d^r}\norm{D_d(x,z)} &\leq \frac{1}{d^r}\frac{\espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0}}{\odet{\ev_{x,y}^d}}\\
&+ \frac{1}{d^r}\frac{\espcond{\odet{\nabla^d_{x}s_d}}{s_d(x)=0}}{\odet{\ev_x^d}}\frac{\espcond{\odet{\nabla^d_{y}s_d}}{s_d(y)=0}}{\odet{\ev_y^d}}.
\end{align*}
Then, let $\beta \in (0,1)$ and $\beta' \in \left(0,\frac{1}{2r+1}\right)$, by Prop.~\ref{prop ultim estimate evxy} and Lem.~\ref{lem bounded cond exp Lambda d} we have:
\begin{align*}
\frac{1}{d^r} \frac{\espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0}}{\odet{\ev_{x,y}^d}} &\leq \frac{\frac{(2r)!}{r!}n^r + O\!\left(d^{\beta-1}\right)}{\left(1-e^{-\Norm{z}^2}\right)^\frac{r}{2}} \left(1 + O\!\left(d^{-\beta'}\right)\right)\\
&\leq C \left(\frac{1}{1-e^{-\Norm{z}^2}}\right)^\frac{r}{2},
\end{align*}
for some large $C$. By a polar change of coordinates similar to~\eqref{integrability} and~\eqref{eq integrability easy}, we show that the integral of this term over $B_{T_xM}\left(0,d^{-\alpha}\right)$ is a $O\!\left(d^{(r-n)\alpha}\right)$. Finally, by Lem.~\ref{lem estimates odet evx} and~\ref{lem estimates cond exp x} we have:
\begin{equation*}
\frac{1}{d^r}\frac{\espcond{\odet{\nabla^d_{x}s_d}}{s_d(x)=0}}{\odet{\ev_x^d}}\frac{\espcond{\odet{\nabla^d_{y}s_d}}{s_d(y)=0}}{\odet{\ev_y^d}} = O\!\left(1\right).
\end{equation*}
Hence the integral of this term over $B_{T_xM}\left(0,d^{-\alpha}\right)$ is a $O\!\left(d^{-n \alpha}\right)$.
\end{proof}
Recall that we defined $\alpha_0 = \dfrac{n-r}{2(2r+1)(2n+1)}$ (see Ntn.~\ref{ntn alpha0}). Let us denote $\alpha_1 = \dfrac{\alpha_0}{n-r}$.
\begin{lem}
\label{lem short distance}
Let $\alpha \in (0,\alpha_1)$, let $\phi \in \mathcal{C}^0(M)$ and $x \in M$, then we have:
\begin{multline*}
\norm{\int_{d^{-\alpha}\leq \Norm{z} < b_n \ln d} \phi\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) \kappa\left(\frac{z}{\sqrt{d}}\right)^\frac{1}{2} \left(\frac{1}{d^r} D_d(x,z) - D_{n,r}(\Norm{z}^2)\right) \dmesure\! z}\\
= \Norm{\phi}_\infty O\!\left(d^{(r-n)\alpha}\right),
\end{multline*}
where the error term does not depend on $x$ or $\phi$.
\end{lem}
\begin{proof}
As in the proof of Lemma~\ref{lem very short distance}, since $\kappa^\frac{1}{2}$ is bounded on $B_{T_xM}\left(0,b_n\frac{\ln d}{\sqrt{d}}\right)$ uniformly in $x \in M$, we only need to prove that:
\begin{equation*}
\norm{\frac{1}{d^r} D_d(x,z) - D_{n,r}(\Norm{z}^2)} = O\!\left(d^{(r-n)\alpha - \alpha'}\right)
\end{equation*}
for some $\alpha' >0$. Then, since $\vol{B_{T_xM}(0,b_n\ln d)} = O\!\left((\ln d)^n\right) = O\left(d^{\alpha'}\right)$, we get the result by integrating over $B_{T_xM}(0,b_n\ln d) \setminus B_{T_xM}(0,d^{-\alpha})$.
Since $\alpha \in (0,\alpha_1)$, we have $0< n\alpha < \frac{1}{2r+1}$ and we can choose a positive $\beta \in \left(n\alpha, \frac{1}{2r+1}\right)$. Let $\beta' \in (0,1)$ be such that:
\begin{equation}
\label{eq def beta'}
1 - 2\alpha(8+r(n+1)) - \beta < \beta' < 1 - 2\alpha(8+r(n+1)) - n\alpha <1.
\end{equation}
We already know that $-\beta < -n\alpha$, so we only need to check that $0 <1-\alpha(16+2rn+2r+n)$ to ensure the existence of such a $\beta'$. This goes as follows:
\begin{equation*}
1-\alpha(16+2rn+2r+n) > 1-2\alpha_1(8+rn+n+r) = \frac{3rn+n+r-7}{(2r+1)(2n+1)}>0.
\end{equation*}
By Lemma~\ref{lem near diagonal cond exp}, for every $x \in M$ and $z \in B_{T_xM}(0,b_n\ln d)$ such that $\Norm{z}\geq d^{-\alpha}$ we have:
\begin{multline}
\label{eq near diag step 1}
\left(\frac{\pi^n}{d^{n+1}}\right)^r \espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0} =\\
\esp{\odet{X(\Norm{z}^2)}\odet{Y(\Norm{z}^2)}} + O\!\left(f(d^{-2\alpha})^{\frac{r(n+1)}{2}+4}d^{\beta'-1}\right),
\end{multline}
where, as usual, $y$ stands for $\exp_x\left(\frac{z}{\sqrt{d}}\right)$. Recall that we have: $f(t) \sim \frac{12}{t^2}$ as $t \to 0$ (cf.~Rem.~\ref{rem norm Lambda z inverse}). Then, we get:
\begin{equation*}
f(d^{-2\alpha})^{\frac{r(n+1)}{2}+4} = O\!\left(d^{2\alpha \left(8 +r(n+1)\right)}\right).
\end{equation*}
We set $\alpha' = 1 - 2\alpha \left(8 +r(n+1)\right)-\beta' - n \alpha$, so that the error term in~\eqref{eq near diag step 1} is a $O\!\left(d^{-n\alpha -\alpha'}\right)$. By~\eqref{eq def beta'}, we have $\alpha'>0$.
By Prop.~\ref{prop ultim estimate evxy}, applied for $\beta$, and eq.~\eqref{eq near diag step 1} we have:
\begin{multline}
\label{eq near diag step 2}
\frac{1}{d^r}\frac{\espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0}}{\odet{\ev_{x,y}^d}}=\\
\frac{\esp{\odet{X(\Norm{z}^2)}\odet{Y(\Norm{z}^2)}} + O\!\left(d^{-n\alpha - \alpha'}\right)}{\left(1-e^{-\Norm{z}^2}\right)^\frac{r}{2}}\left(1 + O\!\left(d^{-\beta}\right)\right),
\end{multline}
for all $x \in M$ and $z \in T_xM$ such that $d^{-\alpha}\leq \Norm{z} < b_n \ln d$. Since
\[\left(1-e^{-d^{-2\alpha}}\right)^{-\frac{r}{2}} = O\!\left(d^{r\alpha}\right),\]
and the numerator of~\eqref{eq near diag step 2} is bounded (cf.~Lemma~\ref{lem cond exp bounded}), the right-hand side of equation~\eqref{eq near diag step 2} equals:
\begin{equation*}
\frac{\esp{\odet{X(\Norm{z}^2)}\odet{Y(\Norm{z}^2)}}}{\left(1-e^{-\Norm{z}^2}\right)^\frac{r}{2}} + O\!\left(d^{(r-n)\alpha - \alpha'}\right) + O\!\left(d^{r \alpha -\beta}\right).
\end{equation*}
Moreover, $n\alpha+\alpha' = 1 - 2\alpha \left(8 +r(n+1)\right)-\beta' < \beta$ (see eq.~\eqref{eq def beta'}), so that we have:
\begin{multline*}
\frac{1}{d^r}\frac{\espcond{\odet{\nabla^d_{x}s_d}\!\odet{\nabla^d_{y}s_d}}{\ev_{x,y}^d(s_d)=0}}{\odet{\ev_{x,y}^d}}=\\
\frac{\esp{\odet{X(\Norm{z}^2)}\odet{Y(\Norm{z}^2)}}}{\left(1-e^{-\Norm{z}^2}\right)^\frac{r}{2}} + O\!\left(d^{(r-n)\alpha - \alpha'}\right).
\end{multline*}
On the other hand, by Lemmas~\ref{lem estimates odet evx} and~\ref{lem estimates cond exp x}, we have:
\begin{multline*}
\frac{1}{d^r}\frac{\espcond{\odet{\nabla^d_{x}s_d}}{s_d(x)=0}}{\odet{\ev_x^d}}\frac{\espcond{\odet{\nabla^d_{y}s_d}}{s_d(y)=0}}{\odet{\ev_y^d}} =\\
(2 \pi)^r \left(\frac{\vol{\S^{n-r}}}{\vol{\S^n}} \right)^2+ O\!\left(d^{-1}\right).
\end{multline*}
Once again, eq.~\eqref{eq def beta'} shows that $n\alpha+\alpha'<\beta<1$. A fortiori $(n-r)\alpha+\alpha' <1$. Thus, for all $x\in M$ and $z \in T_xM$ such that $d^{-\alpha}\leq \Norm{z}< b_n \ln d$, we have:
\begin{equation*}
\norm{\frac{1}{d^r} D_d(x,z) - D_{n,r}(\Norm{z}^2)} = O\!\left(d^{(r-n)\alpha - \alpha'}\right),
\end{equation*}
where $\alpha' >0$ and the error term is independent of $(x,z)$.
\end{proof}
\begin{prop}
\label{prop main estimate near diag}
Let $\alpha \in (0,\alpha_0)$, let $\phi_1$ and $\phi_2 \in \mathcal{C}^0(M)$, we have the following asymptotic as $d \to +\infty$:
\begin{multline*}
\frac{1}{d^r}\int_{x \in M} \left(\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} \phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) D_d(x,z) \kappa\left(\frac{z}{\sqrt{d}}\right)^\frac{1}{2}\dmesure\! z \right) \rmes{M}\\
=\int_{x \in M} \left(\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)}\hspace{-1mm} \phi_1(x)\phi_2\left(\!\exp_x\left(\frac{z}{\sqrt{d}}\right)\!\right) D_{n,r}(\Norm{z}^2) \dmesure\! z \right) \rmes{M}\\
+ \Norm{\phi_1}_\infty\Norm{\phi_2}_\infty O\!\left(d^{-\alpha}\right),
\end{multline*}
where the error term does not depend on $\left(\phi_1,\phi_2\right)$.
\end{prop}
\begin{proof}
Let $\alpha \in (0,\alpha_0)$, we set $\alpha' = \frac{\alpha}{n-r} \in (0,\alpha_1)$. Let $\phi_1,\phi_2 \in \mathcal{C}^0(M)$ and let $x \in M$, we apply Lemmas~\ref{lem very short distance} and~\ref{lem short distance} for $\alpha'$ and $\phi_2$. Then, we have:
\begin{multline}
\label{mult1}
\frac{1}{d^r}\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} \phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) D_d(x,z) \kappa\left(\frac{z}{\sqrt{d}}\right)^\frac{1}{2}\dmesure\! z\\
=\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} \phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) D_{n,r}(\Norm{z}^2) \kappa\left(\frac{z}{\sqrt{d}}\right)^\frac{1}{2}\dmesure\! z\\
+ \norm{\phi_1(x)}\Norm{\phi_2}_{\infty} O\!\left(d^{(r-n)\alpha'}\right),
\end{multline}
and the error term can be rewritten as $O\!\left(d^{-\alpha}\right)$.
Since $\kappa(z)^\frac{1}{2} = 1 + O\left(\Norm{z}^2\right)$ (cf.~\eqref{eq estimate kappa}), there exists $C>0$ independent of $x$ such that for all $z \in B_{T_xM}(0,R)$, $\norm{\kappa(z)^\frac{1}{2}-1} \leq C\Norm{z}^2$. Then, we get:
\begin{multline*}
\norm{\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} \phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) D_{n,r}(\Norm{z}^2) \left(\kappa\left(\frac{z}{\sqrt{d}}\right)^\frac{1}{2}-1\right)\dmesure\! z}\\
\begin{aligned}
&\leq \norm{\phi_1(x)}\Norm{\phi_2}_{\infty} C\frac{(b_n \ln d)^2}{d} \int_{z \in B\left(0,b_n \ln d\right)} \norm{D_{n,r}(\Norm{z}^2)} \dmesure\! z\\
&\leq \norm{\phi_1(x)}\Norm{\phi_2}_{\infty} \frac{C}{2}\frac{(b_n \ln d)^2}{d} \vol{\S^{n-1}} \int_{t=0}^{(b_n \ln d)^2} \norm{D_{n,r}(t)}t^\frac{n-2}{2} \dmesure\! t.
\end{aligned}
\end{multline*}
Since $\norm{D_{n,r}(t)}t^\frac{n-2}{2}$ is integrable on $(0,+\infty)$ (Lem.~\ref{lem integrability Dnr}) and $\alpha <1$, we have:
\begin{multline}
\label{mult2}
\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} \phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) D_{n,r}(\Norm{z}^2) \kappa\left(\frac{z}{\sqrt{d}}\right)^\frac{1}{2}\dmesure\! z=\\
\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} \phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) D_{n,r}(\Norm{z}^2) \dmesure\! z + \norm{\phi_1(x)}\Norm{\phi_2}_{\infty} O\!\left(d^{-\alpha}\right),
\end{multline}
where the error term in independent of $x$. By~\eqref{mult1} and~\eqref{mult2}, we have:
\begin{multline*}
\frac{1}{d^r}\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} \phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) D_d(x,z) \kappa\left(\frac{z}{\sqrt{d}}\right)^\frac{1}{2}\dmesure\! z=\\
\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} \phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) D_{n,r}(\Norm{z}^2) \dmesure\! z + \norm{\phi_1(x)}\Norm{\phi_2}_{\infty} O\!\left(d^{-\alpha}\right),
\end{multline*}
uniformly in $x \in M$. Integrating this relation over $M$ yields the result.
\end{proof}
Now, let $\alpha \in (0,\alpha_0)$, let $\phi_1$ and $\phi_2 \in \mathcal{C}^0(M)$, then by eq.~\eqref{eq variance2}, Prop.~\ref{prop off diagonal is small}, eq.~\eqref{eq double int rescaled} and Prop.~\ref{prop main estimate near diag} we have:
\begin{multline}
\label{variance 3}
\var{\rmes{d}}\left(\phi_1,\phi_2\right)=\\
\frac{d^{r-\frac{n}{2}}}{(2\pi)^r} \int_{x \in M} \left(\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} \phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) D_{n,r}(\Norm{z}^2) \dmesure\! z \right) \rmes{M}\\
+ \Norm{\phi_1}_{\infty}\Norm{\phi_2}_{\infty} O\!\left(d^{r-\frac{n}{2}-\alpha}\right),
\end{multline}
where the error term is independent of $\left(\phi_1,\phi_2\right)$. Then, we have:
\begin{multline*}
\norm{\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} \left(\phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right)-\phi_1(x)\phi_2(x)\right) D_{n,r}(\Norm{z}^2) \dmesure\! z} \\
\leq \Norm{\phi_1}_\infty \varpi_{\phi_2}\left(\frac{b_n \ln d}{\sqrt{d}}\right) \int_{z \in B\left(0,b_n \ln d\right)} \norm{D_{n,r}(\Norm{z}^2)} \dmesure\! z,
\end{multline*}
where $\varpi_{\phi_2}$ is the continuity modulus of $\phi_2$ (see Def.~\ref{def continuity modulus}). Besides, by a polar change of coordinates, we have:
\begin{equation}
\label{polar coordinates}
\int_{z \in B\left(0,b_n \ln d\right)} \norm{D_{n,r}(\Norm{z}^2)} \dmesure\! z = \frac{1}{2}\vol{\S^{n-1}} \int_{t=0}^{(b_n \ln d)^2} \norm{D_{n,r}(t)}t^\frac{n-2}{2} \dmesure\! t,
\end{equation}
and this quantity is bounded, by Lemma~\ref{lem integrability Dnr}. Then,
\begin{multline}
\label{variance 3bis}
\int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} \phi_1(x)\phi_2\left(\exp_x\left(\frac{z}{\sqrt{d}}\right)\right) D_{n,r}(\Norm{z}^2) \dmesure\! z =\\
\phi_1(x)\phi_2(x) \int_{z \in B_{T_xM}\left(0,b_n \ln d\right)} D_{n,r}(\Norm{z}^2) \dmesure\! z + \Norm{\phi_1}_{\infty} \varpi_{\phi_2}\left(\frac{b_n \ln d}{\sqrt{d}}\right) O(1),
\end{multline}
where the error term is independent of $\left(\phi_1,\phi_2\right)$.
Let $\beta \in \left(0,\frac{1}{2}\right)$, then there exists $C_\beta >0$ such that for all $d \in \mathbb{N}^*$, $b_n\frac{\ln d}{\sqrt{d}} \leq C_\beta d^{-\beta}$. Since $\varpi_{\phi_2}$ is a non-decreasing function, we have $\varpi_{\phi_2}\left(b_n\frac{\ln d}{\sqrt{d}}\right) \leq \varpi_{\phi_2}\left(C_\beta d^{-\beta}\right)$. By~\eqref{variance 3}, \eqref{polar coordinates} and~\eqref{variance 3bis}, we obtain:
\begin{multline}
\label{variance 4}
\var{\rmes{d}}\left(\phi_1,\phi_2\right) =\\
d^{r-\frac{n}{2}} \frac{\vol{\S^{n-1}}}{(2\pi)^r}\left(\int_{M} \phi_1\phi_2\rmes{M}\right) \left(\frac{1}{2}\int_{t=0}^{(b_n \ln d)^2} D_{n,r}(t)t^\frac{n-2}{2} \dmesure\! t \right)\\
+ \Norm{\phi_1}_{\infty} \Norm{\phi_2}_{\infty} O\!\left(d^{r-\frac{n}{2}-\alpha}\right) + \Norm{\phi_1}_{\infty}\varpi_{\phi_2}\left(C_\beta d^{-\beta}\right) O\!\left(d^{r-\frac{n}{2}}\right).
\end{multline}
By Lemma~\ref{lem integrability D infinity}, we have: $\norm{D_{n,r}(t)} = O\!\left(te^{-\frac{t}{2}}\right)$. Then there exists some $C>0$ such that, for all $t$ large enough,
\begin{equation*}
\norm{D_{n,r}(t)} t^\frac{n-2}{2} \leq C e^{-\frac{t}{4}}.
\end{equation*}
Then, for $d$ large enough we have:
\begin{equation}
\label{variance 6}
\norm{\int_{t=(b_n \ln d)^2}^{+\infty} D_{n,r}(t)t^\frac{n-2}{2} \dmesure\! t} \leq C \int_{t=(b_n \ln d)^2}^{+\infty} e^{-\frac{t}{4}} \dmesure\! t \leq 4C \exp \left(-\frac{1}{4}b_n^2 (\ln d)^2\right) = O\!\left(d^{-1}\right).
\end{equation}
By equations~\eqref{variance 4} and~\eqref{variance 6}, we get:
\begin{multline}
\var{\rmes{d}}\left(\phi_1,\phi_2\right) = d^{r-\frac{n}{2}} \left(\int_M \phi_1\phi_2\rmes{M}\right) \frac{\vol{\S^{n-1}}}{(2\pi)^r}\left(\frac{1}{2}\int_{0}^{+\infty} \norm{D_{n,r}(t)}t^\frac{n-2}{2} \dmesure\! t \right)\\
+ \Norm{\phi_1}_{\infty}\Norm{\phi_2}_{\infty} O\!\left(d^{r-\frac{n}{2}-\alpha}\right) + \Norm{\phi_1}_{\infty}\varpi_{\phi_2}\left(C_\beta d^{-\beta}\right) O\!\left(d^{r-\frac{n}{2}}\right).
\end{multline}
Finally, recall that we defined ${I}_{n,r}$ by eq.~\eqref{eq def Inr} and $D_{n,r}$ by Def.~\ref{def D}. Hence, we have:
\begin{align*}
\mathcal{I}_{n,r} = \frac{1}{2}\int_{0}^{+\infty} \norm{D_{n,r}(t)}t^\frac{n-2}{2} \dmesure\! t,
\end{align*}
and this quantity is finite by Lemma~\ref{lem integrability Dnr}. This concludes the proof of Theorem~\ref{thm asymptotics variance}.
\section{Proofs of the corollaries}
\label{sec proofs of the corollaries}
\subsection{Proof of Corollary~\ref{cor concentration}}
\label{subsec proof of cor concentration}
Corollary~\ref{cor concentration} is a direct consequence of Thm.~\ref{thm asymptotics variance} and the Markov inequality. Let $\phi \in \mathcal{C}^0(M)$, then, by~\eqref{eq variance of linear statistics} we have:
\begin{equation*}
\var{\prsc{\rmes{d}}{\phi}} = O\!\left(d^{r-\frac{n}{2}}\right),
\end{equation*}
where the error term depends on $\phi$. Now, let $\alpha > \frac{r}{2}-\frac{n}{4}$ and $\epsilon >0$. We have:
\begin{align*}
\P\left( \norm{\prsc{\rmes{d}}{\phi}-\rule{0pt}{4mm}\esp{\prsc{\rmes{d}}{\phi}}} > d^{\alpha} \epsilon \right) &= \P\left( d^{-\alpha}\norm{\prsc{\rmes{d}}{\phi}-\esp{\prsc{\rmes{d}}{\phi}}} > \epsilon \right)\\
&\leq \frac{1}{\epsilon^2} \var{d^{-\alpha}\prsc{\rmes{d}}{\phi}}\\
&\leq \frac{1}{\epsilon^2} d^{-2\alpha}\var{\prsc{\rmes{d}}{\phi}}.
\end{align*}
\subsection{Proof of Corollary~\ref{cor connected components}}
\label{subsec proof of cor connected components}
We obtain Cor.~\ref{cor connected components} as a consequence of Cor.~\ref{cor concentration}. Let $U\subset M$ be an open subset. We denote by $\phi_U \in \mathcal{C}^0(M)$ the function such that $\phi_U(x)$ is the geodesic distance from $x$ to the complement of $U$ in $(M,g)$. Then we have:
\begin{equation*}
U = \left\{ x \in M \mathrel{}\middle|\mathrel{} \phi_U(x) >0 \right\},
\end{equation*}
and $\phi_U$ is non-negative. Hence, $Z_d \cap U = \emptyset$ if and only if $\prsc{\rmes{d}}{\phi_U}=0$. Let $\epsilon >0$ such that:
\begin{equation*}
\epsilon < \frac{1}{2}\left(\int_M \phi_U \rmes{M}\right) \frac{\vol{\S^{n-r}}}{\vol{\S^n}}.
\end{equation*}
Then, by Thm.~\ref{thm reminder expectation test function}, for $d$ large enough we have:
\begin{equation*}
d^{-\frac{r}{2}}\esp{\prsc{\rmes{d}}{\phi_U}} - \epsilon \geq \frac{1}{2}\left(\int_M \phi_U \rmes{M}\right) \frac{\vol{\S^{n-r}}}{\vol{\S^n}} >0.
\end{equation*}
Thus, for $d$ large enough, we have:
\begin{align*}
\P\left( Z_d \cap U = \emptyset \right) &= \P\left( \prsc{\rmes{d}}{\phi_U} = 0 \right)\\
&\leq \P\left( \prsc{\rmes{d}}{\phi_U} < \esp{\prsc{\rmes{d}}{\phi_U}} - d^{\frac{r}{2}}\epsilon\right)\\
&\leq \P\left( \norm{\prsc{\rmes{d}}{\phi_U}-\rule{0pt}{4mm}\esp{\prsc{\rmes{d}}{\phi_U}}} > d^{\frac{r}{2}} \epsilon \right).
\end{align*}
And by Cor.~\ref{cor concentration}, this is a $O\!\left(d^{-\frac{n}{2}}\right)$.
\subsection{Proof of Corollary~\ref{cor as convergence}}
\label{subsec proof of cor as convergence}
In this section we assume that $n \geq 3$. We consider a random sequence $(s_d)_{d \in \mathbb{N}}$ of sections of increasing degree, distributed according to the probability measure $\dmesure\! \nu = \bigotimes_{d \in \mathbb{N}} \dmesure\! \nu_d$ on $\prod_{d \in \mathbb{N}} \mathbb{R} \H$. Strictly speaking, $\rmes{s_d}$ is not defined for small $d$. However, $\dmesure\! \nu$-almost surely, $\rmes{s_d}$ is well-defined for all $d \geq d_1$, so the statement of Cor.~\ref{cor as convergence} makes sense.
Our proof follows the lines of the proof of Shiffman and Zelditch \cite[sect.~3.3]{SZ1999} in the complex case. First, we prove that for every fixed $\phi \in \mathcal{C}^0(M)$ we have:
\begin{equation}
\label{eq as convergence single function}
d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\phi} \xrightarrow[d \to +\infty]{} \frac{\vol{\S^{n-r}}}{\vol{\S^n}} \left(\int_M \phi \rmes{M}\right).
\end{equation}
Then we use a separability argument to get the result. In the complex algebraic setting of~\cite{SZ1999}, the scaled volume of $s_d^{-1}(0) \subset \mathcal{X}$ is a deterministic constant, independent of $d$. In our real algebraic setting this is not the case.
Let $\phi \in \mathcal{C}^0(M)$, then we have:
\begin{equation*}
\esp{\sum_{d \in \mathbb{N}} \left(d^{-\frac{r}{2}}\left(\prsc{\rmes{s_d}}{\phi} - \rule{0pt}{4mm}\esp{\prsc{\rmes{d}}{\phi}}\right)\right)^2} = \sum_{d \in \mathbb{N}} d^{-r} \var{\prsc{\rmes{d}}{\phi}} < +\infty,
\end{equation*}
since $d^{-r} \var{\prsc{\rmes{d}}{\phi}} = O\!\left(d^{- \frac{n}{2}}\right)$ by Cor.~\ref{cor variance of linear statistics}. Hence, $\dmesure\! \nu$-almost surely, we have:
\begin{equation*}
\sum_{d \in \mathbb{N}} \left(d^{-\frac{r}{2}}\left(\prsc{\rmes{s_d}}{\phi} - \rule{0pt}{4mm}\esp{\prsc{\rmes{d}}{\phi}}\right)\right)^2 < +\infty,
\end{equation*}
and
\begin{equation*}
\left(d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\phi} - \rule{0pt}{4mm}d^{-\frac{r}{2}}\esp{\prsc{\rmes{d}}{\phi}}\right) \xrightarrow[d \to +\infty]{} 0.
\end{equation*}
Then, by Thm.~\ref{thm reminder expectation test function}, $\prsc{\rmes{s_d}}{\phi}$ satisfies~\eqref{eq as convergence single function} $\dmesure\! \nu$-almost surely.
Let $\left(\phi_k\right)_{k \in \mathbb{N}}$ be a dense sequence in the separable space $\left(\mathcal{C}^0(M),\Norm{\cdot}_\infty\right)$. Without loss of generality, we can assume that $\phi_0 = \mathbf{1}$, the unit constant function on $M$. Then, $\dmesure\! \nu$-almost surely, we have:
\begin{equation}
\label{eq as convergence countable}
\forall k \in \mathbb{N}, \qquad d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\phi_k} \xrightarrow[d \to +\infty]{} \frac{\vol{\S^{n-r}}}{\vol{\S^n}} \left(\int_M \phi_k \rmes{M}\right).
\end{equation}
Let $\underline{s}=\left(s_d\right)_{d \in \mathbb{N}} \in \prod_{d \in \mathbb{N}} \mathbb{R} \H$ be a fixed sequence such that \eqref{eq as convergence countable} holds. For every $\phi \in \mathcal{C}^0(M)$ and $k \in \mathbb{N}$ we have:
\begin{multline*}
\norm{d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\phi} - \frac{\vol{\S^{n-r}}}{\vol{\S^n}} \left(\int_M \phi \rmes{M}\right)}\\
\begin{aligned}
\leq& \norm{d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\phi}-d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\phi_k}}\\
&+ \frac{\vol{\S^{n-r}}}{\vol{\S^n}}\norm{\int_M \phi_k \rmes{M}-\int_M \phi \rmes{M}}\\
&+ \norm{d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\phi_k} - \frac{\vol{\S^{n-r}}}{\vol{\S^n}} \left(\int_M \phi_k \rmes{M}\right)}\\
\leq& \Norm{\phi-\phi_k}_\infty \left(d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\mathbf{1}}+ \frac{\vol{\S^{n-r}}}{\vol{\S^n}}\vol{M}\right)\\
&+\norm{d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\phi_k} - \frac{\vol{\S^{n-r}}}{\vol{\S^n}} \left(\int_M \phi_k \rmes{M}\right)}.
\end{aligned}
\end{multline*}
Recall that $\phi_0 = \mathbf{1}$. Then, by~\eqref{eq as convergence countable}, the sequence $(d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\mathbf{1}})_{d \in \mathbb{N}}$ converges. Hence it is bounded by some positive constant $K_{\underline{s}}$. Let $\phi \in \mathcal{C}^0(M)$ and let $\epsilon >0$. Let $k \in \mathbb{N}$ be such that:
\begin{equation*}
\Norm{\phi - \phi_k}_\infty \leq \epsilon \left(K_{\underline{s}} +\frac{\vol{\S^{n-r}}}{\vol{\S^n}}\vol{M}\right)^{-1}.
\end{equation*}
Then, for every $d$ large enough we have:
\begin{equation*}
\norm{d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\phi_k} - \frac{\vol{\S^{n-r}}}{\vol{\S^n}} \left(\int_M \phi_k \rmes{M}\right)} \leq \epsilon,
\end{equation*}
and
\begin{equation*}
\norm{d^{-\frac{r}{2}}\prsc{\rmes{s_d}}{\phi} - \frac{\vol{\S^{n-r}}}{\vol{\S^n}} \left(\int_M \phi \rmes{M}\right)} \leq 2\epsilon.
\end{equation*}
Thus, $\phi$ satisfies~\eqref{eq as convergence single function}.
Finally, whenever \eqref{eq as convergence countable} is satisfied we have: for every $\phi \in \mathcal{C}^0(M)$, $\phi$ satisfies~\eqref{eq as convergence single function}. Since the condition~\eqref{eq as convergence countable} is satisfied $\dmesure\! \nu$-almost surely, this proves Cor.~\ref{cor as convergence}.
\bibliographystyle{amsplain}
|
\section{Introduction}
Data from observational studies provide useful information to address health-related questions
and notably estimate treatment effect in real settings \cite{concato_randomized_2000}. However, because individuals are not
randomised, the study groups are often not comparable, which may lead to confounding bias
\cite{cochran_controlling_1973} if these studies are analyzed without appropriate adjustment for confounding.
Propensity scores (PS) have been proposed
as a means to recover balance between groups on observed confounders and so
obtain a consistent estimate of the causal treatment effect \cite{rosenbaum_central_1983}.
The PS is defined as the individual's probability of receiving the treatment rather than the control given
their baseline characteristics \cite{guo_propensity_2009}. One popular method to achieve covariate
balance between treatment groups is to re-weight
individuals according to their PS value. This approach, known as inverse probability of treatment
weighting (IPTW) \cite{hirano_estimation_2001} aims to emulate the sample that would have been observed in a randomised trial. In practice, a major issue
when estimating the PS is the presence of partially observed confounders, as the PS cannot be estimated for individuals with at least one missing covariate value.
A simple solution is to perform the analysis using only data those individuals with complete records
(\textit{i.e.} with complete baseline measurements), but such an approach
can lead to a biased estimate of the treatment effect if the missingness mechanism is associated with the outcome \cite{white_bias_2010}.
Another approach, proposed by Rosenbaum and Rubin \cite{rosenbaum_reducing_1984}, is to use missingness patterns (MP)
to estimate a generalised propensity score: individuals are classified according their pattern of missingness,
and then the PS is estimated within each pattern using the fully observed confounders for these individuals. This approach has
three drawbacks: it relies on strong assumptions, some confounders are ignored for some patterns and it needs to have a sample size large enough in each stratum
to estimate the PS \cite{qu_propensity_2009}. A popular alternative to handle missing data is multiple imputation (MI).
MI is used to fill in missing confounders a given number of times, replacing the missing values by random draws from
the appropriate predictive distribution. The model parameters are estimated in the resulting imputed datasets and
then are combined using Rubin's rules to obtain the overall parameter estimates and their variance accounting for
the presence of missing data \cite{carpenter_multiple_2012}. Because the IPTW estimator is a two-step estimator
(involving models for PS estimation and treatment effect estimation, respectively), Rubin's rules could be applied in several different ways.
First, the IPTW estimator could be applied to each of the imputed datasets and then the resulting estimates averaged (MIte). Second,
each individual's PS could be averaged over the imputed datasets and then these PSs used to calculate a single IPTW estimate (MIps). Third,
the PS parameters and each individual's imputed confounders could be averaged over the imputed datasets and these used
to calculate a single IPTW estimate (MIpar).
MI of partially observed confounders in the PS model has been used in applications
\cite{hayes_using_2008,mattei_estimating_2008}, and
studied through simulations for PS matching \cite{mitra_comparison_2012,hill_reducing_2004}. However, unresolved questions remain about how
to use MI in the context of IPTW. In particular,
it is still unclear if combining the treatment effects across the imputed datasets outperforms the methods of combining the PSs,
as well as if the outcome must be included in the imputation model.
Moreover, little is known about the variance of the IPTW estimator when combining the PS parameters.
Thus, the aim of this work is to study the three different MI strategies described above to handle missing confounders for IPTW, and to compare them with
CC analysis and the MP approach.
This paper is organised as follows: we first present a motivating example looking at the effect of statins on short-term mortality after pneumonia in Section 2. A brief description of IPTW in complete data, its variance and its underlying assumptions is given in Section 3. Section 4 presents different strategies to handle partially missing confounders for PS analysis, focusing in particular on MI. We provide an approximately unbiased estimator of the variance for IPTW when combining the PS parameters. The consistency and balancing properties of the MI approaches are studied in Section 5. Section 6 and 7 present the methods and results of a simulation study assessing the performance of these methods
to estimate treatment effect with IPTW for binary outcomes. The application of these approaches to the statin motivating example is presented in Section 8, followed by a discussion in Section 9.
\section{Motivating example: effect of statin use on short-term mortality after pneumonia}
To illustrate the importance of an adequate handling of missing confounders for PS analysis,
we focused on a published study of the effect of statin use on short term mortality after pneumonia \cite{douglas_effect_2011}.
We utilised the THIN database, which consists of anonymised patient records from general practitioners (GPs) in the UK.
As of the end of 2015, the database represented 3.5 million unique active patients,
or approximately 6\% of the UK population. The database has been found to be broadly representative of the UK population,
and the validity of recorded information has been established in previous studies \cite{lewis_validation_2007,blak_generalisability_2011}.
Douglas \textit{et al} carried out an analysis of 9073 patients who had a pneumonia episode, of whom 1398 were under statin treatment
when pneumonia was diagnosed. In the statin group, 305 patients (21.8\%) died within 6 months, while 2839 (37.0\%) of the non-users
died within 6 months. However, statin users and non users were very different in terms of characteristics,
in particular on characteristics associated with mortality.
In Douglas \textit{et al}, propensity scores were used to recover balance between groups. However, three important potential
confounders were only partially observed: body mass index (BMI), smoking status and alcohol consumption,
with respectively 19.2\%, 6.2\% and 18.5\% of missing data. In the original analysis, a missing indicator method was used. This approach is similar to the missingness pattern approach
described later in this paper.
\section{Inverse probability of treatment weighting}
\subsection{Propensity score estimation and assumptions}
\label{assumptions}
Propensity scores (PS) were first introduced by Rosenbaum and Rubin in the context of observational
studies with confounding \cite{rosenbaum_central_1983}. The PS has become a major tool in causal inference to
estimate the causal effect of a binary treatment $Z$ ($Z=1$ if treated, $Z=0$ otherwise). Formally, PS is the individual's probability
of receiving the treatment conditional on the individual's characteristics. The PS is usually estimated from the data using a logistic regression
model which predicts each individual's probability of receiving the treatment from their baseline characteristics \cite{dagostino_propensity_1998}:
\begin{equation}
\hat{e}_i=\frac{exp(\boldsymbol{x}_i^T \hat{\boldsymbol{\alpha}})}{1+exp(\boldsymbol{x}_i^T \hat{\boldsymbol{\alpha}})},
\label{PS}
\end{equation}
where $\hat{e}_i$ is the estimated PS for individual $i$ $(i=1,...,n)$, $\boldsymbol{x}_i=(1,...,X_{pi})$
the vector of $p$ observed baseline confounders and the intercept for individual $i$ and $\hat{\boldsymbol{\alpha}}$ is the (p+1) vector of
maximum likelihood estimates for the PS parameters. \\
The PS approach can be viewed within the Neyman-Rubin counterfactual framework for causal inference \cite{rubin_estimating_1974}. In this framework,
the causal effect of the treatment is defined as the contrast of the two potential outcomes (counterfactuals) $Y^{Z=0}$ and $Y^{Z=1}$,
which are the outcomes that would have been observed if an individual had been not treated and treated, respectively. Three assumptions
are needed to consistently estimate the causal effect of a treatment, and thus needed for PS analysis:
\begin{enumerate}
\item Positivity assumption: each individual has a non-null probability of receiving either treatment \cite{cole_constructing_2008}
\item Stable unit treatment value assumption (SUTVA): each individual has only one possible potential
outcome value for each treatment \cite{rubin_comment:_1986}. This assumption can be divided in two parts \cite{schwartz_extending_2012}:
\begin{itemize}
\item each of the two potential outcomes is always the same for a given individual, whatever the conditions in which the
treatment has been received, \textit{i.e.} $Y=Y^z$ if $Z=z$, where $Y$ is the observed outcome. In other words, the treatment has the same effect on
the individual's outcome regardless of how the individual came to be treated. This assumption is also called consistency
\item the two potential outcome values for an individual are not affected by the treatment received by other individuals
\end{itemize}
\item Strongly ignorable treatment assignment (SITA): treatment allocation and the potential outcomes
are conditionally independent given the confounders, \textit{i.e.} $(Y^{Z=0},Y^{Z=1}) \perp Z|\boldsymbol{x}$. This
implies that there are no unmeasured confounders.
\end{enumerate}
The key property of the PS is that it is a balancing score. That is, if these assumptions are valid and the PS
model is correctly specified (the functional form of the PS is correct),
the variables included in the PS model are
balanced between treatment groups at any level of the PS. In other words, individuals with close
PS values have similar distribution of their characteristics. This balancing property
of the PS and the three assumptions lead to the consistency of PS-based estimators. Initially, Rosenbaum and Rubin
proposed three different PS-based approaches to estimate causal effects \cite{rosenbaum_central_1983}: PS matching, subclassification
(also known as stratification) and covariate adjustment. Although PS matching is the most common
approach used nowadays, two major limitations exist: matching often discards a substantial number of individual from the analysis \cite{austin_comparison_2013}
and variance estimation after PS matching is not straightforward \cite{an_bayesian_2010}. The two other approaches have
drawbacks as well: residual bias due to heterogeneity within strata can remain with subclassification \cite{rosenbaum_reducing_1984},
whereas covariate adjustment can be biased in some circumstances \cite{hade_bias_2013}. In addition, covariate adjustment provides a conditional,
rather than a marginal effect. Thus, we focus on a fourth PS-based approach \cite{rosenbaum_model-based_1987}:
inverse probability of treatment weighting (IPTW).
\subsection{IPTW estimator and its variance for complete data}
IPTW aims to create a pseudo-population similar to a randomised trial by re-weighting the individuals
according to the inverse of their probability of receiving the treatment they actually received (\textit{i.e.} $\hat{e}_i^{-1}$
for treated individuals and $(1-\hat{e}_i)^{-1}$ for untreated individuals). Thus, the IPTW estimators for the marginal
proportions for a binary outcome $Y$, $\mu_1$ and $\mu_0$, among the treated and the untreated are \cite{williamson_variance_2014}:
\begin{equation}
\hat{\mu}_1=\left( \sum_{i=1}^{n} \frac{Y_i Z_i}{\hat{e}_i}\right) \left( \sum_{i=1}^{n} \frac{Z_i}{\hat{e}_i} \right)^{-1},\mbox{~~~~~}
\hat{\mu}_0=\left( \sum_{i=1}^{n} \frac{Y_i (1- Z_i)}{1-\hat{e}_i}\right) \left( \sum_{i=1}^{n} \frac{1-Z_i}{1-\hat{e}_i} \right)^{-1}.
\label{IPTW}
\end{equation}
$Z_i$ is the treatment indicator for individual $i$ ($Z_i=1$ if treated, 0 otherwise) and $Y_i$ is their outcome value.
From these two marginal estimates, it is possible to estimate
a relative risk $\left(\frac{\hat{\mu}_1}{\hat{\mu}_0}\right)$, an odds ratio $\left(\frac{\hat{\mu}_1/(1-\hat{\mu}_1)}{\hat{\mu}_0/(1-\hat{\mu}_0)}\right)$
or a risk difference $\left(\hat{\mu}_1-\hat{\mu}_0\right)$ for a binary outcome.\\
The IPTW estimator, as with other PS-based estimators, is a "two-step estimator": a first step is needed for the PS estimation and the second step is the
treatment effect estimation. If the uncertainty linked to the PS estimation in the first step is not taken into account
in a second step, the repeated sampling variance of the treatment effect will be overestimated and inference will be conservative
\cite{an_bayesian_2010}. Lunceford and Davidian \cite{lunceford_stratification_2004} and Williamson \textit{et al.}
\cite{williamson_variance_2014} proposed a large-sample variance estimator for the IPTW treatment effect estimator in which a correction term
including the variance/covariance matrix of the estimated PS parameters is applied.
\section{Handling missing data in propensity score analysis}
A major issue in PS estimation from model (\ref{PS}) is the presence of partially observed confounders,
which can lead to bias in the treatment effect estimate if ignored.
In this section, we describe five methods for applying IPTW on incomplete data. We assume the treatment status $Z$ and
outcome $Y$ are fully observed.
\subsection{Complete case analysis}
Complete case analysis (or complete records analysis) is the most basic method in the
presence of missing data. The PS is estimated only within the subgroup of individuals
with observed values for all of the variables included in the PS model, and
only these individuals contribute to the estimation of the treatment effect
\cite{hill_reducing_2004}. Although the complete case analysis
provides an unbiased estimate of the parameters of an outcome regression model when the missingness is independent of the outcome \cite{white_bias_2010}, little
is known about complete case analysis for IPTW. Moreover,
excluding individuals with missing confounders can reduce statistical power,
particularly when the rate of missing data is high, because no use is made of partially
observed records.
\subsection{Missingness pattern approach}
The missingness pattern approach for PS analysis was first introduced by Rosenbaum
and Rubin \cite{rosenbaum_reducing_1984}. They defined a generalized PS $\hat{e}^*$ as
the probability of receiving the treatment given the observed confounders and the pattern of missing data.
In practice, the PS is estimated separately in each stratum defined by missingness patterns. For example,
if two variables $X$ and $W$ are partially observed, 4 different missingness patterns can exist:
X and W are both observed, X is observed but W is missing, X is missing and W is observed,
or X and W are both missing. In each stratum, the PS is estimated from the confounders observed
in that stratum.
Under an extension of the SITA assumption, \textit{i.e.} when treatment allocation is independent of the potential outcomes
given the observed confounders and the missingness pattern, \textit{i.e.} $(Y^{Z=0},Y^{Z=1}) \perp Z|\boldsymbol{x}_{obs},r$
where $\mathbf{x_{obs}}$ is the observed component of $\mathbf{X}$ and $r$ is the missingness pattern, the generalized propensity score balances
the observed component of the partially observed confounders and the missingness indicators.
However, this generalized PS does not balance the unobserved component of the variables \cite{dagostino_estimating_2000}.
\subsection{Multiple imputation}
\subsubsection{Principles}~\\
Multiple imputation (MI) is a popular approach to handle missing data because of its flexibility and its efficiency in
a large variety of contexts and its implementation in many statistical software packages. The principle of
MI is to generate a set of plausible values for the missing variables by drawing from the posterior predictive distribution of
these variables given the observed data. $M$ complete datasets are created and analysed
independently to produce estimates $\boldsymbol{\hat{\theta}_k}$, $(k=1,...,M)$ of $\boldsymbol{\theta}$ the vector of the parameters of interest
(\textit{eg.} regression coefficients) and estimates $\boldsymbol{W}_k$ of their associated variance matrix.
Then $\boldsymbol{\hat{\theta}_k}$ and $\boldsymbol{W}_k$ , $(k=1,...,M)$ are combined across the $M$ imputed datasets.
Following Rubin's rules, the overall estimate $\boldsymbol{\hat{\theta}_{MI}}$ of $\boldsymbol{\theta}$ and estimate of the variance
of $\boldsymbol{\hat{\theta}_{MI}}$, $\widehat{Var}(\boldsymbol{\hat{\theta}_{MI}})$,
are estimated as follows \cite{carpenter_multiple_2012}:
\begin{equation}
\boldsymbol{\hat{\theta}}_{MI}=\frac{1}{M}\sum_{k=1}^M \boldsymbol{\hat{\theta}}_k, \mbox{~~~~~~~} \widehat{Var}(\boldsymbol{\hat{\theta}}_{MI})=\boldsymbol{W}+\left(1+\frac{1}{M}\right)\boldsymbol{B},
\label{Rubins}
\end {equation}
where $\boldsymbol{W}$ is the within-imputation variance-covariance matrix, which reflects the variability of the parameter estimates in each imputed dataset,
and $\boldsymbol{B}$ is the between-imputation variance matrix reflecting the variability in the estimates caused by the missing information.
These two components are defined as:
\begin{equation}
\boldsymbol{W}=\frac{1}{M}\sum_{k=1}^M \boldsymbol{W}_k, \mbox{~~~~~~~} \boldsymbol{B}=\frac{1}{M-1}\sum_{k=1}^M \left( \boldsymbol{\hat{\theta}}_k-\boldsymbol{\hat{\theta}}_{MI}\right)^2.
\end {equation}
Two approaches exist to impute datasets: joint modelling
and chained equations. Joint modelling consists of imputing missing values
from a common parametric joint model for the complete data (often
a multivariate normal model) \cite{he_missing_2010}. However, when variables are of different types,
the existing classes of joint models may not be appropriate. The second approach,
chained equations (also known as fully conditional specification) is more flexible
in case of different variable types since a specific imputation model is
specified for each partially observed variable \cite{azur_multiple_2011}.
Because of its flexibility, we focused on this approach in the context of PS.
\subsubsection{How to apply Rubin's rules for PS analysis}~\\
Hayes and Groner used MI to estimate the PS from partially observed confounders
in a study looking at the effect of seat belt usage on injury severity \cite{hayes_using_2008}. However,
instead of combining information from their 15 imputed datasets, they randomly selected
one complete record per individual and then estimated the PS. Because this approach does not
keep the entire benefit of MI, two approaches to combine information from the imputed datasets have been proposed
in the context of propensity score analyses: applying Rubin's rules on the treatment effect
or applying Rubin's rules on the PS itself. The former approach is the natural MI
approach in which the parameter which is combined is also the parameter of interest: a PS
model is fitted on each imputed dataset, and the resulting PSs used to estimate a treatment effect.
Then formula (\ref{Rubins}) is applied to obtain an overall treatment effect and its variance
\cite{mitra_comparison_2012,hill_reducing_2004}. Seaman and White \cite{seaman_inverse_2014}
showed that Rubin's rule for estimating the variance performs well in practice in this setting, although theoretical justification for Rubin's rules
relies on the parameter of interest being estimated with maximum likelihood, which is not the case for the IPTW method. We will refer to this
approach as MIte hereafter. The second approach used in the literature consists of
combining each individual's PS across the imputed datasets (to obtain an average PS for each individual) and
then using these PSs to estimate a single treatment effect estimate \cite{mitra_comparison_2012,hill_reducing_2004}.
This method is called MIps hereafter. Because the PS is unlikely to follow a normal distribution, taking the average
PS may not be appropriate. Thus, we
propose a third combination method for PS analysis after MI, in which the PS parameter estimates $\hat{\boldsymbol{\alpha}}$
(ie.the regression coefficients of the PS model) are combined rather than the PS itself. Then
these parameters are used to estimate the PS corresponding to an individual's average imputed confounder values:
\begin{equation}
\hat{e}_i\left(\boldsymbol{\bar{x}}\right)=\frac{exp^{\boldsymbol{\bar{\alpha}\bar{x}_i}}}{1+exp^{\boldsymbol{\bar{\alpha}\bar{x}_i}}},
\end{equation}
$(i=1,...,n)$ with $\boldsymbol{\bar{\alpha}}$ the $(p+1)$ vector of the average PS parameter values (for the p confounders and the intercept) and $\boldsymbol{\bar{x}_i}$
a $(p+1)$ vector of the average $p$ confounders across imputed datasets for individual $i$. This method will be referred to as MIpar.
Whereas the MIps estimate of treatment effect is based on the average PS, $\bar{e}_i\left(\boldsymbol{x}_i\right)$,
the MIpar estimate is based on the PS of the average confounders, $e_i\left(\boldsymbol{\bar{x}}\right)$.
The 3 MI approaches are illustrated in Figure \ref{methods_MI}.\\
Because the PS is obtained from $M$ imputations, the standard variance estimator for the IPTW treatment effect for MIps and MIpar is no longer valid
since it does not take into account the presence of missing data. A large-sample estimate
of the variance for MIpar, derived from \cite{williamson_variance_2014}, is detailed in Appendix 1.
The question of which PS strategy is preferable, as well as how to estimate the variance for MIps is still unresolved.
Moreover, the MIte and MIps methods have mainly been studied for PS matching \cite{mitra_comparison_2012,hill_reducing_2004}, suggesting a better reduction
of the bias observed in CC analysis with MIps than with MIte. Qu and Lipkovitch \cite{qu_propensity_2009} and Seaman and White \cite{seaman_inverse_2014}
assessed the performance of MIte additionally including the missingness pattern indicator in the PS model, but they did not
compare this approach to other combination rules after MI. For PS stratification
Crowe \textit{et al.} \cite{crowe_comparison_2010} showed empirically that MIte has good statistical properties. However,
little is known about performance of MIte and MIps in the context of IPTW. \\
To our knowledge,
the only simulation study comparing MIte and MIps with MI for IPTW did not take into account
the outcome in the imputation model \cite{mitra_comparison_2012}, which could explain the bias observed in their study
both for MIte and MIps. When using the PS approach, the fitting of the PS regression model is a first step to estimating
the quantity which is of principal interest, namely the treatment effect. An attraction of this approach is that it enables
the choice of how to adjust for confounding to be made without needing data on the outcome: choosing and fitting the PS model
requires data only on covariates and treatment. This helps the user to avoid the temptation to search for a PS model that
gives a significant treatment effect estimate. Intuition may therefore lead one to believe that imputation of missing
covariates should also be done without using the outcome variable \cite{mitra_comparison_2012}. However, this intuition conflicts with advice to include
the outcome when imputing missing covariates in a regression model whose parameters are the quantities of interest
\cite{moons_using_2006}. In this setting, excluding the outcome causes the parameter estimates to be biased towards the null.
The question remains, therefore, whether or not the outcome should be included when imputing missing covariates in the PS model.
\section {Balancing properties and consistency of IPTW estimator after MI}
Without missing data, Rosenbaum and Rubin showed that the
PS is a balancing score \cite{rosenbaum_central_1983}. A balancing score $b(\boldsymbol{x})$ is defined as
a function of the observed confounders $\boldsymbol{x}$ such that the conditional distribution of $\boldsymbol{x}$ given $b(\boldsymbol{x})$ is the same for $Z=0$ and $Z=1$. Moreover, Rosenbaum and Rubin showed that any balancing score $b(\boldsymbol{x})$ is 'finer' than the true PS, that is $e(\boldsymbol{x})=f \left\{b(\boldsymbol{x}) \right\}$, for some function $f(.)$. The consistency of PS estimators comes from this balancing property. \\
Lunceford and Davidian \cite{lunceford_stratification_2004} studied theoretical properties of the IPTW estimator when data are complete including a proof of consistency of this estimator.
In this section, we study the consistency of the IPTW estimators
obtained from MIte, MIps and MIpar and how this relates to balancing properties of the PS models used in these approaches. We suppose hereafter that (i) the SITA assumption required for IPTW (see \ref{assumptions})
holds, (ii) the missing data are missing at random (MAR) and (iii) the imputation model is correctly specified.
For simplicity, we consider here the estimation of $\theta = \mathbb{E} [Y^{Z=1}].$
\subsection{Combining the treatment effects after MI (MIte)}
Let $\mathbf{X}$, the vector of confounders, be split into fully observed and missing components, $\mathbf{X} = (\mathbf{X}_{obs}, \mathbf{X}_{miss})$. $\mathbf{X}_{m}^{(k)}$ is the imputed value of $\mathbf{X}_{miss}$ in the $k^{th}$ imputed dataset ($k=1,..,M$).
We show (see Appendix 2a) that in each imputed dataset:
\begin{equation} \label{eq: expPS_MIte}
e(\mathbf{X}_{obs}, \mathbf{X}_{m}^{(k)}) = E[Z | \mathbf{X}_{obs}, \mathbf{X}_{m}^{(k)}].
\end{equation}
If $\mathbf{X}_{m}^{(k)}$ is imputed from the true model (\textit{i.e.} correctly specified at the
true parameter values), we can also show (Appendix 2b) that a
SITA-type assumption holds in each imputed dataset, \textit{i.e.}:
\begin{align} \label{eq: condexch_MIte}
&Y^{Z=1} \perp Z \, | \, \mathbf{X}_{obs}, \mathbf{X}_{m}^{(k)} \\
&Y^{Z=0} \perp Z \, | \, \mathbf{X}_{obs}, \mathbf{X}_{m}^{(k)} \nonumber
\end{align}
These two assumptions are the imputed-data version of what Imbens calls weak unconfoundedness \cite{imbens_role_2000}. Note that we do not have the analogy of the usual, stronger, assumption
\begin{align*}
(Y^{z=1}, Y^{z=0}) \perp Z \, | \, \mathbf{X}_{obs}, \mathbf{X}_{m}^{(k)},
\end{align*}
which requires the treatment to be independent of the set of potential outcomes. This is because our imputation
model is a model for $\mathbf{X}_{miss} | Z=z, Y^{Z=z}, \mathbf{X}_{obs}$. The stronger assumption would require our
imputation model to capture $\mathbf{X}_{miss} | Z=z, Y^{Z=0}, Y^{Z=1}, \mathbf{X}_{obs}$. However, it is important to note, as Imbens does, that the weak
unconfoundedness suffices to obtain unbiased estimates of the causal treatment effect.
\textbf{Balancing properties}: We have shown in Appendix 2 that
\begin{align*}
\mathbf{X}_{obs} \perp Z \, | \, e(\mathbf{X}_{obs}, \mathbf{X}_{m}^{(k)}) \\
\mathbf{X}_{m}^{(k)} \perp Z \, | \, e(\mathbf{X}_{obs}, \mathbf{X}_{m}^{(k)}).
\end{align*}
Thus the true PS in each completed dataset balance both the unobserved and the imputed values of the confounders across treatment groups.
This balancing property is what leads to consistency of the MIte estimator. In practice, the PS has to be estimated.
\textbf{Consistency}:
Seaman and White \cite{seaman_inverse_2014} proved that for an infinite number of imputations, the estimator
obtained by combining the treatment effects after MI (MIte) is consistent. To understand how this consistency
relates to the SITA-type assumption above, it is helpful to consider the following expectation:
\begin{align}
E\left[ \frac{Y Z}{e(\mathbf{X}_{obs}, \mathbf{X}_{m}^{(k)})} \right] & =
E \left[E \left[ \left. \frac{Y Z}{e(\mathbf{X}_{obs}, \mathbf{X}_{m}^{(k)})} \right| \mathbf{X}_{obs}, \mathbf{X}_{m}^{(k)} \right] \right] because Z=0,1 \nonumber \\ & =
E \left[ \frac{ \ E [ Y^{Z=1} | \mathbf{X}_{obs}, \mathbf{X}_m^{(k)} ] \ E [ Z | \mathbf{X}_{obs}, \mathbf{X}_m^{(k)} ]}{e(\mathbf{X}_{obs}, \mathbf{X}_m^{(k)})} \right]
\label{eq: MIte_condexch} \\ & = E [E[ Y^{Z=1} | \mathbf{X}_{obs}, \mathbf{X}_m^{(k)}]]
\label{eq: MIte_expPS} \\ & = E [Y^{Z=1}] \nonumber \\ & = \theta \nonumber,
\end{align}
Step \ref{eq: MIte_condexch} requires the SITA-type assumption \ref{eq: condexch_MIte}. Step \ref{eq: MIte_expPS} relies on PS in the $k^{th}$ imputed dataset
being equal to the probability of being treated given the observed and imputed part of the confounders (equation \ref{eq: expPS_MIte}).
\subsection{Combining the PS or the PS parameters after MI (MIps and MIpar)}
For PS methods, consistency comes from the ability of PS to balance confounders between groups. MIps and MIpar create a single
overall PS used to estimate the treatment effect. Thus, consistency for these methods would rely on the ability of these overall PS
to balance both the observed and the missing parts of the confounders.
However, when combining the PS or the PS parameters, the overall PS used for the analysis is not a function of the observed confounders. Thus, the pooled PS (as estimated either in MIps or MIpar)
is not 'finer' than the true PS according to Rosenbaum and Rubin's definition \cite{rosenbaum_central_1983} (\textit{i.e.} the true PS is not a function of the pooled PS).
Consequently, it cannot be a balancing score. Thus we do not have
$\mathbf{X}_{obs}, \mathbf{X}_{miss}\perp Z|(\bar{e})$ (with $\bar{e}$ the relevant pooled PS). Therefore,
neither $\hat{\theta}_{MIps}$ nor $\hat{\theta}_{MIpar}$ are consistent estimators. We illustrate the
lack of consistency with a counter example in Appendix 3. We also discuss the balancing properties of the MP approach in Appendix 4.
However, in practice, the consequences of this inconsistency on the treatment effect estimate is not known.
So we performed a simulation study to assess the performance of these estimators under different scenarios.
\section{Simulation study}
The aim of this simulation study is to assess the performance of the three MI approaches, complete case analysis and missingness pattern approach
for IPTW when the outcome is binary.
\subsection{Data generation}
We generated datasets of sample size $n=2000$, reflecting an observational study comparing a treatment $Z=1$ to a control
treatment $Z=0$ on a fully observed binary outcome $Y$ with 3 measured confounders $\mathbf{X}=(X_1,X_2,X_3)$.
$X_1$ and $X_2$ were continuous and
$X_3$ was binary. $X_2$ was fully observed whereas $X_1$ and $X_3$ were partially observed. We generated the data
as follows:
\begin{itemize}
\item \textbf{Covariates:} The 3 confounders $\mathbf{X}=(X_1,X_2,X_3)$ are generated from a multivariate normal distribution
$\mathbf{X} \sim N_3(\mathbf{0},\mathbf{\Sigma})$, with $\Sigma_{ii}=1$ and $\Sigma_{ij}=\rho$ for $i\neq j$.
$X_3$ is then dichotomised according to a threshold of 0 to obtain a prevalence of 0.5.
We considered two values for $\rho:~\rho=0.3$ and $\rho=0.6$
corresponding to a moderate and strong correlation between confounders.
After dichotomization, $corr(X_1,X_3)=corr(X_2,X_3)=0.24$ for $\rho=0.3$ and 0.48 for $\rho=0.6$.\\
\item \textbf{Treatment assignment} depends on
$\mathbf{X}$ according to the following model:
\begin{equation}
logit(p(Z=1|\mathbf{x}))=-1.15+0.7 x_1+0.6 x_2+0.6 x_3.
\end{equation}
These coefficients give
$E(Z)=p_Z=0.3$ and an important imbalance on confounders between treatment groups, as shown in Figure 1 of the Appendices.\\
\item \textbf{Binary outcome:} The outcome depends on the 3 confounders and the treatment received
according the following model:
\begin{equation}
logit(p(Y=1|Z,\mathbf{x}))=-1.5+0.5 x_1+0.5 x_2+0.3 x_3+\theta_c Z.
\label{outcome2}
\end{equation}
In this model, $\exp{\theta_c}$ is the conditional odds ratio (OR). We used the method
described in \cite{austin_performance_2007} to find the value of $\theta_c$ to obtain the desired relative risk (RR).
In order to have a RR of 1 $\theta_c=0$, and to have a RR of 2, $\theta_c=1.221$ and 1.289 for $\rho=0.3$ and $\rho=0.6$,
respectively.
With the non-null value of $\theta_c$, the marginal OR is 2.894 and the risk difference is 0.236
when $\rho=0.3$; they are 2.949 and 0.243 when $\rho=0.6$.
With this model, the observed prevalence of the outcome is $p_0\approx 20\%$ in the control group. For the treated,
$p_1\approx 33\%$ under the situation of no treatment effect $(\theta_c=0)$, and $58\%$ under a non-null treatment effect situation $(\theta_c\neq=0)$
when $\rho=0.3$, and 35\% and 63\%, respectively for $\rho=0.6$.
\\
\item \textbf{Missingness mechanism:}
In this simulation study, we consider a missing at random (MAR) mechanism. The missingness of
each partially observed covariate ($X_1$ and $X_3$) depends on the fully observed covariate $X_2$,
the treatment received $Z$ and the outcome $Y$ and was the same for $X_1$ and $X_3$:
\begin{equation}
logit(p(R_1=1|Z,x_1,x_2,x_3,y))=\gamma_0+ z+ x_2+ \gamma_Y y,
\label {miss1}
\end{equation}
\begin{equation}
logit(p(R_3=1|Z,x_1,x_2,x_3,y))=\gamma_0+z+ x_2+\gamma_Y y,
\label {miss3}
\end{equation}
where $R_1$ and $R_3$ are the missingness indicators for $X_1$ and $X_3$ (equal to 1 if the value is missing), respectively.
We set two values for $\gamma_Y$: $\gamma_Y=(0,-0.4)$ to have a situation in which the outcome value is not a predictor of missingness
(following Hill \cite{hill_reducing_2004} or Mitra and Reiter \cite{mitra_comparison_2012})
and a situation in which the missingness depends on the outcome (as in Bartlett \textit{et al.} \cite{bartlett_multiple_2015} or
Seaman and White \cite{seaman_inverse_2014}). When $\gamma_Y=0,~ \gamma_0=-1.5$ and when $\gamma_Y=-0.4,~ \gamma_0=-1.3$
to ensure the same missingness rate of 30\% for $X_1$ and $X_3$ across scenarios. \\
\end{itemize}
Because we considered two values for the correlation between confounders, two values for the treatment effect and two values
for the association between missingness and the outcome, a full factorial design leads to 8 scenarios both for binary and continuous outcomes.
We also looked at the impact of the omission of the outcome in the imputation model in the 8 main scenarios.
For one scenario (RR=2, $\rho=0.6$ and the outcome predictor of missingness), we studied the impact of the number of
imputed datasets ($M=5$ and $M=20$), the sample size ($n=500$) and the missingness rate ($10\%$ and $60\%$) on the results.
We also generated 30 \% missing data on the outcome and the treatment indicator on this scenario assuming MCAR.
For each scenario, 5000 datasets were simulated. Simulations were performed in R and the \textit{mi} package
was used for multiple imputation \cite{R_mi}, based on full conditional specification (FCS) to generate $M=10$ imputed datasets.\\
\subsection{Estimated parameters}
For each studied scenario, we estimated the absolute bias of the treatment effect estimator and its
variance estimator, the empirical variance, the coverage rate and the standardized differences of the confounders
after IPTW. In the absence of weighting, standardized differences are defined as:
\begin{equation}
\mbox{SDiff}=\frac{100 \times \left|\bar{X}_{1}-\bar{X}_{0}\right|}{\sqrt{\frac{\hat{s}^2_{1}+\hat{s}^2_{0}}{2}}},
\label{SD1}
\end{equation}
for continuous variables, with $\bar{X}_{0},~ \bar{X}_{1},~ \hat{s}^2_{0}$ and $\hat{s}^2_{1}$ denoting the average value for covariate and its estimated variance
in the control and treatment group, respectively. For the binary confounders, the SD is:
\begin{equation}
\mbox{SD}=\frac{100 \times \left|\hat{P}_1-\hat{P}_0\right|}{\sqrt{\frac{\hat{P}_0(1-\hat{P}_0)+\hat{P}_1(1-\hat{P}_1)}{2}}},
\label{SD2}
\end{equation}
with $\hat{P}_0$ and $\hat{P}_1$ denoting the
estimated proportion of the covariate in the control and treatment groups.
~\\
For PS weighting, standardized differences are calculated by replacing the unweighted means and variances
in (\ref{SD1}) and (\ref{SD2}) by their weighted equivalents (weighted by inverse PS).
For the MIte approach, standardized differences were calculated using the PS estimated from each imputed dataset,
both to assess the balance on the originally simulated complete dataset (before imposing missingness) and on the given imputed dataset.
For MIps and MIpar, standardized differences were calculated using the pooled PS
to assess balance on (i) the original dataset, (ii) on the average value of the confounders across the imputed datasets. For (ii)
we also calculated the standardized differences separately on the observed part of the confounders and the average imputed part.
\section{Results}
Because results were similar for the three measures
of interest, we present the results for relative risks (RR) only in the main text, while results
for odds ratios and risk differences are in the appendices.
\subsection{Bias}
The absolute bias of the log(RR) of the treatment, for $\rho=0.6$, is presented in Figure \ref{bias}. Since results
for $\rho=0.3$ are similar, they are presented in the Appendices. \\
\textbf{Full data, CC and MP analyses}: As expected,
the IPTW estimator on the full data (before generating missingness for $X_1$ and $X_3$) is approximately unbiased and the
complete case (CC) estimator is strongly biased in all scenarios
except those where the outcome is not associated with missingness and there is no treatment effect.
The MP approach is always biased in the situations considered, with a bias which can be even stronger than
that which is observed for the CC approach. The reason for this is an incorrect PS model specification in each pattern of missingness: in the strata
in which a confounder is not observed, the confounder is omitted in the model. \\
\textbf{Multiple imputation}: First, the results show that the imputation model must include the outcome,
even if the outcome is not a predictor of missingness. All 3 MI estimators are strongly biased in all scenarios
when the outcome is not included in the imputation model. Second, when the outcome is included in the imputation model,
the 3 MI approaches lead to a decrease in bias relative to the crude analysis. However,
only the MIte approach leads to an unbiased estimate
in the 8 main scenarios. Combining the PS parameters to estimate the PS of the average confounders (MIpar)
performed better than combining the PS themselves, but both these approaches are slightly biased. \\
\subsection{Standardized differences between groups}
The bias observed for the MP, MIps and MIpar methods can be explained by a remaining imbalance
on the confounders between groups. Standardized differences for each covariate are in Table \ref{tab_SD}.
A covariate is usually considered adequately balanced if its standardized difference is $<10\%$. IPTW on the
full data achieved a very good balance between groups on the 3 confounders (standardized difference $<$ 5\% for each of the 3 confounders).
For the CC approach, groups were balanced but the bias occurs since excluded individuals are different
from included individuals on confounding factors. This can be seen as a selection bias, in which the sample analysed
is not representative of the target population.
The PS obtained from the MP approaches balanced the observed part of the confounders, but not the unobserved
part. This means that within each pattern of missingness,
treated and untreated individuals are balanced for the confounders included in the PS model, but unbalanced
on the missing covariate because this covariate is an unmeasured confounder in the PS model.
Thus, when the missingness rate increases, imbalances (and consequently, bias of the treatment effect estimate)
increase (see section \ref{miss_rate}).
Whereas the PS estimated from each imputed dataset (MIte) balanced both the observed and imputed part of the confounders
on the given imputed dataset, the average PS (MIps), the PS of the average confounders (MIpar) and the
PS estimated with MP recovered the balance only the observed part of the confounders, as for the MP approach.
Moreover, in the MIte approach, the balance achieved on the fully observed covariate ($X_2$) is similar to
the balance observed on full data, whereas the imbalance on this variable with MIps and MIpar is slightly
higher. Finally, the PS estimated in each imputed dataset (MIte) balances observed and imputed confounders in the given imputed
dataset, but this PS did not balance confounders in the original dataset.
\subsection{Coverage rate and standard errors}
Figure \ref{coverage} displays the coverage rate for each method when the outcome is included in the imputation model. Each boxplot represents the coverage distribution for the 8 main scenarios.
Because the CC and MP approaches are strongly biased, their coverage rates are not relevant. The coverage rate
for the MIte approach is close to the nominal value of 95\%, confirming that Rubin's rules perform well in
this context provided that the within-imputation variance takes into account the uncertainty in PS estimation. \\
Table \ref{tab_var} shows the mean estimates from different variance estimators for each method when the outcome is included in the
imputation model. It is important to note that using a variance
estimator for IPTW without taking into account the PS estimation leads to an overestimated variance. For the analysis on full data and for MIte, the
corrected variances are close to the empirical variance, whereas a naive estimator tends to overestimate the variance for these approaches.
A variance formula accounting for the PS estimation has not yet been provided for the MP approach.
For MIps and MIpar, the uncorrected variance estimator whose results are displayed in the table incorporates the variability linked to PS estimation
but not the imputation procedure. The proposed variance estimator (Appendix 1)
performs well in our simulations. In the scenarios presented in Table \ref{tab_var}, the corrected variance is smaller than the
uncorrected variance because the within imputation variance of the PS parameters (reflecting the correlation between
the confounders and treatment; the higher this is the larger gain in precision for IPTW) is higher
than the between imputation variance component (noise due to missing data). However, when the missingness rate increases,
the corrected variance can be higher than the uncorrected, because of a larger heterogeneity between imputed datasets.
\subsection{Sample size}
Table 17 of the Appendices present the results of one scenario with a non-null treatment effect
with a smaller sample size ($n=500$). Results were similar in terms of bias for $n=500$ and $n=2000$. Because the variance
estimator for IPTW has been developed for large samples, we observed slightly underestimated variances for
the full data analysis, MIps and MIparam. This underestimation is more pronounced in the CC analysis because
the sample for the analysis is even smaller (269 on average when $n=500$).
\subsection{Missingness rate}
\label{miss_rate}
Figure \ref{bias_miss} shows the bias when 10\% or 60\% of each partially observed covariate is missing. Full
results are presented in Appendix 6.3.
For a low missingness rate, the CC and MP approaches are still biased but the 3 MI approaches corrected the bias.
For a missingness rate of about 60\% for each covariate, only the MIte approach showed good performance in terms of bias reduction,
confirming the good statistical properties of this approach even with this large amount of missing data.
\subsection{Number of imputed datasets}
In our simulations, increasing the number of imputed datasets did not strongly impact the results
in terms of bias or variance (See Appendix 6.4).
\section{Application to the motivating example}
We applied CC analysis, the MP approach and the three MI strategies to estimate the effect of statin treatment on mortality
after pneumonia from our motivating example dataset. For simplicity, we analysed the primary outcome, mortality within 6 months,
as a binary outcome, and estimated the corresponding relative risk and its 95\% confidence interval. For each approach,
IPTW was used to account for the confounding. We focused the analysis on the 7158 patients without coronary heart disease.
The propensity score was estimated from a logistic regression
modeling statin use as a function of the following covariates: age, sex, body mass index (BMI), alcohol consumption,
smoking status, diabetes, cardiovascular disease circulatory disease, heart failure,
dementia, cancer, hyperlipidaemia, hypertension and prescription of antipsychotics, hormone replacement therapy, antidepressants,
steroids, nitrates,
beta-blockers, diuretics, anticoagulants and use of antihypertensive drugs.
The imbalance between the study groups is illustrated in Figure \ref{THIN}.
Complete case (CC) analysis was conducted on the 5168 individuals with complete records. For the missingness pattern approach (MP),
8 patterns were identified. However, some of these patterns were very rare. For instance, only 6 individuals had only the smoking status missing,
and only 8 had both smoking status and BMI missing. Thus, we considered only 4 groups:
\begin{itemize}
\item complete records (n=5168) for which all the covariates listed above are included
\item individuals with only the alcohol consumption missing (n=455)
\item individuals with only BMI missing (n=575)
\item individuals with the smoking status missing (alone or in addition to BMI and alcohol consumption)
and individuals with both BMI and alcohol consumption missing (n=960)
\end{itemize}
For multiple imputation (MI), 10 imputed datasets were created. The imputation model included statin use, mortality and all the variables listed above.\\
The standardized differences estimated before weighting and after weighting by PS for CC, MP, MIte, MIps and MIpar are presented in Table \ref{tab_SDiff}.
The MP and the 3 MI approaches lead to a similar reduction in imbalance between groups on the observed variables as compared to the crude standardized differences.
Nevertheless, because of the poor overlap of the patients characteristics between groups (Figure \ref{THIN}), some covariates are still unbalanced even after MI.
However, for binary covariates, large standardized differences can occur even for slight imbalance when the prevalence is low.
Estimated RR are presented in Table \ref{res_THIN}. First, all approaches based on IPTW lead to a treatment effect estimate smaller than the unweighted treatment effect. The 3 MI approaches lead to similar RR and these were smaller than the RR obtained from CC and MP analyses. The small differences between the 3 MI approaches
in this example can be explained by a low rate of missing data and the fact that the 3 partially observed covariates were not strong confounders.
\section{Discussion}
This paper aimed to address three main questions about multiple imputation in the context of IPTW:
(i) does the outcome have to be included in the imputation model?
(ii) should we apply Rubin's rules on the IPTW treatment effect estimates or
on the PS estimates themselves? (iii) how should we
estimate the variance of the IPTW estimator after MI?
First, results showed that the outcome must be included in the imputation model,
even if the outcome is not a predictor of missingness. This is well known in the context multivariable regression,
but can be seen as counter intuitive in the PS paradigm since the PS model is built without reference to the outcome.
The simulation results showed a bias in the 3 MI estimators when the outcome was omitted from the imputation model,
even when the outcome was not a predictor of missingness. This may explain the bias observed in Mitra and Reiter's study \cite{mitra_comparison_2012}
for PS matching.
Second, we showed that combining the treatment effects after MI (MIte approach) is the preferred MI strategy in terms of bias reduction
under a MAR mechanism.
This estimator is the only of the 3 MI estimator to be proven consistent and to provide good balancing properties. Even though MIps and MIpar are not consistent estimators of the treatment effect,
they can reduce the bias observed for CC analysis, in particular when the rate of missing data is low.
Combining the PS or the PS parameters has no clear advantage for IPTW,
but may be useful in the context of PS matching: because it involves only one
treatment effect estimation, it could provide computational advantage for large datasets.
In addition, MIte for PS matching implies that the $M$ treatment effect estimates are estimated from
different matched sets, potentially of different sample sizes, leading to a more complex variance estimation because
of these different sample sizes. In our illustrative example, in which the missingness rate was moderate, the relative
risks estimated with the 3 MI estimators were very similar.\\
Third, as
long as the uncertainty in the PS estimation is taken into account in the variance estimation \cite{williamson_variance_2014}, Rubin's rules perform
well for MIte, even for moderate sample size (n=500). For MIpar, the proposed variance approximation (Appendix 1) showed good performance in our simulation study.
The 3 MI approaches differ in terms of their balancing properties. We showed that whereas the PS estimated in each dataset in the MIte approach
can balance confounders between groups in each imputed dataset, this is no longer true for MIps and MIpar. However,
the best method to assess covariate balance after MI remains unknown. With MIte, the aim being to estimate a treatment effect from each
dataset, we require balance between groups within each imputed dataset. In contrast, for MIps and MIpar, we need further investigation
to know if we should assess the balancing properties of the pooled PS on the average confounder values across the imputed dataset or on each dataset.\\
The MP approach, which is widely used in practice to handle missing data for PS analysis, revealed
poor performance under a MAR mechanism. This can be explained by the inability of the generalised propensity score
to balance the missing component of the confounders. Moreover, its application on our real life example was challenging because
of the sample size was not large enough within each missingness pattern to estimate the PS.
This work has some limitations. We generated only 3 confounders in our simulation, whereas PS are
often built from a large number of confounders, despite the recent recommendation of parsimonious models \cite{brookhart_variable_2006}. Moreover,
we studied only the common situation of log-linear relationships between the confounders
and the outcome and between the confounders and the treatment status. In the presence of interactions or quadratic terms
in these two models, the specification of the imputation model can be less straightforward, requiring further efforts
to ensure the imputation model is compatible with the substantive (analysis) model \cite{bartlett_multiple_2015}. \\
In conclusion, for IPTW, multiple imputation followed by pooling of treatment effect estimates is the preferred approach amongst those studied
when data are missing at random, and the outcome must be included in the imputation model.\\~\\
\textbf{Funding}:
This work has been supported by the MRC (project grant MR/M013278/1).\\~\\
\textbf{Supplementary material}:
Complementary results are provided in Appendices
\bibliographystyle{Vancouver}
|
\section*{Acknowledgements}
This research was supported by the project GOA 13/010 Research Fund
KU Leuven and projects G.0489.10, G.0357.12 and G.0922.13 of FWO
(Research Foundation - Flanders).
Bart Bogaerts is supported by the Finnish Center of Excellence in Computational
Inference Research (COIN) funded by the Academy of Finland (grant \#251170).
\small
\bibliographystyle{plainnat}
\section*{\refname}\small\renewcommand\bibnumfmt[1]{##1.}}
\hyphenpenalty=0
\usepackage{tikz}
\usetikzlibrary{arrows, shapes, positioning, backgrounds, fit, decorations.pathreplacing, matrix}
\tikzstyle{litnode}=[draw,circle,minimum size=.75cm]
\tikzstyle{justified}=[dashed]
\begin{document}
\newcommand{\atomof}[1]{\m{|#1|}}
\newcommand\pstruct{\m{\mathcal{I}}}
\newcommand\defs[1]{\m{\mathit{defs}(#1)}}
\newcommand\opens[1]{\m{\mathit{opens}(#1)}}
\newcommand{\mathord{\sim}}{\mathord{\sim}}
\newcommand{\Val}[2]{\m{V_{#1}(#2)}}
\newcommand{\rels}[2]{\m{\mathcal{R}_{{#1},{#2}}}}
\newcommand{\drels}[2]{\m{\mathcal{DR}_{#1}(#2)}}
\newcommand\expl{\m{\mathit{expl}}}
\newcommand{\m{\mathit{dd}}}{\m{\mathit{dd}}}
\newcommand{\litsof}[1]{\m{\overline{#1}}}
\newcommand{\m{\mathit{p_\m{\mathcal{T}}}}}{\m{\mathit{p_\m{\mathcal{T}}}}}
\renewcommand\m{\bot}{\color{red}{DID YOU MEAN lfalse instead of false?}}
\renewcommand\m{\top}{\color{red}{DID YOU MEAN ltrue instead of true?}}
\newcommand{\m{\mathcal{S}}}{\m{\mathcal{S}}}
\newcommand{\theoremref}[1]{Theorem~\ref{#1}}
\newcommand{\exampleref}[1]{Example~\ref{#1}}
\newcommand{\sectionref}[1]{Section~\ref{#1}}
\newcommand{\definitionref}[1]{Definition~\ref{#1}}
\newcommand{\m{\m{\Delta}_{j}}}{\m{\m{\Delta}_{j}}}
\newcommand{\hyph}{-\penalty0\hskip0pt\relax}
\renewcommand{\cite}[1]{\citep{#1}}
\pdfinfo{
/Title (Implementing a Relevance Tracker Module)
/Author (Joachim Jansen, Jo Devriendt, Bart Bogaerts, Gerda Janssens, Marc Denecker)
}
\title{Implementing a Relevance Tracker Module}
\author{Joachim Jansen$^1$, Jo Devriendt$^1$, Bart Bogaerts$^{2,1}$,\\ Gerda Janssens$^1$, Marc Denecker$^1$\\
$^1$\email{<EMAIL>}, $^2$\email{<EMAIL>}}
\institute{$^1$KU Leuven, Leuven, Belgium\\$^2$Aalto University, Espoo, Finland}
\maketitle
\section{Introduction}
Since the addition of conflict-driven clause learning \cite{Marques-SilvaS99},
SAT solvers have made huge leaps forward.
Now that these highly\hyph performant SAT-solvers exist, research often stretches \emph{beyond SAT} by extending the language supported by SAT with richer language constructs.
Research fields such as \SMT \mycite{SMT}, \CP \mycite{CP} in the form of lazy clause generation \mycite{LCG}, or \ASP \mycite{ASP} could be seen as following this approach.
In this paper, we focus on the logic PC(ID): the Propositional Calculus extended with Inductive Definitions~\cite{lpar/MarienWD07}.
The satisfiability problem for PC(ID) encodings is called SAT(ID)~\cite{sat/MarienWDB08}.
SAT(ID) can be formalised as SAT modulo a theory of inductive definitions and is closely related to answer set solving.
In fact, all the work we introduce in this paper is also applicable to so-called generate-define-test answer set programs.
Recently we \mycite{relevance} defined a notion of \emph{relevance} for SAT(ID).
The idea is that we identify a set of literals that can contribute to a solution, and then limit the solver to only make choices on those literals.
We call such literals \emph{relevant}.
Furthermore, instead of searching for a variable assignment that satisfies the PC(ID) theory, we search for a \emph{partial} assignment that contains sufficient information to guarantee satisfiability.
Our approach is based on the notion of \emph{justifications}~\cite{DeneckerS93,lpnmr/DeneckerBS15}.
As a small example, consider the following theory.
\begin{ltheo}
\m{\mathit{p_\m{\mathcal{T}}}}.\\
\begin{ldef}
\LRule{\m{\mathit{p_\m{\mathcal{T}}}}}{a \land b }\\
\LRule{a}{d\lor \lnot e \lor f}\\
\LRule{b}{c \lor \lnot g\lor h}\\
\LRule{e}{f\lor \lnot h\lor i}
\end{ldef}
\end{ltheo}
This theory contains one constraint, that \m{\mathit{p_\m{\mathcal{T}}}} must hold, and a definition (between `\{' and `\}') of \m{\mathit{p_\m{\mathcal{T}}}} in terms of variables $a$ to $i$.
One way to check satisfiability would be to generate an assignment of all variables that satisfies the above theory (this is the classical approach to solving such problems).
What we do, on the other hand, is to search for a \emph{partial} assignment to these variables such that $\m{\mathit{p_\m{\mathcal{T}}}}$ is \emph{justified} in that partial assignment. Consider for example the partial assignment where \m{\mathit{p_\m{\mathcal{T}}}}, $a$, $b$, $c$ and $d$ are true and everything else is unknown.
In this assignment, $a$ and $b$ are \emph{justified} because $d$ and $c$ hold respectively; \m{\mathit{p_\m{\mathcal{T}}}} is \emph{justified} because both $a$ and $b$ are justified. This suffices to determine satisfiability of the theory, without considering the definition of $e$ for instance.
Intuitively, a literal is \emph{relevant} if it can contribute to justifying the theory. In the above example, as soon as $d$ is assigned {\em true}, the variable $e$ becomes \emph{irrelevant}. From that point onwards, search should not take $e$'s defining rule into account.
\mycitet{relevance} have studied relevance extensively from the theoretical perspective and ran experiments to determine how relevance influences the number of choices and the number of conflicts a solver encounters.
These experiments concluded that VSIDS chooses a significant number of relevant literals, and that prevention of decisions on irrelevant literals can lead to significant performance gains in some hand-constructed examples.
The actual implementation has not yet been described in detail.
That is exactly the goal of this paper.
The insights presented in this paper should make it possible to implement relevance in other solvers as well, for instance in ASP solvers.
In fact, our implementation heavily relies on components that are already present in modern ASP solvers.
To be precise, we reuse propagation mechanisms for inductive definitions and an unfounded set detection algorithm.
The main contributions of this paper are
\begin{inparaenum}
\item the introduction of a method to keep track of justification status in a SAT(ID) solver,
\item the introduction of a method to keep track of relevance status in a SAT(ID) solver, and
\item a discussion on the properties of these implementations.
\end{inparaenum}
\section{Preliminiaries: SAT(ID), Relevance}
\label{sec:prelims}
Here we give a short introduction on PC(ID), SAT(ID), justifications, relevance, and recall how relevance can be exploited to improve PC(ID) solvers.
For a more elaborate exposition we refer to~\mycitet{relevance}.
\subsection{PC(ID)}
\label{sec:pcid}
We briefly recall the syntax and semantics of Propositional Calculus extended with Inductive Definitions (PC(ID))~\cite{phd/Marien09}.
A truth value is one of $\{\trval{t},\trval{f},\trval{u}\}$; $\trval{t}$ represents \emph{true}, \trval{f} \emph{false} and \trval{u} \emph{unknown}. The truth order $\m{\leq_t}$ on truth values is given by $\trval{f}\m{\leq_t}\trval{u}\m{\leq_t}\trval{t}$, the precision order $\m{\leq_p}$ is given by $\trval{u} \m{\leq_p} \trval{f}$ and $\trval{u}\m{\leq_p}\trval{t}$.
Let \m{\Sigma} be a finite set of symbols called \emph{atoms}. A \emph{literal} \m{l} is an atom $p$ or its negation $\lnot p$.
In the former case, we call \m{l} \emph{positive}, in the latter, we call \m{l} \emph{negative}.
We use $\litsof{\m{\Sigma}}$ to denote the set of all literals over $\m{\Sigma}$.
If $\m{l}$ is a literal, we use $\atomof\m{l}$ to denote the atom of $\m{l}$, i.e., to denote $p$ whenever $\m{l}=p$ or $\m{l} = \lnot p$.
We use $\mathord{\sim} \m{l}$ to denote the literal that is the negation of $\m{l}$, i.e., $\mathord{\sim} p = \lnot p$ and $\mathord{\sim} \lnot p = p$.
Propositional formulas are defined as usual. We use $\varphi \Rightarrow \psi$ for material implication, i.e., as a shorthand for $\lnot \varphi \lor \psi$.
A \emph{partial interpretation} \pstruct is a mapping from $\m{\Sigma}$ to truth values.
We use the notation $\{p_1^\trval{t},\dots,p_n^\trval{t}, q_1^\trval{f},\dots,q_m^\trval{f}\}$ for the partial interpretation that maps the $p_i$ to \trval{t}, the $q_i$ to \trval{f} and all other atoms to \trval{u}.
We call a partial interpretation \emph{two-valued} if it does not map any atom to \trval{u}.
If $\pstruct$ and $\pstruct'$ are partial interpretations, we say that $\pstruct$ is less precise than $\pstruct'$ (notation $\pstruct\m{\leq_p}\pstruct'$) if for all $p\in \m{\Sigma}$, $\pstruct(p)\m{\leq_p}\pstruct'(p)$.
If $\varphi$ is a propositional formula, we use $\varphi^\pstruct$ to denote the truth value ($\trval{t}$, $\trval{f}$ or $\trval{u}$) of $\varphi$ in $\pstruct$, based on the Kleene truth tables \cite{Kleene38}.
If \pstruct is a partial interpretation and \m{l} a literal, we use $\pstruct[\m{l}:\trval{t}]$ to denote the partial interpretation equal to \pstruct, except that it interprets $\m{l}$ as $\trval{t}$ (and similar for \trval{f}, \trval{u}).
If $\m{\Sigma}'\subseteq \m{\Sigma}$, we use the notation $\pstruct|_{\m{\Sigma}'}$ to indicate the {\it restriction} of $\pstruct$ to symbols in $\m{\Sigma}'$. This restriction is a partial interpretation of $\m{\Sigma}$ and satisfies $\pstruct|_{\m{\Sigma}'}(p) = \trval{u}$ if $p \notin \m{\Sigma}'$ and $\pstruct|_{\m{\Sigma}'}(p) = \pstruct(p)$ otherwise.
In the rest of this text, when we just say \emph{interpretation}, we mean a two-valued partial interpretation.
An interpretation \m{I} is completely characterised by the set of atoms $p$ such that $\m{I}(p)=\trval{t}$. As such, slightly abusing notation, we often identify an interpretation with a set of atoms.
An inductive definition \m{\Delta} over $\m{\Sigma}$ is a finite set of rules of the form $p\leftarrow \varphi$ where $p\in \m{\Sigma}$ and $\varphi$ is a propositional formula over $\m{\Sigma}$.
We call $p$ the head of the rule and $\varphi$ the body of the rule.
We call $p$ \emph{defined in} \m{\Delta} if $p$ occurs as the head of a rule in $\m{\Delta}$. The set of all symbols defined in \m{\Delta} is denoted by $\defs\m{\Delta}$.
All other symbols are called \emph{open in} \m{\Delta}. The set of open symbols in \m{\Delta} is denoted $\opens\m{\Delta}$. We say that a literal $\m{l}$ is \emph{defined in} \m{\Delta} if $\atomof{\m{l}}\in\defs\m{\Delta}$.
We use the \emph{parametrised well-founded semantics} for inductive definitions \cite{lpnmr/DeneckerV07}.
That is, interpretation \m{I} is a model of \m{\Delta} (denoted $\m{I}\models\m{\Delta}$) if $\m{I}$ is the well-founded model of $\m{\Delta}$ in context $\m{I}|_{\opens\m{\Delta}}$.
We call an inductive definition \emph{total} if for every interpretation \m{I} of the open symbols, the well-founded model in context \m{I} is a two-valued interpretation.
A \emph{PC(ID)} theory \m{\mathcal{T}} over \m{\Sigma} is a set of propositional formulas, called constraints, and inductive definitions over \m{\Sigma}. Interpretation \m{I} is a model of \m{\mathcal{T}} if \m{I} is a model of all definitions and constraints in \m{\mathcal{T}}.
Without loss of generality~\cite{phd/Marien09}, we assume that every PC(ID) theory is in {\bf \DEFNF}, where $\m{\mathcal{T}} = \{\m{\mathit{p_\m{\mathcal{T}}}}, \m{\Delta}\}$ and
\begin{compactitem}
\item $\m{\mathit{p_\m{\mathcal{T}}}}$ is an atom,
\item $\m{\Delta}$ is an inductive definition defining \m{\mathit{p_\m{\mathcal{T}}}},
\item every rule in $\m{\Delta}$ is of the form $p \leftarrow l_1\odot \dots\odot l_n$, where $\odot$ is either $\land$ or $\lor$, $p$ is an atom, and each of the $l_i$ are literals,
\item every atom $p$ is defined in at most one rule of $\m{\Delta}$.
\end{compactitem}
A rule in which $\odot$ is $\land$, respectively $\lor$ is called a \emph{conjunctive}, respectively \emph{disjunctive}, rule.
The rules in a definition \m{\Delta} impose a \emph{direct dependency relation}, denoted $\m{\mathit{dd}}_\m{\Delta}$, between literals, defined as follows.
For every rule $p \leftarrow l_1\odot \dots\odot l_n$ in \m{\Delta}, $\m{\mathit{dd}}_\m{\Delta}$ contains $(p, l_i)$ and $(\mathord{\sim} p, \mathord{\sim} l_i)$ for all $1 \leq i \leq n$.
The \emph{dependency graph} of \m{\Delta} is the graph $G_\m{\Delta} = (\litsof{\m{\Sigma}},\m{\mathit{dd}}_\m{\Delta})$.
For the remainder of the paper, we assume that some PC(ID) theory $\m{\mathcal{T}}=\{\m{\mathit{p_\m{\mathcal{T}}}},\m{\Delta}\}$ is fixed.
It has been argued many times before \cite{Denecker98,tocl/DeneckerT08,KR/DeneckerV14} that all sensible definitions in mathematical texts are total definitions. Following these arguments, in the rest of this paper we assume \m{\Delta} to be a total definition.
The satisfiability problem for PC(ID), i.e., deciding whether a PC(ID) theory has a model, is called \emph{SAT(ID)}.
This problem is NP-complete~\cite{sat/MarienWDB08}.
\subsection{Justifications}
Consider a directed graph $G= (V,E)$, with $V$ a set of nodes and $E$ a set of edges.
If $G$ contains an edge from $l$ to $l'$ (i.e., $(l,l') \in E$), we say that $l$ is a parent of $l'$ in $G$ and that $l'$ is a child of $l$ in $G$.
A node $l$ is called a \emph{leaf} of $G$ if it has no children in $G$; otherwise it is called \emph{internal} in $G$.
Let $G' = (V',E')$ be another graph.
We define the union of two graphs (denoted $G\cup G'$) as the graph with vertices $V \cup V'$ and edges $E\cup E'$.
Suppose $\m{l}$ is a literal with $p=\atomof\m{l}$ and $p\in\defs\m{\Delta}$ with defining rule $p \leftarrow l_1\odot \dots\odot l_n$.
A set of literals $J_d$ is a \emph{direct justification} of $\m{l}$ in \m{\Delta} if one of the following holds:
\begin{compactitem}
\item $\m{l}=p$, $\odot$ is $\land$, and $J_d=\{l_1,\dots,l_n\}$,
\item $\m{l}=p$, $\odot$ is $\lor$, and $J_d=\{l_i\}$ for some $i$,
\item $\m{l}=\lnot p$, $\odot$ is $\land$, and $J_d=\{\mathord{\sim} l_i\}$ for some $i$,
\item $\m{l}=\lnot p$, $\odot$ is $\lor$, and $J_d=\{\mathord{\sim} l_1,\dots,\mathord{\sim} l_n\}$.
\end{compactitem}
Note that a direct justification of a literal can only contain children of that literal in the dependency graph.
A \emph{justification}~\cite{DeneckerS93,lpnmr/DeneckerBS15} $J$ of a definition \m{\Delta} is a subgraph of $G_\m{\Delta}$, such that each internal node $\m{l}\in J$ is a defined literal and the set of its children is a direct justification of \m{l} in \m{\Delta}.
We say that $J$ \emph{contains} \m{l} if $\m{l}$ occurs as node in $J$.
A justification is \emph{total} if none of its leaves are defined literals.
A justification can contain \emph{cycles}.\footnote{In this text, we assume that \m{\Delta} is finite; in this case cycles are simply loops in the graph. The infinite case is a bit more subtle, and an adapted definition of cycle is required to maintain all results presented below. }
A cycle is called \emph{positive} (resp. \emph{negative}) if it contains only positive (resp.~negative) literals.
It is called a \emph{mixed} cycle otherwise.
If $J$ is a justification and \pstruct a (partial) interpretation, we define the value of $J$ in $\pstruct$, denoted $\Val{\pstruct}{J}
$ as follows:
\begin{compactitem}
\item $\Val{\pstruct}{J} = \trval{f}$ if $J$ contains a leaf $\m{l}$ with $\m{l}^\pstruct = \trval{f}$ or a positive cycle (or both).
\item $\Val{\pstruct}{J} = \trval{u}$ if $\Val{\pstruct}{J} \neq \trval{f}$ and $J$ contains a leaf $\m{l}$ with $\m{l}^\pstruct = \trval{u}$ or a
mixed cycle (or both).
\item $\Val{\pstruct}{J} = \trval{t}$ otherwise (all leaves are $\trval{t}$ and cycles, if any, are negative).
\end{compactitem}
A literal $\m{l}$ is \emph{justified} (in \pstruct, for \m{\mathcal{T}}) if there exists a total justification $J$ (of \m{\Delta}) that contains $\m{l}$ such that $\Val{\pstruct}{J}=\trval{t}$.
In this case, we say that such a $J$ \emph{justifies} $\m{l}$ (in \pstruct, for \m{\mathcal{T}}).
It follows from the definition that it is not possible that both $l$ and $\neg l$ are justified in the same in the same interpretation.
The {\em justification status} of an atom $p$ (in \pstruct, for \m{\mathcal{T}}) is defined as follows.
The justification status of $p$ is {\em true} if and only if the literal $p$ is justified in \pstruct for \m{\mathcal{T}}.
The justification status of $p$ is {\em false} if and only if the literal $\neg p$ is justified in \pstruct for \m{\mathcal{T}}.
Otherwise the justification status of $p$ is {\em unknown}.
\citet{DeneckerS93} and later also \citet{lpnmr/DeneckerBS15} showed that many semantics of logic programs can be captured by justifications.
\subsection{Relevance}
Now, we recall the central definitions and theorems related to relevance. For a more detailed exposition of the theoretical foundations, we refer to \mycitet{relevance}.
\begin{theorem}[\mycite{relevance}; Theorem 3.1]\label{thm:relmain}
$\m{\mathcal{T}}$ is satisfiable if and only if there exists a partial interpretation $\pstruct$ and a justification $J$ that justifies $\m{\mathit{p_\m{\mathcal{T}}}}$ in $\pstruct$.
\end{theorem}
\begin{definition}[\mycite{relevance}; Definition 3.2] \label{def:relevant}
Given a PC(ID) theory $\m{\mathcal{T}}=\{\m{\mathit{p_\m{\mathcal{T}}}},\m{\Delta}\}$ and a partial interpretation \pstruct, we inductively define the set of relevant literals, denoted $\rels{\m{\mathcal{T}}}{\pstruct}$, as follows
\begin{compactitem}
\item $\m{\mathit{p_\m{\mathcal{T}}}}$ is relevant if $\m{\mathit{p_\m{\mathcal{T}}}}$ is not justified,
\item $l$ is relevant if $\m{l}$ is not justified and there exists some $\m{l}'$ such that $(\m{l}',\m{l})\in \m{\mathit{dd}}_\m{\Delta}$ and $\m{l}'$ is relevant.
\end{compactitem}
\end{definition}
The central theorem regarding relevance shows that any search algorithm that arrives in a state in which \m{\mathit{p_\m{\mathcal{T}}}} is justified by deciding on a literal $l$ that is irrelevant can also arrive in such a state without deciding on $l$.
Hence, if a literal $l$ is irrelevant, choosing on $l$ does not help justifying \m{\mathit{p_\m{\mathcal{T}}}}.
\begin{theorem}[\mycite{relevance}; Theorem 3.5] \label{thm:choice}
Let $\m{\mathcal{T}}=\{\m{\mathit{p_\m{\mathcal{T}}}},\m{\Delta}\}$ be a PC(ID) theory. Suppose \pstruct is a partial interpretation and $\m{l}$ a literal such that $\pstruct(\atomof{\m{l}}) = \trval{u}$ and $\m{l}$ is not relevant in \pstruct.
If $\m{\mathit{p_\m{\mathcal{T}}}}$ is justified in some partial interpretation $\pstruct'$ more precise than $\pstruct$, then $\m{\mathit{p_\m{\mathcal{T}}}}$ is also justified in $\pstruct'[\m{l}:\trval{f}]$ and in $\pstruct'[\m{l}:\trval{t}]$.
\end{theorem}
Consequently, the work suggests adapting SAT(ID) solvers such that they
\begin{inparaenum}
\item do not make decisions on irrelevant literals, and
\item stop searching when there are no relevant literals left.
\end{inparaenum}
This requires the underlying solver to keep track of which literals are relevant.
This task is incremental in nature: small changes to the state of the solver will result in small changes to the relevance of literals.
Since modern solvers work with hunderds of thousands of variables and go through millions of assignment in quick succession, it is essential to do these changes as efficiently as possible.
\section{The Basic Framework}
The aim of this work is to discuss the implementation of an algorithm to keep track of relevant literals.
As said in Theorem~\ref{thm:relmain}, the solver aims to arrive at a state $\m{\mathcal{S}}$ where $\m{\mathit{p_\m{\mathcal{T}}}}$ is justified in $\pstruct$.
The solver does this by making decisions, performing propagation, and backtracking.
To prevent the solver from making ``useless'' decisions, we need to know whether literals are relevant or not in $\pstruct$.
We consider the underlying solver to have an internal state \m{\mathcal{S}} of the form $\m{\mathcal{S}} = \langle \litsof{\m{\Sigma}}, \m{\mathcal{T}}, \pstruct\rangle$, with
\begin{inparaenum}
\item $\litsof{\m{\Sigma}}$ denoting the set of literals used in the solver,
\item $\m{\mathcal{T}} = \{\m{\mathit{p_\m{\mathcal{T}}}}, \m{\Delta}\}$ a \DEFNF theory, and
\item $\pstruct$ the current partial interpretation in the solver.
\end{inparaenum}
During the search process, the (CDCL) solver adds learned conflict clauses to the theory.
However, learned conflict clauses are logical consequences of the theory and because of this we do not consider them to be a part of the theory \m{\mathcal{T}} in \m{\mathcal{S}}.
Instead, \m{\mathcal{T}} is reserved for non-learned clauses.
We assume \m{\mathcal{T}} to remain static during the search process.
This assumption is valid in most {\em ground-and-solve} systems.
Recent work focuses on interleaving this process~\mycite{lazygrounding}.
Extending this work to allow for a changing theory is future work, but should be of limited complexity given the framework we present here.
The relevance tracker needs to take into account changes in the solver state, in particular in \pstruct. Before we define the interface between the relevance tracker and the solver, we discuss the solver state and its changes.
During the search process, the solver iteratively performs one of the following state changes:
\begin{itemize}
\item $\langle \litsof{\m{\Sigma}}, \m{\mathcal{T}}, \pstruct\rangle \mapsto \langle \litsof{\m{\Sigma}}, \m{\mathcal{T}}, \pstruct[l:\trval{t}]\rangle$ a literal $l$ becomes {\em true}, or
\item $\langle \litsof{\m{\Sigma}}, \m{\mathcal{T}}, \pstruct\rangle \mapsto \langle \litsof{\m{\Sigma}}, \m{\mathcal{T}}, \pstruct[l:\trval{u}]\rangle$ a literal $l$ becomes {\em unknown}.
\end{itemize}
Note that this set of operations allows the solver to make a literal $l$ {\em false} by making literal $\mathord{\sim} l$ {\em true}.
In order to get the necessary information about the changes of the solver and to implement the above revision problem, the relevance tracker listens to notifications.
The relevance tracker supports the following interface to the underlying solver.
\begin{description}
\item [notifyBecomesTrue] a literal $l$ becomes {\em true} in \pstruct
\item [notifyBecomesUnknown] a literal $l$ becomes {\em unknown} in \pstruct
\item [isRelevant] query whether a given literal $l$ is relevant (returns a boolean value)
\end{description}
Methods {\bf notifyBecomesTrue} and {\bf notifyBecomesUnknown} must be cal\-led by the underlying solver when a literal has become true, respectively unknown.
The {\bf isRelevant} method is used by the solver to ask the module whether the given literal is relevant.
The relevance information allows the solver to change its underlying heuristic, selecting only relevant literals.
\ignore{
\section{Relevance as a Meta-Definition on Top of Solver State}
\label{sec:metadef}
\bart{Joachim, had jij geen refernties naar revision papers die je hier ging plaatsen. Als we (bijna) niets zeggen over revisie, dan is de waarde van ``relevance as a meta definition'' wel heel klein. Immers... De paper zou gaan over implementatie dingen...}
\joachim{Deze sectie mag er van mij part uit. Ik heb deze gedachtengang wel gevolgd bij het implementeren, maar het is niet nodig dat die in deze paper naar boven komt. Maakt de hele zaak zelfs verwarrender voor ASP mensen zelfs denk ik.}
In this section, we rephrase the (informal) definition of relevance as a set of (formal) rules. To phrase it differently, we give a definition of relevance in the logic \logicname{FO(ID)}~\mycite{fodot}. \logicname{FO(ID)} extends PC(ID) by supporting
\begin{inparaenum}
\item in the vocabulary, {\em types} as a sets of values,
\item in the vocabulary, predicate symbols with a given type signature for thei arguments, and
\item in the theory, existential ($\exists$) or universal ($\forall$) quantification over variables of a given type.
\end{inparaenum}
We introduce a \logicname{FO(ID)} meta-vocabulary to reify the internal state of the underlying solver.
The meta-vocabulary contains a single type {\bf Literal} that contains the set of all literals.
In addition, the meta-vocabulary contains the following predicate symbols (using type {\bf Literal} for all arguments).
\begin{description}
\item [$\m{\mathit{dd}}_\m{\Delta}(l',l)$]: literal $l$ occurs in the body of the rule defining $l'$ in \m{\mathcal{T}}
\item [justified($l$)]: literal $l$ is justified in \pstruct, for \m{\mathcal{T}}
\item [relevant($l$)]: literal $l$ is relevant in \pstruct, for \m{\mathcal{T}}
\end{description}
\noindent
Using this meta-vocabulary, the definition of relevance can then be specified using a meta-definition.
\[\left\{\begin{array}{ll}
relevant(l) \leftarrow & l = \m{\mathit{p_\m{\mathcal{T}}}}~\land~\lnot justified(\m{\mathit{p_\m{\mathcal{T}}}}).\\
relevant(l) \leftarrow & \lnot justified(l)~\land~(\exists l' : relevant(l')~\land~\m{\mathit{dd}}_\m{\Delta}(l',l)).
\end{array}\right\}\]
Note that the relevance of a literal $l$ and also the justifiedness of a literal is completely determined by the theory and the current status of the solver.
As said before we assume \m{\mathcal{T}} and thus $\m{\mathit{dd}}_\m{\Delta}(l',l)$ to be static.
But, whether a literal is justified or not can change when the status of the solver changes.
In the above \logicname{FO(ID)} definition the only open symbol is {\bf justified}.
We can then view the computation of the changed value of $relevant/1$ based on changes in the interpretations of the open symbol {\bf justified($l$)} as a model revision task.
As such, in principle, tracking relevance can be tackled by one of the many existing model revision algorithms \cite{,,,} or database view maintenance algorithms \cite{,,,} \todo{joachim, refs?}. We, however, opted for a standalone implementation \todo{since?}, to be detailed in the following sections.
In order to implement such a revision algorithm, the relevant tracker needs to have information about which literals are justified.
In section~\ref{sec:justtracking} we show how also this information can be extracted from the solver, reusing existing propagation mechanisms.
}
\section{Deriving the justification status of literals}
\label{sec:justtracking}
The definition of relevance relies on knowledge about which literals are \emph{justified} in the solver. In this section, we discuss how we implemented extraction of this information.
We opted to implement a method that re-uses the underlying SAT(ID) solver to keep track of the justification status of literals.
The method creates a new atom, called the ``justification atom'', for each defined atom $p$, denoted as $j(p)$.
We call a literal $j(p)$ or $\lnot j(p)$ a {\em justification literal}.
The intended interpretation of $j(p)$ is that $j(p)$ is true iff $p$ is justified, $j(p)$ is false iff $\lnot p$ is justified and $j(p)$ is unknown otherwise.
To ensure that justification literals indeed get the right value, an extra PC(ID) definition \m{\m{\Delta}_{j}}, denoted the ``justification definition'', is added to the theory \m{\mathcal{T}}.
\m{\m{\Delta}_{j}} is constructed based on the original definition \m{\Delta} in the following manner.
The existing definition \m{\Delta} is copied, except that every defined atom $p$ is replaced with the newly created atom $j(p)$.
Thus, of all the atoms in the original definition, only the open atoms remain.
\begin{example}\label{ex:justdef}
Transforming the original definition
\[\m{\Delta} = \left\{\begin{array}{rrrrrrrrr}
\m{\mathit{p_\m{\mathcal{T}}}} & \leftarrow & c_1 & \land & c_2 & \land & c_3 & \land & c_4\\
c_1 & \leftarrow & \lnot b & \lor & \lnot d\\
c_2 & \leftarrow & a & \lor & b & \lor & \lnot c\\
c_3 & \leftarrow & \lnot b & \lor & e & \lor & \lnot f\\
c_4 & \leftarrow & d & \lor & f & \lor & \lnot a\\
f & \leftarrow & b & \lor & d\\
\end{array}\right\}\]
leads to the justification definition
\[\m{\m{\Delta}_{j}} = \left\{\begin{array}{rrrrrrrrr}
j(\m{\mathit{p_\m{\mathcal{T}}}}) & \leftarrow & j(c_1) & \land & j(c_2) & \land & j(c_3) & \land & j(c_4)\\
j(c_1) & \leftarrow & \lnot b & \lor & \lnot d\\
j(c_2) & \leftarrow & a & \lor & b & \lor & \lnot c\\
j(c_3) & \leftarrow & \lnot b & \lor & e & \lor & \lnot j(f)\\
j(c_4) & \leftarrow & d & \lor & j(f) & \lor & \lnot a\\
j(f) & \leftarrow & b & \lor & d
\end{array}\right\}\]
\end{example}
In addition to the creation of this new definition \m{\m{\Delta}_{j}}, we prohibit the solver from making choices on these justification atoms.
Because of this, the value of all $j(p)$ will be purely the result of the underlying propagation mechanism for definitions.
Our claim is now that existing propagation mechanisms will propagate \emph{exactly} those literals that are justified.
We assume a solver that performs \emph{unit propagation} and \emph{unfounded set propagation}~\cite{lpar/MarienWD07,ai/GebserKS12}, i.e., propagation that makes all atoms in an unfounded set false.
\begin{theorem}
Let $\m{\Delta}$ be a (total) definition and $\pstruct$ a partial interpretation in which all defined symbols of $\m{\Delta}$ are interpreted as \trval{u}.
Let $\m{l}$ be a defined literal in $\m{\Delta}$.
In this case $\m{l}$ is justified in \pstruct if and only if $\m{l}$ is derivable by unit propagation on the completion\footnote{The completion of a rule $p \leftarrow q$ is the underapproximation using the \FO sentence $p \Leftrightarrow q$, which demands that $p$ and $q$ hold equal truth values}\mycite{completion} of \m{\Delta} and unfounded set propagations.
\end{theorem}
\begin{proof}[Sketch of the proof]
Intuitively, from a sequence of propagations, we can create a justification and vice versa: each justification induces a sequence of propagations.
The correspondence is as follows.
First for the completion, if \m{\Delta} contains a rule
$p \leftarrow l_1\land \dots \land l_n$, then this rule propagates $p=\trval{t}$ if and only if each of the $l_i$ is true.
This corresponds to the justification
\[
\begin{tikzpicture}
\def 5 {5}
\def 3cm {3cm}
\def \margin {8}
\node[draw, circle] at (1,0) (pnode) {$p$};
\node[draw, circle] at (0,-1) (l1node) {$l_1$};
\node[draw, circle] at (1,-1) (ldotsnode) {$\dots$};
\node[draw, circle] at (2,-1) (lnnode) {$l_n$};
\draw[->] (pnode) -- (l1node);
\draw[->] (pnode) -- (ldotsnode);
\draw[->] (pnode) -- (lnnode);
\end{tikzpicture}\]
And similar justification constructs can be defined for when $\lnot p$ is propagated or when the rule is disjunctive.
Unfounded set propagation essentially corresponds to a justification of a set of negative facts by a negative cycle.
The condition that a justification can have no mixed or positive cycles corresponds to the fact that propagation must happen in order. E.g., from the rule $p\leftarrow p \lor q$, $p$ can only be propagated if $q$ is true; $p$ cannot be propagated ``because $p$ is true''.
\qed
\end{proof}
The previous theorem establishes that our approach works; a justification literal $j(p)$ will be propagated to true if $p$ is justified (note that $p$ is justified iff $j(p)$ is justified).
It also explains why we use a duplicated definition: the theorem only holds if $\m{\mathcal{I}}$ is an $opens(\m{\Delta})$ interpretation.
Since this cannot be enforced (we don't want to intrude in the solver's search), we make a copy and never make choices on the copied defined symbols.
Thus, we extend our solver state $\m{\mathcal{S}} = \langle \litsof{\m{\Sigma}}, \m{\mathcal{T}}, \pstruct\rangle$ to a $\m{\mathcal{S}}' = \langle \litsof{\m{\Sigma}}', \m{\mathcal{T}}', \pstruct', \m{\Sigma}' \rangle$ with
\begin{itemize}
\item $\m{\Sigma}' = $ set containing the newly introduced justification atoms that the solver cannot decide on
\item $\litsof{\m{\Sigma}}' = \litsof{\m{\Sigma}} \cup \litsof{\m{\Sigma}'}$
\item $\m{\mathcal{T}}' = \{\m{\mathit{p_\m{\mathcal{T}}}}, \m{\Delta}'\}$ if $\m{\mathcal{T}} = \{\m{\mathit{p_\m{\mathcal{T}}}}, \m{\Delta}\}$ and $\m{\Delta}' = \m{\Delta} \cup \m{\m{\Delta}_{j}}$
\item $\pstruct' = $ a partial structure over $\litsof{\m{\Sigma}}'$
\end{itemize}
With all this in place, we derive the interpretation for $justified(l)$ as follows.
\begin{itemize}
\item $justified(p)$ is true if and only if $j(p)$ is true in $\pstruct'$, and
\item $justified(\lnot p)$ is true if and only if $j(p)$ is false in $\pstruct'$.
\end{itemize}
\section{Implementing the relevance tracker}
\label{sec:impl}
The source code that implements the techniques discussed here can be found at \small \url{https://dtai.cs.kuleuven.be/static/krr/files/experiments/idp_relevance_experiments.tar.gz} \normalsize.
The solver maintains a subgraph of the dependency graph in order to keep track of the set of relevant literals.
This subgraph, denoted as the {\em relevance graph} contains all literals that are relevant and all edges between them (in the dependency graph).
As such, the task of the tracker is to determine whether a given literal is a member of this graph or not.
We store this graph using a data structure, denoted {\tt candidate\_parents($l$)} that associates a literal with a set of literals called ``candidate parents''.
The ``candidate parents'' of \m{l} are the literals that are parents of \m{l} in the relevance graph. I.e., if \m{l} is irrelevant, this set is empty, otherwise, it consists of all relevant parents of \m{l} in the dependency graph.
As such, it can be seen that there is an edge $(p,l)$ in the relevance graph iff $p\in~${\tt candidate\_parents($l$)}.
Thus, $l$ is relevant if and only if $l$ has a non-empty set of candidate parents.
We now describe an incremental algorithm to update the set of candidate parents for all literals if the state of the solver changes.
We prefer to keep these changes \emph{local}, i.e., to not reconstruct the entire relevance graph with each solver change.
When the solver state changes, and the set of candidate parents must be updated, care must be taken to detect and remove cyclic dependencies.
These cyclic dependencies can arise when a candidate parent is removed from a literal $l$ and the remaining candidate parents of that literal are not reachable from \m{\mathit{p_\m{\mathcal{T}}}} anymore but still have $l$ as a candidate parent, creating a loop.
A more detailed example is given in \exampleref{ex:cycles}
\begin{example}
\label{ex:cycles}
The following definition has the cyclic dependency of $p \leftarrow q \leftarrow p$.
\[\left\{\begin{array}{rl}
\m{\mathit{p_\m{\mathcal{T}}}} & \leftarrow a \lor p\\
p & \leftarrow q\\
q & \leftarrow p\\
\end{array}\right\}\]
Initially \pstruct = $\emptyset$, thus nothing is justified and all literals are relevant.
Thus, $p$ has candidate parents $\{\m{\mathit{p_\m{\mathcal{T}}}}, q\}$, and $q$ has candidate parents $\{p\}$.
\begin{figure}
\centering
\begin{tikzpicture}
\node [] at (1,2) (start) {};
\node [litnode] at (1,1) (PT) {\m{\mathit{p_\m{\mathcal{T}}}}};
\node [litnode] at (0,0) (A) {$a$};
\node [litnode] at (2,0) (P) {$p$};
\node [litnode] at (4,0) (Q) {$q$};
\draw[->] (start) to (PT);
\draw[->] (PT) to (A);
\draw[->] (PT) to (P);
\draw[->] (P) to[out=20,in=160] (Q);
\draw[->] (Q) to[out=200,in=340] (P);
\end{tikzpicture}
\caption{Relevance graph for $\pstruct = \{\}$}
\end{figure}
Consider the case where $a$ becomes true and hence $\m{\mathit{p_\m{\mathcal{T}}}}$ becomes justified.
\begin{figure}
\centering
\begin{tikzpicture}
\node [litnode] at (1,1) (PT) {\m{\mathit{p_\m{\mathcal{T}}}}};
\node [litnode] at (0,0) (A) {$a$};
\node [litnode] at (2,0) (P) {$p$};
\node [litnode] at (4,0) (Q) {$q$};
\draw[->, dashed] (P) to[out=20,in=160] (Q);
\draw[->, dashed] (Q) to[out=200,in=340] (P);
\end{tikzpicture}
\caption{Relevance graph for $\pstruct = \{a^{\trval{t}}\}$. Remaining loop indicated with dashed edges. }
\end{figure}
Simply removing $\m{\mathit{p_\m{\mathcal{T}}}}$ from the set of candidate parents of $p$ means that $p$ still has candidate parents $\{q\}$, which is actually just a loop leading back to $p$.
Thus, a cycle detection algorithm is needed to force $p$ and $q$ to remain loop-free.
\end{example}
Thus, adding and removing candidate parents is a complicated matter.
In \sectionref{sec:ufsdetection} we discuss how this cycle detection is done.
For now, we use the following interface for adjusting the set of a candidate parents.
\begin{description}
\item [notifyAddCandidateParent($l$,$l'$)] add $l'$ to the candidate parents of $l$
\item [notifyRemoveCandidateParent($l$,$l'$)] remove $l'$ from the candidate parents of $l$
\end{description}
Our definition of candidate parents potentially changes when the following changes take place (note that we already assumed the dependency relation to be non-changeable).
\begin{itemize}
\item A change in the justification status of $l$
\item A change in the relevance status of a parent literal $l'$
\end{itemize}
\noindent
Thus, we extend the interface of the relevance tracker to also support the following methods.
\begin{description}
\item [notifyBecomesJustified($l$)] A literal $l$ goes from unjustified to justified
\item [notifyBecomesUnjustified($l$)] A literal $l$ goes from justified to unjustified
\item [notifyBecomesRelevant($l$)] A literal $l$ goes from irrelevant to relevant
\item [notifyBecomesIrrelevant($l$)] A literal $l$ goes from relevant to irrelevant
\end{description}
\noindent
In the following subsections we present
\begin{inparaenum}
\item an overview of the data structures in the relevance tracker and
\item the algorithms for the methods in our interface.
\end{inparaenum}
\subsection{Data Structures}
The data structures include sets and maps.
Unless specified otherwise, we use hash sets and hash maps.
The implementation uses {\tt std::unordered\_set} and {\tt std::unordered\_map} provided by the {\tt C++} standard library.
Internally, we store the dependency relation $\m{\mathit{dd}}_\m{\Delta}$ using two maps in our module, named {\tt children} and {\tt parents}.
These data structures map a literal to a set of literals.
The first map ({\tt children}) maps a literal to its set of children in $\m{\mathit{dd}}_\m{\Delta}$.
The second map ({\tt parents}) maps a literal to its set of parents in $\m{\mathit{dd}}_\m{\Delta}$.
They are initialised using the {\bf notifyNewRule} method.
Once they are initialised, they remain fixed.
We use a map ({\tt to\_just\_lit}) to transform a normal literal to its justification literal ($p \mapsto j(p)$, $\lnot p \mapsto \lnot j(p)$).
For efficiency reasons, we also maintain the inverse map {\tt to\_nonjust\_lit}~$=$~{\tt to\_just\_lit$^{-1}$}.
These maps are initialised when the justification definition \m{\m{\Delta}_{j}} is created and do not change during execution afterwards.
We maintain a set of atoms ({\tt is\_just\_atom}) to identify the justification atoms that were introduced.
This set are initialised when the justification definition \m{\m{\Delta}_{j}} is created and does not change during execution afterwards.
We use round brackets to indicate the result of a map lookup, e.g., \[\text{\tt to\_just\_lit($p$) = $j(p)$}.\]
We use round brackets to do a containment check of sets.
More precisely, \[{\tt is\_just\_atom(}p{\tt )} = \text{\tt true}\] if and only if $p$ is in the set {\tt is\_just\_atom}.
As mentioned before, the underlying solver is not allowed to make decisions on literals in this set.
We maintain a map {\tt candidate\_parents} with the invariant that it maps a literal $l$ to the set of candidate parents of $l$.
This map is dynamic throughout execution and changes to this map are performed using the {\bf notifyAddCandidateParent} and {\bf notifyRemoveCandidateParent} methods.
\subsection{Notification-based Algorithms}
Given that the invariant of {\tt candidate\_parents} is satisfied in solver state $\m{\mathcal{S}}=\langle \litsof{\m{\Sigma}}, \m{\mathcal{T}}, \pstruct\rangle$,
we wish to perform the necessary changes such that they are satisfied in solver state $\m{\mathcal{S}}' = \langle \litsof{\m{\Sigma}}, \m{\mathcal{T}}, \pstruct[p:tv]\rangle$ with $p$ some atom and $tv$ one of $\{\trval{t}, \trval{f}, \trval{u}\}$.
We initiate our notification-based algorithm as follows.
If $tv = \trval{t}$, then we call {\bf notifyBecomesTrue($p$)}.
If $tv = \trval{f}$, then we call {\bf notifyBecomesTrue($\lnot p$)}.
If $tv = \trval{u}$, then we call {\bf notifyBecomesUnknown($p$)}.
This call, in turn, can trigger other internal notifications.
The implementation of these cascading notifications ensures that {\tt candidate\_parents} will comply with its invariant in interpretation $\m{\mathcal{S}}'$ after the designated call to {\bf notifyBecomesTrue} or {\bf notifyBecomesUnknown} is complete.
The relevance tracker implements {\bf isRelevant($l$)} by checking whether {\tt can\-didate\_parents($l$)} maps to an empty set or not.
This is a correct representation of the relevance status of $l$ if the invariant of {\tt candidate\_parents} is satisfied.
For methods {\bf notifyBecomesTrue($l$), notifyBecomesUnknown($l$)}: the given literal can be a normal literal ($p$ or $\lnot p$) or a justification literal ($j(p)$ or $\lnot j(p)$).
The relevance tracker takes no action for normal literals.
If the given literal is a justification literal, then we retrieve the original normal literal and notify the relevance tracker that this literal has become (un)justified.
Note that we re-use the notation of $\atomof{l}$ to indicate the atom of literal $l$.
\noindent
{\bf notifyBecomesTrue($l$)}: if {\tt is\_just\_atom}($\atomof{l}$), then call {\bf notifyBecomesJustified({\tt to\_nonjust\_lit($l$)})}.
\noindent
{\bf notifyBecomesUnknown($l$)}: if {\tt is\_just\_atom}($\atomof{l}$), then call {\bf notifyBecomesUnjustified({\tt to\_nonjust\_lit($l$)})}.
\noindent
{\bf notifyBecomesJustified($l$)}: call {\bf notifyRemoveAllCandidateParentsOf($l$)}.
\noindent
{\bf notifyBecomesUnjustified($l$)}: for all parents $p$ of $l$ that are relevant, call {\bf notifyAddCandidateParent($l$,$p$)}.
\noindent
{\bf notifyBecomesRelevant($l$)}: for all children $c$ of $l$, call {\bf notifyAddCandidate\-Parent($c$,$l$)}.
\noindent
{\bf notifyBecomesIrrelevant($l$)}: for all children $c$ of $l$, call {\bf notifyRemoveCan\-didateParent($c$,$l$)}.
\subsection{Maintaining watches instead of sets of candidates}
The above methods dictate how the {\tt candidate\_parents} map should be manipulated.
For efficiency reasons, the relevance tracker does not actively maintain this set of candidate parents.
Instead it keeps track of a single candidate parent as ``watched'' parent.
This watched parent is maintained using a map called {\tt watched\_parent($l$)} that maps a literal to a single parent of $l$.
The method {\bf isRelevant($l$)} now checks whether a given literal $l$ has a watched parent or not.
We only keep track of a single watched parent in order to minimize how many times a cycle detection algorithm has to be invoked.
The manipulation of the set of candidate parents, along with the invocation of a cycle detection algorithm is done as follows
\begin{description}
\item [notifyAddCandidateParent($l$,$l'$)]
Check for the following criteria
\begin{itemize}
\item $l$ does not have a watched parent yet
\item $l$ is not justified
\item $l'$ is relevant
\item $l$ is a child of $l'$
\end{itemize}
If they are met, make {\tt watched\_parent($l$) = $l'$} and call {\bf notifyBecomes\-Relevant($l$)}.
Note that a cyclic dependency check between $l$ and $l'$ is not needed, since $l$ could not have been a suitable watch for any other literal, as it was not relevant before.
\item [notifyRemoveCandidateParent($l$,$l'$)]
If $l$ had $l'$ as its watch, remove $l'$ as watched parent of $l$.
Try to find an alternative candidate parent $n$ such that the following hold.
\begin{itemize}
\item $n \neq l'$
\item $l$ is a child of $n$
\item $l$ is not justified
\item $n$ is relevant
\item Use a cycle detection algorithm to verify that setting {\tt watched\_parent($l$) = $n$} would not create a dependency cycle
\end{itemize}
If such $n$ can be found, set {\tt watched\_parent($l$) = $n$}.
If such $n$ cannot be found, call {\bf notifyBecomesIrrelevant($l$)}.
\end{description}
The implementation for the search for an alternative watch is a re-use of an existing ``unfounded set detection'' algorithm.
This algorithm is considered the fastest algorithm to achieve this task to date.
\subsection{Detecting Cycles}
\label{sec:ufsdetection}
For our implementation of the detection of cycles, we re-use parts of the existing {\em unfounded set propagation} algorithm~\cite{lpar/MarienWD07,ai/GebserKS12}.
This algorithm has a subcomponent that searches for cycles over negative literals.
\section{Conclusion and Future Work}
\label{sec:conclusion}
The notion of relevance is related to Magic Sets~\cite{jlp/BeeriR91,pods/BancilhonMSU86} in the field of Logic Programming in the sense that the resulting program of the magic set transformation on a program $P$ and query $Q$ does not execute parts of $P$ that do not contribute towards solving query $Q$.
One area where relevance should give great speedups is stable model counting.
When \m{\mathit{p_\m{\mathcal{T}}}} is justified, the number of solutions that this partial assignment represents is equal to $2^{n}$ with $n$ the number of unassigned open atoms in \m{\Delta}.
Stable model counters generally stop when the \emph{justified residual program} \cite{aaai/AzizCMS15} is empty.
Whenever this occurs, \m{\mathit{p_\m{\mathcal{T}}}} is justified.
However, this does not hold the other way round.
There are other cases where \m{\mathit{p_\m{\mathcal{T}}}} is justified, but the justified residual program is non-empty.
As such, exploiting relevance can ensure cutting out bigger parts of the search tree when model counting.
In this paper we have presented our implementation of a relevance tracker module on top of an existing SAT(ID) solver, which consists of two methods; one to keep track of the set of justified literals and one to keep track of the set of relevant literals.
Each of them reuses existing techniques, namely the following.
\begin{itemize}
\item Our method for keeping track of the justification status of literals. This method reuses existing propagation mechanisms.
We also prove the correctness of this approach if the underlying solver guarantees completeness with respect to rule application and unfounded set propagation.
\item Our method for keeping track of the relevance status of literals reuses the existing unfounded set detection techniques~\cite{lpar/MarienWD07,ai/GebserKS12} for detection of cyclic dependencies between ``candidate parents''.
\end{itemize}
Generate-and-test \ASP programs are the most common form of \ASP programs, as can be witnessed, e.g., from the benchmarks in the latest \ASP competitions~\cite{journals/ai/CalimeriGMR16}.
Generate-and-test \ASP programs closely correspond to PC(ID) theories~\cite{DeneckerLTV12}.
This paper imposes minimal assumptions on the underlying solver, thus making it possible to translate these ideas to an \ASP context.
Since experiments have shown~\mycite{relevance} that exploiting relevance during solving can reduce the number of decisions, as well as the number of conflicts, it would be interesting to see how relevance might possibly affects existing ASP solvers.
|
\subsubsection{#1:}}
\newcommand{\tparagraph}[1]{\noindent\textbf{#1:}}
\DOI{foobar}
\cclogo{\includegraphics{by-nc-nd.pdf}}
\begin{document}
%
\setlength{\textfloatsep}{10pt plus 1.0pt minus 2.0pt}
\newcommand{\specialcell}[2][c]{ \begin{tabular}[#1]{@{}c@{}}#2\end{tabular}}
\author*[1]{Hamza Harkous}
\author[2]{Rameez Rahman}
\author[3]{Bojan Karlas}
\author[4]{Karl Aberer}
\affil[1]{EPFL, \mbox{E-mail: <EMAIL>}}
\affil[2]{EPFL, E-mail: <EMAIL>}
\affil[3]{EPFL, E-mail:
<EMAIL>}
\affil[4]{EPFL, E-mail: <EMAIL>}
\title{\huge The Curious Case of the PDF Converter that Likes Mozart: Dissecting and Mitigating the Privacy Risk of Personal Cloud Apps}
\runningtitle{Article title}
\begin{abstract}
{Third party apps that work on top of personal cloud services, such as Google Drive and Dropbox, require access to the user's data in order to provide some functionality. Through detailed analysis of a hundred popular Google Drive apps from Google's Chrome store, we discover that the existing permission model is quite often misused: around \textit{two-thirds} of analyzed apps are \emph{over-privileged}, i.e., they access more data than is needed for them to function. In this work, we analyze three different permission models that aim to discourage users from installing over-privileged apps. In experiments with 210 real users, we discover that the most successful permission model is our novel ensemble method that we call \emph{Far-reaching Insights}. Far-reaching Insights inform the users about the data-driven insights that apps can make about them (e.g., their topics of interest, collaboration and activity patterns etc.) Thus, they seek to bridge the gap between what third parties can actually know about users and users' perception of their privacy leakage. The efficacy of Far-reaching Insights in bridging this gap is demonstrated by our results, as Far-reaching Insights prove to be, on average, \textit{twice} as effective as the current model in discouraging users from installing over-privileged apps. In an effort to promote general privacy awareness, we deployed {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace, a publicly available privacy-focused app store that uses Far-reaching Insights. Based on the knowledge extracted from data of the store's users (over 115 gigabytes of Google Drive data from 1440 users with 662 installed apps), we also delineate the ecosystem for 3rd party cloud apps from the standpoint of developers and cloud providers. Finally, we present several general recommendations that can guide other future works in the area of privacy for the cloud. To the best of our knowledge, ours is the first work that tackles the privacy risk posed by 3rd party apps on cloud platforms in such depth.}
\end{abstract}
\keywords{cloud computing, usable privacy, apps}
\journalname{Proceedings on Privacy Enhancing Technologies}
\DOI{10.1515/popets-2016-0032}
\startpage{123}
\received{ \vspace{-1\baselineskip} 2016-02-29}
\revised{2016-06-02}
\accepted{2016-06-02}
\journalyear{}
\journalvolume{2016}
\journalissue{4}
\maketitle
\section{Introduction}
Cloud services such as Google Drive, Dropbox, OneDrive etc., have become increasingly popular in recent years. At the same time, such services have raised privacy concerns about users' data. But the danger is graver than it appears. While such cloud services are few in number, and, at least, have clearly defined privacy policies, they also serve as platforms that allow a myriad of 3rd party apps to work on top of users' data. These 3rd party apps provide certain functionalities for which they require access to the users' data. Put simply, users sacrifice some of their privacy in order to get functionalities that such apps provide. However, it appears that often such apps acquire more data than is needed for them to function.
As a case study, we did an analysis of a 100 third-party apps on one of the most popular personal cloud services, namely Google Drive (240 million active users in 2014~\cite{GdriveUsers}), and discovered that almost two-thirds (64\%) of these apps require more permissions than they actually need for functioning.
Thus, users often end up exposing more data than is needed to unaccountable apps. For instance, a user's favorite PDF converter is highly likely to get access to her music library and discover her taste in Mozart or obtain her geo-tagged photos and know where she went on the weekend. Throughout this paper, we refer to such apps as \textit{over-privileged apps}.
As observed in other 3rd party apps ecosystems, giving such over-privileged apps superfluous access can potentially result in users' data being abused. This has recently been the case in the health apps market where the top 20 most visited apps were found to be sharing users' data with 70 analytics and advertising companies~\cite{healthApps}.
Nevertheless, the cloud apps ecosystem has unique features that warrant a particular study of this ecosystem. Unlike studies on mobile app ecosystems, where the permissions concern the user's list of contacts, current location, or photos, the cloud permissions allow the 3rd party apps to get access to any file the user has stored in the cloud. Thus, instead of profiling the current user context, such apps can get far-reaching insights inferred from her documents, concerning her financial, legal, or health-related outlook for example. Put simply, the scale and the quality of data that can be collected is both a privacy nightmare for unaware users and a goldmine for advertisers. Second, collecting the permissions of cloud apps is challenging. Unlike other ecosystems where app permissions of thousands of apps can be easily mined via traditional web crawling, each 3rd party cloud app has a unique interface that links to the service providers. Hence, this limits the corpus of apps that one can study.
In this paper, we aim to dissect and mitigate the privacy risk posed by such over-privileged 3rd party cloud apps. We study Google Drive's app ecosystem from the standpoint of all relevant parties, namely, the users, app developers and cloud providers. Generally, we seek to characterize (a)~influencing factors that can deter users from installing over-privileged apps, (b)~conditions determining developers misbehavior, and (c)~the steps cloud providers can take to mitigate users' privacy risks. To the best of our knowledge, ours is the first work that studies the privacy risk of 3rd party apps in personal cloud ecosystems.
Towards that end, we present three different permission models namely: \textit{(a)~Delta Permissions}, \textit{(b)~Immediate Insights}, and \textit{(c)~Far-reaching Insights}. The first model, i.e., \textit{Delta Permissions}, informs users about the unneeded permissions that over-privileged apps are using.
The second model, \textit{Immediate Insights} presents randomly selected examples from the user's data such as portions of text or image files, photo locations etc., that over-privileged apps can get access to.
The third model, \emph{Far-reaching Insights}, has been motivated by the novel concept of \emph{Inverse Privacy}~\cite{inversePrivacy}. Inverse privacy refers to the situation when a user is not aware of the information that an external entity has about the user. Based on this definition, Far-reaching Insights communicate to users the inferences which can be made by the apps with superfluous permissions using advanced text and image analysis techniques. These include but are not limited to user collaboration and activity patterns; the top faces, locations, and concepts that appear in users' photos, etc.
Overall, we make the following specific contributions in this paper:
\textbf{i. Far-reaching Insights sensitize users with intimate details, and promote privacy aware behavior:}
Through extensive user experiments, we discover that our first simple model, \textit{Delta Permissions} fails to deter users from installing over-privileged apps. Put bluntly, \textit{telling users that their privacy is being infringed does not help.} The second model, \textit{Immediate Insights}, does twice as well in discouraging users from installing over-privileged apps. However, the clear winner is our novel model, Far-reaching Insights, which can be twice as effective in deterring users from installing over-privileged apps as Immediate Insights.
Overall, our analysis reveals various factors that can deter users from installing over-privileged apps. For instance, we discover that within Far-reaching Insights, \textit{Relational Insights} (that reveal users' relations with other people) reduce by half the installation of over-privileged apps, as compared to \textit{Personal Insights} (that reveal information about the \textit{users themselves}) (Section ~\ref{sec:anaUser}).
\textbf{ii. {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace helps us profile developer behavior and helps users safeguard their data:}
In an effort to promote privacy awareness in the general public, and to help users safeguard their privacy, we present {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace, a privacy-focused app store that uses Far-reaching Insights to warn users about over-privileged apps. This store is available for public use and has been used by over 1440 registered users.
A considerable fraction of these users has prior experience of using Google Drive 3rd party apps. By automatically getting their previously installed apps' metadata, we anatomize current developers' behavior, point towards potential avenues of misbehavior, and present suggestions to deter misbehavior (Sections~\ref{sec:developers} and~\ref{sec:privacyProtection}).
\textbf{iii. Shared wisdom:}
Finally, based on our analysis we present several easy to implement practical suggestions that can be adopted by cloud providers and by those others working in the domain of privacy to safeguard users privacy in the cloud (Section~\ref{sec:privacyProtection}).
The remainder of this paper is organized as follows. Section~\ref{sec:ThirdParty} describes the 3rd party cloud app ecosystem, its threat model, and the specific case of Google Drive. In Section~\ref{sec:privacyRisk}, we describe in detail our privacy app permissions review process and results. In Section~\ref{sec:Delta}, we present our three permission models, before evaluating them in Section~\ref{sec:anaUser}. Based on the privacy-focused store we have developed (Section~\ref{sec:privysealStore}), we analyze app developer behavior in Section~\ref{sec:developers}. Finally, we give our recommendations for the community in Section~\ref{sec:privacyProtection}.
\section{Third-party Cloud Apps Ecosystem}
\label{sec:ThirdParty}
There are three entities that interact in the 3rd party cloud app system: (1) a \textit{developer} who programs and manages a \textit{3rd party app}, (2) a \textit{user} who uses that app for achieving a certain service, and (3) a \textit{cloud storage provider (CSP)} hosting the user's \textit{data}. Using the CSP's API, the app gets access to a subset of the user's data after \textit{user authorization}, which is based on the user accepting a list of \textit{permissions} that determines this subset.
\vspace{-\baselineskip}
\subsection{Threat Model}
Upon using 3rd party cloud apps that access their data, users sacrifice some of their privacy for getting some service(s). This tradeoff between privacy and services has been called the \emph{Privacy vs. Services Dilemma} in the literature~\cite{barnes_privacy_2006}. However, as we will see in the next section, there are many apps that require more permissions than are needed for them to function. We call such apps \textit{over-privileged apps} (as opposed to \textit{least-privileged apps} that only request the permissions needed for their functionality).
These apps pose a \textit{risk} which can be potentially exploited, e.g., by selling data to 3rd party advertising providers.
A lot of the highly used 3rd party apps do not have privacy policies or justifications of the requested permissions. Moreover, the users are not usually aware of the API details or the app functionality, especially before installing the apps. Thus, the choice of installation is not well-informed from a privacy perspective.
In this work, we consider the 3rd party apps as the adversary (and not the CSP). We seek to combat the risk posed by over-privileged apps through improving the \textit{risk indicators} (permission models in particular) that users are presented with, during the authorization process.
\vspace{-\baselineskip}
\subsection{The Case of Google Drive}
Towards that end, we have taken Google Drive as a case study, and we have anatomized this ecosystem in detail. Nevertheless, the insights gained from our analysis are applicable to other cloud platforms as well.
To begin with, any developer can register an app that accesses Google Drive API at Google Developers Console for free. She then receives a \emph{Client ID} and \emph{Client Secret} that need to be included in the app code to access Google APIs. The developer can then specify in her code a set of Google permissions (a.k.a. scopes) she wants to obtain.
The app itself can be hosted on any website the developer chooses; i.e., it is not hosted by Google itself. The developer can also submit a request for featuring the app on Google Chrome Web Store, which has a section for apps that work with Google Drive. In the store, apps are presented along with screenshots and descriptions of their functionality (provided by the developer). The store also allows users to rate and review apps.
Apps can be also submitted to other web stores hosted by Google, such as the Add-ons Stores for Google Docs, Google Sheets, or Google Slides and the Google Apps Marketplace for enterprises. However, there are a lot of apps that exist outside these stores.
An app can request permission to access Google Drive data at any time of its operation, and not necessarily at the beginning. For example, the user can be presented with a button in a side menu that reads ``Import file from Google Drive'', and clicking on this button redirects to a Google-hosted page that presents the set of permissions requested by the app, as shown in Figure~\ref{fig:baseline_pdf_converter}. The user has to accept all these permissions to connect the app to her Google Drive. She cannot select a subset of them at installation time or later. However, she may revoke the app authorization completely from her Google account settings. As we see later, the absence of a standard location and interface for hosting apps and triggering the permissions request is one of the reasons that makes the automated, large-scale privacy analysis of apps infeasible.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{figures/baseline_pdf_converter.pdf}
\caption{Example of the current permissions interface of Google Drive}
\label{fig:baseline_pdf_converter}
\end{figure}
The main permissions pertinent to Google Drive are presented in Table~\ref{tab:defaultPermissions}, along with the Google-provided description for each. This short description is also presented to the user, and a longer explanation is available via clicking the info button \circled{i} next to each permission.
\begin{table}
\scriptsize
\centering
\begin{tabular}{L{4.5cm}|L{3cm}}
\hline \textbf{Permission} & \textbf{Short Name} \\
\hline View and manage the \textbf{files} in your Google Drive. & \textsc{drive}\\
\hline View the \textbf{files} in your Google Drive. & \textsc{drive\_readonly}\\
\hline View and manage \textbf{metadata} of files in your Google Drive. & \textsc{drive\_metadata}\\
\hline View \textbf{metadata} for files in your Google Drive. & \textsc{drive\_metadata\_readonly}\\
\hline View and manage Google Drive files that you have \textbf{opened or created with this app}. & \textsc{drive\_file}\\
\hline View your Google Drive {apps}. & \textsc{drive\_apps\_readonly} \\
\hline Add itself to Google Drive. & \textsc{add\_drive} \\
\hline View and manage its {own configuration data} in your Google Drive. & \textsc{drive\_appdata}\\
\hline
\end{tabular}
\newline
\caption{Requested permissions with the short name we use for reference}
\label{tab:defaultPermissions}
\end{table}
As far as files' data is concerned, an app can request access to all files (\textsc{drive} and \textsc{drive\_readonly} permissions) or on a per-file basis (\textsc{drive\_file}). When developers request \textsc{drive\_file} only, the explicit approval for each new file(s) is mediated by an interface provided by Google. For example, the developer presents the user with a Google-hosted file picker popup (Figure~\ref{fig:file_picker}) so that she can select (and thus approve access to) the file. Alternatively, the file can be opened from Google Drive's interface via the ``Open with'' option in the context menu of the file. In the case of full access, an app can access any file directly via Google Drive API without the need for user intervention. For example, this type of access enables an app to obtain all the user's files and download them in the background.
The developer can alternatively request access to file metadata only via \textsc{drive\_metadata} or \textsc{drive\_metadata\_readonly} (allows accessing filenames, editing dates, photos' Exif information, etc.). Additionally, the developer can request access to the list of apps the user has authorized before via \textsc{drive\_apps\_readonly}.
It is worth noting that the permission list is not limited to Google Drive API and that it typically includes permissions from other Google APIs, such as access to user's profile information, email address, contacts list, etc.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/file_picker}
\caption{Google Drive file picker interface }
\label{fig:file_picker}
\end{figure}
\section{Privacy Risk of 3rd Party Google Drive Apps}
\label{sec:privacyRisk}
The question that comes next is: \textbf{``what is the extent of risk that actual users are exposed to?''} To answer this, we examined a sample of 3rd party Google Drive apps to determine the percentage of apps that request extra permissions.
We proceeded to Google Chrome Web Store, which has a section for apps that work with Google Drive.
The store features apps on its main page, that change with time. In total, there are around 420 apps on the store that are labeled as ``\textit{Works with Google Drive}''. We selected 100 featured apps at random from the main page (during May 2015), and we manually reviewed them one by one. Hence, our sample represents around one-fourth of the whole set of apps in the store, which is one of the main avenues for finding Google Drive apps. As we discuss later in Section~\ref{sec:anatomDevCurrent}, we discovered that, in our real word sample of 1440 users, around one-fourth of the installed apps are from the Google Chrome Store.
More details on the apps' dataset are presented in Appendix~\ref{apndx:reviewData}.
\subsection{Permissions Review Process}
\label{sec:revProcess}
We now explain the App Permissions Review (APR) methodology we followed, and we refer the reader to Figure~\ref{fig:review} in Appendix~\ref{apndx:reviewData} for the corresponding flowchart. Our methodology is inspired by Google Drive's guide for choosing authentication scopes.
Each APR aims to get: (a) \textbf{set $P$ of requested permissions}, (b) \textbf{set $S$ of sufficient permissions} for the app functionality.
We start each APR by going to the app's website, linked from the store, and testing the app manually. For each app, we first find the step where the app can be connected to Google Drive (if this is not upon the initial sign up). Then, we record the set $P$ of requested permissions and authorize the app to access a test Google Drive account created for this purpose, and we record the permissions requested. If \textsc{drive\_file} (i.e. minimal per-file access) is the only Google Drive permission requested, the app review is complete (${S=\{\textsc{drive\_file}\}}$). Otherwise, we continue to check the app's interface for all file pickers that allow importing files from Google Drive (in almost all the cases, there is at most one file picker).
In the first case where the app solely uses Google's official file picker of Figure~\ref{fig:file_picker} (e.g., an app that allows users to convert specific files to PDF format), we set ${S=\{\textsc{drive\_file}\}}$.
In the second case where we find that there are no file pickers in the interface and that the app functionality does not require access to any file, the app is labeled as not requiring any file permissions (${S=\{\}}$).
In the case where the app includes a custom file picker, we decide that (a) $S=\{\textsc{drive}\}$ if the app's declared functionality necessitates files' content (e.g., a photo collage app with custom photos browser) or (b) ${S=\{\textsc{drive\_metadata}\}}$ if the functionality does not need the content (e.g., an app that visualizes who has access to a selected folder).
Similarly, if the app has no file picker, we decide that (a) ${S=\{\textsc{drive}\}}$ if the app's declared functionality necessitates file content (e.g., malware scanning apps for Google Drive that do not need a file picker) or (b) ${S=\{\textsc{drive\_metadata}\}}$ if the functionality necessitates file metadata (e.g., an app that visualizes all collaborators with access to user's files).
Finally, we label an \textbf{app as over-privileged} if either (a) the set $S$ is empty and $P$ is not or (b) if the set $P$ contains at least one permission that is more demanding than all permissions in $S$ (the five file-related permissions in Table~\ref{tab:defaultPermissions} are listed from the most demanding down to the least demanding).
We also determine the set of \textbf{unneeded permissions $U$}, composed of each permission in $P$ that is more demanding than all permissions in $S$. The set of \textbf{needed permissions} is given by $N = P \setminus U$.
\subsection{Review Results}
Analyzing the APRs, we found out that \textbf{64 out of 100 apps} request unneeded permissions.
In other words, the developers could have requested less invasive permissions with the current API provided by Google.
In total, 76 out of the 100 apps requested full access to the all the files in the user's Google Drive. Moreover, the 64 over-privileged apps have actually all requested full access. Accordingly, in our sample, around \textbf{84\% (64/76) of apps requesting full access} are over-privileged.
\begin{table}[t]
\centering
\scriptsize
\begin{tabular}
{ |C{4cm}|c|c|}
\hline \textbf{Permission} & \textbf{Unneeded} & \textbf{Needed}\\
\hline \textsc{drive\_readonly} & 17 & 1\\
\hline \textsc{drive} & 55 & 7\\
\hline \textsc{drive\_metadata\_ readonly} & 2 & 1\\
\hline \textsc{drive\_file} & 0 & 41\\
\hline
\end{tabular}
\caption{Permissions' Usage in APRs}
\label{tab:permissionStats}
\end{table}
The top permission that is needlessly requested is the full read and write access to Google Drive (in 55 apps), followed by the full read access permission (in 17 apps). This further increases the magnitude of data that can be exploited with the extra permissions.
On the other hand, the per-file access permission is the top permission that is actually needed when requested. This happens in 41 of the apps. However, in \textbf{16 of these 41 apps}, we have found that the developer also requested full access to the user's data. Accordingly, developers are sometimes mixing full-access with partial access (which is a subset of the former).
We note that such mixing of permissions can either be the result of developer incompetence, or it may be aimed at deceiving the user. Regardless, such apps pose a risk which can be potentially exploited.
Another outcome of the APR was that \textsc{drive\_file} was the top alternative permission (in 48 apps) that could replace the unnecessary permissions requested. \textsc{drive\_metadata\_readonly} was the alternative for one app only. This indicates that, simply, the correct usage of the current Google Drive API (which does provide per-file access), can eliminate the major part of the privacy risk. Nevertheless, it is evident that developers are generally guilty of not doing this.
\subsection{Automating the APR Process}
Being an external party, we do not have access to the full list of Google Drive apps with their permissions. Hence, the \textit{first} task we had to do was to find the position in each app where Google Drive permissions are requested. This is not always on the main page of the app, and sometimes finding it requires navigating multiple menus and/or pages.
Automating this task involves building an advanced web crawler that can retrieve the permissions from a large number of such apps by smartly searching for the sign-in button.
The \textit{second} task was checking the functionality of the app in order to see if it matches the requested permissions. Automating the process of over-privilege detection or of real time private data leakage detection has been tackled in the mobile apps scenario (e.g. in~\cite{Felt:2011:APD} and~\cite{Enck:2010}). However, in the mobile scenario (or any similar architecture), the user's device hosts the data, the 3rd party apps, and the detection/monitoring solution. Cloud apps present a radically different scenario as the data is hosted by the CSP, the 3rd party app is served at a developer-specified location, and any detection/monitoring app would operate from outside. The only part of the code that the 3rd party app exposes is the client side code. Hence, all techniques that check the app's code (e.g., via static/dynamic analysis) or its inputs/outputs cannot be transplanted to the cloud app case as they would evidently underestimate what APIs/permissions the app might need.
The only automatic way we perceive for over-privilege detection is to cluster apps of similar functionality and identify the ones which request more permissions than others in the same cluster. Even then, the data collected manually would be used as the ground-truth to evaluate the automated method. Detecting actual data leakage is much more challenging in the cloud apps scenario as the app can send users' data to other parties from the server side (which is impossible to monitor via external solutions).
Faced with these limitations, manual expert reviews are the closest we can get to assessing the apps' needed permissions. Still, we do not claim that this method is perfectly accurate as a developer might be working on a non-advertised feature that requires new permissions. However, we conjecture that APRs are accurate with the vast majority of the reviewed apps\footnote{From our experience over one year, rarely did apps introduce new features that required new permissions. Moreover, in Section~\ref{sec:privysealStore}, we discuss how to further alleviate this concern in a real-world deployment by allowing developers to submit rebuttals.}.
Finally, as our main purpose in this work is to characterize the ecosystem and suggest alternative permission models, automating both the app permissions collection and the over-privilege detection tasks falls out of the scope of this work. We note though that we are concurrently working on the specific research problem of designing automated APRs.
\section{New Permission Models}
\label{sec:Delta}
In the light of the risk that over-privileged apps pose, we propose in this section three alternatives to the existing permission model in Google Drive before evaluating their efficacy in mitigating the risk in the next section.
\subsection{Delta Permissions}
Our first model is based on the following hypothesis:
\textit{``When users are informed about the unneeded permissions being requested by apps, they are less likely to authorize such apps.''}
\begin{figure}[t]
\centering
\includegraphics[width=0.8\linewidth]{figures/delta_example.pdf}
\caption{Example of Delta Permissions interface}
\label{fig:delta_example}
\end{figure}
Hence, this model replaces the current permissions interface displayed in Figure~\ref{fig:baseline_pdf_converter} with a new interface, presented in Figure~\ref{fig:delta_example}. We call this permission model \textit{Delta Permissions (DP)}, and it reveals to the user the distinction between permissions that are needed for the app functionality and those others (the delta) that are unnecessarily requested.
\subsection{Immediate Insights}
\label{sec:immediateInsights}
The second model is based on the following hypothesis:
\textit{``When users are shown samples of the data that can be extracted from the unneeded permissions granted to apps, they are less likely to authorize these apps.''}
Accordingly, we show users randomly selected data examples, directly extracted from their Google Drive, such as excerpts of text or image files, photo locations, or people she collaborated with.
An instance of this model, which we call \textit{Immediate Insights (IM)}, is given in Figure~\ref{fig:personalizedInsights}. On the left, we have the same previous \textit{DP} interface. On the right, we have the \textit{Insights Area}, where we show a question that says: ``What do the \textbf{unneeded permissions} say about you?'', followed by an answer in the form of a visual with short explanatory text. In this figure, the Insights Area visualizes the location where a randomly chosen user photo was taken.
In the following, we describe the design of the \textit{IM} Insights:
\tparagraph{Image}
We show an image selected at random from the set of user's image files.
\tparagraph{Location}
We randomly choose a photo from the set of user's image files, such that it includes a GPS location in its Exif data. Then we show that photo on a map centered at that location (as in Figure~\ref{fig:personalizedInsights}).
\tparagraph{Text}
We show the user an excerpt from the beginning of a randomly chosen textual file.
\tparagraph{Collaborator}
We show the profile picture and the name of a randomly chosen collaborator.
\begin{figure*}[t!]
\centering
\minipage{0.73\textwidth}
\fbox{\includegraphics[width=\linewidth]{figures/personalized_insights.pdf}}
\caption{Example of Immediate Insights interface; the same layout is used for Far-reaching insights, with the insights area content changing accordingly}
\label{fig:personalizedInsights}
\endminipage\quad
\minipage{0.23\textwidth}
\fbox{\includegraphics[width=\linewidth]{figures/entities.pdf}}
\caption{Named \textit{Entities} insight}
\label{fig:entities}
\endminipage
\end{figure*}
\begin{figure*}[t!]
\centering
\minipage{0.26\textwidth}
\fbox{\includegraphics[width=\linewidth]{figures/sentiments.pdf}}
\caption{\textit{Sentiments} insight}
\label{fig:sentiments}
\endminipage\hfill
\minipage{0.24\textwidth}
\fbox{\includegraphics[width=\linewidth]{figures/sharedInterests.pdf}}
\caption{\textit{SharedInterests} insight}
\label{fig:sharedInterests}
\endminipage\hfill
\minipage{0.23\textwidth}
\fbox{\includegraphics[width=\linewidth]{figures/faceswithcontext.pdf}}
\caption{\textit{FacesWithContext} insight}
\label{fig:faceswithcontext}
\endminipage\hfill
\minipage{0.22\textwidth}
\fbox{\includegraphics[width=\linewidth]{figures/facesonmap.pdf}}
\caption{\textit{FacesOnMap} insight}
\label{fig:facesonmap}
\endminipage
\end{figure*}
\subsection{Far-reaching Insights}
\label{sec:farInsights}
The third model is based on the following hypothesis:
\textit{``When the users are shown the far-reaching information that can be inferred from the unneeded permissions granted to apps, they are less likely to authorize these apps.''}
These are insights that go beyond examples and include what can be inferred by running more involved algorithms, such as sentiments towards entities, objects identified in photos, faces detected, etc. Hence, we denote this model by \textit{Far-reaching Insights} (or shortly \textit{FR Insights}).
The interface layout is the same as that of Figure~\ref{fig:personalizedInsights}, but with the Insights Area containing an \textit{FR} insight instead of an immediate insight.
In this work, we have designed 6 types of \textit{FR} insights that can be extracted from users' data. Below, we briefly describe each of them and we refer the reader to Appendix~\ref{apndx:Insights} for the algorithms used for their generation:\\
\tparagraph{Entities, Concepts, and Topics (ECT)}
The first type of insights we form is based on applying various NLP techniques to extract \textit{named Entities (E)}, \textit{Concepts (C)}, and \textit{Topics (T)} from users' textual files. Entities, extracted via \textit{Named Entity Recognition}, might include people the user works with, companies she talks about, places she plans to visit, etc.
Concepts are extracted tags from the user's documents. For example, the sentence ``My favorite brands are BMW, Ferrari, and Porsche'', would be tagged by the concept ``Automotive Industry''. Topics are higher level abstractions (e.g. technology, art, business, etc.) that can be used for classifying users' documents. Both concepts and topics can serve for profiling users' interests. We combine these together due to the similar nature of these insights. When we use \textit{ECT}, one of $E$, $T$ or $C$ is randomly displayed to the user in the Insights Area.
For purposes of text analysis in this work, we partially used AlchemyAPI's service with random excerpts of text extracted from users' documents. Users in our experiments were informed about this on the main page of the web app they sign in to.
\tparagraph{Sentiments}
We used sentiment analysis in order to identify the entities with the most positive or negative sentiments and then display them to the user in the Insights Area as in Figure~\ref{fig:sentiments}.
\tparagraph{Top Collaborators}
The next insight we added displays the top collaborators a user has, based on the analyzed files. These typically include close work colleagues, intimate friends, or people the user goes out with and shares pictures with afterward.
\tparagraph{Shared Interests}
In this insight, we represent the user's mutual topics of interests with a group of people. As shown in Figure~\ref{fig:sharedInterests}, the Insights Area would then contain a list of people alongside different topics of documents shared with these people.
\tparagraph{Faces with Context}
We now come to the insights that are based on features inside the user's images. The first insight of this type shows a group of faces, representing the most frequent people appearing in the user's images, alongside the concepts that appear in the same images (see Figure~\ref{fig:faceswithcontext}). One can imagine that such information might be valuable, for example, to advertisers that aim to extract the user's interests in certain people, products, or services.
\tparagraph{Faces on Map}
In addition to the image content itself, image metadata can be also sensitive, especially the geographical location where the image is captured. Hence, this insight, shown in Figure~\ref{fig:facesonmap}, relays to the user the places where her photos are taken, in addition to the faces and items in those photos.
\noindent The component diagram of Figure~\ref{fig:insights_generation} summarizes the techniques used for generating each of the \textit{FR} insights.
\tparagraph{\textit{Further Notes}}
\begin{figure}[t]
\centering \includegraphics[width=0.7\linewidth]{figures/insights_generation.pdf}
\caption{Component diagram mapping the used analysis techniques to the generated insights}
\label{fig:insights_generation}
\end{figure}
We note that the reasoning behind designing lightweight models such as \textit{DP} and \textit{IM} was that we wanted to examine whether designing heavyweight insights such as \textit{FR} Insights is worth the effort for us and the potential adopters of our approach, or do users respond equally favorably (or badly) to both the heavy and the lightweight approaches, in which case \textit{FR} insights need not be adopted?
We also note that, for an app that does not request unneeded permissions (even if it requests full access), the Insights Area will simply show a text saying that the app doesn't require any extra permissions.
We also note that we follow Google Drive's approach of requesting permissions ``At Setup''~\cite{schaub2015design} (i.e., at the first time of app authorization).
This is unlike other ecosystems (e.g., iOS or Android M), which require a ``Just in Time'' approach (i.e., permissions are requested only when the actual functionality is needed).
This is because, in Google Drive, many apps are supposed to work with the user's data even when she is offline. Hence, granting access in an interactive manner for individual permissions is not always feasible.
\section{Evaluating the Models}
\label{sec:anaUser}
\subsection{Experimental Setup}
\label{sec:Exp1}
We designed an experiment with actual users in order to test the hypothesis of whether the new models can better deter the users from installing over-privileged apps as compared to the existing one, and to discover factors that influence users' decisions.
\minisec{User Recruitment}
In order to recruit users, we primarily used our university's mailing list.
The users were briefed about an app that is related to protecting the privacy of their data against 3rd party apps on Google Drive. The news about the app was also reported on the university's website and was picked up by several technology websites.
The website described itself as an app for Google Drive that aims at exposing what 3rd party web apps can needlessly get about users.
Via our website, the users can sign in to their Google account and then grant full Google Drive access to our app.
Next to the ``sign in'' button, we linked to our privacy policy, explaining what data the app gets and what it keeps.
Only those users who had at least 10 files containing text or 20 images were allowed to continue. This is to ensure that they possess at least a minimal level of experience with Google Drive. Figure~\ref{fig:userFileStats} of Appendix~\ref{appndx:experiment} shows the density plot of the percentages of users' analyzed files that are textual. Next, users who agreed to participate in our experiment were randomly assigned to one of the groups described below. As a motivation to complete the experiments, the users were enrolled in a lucky draw, where they could win one of five gift cards to a mobile app store of their choice.
\vspace{-\baselineskip}
\minisec{Methodology}
The first goal of the experiment is to investigate the efficacy of the three new permission models by comparing them to the existing Google Drive permission model as well as to each other. The second goal is to perform micro-comparisons among the different types of \textit{IM} and \textit{FR} insights. Accordingly, we went for a mixed between-subject and within-subject design. The reason for not going for a complete between-subject design was the large number of participants needed for statistically significant results with 12 independent groups (for all micro-comparisons). The reason for not going for a complete within-subject design was to avoid any participants' bias that can result from showing the existing interface they are used to and the new interface we developed in the same experiment.
Accordingly, we had four groups in our experiment. A user is assigned to one single group, and the only permission interface she sees during the experiment is that of its group. The groups were:
\begin{enumerate}
\item
\textbf{Baseline group (\textit{BL} group):} Users in this group were presented with a clone of the original interface that Google shows upon installing the app (shown in Figure~\ref{fig:baseline_pdf_converter}). This group serves as the control group, and we briefly refer to it as \textit{BL}
\item
\textbf{Delta Permissions group (\textit{DP} group):} Users in this group were presented with the modified interface, previously shown in Figure~\ref{fig:delta_example}.
\item
\textbf{Immediate Insights Group (\textit{IM} group):} Users in this group were presented with the modified interface of Figure~\ref{fig:personalizedInsights}, with the Insights area containing one of the \textit{IM} insights of Section~\ref{sec:immediateInsights}.
\item
\textbf{Far-reaching Insights Group (\textit{FR} group):} Users in this group were presented with the modified interface, of Figure~\ref{fig:personalizedInsights}, with the Insights Area containing one the \textit{FR} insights described in Section~\ref{sec:farInsights}.
\end{enumerate}
A user experiment was divided into multiple \textit{tasks}.
In each task, the user was requested to select an app with a specified \textit{goal}.
For example, the goal would read ``Select the app which allows you to extract the ZIP files on your Google Drive'', and the corresponding app would be ``ZIP Extractor''.
The user would then choose this app among other apps that are listed in the interface. We show this interface in Figure~\ref{fig:task} of Appendix~\ref{appndx:experiment}, and we note that it is similar to the actual Google Chrome Web Store.
Moreover, only one app of those listed satisfies the given goal, and it is highlighted in the interface.
This part of the setup only serves a gamification purpose in order to keep the user interested.
Once the user selects the app, she is presented with a permission interface that corresponds to its group (i.e. that of Figure~\ref{fig:baseline_pdf_converter} for the \textit{BL} group, Figure~\ref{fig:delta_example} for the \textit{DP} group, and Figure~\ref{fig:personalizedInsights} with a randomly selected visual for the \textit{IM} or \textit{FR} Insights groups).
The user is then presented with a question that says: ``Based on permissions below, would you be likely to install this app?''. She can choose between ``Permissions are too invasive'' (accept) and ``I'm OK with these permissions'' (reject).
We worded the question so that we avoid all users rejecting the installations of all apps. We rather aimed that users would reject apps whose permissions they consider as too invasive, thus allowing us to do within-subject comparisons.
After answering the question, the user is directed to the next task with another app, until she completes the whole set of tasks.
The apps used in the experiment were obtained from the Google Drive section of the Chrome Web Store. For experimental purposes, we modified the permissions requested by these apps to be able to test various conditions.
Unlike in the store, we removed elements such as ratings, user reviews, and screenshots and kept a minimal interface, allowing the users to focus solely on the app permissions.
We also avoided using apps from popular vendors to avoid the bias resulting from users being influenced by famous brands.
These steps were taken to study the effect that the permission model has on the user's decisions, without the influence of extraneous factors\footnote{Incidentally, the user might confront a scenario exactly as in the experiments if she does not find the app from the store, but lands on a certain site that has the option of authenticating with Google Drive.}. Moreover, the apps were presented to the users in randomized order to compensate for the effects of learning and fatigue.
For reference, the permissions that each app requested are presented in Table~\ref{tab:experimentPermissions}.
A user assigned to the \textit{BL} or \textit{DP} groups had to install 5 apps in 5 tasks. For the \textit{IM} Insights and \textit{FR} groups, we added additional apps. This is because we wanted to compare the effects of the different kinds of insights. The permissions of the additional apps were fixed to those of (ZIP Extractor), but the insights displayed were changing. For each user, the insights were assigned at random to each of those added apps. In total, users assigned to the \textit{IM} Insights and \textit{FR} Insights groups had to complete 8 and 10 tasks respectively.
\begin{table}[t]
\scriptsize
\centering
\begin{tabular}{L{2cm}| C{0.5cm} C{1.2cm} C{0.8cm} C{0.9cm}}
\hline App & \textsc{drive} & \textsc{drive\_ metadata}& \textsc{drive\_ file} & group\\
\hline ZIP Extractor & R,U & & R,N & 1,2,3,4\\
\hline Xodo \textsc{pdf} Viewer \& Editor & & R,U & R,N & 1,2,3,4\\
\hline WhoHasAccess & R,U & R,N & & 1,2,3,4\\
\hline Video Converter & R,U & & & 1,2,3,4\\
\hline Cloud Convert & R,U & & NR,N & 1,2,3,4\\
\hline HelloFax & R,U & & R,N& 3,4\\
\hline Heap Note & R,U & & R,N& 3,4\\
\hline Photo to\\ Cartoon & R,U & & R,N& 3,4\\
\hline PDFUnlock & R,U & & R,N& 4\\
\hline HelloSign & R,U & & R,N& 4\\
\hline
\end{tabular}\caption{Permissions of apps in the experiment. \textit{R}: Requested, \textit{N}: Needed, \textit{U}: Unneeded, \textit{NR}: Not Requested}
\label{tab:experimentPermissions}
\end{table}
\minisec{Data Protection and Ethics}
Respecting the user privacy when working with cloud data is of fundamental importance. Our experiments were done according to a code of ethics protecting this privacy. In particular, after generating the insights from a user's files, these files are deleted immediately from our apps' servers. As per our displayed privacy policy, only the insights' data presented to the user is kept in the app database. Moreover, the user is given the option to delete her insights data at any time with a single click in the app's menu. The database dump we ran our analysis on was isolated from the one to which the deployed web server connects. Also, we use the \textit{https} protocol so that users can securely connect to our system. Before data analysis, we anonymized any occurrence of names and emails in the database by applying a one-way MD5 hash on them. At all times, we refrained from manually checking the database for users' insights.
All the images used in this paper are in the public domain, and the insights shown do not belong to real users. For further transparency, all the libraries and frameworks used for building the tool and data analysis were listed and linked to from the main page of the website. Although an IRB review was not performed beforehand, this paper was subsequently reviewed by our university’s IRB, which did not object to publishing the results.
\subsection{Results}
\label{sec:reviewResults}
We got 210 users in total who successfully completed this part of the experiment. Out of them, 55 were in the \textit{BL} Group, 50 in the \textit{DP} Group, 54 in the \textit{IM} Insights group, and 51 in the \textit{FR} group.
We start by interpreting the results of our user experiment and comparing the efficacy of the various permission models.
The metric we used in our evaluation is the \textit{Acceptance Likelihood $AL$}, defined as:
\begin{equation}
AL= \frac{ \# \mbox{(\textit{Accepts}) } }{ \# \mbox{(\textit{Accepts})} + \#\mbox{(\textit{Rejects})} },
\end{equation}
where \textit{Accepts} denotes the cases where users were fine with the permissions, and \textit{Rejects} denotes the cases where they found them too invasive. \textit{Accepts} and \textit{Rejects} are aggregated across users and tasks for the permission model under consideration. The lower the $AL$, the better the performance in deterring users from installing over-privileged apps.
\begin{figure}[t!]
\minipage{0.5\textwidth}
\includegraphics[width=.95\linewidth]{figures/visuals_comparison_embed.pdf}
\centering
\caption{$AL$ for the different types of interfaces; numbers next to each bar are the error values at 95\% confidence interval}
\label{fig:visuals_comparison}
\endminipage\hfill
\minipage{0.5\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{figures/pvaluesTable_mixed_app_embed.pdf}
\caption{$p$\mbox{-values} of pairwise tests; if $p$\mbox{-value}$\le0.05$ we consider the visuals on the corresponding row and column as different; the difference direction is obtained from Figure~\ref{fig:visuals_comparison}}
\label{fig:pvaluesTable}
\endminipage
\end{figure}
In order to compare the effect of different interfaces, we plotted in Figure~\ref{fig:visuals_comparison} the Acceptance Likelihood for the \textit{BL} and \textit{DP} groups and also for each particular insight of the \textit{IM} and \textit{FR} Insights groups.
To evaluate the significance of the $AL$ differences among the interface types, we fit a generalized linear mixed model (GLMM) with the user's decision (Accepting/Rejecting the app installation) as the binary response variable and the interface type as the fixed effect. Participants' IDs and apps' names were fitted as random effects to control for the potential between-participants and between-apps variabilities. The model was fit assuming a binomial distribution and a logit link function, using the glmer function in the \textit{lme4} package in \textit{R}~\cite{lme4}. Visual inspection of residual plots did not reveal any obvious deviations from homoscedasticity or normality. The significance of differences among $AL$ values was determined using Tukey Honest Significant Difference test with the \textit{glht} function of the \textit{multcomp} package~\cite{multicomp}. The difference between the $AL$ of any two interfaces in Figure~\ref{fig:visuals_comparison} is significant if the corresponding row and column intersect at a $p$\mbox{-value}$\le0.05$ in Figure~\ref{fig:pvaluesTable}.
\tparagraph{Inefficacy of Baseline and Delta Permissions}
We first found that the Delta Permissions and Baseline approaches performed closely ($AL$ of 0.42 and 0.39 respectively) without a statistically significant difference ($p$\mbox{-values}$=0.77$). Hence, we found no evidence of any advantage that the \textit{DP} can introduce, which means that telling our experiment's participants explicitly about unneeded permissions did not help deter them from installing over-privileged apps. We also observe that both these interfaces had a significantly higher $AL$ (i.e. $p$\mbox{-values}$\le 0.05$) than all the insights, except for the Collaborator insight. This highlights the fact that showing well-selected insights will result in deterring more users compared to the case of not showing any insights.
\tparagraph{The Power of Relational Insights}
The next interesting outcome is that there is a category of insights (Category 1) composed of \{\textit{Image}, \textit{Text}, \textit{ECT}, and \textit{Sentiments}\} that are all associated with a significantly higher acceptance likelihood than the category composed of \{\textit{FacesWithContext}, \textit{TopCollaborators}, and \textit{SharedInterests}\} (Category 2)\footnote{The number of users who had location-tagged photos was low; hence, we could not obtain highly significant results in the case of \textit{Location} and \textit{FacesOnMap} insights. }.
Since this is a very interesting result, we investigate further to analyze the defining characteristics of these two naturally clustered categories. The main feature of Category 1, which includes both \textit{IM} and \textit{FR} insights, is that insights in this category are restricted to characterizing the user \textit{herself}, such as showing text excerpts from her documents, topics appearing in them, or images she has in her files. Hence, we denote this category as \textit{Personal Insights}.
On the other hand, the defining feature of Category 2 insights, which are all Far-reaching, is that they extend to characterizing the relationships of the user \textit{with other people}. For instance, \textit{FacesWithContext} shows the most important faces in user's photos along with the items appearing with them. \textit{SharedInterests} shows the people who collaborate with the user and the type of topics they share. Also, \textit{TopCollaborators} identifies the most frequent people the user interacts with. We denote these as \textit{Relational Insights}. From our results, we can conclude that Relational Insights promote greater privacy awareness in users, as such insights are more likely to dissuade them from installing over-privileged apps.
\tparagraph{Impact of Face Recognition}
Delving deeper into more results brought forth by the comparison of different insights, one can notice that showing examples of user's images ($AL=0.21$) is significantly less deterring than showing the important faces and listing the concepts in the image ($AL=0.08$) with pairwise comparison $p$\mbox{-value}$ <0.01$. This highlights the fact that users are sensitive towards the output of face detection and object recognition in photos. Given that services such as Google Photos, OneDrive, and Flickr already apply such techniques to facilitate search, the above result highlights that they can also be used by these companies to easily implement solutions such as ours for raising users' privacy awareness when sharing data.
\tparagraph{Influence of High-Level Textual Insights}
Contrary to the case of images, in the case of textual documents, showing the high-level entities or concepts extracted from the text does not seem to have a significant difference as compared to simply showing direct excerpts from the text ($p$\mbox{-value}$=0.94$). Only when the relationship factor is introduced does the $AL$ significantly decrease (as in the case of \textit{SharedInterests}).
\tparagraph{Superiority of Far-reaching Insights}
By aggregating the results over all the experiments with \textit{FR} Insights, we obtained a lower $AL$ value compared to \textit{IM} Insights ($AL=0.161$ and $0.226$ respectively). To check the statistical significance of this difference, we followed the previous methodology and fit a GLMM model, but with the fixed effect being the experimental group instead of the specific interface. We confirmed that the $AL$ difference is significant with a pairwise comparison $p$\mbox{-value}$=0.004$). We also noticed from Figure~\ref{fig:pvaluesTable} that the best Far-reaching insight, \textit{FacesWithContext} ($AL=0.081$), performed more than twice better than the best Immediate Insight, \textit{Text} insight ($AL=0.206$) (ignoring the insights where the difference is not statistically significant).
Overall, these results demonstrate the superiority of our novel approach of \textit{FR} Insights. Nevertheless, \textit{IM} Insights are still significantly better than the \textit{BL} and \textit{DP} models. This goes in line with the findings of~\cite{Harbach:2014}, which showed the goodness of an approach similar to Immediate Insights in the case of Android permissions, even though they didn't have Delta Permissions as a building block.
\subsection{Limitations}
First, our design of the experiment abstracted several other factors that users take into account when installing apps. The interplay between ratings, app's brand, and permissions has been studied before (\cite{Kelley:2013} and~\cite{Harbach:2014}), and it might be worth revisiting in a future work in the context of our new permission models. Second, our experiment's advertisement included a mention of privacy as we wanted participants to focus on the app permissions. Evidently, this might have made participants more alert towards this issue. Both these points can imply that the real values of the $AL$ might be different in reality, where privacy might not be the main factor.
Nevertheless, even if the absolute values of $AL$ have been impacted, the relative advantages of new permission models still hold.
Moreover, we also note that the users in the $FR$ and $IM$ groups had to do more tasks than the $BL$ and $DP$ groups, which might have resulted in more user fatigue and habituation in the $FR$ and $IM$ groups. This was counteracted first via task randomization at design time and second by the very nature of insights that change at every step. For further validation, we computed the $AL$ values of Figure~\ref{fig:visuals_comparison}, considering only the first 5 tasks each user performed. We did not see any major deviation from the results with all tasks included.
Finally, our user recruitment strategy was primarily targeted towards our university's network, and our study was only for English speakers. It would be interesting to see how the results compare in a more general sample (linguistically, demographically and geographically).
\section{PrivySeal: A Privacy-Focused App Store}
\label{sec:privysealStore}
Driven by the magnitude of the risk posed by over-privileged apps in Google Drive, we were motivated to bring the advantages of the Far-reaching Insights interface to the user community of this platform.
One approach towards achieving that would be for Google itself to implement a scheme similar to ours and to integrate it within the app authorization process. However, we decided not to wait and chose an alternative approach, which is independent of the company's plans and is ready for user utilization immediately. We built {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace, a privacy-focused store for Google Drive apps, which is readily available at \mbox{\url{https://privyseal.epfl.ch}}.
{{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace allows users to navigate a list of apps, click on those of interest, and check whether they are over-privileged via our \textit{FR} Insights interface. Users can also search by keyword for apps, specifying criteria such as the app being least-privileged.
The component diagram for {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace is shown in Figure~\ref{fig:architecture}.
Similar to the APRs we conducted, we have included a ``Review Wizard'' inside {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace for indicating the requested, needed and unneeded permissions along with the alternative permissions the developer could have used. This responsibility is currently given to a small set of expert developers and is moderated by the store administrators. Developers who would like to object to existing APRs of their apps can submit rebuttals.
Currently, {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace has 100 apps and more than 1440 registered users, with a geometric mean of around 50 new users per month (whose vast majority is signing up out of interest in the app after reading article(s) about it).
We finally note that {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace gets access, as is the case with other apps, to users' data to generate insights. Hence, users are assumed to trust the provider of such a ``Privacy-as-a-Service'' solution. However, this assumption of trust will hold if a solution such as {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace is hosted by the CSP itself (which already possesses the data), or an enterprise protecting its documents from 3rd party apps. The assumption of trust is also valid if the users choose to trust a \textit{single} entity (such as {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace) to protect themselves from \textit{multiple} other unaccountable over-privileged entities that they would otherwise be forced to trust.
\begin{figure}[t]
\centering
\includegraphics[width=0.7\linewidth]{figures/new_architecture.pdf}
\caption{Component diagram of {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace (components labeled by $S$ are server-side and by $C$ are client-side)}
\label{fig:architecture}
\end{figure}
\section{Anatomizing Developers' Behavior}
\label{sec:developers}
After studying the users' privacy decisions, we now move to investigate the developers' landscape, building on the apps data that our store's users have installed. In total, we obtained data from 1440 registered users of our privacy store.
\subsection{Current Developer Behavior}
\label{sec:anatomDevCurrent}
The ``\textsc{drive\_apps\_readonly}'' permission requested by our app allowed us to get the list of apps previously installed by users, along with the information that Google Drive API gives about the apps\footnote{For details, we refer the reader to: \url{https://developers.google.com/drive/v2/reference/apps}}. We found 662 unique apps installed by users in our dataset. For each app, we obtained the following:
\textbf{i.} \textit{Access Level}:
which indicates whether the app had \textit{Partial Access} or \textit{Full Access} to the user's drive \textit{on authorization time}. Since an app can change the permissions it requests from future users, our dataset had instances of the same app installed with different access levels by different users. We denote such apps as having an access level of \textit{Both}.
\textbf{ii. } \textit{App Location}:
which indicates whether the app is (1) on \textit{Google Chrome Web Store}, (2) on Google's \textit{Other Web Stores} (namely the Add-ons Stores and the Google Apps Marketplace for enterprises), or (3) \textit{Outside Web Stores} of Google.
This categorization is inferred by following the \textit{productUrl} field present in the app information, which either leads to one of the stores or is absent.
Figure~\ref{fig:effectOfStore} shows how the apps in our dataset were distributed over the different locations and the number of apps requesting the different access levels.
From this figure, one can observe the following:
\begin{figure}[t!]
\centering
\includegraphics[width=0.7\linewidth]{figures/effectOfStore_embed.pdf}
\caption{Change of access level with app location}
\label{fig:effectOfStore}
\end{figure}
\tparagraph{Developers Changing Behavior}
The first surprising outcome from this dataset is that around 40\% of apps on Chrome Web Store (63 apps) had \textit{Both} as access level, signifying that a lot of developers have changed the requested permissions at least once. In order to check the current access levels of these apps, we reviewed them one-by-one. We discovered that 59 of these apps (i.e., 94\%) have changed from requesting \textit{Partial Access} in the past, to requesting \textit{Full Access} currently. Hence, we can deduce that when developers change the access level, \textit{there is a high probability that it is associated with getting more data instead of the other way round}. Highlighting this change of access level on installation time can further serve for more informed user decisions.
\tparagraph{Developer Deterrence through Official Stores}
Apps outside the Web Stores requested \textit{Full Access} almost twice as much as they requested \textit{Partial Access} (281 full vs. 155 partial). This was not the case in the Chrome Web Store, where we observe only a slightly higher number of apps with \textit{Full Access} (81 full vs. 76 partial - counting apps that fall under the \textit{Both} access level but which currently request \textit{Full Access}).
So we can see that developers with apps outside the Web Stores are more prone to asking for \textit{Full Access}. This can be explained by the conjecture that the store acts as a medium where the apps receive more exposure. Hence, \textit{developers there are likely to be under the pressure of being evaluated through reviews and ratings, and thus tend to avoid abusing the permissions}, while developers outside Web Stores are under no such pressure. Although \textit{Full Access} does not necessarily mean that apps are over-privileged, our APRs have actually shown that 84\% of apps that request \textit{Full Access} are over-privileged apps.
In the case of the Other Web Stores, the number of apps that requested \textit{Partial Access} is around thrice the number that requested \textit{Full Access} (16 full vs. 47 partial). This is mainly due to the fact that these Add-ons apps are generally expected to provide functionality for Google Docs (or other native Google file types), so deviating from this and requesting permissions for all Google Drive files will be easily detected by the community. Similarly, the community of enterprises, which is highly sensitive towards privacy will deter developers from requesting \textit{Full Access} in the Google Apps Marketplace.
\tparagraph{Deterring Developers in the Wild}
The majority of apps in our sample do not actually come from any Google Web Store (24\% from Google Chrome Store, 10\% from the Other Web Stores, and 66\% from outside the Web Stores). This is also the case for 75\% of the apps requesting full data access. These can be apps on other platforms, such as mobile platforms, for example, where there are other types of application stores. These apps can also be ones that are not present in any store but still have Google Drive integration. Hence, we can infer that improving the Chrome Web Store privacy indicators might not be a sufficient solution for deterring the majority of developers. \textit{There is a need for alternative solutions, focused on Google Drive permissions in specific, and independent of the various stores.}
\subsection{Potential Developer Misbehavior}
Although it is clear that full access to users' data can expose various far-reaching insights about the user, it is not completely apparent what seemingly benign permissions, such as metadata-only access can reveal about the user.
In the previous section, we have shown that the \textit{TopCollaborators} insight, which can be extracted just from the metadata, has resulted in an Acceptance Likelihood of $0.13$, which is around three times lower than what the current Google permission scheme (Baseline) attains.
Hence, users are remarkably deterred by seeing what they expose when they give access to their file metadata. Therefore, it is worthwhile to explore the potential of information leakage through metadata-only access.
In this section, we show that metadata-only access on its own can allow developers to gather deeper insights about the user's topics and concepts of interest. This calls for extending the \textit{FR} Insights with such information in order to better inform the user about the potential risk of giving unneeded access to metadata.
Towards that, we analyze and compare insights inferred from users' file metadata to what can be inferred from file contents (the data).
Upon user \textit{u} signing up to our app, the following operations are executed as part of the data analysis:
\indent \textbf{i. }For each analyzed file, the filename is processed by removing its extension and replacing punctuation marks by spaces.\\
\indent \textbf{iii. }The names of all the analyzed files are grouped into a comma-separated list $L_{FN}(u)$\\
\indent \textbf{iii. }Topic and concept analysis are applied to $L_{FN}(u)$. The service used for text analysis allowed us to extract topics in the form ``$a_1/a_2/ \ldots /a_n$'', representing a hierarchy among the labels (e.g., ``/law, government and politics/espionage and intelligence/surveillance'' or ``/finance/investing/venture capital'').
In this section, we differentiate between \textit{General Topics} where we only consider $a_1$, and \textit{Specific Topics}, where we consider $a_n$.
Accordingly, \textit{General Topics} would indicate user's interest in \textit{law, government and politics} or \textit{finance} for example while \textit{Specific Topics} could indicate the user's interest in \textit{surveillance} or \textit{venture capital}.
At the end of this step, we filter the results to restrict our analysis of metadata to a maximum of 3 \textit{General Topics}, 3 \textit{Specific Topics}, and 5 \textit{Concepts} for each user's list $L_{FN}(u)$.\\
\indent\textbf{iv. }From the user's files' contents, we extract the top 5 \textit{General Topics}, top 10 \textit{Specific Topics}, and top 20 \textit{Concepts}. These choices are motivated by the general observation that one's \textit{Concepts} of interest are usually more in number than the \textit{Specific} abstract topics one cares about, which are in turn more than the \textit{General Topics} of interest.
For each user $u$, we compared the list $D(u)$ of labels (i.e., concepts/topics) extracted from the files' contents with the list $M(u)$ of labels extracted from the list $L_{FN}(u)$ of filenames.
We selected precision as the evaluation metric as we are mainly interested in determining whether labels extracted from metadata serve as a good approximation of labels extracted from the data.
Inspired by the multi-label classification literature~\cite{Multilabel:2010}, we computed precision using the micro-averaging method, i.e., directly across all labels.
A label occurrence is considered as true positive if it belongs to $M(u) \cap D(u)$ and a false positive if it belongs to $M(u) \setminus D(u)$. $tp(l)$ is the number of true positives for a label $l$, and $fp(l)$ is that of false positives, both across all users. Let $LT$ also be the set of all labels found across user's data and metadata. The overall precision is thus given by the following equation:
\begin{equation}
P_{micro}=\frac{\sum_{l \in LT}^{N} tp(l)} {\sum_{l \in LT}^{N} (tp(l)+fp(l))}
\end{equation}
We used this method instead of macro-averaging (i.e., computing the precision per label and then taking the average) because we are interested in estimating the users' interests more than the ability to predict each and every label.
For this experiment, we only considered who signed in to our app and had at least 10 textual files with associated concepts/topics. Hence, our sample contained 200 users.
Interestingly, the results for \textit{General Topics} indicate that 0.78 of the metadata labels across users match with their top 5 topics of interest.
In the case of \textit{Specific Topics}, on average, nearly two of the three extracted metadata labels also appear in the top 10 \textit{Specific Topics} extracted from data ($P_{micro}=0.61$).
Finally, the fraction of metadata \textit{Concepts} that also appear in the data is around one-third ($P_{micro}=0.31$).
However, this does not necessarily imply that the other two-thirds of concepts appearing in the metadata are not relevant to the user. In fact, we have noticed that a lot of these metadata labels are semantically similar to those in the data.
In sum, we have observed that metadata on its own can be considerably accurate in revealing part of users' interests.
It can be easily abused by sophisticated adversaries who conceal their misbehavior through only requesting seemingly benign permissions (for metadata access). Therefore, this calls for extending the \textit{FR} insights in the case of metadata-only access to match the developer's potential. For instance, \textit{SharedInterests}, which was shown earlier to convey inferences from content, can also be used as an insight based on the collaborators and on the potential mutual topics inferred exclusively from the files' metadata.
\section{Recommended Best Practices}
\label{sec:privacyProtection}
In addition to {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace, there are several steps that can serve to mitigate the potential of misbehavior in Google Drive and similar services. These solutions serve to help the user both before and after installing the apps.
\tparagraph{Fine-Grained APIs}
The availability of finer grained permissions (such as access for a specific file type) evidently reduces the amount of data in the hands of the developer and is in line with the principle of least privilege. One disadvantage of such detailed permissions is that they become more difficult for users to comprehend in a short amount of time.
However, providing developers with the means to request such fine-grained controls should not necessarily result in a more complicated interface. This can be achieved via multi-layered interfaces~\cite{schaub2015design}. For example, instead of the app indicating that it needs to ``View the files on your Google Drive'', it can indicate that it needs to ``View files of specific types in your Google Drive''. Users that are interested in knowing these file types can then click on an additional button (such as the info button \circled{i} in the current interface of Figure~\ref{fig:baseline_pdf_converter}).
\tparagraph{Transparency Dashboard}
A post-installation technique which can potentially deter developers from actually abusing the users' data is for the cloud platform to provide what we call ``Transparency Dashboards''. These dashboards allow the user to see which files have been downloaded by each 3rd party app and when such operations took place. Such a monitoring solution for all apps can only be achieved by the platform itself.
\tparagraph{Insights Based on Used Data}
Unlike external solutions (e.g., ours) that can only determine what data can be \textit{potentially} accessed, the cloud platform can provide users with insights based on the data that developers have \textit{previously} downloaded. Such an interface will help users better pinpoint adversarial apps that needlessly retrieve files outside the scope of their functionality.
\tparagraph{A Privacy Preserving API Layer}
It is not uncommon nowadays to find APIs that work as an additional layer on top of one or more existing cloud APIs (e.g., Cloud Elements Documents Hub).
Hence, one solution to build a privacy-preserving API is to create it as a layer on top of one or multiple existing platforms' APIs. This new API can provide finer grained access control, allow permissions reviews from the community, and implement transparency dashboards. By building this layer on top of existing cloud APIs that already offer various services, one can circumvent the problem of attracting developers who might otherwise be loathe to using a solution that only serves to protect privacy.
\section{Related Work}
To our knowledge, this is the first work that studies the problem of user privacy in the context of 3rd party apps on top of cloud storage providers. Other works have previously studied the problem of direct information sharing to providers themselves (e.g.,~\cite{harkous2014c3p}).
\subsection{Privacy in Other App Ecosystems}
In the case of other ecosystems, there are related works that have studied the current state of privacy notices (e.g. ~\cite{Chia:2012, Huber:2013, Felt:2011, Pandita:2013}).
For instance, Chia et al., conducted a large-scale analysis of Facebook apps, Chrome extensions, and Android apps to study the effectiveness of user-consent permissions systems~\cite{Chia:2012}. They observed that the community ratings are not reliable indicators of app privacy in these ecosystems and showed evidence of attempts at misleading users into granting permissions via free apps or apps with mature content.
Huber et al., developed \textit{AppInspect}, a framework for automating the detection of malpractices in 3rd party apps within Facebook's ecosystem and used network traffic analysis to spot web trackers and identify leaks of sensitive information to other third parties~\cite{Huber:2013}.
The case of 3rd party apps in Google Drive differs from these platforms in that it is not possible to perform large-scale analysis, firstly due to the absence of a standard application format and secondly due to the difficulty of automatically finding the triggering button for permission requests in different apps.
Aside from the above, client-side traffic analysis is not sufficient to detect all cloud data leaks as the apps can send data to third parties after it arrives at the server side, to which outsiders do not have access.
\subsection{Improving Current Privacy Notices}
A few works have recently suggested improvements to the existing permissions schemes, with a special focus on the case of Google Play Store.
Kelly et al., argued that the privacy information should be a part of the app decision-making process and should not be left till after the user makes her decision~\cite{Kelley:2013}. Hence, they appended a list of ``Privacy Facts'' to the app description screen, textually indicating that the app, for example, collects contacts, location, photos, credit card details, etc., and found that it assisted users in choosing apps that request fewer permissions.
Harbach et al., proposed to integrate examples from user's data in the permissions request screen to expose the data apps can get access to~\cite{Harbach:2014}. This involved showing random pictures, call logs, location, and contacts from user's data that correspond to each permission.
Another related work in the context of Facebook is that by Wang et al.~\cite{Wang2013}, who introduced the ``Privacy Nudges'' technique to aid users while posting statuses to Facebook through showing random profile pictures of friends who can see the post, introducing a time interval before the actual post is sent, or showing the post sentiment.
In this work, we go further, and we show that well-crafted visuals showing far-reaching insights extracted from users' data can be more effective than randomly selected data. We also show through pairwise comparisons among the insights themselves that the choice of the displayed insight highly affects the interface's effectiveness.
It is also worth mentioning that, in our experiments the number of users who were involved with their personal accounts in the experiment was more than five times the number of users in~\cite{Harbach:2014} and~\cite{Kelley:2013}. Furthermore, we also provide a readily available solution for the public in the form of a privacy-focused app store.
Moreover, our work is in line with the best practices recommended by the recent work of Schaub et al., who developed a design space for privacy notices to assist researchers in increasing the impact of their schemes~\cite{schaub2015design}. For instance, we implemented the multi-layered notice concept by showing data of textual and visual modalities. We also developed various visuals to ensure that the permissions dialog is \textit{polymorphic}, which was also shown recently to have an effect on reducing the habituation effect in the user's brain~\cite{Anderson:2015}.
Personalizing warning notices, as we do in this paper, has been studied before in the context of LED signs~\cite{Wogalter1994233} and was shown to significantly increase compliance compared to impersonal signs.
\section{Conclusion and Future Work}
In this paper, we characterized the various factors that have an impact on user privacy in the ecosystem of 3rd party apps for the cloud. We considered Google Drive as an example case study and comprehensively anatomized the ecosystem from the viewpoint of users, developers, and the cloud provider. For users, we carefully devised a set of experiments and tested existing and novel risk communication models to analyze the factors that influence users' decisions in app installation. Our results provide interesting insights into how user privacy can be improved and how CSPs can develop better risk indicators. We also presented a privacy aware store for cloud apps, which already has over 1440 registered users. From our store users and people who took part in our experiments, we had the unique and unprecedented opportunity to first-hand study real users cloud data. Based on this data, we were able to characterize the current behavior of 3rd party app developers and also point out avenues for developer misbehavior. Finally, based on our analysis, we provided several suggestions for CSPs that can help in safeguarding users' privacy and protecting their data from needless leakage and exploitation.
In the future, we aim to build on {{\href{https://privyseal.epfl.ch}{\mbox{\tt{PrivySeal}}}}}\xspace and develop a recommendation system that suggests apps of similar functionality but superior privacy.
We are also in the process of integrating Personalized Insights in the scenario of user-to-user sharing privacy in \textit{StackSync}, which is an open source cloud platform with a significant number of users.
Finally, it would be interesting to study how our findings on the best risk indicators generalize to other ecosystems, such as Android or iOS.
\section*{Acknowledgments}
We would like to thank the anonymous reviewers for their insightful comments. The research leading to these results has received funding from the EU in the context of the project \textit{CloudSpaces}: Open Service Platform for the Next Generation of Personal clouds (FP7-317555).
|
\section{Introduction}
\label{sec: intro}
Landau-Ginzburg symmetry breaking formalism is one of the fundamental building blocks for conventional condensed matter physics\cite{landau1937theory,ginzburg1950theory,landau1958lifshitz}. Historically, it applies successfully to understanding magnets, BCS superconductor, and etc.
Within the Landau-Ginzburg paradigm, phases are characterized by their global symmetries and which of those symmetries are spontaneously broken. The result is that phases can be paramagnetic with short ranged correlations, or have long range correlations associated with spontaneously broken symmetries. More recently, it has been found that this characterization is a) too coarse, and b) needs to be extended. First, not all paramagnetic phases with a certain global symmetry group are identical. Indeed, there exist paramagnetic symmetry protected topological (SPT) phases which have a global symmetry which is not broken spontaneously \cite{chen2011classification,chen2010local,levin2012braiding,burnell2014exactly,chen2011complete,chen2011two,chen2012chiral,chen2012symmetry,chen2013symmetry,liu2011symmetry,gu2015multi,wang2015bosonic,wang2015field,wang2015non,ye2013symmetry,ye2015vortex,pollmann2012symmetry,fidkowski2010effects,gu2009tensor,wen2013classifying,wen2013topological,lu2012theory,haldane1983nonlinear,affleck1987critical,haldane1983continuum, vishwanath2013physics,fidkowski2011topological,pollmann2010entanglement}, but which nevertheless cannot be adiabatically connected to one another in the presence of the protecting global symmetry. Secondly, the Landau-Ginzburg paradigm does not account for topologically ordered states\cite{WenNiu90,Wen90top2}, which are not characterized by the spontaneous breaking of global symmetries and the long distance correlations of local order parameters.
The mentioned SPTs and topological phases both have a connection to topological quantum field theories (TQFTs). The low-energy long-distance behavior of topological phases is described by TQFTs. Bosonic SPTs on the other hand, can be characterized by gauging their protecting global symmetry as done in Ref.~\onlinecite{levin2012braiding}. After gauging the global symmetry, any SPT phase becomes topologically ordered, and it turns out that distinct SPT phases become distinct topological orders when gauged. For example, there are two different SPTs with a $\mathbb{Z}_2$ global symmetry in (2+1)D. After gauging the $\mathbb{Z}_2$ symmetry of these two SPTs, one becomes deconfined $\mathbb{Z}_2$ gauge theory and the other becomes the distinct double semion topological order\cite{levin2012braiding}. The classification of a large class of SPTs then simply becomes the classification of different possible topological orders or TQFTs with gauge group $G$. When $G$ is unitary and finite, such TQFTs are called ``Dijkgraaf-Witten" (DW) models, which are classified by the cohomology group $H^{d+1}(G,U(1))$\cite{dijkgraaf1990topological}, where $G$ is the given on-site global symmetry and $d$ is the spatial dimension. Hence, bosonic SPTs with finite on-site global unitary symmetry group $G$ are classified by an element of $H^{d+1}(G,U(1))$ (called a ``cocycle"), and are in 1-to-1 correspondence with DW models. In addition to their connection to bosonic SPTs, DW models are themselves interesting examples of topological orders. Therefore by studying DW models we learn something about SPTs, and gain insight into a wide range of possible topological orders.
We are thus motivated to study DW field theories, and in this work we focus on (2+1)D. We further focus on DW theories with finite abelian gauge groups. Topological field theories with non-semisimple gauge groups were studied in Ref~\onlinecite{Ferrari2003,Ferrari2007,Ferrari2015}. On the one hand, the DW theories were originally introduced as field theories with topological terms directly related to the cocycle in question\cite{dijkgraaf1990topological}. Subsequently, the same cocycle data was found to encode an algebraic structure called a quasi-quantum double\cite{propitius1995topological}, which it was proposed should describe the algebra of anyon excitations in the DW model. The goal of the present work is to provide a clearer bridge between the field theory for the DW model, and the algebraic theory of its anyonic excitations. We do this by explicitly constructing line operators in the DW field theory, and calculating their braiding and fusion rules.
Some DW theories (called type I and type II\cite{propitius1995topological}) can be thought of as continuum `K-matrix' Chern-Simons theories, and their line operators are already well understood\cite{lu2012theory,cheng2014topological}. However, the line operators in more exotic `type III' DW field theories have remained elusive (Field theories of type III DW model were also discussed in \onlinecite{PhysRevLett.112.231602,kapustin2014anomalies,wang2015bosonic,gu2015multi}). In this work we focus on the very simplest such theory -- the type III twisted DW model with gauge group $\mathbb{Z}_2^{\otimes 3}$. We construct all line operators (Wilson lines, flux insertion lines and their composites). Instead of the naive 64 line operators (8 Wilson operators, 8 flux insertion operators and their composites), we find that the number of distinct line operators is only 22, reproducing an algebraic result in Ref.~\onlinecite{propitius1995topological}. Moreover, we can compute the correlation functions and fusion rules for these operators and confirm that the type III twisted $\mathbb{Z}_2^{\otimes 3}$ theory is a non-abelian topologically ordered phase. We also explain how these results extend to more general abelian gauge groups. Our results should not be considered as completely mathematically rigorous -- we use a continuum field theory formalism on the understanding that at certain key points in the calculation the lattice regularization needs to be considered carefully. In this manner, our approach is of a similar level of rigor to other continuum approaches used to understand similar models\cite{Ye:2015eba,kapustin2014coupling,gu2015multi}. The utility of our formalism is that it readily allows us to derive a number of non-trivial results in a reasonably intuitive manner, without the need for understanding some the more abstract algebraic machinery behind the existing group cohomology results.
The rest of the paper is organized as follows: In Sec.~\ref{sec: 2+1D}, we briefly summarize the $K$-matrix formulation of type I,II cocycles. Then we attack the problem of type III cocycles. We begin by constructing all of the line operators for the type III twisted $\mathbb{Z}_2^{\otimes 3}$ theory in (2+1)D and work out their correlation functions, and fusion rules. This theory turns out to be a non-abelian topologically ordered state. Sec.~\ref{sec. ZN examples} further generalizes these results to the type III twisted $\mathbb{Z}_N^{\otimes3}$ theory. We also provide three appendices for further details: App.~\ref{app. coefficient} explains how we fix the coefficients in the Lagrangians; App.~\ref{app. path integral} explains how to calculate the path integrals with a lattice regularization; App.~\ref{app.sec. gauge} varifies the gauge invariance of flux insertion operator $V_{n_1n_2n_3}$ on lattice explicitly; App.~\ref{app. ModMat}, we provide the modular $S$, $T$ matrices for the type III twisted $\mathbb{Z}_2^{\otimes 3}$ DW theory; App.~\ref{app. QD} shows quantum double calculations for type III $\mathbb{Z}_N^{\otimes 3}$ theory used to verify the field theoretic results from the main text.
\section{DW models in continuum formalism }
\label{sec: 2+1D}
DW theories were first formulated as lattice gauge theories\cite{dijkgraaf1990topological}. Consider a (2+1)D theory with abelian gauge group $G$. The DW theory action is encoded by some 3-cocycle $\omega: G\times G \times G \mapsto \text{U}(1)$. The DW action is obtained by performing a simplicial decomposition of the manifold in question and orders the vertices, to write down a partition function weight
\begin{equation}
\prod_t \omega(\vec{A}^{01}_{t},\vec{A}^{12}_{t}, \vec{A}^{23}_{t}),
\end{equation}
where $\vec{A}^{e}_{t}$ is a $G$-valued flat gauge field living on an edge $e$; $t$ are 3-simplices obtained by triangularizing the spacetime manifold; and $01$, $12$, $23$ are specific edges in $t$ determined by the ordering on the simplicial decomposition\footnote{Gauge invariance is ensured by the so-called cocycle condition on $\omega$ -- see Ref.~\onlinecite{dijkgraaf1990topological}.}. Note that in this construction, $\vec{A}$ is assumed to be flat; one goal in this work is to extend the construction above so as to relax this constraint on $\vec{A}$ whilst maintaining gauge invariance. This in turn allows us to examine the full spectrum of the DW model, and explicitly construct all the line operators in the theory. In addition, the lattice actions considered in Ref.~\onlinecite{dijkgraaf1990topological} are difficult to work with. A second goal of this work is to formulate in detail a more convenient and transparent continuum version of these field theories much like those in Ref.~\onlinecite{kapustin2014coupling} -- it will turn out that the explicit regularization of the models is for many purposes unimportant.
Our discussion is organized as follows: in \secref{subsec: type I II} we briefly discuss the type I and type II cocycles which leads to abelian topological phases; in \secref{subsec: continuum} we construct the continuum action for type III DW models; in Sec. \ref{subsec: 22 line operators}, we argue that there are only 22 distinct line operators (in agreement with Ref.~\onlinecite{propitius1995topological}), instead of the naive 64 operators one expects in a $\mathbb{Z}^{\otimes 3}_2$ gauge theory. In Sec. \ref{subsec: fusion rules 2+1} and \ref{subsec: correlation 2+1}, we calculate correlation functions and fusion rules of these line operators. Further generalizations to $\mathbb{Z}_N^{\otimes 3}$ can be found in Sec.~\ref{sec. ZN examples}.
\subsection{Type I and Type II Cocycles}\label{subsec: type I II}
Before we delve into the field theory for type III cocycle, let's briefly comment on the type I and type II cocycles. The cocycles of abelian discrete groups $\prod_i \mathbb{Z}_{k_i}$ have been categorized into three types\cite{propitius1995topological,wang2015field}. Type III cocycle is the focus of this paper and will be explained in the following texts. Here we only briefly discuss continuum field theories corresponding to the type I and type II cocycles, as they are just special cases of abelian $K$-matrix theories.
As an example, for a discrete group in the form of $(\mathbb{Z}_{N})^{\otimes L}$, the type I and type II cocyles can be written uniformly as\cite{propitius1995topological}
\begin{equation}\label{eq. type I II}
\omega^{ij}(A,B,C) = \exp\left( \frac{2\pi\mathrm{i} M^{ij}}{N^2} a^{i}(b^{j}+c^{j}-[b^j+c^j])\right),
\end{equation}
where $A, B, C \in \mathbb{Z}_{N}^{\otimes L}$; $a^i,b^i,c^i \in\{0,1\ldots,N-1\}$ for $i=1,2,\ldots L$ label the $i$-th component of these group elements in the $L$ copies of $\mathbb{Z}_{N}$ respectively; $M^{ij}$ are integers valued in $\{0,1,\ldots,N-1\}$. The bracket notation is defined by $[x]:= x \mod{N}$ with $[x]\in \{0,1,\ldots,N-1\}$
The continuum field theories corresponding to the type I and II cocycles Eq.~\eqref{eq. type I II}, have action $\frac{1}{4\pi}\int K^{i j} a^i d a^j$ where $i\in1,\ldots ,2L$ and each $a^i$ is a compact $\text{U}(1)$ connection 1-form. The corresponding $K$-matrix is
\begin{equation}
\left(
\begin{matrix}
0 & N \mathbb{I}_{L} \\
N \mathbb{I}_{L} & M+M^T
\end{matrix}
\right),
\end{equation}
where $\mathbb{I}_{L}$ is $L$-by-$L$ identity matrix, $M$ is an $L$-by-$L$ integer matrix whose elements are just $M^{ij}$ in Eq.~\eqref{eq. type I II}. The type I and type II theories hence only produce abelian topological order and all line operators and their statistics/correlations are well known\footnote{See Ref.~\onlinecite{Wen04,WenReview2016} for review articles, and references therein.}.
\subsection{Type III Cocyles}\label{subsec: continuum}
Having summarized the story for type I and II cocycles, we describe the so-called type III twisted DW theory. These are characterized by a 3-cocycle of form
\begin{equation}\label{eq:3cocycle}
\omega(A,B,C) = e^{2 \pi i p a^{1} b^{2} c^{3}/k_1 k_2 k_3},
\end{equation}
where $A, B, C \in \mathbb{Z}_{k_1}\times\mathbb{Z}_{k_2} \times\mathbb{Z}_{k_3} $ and $a^i,b^i,c^i =0,1\ldots,k_i-1$ for $i=1,2,3$ label the components of these three group elements in the three copies of $\mathbb{Z}_{k_1},\mathbb{Z}_{k_2},\mathbb{Z}_{k_3}$ respectively. Here,
\begin{equation}
p=n k_1k_2k_3/\gcd(k_1,k_2,k_3),
\end{equation}
where $n \in \mathbb{Z}_{\gcd(k_1,k_2,k_3)}$ labels the distinct possible choices of cocycle. Using the above prescription, the DW models are rigorously formulated on the lattice. However, many of the known abelian examples of these theories are more conveniently formulated in the continuum. For instance Refs.~\onlinecite{tiwari2016wilson,kapustin2014coupling,wang2015field,ye2013symmetry,gu2015multi} characterize certain abelian DW topological orders in terms of continuum toy models. In this spirit, we start by writing down the most naive interpretation of the 3-cocycle \eqnref{eq:3cocycle} in the continuum and examine under which conditions it is gauge invariant. The Lagrangian for $\mathbb{Z}_{k_1}\times \mathbb{Z}_{k_2} \times \mathbb{Z}_{k_3}$ theory is:
\begin{equation}\label{eq: lag 2+1}
\mathcal{L}= \frac{k_i}{2\pi} b_i \wedge dA_i + \frac{p\epsilon^{ijk}}{6(2\pi)^2}A_i \wedge A_j \wedge A_k,
\end{equation}
where the repeated indices imply summation. We now clarify the above notation: $A_{i=1,2,3}$ are the components of $\vec{A}$ in $\mathbb{Z}_{k_1}\times \mathbb{Z}_{k_2} \times \mathbb{Z}_{k_3}$, and where $i,j,k$ are summed over $\{1,2,3\}$. The first term is a $bF$ term which enforce the flatness condition of $A_i$ fields in the partition function, and the second term is the type III twist term. For $G=\mathbb{Z}_2^{\otimes 3}$ we have $k_i=2~(i=1,2,3)$. In this case there are two possible choices for $p$: $p=0$ corresponds to plain $\mathbb{Z}_2^{\otimes 3}$ gauge theory (3 copies of $\mathbb{Z}_2$ model), while $p=4$ we refer to as `twisted' $\mathbb{Z}_2^{\otimes 3}$ gauge theory. Following previous work on these theories\cite{wang2015field}, we detail how to fix the possible values of the coefficients of the twist terms in \eqnref{eq: lag 2+1} in \appref{app. coefficient}.
As written, the action is invariant under transformations
\begin{equation}\label{eq: gauge transf 2+1}
\begin{split}
& b_i\rightarrow b_i+d\beta_i+\frac{p\epsilon^{ijk}}{2\pi k_i}(A_j\alpha_k-\frac{1}{2}\alpha_jd\alpha_k) \\
& A_i\rightarrow A_i+d\alpha_i,\;i=1,2,3\;.
\end{split}
\end{equation}
where $\alpha,\beta$ is a scalar field and we have omitted wedge products for brevity\footnote{The noncommutative gauge transformations have indicated that the theory is actually non-Abelian. We thank E. Witten for pointing it out. Also see Ref.~\onlinecite{gu2015multi}.}. In addition both gauge fields are presumed to be compact insofar as
\begin{align}
b_i &\equiv b_i+ 2\pi \nonumber\\
A_i &\equiv A_i +2\pi\label{eq: comp transf 2+1}.
\end{align}
Here $b_i$ and $A_i$ are understood as the value of gauge field on a bond of spacetime lattice.
As mentioned before, in the partition function,
\begin{eqnarray}
\mathcal{Z}:=\int D[A_i]D[b_i] \exp(\mathrm{i}\int\mathcal{L}) \;,
\end{eqnarray}
$b_i$ fields play the role of Lagrangian multipliers and enforce the flatness constraint on $A_i$. However, once $b$ sources are inserted in the path integral, $A_i$ fields are no longer flat. To see this, note that in the presence of $b$ sources, the path integral takes the form
\begin{equation}
\mathcal{Z}[\gamma]=\int D[A_i]D[b_i] \exp\db{\mathrm{i}\int\mathcal{L}+\mathrm{i}\oint_{\gamma} b_i \ldots}.
\end{equation}
Once the $b_i$ fields are integrated out, $dA_i$ is enforced to be nonzero on $\gamma$ i.e., $\textstyle{dA_i=\frac{2\pi}{k_i}\gamma^{(2)}}$ where $\gamma^{(2)}$ is the 2-form Hodge dual to the contour $\gamma$. However, a single term, $\exp\left(\mathrm{i}\oint_{\gamma} b_i\right)$, is not gauge invariant. Hence, it cannot be a valid operator for the twisted type III $\mathbb{Z}_2^{\otimes3}$ theory. We discuss the all valid line operators in the following section. And we coin the operators involving $\exp\left(\mathrm{i}\oint_{\gamma} b_i\right)$ ``flux insertion operators" for the following text.
\subsection{Line Operators}
\label{subsec: line operators 2+1}
In this section, we construct all the Wilson operators and flux insertion operators on a given loop $\gamma$, for the type III twisted $\mathbb{Z}_2^{\otimes 3}$ field theory mentioned above, Eq.~\eqref{eq: lag 2+1}. We adopt the notation $U_{pqr}$ for Wilson operators, and $V_{pqr}$ for flux insertion operators, where $p,q,r=0,1$. We will see that when $p,q,r\equiv 0 \mod 2 $, the resulting operators are trivial in the sense that they have trivial correlations with other operators.
An essential requirement of constructing these loop operators is that they are invariant under gauge transformation Eq.~\eqref{eq: gauge transf 2+1}. Moreover, the line operators should also be invariant under $A_i \mapsto A_i + 2\pi$ and $b_i \mapsto b_i + 2\pi$, because the gauge fields are assumed to be compact with $2\pi$ periodicity. Following the gauge invariance principle, the Wilson operators can be written as
\begin{equation}
U_{pqr}(\gamma) = \exp\Big(\mathrm{i}\oint_{\gamma} pA_1+qA_2+rA_3\Big),~p,q,r=0,1
\end{equation}
which are gauge invariant under gauge transformations Eq.~\eqref{eq: gauge transf 2+1}. Using the form of the Lagrangian \eqnref{eq: lag 2+1}, the compactness condition on $b_i$ in \eqnref{eq: comp transf 2+1} breaks $A_i$ down to $\mathbb{Z}_2$, so the operator $U_{pqr}$ only depends on the values of $p,q,r$ modulo $2$.
Flux insertion operators are more complicated, since a single term $\exp(\mathrm{i}\oint_{\gamma} b_i),~i=1,2,3$ is not gauge invariant under transformation \eqnref{eq: gauge transf 2+1}. One can construct the flux insertion operators by introducing auxiliary fields $\phi_i$ and $\lambda_i$ living on the loop $\gamma$, Indeed, we find that the following operator defined via a path integral is gauge invariant
\begin{equation}\label{eq.100}
\begin{split}
V_{100}(\gamma)=&\frac{1}{\mathcal{N}} \int D[\phi_2]D[\phi_3]D[\lambda_2]D[\lambda_3]\\ \exp\Bigg(\mathrm{i}\oint_{\gamma}
& b_1 + \sum_{ij=2}^{3} \frac{\epsilon^{1ij}}{\pi} (\frac{1}{2} \phi_i d \phi_j + (d\phi_i - A_i) \lambda_j )\Bigg)\punc{,}
\end{split}
\end{equation}
where $i$ and $j$ are actually summed over $\{2,3\}$ because of $\epsilon^{1ij}$; $\mathcal{N}$ is a normalization factor which we determined later in \secref{subsec: fusion rules 2+1} by insisting on a consistent set of fusion rules for the flux insertion operators. The operator of Eq.~\eqref{eq.100} is gauge invariant under gauge transformation Eq.~\eqref{eq: gauge transf 2+1} with additional transformations, $\phi_i \mapsto \phi_i + \alpha_i$ and $\lambda_i \mapsto \lambda_i + \alpha_i$.
The auxiliary fields in Eq.~\eqref{eq.100}, $\phi_2$, $\phi_3$, $\lambda_2$ and $\lambda_3$ can be integrated out exactly -- the details of the calculation can be found in the App.~\ref{app. path integral}. The result is conveniently expressed as
\begin{equation}\label{eq.100 close form}
V_{100}=2 \exp\Big(\mathrm{i}\oint_{\gamma} b_1 + \sum_{i,j=1}^{3} \frac{\epsilon^{1ij}}{2\pi} \omega_i d \omega_j\Big) \delta(\bar{\omega}_2|_{\gamma})\delta(\bar{\omega}_3|_{\gamma}),
\end{equation}
where $\omega_i$ is the holonomy function for $A_i$ which is defined explicitly on the loop $\gamma$ as
\be\label{eq:wi}
\omega_i (x):= \int_{\gamma, x_0}^x A_i\punc{,}
\ee
while $\bar{\omega}_i:=\oint_{\gamma} A_i$, $i=1,2,3$. The choice of origin of integration $x_0$ is arbitrary. The $\delta$ functions appearing in \eqnref{eq.100 close form} project onto configurations for which the $A_2,A_3$ fluxes threading $\gamma$ are zero. They are not the usual $\delta$ functions encountered in the continuum -- rather they are defined to be a projector to the trivial holonomy state: $\delta(\bar{\omega}_i |_\gamma):=\frac{1}{2} (1+ \exp(\mathrm{i} \bar{\omega}_i |_\gamma))$. By trivial holonomy, we mean $ \bar{\omega}_i=2\pi n$ for any $n$, where $n$ is an integer. We will come back to the overall factor of $2$ in \eqnref{eq.100 close form} when fusion rules are discussed in Sec. \ref{subsec: fusion rules 2+1}. Note that the expression $\oint_{\gamma} \omega_i d \omega_j$ resulting from integrating out the scalar fields is not local in terms of the gauge fields $A_i$. The other flux insertion operators have similar expressions
\begin{widetext}
\begin{equation}\label{eq.010}
\begin{split}
V_{010}(\gamma)=&\frac{1}{\mathcal{N}} \int D[\phi_1]D[\phi_3]D[\lambda_1]D[\lambda_3] \exp\Big(\mathrm{i}\oint_{\gamma} b_2 + \sum_{i,j=1}^{3} \frac{\epsilon^{2ij}}{\pi} (\frac{1}{2} \phi_i d \phi_j + (d\phi_i - A_i) \lambda_j )\Big) \\
=& 2 \exp\Big(\mathrm{i}\oint_{\gamma} b_2 + \sum_{i,j=1}^{3} \frac{\epsilon^{2ij}}{2\pi} \omega_i d \omega_j\Big) \delta(\bar{\omega}_1|_{\gamma})\delta(\bar{\omega}_3|_{\gamma}) \;,
\end{split}
\end{equation}
\begin{equation}\label{eq.001}
\begin{split}
V_{001}(\gamma)=&\frac{1}{\mathcal{N}} \int D[\phi_1]D[\phi_2]D[\lambda_1]D[\lambda_2] \exp\Big(\mathrm{i}\oint_{\gamma} b_3 + \sum_{i,j=1}^{3} \frac{\epsilon^{3ij}}{\pi} (\frac{1}{2} \phi_i d \phi_j + (d\phi_i - A_i) \lambda_j )\Big) \\
=& 2 \exp\Big(\mathrm{i}\oint_{\gamma} b_3 + \sum_{i,j=1}^{3} \frac{\epsilon^{3ij}}{2\pi} \omega_i d \omega_j\Big) \delta(\bar{\omega}_1|_{\gamma})\delta(\bar{\omega}_2|_{\gamma}).
\end{split}
\end{equation}
\end{widetext}
Before moving on to the remaining flux insertion operators, let us further motivate the path integral form of the operators $V_{100}$, Eq.~\eqref{eq.100}, and similarly for $V_{010}$ Eq.~\eqref{eq.010} and $V_{001}$ Eq.~\eqref{eq.001}. Gauge invariance strongly constrains the forms of these operators. If we write down an operator of the form
\begin{equation}
V_{100} = \exp(\mathrm{i}\oint_\gamma b_1) \mathfrak{g}(A),
\end{equation}
and insist on gauge invariance, we find that the functional $\mathfrak{g}$ is necessarily a non-local functional of $A$ -- it must have something like the $A_i$ dependence of \eqnref{eq.100 close form}, involving constraints $\bar{\omega}_2=\bar{\omega}_3=0$, and phase terms like $\int_\gamma \omega_i d \omega_j$. In order to realize the operator Eq.~\eqref{eq.100 close form} in a local form, one possible solution is to introduce auxilliary fields into the path integral living on $\gamma$ which once integrated out, realize \eqnref{eq.100 close form}. This is the approach which led to \eqnref{eq.100}.
While the introduction of these auxiliary fields may seem ad hoc, there is a neat underlying physical interpretation for this procedure. To understand this interpretation, we briefly return to the quantum double theory approach of Ref.~\onlinecite{propitius1995topological}. Within that algebraic framework, the flux quasi-particles for the $\mathbb{Z}_2\times \mathbb{Z}_2 \times \mathbb{Z}_2$ theory considered here carry a projective representation. In other words, the fluxes carry an internal degree of freedom which transforms projectively under the gauge group. The flux insertion operators we consider insert precisely such fluxes, so should also carry some such internal degree of freedom. And indeed they do: One way of interpreting the $\phi,\lambda$ fields is that they are matter fields which on net transform projectively under the gauge group.
To further substantiate this idea, note that in the study of SPT phases, the boundary of a (1+1)D SPT bulk transforms projectively under the bulk symmetry\cite{pollmann2012symmetry,chen2011classification,fidkowski2011topological}. This statement, curiously enough, is helpful in interpreting our line operators. Suppose we have an abstract form of $V_{100}$ as follows:
\begin{equation}
V_{100} = \int D[\phi] D[\lambda] \ldots \exp(\mathrm{i}\oint_\gamma b_1 + f(A,\phi,\lambda,\ldots)),
\end{equation}
where $f$ is a function of $A$ and auxiliary fields such as $\phi$, $\lambda$ etc. And we try to calculate the expectation value of $V_{100}$. We need:
\begin{widetext}
\begin{equation}\label{eq. DecoratedDomainWall}
\begin{split}
\langle V_{100} \rangle =& \int D[b_i] D[A_i] D[\phi] D[\lambda] \exp(\mathrm{i} \int \mathcal{L} + \mathrm{i} \oint_\gamma (b_1 + f(A,\phi,\lambda,\ldots ))) \\
=& \int D[A_i] D[\phi] D[\lambda] \exp(\mathrm{i} \int \frac{1}{\pi^2} \bar{A}_1\bar{A}_2\bar{A}_3 + \mathrm{i} \oint_\gamma f(\bar{A},\phi,\lambda,\ldots )) \\
=& \int D[A_i] D[\phi] D[\lambda] \exp(\mathrm{i} \int_{[\bar{A}_1]} \frac{1}{\pi} \bar{A}_2\bar{A}_3 + \mathrm{i} \oint_\gamma f(\bar{A},\phi,\lambda,\ldots )).
\end{split}
\end{equation}
\end{widetext}
We have omitted all the wedges ``$\wedge$" in the above, and will continue this convention in the following texts if without misunderstanding.
The second equality comes from integrating out all $b_i$ fields. In this case, $A_2$ and $A_3$ will be flat and thus be exact on a simple spacetime manifold, while $A_1$ will not. And these fields after integrating out $b_i$ fields are denoted as $\bar{A}_1$, $\bar{A}_2$ and $\bar{A}_3$. Note that the integral $\int \bar{A}_1 \bar{A}_2\bar{A}_3$ can be written as the integral over $A_1$ flux sheet $[\bar{A}_1]$ (see Ref.~\onlinecite{Ye:2015eba} for a similar discussion) whose boundary is $\partial[\bar{A}_1]=\gamma$:
\begin{eqnarray}
\int \bar{A}_1\bar{A}_2\bar{A}_3 = \int_{[\bar{A}_1]} \pi\bar{A}_2\bar{A}_3 \;.
\end{eqnarray}
where $\pi$ comes from the normalization. We still need the rest of the terms in the second equality of Eq.~\eqref{eq. DecoratedDomainWall} to be gauge invariant. Then the gauge anomalies of two integrals $\int_{[\bar{A}_1]} \frac{1}{\pi} \bar{A}_2\bar{A}_3$ and $\oint_\gamma f(\bar{A},\phi,\lambda,\ldots)$ need to cancel each other. Notice that $\int_{[\bar{A}_1]} \bar{A}_2\bar{A}_3$ is just the SPT Lagrangian on the manifold $[\bar{A}_1]$ with symmetry $\mathbb{Z}_2^{\otimes2}$, when $\bar{A}_2$ and $\bar{A}_3$ are both exact. Hence $f$ should look like the boundary action of an SPT, insofar as it should transform to compensate for the gauge anomaly from the bulk action. Indeed, the particular $f$ chosen in \eqnref{eq.100} looks very much like the boundary action of the SPT in Ref.~\onlinecite{kapustin2014coupling}.
We can similarly write down the direct generalizations of flux insertion operators from Eq.~\eqref{eq.100}, which insert two types of fluxes and three types of flues.
\begin{widetext}
\begin{equation}\label{eq.110}
\begin{split}
V_{110}=&\frac{1}{\mathcal{N}'}\int D[\phi_i]D[\lambda_i] \exp\Big(\mathrm{i}\oint_{\gamma} b_1 + b_2 + \sum_{i=1,2}\sum_{j,k=1}^{3}\frac{\epsilon^{ijk}}{\pi} (\frac{1}{2}\phi_j d \phi_k + (d\phi_j-A_j)\lambda_k)\Big) \\
=& 2 \exp\Big(\mathrm{i}\oint_{\gamma} b_1 + b_2 + \sum_{i=1,2}\sum_{j,k=1}^{3} \frac{\epsilon^{ijk}}{2\pi}\omega_j
d\omega_k\Big) \delta(\bar{\omega}_1|_{\gamma}-\bar{\omega}_2|_{\gamma})\delta(\bar{\omega}_3|_{\gamma}),
\end{split}
\end{equation}
\begin{equation}\label{eq.111}
\begin{split}
V_{111}=&\frac{1}{\mathcal{N}''}\int D[\phi_i]D[\lambda_i] \exp\Big(\mathrm{i}\oint_{\gamma} b_1 + b_2 + b_3 + \sum_{i,j,k=1,2,3}\frac{\epsilon^{ijk}}{\pi} (\frac{1}{2}\phi_j d \phi_k + (d\phi_j-A_j)\lambda_k)\Big) \\
=& 2 \exp\Big(\mathrm{i}\oint_{\gamma} b_1 + b_2 + b_3 + \sum_{i,j,k=1,2,3} \frac{\epsilon^{ijk}}{2\pi}\omega_j
d\omega_k\Big) \delta(\bar{\omega}_1|_{\gamma}-\bar{\omega}_2|_{\gamma})\delta(\bar{\omega}_2|_{\gamma}-\bar{\omega}_3|_{\gamma}).
\end{split}
\end{equation}
\end{widetext}
The notations are the same as in Eq.~\eqref{eq.100} and \eqref{eq.100 close form}. The second equalities of both the above equations follow by integrating out all $\phi_i$ and $\lambda_i$ fields. Start by integrateing out $\lambda_1$. Then we have a constraint $d\phi_2 - d\phi_3 = A_2 - A_3$, the solution of which can be written as $\phi_2=\phi_3 + \omega_2 - \omega_3 + C_2$ where $C_2$ is a constant. Similarly integrating out $\lambda_2$ yields $\phi_1=\phi_3 + \omega_1 - \omega_3 + C_1$. Lastly, the constraint obtained by integrating out $\lambda_3$ is automatically satisfied. Plugging these two solutions back in produces the second equality, where $C_2$ and $C_3$ have been shifted away. Notice that the solutions for the two constraints exists with the condition that $\bar{\omega}_1=\bar{\omega}_2=\bar{\omega}_3$, which is actually a similar phenomenon in the cases of $V_{110}$ and $V_{111}$.
The operators $V_{100}$, $V_{110}$ and $V_{111}$ share formal similarities as we have seen from their closed form. However, they differ from $V_{100}$, $V_{010}$ and $V_{001}$ in that the projector $\delta$ function changes. In the operator $V_{110}$, Eq.~\eqref{eq.110}, we need the projector that forces $\bar{\omega}_1=\bar{\omega}_2$, while $\bar{\omega}_3$ is forced to be trivial. Similarly for $V_{101}$ and $V_{011}$. In the operator $V_{111}$, Eq.~\eqref{eq.111}, $\bar{\omega}_1$, $\bar{\omega}_2$ and $\bar{\omega}_3$ are forced to be the same by the $\delta$ function. These $\delta$ functions will be essential when we compute the correlation functions in Sec.~\ref{subsec: correlation 2+1}.
We have listed all possible Wilson operators and flux insertion operators, using gauge invariance and locality as our principle constraints. Our ansatz is inspired by considering anomaly inflow in lower dimensions. We have found 8 types of Wilson operators $U_{pqr}$, $(p,q,r=0,1)$, and 8 types of flux insertion operators $V_{pqr}$, $(p,q,r=0,1)$. Therefore there should be 64 types line operators including all the composites of Wilson and flux insertion operators. However, in the next section, Sec.~\ref{subsec: 22 line operators}, we will show that many operators are identified due to the $\delta$ function, and there are only 22 distinguishable line operators in total. This agrees with the quantum double calculation in Ref.~\onlinecite{propitius1995topological}.
\subsection{22 Distinguishable Line Operators}
\label{subsec: 22 line operators}
In this section, we will show that in the type III twisted $\mathbb{Z}_2^{\otimes 3}$ field theory Eq.~\eqref{eq: lag 2+1}, there are only 22 distinguishable line operators rather than naively 64 line operators. To show this, we argue that some operators always have the same correlation functions. Hence, many of the naive $64$ operators should be identified since they have identical correlation functions with all other operators. The essential point is that $V_{pqr}$ are always associated with certain constraints ($\delta$ functions) on the gauge fields $A_i$. See the $\delta$ functions in Eq.~\eqref{eq.100 close form}, \eqref{eq.110} and \eqref{eq.111} resulting from integrating out the matter fields $\lambda_i,\phi_i$. As a result, the insertion of a flux insertion operator along loop $\mathcal{C}$ fixes certain combinations of holonomies of the gauge fields along the same loop $\mathcal{C}$. The flux insertion operator then has trivial fusion rules with Wilson lines corresponding to the mentioned holonomies, simply because the flux insertion operator fixes the values of the Wilson lines. So fusing the flux insertion line with certain Wilson lines is precisely the same as inserting just the flux insertion line.
Let us argue more concretely with an example. We have already listed 8 pure Wilson operators $U_{pqr}$ which insert charges, and 8 pure flux insertion operators $V_{pqr}$ which insert fluxes. We now consider composites of the two kinds of operator. First consider $V_{100}$ along loop $\gamma$ and fuse it with Wilson operator $U_{p q r}$. One can compute the correlation function of the composite operator with arbitrary operator $\langle \mathcal{O}V_{100}\times U_{p q r}\rangle$, and measure the effect of the additional Wilson operator. We assume that the support of the operator $\mathcal{O}$ excludes $\gamma$.
Multiplying $V_{100}$ by $U_{010}(\gamma)$ or $U_{001}(\gamma)$ or their combination will not change the correlation function, because $A_2$ and $A_3$ fields have trivial holonomy along $\gamma$ -- this follows from the $\delta$ function constraint in \eqnref{eq.100 close form}, which directly implies $\exp(\mathrm{i}\oint_{\gamma}A_2)$ and $\exp(\mathrm{i}\oint_{\gamma}A_3)$ equal 1. Then $U_{010}=1$ and $U_{001}=1$ within the correlation functions $\langle V_{100}(\gamma) U_{010}(\gamma) \mathcal{O} \rangle$ and $\langle V_{100}(\gamma)U_{001}(\gamma)\mathcal{O} \rangle$ for any $\mathcal{O}$.
On the other hand, there is no constraint on $\bar{\omega}_1$ in $V_{100}$, so the holonomy of $A_1$ around $\gamma$ is unconstrained, and indeed we can (and will) construct operators $\mathcal{O}$ such that $V_{100} \times U_{100}\mathcal{O}\neq V_{100} \mathcal{O}$ within a correlation function. To summarize, we find that
\begin{equation}
\begin{split}
\langle V_{100}(\gamma)\mathcal{O}\rangle
=&\langle V_{100}(\gamma)U_{010}(\gamma)\mathcal{O}\rangle \\
=&\langle V_{100}(\gamma)U_{001}(\gamma)\mathcal{O}\rangle, \\
=&\langle V_{100}(\gamma)U_{010}(\gamma)U_{001}(\gamma)\mathcal{O}\rangle \\
\langle V_{100}U_{100}(\gamma)\mathcal{O}\rangle
=&\langle V_{100}U_{100}(\gamma)U_{010}(\gamma)\mathcal{O}\rangle \\
=&\langle V_{100}U_{100}(\gamma)U_{001}(\gamma)\mathcal{O}\rangle \\
=&\langle V_{100}U_{100}(\gamma)U_{010}(\gamma)U_{001}(\gamma)\mathcal{O}\rangle\punc{.}
\end{split}
\end{equation}
Therefore, all of the distinguishable operators associated with $V_{100}$ are divided into two equivalence classes --that is, they have the same correlation functions as one of $V_{100}(\gamma)$, $V_{100}(\gamma)U_{100}(\gamma)$. We adopt the quantum double notation by denoting the two classes of $V_{100}(\gamma)$ and $V_{100}(\gamma)U_{100}(\gamma)$ as $(100,\alpha_{\pm}^1)$ respectively, where $100$ represents $V_{100}$ and the plus sign corresponds to $V_{100}(\gamma)$ while the minus sign corresponds to $V_{100}(\gamma)U_{100}(\gamma)$.
Similar arguments can also be applied for $V_{010}$ and $V_{001}$, where the operators are denoted similarly by $(010,\alpha_{\pm}^2)$, $(001,\alpha_{\pm}^3)$.
Using the same ideas, we consider fusing $V_{110}(\gamma)$ with various Wilson lines. Once again the $\delta$ function constraints arising from integrating out the $\phi,\lambda$ matter fields Eq.~\eqref{eq.110} are useful. In this case, the constraints imply that $A_1$ and $A_2$ share the same holonomy along $\gamma$, while $A_3$ has no holonomy along $\gamma$. As a result $V_{110}U_{100}$ always gives the same correlation functions as $V_{110}U_{010}$ does, and $V_{110}U_{110}$ and $V_{110}U_{001}$ give the same correlation functions as $V_{110}$. Therefore, we there are two equivalence classes of $V_{110}$ operator with representatives (for example) $V_{110}$ and $V_{110}U_{100}$. We denote them by $(110, \beta^3_{\pm})$ respectively. The same line of reasoning also applies to $V_{011}$ and $V_{101}$. We denote the operators by a similar notation, $(011, \beta^1_{\pm})$ and $(101, \beta^2_{\pm})$.
Finally, let us consider the possible fusions of $V_{111}(\gamma)$ with Wilson lines. Using the constraint in \eqnref{eq.111}, $A_1$, $A_2$ and $A_3$ must share the same holonomy along $\gamma$. Therefore, we find that $U_{100}$, $U_{010}$, $U_{001}$ and $U_{111}$ are equivalent along $\gamma$. Moreover, $U_{110}$, $U_{011}$ and $U_{101}$ are all equal to $1$ and do not contribute any phases to the correlation functions. Therefore, once again there are two equivalence classes of line operators which we denote $(111,\gamma_{\pm})$, where ``$+$" sign corresponds to $V_{111}$ itself, or its decorations by $U_{110}$, $U_{011}$ and $U_{101}$, and ``$-$" sign corresponds to the equivalence class $V_{111}U_{100}$, $V_{111}U_{010}$, $V_{111}U_{001}$ and $V_{111}U_{111}$.
In summary, we have $22$ distinguishable operators in total: $U_{pqr}~(p,q,r=1,2,3)$, $(100,\alpha^1_{\pm})$, $(010,\alpha^2_{\pm})$, $(001,\alpha^3_{\pm})$, $(011, \beta^1_{\pm})$, $(101, \beta^2_{\pm})$, $(110, \beta^3_{\pm})$, and $(111,\gamma_{\pm})$. The same result also arises from quantum double calculation with type III cocycles. C.f. Ref.~\onlinecite{propitius1995topological}. We have therefore established a 1-to-1 map between field theoretical operators and the projective representations in quantum double models.
\subsection{Fusion Rules of Line Operators}
\label{subsec: fusion rules 2+1}
Having identified the various possible gauge invariant line operators, we calculate their fusion rules. This allows us to motivate the normalizations used in defining the line operators, e.g., the factor of two appearing in Eq.~\eqref{eq.100 close form}.
In quantum field theory, the fusion of two line operators is defined via the process of dragging two lines operators close to each other. The outcome of the product of two line operators can be decomposed as a sum of a set of line operators. If the fusion outcome can only contain one operator, we will call such theories and operators, ``abelian" theories and ``abelian particles" respectively . Similarly, we will call them ``non-abelian" theories and ``non-abelian" particles if there exists more than one fusion outcome.
To begin with, we can calculate fusion rules of Wilson operators quite straightforwardly
\begin{equation}\label{eq.U fuse U}
U_{pqr}(\gamma) \times U_{xyz}(\gamma) = U_{(q+x)(q+y)(r+z)}(\gamma)\punc{,}
\end{equation}
where the sums are defined modulo 2. The fusion rules, Eq.~\eqref{eq.U fuse U}, also demonstrate that all Wilson operators are abelian.
Next we address the flux insertion operators. Henceforth, for simplicity, we adopt the closed form of flux insertion operators $V_{pqr}$ written in terms of the holonomy functions $\omega_i$'s. For example Eq.~\eqref{eq.100 close form}. To begin, let us fuse the same two flux insertion operators $(100,\alpha^1_+)$
\begin{equation}
(100,\alpha^1_+)\times(100,\alpha^1_+) \equiv V_{100}\times V_{100}\punc{,}
\end{equation}
where we have used the fact established in the last section that $V_{100}$ is a representative in the class of $(100,\alpha^1_+)$. By definition of $V_{100}$ in Eq.~\eqref{eq.100 close form}, we have
\begin{equation}\label{eq.100 fuse 100}
\begin{split}
& V_{100}\times V_{100} \\
=& 4\exp(\mathrm{i}\oint_{\gamma} 2b_1+\frac{2}{\pi}\omega_2d\omega_3)) (\delta(\bar{\omega}_2|_{\gamma}))^2(\delta(\bar{\omega}_3|_{\gamma}))^2 \\
=& 4\delta(\bar{\omega}_2|_{\gamma})^2\delta(\bar{\omega}_3|_{\gamma})^2 \\
=& 4\delta(\bar{\omega}_2|_{\gamma})\delta(\bar{\omega}_3|_{\gamma}) \\
=& (1+\exp(\mathrm{i}\bar{\omega}_2))(1+\exp(\mathrm{i}\bar{\omega}_3)) \\
=& 1+\exp(\mathrm{i}\bar{\omega}_2)+\exp(\mathrm{i}\bar{\omega}_3)+\exp(\mathrm{i}\bar{\omega}_2)\exp(\mathrm{i}\bar{\omega}_3) \\
=& U_{000}+U_{010}+U_{001}+U_{011}.
\end{split}
\end{equation}
This rather bizarre looking calculation requires some explanations. The first equality just follows from definition of $V_{100}$, and the $\delta$ function is actually a projector that projects into zero flux state (more explicitly, $\delta(\bar{\omega}_i|_{\gamma}):=\frac{1}{2}(1+\exp(\mathrm{i}\bar{\omega}_i)),~(i=1,2,3)$ as noted below \eqnref{eq:wi}); the second equality follows from the fact that all variables are $\mathbb{Z}_2$ variables valued in $\{0,\pi\}$, then the exponential is actually trivial because it is always $2\pi$ \footnote{ It may seem strange to discuss a $\mathbb{Z}_2$ in the continuum. In fact the expressions here are short-hand for a more careful (but more cumbersome) calculation on the lattice, where there is certainly no obstacle to considering $\mathbb{Z}_2$ and more generally discrete valued fields.}. The third equality follows from the fact that the $\delta$ function satisfies $(\delta(\bar{\omega}_i))^2=\delta(\bar{\omega}_i)$; the fourth equality just expresses the $\delta$ functions explicitly as $\delta(\bar{\omega}_i|_{\gamma})=\frac{1}{2}(1+\exp(\mathrm{i}\bar{\omega}_i))$.
Using quantum double notation, Eq.~\eqref{eq.100 fuse 100} is expressed as
\begin{equation}\label{eq:integerfusion}
(100,\alpha^1_+)\times(100,\alpha^1_+)=1 + U_{010} + U_{001} + U_{011}.
\end{equation}
As promised in \secref{subsec: line operators 2+1}, we need to motivate the normalization factors for the flux insertion operators. Indeed, the fact insisting that fusion rules like \eqnref{eq:integerfusion} involve positive integer combinations of line operators fixes the overall normalization factors (e.g., the $2$ factor in Eq.~\eqref{eq.100}).
As another example, consider fusion rule
\begin{equation}
(010,\alpha^2_+)\times(001,\alpha^3_+)\equiv V_{010}\times V_{001}\punc{.}
\end{equation}
To see how this comes about, we use our explicit expressions for the line operators (Eq.~\eqref{eq.010} and Eq.~\eqref{eq.001})
\begin{equation}
\begin{split}\label{eq:010x001}
V_{010}\times V_{001}=&4\exp\Big[\mathrm{i}\oint_{\gamma} b_2+b_3+\frac{1}{\pi}(-\omega_1 d \omega_3 + \omega_1 d \omega_2)\Big] \\
& \delta(\bar{\omega}_1|_{\gamma})\delta(\bar{\omega}_2|_{\gamma})\delta(\bar{\omega}_3|_{\gamma})\punc{.}
\end{split}
\end{equation}
The right hand side of this equation can be manipulated into the form
\begin{equation}
V_{010}\times V_{001}=V_{011}+V_{011}U_{010}=V_{011}+V_{011}U_{001}\punc{.}
\end{equation}
To see why, in \eqnref{eq:010x001} rewrite $\delta(\bar{\omega}_2 |_\gamma) \delta(\bar{\omega}_3 |_\gamma) = \delta(\bar{\omega}_2 |_\gamma - \bar{\omega}_3 |_\gamma) \delta(\bar{\omega}_3 |_\gamma)$ and expand $\delta(\bar{\omega}_3|_{\gamma})=\frac{1}{2} (1+ \exp(\mathrm{i} \bar{\omega}_3 |_\gamma))$. Then compare the result with the definition of $V_{011}$ from Eq.~\eqref{eq.110}
\begin{equation}
\begin{split}
V_{011}=& 2\exp\Big[i\oint_{\gamma} b_2+b_3 +\frac{1}{\pi}(-\omega_1 d \omega_3 + \omega_1 d \omega_2)\Big] \\
&\delta(\bar{\omega}_1|_{\gamma})\delta(\bar{\omega}_2|_{\gamma}-\bar{\omega}_3|_{\gamma})\punc{.}
\end{split}
\end{equation}
In terms of the quantum double notation we have shown that
\begin{equation}
(010,\alpha^2_+) \times (001,\alpha^3_+) = (011,\beta^1_+) + (011,\beta^1_-)\punc{.}
\end{equation}
We can readily find the quantum dimensions of all of the operators above. The Wilson lines all have quantum dimension $1$. This follows most readily from the fact that $U_{pqr}\times U_{pqr} = 1$ in \eqnref{eq.U fuse U}, along with some general constraints on the structure of fusion algebras \footnote{Use Eq. 14 of Ref.~\onlinecite{bernevig2015topological}.}. Moreover, the fusion $(100, \alpha^+)\times (100, \alpha^+)$ in \eqnref{eq:integerfusion} gives a sum of four Wilson lines. Again using Ref.~\onlinecite{bernevig2015topological}, this implies that the quantum dimension of $(100, \alpha^+)$ is 2. Similarly, for other flux insertion operators. The overall factor $2$ in the definitions of the flux insertion lines operators, for example Eq.~\eqref{eq.100 close form}, is actually the quantum dimension for the operators.
In summary, we have worked out several examples of fusion rules and quantum dimensions of the line operators using our field theoretic formalism. We have demonstrated that the type III twisted $\mathbb{Z}_2^{\otimes3}$ theory is a non-abelian topological phase, even though it is a topological field theory involving only abelian gauge fields.
\subsection{Correlation Functions of Line Operators}
\label{subsec: correlation 2+1}
In this section, we calculate correlation functions for line operators that link one another. Typically, we will consider two line operators forming a Hopf link in (2+1)D, Fig.~\ref{fig. Hopf}. If we have a link of two t' Hooft operators corresponding to gauge fluxes $\phi_1,\phi_{2}$, then the holonomy along the first loop is $\phi_1$ while that along the second loop is $\phi_{2}$. As we have seen before, the flux insertion operators are associated with constraints on the holonomies of $A_i$ along the loop. As a result, we will see that for many of the possible links the holonomies are not compatible with the constraints, so the expectation value for the link is simply zero.
\begin{figure}[b]
\centering
\includegraphics[width=0.4\textwidth]{hopflink.pdf}
\caption{An illustration of a linking correlation of $V_{100}(\gamma_1)$ and $V_{010}(\gamma_2)$.}
\label{fig. Hopf}
\end{figure}
To demonstrate this point, examine a link of $V_{100}(\gamma_1)$ and $V_{010}(\gamma_2)$: $V_{010}(\gamma_2)$ will insert a holonomy $\pi$ of $A_2$ along $\gamma_1$. See Fig~\ref{fig. Hopf}. However, we know from \eqnref{eq.100} that $V_{100}(\gamma_1)$ is associated with two constraints $\oint_{\gamma_1} A_2=\oint_{\gamma_1} A_3=0$. The mismatch between the holonomy and the constraint leads to a zero expectation value. For explicit path integral calculation details for the linking correlation of $V_{100}(\gamma_1)$ and $V_{010}(\gamma_2)$, please refer to App.~\ref{app. path integral}.
The mismatching of the constraints and flux insertion mean that most of the Hopf links we consider disappear. Here in the main text, we present only one subtle calculation, the corellator $\langle V_{111}(\gamma_1)V_{111}(\gamma_2)\rangle $ where $\gamma_1$ and $\gamma_2$ form a link in (2+1)D.
\begin{widetext}
\begin{equation}\label{eq.111correlation}
\begin{split}
&\langle V_{111}(\gamma_1) V_{111}(\gamma_2) \rangle \\
=& 4 \int Db_i DA_i \exp(\mathrm{i} S_0) \exp(\mathrm{i}\oint_{\gamma_1} (b_1+b_2+b_3)+\frac{\mathrm{i}}{2\pi}\oint_{\gamma_1}\sum_{ijk}\epsilon^{ijk} \omega_i d \omega_j)
\delta(\bar{\omega}_1|_{\gamma_1}-\bar{\omega}_2|_{\gamma_1})\delta(\bar{\omega}_2|_{\gamma_1}-\bar{\omega}_3|_{\gamma_1}) \\
& \exp(\mathrm{i}\oint_{\gamma_2} (b_1+b_2+b_3)+\frac{\mathrm{i}}{2\pi}\oint_{\gamma_2}\sum_{ijk}\epsilon^{ijk} \omega_i d w_j)
\delta(\bar{\omega}_1|_{\gamma_2}-\bar{\omega}_2|_{\gamma_2})\delta(\bar{\omega}_2|_{\gamma_2}-\bar{\omega}_3|_{\gamma_2}) \\
=& 4 \exp(\mathrm{i}\int \frac{1}{\pi^2} \tilde{A}_1\tilde{A}_2\tilde{A}_3) \exp(\frac{\mathrm{i}}{2\pi}\oint_{\gamma_1}\sum_{ijk}\epsilon^{ijk} \tilde{\omega}_i d \tilde{\omega}_j) \exp(\frac{\mathrm{i}}{2\pi}\oint_{\gamma_2}\sum_{ijk}\epsilon^{ijk} \tilde{\omega}_i d \tilde{\omega}_j) \\
=& 4 \exp(\frac{\mathrm{i}}{2\pi}\oint_{\gamma_1}\sum_{ijk}\epsilon^{ijk} \tilde{\omega}_i d \tilde{\omega}_j) \exp(\frac{\mathrm{i}}{2\pi}\oint_{\gamma_2}\sum_{ijk}\epsilon^{ijk} \tilde{\omega}_i d \tilde{\omega}_j).
\end{split}
\end{equation}
\end{widetext}
The first equality follows from our definitions of the line operators. The second equality is obtained by integrating out all $b_i$ fields. This yields the constraints
\begin{equation}
\frac{1}{\pi}dA_i=\star (j(\gamma_1)+j(\gamma_2)),
\end{equation}
where $j(\gamma_1)$ and $j(\gamma_2)$ are the unit vectors tangential to $\gamma_1$ and $\gamma_2$ respectively. Here $\star$ is the Hodge dual. For positions away from the support of the loops $\gamma_{1,2}$ we have $dA_i=0$. We denote by $\tilde{A}_i$ a particular reference solution to the constraints on $A_i$. One thing we certainly know about the $\tilde{A}_i$ is that they have nontrivial holonomies along $\gamma_1$ and $\gamma_2$, i.e.
\begin{equation}
\oint_{\gamma_1} \tilde{A}_i=\oint_{\gamma_2} \tilde{A}_i=\pi \mod 2\pi ,~\forall i=1,2,3.
\end{equation}
Therefore, the $\delta$ functions in the first equality in \eqnref{eq.111correlation} is automatically satisfied. Moreover, since all $A_i$ obey the same equations of motion set by integrating out $b_i$ fields, their solutions $\tilde{A}_i$ are the same up to gauge transformations. Therefore, we are free to choose a gauge for which the $\tilde{A}_i$ are identical, so that the integral of $\frac{1}{\pi^2} \tilde{A}_1\tilde{A}_2\tilde{A}_3$ vanishes at least in this quasi-continuum setting. All we need to do now is to evaluate the last two line integrals in the last line of Eq.~\eqref{eq.111correlation}.
In the last line, $\tilde{\omega}_i$'s are the holonomy functions for each field $A_i$. As we have chosen a gauge for which the $\tilde{A}_i$ are identical, the corresponding holonomy functions are identical. As a result, on each loop the $\tilde{\omega}_i$ are identical multi-valued staircase functions which sharply step up by $\pi$ modulo $2\pi$ upon moving around the loop once.
In order to evaluate these line integrals, we need to think more carefully about the regularization of the field theory. To this end, we consider using discrete derivatives on a lattice of form
\begin{equation}
\begin{split}
d\tilde{\omega}_i(r) =& \tilde{\omega}_i(r+1) - \tilde{\omega}_i(r) \\
\bar{d}\tilde{\omega}_i(r) =& \tilde{\omega}_i(r) - \tilde{\omega}_i(r-1),
\end{split}
\end{equation}
where $\bar{d}$ is the adjoint operator to $d$ on the lattice, $r$ is the position on the lattice. In terms of these operators, the line integrals over $\gamma_1$ and $\gamma_2$ are regularized as
\begin{equation}
\begin{split}
& \oint_{\gamma_1/\gamma_2} \tilde{\omega}_i d \tilde{\omega}_j - \tilde{\omega}_j \bar{d} \tilde{\omega}_i \\
=&\sum_{r=-L/2}^{L/2} \lbrack \tilde{\omega}_i(r)(\tilde{\omega}_j(r+1)-\tilde{\omega}_j(r)) \\
&- \tilde{\omega}_j(r)(\tilde{\omega}_i(r)-\tilde{\omega}_i(r-1)) \rbrack \\
=& \tilde{\omega}_i(-1)\pi - \tilde{\omega}_j(0)\pi \\
=& -\pi^2.
\end{split}
\end{equation}
More details of the particular choice of derivatives can be found in the Sec.~\ref{sec. ZN examples} when we verify that the operators for type III twisted $\mathbb{Z}_N^{\otimes3}$ theories are gauge invariant on lattice. Hence we have
\begin{equation}
\begin{split}
\sum_{ijk} \epsilon^{ijk} \tilde{\omega}_i d\tilde{\omega}_j \equiv & (\tilde{\omega}_2 d\tilde{\omega}_3 - \tilde{\omega}_3 \bar{d}\tilde{\omega}_2) + (\tilde{\omega}_3 \bar{d}\tilde{\omega}_1 - \tilde{\omega}_1 d\tilde{\omega}_3) \\
&+ (\tilde{\omega}_1 d\tilde{\omega}_2 - \tilde{\omega}_2 \bar{d}\tilde{\omega}_1) \\
=& -\pi^2+\pi^2-\pi^2 \\
=& -\pi^2.
\end{split}
\end{equation}
Substituting these results back into the line integral, we obtain
\begin{equation}
\begin{split}
\langle V_{111}(\gamma_1) V_{111}(\gamma_2) \rangle =& 4 \exp(\frac{i}{2\pi}\oint_{\gamma_1+\gamma_2}\sum_{ijk}\epsilon^{ijk} \bar{\omega}_i d \bar{\omega}_j) \\
=& -4\punc{.}
\end{split}
\end{equation}
The correlation function suggests that the topological spin of $V_{111}$ is either $\mathrm{i}$ or $-\mathrm{i}$. A similar calculation for the linking correlation $\langle V_{n_1n_2n_3}(\gamma_1) V_{m_1m_2m_3}(\gamma_2) \rangle$ for type III twisted $\mathbb{Z}_N^{\otimes3}$ theory can be found in Sec.~\ref{sec. ZN examples}. And we provide more comments on the topological spins there, and find the topological spin can actually fixed to be $-\mathrm{i}$ for $V_{111}$, and hence $\mathrm{i}$ for $V_{111}U_{111}$
In summary for this section (and App.~\ref{app. path integral}) we have calculated the correlation functions of linked line operators. The vanishing correlation functions (App.~\ref{app. path integral}) are further indications that type III twisted $\mathbb{Z}_2^{\otimes3}$ is a non-abelian topological theory. The modular matrices of type III twisted $\mathbb{Z}_2^{\otimes3}$ are explicitly written down in App.~\ref{app. ModMat}.
\section{Type III Twisted $\mathbb{Z}_N^{\otimes3}$ Theory}
\label{sec. ZN examples}
In this section, we generalize the gauge group from $\mathbb{Z}_2^{\otimes 3}$ to $\mathbb{Z}_N^{\otimes 3}$. More explicitly, we construct the line operators and their correlation functions etc for the type III twisted $\mathbb{Z}_N^{\otimes3}$ theories in (2+1)D. The basic idea of the constructing these line operators is still introducing the auxiliary fields and gauge invariance. Once we obtain the valid line operators, we can obtain their linking correlation function by path integral.
This section is divided into the following: In Sec.~\ref{app.sec. lag}, we again introduce the Lagrangian and gauge transformations; In Sec.~\ref{app.sec oper}, we list our line operators; In Sec.~\ref{app.sec. correlation}, we work out correlation functions of flux insertion operators.
Moreover, two appendices are associated with this section: In App.~\ref{app.sec. gauge}, we verify the gauge invariance of flux insertion operators with lattice regularization; And in App.~\ref{app. QD}, we provide with a quantum double calculation which gives the same results of correlation functions as in the field theory approach derived in the following main text.
\subsection{Lagrangian and Gauge Transformation}
\label{app.sec. lag}
In this section, we introduce the Lagrangian for the twisted $\mathbb{Z}_N^{\otimes3}$ theory and its gauge transformation, as a preparation for the following sections. The Lagrangian for the theory is
\begin{equation}\label{lag. 2+1 ZN}
\mathcal{L}=\frac{N}{2\pi} b_i dA_i + \frac{pN^2}{(2\pi)^2} A_1 A_2 A_3,
\end{equation}
where $p\in\mathbb{Z}_N=\{0,1,2,\ldots,N-1\}$, which can be determined by the same method in App.~\ref{app. coefficient}. The gauge transformations, by Eq.~\eqref{eq: gauge transf 2+1}, are
\begin{equation}\label{eq. gauge transf 2+1 ZN}
\begin{split}
b_i &\rightarrow b_i+d\beta_i+\frac{pN\epsilon^{ijk}}{2\pi}(A_j\alpha_k-\frac{1}{2}\alpha_jd\alpha_k) \\
A_i &\rightarrow A_i+d\alpha_i, \;\; i=1,2,3.
\end{split}
\end{equation}
In the following sections, we will find out all line operators that are gauge invariant under the gauge transformation Eq.~\eqref{eq. gauge transf 2+1 ZN}, and work out some of their fusion rules and correlation functions etc.
\subsection{Line Operators}
\label{app.sec oper}
In this section, we find out all possible gauge invariant line operators. First of all, by gauge invariance, the Wilson line operators are:
\begin{equation}
U_{n_1n_2n_3}=\exp(\mathrm{i} \oint_{\gamma} n_iA_i), n_i\in\{0,1,.\ldots,N-1\}.
\end{equation}
Similar to the type III twisted $\mathbb{Z}_2^{\otimes 3}$ example in Sec.~\ref{sec: 2+1D}, the flux insertion operators can be constructed by introducing the auxiliary fields, $\phi_i$ and $\lambda_i$. From the decorated domain wall picture, the single type flux can be inferred from SPTs boundary with symmetry group $\mathbb{Z}_{N}^{\otimes 2}$ instead of $\mathbb{Z}_N^{\otimes 3}$. The reason is that in the path integral with a single flux insertion operator, for example $V_{100}$, we have flat connections $A_2$ and $A_3$ and non-flat connection $A_1$. Thus, the flux sheet of $A_1$ is actually a $\mathbb{Z}_N^{\otimes 2}$ SPT whose boundary is the flux loop. Therefore, by gauge anomaly inflow, we can construct the operator $V_{100}$ explicitly where auxiliary fields need to be introduced.
The auxiliary fields can be integrated out to produce a closed form for these flux insertion operators. However, instead of writing them down directly, we explain from the single type flux insertion operators to the triple type fluxes, mainly because the quantum dimensions of these operators are not written in a uniform way.
The single type of flux insertion operators $V_{r00}$ are:
\begin{widetext}
\begin{equation}
\label{eq.r00}
\begin{split}
V_{r00}=&\frac{1}{\mathcal{N}}\int D[\phi_2]D[\phi_3]D[\lambda_2]D[\lambda_3] \exp(\mathrm{i} \oint_{\gamma} rb_1+\frac{\mathrm{i} rpN}{2\pi}\oint_{\gamma}\epsilon^{1ij}(\frac{1}{2}\phi_i d\phi_j + (d\phi_i - A_i)\lambda_j )) \\
=& \mathcal{N}_{r00} \exp(\mathrm{i} \oint_{\gamma} rb_1+\frac{\mathrm{i} rpN}{4\pi}\oint_{\gamma}\epsilon^{1ij}\omega_id\omega_j) \delta(rp\bar{\omega}_2)\delta(rp\bar{\omega}_3),
\;\; r=0,1,\ldots, N-1.
\end{split}
\end{equation}
\end{widetext}
It is clearly gauge invariant as the $\phi_i$, $\lambda_i$ fields transform exactly the same as in $\mathbb{Z}_2^{\otimes 3}$ case. And the integration over the auxiliary fields are also the same. We emphasis that delta function $\delta(x)$ is still a projector, imposing any element $x \in 2\pi \mathbb{Z}$. The subtle difference from $\mathbb{Z}_2^{\otimes 3}$ is the normalization constant $\mathcal{N}_{r00}$ which is determined by fusion rules, for example $V_{r00} \times V_{(N-r)00}$. The flux is trivial after fusion. Hence we only expect charges appear in the fusion channels if the fusion is possibly nontrivial.
$\mathcal{N}_{r00}$ is fixed to be:
\begin{equation}\label{eq.r00 norm}
\mathcal{N}_{r00}=\frac{N}{\mathrm{gcd}(N,rp)} \;\;.
\end{equation}
The reason for it is that the fusion rule of $V_{r00}$ and $V_{(N-r)00}$ is with Eq.~\eqref{eq.r00 norm}:
\begin{equation}\label{eq: fusion rule ZN}
V_{r00}\otimes V_{(N-r)00}=\bigoplus_{i,j=0}^{{\mathcal{N}_{r00}}-1} U_{0(ipr)(jpr)},
\end{equation}
where the fusion channels on the RHS has the greatest common divisor $1$, and the identity operator $U_{000}$ only appears once. The fusion rule is derived by taking the product of $V_{r00}$ and $V_{(N-r)00}$, canceling the exponential phases and expanding the $\delta$ function as follows:
\begin{equation}\label{eq. ZN delta}
\delta(rp\bar{\omega}_j) = \frac{\mathrm{gcd}(N,rp)}{N} \sum\limits_{m=0}^{\frac{N}{\mathrm{gcd}(N,rp)}-1} \exp\left(\mathrm{i} m rp \bar{\omega}_j \right),
\end{equation}
where $j=2,3$. And we also use the fact for the derivation of fusion rule Eq.~\eqref{eq: fusion rule ZN}:
\begin{equation}
\mathcal{N}_{r00} = \frac{N}{\mathrm{gcd}(N,rp)} = \frac{N}{\mathrm{gcd}(N,(N-r)p)} = \mathcal{N}_{(N-r)00}.
\end{equation}
For the $\mathbb{Z}_2^{\otimes 3}$ example discussed in the previous section, where $N=2,p=1$, we have:
\begin{equation}
\mathcal{N}_{100}=\frac{2}{\mathrm{gcd}(2,1)}=2 \;.
\end{equation}
Having fixed the normalization for $V_{r00}$, the quantum dimensions for these flux insertion operators are just $\mathcal{N}_{r00}$, which can be manifested by calculating $\langle V_{r00} \rangle$. (More rigorously, the quantum dimension is obtained via fusion rules.) The explicit calculation is omitted here since it is exactly the same as in $\mathbb{Z}_2^{\otimes 3}$ situation.
Other types of single flux insertion operators $V_{0r0}$ and $V_{00r}$ can be obtained by simply permuting the indices, as in the previous section, Sec.~\ref{sec: 2+1D}. Hence we omit their expressions here for simplicity.
One can also consider inserting two types of fluxes and three types of fluxes. We follow the same prescription as in Eq.~\eqref{eq.111} by introducing the auxiliary fields $\phi_i$ and $\lambda_i$. And integrating the auxiliary fields out yields a closed form of flux insertion operators in terms of the holonomy functions $\omega_i$'s:
\begin{widetext}
\begin{equation}
\begin{split}
V_{n_1n_20}=&\frac{1}{\mathcal{N}'}\int D[\phi_i]D[\lambda_i] \exp\left( \mathrm{i}\oint_{\gamma} n_1 b_1+n_2 b_2+\frac{n_1pN}{2\pi}\epsilon^{1ij}(\frac{1}{2}\phi_i d\phi_j + (d\phi_i-A_i)\lambda_j) + \frac{n_2pN}{2\pi}\epsilon^{2ij}(\frac{1}{2}\phi_i d\phi_j + (d\phi_i-A_i)\lambda_j) \right) \\
=& \mathcal{N}_{n_1n_20}\exp\left( \mathrm{i}\oint_{\gamma} n_1 b_1 + n_2 b_2 - \frac{\mathrm{i} Np}{2\pi}\oint_{\gamma}(n_2\omega_1-n_1\omega_2)d\omega_3 \right) \delta(pn_2\bar{\omega}_1-pn_1\bar{\omega}_2)
\delta(n_1p\bar{\omega}_3)\delta(n_2p\bar{\omega}_3)
\end{split}
\end{equation}
\begin{equation}\label{eq. n1n2n3}
\begin{split}
V_{n_1n_2n_3}=&\frac{1}{\mathcal{N}''}\int D[\phi_i]D[\lambda_i] \exp\left(\mathrm{i}\oint_{\gamma} (n_i b_i +\frac{n_ipN}{2\pi}\epsilon^{ijk}(\frac{1}{2}\phi_j d\phi_k + (d\phi_j-A_j)\lambda_k))\right) \\
=& \mathcal{N}_{n_1n_2n_3} \exp\left(\mathrm{i}\oint_{\gamma} (n_i b_i + n_i\frac{Np}{4\pi}\epsilon^{ijk}\omega_jd\omega_k)\right) \delta(n_2p\bar{\omega}_1-n_1p\bar{\omega}_2)\delta(n_3p\bar{\omega}_2-n_2p\bar{\omega}_3)\delta(n_1p\bar{\omega}_3-n_3p\bar{\omega}_1),
\end{split}
\end{equation}
\end{widetext}
where in Eq.~\eqref{eq. n1n2n3}, only two of the three $\delta$ functions are independent; and $\mathcal{N}_{rs0}$ and $\mathcal{N}_{n_1n_2n_3}$ can be determined similarly as in $\mathcal{N}_{r00}$. More explicitly:
\begin{equation}\label{eq.rs0 norm}
\begin{split}
\mathcal{N}_{n_1n_20} &= \frac{N}{\mathrm{gcd}(N,pn_1,pn_2)} \;\;, \\
\mathcal{N}_{n_1n_2n_3}&=\frac{N}{\mathrm{gcd}(N,pn_1,pn_2,pn_3)} \;\;.
\end{split}
\end{equation}
And similarly as before, the quantum dimensions of $V_{n_1n_20}$ and $V_{n_1n_2n_3}$ are $\mathcal{N}_{n_1n_20}$ and $\mathcal{N}_{n_1n_2n_3}$ respectively. They are also consistent with the case of $N=2, p=1$. We write them in a uniform way:
\begin{equation}
\langle V_{n_1n_2n_3} \rangle = \mathcal{N}_{n_1n_2n_3} \;.
\end{equation}
In the following, we will use the natural convention:
\begin{equation}
\begin{split}
\mathrm{gcd}(a,b,0,0) &\equiv \mathrm{gcd}(a,b) \\
\mathrm{gcd}(a,b,c,0) &\equiv \mathrm{gcd}(a,b,c)
\end{split}
\end{equation}
to simplify our notations and discussions below. We can write all the flux insertion operators uniformally by using this notation.
In summary, we have determined the flux insertion operators as in Eq.~\eqref{eq. n1n2n3}. They are constructed by introducing the auxiliary fields on the fluxes, and integrating out the auxiliary fields yields the closed forms for these operators in terms of the holonomy functions $\omega_i$. The quantum dimensions for the flux insertion operators are determined by the overall coefficients in the closed form of flux insertion operators. See Eq.~\eqref{eq.r00 norm} and \eqref{eq.rs0 norm}. Moreover, in App.~\ref{app.sec. gauge}, we verify the gauge invariance of these operators Eq.~\eqref{eq. n1n2n3} explicitly with lattice regularizations.
We do not elaborate on how many different line operators here, but only comment that because of the $\delta$ functions in the flux insertion operators $V_{n_1n_2n_3}$, attaching a Wilson line onto $V_{n_1n_2n_3}$ may actually contribute nothing trivial phases to the correlation functions. Hence, some operators are identified in the sense of producing the same correlation functions, although their appearances are different.
\subsection{Correlation Functions}
\label{app.sec. correlation}
In this section, we provide with general linking correlation functions for two flux insertion operators, $\langle V_{n_1n_2n_3}(\gamma_1)V_{m_1m_2m_3}(\gamma_2) \rangle$, for the type III twisted $\mathbb{Z}_N^{\otimes3}$ theory. Before we calculate $\langle V_{n_1n_2n_3}(\gamma_1)V_{m_1m_2m_3}(\gamma_2) \rangle$, we first comment on other simpler linking correlation function, for example the linking of two Wilson lines, or Wilson lines and flux insertion operators.
The linking correlations between any two Wilson lines are simply identity. And the linking correlations between flux insertion operators and Wilson line operators remain to be simple. Flux insertion operator $V_{n_1n_2n_3}$ simply inserts $n_1$ units of $A_1$ flux, $n_2$ units of $A_2$ flux and $n_3$ units of $A_3$ flux. Thus Wilson lines that are linked to $V_{n_1n_2n_3}$ simply take three Aharonov-Bhom phases according to the charges of the Wilson lines. However, as we have seen in $\mathbb{Z}_2^{\otimes 3}$ section, the linking correlation functions may vanish due to the constraint part of these operators, or pick up nontrivial phases from the $\omega_i d\omega_j$ terms.
In the following formulas, we do not distinguish the lattice derivatives $d$ and $\bar{d}$ until necessary. And more importantly, we assume $\mathrm{gcd}(N,p)=1$ which simplifies the calculations. The explanations for the assumption will be explained after the calculations. The detailed calculation goes as follows:
\begin{widetext}
\begin{equation}\label{eq. ZN linking}
\begin{split}
&\langle V_{n_1n_2n_3}(\gamma_1) V_{m_1m_2m_3}(\gamma_2) \rangle \\
=& \mathcal{N}_{n_1n_2n_3}\mathcal{N}_{m_1m_2m_3} \int Db_i DA_i \exp\left(\mathrm{i} \int \frac{N}{2\pi} b_i dA_i + \frac{pN^2}{(2\pi)^2} A_1 A_2 A_3 \right) \\
&\exp\left(\mathrm{i}\oint_{\gamma_1}n_ib_i+\frac{\mathrm{i} Np}{4\pi}\oint_{\gamma_1}\epsilon^{ijk} n_i \omega_j d \omega_k\right) \prod_i \delta(\epsilon^{ijk}pn_i\bar{\omega}_k|_{\gamma_1}) \\
&\exp\left(\mathrm{i}\oint_{\gamma_2}m_ib_i+\frac{\mathrm{i} Np}{4\pi}\oint_{\gamma_2}\epsilon^{ijk} m_i \omega_j d \omega_k\right) \prod_i \delta(\epsilon^{ijk}pm_i\bar{\omega}_k|_{\gamma_2}) \\
=& \mathcal{N}_{n_1n_2n_3}\mathcal{N}_{m_1m_2m_3} \exp\left(\mathrm{i}\frac{pN^2}{(2\pi)^2}\int\tilde{A}_1\tilde{A}_2\tilde{A}_3\right)
\exp\left(\frac{\mathrm{i} Np}{4\pi}\oint_{\gamma_1}\epsilon^{ijk} n_i \tilde{\omega}_j d \tilde{\omega}_k\right)
\exp\left(\frac{\mathrm{i} Np}{4\pi}\oint_{\gamma_2}\epsilon^{ijk} m_i \tilde{\omega}_j d \tilde{\omega}_k\right) \\
&\prod_i \delta(\epsilon^{ijk}pn_j\frac{2\pi m_k}{N}) \prod_i \delta(\epsilon^{ijk}pm_j\frac{2\pi n_k}{N}) \\
=& \mathcal{N}_{n_1n_2n_3}\mathcal{N}_{m_1m_2m_3} \exp\left( \frac{\mathrm{i} Np}{4\pi}\oint_{\gamma_1}\epsilon^{ijk} n_i \tilde{\omega}_j d \tilde{\omega}_k\right) \exp\left( \frac{\mathrm{i} Np}{4\pi}\oint_{\gamma_2}\epsilon^{ijk} m_i \tilde{\omega}_j d \tilde{\omega}_k\right) \prod_i \delta(\epsilon^{ijk}\frac{2\pi p}{N} n_jm_k)\prod_i \delta(\epsilon^{ijk}\frac{2\pi p}{N} m_jn_k) \\
=& \mathcal{N}_{n_1n_2n_3}\mathcal{N}_{m_1m_2m_3}
\exp\left(-\frac{\mathrm{i} p\pi}{N}\epsilon^{ijk}(n_im_jm_k + m_in_jn_k)\right)
\prod_i \delta(\epsilon^{ijk}\frac{2\pi p}{N} n_jm_k) \prod_i \delta(\epsilon^{ijk}\frac{2\pi p}{N} m_jn_k).
\end{split}
\end{equation}
\end{widetext}
The equations deserves certain explanations: The first equality just lists all the terms, following the same convention as before. In the second equality, we integrate out all $b_i$ fields, which yields the equations of motion as follows:
\begin{equation}\label{eq. ZN eom}
dA_i = \frac{2\pi}{N}(n_i\star j_1 + m_i\star j_2) \mod{2\pi},~i=1,2,3,
\end{equation}
where $j_1$ and $j_2$ are the currents representing $\gamma_1$ and $\gamma_2$ respectively, $\star$ is the Hodge dual. The solutions of such equations of motion are denoted as $\tilde{A}_i,~i=1,2,3$.
As a result of the equations of motion, we have:
\begin{equation}
\begin{split}
\bar{\omega}_i(\gamma_1)&\equiv\oint_{\gamma_1}A_i=\frac{2\pi}{N} m_i, \\
\bar{\omega}_i(\gamma_2)&\equiv\oint_{\gamma_2}A_i=\frac{2\pi}{N} n_i,~i=1,2,3.
\end{split}
\end{equation}
Now notice that due to the $\delta$ function constraints, the correlation will vanish if they are not satisfied. One subtlety is that the $\delta$ functions associated with $V_{n_1n_2n_3}$ and $V_{m_1m_2m_3}$ are slightly different, because the summation periods as in Eq.~\eqref{eq. ZN delta} are determined by $\mathcal{N}_{n_1n_2n_3}$ and $\mathcal{N}_{m_1m_2m_3}$ respectively:
\begin{equation}
\begin{split}
& \delta(\epsilon^{ijk}\frac{2\pi p}{N}n_jm_k) \\
=& \frac{1}{\mathcal{N}_{n_1n_2n_3}} \sum\limits_{q=0}^{\mathcal{N}_{n_1n_2n_3}-1} \exp\left(\mathrm{i} q \epsilon^{ijk}\frac{2\pi p}{N}n_jm_k\right) \\
& \delta(\epsilon^{ijk}\frac{2\pi p}{N}m_jn_k) \\
=& \frac{1}{\mathcal{N}_{m_1m_2m_3}} \sum\limits_{q=0}^{\mathcal{N}_{m_1m_2m_3}-1} \exp\left(\mathrm{i} q \epsilon^{ijk}\frac{2\pi p}{N}m_jn_k\right).
\end{split}
\end{equation}
These $\delta$ functions leads to the following equations:
\begin{equation}\label{eq. constraint result}
\begin{split}
& n_2m_3-n_3m_3 = 0 \mod{\frac{N}{\mathrm{gcd}(N,p)}}, \\
& \text{and permuted equations.}
\end{split}
\end{equation}
For the convenience of the following calculations, we assume that $\mathrm{gcd}(N,p)=1$. Thus the above equations Eq.~\eqref{eq. constraint result} are valid mod $N$. Therefore, the RHS of the equations of motions Eq.~\eqref{eq. ZN eom} are proportional to each other, for $i=1,2,3$. Hence, the solutions, $\tilde{A}_i,~i=1,2,3$, can be set to proportional to each other, up to gauge transformations. Hence, the first integral of the second equality will vanish at least in the continuous limit, leading to the third equality. Note that if $\mathrm{gcd}(N,p)\neq 1$, the argument that $\int \tilde{A}_1\tilde{A}_2\tilde{A}_3$ vanishes may not be true.
The fourth equality is obtained by using the lattice derivatives, $d$ and $\bar{d}$. For example:
\begin{equation}
\oint_{\gamma_1} \tilde{\omega}_2 d\tilde{\omega}_3 - \tilde{\omega}_3 \bar{d} \tilde{\omega}_2 = -\left(\frac{2\pi}{N}\right)^2 m_2m_3.
\end{equation}
Thus, we have completed the calculation of $\langle V_{n_1n_2n_3}(\gamma_1)V_{m_1m_2m_3}(\gamma_2) \rangle$.
One particular simple and non-vanishing example of these linking correlations is
\begin{equation}\label{eq. linking n1n2n3}
\begin{split}
&\langle V_{n_1n_2n_3}(\gamma_1) V_{n_1n_2n_3}(\gamma_2)\rangle \\
=&\mathcal{N}_{n_1n_2n_3}^2 \exp\left(\frac{-2\pi\mathrm{i} p}{N}n_1n_2n_3\right) \;.
\end{split}
\end{equation}
where we need to recall that:
\begin{equation}
\mathcal{N}_{n_1n_2n_3} = \frac{N}{\mathrm{gcd}(N,pn_1,pn_2,pn_3)}.
\end{equation}
The linking correlation function Eq.~\eqref{eq. linking n1n2n3} also suggests the topological spin for $V_{n_1n_2n_3}$ is:
\begin{equation}\label{eq. spin ZN}
\Theta(V_{n_1n_2n_3}) = \exp\left( -\frac{\pi\mathrm{i} p}{N}n_1n_2n_3 \right) \;,
\end{equation}
although it is a non-abelian topological phase. The reason is the following: Suppose we have a ``self-twist" loop $\gamma$. See Fig.~\ref{fig.ribbon} for an illustration of self-twist. Then $V_{n_1n_2n_3}(\gamma)$ itself inserts $n_{1,2,3}$ units of $A_{1,2,3}$ fluxes through $\gamma$. Hence in the path integral of $\langle V_{n_1n_2n_3}(\gamma) \rangle$, we only have one contribution for the phase, instead of two contributions as in the last two equality of Eq.~\eqref{eq. ZN linking}. The $\delta$ functions in the $V_{n_1n_2n_3}$ are automatically satisfied and hence do not contribute.
\begin{figure}[t]
\centering
\includegraphics[width=0.3\textwidth]{ribbon.pdf}
\caption{An illustration of self-twist. The left panel is a ribbon without self-twists. The dotted line is the illustration of framing for the solid line. The right panel is a ribbon with one self-twist. The dotted line, the framing, winds around the solid line once\cite{witten1989}. We can simply view the right panel as a link of the solid and dotted line.}
\label{fig.ribbon}
\end{figure}
Restricting to the $\mathbb{Z}_2^{\otimes 3}$ situation where $N=1,p=1$, we actually have $\Theta(V_{111}) = -\mathrm{i}$\cite{propitius1995topological}. In terms of quantum double notations in Sec.~\ref{subsec: fusion rules 2+1}, $\Theta((111,\gamma_{+}))=-\mathrm{i}$. Moreover, when the Wilson line contributing a minus sign in the path integral is attached to $V_{111}$, the topological spin obtains one more minus sign. Hence, we find that $\Theta((111,\gamma_{-}))=\mathrm{i}$\cite{propitius1995topological}.
In App.~\ref{app. QD}, we provide a quantum double calculation and calculate the projective representations determined by the slant product of type III cocycles with group $\mathbb{Z}_N^{\otimes 3}$. And we show that it gives the same correlation functions etc as by the field theoretical approach above.
In summary for this section, we calculated the linking correlation functions of flux insertion operators explicitly, Eq~\eqref{eq. ZN linking}, with an assumption $\mathrm{gcd}(N,p)=1$. And as a consequence, we obtain the topological spins for the flux insertion operators, Eq~\eqref{eq. spin ZN}.
\section{Conclusion}
In this work we considered a continuum formulation of abelian Dijkgraaf-Witten field theories in (2+1)D. These theories come in three varieties: types I, II and III. We constructed all the possible gauge invariant line operators, which correspond to the possible quasi-particle excitations. The quasi-particles of type I and type II theories are readily understood using a $K$-matrix Chern-Simons theory approach. We mostly consider the subtler issue of type III DW models focussing on type III twisted $\mathbb{Z}_2^{\otimes 3}$ and more generally $\mathbb{Z}_N^{\otimes 3}$ DW theory. Despite the fact these theories have abelian gauge groups, their excitations have non-abelian fusions and statistics. We demonstrated this by directly constructing all Wilson and flux insertion operators, and computing all of their associated braiding and fusion rules. The guiding principle in constructing these operators is gauge invariance which, once imposed, leads to the introduction of auxiliary fields which live on the line operators in question. These auxiliary fields can be viewed as internal degrees of freedom of the particle in question.
Our work thus provides a field-theoretical platform for analyzing non-abelian (2+1)D SPTs and topological orders. It would be useful to extend some of the constructions here to higher dimensions, where topological phases are less well understood\cite{moradi2015,wang2015non,juvenwang2016quantum,wang2014,wang2015,lin2015,wang2016,jiang2014,wan2015twisted}.
\acknowledgements
HH and YZ thank B. A. Bernevig, E. Witten for useful comments and discussions.
CvK acknowledges the support of the Princeton Center for Theoretical Science.
HH and YZ acknowledge the support from Department of Physics, Princeton University.
|
\section{Introduction}
Semantic understanding of visual scenes is one of the holy grails of computer vision. The emergence of large-scale image datasets like ImageNet \cite{ILSVRC15}, COCO \cite{lin2014microsoft} and Places \cite{zhou2014learning}, along with the rapid development of the deep convolutional neural network (ConvNet) approaches, have brought great advancements to visual scene understanding. Nowadays, given a visual scene of a living room, a robot equipped with a trained ConvNet can accurately predict the scene category. However, to freely navigate in the scene and manipulate the objects inside, the robot has far more information to digest. It needs to recognize and localize not only the objects like sofa, table, and TV, but also their parts, e.g., a seat of a chair or a handle of a cup, to allow proper interaction, as well as to segment the stuff like floor, wall and ceiling for spatial navigation.
Scene parsing, or recognizing and segmenting objects and stuff in an image, remains one of the key problems in scene understanding. Going beyond the image-level recognition, scene parsing requires a much denser annotation of scenes with a large set of objects. However, the current datasets have limited number of objects (e.g., COCO~\cite{lin2014microsoft}, Pascal~\cite{everingham2010pascal}) and in many cases those objects are not the most common objects one encounters in the world (like frisbees or baseball bats), or the datasets only cover a limited set of scenes (e.g., Cityscapes~\cite{Cordts2015Cvprw}). Some notable exceptions are Pascal-Context~\cite{mottaghi2014role} and the SUN database~\cite{xiao2010sun}. However, Pascal-Context still contains scenes primarily focused on 20 object classes, while SUN has noisy labels at the object level.
Our goal is to collect a dataset that has densely annotated images (every pixel has a semantic label) with a large and an unrestricted open vocabulary. The images in our dataset are manually segmented in great detail, covering a diverse set of scenes, object and object part categories. The challenge for collecting such annotations is finding reliable annotators, as well as the fact that labeling is difficult if the class list is not defined in advance. On the other hand, open vocabulary naming also suffers from naming inconsistencies across different annotators. In contrast, our dataset was annotated by a single expert annotator, providing extremely detailed and exhaustive image annotations. On average, our annotator labeled 29 annotation segments per image, compared to the 16 segments per image labeled by external annotators (like workers from Amazon Mechanical Turk). Furthermore, the data consistency and quality are much higher than that of external annotators. Fig.~\ref{fig:teaser} shows examples from our dataset.
The paper is organized as follows. Firstly we describe the ADE20K dataset, the collection process and statistics. We then introduce a generic network design called Cascade Segmentation Module, which enables neural networks to segment stuff, objects, and object parts in cascade. Several semantic segmentation networks are evaluated on the scene parsing benchmark of ADE20K as baselines. The proposed Cascade Segmentation Module is shown to improve over those baselines. We further apply the scene parsing networks to the tasks of hierarchical semantic segmentation and automatic scene content removal.
\subsection{Related work}
Many datasets have been collected for the purpose of semantic understanding of scenes. We review the datasets according to the level of details of their annotations, then briefly go through the previous work of semantic segmentation networks.
{\bf Object classification/detection datasets.} Most of the large-scale datasets typically only contain labels at the image level or provide bounding boxes. Examples include Imagenet~\cite{ILSVRC15}, Pascal~\cite{everingham2010pascal}, and KITTI~\cite{Geiger2012CVPR}. Imagenet has the largest set of classes, but contains relatively simple scenes. Pascal and KITTI are more challenging and have more objects per image, however, their classes as well as scenes are more constrained.
{\bf Semantic segmentation datasets.} Existing datasets with pixel-level labels typically provide annotations only for a subset of foreground objects (20 in PASCAL VOC~\cite{everingham2010pascal} and 91 in Microsoft COCO~\cite{lin2014microsoft}). Collecting dense annotations where all pixels are labeled is much more challenging. Such efforts include Pascal-Context~\cite{mottaghi2014role}, NYU Depth V2~\cite{Silberman:ECCV12}, SUN database~\cite{xiao2010sun}, SUN RGB-D dataset~\cite{song2015sun}, CityScapes dataset~\cite{Cordts2015Cvprw}, and OpenSurfaces~\cite{bell13opensurfaces,bell2015material}.
{\bf Datasets with objects, parts and attributes.} Recently, two datasets were released that go beyond the typical labeling setup by also providing pixel-level annotation for the object parts, i.e. Pascal-Part dataset~\cite{chen2014parts}, or material classes, i.e. OpenSurfaces~\cite{bell13opensurfaces,bell2015material}. We advance this effort by collecting very high-resolution imagery of a much wider selection of scenes, containing a large set of object classes per image. We annotated both stuff and object classes, for which we additionally annotated their parts, and parts of these parts. We believe that our dataset, ADE20K, is one of the most comprehensive datasets of its kind. We provide a comparison between datasets in Sec.~\ref{sec:datasetcomparison}.
{\bf Semantic segmentation networks.} With the success of convolutional neural networks (CNN) for image classification \cite{krizhevsky2012imagenet}, there is growing interest for semantic pixel-wise labeling using CNNs with dense output, such as the fully CNN \cite{long2015fully}, deconvolutional neural networks \cite{noh2015learning}, encoder-decoder SegNet \cite{badrinarayanan2015segnet}, multi-task network cascades \cite{dai2015instance}, and DilatedNet \cite{CP2016Deeplab, YuKoltun2016}. They are benchmarked on Pascal dataset with impressive performance on segmenting the 20 object classes. Some of them \cite{long2015fully,badrinarayanan2015segnet} are evaluated on Pascal-Context~\cite{mottaghi2014role} or SUN RGB-D dataset~\cite{song2015sun} to show the capability to segment more object classes in scenes. Joint stuff and object segmentation is explored in~\cite{dai2015convolutional} which uses pre-computed superpixels and feature masking to represent stuff. Cascade of instance segmentation and categorization has been explored in \cite{dai2015instance}. In this paper we introduce a generic network module to segment stuff, objects, and object parts jointly in a cascade, which could be integrated in existing networks.
\section{ADE20K: Fully Annotated Image Dataset}
In this section, we describe our ADE20K dataset and analyze it through a variety of informative statistics.
\begin{figure*}
\centering
{\bf a)}\includegraphics[width=.695\textwidth]{figures/labelme_122.pdf}\hspace{1mm}
{\bf b)}\includegraphics[width=0.212\textwidth]{figures/partsTress.jpg}
\caption{a) Annotation interface, the list of the objects and their associated parts in the image. b) Section of the relation tree of objects and parts for the dataset (see the dataset webpage for the full relation tree)}
\label{fig:labelme}
\end{figure*}
\subsection{Dataset summary}
There are 20,210 images in the training set, 2,000 images in the validation set, and 3,000 images in the testing set. All the images are exhaustively annotated with objects. Many objects are also annotated with their parts. For each object there is additional information about whether it is occluded or cropped, and other attributes. The images in the validation set are exhaustively annotated with parts, while the part annotations are not exhaustive over the images in the training set. The annotations in the dataset are still growing. Sample images and annotations from the ADE20K dataset are shown in Fig.~\ref{fig:teaser}.
\subsection{Image annotation}
For our dataset, we are interested in having a diverse set of scenes with dense annotations of all the objects present. Images come from the LabelMe~\cite{russell2008labelme}, SUN datasets~\cite{xiao2010sun}, and Places~\cite{zhou2014learning} and were selected to cover the 900 scene categories defined in the SUN database. Images were annotated by a single expert worker using the LabelMe interface~\cite{russell2008labelme}. Fig.~\ref{fig:labelme}.a shows a snapshot of the annotation interface and one fully segmented image. The worker provided three types of annotations: object segments with names, object parts, and attributes. All object instances are segmented independently so that the dataset could be used to train and evaluate detection or segmentation algorithms. Datasets such as COCO~\cite{lin2014microsoft}, Pascal~\cite{everingham2010pascal} or Cityscape~\cite{Cordts2015Cvprw} start by defining a set of object categories of interest. However, when labeling all the objects in a scene, working with a predefined list of objects is not possible as new categories appear frequently (see fig.~\ref{fig:stats2}.d). Here, the annotator created a dictionary of visual concepts where new classes were added constantly to ensure consistency in object naming.
Object parts are associated with object instances. Note that parts can have parts too, and we label these associations as well. For example, the `rim' is a part of a `wheel', which in turn is part of a `car'. A `knob' is a part of a `door' that can be part of a `cabinet'. This part hierarchy in Fig.~\ref{fig:labelme}.b has a depth of 3.
\subsection{Annotation consistency}
Defining a labeling protocol is relatively easy when the labeling task is restricted to a fixed list of object classes, however it becomes challenging when the class list is open-ended. As the goal is to label all the objects within each image, the list of classes grows unbounded. Many object classes appear only a few times across the entire collection of images. However, those rare object classes cannot be ignored as they might be important elements for the interpretation of the scene. Labeling in these conditions becomes difficult because we need to keep a growing list of all the object classes in order to have a consistent naming across the entire dataset. Despite the annotator's best effort, the process is not free of noise.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{figures/consistency_fig_new}
\caption{Analysis of annotation consistency. Each column shows an image and two segmentations done by the same annotator at different times. Bottom row shows the pixel discrepancy when the two segmentations are subtracted, while the number at the bottom shows the percentage of pixels with the same label. On average across all re-annotated images, $82.4\%$ of pixels got the same label. In the example in the first column the percentage of pixels with the same label is relatively low because the annotator labeled the same region as `snow' and `ground' during the two rounds of annotation. In the third column, there were many objects in the scene and the annotator missed some between the two segmentations.}
\label{fig:consistency}
\end{figure*}
To analyze the annotation consistency we took a subset of 61 randomly chosen images from the validation set, then asked our annotator to annotate them again (there is a time difference of six months). One expects that there are some differences between the two annotations. A few examples are shown in Fig~\ref{fig:consistency}. On average, $82.4\%$ of the pixels got the same label. The remaining 17.6\% of pixels had some errors for which we grouped into three error types as follows:
\begin{itemize}
\item {\bf Segmentation quality}: Variations in the quality of segmentation and outlining of the object boundary. One typical source of error arises when segmenting complex objects such as buildings and trees, which can be segmented with different degrees of precision. 5.7\% of the pixels had this type of error.
\item {\bf Object naming}: Differences in object naming (due to ambiguity or similarity between concepts, for instance, calling a big car a `car' in one segmentation and a `truck' in the another one, or a `palm tree' a `tree'. 6.0\% of the pixels had naming issues. These errors can be reduced by defining a very precise terminology, but this becomes much harder with a large growing vocabulary.
\item {\bf Segmentation quantity}: Missing objects in one of the two segmentations. There is a very large number of objects in each image and some images might be annotated more thoroughly than others. For example, in the third column of Fig~\ref{fig:consistency} the annotator missed some small objects in different annotations. 5.9\% of the pixels are due to missing labels. A similar issue existed in segmentation datasets such as the Berkeley Image segmentation dataset \cite{MartinFTM01}.
\end{itemize}
The median error values for the three error types are: 4.8\%, 0.3\% and 2.6\% showing that the mean value is dominated by a few images, and that the most common type of error is segmentation quality.
To further compare the annotation done by our single expert annotator and the AMT-like annotators, 20 images from the validation set are annotated by two invited external annotators, both with prior experience in image labeling. The first external annotator had 58.5\% of inconsistent pixels compared to the segmentation provided by our annotator, and the second external annotator had 75\% of the inconsistent pixels. Many of these inconsistencies are due to the poor quality of the segmentations provided by external annotators (as it has been observed with AMT which requires multiple verification steps for quality control~\cite{lin2014microsoft}). For the best external annotator (the first one), 7.9\% of pixels have inconsistent segmentations (just slightly worse than our annotator), 14.9\% have inconsistent object naming and 35.8\% of the pixels correspond to missing objects, which is due to the much smaller number of objects annotated by the external annotator in comparison with the ones annotated by our expert annotator. The external annotators labeled on average 16 segments per image while our annotator provided 29 segments per image.
\subsection{Dataset statistics}
Fig.~\ref{fig:stats}.a shows the distribution of ranked object frequencies. The distribution is similar to a Zipf's law and is typically found when objects are exhaustively annotated in images \cite{spainPerona10,xiao2010sun}. They differ from the ones from datasets such as COCO or ImageNet where the distribution is more uniform resulting from manual balancing.
Fig.~\ref{fig:stats}.b shows the distributions of annotated parts grouped by the objects they belong to and sorted by frequency within each object class. Most object classes also have a non-uniform distribution of part counts. Fig.~\ref{fig:stats}.c and Fig.~\ref{fig:stats}.d show how objects are shared across scenes and how parts are shared by objects. Fig.~\ref{fig:stats}.e shows the variability in the appearances of the part `door'.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{figures/histograms_new.pdf}\\[1mm]
\includegraphics[width=1\textwidth]{figures/shareability.pdf}
\caption{a) Object classes sorted by frequency. Only the top 270 classes with more than 100 annotated instances are shown. 68 classes have more than a 1000 segmented instances. b) Frequency of parts grouped by objects. There are more than 200 object classes with annotated parts. Only objects with 5 or more parts are shown in this plot (we show at most 7 parts for each object class). c) Objects ranked by the number of scenes they are part of. d) Object parts ranked by the number of objects they are part of. e) Examples of objects with doors. The bottom-right image is an example where the door does not behave as a part. }
\label{fig:stats}
\vspace{3mm}
\end{figure*}
The mode of the object segmentations is shown in Fig.~\ref{fig:stats2}.a and contains the four objects (from top to bottom): `sky', `wall', `building' and `floor'. When using simply the mode to segment the images, it gets, on average, 20.9$\%$ of the pixels of each image right. Fig.~\ref{fig:stats2}.b shows the distribution of images according to the number of distinct classes and instances. On average there are 19.5 instances and 10.5 object classes per image, larger than other existing datasets (see Table~\ref{table:headings}). Fig.~\ref{fig:stats2}.c shows the distribution of parts.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{figures/stats_new}
\vspace{-3mm}
\caption{a) Mode of the object segmentations contains `sky', `wall', `building' and `floor'. b) Histogram of the number of segmented object instances and classes per image. c) Histogram of the number of segmented part instances and classes per object. d) Number of classes as a function of segmented instances (objects and parts). The squares represent the current state of the dataset. e) Probability of seeing a new object (or part) class as a function of the number of instances.}
\label{fig:stats2}
\end{figure*}
As the list of object classes is not predefined, there are new classes appearing over time of annotation. Fig.~\ref{fig:stats2}.d shows the number of object (and part) classes as the number of annotated instances increases. Fig.~\ref{fig:stats2}.e shows the probability that instance $n+1$ is a new class after labeling $n$ instances. The more segments we have, the smaller the probability that we will see a new class. At the current state of the dataset, we get one new object class every 300 segmented instances.
\subsection{Comparison with other datasets}
\label{sec:datasetcomparison}
We compare ADE20K with existing datasets in Tab.~\ref{table:headings}. Compared to the largest annotated datasets, COCO~\cite{lin2014microsoft} and Imagenet~\cite{ILSVRC15}, our dataset comprises of much more diverse scenes, where the average number of object classes per image is 3 and 6 times larger, respectively. With respect to SUN~\cite{xiao2010sun}, ADE20K is roughly 35\% larger in terms of images and object instances. However, the annotations in our dataset are much richer since they also include segmentation at the part level. Such annotation is only available for the Pascal-Context/Part dataset~\cite{mottaghi2014role,chen2014parts} which contains 40 distinct part classes across 20 object classes. Note that we merged some of their part classes to be consistent with our labeling (e.g., we mark both \emph{left leg} and \emph{right leg} as the same semantic part \emph{leg}). Since our dataset contains part annotations for a much wider set of object classes, the number of part classes is almost 9 times larger in our dataset.
An interesting fact is that any image in ADE20K contains at least 5 objects, and the maximum number of object instances per image reaches 273, and 419 instances, when counting parts as well. This shows the high annotation complexity of our dataset.
\begin{table*}[t!]
\begin{center}
\caption{Comparison with existing datasets with semantic segmentation.}
\label{table:headings}
\begin{tabular}{lcccccc}
\hline\noalign{\smallskip}
& Images & Obj. inst. & Obj. classes & Part inst. & Part classes & Obj. classes per image \\
\noalign{\smallskip}
\hline
\noalign{\smallskip}
COCO & 123,287 & 886,284 & 91 & 0 & 0 & 3.5\\
ImageNet$^*$ & 476,688 & 534,309 & 200 & 0 & 0 & 1.7\\
NYU Depth V2 & 1,449 & 34,064 & 894 & 0 & 0 & 14.1 \\
Cityscapes & 25,000 & 65,385 & 30 & 0 & 0 & 12.2 \\
SUN & 16,873 & 313,884 & 4,479 & 0 & 0 & 9.8\\
OpenSurfaces & 22,214 & 71,460 & 160 & 0 & 0 & N/A\\
PascalContext & 10,103 & $\sim$104,398$^{**}$ & 540 & 181,770 & 40 & 5.1\\
ADE20K & 22,210 & 434,826 & 2,693 & 175,961 & 476 & 9.9\\
\hline
\end{tabular}
\end{center}
\vspace{-3mm}
\scriptsize $^*$ has only bounding boxes (no pixel-level segmentation). Sparse annotations.\\
\scriptsize $^{**}$ PascalContext dataset does not have instance segmentation. In order to estimate the number of instances, we find connected components (having at least 150pixels) for each class label.
\end{table*}
\setlength{\tabcolsep}{1.4pt}
\section{Cascade Segmentation Module}
\begin{figure*}
\begin{center}
\vspace{-3mm}
\includegraphics[width=0.9\textwidth]{figures/framework_new3.pdf}
\end{center}
\vspace{-7pt}
\caption{The framework of Cascade Segmentation Module for scene parsing. \emph{Stuff} stream generates the stuff segmentation and objectness map from the shared feature activation. The \emph{object} stream then generates object segmentation by integrating the objectness map from the stuff stream. Finally the full scene segmentation is generated by merging the object segmentation and the stuff segmentation. Similarly, \emph{part} stream takes object score map from object stream to further generate object-part segmentation. Since not all objects have part annotation, the part stream is optional. Feature sizes are based on the Cascade-dilatedNet, the Cascade-SegNet has different but similar structures.}
\label{framework}
\end{figure*}
While the frequency of objects appearing in scenes follows a long-tail distribution, the pixel ratios of objects also follow such a distribution. For example, the stuff classes like `wall', `building', `floor', and `sky' occupy more than 40\% of all the annotated pixels, while the discrete objects, such as `vase' and `microwave' at the tail of the distribution (see Fig. \ref{fig:stats}b), occupy only 0.03\% of the annotated pixels. Because of the long-tail distribution, a semantic segmentation network can be easily dominated by the most frequent stuff classes.
On the other hand, there are spatial layout relations among stuff and objects, and the objects and the object parts, which are ignored by the design of the previous semantic segmentation networks. For example, a drawing on a wall is a part of the wall (the drawing occludes the wall), and the wheels on a car are also parts of the car.
To handle the long-tail distribution of objects in scenes and the spatial layout relations of scenes, objects, and object parts, we propose a network design called Cascade Segmentation Module. This module is a generic network design which can potentially be integrated in any previous semantic segmentation networks. We first categorize semantic classes of the scenes into three macro classes: \emph{stuff} (sky, road, building, etc), foreground \emph{objects} (car, tree, sofa, etc), and object \emph{parts} (car wheels and door, people head and torso, etc). Note that in some scenarios there are some object classes like `building' or `door' could belong to either of two macro classes, here we assign the object classes to their most likely macro class.
In the network for scene parsing, different streams of high-level layers are used to represent each macro class and recognize the assigned classes. The segmentation results from each stream are then fused to generate the segmentation. The proposed module is illustrated in Fig.~\ref{framework}.
More specifically, the stuff stream is trained to classify all the stuff classes plus one foreground object class (which includes all the non-stuff classes). After training, the stuff stream generates stuff segmentation and a dense objectness map indicating the probability that a pixel belongs to the foreground object class. The object stream is trained to classify the discrete objects. All the non-discrete objects are ignored in the training loss function of the object stream. After training, the object stream further segments each discrete object on the predicted objectness map from the stuff stream. The result is merged with the stuff segmentation to generate the scene segmentation. For those discrete objects annotated with parts, the part stream can be jointly trained to segment object parts. Thus the part stream further segments parts on each object score map predicted from the object stream.
The network with the two streams (stuff+objects) or three streams (stuff+objects+parts) could be trained end-to-end. The streams share the weights of the lower layers. Each stream has a training loss at the end. For the stuff stream we use the per-pixel cross-entropy loss, where the output classes are all the stuff classes plus the foreground class (all the discrete object classes are included in it). We use the per-pixel cross-entropy loss for the object stream, where the output classes are all the discrete object classes. The objectness map is given as a ground-truth binary mask that indicates whether a pixel belongs to any of the stuff classes or the foreground object class. This mask is used to exclude the penalty for pixels which belong to any of the stuff classes in the training loss for the object stream. Similarly, we use cross-entropy loss for the part stream. All part classes are trained together, while non-part pixels are ignored in training. In testing, parts are segmented on their associated object score map given by the object stream. The training losses for the two streams and for the three streams are $L= L_{stuff} + L_{object}$ and $L= L_{stuff} + L_{object} + L_{part}$ respectively.
The configurations of each layer are based on the baseline network being used. We integrate the proposed module on two baseline networks Segnet \cite{badrinarayanan2015segnet} and DilatedNet \cite{CP2016Deeplab,YuKoltun2016}. In the following experiments, we evaluate that the proposed module brings great improvements for scene parsing.
\section{Experiments}
To train the networks for scene parsing, we select the top 150 objects ranked by their total pixel ratios\footnote{As the original images in the ADE20K dataset have various sizes, for simplicity we rescale those large-sized images to make their minimum heights or widths as 512 in the SceneParse150 benchmark.} from the ADE20K dataset and build a scene parsing benchmark of ADE20K, termed as SceneParse150. Among the 150 objects, there are 35 stuff classes (i.e., wall, sky, road) and 115 discrete objects (i.e., car, person, table). The annotated pixels of the 150 objects occupy 92.75\% of all the pixels in the dataset, where the stuff classes occupy 60.92\%, and discrete objects occupy 31.83\%.
We map the wordnet synsets with each one of the object names, then build up a wordnet tree through the hypernym relations of the 150 objects shown in Fig.~\ref{wordnet_tree}. We can see that these objects form several semantic clusters in the tree, such as the \textit{furniture} synset node containing cabinet, desk, pool table, and bench, the \textit{conveyance} node containing car, truck, boat, and bus, as well as the \textit{living thing} node containing shrub, grass, flower, and person. Thus, the structured object annotation given in the dataset bridge the image annotation to a wider knowledge base.
\begin{figure*}
\begin{center}
\includegraphics[width=1\textwidth]{figures/semantic_tree150.pdf}
\end{center}
\vspace{-5mm}
\caption{Wordnet tree constructed from the 150 objects in the SceneParse150 benchmark. Clusters inside the wordnet tree represent various hierarchical semantic relations among objects.}
\label{wordnet_tree}
\end{figure*}
\subsection{Scene parsing}
As for baselines of scene parsing on SceneParse150 benchmark, we train three semantic segmentation networks: SegNet~\cite{badrinarayanan2015segnet}, FCN-8s~\cite{long2015fully}, and DilatedNet~\cite{CP2016Deeplab, YuKoltun2016}. SegNet has encoder and decoder architecture for image segmentation; FCN upsamples the activations of multiple layers in the CNN for pixelwise segmentation; DilatedNet drops \textit{pool4} and \textit{pool5} from fully convolutional VGG-16 network, and replaces the following convolutions with dilated convolutions (or atrous convolutions).
We integrate the proposed cascade segmentation module on the two baseline networks: SegNet and DilatedNet. We did not integrate it with FCN because the original FCN requires a large amount of GPU memory and has skip connections across layers. For the Cascade-SegNet, two streams share a single encoder, from \texttt{conv1\_1} to \texttt{conv5\_3}, while each stream has its own decoder, from \texttt{deconv5\_3} to \texttt{loss}. For the Cascade-DilatedNet, the two streams split after \texttt{pool3}, and keep spatial dimensions of their feature maps afterwards. For a fair comparison and benchmark purposes, the cascade networks only have stuff stream and object stream. We train these network models using the Caffe library~\cite{jia2014caffe} on NVIDIA Titan X GPUs.
Results are reported in four metrics commonly used for semantic segmentation~\cite{long2015fully}:
\begin{itemize}
\item \textbf{Pixel accuracy} indicates the proportion of correctly classified pixels;
\item \textbf{Mean accuracy} indicates the proportion of correctly classified pixels averaged over all the classes.
\item \textbf{Mean IoU} indicates the intersection-over-union between the predicted and ground-truth pixels, averaged over all the classes.
\item \textbf{Weighted IoU} indicates the IoU weighted by the total pixel ratio of each class.
\end{itemize}
Since some classes like `wall' and `floor' occupy far more pixels of the images, pixel accuracy is biased to reflect the accuracy over those few large classes. Instead, mean IoU reflects how accurately the model classifies each discrete class in the benchmark. The scene parsing data and the development toolbox are released in the Scene Parsing Challenge hosted at ILSVRC'16\footnote{\url{http://sceneparsing.csail.mit.edu}}. We take the average of the pixel accuracy and mean IoU as the evaluation criteria in the challenge.
\begin{table}
\caption{Performance on the validation set of SceneParse150.}
\label{performance_segmentation}
\centering
\footnotesize
\addtolength{\tabcolsep}{0.8pt}
\begin{tabular}{| l | c | c | c | c |}
\hline
Networks &Pixel Acc. & Mean Acc. & Mean IoU & Weighted IoU\\
\hline
FCN-8s & 71.32\% & 40.32\% & 0.2939 & 0.5733 \\
SegNet & 71.00\% & 31.14\% & 0.2164 & 0.5384 \\
DilatedNet & 73.55\% & 44.59\% & 0.3231 & 0.6014 \\
\hline
Cascade-SegNet & 71.83\% & 37.90\% & 0.2751 & 0.5805 \\
Cascade-DilatedNet & \textbf{74.52}\% & \textbf{45.38}\% & \textbf{0.3490} & \textbf{0.6108} \\
\hline
\end{tabular}
\end{table}
The segmentation results of the baselines and the cascade networks are listed in Table~\ref{performance_segmentation}. Among the baselines, the DilatedNet achieves the best performance on the SceneParse150. The cascade networks, Cascade-SegNet and Cascade-DilatedNet both outperform the original baselines. In terms of mean IoU, the improvement brought by the proposed cascade segmentation module for SegNet is 6\%, and for DilatedNet is 2.5\%. We further decompose the performance of networks on 35 stuff and 115 discrete object classes respectively, in Table~\ref{performance_object115stuff35}. We observe that the two cascade networks perform much better on the 115 discrete objects compared to the baselines. This validates that the design of cascade module helps improve scene parsing for the discrete objects as they have less training data but more visual complexity compared to those stuff classes.
Segmentation examples from the validation set are shown in Fig.~\ref{segmentation-result}. Compared to the baseline SegNet and DilatedNet, the segmentation results from the Cascade-SegNet and Cascade-DilatedNet are more detailed. Furthermore, the objectness maps from the stuff stream highlight the possible discrete objects in the scenes.
\begin{table}
\caption{Performance of stuff and discrete object segmentation.}
\label{performance_object115stuff35}
\footnotesize
\centering
\addtolength{\tabcolsep}{1.7pt}
\begin{tabular}{| l | c | c | c | c |}
\hline
& \multicolumn{2}{|c|}{35 stuff} & \multicolumn{2}{|c|}{115 discrete objects} \\
\hline
Networks & Mean Acc. & Mean IoU & Mean Acc. & Mean IoU \\
\hline
FCN-8s & 46.74\% & 0.3344 & 38.36\% & 0.2816 \\
SegNet & 43.17\% & 0.3051 & 27.48\% & 0.1894 \\
DilatedNet & 49.03\% & 0.3729 & 43.24\% & 0.3080 \\
\hline
Cascade-SegNet & 40.46\% & 0.3245 & 37.12\% & 0.2600 \\
Cascade-DilatedNet & \textbf{49.80}\% & \textbf{0.3779} & \textbf{44.04}\% & \textbf{0.3401} \\
\hline
\end{tabular}
\end{table}
\begin{figure*}
\begin{center}
\includegraphics[width=1.0\textwidth]{figures/segmentation_result_new3.pdf}
\end{center}
\caption{Ground-truths, segmentation results given by the baselines and the cascade networks, and the objectness map and stuff segmentation given by the Cascade-DilatedNet.}
\label{segmentation-result}
\end{figure*}
\subsection{Part segmentation}
For part segmentation, we select the eight object classes frequently annotated with parts: `person', `building', `car', `chair',`table', `sofa', `bed', `lamp'. After we filter out the part classes of those objects with instance number smaller than 300, there are 36 part classes included in the training and testing. We train the part stream on the Cascade-DilatedNet used in the scene parsing.
The results of joint segmentation for stuff, objects, and object parts are shown in Fig.~\ref{part_example}. In a single forward pass the network with the proposed cascade module is able to parse scenes at different levels. We use the accuracy instead of the IoU as the metric to measure the part segmentation results, as the parts of object instances in the dataset are not fully annotated. The accuracy for all the parts of the eight objects is plotted in Fig.\ref{part_acc_all}. The average accuracy is 55.47\%.
\begin{figure*}
\begin{center}
\includegraphics[width=1.0\textwidth]{figures/part_result2.pdf}
\end{center}
\vspace{-6mm}
\caption{Part segmentation results. The middle row is the object score map predicted by the object stream for each object. The part stream further segments the object score map into different parts. }
\label{part_example}
\vspace{-2mm}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=0.45\textwidth]{figures/part_acc_plot.pdf}
\end{center}
\vspace{-3mm}
\caption{The part segmentation accuracy grouped by the objects.}
\label{part_acc_all}
\vspace{-2mm}
\end{figure}
\subsection{Further applications}
Accurate scene parsing leads to wider applications. Here we take the hierarchical semantic segmentation and the automatic scene content removal as exemplar applications of the scene parsing networks.
\textbf{Hierarchical semantic segmentation.} Given the wordnet tree constructed on the object annotation shown in Fig.\ref{wordnet_tree}, the 150 objects are hierarchically connected and have hyponyms relations. Thus we could gradually merge the objects into their hyponyms so that objects with similar semantics are merged at the early levels. Through this way, we generated a hierarchical semantic segmentation of the image shown in Fig.~\ref{The hierarchical_segmentation_example}. The tree also provides a principled way to segment more general visual concepts. For example, to detect all furniture in a scene, we can simply merge the hyponyms associated with each synset, such as the chair, table, bench, and bookcase.
\textbf{Automatic image content removal}. Image content removal methods typically require the users to annotate the precise boundary of the target objects to be removed. Here, based on the predicted object probability map from Cascade-SegNet, we automatically identify the image region of the target objects. After cropping out the target objects using the predicted object probability map, we simply use image completion/inpainting methods to fill the holes in the image. Fig.~\ref{image_completion} shows some examples of the automatic image content removal. It can be seen that with the predicted object score maps, we are able to crop out the objects from the image in a precise way. We used the image completion technique described in ~\cite{huang2014image}.
\begin{figure}
\begin{center}
\vspace{-3mm}
\includegraphics[width=0.48\textwidth]{figures/hierarchical_new2.pdf}
\end{center}
\vspace{-4mm}
\caption{The examples of the hierarchical semantic segmentation. Objects with similar semantics like furnitures and vegetations are merged at early levels following the wordnet tree.}
\label{The hierarchical_segmentation_example}
\vspace{-3mm}
\end{figure}
\begin{figure}
\includegraphics[width=0.48\textwidth]{figures/imageCompletion.pdf}
\caption{\small Automatic image content removal using the predicted object score maps given by the scene parsing network. We are not only able to remove individual objects such as person, tree, car, but also groups of them or even all the discrete objects. For each row, the first image is the original image, the second is the object score map, and the third one is the filled-in image.}
\vspace{-3mm}
\label{image_completion}
\end{figure}
\section{Conclusion}
In this paper, we introduced a new densely annotated dataset with the instances of stuff, objects, and parts, covering a diverse set of visual concepts in scenes. The dataset was carefully annotated by a single annotator to ensure precise object boundaries within the image and the consistency of object naming across the images. A generic network design called Cascade Segmentation Module was proposed for scene parsing. It enables the convolutional neural networks to parse scenes into stuff, objects, and object parts in cascade with the state-of-the-art performance.
{\small
\bibliographystyle{ieee}
|
\section{Introduction}
Modelling of the distribution and kinematics of the neutral hydrogen in
galaxies beyond our own, and the most nearby systems M\,31 and M\,33, took
off in the 1970s, with the addition of spectroscopic capability to the
Westerbork Synthesis Radio Telescope \citep{Allen1974A}.
This new tool gave observers sufficient spatial and velocity resolution
to resolve the \textsc{Hi} structure and kinematics for a large number of galaxies.
In edge-on galaxies, it thus became possible to trace the outer envelope
of the position-velocity (XV) diagram in order to get the rotation curve
of the galaxy \citep{Sancisi1979A}.
Various strategies for this exist, such as fitting for the peak intensity
or fitting a one-dimensional Gaussian with fixed
velocity dispersion to the outer edge of the XV-diagram
\citep[e.g.][]{Garcia-Ruiz2002A}.
We refer the reader to \citet{OBrien2010B} for a detailed treatment on the
various methods for deriving the rotation curve from a XV-diagram.
For face-on galaxies, fitting the structure and kinematics was first done by
modelling the zeroth and first-moments maps, based on the channel maps in the
\textsc{Hi} data cube.
These gave rise to the famous spider diagrams,
in which the iso-velocity contours showed the mean velocity of the gas at each
position in the galaxy \citep{Rogstad1971A}.
With the assumption of a thin disc, it is then possible to measure the
rotation of the gas in rings centred on the galaxy.
The position angle and inclination could be varied as function of the
rings. Assuming a rotation curve,
\citet{Rogstad1974A} used it to model the
peculiar velocity field of galaxy M\,83, and found the presence of a warp.
\citet{Bosma1978} improved this by minimizing line of sight
velocity residuals for different rotation velocity in order to fit
a rotation curve to the data.
This method proved quite successful at modelling galaxies. It has been
extended by many authors into many software packages
\citep{Begeman1987,Schoenmakers1999,Spekkens2007,Krajnovic2006}.
We refer to \citet{Jozsa2007} for a detailed discussion on tilted-ring
fitting to velocity fields.
The ever-continuing technological improvements led to ever-higher quality
observations, which led to ever-better resolved galaxies, both spatially as
well as in velocity.
A natural consequence of this was the attempt to model the \textsc{Hi} data cube itself.
This was first attempted by \citet{Irwin1991A,Irwin1993A}, who developed the
\textsc{Cubit} code\footnote{Which can be found at
www.astro.queensu.ca/$\sim $irwin/}.
\textsc{Cubit} offers the user parameterized functions, which could be
automatically fit to the galaxy.
The program has later been expanded to fit more properties of galaxies
\citep{Irwin1994A}.
\begin{figure*}
\includegraphics[width=0.32\textwidth]{Petersetal-IIfig1a.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIfig1b.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIfig1c.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIfig1d.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIfig1e.jpg}
\includegraphics[width=0.32\textwidth]{Petersetal-IIfig1f.jpg}
\caption[Illustrating the coordinate system generation]{A cartoon illustrating the generation of the galaxy coordinate system. The model demonstrate the generation of an edge-on galaxy at a position angle of $135^\circ$. The boxes on each side represent the outer edges of the voxels. From top-left to bottom-right: The initial coordinate system. Centering the coordinate system (Equation \ref{eqn:cartoon1}). The shift parameters can be used to move the central position of the galaxy along the horizontal and vertical axes (Equation \ref{eqn:cartoon2}). The position angle is applied by rotating along the $z$ axis (Equation \ref{eqn:cartoon3}). The inclination is applied by rotating along the $y$ axis (Equation \ref{eqn:cartoon4}). In the final step, we switch to cylindrical coordinates. }\label{fig:cartoon}
\end{figure*}
Another software package is \textsc{GALMOD},
designed by T.S.~van~Albada and F.J.~Sicking, and incorporated
into the \textsc{GIPSY} data reduction system, \citep{GIPSY}.
Based on a tilted-ring geometry, \textsc{galmod} randomly projects a very
large amount of `clouds' in the \textsc{Hi} data cube.
As these clouds follow the specified kinematics and distribution, after
sufficient samples have been drawn, the \textsc{Hi} data cube has been built and the
only task remaining is scaling the cube to the desired intensity.
\textsc{Galmod} does not support automatic fitting and users thus have to
fit a cube using the hand-and-eye strategy.
\textsc{TiRiFiC}\footnote{Currently available at
www.astron.nl/$\sim $jozsa/tirific/index.html} is another code,
which can automate
the fit to a cube \citep{Jozsa2007}.
It is based on a tilted-ring model, but is expanded compared to
\textsc{galmod}, to account for inhomogeneity in galaxy discs.
Recently, it has used to fit features resembling spiral arms in
edge-on galaxies \citep{Kamphuis2013A}.
Finally, we note the \textsc{GalAPAGOS} project, which can automatically
fit rotation curves in galaxies \citep{Wiegert2011}.
Out of all software codes currently publicly available, \textsc{TiRiFiC}
seems to be the only one under active development.
A fully automated three-dimensional
fitting routine \textsc{FAT} for \textsc{Fully Automated TiRiFiC}
\citep{Kamphuis2015} and another routine, called $^{\rm 3D}$BAROLO that also
estimates velocity dispersions \citep{Teodoro2015} has become available.
In Paper I \citep{Peters2015A},
we presented the \textsc{Hi} observations for eight edge-on galaxies.
One of the key conclusions of that paper was that self-absorption could
well play an important role.
The effect of \textsc{Hi} self-absorption has long been a topic of concern, for
example in the work by \citet{Sancisi1979A} and of \citet{Braun2009}.
The latter estimated quite significant corrections for self-absorption in the
case of the Andromeda Nebula M\,31.
Yet in most of the publications featuring models of neutral hydrogen in
galaxies, the assumption remains that of an optically thin gas.
None of the above software packages supports modelling of self-absorption
of the gas.
This, combined with some other research questions we had concerning the
neutral hydrogen, such as the behavior of the velocity dispersion as function
of height above the plane,
led us to develop a new tool for generating \textsc{Hi} data cubes that could treat for
self-absorption.
Rather than using a tilted-ring model, we model the galaxy as a single plane
on which harmonic offsets above or below the central plane represent the warps.
The code was developed in the \textsc{Python} and \textsc{C++} programming
languages,
with the computationally most intensive parts implemented as multi-threaded.
The software is called \textsc{Galactus}. It has been publicly released under
an open-source license and is listed in the Astrophysics Source Code Library
as \citet[ascl.net/1303.018,][]{GalactusASCL}.
The code can be downloaded at sourceforge.net/p/galactus/.
The purpose of this paper is twofold.
In Section \ref{sec:galactusdesign}, we present the inner workings of
\textsc{Galactus}.
Section \ref{sec:notes} discusses some important features of the tool
in more detail.
In Section \ref{sec:effectiveTspin}, we explore the effective spin
temperature in more detail.
We test the program on face-on galaxy NGC\,2403 in Section \ref{sec:NGC2403},
using both an optically thin and a self-absorbing model.
In Section \ref{sec:NGC2403edgeon}, the fits from NGC\,2403 are projected to
an edge-on galaxy to show that the maximum surface brightness profiles seen
in Paper I should indeed have been higher.
We test how self-absorption can lower the apparent total \textsc{Hi} mass, depending
on the inclination, in Section \ref{sec:effectofinclination}.
The Tully-Fisher relation
is shown to be independent of \textsc{Hi} self-absorption, in
Section \ref{sec:TullyFisher}.
\section{\textsc{Galactus} Design}\label{sec:galactusdesign}
In this section, we discuss how \textsc{Galactus} generates a \textsc{Hi} data cube.
This generation consists of two steps.
In the first phase, the program creates a 3D model of the galaxy, while the
second phase performs the radiative transfer.
To generate the model the software creates a three dimensional grid.
The first two axes are aligned with, and have the same lengths as, the
horizontal and vertical
axes of the \textsc{Hi} data cube channels\footnote{Note that we assume the size of the
pixel is not frequency dependent. This is not true by default for \textsc{Hi} data cubes
created using \textsc{Miriad}.}.
The third axis does not denote velocity, but represents physical depth.
The length of the third axis is the same as that of
the horizontal axis of the \textsc{Hi} data
cube\footnote{When modelling only one side of a galaxy, this length is
doubled.}.
Each position in this model is a three dimensional pixel, also known as a
volumetric pixel, or voxel for short.
These voxels all represent a small volume of the galaxy.
Since the voxels are perfectly aligned with the \textsc{Hi} data cube channels,
calculating the brightness of each pixel in a channel can be done by
performing the radiative transfer along the depth axis of the voxels
associated with that pixel.
Each voxel is assigned coordinates $x$, $y$ and $z$.
These are initially aligned with the horizontal, vertical and depth axis.
We will subsequently bend and twist these coordinates to match the position
and angle of the galaxy as seen in the cube.
We emphasise however that these are \emph{internal} coordinates, the grid of
voxels itself does not change and \emph{always} remains aligned with the
\textsc{Hi} data cube.
The internal coordinates will represent the coordinate system of the galaxy
in question.
\subsection{Generating the Galaxy Coordinate System}
We start with the coordinates $x,y,z$ assigned to the grid of
voxels\footnote{As an aid to understanding the coordinate system, we
illustrate the steps these equations perform in Figure \ref{fig:cartoon}.}.
We first shift the coordinates, such that position (0,0,0) aligns with the
centre of the model, i.e.
\begin{equation}
\begin{pmatrix}x\\y\\z\end{pmatrix} = \begin{pmatrix}x - \max(x)/2 \\ y - \max(y)/2\\ z - \max(z)/2\end{pmatrix}\,\,.\label{eqn:cartoon1}
\end{equation}
When performing a fit, the actual galaxy is often not perfectly aligned with
the centre of the \textsc{Hi} data cube.
It is thus necessary to shift the coordinates, based on parameters
$\mathrm{shift}_x$ and $\mathrm{shift}_y$,
\begin{equation}
\begin{pmatrix}x\\y\end{pmatrix} = \begin{pmatrix}x - \mathrm{shift}_x\\y - \mathrm{shift}_y\end{pmatrix}\,\,.\label{eqn:cartoon2}
\end{equation}
Next, the position angle $PA$ is applied. This is done using the coordinate
transforms
\begin{equation}
\begin{pmatrix}x\\y\end{pmatrix} = \begin{pmatrix} x \cos(PA) - y \sin(PA) \\ x \sin(PA) + y \cos(PA)\end{pmatrix}\,\,.\label{eqn:cartoon3}
\end{equation}
Followed by the inclination $i$,
\begin{equation}
\begin{pmatrix}x\\z\end{pmatrix} = \begin{pmatrix} x \cos(i) + z \sin(i) \\ -x \sin(i) + z \cos(i)\end{pmatrix}\,\,.\label{eqn:cartoon4}
\end{equation}
We can now convert to the final cylindrical coordinates system that will be
used for the galaxy,
\begin{equation}
\begin{pmatrix}R\\ \theta\\z\end{pmatrix} = \begin{pmatrix} \sqrt{x^2+y^2} \\ \arctan(y/x)\\ z \end{pmatrix}\,\,.\label{eqn:cartoon5}
\end{equation}
Warps are known to exist in many galaxies, yet the actual physics is not well
understood.
The traditional way of dealing with warps is by applying a tilted-ring model.
This is however impossible in our case, as the tilted-ring model does not map
all coordinates to a unique physical position.
Some voxels would then be mapped to multiple values of $R$ simultaneously,
which is undesirable.
We thus choose a different strategy to account for warps.
Following \citet{Binney1987A}, we define a warp as a harmonic oscillation
along the disc, in which the mid-plane $z=0$ is raised, or lowered, above the
initial mid-plane, as we go around the galaxy following $\theta$.
The warp itself has a maximum amplitude which is described by super-function
$z_\mathrm{warp}(R)$ (see Section \ref{sec:superfunctions}), which occurs at
pitch angle $\theta_\mathrm{warp}$.
We do not account for radial changes in pitch angle, so $\theta_\mathrm{warp}$
is a single value. Then
\begin{equation}
z = z + z_\mathrm{warp}(R) \times \cos(\theta + \theta_\mathrm{warp})\,\,.
\end{equation}
\subsection{Generating the Physics}
At this point, we have successfully generated the coordinates for the galaxy.
The next step is to calculate the actual physical properties for each voxel.
First, the line-of-sight velocity $v_\mathrm{obs}$ is calculated, based on
rotation velocity $v_\mathrm{rot}$ and systemic velocity $v_\mathrm{sys}$.
The rotation velocity $v_\mathrm{rot}$, along with the scale-height of the disc
$z_0(r)$, velocity dispersion $\sigma(R)$ and face-on surface density
$A_\textrm{\textsc{Hi} }(R)$ are super-functions, which are prepared by the user (see
Section \ref{sec:superfunctions}). Then
\begin{align}
v &= v_\mathrm{rot}(R) - L \times \!\mid \!z\!\mid \label{eqn:lagginghalo}\,\,\,\,\,\,\,(v\geq0),\\
v_\mathrm{obs} &= v \times \sin(\theta) \times \sin(i) + v_\mathrm{sys}\,\,.
\end{align}
Equation \ref{eqn:lagginghalo} offers the possibility to model lagging haloes.
The lag is expected to decrease linearly the observed rotation with value $L$
as the height $\mid \!z\! \mid$ above the plane increases.
The model automatically prevents $v<0$ and forces these to $v=0$.
When no halo lagging is modelled $L$ is set to zero.
The observed velocity dispersion $\sigma$ is calculated separately for the
three (cylindrical) directions $\sigma_{z,\mathrm{obs}}$, $\sigma[B_{R,\mathrm{obs}}$
and $\sigma_{\theta,\mathrm{obs}}$, and then combined into a single observed
velocity dispersion $\sigma_\mathrm{obs}$.
The $a_R, a_z, a_\theta$ components are scalars to account for a possible
anisotropic velocity-dispersion tensor. \label{sec:tensordispersion}
There is currently no known case of anisotropy, so we set $a_R, a_z, a_\theta$
to one by default.
We also correct for instrument broadening using dispersion
$\sigma_\textrm{inst}$, based on the channel-width $dv$. Then
\begin{align}
\sigma_{z,\mathrm{obs}} &= a_z \sigma(r) \times \sin(\theta) \times \sin(i)\,\,, \\
\sigma_{R,\mathrm{obs}} &= a_R \sigma(r) \times \cos(i) \,\,, \\
\sigma_{\theta,\mathrm{obs}} &= a_\theta \sigma(r) \times \sin(i) \times \sin(\theta)\,\,, \\
\sigma_\textrm{inst} &= dv / 2.355\,\,,\\
\sigma &= \sqrt{\sigma_{z,\mathrm{obs}}^2 + \sigma_{R,\mathrm{obs}}^2 + \sigma_{\theta,\mathrm{obs}}^2 + \sigma_\textrm{inst}^2}\,\,.
\end{align}
The system takes the face-on surface density $A_\textrm{\textsc{Hi} }(R)$ and, using the
flaring $z_0(R)$, converts it into a mid-plane density $\rho_\mathrm{m}(R)$.
From this, the actual density $\rho$ for each voxel is calculated.
We treat the gas flaring as a Gaussian distribution.
There exist alternative distributions, but the Gaussian has some advantages
for the analysis of the hydrostatics, in Paper V \citep{Peters2015E}.
Based on \citet{Olling1995A}, the errors due to this are expected to be minor.
Then
\begin{align}
\rho_\mathrm{m} &= \frac{A_\textsc{Hi} (R)}{\sqrt{2 \pi z_0(R)^2}} \,\,,\\
\rho &= \rho_\mathrm{m} \exp\left[-\frac{z^2}{2 \,z_0(R)^2}\right]\,\,.\label{eqn:gaussiandisc}
\end{align}
The densities $\rho$ are all in units of \textsc{Hi} atoms per cm$^{3}$.
In the final step, we convert to a local column density $N_\mathrm{HI}$
cm$^{-2}$ by multiplying with the physical length $ds$ that a single voxel
represents along its depth axis.
\begin{equation}
N_\mathrm{\textsc{Hi} } = \rho \,\, ds
\end{equation}
\subsection{Generating the \textsc{Hi} data Cube}
We have now prepared the 3D cube and it is time to perform a ray-trace through
the voxels along the depth axis and generate the final \textsc{Hi} data cubes.
\textsc{Galactus} supports two integration modes: optically thin and
self-absorption.
\subsubsection{Optically thin}
We assume that the \textsc{Hi} follows a Gaussian velocity distribution,
\begin{equation}
N_\mathrm{\textsc{Hi} } dv = \frac{ N_\mathrm{\textsc{Hi} }}{\sqrt{2\pi\sigma^2}} \exp\left[-\frac{(v-v_\mathrm{obs})^2}{2\sigma^2}\right] \,\,.\label{eqn:phi_v}
\end{equation}
For a single channel in the \textsc{Hi} data cube, starting at velocity $v_1$ and with a
channel width $dv$, the number of atoms $N$ in a voxel that are contributing
to the emission in a single channel, can be calculated as
\begin{equation}
N = \int_{v_1}^{v_1+dv}{N_\mathrm{\textsc{Hi} } \,dv}\,\,.
\end{equation}
We solve this integral using the error function \emph{erf}, i.e.
\begin{equation}
N = \frac{N_\mathrm{\textsc{Hi} }}{2 dv} \left[ {\rm erf }\left(\frac{v_1 + dv - v_\mathrm{obs}}{\sqrt{2\pi\sigma^2}}\right) - {\rm erf }\left(\frac{v_1 - v_\mathrm{obs}}{\sqrt{2\pi\sigma^2}}\right) \right] \label{eqn:Npervoxel}
\end{equation}
To calculate the optically thin surface-brightness of each voxel, we use the
optically thin limit of the radiative transfer equation
\citep[Equation 8.26]{Draine2011}, i.e.
\begin{equation}
T_{B,\textrm{thin}} = \frac{N}{ 1.8127\times10^{18}\,dv}\,\,.\label{eqn:Tpervoxel}
\end{equation}
As there is no self-absorption, the intensity of a pixel can simply be
calculated by taking the sum of $T_B$ in all voxels along the depth axis
at the position of that pixel.
\subsubsection{Self-absorption}
When running in self-absorption mode, the calculation gets more complex.
We follow the method from the previous subsection up to equation
\ref{eqn:Tpervoxel}, which gives us the emitted temperature in the velocity
range of the channel per voxel for an optically thin model.
From this, we can then calculate the optical depth $\tau_\nu$ of that voxel
as\footnote{Note that this is just a rewrite of Equation \ref{eqn:Tpervoxel}
with Equation 8.11 in \citet{Draine2011}.},
\begin{equation}
\tau_\nu = T_{B,\textrm{thin}} / T_\mathrm{spin}\label{eqn:tauTT}\,\,.
\end{equation}
We assume that the galaxy has no internal continuum absorption, nor a
background continuum source.
This assumption is at least true for the galaxies in Paper I, although it
does not hold for every galaxy.
To calculate the brightness of a pixel, we begin for the voxel at the back
of the depth axis associated with that pixel, and calculate the radiative
transfer equation with $T_\mathrm{bg}=0$ \citep[Equation 8.22]{Draine2011}.
\begin{equation}
T_{B,\textrm{self-absorbing}} = T_\mathrm{bg} \textrm{e}^{-\tau_\nu} + T_\mathrm{spin} \left(1 - \textrm{e}^{-\tau_\nu}\right)\label{eqn:self-absorption}
\end{equation}
For subsequent voxels $T_\textrm{bg}$ is the $T_B$ from the previous voxel.
We continue this for all voxels along the depth axis until the final value
$T_B$, which is adopted as the brightness for that pixel.
In our own Galaxy, many lines of sight are optically thick within a few
hundred parsec \citep{Allen2012A}.
To ensure sufficient accuracy, it is therefore advised to choose the physical
length of each voxel along the depth axis, such that it does not represent
more than a hundred parsec.
This can be done using the accuracy parameter (see Section
\ref{sec:supersampling}).
We note that the use of a constant spin temperature $T_\textrm{spin}$ and a
uniform density $\rho$ per voxel serves only as an
approximation.\label{sec:uniformdensity}
In reality the \textsc{Hi} in galaxies consists of gas at a range of spin
temperatures, not to mention varying densities (e.g. clouds) on scales
much smaller than a hundred parsec.
We discuss this in more detail in Section \ref{sec:effectiveTspin}.
\subsection{Beam Smearing}
After the \textsc{Hi} data cube has been generated, we convolve the cube with the beam
$\theta_\textrm{FWHM}$.
Note that we require the beam to be circular and the cells to be square.
\subsection{Simulating Noise}
It is possible to introduce artificial noise into a cube.
This can be done using the addition of the \texttt{-N} argument to the command
line call to \textsc{Galactus}.
The noise $\sigma$ in Table 3 of Paper I was measured as
the standard deviation per pixel.
In reality however, the noise is correlated between neighbouring pixels due to
the beam smearing.
\textsc{Galactus} automatically corrects for this, such that the noise
$\sigma$ is equal to that as measured from the standard deviation of the
convolved field.
\subsection{Input}
\textsc{Galactus} can be used in two modes, stand-alone and as a software
library.
To get the stand-alone version, the user can download the source code of the
program from the dedicated website at sourceforge.net/p/galactus.
After compiling the required \textsc{C++} libraries, the program can then be
run from the program folder using the command \texttt{./Galactus ini-file}.
\textsc{Galactus} uses ini-files as its main source of information, and any
run of the program thus needs to be set up using them.
We provide a list of all the options in the online Appendix.
It is also possible to import \textsc{Galactus} into your own \textsc{Python}
programs as a library.
All libraries required to run the program are contained in the
\texttt{support/} module.
It is beyond the scope of this paper to discuss this in detail
and we refer to the documentation in the code itself for more details.
In short, to generate a basic model, a programmer would first need
to import the \texttt{Galaxy} class from the \texttt{support.parameters}
module, initialise a version of it, and then use the
\texttt{parse\_ini(filename)} method to initialise all parameters
in the \texttt{Galaxy} object.
The \textsc{Hi} data cube can then be generated by calling the
\texttt{support.mainmodel.model()} method.
The source code has been extensively documented, and we invite
developers to have a look and contribute to the code-base.
\subsubsection{Superfunctions}\label{sec:superfunctions}
While most variables in \textsc{Galactus} are single valued, the
rotation curve and face-on surface density are examples of so-called
super-functions.
Super-functions are a special class of functions that can behave
differently based on the preference of the user.
In their 1 or 2 modes, (see Section \ref{tbl:Galactusinput}) they
act as parametrised functions, while in the 3 and 4 modes they act
as tabulated functions.
In tabulated mode, the user specifies the values of the function that
the function has at positions $R$.
During runtime, \textsc{Galactus} performs a linear interpolation
over these points to calculate the value of intermediate radii.
We have experimented with alternative types of interpolation, such
as spline interpolation.
While favourable from a theoretical point of view, in practise we found that
during fitting a spline could start to show extreme spikes (both negative and
positive), in an attempt to still reach all the user specified $R$ and $f(R)$
values.
This made attaching boundary conditions very hard.
In the end, we have settled for linear interpolation between values.
In parametrised mode, the functions are simplified to
an analytical form that best fits their common profiles.
For example, the velocity curve $v_\textrm{rot}$ has a parametrised form of
$v_\textrm{rot}(R) = a + a / (R/b+1)$.
The advantage of parametrizing functions is that it limits the amount of
free parameters, in this case only two.
The form of each super-function is shown in Table \ref{tbl:Galactusinput}.
Another advantage to parametrised functions is that they are less sensitive
to noise -- although at the expense of accuracy -- and give the ability
for a very quick first estimate of the parameters.
We provide a function \texttt{all\_to\_spline},
which can convert parametrised curves to tabulated ones.
\subsection{Output}
\textsc{Galactus} is capable of producing a range of products,
which can be selected in the parameter file.
The main output of the program is the generated \textsc{Hi} data cube.
There are however, more options. In this subsection, we will highlight the
most important ones.
\subsubsection{Total, visible and hidden matter cube}
Since the brightness can depend on opacity, the generated \textsc{Hi} data cube no
longer represents the total amount of atoms.
The total matter cube gives the total amount of atoms in a velocity range,
before any self-absorption is applied.
This starts with the result of Equation \ref{eqn:Npervoxel}, which is a cube
with the amount of \textsc{Hi} per voxel, for a given channel.
Rather than converting the data into temperatures with Equation
\ref{eqn:Tpervoxel}, we instead sum the voxels along the depth axis
and thus get the number of atoms.
This process is repeated for each channel in the \textsc{Hi} data cube.
In contrast, the visible matter cube uses the generated \textsc{Hi} data cube and
converts it back into atoms, using the inverse of the optically thin
approximation from Equation \ref{eqn:Tpervoxel}.
When running in optically thin mode, the visible and total matter cubes
are equal.
The difference comes into play when running in self-absorption mode.
This difference can be visualised with the hidden matter cube, which is
a cube showing the difference between the total and visible matter cubes.
\subsubsection{Moment maps}
Besides channel maps, \textsc{Galactus} can also output various moment
maps, specifically the zeroth, first, second and third moments.
These follow the same moment definitions as \textsc{Gipsy} \citep{GIPSY}.
\subsubsection{XV map and integrated profile}
\textsc{Galactus} is also capable of plotting the integrated XV (otherwise
known as integrated PV) diagrams of the \textsc{Hi} data cube, by integrating the
generated \textsc{Hi} data cube along the vertical axis.
By also integrating along the velocity axis, the program can generate an
integrated profile.
\subsection{Fitting}
Fitting a galaxy model to a \textsc{Hi} data cube is a very complex problem, and great
care needs to be taken in order to fit the data properly.
To give the user the best capabilities, we have included a number of fitting
routines in the program.
Suppose one is fitting a well-resolved face-on galaxy with fixed velocity
dispersion and no beam smearing.
In this case, a line-of-sight corresponds to a single radius inside the galaxy.
This radius is sampled by many pixels as we look around the galaxy.
The parameter space is thus very limited, as each radius only has to fit a
single face-on surface density and rotation curve.
A simple fitting routine, such as the Levenberg-Marquardt algorithm
\citep{Press1992}, can quickly converge to the optimal solution.
We make use of the \textsc{Scipy} package \texttt{optimise.leastsq}.
When beam smearing is involved, the light from various radii starts to
overlap, which causes the parameter space to get far more complex.
Even worse, local optimal solutions start to appear.
For example, where the true face-on surface density at a radius might be
$5\times10^{20}$ atoms/cm$^2$, the model could be happily converging towards
$3\times10^{20}$ and $7\times10^{20}$ atoms/cm$^2$
in two adjacent radii, with the beam
taking care of the smoothing back to $5\times10^{20}$ atoms/cm$^2$ on average.
The velocity dispersion also creates many local optima, especially in edge-on
galaxies where the use of a very high velocity dispersion at a high radius
$R$ can mask out the velocity dispersion at a slightly lower radius, thus
convincing the program that low velocity dispersions at that radius are also
acceptable.
To find the global optimum, a more powerful algorithm is required.
We have extensively tested many algorithms.
Our preferred method is the \textsc{PSwarm} algorithm, which is a combination
of pattern search and particle swarm
\citep{Vaz07,Vaz09,Thi12}\footnote{Website: www.norg.uminho.pt/aivaz/pswarm/}
and is used through the \textsc{OpenOpt}
framework\footnote{Available at openopt.org/.}.
Solving for a global optimization problem requires a far more thorough search
of the parameter space and thus requires much more time to converge, compared
to the Levenberg-Marquardt algorithm.
We have also tested various implementations of genetic algorithms, but found
that the random nature of the improvements led to unacceptably long
convergence times.
We also find that when fitting for the velocity dispersion, it is often
misused by an algorithm to 'average over' local structure in the XV
diagram, by using a very high velocity dispersion.
This can happen in particular when warps and self-absorption are
present, but are not allowed for in the modelling.
In those cases, it is not even possible to model the XV diagram
accurately. The algorithm will then start to misuse the velocity dispersion.
To estimate the $\chi^2$ cost of some parameter set $p$ on the model
$T_\textrm{model}(p)$, we use least-squares deviations.
We show this in Equation \ref{eqn:lossfunction}, where we use the standard deviation of the noise $\sigma$ from the observed \textsc{Hi} data cube $T_\textrm{observation}$.
Index $i$ runs over all pixels in the \textsc{Hi} data cube,
\begin{equation}
\chi^2(p) = \sum_{i=0}^{N_\textrm{pixels}}{\frac{\left[T_\textrm{model}(i;p) - T_\textrm{observation}(i)\right]^2}{\sigma^2}}\,\,. \label{eqn:lossfunction}
\end{equation}
Lastly, we implement a Monte-Carlo Markov-Chain (MCMC) version of the model.
MCMC is a different approach to model optimisation. \label{sec:Galactus-MCMC}
Rather than look for a single parameter set that represents the global optimal
solution, MCMC explores the parameter space and draws samples of parameter
sets from the region where, based on the probability distribution, the optimum
is expected.
We make use of the \textsc{emcee} library \citep{emcee}, which implements the
affine-invariant ensemble sampler proposed by \citep{emcee2}.
Following the cost function $\chi^2$, given in Equation \ref{eqn:lossfunction},
the log-likelihood $\mathcal{L}$ function for a particular parameter-set $p$
is calculated as
\begin{align}
\mathcal{L}(p) &= \log{\left(-\frac{\chi^2(p)}{2}\right)}\,\,,\\
&= \log{\sum_{i=0}^{N_\textrm{pixels}}{-\frac{\left[T_\textrm{model}(i;p) - T_\textrm{observation}(i)\right]^2}{2\sigma^2}} }\,\,.\label{eqn:laplacedistribution}
\end{align}
\textsc{emcee} uses multiple MCMC samplers in parallel, each of which generates
a chain of samples.
For the number of samplers we choose double the amount of free parameters,
which for tabulated fitting implies 150-300 chains.
After a sufficiently long burn-in period, we collect a large group of $N$
samples from the collection of chains.
For each free parameter that is being fitted,
the group thus has $N$ values
that this parameter would most likely have.
These values thus form a distribution.
In this and subsequent papers, we make use of the central 68, 95 and 99.7\%
of the likelihood distribution to denote the errors.
Note that if the distribution in a parameter follows a Gaussian distribution,
this would be equal to the one, two and three sigma deviations of that Gaussian.
\section{Notes}\label{sec:notes}
\subsection{Masking the Data}
\textsc{Galactus} can deal with masked pixels.
Pixels that are masked in the input \textsc{Hi} data cube, specified with the \
texttt{image} parameter (see also Section \ref{tbl:Galactusinput}), are
treated as pixels with a value of zero.
\textsc{Galactus} will still render the corresponding pixels in the
output model, as the user could still be interested in the model at
these pixels, but has masked them on purpose for some reason.
The \texttt{mask} parameter can be used to specify which pixels
\textsc{Galactus} does not need to render.
The parameter needs to point to a \textsc{Hi} data cube, which can be the
same as the one used in the \texttt{image} parameter.
All pixels in the channel maps that make up the \textsc{Hi} data-cube that
have value NaN\footnote{NaN stands for Not A Number. By default,
\textsc{Miriad} masking gives pixels this value.} are selected
as the masked pixels.
\textsc{Galactus} calculates which voxels contribute to which pixels.
If a voxel only contributes to masked pixels, it is removed from the
calculation.
This can lead to drastic performance increases.
The last type of mask is the \texttt{boundaries} mask.
This can be used to specify in which channel maps the line-of-nodes
velocity $v_\textrm{rot}(R)$ is expected to lie.
It is then possible to mask everything except the outer envelope of a XV
diagram.
The fitting algorithms are then constrained in the $v_\textrm{rot}(R)$
parameter.
\subsection{Concerning the Resolution}\label{sec:supersampling}
By default, \textsc{Galactus} traces only one ray per pixel.
The initial cube has voxels that are one pixel wide on all sides.
Thus, if the output cube has pixels with a (projected) size of one kpc,
the voxels are one kpc on each side as well, and the ray is thus sampled
every one kpc as it crosses the galaxy.
For a galaxy that has a diameter of 20 kpc, the rays will thus at most only
sample 20 positions inside the galaxy.
This can lead to a jagged looking result.
To provide a better representation of the galaxy, we have introduced two
additional user options.
The first is the \texttt{accuracy} level, which controls the number of
samples along the ray.
It works as a scalar on the original number of samples along the depth
axis, such that if one originally measured 20 samples, an accuracy level
of four would lead to 80 samples.
Following the above example, these would thus be separated 250 pc apart
along the ray.
The second option is \emph{super-sampling}, a technique common in the 3D
graphics industry.
In super-sampling, each output pixel is represented by multiple rays at
slightly different positions inside the pixel, with the result being the
average of all rays.
For example, suppose the original ray runs a trace centred on the middle
of the pixel, thus at position 1/2.
Then with super-sampling, we have two rays, both at 1/3 and 2/3.
Suppose that a significant change would beyond position 1/2. The normal
model would not represent this, while super-sampled pixel would.
Super-sampling works in both the horizontal and vertical directions.
A super-sampling of factor two means each pixel is sampled by four rays.
An increased accuracy thus increases the number of voxels along the depth axis.
An increase in super-sampling decreases the size of an individual voxel
along the horizontal and vertical axis and thus increases the number of
voxels quadratically.
The increases in accuracy and super-sampling are thus increases in the
number of voxels that need to be calculated, which come with a performance
penalty.
It is thus important to realise that there is a trade-off between the
desired accuracy and the time available.
In Figure \ref{fig:AA}, we visualise this trade-off.
Shown are the timing results of an edge-on model, at multiple combinations
of super-sampling and accuracy levels, with their relative error
compared to a very high-resolution model with an accuracy of five and a
super-sampling of five, which in this case implies a model with voxels
that each represent 27 pc on all side.
In a perfect world, the accuracy and super-sampling could be set to
$\infty$ and the results would be perfect.
In practice however, due to time constraints, there is no single right setting,
it is thus important to think about the problem and decide what the
best setting should be.
For example, in a face-on galaxy one quickly runs the risk of 'overshooting'
the disc,
if there were not enough samples along the ray, and a higher accuracy would
then be desirable.
In an edge-on, overshooting the disc is less important, but one might want to
super-sample
the galaxy as to model the vertical structure more accurately.
A super-sampling of two incorporates a special feature, which positions the
four rays in an optimal distribution.
In this way, both the horizontal and vertical axis of the pixel in question
are traced at four unique positions, rather than the default of two positions
per axis.
The effect of this can be seen in Figure \ref{fig:AA}, where the results
from a super-sampling of 2 are nearly equal to the far more computationally
expensive super-sampling of four (i.e. 16 rays).
A super-sampling of two outperforms a super-sampling of three.
Figure \ref{fig:AA} also demonstrates that increasing the accuracy does not
drastically improve the model.
We recommend that the accuracy be chosen such that the length of the voxel
along the depth axis is at most 100 parsec, such that the radiative transfer
equation is calculated with sufficient precision.
\begin{figure}
\resizebox{0.48\textwidth}{!}{\includegraphics{Petersetal-IIfig2.jpg}}
\caption[Resolution trade-off in Galactus]{The accuracy versus time trade-off visualized. The various tracks show the various super-sampling modes, while the dots along each track represent the accuracy level used. Along each track, the top left point is always accuracy level of one, progressing to an accuracy level of four, at the right-hand side. Overall the rule holds that the closer to (0,0), the better.}
\label{fig:AA}
\end{figure}
\begin{figure*}
\includegraphics[width=0.38\textwidth]{Petersetal-IIfig3a.jpg}
\includegraphics[width=0.38\textwidth]{Petersetal-IIfig3b.jpg}
\caption[Effective spin temperature assuming $p_\textrm{CNM}=0.8$]{Effective spin temperature assuming $40\%$ (left panel) or $80\%$ (right panel) of the gas is in the CNM. Colour bars represent $68.5\%$, $95.45\%$ and $99.73\%$ of the distribution at each point. In white the $50\%$ (median) of the sample is shown. The red dashed line represents the fitted effective spin temperature.}
\label{fig:effectivespin8}\label{fig:effectivespin4}
\end{figure*}
\begin{figure*}
\includegraphics[width=0.80\textwidth]{Petersetal-IIfig4.jpg}
\caption{Maximum surface brightness map for NGC\,2403. The values for the bar on the right is brightness temperature in Kelvin.}
\label{fig:NGC2403-maxtempmap}
\end{figure*}
\section{The Effective Spin Temperature}\label{sec:effectiveTspin}
In a real galaxy, the \textsc{Hi} never has a single spin temperature.
The gas is balanced between the phases of the CNM, which has a median spin
temperature of 80\,K
and the WNM, with temperatures between 6000 and 10000\,K.
The fraction of the mass contained in either state remains unclear,
but it was estimated that $40\%$ of the neutral hydrogen mass of the Galaxy
is in the CNM \citep{Draine2011}.
With such a wide range in temperatures, how reliable is the use of a single
effective spin temperature?
To test this, we have run a range of Monte Carlo simulations.
For a fixed CNM fraction $p$, we calculate the effective spin temperature over
a range of column densities.
For column density $N_\textrm{\textsc{Hi} }$, we divide it up in $N_\textrm{clouds}$ equal
sized bins, each of which represents a cloud of gas with a column density
$N_\textrm{cloud}$.
The number of clouds is randomly picked between 100 and 1000, to represent the
variation in crossing a real galaxy.
Based on fraction $p$, we randomly choose if a cloud is part of either the CNM
or the WNM.
If it is a CNM cloud, it is assigned a random spin temperature
$T_\textrm{spin}$, chosen from a uniform distribution between 50 and 100\,K.
Otherwise, it is assigned a $T_\textrm{spin}$ chosen from a distribution
between 6000 and 8000\,K.
We then perform a line-of-sight integration, starting with background
temperature $T_\textrm{bg}=0$\,K, using Equations \ref{eqn:Tpervoxel},
\ref{eqn:tauTT} and \ref{eqn:self-absorption}.
Taking 1000 samples at each column density $N_\textsc{Hi} $, we get a distribution
of observed temperatures $T_B$.
We show the results for two mass fractions in Figures \ref{fig:effectivespin4}a
and \ref{fig:effectivespin8}b.
We fit a single effective $T_\textrm{spin}$ to the median of the distributions,
using
\begin{equation}
T = T_\textrm{spin}\left(1 - e^{-\tau}\right)\,\,.
\end{equation}
The results are over-plotted in red in Figures \ref{fig:effectivespin4}a and
\ref{fig:effectivespin8}b.
For the $40\%$ fraction, we find an effective spin temperature of 177.2\,K,
while for the $80\%$ fraction we find 89.5\,K.
Clearly, the balance between the two populations is important.
More importantly, in both cases we see that the median spin temperature
levels out, and a single effective spin temperature gives a decent fit
to the median one.
Provided the right value is chosen, it is thus possible to use a single
effective spin temperature when modelling a galaxy.
What the exact value for this effective spin temperature is, is hard to determine
and is beyond the scope of this work.
During fitting, we shall use a spin temperature of 100\,K,
a value in good agreement with the figures in Paper I.
\begin{figure}
\includegraphics[width=.48\textwidth]{Petersetal-IIfig5.jpg}
\caption{Radial maximum surface brightness of NGC\,2403}
\label{fig:NGC2403-maxtempalongr}
\end{figure}
\section{Face-on Galaxy NGC~2403}\label{sec:NGC2403}
In the following sections we test the program for a number of cases.
To begin with, we wish to test our fitting routines on a face-on galaxy to
confirm the accuracy of the rotation curve and surface density modelling.
For this purpose, we have chosen the galaxy {NGC\,2403}.
Like the galaxies in our sample it is a nearby, late-type Scd galaxy
\citep{Tully2000A}. \textsc{Hi} observations of high spatial and velocity
resolution are available
for it. Even though the rotation velocity is slight higher (130 versus on
average 90{km s$^{-1}$}) it is suitable for
comparing to our sample of edge-on, late-type galaxies (Paper I).
Using the tip of the red giant branch, \citet{Dalcanton2009A} estimated the
distance at 3.2\,Mpc.
The \textsc{Hi} kinematics have have been studied in detail by \citet{Sicking1997A},
\citet{Fraternali2001A,Fraternali2002A} and \citet{DeBlok2008A}.
The galaxy is part of the public-data release from The \textsc{Hi} Nearby Galaxy
Survey (THINGS), which was based on B-, C- and D- configuration Very Large
Array (VLA) observations\footnote{The public data from the THINGS survey is
available at www.mpia-hd.mpg.de/THINGS/Data.html.} \citep{Walter2008A}.
The THINGS observations were previously studied in detail by
\citet{DeBlok2008A}, who measured a systemic velocity $v_\mathrm{sys}$ of
132.8\,km\,s$^{-1}$ and an inclination $i$ of $62.9^\circ$.
We test \textsc{Galactus} using the naturally weighted \textsc{Hi} data cube from the
THINGS website.
The original cube has a size of $2048\times2048$ pixels and 61 channels,
which is too large for \textsc{Galactus} to handle.
\textsc{Miriad} task \texttt{smooth} was used to create a circular beam
with a full-width half-maximum of 12.4\, arcseconds.
We used the task \texttt{regrid} to downsize the resolution from 1.0 to
8.8\,arcsec/pixel, resulting in somewhat less than 2 pixels/beam.
The initial beam was $8.75\times7.65$ arcsec, with a position angle of
25.2 degrees.
We used the task \textsc{regrid} to downsize the sampling of the data
from 1.0 to 8.8 arcsec/pixel, resulting in somewhat less than 2 pixels/beam.
We have chosen this sampling as this corresponds to roughly 100 parsec per
pixel.
As each pixel is supersampled
by 4 rays, the area of the FWHM of the beam will be modeled by of order
20 rays,
which is sufficient to prevent undersampling. While an even higher resolution
would have been desirable, it was found to be computationally infeasable, as
the required memory and computation time increase with the power 3.
Using task \texttt{cgcurs}, a mask was drawn by hand around all regions
containing flux.
With task \texttt{imsub}, we used this mask to create a final, masked cube
that only contains parts of the cube where flux was detected.
The final cube has a size of $232\times228$ pixels and 55 channels at
5.169\,km\,s$^{-1}$ resolution.
In Figure \ref{fig:NGC2403-maxtempmap}, the maximum surface brightness has
been calculated from the original THINGS image.
Here we have converted the intensity $I$ from Jy/beam to Kelvin using
Equation 4 of Paper I:
\begin{equation}
T_B = \frac{\lambda^2 S}{2 k_B \Omega} = \frac{606000 S}{\theta^2}\,\,.\label{eqn:JytoK}.
\end{equation}
A clear spiral structure is visible in the figure.
Most of the brightness hovers between 45 and 60\,K and appears as spiral arm
structure, with local regions peaking at over 120\,K.
We rotated the galaxy such that the major axis is lined up with
the horizontal x-axis.
If we only look at the maximum brightness along the horizontal axis, such that
each position reflects the maximum brightness independent of location along
vertical axis, we get the result as seen in Figure
\ref{fig:NGC2403-maxtempalongr}.
This produces a result similar to the maxima along the major axis in edge-on
galaxies as seen in Paper I.
There we found a plateau around 80\,Kelvin in all edge-on galaxies for a
significant part of their discs.
There is no clear plateau in Figure \ref{fig:NGC2403-maxtempalongr}.
The wings of this profile are much extended, running continuously towards
the background.
This is different compared to our edge-on galaxies, where we observed very
sharp wings.
In Paper I, we argued that the plateau in the profile was due to
self-absorption in long lines-of-sight.
The shorter lines-of-sight in {NGC\,2403} would therefore not be
expected to exhibit (strong) signs of self-absorption.
Only in the dense inner region do the densities still reach levels
high enough for self-absorption to significantly affect them, as
exhibited by the peak around 80\,K.
We thus argue that with the exception of the central part, most of
the galaxy is not affected by significant self-absorption.
We note in general, that observations of self-absorption can only
occur when the beam has sufficient spatial resolution. Otherwise,
the effect will wash out due to smearing.
We fit {NGC\,2403} using both an optically thin model and a self-absorption
model.
For the self-absorption model, we use a spin temperature $T_\mathrm{spin}$
of 100\,K.
We fit the galaxy using a constant velocity dispersion of 10\,km/s.
We constrain the disc to a thickness of 700\,pc in the inner part and
flaring out to 1\,kpc at 1400'' ($23$\,kpc).
Warps or lagging haloes are beyond the scope of the current test.
A double-pass strategy is used to fit the data.
We first use the parametrised functions for the face-on surface
density, flaring and rotation curve.
This pass is fitted using the Levenberg-Marquardt algorithm.
After that, we interpolate over the fitted functions and tabulate
them into 37 chunks with a separation of 38''.
We then run a second pass.
As the galaxy is face-on and very well resolved there is no strong
danger of running into local minima, and we thus again use the quick
Levenberg-Marquardt algorithm.
The results for this fitting are shown in Figure \ref{fig:NGC2403-fits}.
The rotation curves of the optically thin and self-absorbing fits are
effectively identical.
It is reassuring to find that in both cases the same rotation curve
is found, proving that the observed $v_\textrm{rot}$ in face-on
galaxies is not affected by self-absorption.
The profile also matches up well with the profile from \citet{Sicking1997A}.
Only in the outer radii do we suffer more from signal to noise issues.
Given the detailed tilted ring and warp fitted by \citet{Sicking1997A}, which
we are ignoring, this is expected.
Now consider the mass models.
Unsurprisingly, the optically thin profile is clearly lower than the
optically thick profile.
The first has a total mass of $2.9\times10^9$\,M$_\odot$.
The self-absorbing mass totals at $3.2 \times 10^9$\,M$_\odot$, of which
$\sim10\%$ is hidden by the self-absorption.
The overall shape of the profiles agree well, with all local features
visible in both profiles.
In comparison, \citet{Sicking1997A} found slightly more mass than our
optically thin model.
They report a total mass of $3.27\times10^9$ solar masses for {NGC\,2403}.
\begin{figure*}
\resizebox{0.48\textwidth}{!}{\includegraphics{Petersetal-IIfig6a.jpg}}
\resizebox{0.48\textwidth}{!}{\includegraphics{Petersetal-IIfig6b.jpg}}
\caption[Kinematic and mass models for NGC\,2403]{Kinematic and mass models for {NGC\,2403}. The blue profiles represent the optically thin model. The green profiles are the self-absorption fit. The dashed red profiles represent the data from \citet{Sicking1997A}.}
\label{fig:NGC2403-fits}
\end{figure*}
\section{The Maximum Surface Brightness in an Edge-on Galaxy}\label{sec:NGC2403edgeon}
In Section 3 of Paper I, we presented a simple toy model to demonstrate
that a plateau of constant maximum surface brightness is an indication of
self-absorption.
Using \textsc{Galactus}, we can now extend this toy model to a more
physical basis.
We project the previous fit of {NGC\,2403} to a full edge-on orientation.
We use the self-absorption fit as the face-on surface-density and
rotation curves as input parameters (see Figure \ref{fig:NGC2403-fits}).
We will compare between an optically thin model and a self-absorption
run at a spin temperature of 100\,K.
The results for this test are shown in Figure \ref{fig:model-maxT}.
The model for the galaxy at its original inclination peaks at roughly
40\,K (blue line).
This is lower than the observed maximum brightness maps as seen in
Figure \ref{fig:NGC2403-maxtempalongr}.
This is expected, as the model cannot recreate the small-scale
bright regions seen in the actual observation.
When we rotate the galaxy to full edge-on, we see that the long
lines-of-sight carry the integration well above 300\,K (green line).
Especially near the inner part at $x=0$ the model keeps rising.
As we have argued in Paper I, this is never observed.
Comparing this to the self-absorption model at $T_\mathrm{spin}=100$~K,
we see that the maximum surface brightness shows a clear plateau
around 80 Kelvin with sharp edges.
Comparing this to the galaxy observations in Paper I, the self-absorption
clearly matches the data better than the optically thin model.
Compared to the toy model from Section 3 of Paper I, the result is
less dramatic than in Figure 2 in that paper, where we applied a
far higher face-on column density distribution than used here.
\section{Effect of Inclination on the Visible Mass}\label{sec:effectofinclination}
Having established that an edge-on galaxy can easily hide a considerable
fraction of its neutral hydrogen, we wish to test the relation between
inclination and self-absorption.
We test the effect of self-absorption on galaxies in the range of
inclinations from 60 to 90 degrees.
We have selected $T_\mathrm{spin}=100$\,K and $T_\mathrm{spin}=150$\,K
as the spin temperatures we wish to model, and will produce an
optically thin model at each inclination to compare.
We use the self-absorption results of {NGC\,2403} from Section
\ref{sec:NGC2403} as the basis for all three models.
The results for this fit are shown in Figure \ref{fig:effect-of-inclination}.
It is clear that in both cases self-absorption is present at all inclinations.
However, beyond an inclination of about 82 degrees, the effects starts
to increase drastically.
Similar to Paper I, we find that for a spin temperature of 100\,K that
$30\%$ of the \textsc{Hi} remains undetected.
The milder case of a spin temperature of 150\,K still has a significant
absorption, with $\sim23\%$ of the \textsc{Hi} going unseen.
A natural prediction from Figure \ref{fig:effect-of-inclination} would
be that in a statistical study the edge-on galaxies would have a lower
average \textsc{Hi} mass than the face-on to moderately inclined galaxies.
We have tried to investigate this by comparing the K-band magnitude
(which is expected not to be much affected by inclination effects) with
the apparent \textsc{Hi} mass, using a combination of literature surveys.
Unfortunately, we were unable to confirm or disprove the results, as
there were too few edge-on galaxies in the available surveys with
exact inclination measured.
It is beyond the scope of this project to investigate this
in more detail.
\section{The Baryonic Tully-Fisher Relation}\label{sec:TullyFisher}
\begin{figure}
\includegraphics[width=0.48\textwidth]{Petersetal-IIfig7.jpg}
\caption[Theoretical maximum surface brightness temperatures for an edge-on galaxy]{Maximum surface brightness along the major axis based on modelling for {NGC\,2403}.
The lower blue line is the galaxy at its normal inclination.
The top green line is the optically thin case as seen edge-on.
The middle red line is the self-absorption case with a spin temperature of 100 K.}
\label{fig:model-maxT}
\end{figure}
Having established that edge-on galaxies can be affected by significant \textsc{Hi}
self-absorption, what is the effect on the baryonic Tully-Fisher (TF)
relationship?
The TF relation relates the dynamical mass of a galaxy to its
luminosity and provides a test of theories of galaxy formation and
evolution \citep{Verheijen1997A,Bell2001A}.
We use the two edge-on models from Section \ref{sec:NGC2403edgeon}
as a basis for this analysis.
The integrated spectra are shown in Figure \ref{fig:TF}.
The T-F relation is based on the width $w_{20}$ of the profile at
20\% of the maximum in the integrated spectra \citep{Tully1977A}.
Clearly, this height varies drastically between the two cases, yet
the difference between the two $w_{20}$ measurements is minimal.
This is due to the sharp drop-off that both profiles show at the
extremes, where self-absorption is less important.
For the optically thin case a value of 285.0 km/s is found and
for the self-absorption case 291.0\,km/s, an increase of only $2\%$.
We thus conclude the $w_{20}$ parameter is not affected
significantly by self-absorption.
The baryonic TF relation is however also based on the
total baryonic mass $M_\textrm{bar}$, and as Figure \ref{fig:TF}
clearly illustrates, self-absorption can significantly affect that property.
The TF-relation has not been well set, with various authors
finding slopes between
$M_\textrm{bar}\propto V^3$ and $M_\textrm{bar}\propto V^4$ (\citet{kf11} and
references therein).
Although beyond the scope of this project, it will be interesting to see
if the intrinsic scatter of the T-F relation decreases when
self-absorption of the \textsc{Hi} is taken into account.
\begin{figure}
\includegraphics[width=0.48\textwidth]{Petersetal-IIfig8.jpg}
\caption[Visible \textsc{Hi} mass as function of inclination]{Visible \textsc{Hi} in an edge-on model based on {NGC\,2403}.
The top blue line is the galaxy as if there is no self-absorption present.
The bottom green line has self-absorption at $T_\mathrm{spin}=100$\,K.
The middle red line is the self-absorption at $T_\mathrm{spin}=150$\,K.}
\label{fig:effect-of-inclination}
\end{figure}
\begin{figure}
\includegraphics[width=0.48\textwidth]{Petersetal-IIfig9.jpg}
\caption[Tully-Fisher with self-absorption]{Impact on the width $w_{20}$ as measured from the spectra from an edge-on galaxy affected by in an optically thin (top, blue) and self-absorption with $T_\textrm{spin} =100$~K (bottom, green) case.}
\label{fig:TF}
\end{figure}
\section*{Acknowledgments}
SPCP is grateful to the Space Telescope Science Institute, Baltimore, USA, the
Research School for Astronomy and Astrophysics, Australian National University,
Canberra, Australia, and the Instituto de Astrofisica de Canarias, La Laguna,
Tenerife, Spain, for hospitality and support during short and extended
working visits in the course of his PhD thesis research. He thanks
Roelof de Jong and Ron Allen for help and support during an earlier
period as visiting student at Johns Hopkins University and
the Physics and Astronomy Department, Krieger School of Arts and Sciences
for this appointment.
PCK thanks the directors of these same institutions and his local hosts
Ron Allen, Ken Freeman and Johan Knapen for hospitality and support
during many work visits over the years, of which most were
directly or indirectly related to the research presented in this series op
papers.
Work visits by SPCP and PCK have been supported by an annual grant
from the Faculty of Mathematics and Natural Sciences of
the University of Groningen to PCK accompanying of his distinguished Jacobus
C. Kapteyn professorhip and by the Leids Kerkhoven-Bosscha Fonds. PCK's work
visits were also supported by an annual grant from the Area of Exact
Sciences of the Netherlands Organisation for Scientific Research (NWO) in
compensation for his membership of its Board.
|
\section{Motivation}
Let $v$ be a polynomial vector field on $\mathbb{C}^2$ having only non-degenerate singularities. Assume $v$ is \emph{non-dicritic}, i.e.~the line at infinity $L$ is invariant for the induced foliation $\mathcal{F}_v$ on $\P^2$ and assume also that the singularities of $\mathcal{F}_v$ along $L$ are non degenerate -- note that all these assumptions are generic conditions for polynomial vector fields on $\mathbb{C}^2$. Each singular point $p$ of $v$ carries two numeric invariants, the eigenvalues of the linearization matrix $Dv(p)$. On the other hand, each singular point at infinity carries only one invariant, its Camacho-Sad index. We know that there are certain relations among all these invariants, namely the \emph{Euler-Jacobi relations}:
\begin{eqnarray}
\sum_{v(p)=0}\frac{1}{\det(Dv(p))}=0, \label{eq:EJ1} \\
\sum_{v(p)=0}\frac{\operatorname{tr}(Dv(p))}{\det(Dv(p))}=0, \label{eq:EJ2}
\end{eqnarray}
the Baum-Bott formula:
\begin{equation} \label{eq:BB}
\sum_{\operatorname{Sing }\mathcal{F}_v}\operatorname{BB}(\mathcal{F}_v,p)=(d+2)^2,
\end{equation}
and the Camacho-Sad formula for the invariant line $L$:
\begin{equation}
\sum_{L\cap\operatorname{Sing }\mathcal{F}_v}\operatorname{CS}(\mathcal{F}_v,L,p)=1.
\end{equation}
The above relations are well-known, yet a simple dimension count shows that there must be more, currently unknown, such relations. The objective of this paper is to show, in \hyperref[sec:foliations]{\S\ref*{sec:foliations}} and \hyperref[sec:vectorfields]{\S\ref*{sec:vectorfields}}, how all the above relations can be reduced to a particular case of the Woods Hole formula. Our hope is to use the Woods Hole formula in the future to discover our missing relations. Some of these motivating questions are discussed in \cite{TwinVectorFields} and \cite{Guillot2004}.
In order to deduce the above index-theorems from Woods Hole we will first introduce the formula in \hyperref[sec:WoodsHole]{\S\ref*{sec:WoodsHole}}, and use use it to get index theorems for regular endomorphisms of $\P^2$ in \hyperref[sec:endomorphisms]{\S\ref*{sec:endomorphisms}}.
\section{The Woods Hole formula}\label{sec:WoodsHole}
The so-called \emph{Woods Hole formula} (also known as Atiyah-Bott fixed point theorem) is a generalization of the Lefschetz fixed point formula due to Atiyah and Bott in the complex analytic case \cite{AtiyahBott1966,AtiyahBott1967,AtiyahBott1968}, and to Verdier in the algebraic case \cite{Verdier1967} (see also \cite{Taelman2015,Beauville1972}).
\medski
Let $X$ be a compact complex manifold and $f$ a holomorphic endomorphism of $X$. Let $\O_X$ denote the structure sheaf on $X$ and let $\mathscr{F}$ be a coherent analytic sheaf.
\begin{definition}\label{def:lift}
A \emph{lift} of $f$ to $\mathscr{F}$ is a morphism of $\O_X$-modules
\[ \varphi\colon f^*\mathscr{F} \to \mathscr{F}. \]
\end{definition}
Note that a lift of $f$ to $\mathscr{F}$ induces linear maps on cohomology $\varphi^{k}\colon H^{k}(X,f^*\mathscr{F}) \to H^{k}(X,\mathscr{F})$. In addition, the pullback functor $f^*$ gives maps $H^{k}(X,\mathscr{F}) \to H^{k}(X,f^*\mathscr{F})$. Composition of these maps gives endomorphisms on the cohomology groups:
\[ \xymatrix{
\widetilde{\varphi}^k\colon H^{k}(X,\mathscr{F}) \ar[r]^{f^*} & H^{k}(X,f^*\mathscr{F}) \ar[r]^{\varphi^k} & H^{k}(X,\mathscr{F})}.
\]
\begin{definition}\label{def:Lefschetznumber}
The \emph{Lefschetz number} $\mathrm{L}(f,\varphi,\mathscr{F})$ of a lift $\varphi$ is defined by
\[ \mathrm{L}(f,\varphi,\mathscr{F}) = \sum_k (-1)^k\, \operatorname{tr}(\widetilde{\varphi}^k,\,H^{k}(X,\mathscr{F})). \]
\end{definition}
Let $x\in X$. We denote by $\mathscr{F}_x$ the stalk of $\mathscr{F}$ at $x$, and by $\mathscr{F}(x)$ the fiber of $\mathscr{F}$ at $x$, namely
\[ \mathscr{F}(x)=\mathscr{F}_x\otimes_{\O_{X,x}}(\O_{X,x}/\mathfrak{m}_x). \]
Note that each fiber is a finite dimensional complex vector space. Since $\varphi$ induces morphisms on stalks
\[ \varphi_x\colon (f^*\mathscr{F})_x \cong \mathscr{F}_{f(x)} \longrightarrow \mathscr{F}_x, \]
we see that if $p$ is a fixed point of $f$ we obtain a morphism $\varphi_p\colon\mathscr{F}_p\to\mathscr{F}_p$. In a similar way we obtain linear endomorphisms on the fibers $\varphi(p)=\mathscr{F}(p)\to\mathscr{F}(p)$. Finite dimensionality of the fibers guarantees that these endomorphisms have a well defined trace.
\begin{definition}\label{def:transversal}
We say that an endomorphism $f$ of $X$ is \emph{transversal} if its graph is transversal to the diagonal in $X\times X$. Equivalently, $f$ is transversal if and only if it has isolated simple fixed points only.
\end{definition}
\begin{theorem}[Woods Hole formula]\label{thm:WoodsHole}
If $f$ is a transversal endomorphism then
\[ \sum_{f(p)=p} \frac{\operatorname{tr}(\varphi(p),\,\mathscr{F}(p))}{\det(1-df(p),\,\Omega^1_X(p))} = \mathrm{L}(f,\varphi,\mathscr{F}), \]
where $df\colon f^*\Omega^1_X\to\Omega^1_X$ denotes the differential of $f$.
\end{theorem}
The Woods Hole formula is commonly used in the particular cases where $\mathscr{F}=\Omega^k_X$ and $\varphi=\wedge^k df$ (being the case $k=0$ the classical holomorphic Lefschetz formula). In such case, if $J_f$ denotes the Jacobian matrix of $f$, the formula becomes
\[ \sum_{f(p)=p}\frac{\operatorname{tr}(\wedge^k J_f(p))}{\det(I-J_f(p))} = \operatorname{L}(f,\wedge^k df,\Omega^k_X). \]
Note that the indices on the left hand side depend only on the eigenvalues of $J_f$ and thus the above equation gives a relation on the spectra of the differentials of $f$ at the fixed points. On the other hand, the Lefschetz number $\operatorname{L}(f,\wedge^k df,\Omega^k_X)$ depends only on the action of $f$ on the Dolbeault cohomology of $X$ in virtue of the standard identification $H^q(X,\Omega^p_X)\cong H^{p,q}_{\bar{\partial}}(X)$. In the particular case $X=\P^n$ this action on the Dolbeault cohomology depends only on the degree $d$ of $f$, and in fact, it is not hard to see that it is just multiplication by an appropiate power of $d$. This gives the following result.
\begin{corollary}\label{coro:indicesPn}
Let $f$ be a transversal holomorphic endomorphism of $\P^n$ of degree $d$. Then
\begin{equation}\label{eq:LefschetzPn}
\sum_{f(p)=p}\frac{\operatorname{tr}(\wedge^k J_f(p))}{\det(I-J_f(p))} = (-1)^k d^k .
\end{equation}
\end{corollary}
\begin{remark}
The Woods Hole formula can also be used to count the number of fixed points or singular points in different settings. For example, we can choose $\mathscr{F}=\wedge^\ast\Omega^1_{\P^n}$, the whole exterior algebra of $\Omega^1_{\P^n}$ with $\varphi$ acting as $\wedge^k df$ on the graded componen of degree $k$. In such case, it follows from \hyperref[coro:indicesPn]{Corollary \ref*{coro:indicesPn}} and from the fact that $\det(I-A) = \sum_k (-1)^k \operatorname{tr}(\wedge^k A)$ that a degree $d$ endomorphism of $\P^n$ has $1+d+\ldots+d^n=\frac{d^{n+1}-1}{d-1}$ fixed points.
\end{remark}
\section{Indices for transversal endomorphisms of \texorpdfstring{$\P^n$}{Pn}}\label{sec:endomorphisms}
Besides the index theorems stated in \hyperref[coro:indicesPn]{Corollary \ref*{coro:indicesPn}} we can use the Woods Hole formula to prove the following index theorem.
Recall that any invariant polynomial function on $\mathfrak{gl}_n(\mathbb{C})$ can be uniquely expressed as a polynomial on the \emph{elementary symmetric polynomials} $\sigma_1, \ldots, \sigma_n$.
\begin{application}\label{app:indicesPn}
Let $B$ be an invariant polynomial function on $\mathfrak{gl}_n(\mathbb{C})$ of degree at most $n$. Let $Q\in\mathbb{C}[z_1,\ldots,z_n]$ be the unique polynomail that satisfies
\[ B=Q(\sigma_1,\ldots,\sigma_n). \]
Then if $f$ is a transversal endomorphism of $\P^n$ of degree $d$ we have
\begin{equation}\label{eq:indicesPn}
\sum_{f(p)=p}\frac{B(J_f(p))}{\det(I-J_f(p))} = Q(-d,(-d)^2,\ldots,(-d)^n).
\end{equation}
\end{application}
Note that if in particular $B=\sigma_k$ then the right hand side of the above formula is simply $(-1)^k d^k$ and we recover equation (\ref{eq:LefschetzPn}). This result refines the main result in \cite{Guillot2004} by computing the value of the sum of indicies, and is also in accordance with the results in \cite{Ueda1995,Abate2014}.
\begin{proof}[Sketch of the proof]
Both side of equation (\ref{eq:indicesPn}) depend linearly on the polynomial $B$, hence it is enough to prove the equality in the case $B$ is a monic monomial. Let us assume then that $B$ is given by
\[ B=\sigma_1^{a_1}\ldots\sigma_n^{a_n}. \]
Since $B$ has degree at most $n$ we must have $a_1+2a_2+\ldots+na_n\leq n$. Define \[ \mathscr{F}=(\Omega^1_{\P^n})^{\otimes a_1}\otimes\ldots\otimes(\Omega^n_{\P^n})^{\otimes a_n}, \]
and $\varphi=(df)^{\otimes a_1}\otimes\ldots\otimes(\wedge^n df)^{\otimes a_n} \colon f^*\mathscr{F}\to\mathscr{F}$ and apply the Woods Hole formula. Since the trace of a tensor product is the product of the traces it is clear that the trace of $\varphi(p)$ acting on $\mathscr{F}(p)$ will be precisely $B(J_f(p))$. Let $k=\operatorname{deg}\,B=a_1+2a_2+\ldots+na_n$. We need to prove now that the Lefschetz number $\operatorname{L}(f,\varphi,\mathscr{F})$ equals $Q(-d,(-d)^2,\ldots,(-d)^n)=(-d)^{a_1}\ldots(-d)^{na_n}=(-d)^k$. Formula (\ref{eq:indicesPn}) follows immediately from the next lemma.
\begin{lemma}\label{lemma:cohomology}
If $k=a_1+2a_2+\ldots+na_n\leq n$ then the natural map
\[ \wedge \colon \mathscr{F}=(\Omega^1_{\P^n})^{\otimes a_1}\otimes\ldots\otimes(\Omega^n_{\P^n})^{\otimes a_n} \longrightarrow \Omega^k_{\P^n} \]
induces isomorphisms in all cohomology groups. In particular, the cohomology of $\mathscr{F}$ vanishes in all degrees except $k$ and $H^k(\P^n,\mathscr{F})\cong\mathbb{C}$. Moreover the following diagram commutes:
\[ \xymatrix{
H^k(\P^n,\mathscr{F}) \ar[r]^{\varphi} \ar[d]_{\wedge} & H^k(\P^n,\mathscr{F}) \ar[d]^{\wedge} \\
H^k(\P^n,\Omega^k_{\P^n}) \ar[r]_{\wedge^k df} & H^k(\P^n,\Omega^k_{\P^n}) .
}
\]
\end{lemma}
Since $\wedge\colon H^k(\P^n,\mathscr{F})\to H^k(\P^n,\Omega^k_{\P^n})$ is an isomorphism it follows from the commutativity of the above diagram that $\operatorname{tr}(\varphi,\,H^k(\P^n,\mathscr{F}))=\operatorname{tr}(\wedge^k df,\,H^k(\P^n,\Omega^k_{\P^n}))=d^k$ and so $\operatorname{L}(f,\varphi,\mathscr{F})=(-1)^k d^k$, which is what we wanted to prove.
\end{proof}
The main step in the proof of \hyperref[lemma:cohomology]{Lemma \ref*{lemma:cohomology}} is proving that $\mathscr{F}$ actually has the same cohomology groups as $\Omega^k_{\P^n}$ (once this is achieved it is not hard to see that the map $\wedge$ induces an isomorphism in degree $k$ cohomology which is the only non-trivial degree). Such step is proved by an induction argument -- we shall prove it here only for $\mathscr{F}=\Omega^p_{\P^n}\otimes\Omega^q_{\P^n}$.
\begin{proof}[Proof of Lemma \ref*{lemma:cohomology}]
Let us prove that $\Omega^p_{\P^n}\otimes\Omega^q_{\P^n}$ and $\Omega^{p+q}_{\P^n}$ have isomorphic cohomology groups by inducting on $p$. The statement is clearly true for $p=0$. Assume now that it holds for $p-1$ and consider the generalized Euler sequence
\begin{equation}\label{eq:Eulersequence}
0 \longrightarrow \Omega^p_{\P^n} \longrightarrow E\otimes \O_{\P^n}(-p) \longrightarrow \Omega^{p-1}_{\P^n} \longrightarrow 0,
\end{equation}
where $E$ denotes the trivial vector bundle of rank $\binom{n+1}{p}$. Since $\Omega^q_{\P^n}$ is locally free, the stalks $(\Omega^q_{\P^n})_x$ are flat $\O_{\P^n,x}$-modules and so tensoring (\ref{eq:Eulersequence}) with the sheaf $\Omega^q_{\P^n}$ preserves the exactness of the sequence. We obtain
\[ 0 \longrightarrow \Omega^p_{\P^n}\otimes\Omega^q_{\P^n} \longrightarrow E\otimes \Omega^q_{\P^n}(-p) \longrightarrow \Omega^{p-1}_{\P^n}\otimes\Omega^q_{\P^n} \longrightarrow 0. \]
It is known that if $p+q\leq n$ then $\Omega^q_{\P^n}(-p)$, and thus $E\otimes\Omega^q_{\P^n}(-p)$, has vanishing cohomology in every dimension. This implies that the connecting homomorphisms
\[ \delta\colon H^k(\P^n,\Omega^{p-1}_{\P^n}\otimes\Omega^q_{\P^n})\to H^{k+1}(\P^n,\Omega^p_{\P^n}\otimes\Omega^q_{\P^n}) \]
are isomorphisms for all $k$. This proves that that $\Omega^p_{\P^n}\otimes\Omega^q_{\P^n}$ and $\Omega^{p+q}_{\P^n}$ have isomorphic cohomology groups.
Lastly, the composition
\[ \xymatrix{
H^p(\P^n,\Omega^p_{\P^n}) \otimes H^q(\P^n,\Omega^q_{\P^n}) \ar[r]^-{\cup} & H^{p+q}(\P^n,\Omega^p_{\P^n}\otimes \Omega^q_{\P^n}) \ar[r]^-{\wedge} & H^{p+q}(\P^n,\Omega^{p+q}_{\P^n})
}
\]
of the cup product and the homomorphism induced by $\wedge$ can be understood as the intersection of analytic cycles in the Chow ring of projective space and is thus an isomorphism. We conclude that $\wedge$ induces isomorphisms in $H^{p+q}$ and hence in all cohomology groups.
\end{proof}
\section{The Baum-Bott theorem for foliations on \texorpdfstring{$\P^2$}{P2}} \label{sec:foliations}
There are two important index theorems for foliations on $\P^2$: the Baum-Bott theorem and the Camacho-Sad theorem for invariant curves. The former can be reduced, under a non-degeneracy assumption, to a particular case of the Woods Hole formula. In order to do so we need to be able to associate to any holomorphic foliation an endomorphism of $\P^2$.
\begin{remark}
Even though we will be able to associate an endomorphism of $\P^2$ to any holomorphic foliation, the choice will not be unique. Also, it is important to point out that a curve that is invariant for the foliation need not be invariant for the associated endomorphism. Indeed, invariant curves are extremely rare for endomorphisms and in particular a smooth, totally invariant curve has to be a line \cite{FornaessSibony1994,CerveauLinsNeto2000}. In the very particular case of an exceptional line we can recover the Camacho-Sad theorem from the Woods Hole formula.
\end{remark}
We now proceed as follows: let $\mathcal{F}$ be a holomorphic foliation on $\P^2$ having only isolated non-degenerate singularities. We can always choose a homogeneous vector field $\xi=P_0\frac{\partial}{\partial x_0} + P_1\frac{\partial}{\partial x_1} + P_2\frac{\partial}{\partial x_2}$ on $\mathbb{C}^3\setminus\{0\}$ that descends to a (singular) line field on $\P^2$ which induces foliation $\mathcal{F}$. The choice is not unique, but every two such choices differ only by a multiple of the radial vector field. Define an endomorphism $f_\xi$ of $\P^2$ by setting
\[ f_\xi(x_0,x_1,x_2) = [P_0:P_1:P_2] .\]
It is not hard to see that the fixed points of $f_\xi$ correspond to the singularities of $\mathcal{F}$ and thus all these are simple and isolated. Moreover we have the following fundamental observations:
\begin{proposition}\label{prop:fundamental}
In affine coordinates $(x,y)=(\frac{x_1}{x_0},\frac{x_2}{x_0})$ foliation $\mathcal{F}$ is induced by
\[ (P_1(1,x,y)-xP_0(1,x,y)\frac{\partial}{\partial x} + (P_2(1,x,y)-yP_0(1,x,y)\frac{\partial}{\partial y}. \]
Therefore if $p=[p_0:p_1:p_2]$ is a singular point of $\mathcal{F}$ with $p_0\neq 0$ there exists a small open neighborhood of $p$ where $P_0(1,x,y)$ does not vanish and so foliation $\mathcal{F}$ is induced by the vector field
\[ \tilde{\xi} = \left(\frac{P_1(1,x,y)}{P_0(1,x,y)}-x\right)\frac{\partial}{\partial x} + \left(\frac{P_2(1,x,y)}{P_0(1,x,y)}-y\right)\frac{\partial}{\partial y}, \]
locally around $p$.
\end{proposition}
Note that around such a singular point $p$ the endomorphism $f_\xi$ is locally given by $(x,y)\mapsto (\frac{P_1}{P_0},\frac{P_2}{P_0})$ and thus we have the relation
\[ D\tilde{\xi}(p)=Df_\xi (p)-I. \]
\begin{application}[Baum-Bott theorem]\label{app:BB}
Let $\mathcal{F}$ be a homlomorphic foliation of degree $d$ on $\P^2$ having isolated non-degenerate singularities only. Then
\[ \sum_{\operatorname{Sing }\mathcal{F}}\operatorname{BB}(\mathcal{F},p)=(d+2)^2. \]
\end{application}
\begin{proof}
Let $\mathscr{F}=\Omega^1_{\P^2}\otimes\Omega^1_{\P^2}$ and $\varphi$ the lift of $f_\xi$ to $\mathscr{F}$ given by $\varphi=df_\xi\otimes df_\xi$ and apply the Woods Hole formula.
\end{proof}
\section{Indices for polynomial vector fields on \texorpdfstring{$\mathbb{C}^2$}{C2}} \label{sec:vectorfields}
We have seen in the previous section that whenever we have a foliation $\mathcal{F}$ on $\P^2$ we can always associate to it a homogeneous vector field $\xi$ on $\mathbb{C}^3$ and an endomorphism $f_\xi$ of $\P^2$, however, this endomorphism is not uniquely determined by $\mathcal{F}$. Suppose now that we are given a degree $d$ polynomial vector field $v=P(x,y)\frac{\partial}{\partial x}+Q(x,y)\frac{\partial}{\partial y}$ on $\mathbb{C}^2$. In this case there is a natural choice for the associated endomorphism given by
\begin{equation} \label{eq:fv}
f_v(x_0,x_1,x_2) = [x_0^d:\tilde{P}(x_0,x_1,x_2)+x_0^{d-1}x_1:\tilde{Q}(x_0,x_1,x_2)+x_0^{d-1}x_2],
\end{equation}
where $\tilde{P}$ and $\tilde{Q}$ are the homogenizations of $P,Q$ with respect to the variable $x_0$. At every point $p$ on $\mathbb{C}^2$ we have the relation
\[ Dv(p) = Df_v(p)-I. \]
Moreover, if $v$ has an invariant line at infinity and non-degenerate singularities then $f_v$ has simple isolated fixed points only. With this construction we can now prove the Euler-Jacobi formulae as a particular case of the Woods Hole formula.
\begin{application}[$1^{\mathrm{st}}$ Euler-Jacobi relation]\label{app:EJ1}
Let $v$ be a polynomial foliation on $\mathbb{C}^2$ of degree $d$ having $d^2$ non-degenerate singularities. Then
\[ \sum_{v(p)=0}\frac{1}{\det(Dv(p))}=0. \]
\end{application}
\begin{proof}
Let $L$ denote the line at infinity and let $\mathcal{I}_L$ denote the sheaf of holomorphic functions on $\P^2$ that vanish on $L$. Under the hypotheses of the theorem the endomorphism $f_v$ has degree $d$, preserves the line $L$, and the sheaf $f^*\mathcal{I}_L$ equals $\mathcal{I}_L^{\otimes d}$ (i.e.~the sheaf of functions of vanishing order at least $d$ along $L$). There is a natural inclusion $\iota\colon f^*\mathcal{I}_L\cong \mathcal{I}_L^{\otimes d} \to \mathcal{I}_L$. Applying the Woods Hole formula to the particular case $\mathscr{F}=\mathcal{I}_L$, $\varphi=\iota$ gives the Euler-Jacobi formula stated above.
\end{proof}
\begin{application}[$2^{\mathrm{nd}}$ Euler-Jacobi relation]\label{app:EJ2}
Let $v$ be a polynomial foliation on $\mathbb{C}^2$ of degree $d$ having $d^2$ non-degenerate singularities. Then
\[ \sum_{v(p)=0}\frac{\operatorname{tr}(Dv(p))}{\det(Dv(p))}=0. \]
\end{application}
\begin{proof}
Consider $\mathscr{F}=\Omega^1_{\P^2}\otimes\mathcal{I}_L$, the sheaf of 1-forms on $\P^2$ that are identically zero on $L$. The second Euler-Jacobi relation follows from the Woods Hole formula in the particular case where $\mathscr{F}=\Omega^1_{\P^2}\otimes\mathcal{I}_L$, $\varphi = df\otimes\iota$.
\end{proof}
Again, let $L$ denote the line at infinity and assume $v$ is non-dicritic. From the expression for $f_v$ in (\ref{eq:fv}) it is clear that the line $L$ is totally invariant for the endomorphism $f_v$ (that is, $f_v^{-1}(L)=L$). Under this special situation we can recover the Camacho-Sad theorem from the Woods Hole formula.
\begin{application}[Camacho-Sad theorem, particular case]\label{app:CS}
If the foliation $\mathcal{F}_v$ defined by $v$ on $\P^2$ has non-degenerate singularities along the invariant line at infinity $L$ we have
\[ \sum_{L\cap\operatorname{Sing }\mathcal{F}_v}\operatorname{CS}(\mathcal{F}_v,L,p)=1. \]
\end{application}
\begin{proof}
Let $N_L$ denote the normal bundle of $L$ in $\P^2$ and recall that $\operatorname{deg }N_L=N_L\cdot N_L=1$. Let $\mathcal{N}^*_L$ denote the conormal sheaf of $L$, that is, the sheaf of sections of the dual of the normal bundle $N_L$. A section of $\mathcal{N}^*_L$ can be understood as a holomorphic 1-form on $\P^2$ that vanishes along the tangent bundle of $L$. If $L$ is totally invariant by $\mathcal{F}$ we have that $f_\xi(L)=L$ and if $\omega$ is a 1-form that vanishes along $TL$ so does $f^*\omega$. This implies that $df_\xi$ acts on $\mathcal{N}^*_L$ in a natural way. Let us apply Woods Hole on the line $L$ by making $\mathscr{F}=\mathcal{N}^*_L$ and lifting $f_\xi|_L$ to $\mathcal{N}^*_L$ by $\varphi=df_\xi$. The Woods Hole formula applied to this particular case implies the Camacho-Sad theorem.
\end{proof}
\section*{Table: Summary of applications}
\renewcommand{\arraystretch}{1.25}
\begin{center}
\begin{tabular}{ | l | c | c | c | c |}
\hline
Formula & Manifold $X$ & Map $f$ & Sheaf $\mathscr{F}$ & Lift $\varphi$ \\ \hline
Generalized Lefschetz & Any $X$ & $f$ transversal & $\Omega^k_X$ & $\wedge^k df$ \\ \hline
Guillot's relations & $\P^n$ & $f$ transversal & $\bigotimes(\Omega^k_{\P^n})^{\otimes a_k}$ & $\bigotimes (\wedge^k df)^{\otimes a_k}$ \\ \hline
Baum-Bott & $\P^2$ & $f_\xi$ & $\Omega^1_{\P^2}\otimes\Omega^1_{\P^2}$ & $df_\xi\otimes df_\xi$ \\ \hline
Euler-Jacobi 1 & $\P^2$ & $f_v$ & $\mathcal{I}_L$ & $\iota$ \\ \hline
Euler-Jacobi 2 & $\P^2$ & $f_v$ & $\Omega^1_{\P^2}\otimes\mathcal{I}_L$ & $df\otimes\iota$ \\ \hline
Camacho-Sad & $L\hookrightarrow\P^2$ & $f_\xi|_L$ & $\mathcal{N}^*_L$ & $df_\xi$ \\ \hline
\end{tabular}
\end{center}
\vspace{4mm}
\bibliographystyle{alpha}
|
\section{Introduction}
\label{sec:intro}
The importance of force-field models is evident from recent hallmarks of atomistic force-field simulations in biology, such as the full characterization of $\beta$2 adrenergic receptor with a Markov state model \cite{kohlhoff2014} and the Anton 2 computer that can perform 10 $\mu$s molecular dynamics simulations per day on a system containing one million atoms. \cite{shaw2014} In many other domains, impressive scientific breakthroughs were also realized with atomistic force-field simulations, such as the virtual screening of $87\ 000$ zeolites for the selective adsorption of CO$_2$. \cite{kim2013}
An efficient and reliable model for electrostatic interactions is a fundamental component of a force-field model. For example, molecular recognition in proteins can be driven by electrostatic interactions. \cite{macchiarulo2003} Several authors studied partial charges derived from electronic wavefunctions to model electrostatic interactions in metal-organic frameworks. \cite{ramsahye2007, vanduyfhuys2012, kadantsev2013, gabrieli2015} Depending on the framework type, the values of the partial charges can strongly affect the predicted adsorption isotherms and self-diffusion coefficients. \cite{hamad2015} Energy decomposition methods have also shown that electrostatic interactions are one of the main driving forces in the formation of hydrogen bonds. \cite{wu2009}
In this work, we propose a new and transparent method to derive, from an electronic wavefunction, a robust, compact and reliable model for electrostatic interactions that is easily included in a force-field model. The goal is thus an efficient computation of electrostatic interactions between the molecules in the frozen density approximation, \cite{wesolowski1993, wu2009, tafipolsky2011} i.e.\ without accounting for induction or polarization effects. Although it is important and challenging to account for polarization in force fields, \cite{warshel2007, verstraelen2013_acks2, verstraelen2014} the development of polarizable force fields goes beyond the scope of this paper. We also do not consider so-called polarized force fields, where polarization is described effectively by computing the charges from an electronic structure calculation with a polarizable continuum model. \cite{ji2014}
The most basic and widespread electrostatic force-field model consists of interacting atomic point charges placed at the positions of the nuclei. In older works, e.g.\ the TraPPE force field for CO$_2$, \cite{potoff2001} the partial charges are fitted to experimental thermodynamic reference data. More recently, e.g.\ as in the TraPPE-EH models, \cite{rai2013} charges are often derived from electronic wavefunctions. Plenty of methods exist to compute such partial charges but usually, for force-field purposes, they are fitted to the electrostatic potential around model compounds of interest, \cite{fox1998} e.g.\ extensions of the AMBER force field often use the RESP method for partial charges. \cite{bayly1993} The point-charge model is only a very crude representation of the molecular charge distribution; it does not account for finer details such as atomic multipoles \cite{stone1985} and the spatial distribution of the electron cloud. The spatial distribution becomes important when electron densities of two atoms or molecules begin to overlap: in that regime point-multipole models neglect a relatively large attractive electrostatic force, which is known as the penetration effect. \cite{kairys1999, krapp2006, spackman2006, tafipolsky2011, lu2011, wang2015}
In principle, atomic multipoles are easily computed with an atoms-in-molecules (AIM) method. In some works, the acronym AIM is used exclusively for Quantum Theory of Atoms in Molecules \cite{bader1991} (QTAIM). Here it is used more generally, to refer to any method that partitions the molecular electron density, $\rho(\mathbf{r})$, into atomic contributions, $\rho_A(\mathbf{r})$, from which e.g.\ atomic multipole moments can be derived. The spatial distribution of the electron density is sometimes also modeled with density-fitting techniques, e.g.\ as in the Gaussian Electrostatic Model \cite{piquemal2006, cisneros2006} or related methods. \cite{misquitta2014, wang2014, ohrn2016} This leads to very accurate models of the electronic density but the use of such advanced charge distributions in force-field simulations poses some difficulties: the conformational dependence of atomic multipoles can be very complex and it is far from trivial to include torques acting on higher moments in a force-field model. Several authors have proposed methods to overcome these challenges, e.g.\ with rigid molecules \cite{leslie2008} or with machine learning methods. \cite{popelier2015} Such advanced techniques are not always feasible for large-scale simulations. In this work, we propose a mathematically elegant and compact approximation of the electron distribution that results in relatively accurate electrostatic interactions in force-field models, without compromising computational efficiency. Only spherically symmetric models for atoms are considered and generalizations toward non-spherical atoms will be studied in future work.
Our new method minimizes the Kullback-Leibler (KL) Divergence of a pro-density, a minimal expansion in atom-centered s-type Slater functions, from a given molecular electron density. \cite{ghillemijn2011, heidarzadeh2015} This approach is closely related to the Iterative Stockholder (IS) method, \cite{lillestolen2008} where the pro-molecule density is a sum of spherical non-negative pro-atom densities, without any restrictions on their radial dependence. As will be discussed in section \ref{sec:mbis}, the algorithm to optimize our pro-density parameters is also very similar to IS. Hence, we refer to our new method as Minimal Basis Iterative Stockholder (MBIS).
MBIS can be perceived in two different ways. In the first place, it is a variant of the Hirshfeld method: \cite{hirshfeld1977} a partitioning of the molecular electron density inspired by information theory. \cite{nalewajski2000} Second, it can also be seen as a density fitting technique that uses the KL-divergence, \cite{heidarzadeh2015} instead of the more common least-squares approach with a Coulomb metric, \cite{elking2010} to optimize the model density. This duality permits many applications, also beyond the scope of modeling electrostatic interactions. For example, the Hirshfeld method is extensively used in different dispersion corrections for Density Functional Theory computations. \cite{becke2007, tkatchenko2009, steinmann2011} Furthermore, AIM populations are widely used in conceptual density functional theory to compute condensed reactivity indicators. \cite{geerlings2003}
Several related AIM methods were proposed in the literature, each trying to improve certain properties of their predecessors. The original Hirshfeld method \cite{hirshfeld1977} has some well-known weaknesses, such as the relatively low partial charges \cite{davidson1992} and some deficiencies in its motivation from information theory. \cite{bultinck2007}
These issues were mostly fixed in the Iterative Hirshfeld (HI) method: \cite{bultinck2007} charges computed with this method reproduce well the electrostatic potential around a molecule. \cite{vandamme2009, verstraelen2009, verstraelen2011} Compared to ESP-fitted charges, HI charges are also relatively robust with respect to conformational changes, choice of basis set, etc. \cite{bultinck2007_basis, verstraelen2011} Unfortunately, also Iterative Hirshfeld has its deficiencies. For example, when the method is applied to highly polar oxides, it requires the spherically averaged density of the non-existing oxygen dianion as input. \cite{verstraelen2013} When this dianion density is computed with a localized basis set, iterative Hirshfeld charges severely overestimate electrostatic potentials of metal oxides. \cite{verstraelen2012_silica, verstraelen2013}
The Iterative Stockholder (IS) analysis was developed independently from the Iterative Hirshfeld method and it addresses most of the issues mentioned so far. \cite{lillestolen2009, verstraelen2013} However, IS charges are not very robust with respect to conformational changes, similar to ESP-fitted charges. \cite{verstraelen2012} A recent analysis revealed that the lack of robustness is strongly related to the ill-defined density tails of the IS pro-atoms, while the core region of the IS pro-atom is usually well defined. \cite{misquitta2014} Several authors have presented solutions to overcome the weaknesses of the Iterative Hirshfeld and Iterative Stockholder methods. \cite{manz2010, manz2012, verstraelen2012, verstraelen2013, vanpoucke2013, misquitta2014} A general difficulty with these recent efforts is that they all significantly increase the algorithmic complexity and/or introduce many tuned parameters that are needed as extra input for the partitioning. In this work, we will reverse this trend and propose a method that is mathematically elegant, straightforward to implement for large systems and free from empirical input (like atomic radii) or pre-computed pro-atoms.
In the development of the MBIS method, we payed special attention to its applicability to condensed phases and extended systems. One of the applications of interest is the automatic derivation of environment specific force-field parameters for supramolecular systems \cite{cole2016} and porous materials. \cite{vanduyfhuys2012, haldoupis2012} In such applications, density partitioning is applied to DFT calculations of large atomistic models, from which force-field parameters are derived. Besides the obvious requirement that an accurate model for electrostatics must be obtained, it is also essential that the atoms-in-molecules method is computationally feasible for large systems. In practice, this means that the computational cost must scale linearly with the system size. This is achieved in MBIS by using only well-behaved integrals over atomic regions whose cost is independent of the system size.
The paper is organized as follows. In section \ref{sec:mbis}, the Minimal Basis Iterative Stockholder (MBIS) method is derived using arguments from information theory, followed by more practical aspects such as numerical algorithms and software implementations. Section \ref{sec:example} showcases typical MBIS results with two brief applications. Section \ref{sec:comp} compares MBIS to 14 other AIM methods, assessing the robustness of charges and the accuracy of electrostatic potentials and electrostatic interactions. Some specific advantages of MBIS over (Iterative) Hirshfeld, are presented in section \ref{sec:disp}, by testing different variants of the Tkatchenko-Scheffler dispersion model. \cite{tkatchenko2009} Finally, our conclusions and an outlook on future work are given in section \ref{sec:concl}.
\section{Minimal Basis Iterative Stockholder Method}
\label{sec:mbis}
\subsection{Information theory approach to Hirshfeld partitioning}
It is instructive to review the information theory arguments \cite{nalewajski2000} that support the Hirshfeld method. \cite{hirshfeld1977} The amount of information lost when atoms-in-molecules (AIM) densities are approximated by pro-atoms, can be expressed as the sum of the KL-divergence for every atom:
\begin{equation}
\label{eq:kld}
\Delta S [\{\rho_A\};\{\rho_A^0\}]
= \sum_{A=1}^{N_\text{atoms}} \int \rho_A(\mathbf{r}) \ln\frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} d\mathbf{r},
\end{equation}
Traditionally, the pro-atoms, ${\rho_A^0}$, are fixed and the AIM densities, ${\rho_A}$, are the unknowns to be determined. In the original Hirshfeld method, spherically averaged isolated neutral atoms are used as pro-atoms. To obtain AIM densities that are maximally similar to the pro-atoms, one minimizes the information loss with the constraint that the AIM densities have to add up to the total density: $\sum_{A=1}^{N_\text{atoms}} \rho_A(\mathbf{r}) = \rho(\mathbf{r})$. Hence, the optimal AIM densities are a stationary point of the following Lagrangian:
\begin{equation}
\label{eq:l0}
\begin{split}
& L_0[\{\rho_A\},\lambda(\mathbf{r});\{\rho_A^0\}] = \\
& \qquad \sum_{A=1}^{N_\text{atoms}} \int \rho_A(\mathbf{r}) \ln\frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} d\mathbf{r} \\
& \quad + \int \lambda(\mathbf{r}) \left( \sum_{A=1}^{N_\text{atoms}} \rho_A(\mathbf{r}) - \rho(\mathbf{r}) \right) d\mathbf{r}.
\end{split}
\end{equation}
with the Lagrange multiplier $\lambda(\mathbf{r})$ and with fixed pro-atom densities $\rho_A^0(\mathbf{r})$. The Lagrange equations take the following form
\begin{equation}
0 = \frac{\delta L_0}{\delta \rho_A(\mathbf{r})} = \ln \frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} + 1 + \lambda(\mathbf{r}) \quad \forall A.
\end{equation}
The solution is:
\begin{equation}
\frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} = \frac{\rho_B(\mathbf{r})}{\rho_B^0(\mathbf{r})} \quad \forall A \neq B.
\end{equation}
After multiplication by $\rho_A^0(\mathbf{r}) \rho_B^0(\mathbf{r})$ and summing over all atoms $B$, one obtains the well-known stockholder partitioning: \cite{nalewajski2000}
\begin{align}
\label{eq:stockholder}
\rho_A(\mathbf{r}) = \rho(\mathbf{r}) \frac{\rho_A^0(\mathbf{r})}{\rho^0(\mathbf{r})}
\quad \text{with} \quad \rho^0(\mathbf{r}) = \sum_B \rho_B^0(\mathbf{r}),
\end{align}
which corresponds to the definition originally given by Hirshfeld. \cite{hirshfeld1977} The name \textit{stockholder} comes from the ratio $\rho_A^0(\mathbf{r}) / \rho^0(\mathbf{r})$: at every point in space it represents the \textit{share} of pro-atom $A$ in the total pro-density. It can be interpreted as an atomic weight function that assigns part of the total electron density to atom $A$. In most Hirshfeld variants,\cite{hirshfeld1977, bultinck2007, lillestolen2009, verstraelen2012, verstraelen2013} the weight function varies smoothly over the range $[0,1]$. In QTAIM, \cite{bader1991} a similar atomic weight function, derived from the topology of $\rho(\mathbf{r})$, is either 1 inside the atomic basin or 0 elsewhere. Due to the minimization of the KL-divergence, the Hirshfeld AIM densities are maximally similar to the pro-atoms, ensuring some degree of transferability between AIM densities in different molecules. \cite{ayers2000}
The use of fixed pro-atoms has some important disadvantages. Results obtained with the Hirshfeld partitioning method depend largely on the choice of the fixed pro-atoms, which is essentially arbitrary. \cite{bultinck2007} Furthermore, $\rho_A^0(\mathbf{r})$ and $\rho_A(\mathbf{r})$ do not necessarily have the same norm ($N_A^0 \neq N_A$), such that the KL-divergence cannot be used as a proper measure for information loss. \cite{parr2005} This shortcoming was one of the motivations to develop the iterative Hirshfeld (HI) method. \cite{bultinck2007} In HI, the pro-atoms are not fully fixed a priori but rather updated iteratively to achieve consistency between the charge of the pro-atom and the AIM density.
\subsection{Definition of the MBIS partitioning}
In this paper, we will make use of the information theory concepts reviewed in the previous subsection, yet with a different model for the pro-atomic density:
\begin{equation}
\rho_A^0(\mathbf{r}) = \sum_{i=1}^{m_A} \rho_{Ai}^0(\mathbf{r}),
\end{equation}
with
\begin{equation}
\label{eq:proslater}
\rho_{Ai}^0(\mathbf{r}) = N_{Ai} f_{Ai}(\mathbf{r}) = \frac{N_{Ai}}{\sigma_{Ai}^3 8\pi} \exp\left(-\frac{|\mathbf{r}-\mathbf{R}_A|}{\sigma_{Ai}}\right),
\end{equation}
where the number of Slater functions, $m_{A}$, is the number of shells of atom $A$, i.e.\ its row in the periodic table. Both the population, $N_{Ai}$, and the width, $\sigma_{Ai}$, of each atomic shell are free variables. The shape functions, $f_{Ai}(\mathbf{r})$, are normalized 1s Slater-type density functions ($\int f_{Ai}(\mathbf{r}) d\mathbf{r} = 1$) and hence the population of a pro-atom is simply $N_A^0 = \sum_{i=1}^{m_A} N_{Ai}$. Figure \ref{fig:concept}a illustrates the expansion of the density in Slater functions, for the case of a carbon dioxide molecule.
It is clear that the pro-atom parameterization with s-type Slater functions is only applicable to (reconstructed) all-electron densities. Regardless of this requirement, the MBIS method has many advantages over existing methods, as will be extensively shown in the remainder of the paper. Future work will focus on more advanced pro-atom models, e.g.\ to make them also suitable for pseudo densities, while still maintaining a numerically robust algorithm. In this work, only the most minimal, yet very effective, parameterization of the pro-atoms is considered.
\begin{figure}
\includegraphics{concept_small.pdf}
\caption{(a) An expansion of the molecular electron density of carbon dioxide (solid gray curve, $\rho$) in a minimal number of 1s Slater-type density functions (black dashed curves, $\rho^0_{Ai}$). The sum of all Slater functions is the pro-molecular density (solid black curve, $\rho^0$). (b) A reduction suitable for force-field models: for every atom $A$, the nuclear charge and the core Slater functions are condensed into an effective core charge (solid vertical line, $q_{A,c}$), while the valence Slater function (dashed black curve, parameters $N_{A,v}$ and $\sigma_{A,v}$) is retained.}
\label{fig:concept}
\end{figure}
All the pro-atom parameters, $\{N_{Ai}\}$ and $\{\sigma_{Ai}\}$, and the AIM densities, $\rho_A(\mathbf{r})$ will be optimized by minimizing the information loss. The main difference with the conventional Hirshfeld method is that also a set of pro-atom parameters is varied, such that the pro-atom densities become a good approximation of the AIM densities. These additional degrees of freedom also allow us to constrain the population of each pro-atom and corresponding AIM to be equal, avoiding any ambiguity in the statistical interpretation of Eq.\ \eqref{eq:kld}. \cite{parr2005} The Lagrangian for this problem is an extension of Eq.\ \eqref{eq:l0} with additional variables and constraints:
\begin{equation}
\label{eq:l1}
\begin{split}
& L_1[\{\rho_A\}, \lambda(\mathbf{r}), \{N_{Ai}\}, \{\sigma_{Ai}\}, \{\mu_A\}] = \\
& \qquad \sum_{A=1}^{N_\text{atoms}} \int \rho_A(\mathbf{r})
\ln\frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} d\mathbf{r} \\
& \quad + \int \lambda(\mathbf{r}) \left(
\sum_{A=1}^{N_\text{atoms}} \rho_A(\mathbf{r}) - \rho(\mathbf{r})
\right) d\mathbf{r} \\
& \quad + \sum_{A=1}^{N_\text{atoms}} \mu_A \int \bigl(
\rho^0_A(\mathbf{r}) - \rho_A(\mathbf{r}) \bigr) d\mathbf{r},
\end{split}
\end{equation}
where $\mu_A$ are new Lagrange multipliers associated with the consistency of the pro-atom and AIM populations.
Independent variation of the Lagrangian $L_1$ with respect to each variable ($\rho_A(\mathbf{r})$, $N_{Ai}$ or $\sigma_{Ai}$) leads to a set of Lagrange equations, which, together with the constraints, determine the MBIS AIM and pro-atom densities and the Lagrange multipliers $\lambda(\mathbf{r})$ and $\{\mu_A\}$.
We first consider the derivative of $L_1$ toward $N_{Ai}$:
\begin{align}
0 = \frac{\partial L_1}{\partial N_{Ai}}
& = \int \frac{\delta L_1}{\delta \rho_A^0(\mathbf{r})} \frac{\partial \rho_A^0(\mathbf{r})}{\partial N_{A,i}} d\mathbf{r} \\
& = \int \left( -\frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} + \mu_A \right) f_{Ai}(\mathbf{r}) d\mathbf{r} \\
& = \mu_A - \int \frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} f_{Ai}(\mathbf{r}) d\mathbf{r}. \label{eq:lvarn}
\end{align}
When we multiply by $N_{Ai}$ and sum over the shells $i$ of atom $A$, we get:
\begin{align}
0 = \sum_i^{m_A} N_{Ai} \frac{\partial L_1}{\partial N_{Ai}}
& = \mu_A N_A^0 - \int \frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} \sum_i^{m_A} N_{Ai} f_{Ai}(\mathbf{r}) \\
& = \mu_A N_A^0 - N_A.
\end{align}
Due to the constraint $N_A = N_A^0$, we have $\mu_A = 1$ for each atom. Next, we take the functional derivative of $L_1$ toward $\rho_A(\mathbf{r})$ and make use of $\mu_A = 1$:
\begin{equation}
0 = \frac{\delta L_1}{\delta \rho_A(\mathbf{r})} =
\ln \frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} + \lambda(\mathbf{r}),
\end{equation}
whose solution is the stockholder partitioning formula in Eq.~\eqref{eq:stockholder}. Finally, we consider the derivative of $L_1$ toward $\sigma_{Ai}$:
\begin{align}
0 = \frac{\partial L_1}{\partial \sigma_{Ai}}
& = \int \frac{\delta L_1}{\delta \rho_A^0(\mathbf{r})} \frac{\partial \rho_A^0(\mathbf{r})}{\partial \sigma_{A,i}} d\mathbf{r} \\
& = -\int \frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} \frac{\partial \rho_A^0(\mathbf{r})}{\partial \sigma_{A,i}}
+ \mu_A N_{Ai} \frac{\partial}{\partial \sigma_{Ai}} \int f_{Ai}(\mathbf{r}) d\mathbf{r} \\
\label{eq:lvarsigma}
& = \int \frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} \left( \frac{3}{\sigma_{Ai}} - \frac{|\mathbf{r} - \mathbf{R}_{A}|}{\sigma_{Ai}^2} \right) \rho_{Ai}^0(\mathbf{r}) d\mathbf{r},
\end{align}
where we made use of $\int f_{Ai} (\mathbf{r}) d\mathbf{r} = 1$.
The AIM densities can be eliminated from the Lagrange equations \eqref{eq:lvarn} and \eqref{eq:lvarsigma} by making use of $\frac{\rho_A(\mathbf{r})}{\rho_A^0(\mathbf{r})} = \frac{\rho(\mathbf{r})}{\rho^0(\mathbf{r})}$. They can be rewritten in the following form:
\begin{align}
\label{eq:updaten}
N_{Ai} &= \int \rho(\mathbf{r})
\frac{\rho_{Ai}^0(\mathbf{r})}{\rho_0(\mathbf{r})}
d\mathbf{r}, \\
\label{eq:updatesigma}
\sigma_{Ai} &= \frac{1}{3N_{Ai}}
\int \rho(\mathbf{r})
\frac{\rho_{Ai}^0(\mathbf{r})}{\rho_0(\mathbf{r})}
|\mathbf{r} - \mathbf{R}_A|
d\mathbf{r}.
\end{align}
These identities form the basis for the self-consistent algorithm that will be explained in the next subsection.
\subsection{Self-consistent algorithm}
Fig.\ \ref{fig:flowchart} depicts a flow chart of the self-consistent algorithm discussed in this subsection. The individual steps are described in more detail below.
\begin{figure}
\includegraphics{flow_chart.pdf}
\caption{Flow chart of the self-consistent algorithm for the refinement of the MBIS parameters.}
\label{fig:flowchart}
\end{figure}
In order to find all the pro-atom parameters, $\{N_{Ai}\}$ and $\{\sigma_{Ai}\}$, an initial guess is generated first, which will be refined later. Because the parameters $\{\sigma_{Ai}\}$ are non-linear, it is not guaranteed that $L_1$ is convex or has a unique minimum. Hence, multiple stationary points may exist and a reasonable initial guess is needed to find the solution of interest. The initial values of the parameters $\{N_{Ai}\}$ of atom $A$ are set to the number of electrons in each shell of the corresponding neutral isolated atom. The initial guess of $\{\sigma_{Ai}\}$ is inspired by hydrogenic s-type orbitals. For the innermost and outermost shell of atom $A$, we take $a_0/2Z_A$ and $a_0/2$, respectively, where $Z_A$ is the atomic number. Initial values of $\sigma_{Ai}$ for the intermediate shells are fixed by geometric interpolation: $\sigma_{Ai} = a_0/2 Z_A^{\left(1-\frac{i-1}{m_A-1}\right)}$.
Given the initial guess, the parameters are refined iteratively with a self-consistent update. In a single iteration, Eqs.\ \eqref{eq:updaten} and \eqref{eq:updatesigma} are evaluated, using the ``old'' parameters in the right-hand side, yielding the ``new'' parameters in the left-hand side. These iterations are repeated until the pro-atom parameters no longer change significantly. In this work, the iterative algorithm (of MBIS and other Iterative Hirshfeld flavors) is stopped after the root-mean-square deviation between the pro-atom densities of the last and the previous iteration drops below a threshold of $10^{-8} \text{ a.u}$:
\begin{equation}
\max_A \sqrt{\int d\mathbf{r} [ \rho_{A,\text{last}}^0(\mathbf{r}) - \rho_{A,\text{previous}}^0(\mathbf{r}) ]^2} < 10^{-8}.
\end{equation}
Other convergence criteria could be used as well, e.g. based on the gradient of the Lagrangian $L_1$.
It should also be possible to optimize the pro-atom parameters with a quasi-Newton optimizer. However, robust quasi-Newton optimizers that can handle various types of equality and inequality constraints (to fix the total population and to keep all parameters positive) are non-trivial. The algorithm sketched above satisfies all constraints at every iteration, is much easier to implement and converges smoothly, even with very tight convergence settings. The Levenberg-Marquardt algorithm is not applicable because it is specifically designed for least-squares objective functions, while MBIS uses the KL-divergence as the objective function.
The integrals in Eqs.\ \eqref{eq:updaten} and \eqref{eq:updatesigma} must be evaluated numerically. Because the integrands are well localized on one atom, it is possible to implement the self-consistent update with a cost that scales linearly with the number of atoms.
\subsection{Relevant pro-atom parameters for modeling electrostatic interactions with force fields}
\label{subsec:ff}
After the optimization of pro-atom parameters, one may reduce the pro-density to a simpler picture, which is suitable for force-field models. The nuclear charge and the Slater functions associated with core electrons can be condensed into a single effective \textit{core charge}, $q_{A,c}$. This is illustrated in Fig.\ \ref{fig:concept}b. The remaining \textit{valence} Slater function is characterized by two parameters, its \textit{valence population}, $N_{A,v}$, and its \textit{valence width}, $\sigma_{A,v}$. The net atomic charge, $q_A = q_{A,c} - N_{A,v}$ can be used to approximate long-range electrostatic interactions. The two remaining degrees of freedom can be used to model the penetration effect, \cite{kairys1999, krapp2006, spackman2006, tafipolsky2011, lu2011, wang2015} i.e.\ the deviation of the electrostatic interactions from the simple point-charge model when the electronic densities begin to overlap. It can be computed efficiently with analytic expressions for the Coulomb interaction between Slater densities. \cite{lu2011, ohrn2016}
\subsection{MBIS Implementation}
\label{subsec:impl}
In the remainder of this work, MBIS will be tested extensively with applications to theoretical electron densities of molecules and condensed phases. In these applications, the all-electron density is first computed on an integration grid suitable for the numerical evaluation of Eqs.\ \eqref{eq:updaten} and \eqref{eq:updatesigma}. The implementation of these numerical integrals differs significantly between isolated molecules and periodic systems.
For isolated molecules, all-electron densities are computed with Gaussian09 \cite{g09}, using Density Functional Theory (DFT). Different functionals and Gaussian basis sets were used, as will be explained in the following sections. The MBIS partitioning of isolated molecule densities is carried out with HORTON 2.0.0, \cite{horton} which uses a standard atom-centered Becke-Lebedev integration grid. \cite{becke1988} This implementation can be combined with any level of theory in Gaussian09 that produces an all-electron 1-particle reduced density matrix (1RDM) with the ``density=current'' option. Because Gaussian09 does not write out the 1RDM when relativistic corrections are used, our tests on isolated molecules are limited to molecules containing no elements heavier than krypton.
Electron densities of periodic crystals are computed with the Projector Augmented Wave (PAW) method \cite{blochl2003} as implemented in GPAW-0.11.0. \cite{mortensen2005, enkovaara2010, bahn2002} Integrals involving the all-electron density of periodic systems are carried out as follows. In the PAW formalism, the total electron density is separated in a smoothly varying part, denoted as $\tilde{\rho}(\mathbf{r})$, and a correction for every atom $A$ in the so-called augmentation sphere. \cite{blochl2003}
In GPAW, the smooth density is represented on an equidistant real-space grid and the corrections are evaluated on atom-centered grids in spherical coordinates. This combination of integration grids makes it possible to perform very accurate integrations involving (reconstructed) all-electron densities of periodic systems. Our second MBIS implementation can handle any type of integration grid and we therefore used the same grid structure as in GPAW for periodic calculations. The advanced numerical techniques in this implementation, such as linear-scaling computational cost \cite{lee2013} and convergence acceleration, will be discussed in future work.
\subsection{Relation to other partitioning methods}
The MBIS pro-atom model has been used previously, however not yet in the context of Hirshfeld partitioning. For example, a similar pro-atom model (with fixed parameters) was also used in an ESP fitting scheme. \cite{hu2007} A similar density model is also used in the Stewart-Slater method. \cite{gill1996} Although our pro-atom model is obviously inspired by Slater's work on atomic shielding constants, \cite{slater1930} the typical polynomial prefactors are omitted. This omission is inspired by the piece-wise exponential ansatz from statistical models for atomic densities. \cite{wang1977, wang1982, fernandezpacios1991} The reduced model for force-field applications in subsection \ref{subsec:ff} has also been used before in the development of force-field models. \cite{donchev2005, wang2014, ohrn2016}
Our approach is comparable to the Iterative Stockholder (IS) method. \cite{lillestolen2009} In IS, spherical pro-atoms are defined by generic radial functions without further restrictions in terms of density basis functions; in practice they are represented by function values on a radial grid. A self-consistent update, in the same spirit as Eqs.~\eqref{eq:updaten} and \eqref{eq:updatesigma}, guarantees that the optimal IS atoms minimize the KL-divergence over all possible spherically symmetric pro-atoms. \cite{bultinck2009, lillestolen2009} Even though this is a convex problem, a well-documented weakness of IS is that the density tails of the pro-atoms are ill-defined, which leads to numerical instabilities and poorly defined atomic charges. \cite{bultinck2009, verstraelen2012, misquitta2014} In MBIS, this is resolved by modeling the density tail of each atom with only a single Slater function, which is comparable to the BS-ISA+DF method. \cite{misquitta2014} The MBIS self-consistent update algorithm is also very similar to the iterative Hirshfeld (HI) algorithm. \cite{bultinck2007} The main difference with HI is that MBIS makes use of an analytic ansatz for each pro-atom with several parameters per atom, i.e.\ the populations and widths of all shells in each atom, while HI varies just one population parameter per atom and makes use of pre-computed isolated atom densities. Furthermore, HI cannot be derived by replacing in Lagrangian $L_1$ the MBIS pro-atom by its HI counterpart. \cite{ghillemijn2011}
It is also important to realize that density fitting \cite{baerends1973, dunlap1979, fonsecaguerra1998} is closely related to MBIS. This connection becomes clear by considering the following Lagrangian:
\begin{equation}
\label{eq:l2}
L_2[\{N_{Ai}\}, \{\sigma_{Ai}\}, \mu]
= \int \rho(\mathbf{r}) \ln\frac{\rho(\mathbf{r})}{\rho^0(\mathbf{r})} d\mathbf{r}
+ \mu \int \rho^0(\mathbf{r}) - \rho(\mathbf{r}) d\mathbf{r}.
\end{equation}
The self-consistent update equations \eqref{eq:updaten} and \eqref{eq:updatesigma} can also be derived from $L_2$. This shows that the optimal MBIS pro-atom parameters can also be found by minimizing the KL-divergence of the pro-molecule density, $\rho^0(\mathbf{r})$, from a given molecular density, $\rho(\mathbf{r})$. This interpretation is similar to density fitting in force-field development, \cite{piquemal2006, cisneros2006} except for the following two points. First, the MBIS pro-density is expanded in Slater functions while density-fitting techniques usually rely on contracted Gaussian functions, also with higher multipoles. \cite{elking2010} Second, MBIS uses the KL-divergence as a cost function to fit the pro-atom parameters, while conventional density-fitting makes use of a least-squares cost, often with a Coulomb metric. The least-squares cost function was also used in other related works, e.g.\ the least-squares analog of IS is known as Stewart atoms \cite{stewart1977} and the least-squares analog of MBIS is very similar to Stewart-Slater atoms. \cite{gill1996} Hybrid approaches, combining least-squares and KL-divergence cost functions, were also proposed, such as Hirshfeld-E, \cite{verstraelen2013} Gaussian ISA \cite{verstraelen2012} and BS-ISA+DF. \cite{misquitta2014} In the development of the MBIS method, a least-squares cost function was avoided because it was recently found to lead to non-local AIM densities. \cite{heidarzadeh2015} Finally, note that the pro-atom parameters are sufficient to construct monopolar electrostatic force fields. We therefore expect that the direct optimization of the pro-atoms with a Lagrangian similar to Eq.~\eqref{eq:l2}, i.e.\ without constructing AIM densities, can be an attractive alternative to conventional AIM methods.
\section{Example MBIS applications}
\label{sec:example}
This section provides two illustrative applications of the MBIS method. Their main purpose is to show the applicability of MBIS in very different scenarios and to provide the reader with some typical results. The first example discusses the robustness of MBIS and its compatibility with chemical intuition, when applied to rather extreme variations of the oxidation states of oxygen. The second example shows that MBIS is also sufficiently robust when studying subtle variations of the electron density of water between the gas, liquid and solid phase.
The examples below only illustrate the usefulness of the MBIS method. A more systematic assessment can be found in section \ref{sec:comp}.
\subsection{Oxygen in different oxidation states}
\label{subsec:oxygen}
In previous studies, Hirshfeld-I (HI) partitioning was criticized for its poor applicability to oxides. \cite{manz2010, verstraelen2012_silica, verstraelen2013} During the iterative convergence of the charges, HI requires reference densities for the oxygen dianion (or sometimes even trianion) in vacuum, which does not exist. \cite{manz2010} When the oxygen dianion is computed with a finite basis, one obtains a very diffuse density that is not representative for the oxygen atom in a molecule or crystal. This mismatch results in very large absolute values for the atomic charges in oxides, overestimating the polarity of oxide clusters or the electrostatic potential in solid oxides. \cite{manz2010, verstraelen2012_silica} Many modifications of HI were proposed to surmount this limitation, \cite{manz2010, manz2012, vanpoucke2013, verstraelen2013, bucko2013, bucko2014, gould2016} often using different (somewhat arbitrary) techniques for the computation of unstable anions. The MBIS method does not need (unstable) ion densities as input and one would therefore expect that it does not suffer from the same overpolarization issues as HI.
Table \ref{tab:examples} compares MBIS and HI results for the oxygen element, in a series of systems where the oxidation state of oxygen varies from -2 to +3: MgO, chabazite, MIL-53(Al), quartz, \ce{H2O}, \ce{CO2}, \ce{H2O2}, \ce{LiO2}, \ce{O3}, \ce{CO}, \ce{O2} and \ce{OF2}. Results for other elements in these systems are given as well for the sake of completeness. Also the isolated oxygen cation, atom and anion are included because these have different valence electron densities that result in different parameters for the outer shell in the MBIS pro-density. Atom types, which are used to differentiate all non-equivalent atoms, are defined in section S1 in the supporting information. All electron densities are computed at the PBE level of theory. For isolated molecules, Gaussian09 \cite{g09} was used with the 6-311+G(2df,p) basis and charges were computed with HORTON. \cite{horton} Electron densities of crystal unit cells were computed with GPAW, \cite{mortensen2005, enkovaara2010, bahn2002} using a grid spacing of $0.1 \textrm{ \AA}$ and charges were derived from the periodic densities with a second implementation of MBIS. (See subsection \ref{subsec:impl}.)
\renewcommand{\arraystretch}{1.0
\begin{table}
\centering
\caption{Comparison of HI and MBIS charges for a selection of molecules and solids. (See text.) Results are grouped per element and sorted by the MBIS charge within each group. ON stands for oxidation number. For the MBIS method, also the quantities from subsection \ref{subsec:ff} are reported: core charge ($q_{A,c}$), valence charge ($q_{A,v}$) and valence width ($\sigma_{A,v}$).}
\label{tab:examples}
\begin{tabular}{llrddddd}
\hline
Molecule or solid & Atom$_\text{(type)}$ & ON & \multicolumn{1}{c}{HI} & \multicolumn{1}{c}{MBIS} & \multicolumn{1}{c}{~} & \multicolumn{1}{c}{~} & \multicolumn{1}{c}{~} \\
~ & ~ & ~ & \multicolumn{1}{c}{$q$ [e]} & \multicolumn{1}{c}{$q$ [e]} & \multicolumn{1}{c}{$q_{A,c}$ [e]} & \multicolumn{1}{c}{$q_{A,v}$ [e]} & \multicolumn{1}{c}{$\sigma_{A,v}$ [\AA]} \\ \hline
MIL-53(Al) & H$_\text{ph}$ & 0 & 0.132 & 0.155 & 1.0 & -0.845 & 0.198 \\
\ce{H2O2} & H & 1 & 0.387 & 0.414 & 1.0 & -0.586 & 0.186 \\
\ce{H2O} & H & 1 & 0.436 & 0.443 & 1.0 & -0.557 & 0.187 \\ \hline
MIL-53(Al) & H$_\text{hy}$ & 1 & 0.565 & 0.519 & 1.0 & -0.481 & 0.175 \\
\ce{LiO2^{.}} & Li & 1 & 0.912 & 0.825 & 1.076 & -0.251 & 0.387 \\ \hline
MIL-53(Al) & C$_\text{pc}$ & 0 & -0.132 & -0.148 & 4.359 & -4.507 & 0.266 \\ \hline
MIL-53(Al) & C$_\text{ph}$ & 0 & -0.087 & -0.110 & 4.354 & -4.463 & 0.266 \\
\ce{CO} & C & 3 & 0.144 & 0.108 & 4.327 & -4.218 & 0.270 \\
MIL-53(Al) & C$_\text{ca}$ & 4 & 0.910 & 0.849 & 4.340 & -3.491 & 0.246 \\
\ce{CO2} & C & 4 & 0.847 & 0.863 & 4.340 & -3.477 & 0.243 \\ \hline
\ce{MgO} & O & -2 & -2.220 & -1.934 & 6.243 & -8.177 & 0.247 \\
Chabazite & O$_\text{b}$ & -2 & -1.497 & -1.261 & 6.335 & -7.596 & 0.222 \\
Chabazite & O$_\text{r}$ & -2 & -1.480 & -1.250 & 6.337 & -7.588 & 0.222 \\
MIL-53(Al) & O$_\text{hy}$ & -2 & -1.952 & -1.220 & 6.329 & -7.548 & 0.223 \\
Chabazite & O$_\text{x}$ & -2 & -1.431 & -1.219 & 6.344 & -7.563 & 0.220 \\
Quartz & O & -2 & -1.473 & -1.213 & 6.343 & -7.555 & 0.221 \\
\ce{O^{-}} & O & -1 & -1.0 & -1.0 & 6.194 & -7.194 & 0.246 \\
\ce{H2O} & O & -2 & -0.872 & -0.885 & 6.333 & -7.219 & 0.220 \\
MIL-53(Al) & O$_\text{ca}$ & -2 & -0.781 & -0.748 & 6.354 & -7.103 & 0.214 \\
\ce{CO2} & O & -2 & -0.424 & -0.431 & 6.380 & -6.811 & 0.208 \\
\ce{H2O2} & O & -1 & -0.387 & -0.414 & 6.351 & -6.765 & 0.212 \\
\ce{LiO2^{.}} & O & -1/2 & -0.456 & -0.412 & 6.330 & -6.743 & 0.215 \\
\ce{O3} & O$_\text{t}$ & 0 & -0.194 & -0.177 & 6.364 & -6.541 & 0.207 \\
\ce{CO} & O & -3 & -0.144 & -0.108 & 6.398 & -6.506 & 0.201 \\
\ce{O} & O & 0 & 0.0 & 0.0 & 6.348 & -6.348 & 0.207 \\
\ce{O2} & O & 0 & 0.0 & 0.0 & 6.371 & -6.371 & 0.203 \\
\ce{OF2} & O & 2 & 0.156 & 0.121 & 6.367 & -6.247 & 0.203 \\
\ce{O3} & O$_\text{c}$ & 0 & 0.389 & 0.354 & 6.386 & -6.032 & 0.197 \\
\ce{O^{+}} & O & 1 & 1.0 & 1.0 & 6.431 & -5.431 & 0.182 \\ \hline
\ce{OF2} & F & -1 & -0.078 & -0.060 & 7.381 & -7.441 & 0.182 \\ \hline
\ce{MgO} & Mg & 2 & 2.220 & 1.934 & 10.551 & -8.617 & 0.117 \\ \hline
MIL-53(Al) & Al & 3 & 2.780 & 2.111 & 3.215 & -1.104 & 0.352 \\ \hline
Quartz & Si & 4 & 2.946 & 2.425 & 4.425 & -2.000 & 0.314 \\
Chabazite & Si & 4 & 2.944 & 2.490 & 4.395 & -1.905 & 0.316 \\ \hline
\end{tabular}
\end{table}
The main trend in Table \ref{tab:examples} is the strong correlation between HI and MBIS charges. For systems where HI was found to be useful for force-field development, MBIS gives very comparable results. However, when oxygen has an oxidation state of -2 and has (semi-ionic) bonds to cations with a high oxidation number, MBIS charges for oxygen are less negative, making them more suitable for force-field development.
A reasonable correlation between atomic charges and oxidation numbers is found. Such correlations are not expected to be perfect because the oxidation number is based on simple counting rules that do not account for the (partial) covalent character of chemical bonds. It may be surprising that the core charge, $q_{A,c}$, is systematically larger than the integer value one would get by combining the nuclear charge and an integer number of core electrons. Because the valence Slater function does not decay toward the nucleus, it also contributes to the core region, which is compensated by a slightly more positive core charge. All the variations in the net charge are reflected in the valence charge, $q_{A,v}$. The valence width, $\sigma_{A,v}$, linearly correlates with the net charge: within each group of a given element, more negative atoms tend to have a slightly larger valence width.
\subsection{Application of MBIS to the three phases of water}
The MBIS method will first be illustrated with an application to an isolated water molecule, 38 clusters of water molecules, a model for the hexagonal phase of ice and 10 snapshots of a liquid water MD simulation. The electron densities of the isolated systems were computed with Gaussian09 \cite{g09} at the BLYP/6-311+G(2df,p) level of theory. \cite{lee1988, ditchfield1971} The 38 clusters, ranging from 2 to 10 water molecules in size, were taken from the work of Temelso \textit{et al}. \cite{temelso2011} We used the 3x3x2 model for the ice-1h phase of water from the work of Hayward and Reimers. \cite{hayward1997} This model contains 96 water molecules and is tuned for computational applications: the water molecules have realistic randomized orientations, yet the net dipole moment of the unit cell is constrained to zero. The geometry of the ice-1h model is refined with CP2K-2.6.0 \cite{cp2k, vandevondele2003, vandevondele2005, hutter2014} at the BLYP-D3 level of theory \cite{becke1988, lee1988, grimme2010, grimme2011} using the MOLOPT-DZVP-SR-GTH \cite{vandevondele2007} basis set and GTH pseudopotentials. \cite{goedecker1996, krack2005} CP2K was also used to generate periodic structures of liquid water (32 molecules per unit cell). These geometries were sampled every 10 ps from a 100 ps NVT \cite{bussi2007} molecular dynamics run at 300 K and at the experimental density, using the same level of theory. The electron densities of all periodic structures (ice and liquid water snapshots) were computed with GPAW \cite{mortensen2005, enkovaara2010, bahn2002} using the BLYP functional and a grid spacing of 0.1 \AA, as explained above. (CP2K was not used for this purpose because it cannot print out a reconstructed all-electron density on suitable integration grids.)
Figure~\ref{fig:water} displays the key MBIS results for the water systems in this section. The partitioning of the density into atomic contributions, $\rho_A(\mathbf{r})$, is first used to construct electron densities of separate water molecules from which multipoles can be derived (relative to the molecular center of mass). The most obvious result is the increase of the molecular dipole moment as water forms hydrogen bonds with surrounding molecules (Figure~\ref{fig:water}a). Similar trends are usually found in simulations of water with polarizable force fields. \cite{horn2004, piquemal2007, burnham2008, yu2013} This increase is seen throughout all water clusters and the solid ice 1h phase. Liquid water exhibits relatively large random fluctuations in the molecular dipole moment, due to variations in the water geometry and its local environment. Another clear trend is that water molecules tend to exchange a small fraction of an electron with their surrounding, leading to non-zero molecular charges (Figure~\ref{fig:water}b).
Atomic charges and dipole moments are directly derived from the AIM densities, $\rho_A(\mathbf{r})$. The increased polarization of water in larger clusters (Figure~\ref{fig:water}c) is due to the decrease of the (negative) oxygen charge, while the norm of the oxygen dipole moment follows the opposite trend (Figure~\ref{fig:water}d), slightly reducing the overall polarization. Hydrogen atoms have a small and constant dipole moment, showing that they are only weakly polarizable.
Because the MBIS pro-molecular density is a sum of spherical atoms, it was to be expected that AIM dipole fluctuations play a minor role compared to atomic and molecular charge fluctuations. In general, the partitioning of the total polarization into contributions from atomic charges and/or dipoles is inherently ambiguous and can depend strongly on the AIM method. However, Mei \textit{et al.}\ observed, for a large set of molecules and for all AIM methods tested in their work, that the overall polarization always involves a significant amount of charge fluctuations. \cite{mei2015}
MBIS pro-atom parameters reveal additional trends that are not easily observed with other methods. The \textit{valence width}, $\sigma_{A,v}$ as introduced in subsection~\ref{subsec:ff}, is also sensitive to the molecular environment, which is most notable for the hydrogen atoms while oxygen has a more constant valence width (Figure~\ref{fig:water}e). Polarizable force fields usually consider fluctuating atomic charges and/or dipoles, but fluctuations in the width of the atomic electron distribution are rarely included. Our results indicate that these may also be relevant to model electronic polarization.
Finally, Figure~\ref{fig:water}f shows the \textit{atomic core charge}, $q_{A,c}$ defined in subsection~\ref{subsec:ff}. This quantity varies relatively little, in line with the expectation that the properties of core electrons should be transferable. There is a small but notable difference between the core charge for Gaussian09 (1-10 H$_2$O) and GPAW (Ice 1h, Liquid) calculations. GPAW calculations on the isolated clusters confirm that this is due to the different treatment of the core electrons in both programs (results not shown).
In general, the MBIS results for different phases of water show that the method is robust enough to uncover several subtle trends in the electronic polarization, which is very helpful for the interpretation of these trends and the construction of polarizable force fields.
\begin{figure}
\includegraphics{overview_water.pdf}
\caption{MBIS results for different phases of water (isolated water molecule, water clusters, ice 1h and liquid water) derived from all-electron densities computed at the BLYP level of theory: (a) the norm of the molecular dipole moment of each water molecule (b) the net charge of each water molecule (c) the oxygen charge of each water molecule, (d) the norm of each atomic dipole moment, (e) the valence width of each atom ($\sigma_{A,v}$, see subsection \ref{subsec:ff}), (f) the atomic core charge of each oxygen atom ($q_{A,c}$, see subsection \ref{subsec:ff}).}
\label{fig:water}
\end{figure}
\section{Systematic comparison to other AIM methods}
\label{sec:comp}
In this section, the MBIS method is compared to a series of other AIM methods, using several molecular datasets. This assessment focuses on properties that are relevant for modeling electrostatic interactions in force fields: the quality of the electrostatic potential (ESP), the accuracy of electrostatic interactions and the robustness of the charges.
\subsection{Molecular datasets}
\label{subsec:compmol}
Several datasets of molecular dimers and isolated molecules are considered in this section: five were taken from the literature and two new datasets are introduced below. All molecular electron densities were computed at the B3LYP/6-311+G(2df,p) level of theory with Gaussian09. \cite{g09}
Three sets of molecular dimers were taken from the work of Hobza et al., namely S66 (diverse non-covalent interactions between neutral organic molecules), \cite{rezac2011} IHB15 (ionic hydrogen bonds) \cite{rezac2012} and X40 (halogen bonds). \cite{rezac2012b} From the X40 set, dimers containing iodine were omitted because proper all-electron densities for such heavy elements can only be computed with relativistic corrections. (See subsection \ref{subsec:impl}.) Also, a new set of molecular dimers is introduced, i.e.\ ZG237, a set of 237 dimers between silica clusters and typical guest/template molecules for porous media. Neutral and anionic silica clusters are present in ZG237 and the guest molecules include noble gases, neutral and cationic organic molecules. (More details are provided in section S2 of the supporting information.) The goal of this assessment with molecular dimers is to test how well atomic charges obtained with different AIM methods can reproduce the electrostatic interaction.
Three datasets of larger isolated molecules are also used in the tests below, of which two were taken from earlier work: PENTA103 (103 random penta-alanine conformers) \cite{verstraelen2011} and SILICA245 (topologically different hydrogen-terminated silica clusters containing up to 8 Si atoms). \cite{verstraelen2013} One new set, MIL53(M)10 was created based on our experience with the development of a flexible force field for the metal-organic framework MIL-53(AL). \cite{vanduyfhuys2012} This set contains ten organometallic clusters with the same structure, see Fig.\ \ref{fig:muoh}, but with different metals in oxidation state III: Al, Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Ga. The spin multiplicity of each cluster was fixed by coupling the spins of the transition metals to obtain a maximal $\langle \hat{S}_z \rangle$ value. The $\mu$-OH group is located at the center and the cluster is carefully terminated by four malondialdehyde anions and one formic acid anion. This neutral configuration is stable for many first-row transition metals (excluding Cu and Zn) and it resembles well the metal-oxide structure found in the MIL-53 framework. \cite{ferey2003}
\begin{figure}
\includegraphics{muoh.pdf}
\caption{Metal-oxide cluster where M is a metal atom in oxidation state III (Al, Sc, Ti, V, Cr, Mn, Fe, Co, Ni, Ga).}
\label{fig:muoh}
\end{figure}
\subsection{Selection of AIM methods}
\label{subsec:compaim}
Three categories of AIM methods are used for comparison: ESP-fitted charges, density partitioning methods (Hirshfeld variants and QTAIM) and Hilbert-space partitioning methods.
Atomic charges fitted to the electrostatic potential (ESP) are among the most ubiquitous for the development of force field models. We selected four such variants: Merz-Singh-Kollman (MSK), \cite{singh1984} CHELPG, \cite{breneman1990} Restrained ESP (RESP) \cite{bayly1993} and Hu-Lu-Yang (HLY). \cite{hu2007} MSK, CHELPG and HLY differ in the way the volume around the molecule is sampled in the fitting procedure but they all make use of a standard least-squares procedure. The RESP method extends the MSK cost function with hyperbolic restraints to penalize large absolute atomic charges.
A large selection of Hirshfeld variants is used in our comparison, starting with the original Hirshfeld method (H). \cite{hirshfeld1977} CM5 is a popular empirical correction to the Hirshfeld method to better reproduce experimental dipole moments of small molecules. \cite{marenich2012} Another popular related method is Hirshfeld-I (HI). \cite{bultinck2007} While the Hirshfeld method simply uses spherically averaged neutral atoms as pro-atoms, HI iteratively updates the pro-atoms to enforce consistency between the AIM charges and the charges of the pro-atoms. Charged pro-atoms are constructed by a linear interpolation between spherically averaged densities of isolated neutral atoms and ions. A particular improvement of HI over Hirshfeld is that HI charges make a good estimate of the electrostatic potential of organic molecules. \cite{vandamme2009} This is no longer the case for metal oxides, e.g.\ the ESP in the pores of zeolites, which inspired several groups to further improve the method, leading to variants such as Hirshfeld-E \cite{verstraelen2013} (HE) and DDEC4. \cite{manz2012} The Iterative Stockholder (IS) analysis is another variant of the Hirshfeld method, proposed independently of HI but with many similarities. \cite{lillestolen2008} IS also iteratively updates its pro-atoms but just uses the spherical averages of the AIM densities from the previous iteration as the new pro-atoms. Besides all the Hirshfeld variants mentioned so far (H, CM5, HI, HE, DDEC4 and IS), one more density-based method, QTAIM, \cite{bader1991} is also included in the comparison.
The third group of methods are Hilbert-space methods, which partition the density matrix instead of the density. Mulliken (M) is the oldest AIM method \cite{mulliken1955} and two popular improvements of this scheme are also widely used: L\"owdin (L) \cite{lowdin1950} and Natural (N) charges. \cite{reed1985} These three Hilbert-space methods assume that each orbital basis function is centered on one of the atoms, which is always the case when using standard Gaussian basis sets. However, in several popular periodic DFT codes, e.g.\ VASP, CPMD and GPAW, such information is not available, because they use delocalized basis sets.
Results for methods MSK, CHELPG, CM5, M, L and N were obtained with Gaussian09. \cite{g09} The MSK or CHELPG methods make use of van der Waals radii but do not define them for all elements, in which case UFF radii were used instead. Gaussian formatted checkpoint files were used to post-process the densities with HORTON-2.0.0 \cite{horton} to compute the HLY, H, HI, HE, IS and MBIS charges. The RESP program from the Antechamber program \cite{wang2006} was used to compute the RESP charges. DDEC4 charges were computed with Chargemol-09.15.2014, \cite{chargemol} QTAIM charges with AIMAll-11.06.19. \cite{aimall} Practically all MBIS results below are obtained from the pro-atom parameters discussed in section \ref{subsec:ff}. MBIS AIM densities as such are not used unless noted otherwise.
\subsection{Quality of the electrostatic potential (ESP)}
\label{subsec:compesp}
In the context of force-field development, it is assumed that whenever atomic charges accurately reproduce the electrostatic potential (ESP) around a molecule, they also make good predictions of the electrostatic interactions. \cite{fox1998} Hence, one of the desirable properties of atomic charges is their ability to reproduce the ESP as well as possible, which is the topic of this subsection. A direct assessment of the quality of electrostatic interactions is discussed in subsection \ref{subsec:compint}.
We have tested the quality of the ESP for all sets of isolated molecules discussed in section \ref{subsec:compmol} and also for the monomers present in all dimer datasets. The HLY ESP cost function is used to measure the quality of the ESP and the results would not change much if we had used an MSK or CHELPG cost function instead. The HLY cost functions is an integration over a volume surrounding the molecule, \cite{hu2007} which we converted to an RMSE value as follows:
\begin{equation}
\label{eq:rsmdesp}
\text{RMSE}_\text{ESP} = \sqrt{\frac{
\int w_\text{HLY}(\mathbf{r}) \left(
V_\text{DFT}(\mathbf{r})
- \sum_A \frac{q_A}{4 \pi \epsilon_0|\mathbf{r} - \mathbf{R}_A|}
\right)^2 d\mathbf{r}
}{
\int w_\text{HLY}(\mathbf{r}) d\mathbf{r}
}}
\end{equation}
where $w_\text{HLY}(\mathbf{r})$ is the weight function designed by Hu, Lu and Yang: it becomes one in the region surrounding a molecule and goes smoothly to zero inside the molecule and at larger distances. \cite{hu2007} $V_\text{DFT}(\mathbf{r})$ is the reference ESP from the DFT calculation and $q_A$ are the atomic charges. The smoothness of the weight function guarantees that the ESP cost is not sensitive to the exact position of the grid points, which is a clear advantage over other ESP fitting methods. RMSE$_\text{ESP}$, which we computed for every molecule and every AIM method, is a measure for the error on the frozen-density interaction energy \cite{wesolowski1993, wu2009, tafipolsky2011} of a unit charge with the molecule when it is placed near its van der Waals surface.
Fig.\ \ref{fig:esptab} compares for every AIM method the average of RMSE$_\text{ESP}$ within five groups of isolated molecules: PENTA103 and MIL53(M)10 are those discussed in section \ref{subsec:compmol}. SILICA contains all those of the SILICA245 set plus all silica clusters from the ZG237 set. X40HMONO contains all halogenide molecules present in the X40 set of dimers. Finally, ORGANIC contains all other monomers from the dimer sets S66, IHB15, X40 and ZG237. (The noble gas atoms from ZG237 are not included.)
\begin{figure}
\includegraphics{esp_table.pdf}
\caption{RMSE$_\text{ESP}$ in kJ\,mol$^{-1}$ computed for all AIM methods in this work, averaged over groups of isolated molecules. (See text for definition of groups).}
\label{fig:esptab}
\end{figure}
Obviously, the ESP-fitted charges (MSK, RESP, CHELPG and HLY) perform well in this test as they are optimized to reproduce the ESP surrounding each molecule. RESP charges are not as optimal as the other three because of the hyperbolic restraints, which becomes very pronounced for large molecules as in the PENTA set. Lowering the strength of the restraints could relieve this issue but it would also result in less robust charges. The halogenides in X40HMONO have an ESP that is relatively difficult to reproduce with point charges: the sigma-hole of the halogen atom corresponds to a large and local dipole moment whose effect on the ESP cannot be explained in terms of atomic monopoles. \cite{ibrahim2011, cole2016}
The original Hirshfeld method (H) usually predicts poor ESPs because the absolute values of the atomic charges are too low. \cite{davidson1992} All variants of the Hirshfeld method (HI, CM5, DDEC4, HE, IS and MBIS) produce more accurate ESPs, except for CM5 and HI when tested with the SILICA set. The good performance of the IS method is not surprising: it partitions the electron density in AIMs that are as spherical as possible, sometimes by introducing an unreasonable radial dependence, \cite{verstraelen2009} thus having small atomic dipole and higher multipole moments. In fact, any method that performs better than IS likely biases the charges to mimic effects of atomic multipoles. Such overfitting clearly occurs in the ESP-fitting methods. In case of MBIS, we only considered MBIS point charges and not the more advanced model with valence Slater functions, see Fig.\ \ref{fig:concept}b, simply because Eq.\ \eqref{eq:rsmdesp} only tests the ESP outside the molecule where the density is very low. In this region, the ESP generated by the Slater functions is very well approximated by that of point charges. Of all Hirshfeld variants, DDEC4, IS and MBIS are comparably good.
In line with previous observations, QTAIM charges are inadequate for the purpose of modeling ESPs. \cite{verstraelen2013} This can only be fixed by including higher atomic QTAIM multipoles, as is often done in QTAIM-based force fields. \cite{popelier2015} Mulliken (M), L\"owdin (L) and Natural (N) charges rarely produce useful ESPs in our tests.
Fig.\ \ref{fig:isosurfaces} shows ESP maps plotted on the $\rho=0.002 \text{ a.u.}$ isosurface of two representative molecules: methylacetamide and methylsilanetriol, which have an electrostatic potential that is respectively easy and difficult to reproduce with point charges. The isosurface approximates the molecular van der Waals surface, \cite{bader1987} which is convenient for visualizing non-covalent interactions. In addition to the ESP of the DFT calculation, the ESPs obtained with a subset of atomic charge methods, and their deviation from the DFT result, are shown. The isosurfaces sample the ESP at a higher density than the HLY cost function, which has two important consequences. First, the scale of the ESP deviations is large compared to the reported RMSE$_\text{ESP}$ values. Second, the ESP maps are different for MBIS point charges and MBIS core charges with delocalized valence shells, the latter accounting for the penetration effect.
\begin{figure}
\includegraphics{isosurfaces.pdf}
\caption{ESP maps plotted on the $\rho=0.002 \text{ a.u.}$ isosurface of (a) methylacetamide and (b) methylsilanetriol. In the left column for each molecule, the DFT ESP and the ESP due to selected point charge methods (HI, MBIS and RESP) and the core charge + valence shell (MBIS) are shown. For each approximate ESP, the deviation from the DFT reference is shown in the right column, for each molecule.}
\label{fig:isosurfaces}
\end{figure}
The main observation is that all model ESPs qualitatively agree with the DFT result. (See left column in Fig.\ \ref{fig:isosurfaces}a and Fig.\ \ref{fig:isosurfaces}b.) Some quantitative differences are present but they only appear clearly in the isosurfaces on which the deviations from the DFT ESP are shown. (See right column in Fig.\ \ref{fig:isosurfaces}a and Fig.\ \ref{fig:isosurfaces}b.) Even though RESP charges have a relatively low RMSE$_\text{ESP}$, the deviations from the DFT ESP are not significantly smaller than for the other methods. When MBIS core charges and valence shells are used to estimate the ESP, a better visual agreement is found, because the penetration effect is already significant at the selected isodensity surface. Also note that the ESP obtained with HI charges for methylsilanetriol deviates the most from the DFT reference, in line with the limitations of HI for oxides, which were also discussed in subsection \ref{subsec:oxygen} and which is also seen in Fig.\ \ref{fig:esptab}. Finally, note that this visualization of two representative molecules merely serves as an illustration. Solid conclusions can only be drawn from a thorough statistical analysis involving many molecules, such as the one presented in Fig.\ \ref{fig:esptab}.
\subsection{Accuracy of electrostatic interactions}
\label{subsec:compint}
A common assumption in force-field development is that ESP-fitted charges also reproduce electrostatic interaction energies in general. Here, we assess the validity of this assumption for molecular dimers: the electrostatic interaction in the frozen-density approximation, \cite{wesolowski1993, wu2009, tafipolsky2011} $E_\text{FD}$, will be used as a reference to test approximate electrostatic interactions obtained with atomic point charges from different AIM methods. The frozen-density approximation does not include any effects from polarization or charge-transfer. Such effects should be modeled with a polarizable (or polarized) force field, which is beyond the scope of this test.
The four sets of molecular dimers described in section \ref{subsec:compmol} (S66, IHB15, X40 and ZG237) cover a large variety of electrostatic interactions, from as little as $-0.06$ kJ\,mol$^{-1}$ to rather extreme values of $-664$ kJ\,mol$^{-1}$. Especially in the ZG237 set, it is often hard to classify dimers into specific interaction types, like hydrogen bonding, salt bridge, etc. To facilitate the interpretation of the results, we have classified the dimers more conveniently, just using thresholds on the strength of the electrostatic interaction in the frozen-density approximation: ``Weak'' ($E_\text{FD} > -10\,\text{kJ\,mol}^{-1}$), ``Medium'' ($-10\,\text{kJ\,mol}^{-1} \ge E_\text{FD} > -50\,\text{kJ\,mol}^{-1}$) and ``Strong'' ($E_\text{FD} \le -50\,\text{kJ\,mol}^{-1}$). Fig.\ \ref{fig:inttab}a shows the numbers of dimers from each dataset in each class. The IHB15 set contributes exclusively to the ``Strong'' class while all other sets have dimers in each class of interaction strength. Fig.\ \ref{fig:inttab}b shows the root-mean-square error (RMSE) on the electrostatic interaction energy for each of the three classes and for each AIM method. The label ``MBIS-S`` refers to the interaction energy computed using effective core charges and valence Slater density functions, as shown schematically in Fig.\ \ref{fig:concept}b.
\begin{figure}
\includegraphics{int_table.pdf}
\caption{The accuracy of the electrostatic interactions for three classes of electrostatic interactions: Weak ($E_\text{FD} > -10\,\text{kJ\,mol}^{-1}$), Medium electrostatics ($-10\,\text{kJ\,mol}^{-1} \ge E_\text{FD} > -50\,\text{kJ\,mol}^{-1}$) and Strong ($E_\text{FD} \le -50\,\text{kJ\,mol}^{-1}$). Part (a) shows the number of dimers each dataset contributes to each class. Part (b) shows the RMSE of the electrostatic interaction in kJ\,mol$^{-1}$, obtained with atomic charges from different AIM methods, for each group. MBIS-S goes beyond the simple point-charge model and uses spherical valence Slater functions to model the penetration effect.}
\label{fig:inttab}
\end{figure}
A surprising result in Fig.\ \ref{fig:inttab} is that electrostatic interactions computed with ESP-fitted charges are not the most accurate. MBIS point charges perform better in all three classes (Weak, Medium and Strong) than the best ESP-fitting method (HLY). This can be understood as follows: as explained in subsection \ref{subsec:compesp}, ESP-fitted charges are biased to reproduce effects of atomic multipoles on the ESP. Although this may improve the accuracy with which the ESP is reproduced, it is a form of overfitting that may deteriorate other results obtained with ESP-fitted charges, as we observe here.
The second important result is that the MBIS method is a very effective model for the penetration effect. The MBIS-S results in Fig.\ \ref{fig:inttab} are obtained by describing every atom with an effective core charge and a valence Slater function. Even though this is a very simple (and thus computationally efficient) approach, it already reduces the RMSE by more than 50\% in the classes ``Weak'' and ``Medium''. Only for ``Strong'' electrostatic interactions, there are no apparent benefits from using such Slater density functions. A more detailed analysis, in which we computed electrostatic interactions with multipole expansions of MBIS AIM densities, showed that the largest error on the ``Strong`` electrostatic interactions is due to the neglect of atomic dipole moments.
\subsection{Robustness of the atomic charges}
\label{subsec:comprob}
For the development of an electrostatic force-field model or for the chemical interpretation of atomic charges, it is desirable that the charges are robust, i.e.\ not too sensitive to small details in the electronic structure calculations from which they are derived. Robustness is a prerequisite for transferability, i.e.\ the assumption that parameters derived from a molecule remain valid when that molecule is embedded (non-)covalently in a molecular environment. Even for environment-specific force-field parameters, \cite{grimme2014, cole2016} a robust partitioning is of interest to assure that such parameters remain valid as far as possible from the reference point for which they were computed. In this subsection, three kinds of sensitivity (the inverse of robustness) of atomic charges are investigated: sensitivity to conformational changes, to chemical changes in the environment and to changes in the basis set.
Fig.\ \ref{fig:senstab}a shows the sensitivity of the atomic charges to conformational changes of the penta-alanine chain, for all AIM methods. For a given AIM method, the standard deviation of the atomic charges in the PENTA103 set are computed with respect to the average charge of each atom over all 103 conformations. These 103 conformers are randomly generated meta-stable structures. \cite{verstraelen2011} Although some fluctuation of the charges may be expected due to internal polarization, some methodological artifacts will cause larger fluctuations without physical origin.
Fig.\ \ref{fig:senstab}b compares an averaged standard deviation of the Si charges in the SILICA245 set. The Si atoms are divided into three groups, based on the number of terminating hydrogen atoms they are bonded to (ranging from 1 to 3). Within each group, the standard deviation on the Si charge is computed and the average over the three groups is shown in Fig. \ref{fig:senstab}b. This standard deviation should be small because the Si atoms within one group have a very comparable chemical environment.
Finally, Fig.\ \ref{fig:senstab}c shows the sensitivity of the metal atom charge in the MIL53(M)10 clusters to the basis set. Their electron densities were computed with 6-311+G(2dp,f), 6-311+G*, 6-31+G*, 6-31+G or 6-31G*. The standard deviation is computed relative to the average charge of each transition metal over all basis sets. Note that the sensitivity values for Mulliken and L\"owdin fall literally off the chart and the corresponding bars in the bar plot were truncated for the sake of clarity.
\begin{figure}
\includegraphics{robust_table.pdf}
\caption{Standard deviations on atomic charges in e, within different sets of molecules. (a) The fluctuation on the atomic charges of 103 penta-alanine conformers. (b) The fluctuation on three different types of Si charges in a set if 245 silica clusters. (See text for definition of groups of Si atoms.) (c) The fluctuation on the metal charge in the MIL53(M)10 clusters, due to changes in basis set.}
\label{fig:senstab}
\end{figure}
The ESP-fitted and Mulliken charges have a very high sensitivity in all three cases, in line with earlier work. \cite{verstraelen2011} This is problematic because it is almost impossible to provide definitive charges with such methods. The basis set sensitivity seems to be the most difficult to control: the standard deviation is larger than 0.1 e for the methods MSK, RESP, CHELPG, HLY, HI, HE, and M, L and N.
The basis set robustness of H, HI and HE can be improved as follows. Currently we used consistent levels of theory for pro-atom and molecular electron densities. If the pro-atoms were computed with a single level of theory and basis set, independent of the settings of the molecular calculation, the robustness would significantly improve. This is noticeable in the low sensitivity of the DDEC4 and especially the CM5 charges. Both CM5 and DDEC are implemented with a unique set of pro-atoms.
The MBIS charges are more robust than the IS charges. This is simply because MBIS pro-atoms have fewer degrees of freedom than IS pro-atoms, in line with previous observations. \cite{verstraelen2012}
The restraints in the RESP method only have a marginal impact on the robustness compared to MSK (the same method without restraints), showing that the restraints do not meet their purpose while they may cause a very poor fit to the ESP (see Fig.\ \ref{fig:esptab}). The HLY method, a rather recent ESP-fitting method with a more carefully constructed cost function, does not guarantee robust results either.
QTAIM charges are among the most robust in our test, which is in line with previous studies assessing the transferability of QTAIM results. \cite{devereux2009, popelier2015} Still, several Hirshfeld variants, such as CM5, MBIS and DDEC4 are comparably robust.
\subsection{Pareto analysis}
\label{subsec:pareto}
In our comparative analysis, we have considered three main criteria that atomic charges should meet for the development of force fields: accuracy of the ESP, accuracy of electrostatic interactions and robustness. Ideally, an AIM method should combine all these features, especially the last two. The Pareto plots in Fig.\ \ref{fig:pareto} visualize the trade-offs between different criteria discussed in the previous subsections. It is unavoidable that some subjective choices slightly affect the Pareto analysis, such as the selected molecules in the datasets, their classification into groups, the ESP cost function, etc. Nevertheless, some clear trends can be observed.
Fig.\ \ref{fig:pareto}a compares the average RMSE$_\text{ESP}$ over all groups in Fig.\ \ref{fig:esptab} (Y-axis) to the average RMSE of the weak and medium electrostatic interactions from Fig.\ \ref{fig:inttab} (X-axis). The strong electrostatic interactions are not included because we found that these can never be reproduced reliably with any model using just atomic monopoles. The datapoint for QTAIM was omitted due to its excessively large average RMSE$_\text{ESP}$. The Pareto front only considers genuine point-charge models. Obviously, the MBIS-S method performs far better for electrostatic interactions as it goes beyond the simple point-charge model. This figure mainly shows that an accurate electrostatic potential does not guarantee accurate electrostatic interactions, and vice versa. If both qualities are of interest, the methods N, MBIS, IS and HLY are Pareto optimal. Obviously, HLY is Pareto optimal as its cost function was used to compute $\langle\text{RMSE}_\text{ESP}\rangle$, which is the reason for its advantage over MSK and CHELPG.
\begin{figure}
\includegraphics{pareto.pdf}
\caption{Pareto plots showing the trade-offs between different desirable properties of AIM methods: (a) The quality of the ESP versus the accuracy of electrostatic interactions and (b) the sensitivity of the atomic charges versus the accuracy of the electrostatic interactions. Red datapoints are on the Pareto front for all models using just point charges. The MBIS-S method goes beyond point charges by introducing a model for the distributed valence electron density.}
\label{fig:pareto}
\end{figure}
Fig.\ \ref{fig:pareto}b uses the same X-axis as figure \ref{fig:pareto}a but has the average of the three sensitivity values from Fig.\ \ref{fig:senstab} on the Y-axis. The datapoint for the Mulliken method was omitted due to its excessively large average sensitivity value. The Pareto-optimal point-charge models are N, MBIS and CM5. Again, when going beyond point charges, MBIS-S has a very attractive performance. The poor performance of ESP-fitted charges in Fig.\ \ref{fig:pareto}b is striking. The RESP method has been the method of choice in the development of many force field models, most notably in the AMBER community. Our results indicate that RESP and other ESP fitting methods are relatively poor methods for modeling electrostatic interactions in force fields.
\section{Application to density-dependent dispersion models}
\label{sec:disp}
Several dispersion models, typically used to correct DFT calculations, make use of the Hirshfeld partitioning method to estimate AIM polarizabilities. The polarizability of atom $A$ in a molecule, $\alpha_A$, is obtained by rescaling the experimental value of the free neutral atom using the third radial moment of AIM density, $\rho_A(\mathbf{r})$:
\begin{equation}
\alpha_A = \alpha_{A,\text{free}} \frac{\langle r^3 \rangle_A}{\langle r^3 \rangle_{A,\text{free}}},
\end{equation}
where
\begin{equation}
\langle r^3 \rangle_A = \int d\mathbf{r} |\mathbf{r} - \mathbf{R}_A|^3 \rho_A(\mathbf{r})
\end{equation}
and similarly for $\langle r^3 \rangle_{A,\text{free}}$. These rescaled polarizabilities are used in various methods \cite{becke2007, tkatchenko2009, steinmann2011} to obtain environment-specific atomic $C_6$ coefficients. The original Hirshfeld method is most often used in this context. The Iterative Hirshfeld method is sometimes used instead \cite{steinmann2010} and found to improve dispersion-corrected DFT calculations for ionic systems. \cite{bucko2013, bucko2014} Density-based dispersion models are not only used for correcting DFT calculations but were recently also employed in force-field development. \cite{cole2016} In this section, we will directly compare the accuracy of molecular $C_6$ coefficients when the Tkatchenko-Scheffler method is used in combination with different Hirshfeld variants. \cite{tkatchenko2009} We expect that similar results can be obtained with the Exchange-hole dipole model (XDM) \cite{becke2007} and related approaches such as dDsC. \cite{steinmann2011}
The expectation value $\langle r^3 \rangle_{A,\text{free}}$ can be computed in two different ways. One may use the symmetry-broken ground state density of the free atom or one may constrain the atom to be spherically symmetric and closed-shell. (In both cases, the same level of theory is used as for the molecule.) The second choice is the most common in the context of dispersion models. For some elements however, this results in higher-energy states with fractionally occupied orbitals. For the sake of consistency, compatible choices are made when computing the reference atoms for the (Iterative) Hirshfeld method. In this section, we consider in total six variants of the TS dispersion model, using three different partitioning methods: Hirshfeld (H), Iterative Hirshfeld (HI) and MBIS. For each partitioning method, ground state reference atoms (GS) or spherical closed-shell atoms (SCS) are used. The molecular $C_6$ coefficients are also computed with Grimme's D3 model. \cite{grimme2010}
The dispersion models are tested with the database of $C_6$ coefficients for dimers of neutral molecules by Tkatchenko and Scheffler, \cite{tkatchenko2009} which are derived from experimental dipole oscillator strengths. B3LYP/6-311+G(2df,p) densities \cite{becke1993b, ditchfield1971} are computed for all molecules in this database and the atoms they contain. Gaussian09 \cite{g09} is used for the B3LYP calculations, except for the SCS atoms, for which a new SCF program was written. Dimers with the xenon atom are omitted because their all-electron density can only be computed properly with relativistic corrections. As explained in subsection \ref{subsec:impl}, the density cannot be written out by Gaussian09 when relativistic corrections are used.
Fig.\ \ref{fig:c6} shows the scatter plots of the model $C_6$ values versus the experimental reference data, including the mean percentage errors (MPE) and the root-mean-square percentage errors (RMSPE). The original TS model corresponds to TS-H-SCS, which is one of the better variants. When using the Iterative Hirshfeld method instead (TS-HI-SCS), the RMSPE increases, indicating that the model becomes less accurate. In both H and HI variants, the use of ground-state atoms, i.e.\ TS-H-GS or TS-HI-GS, leads to a systematic overestimation of the reference $C_6$ coefficients. The situation is reversed when using MBIS partitioning, i.e.\ the dispersion model is most accurate when using ground state reference atoms (TS-MBIS-GS), while the use of spherical reference atoms (TS-MBIS-SCS) is clearly inferior. Finally, it is worth noting that Grimme's D3 model performs slightly better than any TS variant, which is impressive given that it only makes use of the nuclear coordinates and not the electron density.
\begin{figure}
\includegraphics{molc6_scatter.pdf}
\caption{Scatter plots of model molecular $C_6$ coefficients versus experimental reference values. Except for the D3 results, all data points are shifted up to avoid overlap between results from different models. The mean percentage error (MPE) and the root-mean-square percentage error (RMSPE) between model and reference molecular $C_6$ coefficients are printed just above the corresponding data points.}
\label{fig:c6}
\end{figure}
We will now analyze why the TS model only works well for certain methodological combinations. First, the spherical closed-shell atoms have slightly larger (or equal) $\langle r^3 \rangle_{A,\text{free}}$ values compared to ground state atoms, as shown in Table \ref{tab:rcubed}. It turns out that, for completely different reasons, the Hirshfeld method exhibits some artifacts in the partitioning that also result in increased values of $\langle r^3 \rangle_A$, as explained below. For most molecules, these two effects balance out, except for the \ce{H2} molecule, which causes some outliers in Fig.\ \ref{fig:c6} for TS-H-SCS and TS-HI-SCS at low $C_6$ values.
\begin{table}
\centering
\caption{The value of $\langle r^3 \rangle_{A,\text{free}}$ in $\mathrm{a}_0^3$ for the elements present in the molecules in the Tkatchenko-Sheffler set, \cite{tkatchenko2009} computed in two ways: using ground-state electron densities (GS) and using densities of atoms that are constrained to be spherical and closed-shell (SCS).}
\label{tab:rcubed}
\begin{tabular}{lddd}
\hline
Element &
\multicolumn{1}{c}{$\langle r^3 \rangle_{A,\text{free,GS}}$ [$\mathrm{a}_0^3$]} &
\multicolumn{1}{c}{$\langle r^3 \rangle_{A,\text{free,SCS}}$ [$\mathrm{a}_0^3$]} &
\multicolumn{1}{c}{$\frac{\langle r^3 \rangle_{A,\text{free,SCS}}}{\langle r^3 \rangle_{A,\text{free,GS}}}$} \\
\hline
H & 7.9 & 9.7 & 1.23 \\
Li & 89.0 & 105.6 & 1.19 \\
C & 35.7 & 41.4 & 1.16 \\
N & 27.0 & 30.9 & 1.14 \\
O & 22.7 & 23.8 & 1.05 \\
F & 18.6 & 18.9 & 1.01 \\
Ne & 15.4 & 15.4 & 1.00 \\
Si & 103.0 & 114.3 & 1.11 \\
S & 77.0 & 79.7 & 1.04 \\
Cl & 66.7 & 67.4 & 1.01 \\
Ar & 57.2 & 57.2 & 1.00 \\
Br & 97.0 & 97.8 & 1.01 \\
Kr & 89.1 & 89.1 & 1.00 \\
\hline
\end{tabular}
\end{table}
The artifact of the Hirshfeld method is very clear in Fig.\ \ref{fig:hf}: it shows the AIM density, $\rho_A$, of the hydrogen atom in hydrogen fluoride, computed with the Hirshfeld and MBIS methods. The Hirshfeld AIM density (solid blue line) is asymmetric, with more electron density toward the fluoride. This can be understood as follows. Any variant of the Hirshfeld method exhibits the same similarity principle, \cite{ayers2000} due to Eq.\ \eqref{eq:stockholder}: the Hirshfeld AIM density will be as close as possible to that of the pro-atom. Because the density tail of isolated hydrogen (dashed blue line) is so different from that of hydrogen in HF, the Hirshfeld AIM density is aspherical with too much density in the bonding region. This accumulation of density, relatively far away from the hydrogen nucleus, leads to larger values of high radial moments, such as $\langle r^3 \rangle_A$. This is a general feature of the Hirshfeld method, also seen in other molecules. In the MBIS method, no such asymmetries are found because the parameters $\sigma_{A,i}$ in Eq.\ \eqref{eq:proslater}, i.e.\ the widths of the Slater functions, are also optimized to match the molecular electron density. By consequence, the MBIS AIM density (solid green curve) almost coincides with the corresponding pro-atom (dashed green curve) and is therefore close to symmetric.
Fig.\ \ref{fig:hf} also shows a side effect of the MBIS method: where the MBIS AIM density (solid green curve) passes through the nucleus of fluoride, some ripples can be seen, because some details in the fluoride core electron density cannot be reproduced by the Slater functions. These ripples are very local and therefore have a negligible effect on $\langle r^3 \rangle_A$.
\begin{figure}
\includegraphics{diatomic_hf.pdf}
\caption{Total electron density (black) of hydrogen fluoride along the internuclear axis, including the AIM density (solid) and the pro-atomic density (dashed) of hydrogen for the Hirshfeld (blue) and MBIS method (green).}
\label{fig:hf}
\end{figure}
Fig.\ \ref{fig:vw} further illustrates the mismatch between the density tails of the (Iterative) Hirshfeld pro-atoms and the molecular densities in the TS set. It shows the MBIS \textit{valence width} of the carbon atoms of all molecules in the TS set \cite{tkatchenko2009} versus their MBIS charge. This figure also includes the data points for free carbon atoms and ions, which are used as pro-atoms in the (iterative) Hirshfeld method. The trends in this figure are general: neutral atoms and anions, have density tails that decay slower (i.e.\ higher valence width) than molecular electron densities. This leads to the asymmetric (Iterative) Hirshfeld AIM densities as the one observed in Fig.\ \ref{fig:hf}. Especially when computing higher radial or multipole moments with the (Iterative) Hirshfeld method, this may result in undesirable artifacts. In previous work, the erroneous density tails of unstable anions were corrected, e.g. by computing these atoms in a Watson sphere \cite{watson1958}. For example, such corrections were used by Bu\v{c}ko \textit{et al.} \cite{bucko2013, bucko2014} in their tests of the TS-HI variant. The results in Fig.\ \ref{fig:vw} suggest that it is also advantageous to reduce the density tails of stable anions and neutral atoms.
\begin{figure}
\includegraphics{valence_widths.pdf}
\caption{The MBIS \textit{valence width} ($\sigma_{A,v}$, see subsection \ref{subsec:ff}) of carbon atoms in the Tkatchenko-Sheffler set of molecules \cite{tkatchenko2009} versus their MBIS charge (black plus). The valence widths of isolated carbon atoms and ions are also included: ground states atoms (red circle) and spherical closed-shell atoms (blue square).}
\label{fig:vw}
\end{figure}
Finally, note that a few other Hirshfeld variants were proposed, in which the density tails of the pro-atoms are optimized to match the molecular electron density, most notably the Iterative Stockholder (IS) method \cite{lillestolen2008, lillestolen2009} and some of its variants including the Gaussian ISA \cite{verstraelen2012} and BS-ISA+DF. \cite{misquitta2014}
\section{Conclusions and outlook}
\label{sec:concl}
The MBIS method is a new density-based AIM method that is particularly suitable for the development of efficient and relatively accurate electrostatic force-field models. MBIS belongs to the family of Hirshfeld methods. Its pro-density is expanded in a minimal set of atom-centered s-type Slater density functions, whose parameters are fitted to a given molecular electron density by minimizing the Kullback-Leibler divergence. In that sense, it can also be interpreted as an information theory density-fitting method, where the Slater functions as such are used in applications, rather than the atoms-in-molecules densities.
The MBIS method is extensively tested for the development of electrostatic force-field models. When it is just used for the purpose of deriving atomic charges, it is one of the best methods available to date, in terms of robustness and accuracy of the electrostatic interactions. When the MBIS Slater functions are used to describe the valence electron density in a force field, the error on the electrostatic interactions can be reduced by 50\%, if the electrostatic interaction is not too strong. This is a computationally efficient approach to describe the so-called penetration effect, i.e.\ the deviation of interatomic electrostatic interactions from that of point charges, when the atomic densities begin to overlap. MBIS is also useful beyond the scope of frozen-density electrostatics, e.g.\ when modeling dispersion interactions, or to analyze density fluctuations that a polarizable force field should reproduce.
In future work, we will focus on improving our method, its implementation and more applications in different areas. The obvious methodological improvement is a better model for the pro-molecular density, e.g. by including atomic multipoles \cite{misquitta2014, ohrn2016} or by making it compatible with pseudo-densities. An improved pro-molecule model should not merely result in a better fit to a given electron density; one should also avoid too many degrees of freedom for the sake of robustness. Moreover, the use of an improved pro-density model in force fields should remain computationally efficient.
\begin{acknowledgement}
The authors would like to thank F. De Proft and P. Bultinck for inspiring discussions and useful suggestions. T.V., S.V.\ and L.V.\ acknowledge the Foundation of Scientific Research - Flanders (FWO). T.V., L.V., M.W.\ and V.V.S.\ thank the Research Board of Ghent University (BOF), and BELSPO in the frame of IAP/7/05 for their financial support. V.V.S.\ acknowledges the European Research Council for funding the European Union’s Horizon 2020 research and innovation programme [consolidator ERC grant agreement no. 647755 -- DYNPOR (2015--2020)]. The computational resources and services used were provided by Ghent University (Stevin Supercomputer Infrastructure). F.H.-Z.\ was supported by a Vanier-CGS fellowship and a Michael Smith Foreign Study Supplement from the National Sciences and Engineering Research Council of Canada (NSERC); P.W.A.\ was supported by a Discovery Grant and a E.W.R.\ Steacie Fellowship form NSERC.
\end{acknowledgement}
\begin{suppinfo}
Atom types in MIL-53(Al), chabazite and ozone.
Description of the ZG237 set of molecular dimers and Cartesian coordinates of the MP2/cc-pVTZ+CP optimized dimer geometries in this set.
\end{suppinfo}
|
\section{Introduction} \label{s:intro}
The Hitchin system \cite{Hitchin1, Hitchin} lies at the crossroad of geometry, Lie theory, and mathematical physics. It has found remarkable applications to non-abelian Hodge theory \cite{Simpson, SIM} and the Langlands program \cite{BD, Ngo}. The goal of this paper is to prove the Hitchin map continues to have many of its crucial properties if $G$ is replaced by a Bruhat-Tits \emph{parahoric group scheme}.
In this paper, we work with parahoric group schemes whose generic fibres are simple and simply connected; thus, we do not treat parahoric group schemes whose generic fibres are not split. Moreover, we restrict to characteristic zero. The reason is we rely on \cite{BS} and \cite{BGPM} which assume these restrictions. We expect, however, that most of the results of this paper hold under less rigid assumptions.
\subsection{Parahoric group schemes} Let $X$ be a smooth projective curve over $\bC$.
For each $x\in X$, we let ${\cO_x}$ denote the algebra of functions on the completed formal neighbourhood of a point and $\cK_x$ denote the fraction field. If we choose a uniformiser $t$, then ${\cO_x} \simeq \bC[\![t]\!]$ and $\cK_x \simeq \bC(\!(t)\!)$.
Let $G$ be a simple simply connected algebraic group over $\bC$. An \emph{integral model} for $G$ is a smooth connected affine group scheme over $X$ whose generic fibre is isomorphic to $G$. If $\cG$ is an integral model, then for all but finitely many $x\in X$, called \emph{points of bad reduction} or \emph{ramification points}, one has an isomorphism $\cG({\cO_x})\simeq G({\cO_x})$.
Among all integral models, a special role is played by \emph{parahoric group schemes}. These are integral models $\cG$ such that at the points of bad reduction, $\cG({\cO})$ is a parahoric subgroup of $G(\cK)$.
To give the reader an idea about parahoric subgroups of the loop group $G(\cK)$, we compare them with their cousins, namely, parabolic subgroups of $G$. For details, cf. \cite{KumarBook}.
\begin{enumerate}
\item[(a)] Recall that parabolic subgroups are exactly those subgroups $P\subset G$ such that $G/P$ is proper. Similarly, parahoric subgroups are exactly those subgroups of the loop group $\tP\subset G(\cK)$ such that the quotient $G(\cK)/\tP$ is \emph{ind-proper}.
\item[(b)] Given a subset of simple roots of $G$, the group generated by the corresponding root subgroups, together with the root subgroups corresponding to all negative simple roots and the maximal torus, is a parabolic subgroup of $G$. This establishes a bijection between subsets of simple roots and conjugacy classes of parabolics. A similar procedure provides a bijection between \emph{proper} subsets of \emph{affine} simple roots and conjugacy classes of parahorics in $G(\cK)$.
\end{enumerate}
Among parahoric subgroups, the most familiar ones are those constructed from parabolics in the following manner. Consider the canonical evaluation map $ev:G({\cO})\ra G$ defined by $t\mapsto 0$. Given a parabolic $P\subset G$, we have that $ev^{-1}(P)$ is a parahoric subgroup of $G(\cK)$. We refer to these as \emph{parahorics of parabolic type}. Equivalently, this is when the subset of affine simple root corresponding to the parahoric is really a subset of the usual (= non-affine) simple roots.
Given parahoric subgroups $\tP_1,\cdots, \tP_l\subset G(\cK)$ and finitely many points $x_1,\cdots, x_l$ on the curve $X$, there exists integral models $\cG$ over $X$ with bad reduction exactly at the $x_i$'s satisfying $\cG({\cO_{x_i}})\simeq \tP_i$. This is a consequence of the work of Bruhat and Tits \cite{BT1, BT2}; see also \cite{Yu, BS}. Throughout this paper we assume that we only have one point $x \in X$ of bad reduction. This is done in order to keep our notation simple. As we shall see, none of our proofs depend on this assumption, so the result generalises easily to the case of finitely many points.
\subsection{Parahoric Hitchin system} Henceforth, we assume that genus of $X$ is greater than one.
Let $\cG$ be a parahoric group scheme over $X$ and let $\Omega=\Omega_X$ denote the canonical bundle. Given a $\cG$-bundle $\cE$, one may construct an adjoint bundle $\ad(\cE)$ of $\cE$, which is a vector bundle with a Lie algebra structure on its sheaf of sections. The co-adjoint bundle $\ad^*(\cE)$ is the vector bundle dual to $\ad(\cE)$. A $\cG$-Higgs bundle is a pair $(\cE, \phi)$ consisting of a $\cG$-bundle together with an element $\phi\in \Gamma(X, \ad^*(\cE)\otimes \Omega)$. Let $\cM_\cG$ denote the moduli stack of $\cG$-Higgs bundles. Using Serre duality, one can show that
\[
\cM_\cG\simeq T^*\Bun_\cG,
\]
where $T^*$ should be interpreted as in \cite[\S 1]{BD}.
Now we discuss the Hitchin map. Let $Q_1,\cdots, Q_\ell$ denote algebraically independent homogenous generators of the invariant ring $\bC[\fg]^G$ and let $d_i$ denote the degree of $Q_i$. Using local trivializations of the $\mathcal{G}$-bundle $\mathcal{E}$, one can locally identify $\phi$ with a section of $\fg^* \otimes \Omega(x)$, or via the Killing form, a local section of $\fg \otimes \Omega(x)$. Since changes of local trivialization act on $\phi$ via the adjoint action, we may apply the invariant polynomial $Q_i$ to $\phi$ to obtain a well-defined global section $Q_i(\phi) \in \Gamma( X , \Omega^{d_i}(d_i.x) )$. In this way we may define the Hitchin map on $T^*\Bun_\cG$ as follows:
\begin{equation} \label{eq:parahoricMap}
(\cE,\phi) \mapsto (Q_1,Q_2, \cdots, Q_\ell)(\phi) \in \bigoplus_i \Gamma(X , \Omega^{d_i}(d_i . x)).
\end{equation}
This map takes values in $\bigoplus_i \Gamma(X, \Omega^{d_i}(d_i.x))$. However, contrary to the usual Hitchin map, this map is not surjective. Let $\cA_\cG$ denote the closure of its image. This is an affine subvariety of $\bigoplus_i \Gamma(X, \Omega^{d_i}(d_i.x))$.\footnote{Determining the precise description of $\cA_\cG$ is a delicate problem, which we treat elsewhere \cite{BK}.}
\begin{defe}\label{d:parhit} The parahoric Hitchin map is the map
\[
h_\cG: T^*\Bun_\cG \ra \cA_\cG.
\]
defined by \eqref{eq:parahoricMap}.
\end{defe}
Let $\cN ilp_\cG=h_\cG^{-1}(0)$ denote the parahoric global nilpotent cone.
Our first main result is that
$\cN ilp_\cG$ is isotropic in $T^*\Bun_\cG$. To prove this, we adapt arguments of Ginzburg to the parahoric setting. A key technical tool we use is the description of parahoric torsors via ramified covers \cite{BS}.
Next, we show that $\Bun_\cG$ is very good in the sense of Beilinson and Drinfeld. In particular, this implies that $\dim(T^*\Bun_\cG)=2\dim(\Bun_\cG)$.
Our proof is a generalisation of the argument of Beilinson and Drinfeld (who worked with $\cG=G\times X$) to the parahoric case. The property of being very good means that $\Bun_\cG$ is in a sense not too far from being a Deligne-Mumford stack. The upshot of this is that one can naively apply various notions of symplectic geometry to $T^*\Bun_\cG$ without having to work in the setting of derived geometry. For instance, this allows us to adapt an argument of Bottacin to prove that
$h_\cG$ is a Poisson map. Here the image of $h_\cG$ is equipped with the trivial Poisson structure, so to say that $h_\cG$ is Poisson is equivalent to saying that the pullbacks of functions on the base are Poisson commuting.
The fact that the global nilpotent cone is isotropic together with the natural contracting $\bC^*$ action present implies that every fibre of $h_\cG$ has dimension at most $\dim(\Bun_\cG)$. On the other hand, the fact that $h_\cG$ is Poisson implies that every fibre is co-isotropic. Thus, we have a Lagrangian fibration. We prove that this Lagrangian fibration is a completely integrable system, by showing that generic fibres of $h_\cG$ are abelian varieties. A crucial step here is establishing that the Hitchin map on
the moduli of \emph{polystable} parahoric Higgs bundles is proper. We do this by appealing to the parahoric non-abelian Hodge theory established in \cite{BGPM}. It should also be possible to establish properness by proving semistable reduction for parahoric Higgs bundles. We leave this as an interesting topic for future work.
\subsection{What was known before?} Let $\cG$ be a parahoric group scheme on $X$. If $\cG$ has no point of bad reduction (i.e., $\cG\simeq G\times X$), then we are reduced to the system considered by Hitchin. Next, if the parahoric group scheme is parabolic of Borel type, then the above results are treated in \cite[\S V]{Faltings}. In type $A$, complete integrability of the parabolic Hitchin map was established by Scheinost and Schottenloher \cite{SS}, but in the setting of semistable moduli spaces as opposed to moduli stacks.
Let us mention that strongly parabolic Higgs bundles have a variant called \emph{weakly} parabolic Higgs bundles. These are pairs $(\cE, \phi)$ consisting of $G$-bundle $\cE$ with parabolic reduction at a point $x$ and an endomorphism $\phi\in \Gamma( \cE \times_G \fg \otimes \Omega(x))$ whose residue lies in $\fp$. (By comparison, in the strongly parabolic case considered in this article, the residue lies in the nilpotent radical $\fn$.) If $G$ is of type $A$, then Logares and Martens \cite{LogaresMarten} have proved that the Hitchin map on the semistable moduli space of weakly parabolic Higgs bundles defines a generalised completely integrable system, see also \cite{Markman, Bot}. Note that this implies that {\em generic} symplectic leaves are integrable systems. However, this does not imply that any particular symplectic leaf is an integrable system. Thus, one can not deduce integrability of moduli spaces of strongly parabolic Higgs bundles from these results, which in any case have only been proven for parabolics of type $A$.
\subsection{Acknowledgements} The idea of considering the Hitchin map for $\Bun_\cG$ is due to Xinwen Zhu. We thank him for sharing his insights with us. We would also like to thank Dima Arinkin, David Ben-Zvi, Tsao-Hsien Chen, Sergei Gukov, Arun Ram
and Zhiwei Yun for helpful discussions. DB and MK were supported by the Australian Research Council DECRA Fellowships and RV was supported by a post-doctoral fellowship at TIFR, Mumbai.
\section{Parahoric Higgs bundles}
\subsection{Basic Lie theory}\label{ss:lietheory}
Let $G$ be a simple, simply-connected complex algebraic group of rank $\ell$. Fix a maximal torus $T \subset G$ and let $\ft \subset \fg$ be the corresponding Lie algebras. Let $X(T) = Hom( T , \mathbb{G}_m )$ be the character group, $Y(T) = Hom( \mathbb{G}_m , T )$ the group of $1$-parameter subgroups of $T$ and denote by $( \, , \, ) : Y(T) \times X(T) \to \mathbb{Z}$ the canonical pairing. Let $R \subset X(T)$ be the root system associated to $G$ and fix a choice of positive roots $R^+ \subset R$. The choice of positive roots $R^+$ determines a Borel subgroup $B \subset G$ with unipotent radical $U$. Associated to each root $r \in R$, we have a root space $\fg_r \subset \fg$, a root homomorphism $u_r : \mathbb{G}_a \to G$ and a corresponding subgroup $U_r \subset G$, called the root group corresponding to $r$.
Let $\alpha_1 , \dots , \alpha_{\ell}$ be the simple roots and $\alpha_{max}$ the highest root. We let $\mathfrak{U}$ denote the rational Weyl alcove
\begin{equation*}
\mathfrak{U} := \{ \theta \in Y(T) \otimes \mathbb{Q} \; | \; (\theta , \alpha_{max}) \le 1, \; (\theta , \alpha_i) \ge 0, \; \forall \text{ positive roots } \alpha_i \}
\end{equation*}
and let $\mathfrak{U}^o$ denote the points $\theta \in \mathfrak{U}$ which satisfy $(\theta , \alpha_{max}) < 1$.
\subsubsection{Parahoric subgroups}
Let $\cO = \bC[\![t]\!]$ be the local ring of formal power series in $t$ and $\cK = \bC(\!(t)\!)$ the fraction field of $\cO$. For any $\theta \in \mathfrak{U}$ and root $r \in R$, consider the integer
\begin{equation*}
m_r(\theta) = -\lfloor ( \theta , r ) \rfloor.
\end{equation*}
The subgroup
\begin{equation}\label{eq:parahoric}
\mathcal{P}_\theta = \langle T(\cO) , \{ z^{m_r(\theta)} U_r(\cO) \}_{r \in R} \rangle \subset G(\cK)
\end{equation}
is a parahoric subgroup in the sense of Bruhat-Tits. Note that $\mathcal{P}_\theta$ and the constructions that follow depend only on the choice of $\theta$ through the facet $\Omega$ of the affine apartment $Y_*(T) \otimes \mathbb{R}$ on which $\theta$ lies, but to keep the notation simple we will avoid any further mention of facets.\footnote{Thus, we are really considering \emph{weighted} parahorics in the sense of \cite{BoalchParahoric}. The weight does not play a role in defining the moduli stack of parahoric torsors, but is essential for defining stability.}
\subsubsection{Parahoric group schemes}
From Bruhat-Tits theory, we get a smooth affine group scheme $\mathcal{G}_\theta$ over $Spec(\cO)$, which satisfies:
\begin{itemize}
\item[(i)]{$\mathcal{G}_\theta \times_{Spec(\cO)} Spec(\cK) \cong G \times Spec(\cK)$.}
\item[(ii)]{$\mathcal{G}_\theta(\cO) = \mathcal{P}_\theta$.}
\end{itemize}
\subsubsection{Parahorics of parabolic type}
If $\theta \in \mathfrak{U}^o$ then $\theta$ also determines a parabolic subgroup $P_\theta \subset G$ which is generated by $T$ and the root groups $U_r$ for which $m_r(\theta) = 0$. Moreover, the parahoric $\mathcal{P}_\theta$ and the parabolic $P_\theta$ are related by
\begin{equation*}
\mathcal{P}_\theta = ev^{-1}( P_\theta ),
\end{equation*}
where $ev : G(\cO) \to G$ is the evaluation map sending $t$ to zero. We refer to the case $\theta \in \mathfrak{U}^o$ as the {\em parabolic case} and we call $\mathcal{P}_\theta$ a parahoric subgroup of {\em parabolic type}.
\subsubsection{Parahoric lie algebras}
Associated to $\theta \in \mathfrak{U}$ we also have the corresponding parahoric subalgebra $\tp_\theta \subset \fg(\cK)$, which may be defined as:
\begin{equation}\label{eq:parahoricsubalgebra}
\tp_\theta = \ft(\cO) \oplus \bigoplus_{r \in R} t^{m_r(\theta)} \fg_r(\cO).
\end{equation}
Let $\kappa$ be the Killing form on $\fg$. We define the dual $\tp_\theta^\perp$ as
\[
\tp_\theta^\perp = \{ u \in \fg(\cK) \; | \; \kappa( u , v ) \in \cO \; \; \forall v \in \tp_\theta \}.
\]
We have a natural isomorphism
\begin{equation*}
\Psi : \tp_\theta^\perp \to Hom_\cO( \tp_\theta , \cO)
\end{equation*}
which sends $u \in \tp_\theta^\perp$ to the homomorphism $v \mapsto \kappa(u , v )$. This isomorphism identifies $\tp_\theta^\perp$ with the dual of $\tp_\theta$ as $\cO$-modules. From (\ref{eq:parahoricsubalgebra}) we see that:
\begin{equation*}
\tp_\theta^\perp = \ft(\cO) \oplus \bigoplus_{r \in R} t^{-m_r(\theta)} \fg_{-r}(\cO).
\end{equation*}
Elements of $\tp_\theta^\perp$ should be interpreted as local parahoric Higgs fields.
\subsection{Parahoric torsors}\label{ss:parabolicBundles}
Fix a marked point $x \in X$ and a parahoric subgroup $\mathcal{P}_\theta \subset G(\cK)$ associated to some $\theta \in \mathfrak{U}$. Recall that $\cO_x$ denotes the completed local ring at $x$. Let $D_x = Spec(\cO_x)$ be the completed formal neighbourhood of $x$. By a {\em parahoric Bruhat-Tits group scheme} on $X$, we mean a group scheme $\mathcal{G}_{X,x,\theta}$ over $X$ such that:
\begin{itemize}
\item[(i)]{$\mathcal{G}_{X,x,\theta} |_{X - x} \cong G \times (X - x)$.}
\item[(ii)]{$\mathcal{G}_{X,x,\theta}|_{D_{x}} \cong \mathcal{G}_{\theta}$.}
\end{itemize}
Such a group scheme exists for any $(X,x,\theta)$ \cite{BS}. To simplify notation we will omit the subscripts $X,x,\theta$ and write $\cG = \cG_{X,x,\theta}$. We will also refer to $\cG$ simply as a {\em parahoric group scheme} on $X$. Denote by $Bun_{\cG}$ the stack of $\mathcal{G}$-torsors on $X$. We note here that the group scheme $\cG$ is not uniquely determined by the data $(X,x,\theta)$, but it follows from \cite{BS} that the corresponding stacks $Bun_{\cG}$ are all equivalent once $(X,x,\theta)$ are fixed.
In \cite{Heinloth}, it is shown that $Bun_{\cG}$ is a smooth algebraic stack locally of finite type. One can show that the dimension of $Bun_{\cG}$ is given by \cite{BS}:
\begin{equation*}
\dim( Bun_{\cG} ) = (g-1)\dim(G) + \# \{ r \in R^+ \; | \; \langle \theta , r \rangle \neq 0,1 \}.
\end{equation*}
In the parabolic case, i.e. when $\theta \in \mathfrak{U}^o$, torsors for $\cG$ correspond to the more familiar notion of (quasi-)parabolic bundles. We recall the definition \cite{LS}:
\begin{defe}
Let $P \subset G$ be a parabolic subgroup of $G$. A quasi-parabolic $G$-bundle on $X$ with parabolic structure of type $P$ at $x$ is a pair $(\cE,\cE_x^P)$, where $\cE$ is a principal $G$-bundle on $X$ and $\cE_x^P$ is a $P$-reduction of $\cE$ at $x$; i.e. an element of $\cE_x \times_G (G/P)$. A parabolic $G$-bundle on $X$ is a quasi-parabolic bundle $(\cE , \cE_x^P)$ together with a choice of $\theta \in \mathfrak{U}^o$ such that $P = P_\theta$. One refers to $\theta$ as the weights of the parabolic bundle at $x$.
\end{defe}
\subsection{Parahoric Higgs bundles} \label{ss:HiggsBundles}
Associated to a parahoric group scheme $\cG = \cG_{X,x,\theta}$, we have a bundle of Lie algebras $\on{Lie}( \cG) \to X$ over $X$. This is a vector bundle equipped with a Lie bracket on its sheaf of sections. Away from $x$ it is just given by the trivial bundle $(X - x ) \times \fg$, while in a formal neighbourhood $D_x = Spec( {\cO}_x)$ of $x$, it may be identified with the Lie algebra $\fp_\theta = \on{Lie}(\mathcal{P}_\theta)$, as given by Equation \ref{eq:parahoricsubalgebra}.
Suppose that $\cE \to X$ is a parahoric $\cG$-torsor. We define the adjoint bundle of $\cE$ by
\begin{equation*}
ad(\cE) = \cE \times_{\cG} Lie( \cG),
\end{equation*}
where $\cG$ acts on $Lie(\cG)$ by the adjoint action. It follows that $ad(\cE)$ is a vector bundle over $X$ and is equipped with a Lie bracket on its sheaf of sections. We let $ad^*(\cE)$ denote the dual bundle.
\begin{defe} Let $\cE \in Bun_{\cG}$ be a parahoric torsor. A parahoric Higgs field is a section $\phi\in \Gamma(X, ad(\cE)^* \otimes \Omega)$. A parahoric Higgs bundle is a pair $(\cE, \phi)$ consisting of a parahoric torsor $\cE$ and a parahoric Higgs field $\phi$.
\end{defe}
More generally, to define the stack $\mathcal{M}_{\cG}$ of parahoric Higgs bundles, let $S$ be an arbitrary scheme over $\bC$ and define
\begin{equation*}
\mathcal{M}_{\cG}(S) = \{ (\cE , \phi ) \; | \; \cE \in Bun_{\cG}(S), \; \phi \in \Gamma( X \times S , ad(\cE)^* \otimes \Omega^1_{X \times S / S}) \},
\end{equation*}
with the obvious notion of isomorphism between objects. We recall that $Bun_{\cG}$ is a smooth equidimensional algebraic stack. From standard deformation theory, we deduce an isomorphism of stacks
\begin{equation*}
\mathcal{M}_{\cG} \cong T^*Bun_{\cG}.
\end{equation*}
Thus, $\mathcal{M}_{\cG}$ is a symplectic stack; in particular, we may speak of the Poisson bracket and of isotropic substacks. Note that in the setting of stacks, care has to be taken to study the above notions; see \cite[\S 1]{BD} for details.
In the parabolic case, where $\theta \in \mathfrak{U}^o$ the notion of parahoric Higgs bundles reduces to the more familiar notion of (quasi-)parabolic Higgs bundles, as we now recall. Note that some authors call such Higgs bundles {\em strongly parabolic}.
\begin{defe} Let $(\cE , \cE_x^P)$ be a parabolic bundle. Let $\fp = \on{Lie}(P)$ be the Lie algebra of the parabolic subgroup $P \subset G$ and $\fn \subset \fp$ the nilpotent radical of $\fp$. A parabolic Higgs field is a section $\phi \in \Gamma(X, \cE \times_G \fg^* \otimes \Omega(x) )$ such that $Res_x(\phi) \in \fn$. A quasi-parabolic Higgs bundle is a triple $(\cE, \cE_x^P , \phi)$ consisting of a parabolic bundle $(\cE , \cE_x^P)$ and a parabolic Higgs field $\phi$. A parabolic Higgs bundle is a quasi-parabolic Higgs bundle $(\cE , \cE_x^P , \phi)$ together with a choice of weight $\theta \in \mathfrak{U}^o$ for which $P = P_\theta$.
\end{defe}
\subsection{Relation to equivariant bundles}\label{ss:equivariant}
In this section we remark on the relation between parahoric torsors and parahoric Higgs bundles to equivariant torsors and equivariant Higgs bundles on a Galois cover of $X$.
Suppose that $p : Y \to X$ is a Galois cover with Galois group $\Gamma$ such that $p$ has ramification over the marked point $x \in X$ and nowhere else. Let $y \in Y$ be one of the points lying over $x$ and let $\Gamma_y \subseteq \Gamma$ be the stabiliser, which is cyclic of order $n$ say. Let $g \in \Gamma_y$ be a generator. Following \cite{BS}, we define a $(\Gamma , G)$-bundle on $Y$ to be a principal $G$-bundle $E \to Y$ with $G$ acting on the right, equipped with a lift of $\Gamma$ to an action on $E$ commuting with the $G$-action. In other words, $E$ is a $\Gamma$-equivariant principal $G$-bundle.
Given a local trivialization of $E$ around $y$, we can identify the fibre $E_y$ with $G$ such that the right action is given by group multiplication on the right. Then $\Gamma_y$ acts on $E_y \cong G$ commuting with the right $G$-action. This implies that there is a homomorphism $\rho_y : \Gamma_y \to G$ such that the action of $\gamma \in \Gamma_y$ on $E_y \cong G$ is left multiplication by $\rho_y(\gamma)$. Since $\Gamma_y = \langle g \rangle $ is a cyclic group of order $n$, the homomorphism $\rho_y : \Gamma_y \to G$ corresponds to choosing an element $h \in G$ whose order divides $n$, so that $\rho_y(g) = h$. If we choose a different local trivialization of $E$, then $\rho_y$ and hence $h \in G$ change by conjugation. Thus associated to the $(\Gamma , G)$-bundle $E$ is a well defined conjugacy class $\mathcal{C} \subset G$ of finite order and we say that $E$ has {\em type} $\mathcal{C}$. If $\tau \in \mathcal{C}$ is any representative of the conjugacy class $\mathcal{C}$, we will also say that $E$ has type $\tau$. Note that for any given representative $\tau \in \mathcal{C}$, we may choose our local trivialization of $E$ such that $\rho_y(g) = \tau$.
Recall that each element $\theta \in \mathfrak{U}$ of the rational Weyl alcove corresponds to a conjugacy class $\mathcal{C}_\theta \subset G$ of finite order via the exponential map. Let $\tau \in \mathcal{C}_\theta$ be a representative of this conjugacy class. We let $Bun_Y^\tau(\Gamma , G)$ denote the stack of $\Gamma$-equivariant principal $G$-bundles on $Y$ of type $\tau$. According to Balaji-Seshadri \cite[Theorem 5.3.1]{BS}, this stack is equivalent to $Bun_{\mathcal{G}_{X,x,\theta}}$, where $\cG_{X,x,\theta}$ is a parahoric group scheme on $X$. We will often find it useful to work with equivariant bundles on $Y$ in place of parahoric bundles on $X$. We note that from the equivariant bundle point of view there is clearly no difficulty in allowing $G$ to be any connected reductive group. This will be important for us in the proof of Theorem \ref{t:verygood}. However, when $G$ is non-simply connected the relation between equivariant bundles and parahoric bundles requires more care, see \cite[\S 8.1.13]{BS}.
If $E \to Y$ is an equivariant bundle of type $\tau$, an equivariant Higgs field on $E$ is defined to be a $\Gamma$-equivariant section of $ad^*(E) \otimes \Omega_Y$. We may then speak of equivariant Higgs bundles of type $\tau$. One sees that the stack of all such Higgs bundles is precisely $T^* Bun_Y^\tau(\Gamma , G)$. The equivalence $Bun_{\mathcal{G}_{X,x,\theta}} \cong Bun_Y^\tau(\Gamma , G)$ gives an equivalence $T^* Bun_{\mathcal{G}_{X,x,\theta}} \cong T^* Bun_Y^\tau(\Gamma , G)$. In other words, we obtain an equivalence between parahoric Higgs bundles on $X$ for the group scheme $\mathcal{G}_{X,x,\theta}$ and equivariant Higgs bundles on $Y$ of type $\tau$. We note that the covering $p : Y \to X$ and the group $\Gamma$ depend on the choice of $\theta \in \mathfrak{U}$, but are not uniquely determined by $\theta$.
\subsection{Parahoric Hitchin map}
Recall that $\cO_x \cong \bC [\![t]\!]$ denotes the completed local ring of $x$ and $D_x = Spec(\cO_x)$. Suppose that $(\cE , \phi)$ is a parahoric Higgs bundle for $\cG = \mathcal{G}_{X,x,\theta}$. Away from $x$, we may identify $\cE$ with a principal $G$-bundle and $\ad(\cE)|_{X - x}$ is the usual adjoint bundle of $\cE |_{X - x}$. Moreover, the Killing form on $\fg$ gives an isomorphism $\ad^*(\cE)|_{X - x} \cong \ad(\cE)|_{X - x}$. Therefore, away from $x$, $\phi$ is a $1$-form valued section of the adjoint bundle. On the other hand, $\phi |_{D_x}$ may be identified with an element of $\tp_\theta^\perp \otimes \Omega^1_{D_x}$. Since we are assuming $\theta \in \mathfrak{U}$, it is easy to see that $m_r(\theta) \le 1$ for all roots $r \in R$ and hence $\tp_\theta^\perp \subset t^{-1} \fg(\cO_x)$. It follows that at all points of $X$, $\phi$ may locally be identified with a section of $\fg \otimes \Omega^1_X(x)$, i.e. $\phi$ is locally a $\fg$-valued $1$-form with at most a first order pole at $x$. A change in local trivialization of $\cE$ acts on $\phi$ via the adjoint action, and hence if $Q_i$ is an invariant polynomial on $\fg$ of degree $d_i$, we may apply $Q_i$ to $\phi$ to obtain a well-defined section $Q_i(\phi) \in \Gamma( X , \Omega^{d_i}(d_i . x))$. As explained in the introduction, this defines a map
\begin{equation*}
(\cE,\phi) \mapsto (Q_1,Q_2, \cdots, Q_\ell)(\phi) \in \bigoplus_i \Gamma(X , \Omega^{d_i}(d_i . x))
\end{equation*}
where $Q_1 , Q_2 , \cdots , Q_\ell$ are algebraically independent homogeneous generators for the ring of invariant polynomials on $\fg$. This gives a map $\hat{h}_\cG : \mathcal{M}_\cG \to \bigoplus_i \Gamma(X , \Omega^{d_i}(d_i . x))$. We let $\cA_\cG$ denote the closure of the image of the map $\hat{h}_\cG$. Therefore $\cA_\cG$ is an affine subvariety of $\bigoplus_i \Gamma(X, \Omega^{d_i}(d_i.x))$. The map $\hat{h}_\cG$ factors through $\cA_\cG \to \bigoplus_i \Gamma(X , \Omega^{d_i}(d_i . x))$ and thus defines a map
\[
h_\cG: \mathcal{M}_\cG \ra \cA_\cG.
\]
As per Definition \ref{d:parhit}, we call $h_\cG$ the {\em parahoric Hitchin map}.
\section{Global nilpotent cone and the very good property}\label{ss:parabolicHitchin}
We continue to use the notation of the previous section. Thus, $\cG = \cG_{X,x,\theta}$ is a parahoric group scheme over $X$ with one ramification point $x$ such that $\cG(\cO_x)$ is a parahoric subgroup of type $\theta$.
\subsection{Global nilpotent cone}\label{ss:globalnilpotent}
Recall that for any scheme $S$ over $\bC$, we have
\[
\cM_{\cG}(S) = \{(\cE,s) \mid \cE \in Bun_{\cG}(S), \phi \in \Gamma(X \times S, ad(E)^*\otimes \Omega^1_{X\times S/S}) \}.
\]
Since $\cG \mid_{(X-x)} \cong G \times (X-x)$,
we have, via the isomorphism $\mathfrak{g} \cong \mathfrak{g}^*$
induced by the killing form,
$ad(\cE)^* \otimes \Omega^1_{X \times S/S} \mid_{(X -x) \times S} \cong ad(\cE)
\otimes \Omega^1_{X\times S/S} \mid_{(X -x) \times S}$. Thus for
any point $y \in (X-x) \times S$, we have
\[
\phi(y) \in \mathfrak{g} \otimes k(y).
\]
\begin{defe} We say $\phi \mid_{(X-x) \times S}$ is nilpotent if
\[
\forall y \in (X-x) \times S, \,\ \phi(y) \in \mathfrak{g} \otimes k(y) \,\ \text{is nilpotent}.
\]
The \textit{global nilpotent cone} is the substack
$\cN ilp_{\cG}$ of $\cM_{\cG}$, given by
\[
\cN ilp_{\cG}(S) = \{ (\cE,\phi) \in \cM_{\cG} \mid \phi \mid_{(X-x) \times S} \text{is nilpotent} \}.
\]
Equivalently, $\cN ilp_{\cG}=h_{\cG}^{-1}(0)$.
\end{defe}
We have the following key result:
\begin{thm}
$\cN ilp_{\cG}$ is an isotropic substack of $\mathcal{M}_{\cG}$.
\end{thm}
\begin{proof}
We use the Balaji-Seshadri description of parahoric torsors as equivariant bundles (\S \ref{ss:equivariant}), and adapt the proof of \cite{Ginzburg} to this setting. Thus,
we will restrict ourselves to explaining the necessary modifications involved. Let $\cE$ be a $\mathcal{G} = \cG_{X,x,\theta}$-torsor on $X$
and let $\phi \in \Gamma(X,ad(\cE)^* \otimes \Omega_X)$ be nilpotent. Let $\mathcal{B} \subset \cG$ be the Borel subgroup scheme,
defined as the flat closure of $(X-x) \times B$ in
$\cG$, for a Borel subgroup $B \subset G$.
From Heinloth~\cite[Lemma 23]{Heinloth}, we have that the natural morphism
$f: Bun_{\mathcal{B}} \rightarrow Bun_{\cG}$ is surjective.
Now we can find a finite Galois cover $p : Y \rightarrow X$,
with Galois group $\Gamma$, such that the stack of $\Gamma$-equivariant principal
$G$-bundles on $Y$ of a fixed local type determined by $\theta$,
is equivalent to the stack of $\cG$ torsors on
$X$ \cite[Theorem 5.3.1]{BS}. As explained in \cite{BS}, if we choose one such equivariant bundle $\cF$ on $Y$, we have an isomorphism of group schemes on $X$:
\[
\mathfrak{R}^\Gamma_{Y/X}(Ad(\cF)) \cong \cG,
\]
where $Ad(\cF)$ denotes the adjoint bundle of groups associated to $\cF$ and $\mathfrak{R}^\Gamma_{Y/X}()$ denotes the invariant direct image functor.
The equivalence of the categories between equivariant bundles on $Y$ of fixed local type
and $\mathcal{G}$-torsors is obtained by
\[
\mathcal{F}^{'} \mapsto \mathfrak{R}_{Y/X}^{\Gamma}(\mathcal{F}^{'} \wedge^G \cF^{op}),
\]
where for any right $G$-torsors $E, F$ on $Y$, $E \wedge^G F^{op}$ denotes the contracted product \cite{BS}:
\[
E \wedge^G F^{op} := \frac{E \times_Y F}{(xg , y) \sim (x , yg^{-1})}.
\]
Next, we have a $\Gamma$-equivariant $B$ reduction $\cF_B$ of $\cF$, such
that $\mathfrak{R}_{Y/X}^{\Gamma}(Ad(\cF_B)) \cong \mathcal{B}$. Let $\mathcal{E}^{'}$ be the equivariant bundle on $Y$, which corresponds to $\cE$
under the equivalence mentioned above. Then we have an equivariant section
$\tilde{\phi}$ of $ad(\mathcal{E}^{'}) \otimes \Omega_Y$, which descends to $\phi$. As in \cite{Ginzburg}, there exists an equivariant $B$ reduction of $\mathcal{E}^{'}$ over the generic point
of $Y$, for which $\tilde{\phi}$ is a section of $\mathfrak{n}_{\mathcal{E}^{'}} \otimes \Omega_Y$, where $\mathfrak{n}$ denotes the nilradical of the Borel subalgebra and
$\mathfrak{n}_{\mathcal{E}^{'}}$ denotes the corresponding bundle of Lie algebras determined by the $B$ reduction of $\mathcal{E}^{'}$. Since $G/B$ is projective, we can extend this $B$
reduction $\mathcal{E}^{'}_B$ of $\mathcal{E}^{'}$
to the whole of $Y$. Thus we have a $\mathcal{B}$ reduction $\cE_{\mathcal{B}}$ of
$\cE$ given by
\[
\cE_{\mathcal{B}} = \mathfrak{R}_{Y/X}^{\Gamma}(\mathcal{E}^{'}_B \wedge^G \cF_B^{op})
\]
such that
over the generic point, $\phi \in \mathfrak{n}_{\cE_{\mathcal{B}}} \otimes \Omega_X$.
As in \cite{Ginzburg}, this implies that $f^*(\phi) = 0 \in \Gamma(X,ad(\cE_{\mathcal{B}})^* \otimes \Omega_X)$
as it vanishes on the generic point.
Following the notations of \cite[Lemma 3]{Ginzburg}, for $N_1 = Bun_{\mathcal{B}}$ and $N_2 = Bun_{\cG}$, we have
$\cN ilp_{\cG} = pr_2(Y_f)$. The rest of the proof
can be done exactly as in the proof of \cite[Lemma 5]{Ginzburg}.
\end{proof}
\subsection{Parahoric torsors form a very good stack}
Recall \cite[\S 1]{BD}, that a smooth equidimensional algebraic stack $\cY$ is said to be {\em good} if $\dim( T^* \cY) = 2 \dim( \cY)$ and is said to be {\em very good} if
\[
\text{codim}\{ y \in \mathcal{Y} \; | \; \dim( G_y ) = d \} > d \quad \text{for all } d > 0,
\]
where $G_y$ denotes the automorphism group of $y$. Equivalently $\mathcal{Y}$ is very good if $T^*\cY^0$ is dense in $T^*\cY$, where $\cY^0$ is the largest Deligne-Mumford substack of $\cY$. Note that very good implies good. Using the fact that $\cN ilp_{\cG}$ is isotropic, we prove:
\begin{thm}\label{t:verygood}
The stack $Bun_{\cG}$ is very good.
\end{thm}
\begin{proof}
As mentioned in \S \ref{ss:equivariant}, it is more convenient to work with equivariant principal bundles of type $\tau$ on a Galois cover $p : Y \to X$, as opposed to parahoric bundles on $X$. We will allow $G$ to be a connected semisimple group throughout this proof. We will prove that $Bun_Y^\tau(\Gamma , G)$ is very good by adapting the argument used by Beilinson-Drinfeld \cite[\S 2.10.5]{BD} in the non-equivariant setting. The main difference is that we need to be more careful in estimating the dimensions of various spaces involved in the proof. We also need the fact that the nilpotent cone in $T^*Bun_Y^\tau(\Gamma , G)$ is isotropic for any semisimple $G$, which was proved in the previous subsection.
We must show that $Bun_Y^\tau(\Gamma , G)$ satisfies:
\[
\text{codim}\{ E \in Bun_Y^\tau(\Gamma , G) \; | \; \dim(H^0_\Gamma(Y , ad(E) ) = d \} > d \quad \text{for all } d > 0,
\]
Where $H^0_\Gamma( Y , ad(E) )$ denotes the space of $\Gamma$-equivariant sections of $ad(E)$. This is equivalent to showing that
\begin{equation}\label{39}
\dim( A(G) \setminus A^0(G) ) < \dim( Bun_Y^\tau(\Gamma , G) ),
\end{equation}
where $A(G)$ is the stack of pairs $(E , s )$, $E \in Bun_Y^\tau(\Gamma , G)$, $s \in H^0_\Gamma( Y , ad(E) )$ and $A^0(G) \subseteq A(G)$ is the closed substack of pairs with $s = 0$.
Let $\fc = Spec( \mathbb{C}[\mathfrak{g}]^G)$ be the affine space whose coordinate ring is the ring of invariant polynomials on $\mathfrak{g}$. Recall that
\[
\fc = \mathfrak{g}/\!/G = \mathfrak{t}/W,
\]
where $W$ is the Weyl group. The morphism $\mathfrak{g} \to \fc$ induces a map $H^0_\Gamma( Y , ad(E)) \to Mor( Y , \fc) = \fc$, in other words, we apply invariant polynomials to $s \in H^0_\Gamma( Y , ad(E))$. This gives a natural map $f : A(G) \to \fc$.
For $h \in \mathfrak{t}$, let $\overline{h} \in \fc$ be the image of $h$ under $\mathfrak{t} \to \fc$ and set $A_h(G) = f^{-1}( \overline{h})$. Set $G^h = \{ g \in G \; | \; Ad_g h = h\}$ and $\mathfrak{g}^h = Lie(G^h) = \{ a \in \mathfrak{g} \; | \; [a,h] = 0\}$. Note that $h \in \mathfrak{t}$ is semisimple and so $G^h$ is reductive. In fact $G^h$ is the Levi of a parabolic subgroup of $G$. Denote by $\mathfrak{z}^h$ the center of $\mathfrak{g}^h$. Since $h \in \mathfrak{z}^h$ and there are a finite number of subalgebras of $\mathfrak{g}$ of the form $\mathfrak{z}^h$ as $h$ varies over $\mathfrak{t}$, (\ref{39}) follows from the inequality $\dim( A_h(G) \setminus A^0(G) ) < \dim(Bun_Y^\tau(\Gamma , G) ) - \dim( \mathfrak{z}^h )$. So it is enough to show that
\begin{equation}\label{40}
\dim( A_h(G) ) < \dim( Bun_Y^\tau(\Gamma , G) ) - \dim( \mathfrak{z}^h ) \quad \text{for } h \neq 0
\end{equation}
\begin{equation}\label{41}
\dim( A_0(G) \setminus A^0(G) ) < \dim( Bun_Y^\tau(\Gamma , G) ) .
\end{equation}
Consider $(E , s ) \in A_h(G)$. Then $s \in H^0_\Gamma( Y , ad(E) )$ and $s$ maps to $\overline{h} \in \fc$. Let us write $s = s_s + s_n$, where $s_s$ is semisimple, $s_n$ is nilpotent and $[s_s , s_n] = 0$. With respect to a local trivialisation of $E$ around $y$, we may identify $s_s(y)$ with an element of $\mathfrak{g}$. We may choose the trivialization so that the generator of the stabilizer group $\Gamma_y$ acts as left multiplication by $\tau$. By $\Gamma$-equivariance of $s$, we have that $Ad_{\tau}( s_s(y) ) = s_s(y)$. Choosing a different trivialisation will change $\tau$ and $s_s(y)$ by conjugation. Next we note that since $s_s(y)$ maps to $\overline{h} \in C \cong \mathfrak{t}/W$, we can choose a local trivialisation in which $s_s(y) = h$. Replacing $\tau$ by a conjugate of $\tau$ if necessary, we obtain a trivialization in which $s_s(y) = h$ and the generator of $\Gamma_y$ acts as left multiplication by $\tau$. In this trivialisation the equation $Ad_{\tau}(s_s(y)) = s_s(y)$ becomes $Ad_{\tau}(h) = h$, so $\tau \in G^h$. Denote by $Z^h$ the center of $G^h$. Note that different trivializations in which $s_s(y) = h$ differ by elements of $G^h$. Thus $\tau$ is determined up to conjugation in $G^h$. Therefore, the conjugacy class of $\tau$ in $G^h$ and also in $G^h/Z^h$ is well-defined independent of our choice of local trivialization. Thus it makes sense to speak of the stacks $Bun_Y^\tau( \Gamma , G^h), Bun_Y^\tau(\Gamma , G^h/Z^h)$ and also of $A_h( G^h ), A_0(G^h), A_0(G^h/Z^h)$. We will now argue that (\ref{40}) follows from (\ref{41}) with $G$ replaced by $G^h/Z^h$.
Next we observe that since $h \in \mathfrak{z}^h$, there is an obvious isomorphism $A_0(G^h) \cong A_h(G^h)$ and a further isomorphism $A_h( G^h) \cong A_h(G)$. The latter follows from the fact that if $(E , s) \in A_h(G)$, then $s_s$, the semisimple part of $s$ defines a reduction of structure group to $G^h$.
Observe that there is a natural morphism $\varphi : A_0(G^h) \to A_0(G^h/Z^h)$. A non-empty fibre of $\varphi$ is isomorphic to $Bun_Y^e(\Gamma , Z^h) \cong Bun_X(Z^h)$, where $e$ denotes the identity element of $Z^h$. So
\[
\dim( A_h(G) ) \le \dim( Bun_X(Z^h) ) + \dim( A_0( G^h/Z^h)).
\]
Let us define $e(\theta)$ and $e(\theta , h)$ by:
\begin{equation*}
\begin{aligned}
e(\theta) &= \# \{ r \in R^+ \; | \; \langle \theta , r \rangle \neq 0,1 \}, \\
e(\theta, h) &= \# \{ r \in R^+ \; | \; \langle \theta , r \rangle \neq 0,1 \text{ and } \langle h , r \rangle = 0 \}
\end{aligned}
\end{equation*}
Then one can show \cite{BS} that:
\begin{equation*}
\begin{aligned}
\dim( Bun_Y^\tau(\Gamma , G )) &= (g-1)\dim(G) + e(\theta), \\
\dim( Bun_Y^\tau(\Gamma , G^h/Z^h)) &= (g-1)\dim(G^h/Z^h) + e(\theta , h).
\end{aligned}
\end{equation*}
Now since $\dim( Bun_X(Z^h)) = (g-1)\dim(\mathfrak{z}^h)$ and (\ref{41}) implies that $\dim( A_0(G^h/Z^h)) < (g-1)\dim( \mathfrak{g}^h / \mathfrak{z}^h) + e(\theta , h)$, we get:
\begin{equation*}
\begin{aligned}
\dim( A_h(G) ) &\le (g-1)\dim(\mathfrak{z}^h) + (g-1)\dim( \mathfrak{g}^h / \mathfrak{z}^h) + e(\theta , h) \\
&= (g-1)\dim(G) - (g-1)\dim( \mathfrak{g}/\mathfrak{g}^h) + e(\theta , h) \\
&\le (g-1)\dim(G) + e(\theta) - (g-1)\dim( \mathfrak{g}/\mathfrak{g}^h) \\
&= \dim( Bun_Y^\tau(\Gamma , G)) - (g-1)\dim( \mathfrak{g}/\mathfrak{g}^h).
\end{aligned}
\end{equation*}
But $\dim(\mathfrak{g}/\mathfrak{g}^h) \ge 2 \dim(\mathfrak{z}^h) > \dim(\mathfrak{z}^h)$, if $h \neq 0$. This shows (\ref{40}).\\
To prove (\ref{41}) we will show that if $V \subseteq A_0(G)$ is a locally closed reduced irreducible substack, then $\dim(V) \le \dim( Bun_Y^\tau(\Gamma , G))$ and $\dim(V) = \dim( Bun_Y^\tau(\Gamma , G))$ if and only if $V \subseteq A^0(G)$. Given such a $V$, let $A_0(G)_V$ denote the irreducible component of $A_0(G)$ containing $V$. Let $\cN ilp$ denote the nilpotent cone in $Bun_Y^\tau(\Gamma , G)$. We have shown that $\cN ilp$ is isotropic, hence $\dim(\cN ilp) \le \dim( Bun_Y^\tau(\Gamma , G))$.\\
For $\xi \in H^0_\Gamma( Y , \Omega_Y) \cong H^0(X , \Omega_X)$ consider the morphism $m_\xi : A_0(G)_V \to \cN ilp$ defined by $(E , s) \mapsto (E , s \xi)$. The morphisms $m_\xi$ define $m : A_0(G)_V \times H^0( X , \Omega) \to \cN ilp$. The image of $m$ is contained in some locally closed reduced irreducible substack $Z \subseteq \cN ilp$. If $\xi \neq 0$ then $m_\xi$ induces an embedding $V \to Z_\xi$, where $Z_\xi$ is the closed substack of $Z$ consisting of pairs $(E, \phi)$ such that the restriction of $\phi$ to the subspace $D_\xi = \{ y \in Y \; | \; \xi(y) = 0\}$ is zero. So $\dim(V) \le \dim(Z_\xi) \le \dim(Z) \le \dim(\cN ilp) \le \dim( Bun_Y^\tau(\Gamma , G))$. If $\dim(V) = \dim( Bun_Y^\tau(\Gamma , G))$, then $Z_\xi = Z$ for all non-zero $\xi \in H^0(X , \Omega_X)$. This means that $\phi = 0$ for all $(E , \phi) \in Z$ and therefore $s = 0$ for all $(E , s) \in V$, i.e. $V \subseteq A^0(G)$.
\end{proof}
\section{Poisson commutativity and Lagrangian fibration}
We continue using the notation of the previous section. Thus, $\cG$ is a parahoric group scheme and $h_\cG: \mathcal{M}_{\cG} \ra \cA_{\cG, X, x}$ is the parahoric Hitchin map.
\subsection{Poisson commutativity}
Note that $h^*_\cG$ is a map
\[
h^*_\cG: \bC[\cA_{\cG}]\ra \Gamma(\cM_{\cG},\cO).
\]
\begin{thm}\label{t:poissoncommute}
The image of the pullback $h^*_\cG$ consists of Poisson commuting functions.
\end{thm}
Our proof is analytic as we will make use of Dolbeault cohomology. One can alternatively give a purely algebraic proof by replacing instances of Dolbeault cohomology with \v{C}ech cohomology. This would lead to a proof along the same lines as given by Bottacin for parabolics of type $A$ \cite{Bot}. However, we found the analytic approach to be more straightforward.
\begin{proof}
Let $(\cE,\phi)$ be a parahoric Higgs bundle. Thus $\phi \in H^0( X , ad(\cE)^* \otimes \Omega )$. If $\alpha$ is a section of $ad(\cE)$ and $\beta$ is a section of $ad(\cE)^*$, then $[\alpha , \beta]$ is a section of $ad(\cE)^*$. We therefore have the following two-term complex:
\begin{equation*}\xymatrix{
ad(\cE) \ar[r]^-{[\phi , \, . \, ]} & ad(\cE)^* \otimes \Omega.
}
\end{equation*}
Borrowing notation from \cite{Bot}, we denote this complex by $[ \phi , \, . \, ]$.
For any parahoric Higgs bundle $(\cE , \phi)$, we have that $\mathbb{H}^0( X , [\phi , \, . \, ])$ is the Lie algebra of infinitesimal automorphisms of $(\cE,\phi)$. Observe that the dual complex $Hom( [\phi , \, . \, ] , \bC)$ tensored by $\Omega$ is canonically isomorphic to $[\phi , \, . \, ]$. Thus, by the extension of Serre duality to hypercohomology, we see that $\mathbb{H}^j(X , [ \phi , \, . \, ] ) \cong \mathbb{H}^{2-j}(X , [ \phi , \, . \, ] )^*$. Therefore $\mathbb{H}^0(X , [\phi , \, . \, ] ) =0$ implies that $\mathbb{H}^2( X , [\phi , \, . \, ] ) = 0$ as well. Thus if $\mathbb{H}^0( X , [\phi , \, . \, ]) = 0$, then around $(\cE , \phi)$, the moduli stack is smooth, Deligne-Mumford and $\mathbb{H}^1( X , [\phi , \, . \, ] )$ is the tangent space at $(\cE,\phi)$. Moreover, Serre duality gives a symplectic pairing on $\mathbb{H}^1( X , [ \phi , \, . \, ] )$, which coincides with the canonical symplectic form $\omega$ on the moduli stack of Higgs bundles.
In Dolbeault cohomology an element of $\mathbb{H}^1(X , [\phi , \, . \, ])$ is represented by a pair $( \alpha , \beta )$, where $\alpha \in \Omega^{0,1}(X , ad(\cE) )$ and $\beta \in \Omega^{0,0}(X , ad(\cE)^* \otimes \Omega)$ satisfying:
\begin{equation*}
\overline{\partial}_{\cE} \beta = [\phi , \alpha],
\end{equation*}
where $\overline{\partial}_{\cE}$ denotes the $\overline{\partial}$-operator on $ad(\cE)^*$. Moreover, the symplectic form $\omega$ on the moduli space of Higgs bundles is given by:
\begin{equation}\label{eq:symplecticform}
\omega( (\alpha_1 , \beta_1) , (\alpha_2 , \beta_2) ) = \int_X \kappa(\alpha_1 , \beta_2 ) - \kappa( \alpha_2 , \beta_1),
\end{equation}
where $\kappa$ denotes the Killing form. This is clear since the above formula is simply the Dolbeault realisation of Serre duality.
Now let $Bun^0_{\cG}$ be the substack of parahoric torsors $\cE$ with $H^0( X , ad(\cE) ) = 0$, i.e. $Bun^0_{\cG}$ is the largest Deligne-Mumford substack of $Bun_{\cG}$. As $Bun_{\cG}$ is very good, we have that $T^*Bun^0_{\cG}$ is dense in $T^*Bun_{\cG}$. Thus, to show Poisson commutativity of functions in the image of $h^*_\cG$, it is enough to show such functions Poisson commute on $T^*Bun^0_{\cG}$. We therefore restrict attention to parahoric Higgs bundles $(\cE , \phi)$ with $H^0( X , ad(\cE) ) = 0$. This also implies that $\mathbb{H}^0( X , [\phi , \, . \, ]) = 0$.
Let $\rho$ be an invariant polynomial on $\mathfrak{g}$ of degree $d_\rho$. Then for any parahoric Higgs bundle $(\cE,\phi)$ we can regard $\rho(\phi)$ as a holomorphic section of $\Omega^{d_\rho}(d_\rho x)$. Let $\mu \in H^1( X , \Omega^{1-d_\rho}(-d_\rho x))$. By Serre duality we can pair $\rho(\phi)$ with $\mu$ to get a complex number
\begin{equation*}
f_{\rho , \mu}(\cE,\phi) = \langle \rho(\phi) , \mu \rangle = \int_X \rho(\phi) \mu \in \mathbb{C}.
\end{equation*}
This defines a regular function $f_{\rho , \mu}$ on the Hitchin base and such functions generate the ring of all regular functions on the base. Thus we just need to show that any two such functions $f_{\rho , \mu} , f_{\rho' , \mu'}$ Poisson commute.
To prove Poisson commutativity, we work out the Hamiltonian vector field $X_{\rho , \mu}$ associated to $f_{\rho , \mu}$. By definition this is the vector field whose value at $(\cE,\phi)$ satisfies
\begin{equation}\label{eq:ham}
df_{\rho , \mu}(\cE,\phi)(Y) = \omega( X_{\rho , \mu}(\cE,\phi) , Y)
\end{equation}
for all tangent vectors $Y$ at $(\cE,\phi)$.
Suppose we represent $Y$ in Dolbeault cohomology as $( \alpha_Y , \beta_Y)$ and $X_{\rho , \mu}(\cE,\phi)$ as $(\alpha , \beta)$. Differentiating $\int_X f_{\rho , \mu}(\phi) \mu$ in the $Y$-direction gives
\begin{equation*}
df_{\rho , \mu}(\cE,\phi)(Y) = \int_X d_\rho \rho( \phi , \dots , \phi , \beta_Y) \mu.
\end{equation*}
From (\ref{eq:symplecticform}), we also have:
\begin{equation*}
\omega( X_{\rho , \mu}(\cE,\phi) , Y) = \int_X \kappa( \alpha , \beta_Y) - \kappa( \alpha_Y , \beta).
\end{equation*}
Comparing these expressions, we see that Equation (\ref{eq:ham}) will be satisfied if we choose $\beta = 0$ and let $\alpha$ be defined by:
\begin{equation*}
\kappa( \alpha , \, \_ \, ) = d_\rho \rho( \phi , \phi , \dots , \phi , \, \_ \, ) \mu.
\end{equation*}
In order for this to be a representative in Dolbeault cohomology, we need to check that $(\alpha , \beta)$ chosen in this way is a cocycle, i.e. $[\phi , \alpha] = 0$. For all sections $y$ of $ad(\cE)$, we have:
\begin{equation*}
\begin{aligned}
\kappa( [\phi , \alpha] , y ) &= -\kappa( \alpha , [\phi , y]) \\
&= -d_\rho \rho( \phi , \phi , \dots , \phi , [\phi , y]) \mu \\
&= 0,
\end{aligned}
\end{equation*}
where in the last line we use $ad$-invariance of $\rho$. Hence $[\phi , \alpha] = 0$, as required.
We have shown that the Hamiltonian vector field $X_{\rho , \mu}$ of $f_{\rho , \mu}$ evaluated at the point $(\cE,\phi)$ may be represented by a Dolbeault cocycle of the form $( \alpha_{\rho , \mu} , 0)$. Similarly if we have another such function $f_{\rho' , \mu'}$ then its Hamiltonian vector field $X_{\rho' , \mu'}$ evaluated at $(\cE,\phi)$ may be represented in the form $( \alpha_{\rho' , \mu'} , 0)$. From the definition of the Poisson bracket we have:
\begin{equation*}
\begin{aligned}
\{ f_{\rho , \mu} , f_{\rho' , \mu'} \}(\cE,\phi) &= \omega( X_{\rho , \mu}(\cE,\phi) , X_{\rho',\mu'}(\cE,\phi) ) \\
&= \int_X \kappa( \alpha_{\rho , \mu} , 0) - \kappa( \alpha_{\rho',\mu'} , 0) \\
&= 0.
\end{aligned}
\end{equation*}
This proves Poisson commutativity of the functions $\{ f_{\rho , \mu} \}_{\rho , \mu}$ and hence any two functions in the image of the pullback $h^*_\cG$ will Poisson commute.
\end{proof}
\begin{rem} The only fact about parahoric group schemes used in the above proof is that $\Bun_\cG$ is very good. In other words, Poisson commutativity property of the Hitchin map holds for any integral model $\cG$ such that $\Bun_\cG$ is very good.
\end{rem}
\subsection{Lagrangian fibration} In this subsection, we prove that the parahoric Hitchin map
\[
h_\cG:T^*\Bun_{\cG} \ra \cA_{\cG}
\]
defines a Lagrangian fibration. More precisely, we have:
\begin{thm}
\begin{itemize}
\item[(i)]{Any irreducible component of any fibre of $h_\cG$ has dimension $\dim( Bun_{\cG})$. In particular, $\cN ilp_{\cG}$ is a Lagrangian substack of $\cM_{\cG}$.}
\item[(ii)]{The image $\cA_{\cG}$ is irreducible and has dimension $\dim( Bun_{\cG})$.}
\end{itemize}
\end{thm}
\begin{proof}
Our proof is largely modelled on \cite[Proposition 1]{Ginzburg}. First we show that $\cA_{\cG}$ is irreducible. Indeed, since we know that $Bun_{\cG}$ is very good and connected we deduce that $\cM_{\cG} = T^*Bun_{\cG}$ is irreducible \cite[\S 1]{BD}. It follows that the image of $h_\cG$ is an irreducible topological space and hence its closure, which is $\cA_{\cG}$, is also irreducible.
Since $Bun_{\cG}$ is good, we have $\dim(\cM_{\cG}) = \dim( T^*Bun_{\cG} ) = 2 \dim( Bun_{\cG})$. It follows that each irreducible component of any fibre of $h_\cG$ has dimension greater than or equal to $2 \dim( Bun_{\cG} ) - \dim( \cA_{\cG} )$. In particular, if $\cN$ is an irreducible component of $\cN ilp_{\cG}$, then:
\begin{equation*}
\dim( \cN ) \ge 2 \dim( Bun_{\cG} ) - \dim( \cA_{\cG} ).
\end{equation*}
On the other hand, since $\cN ilp_{\cG}$ is isotropic, we have
\begin{equation*}
\dim( \cN ) \le \dim( Bun_{\cG} ).
\end{equation*}
Note that $\cN ilp_{\cG}$ is Lagrangian if and only if the above is an equality for each irreducible component $\cN$ of $\cN ilp_{\cG}$.
Next, let $F$ be a generic non-singular fibre of $h_\cG$. Because the Hitchin map is a Poisson map, we have that $F$ is a coisotropic substack of $T^*Bun_{\cG}$ and it follows that every irreducible component of $F$ has dimension at least $\dim( Bun_{\cG})$. Using the natural $\bC^*$-action on $T^*Bun_{\cG}$, we can put the fibre $F$ into a family parametrised by $\mathbb{A}^1$, so that the central fibre is $\cN ilp_{\cG}$ and all other fibres are isomorphic to $F$. One deduces from this that $\dim(F) \le \dim( \cN ilp_{\cG} )$. Combining these dimension estimates, we have
\begin{equation*}
\dim( Bun_{\cG} ) \le \dim(F) \le \dim( \cN ilp_{\cG} ) \le \dim( Bun_{\cG}).
\end{equation*}
We must have equality throughout and hence $\dim(F) = \dim( \cN ilp_{\cG}) = \dim( Bun_{\cG})$. Since the generic fibres of $h_\cG$ have dimension $\dim( Bun_{\cG})$ and $\cM_{\cG}$ is irreducible, we deduce that $\cA_{\cG}$ has dimension $\dim( T^*Bun_{\cG} ) - \dim( Bun_{\cG}) = \dim( Bun_{\cG})$, which proves (ii). Repeating the argument using the $\bC^*$-action shows that every fibre of $h_\cG$ has dimension at most $\dim( \cN ilp_{\cG} ) = \dim( Bun_{\cG})$. On the other hand, any irreducible component of any fibre of $h_\cG$ has dimension at least $\dim( T^* Bun_{\cG}) - \dim( \cA_{\cG}) = \dim( Bun_{\cG})$. Hence any irreducible component of any fibre has dimension equal to $\dim( Bun_{\cG} )$, proving (i).
\end{proof}
\begin{cor}\label{c:flat}
\begin{enumerate}
\item[(i)] If $\cA_{\cG}$ is smooth, then the Hitchin map is flat and surjective as a map to $\cA_{\cG}$.
\item[(ii)] More generally, let $\cA^\reg_\cG$ denote the smooth subscheme of $\cA_{\cG}$. Then $h_\cG: h^{-1}_\cG(\cA^{\reg}_\cG) \ra \cA^{\reg}_\cG$ is flat.
\end{enumerate}
\end{cor}
\begin{proof} The very good property implies that $T^*\Bun_{\cG}$ is a local complete intersection, therefore Cohen-Macauley. By the previous theorem, fibres of the the hitchin map $h_\cG$ have the same dimension. As the base is regular, we can use miracle flatness to conclude that $h_\cG$ is flat.
As $h_\cG$ is flat we deduce that the image of $h_\cG$ is an open subset of $\cA_{\cG}$. Let us denote the vector space $\bigoplus_i \Gamma(X, \Omega^{d_i}(d_i.x))$ by $\mathbb{W}$ and let $0 \in \mathbb{W}$ denote the origin. Note that $\cA_{\cG} \subset \mathbb{W}$ is a closed, $\bC^*$-invariant subvariety and $0 \in \cA_{\cG}$. The image of $h$ is open in $\cA_{\cG}$, hence has the form $U \cap \cA_{\cG}$, where $U$ is an open subset of $\mathbb{W}$ containing $0$. It follows that $\bC^* U = \mathbb{W}$. Clearly we also have $\bC^* \cA_{\cG} = \cA_{\cG}$ and hence it follows that $\bC^*( U \cap \cA_{\cG} ) = (\bC^* U) \cap \cA_{\cG} = \mathbb{W} \cap \cA_{\cG} = \cA_{\cG}$. However the image of $h$ is clearly $\bC^*$-invariant, so we get $U \cap \cA_{\cG} = \bC^*(U \cap \cA_{\cG}) = \cA_{\cG}$, proving surjectivity. Part (ii) similarly follows by miracle flatness.
\end{proof}
\begin{rem} The image of the parahoric Hitchin map is smooth (in fact, an affine space) for many parahorics. However, for certain parahorics in type $D$, this image can be singular; see \cite{BK} for details.
\end{rem}
\section{Properness and complete integrability} \label{s:integrability}
Our goal in this section is to show that generic fibres of the parahoric Hitchin map are abelian varieties. For this, we need to show that the Hitchin map is proper over some open subset in the base $\cA_{\cG}$. In fact, we will show that the Hitchin map on the moduli space of {\em polystable} parahoric Higgs bundles is proper.
\subsection{Properness of the Hitchin map}
Let $\mathcal{M}_\cG$ denote the moduli stack of $\cG$-Higgs bundles on $X$. The notions of stability and polystability and the corresponding coarse moduli spaces were defined for parahoric bundles in \cite{BS} and extended to parahoric Higgs bundles in \cite{BGPM}. We denote by $\mathcal{M}_{\cG}^{s}$ (resp. $\mathcal{M}_{\cG}^{ps}$) the open substack of stable (resp. polystable) $\cG$-Higgs bundles. Denote by $M^s_{\cG}$ (resp. $M^{ps}_{\cG}$) the underlying coarse moduli spaces of stable (resp. polystable) $\cG$-Higgs bundles on $X$. The Hitchin map $h_\cG : \mathcal{M}_\cG \to \mathcal{A}_\cG$ restricted to $\mathcal{M}^{ps}_{\cG}$ factors through the map $\mathcal{M}^{ps}_\cG \to M^{ps}_{\cG}$ and thus defines a map $h_\cG : M^{ps}_\cG \to \mathcal{A}_{\cG}$.
In this subsection, we prove that:
\begin{thm}\label{t:properss}
The parahoric Hitchin map $h_{\cG} : M^{ps}_{\cG} \to \mathcal{A}_{\cG}$ on the moduli space of polystable parahoric $\cG$-Higgs bundles is a proper map.
\end{thm}
\begin{proof}
Let $\mathcal{C}_\theta \subset G$ be the conjugacy class corresponding to $\theta$ and let $m$ be the order of $\mathcal{C}_\theta$. We regard $X$ as an orbifold, where the point $x \in X$ has order $m$ and there are no other orbifold points (of course, this can be extended to the case of multiple marked points).
Let $\pi = \pi_1^{orb}(X)$ be the orbifold fundamental group. A presentation of $\pi$ is given by generators
\[
\alpha_1 , \dots , \alpha_g, \quad \beta_1 , \dots , \beta_g , \quad \gamma,
\]
and relations
\[
[\alpha_1 , \beta_1] \dots [\alpha_g , \beta_g] \gamma = 1, \quad \quad \gamma^m = 1.
\]
In this presentation, $\gamma$ corresponds to a loop around the orbifold point $x$. Let $Hom( \pi , G)$ denote the space of homomorphisms $\rho : \pi \to G$. This is an affine algebraic variety and the group $G$ acts by conjugation. Let $Rep( \pi , G) = Hom( \pi , G)//G$ be the affine GIT quotient. Let $Rep_{\mathcal{C}}( \pi , G) \subset Rep(\pi , G)$ be the closed subvariety of representations $\rho : \pi \to G$ such that $\rho(\gamma) \in \mathcal{C}$.
In \cite{BGPM}, the non-abelian Hodge correspondence is extended to the case of parahoric Higgs bundles. In particular, as a special case of \cite[Theorem 7.8]{BGPM}, we obtain a homeomorphism
\[
M^{ps}_{\cG} \cong Rep_{\mathcal{C}}( \pi , G ),
\]
where $M^{ps}_{\cG}$ is the moduli space of polystable parahoric $\cG$-Higgs bundles on $X$.
Choose an integer $n$ such that $G$ is a closed subgroup of $GL(n,\mathbb{C})$. The inclusion $j : G \to GL(n,\mathbb{C})$ induces a map $Rep( \pi , G) \to Rep(\pi , GL(n,\mathbb{C}))$. By \cite[Corollary 9.16]{SIM}, this map is finite, hence it is also a proper map. Let $\mathcal{C}' \subset GL(n,\mathbb{C})$ be the conjugacy class in $GL(n,\mathbb{C})$ containing $j(\mathcal{C})$. Then we obtain a commutative diagram
\begin{equation*}
\xymatrix{
Rep_{\mathcal{C}}(\pi , G) \ar[r] \ar[d]^-j & Rep(\pi , G) \ar[d]^-j \\
Rep_{\mathcal{C}'}(\pi , GL(n,\mathbb{C})) \ar[r] & Rep(\pi , GL(n,\mathbb{C})) \\
}
\end{equation*}
In this diagram the horizontal maps are closed immersions and the second vertical map is proper, hence the map $j : Rep_{\mathcal{C}}(\pi , G) \to Rep_{\mathcal{C}'}(\pi , GL(n,\mathbb{C}))$ is also proper. The conjugacy class $\mathcal{C}'$ has finite order and therefore is of the following form: there are rational numbers $0 \le \alpha_1 \le \alpha_2 \le \dots \le \alpha_n < 1$ such that $g \in \mathcal{C}'$ if and only if $g$ is conjugate to ${\rm diag}( {\rm exp}(2\pi i \alpha_1) , {\rm exp}(2\pi i \alpha_2) , \dots , {\rm exp}(2\pi i \alpha_n ) )$. We may view $\alpha = (\alpha_1 , \alpha_2 , \dots , \alpha_n)$ as a set of parabolic weights.
By a second application of the non-abelian Hodge correspondence, we have a homeomorphism
\[
Rep_{\mathcal{C}'}(\pi , GL(n , \mathbb{C})) \cong M^{ps}_0( X , x , \alpha , GL(n,\mathbb{C}))
\]
where $M^{ps}_0( X , x , \alpha , GL(n,\mathbb{C}))$ denotes the moduli space of strongly parabolic $GL(n,\mathbb{C})$-Higgs bundles of parabolic degree $0$ and one marked point $x$ with parabolic weights $\alpha$.
As the non-abelian Hodge correspondence commutes with the group homomorphism $j : G \to GL(n,\mathbb{C})$, it is easy to see that we have a commutative diagram of the form
\begin{equation*}
\xymatrix{
M^{ps}_{\cG} \ar[d]^-{h_{\cG}} \ar@{}[r]|*[@]{\cong} & Rep_{\mathcal{C}}(\pi , G) \ar[r]^-j & Rep_{\mathcal{C}'}(\pi , GL(n,\mathbb{C})) \ar@{}[r]|*[@]{\cong} & M^{ps}_0(X , x , \alpha , GL(n,\mathbb{C})) \ar[d] \\
\bigoplus_i \Gamma( X , \Omega^{d_i}(d_i . x) ) \ar[rrr]^-j & & & \bigoplus_{i=1}^n \Gamma( X , \Omega^{i}(i . x ) )
}
\end{equation*}
where the vertical maps are Hitchin maps. We have seen that the upper horizontal map is proper. The lower vertical map is clearly a closed immersion. Therefore to show that the Hitchin map $h_{\cG} : M^{ps}_{\cG} \to \bigoplus_i \Gamma( X , \Omega^{d_i}(d_i . x))$ is proper, it is enough to show that the Hitchin map $M^{ps}_0(X , x , \alpha , GL(n,\mathbb{C})) \to \bigoplus_{i=1}^n \Gamma(X , \Omega^{i}(i.x) )$ is proper. However it is well known that the Hitchin map for parabolic $GL(n,\mathbb{C})$-Higgs bundles is proper \cite{YOK}, and so the proof is complete.
\end{proof}
\begin{rem}
Alternatively, to see that the Hitchin map $M^{ps}_0(X , x , \alpha , GL(n,\mathbb{C})) \to \bigoplus_{i=1}^n \Gamma(X , \Omega^{i}(i.x) )$ is proper, we may identify $M^{ps}_0(X , x , \alpha , GL(n,\mathbb{C}))$ with a moduli space of semistable orbifold $GL(n,\mathbb{C})$-Higgs bundles on $X$. Then by \cite{VAR}, there exists an elliptic surface $S \to X$ over $X$ such that semistable orbifold Higgs bundles on $X$ can be pulled back to give semistable Higgs bundles on $S$ in the usual sense. One sees that this gives rise to a commutative diagram of Hitchin maps
\begin{equation*}
\xymatrix{
M^{ps}_0(X , x , \alpha , GL(n,\mathbb{C})) \ar[r] \ar[d] & M^{ps}( S , GL(n,\mathbb{C}) ) \ar[d] \\
\bigoplus_{i=1}^n \Gamma( X , \Omega^i(i.x)) \ar[r] & \bigoplus_{i=1}^n \Gamma( S , Sym^i( T^*S ) )
}
\end{equation*}
where the horizontal arrows are closed immersions. Therefore properness of $M^{ps}_0(X , x , \alpha , GL(n,\mathbb{C})) \to \bigoplus_{i=1}^n \Gamma(X , \Omega^{i}(i.x) )$ follows from properness of the ordinary Hitchin map $M^{ps}( S , GL(n,\mathbb{C}) ) \to \bigoplus_{i=1}^n \Gamma( S , Sym^i( T^*S ) )$ on the surface $S$, which is well known \cite[Theorem 6.11]{SIM}.
\end{rem}
\subsection{Complete integrability}
\begin{lem}\label{l:aut}
Let $(E , \phi)$ be a polystable $\Gamma$-equivariant $G$-Higgs bundle on $Y$. Under the non-abelian Hodge correspondence, $(E,\phi)$ corresponds to a reductive representation $\rho : \pi_1^{orb}(X) \to G$ of the orbifold fundamental group. Let $Aut_\Gamma(E,\phi)$ be the group of automorphisms of $(E,\phi)$ commuting with the action of $\Gamma$ and let $Aut(\rho) = \{ g \in G \; | \; g \rho g^{-1} = \rho\}$. Then we have an inclusion $Aut_\Gamma(E,\phi) \subseteq Aut(\rho)$.
\end{lem}
\begin{proof}
Let $V$ be any representation of $G$ and set $V_E = E \times_G V$. Then $\phi$ induces a bundle endomorphism $\phi_E : V_E \to V_E \otimes \Omega^1_Y$. Polystability of $(E,\phi)$ implies polystability of $(V_E , \phi_E)$. Thus there exists a hermitian metric $h( \; , \; )$ on $V_E$ (taken to be conjugate linear in the second variable) such that $(V_E , \phi_E)$ satisfies the Hitchin equations
\[
F_A + [\phi_E , \phi^*_E] = 0,
\]
where $F_A$ is the curvature of the Chern connection $\nabla_A = \partial_A + \overline{\partial}$ on $V_E$ induced by $h$. Now suppose $s$ is a holomorphic section of $V_E$ such that $\phi_E(s) = 0$. We claim that $\partial_A s = \phi^*_Es = 0$. To see this consider the following:
\begin{equation*}
\begin{aligned}
0 & =i\int_Y h( F_A s + [\phi_E , \phi^*_E]s , s ) \\
& = i\int_Y h( (\overline{\partial}_A \partial_A s , s) + i\int_Y h( \phi_E \phi^*_E s , s) \\
&= i\int_Y h( \partial_A s , \partial_A s ) - i\int_Y h( \phi^*_E s , \phi^*_E s) \\
&= || \partial_A s ||^2 + || \phi^*_E s ||^2.
\end{aligned}
\end{equation*}
Thus $\partial_A s = \phi^*_E s = 0$, as claimed. Under the non-abelian Hodge correspondence, $(E,\phi)$ corresponds to a flat $G$-connection $\nabla$. The flat connection $\nabla$ acts as the flat connection on $V_E$ given by $\nabla = \nabla_A + \phi_E + \phi^*_E$. Therefore $\nabla s = 0$ and so $s$ is covariantly constant. To prove the lemma, consider the case where $V = Hom( \fg , \fg )$ and $s : ad(E) \to ad(E)$ is an automorphism $s \in Aut_\Gamma(E , \phi)$. The fact that $s$ is an automorphism of $(E,\phi)$ means that $\phi_E (s) = [\phi , s] = 0$. By the above argument, $s$ is an automorphism of $\nabla$ which moreover commutes with $\Gamma$, i.e. $s \in Aut( \rho)$.
\end{proof}
\begin{lem}\label{l:openzg}
Let $Z(G) \subset G$ be the center of $G$. The open substack of $\cM_\cG$ consisting of pairs $(\cE , \phi)$ whose automorphism group is $Z(G)$ is non-empty.
\end{lem}
\begin{proof}
Let $p : Y \to X$ be a Galois covering of $X$ with Galois group $\Gamma$. Recall that we can identify $\cG$-Higgs bundles on $X$ with $\Gamma$-equivariant $G$-Higgs bundles on $Y$ of type $\tau$ and this is an equivalence of stacks. Therefore it is enough to show that there exists a $\Gamma$-equivariant $G$-Higgs bundle on $Y$ of type $\tau$ whose automorphism group (as a $\Gamma$-equivariant Higgs bundle) is $Z(G)$. By the non-abelian Hodge correspondence, polystable $\Gamma$-equivariant $G$-Higgs bundles of type $\tau$ correspond to reductive representations of the orbifold fundamental group of $X$ for which the holonomy of a loop around $x$ takes values in the conjugacy class containing $\tau$. It is easy to see there exists such representations $\rho$ whose image is Zariski dense in $G$. Such a representation has automorphism group $Z(G)$. By Lemma \ref{l:aut}, the $\Gamma$-equivariant Higgs bundle $(E,\phi)$ on $Y$ corresponding to $\rho$ will have automorphism group contained in $Z(G)$. However, every element of $Z(G)$ clearly acts as a $\Gamma$-equivariant Higgs bundle automorphism of $(E,\phi)$, hence $Aut_\Gamma(E,\phi) = Z(G)$.
\end{proof}
\begin{thm}\label{t:fibres}
Let $h_\cG :T^*\Bun_{\cG} \ra \cA_{\cG}$ denote the parahoric Hitchin map. Then the generic fibres of $h_\cG$ are $Z(G)$-gerbes over a disjoint union of abelian varieties. More precisely, there exists a non-empty Zariski open set $U\subset \cA_{\cG}$ such that the morphism $h^{-1}_\cG(U)\ra U$ is proper and smooth and each fibre is a $Z(G)$-gerbe over a disjoint union of abelian varieties.
\end{thm}
\begin{proof}
Clearly there exists a non-empty smooth Zariski open set $U \subset \cA_{\cG}$ such that the morphism $h^{-1}_\cG(U) \ra U$ is smooth. By Corollary \ref{c:flat} (ii), the restriction of $h_\cG$ to $h^{-1}_\cG(U)$ is flat and in particular an open map. Restricting $U$ if necessary, we can assume that $h_\cG : h^{-1}_\cG(U) \to U$ is surjective. Since $h_\cG |_{h^{-1}_\cG(U)}$ is open, we can further restrict $U$ so that $U \subseteq h_\cG( \cM^{s}_\cG)$. Note that $\cM^s_{\cG}$ is a non-empty open subset. To see this, note that it is known that there exists stable $\cG$-bundles \cite{BS}. If $\cE$ is a stable $\cG$-bundle, then $(\cE , 0 )$ is a stable $\cG$-Higgs bundle, so $\cM^s_\cG$ is non-empty.
Next, let $\mathcal{W} \subseteq \cM_\cG$ be the open substack of pairs $(\cE , \phi)$ whose automorphism group is $Z(G)$, which is non-empty by Lemma \ref{l:openzg}. Therefore, we may assume that $U$ is chosen with $U \subseteq h_\cG(\mathcal{W})$.
With our assumptions on $U$ we have that every fibre of $h^{-1}_\cG(U)$ contains a stable Higgs bundle and also contains a Higgs bundle with automorphism group $Z(G)$. We will give an argument using Hamiltonian flows to show that in fact every Higgs bundle in $h^{-1}_\cG(U)$ is stable and has automorphism group $Z(G)$. To carry out this argument we identify parahoric $\cG$-Higgs bundles with $\Gamma$-equivariant $G$-Higgs bundles on $Y$ of type $\tau$.
Let $(E,\phi)$ be a $\Gamma$-equivariant $G$-Higgs bundle. Thus $\phi \in H^0( Y , ad(E) \otimes \Omega_Y )^\Gamma$. Since we are working with $\Gamma$-equivariant Higgs bundles, one finds that the tangent space at smooth points is given as follows. Consider the $2$-term complex
\begin{equation*}\xymatrix{
ad(E) \ar[r]^-{[\phi , \, . \, ]} & ad(E) \otimes \Omega_Y.
}
\end{equation*}
Let us denote this complex by $[ \phi , \, . \, ]$. We have that $\Gamma$ acts on this complex and the tangent space to $(E,\phi)$ may be identified with $\mathbb{H}^1( Y , [ \phi , \, . \, ] )^\Gamma$. Let $\rho$ be an invariant polynomial on $\fg$ of degree $d_\rho$ and let $\mu \in H^1( Y , \Omega^{1-d_\rho}(-d_\rho x))^\Gamma$. Since $\Gamma$ is a finite group, by averaging over $\Gamma$, we may represent $\mu$ by a $\Gamma$-invariant section of $\Omega^{0,1}( Y , \Omega^{1-d_\rho}(-d_\rho x) )$. Define a function $f_{\rho , \mu}$ as in the proof of Theorem \ref{t:poissoncommute}. In other words, we use Serre duality to pair $\rho(\phi)$ with $\mu$ to get a complex number
\begin{equation*}
f_{\rho , \mu}(\cE,\phi) = \langle \rho(\phi) , \mu \rangle = \int_Y \rho(\phi) \mu \in \mathbb{C}.
\end{equation*}
Let $X_{\rho , \mu}$ be the Hamiltonian vector field associated to $f_{\rho , \mu}$. By essentially the same calculation as given in the proof of Theorem
\ref{t:poissoncommute}, we find that $X_{\rho ,\mu}$ evaluated at the point $(E , \phi)$ may be represented in Dolbeault cohomology by a cocycle of the form $(\alpha_{\rho , \mu}(\phi) , 0 )$, where $\alpha_{\rho , \mu}(\phi)$ is defined by:
\begin{equation*}
\kappa( \alpha_{\rho,\mu}(\phi) , \, \_ \, ) = d_\rho \rho( \phi , \phi , \dots , \phi , \, \_ \, ) \mu.
\end{equation*}
Moreover, since $\mu$ was chosen to be $\Gamma$-invariant, $\alpha_{\rho,\mu}$ is also $\Gamma$-invariant. For $t \in \mathbb{R}$, let $(E_t , \phi_t)$ be the $\Gamma$-equivariant Higgs bundle obtained by flowing $(E , \phi)$ along $X_{\rho , \mu}$ for time $t$ and let $\overline{\partial}_{E_t}$ be the $\overline{\partial}$-operator on $ad(E)$ defining the holomorphic bundle $E_t$. Since $X_{\rho , \mu}(E_t , \phi_t) = (\alpha_{\rho , \mu}(\phi_t) , 0)$, we see that $\partial_t \phi_t = 0$, or $\phi_t = \phi$ is constant. Thus $\alpha_{\rho , \mu}(\phi_t) = \alpha_{\rho , \mu}(\phi)$ is $t$-independent and it follows that
\begin{equation*}
(\overline{\partial}_{E_t} , \phi_t ) = (\overline{\partial}_{E} + t \alpha_{\rho , \mu}(\phi) , \phi).
\end{equation*}
Since $\alpha_{\rho , \mu}(\phi)$ is $\Gamma$-invariant, we see that the $\Gamma$-action on $E$ acts by automorphisms of $(\overline{\partial}_{E_t} , \phi_t)$ for any $t$. In this way, the $(\overline{\partial}_{E_t} , \phi_t)$ become $\Gamma$-equivariant Higgs bundles on $Y$. Now suppose that $\psi : E \to E$ is a $\mathcal{C}^\infty$ principal bundle isomorphism that preserves $\phi$. Then as $\psi$ preserves the Killing form $\kappa$ and the invariant polynomial $\rho$, we find that $\psi$ preserves $\alpha_{\rho , \mu}(\phi)$. Therefore $\psi$ preserves $\overline{\partial}_{E}$ if and only if it preserves $\overline{\partial}_{E_t}$ for any given $t \in \mathbb{R}$. In particular this means that the automorphism group of $(E_t , \phi_t)$ is independent of $t$. Suppose we had chosen $(E , \phi)$ with automorphism group $Z(G)$. Then for each $t$, the automorphism group of $(E_t , \phi_t)$ is also $Z(G)$. In a similar manner we see that for each $t$, $(E_t , \phi_t)$ is stable if and only if $(E , \phi )$ is stable.
Observe that the Hamiltonian flows above are complete, i.e., the flows exist for all time $t$. This implies that each connected component of each fibre of $h^{-1}_\cG(U) \to U$ has universal cover given by $\mathbb{C}^n$, such that the Hamiltonian flows are given by translations. In particular, it follows that any two points in the same connected component of a fibre of $h^{-1}_\cG(U) \to U$ are connected by a series of Hamiltonian flows along vector fields of the form $X_{\rho_j , \mu_j}$. Thus any two Higgs bundles in the same connected component of a fibre have the same automorphism group. This shows that $\mathcal{W} \cap h^{-1}_\cG(U)$ is non-empty and is open and closed in the analytic topology. But $h^{-1}_\cG(U)$ is connected since it is a Zariski open subset of $T^*Bun_{\cG}$, which is irreducible as $Bun_{\cG}$ is very good. Therefore $\mathcal{W} \cap h^{-1}_\cG(U) = h^{-1}_\cG(U)$ and so every Higgs bundle in $h^{-1}_\cG(U)$ has automorphism group $Z(G)$, as claimed. A similar argument shows that $\cM^{s}_\cG \cap h^{-1}_\cG(U) = h^{-1}_\cG(U)$, so every Higgs bundle in $h^{-1}_\cG(U)$ is stable.
The underlying space of $\cM^s_\cG$ can be identified with the coarse moduli space $M^s_\cG$ of stable parahoric Higgs bundles. The above results show that $h^{-1}_\cG(U)$ is a $Z(G)$-gerbe over its underlying coarse moduli space, which is an open subset in $M^s_\cG$. Next we use that the Hitchin map on the coarse moduli space of semistable parahoric Higgs bundles is proper to see that the restriction $h_\cG : h^{-1}_\cG(U) \to U$ is proper.
Let $F$ be the underlying space of a connected component of a fibre of $h^{-1}_\cG(U) \to U$. Then $F$ is compact, since $h$ is proper. The Hamiltonian vector fields $X_1, \dots , X_m$ associated to the coordinates $h_1 , \dots , h_m$ of the Hitchin map give a global frame of commuting holomorphic vector fields on $F$. Since $F$ is compact in the analytic topology, then the existence of such vector fields implies that $F$ is biholomorphic to a complex torus. Moreover we have that $T^*Bun_{\cG}$, $\cA_{\cG}$ and $h_\cG$ are all algebraic, hence $F$ is itself algebraic. It is well known that a complex torus which is algebraic is an abelian variety, so we conclude that $F$ is an abelian variety.
\end{proof}
\begin{rem}
We have shown that each connected component of a generic fibre of the Hitchin map is a $Z(G)$-gerbe over an abelian variety. It is natural to conjecture that this is a trivial gerbe, so that each connected component is the product of an abelian variety with the classifying stack of $Z(G)$. Indeed, this is the case in the usual setting \cite{DP} and is a requirement for Langlands duality of Hitchin systems for Langlands dual groups. Triviality in type $A$ can be seen using the spectral data description of fibres of the Hitchin system \cite{SS}. Showing triviality of the gerbe along the fibres for more general parahorics presumably requires the development of a theory of ``parahoric cameral data". We leave this as an interesting topic for future work.
\end{rem}
\begin{bibdiv}
\begin{biblist}
\bib{BS}{article}
{
Author = {Balaji, V.},
Author = {Seshadri, C.S.},
Title = {Moduli of parahoric $\mathcal{G}$-torsors on a compact Riemann surface}
Year = {2015}
Journal = {J. Algebraic Geom.}
Volume = {24}
Number= {1}
Pages = {1--49}
}
\bib{BK}{article}
{
Author = {Baraglia, D.},
Author = {Kamgarpour, M.},
Title = {On the image of the parabolic Hitchin map},
Journal = {in preparation},
Year={2017},
}
\bib{BD}{article}
{
AUTHOR = {Beilinson, A.},
Author={Drinfeld, V.},
TITLE = {Quantization of Hitchin's integrable system and Hecke eigensheaves},
YEAR = {1997},
}
\bib{BGPM}{article}
{
Author = {Biquard, O.}
Author= {Garc\'ia-Prada, O.}
Author = {Mundet i Riera, I.}
Title = {Parabolic Higgs bundles and representations of the fundamental group of a punctured surface into a real group},
Year = {2015}
Journal={arXiv preprint arXiv:1510.04207v3},
}
\bib{BoalchParahoric}{article}
{
AUTHOR = {Boalch, P. P.},
TITLE = {Riemann-{H}ilbert for tame complex parahoric connections},
JOURNAL = {Transform. Groups},
FJOURNAL = {Transformation Groups},
VOLUME = {16},
YEAR = {2011},
NUMBER = {1},
PAGES = {27--50},
}
\bib{Bot}{article}
{
AUTHOR = {Bottacin, F.},
TITLE = {Symplectic geometry on moduli spaces of stable pairs},
YEAR = {1995},
Journal={Ann. Sci. \'{E}cole Norm. Sup.}
Volume={(4) 28}
Number={4}
Pages={391--433}
}
\bib{BT1}{article}
{
Title={Groupes r{\'e}ductifs sur un corps local},
Author={Bruhat, F.},
Author={Tits, J.},
Journal={Publ. Math. Inst. Hautes Etudes Sci.},
Volume={60},
Number={1},
Pages={5--184},
Year={1984},
Publisher={Springer}
}
\bib{BT2}{article}
{
title={Groupes r{\'e}ductifs sur un corps local},
Author={Bruhat, F.},
Author={Tits, J.},
Journal={Publ. Math. Inst. Hautes Etudes Sci.},
Volume={41},
Number={1},
Pages={5--251},
Year={1972},
Publisher={Springer}
}
\bib{DP}{article}
{
Author = {Donagi, R.}
Author = {Pantev, T.}
Title = {Langlands duality for Hitchin systems}
Year = {2012}
Journal = {Invent. Math.}
Volume = {189}
Number = {3}
Pages = {653--735}
}
\bib{Faltings}{article}
{
AUTHOR = {Faltings, G.},
TITLE = {Stable $G$-bundle and projective connections},
YEAR = {1993},
Journal = {J. Algebraic Geom.}
Volume = {2}
Number = {3}
Pages = {507--568}
}
\bib{Ginzburg}{article}
{
Author={Ginzburg, V.},
Title = {The global nilpotent variety is Lagrangian}
Year={2001}
Journal = {Duke Math. J.}
Volume = {109}
Number = {3}
pages = {511--519}
}
\bib{Heinloth}{article}
{
Author={Heinloth, J.},
Title = {Uniformization of $\mathcal{G}$-bundles}
Year={2010}
Journal = {Math. Ann.}
Volume = {347}
Number = {3}
Pages = {499--528}
}
\bib{Hitchin1}{article}
{
AUTHOR = {Hitchin, N. J.},
TITLE = {The self-duality equations on a {R}iemann surface},
JOURNAL = {Proc. London Math. Soc. (3)},
FJOURNAL = {Proceedings of the London Mathematical Society. Third Series},
VOLUME = {55},
YEAR = {1987},
NUMBER = {1},
PAGES = {59--126},}
\bib{Hitchin}{article}
{
Author={Hitchin, N.},
Title={Stable bundles and integrable systems},
Year={1987}
Journal = {Duke Math. J.}
Volume = {54}
Number = {1}
Pages = {91--114}
}
\bib{KumarBook} {book}
{
Title={Kac-Moody Groups, their Flag Varieties and Representation Theory},
Author={Kumar, S.},
Volume={204},
Year={2002},
Publisher={Springer Science \& Business Media}
}
\bib{LS}{article}{
AUTHOR = {Laszlo, Y.},
Author={Sorger, C.},
TITLE = {The line bundles on the moduli of parabolic {$G$}-bundles over
curves and their sections},
JOURNAL = {Ann. Sci. \'Ecole Norm. Sup. (4)},
FJOURNAL = {Annales Scientifiques de l'\'Ecole Normale Sup\'erieure.
Quatri\`eme S\'erie},
VOLUME = {30},
YEAR = {1997},
NUMBER = {4},
PAGES = {499--525}
}
\bib{LogaresMarten}{article}
{
Title={Moduli of parabolic Higgs bundles and Atiyah algebroids},
Author={Logares, M.},
Author={Martens, J.},
Journal={J. Reine Angew. Math.},
Volume={2010},
Number={649},
Pages={89--116},
Year={2010}
}
\bib{Markman}{article}
{
Author ={Markman, E.},
Title={Spectral curves and integrable systems},
Year={1994}
Journal={Compositio Math.}
Volume={93}
Number={3}
Pages={255--290}
}
\bib{Ngo}{article}
{
author = {Ng\^o, B. C.}
title = {Le lemme fondamental pour les alg\`ebres de Lie.}
journal = {Publ. Math. Inst. Hautes \'Etudes Sci.}
volume = {111}
year = {2010}
pages = {1--169}
}
\bib{SS}{article}
{
Author={Scheinost, P.},
Author={Schottenloher, M.},
Title={Metaplectic quantization of the moduli spaces of flat and parabolic bundles},
Year={1995},
Journal={J. Reine Angew. Math.}
Volume={466},
Pages={145--219}
}
\bib{Simpson}{article}
{
AUTHOR = {Simpson, C. T.},
TITLE = {Higgs bundles and local systems},
JOURNAL = {Inst. Hautes \'Etudes Sci. Publ. Math.},
FJOURNAL = {Institut des Hautes \'Etudes Scientifiques. Publications
Math\'ematiques},
NUMBER = {75},
YEAR = {1992},
PAGES = {5--95},
}
\bib{SIM}{article}
{
Author = {Simpson, C. T.}
Title = {Moduli of representations of the fundamental group of a smooth projective variety. II.}
Journal ={Inst. Hautes \'Etudes Sci. Publ. Math.},
Volume ={80},
Year = {1994},
Pages = {5-79},
}
\bib{VAR}{article}
{
Author = {Varma, R.},
Title = {On Higgs bundles on elliptic surfaces},
Journal = {Q. J. Math.},
Volume = {66},
Year = {2015}
Number = {3},
Pages = {991-1008}
}
\bib{YOK}{article}
{
Author = {Yokogawa, K.},
Title = {Compactification of moduli of parabolic sheaves and moduli of parabolic Higgs sheaves},
Journal = {J. Math. Kyoto Univ.},
Volume = {33},
Year = {1993},
Number = {2},
Pages = {451-504}
}
\bib{Yu}{article}
{
title = {Smooth models associated to concave functions in Bruhat-Tits theory}
author = {Yu, J.-K.}
journal = {preprint}
year = {2002}
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{Introduction}
V-T theory of liquid dynamics is being developed within the framework of many-body theory. The liquid atomic motion has two contributions, many-body vibrational motion described by a first-principles liquid-specific Hamiltonian, and transit motion described by a parameterized model. Vibrations express the dominant part of the liquid thermal energy, while transits provide the liquid diffusion. It was recently shown that a proper adjustment of the transit parameters yields an extremely accurate account of molecular dynamics (MD) data for the mean square displacement time-correlation function (MSD) for liquid Na~\cite{MSD}. In the present study we employ precisely the same atomic motion to calculate the self-intermediate scattering function (SISF) of the same liquid Na system. While the MSD constitutes a single scalar measure of the motional decorrelation process, the SISF measures the complete Fourier transform of that process. This extra complication of the SISF, plus the use of an independent calibration of the motion, poses a stringent test of the V-T theoretical formulation. Our aim here is to carry out this test by comparing theory with MD, and assess the results.
Unless otherwise stated, we consider monatomic systems. By hypothesis, the potential surface underlying the liquid atomic motion is overwhelmingly dominated by the random class of $3N$-dimensional potential energy valleys~\cite{DCW1997}. These valleys are macroscopically uniform, so that a single such valley is sufficient for statistical mechanical calculations~\cite{DCW1997}. This hypothesis has been verified extensively for the present Na system~\cite{WC1999, CW1999, Holm2009, Holm2010}, and has been verified by density functional theory (DFT) calculations for Na and Cu~\cite{Nick2010}, for Ga~\cite{Sven2014}, and for a five component metallic glass~\cite{Holm2010}. The hypothesis is verified in each instance where a single-valley vibrational contribution rationalizes MD data, as in the MSD graphs~\cite{MSD}, and in the SISF graphs of the present study. A more detailed discussion of the liquid potential energy surface, and of the technique for calculating the vibrational Hamiltonian, is provided in~\cite{MSD}.
Ninety years ago, Frenkel argued that a liquid atom oscillates for a time about one equilibrium position, then jumps to a new one and oscillates there~\cite{F1926, F1946}. In developing V-T theory, we have added much crucial information to this description. First, the vibrational motion is now the fully correlated $3N$-dimensional normal-mode motion. Second, transits replace jumps. Individual transits were observed in low-temperature equilibrium MD trajectories, in metastable liquid states of Ar and Na~\cite{WCC2001}. Each observed transit is the highly correlated motion of a small local group of atoms that carries the system across the boundary between two liquid potential energy valleys. These transits are mechanical motion, not comparable to the much larger statistical-mechanical cooperatively rearranging regions of Adam and Gibbs~\cite{AG1965}. The transit-induced motion of the atomic equilibrium positions, abbreviated transit motion, is continuous and has physically meaningful time dependence (section~III; see also~\cite{MSD}).Transit theory is still under development, and the present work is a step in that direction.
In section~II, vibrational theory for the SISF is outlined, and the complete system correlation components and decorrelation processes are mathematically defined. Transit theory for the SISF is outlined in section~III, and the transit damping coefficients are derived. Equations for the theoretical SISF are derived in section~IV, and their accuracy is verified for a representative $q$ value. In section~V, the complete theory is applied to a set of $q$ spanning all physical behaviors, from hydrodynamic at small $q$ to free particle at large $q$. The deviation of the theory from MD data is tabulated and discussed. Mode coupling theory (MCT) and V-T theory are compared in terms of their respective descriptions of the underlying atomic motion in section~VI. Major conclusions are outlined in section~VII.
\section{Vibrational theory}
We study a system representing liquid Na at $395$K, based on a well-tested interatomic potential from pseudopotential perturbation theory (see figure 1 in~\cite{WC1999}). The system has $N$ atoms in a cubical box with periodic boundary conditions, with $N=500$ here. The atomic positions are $\bm{r}_{K}(t)$ at time $t, K=1,\dots,N$, and $\{\bm{q}\}$ is the set of wavevectors commensurate with the box periodicity. The SISF is~\cite{HMcD_3rded,BZ1994}
\begin{equation} \label{eq1}
F^{s}(q,t) = \frac {1}{N} \left < \sum_{K} e^{-i \bm{q}\cdot (\bm{r}_{K}(t)-\bm{r}_{K}(0))}
\right >.
\end{equation}
The brackets indicate the average over the atomic motion in an equilibrium state, plus the average over all $\bm{q}$ for each $q$ magnitude. The $\bm{q}$ average is the last to be done, and we shall omit its explicit notation. We work in classical statistical mechanics, which allows comparison with MD data and which is accurate for most elemental liquids, including Na. Equation~(\ref{eq1}) is used for direct calculation of the MD function
$F_{MD}^{s}(q,t)$.
Vibrational theory for the SISF is derived and discussed in~\cite{G1}; a brief summary is helpful here.
The vibrational contribution to the liquid atomic motion is harmonic vibrations in a single random valley. Each atom moves with displacement $ \bm{u}_{K}(t)$ away from the fixed equilibrium position $\bm{R}_{K}$, so that
\begin{equation} \label{eq2}
\bm{r}_{K}(t)=\bm{R}_{K} + \bm{u}_{K}(t).
\end{equation}
For this motion, equation~(\ref{eq1}) is
\begin{equation} \label{eq3}
F_{vib}^{s}(q,t)=\frac {1}{N} \left < \sum_{K} e^{-i \bm{q}\cdot (\bm{u}_{K}(t)-\bm{u}_{K}(0))}
\right >_{vib},
\end{equation}
where $<\dots>_{vib}$ indicates the average over the vibrational motion. With Bloch's identity (equation N15 of \cite{AM1976}), equation~(\ref{eq3}) becomes
\begin{equation} \label{eq4}
F_{vib}^{s}(q,t)=\frac {1}{N} \sum_{K} e^{-2W_{K}(\bm{q})}\;
e^{\left < \bm{q}\cdot\bm{u}_{K}(t) \;\;\bm{q}\cdot\bm{u}_{K}(0)\right >_{vib}},
\end{equation}
where
\begin{equation} \label{eq5}
2 W_{K}(\bm{q})=\left < (\bm{q}\cdot\bm{u}_{K}(0))^{2}\right > _{vib}.
\end{equation}
To evaluate the average in equation~(\ref{eq4}), the vibrational Hamiltonian is diagonalized in terms of the vibrational normal modes, labeled $\lambda=1,\dots,3N-3$, where the three zero frequency modes are omitted from all calculations. Following the algebra in section 5.1 of \cite{CW2001}, we find
\begin{equation} \label{eq6}
\left < \bm{q}\cdot\bm{u}_{K}(t)\;\; \bm{q}\cdot\bm{u}_{K}(0)\right >_{vib} =
\frac{kT}{M} \sum_{\lambda} (\bm{q}\cdot \bm{w}_{K\lambda})^{2}
\; \frac{\cos \omega_{\lambda}t}{\omega_{\lambda}^{2}},
\end{equation}
where $T$ is temperature, $M$ is the atomic mass, $\bm{w}_{K\lambda}$ is the Cartesian vector of the $K$ component of eigenvector $\lambda$, and $\omega_{\lambda}$ is the corresponding frequency. The functions $e^{-2W_{K}(\bm{q})}$ in equation~(\ref{eq4}) are Debye-Waller factors, and the left side of equation~(\ref{eq6}) is the vibrational time-correlation function.
In this section and the next two, we shall illustrate the theory for a single representative $q$, namely $q_{1}$ at the first peak of the structure factor $S(q)$. Figure~\ref{fig1} shows $F_{MD}^{s}(q,t)$ and $F_{vib}^{s}(q,t)$, together with the ultimate constant value $F_{vib}^{s}(q,\infty)$ of $F_{vib}^{s}(q,t)$.
\begin{figure} [h]
\includegraphics [width=0.70\textwidth]{Figure1.eps}
\caption{Dots show $F_{MD}^{s}(q,t)$ at MD time steps, and line is $F_{vib}^{s}(q,t)$. Dashed line is $F_{vib}^{s}(q,\infty)$, the $t\rightarrow \infty$ limit of $F_{vib}^{s}(q,t)$. The system correlation per atom at $t=0$ has the marked vibrational and structural components. $\tau_{RW}$ is explained in the text.}
\label{fig1}
\end{figure}
At $t=0$, the contribution to equation~(\ref{eq1}) is $1$ for every atom, and this is maintained in equations~(\ref{eq4}) and (\ref{eq5}). As $t$ increases from zero, the theoretical motion is free particle (ballistic), for which $\bm{r}_{K}(t)-\bm{r}_{K}(0)=\dot{\bm{r}}_{K}(0) t$. With this motion the leading $t$-dependence of equation~(\ref{eq1}) is
\begin{equation} \label{eq7}
F_{free}^{s}(q,t) = e^{-a(q)t^{2}},
\end{equation}
where $a(q)=kTq^{2}/2M$. Precisely this leading $t$-dependence, to exponential order, is contained in equations~(\ref{eq4}) and (\ref{eq6}). It is shown below that the vibrational motion is pure ballistic for only a very short time.
Following the ballistic regime, the $t$-dependence of $F_{vib}^{s}(q,t)$ is controlled by vibrational dephasing, which is the interference of the $\cos \omega_{\lambda}t$ terms in equation~(\ref{eq6}). This process continues until the vibrational time correlation functions are zero. Then from equation~(\ref{eq4}), the ultimate value of $F_{vib}^{s}(q,t)$ is
\begin{equation} \label{eq8}
F_{vib}^{s}(q,\infty) = \frac{1}{N} \sum_{K}e^{-2W_{K}(\bm{q})}, \;\; t\geq t_{\infty}.
\end{equation}
From extension of figure~\ref{fig1} to longer times we find $t_{\infty}\approx 300 \delta{t}$. $t_\infty$ is not a calibration time in the present study.
We can define the correlations and the decorrelation processes involved in the SISF at all $q$ and $t$. The correlations are between $\bm{r}_{K}(t)$ and $\bm{r}_{K}(0)$, averaged over the motion, as expressed in equation~(\ref{eq1}). In V-T theory, these correlations have two independent components, vibrational and structural. At $t=0$, the vibrational correlation is $1-F_{vib}^{s}(q,\infty)$ per atom, and the structural correlation is $F_{vib}^{s}(q,\infty)$. This decomposition is marked in figure~\ref{fig1}.
The Debye-Waller factor measures the mean structural correlation for a system of atoms vibrating about fixed equilibrium positions. For such a system, e.g. an amorphous solid, the structural correlation is not damped. Vibrational dephasing damps the vibrational correlation to zero while the structural correlation remains. This is shown by $F_{vib}^{s}(q,t)$ in figure~\ref{fig1}. The $q$-dependence of this process is controlled by the $q^2$ factor in equation~(\ref{eq5}), which appears in the exponent of equation~(\ref{eq8}). This makes $F_{vib}^{s}(q,\infty)$ approach 1 as $q\rightarrow 0$, and carries $F_{vib}^{s}(q,\infty)$ to zero as $q$ increases. The progression is shown in figure~\ref{fig2}. The representative $q$ chosen here is $q_{1}=1.1$, which divides the initial correlation nearly evenly between vibrational and structural components (see figure~\ref{fig1}). The primary function of transits in the present theory is to damp the structural correlation to zero.
\begin{figure} [h]
\includegraphics [width=0.70\textwidth]{Figure2.eps}
\caption{For $q$ from $0.3$ to $6.0$ a$_{0}^{-1}$: $F_{vib}^{s}(q,\infty)$ is near $1$ at $q=0.3$a$_{0}^{-1}$ and goes to zero around $\tau_{D}$ at $q=3.5$. As $q$ increases from $3.5$, $F_{vib}^{s}(q,t)$ goes to zero in an ever shorter time.}
\label{fig2}
\end{figure}
\section{Transit Theory}
\begin{figure} [h]
\includegraphics [width=0.70\textwidth]{Figure3.eps}
\caption{Dots are $F_{MD}^{s}(q,t)$, line is $F_{vib}^{s}(q,t)$, and their difference is due to transit damping. The slope discontinuity in $F_{vib}^{s}(q,t)$ at $\tau_{RW}$ is due to a small transit effect. $\tau_{B}$ is explained in the text.}
\label{fig3}
\end{figure}
Figure~\ref{fig3} compares $F_{MD}^{s}(q,t)$ and $F_{vib}^{s}(q,t)$ on three time intervals. The curves in figure~\ref{fig3} approximate an inversion of the corresponding MSD curves~\cite{MSD}. Our MSD study provides the following brief summary of the atomic motion underlying figure~\ref{fig3}. On the first interval, until the delay time $\tau_{D}$, the MD system measures only pure vibrational motion. On the crossover interval, the vibrational contribution saturates to its ultimate constant value while the MD curve continues to decay toward zero. The difference between the MD and vibrational curves is transit damping, which is first resolved at $\tau_{D}$ and reaches its full steady state at $\tau_{RW}$. Besides damping the structural correlation, an additional small transit effect is to damp away the final vibrational correlation, and make $F_{vib}^{s}(q,t)$ constant at $t \geq \tau_{RW}$. The approximation for this effect introduces the small slope discontinuity in $F_{vib}^{s}(q,t)$ at $\tau_{RW}$ (for details see~\cite{MSD}).
The commonly studied time correlation functions, including the ones treated here, measure the trajectory of one atom at a time, as in equation~(\ref{eq1}). We therefore need only the separate transit motion of a single atom, without explicit accounting of its motional correlation with other atoms. The single atom behavior is encoded in the V-T decomposition of the liquid motion:
\begin{equation} \label{eq9}
\bm{r}_{K}(t)=[\bm{R}_{K}(t)-\bm{R}_{K}(0)] + [\bm{u}_{K}(t)+\bm{R}_{K}(0)].
\end{equation}
The second bracket is vibrational motion about the permanent equilibrium position $\bm{R}_{K}(0)$, as in equation~(\ref{eq2}). The first bracket expresses the transit induced motion of the atomic equilibrium positions. The distance moved by the equilibrium position of one atom in one transit, averaged over transits, is denoted $\delta{R}$. The motion $\delta{R(t)}$ starts from zero, and the liquid system cannot sense the transit motion until it has reached a sufficient magnitude. The corresponding time required for the MD system in figure~\ref{fig3} to resolve a transit, and begin measuring it, is identified as $\tau_{D}$. This explains why the MD system measures only vibrational motion for $t$ up to $\tau_{D}$.
Transits are proceeding uniformly throughout the liquid at a high rate. The one-atom transit rate is $\nu$, and we make the following uniform transit timing approximation. Every atom transits once in the transit period $\nu^{-1}$, the transits are uniformly distributed over time, and the same sequence of one-atom transits occurs in every succeeding transit period. In an increment of time, a time correlation function averages this motion over a great many transits.
By some time after $t=0$, every transit motion is resolved from its beginning and measured to its completion. Then the net effect of the transit motion is to move the equilibrium position of every atom a randomly directed distance $\delta{R}$ in a time $\nu^{-1}$. This motion constitutes a steady-state random walk of each atomic equilibrium position. It is no longer necessary to account for the $t$-dependence of the transit motion; we can simply record each transit as it is completed.
The damping factor for the steady-state transit random walk is derived in~\cite{G1}. Now we need damping factors for two successive time intervals. To set this up, we consider $\bm{R}_{K}(t)-\bm{R}_{K}(0)$, from equation~(\ref{eq9}), to be a random walk of step rate equal to the transit rate $\nu$, but of arbitrary step distance $S$ and start time $\tau$. Equation~(\ref{eq1}) then takes the form
\begin{equation} \label{eq10}
F_{VT}^{s}(q,t) = \frac {1}{N} \sum_{K} \left < \left <e^{-i \bm{q}\cdot (\bm{R}_{K}(t)-\bm{R}_{K}(0))}\;\;e^{-i \bm{q}\cdot (\bm{u}_{K}(t)-\bm{u}_{K}(0))}
\right >_{trans} \right>_{vib},
\end{equation}
where $<\dots>_{trans}$ is the average over transits. We shall neglect correlations between the two motions, and average the first exponential uniformly over all system transits. This average removes the $K$ dependence of the first factor and also removes its dependence on the direction of $\bm{q}$. The required average can therefore be written
\begin{equation} \label{eq11}
A(q,t)=\left <e^{-i \bm{q}\cdot (\bm{R}(t)-\bm{R}(0))} \right >_{trans}.
\end{equation}
Equation~(\ref{eq10}) becomes
\begin{equation} \label{eq12}
F_{VT}^{s}(q,t)=A(q,t)\; F_{vib}^{s}(q,t)
\end{equation}
where $F_{vib}^{s}(q,t)$ is given by equations~(\ref{eq3}) and (\ref{eq4}).
In a time increment $\delta{t}$, we have
\begin{equation} \label{eq13}
\delta A(t )= \left< \left[e^{-i \bm{q}\cdot \bm{R}(t+\delta t)}-e^{-i\bm{q}\cdot \bm{R}(t)}\right]
e^{i \bm{q}\cdot\bm{R}(0)} \right>_{trans}.
\end{equation}
We shall record each transit step in the $\delta{t}$ in which it is completed. The probability a given atom completes a step in $\delta{t}$ is $\nu \delta{t}$. If the atom completes a step in $\delta{t}$, $\bm{R}(t+\delta t) = \bm{R}(t) + \bm{S}$; otherwise $\bm{R}(t+\delta t) = \bm{R}(t)$. Therefore
\begin{equation} \label{eq14}
\delta A(t)=\left< [ e^{-i \bm{q}\cdot \bm{S}}-1]
\;\;e^{-i\bm{q}\cdot (\bm{R}(t)-\bm{R}(0))} \right>_{trans} \nu \delta t.
\end{equation}
The average over transits is now an average over the uniformly distributed directions of $\bm{S}$. Equation~(\ref{eq14}) becomes
\begin{equation} \label{eq15}
\frac{\delta A(q,t)}{\delta t}= - \gamma(qS)\;A(q,t),
\end{equation}
where
\begin{equation} \label{eq16}
\gamma(x) = \nu \left[1-\frac{\sin x}{x}\right].
\end{equation}
Since the random walk begins at $\tau$, the solution of equation~(\ref{eq15}) is
\begin{equation} \label{eq17}
A(q,t) = \begin{cases}
1 &\text{for $t\leq \tau$,} \\
e^{-\gamma(qS)(t-\tau)} &\text{for $t\geq \tau$.}
\end{cases}
\end{equation}
The above equations will be assembled as needed to produce transit damping in the crossover and random walk intervals.
\section{Calibration and verification of theory}
We begin with the early calibration of two transit parameters, $\nu$ and $\delta{R}$. In our study of the MD trajectories of transiting atoms~\cite{WCC2001}, we recorded the mean $\delta{R}$ as around $0.40 R_{1}$ for Ar and $0.25 R_{1}$ for Na, where $R_{1}$ is the nearest neighbor distance. We have since used this $\delta{R}$ for liquid Na near the melting temperature $T_{m}$, since $\delta{R}$ is expected to have little $T$-dependence. The study was not exhaustive, and $\delta{R}$ can be in error by $25\%$ or so.
In order to calibrate the theoretical $F_{VT}^{s}(q,t)$, $\nu$ and $\delta{R}$ are needed separately in $\gamma(q\delta{R})$ (see equation~(\ref{eq16}). From an earlier calculation of the self-diffusion coefficient $D$ for liquid Na, we used the Einstein relation $D=\frac{1}{6} \nu(\delta{R})^2$, along with the above value of $\delta{R}$, to find $\nu$~\cite{G1}. The parameter combination $\nu(\delta{R})^2$ is extremely accurate, since the evaluation of $D$ is accurate to $1\%$ for our liquid Na system. These parameter calibrations, along with the others described below, are reported in~\cite{MSD} and are listed in Table I. We shall now discuss the remaining calibrations and the verification of theory for each interval in turn.
\begin{table}[ht]
\caption{\label{table1}Values for the parameters as defined in the text. The MD time step is $\delta t = 7.00288$~fs.}
\begin{ruledtabular}
\begin{tabular}{ccccc}
$\nu$&$\delta R$&$S$&$\tau_{D}$&$\tau_{RW}$\\
\hline
3.9~ps$^{-1}$&1.75~a$_{0}$&1.46~a$_{0}$&$28~\delta t$&$60~\delta t$\\
\end {tabular}
\end{ruledtabular}
\label{Table1}
\end{table}
\subsection{Vibrational interval:}
\begin{figure} [h]
\includegraphics [width=0.70\textwidth]{Figure4.eps}
\caption{Dots are $F_{MD}^{s}(q,t)$, line is $F_{vib}^{s}(q,t)$, and dashed line is the ballistic contribution, which is accurate only to $\tau_{B}=9\delta{t}$.}
\label{fig4}
\end{figure}
In our MSD study, $\tau_{D}$ is defined operationally as the time when $X_{MD}(t)$ begins to move away from $X_{vib}(t)$, and $\tau_{D}$ is calibrated from the comparison graph of the two functions. To apply the MSD calibration to $F_{VT}^{s}(q,t)$, we write
\begin{equation} \label{eq18}
F_{VT}^{s}(q,t)=F_{vib}^{s}(q,t), \;\; 0\leq t\leq \tau_{D}.
\end{equation}
The comparison with $F_{MD}^{s}(q,t)$ in figure~\ref{fig4} shows excellent agreement for the function and the time interval. Other details identical between MSD and SISF are: a) The ballistic contribution, shown in figure~\ref{fig4}, is accurate only to $\tau_{B}=9\delta{t}$, and accounts for only a very small part of the vibrational decorrelation; b) In figure~\ref{fig4}, $F_{VT}^{s}(q,t)$ shows a tiny positive departure from $F_{MD}^{s}(q,t)$, through the center of the interval, similar to the negative MSD departure in figure 2 of~\cite{MSD}, and negligible in both cases.
\subsection{Crossover interval:}
\begin{figure} [h]
\includegraphics [width=0.70\textwidth]{Figure5.eps}
\caption{Dots are $F_{MD}^{s}(q,t)$ and line is $F_{VT}^{s}(q,t)$ up to $\tau_{RW}$. The line beyond $\tau_{RW}$ is a continuation of crossover theory.}
\label{fig5}
\end{figure}
On the crossover interval, the MSD transit contribution is very well fitted by a straight-line segment from $\tau_{D}$ to $\tau_{RW}$ (figure (4) of~\cite{MSD}). Because it is linear in $t$, the straight line represents a random walk. However, this random walk is incomplete because its step distance $S$ is less than the complete $\delta{R}$ (see Table I). To calibrate
$F_{VT}^{s}(q,t)$ we must employ the same ``crossover walk" expressed in equation (10) of ~\cite{MSD}. For a random walk of step rate $\nu$, step distance $S$, and starting time $\tau$, $F_{VT}^{s}(q,t)$ is given by equations~(\ref{eq12}), (\ref{eq16}) and (\ref{eq17}). The present result for the crossover walk is therefore
\begin{equation} \label{eq19}
F_{VT}^{s}(q,t)=F_{vib}^{s}(q,t) \;\;e^{-\gamma(qS)(t-\tau_{D})} , \;\; \tau_{D}\leq t\leq \tau_{RW}.
\end{equation}
Comparison with $F_{MD}^{s}(q,t)$ in figure~\ref{fig5} shows excellent agreement.
Going back to the MSD, it is possible to achieve better agreement between theory and MD on the crossover. However, we chose to fit the MSD crossover with a simple random walk, because only that motion through equations~(\ref{eq10})-(\ref{eq17}) leads to the physically simple and tractable damping solution on the right side of equation~(\ref{eq19}).
\subsection{Random walk interval:}
In our MSD study, $\tau_{RW}$ is defined operationally as when the mean square displacement, $X_{MD}(t)$, begins its ultimate linear-in-$t$ straight line, whose slope is $6D=\nu(\delta{R})^{2}$. The calibration of $\tau_{RW}$ is possible because $X_{MD}(t)$ can be calculated to extreme accuracy for short times, with virtually no scatter, to $t$ well beyond $\tau_{RW}$. Applying this calibration to the SISF, the transit motion is a random walk of step rate $\nu$ and step distance $\delta{R}$, beginning at $\tau_{RW}$. The damping factor is given by equation~(\ref{eq17}), so that
\begin{equation} \label{eq20}
F_{VT}^{s}(q,t)=F_{VT}^{s}(q,\tau_{RW}) \;\;e^{-\gamma(q\;\delta R)(t-\tau_{RW})} , \;\; t\geq \tau_{RW}.
\end{equation}
The factor $F_{VT}^{s}(q,\tau_{RW})$ is given by equation~(\ref{eq19}) evaluated at $t=\tau_{RW}$. Comparison of equation~(\ref{eq20}) with $F_{MD}^{s}(q,t)$ in figure~\ref{fig6} shows excellent agreement at all $t$ beyond $\tau_{RW}$. Just as for the MSD, the theory is rather insensitive to the choice of $\tau_{RW}$. The reason is evident from figure~\ref{fig5}, which shows the crossover theory and MD nearly parallel for a while beyond $\tau_{RW}$. The steady-state random walk produces exponential damping for $t\geq\tau_{RW}$.
\begin{figure} [h]
\includegraphics [width=0.70\textwidth]{Figure6.eps}
\caption{Dots are $F_{MD}^{s}(q,t)$ and line is $F_{VT}^{s}(q,t)$. The deviation is insignificant at all $t$.}
\label{fig6}
\end{figure}
\section{Extension to all $q$}
We shall now apply the theory to a wide range of $q$, a range for which the initial structural correlation $F_{vib}^{s}(q,\infty)$ varies from near one at small $q$ to zero at large $q$. Comparison of $F_{VT}^{s}(q,t)$ with $F_{MD}^{s}(q,t)$ for seven $q$ is shown in figure~\ref{fig7}. The deviation function (the error) is
$\Delta{F^{s}(q,t)}$, defined by
\begin{equation} \label{eq22}
\Delta F^{s}(q,t)=F_{VT}^{s}(q,t) - F_{MD}^{s}(q,t).
\end{equation}
The deviation and other relevant parameters are listed for eight $q$ in Table \ref{Table2}. The accurate value of $q_1$ is $1.1050$ a$_{0}^{-1}$.
\begin{figure} [h]
\includegraphics [width=0.70\textwidth]{Figure7.eps}
\caption{For $q$ from $0.3$ to $6.0$ a$_{0}^{-1}$: Dots are $F_{MD}^{s}(q,t)$ and lines are $F_{VT}^{s}(q,t)$, where the $t$ scale is logarithmic. Deviation is so small it can be detected only at short segments on a few curves.}
\label{fig7}
\end{figure}
\begin{table}[ht]
\caption{\label{table2} Analysis of the deviation function. $F_{vib}^{s}(q,\infty)$ measures the structural correlation at $t=0$. $\gamma(q\delta R)$ is the transit damping coefficient on the random walk interval, equation~(\ref{eq20}). $t_{\epsilon}(q)$ is when $F_{MD}^{s}(q,t)$ reaches $\epsilon=0.002$, our numerical measure of zero. Columns 5-7 respectively list the largest deviation of the negative dip (vibrational interval), spread of deviation (crossover interval), and largest deviation of the negative dip (random walk interval).}
\begin{ruledtabular}
\begin{tabular}{ccccccc}
$q$(a$_{0}^{-1})$&$F_{vib}^{s}(q,\infty)$&$\gamma (q\delta R) $(ps$^{-1})$&$t_{\epsilon}(q)$(ps)&$0\leq t \leq \tau_{D}$&$\tau_{D}\leq t \leq \tau_{RW}$& $t \geq\tau_{RW}$\\
\hline
0.29711&0.94185&0.1733&30 & 0.000&-0.001 to 0.000&-0.003\\
0.70726&0.71699&0.9222&7.0 &-0.002&-0.005 to 0.004&-0.005\\
1.1050&0.45332&2.0146&3.1 &-0.003&-0.006 to 0.006&-0.003\\
1.5052&0.24047&3.1805&1.8 &-0.006&-0.004 to 0.007&-0.001\\
2.0041&0.08776&4.2975&1.0 &-0.007&-0.002 to 0.005&0.000\\
2.5064&0.02573&4.7423&0.60&-0.006& 0.000 to 0.003&0.000\\
3.5008&0.00126&3.9994&0.22&-0.003& 0.000 to 0.000&0.000\\
6.0013&0.00000&4.2271&0.08& 0.000& 0.000 to 0.000&0.000\\
\end {tabular}
\end{ruledtabular}
\label{Table2}
\end{table}
We shall describe the observable deviation features and assess their significance. The deviation at each $q$ exhibits a negative dip during the vibrational interval, which is visible in figure~\ref{fig4} and whose maximum negative deviation is listed in Table II. The likely cause of this dip is a small negative transit contribution omitted from equation~(\ref{eq18}); however, something like finite-$N$ error cannot be ruled out.
In fitting a random walk to the MSD crossover, and applying the same approximation to the SISF crossover, we are neglecting explicit nonlinear $t$ dependence of the transit motion. The corresponding errors in Table II are negligible, but this need not always be the case. We expect the crossover to show more complex behavior with increasing $T$.
Finally, the deviation exhibits a negative dip during the random walk interval for small $q$. The largest magnitude is at $q=0.70726$a$_{0}^{-1}$, Table II, and lies at $t$ around $400 \delta{t}$ in figure~\ref{fig7}. Our overall assessment of Table II is as follows. At all $q$ and $t$, the deviation is small enough that it indicates no significant error, numerical or theoretical, and no attempt to refine the numerical accuracy is warranted at this time.
In our original study of the SISF~\cite{G1}, the crossover interval was not resolved and theory was based on two time intervals, vibrational and random walk. The self dynamic structure factor was shown moderately more accurate then benchmark mode coupling theories~\cite{LV1970,WS1982}. In the present formulation, $\Delta(q,t)$ is everywhere smaller than the original by a factor of $3$ or more.
The original formulation is still correct, accurate, and a source of useful analysis. Two important limits were shown by analysis to derive from V-T theory~\cite{G1}, the hydrodynamic limit at small $q$ and the free particle limit at large $q$. Table II shows the vanishing of deviations as $q$ decreases toward the hydrodynamic regime. On the other hand, the free particle limit provides information not previously available for liquids on the overall convergence to the large $q$ limit~\cite{G1}. Notice the free particle motion is the leading $t$ dependence of vibrational theory, equation~(\ref{eq7}). Notice also as $q$ increases,
$F_{vib}^{s}(q,\infty)$ decreases to zero, figure~\ref{fig2}. When we follow $F_{VT}^{s}(q,t)$ as $q$ increases from the hydrodynamic regime, say via figure~\ref{fig7}, we see that $F_{VT}^{s}(q,t)$ goes to zero at an ever decreasing time; this time decreases through $\tau_{RW}$, then through $\tau_{D}$, and ultimately through $\tau_{B}$. The damping process changes dramatically as $q$ increases, but the initial motion is always vibrational and its initial segment is always free particle. This behavior is shown for both $F_{VT}^{s}(q,t)$ and its Fourier transform $S_{VT}^{s}(q,\omega)$ in figures~ (6)-(11) of~\cite{G1}. It is also shown in Table~II as the vanishing of the deviation as $q$ increases toward the free particle regime.
\section{Comparison of theories}
We begin by comparing the goal and the operational techniques of V-T theory with those of the broader field of research in dynamics of noncrystalline materials. In the past few decades, the broader field has focused on understanding the glass transition and the glassy state, and has coalesced around the technique of mode coupling theory (MCT). This technique has the remarkable ability to treat simple and complex systems with equal ease, from monatomic to molecular liquids, molten salts, colloids, polymers, and large organic molecules in the liquid phase. Diverse and valuable reviews are available~\cite{Gotze1999, Das2004, BK2005, BB2011}.
V-T theory specifically addresses liquid dynamics, and applies to equilibrium liquids at all temperatures, allowing comparison with MCT studies of equilibrium supercooled liquids. On the face of it, however, that comparison is not straightforward. For a given time correlation function, MCT evaluates the contributions to a generalized Langevin equation, while V-T theory evaluates the function's defining equation. The working measures of the atomic motion are quite different in the two theories. But the actual atomic motion is unique, and we can base a comparison of theories on their respective descriptions of that motion.
For normal and supercooled liquids, MD data for the MSD time evolution~\cite{KA1995} has a well established interpretation in terms of the atomic motion~\cite{Kob1999}. Here we apply the same interpretation to the SISF, as shown in figure~\ref{fig8}. There are three successive time intervals, characterized respectively by ballistic, ``cage-jump'' and diffusive motions. Upon supercooling the liquid, the primary change in figure~\ref{fig8} is the appearence of a plateau in the middle interval, where the atomic dynamics is slowed by cage motion. G{\"o}tze tells us MCT for density fluctuation dynamics was developed originally to deal with the cage effect (section 1 of ~\cite{Gotze1999}. The idea is that a particle remains trapped for a while in the cage of its neighbors. The process of exiting a cage is a jump, and cage-jump motion characterizes the middle interval. The time evolution of figure~\ref{fig8} is commonly observed, as for example in binary LJ and silica systems (figure 3 of~\cite{Kob1999}, complex hard sphere systems~\cite{Toku2008}, a one-component LJ system~\cite{Hoang2013}, and Al~\cite{JP2013}.
Figure~\ref{fig8} compares directly with the V-T time evolution in figure~\ref{fig3}. In the comparison we can see two theoretical advantages in working with vibrational motion in place of the ballistic motion: vibrational motion maintains agreement with MD through a much greater amount of damping, until $\tau_{D}$, and the vibrational contribution goes on to saturate near the level of zero vibrational correlation. On the other hand, these theoretical advantages can be lost in complicated systems, due to the increased complexity of the vibrational Hamiltonian. Finally in comparing figures~\ref{fig8} and \ref{fig3}, we note that the two plans are the same on $0 \leq t \leq \tau_{B}$; are the same on $t \geq \tau_{RW}$ when the diffusive motion is represented by the steady state transit random walk; and that cage-jump and vibration-transit motions have a logical correspondence on $\tau_{B} \leq t \leq \tau_{RW}$.
\begin{figure} [h]
\includegraphics [width=0.70\textwidth]{Figure8.eps}
\caption{The time evolution commonly observed in MD data for the MSD in normal and supercooled liquids, shown for the SISF of liquid Na. Dots are $F_{MD}^{s}(q,t)$ and short line to $\tau_{B}$ is the ballistic contribution.}
\label{fig8}
\end{figure}
The motional correspondence just mentioned is apparent in the literature. Cage breaking processes in binary LJ systems can be related to long term diffusion~\cite{dSW2008}, and can be modeled by a correlated random walk~\cite{dSW2009}. This description matches transits in monatomic liquids, where random walk correlations are minimal because all atoms are equivalent~\cite{G1}. It has been found that atoms in a binary LJ system involve the sensible displacement of a small group of particles, and that these jumps are diffusive at high $T$, subdiffusive at low $T$~\cite{CPC2016}. The transits observed in Na and Ar also occur in correlated groups, but are diffusive at all $T$~\cite{WCC2001}. Models which connect fast and slow degrees of freedom of viscous liquids have been discussed~\cite{Dyre2006}, and distinct fast and slow channels in a LJ fluid have been found~\cite{BNGBB2015}. The fast-slow connection in V-T theory is the three-interval evolution upon which the present work is based, from the initial pure vibrational motion (fast) to the ultimate pure transit motion (slow) (see also ~\cite{MSD}). Finally, an example of the formidable complications MCT is able to address is a study of flexible trimers rattling in the cages of their neighbors~\cite{BL2016}.
\section{Conclusions}
In this paper and the previous one~\cite{MSD}, V-T theory has demonstrated its analytical tractability and mathematical simplicity. In this paper, the theory with \emph{apriori} vibrational calibration, and with transit calibration from the MSD, accounts for the SISF to extreme accuracy at all $q$ and $t$. Two questions lead us to deeper insight into what all this means.
(a) Considering the massive $q$ dependence of $F_{MD}^{s}(q,t)$, figure~\ref{fig7}, how can we have a $q$-independent calibration? It is because we calibrate the motion, not the time correlation function. The atomic motion has no $q$ dependence. Vibrational motion is given by the vibrational Hamiltonian, and transit motion is given by parameters calibrated from MD (Table I). All $q$ dependence is in equation~(\ref{eq1}), and carries over to the vibrational time correlation functions and Debye-Waller factors, equations~(\ref{eq4})-(\ref{eq6}), and to the transit damping coefficients of equations~(\ref{eq19}) and (\ref{eq20}).
(b) What makes the theory uniformly accurate over all $q$ and $t$? With reference to figure~\ref{fig7}, the answer has two parts. (i) The vibrational motion is accurate because it is calculated from one of the liquid's own $3N$-dimensional random valleys, and the liquid measures only vibrational motion at $t\leq \tau_{D}$ and all $q$. (ii) The transit motion is accurate because the steady-state transit random walk is calibrated from the MSD measurement of $D$, and the liquid measures only random walk motion at $t\geq \tau_{RW}$ and all $q$. This leaves the crossover interval accurately calibrated at its end points at all $q$.
A common three interval time evolution plan, figure~\ref{fig8}, can be applied to the MSD and SISF, in MCT or V-T theory. The atomic motion for the two theories is either the same or logically equivalent on each interval (section VI).
We are currently applying the present theory to critical slowing down in supercooled liquid Na.
\acknowledgments{We are pleased to thank Brad Clements for helpful and encouraging discussions. This research is supported by the Department of Energy under Contract No. DE-AC52-06NA25396.}
|
\section{Introduction}
The grand unified theory (GUT)~\cite{gut}, in which quarks and leptons are treated in the same footing,
is an attractive idea and has been investigated in various aspects since its proposal.
The gauge group of the standard model, $SU(3)\times SU(2)\times U(1)$ is embedded into a GUT
group, so that the gauge symmetry of the GUT must be broken down to the
$SU(3)\times SU(2)\times U(1)$ by some mechanism. This is not understood well at the
present. One often assumes the Higgs mechanism to break the gauge symmetry. In fact, the prototype $SU(5)$
GUT~\cite{gut} introduces the Higgs scalar field belonging to the adjoint representation
under the $SU(5)$ by hand, and the scalar field is assumed to have the vacuum
expectation value in such a way that the $SU(5)$ breaks down to the gauge group of the standard model.
There is not any evidence, however, that it should be the Higgs mechanism, and thus
it is important to seek the mechanism of the gauge symmetry breaking.
The GUT in a framework of the higher dimensional gauge theory has also been
proposed and studied~\cite{orbifoldGUT, ghgut}, where the unified gauge symmetry is broken by the
boundary conditions with respect to the compactified extra dimensions~\cite{orbifoldGUT}.
Meanwhile, one
of the interesting points of the higher dimensional gauge theory
is that it can provide the mechanism of the gauge symmetry breaking which is different
from the Higgs mechanism, that is, the Hosotani mechanism~\cite{hosotani}.
In the Hosotani mechanism, component gauge fields for compactified directions, which behave like the
Higgs scalar fields at low energy and are closely related with the Wilson line phases, develop
the vacuum expectation values to induce the gauge symmetry breaking.
An advantage of the Hosotani mechanism is that one can compute the effective potential
for the Wilson line phases, with no need to introduce additional counterterms~\cite{hierarchy},
and the vacuum expectation values for the component gauge fields are
dynamically determined by minimizing it once one fixes matter content of the theory.
Then, one understands definite origin of the potential which induces the gauge symmetry breaking.
It should be noted that the periodic nature of the phases
enables us to study the global structure of the
effective potential even if we consider physics around the GUT and the Planck scale.
This is in sharp contrast to the usual effective potential in the quantum field theory, where the
potential is given by the polynomial of the scalar fields, so that one can study only the
local structure of the potential near the origin.
When the number of the order parameters in the effective potential increases, it becomes
difficult to minimize the effective potential analytically. In recent years, however, the environment
of numerical study is dramatically improved, which makes easier to find the global minimum
of the effective potential. In addition, the idea of GUTs with the Hosotani
mechanism is a very attractive from phenomenological and theoretical points of view as
discussed below, so that it should be investigated extensively.
When one tries to apply the Hosotani mechanism to break the $SU(5)$ gauge symmetries of
GUT models,
one immediately encounters a difficulty.
Namely, the zero modes of the component gauge fields belonging to the adjoint
representation under the gauge group that remains unbroken against the boundary conditions
tend to be projected out by the boundary conditions in models with chiral fermions.
Thus, in most GUT models with the Hosotani mechanism, the $SU(5)$ symmetry that contains
the standard model gauge symmetry is broken by the boundary condition~\cite{ghgut}.
We have shown, however, in the paper~\cite{kty} that the difficulty is overcome
by applying the so-called diagonal embedding method~\cite{diagonal}, which is invented in the
context of string theory, to the higher dimensional
gauge theory.
The method enables us to have GUT models with the Hosotani breaking of the $SU(5)$ gauge
symmetry without contradicting with the chiral fermions and
study the gauge symmetry breaking of the GUT.
What is striking is that the effective potential obtained by the diagonal embedding method
has basically
the same form with the one obtained for the $S^1$ compactification.
In other words, the diagonal embedding
method can be seen as a way to introduce chiral fermions in $S^1$ compactification.
In this paper, we study the gauge symmetry breaking patterns of the models of the
five-dimensional $SU(5)$ GUT
compactified on an $S^1/\Z2$ by utilizing the diagonal embedding method.
We refer to the models as the $SU(5)$ grand gauge-Higgs unification.
The $SU(5)$ gauge symmetry is broken by the Hosotani mechanism to one of its regular subgroups
according to the matter content of the theory.
We present examples of the matter content that realize the vacuum configuration
with each regular subgroup of the $SU(5)$.
We emphasize that matter contents, for which the
$SU(3)\times SU(2)\times U(1)$ gauge symmetry is realized on the global minimum of the theory,
are found. One of the matter contents that realize the gauge symmetry of the standard model at the vacuum
does not include the periodic fermions belonging to the adjoint representation,
contrary to the expectation~\cite{congru}.
It has been known that the finite temperature field theory provides us a useful tool to study
the phase transition, in the early Universe, {\it etc}~\cite{dj}. We also study the finite temperature phase transition
for the matter contents that realize the gauge symmetry of the standard model at zero temperature
in our $SU(5)$ models.
We obtain the critical temperature and find the order of the phase transition. In the limit of
high temperature, the
fermions do not contribute to the effective potential, while the bosons do to yield
the symmetry restoration of the $SU(5)$ at a certain temperature (if light scalar fields are absent).
We apply the analysis of the finite temperature to the supersymmetric (SUSY) version of
the $SU(5)$ grand gauge-Higgs unification.
This scenario with the SUSY breaking scale being $O(1)$~TeV is phenomenologically
very interesting in itself.
For example, it generally predicts the existence of the light adjoint chiral supermultiplets
with masses of the same scale and we may find a hint of the breaking of the GUT gauge symmetry
at the $\mbox{TeV}$ scale experiments~\cite{gGHU-DTS, gGHU-pheno}.
In addition, if a specific vacuum which is realized as a local minimum in Ref.~\cite{kty} is
assumed, the so-called doublet-triplet splitting problem can be naturally solved~\cite{gGHU-DTS}.
In the SUSY version, one may also expect different behaviors of phase transitions
from the non-SUSY cases because scalar fields are introduced in contrast to the latter.
We find models of the SUSY $SU(5)$ grand gauge-Higgs unification in
which the desired vacuum, where the doublet-triplet splitting problem can be naturally
solved, is realized for the wide range of the temperature.
The models may provide us an interesting possibility for the vacuum selection in the early
Universe. Namely, the desired vacuum is dynamically selected
in the epoch with very high temperature which may be exist
before the inflation
and continues to stay there up to the present.
This paper is organized as follows. In section $2$, after the brief introduction of the diagonal
embedding method, we study the gauge symmetry breaking patterns
of the $SU(5)$ grand gauge-Higgs unification by the analyses of the one-loop effective
potential. In section $3$, the finite temperature phase transition is studied for some models introduced
in the section $2$. We also consider the supersymmetric version of the scenario and study
the behavior at finite temperature and address the vacuum selection in the early Universe.
The final section is devoted to conclusions and discussions.
We present some formulae necessary in the discussion of the section $3$ in the appendix.
\section{Gauge symmetry breaking of an $SU(5)$ grand gauge-Higgs unification}
\subsection{Diagonal embedding method}
Let us first review quickly the diagonal embedding method
which makes it possible for the Hosotani mechanism to break the $SU(5)$ gauge symmetry
consistent with chiral fermions. The simplest setting is a five-dimensional $SU(5)$ model on
an orbifold $S^1/\Z2$ compactification with its radius being of the GUT scale.
The detailed discussions are given in our previous paper~\cite{kty}, and readers
who are familiar with the method can go to the next subsection.
In the simplest setting of the method, we prepare two copies of the gauge symmetry
which are exchanged by a discrete symmetry $\Z2$. We consider the theory
with $SU(5)_1\times SU(5)_2\times \Z2$ symmetry, where the five-dimensional gauge field
for the $SU(5)_i~(i=1,2)$ is denoted by $A_M^{(i)}~(i=1,2)$. The $A_M^{(1)}$ and $A_M^{(2)}$
are related with each other by the $\Z2$. Here, $M=(\mu, y)=(0\mbox -3, 5)$ is a
five-dimensional Lorentzian index, where the coordinate of the extra dimension is
denoted by $y$ and the circumference of the $S^1$ is
$L=2\pi R$.
We impose the boundary conditions on the gauge fields at the two fixed
points, $y_0=0, y_{\pi}=\pi R$ of the $S^1/\Z2$ as
\begin{equation}
A_{\mu}^{(1)}(y_i - y)=A_{\mu}^{(2)}(y_i + y),\qquad
A_y^{(1)}(y_i -y) =-A_y^{(2)}(y_i + y),
\label{shiki1}
\end{equation}
where we have used the notation $y_i (i=0, \pi)$.
We define the eigenstate under the operation of the $\Z2$ by
$X^{(\pm)}\equiv (X^{(1)}\pm X^{(2)})/\sqrt{2}$. Then, we see that
$A_{\mu}^{(+)}$ and $A_y^{(-)}$ satisfy the Neumann boundary condition
at the both fixed points and thus have the zero modes. It implies that
the $SU(5)_1\times SU(5)_2$ breaks down to their diagonal part $SU(5)_{\rm diag}$
whose gauge field in the four-dimensional effective theory
is $A_{\mu}^{(+)}$. One says that our $SU(5)$ GUT symmetry
is embedded in the diagonal part. At the same time, one obtains the zero mode of the adjoint scalar field
$A_y^{(-)}$ under the $SU(5)_{\rm diag}$, and as shown in the previous paper,
it composes the Wilson line phase,
\begin{equation}
W={\cal P}~{\rm exp}\biggl(
ig\int_0^{2\pi R}dy~~\frac1{\sqrt{2}}A_y^{(-)a}(T^a_1 -T^a_2)
\biggr),
\label{shiki2}
\end{equation}
where $\cal P$ stands for the path-ordered integral, $g$ is the common gauge coupling
constant, $T_i^a(i=1,2)$ is the generator of the $SU(5)_i (i=1,2)$ symmetry, and $a$ is an $SU(5)$
adjoint index. If we consider the fundamental representation ${\bf R_1}={\bf 5}, {\bf R_2}={\bf 1}$
for concreteness, we can parametrize the vacuum expectation value of $A_y^{(-)}$ as
\begin{equation}
\frac1{\sqrt{2}}gL\vev{A_y^{(-)a}T^a_{\bf 5}}
=2\pi~{\rm diag.} (\alpha_1, \alpha_2, \alpha_3, \alpha_4, \alpha_5)=2\theta^aT^a
\quad \mbox{with}\quad
\sum_{j=1}^5\alpha_j=0,
\label{shiki3}
\end{equation}
where we have used the remaining $SU(5)$ degrees of freedom to diagonalize the $A_y^{(-)a}T_{\bf 5}^a$.
Thus, the Wilson line phase is written as
\begin{equation}
W={\rm exp}~\biggl(2\pi i ~{\rm diag.}\left(\alpha_1, \alpha_2, \alpha_3, \alpha_4, \alpha_5\right)\bigg).
\label{shiki4}
\end{equation}
We see that the order parameter $W$ is invariant under the shift $\alpha_j\to\alpha_j+1$.
Let us next introduce the fermion fields, $\Psi^{(1)}({\bf R}, {\bf 1})$ and
$\Psi^{(2)}({\bf 1}, {\bf R})$, where
${\bf R}$ denotes a representation under the $SU(5)$ and are the $\Z2$ partner each other.
We do not consider matter fields that are non singlet under the both gauge groups,
$SU(5)_1$ and $SU(5)_2$, for simplicity.
We impose the boundary condition,
\begin{equation}
\Psi^{(1)}(y_i -y) =\eta_{\Psi}^i\gamma^5\Psi^{(2)}(y_i +y),
\label{shiki5}
\end{equation}
where the parameter $\eta_{\Psi}^{i=0,\pi}$, which is associated with each fermion, takes
$+1$ or $-1$, and the product $\eta_{\Psi}^0\eta_{\Psi}^{\pi}$
gives the periodicity for the $S^1$ direction. The eigenstate $\Psi^{(\pm)}$ of the $\Z2$
has the zero mode for $\eta_{\Psi}^0\eta_{\Psi}^{\pi}=1$, while it does not
for $\eta_{\Psi}^0\eta_{\Psi}^{\pi}=-1$.
Note that the zero modes are vector-like as $\Psi_L^{(\pm)}$ obeys the same boundary
conditions as $\Psi_R^{(\mp)}$. The chiral
fermions, for example the quarks and leptons in the standard model, can
be put on the boundaries, though we do not consider their effects
in this paper for simplicity.
As studied in the paper~\cite{kty}, the Kaluza-Klein (KK) mass spectrum for the $\Psi^{(\pm)}$
with $\eta_{\Psi}^0\eta_{\Psi}^{\pi}=1$ is given by
\begin{equation}
m_{\rm KK} ^{(n)}R=n+\frac{gR}{\sqrt{2}}\vev{A_y^{(-)a}} T^a_{\bf R}
=n+\frac{\theta^a}{\pi}T^a_{\bf R},\quad n\in {\mathbb Z},
\label{shiki6}
\end{equation}
and that for the $\Psi^{(\pm)}$ with $\eta_{\Psi}^0\eta_{\Psi}^{\pi}=-1$ by the same form
in Eq.~(\ref{shiki6}) with replacement $n\to n+1/2$, which
implies that the KK mass spectrum is basically the same form with the one obtained for
the $S^1$ compactification. Thus, each contribution from the $\Psi^{(+)}$ and $\Psi^{(-)}$ on the
orbifold $S^1/\Z2$ forms the same contribution from a fermion field on the $S^1$.
Then, once we fix the matter content whose representation is $\bf R$ under the $SU(5)$, one can
immediately write down the contribution from the fermion to the effective potential by the help of the
knowledge of the $S^1$ compactification. And one can study the breaking of the $SU(5)$
gauge symmetry through the Hosotani mechanism.
In non-SUSY models, we do not introduce the scalar fields in this paper,
as they are generally expected to be superheavy due to the quantum corrections.
In cases where the corrections are canceled to realize light scalars $\Phi^{(1)}({\bf R}, {\bf 1})$ and
$\Phi^{(2)}({\bf 1}, {\bf R})$, by fine-tuning or the SUSY, we should impose the boundary condition,
\begin{equation}
\Phi^{(1)}(y_i -y) =\eta_{\Phi}^i\Phi^{(2)}(y_i +y),
\label{shiki5-2}
\end{equation}
where the parameter $\eta_{\Phi}^{i=0,\pi}$ takes $+1$ or $-1$, and the KK mass spectrum for the
$\Phi^{(\pm)}$ is the same as the one in Eq.~(\ref{shiki6}).
\subsection{Matter content and gauge symmetry breaking pattern}
Let us study the gauge symmetry breaking patterns of the $SU(5)$ grand gauge-Higgs
unification. We introduce fermions whose representations under the $SU(5)$ gauge group
are ${\bf 24, 5, 10}$ and ${\bf 15}$ \footnote{
We note that the fermions in the conjugate representation, {\it e.g.} ${\bf \bar 5}$,
give the same contributions as those in the corresponding representation, since the
five-dimensional models are vector-like.}
and whose periodicities
are $\eta_{\Psi}^0\eta_{\Psi}^{\pi}=+1~(-1)$ corresponding to the
(anti) periodic boundary condition. Following the standard prescription~\cite{hosotani} to
calculate the one-loop effective potential in the background in Eq.~(\ref{shiki3}), we
obtain the contribution of each degree of freedom in the aforementioned four representations to the
potential as, up to the overall sign $(-1)^{f+1}$,
\begin{eqnarray}
V_{24}(\alpha_i, z, \delta, f)&=&
C\sum_{i\neq j=1}^5 F\Bigl(\alpha_i - \alpha_j +\frac{\delta}{2},z,f\Bigr),
\label{shiki7}\\
V_{5}(\alpha_i, z, \delta, f)&=&
C\sum_{i=1}^5F\Bigl(\alpha_i+\frac{\delta}{2},z,f\Bigr),
\label{shiki8}\\
V_{10}(\alpha_i, z, \delta, f)&=&
C\sum_{1\leq i < j\leq 5}
F\Bigl(\alpha_i + \alpha_j +\frac{\delta}{2},z,f\Bigr),
\label{shiki9}\\
V_{15}(\alpha_i, z, \delta, f)&=&
C\sum_{1\leq i \leq j\leq 5}~
F\Bigl(\alpha_i + \alpha_j +\frac{\delta}{2},z,f\Bigr),
\label{shiki10}
\end{eqnarray}
where we accommodate the temperature $T$ (normalized by $L$ as $z=LT$) dependence for later
convenience, which should be set to zero in this section.
Here we have defined the function,
\begin{equation}
F(x,z=0,f)\equiv \sum_{w=1}^{\infty}\frac{1}{w^5}\cos(2\pi w x),
\label{shiki11}
\end{equation}
which represent{s} the contribution of each KK tower,
and the overall constant $C$, which we shall ignore for the numerical analyses, is
\begin{equation}
C\equiv \frac{3}{4\pi^2}\frac{1}{L^5}.
\label{shiki12}
\end{equation}
The parameter $f$ stands for the fermion number and takes $f=1~(0)$ for fermions (bosons), and
the fields satisfying the (anti) periodic boundary condition take $\delta=0~(1)$.
We incorporate the suppressed overall sign $(-1)^{f+1}$ into the numbers of
the degrees of freedom.
Then, the total effective potential is given by
\begin{eqnarray}
V_{\rm eff}(\alpha_i,z)&=&
N_gV_{24}(\alpha_i, z, 0, 0)+4N_{24}^{(+)}V_{24}(\alpha_i, z, 0, 1)
+4N_{24}^{(-)}V_{24}(\alpha_i, z, 1, 1)\nonumber\\
&&
+4N_{5}^{(+)}V_{5}(\alpha_i, z, 0, 1)
+4N_{5}^{(-)}V_{5}(\alpha_i,, z, 1, 1)
+4N_{10}^{(+)}V_{10}(\alpha_i, z, 0, 1)\nonumber\\
&&
+4N_{10}^{(-)}V_{10}(\alpha_i, z, 1, 1)
+4N_{15}^{(+)}V_{15}(\alpha_i, z, 0, 1)
+4N_{15}^{(-)}V_{15}(\alpha_i, z, 1, 1),
\label{shiki13}
\end{eqnarray}
where $N_g$ and the coefficients $4$ are the on-shell degrees of freedom for the gauge boson
and fermion in the five dimensions, respectively.
In the non-SUSY case at zero temperature, $N_g=-3$ and $z=0$.
Here we have denoted the flavor numbers specifying the matter content as
\begin{equation}
(N_{24}^{(+)}, N_{24}^{(-)}, N_{5}^{(+)}, N_{5}^{(-)}, N_{10}^{(+)}, N_{10}^{(-)}, N_{15}^{(+)}, N_{15}^{(-)}).
\label{shiki14}
\end{equation}
The $(+), (-)$ denotes the periodicity of the field which corresponds to $\delta=0, 1$, respectively.
The subscript number stands for the representation under the $SU(5)$ gauge group.
\begin{table}[]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}\hline
Case & $N_{24}^{(+)}$ &$N_{24}^{(-)}$ &$N_5^{(+)}$ &$N_5^{(-)}$ &$N_{10}^{(+)}$ &$N_{10}^{(-)}$ &$N_{15}^{(+)}$ &$N_{15}^{(-)}$ \\\hline\hline
1 & 1 & 0 & 0 & 0& 1& 0& 0& 1 \\\hline
2 & 0 & 0 & 0 & 0& 7& 0& 0& 5 \\\hline
3 & 1 & 0 & 0 & 0& 0& 0& 1& 1 \\\hline
4 & 1 & 0 & 1 & 0& 0& 1& 1& 0 \\\hline \hline
5 & 0 & 0 &0 & 0& 0 & 0& 0& 0 \\\hline
6 & 0 & 0 & 1 & 1& 1& 0& 0& 3 \\\hline\hline
7 & 1 & 0 & 1 & 0& 0& 0& 0& 1 \\\hline
8 & 1 & 0 & 1 & 0& 0& 0& 1& 0 \\\hline\hline
9 & 1 & 0 & 0 & 1& 0& 0& 1& 1 \\\hline\hline
10 & 1 & 0 & 2 & 0& 0& 0& 0& 0 \\\hline\hline
11 & 1 & 0 & 0 & 1& 1& 1& 0& 0 \\\hline \hline
12 & 1 & 0 & 0 & 0& 0& 0& 0& 0 \\\hline
\end{tabular}
\caption{Matter contents in the $SU(5)$ grand gauge-Higgs unification.}
\end{center}
\end{table}
\begin{table}[]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}\hline
Case & $\alpha_1$ &$\alpha_2$ &$\alpha_3$ &$\alpha_4$ &$\alpha_5$ &Gauge symmetry\\\hline\hline
1 & 0.5 & 0.5 & 0 & 0& 0 &\multirow{4}{*}{$SU(3)\times SU(2)\times U(1)$} \\\cline{1-6}
2 & 0.5 & 0.5 & 0 & 0& 0 & \\\cline{1-6}
3 & 0.1 & 0.1 & 0.6 & 0.6 & 0.6 &\\\cline{1-6}
4 & 0.223158 &0.223158 &0.223158 &0.665263 & 0.665263 &\\\hline \hline
5& $k/5$& $k/5$ & $k/5$ & $k/5$& $k/5$ &\multirow{2}{*}{$SU(5)$}\\\cline{1-6}
6 & 0 & 0 & 0 & 0 & 0 &\\\hline \hline
7& 0.5 &0.5 & 0.5 &0.5 &0 & \multirow{2}{*}{$SU(4)\times U(1)$}\\\cline{1-6}
8& 0.287091& 0.287091 & 0.287091& {0.287091} & 0.851637 &\\\hline \hline
9& 0.139502 & 0.139502 & 0.139502 &0.679867 & 0.901628 &$ SU(3)\times U(1)^2$\\\hline\hline
10& 0.465989 &0.465989 & 0.534011&0.534011&0&$ SU(2)^2\times U(1)^2$\\\hline \hline
11& 0.154787 & 0.25036& 0.700439 & 0.947207& 0.947207 & $SU(2)\times U(1)^3$\\\hline \hline
12& 0.2& 0.4 & 0.6 & 0.8 &0 &$U(1)^4$\\\hline
\end{tabular}
\caption{Vacuum expectation values and gauge symmetry breaking patterns for the cases
given in Table $1$. The parameter $k$ takes $0, 1, 2, 3$ and $4$.
Here, we have shifted $\alpha_j$ into an interval $[0,1)$, instead of imposing the traceless
condition which can be recovered by using the periodicity $\alpha_j\sim\alpha_j+1$.}
\end{center}
\end{table}
The parameter $\alpha_j$ is related with the Wilson line phase in Eq.~(\ref{shiki4}),
so that the physical region is compact,
reflecting its phase nature. Hence, one can
study the global structure of the effective potential in Eq.~(\ref{shiki13})
even if we consider physics around the Planck scale.
This is very contrast to the usual effective potential which is given by the
polynomial of the order parameters of the scalar fields and the analyses of
the potential for the huge scale region around the Planck scale is unreliable.
This is a great advantage of our approach to the breaking of the GUT gauge symmetry.
We show the matter contents we have chosen in the Table $1$. For each
set of the matter content given in Table $1$, we minimize the potential numerically to obtain
the vacuum expectation values $\alpha_j$. Then, we understand how the $SU(5)$ is
broken down to its regular subgroup via the Hosotani mechanism,
by evaluating the Wilson line phase in Eq.~(\ref{shiki4}).
In particular, we are very much interested in the matter content that can realize
the $SU(3)\times SU(2)\times U(1)$ gauge symmetry of the standard model
on the global minimum of the effective potential.
We summarize our results in the Table $2$ in which the vacuum expectation values and
the realized gauge symmetry on the global minimum are specified for each set of the matter content.
Let us note that the example of the matter
content studied in~\cite{kty}, the case $6$, realizes the standard model gauge symmetry on a
local minimum of the effective potential. It may be
worth to note here that the case $2$ realizes that the $SU(5)$ breaks down to
the $SU(3)\times SU(2)\times U(1)$ even though it {does}
not contain the fermions belonging to the ${\bf 24}$
representation, contrary to the understanding until now~\cite{congru}. We
find examples of matter contents,
for which each regular subgroup
of $SU(5)$ is realized on the global minimum of the effective potential.
\section{Grand gauge-Higgs unification at finite temperature}
In this section, we study the finite temperature phase transition for the models with the standard
model gauge group on the global minimum studied in the previous section and for some SUSY models.
At high temperature, the global minimum of the potential and vacuum configuration of
the Wilson line phase would be altered due to existence of temperature dependent contribution to
the effective potential.
Thus, if the Universe has an epoch with very high temperature,
the finite temperature phase transitions may have implications for
grand unified models.
We follow the standard prescription, say, the Euclidean time formulation, to
obtain the effective potential at finite temperature~\cite{dj}. The Euclidean time $\tau$ direction is,
then, compactified on
$S^1_{\tau}$ whose circumstance is $1/T$, where $T$ is the temperature.
We shall ignore the Wilson line
degrees of freedom arising from the $S^1_{\tau}$, since no nontrivial vacuum configuration of
the Wilson line phase is expected; the detailed discussion on the
phase is given in Ref.~\cite{sakatake}.
\subsection{Non supersymmetric models at finite temperature}
If we turn on the temperature $T$, the effective potential in one-loop approximation
from the fermions belonging to the ${\bf 24}, {\bf 5}, {\bf 10}$ and ${\bf 15}$ representations
are given by Eqs.~(\ref{shiki7})-(\ref{shiki10}) respectively, with the function for
nonzero $z=LT$ which is treated as a free parameter in the theory,
\begin{equation}
F(x, z, f)\equiv \sum_{\omega=1}^{\infty}{\frac1{\omega^5}}\cos(2\pi \omega x)
+2 {z^5}
\sum_{\omega=1}^{\infty}\sum_{l=1}^{\infty}
\frac{(-1)^{fl}}{[(\omega z)^2 +l^2]^{\frac52}}\cos(2\pi \omega x),
\label{shiki15}
\end{equation}
as derived in Appendix~\ref{Sec:EffPotAtFT}.
The first (second) term in the right-hand side in Eq.~(\ref{shiki15})
is the zero (finite) temperature contribution.
Let us note that the (anti) periodicity of the bosons (fermions) for the Euclidean time direction
manifestly appears as $(-1)^{fl}$ in the summation of the second term in
Eq.~(\ref{shiki15}), which originally takes the form of $\cos(\pi fl)$
in Eq.~(\ref{Eq:EffPotAtFT}). The
total effective potential is given by Eq.~(\ref{shiki13}), again with $N_g=-3$ but nonzero $z$.
We are interested in the behavior of the models at finite temperature
for the cases, $1,2,3$ and $4$ where the $SU(3)\times SU(2)\times U(1)$ is realized
on the global minimum
at zero temperature. We numerically obtain the critical temperature $z_c$ for the four
cases.\footnote{For the numerical analyses, we set the upper bound on the summation
with respect to $\omega, l$ in Eq.~(\ref{shiki15}) to $100$. }
Let us present the results for the cases, $1, 2, 3$ and $4$ in Tables $3, 4, 5$ and $6$, respectively.
We write there the vacuum expectation values $\alpha_j$, the depths and
the realized gauge symmetries for the global minimum and for some energetically lower local minima
at two temperatures near the critical temperature.
The critical temperatures for the cases $1,2$ and $3$ are commonly obtained as $z_c=0.720975$
by solving numerically the concrete form of an equation shown in Eq.~(\ref{Tc}) which is resulted by
the differences among the contributions in Eqs.~(\ref{shiki7})-(\ref{shiki10})
with the vacuum expectation values $\alpha_j$ corresponding to $G321, G41$ and $G5$.
At this temperature, the three different vacua shown in the tables are
degenerate with each other, and thus the phase transitions are of the
first order.
The detailed discussion is given in Appendix~\ref{appendct}.
As for the case $4$, the phase transition occurs at $z=0.585788$ and is of the first order.
It can be shown that there is another phase transition at $z=0.830305$ above which
the global minimum resides on $(0.4,0.4,0.4,0.4,0.4)$.
We confirm the expectation that the $SU(5)$ gauge symmetry is restored in the
limit of high temperature because in the limit, the
fermions are decoupled from the system, so that the gauge bosons give dominant
contributions to the effective potential~\cite{sakatake2}.
\begin{table}[t]
\begin{center}
\begin{tabular}{|c|c||c|c|c|c|c|c|}\hline
$z=LT$&Depth &$\alpha_1$ &$\alpha_2$ &$\alpha_3$ &$\alpha_4$ &$\alpha_5$ &Gauge symmetry\\\hline\hline
& $-19.2997$ &0.5 & 0.5 & 0 & 0& 0 &$G321$ \\\cline{2-8}
0.7209& $-19.2971$ &0.5 & 0.5 & 0.5 & 0.5& 0 &$G41$ \\\cline{2-8}
& $-19.2917$ &0 & 0 & 0 & 0 & 0 &$G5$\\\hline \hline
& $-19.3007$ & 0& 0 & 0 & 0& 0 &$G5$\\\cline{2-8}
0.7210& $-19.2989$ & 0.5 & 0.5 & 0.5 & 0.5 & 0 &$G41$\\\cline{2-8}
& $-19.2980$ &0.5 &0.5 & 0 &0 &0 & $G321$\\\hline
\end{tabular}
\caption{Global and two local minima for the case $1$ near the
critical temperature.
Depths of the minima, which is
values of the potential in Eq.~(\ref{shiki13}) normalized by overall factor $C$
in Eq.~(\ref{shiki12}), and values of $\alpha_i$ are shown.
$G321\equiv SU(3)\times SU(2)\times U(1), G41\equiv SU(4)\times U(1)$ and $G5\equiv SU(5)$.}
\end{center}
\end{table}
%
%
\begin{table}[t]
\begin{center}
\begin{tabular}{|c|c||c|c|c|c|c|c|}\hline
$z=LT$& Depth &$\alpha_1$ &$\alpha_2$ &$\alpha_3$ &$\alpha_4$ &$\alpha_5$ &Gauge symmetry\\\hline\hline
& $-84.4617$ &0.5 & 0.5 & 0 & 0& 0 &$G321$ \\\cline{2-8}
0.7209 & $-84.4590$ &0.5 & 0.5 & 0.5 & 0.5& 0 &$G41$ \\\cline{2-8}
& $-84.4537$ &0 & 0 & 0 & 0 & 0 &$G5$\\\hline \hline
& $-84.4595$ & 0& 0 & 0 & 0& 0 &$G5$\\\cline{2-8}
0.7210 & $-84.4577$ & 0.5 & 0.5 & 0.5 & 0.5 & 0 &$G41$\\\cline{2-8}
& $-84.4568$ & 0.5 &0.5 & 0 &0 &0 & $G321$\\\hline
\end{tabular}
\caption{Global and two local minima for the case $2$ near the
critical temperature. Depths of the minima and values of $\alpha_i$ are shown.
$G321\equiv SU(3)\times SU(2)\times U(1), G41\equiv SU(4)\times U(1)$ and $G5\equiv SU(5)$.}
\end{center}
\end{table}
\begin{table}[t]
\begin{center}
\begin{tabular}{|c|c||c|c|c|c|c|c|}\hline
$z=LT$& Depth &$\alpha_1$ &$\alpha_2$ &$\alpha_3$ &$\alpha_4$ &$\alpha_5$ &Gauge symmetry\\\hline\hline
& $-4.11202$ &0.1 & 0.1 & 0.6 & 0.6& 0.6 &$G321$ \\\cline{2-8}
0.7209 & $-4.10934$ &0.1 & 0.1 & 0.1 & 0.1& 0.6 &$G41$ \\\cline{2-8}
& $-4.10398$ &0.6 & 0.6 & 0.6 & 0.6 & 0.6 &$G5$\\\hline \hline
& $-4.11335$ & 0.6& 0.6 & 0.6 & 0.6& 0.6 &$G5$\\\cline{2-8}
0.7210 & $-4.11159$ & 0.1 & 0.1 & 0.1 & 0.1 & 0.6 &$G41$\\\cline{2-8}
& $-4.11070$ & 0.1 &0.1 & 0.6 &0.6 &0.6 & $G321$\\\hline
\end{tabular}
\caption{Global and two local minima for the case $3$ near the
critical temperature. Depths of the minima and values of $\alpha_i$ are shown.
$G321\equiv SU(3)\times SU(2)\times U(1), G41\equiv SU(4)\times U(1)$ and $G5\equiv SU(5)$.}
\end{center}
\end{table}
\begin{table}[t]
\begin{center}
\begin{tabular}{|c|c||c|c|c|c|c|c|}\hline
$z=LT$& Depth &$\alpha_1$ &$\alpha_2$ &$\alpha_3$ &$\alpha_4$ &$\alpha_5$ &Gauge symmetry\\\hline\hline
\multirow{2}{*}{0.5857} & $-19.2778$ &0.224824 & 0.224824 & 0.224824 & 0.662764 & 0.662764 &$G321$ \\\cline{2-8}
& $-19.2759$ &0.311785 & 0.311785 & 0.311785 & 0.311785 & 0.752860 &$G41$ \\\hline \hline
\multirow{2}{*}{0.5858} & $-19.2773$ &0.311786 & 0.311786 & 0.311786 & 0.311786 & 0.752857 &$G41$\\\cline{2-8}
& $-19.2770$ &0.224826 & 0.224826 & 0.224826 & 0.662761 & 0.662761 &$G321$\\\hline
\end{tabular}
\caption{Global and the energetically lowest local minimum for
the case $4$ near the critical temperature. Depths of the minima and values of $\alpha_i$ are shown.
$G321\equiv SU(3)\times SU(2)\times U(1)$ and $G41\equiv SU(4)\times U(1)$.
}
\end{center}
\end{table}
\subsection{Supersymmetric models at finite temperature}
\label{Sec:SUSY-FT}
As stated in the introduction, the SUSY version of our $SU(5)$ model with the SUSY breaking scale
$M_{\rm SB}$ being around the $\mbox{TeV}$ scale is very attractive, especially on the desired vacuum
shown below in Eq.~(\ref{shiki18}), since the doublet-triplet splitting
problem can be naturally solved there~\cite{gGHU-DTS}.
In Refs.~\cite{gGHU-DTS,gGHU-pheno}, it is required that the desired vacuum is just one of
minima, which is sufficient at zero temperature since the transition rate is extremely
suppressed even if there are deeper vacua~\cite{tunnel}. The stability
arises because the loop induced effective potential is suppressed by the SUSY
breaking scale which is now much smaller than the compactification
scale.\footnote{Hence, the mass scale of light adjoint chiral supermultiplets from zero-modes of $A_y$
and their superpartners is typically $O(M_{\rm SB})$ in this scenario.}
The smallness of the effective potential, at the same time, means that the stability of the vacuum
can be easily affected by the finite temperature effects at high temperature $T\gg M_{\rm SB}$.
Thus, it is important to study the finite temperature phase transition of the SUSY cases.
Because of the existence of the scalar fields,
it is expected that the behavior of the effective potential should be modified from the non-SUSY
version in which the only bosonic field is the gauge one.
We are, in particular, interested in the matter content that realizes the
desired vacuum in Eq.~(\ref{shiki18}) as the global minimum for the wide range of the temperature.
We set the scale of the extra dimension $1/R$ around the GUT scale $M_{\rm GUT}$ as before.
Then, $M_{\rm SB}$ is negligibly small compared with it.
At high temperature $T\gg M_{\rm SB}$, we may neglect the effects of the
zero-temperature SUSY breaking, and the SUSY breaking entirely comes from the finite
temperature effect, say, the
difference of the boundary conditions of bosons and fermions for the Euclidean time direction.
With this approximation, the zero temperature effective potential becomes vanishing.
The diagonal embedding method can be applied to supersymmetric models as well if we replace
all the fields by the corresponding superfields. We again
follow the standard prescription to calculate the finite temperature effective potential in
one-loop approximation in the background in Eq.~(\ref{shiki3}).
Since we neglect the effects of the zero-temperature SUSY breaking,
the contribution from the vectormultiplet is common with that from the adjoint hypermultiplet.
The contributions from each degree of freedom of supermultiplets of ${\bf 24}, {\bf 5}, {\bf 10}$
and ${\bf 15}$ representations under the $SU(5)$
to the one-loop effective potential are given again by
Eqs.~(\ref{shiki7})-(\ref{shiki10}) respectively, but with replacing the function $F(x,z,f)$ by
the SUSY version $F_S(x,z)$ defined as
\begin{equation}
F_S(x, z)\equiv -F(x, z,0)+F(x, z,1)=-2z^5\sum_{\omega=1}^{\infty}
\sum_{l=1}^{\infty}\frac{\bigl(1-(-1)^l\bigr)}{\bigl[(\omega z)^2+l^2\bigr]^{\frac52}}\cos(2\pi \omega x).
\label{shiki16}
\end{equation}
The last variable of the contributions
in Eqs.~(\ref{shiki7})-(\ref{shiki10}), the fermion number $f$, is dummy in the SUSY cases.
Then, the total effective potential is given by Eq.~(\ref{shiki13}) with $N_g=4$.
Let us present an example of the matter content,
for which the $SU(3)\times SU(2)\times U(1)$ gauge symmetry
is realized on the global minimum of the effective potential at finite temperature,
\begin{eqnarray}
(N_{24}^{(+)}, N_{24}^{(-)}, N_{5}^{(+)}, N_{5}^{(-)}, N_{10}^{(+)}, N_{10}^{(-)}, N_{15}^{(+)}, N_{15}^{(-)})
&=&(0, 1, 0, 0, 0, 1, 0, 0).
\label{shiki17}
\end{eqnarray}
We numerically find that for the above case, the
phase transition occurs at $z_c\simeq 0.382008$, below which
the gauge symmetry of the standard model is realized on the global minimum whose vacuum
expectation values are given by
\begin{equation}
(\alpha_1, \alpha_2, \alpha_3, \alpha_4, \alpha_5)=(0.5, 0.5, 0, 0, 0).
\label{shiki18}
\end{equation}
As mentioned in the beginning of this subsection, we call it the desired vacuum. It may be
worth to repeat that the vacuum is desirable
in the sense that it provides us an interesting solution to the doublet-triplet
splitting problem in certain SUSY models as studied in Ref.~\cite{gGHU-DTS}.
Note that in the present case, the desired vacuum is obtained as
the global minimum of the potential, not as its local minimum.
This model has the same nature of the finite temperature phase transition as
the models without the supersymmetry. If the temperature increases further beyond
the critical temperature, the gauge symmetry tends to be restored to
the $SU(5)$ gauge symmetry for which the vacuum expectation values are given by
\begin{equation}
(\alpha_1, \alpha_2, \alpha_3, \alpha_4, \alpha_5)=(0.2, 0.2, 0.2, 0.2, 0.2).
\label{shiki19}
\end{equation}
We also confirm the result in Eq.~(\ref{shiki19}) by minimizing numerically the
effective potential consist of the bosonic degrees of freedom alone
in the matter content in Eq.~(\ref{shiki17}) for the high temperature limit~\cite{sakatake2}. We
summarize the results in Table {7} whose contents are the same as the ones in Tables {3,4,5 and 6}.
\begin{table}[t]
\begin{center}
\begin{tabular}{|c|c||c|c|c|c|c|c|}\hline
$z=LT$& Depth &$\alpha_1$ &$\alpha_2$ &$\alpha_3$ &$\alpha_4$ &$\alpha_5$ &Gauge symmetry\\\hline\hline
&{$-2.77733$}&0.5 & 0.5 & 0 & 0& 0 &$G321$ \\\cline{2-8}
0.3820 &{$-2.77731$}&0.2 & 0.2 & 0.2 & 0.2& 0.2 &$G5$ \\\cline{2-8}
&{$-2.32086$}&0.5 & 0.5 & 0.5 & 0.5 & 0 &$G41$\\\hline \hline
&{$-2.77974$}&{0.2}&{0.2}&{0.2}&{0.2}&{0.2}&$G5$\\\cline{2-8}
0.3821 &{$-2.77948$}& 0.5 & 0.5 & 0 & 0 & 0 &$G321$\\\cline{2-8}
&{$-2.32254$}& 0.5 &0.5 & 0.5 &0.5 &0 & $G41$\\\hline
\end{tabular}
\caption{Global and two local minima for the matter content in Eq.~(\ref{shiki17}).
Depths of the minima and values of $\alpha_i$ are shown.
$G321\equiv SU(3)\times SU(2)\times U(1),
G41\equiv SU(4)\times U(1), G5\equiv SU(5)$.}
\end{center}
\end{table}
Let us next consider another model whose matter content is given by
\begin{eqnarray}
(N_{24}^{(+)}, N_{24}^{(-)}, N_{5}^{(+)}, N_{5}^{(-)}, N_{10}^{(+)}, N_{10}^{(-)}, N_{15}^{(+)}, N_{15}^{(-)})
&=&(0, 2, 1, 0, 0, 0, 0, 0).
\label{shiki20}
\end{eqnarray}
We numerically find that the $SU(3)\times SU(2)\times U(1)$ gauge symmetry persists as the
global minimum of the effective potential for the wide range of the parameter $z < 1$ and $z \sim 10$
as far as our numerical analyses are concerned.
The vacuum expectation values for both parameter regions are
given by the desired vacuum in Eq.~(\ref{shiki18}). We also check that the desired vacuum is
the global minimum of the effective potential consist of only the bosonic degrees of freedom
in the matter content in Eq.~(\ref{shiki20}) for the high temperature limit~\cite{sakatake2}.
The above numerical analysis is performed in the range $0.1\leq z\leq10$, which corresponds to
the temperature around the GUT scale.
Although the reheating temperature can not be so high to avoid the so-called gravitino
problem~\cite{gravitinoproblem}, the result obtained above may give us an interesting scenario
of the vacuum selection in the evolution of the very early Universe before the inflation.
Supposing that the Universe starts with very high temperature, say the Planck scale,
there is an epoch with temperature corresponding to $z=O(1)$. Our
result shows an expectation that the dynamics in the epoch naturally selects the desired vacuum.
Of course, in order to claim that the vacuum continues to stay there up to the present, we have to
examine the potential for $z\ll 1$, where the one-loop effective potential is not
reliable due to large logarithmic corrections
as usual. For reliable analysis, such logarithmic terms should be resummed to result in the RG improved
potential, which is not an easy task in the framework of the five dimensions.
It is, however, instructive to see the behavior of the one-loop effective
potential for $z\ll 1$ as shown below.
In this region, the {$w$-}dependence in the fraction in {Eq.}~%
(\ref{shiki16}) can be neglected for
small $\omega$ $(\ll l/z)$, and then the function becomes
\begin{equation}
F_S(x, z)\sim-2z^5\sum_{l}\frac{(1-(-1)^l)}{l^5}\sum_{\omega}\cos(2\pi \omega x),
\label{shiki21}
\end{equation}
aside form the overall factor. If we carried out the summation
over $\omega$ to infinity, this function would become proportional
to the periodic delta function, which vanishes except for the points $x=0$ (mod 1).
Although this expression is not correct for large $\omega$, the contributions of the large $\omega$ tend
to oscillate rapidly to cancel each other out approximately for $x\neq0$ (mod 1).
Thus, we can see that it almost vanishes except for the neighborhoods of the points $x=0$ (mod 1).
Noting that each KK tower that couples with $A_y$ gives a contribution expressed by the function
$F_S(x, z)$ and the tower contains a massless mode on the points $x=0$ (mod 1),
we come up with an anticipation that the effective potential is dominated only by the zero mode
contribution. In fact, the decoupling theorem~\cite{decoupling} tells us that in the region away
from the points $x=0$ (mod 1)
all the modes should be decoupled to give a vanishing contribution as they are much heavier than
the temperature, which is consistent with the above behavior.
This anticipation can be confirmed by comparing $F_S(0,z)$ with the zero mode contribution (and also
their derivatives) in low temperature limit $z\ll1$. We present some formulae necessary in the
discussions given here in Appendix~\ref{Sec:EffPotAtFT}.
The above discussion justifies the evaluation of the the effective potential
using the four-dimensional effective theory that consists only of the zero modes,
as is the case with the Higgs quartic self coupling discussed in Ref.~\cite{EffTheo}.
This fact is, of course, nothing but what is required in the field theory and thus may not be
surprising at all, but it is quite nontrivial and amazing to see that it also holds in the
nonrenormalizable five-dimensional models.
This consideration is not only academically interesting but also useful in the analysis of the
effective potential for $z\ll1$, as it is much easier to study the RG improved
effective potential in the four-dimensional effective theory.
Although it appears to be difficult, at first glance, to carry out such a study model-independently
even in the four dimensional effective theory,
the SUSY allows us to derive a nontrivial conclusion.
At (relatively) high temperature $M_{\rm SB}\ll T~(\ll1/R ~\sim M_{\rm GUT})$ where the zero temperature SUSY
breaking is negligible, the interactions between the $A_y$ and the bosonic fields are determined
by the gauge interaction and the three point interaction in the superpotential.
It is well-known that the coupling constants of these interactions do not destabilize the system
independently of their values, and thus of the RG running effects.
At lower temperature $T\lesssim M_{\rm SB}$, the effective potential becomes dominated by the
zero temperature contribution.
This contribution is less controlled but we should impose that the desired vacuum is at
least a local minimum if we would like to work there, and such parameters are chosen
in Ref.~\cite{gGHU-pheno} actually.
In this way, we can see that the desired vacuum in Eq.~(\ref{shiki18}) does not become unstable
for $z\ll1$, while it
is not clear if it keeps being the global minimum as the global structure of the effective
potential is hard to see in the effective-theoretical approach.
Even if other vacua becomes deeper than the desired one, however, the transition rate is highly
suppressed because the depth of the potential is very shallow compared with the distance
between the minima, the order of $M_{\rm GUT}$, for $z\ll1$.
Let us note that the condition $z\ll 1$ should always hold after the inflation in the history of the
Universe. Thus, we conclude that in the models where the desired vacuum is the global minimum
at $z\gtrsim1$ discussed above the vacuum is likely selected by the dynamics before the inflation
and keeps staying there up to the present.
\section{Conclusions and Discussions}
We have studied the gauge symmetry breaking patterns by the Hosotani mechanism
of the models of the $SU(5)$ grand gauge-Higgs unification
in the framework of the diagonal embedding method. The method
enables us to have the zero mode for the adjoint scalar field and the chiral fermion simultaneously and to
utilize the one-loop effective potential obtained in the case of the $S^1$
compactification. We have investigated the vacuum structure of our models
for the given matter contents through the analyses of the effective potential numerically.
We have found matter contents that realize the symmetry breaking down to
each regular subgroup of the $SU(5)$. In
particular, we have presented those that give us the desired vacuum
with the $SU(3)\times SU(2)\times U(1)$ gauge symmetry as the global minimum.
One of the matter contents that lead to
the desired gauge symmetry breaking pattern
does not include the periodic fermions belonging to the adjoint representation,
contrary to the previous study~\cite{congru}.
We have also studied the finite temperature phase transition of the models of the
$SU(5)$ grand gauge-Higgs unification with the $SU(3)\times SU(2)\times U(1)$ gauge
symmetry being realized on the global minimum
at zero temperature. We have numerically obtained the
critical temperature and found that the phase transition is of the first order.
The restoration of the $SU(5)$ gauge symmetry in
the high temperature limit is observed as it should.
Then, we examine the SUSY version of the $SU(5)$ grand gauge-Higgs unification at finite
temperature. The SUSY version is phenomenologically fascinating when the SUSY breaking scale
is set around the
$\mbox{TeV}$ scale as the existence of the light adjoint chiral supermultiplets with masses of the same
scale are generally predicted to provide us a clue as to the GUT breaking
accessible at the $\mbox{TeV}$ scale experiments~\cite{gGHU-DTS,gGHU-pheno}.
Furthermore, the doublet-triplet splitting problem can be naturally solved~\cite{gGHU-DTS}
if the desired vacuum in Eq.~(\ref{shiki18}) is selected.
Since the effective potential is suppressed by the SUSY breaking scale at zero temperature,
the finite temperature
effects are crucial to study if the desired vacuum is actually selected dynamically.
At enough high temperature, the SUSY breaking at zero temperature can be neglected and
we have taken account of only the finite temperature effect for the SUSY breaking.
We have analyzed the one loop effective potential to find a matter content for which the phase
transition from the $SU(5)$ symmetric vacuum to the desired vacuum occurs at $z_c\sim 0.382$
as the temperature decreases. We have presented another matter content
with which the desired vacuum stays as the global minimum even for $z\gtrsim1$.
We have also examined the effective potential in lower temperature
region $M_{\rm SB}\ll T\ll1/R$ and found
that its behavior is consistent with the decoupling theorem also in our nonrenormalizable
five-dimensional setup.
Though the effective potential without the RG improvement is not reliable in this region,
the above observation shows that the effective potential is dominated by the zero mode contributions
and thus justifies the use of the four-dimensional effective theory, as usual.
Then, we have seen that the SUSY controls the finite temperature contribution well so that this
contribution stabilizes the desired vacuum irrelevantly to the RG running.
Thus, we conclude that in the models discussed in Sec.~\ref{Sec:SUSY-FT}, the vacuum selected
before the inflation keeps staying there up to the present.
Given that in the model whose matter content is shown in Eq.~(\ref{shiki20}) the desired vacuum is
the global minimum also for $z\gtrsim1$ and thus it is likely selected dynamically before the inflation,
this may provide us an interesting scenario of the vacuum selection in the evolution of the early
Universe.
Finally, we note that the examples analyzed in this paper are
kind of toy models
where we neglect the
contributions from the matter sector of the standard model and do not discuss the gauge coupling unification.
When we construct a realistic model, we should redo our analysis.
The present study creates an expectation that there exist several examples of appropriate matter
contents.
\begin{center}
{\bf Acknowledgement}
\end{center}
The authors would like to thank Dr. Y. Ookouchi (Kyushu Uni{v}.) for valuable discussions.
|
\section{Introduction}
Finite topological spaces have attracted increasing attention in the last years, principally from works by J. Barmak and G. Minian \cite{barmak2007minimal,barmak2008one,barmak2008simple,barmak2012strong}. One of the main interests of the theory of finite spaces is that they serve as models for weak homotopy types of compact polyhedra. More precisely, for every compact polyhedron $K$ there exists a finite T$_0$--space $\mathcal{X}(K)$ together with a weak homotopy equivalence $K\to\mathcal{X}(K)$ \cite{mccord1966singular}. Moreover, there is a functorial correspondence between finite T$_0$--spaces and finite posets \cite{alexandroff1937diskrete} which endows the theory of finite topological spaces with a natural combinatorial flavour. This allows the study of compact polyhedra (and often of general topological spaces) by means of combinatorial tools and gives a new insight into relevant topological questions \cite{barmak2011algebraic}.
A natural problem of this theory is to find finite topological spaces with the minimum number of points which have certain weak homotopy type. One of the first questions of this type was asked by J.P. May in \cite{may2003finite}, where he conjectures that, for all $n\in\mathbb{N}$, the $n$--fold non-Hausdorff suspension of the $0$--sphere, denoted $\mathbb{S}^n S^0$, is a \emph{minimal finite model} of the $n$--sphere, that is, a finite topological space which is weak homotopy equivalent to the $n$--sphere with the minimum possible cardinality. This question was answered positively by Barmak and Minian in \cite{barmak2007minimal}. Moreover, they also prove that $\mathbb{S}^n S^0$ is the only minimal finite model of the $n$--sphere. In the same article they give a characterization of the minimal finite models of the finite graphs.
Further minimality questions were formulated in \cite{hardie2002nontrivial} and in \cite{barmak2011algebraic} regarding finite models of the real projective plane and the torus. These problems were solved in \cite{cianci2015splitting}, where a characterization of all the minimal finite models of these spaces is given.
Also, in \cite{rival1976fixed} and in \cite{barmak2011algebraic}, a homotopically trivial non-contractible space of nine points is given. Thus, another natural question to pose in this theory is whether this is the minimum number of points that a homotopically trivial non-contractible space can have \cite[problem 3.5.4]{may2016finite}. In this article we give an affirmative answer to this question showing that a homotopically trivial non-contractible space must have at least nine points. Moreover, we find all the homotopically trivial non-contractible spaces of nine points.
\section{Preliminaries}
In this section we will recall the basic notions of the theory of finite topological spaces and fix notation. For a comprehensive exposition on finite spaces the reader may consult \cite{barmak2011algebraic}.
If $X$ is a finite topological space and $x\in X$, the intersection of all the open subsets of $X$ which contain $x$ is clearly an open subset and is denoted by $U_x$. Any finite T$_0$--space $X$ can be endowed with a partial order which is defined as follows: $x_1\leq x_2$ if and only if $U_{x_1}\subseteq U_{x_2}$. This defines a correspondence between finite T$_0$--spaces and finite posets which was first observed by Alexandroff \cite{alexandroff1937diskrete}. Moreover, under this correspondence continuous maps between finite T$_0$--spaces correspond to order-preserving morphisms between the respective posets. Hereafter, any finite T$_0$--space will be regarded also as a poset without further notice.
Let $X$ be a finite T$_0$--space and let $x\in X$. From the definition of the associated partial order it follows that $U_x=\{a\in X \;/\; a\leq x\}$. In a similar way, the smallest closed set which contains $x$ is $\overline{\{x\}}=\{a\in X \;/\; a\geq x\}$ and is denoted by $F_x$. It is also standard to define $\widehat{U}_x=\{a\in X \;/\; a< x\}$, $\widehat{F}_x=\{a\in X \;/\; a> x\}$, $C_x=U_x\cup F_x$ and $\widehat{C}_x=C_x-\{x\}$. We say that the point $x\in X$ is an \emph{up beat point} (resp. \emph{down beat point}) of $X$ if the subposet $\widehat{F}_x$ has a minimum (resp. if the subposet $\widehat{U}_x$ has a maximum) \cite{stong1966finite,may2003finite,barmak2011algebraic}.
Stong proves in \cite{stong1966finite} that if $x$ is a beat point of $X$ then $X-\{x\}$ is a strong deformation retract of $X$ and that two finite T$_0$--spaces are homotopy equivalent if and only if one obtains homeomorphic spaces after successively removing their beat points. Using the results of Stong it is easy to prove that a finite T$_0$--space which has a maximum or a minimum is contractible.
If $X$ is a finite T$_0$--space, $\mathcal{K}(X)$ will denote the \emph{order complex} of $X$, that is, the simplicial complex of the non-empty chains of $X$. Also, $X^\textnormal{op}$ will denote the poset $X$ with the inverse order and will be called the \emph{opposite} space of $X$.
McCord proves in \cite{mccord1966singular} that if $X$ is a finite T$_0$--space then there exists a weak homotopy equivalence from the geometric realization of $\mathcal{K}(X)$ to $X$. In particular, any finite T$_0$--space is weak homotopy equivalent to its opposite space since their order complexes coincide. Note also that the aforementioned result of McCord implies that the singular homology groups of a finite T$_0$--space $X$ are isomorphic to the simplicial homology groups of $\mathcal{K}(X)$.
The \emph{non-Hausdorff suspension} of a topological space $X$ is defined as the space $\mathbb{S} X$ whose underlying set is $X\amalg \{+,-\}$ and whose open sets are those of $X$ together with $X\cup\{+\}$, $X\cup\{-\}$ and $X\cup\{+,-\}$ \cite{mccord1966singular}. Note that, if $X$ is a finite T$_0$--space then the partial order in $\mathbb{S} X$ is induced by the partial order of $X$ together with the relations $x\leq +$ and $x\leq -$ for all $x\in X$. McCord proves that for every topological space $X$ there exists a weak homotopy equivalence between the suspension of $X$ and $\mathbb{S} X$ \cite{mccord1966singular}. As an example, he shows that, for all $n\in\mathbb{N}$, the $n$--sphere $S^n$ is weak homotopy equivalent to the $n$--fold non-Hausdorff suspension of the $0$--sphere $S^0$. Observe that $\mathbb{S}^n S^0$ is a finite T$_0$--space of $2n+2$ points.
May asked in \cite{may2003finite} if $\mathbb{S}^n S^0$ was the smallest space which is weak homotopy equivalent to the $n$--sphere. This question was answered by Barmak and Minian in \cite{barmak2007minimal}. More precisely, they proved the following theorem from which the affirmative answer to May's question follows.
\begin{theo} \label{theo_height_and_cardinality}
Let $X \neq\ast$ be a finite topological space without beat points. Then $X$ has at least $2h(X)+2$ points.
Moreover, if $X$ has exactly $2h(X)+2$ points, then it is homeomorphic to $\mathbb{S}^{h(X)} S^0$.
\end{theo}
In \cite{cianci2014homology} we studied the homology groups of finite T$_0$--spaces obtaining several results and applications. Among them we mention the following proposition, which will be needed later. In what follows, homology will always mean homology with integer coefficients. Thus, the group of coefficients will be omitted from the notation.
\begin{prop}\label{prop_relative_homology}
Let $X$ be a finite $T_0$--space and let $D$ be an antichain in $X$. Then $H_n(X,X-D)\cong\bigoplus\limits_{x\in D}\tilde{H}_{n-1}(\hat{C}_x)$ for every $n\in \mathbb{Z}$.
\end{prop}
If $X$ is a finite T$_0$--space, $\textnormal{mxl}(X)$ and $\textnormal{mnl}(X)$ will denote the subsets of maximal and minimal points of $X$ respectively. The following remark states some simple facts concerning the maximal and minimal points of a finite T$_0$--space. The first two items already appeared in \cite{cianci2015splitting}.
\begin{rem} \label{rem_mxl_mnl}
\
\begin{enumerate}
\item Let $X$ be a connected and finite T$_0$--space with more than one point. Then $\textnormal{mxl}(X)\cap\textnormal{mnl}(X)=\varnothing$.
\item Let $X$ be a finite T$_0$--space without beat points. If $a\in X-\textnormal{mxl}(X)$ then $\#(\widehat F_a \cap \textnormal{mxl}(X) )\geq 2$. Similarly, if $b\in X-\textnormal{mnl}(X)$ then $\#(\widehat U_b \cap \textnormal{mnl}(X) )\geq 2$.
\item Let $X$ be a finite T$_0$--space without beat points. If $\#\textnormal{mxl}(X)=2$ then $X$ is homeomorphic to $\mathbb{S}(X-\textnormal{mxl}(X))$.
\end{enumerate}
\end{rem}
Finally, if $X$ is a finite T$_0$--space, we define the \emph{height} of $X$ as
\begin{displaymath}
h(X)=\max\{\#c-1\;/\; \textnormal{$c$ is a chain of $X$}\}.
\end{displaymath}
Note that $h(X)=\dim \mathcal{K}(X)$.
\section{Results}
In this section we will prove that a homotopically trivial non-contractible space must have at least nine points. In addition, we will find all the homotopically trivial non-contractible spaces of nine points.
In figure \ref{fig_ht_non-contractible} we exhibit a homotopically trivial non-contractible T$_0$--space of nine points, which was also considered in \cite[Figure 2]{rival1976fixed} and in \cite[Example 4.3.3]{barmak2011algebraic}. Observe that this space is not contractible since it does not have beat points and is homotopically trivial since the geometric realization of its order complex is contractible.
\begin{figure}[h!]
\begin{tikzpicture}[x=3cm,y=3cm]
\tikzstyle{every node}=[font=\footnotesize]
\foreach \x in {1,...,3} \draw (0.5*\x,1) node(a\x){$\bullet$} node[above=1]{$a_{\x}$};
\foreach \x in {1,...,3} \draw (0.5*\x,0.5) node(b\x){$\bullet$} node[right=1]{$b_{\x}$};
\foreach \x in {1,...,3} \draw (0.5*\x,0) node(c\x){$\bullet$} node[below=1]{$c_{\x}$};
\foreach \x in {1,2} \draw (a1)--(b\x);
\foreach \x in {1,3} \draw (a2)--(b\x);
\foreach \x in {2,3} \draw (a3)--(b\x);
\foreach \x in {1,2} \draw (c1)--(b\x);
\foreach \x in {1,2,3} \draw (c2)--(b\x);
\foreach \x in {2,3} \draw (c3)--(b\x);
\end{tikzpicture}
\caption{Homotopically trivial non-contractible space of 9 points.}
\label{fig_ht_non-contractible}
\end{figure}
We will give now several lemmas which will be useful to prove the main results of this article.
\begin{lemma} \label{lemma_card_Ua_Ub}
Let $X$ be a finite T$_0$--space without beat points. Let $a,b\in X$ with $a>b$. Then $\#\widehat{U}_a\geq \#\widehat{U}_b+2$ and $\#\widehat{F}_b\geq \#\widehat{F}_a+2$.
\end{lemma}
\begin{proof}
Note that $\widehat U_a \supseteq U_b$ since $b<a$. And since $a$ is not a beat point of $X$ we obtain that $\widehat{U}_a \supsetneq U_b$. Therefore $\# \widehat U_a\geq \# U_b + 1=\#\widehat{U}_b+2$.
Applying this result to $X^\textnormal{op}$ we obtain that $\#\widehat{F}_b\geq \#\widehat{F}_a+2$.
\end{proof}
\begin{lemma} \label{lemma_subspace_with_non_trivial_H2}
Let $X$ be a finite T$_0$--space such that $h(X)=2$. If $X$ is homotopically trivial then $H_2(A)=0$ for all subspaces $A\subseteq X$.
\end{lemma}
\begin{proof}
Let $A$ be a subspace of $X$. Note that $H_3(X,A)=0$ since $h(X)=2$. The result then follows from the exact sequence $H_3(X,A) \longrightarrow H_2(A) \longrightarrow H_2(X)$.
\end{proof}
\begin{lemma} \label{lemma_no_subspace_finite_S2}
Let $X$ be a homotopically trivial finite T$_0$--space such that $h(X)=2$. Let $b,b'\in X-(\textnormal{mxl}(X)\cup\textnormal{mnl}(X))$. If $\#(\widehat{F}_{b}\cap \widehat{F}_{b'})\geq 2$ then $\#(\widehat{U}_{b}\cap \widehat{U}_{b'})\leq 1$.
\end{lemma}
\begin{proof}
Note that $\{b,b'\}$ must be an antichain. Suppose that $\#(\widehat{U}_{b}\cap \widehat{U}_{b'})\geq 2$. Then there exist distinct elements $a,a'\in \widehat{F}_{b}\cap \widehat{F}_{b'}$ and distinct elements $c,c'\in \widehat{U}_{b}\cap \widehat{U}_{b'}$. Note that $\{a,a'\}\subseteq \textnormal{mxl}(X)$ and that $\{c,c'\}\subseteq \textnormal{mnl}(X)$. Hence the subspace $A=\{a,a',b,b',c,c'\}$ is homeomorphic to $\mathbb{S}^2 S^0$ and then $H_2(A)\neq 0$, contradicting \ref{lemma_subspace_with_non_trivial_H2}. Thus $\#(\widehat{U}_{b}\cap \widehat{U}_{b'})\leq 1$.
\end{proof}
\begin{lemma} \label{lemma_homotopically_trivial_suspension}
Let $X$ be a topological space such that, for some $x_0\in X$, $\pi_1(X,x_0)$ is not a non-trivial perfect group. If $\mathbb{S} X$ is homotopically trivial then $X$ is homotopically trivial.
\end{lemma}
\begin{proof}
Let $\Sigma X$ be the suspension of $X$. Since $\Sigma X$ is weak homotopy equivalent to $\mathbb{S} X$ we obtain that $\Sigma X$ is homotopically trivial. Thus $H_n(\Sigma X)=0$ for all $n\in\mathbb{N}$. Hence $X$ is path-connected and $H_n(X)=0$ for all $n\in\mathbb{N}$. Thus $\pi_1(X,x_0)$ is a perfect group. Therefore $\pi_1(X,x_0)$ must be the trivial group. The result then follows from Hurewicz's theorem.
\end{proof}
The following proposition shows that the height of a homotopically trivial non-contractible finite T$_0$--space must be greater than 1.
\begin{prop} \label{prop_ht_of_height_1}
Let $X$ be a homotopically trivial finite T$_0$--space with $h(X)\leq 1$. Then $X$ is contractible.
\end{prop}
\begin{proof}
Suppose that $X$ is not contractible. Without loss of generality, we may assume that $X$ does not have beat points. Let $E$ denote the set of edges of the Hasse diagram of $X$. Let $\mathcal{R}\subseteq X\times E$ be the relation defined by $x\mathcal{R} a$ if and only if the point $x$ belongs to the edge $a$.
Note that $\#\mathcal{R}=2\#E$. And since $X$ is a path-connected space, $\#X\geq 2$ and $X$ does not have beat points, from \ref{rem_mxl_mnl} we obtain that $\#\mathcal{R}\geq 2\#X$. Thus $\#E\geq \#X$. But $1=\chi(X)=\# X - \#E \leq 0$ which entails a contradiction.
\end{proof}
We will prove now one of the main results of this article.
\begin{theo} \label{theo_ht_non_contractible_1}
Let $X$ be a homotopically trivial non-contractible topological space. Then $\#X \geq 9$.
\end{theo}
\begin{proof}
By the works of McCord \cite{mccord1966singular} and Stong \cite{stong1966finite} we may assume that $X$ is a finite T$_0$--space without beat points.
By \ref{theo_height_and_cardinality}, if $h(X)\geq 4$ then $\#X\geq 10$. And by \ref{prop_ht_of_height_1}, $h(X)\geq 2$. Thus $h(X)=2$ or $h(X)=3$.
\underline{Case 1}: $h(X)=3$. By \ref{theo_height_and_cardinality}, $\#X\geq 8$. If $\#X=8$ then, again by \ref{theo_height_and_cardinality}, $X$ is homeomorphic to $\mathbb{S}^{3} S^0$ which is weak equivalent to $S^3$. But this is a contradiction since $X$ is homotopically trivial. Thus $\#X\geq 9$.
\underline{Case 2}: $h(X)=2$. If $\#\textnormal{mxl}(X)=2$ then, from \ref{rem_mxl_mnl}, we obtain that $X=\mathbb{S} Y$ with $Y=X-\textnormal{mxl}(X)$ and $h(Y)=1$. Thus $\pi_1(Y,y_0)$ is a free group for all $y_0\in Y$. By \ref{lemma_homotopically_trivial_suspension}, $Y$ is homotopically trivial. Since $X$ does not have beat points, the same holds for $Y$. But this is not possible by \ref{prop_ht_of_height_1}. Thus $\#\textnormal{mxl}(X)\geq 3$.
Working with $X^\textnormal{op}$ in a similar way, we obtain that $\#\textnormal{mnl}(X)\geq 3$. Let $\mathcal{B}_X=X-(\textnormal{mxl}(X)\cup\textnormal{mnl}(X))$. Since $h(X)=2$ we obtain that $\mathcal{B}_X$ is a non-empty antichain of $X$. If $\#\mathcal{B}_X\geq 3$ then $\# X \geq 9$. Thus we may assume that $\#\mathcal{B}_X\leq 2$.
\underline{Case 2.1}: $\#\mathcal{B}_X = 1$. Suppose that $\mathcal{B}_X=\{b\}$. Note that $\widehat{F}_b\subseteq \textnormal{mxl}(X)$ and $\widehat{U}_b\subseteq \textnormal{mnl}(X)$. Let $\alpha_b=\# \widehat{F}_b$ and $\beta_b=\#\widehat{U}_b$. By \ref{rem_mxl_mnl}, $\alpha_b \geq 2$ and $\beta_b\geq 2$.
Since $X$ is path-connected, from the first item of \ref{rem_mxl_mnl} we obtain that $\textnormal{mxl}(X)\cap\textnormal{mnl}(X)=\varnothing$ and thus $\# \widehat{U}_a\geq 2$ for all $a\in\textnormal{mxl}(X)$ by the second item of \ref{rem_mxl_mnl}. From \ref{lemma_card_Ua_Ub} we obtain that $\# \widehat{U}_a\geq \beta_b + 2$ for all $a\in\widehat{F}_b$.
Let $l$ denote the number of $1$--chains of $X$ and let $m=\#\textnormal{mxl}(X)$. We have that
\begin{displaymath}
l=\# \widehat{U}_b + \!\! \sum_{a\in \textnormal{mxl}(X)}\!\!\!\!\!\# \widehat{U}_a \geq \beta_b + \alpha_b(\beta_b+2)+ 2(m-\alpha_b) = \beta_b + \alpha_b\beta_b + 2m\geq \alpha_b\beta_b + 8.
\end{displaymath}
On the other hand, note that the number of $2$--chains of $X$ is $\alpha_b\beta_b$.
Hence
\begin{displaymath}
1=\chi(X)=\#X-l+\alpha_b\beta_b \leq \#X - (\alpha_b\beta_b + 8) + \alpha_b\beta_b = \#X -8 \ .
\end{displaymath}
Therefore $\#X\geq 9$.
\underline{Case 2.2}: $\#\mathcal{B}_X = 2$. If $\#\textnormal{mxl}(X)\geq 4$ or $\#\textnormal{mnl}(X)\geq 4$ then $\#X\geq 9$. Thus we may assume that $\#\textnormal{mxl}(X)=\#\textnormal{mnl}(X)=3$. As above, note that $\widehat{F}_b\subseteq \textnormal{mxl}(X)$ and $\widehat{U}_b\subseteq \textnormal{mnl}(X)$ for all $b\in \mathcal{B}_X$ since $\mathcal{B}_X$ is an antichain.
First, we will prove that $\# \widehat{U}_b \neq 3$ for all $b\in \mathcal{B}_X$. Let $b\in \mathcal{B}_X$ and suppose that $\# \widehat{U}_b = 3$. Then $\widehat{U}_b=\textnormal{mnl}(X)$. Let $b'$ be the element of $\mathcal{B}_X-\{b\}$. We claim that $\widehat{F}_b\subseteq \widehat{F}_{b'}$. Indeed, let $a\in \widehat{F}_b$. Then $\widehat{U}_a \supseteq U_b$ and since $a$ is not a beat point of $X$ we obtain that $\widehat{U}_a \supsetneq U_b$. But since $U_b \supseteq \textnormal{mnl}(X)$ and $a\in\textnormal{mxl}(X)$ it follows that $b'\in \widehat{U}_a$ and hence $a\in \widehat{F}_{b'}$.
By \ref{rem_mxl_mnl}, $\# \widehat{F}_b \geq 2$ and $\# \widehat{U}_{b'} \geq 2$. Thus $\#(\widehat{F}_b\cap \widehat{F}_{b'})\geq 2$ and $\#(\widehat{U}_b\cap \widehat{U}_{b'})\geq 2$ contradicting \ref{lemma_no_subspace_finite_S2}. Therefore $\# \widehat{U}_b \neq 3$ for all $b\in \mathcal{B}_X$.
Hence $\# \widehat{U}_b = 2$ for all $b\in \mathcal{B}_X$ by \ref{rem_mxl_mnl}. In a similar way, we obtain that $\# \widehat{F}_b = 2$ for all $b\in \mathcal{B}_X$.
Let $\textnormal{mxl}(X)=\{a_1,a_2,a_3\}$, $\mathcal{B}_X=\{b_1,b_2\}$ and $\textnormal{mnl}(X)=\{c_1,c_2,c_3\}$. Without loss of generality, we may assume that $\widehat{F}_{b_1}=\{a_1,a_2\}$ and $\widehat{U}_{b_1}=\{c_1,c_2\}$. We will prove that $\#\widehat{U}_{a_1}+\#\widehat{U}_{a_2}+\#\widehat{U}_{a_3}\geq 12$ analizing two cases: $b_2<a_3$ and $b_2\nless a_3$.
If $b_2<a_3$ then $\#\widehat{U}_{a_3}\geq\#\widehat{U}_{b_2}+2= 4$ by \ref{lemma_card_Ua_Ub}. And since $b_1<a_1$ and $b_1<a_2$, we also obtain that $\#\widehat{U}_{a_1}\geq 4$ and $\#\widehat{U}_{a_2}\geq 4$ by \ref{lemma_card_Ua_Ub}. Thus $\#\widehat{U}_{a_1}+\#\widehat{U}_{a_2}+\#\widehat{U}_{a_3}\geq 12$.
If $b_2\nless a_3$ then $\widehat{F}_{b_2}=\{a_1,a_2\}=\widehat{F}_{b_1}$. By \ref{lemma_no_subspace_finite_S2}, $\widehat{U}_{b_2}\neq \widehat{U}_{b_1}=\{c_1,c_2\}$. Hence $c_3<b_2$. Thus $\widehat{U}_{a_1}=\widehat{U}_{a_2}=\{b_1,b_2,c_1,c_2,c_3\}$ and since $\#\widehat{U}_{a_3}\geq 2$ by \ref{rem_mxl_mnl} we obtain that $\#\widehat{U}_{a_1}+\#\widehat{U}_{a_2}+\#\widehat{U}_{a_3}\geq 12$.
Therefore $\#\widehat{U}_{a_1}+\#\widehat{U}_{a_2}+\#\widehat{U}_{a_3}\geq 12$ in any case. As above, let $l$ denote the number of $1$--chains of $X$. Then, applying \ref{rem_mxl_mnl}, we get
\begin{displaymath}
l=\#\widehat{U}_{a_1}+\#\widehat{U}_{a_2}+\#\widehat{U}_{a_3}+\#\widehat{U}_{b_1}+\#\widehat{U}_{b_2}\geq 12 + 2 + 2 = 16.
\end{displaymath}
Now, note that the number of $2$--chains of $X$ is $\#\widehat{F}_{b_1}\#\widehat{U}_{b_1}+\#\widehat{F}_{b_2}\#\widehat{U}_{b_2}=8$. Thus
\begin{displaymath}
\chi(X)= \#X-l+8 \leq 8 - 16 + 8 = 0
\end{displaymath}
and hence the space $X$ is not homotopically trivial.
\end{proof}
As a corollary of the previous theorem we obtain that the space of figure \ref{fig_ht_non-contractible} is a homotopically trivial non-contractible space with the minimum possible number of points. In the following theorem we find all the homotopically trivial non-contractible spaces of this minimum number of points.
\begin{theo}
Let $X$ be a homotopically trivial non-contractible topological space such that $\#X = 9$. Then $X$ is homeomorphic to either the space of figure \ref{fig_ht_non-contractible} or its opposite.
\end{theo}
\begin{proof}
By \ref{theo_ht_non_contractible_1} we may assume that $X$ is a finite T$_0$--space without beat points. By \ref{rem_mxl_mnl}, $\#\textnormal{mxl}(X)\geq 2$. If $\#\textnormal{mxl}(X)= 2$, $X$ is homeomorphic to $\mathbb{S}(X-\textnormal{mxl}(X))$ by \ref{rem_mxl_mnl}. Note that $X-\textnormal{mxl}(X)$ does not have beat points. Since $\#(X-\textnormal{mxl}(X))=7$, from \cite[theorem 5.7]{cianci2015splitting} we obtain that $\pi_1(X-\textnormal{mxl}(X),x_0)$ is a free group for all $x_0\in X-\textnormal{mxl}(X)$. Thus $X-\textnormal{mxl}(X)$ is homotopically trivial by \ref{lemma_homotopically_trivial_suspension}, which contradicts \ref{theo_ht_non_contractible_1}. Therefore $\#\textnormal{mxl}(X)\geq 3$. And applying this argument to $X^\textnormal{op}$ we obtain that $\#\textnormal{mnl}(X)\geq 3$.
Let $\mathcal{B}_X=X-(\textnormal{mxl}(X)\cup\textnormal{mnl}(X))$. From the first item of \ref{rem_mxl_mnl} it follows that $\#\mathcal{B}_X\leq 3$. By \ref{prop_ht_of_height_1}, $h(X)\geq 2$. Thus $\mathcal{B}_X\neq\varnothing$.
We will analize three cases which correspond to the possible cardinalities of the subset $\mathcal{B}_X$.
\underline{Case 1}: $\# \mathcal{B}_X=1$. Let $b$ be the only element of $B$. Let $n=\#\textnormal{mnl}(X)$, $\alpha=\#\widehat{F}_b$ and $\beta=\#\widehat{U}_b$. By \ref{rem_mxl_mnl}, $\alpha\geq 2$ and $\beta\geq 2$. Let $\mathcal{R}\subseteq X\times X$ be the order relation of $X$ and let $\mathcal{S}=\mathcal{R}\cap(\textnormal{mnl}(X)\times\textnormal{mxl}(X))$.
Let $\mathcal{S}_1=\mathcal{R}\cap(\widehat{U}_b\times\widehat{F}_b)$, $\mathcal{S}_2=\mathcal{R}\cap(\widehat{U}_b\times(\textnormal{mxl}(X)-\widehat{F}_b))$ and $\mathcal{S}_3=\mathcal{R}\cap((\textnormal{mnl}(X)-\widehat{U}_b)\times\textnormal{mxl}(X))$. Clearly, $\mathcal{S}_1$, $\mathcal{S}_2$ and $\mathcal{S}_3$ are pairwise disjoint and $\mathcal{S}=\mathcal{S}_1\cup \mathcal{S}_2\cup \mathcal{S}_3$.
Note that $\#\mathcal{S}_1=\alpha\beta$. Also, if $z\in \widehat{U}_b$ then $\widehat{F}_z\supsetneq F_b$ since $z$ is not a beat point of $X$. Hence $\widehat{F}_z\cap (\textnormal{mxl}(X)-\widehat{F}_b)\neq\varnothing$ for all $z\in \widehat{U}_b$. Thus $\#\mathcal{S}_2\geq \beta$. On the other hand, from \ref{rem_mxl_mnl} we obtain that $\#\mathcal{S}_3\geq 2\#(\textnormal{mnl}(X)-\widehat{U}_b)=2(n-\beta)$.
Thus
\begin{displaymath}
\#\mathcal{S}\geq \alpha\beta + \beta + 2(n-\beta)\geq \alpha\beta + 2 + 2(n-\beta) \ .
\end{displaymath}
Proceeding in a similar way we also obtain that
\begin{displaymath}
\#\mathcal{S} \geq \alpha\beta + 2 + 2 \# (\textnormal{mxl}(X)-\widehat{F}_b)= \alpha\beta + 2 + 2 (8-n-\alpha) \ .
\end{displaymath}
Hence
\begin{displaymath}
\#\mathcal{S} \geq \alpha\beta + 2 + 2\max\{n-\beta,8-n-\alpha\}\geq \alpha\beta + 2 + (n-\beta) + (8-n-\alpha) = \alpha\beta+10-\alpha-\beta \ .
\end{displaymath}
Thus we obtain that
\begin{displaymath}
\chi(X)=9-(\#\mathcal{S}+\alpha+\beta)+\alpha\beta \leq 9 - \alpha\beta - 10 + \alpha\beta = -1 \ .
\end{displaymath}
Hence $X$ is not homotopically trivial. Therefore this case is not possible.
\underline{Case 2}: $\# \mathcal{B}_X=2$. Without loss of generality we may assume that $\#\textnormal{mxl}(X)=3$ and $\#\textnormal{mnl}(X)=4$. Let $b_1$ and $b_2$ be the elements of $\mathcal{B}_X$.
We will prove that $\mathcal{B}_X$ is an antichain. Indeed, if $b_1<b_2$ then $\# \widehat{F}_{b_1} \geq \#\widehat{F}_{b_2}+2 \geq 4$ by \ref{lemma_card_Ua_Ub} and \ref{rem_mxl_mnl}. Thus $\widehat{F}_{b_1}=\{b_2\}\cup\textnormal{mxl}(X)$.
Let $c\in\textnormal{mnl}(X)\cap U_{b_1}$. Then $F_{b_1}\subseteq \widehat{F}_c \subseteq X-\textnormal{mnl}(X) = \mathcal{B}_X\cup\textnormal{mxl}(X) = F_{b_1}$. Hence $\widehat{F}_c=F_{b_1}$ and then $c$ is a beat point of $X$, which contradicts our assumptions. Therefore $\mathcal{B}_X$ must be an antichain.
Hence $h(X)=2$. For $j\in\{1,2\}$, let $\alpha_j=\#\widehat{F}_{b_j}$ and $\beta_j=\#\widehat{U}_{b_j}$. Note that $\alpha_j\geq 2$ and $\beta_j\geq 2$ for all $j\in\{1,2\}$ by \ref{rem_mxl_mnl}. As in the previous case, let $\mathcal{R}\subseteq X\times X$ be the order relation of $X$ and let $\mathcal{S}=\mathcal{R}\cap(\textnormal{mnl}(X)\times\textnormal{mxl}(X))$. Thus
\begin{displaymath}
\chi(X)=9-(\alpha_1+\beta_1+\alpha_2+\beta_2+\#\mathcal{S})+\alpha_1\beta_1+\alpha_2\beta_2=7-\#\mathcal{S}+\sum_{j=1}^2 (\alpha_j-1)(\beta_j-1)\ .
\end{displaymath}
We will analyze two subcases.
\underline{Case 2.1}: $\#(\widehat{F}_{b_1}\cap \widehat{F}_{b_2})=1$. In this case, $\alpha_1=\alpha_2=2$ and $\widehat{F}_{b_1}\cup \widehat{F}_{b_2}=\textnormal{mxl}(X)$. Let $a$ be the only point of $\widehat{F}_{b_2}-\widehat{F}_{b_1}$.
We will prove now that $\beta_2\leq 3$. Suppose that $\beta_2>3$. Then $\beta_2=4$ and $\widehat{U}_{b_2}=\textnormal{mnl}(X)$. Thus $\widehat{U}_{a}=U_{b_2}$ and hence $a$ is a beat point of $X$, which entails a contradiction. Therefore $\beta_2\leq 3$.
We will prove now that $\textnormal{mxl}(X)\subseteq F_c$ for all $c\in \widehat{U}_{b_1}$. Let $c\in \widehat{U}_{b_1}$. If $b_2>c$ then $F_c\supseteq F_{b_1}\cup F_{b_2}\supseteq \textnormal{mxl}(X)$. If $b_2 \ngtr c$ then $F_{b_1}\subseteq \widehat{F}_c \subseteq \{b_1\}\cup\textnormal{mxl}(X)= F_{b_1}\cup \{a\}$. And since $c$ is not a beat point of $X$, we obtain that $\widehat{F}_c = F_{b_1}\cup \{a\}$ and hence $\textnormal{mxl}(X)\subseteq F_c$.
Since $\textnormal{mxl}(X)\subseteq F_c$ for all $c\in \widehat{U}_{b_1}$ and applying \ref{rem_mxl_mnl}, we obtain that
\begin{displaymath}
\#\mathcal{S}\geq 3\#\widehat{U}_{b_1} + 2 \#(\textnormal{mnl}(X)-\widehat{U}_{b_1})= 3 \beta_1 + 2 (4-\beta_1)=\beta_1+8 \ .
\end{displaymath}
Thus
\begin{displaymath}
\chi(X)=7-\#\mathcal{S}+\sum_{j=1}^2 (\alpha_j-1)(\beta_j-1)\leq 7 - \beta_1 - 8 + \beta_1-1+\beta_2-1=\beta_2-3 \leq 0
\end{displaymath}
and hence $X$ is not homotopically trivial.
\underline{Case 2.2}: $\#(\widehat{F}_{b_1}\cap \widehat{F}_{b_2})\geq 2$. By \ref{lemma_no_subspace_finite_S2}, $\#(\widehat{U}_{b_1}\cap \widehat{U}_{b_2})\leq 1$.
Since $\beta_1\geq 2$ and $\beta_2\geq 2$ we obtain that $\widehat{U}_{b_1}-\widehat{U}_{b_2}\neq\varnothing$ and $\widehat{U}_{b_2}-\widehat{U}_{b_1}\neq\varnothing$. Let $c \in \widehat{U}_{b_1}-\widehat{U}_{b_2}$. We have that $F_{b_1}\subseteq\widehat{F}_{c}\subseteq \{b_1\}\cup\textnormal{mxl}(X)$. And since $c$ is not a beat point of $X$ we obtain that $\widehat{F}_{c}\neq F_{b_1}$. Therefore $\textnormal{mxl}(X)\nsubseteq F_{b_1}$ which implies that $\alpha_1 = 2$. Thus $\#(\textnormal{mxl}(X)-F_{b_1})= 1$ and since $\widehat{F}_{c}\neq F_{b_1}$ we obtain that $\textnormal{mxl}(X)\subseteq F_c$.
In a similar way we obtain that $\alpha_2 = 2$ and that $\textnormal{mxl}(X)\subseteq F_c$ for all $c\in \widehat{U}_{b_2}-\widehat{U}_{b_1}$. Thus, applying \ref{rem_mxl_mnl}, we obtain that $\#\mathcal{S}\geq 3+3+2+2 = 10$.
Hence
\begin{displaymath}
\chi(X)=7-\#\mathcal{S}+\sum_{j=1}^2 (\alpha_j-1)(\beta_j-1)\leq 7 -10 + \beta_1-1+\beta_2-1=\beta_1+\beta_2-5 \leq 0
\end{displaymath}
where the last inequality follows from the fact that $\#(\widehat{U}_{b_1}\cap \widehat{U}_{b_2})\leq 1$. Therefore $X$ is not homotopically trivial.
\underline{Case 3}: $\# \mathcal{B}_X=3$. Note that $\#\textnormal{mxl}(X)=\#\textnormal{mnl}(X)=3$. Let $b_1$, $b_2$ and $b_3$ be the elements of $\mathcal{B}_X$.
Suppose that $\mathcal{B}_X$ is a chain. Without loss of generality we may assume that $b_1<b_2<b_3$. Applying \ref{lemma_card_Ua_Ub} and \ref{rem_mxl_mnl} we obtain that
\begin{displaymath}
\# \widehat{F}_{b_1}\geq \# \widehat{F}_{b_2} +2 \geq \# \widehat{F}_{b_3} + 4 \geq 6
\end{displaymath}
but this can not be possible since $\widehat{F}_{b_1}\subseteq \{b_2,b_3\}\cup \textnormal{mxl}(X)$. Thus $\mathcal{B}_X$ is not a chain.
Hence, without loss of generality, we may assume that $b_1$ and $b_3$ are incomparable.
Suppose now that $\mathcal{B}_X$ is not an antichain. Without loss of generality we may assume that $b_2$ and $b_1$ are comparable elements, and considering $X^\textnormal{op}$ if necessary, we may suppose that $b_2>b_1$. Note that $b_2\nless b_3$ since $\mathcal{B}_X$ is not a chain. Hence $\widehat{F}_{b_2}\subseteq \textnormal{mxl}(X)$. By \ref{rem_mxl_mnl}, $\#\widehat{F}_{b_2}\geq 2$. Let $a_1$ and $a_2$ be distinct elements of $\widehat{F}_{b_2}$ and let $a_3$ be the remaining maximal element of $X$. By \ref{lemma_card_Ua_Ub}, $\#\widehat{F}_{b_1}\geq \#\widehat{F}_{b_2}+2 \geq 4$ and since $b_1$ and $b_3$ are incomparable it follows that $\widehat{F}_{b_1}=\{b_2,a_1,a_2,a_3\}$ and $\widehat{F}_{b_2}=\{a_1,a_2\}$.
Now, observe that $\widehat{F}_{b_3}\subseteq \{b_2,a_1,a_2,a_3\}$ since $b_1$ and $b_3$ are incomparable. We claim that $\widehat{F}_{b_3}$ is not path-connected. Indeed, if $b_3<b_2$ then proceeding as in the previous paragraph we obtain that $\widehat{F}_{b_3}=\{b_2,a_1,a_2,a_3\}$ which is not path-connected. And if $b_3\nless b_2$ then $\widehat{F}_{b_3}\subseteq\textnormal{mxl}(X)$ and hence $\widehat{F}_{b_3}$ is a discrete subspace and $\#\widehat{F}_{b_3}\geq 2$ by \ref{rem_mxl_mnl}. Thus $\widehat{F}_{b_3}$ is not path-connected.
By \ref{rem_mxl_mnl}, $\# (\widehat{U}_{b_1}\cap\textnormal{mnl}(X))\geq 2$. Let $c_1$ and $c_2$ be distinct elements of $\widehat{U}_{b_1}\cap\textnormal{mnl}(X)$ and let $c_3$ be the remaining minimal element of $X$. Since $c_1<b_1$ and $c_1$ is not a beat point of $X$ we obtain that $\{b_1,b_2,a_1,a_2,a_3\}=F_{b_1}\subsetneq\widehat{F}_{c_1}\subseteq X-\textnormal{mnl}(X)$. Hence $\widehat{F}_{c_1}=\{b_1,b_2,b_3,a_1,a_2,a_3\}$. In a similar way, $\widehat{F}_{c_2}=\{b_1,b_2,b_3,a_1,a_2,a_3\}$.
From \ref{prop_relative_homology} we obtain that
$H_1(\widehat{F}_{c_1})\cong H_1(\widehat{F}_{c_1},F_{b_1}) \cong \widetilde{H}_0(\widehat{F}_{b_3})\neq 0$
since $\widehat{F}_{b_3}$ is not path-connected.
Applying \ref{prop_relative_homology} again we obtain that
$H_2(X)\cong H_2(X,F_{c_2})\cong H_1(\widehat{F}_{c_1})\oplus H_1(\widehat{F}_{c_3}) \neq 0$
and hence $X$ is not homotopically trivial. Therefore $\mathcal{B}_X$ must be an antichain.
Hence $h(X)=2$. For $j\in\{1,2,3\}$, let $\alpha_j=\#\widehat{F}_{b_j}$ and $\beta_j=\#\widehat{U}_{b_j}$. Note that $\alpha_j\geq 2$ and $\beta_j\geq 2$ for all $j\in\{1,2,3\}$ by \ref{rem_mxl_mnl}. As in the previous cases, let $\mathcal{R}\subseteq X\times X$ be the order relation of $X$ and let $\mathcal{S}=\mathcal{R}\cap(\textnormal{mnl}(X)\times\textnormal{mxl}(X))$. Thus
\begin{displaymath}
1=\chi(X)=9-\left(\sum_{j=1}^3\alpha_j+\sum_{j=1}^3\beta_j +\#\mathcal{S}\right)+\sum_{j=1}^3\alpha_j\beta_j = 6-\#\mathcal{S}+\sum_{j=1}^3 (\alpha_j-1)(\beta_j-1)\ .
\end{displaymath}
Hence
\begin{displaymath}
\#\mathcal{S}= 5+\sum_{j=1}^3 (\alpha_j-1)(\beta_j-1)\geq 8\ .
\end{displaymath}
Thus $\#\mathcal{S}=8$ or $\#\mathcal{S}=9$ and hence $\sum\limits_{j=1}^3 (\alpha_j-1)(\beta_j-1)\in\{3,4\}$.
Therefore at least five of the numbers $\alpha_1$, $\alpha_2$, $\alpha_3$, $\beta_1$, $\beta_2$ and $\beta_3$ are equal to $2$ and the remaining one might be $2$ or $3$.
Without loss of generality and considering $X^\textnormal{op}$ if necessary, we may assume that $\alpha_1=\alpha_2=\alpha_3=2$.
\underline{Claim 1}: $\#(\widehat{F}_{b_k}\cap\widehat{F}_{b_l})=1$ for all $k,l\in\{1,2,3\}$ with $k\neq l$.
Suppose that there exist $k,l\in\{1,2,3\}$ with $k\neq l$ such that $\#(\widehat{F}_{b_k}\cap\widehat{F}_{b_l})=2$. Without loss of generality we may assume that $\#(\widehat{F}_{b_1}\cap\widehat{F}_{b_2})=2$.
Thus $\#(\widehat{U}_{b_1}\cap \widehat{U}_{b_2})\leq 1$ by \ref{lemma_no_subspace_finite_S2}. And since $\#\textnormal{mnl}(X)=3$ we obtain that $\#(\widehat{U}_{b_1}\cap \widehat{U}_{b_2})=1$. Hence $\beta_1=\beta_2=2$. Thus $\widehat{U}_{b_1}\cup \widehat{U}_{b_2}=\textnormal{mnl}(X)$.
Let $a_1$ and $a_2$ be the elements of $\widehat{F}_{b_1}\cap\widehat{F}_{b_2}$ and let $a_3$ be the remaining maximal element of $X$. Note that $U_{a_1}\supseteq \textnormal{mnl}(X)$ and $U_{a_2}\supseteq \textnormal{mnl}(X)$.
If $\{a_1,a_2\}\subseteq F_{b_3}$ then $\#(\widehat{U}_{b_1}\cap \widehat{U}_{b_3})\leq 1$ and $\#(\widehat{U}_{b_2}\cap \widehat{U}_{b_3})\leq 1$ by \ref{lemma_no_subspace_finite_S2}. Thus $\#(\widehat{U}_{b_1}\cap \widehat{U}_{b_3})=1$ and $\#(\widehat{U}_{b_2}\cap \widehat{U}_{b_3})=1$. Hence $|\mathcal{K}(X-\{a_3\})|$ is homeomorphic to $S^2$ contradicting \ref{lemma_subspace_with_non_trivial_H2}. Thus $\{a_1,a_2\}\nsubseteq F_{b_3}$.
Hence, $a_3\in \widehat{F}_{b_3}$. Since $\alpha_1=\alpha_2=2$, we obtain that $b_1\nless a_3$ and $b_2\nless a_3$. And since $a_3$ is not a beat point of $X$ we obtain that $U_{b_3}\subsetneq\widehat{U}_{a_3}\subseteq \{b_3\}\cup\textnormal{mnl}(X)$. But $\# U_{b_3}\geq 3$. Thus $\# U_{b_3}= 3$, $\beta_3=2$ and $\widehat{U}_{a_3}=\{b_3\}\cup\textnormal{mnl}(X)$. Hence $\mathcal{S}=\textnormal{mnl}(X)\times\textnormal{mxl}(X)$. Thus
\begin{displaymath}
\chi(X)=6-\#\mathcal{S}+\sum\limits_{j=1}^3 (\alpha_j-1)(\beta_j-1)=6-9+3=0
\end{displaymath}
which entails a contradiction. This proves claim 1.
Hence $X-\textnormal{mnl}(X)$ is homeomorphic to the following space
\begin{displaymath}
\begin{tikzpicture}[x=3cm,y=3cm]
\tikzstyle{every node}=[font=\footnotesize]
\foreach \x in {1,...,3} \draw (0.5*\x,1) node(a\x){$\bullet$};
\foreach \x in {1,...,3} \draw (0.5*\x,0.5) node(b\x){$\bullet$};
\foreach \x in {1,2} \draw (a1)--(b\x);
\foreach \x in {1,3} \draw (a2)--(b\x);
\foreach \x in {2,3} \draw (a3)--(b\x);
\end{tikzpicture}
\end{displaymath}
\underline{Claim 2}: If $k,l\in\{1,2,3\}$ are distinct elements such that $\beta_k=\beta_l=2$ then $\#(\widehat{U}_{b_k}\cap\widehat{U}_{b_l})=1$.
Suppose that there exist distinct elements $k,l\in\{1,2,3\}$ such that $\beta_k=\beta_l=2$ and $\#(\widehat{U}_{b_k}\cap\widehat{U}_{b_l})=2$. Let $m$ be the remaining element of $\{1,2,3\}$ and let $c_1$ and $c_2$ be the elements of $\widehat{U}_{b_k}\cap\widehat{U}_{b_l}$. Note that $\textnormal{mxl}(X)\subseteq F_{c_1}\cap F_{c_2}$. If $\{c_1,c_2\}\subseteq \widehat{U}_{b_m}$ then $|\mathcal{K}(X-\{c_3\})|$ is homeomorphic to $S^2$ contradicting \ref{lemma_subspace_with_non_trivial_H2}. Thus $\{c_1,c_2\}\nsubseteq \widehat{U}_{b_m}$. Hence $\beta_m=2$ and $c_3\in U_{b_m}$. If either $c_3<b_k$ or $c_3<b_l$ then $\textnormal{mxl}(X)\subseteq F_{c_3}$. Otherwise, since $c_3$ is not a beat point of $X$ we obtain that $F_{b_m}\subsetneq\widehat{F}_{c_3}\subseteq \{b_m\}\cup\textnormal{mnl}(X)$. As $\# F_{b_m}=3$, it follows that $\widehat{F}_{c_3}=\{b_m\}\cup\textnormal{mnl}(X)$ and hence $\textnormal{mxl}(X)\subseteq F_{c_3}$. Thus $\textnormal{mxl}(X)\subseteq F_{c_3}$ in any case.
Hence $\mathcal{S}=\textnormal{mnl}(X)\times\textnormal{mxl}(X)$. Thus
\begin{displaymath}
\chi(X)=6-\#\mathcal{S}+\sum_{j=1}^3 (\alpha_j-1)(\beta_j-1)=6-9+3=0
\end{displaymath}
which entails a contradiction. This proves claim 2.
Now, let $a_1$, $a_2$ and $a_3$ be the only elements of $\widehat{F}_{b_1}\cap\widehat{F}_{b_2}$, $\widehat{F}_{b_1}\cap \widehat{F}_{b_3}$ and $\widehat{F}_{b_2}\cap \widehat{F}_{b_3}$ respectively. Without loss of generality we may assume that $\beta_1=\beta_3=2$. Thus $\#(\widehat{U}_{b_1}\cap\widehat{U}_{b_3})=1$ by claim 2. Let $c_1$, $c_2$ and $c_3$ be the only elements of $\widehat{U}_{b_1}-\widehat{U}_{b_3}$, $\widehat{U}_{b_1}\cap \widehat{U}_{b_3}$ and $\widehat{U}_{b_3}-\widehat{U}_{b_1}$ respectively.
If $\beta_2=2$ then $\#(\widehat{U}_{b_2}\cap\widehat{U}_{b_1})=1$ and $\#(\widehat{U}_{b_2}\cap\widehat{U}_{b_3})=1$ by claim 2. Then $X$ is homeomorphic to the following space
\begin{displaymath}
\begin{tikzpicture}[x=3cm,y=3cm]
\tikzstyle{every node}=[font=\footnotesize]
\foreach \x in {1,...,3} \draw (0.5*\x,1) node(a\x){$\bullet$};
\foreach \x in {1,...,3} \draw (0.5*\x,0.5) node(b\x){$\bullet$};
\foreach \x in {1,...,3} \draw (0.5*\x,0) node(c\x){$\bullet$};
\foreach \x in {1,2} \draw (a1)--(b\x);
\foreach \x in {1,3} \draw (a2)--(b\x);
\foreach \x in {2,3} \draw (a3)--(b\x);
\foreach \x in {1,2} \draw (c1)--(b\x);
\foreach \x in {1,3} \draw (c2)--(b\x);
\foreach \x in {2,3} \draw (c3)--(b\x);
\end{tikzpicture}
\end{displaymath}
Hence $|\mathcal{K}(X)|$ is homotopy equivalent to $S^1$ and then $X$ is not homotopically trivial.
Thus $\beta_2=3$ and hence $X$ is homeomorphic to the space of figure \ref{fig_ht_non-contractible}.
\end{proof}
\bibliographystyle{acm}
|
\section{Introduction}
\label{sec:intro}
Using perturbative QCD, we would like to study particle production in high-energy
proton-nucleus ($pA$) collisions in the kinematical regime where the produced
particle is {\em semi-hard to hard} (meaning that its transverse momenta can be
larger than the nuclear saturation momentum $Q_s$, but not {\em much} larger)
and it propagates at {\em forward rapidity} in the proton fragmentation region
(that is, it makes a very small angle w.r.t. the collision axis)
\cite{Kovchegov:1998bi,Kovchegov:2001sc,Dumitru:2002qt,Albacete:2003iq,Kharzeev:2003wz,Iancu:2004bx,Blaizot:2004wu,Blaizot:2004wv,Dumitru:2005gt,Albacete:2010bs,Tribedy:2011aa,Rezaeian:2012ye,Lappi:2013zma}.
What is special about this
kinematics is that the scattering probes the small-$x$ part of the nuclear wavefunction,
but the large-$x$ part of the proton wavefunction, so it acts as a clean probe
of the nuclear gluon distribution in the interesting regime where one expects
large gluon occupation numbers and strong non-linear phenomena, like gluon saturation.
This probe is `clean' since the large-$x$ part of the proton wavefunction is
very dilute and hence well described by the standard QCD parton picture
and the associated collinear factorization. Accordingly, the overall process can be
depicted as follows: a collinear parton from the proton undergoes
multiple scattering off the dense gluon system in the nuclear target and hence acquires
some transverse momentum $\kt$, before eventually fragmenting into the hadrons
that are measured in the final state.
The above physical picture naturally lends itself to a {\em hybrid
factorization} scheme \cite{Dumitru:2002qt,Dumitru:2005gt} for the calculation of the single-inclusive hadron multiplicity,
which combines the {\em collinear factorization} for the parton distribution of the
incoming proton and also for the fragmentation of the produced quark or gluon
\cite{Ellis:1991qj},
with the {\em CGC factorization} for the high-energy scattering between
the collinear parton and the nucleus. The `CGC' refers to
the Color Glass Condensate effective theory, which is the appropriate pQCD framework
to address the problem of high-energy scattering in the presence of high gluon densities
\cite{Iancu:2002xk,Iancu:2003xm,Gelis:2010nm,Kovchegov:2012mbw}.
This is essentially a theory for the gauge-invariant correlations of Wilson lines
and their evolution with increasing energy. A {\em Wilson line} (a unitary matrix in
the color group SU$(N_c)$) is the $S$-matrix of an
energetic parton which undergoes multiple scattering off a strong color field
representing the gluon distribution of the target.
The CGC factorization\footnote{The CGC factorization can be viewed as the
generalization to high gluon density of the $\kt$-factorization \cite{Catani:1990eg,Catani:1994sq},
which deals with the `unintegrated gluon distribution' and the associated BFKL evolution
\cite{Kovchegov:2012mbw}. The $\kt$-factorization
applies so long as the gluon density is moderately low and non-linear effects
like gluon saturation and multiple scattering can be still neglected.}
for `dilute-dense' scattering associates
one such a Wilson line to each of the partons partaking in the collision,
separately in the direct amplitude and the complex conjugate amplitude. Cross-sections
are obtained by averaging over all the configurations of the color fields in the target, a
procedure which generates the Wilson-line correlators aforementioned.
In the simplest case, that is for single-inclusive particle production at leading order,
this correlator involves the trace of the product of two Wilson lines\footnote{The Wilson
lines are in the fundamental representation of SU$(N_c)$ if the colliding parton is
a quark and in the adjoint representation if this parton is a gluon. Accordingly, the color dipole
is either a quark-antiquark pair, or a pair of two gluons, in an overall
color singlet state.}, which can be identified with the elastic $S$-matrix of a
{\em color dipole} which scatters off the nuclear target.
Whereas this hybrid factorization may indeed look natural, in view of the underlying physical
picture, its foundation in pQCD is not obvious, nor easy to establish. In order to make sense
beyond tree-level, this scheme must be consistent with the QCD radiative corrections
and notably with the collinear and high-energy evolutions. As we shall shortly explain,
this issue is already non-trivial at leading-order (LO) and it becomes even more so at
next-to-leading order (NLO) and beyond.
The LO version of the hybrid factorization for single-inclusive hadron production
\cite{Dumitru:2002qt,Dumitru:2005gt}
includes the LO DGLAP evolution for the parton distribution in the proton and for the parton
fragmentation in the final state. It furthermore
includes the LO B-JIMWLK evolution of the dipole $S$-matrix. The B-JIMWLK
(from Balitsky, Jalilian-Marian, Iancu, McLerran, Weigert, Leonidov and Kovner) equations
\cite{Balitsky:1995ub,JalilianMarian:1997jx,JalilianMarian:1997gr,Kovner:2000pt,Iancu:2000hn,Iancu:2001ad,Ferreiro:2001qy} form an infinite hierarchy of coupled equations which describes
the non-linear evolution of the $n$-point correlations of the Wilson lines.
(Operators with different number of Wilson lines couple under the evolution due to multiple scattering.)
This hierarchy drastically simplifies in the limit of a large number of colors $N_c\gg 1$, in which
expectation values of gauge-invariant operators factorize from each other. In that limit,
the evolution of the dipole $S$-matrix is governed by a closed non-linear equation,
known as the Balitsky-Kovchegov (BK) equation \cite{Balitsky:1995ub,Kovchegov:1999yj}.
A first subtle point, which arises already at LO, refers to the relation between the cross-section for
parton-nucleus scattering on one hand, and the dipole scattering amplitude on the other hand.
In general, cross-sections and amplitudes are different quantities
(e.g. they have different analytic properties) and it is only due to the high-energy approximations
--- notably, due to the fact that the high-energy amplitudes are
purely absorptive --- that such an identification becomes possible in the problem at hand.
Yet, the consistency between this relation and the high-energy evolution
is far from being trivial (see the discussion in \cite{Mueller:2012bn}).
So far, this has been demonstrated up to next-to-leading order
\cite{Kovchegov:2001sc,Mueller:2012bn} and there is no obvious reason
why it should remain true in higher orders.
With this in mind, we can address the calculation of single-inclusive hadron production
in $pA$ collisions at NLO. The NLO version of the DGLAP equation is known since long
(see e.g. the textbook \cite{Ellis:1991qj} for a pedagogical discussion). Recently, the BK
equation and the full B-JIMWLK hierarchy have been promoted to NLO accuracy as well
\cite{Balitsky:2008zza,Balitsky:2013fea,Kovner:2013ona}.
By itself, the NLO approximation turns out to be unstable
\cite{Avsar:2011ds,Lappi:2015fma,Iancu:2015vea},
due to the presence of large NLO corrections enhanced by transverse logarithms.
A similar difficulty was already encountered for the NLO version of the BFKL equation
(the linearized version of the BK equation valid when the scattering is week; see e.g. the textbook
\cite{Kovchegov:2012mbw}). As in that case
\cite{Kwiecinski:1997ee,Salam:1998tj,Ciafaloni:1999yw,Altarelli:1999vw,Ciafaloni:2003rd,Vera:2005jt},
resummation schemes have been devised also for the non-linear, BK and B-JIMWLK,
equations \cite{Beuf:2014uia,Iancu:2015vea,Iancu:2015joa,Hatta:2016ujq},
to restore the convergence of perturbation theory. In particular, the collinearly-improved BK equation
\cite{Iancu:2015vea,Iancu:2015joa},
which resums to all orders the double-collinear logarithms together with a subset of
the single-collinear logarithms and with the running coupling corrections, appears to be
a convenient tool for the phenomenology \cite{Iancu:2015joa,Albacete:2015xza}.
Moreover, the full NLO BK equation with collinear
improvement has recently been shown to be stable and tractable via numerical methods
\cite{Lappi:2016fmu}.
Besides the NLO evolution, a calculation of the particle production to NLO accuracy must
include an equally accurate version of the {\em impact factor}. The `impact factor' refers to
the partonic subprocess and for the present purposes can be simply defined as the
cross-section for parton-nucleus scattering in the absence of any QCD evolution.
For more clarity, from now on, we shall assume that the parton from the proton which participates
in the collision is a quark.
At LO, the impact factor is simply the cross-section for the scattering between a bare quark
and a nucleus or, equivalently, the $S$-matrix for a bare dipole.
At NLO, the wavefunction of the incoming quark (or dipole) may contain an additional gluon,
to be referred to as the `primary gluon' in what follows. This
gluon can be released in the final state (`real correction'), or not (`virtual correction'), but
in any case its emission modifies the cross-section (or the dipole amplitude) w.r.t. to its LO value.
Since the kinematics of this primary gluon is integrated over, the corresponding correction
to the impact factor is truly a one-loop effect. So far, this correction has been computed
via two different approaches, \cite{Chirilli:2011km,Chirilli:2012jd} and
respectively \cite{Altinoluk:2011qy,Altinoluk:2014eka}, with results which are quite
difficult to compare with each other, but {\em a priori} look different at NLO accuracy.
In what follows, we shall mostly refer to the NLO calculation in
Refs.~\cite{Chirilli:2011km,Chirilli:2012jd}. This is better suited for our
new developments in this paper and this is also the context in which emerged the
problem of the negativity of the cross-section \cite{Stasto:2013cha}, which attracted
our interest on this topic. But our general philosophy for attacking this problem
is perhaps closer in spirit to that in \cite{Altinoluk:2014eka}, in that it involves
no subtraction for the `rapidity divergence' (see below).
When performing the one-loop integration alluded to above,
one should keep in mind that there are regions in phase-space that have already
been included at LO, at least approximately, via the collinear and the high-energy evolutions.
In the absence of physical cutoffs, these regions would generate logarithmic divergences.
The physical cutoffs are truly needed when solving the evolution equations
(they define the boundaries of the corresponding phase-space),
but they can often be avoided when computing the NLO correction to the impact
factor. Namely, one can directly subtract the would-be
divergences by using a suitable `renormalization prescription', tuned to match the
resummation performed by the LO evolution equations. This is the strategy followed
by the authors of Refs.~\cite{Chirilli:2011km,Chirilli:2012jd}.
Specifically, Refs.~\cite{Chirilli:2011km,Chirilli:2012jd} used dimensional
regularization plus minimal subtraction to `remove' the collinear divergences.
This is a rather standard procedure in the context of the collinear factorization
and relies on the fact that the collinear divergences can be factorized from
the transverse integrations, as they refer to the renormalization
of the {\em integrated} parton distributions and fragmentation functions.
Refs.~\cite{Chirilli:2011km,Chirilli:2012jd}
furthermore proposed a `plus' prescription in order to subtract the `rapidity
divergence', i.e. the would-be divergence\footnote{This is also known as the
`soft divergence', or the `small-$x$ divergence';
Refs.~\cite{Chirilli:2011km,Chirilli:2012jd} used the variable $\xi\equiv 1-x$,
so for them the `rapidity divergence' appears in the limit $\xi\to 1$.}
at $x\to 0$, where $x$ is the longitudinal
momentum fraction of the primary gluon w.r.t. the incoming quark.
This prescription is quite common in the context of the $k_\perp$-factorization
at next-to-leading order (see e.g. \cite{Ivanov:2012iv} and references therein)
and its `non-linear' extension to the CGC factorization may look natural.
Recall however that, underlying this prescription, there is
the strong assumption that cross-sections in perturbative QCD at high-energy can
be factorized in rapidity. This assumption is highly non-trivial (and still unproven
in the general case and beyond LO accuracy) because the perturbative corrections
are truly non-local in rapidity. Accordingly, it is {\em a priori} not obvious that the small-$x$
divergence can be factorized from the high-energy evolution of the various scattering
operators --- the `dipole $S$-matrices' which describe the eikonal scattering between
the quark-gluon projectile and the target.
This being said, we shall explicitly demonstrate
in this paper that, via an appropriate reorganization of the perturbative corrections,
one can indeed obtain such a factorized expression, valid to NLO,
which involves the `plus' prescription and agrees with the
proposal in Refs.~\cite{Chirilli:2011km,Chirilli:2012jd}.
However, we shall also argue that the manipulations associated with this
reorganization --- namely, with the subtraction of the high-energy evolution
from the NLO impact factor --- involve a considerable amount of fine-tuning,
which may be dangerous in practice.
This fine-tuning might be at the origin of the negativity problem observed in explicit numerical
calculations based on the factorization scheme in \cite{Chirilli:2011km,Chirilli:2012jd}:
the cross-section for single-inclusive particle production suddenly turns negative
for transverse momenta
slightly larger than the target saturation momentum $Q_s$ \cite{Stasto:2013cha,Stasto:2014sea}.
Various proposals to circumvent this problem, by either introducing a cutoff in the
rapidity subtraction scheme \cite{Kang:2014lha,Xiao:2014uba,Ducloue:2016shw},
or via a more careful implementation of the kinematics
\cite{Stasto:2014sea,Altinoluk:2014eka,Watanabe:2015tja},
have managed to alleviate the problem (by pushing it to somewhat larger values of the
transverse momentum), but without offering a fully satisfactory solution,
at either conceptual or practical level (see also the discussion in the recent
review paper \cite{Stasto:2016wrf}). Whereas high-energy approximations are expected
to become less accurate at sufficiently large transverse momenta $k_\perp\gg Q_s$,
we find it surprising that they fail already in the transition region towards saturation
($k_\perp\gtrsim Q_s$) --- a region whose description is in fact the main focus of the CGC
effective theory \cite{Iancu:2002xk,Iancu:2003xm,Gelis:2010nm,Kovchegov:2012mbw}.
This negativity problem encouraged us to reconsider the overall calculation from a new perspective
and thus propose a new factorization scheme for the high-energy aspects of the problem. In presenting
this proposal below, we shall restrict ourselves to quark production, that is, we shall ignore other
partonic channels and also the fragmentation of the quark into hadrons in the final state. Also,
we shall omit all the NLO corrections associated with the collinear resummation, i.e. the finite
terms which remain after subtracting the collinear divergence. These terms can
be unambiguously distinguished from those referring to the high-energy factorization
\cite{Ducloue:2016shw} and can be simply added to our final results.
Our main observations and new results can be summarized as follows:
\texttt{(i)} In our opinion, the negativity problem
is most likely related to the severe fine-tuning inherent in the rapidity subtraction.
The `fine-tuning' refers to the delicate balance between the NLO corrections
which are included in the evolution of the dipole $S$-matrix and those which are subtracted via
the `plus' prescription in order to construct the NLO correction to the impact factor.
As we shall explain in detail in Sect.~\ref{sec:sub}, this subtraction amounts to a
reorganization of the perturbation theory which exploits the integral
representation for the solution to the BK equation. Any approximation in solving
this equation, as well as the subsequent approximations which are in practice needed
to derive the `plus' prescription, will lead to an imbalance between the
large, `added' and `subtracted', contributions and thus possibly to unphysical
results.
\texttt{(ii)} The `plus' prescription is actually not needed: the would-be `rapidity divergence' is truly
cut off by physical mechanisms, namely by energy conservation for the `real' corrections and
by probability conservation for the `virtual' ones. The role of the energy conservation in
constraining the longitudinal phase-space for the primary emission is in fact well appreciated
\cite{Stasto:2014sea,Altinoluk:2014eka}. This constraint has been used to cut off the
soft divergence in Ref.~\cite{Altinoluk:2014eka}
and to alleviate the negativity problem in the context of the `plus' prescription in
Ref.~\cite{Watanabe:2015tja}. The corresponding constraint on the `virtual' corrections has not been
discussed to our knowledge, so we shall devote an appendix to an explicit NLO calculation
which demonstrates this. Specifically, in App.~\ref{app:canc} we show
that the `virtual' corrections with very short lifetimes --- outside the physical range
for the `real' corrections --- mutually cancel each other. This result is in fact natural:
the `real' and `virtual' corrections must have the same support in longitudinal
phase-space, since they should combine with each other to ensure probability conservation.
\texttt{(iii)} To NLO accuracy, the calculation of the single-inclusive forward particle production in
dilute-dense collisions can be given a different factorization, cf. \eqn{NLO},
in which the small-$x$ logarithm associated with the primary gluon emission is not included
in the high-energy evolution, but is implicitly kept within the impact factor\footnote{At this
level, we use the notion of `impact factor' in a rather informal way, as a proxy for the
dilute projectile made with the incoming quark and its primary gluon. The more conventional
impact factor which is defined order-by-order in perturbation theory and involves no
high-energy evolution (i.e. no small-$x$ logarithms),
will be computed to NLO in Sect.~\ref{sec:sub}, after separating
our general result into leading-order plus next-to-leading order contributions.}.
The latter includes the incoming quark and its (not necessarily soft) primary gluon,
which together scatter off the gluon distribution in the nucleus and thus measure its high-energy
evolution. This is the same picture as for the CGC calculation of
di-hadron production \cite{Marquet:2007vb,Albacete:2010pg,Dominguez:2011wm,Iancu:2013dta}, except that the kinematics of the primary gluon is now integrated over
and one must add the `virtual' corrections.
\comment{This picture was also the starting
point of the NLO calculation of the impact factor in \cite{Chirilli:2011km,Chirilli:2012jd}.
But as compared to that previous work, we show that by carefully specifying the integration
limits in $x$ and also the ($x$-dependent) rapidity variables of the various $S$-matrices,
we can promote that calculation to a new factorization scheme, valid through NLO, in which the
high-energy evolution of the target is factorized from the quark-gluon
impact factor described above.
}
\texttt{(iv)} In the limit where the
primary gluon is soft and treated in the eikonal approximation, our general formula in
Eq.~\eqref{NLO} reduces to the integral representation
\eqref{rcBKint} of the solution to rcBK.
This is indeed the correct result for the quark multiplicity at LO.
Vice-versa, our formula may be viewed as a generalization of the LO BK evolution
in which the very first gluon emission (and that emission only) is treated beyond the
eikonal approximation. In view of that, we expect our result for the cross-section to be
positive semi-definite, albeit we have not been able to prove this explicitly.
\texttt{(v)} Our general formula \eqref{NLO} is probably too cumbersome to be
used in practice, due to the complicated structure of the transverse and longitudinal
integrations, which are entangled with each other. Fortunately though, the problem
can be considerably simplified in the interesting regime where the transverse momentum $\kt$
of the produced quark is sufficiently hard, $\kt\gtrsim Q_s$.
In that regime, the primary gluon is relatively hard as well,
with a transverse momentum $\pt\sim \kt$ (see the discussion in Sect.~\ref{sec:dijet}).
This allows us to replace $\pt\sim \kt$ within the rapidity variables in \eqn{NLO} and
thus deduce a much simpler result, \eqn{NLOkt}, which is of the same degree of difficulty
as the formulae used within previous numerical simulations
\cite{Zaslavsky:2014asa,Stasto:2013cha,Stasto:2014sea,Watanabe:2015tja,Ducloue:2016shw},
while at the same time avoiding the rapidity subtraction and the associated fine-tuning.
\texttt{(vi)} To ensure the desired NLO accuracy of the overall scheme, the high-energy
evolution of the color dipoles must be computed to NLO as well. In Sect.~\ref{sec:evol}
we complete our factorization scheme by specifying the NLO corrections associated with
the high-energy evolution. As we also explain there, the inclusion of the NLO evolution
in the problem at hand is {\em a priori} problematic, for two reasons:
(a) the evolution of a dense wavefunction, like a nucleus, is not known beyond
leading-order, and (b) the strict NLO approximation is expected to be unstable, due to large
corrections enhanced by collinear logarithms.
We provide solutions to these problems by relating the target evolution
to that of the dilute projectile, which is indeed known to
NLO accuracy \cite{Balitsky:2008zza}, including the all-order resummation of the
collinear logarithms \cite{Iancu:2015vea,Iancu:2015joa}.
This is furthermore discussed in Appendices \ref{sec:proj} and \ref{sec:NLOBK}.
\texttt{(vii)} To make contact with the formalism in
\cite{Chirilli:2011km,Chirilli:2012jd}, we consider in Sect.~\ref{sec:sub} the decomposition
of our general result \eqref{NLOkt} between NLO dipole evolution and NLO corrections
to the impact factor. This decomposition, shown in schematic notations in \eqn{toyNLO1},
relies in an essential way on the fact that the dipole $S$-matrix obeys a specific evolution
equation --- either the LO BK equation with running-coupling
(rcBK) \cite{Kovchegov:2006wf,Kovchegov:2006vj,Balitsky:2006wa}, or the NLO BK equation
\cite{Balitsky:2008zza} with collinear improvement \cite{Iancu:2015vea,Iancu:2015joa},
depending upon the desired accuracy. Indeed, the integral version of
the evolution equation is used to reshuffle the largest contribution to the cross-section,
that associated with the LO evolution.
\eqn{toyNLO1} is very similar, but not fully identical, to the factorization scheme
proposed in \cite{Chirilli:2011km,Chirilli:2012jd}, which is schematically shown
in \eqn{toyCXY}. As explained in Sect.~\ref{sec:sub},
the differences between Eqs.~\eqref{toyNLO1} and \eqref{toyCXY} are irrelevant
to NLO accuracy, but they can be nevertheless important in practice, as they
introduce an imbalance between the terms included in the dipole evolution and
those subtracted via the `plus' prescription. As already mentioned at point \texttt{(i)},
we believe that this imbalance is responsible for the problem of the negativity
of the cross-section.
To summarize, all the potential difficulties with the subtraction method can be avoided
by computing the cross-section directly from our formula \eqref{NLOkt},
which involves no subtraction at all. This formula can be evaluated with a suitable
approximation for the high-energy evolution, like rcBK or the more elaborated approximations
described in Sect.~\ref{sec:evol} and in Appendix \ref{sec:proj}.
This paper includes two major sections,
devoted to the LO and the NLO calculations respectively, and three appendices.
Sect.~\ref{sec:LO} starts with a discussion of the kinematics and of
the importance of the choice of a Lorentz frame for building a physical picture. Such a picture is
first developed in the target infinite momentum frame (in Sects.~\ref{sec:kin} and \ref{sec:dip}),
then extended to a mixed frame, where one of the evolution gluons (the `primary gluon')
is viewed as an emission by the incoming quark, whereas all the subsequent ones are
included in the evolution of the nuclear target (in Sect.~\ref{sec:TvsP}). In Sect.~\ref{sec:dijet},
we discuss the di-jet configurations which control the final state in the regime where the
produced quark has a large transverse momentum $\kt \gg Q_s$. The first subsection of
Sect.~\ref{sec:NLO} summarizes the result for the NLO impact factor
obtained in \cite{Chirilli:2011km,Chirilli:2012jd} and also extends that result
by specifying the rapidity
variables for the evolution of the various dipole $S$-matrices. This discussion motivates
our main result in this paper, that is, the NLO factorization displayed in \eqn{NLO}.
This general but rather cumbersome expression is rendered more tractable and also more explicit
in Sects.~\ref{sec:approx}, \ref{sec:evol} and in Appendix \ref{sec:proj},
where we simplify the kinematics via approximations appropriate
at large $\kt\gtrsim Q_s$ and we replace the unknown NLO evolution of the nuclear target by that
of the dilute quark-gluon projectile (with collinear improvement).
In Sect.~\ref{sec:sub}, we isolate LO from NLO contributions,
as described at point \texttt{(vii)} above. Sect.~\ref{sec:conc} contains our conclusions.
In Appendix \ref{app:canc} we demonstrate the
mutual cancellation of the `virtual' fluctuations whose lifetime is shorter than the longitudinal
extent of the target. Finally, Appendices \ref{sec:proj} and \ref{sec:NLOBK} give more
details on the NLO evolution of color dipoles.
\section{The leading order calculation}
\label{sec:LO}
In this section, we shall briefly review the leading-order (LO) calculation of
single-inclusive quark production in high-energy proton-nucleus ($pA$)
at forward rapidities (i.e. in the proton fragmentation region).
This calculation relies on a hybrid factorization scheme \cite{Dumitru:2005gt} which involves
collinear factorization at the level of the proton wavefunction together with the dipole
picture for the scattering between a collinear quark from the proton and the nuclear
gluon distribution.
\subsection{General picture and kinematics}
\label{sec:kin}
To LO in perturbative QCD and in a suitable Lorentz frame,
the forward production of a quark in $pA$ collisions
proceeds via the {\em transverse momentum broadening} of one of the quarks from the incoming
proton: the quark, which was originally collinear with the proton,
acquires a transverse momentum $\bm{k}$ via scattering off the small-$x$ gluons in the
nuclear wavefunction and thus emerges at a small angle $\theta \simeq \kt/k^+$
w.r.t. the collision axis. The typical situation
is such that the quark undergoes multiple soft scattering and thus accumulates
a transverse momentum of order $Q_s$ --- the target saturation momentum
at the longitudinal resolution probed by the scattering (see below).
But the $\kt$-distribution of the produced quark also features a power-like
tail at high momenta $\kt\gg Q_s$, which is the result of a single, relatively hard,
Coulomb scattering off the color sources in the target.
The physical picture actually depends upon the choice of a Lorentz frame.
The picture that we have just described only holds in a `target infinite momentum frame',
where the nuclear target carries most of the total energy, so the high-energy
evolution via the successive emissions of soft gluons is fully encoded in the nuclear
gluon distribution. On the other hand, the picture would be different in a frame where
the projectile proton carries most of the total energy; in that case, the wavefunction
of the incoming quark is highly evolved, in the sense that it contains many soft gluons,
which can be put on-shell by their scattering off the (un-evolved) nucleus.
The final transverse momentum $\bm{k}$ acquired by the quark is then
the result of the recoil from this induced gluon radiation.
To transform these pictures into actual calculations, we need to better
specify the kinematics. We work in a Lorentz frame where the proton is a right mover,
with longitudinal momentum $Q^+$, while the nuclear target is a left mover, with longitudinal
momentum $P^-$ per nucleon. The high-energy regime corresponds to the situation where the
center-of-mass energy $\sqrt{s}$, with $s=2Q^+P^-$, is much larger than any of the transverse
momentum (or virtuality) scales in the problem; in particular, $s\gg k_\perp^2$ and
$s\gg Q_s^2$. For our purposes, the longitudinal momentum $k^+$ of the produced quark is
most conveniently parametrized in terms of the boost-invariant ratio $x_p\equiv k^+/Q^+\le 1$
(the quark longitudinal momentum fraction w.r.t. the incoming proton).
Let us start by choosing a {\em target infinite-momentum frame}, where the scattering
involves a bare quark from the proton and the highly-evolved gluon distribution of the nucleus.
Prior to the collision, the quark has only a `plus' momentum $q^+_0$.
After the scattering, which can involve
one or several gluon exchanges with color sources from the target,
the quark emerges with the same longitudinal momentum, $k^+= q^+_0$
(since gluons from the target have negligible `plus' momenta), but
it acquires a transverse momentum $\bm{k}$
and also a `minus' component $k^-$, which is needed for the produced quark to be
on-shell: $k^-=\kt^2/2k^+$. This condition fixes the total longitudinal momentum
fraction carried by the gluons from the target that were involved in the collision\footnote{Clearly,
in the case of multiple scattering, the light-cone energy $q^-$ which is individually carried by
the exchanged gluons can be smaller than this overall value $k^-=\kt^2/2k^+$, as emphasized
in Ref.~\cite{Altinoluk:2014eka}. However, we disagree with the conclusion there that the
longitudinal fraction $X_g$ which counts for the target evolution can be {\em
parametrically} different from the estimate \eqref{Xg} (and in particular independent of the
quark transverse momentum $\kt$). Indeed, the relevant value of $X$ is the one
which controls the energy dependence of the target saturation momentum $Q_s(X)$.
As well known, the latter is fully determined by the condition that the amplitude
for a {\em single} scattering become of order one \cite{Iancu:2002tr,Mueller:2002zm}.
} :
\begin{equation}\label{Xg}
X_g= \,\frac{k^-}{P^-}\,=\,\frac{\kt^2}{2k^+P^-}\,=\,\frac{\kt^2}{x_p s}
\,\equiv\,\frac{\kt^2}{\hat s}\,.\end{equation}
To relate to the experimental situation, it is customary to express the longitudinal fractions
$x_p$ and $X_g$ in terms of the
rapidity $\eta\equiv (1/2)\ln(k^+/k^-)$ of the produced quark in the center-of-mass frame
(where $Q^+=P^-= \sqrt{s/2}$). Using $x_p=k^+/Q^+$ and $k^+=(\kt/\sqrt{2}){\rm e}^\eta$, one finds
\begin{equation}
\label{COM}
x_p=\frac{\kt}{\sqrt{s}}\,{\rm e}^\eta\,,\qquad X_g=\frac{\kt}{\sqrt{s}}\,{\rm e}^{-\eta}\,.
\end{equation}
The {\em forward kinematics} corresponds to the situation where $\eta$
is positive and large. Then \eqn{COM} makes it clear that $X_g\ll x_p < 1$, thus confirming
that forward particle production explores the small-$X_g$ part of the nuclear wavefunction,
as anticipated in the Introduction.
\subsection{Dipole picture}
\label{sec:dip}
To LO in the CGC effective theory, the `quark multiplicity' (i.e. the
distribution of the produced quarks in transverse momentum $\bm{k}$ and COM rapidity $\eta$)
is computed as follows
\begin{equation}\label{LO}
\frac{{\rm d} N^{pA\to qX}}{{\rm d}^2\bm{k}\, {\rm d} \eta}\bigg|_{{\rm \scriptscriptstyle LO}}
=\, \frac{1}{(2\pi)^2}\,
x_p q(x_p)\,
{\mathcal{S}}(\bm{k},X_g)\,,
\end{equation}
where the kinematic variables $\bm{k}$, $\eta$, $x_p$, and $X_g$ have already been introduced,
$x_p q(x_p)$ is the quark distribution in the proton for a collinear quark with
longitudinal momentum fraction $x_p$,
and $\mathcal{S}(\bm{k},X_g)$ is the Fourier transform of the elastic $S$--matrix for the scattering
between a color dipole in the fundamental representation and the nucleus:
\begin{equation}\label{SF}
{\mathcal{S}}(\bm{k},X_g)=\int {\rm d}^2\bm{r}\, {\rm e}^{-{\rm i} \bm{k} \cdot \bm{r}} {S}(\bm{r},X_g)\,.
\end{equation}
The `color dipole' is a quark-antiquark pair in a color-singlet state. In the present context,
this appears as merely a mathematical representation for the cross-section for the scattering
between the produced quark and the nucleus: the `quark' component of the dipole
is the colliding quark viewed in the direct amplitude (DA) and the `antiquark' is the same physical
quark, but viewed in the complex conjugate amplitude (CCA).
To the accuracy of interest, the dipole-nucleus scattering can be computed
in the eikonal approximation, i.e. the transverse coordinates of the quark ($\bm{x}$) and
the antiquark ($\bm{y}$) can be treated as fixed during the collision. Then the only effect of the
collision are color rotations of the two fermions, as described by Wilson lines extending
along their trajectories:
\begin{equation}\label{Sdip}
S(\bm{x},\bm{y}; X)\,\equiv\,\frac{1}{N_c}\left\langle {\rm tr}
\big[U(\bm{x}) U^\dagger(\bm{y})\big] \right\rangle_{X}.
\end{equation}
Here, $U(\bm{x})$ and $U^\dagger(\bm{y})$ are Wilson lines in the fundamental representation,
e.g.,
\begin{align}\label{Udef}
U(\bm{x}) = {\rm P} \exp\left[{\rm i} g \int {\rm d} x^+ A^-_a(x^+,\bm{x}) t^a\right],
\end{align}
and $A^-_a(x)$ is (the relevant component of) the color field representing the gluons from the target
with longitudinal momentum fraction $X\equiv q^-/P^-$. In general, this field is strong
(corresponding to large gluon occupation numbers) and the path-ordered phase in \eqn{Udef}
resums multiple scattering to all orders. In the Fourier transform in \eqn{SF}, the transverse momentum
$\bm{k}$ of the produced quark is conjugated to the dipole size $\bm{r}\equiv \bm{x}-\bm{y}$. Both the l.h.s. and
the r.h.s. of \eqn{SF} depend upon the impact parameter $\bm{b}\equiv (\bm{x}+\bm{y})/2$, but this
dependence is unessential for what follows and will be omitted: for our purposes, the target
can be treated as quasi-homogeneous in the transverse plane.
The brackets in the r.h.s. of \eqn{Sdip} denote the target
average over the color field $A^-_a(x)$, as computed with the CGC weight
functional \cite{Iancu:2002xk,Iancu:2003xm,Gelis:2010nm}.
By using the JIMWLK equation for the latter, or directly the Balitsky equations for the
color dipole operator, one finds an equation for the evolution of the dipole $S$-matrix
with decreasing $X$. In general, this is just the first equation from an infinite hierarchy, but the
situation simplifies in the limit of a large number of colors $N_c\gg 1$, where the dipole $S$-matrix
obeys a closed, non-linear, equation, known as the Balitsky-Kovchegov (BK) equation \cite{Balitsky:1995ub,Kovchegov:1999yj}. This equation will be later needed, so let us display it here:
\begin{align}\label{BK}
\frac{\partial }{\partial Y} \,S(\bm{x},\bm{y}; Y)=
\frac{\bar{\alpha}_s}{2\pi}\, \int {\rm d}^2\bm{z}\,
\frac{(\bm{x}-\bm{y})^2}{(\bm{x}-\bm{z})^2(\bm{z}-\bm{y})^2}\,\Big[
S(\bm{x},\bm{z}; Y) S(\bm{z}, \bm{y}; Y)
-S(\bm{x},\bm{y}; Y)\Big]\,,
\end{align}
where $\bar{\alpha}_s\equiv\alpha_s N_c/\pi$ and
$Y\equiv \ln(1/X)$. The integration variable $\bm{z}$ in \eqn{BK} represents
the transverse coordinate of a soft gluon with longitudinal fraction $X= q^-/P^-\ll 1$,
which is emitted by `fast' color sources from the target (valence quarks and gluons
from the previous generations, with momentum fractions $X'\gg X$) and
absorbed by the projectile dipole. \eqn{BK} must be integrated from some
lower value $Y_0\equiv \ln(1/X_0)$, where one can use a low-energy model for
the nuclear gluon distribution,
up to $Y_g\equiv \ln(1/X_g)=\ln(\hat s/\kt^2)$, where we compute the quark production.
Typically, $X_0\sim 1\gg X_g$. For instance, if one
uses a valence-quark model for the nucleus, like the McLerran-Venugopalan (MV) model
\cite{McLerran:1993ni,McLerran:1993ka},
then $X_0$ must satisfy $\bar{\alpha}_s \ln(1/X_0)\ll 1$ and the whole gluon distribution
is built up via evolution.
In the above discussion, we have privileged the viewpoint of {\em target evolution},
that is, we have described \eqn{BK} as the result of a change in the gluon distribution of the target.
The complementary point of view, that of {\em projectile} evolution, will be useful too for what
follows and will be introduced in the next subsection.
Also, we implicitly assumed that the relation \eqref{LO} between
the quark transverse momentum broadening and the dipole $S$-matrix remains valid in the
presence of the high-energy evolution; that is, both sides of \eqn{LO} evolve in exactly
the same way with increasing energy (i.e. with increasing $\eta$,
or decreasing $X_g$). This is known to be true,
at least, up to NLO accuracy, as demonstrated in Ref.~\cite{Kovchegov:2001sc} for the
LO BK evolution and in Ref.~\cite{Mueller:2012bn} for the NLO one. However, \eqn{LO}
is not complete beyond leading order: to this `dipole' piece, one must add the
`corrections to the impact factor', that is, the contributions from
partonic configurations which do not reduce to either the dipole, or its
high-energy evolution. Such corrections will represent a main topic of the NLO discussion
in Sect.~\ref{sec:NLO}.
\subsection{Target versus projectile evolution}
\label{sec:TvsP}
Previously, we have insisted that the physical picture of the high-energy evolution
depends upon the choice of a frame, but as a matter of facts the
BK equation \eqref{BK} holds exactly as written in {\em any} frame that is obtained from
the COM frame via a boost. This includes the infinite momentum frame of the nucleus that
we have considered so far, but also the corresponding frame for the projectile, where the
incoming proton carries most of the total energy and the high-energy evolution of interest
refers to the emission of soft gluons in the wavefunction of the colliding dipole (or quark).
By `soft gluons' in this case, one means gluons which carry small fractions
$x\equiv p^+/q^+_0\ll 1$ of the longitudinal momentum $q^+_0$ of the parent quark.
This `boost invariance' of the LO BK equation is not an automatic consequence of
the underlying Lorentz symmetry of the problem --- after all, the respective evolution variables are
different: $Y=\ln(1/X)$ for the target evolution and $y\equiv \ln(1/x)$ for that of the projectile.
Rather, it reflects approximations specific to the LLA at hand, whose effect is to identify
these two variables $Y$ and $y$ to the accuracy of interest (up to a change of sign). In other
terms, at LLA, the fact of decreasing $Y$ is indeed equivalent with increasing $y$, meaning that
the evolution can be progressively transferred from the projectile to the target, and back.
This point will play an important role in our
subsequent discussion of the NLO contribution to particle production.
In preparation for that, let us briefly remind here the kinematical assumptions
underlying the LLA and thus expose their limitations.
To that aim, we consider the situation where only one of the
soft gluons has been emitted by the quark, while all the other ones belong to
the wavefunction of the target (see Fig.~\ref{QtoQG}).
The gluon that has been singled out in this way is the one to be closest
in rapidity ($y$) to the incoming quark; we shall refer to it as the {\em
primary gluon} and write its longitudinal and transverse momenta
as $p^+=xq^+_0$ and respectively $\bm{p}$.
Consider a `real' graph in which the primary gluon, albeit unmeasured, is
released in the final state\footnote{The associated `virtual' graphs are needed for the
conservation of probability, hence
one can naturally assume that they must involve the same phase-space for gluon
emission as the `real' ones. We shall later return to a more elaborate
discussion of this point, including an explicit computation of the `virtual' corrections
in App.~\ref{app:canc}.}.
Then longitudinal momentum conservation implies
$k^+=(1\!-\! x)q^+_0=x_p Q^+$ and therefore $q_0^+/Q^+=x_p/(1\!-\! x)$.
[Recall that $x_p$ is defined as the boost-invariant ratio $x_p\equiv k^+/Q^+$,
which in the COM frame takes the form shown in \eqn{COM}.]
Both the quark and the primary gluon must be on mass-shell in the final state.
Then, light-cone energy conservation implies that the scattering
off the nuclear target must transfer a total `minus' component $q^-=X(x,\pt)P^-$ with
(recall that $\hat s=x_p s$)
\begin{equation}\label{Xx}
X(x,\pt)\,=\,\frac{1}{P^-}\left(\frac{\kt^2}{2k^+}+\frac{\pt^2}{2p^+}\right)
=\frac{1}{\hat s}\left[\kt^2+\frac{1\!-\! x}{x}\, {\pt^2}\right]\,.\end{equation}
The LLA essentially relies on the two following kinematical assumptions:
\texttt{(i)} The longitudinal fraction of the emitted gluon is small: $x\ll 1$. This allows one
to simplify the calculation, notably by computing
the quark-gluon vertex in Fig.~\ref{QtoQG} in the eikonal approximation.
\texttt{(ii)} The transverse momenta of the successive emissions are parametrically
of the same order: $\kt\sim \pt$ or, more precisely, $\ln(1/x)\gg |\ln(\kt^2/\pt^2)|$. This condition
is necessary to simplify the energy denominators (by neglecting $k^- \equiv
{\kt^2}/{2k^+}$ compared to $p^-\equiv {\pt^2}/{2p^+}$) in the study of soft successive emissions
and thus obtain the LO BK (or BFKL) equation.
\begin{figure}[t] \centerline
\includegraphics[width=.55\textwidth]{QG_1step-proj.pdf}}
\caption{\small An illustration of the LO high-energy evolution of the `cross-section for quark production'
$\mathcal{S}(\bm{k},X_g)$ (the Fourier transform of the dipole $S$--matrix). The `primary gluon' (the first gluon
emission by the quark, which carries a longitudinal momentum fraction $x\ll 1$ is) viewed as a part of the wavefunction of the incoming quark (a right mover). The subsequent emissions are rather associated with
the evolution of the gluon distribution in the nuclear target (a left mover),
within the range $X(x) < X < X_0$, with $X(x)=X_g/x$ and $X_0\sim 1$. Alternatively,
and equivalently at LLA, they can be associated with the evolution of the wavefunction of
the primary quark-gluon pair (a right mover) within the `plus' momentum range $x_g < x' < x$. }
\label{QtoQG}
\end{figure}
Under these assumptions, the second term in the r.h.s. of \eqn{Xx}
(the light-cone energy of the primary gluon) dominates over
the first one and, moreover, one can ignore the difference between $\pt$ and $\kt$
when computing the evolution variables $Y=\ln(1/X)$ and $y=\ln(1/x)$.
The above argument can be immediately extended to
an arbitrary separation of the LO high-energy evolution between the quark projectile and the
nuclear target:
successive emissions in the projectile are strongly ordered in $x$ but have comparable
transverse momenta, hence both energy conservation and the energy
denominators are controlled by the last emitted gluon --- the one with the smallest value of $x$
and a transverse momentum of order $\kt$.
Accordingly, instead of \eqn{Xx}, one can use the following, simpler, relation,
\begin{equation}\label{XxLO}
X(x)\,=\,\frac{\kt^2}{x \hat s}\,=\,\frac{X_g}{x}\,,\end{equation}
(or $Y=Y_g-y$) in order to connect the LO evolution of the projectile to that of the target.
The differences between \eqref{Xx} and \eqref{XxLO} become however important
starting with NLO, as we shall see.
The above discussion can be
summarized by the following integral representation of the solution to the BK equation,
illustrated in Fig.~\ref{QtoQG},
in which the total evolution is explicitly split between exactly one soft gluon ($x\ll 1$) in the quark
wavefunction and an arbitrary number of soft gluons ($X\ll 1$) in the wavefunction of the target:
\begin{align}\label{BKint}
S\big(\bm{x},\bm{y}; X_g\big)=S_0(\bm{x},\bm{y})+
\frac{\bar{\alpha}_s}{2\pi}\, & \int_{x_g}^{1}\frac{{\rm d} x}{x}\int {\rm d}^2\bm{z}\,
\frac{(\bm{x}-\bm{y})^2}{(\bm{x}-\bm{z})^2(\bm{z}-\bm{y})^2}\nonumber\\
& \,\Big[
S\big(\bm{x},\bm{z}; X(x)\big) S\big(\bm{z}, \bm{y}; X(x)\big)
-S\big(\bm{x},\bm{y}; X(x)\big)\Big]\,.
\end{align}
In this equation, $S_0$ is the initial condition at $X_0\simeq 1$ and the lower limit
$x_g\equiv \kt^2/\hat s$ for the integral over $x$ corresponds via \eqref{XxLO} to $X=1$,
i.e. to the situation where the soft gluon from the projectile probes bare nucleons from the target.
\eqn{BKint} can also be viewed as purely target evolution provided one
changes the integration variable from $x$ to $X\equiv X(x)$. Then it becomes obvious that
this is the same as \eqn{BK} integrated over $Y$, from $Y_0=0$ up to $Y_g$.
\comment{
Alternatively, \eqn{BKint} can be also viewed as purely projectile evolution; in that interpretation,
the rapidity argument $x_g/x$ of the dipole $S$-matrices is recognized as the ratio of two
`plus' momenta: $x_g/x=p^+_g/p^+$
if one rewrites the rapidity argument of the dipole $S$-matrices
in the r.h.s as $X(x)\to x_g/x$ (recall that $X_g=x_g$) and one observes that e.g.
$S(\bm{x},\bm{z}; x_g/x)$ describes the evolution of the wavefunction of the right-moving dipole $(\bm{x},\,\bm{z})$
from a longitudinal momentum scale $xq^+_0$ (for the valence $q\bar q$ pair) down to
$x_gq^+_0$ (for the softest gluons that matter for the scattering).
}
\begin{figure*}
\begin{center}
\begin{minipage}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=0.9\textwidth,angle=0]{real1}\\(a)\vspace{0.5cm}
\end{center}
\end{minipage}
\begin{minipage}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=0.9\textwidth,angle=0]{real2}\\(b)\vspace{0.5cm}
\end{center}
\end{minipage}
\begin{minipage}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=0.92\textwidth,angle=0]{virtual2}\\(c)\vspace{0cm}
\end{center}
\end{minipage}
\begin{minipage}[b]{0.49\textwidth}
\begin{center}
\includegraphics[width=0.92\textwidth,angle=0]{virtual1}\\(d)\vspace{0cm}
\end{center}
\end{minipage}
\end{center}
\caption{\label{fig:shocks} \small
Typical diagrams contributing to the high-energy evolution of the quark production at leading order. (a) Real diagram, in which all possible interactions of the gluon with the target cancel one another. (b) Real diagram in which the gluon in the CCA is emitted before the collision with the shockwave, while in the DA is emitted after the collision. (c,d) Virtual diagrams in which the gluon interacts, or not, with the shockwave. Only transverse momenta are shown in all the graphs.}
\end{figure*}
In Fig.~\ref{fig:shocks}, we present some Feynman graphs which contribute to the integral
term in \eqn{BKint}.
We do not use the dipole picture, rather we show graphs which enter the cross-section for quark production (so, in particular, we use the transverse momentum representation).
The nuclear target evolved up to $X(x)$ is represented as a shockwave (recall that we are still
in a frame where the target is ultrarelativistic). There are two types of graphs: `real', where
the primary gluon appears in the final state --- it is emitted in the direct amplitude (DA) and
reabsorbed in the complex conjugate amplitude (CCA) --- and `virtual', where the gluon
is both emitted and reabsorbed on the same side of the cut (either in the DA, or in the CCA).
It is instructive to notice how such graphs are generated from the BK equation
\eqref{BKint}: decomposing the dipole kernel there as
\begin{equation}\label{dipole}
{\mathcal M}_{\bm{x}\bm{y}\bm{z}}\equiv
\frac{(\bm{x}-\bm{y})^2}{(\bm{x}-\bm{z})^2(\bm{y}-\bm{z})^2}=\frac{1}{(\bm{x}-\bm{z})^2} +
\frac{1}{(\bm{y}-\bm{z})^2} - 2\,\frac{x^i-z^i}{(\bm{x}-\bm{z})^2}\frac{y^i-z^i}{(\bm{y}-\bm{z})^2}\,,
\end{equation}
one can check that the `virtual' terms are generated by the first 2 terms in the r.h.s.
of \eqn{dipole}, whereas the `real' terms come from the third one. Note that, for the particular
`real' term where the gluon crosses the shockwave twice, cf. Fig.~\ref{fig:shocks}.a, the gluon
interactions cancel between the DA and the CCA, by unitarity, hence the respective
contribution to the r.h.s. of \eqn{BKint} involves only the scattering of the original
quark (as described by the dipole $S$-matrix $S\big(\bm{x},\bm{y}; X(x)\big)$).
Given the `boost-invariance' of the (LO) BK equation alluded to above, one may wonder
what can be the utility of dividing the evolution between target and projectile, as we
did above. As a matter of facts, there are several advantages for doing that. First, one
should keep in mind that the laboratory frame for `dilute-dense' (d+Au or p+Pb) collisions
at RHIC and the LHC coincides with the COM frame (at RHIC), or is close to it
(at the LHC). Hence the picture of the high-energy evolution which is directly visible
in the experiments is that of an evolution shared by the two incoming hadrons. Second,
we shall shortly argue that the first gluon emission by the incoming quark (the `primary
gluon') plays in fact a special role, at least for relatively large $\kt \gtrsim Q_s$. Because
of that, it is preferable (and even compulsory, starting with NLO) to view this gluon
as a part of the quark evolution, like in \eqn{BKint}. Still beyond LO, it is
conceptually simpler to associate the high-energy evolution with the target wavefunction.
As we shall see, the complete result for the quark multiplicity to NLO, to be
presented in Sect.~\ref{sec:NLO}, can be viewed as a natural generalization of \eqn{BKint}.
\subsection{Hard transverse momentum and di-jet events}
\label{sec:dijet}
In this subsection, we shall discuss the physical picture of forward quark production
in the IMF of the projectile, or, more generally, in any `mixed' frame, like
that illustrated in Fig.~\ref{QtoQG}, where the quark wavefunction contains at least
one soft gluon. We would like to show that, in any such a frame, the tail of the quark distribution
at relatively high $\kt$ comes from the recoil in the emission of the {\em primary gluon}
(the first gluon emitted by the quark). That is, a forward quark with large transverse momentum
$\kt \gtrsim Q_s$ is produced in a {\em di-jet event} where the quark is accompanied by a recoil
gluon and the two particles propagate back-to-back in the transverse plane.
This point is important in that it will affect the NLO
calculation of the quark production at relatively high $\kt$, where the negativity problem in
the cross-section has been observed.
At a first sight, the prominence of the di-jet configuration
at large $\kt$ might look rather obvious, as an immediate
consequence of transverse momentum conservation at the emission vertex. But the situation
is a bit more subtle, since a large transverse momentum can also be transferred by the target,
via a sufficiently hard scattering. As a matter of facts, in the classical approximation at
low energy (i.e. in the absence of any evolution), a power-like tail $\propto 1/\kt^4$
in the quark distribution at high $\kt$ is generated via Coulomb scattering (see below).
The same physical picture would also hold at high energy (within the limits of the LLA),
but only in the target IMF, where there is no gluon emission by the quark.
But in a frame where the quark itself is allowed to
radiate, the phase-space for high-energy evolution at high $\kt$ favors configurations
where the momentum $\lt$ transferred from the target to the projectile (the quark together
with its small-$x$ radiation) is relatively low, $\lt\ll\kt$.
Because of that, the only way to produce a quark with very large $\kt$ is via a di-jet event,
as anticipated.
Consider first the semi-classical approximation (no evolution), that we shall treat within the
MV model. Since we are interested in a relatively hard quark with $\kt\gg Q_s$, we can limit
ourselves to the single-scattering approximation, as obtained by expanding the Wilson
lines in \eqn{Sdip} up to second order in the target color fields $A^-$ (see Fig.~\ref{Dipole_2g}). Writing
$S=1-T$, one finds the dipole scattering amplitude in the 2-gluon exchange approximation as
(recall that $\bm{r}=\bm{x}-\bm{y}$)
\begin{align}\label{T2g}
\hspace*{-1.5cm}
T_{0}(\bm{r})&=
\frac{g^2}{2N_c} \left\langle\big(A^-_a(\bm{x})-
A^-_a(\bm{y})\big)^2\right\rangle\nonumber\\
&=g^2C_F\!
\int\frac{{\rm d}^2\bm{q}}{(2\pi)^2}
\frac{\mu^2}
{q_\perp^4}\big[1-{\rm e}^{i\bm{q}\cdot\bm{r}}\big]
\simeq\frac{\alpha_sC_F}{4}\,r^2\mu^2\ln\frac{1}{r^2\Lambda^2}\,,
\end{align}
where in the second line
we have used the MV model expression for the 2-point correlator of the color fields
in a dense nucleus (with atomic number $A$ and transverse area $\pi R_A^2$), namely
\begin{align}\label{correlmed}
\int{\rm d} x^+{\rm d} y^+
\left\langle A^-_a(x^+,\bm{q})\,A^{-}_b(y^+,\bm{\ell})\right\rangle_{{\rm \scriptscriptstyle MV}}=
(2\pi)^2\delta^{(2)}(\bm{q}+\bm{\ell})\delta_{ab}\,\frac{\mu^2}{\bm{q}^4}\,,\qquad
\mu^2= \frac{g^2C_F AN_c}{(N_c^2-1)\pi R_A^2}\,.
\end{align}
The quantity $\mu^2$ represents the color charge squared of the $AN_c$ valence quarks
(treated as uncorrelated color sources) per unit transverse area.
The variable $\bm{q}$ that is integrated over in \eqn{T2g} is the transverse
momentum transferred from the target to the dipole and $\Lambda$ is an infrared cutoff
(say, the confinement scale).
The unit term within the square brackets corresponds to the case where the
two exchanged gluons are attached to a same quark leg within the dipole, while the
exponential ${\rm e}^{i\bm{q}\cdot\bm{r}}$ refers to attachments to both legs (see Fig.~\ref{Dipole_2g}).
For relatively small dipole sizes $r\ll 1/\Lambda$, the integral over $\qt$ develops a transverse
logarithm which can be isolated by expanding out the exponential to second order. This yields
the final result shown in \eqn{T2g}. When this result becomes of $\order{1}$, multiple scattering
becomes important and the above approximation breaks down.
This condition defines the target saturation momentum $Q_0$ at low energy:
$T_{0}(r)\sim 1$ for $r\simeq 1/Q_0$.
\begin{figure}[t] \centerline
\includegraphics[width=.7\textwidth]{Dipole_2g_exchange}}
\caption{\small Diagrams for the elastic scattering of the dipole in the single scattering
approximation, or 2 gluon exchange. The blob at the bottom of the diagram refers to the
average over the color fields in the target, which effectively generates the gluon
distribution on the resolution scale $r$ of the dipole projectile.}
\label{Dipole_2g}
\end{figure}
This simple calculation makes it clear that the scattering of a small dipole ($1/r\gg Q_0$) is
controlled by relatively soft gluon exchanges ($\Lambda \ll \qt\ll 1/r$) with the target.
Let us similarly compute the
quark production, for a quark with large transverse momentum $\kt \gg Q_0$.
When taking the Fourier transform of $T_{0}(\bm{r})$, the unit term within the square brackets in
\eqn{T2g} does not matter (this would describe an elastic scattering without net momentum
transfer; see Fig.~\ref{Quark_2g}), whereas the exponential term there selects $\bm{q}=\bm{k}$. This is simply
the expression of momentum conservation and confirms that one needs a hard (inelastic)
scattering in order to produce a hight-$\kt$ quark. One thus finds\footnote{Notice that the
Fourier transform of the dipole scattering amplitude is defined with a minus sign, $T(\bm{r}) \to
-\mathcal{T}(\bm{k})$, in such a way that $\mathcal{T}(\bm{k})$ has the same sign as $\mathcal{S}(\bm{k})$ ; one therefore has
$\mathcal{S}(\bm{k}) = (2\pi)^2 \delta^{(2)}(\bm{k}) + \mathcal{T}(\bm{k})$.}
$\mathcal{T}_0(\kt)=g^2C_F\mu^2/\kt^4$,
which is recognized as the Rutherford cross-section for the Coulomb scattering between the quark
and the nucleus; therefore,
\begin{equation}\label{MV}
\frac{{\rm d} N^{pA\to qX}}{{\rm d}^2\bm{k}\, {\rm d} \eta}\bigg|_{{\rm \scriptscriptstyle MV}}
\simeq \, x_p q(x_p)\,
\frac{\alpha_s C_F \mu^2}{\pi\kt^4}\,\qquad\mbox{for}\quad \kt\gg Q_0\,.
\end{equation}
We shall now study the high-energy evolution of the above results, in the double logarithmic
approximation (DLA) which is appropriate for sufficiently small dipole sizes, or large $\kt$. For the
present purposes, it is convenient to work in a frame where this is viewed as {\em projectile} evolution;
that is, the soft gluons belong to the wavefunction of the quark and they are all right movers.
\begin{figure}[t] \centerline
\includegraphics[width=.8\textwidth]{Quark_2g-ex}}
\caption{\small Diagrams for quark production in the 2-gluon exchange approximation.
The diagram on the left describes an
elastic scattering in the DA and no scattering in the CCA; hence it contributes
to quark production only for $\bm{k}=0$. The diagram on the right describes an inelastic
scattering in the DA and another one in the CCA. The final momentum $\bm{k}$ of the
produced quark is transferred by the target.
Momenta are flowing from left to right both in the DA and in the CCA
and from bottom to top in the exchange (red) gluons.}
\label{Quark_2g}
\end{figure}
In transverse coordinate space, the DLA corresponds to the splitting of the original dipole
$(\bm{x}, \,\bm{y})$ into two daughter dipoles, $(\bm{x}, \,\bm{z})$ and $(\bm{z}, \,\bm{y})$, whose transverse
sizes are much larger, but still small enough to undergo only single scattering:
$r \ll \bar z \ll 1/Q_0$, with $\bar z\equiv |\bm{x}-\bm{z}| \simeq |\bm{z}-\bm{y}|$. The respective evolution
equation is obtained from the general BK equation \eqref{BK} by \texttt{(i)} linearizing w.r.t. $T=1-S$
(by itself, this step yields the BFKL equation), then \texttt{(ii)} approximating the dipole kernel as
${\mathcal M}_{\bm{x}\bm{y}\bm{z}}\simeq r^2/\bar z^4$, and
\texttt{(iii)} keeping only the scattering amplitudes $T(\bm{x},\bm{z})+T(\bm{z},\bm{y})\simeq 2T(\bar z)$
for the two daughter dipoles, whose scattering is stronger (since $T(r)\propto r^2$
in this physical regime). One thus finds (as before, we use $y=\ln(1/x)$ for the evolution
`time' of the projectile)
\begin{align}\label{DLA}
\frac{\partial }{\partial y} \,T(r, y)=
{\bar{\alpha}_s}\,r^2 \int_{r^2}^{1/Q_0^2} \frac{{\rm d}\bar z^2}{\bar z^4}\,
T(\bar z, y)\,.
\end{align}
Since $T(\bar z)\propto \bar z^2$, the integral in the r.h.s. is clearly logarithmic.
The first iteration of this equation, as obtained by evaluating its r.h.s. with the
amplitude $T_{0}$ from \eqn{T2g}, describes the first gluon emission by the parent dipole.
The physical picture of this emission follows from the previous discussion:
the original dipole with size $r$ emits a relatively soft gluon with transverse momentum $\pt\sim 1/\bar z$
within the range $Q_0\ll \pt\ll 1/r$, which then suffers an even softer scattering off the nuclear
target, with transferred momentum $\Lambda\lesssim \qt\ll \pt$ (see Fig.~\ref{fig:DLA} left). This picture extends
to the whole gluon cascade generated by iterating \eqn{DLA}:
successive gluon emissions are strongly ordered not only in $x$ but also in
transverse momenta, and the final exchange with the target is even softer.
\begin{figure}[t]
\begin{minipage}[b]{0.49\textwidth}
\begin{center}
\includegraphics[scale=0.51]{Dipole_1step_DLA}
\end{center}
\end{minipage}
\begin{minipage}[b]{0.49\textwidth}
\begin{center}
\includegraphics[scale=0.53]{QG_1step_DLA}
\end{center}
\end{minipage}
\caption{\small Left: one step in the DLA evolution of a small dipole, with size $r\ll 1/Q_s$.
The daughter gluon is
typical soft and thus emitted at a large distance $|\bm{z}-\bm{x}| \simeq |\bm{z}-\bm{y}| \gg r$ from the
parent dipole. The gluon exchange $\bm{q}$ with the nuclear target is even softer.
Right: one step
in the DLA evolution of the cross-section for quark production, at large transverse
momentum $\kt\gg Q_s$.
The primary gluon is as hard as the produced quark and they are both much harder
than the gluon exchanged with the target: $\kt\simeq\pt \gg \qt=|\bm{k}+\bm{p}|$.
The other diagrams contributing to this process at the level of the amplitude
are shown in Fig.~\ref{Recoil}.}
\label{fig:DLA}
\end{figure}
We now turn to the corresponding picture in transverse momentum space, that is,
to the problem of quark production (see Fig.~\ref{fig:DLA} right and also Fig.~\ref{Recoil}).
The momentum-space DLA equation reads
\begin{align}\label{DLAk}
\frac{\partial }{\partial y} \,\mathcal{T}(\kt, y)=
\frac{\bar{\alpha}_s}{\kt^4} \int_{Q_0^2}^{\kt^2} {\rm d}\qt^2\,\qt^2 \mathcal{T}(\qt, y)\,,
\end{align}
where the integral in the r.h.s. is indeed logarithmic, since $\mathcal{T}(\qt, y)\propto 1/\qt^4$.
Within this integral, $\mathcal{T}(\qt, y)$ should be interpreted as the cross-section for a single scattering,
with transferred momentum $\qt$, between partons in the quark wavefunction
and the target. The factor ${\bar{\alpha}_s}/{\kt^4}$ in front of the integral does not represent
anymore a $t$-channel exchange with the target, as in \eqn{MV}, but rather it comes
from the propagator of the intermediate quark, or gluon, in the $s$-channel
(see Fig.~\ref{Recoil}). Hence, the physical picture of the first emission is now as follows:
the original quark with zero transverse momentum emits a gluon
with momentum $\bm{p}$ and turns into a final quark with momentum $\bm{k}$,
while at the same time receiving a momentum transfer $\bm{q}$ from the target
(via a scattering that can occur either before, or after the splitting). Transverse
momentum conservation requires $\bm{q}=\bm{k}+\bm{p}$. But the overall cross-section, as described
by \eqn{DLAk}, favors {\em soft} scattering, with transferred momenta $\qt\ll \kt$. Accordingly,
the first emitted gluon must be hard, $\pt\simeq \kt$, to balance the momentum
of the produced quark.
\begin{figure}[t] \centerline
\includegraphics[width=.96\textwidth]{Recoil_QG_highkT}}
\caption{\small The 3 diagrams which contribute to the production of a quark-gluon pair
in the final state in the regime where both the quark and the gluon are relatively hard
($\kt\simeq\pt \gg \qt=|\bm{k}+\bm{p}|$).}
\label{Recoil}
\end{figure}
As for the subsequent gluon emissions, starting with the second one,
they follow the standard DLA ordering, in both $x$ and $\pt$, as in the respective calculation
in coordinate space, cf. \eqn{DLA}. This argument too shows that, when computing particle
production, it is quite natural to associate the primary gluon with the wavefunction of the produced
particle, whereas the other gluons are more conveniently included in the gluon distribution of
the target, as measured by the hard splitting process. While natural already at LLA, this viewpoint
becomes almost unavoidable when moving to the next-to-leading order calculation, where
the primary gluon is also allowed to have a large longitudinal momentum $p^+\sim q^+_0$.
The NLO calculation will be discussed in the next section.
\section{Next-to-leading order}
\label{sec:NLO}
In order to move on to next-to-leading order (NLO) accuracy,
one must relax some of the previous approximations and add new contributions
which start at NLO. By inspection of the LO result \eqref{LO}, it is clear that one
ingredient required in that sense is the NLO version of the B-JIMWLK (or BK) equations
\cite{Balitsky:2008zza,Balitsky:2013fea,Kovner:2013ona}, together with their all-order
`collinear' resummations \cite{Beuf:2014uia,Iancu:2015vea,Iancu:2015joa,Hatta:2016ujq}.
This in particular means that some gluon emissions must
be computed beyond the eikonal approximation: besides the effect of order $\alpha_s Y$,
which dominates at high energy, one must also keep, for each such an emission,
the `pure-$\alpha_s$' corrections which are not enhanced by the rapidity logarithm
$Y$ (but may be accompanied by transverse logarithms). So long as these
NLO corrections refer to generic gluons inside the cascade, they can be absorbed
into a renormalization of the kernel of the evolution equation.
The same is true for the quark-antiquark loop which at NLO can be inserted within any
of the gluon lines. But the NLO corrections associated with the `primary gluon'
(the very first emission by the leading quark) must rather be used to renormalize
the `impact factor', i.e. the value of the cross-section in the absence of high-energy evolution.
At LO, the impact factor is the cross-section for the inelastic scattering between
the leading quark and the low energy nucleus (say, as described by the MV model).
Equivalently (to the accuracy of interest), it can be written as the $S$--matrix for the
elastic scattering of a $q\bar q$ dipole. At NLO, one must add the impact factor encoding
the inelastic scattering of the quark-gluon pair made with
the leading quark and the primary gluon. Unlike the emission of the primary gluon,
which must be computed exactly, the scattering between the quark-gluon pair and the target
can still be computed in the eikonal approximation and thus related to
elastic scattering amplitudes for color multipoles
\cite{Marquet:2007vb,Dominguez:2011wm,Chirilli:2012jd}.
So, it may look like, in order to compute quark production at NLO, one must dress
the two contributions to the impact factor aforementioned with the high-energy
evolutions of the respective scattering amplitudes
(themselves computed at NLO) and then add the results. But a moment of
thinking reveals that the two pieces of the impact factor mix with each other under the
high-energy evolution: a part of the primary gluon emission that we have explicitly
included in the NLO impact factor is also included (within the limits of the eikonal
approximation) as the first small-$x$ gluon in the evolution of the dipole $S$--matrix
from the LO cross-section \eqref{LO}. This is the problem of over-counting.
Previous papers in the literature \cite{Chirilli:2011km,Chirilli:2012jd} proposed a solution
to this problem, in the form of a `plus' prescription which subtracts the LO
evolution from the NLO impact factor. This prescription however appears
to be responsible for the problem with the negativity of the cross-section
discussed in the Introduction.
In what follows, we shall propose a different way to organize the calculation,
which avoids the over-counting without performing any subtraction. Our strategy
will naturally exploit the structure of perturbation theory at high energy.
As we shall see, the contribution to the cross-section which includes the NLO correction
to the impact factor does also encode, completely and faithfully, the LO evolution
of the dipole $S$-matrix. Hence, by computing this contribution as it stands,
one can simultaneously include both effects without any ambiguity, or over-counting.
On top of that, there is a NLO correction to the evolution of the color dipole; this
will be clearly identified and related to recent results concerning the NLO
version of the BK equation \cite{Balitsky:2008zza} and its collinear resummations
\cite{Iancu:2015vea,Iancu:2015joa}.
\subsection{Revisiting the NLO calculation by Chirilli, Xiao, and Yuan}
\label{sec:CXY}
In this subsection, we shall exhibit, discuss, and adapt to our present purposes the result of the NLO
calculation of the impact factor by Chirilli, Xiao, and Yuan \cite{Chirilli:2011km,Chirilli:2012jd}.
First, we shall display their `bare', or `unsubtracted', result, where the soft
divergence\footnote{We recall that $x=p^+/q^+_0$ is the longitudinal
momentum fraction of the primary gluon relative to the incoming quark. In
Refs.~\cite{Chirilli:2011km,Chirilli:2012jd}, one has rather used the variable $\xi\equiv 1-x$,
hence our `soft divergence' at $x=0$ appears there as the `rapidity divergence' at $\xi=1$.
To facilitate the comparison, in this section we shall use both notations, $x$ and $\xi$.} at $x\to 0$ is explicit. Then we shall briefly mention the `plus' prescription advocated in
Refs.~\cite{Chirilli:2011km,Chirilli:2012jd} in order to subtract the rapidity divergence.
(We shall return to this point in Sect.~\ref{sec:sub}.) Finally, we shall explain our strategy
to deal with this problem, which is to use kinematical constraints like energy
conservation in order to cut off the soft divergence and at the same time fix the rapidity variables
for the evolution of the dipole $S$-matrices. The only subtle point here is the treatment of the
virtual corrections, where the phase-space for the emission of the primary gluon is not directly
constrained by the kinematics. Yet, as we shall demonstrate via explicit calculations
(in Appendix \ref{app:canc}), the same
lower limit on $x$ applies in that case too, albeit its emergence is now {\em dynamical}.
\begin{figure}[t] \centerline
\includegraphics[width=.55\textwidth]{QG_NLO.pdf}}
\caption{\small Pictorial representation of a typical amplitude contributing to the NLO piece of the impact factor.
This is a `real' amplitude,
in the sense that the primary gluon is released in the final state. }
\label{fig:NLO}
\end{figure}
The NLO result in Refs.~\cite{Chirilli:2011km,Chirilli:2012jd} has been obtained by evaluating
Feynman graphs like that illustrated in Fig.~\ref{fig:NLO} in which the emission
of the primary gluon is treated exactly. There is a similar graph where the
gluon emission occurs after the scattering between the quark and the target. And there
are of course virtual graphs, whose evaluation is somewhat subtle as just mentioned
and that we shall deal with in some detail. (See Figs.~\ref{fig:real} and \ref{fig:virtual}
below for more examples of Feynman graphs.)
After the scattering, both the quark and the gluon will fragment into hadrons and
thus contribute to single-inclusive hadron production. There is also
another channel where the original collinear parton is a gluon, which splits into
a pair of gluons, or into a quark-antiquark pair,
in the process of scattering. As before, we shall omit the discussion of the
fragmentation process and concentrate on quark production alone (see
Refs.~\cite{Chirilli:2011km,Chirilli:2012jd} for a complete discussion and also
\cite{Altinoluk:2014eka} for an alternative calculation, whose precise relation to the
original results in \cite{Chirilli:2011km,Chirilli:2012jd} is still unclear). That is, the primary
gluon is not measured, so one needs to integrate out its kinematics --- the longitudinal
momentum fraction $x=p^+/q_0^+$ and the transverse momentum $\bm{p}$.
The NLO result in Refs.~\cite{Chirilli:2011km,Chirilli:2012jd} can be conveniently
written as the sum of 2 pieces\footnote{\label{footFierz}
These 2 different color structures are generated when
using Fierz identities to rewrite the adjoint Wilson lines which describe the eikonal scattering
of the primary gluon in terms of fundamental Wilson lines. Accordingly, all the scattering operators
which appear in the final result are built with fundamental Wilson lines alone. At large $N_c$,
they are either linear, or bi-linear, in the dipole $S$-matrix
(see Eqs.~\eqref{Jbare} and \eqref{Jvbare} below).}:
\texttt{(A)} A piece proportional to the quark Casimir $C_{\rm F}$ which develops no logarithm
at small $x$ (the respective integrand vanishes as $x\to 0$), but has collinear divergences
in the transverse momentum integrations. In \cite{Chirilli:2011km,Chirilli:2012jd}, these divergences
have been isolated with the help of dimensional regularization and reabsorbed into the leading-order
DGLAP evolution of the quark distribution function $q(x_p)$ (if the primary emission occurs prior
to scattering) and of the quark-to-hadrons fragmentation function (if the emission occurs after
the scattering). This prescription leaves a finite remainder of NLO order whose explicit
evaluation poses no special problem.
\texttt{(B)} A piece proportional to the gluon Casimir $N_{\rm c}$ which is free of collinear problems but
develops a logarithm at small $x$ (the respective integral over $x$ exhibits a logarithmic divergence
at $x\to 0$ in the absence of any physical regulator). The proper way to deal with this
`rapidity divergence' at small $x$ represents our main concern in this paper.
To better focus on this problem while avoiding cumbersome notations,
we shall omit the piece proportional to $C_{\rm F}$ in what follows.
(This piece can be easily added to our main result shown in \eqn{NLO} below.)
As for the second piece, proportional to $N_{\rm c}$, we start by displaying the original result,
as presented in Refs.~\cite{Chirilli:2011km,Chirilli:2012jd} :
\begin{equation}
\label{eq:nlosigma}
\frac{{\rm d} N^{pA\to qX}}{{\rm d}^2\bm{k}\, {\rm d} \eta}\bigg|_{{\rm \scriptscriptstyle NLO}}^{\rm unsub}
= \frac{\alpha_s N_{\rm c}}{(2\pi)^2}
\int_{0}^1 {\rm d} \xi \,\frac{1+\xi^2}{1-\xi}\left\{
\frac{x_p}{\xi} q\left(\frac{x_p}{\xi}\right)\mathcal{J}(\bm{k} ,\xi)-
x_p q\left(x_p \right) \mathcal{J}_v(\bm{k} ,\xi) \right\},
\end{equation}
where $\xi\equiv 1-x$
and the two functions $\mathcal{J}(\bm{k} ,\xi)$
and $\mathcal{J}_v(\bm{k} ,\xi) $ correspond to real and virtual contributions
to the process illustrated in Fig.~\ref{fig:NLO}. They read
(our present notations are slightly different from the original ones
Refs.~\cite{Chirilli:2011km,Chirilli:2012jd}, but follow closely the recent paper
\cite{Ducloue:2016shw})
\begin{equation}\label{Jbare}
\mathcal{J}(\bm{k} ,\xi) =
\!\int \frac{{\rm d}^2 \bm{q}}{(2\pi)^2}
\frac{2(\bm{k}-\xi\bm{q})\cdot(\bm{k}-\bm{q})}{(\bm{k}-\xi\bm{q})^2(\bm{k}-\bm{q})^2}
\mathcal{S}(\bm{q})
-\int\! \frac{{\rm d}^2 \bm{q}}{(2\pi)^2} \frac{ {\rm d}^2\bm{\ell}}{(2\pi)^2}
\frac{2(\bm{k}-\xi\bm{q})\cdot(\bm{k}-\bm{\ell})}{(\bm{k}-\xi\bm{q})^2(\bm{k}-\bm{\ell})^2}
\mathcal{S}(\bm{q})\mathcal{S}(\bm{\ell}),
\end{equation}
and respectively
\begin{equation}\label{Jvbare}
\mathcal{J}_v(\bm{k} ,\xi) =
\mathcal{S}(\bm{k})
\left[
\int\! \frac{{\rm d}^2 \bm{q}}{(2\pi)^2}
\frac{2(\xi\bm{k}-\bm{q})\cdot(\bm{k}-\bm{q})}{(\xi\bm{k}-\bm{q})^2(\bm{k}-\bm{q})^2}
-\int \!\frac{{\rm d}^2 \bm{q}}{(2\pi)^2} \frac{ {\rm d}^2\bm{\ell}}{(2\pi)^2}
\frac{2(\xi\bm{k}-\bm{q})\cdot(\bm{\ell}-\bm{q})}{(\xi\bm{k}-\bm{q})^2(\bm{\ell}-\bm{q})^2}
\mathcal{S}(\bm{\ell})
\right].
\end{equation}
As before, the dipole $S$-matrices like $\mathcal{S}(\bm{k})$ or $\mathcal{S}(\bm{q})$ refer to dipoles in the fundamental
representation (cf. footnote \ref{footFierz}).
To simplify writing, we have considered the large $N_{\rm c}$ limit, in which the
scattering of a system of two dipoles factorizes as the product of
two individual dipole $S$-matrices, but this limit is not essential for what follows.
The variables $\bm{q}$ and $\bm{\ell}$ which appear in the above integrations
represent transverse momenta exchanged between the target and the quark-gluon pair.
For what follows, it is important to understand their precise meaning and notably
their relation with the transverse momentum $\bm{p}$ taken by the primary gluon.
By following the derivation of these results in Refs.~\cite{Chirilli:2011km,Chirilli:2012jd},
one can check
that $\bm{q}=\bm{p}+\bm{k}$ whereas $\bm{\ell}$ is independent of $\bm{p}$.
For more clarity, let us briefly discuss the physical interpretation
of the various terms in Eqs.~\eqref{Jbare} and \eqref{Jvbare}.
The `real' terms in \eqn{Jbare} represent processes where the
primary gluon, albeit not measured, is released in the final state (see Fig.~\ref{fig:real}).
For such processes, longitudinal momentum conservation implies $\xi= k^+/q^+_0$.
The first term in \eqn{Jbare}, which is linear in $\mathcal{S}(\bm{q})$, represents situations where the
hard splitting occurs either after the collision, or prior to it, in {\em both} the DA and the CCA.
In these cases, the gluon either does not interact with the target at all (emissions after the
collision), or the effects of its interaction cancel out from the final result, by unitarity, because
the gluon is not measured (emissions before the collision). Accordingly, there
is only one dipole $S$-matrix, $\mathcal{S}(\bm{q})$, which physically describes the {\em inelastic} scattering
of the quark. This scattering transfers a non-zero transverse momentum $\bm{q}$ to the quark;
then momentum conservation implies $\bm{q}=\bm{p}+\bm{k}$, as aforementioned.
\begin{figure}[t]
\begin{minipage}[b]{0.49\textwidth}
\begin{center}
\includegraphics[scale=0.53]{realsq}\\{\small (a)}
\end{center}
\end{minipage}
\begin{minipage}[b]{0.49\textwidth}
\begin{center}
\includegraphics[scale=0.53]{realsqsl}\\{\small (b)}
\end{center}
\end{minipage}
\caption{\small Production of a quark with a transverse momentum $\bm{k}$. Typical real diagrams, i.e.~diagrams in which the gluon is crossing the cut, but at the same time is integrated (cf. \eqn{Jbare}).
The primary gluon is represented as a
$q\bar q$ pair, as appropriate at large $N_c$.
(a) Left: Diagram contributing to the real term proportional to $\mathcal{S}(\bm{q})$ and which originates from $S(\bm{x},\bm{y})$ in coordinate space. All possible interactions of the gluon with the target cancel each other. (b) Right: Diagram contributing to the real term proportional to $\mathcal{S}(\bm{q}) \mathcal{S}(\bm{\ell})$ and which originates from $S(\bm{x},\bm{z}) S(\bm{z},\bm{y})$ in coordinate space. In both diagrams the target transfers momentum $\bm{q} = \bm{p} + \bm{k}$ to the final state. Momenta are flowing from left to right both in the DA and in the CCA and from bottom to top in the exchange (red) gluons.}
\label{fig:real}
\end{figure}
\begin{figure}[t]
\begin{minipage}[b]{0.49\textwidth}
\begin{center}
\includegraphics[scale=0.53]{virtualsk}\\{\small (a)}
\end{center}
\end{minipage}
\begin{minipage}[b]{0.49\textwidth}
\begin{center}
\includegraphics[scale=0.53]{virtualsksl}\\{\small (b)}
\end{center}
\end{minipage}
\caption{\small
Production of a quark with a transverse momentum $\bm{k}$. Typical virtual diagrams, i.e.~diagrams in which the gluon is not crossing the cut (cf. \eqn{Jvbare}). (a) Left: Diagram contributing to the virtual term proportional to $\mathcal{S}(\bm{k})$, which originates from $S(\bm{x},\bm{y})$ in coordinate space (b) Right: Diagram contributing to the real term proportional to $\mathcal{S}(\bm{k}) \mathcal{S}(\bm{\ell})$ and which originates from $S(\bm{x},\bm{z}) S(\bm{z},\bm{y})$ in coordinate space. Momenta are flowing from left to right both in the DA and in the CCA and from bottom to top in the exchange (red) gluons.}
\label{fig:virtual}
\end{figure}
The second term in \eqn{Jbare}, bilinear in the dipole $S$-matrix, corresponds to interference
processes, where the primary gluon is emitted prior to scattering in the
direct amplitude (DA) and after the scattering in the complex conjugate
amplitude (CCA), or vice-versa. In such processes, both the quark and the gluon can
participate in the collision. At large $N_c$, this yields 2 dipole $S$-matrices:
one made with the quark in the DA and the antiquark piece of the gluon, the other
one with the quark piece of the gluon and the antiquark in CCA. One of these $S$-matrices,
denoted as $\mathcal{S}(\bm{\ell})$ in \eqref{Jbare}, describes the {\em elastic} scattering
of a {\em physical} dipole --- i.e. a dipole whose both fermion legs exist on the same side
of the cut (either in the DA, or in the CCA). For this elastic scattering, there is no net transfer
of transverse momentum; e.g., if $\mathcal{S}(\bm{\ell})$ is computed in the 2-gluon exchange approximation,
then the momentum $\bm{\ell}$ transferred by the first exchanged gluon towards the dipole is
subsequently taken back by the second exchanged gluon. The other
dipole $S$-matrix, $\mathcal{S}(\bm{q})$, describes an inelastic scattering with net momentum transfer
$\bm{q}=\bm{p}+\bm{k}$.
Consider similarly the `virtual' contributions encoded in \eqref{Jvbare} (see Fig.~\ref{fig:virtual}).
In that case,
the primary gluon is both emitted and reabsorbed on the same side of the cut, hence
the momentum $\bm{k}$ of the produced quark fully comes via inelastic scattering (and
$k^+=q_0^+$).
In the first term in \eqref{Jvbare}, the gluon fluctuation has no overlap with the target,
hence the (inelastic) scattering refers to the quark alone. In the second term, the gluon
can scatter too. Accordingly, this term involve 2 dipole $S$-matrices, one describing
an elastic scattering ($\mathcal{S}(\bm{\ell})$), the other one an inelastic one ($\mathcal{S}(\bm{k})$).
The following observations will be useful for the subsequent arguments:
\texttt{(i)} In \eqn{eq:nlosigma} one recognizes the full LO DGLAP quark-to-quark
splitting function $P_{qq}(\xi)$, in line with the fact that the gluon emission has been
treated exactly, and not in the eikonal approximation.
\texttt{(ii)} In Eqs.~\eqref{Jbare} and \eqref{Jvbare}, the splitting fraction $\xi$ is visible only
in the various kernels describing the transverse momentum structure of the hard splitting,
which in turn have been generated by combining the light-cone energy denominator with
factors coming from the splitting vertex.
\texttt{(iii)} The various dipole $S$-matrices in Eqs.~\eqref{Jbare} and \eqref{Jvbare} are
supposed to describe scattering off the nuclear gluon distribution evolved up to the right
`rapidity' ($Y=\ln(1/X)$) scale, but this scale is left unspecified in the above equations.
For the `real' contributions at least, we know by now what is the typical longitudinal momentum
fraction $X$ of the gluons from the target which are probed by this scattering: this is the value
$X(x,\pt)$ given by \eqn{Xx}. Hence, the dipole $S$-matrices in
Eq.~\eqref{Jbare} must be evaluated at $X\simeq X(x,\pt)$,
where it is understood that $\bm{p}=\bm{q}-\bm{k}$. We shall later demonstrate that $X(x,\pt)$
with $\bm{p}=\bm{q}-\bm{k}$ is also the appropriate choice
for the rapidity argument of
$S$-matrices which enter the `virtual' terms in \eqn{Jvbare}. This means
that, strictly speaking, one cannot factorize the $S$-matrix $\mathcal{S}(\bm{k})$ in front of
the integrals in \eqn{Jvbare}, in contrast to the results in \cite{Chirilli:2011km,Chirilli:2012jd}.
\texttt{(iv)} The integral over $\xi$ in \eqn{eq:nlosigma}
seems to develop a logarithmic singularity at $\xi=1$, meaning an infrared
divergence associated with the emission of very soft ($x\to 0$) gluons. (This is the meaning
of the upper label `unsub' in the l.h.s. of \eqn{eq:nlosigma}.) As already mentioned,
Refs.~\cite{Chirilli:2011km,Chirilli:2012jd} proposed to eliminate this divergence via
the `plus' prescription, defined as (for a generic function $F(\xi)$)
\begin{equation}\label{plus}
\int_{0}^1 {\rm d} \xi \,\frac{F(\xi)}{1-\xi}\,\longrightarrow\,
\int_{0}^1 {\rm d} \xi \,\frac{F(\xi)}{(1-\xi)_+}\,\equiv\,
\int_{0}^1 {\rm d} \xi \,\frac{F(\xi)-F(1)}{1-\xi}\,.
\end{equation}
After this subtraction, the result in Eqs.~\eqref{eq:nlosigma}-- \eqref{Jvbare} is supposed
to represent a purely NLO correction, to be added to the respective LO result in \eqn{LO}.
We shall further discuss this particular prescription in Sect.~\ref{sec:sub},
but already at this level it should be clear that, as a matter of facts,
there is no physical singularity in \eqn{eq:nlosigma}: for the `real' terms at least,
the integral over $\xi$ is cut off near $\xi=1$ by energy conservation, cf. \eqn{Xx}.
Specifically, by using \eqn{Xx} together with the kinematical limit $X(x,\pt)\le 1$,
one finds the following lower limit on $x=1-\xi$:
\begin{equation}\label{xmM}
x\,\gtrsim\, x_{\rm m}(\pt) \equiv\, \frac{\pt^2}{\hat s}
\,,\end{equation}
where we also used $x_{\rm m}\ll 1$. Still for the `real' terms, there is
also an upper limit $x \le 1-x_p$, coming from the condition
$x_p/(1-x) \le 1$ on the longitudinal fraction of the incoming quark.
This lower limit $x\gtrsim x_{\rm m}$, can be recognized as the condition that the lifetime
$\Delta x^+\sim 2p^+/\pt^2$ of the softest primary
gluon emission be at least as large as the longitudinal width $1/P^-$ of the target (a
necessary condition for having significant scattering). This condition has been previously
emphasized in \cite{Altinoluk:2014eka} (the `Ioffe time') and numerically implemented in
\cite{Watanabe:2015tja} (where however the dependence of the various $S$-matrices upon
the target rapidity $X(x,\pt)$ has not been taken into account).
The existence of a physical lower limit on $x$ is indeed crucial for our subsequent
construction, which will not involve the `plus' prescription, or any other infrared regularization
of the integral over $x$. It is therefore important to demonstrate
that such a limit exists also for the `virtual' terms in \eqn{Jvbare},
for which the previous argument on energy conservation does not apply.
We shall do that in Appendix \ref{app:canc}, where we demonstrate that the same lower limit
$x\gtrsim x_{\rm m}$ holds also for the `virtual' terms, as a consequence
of fine cancellations among the virtual gluon graphs
which occur in the complementary region at $x < x_{\rm m}$.
Whereas mathematically subtle, the occurrence of such cancellations has a clear
physical interpretation: gluon fluctuations with $x<x_{\rm m}$ cannot interact with the target,
since their lifetime is too short. Accordingly the
emissions of such short-lived gluons cannot modify the $S$-matrix of the
projectile. Since `real' emissions with $x < x_{\rm m}$ are anyway forbidden by energy
conservation, it follows that the respective `virtual' graphs must cancel among themselves.
The precise way how such cancellations occur is demonstrated in Appendix \ref{app:canc}
(for the somewhat simpler, but general enough, situation where the target
itself is a quark).
\texttt{(v)} If one takes the limit $\xi\to 1$ (i.e. $x\to 0$) in the {\em transverse kernels}
in Eqs.~\eqref{Jbare} and \eqref{Jvbare} (but not necessarily also in their implicit dependence
upon $x$ via the rapidity cutoff $X(x,\pt)$), then the combination of these two terms which
enters \eqn{eq:nlosigma} with $\xi\to 1$ reduces
to the Fourier transform of the r.h.s. of the BK equation \eqn{BK}. Specifically,
\begin{equation}\label{xi1}
2\pi\bar{\alpha}_s\big[\mathcal{J}(\bm{k} ,1)- \mathcal{J}_v(\bm{k} ,1)\big]=
\int {\rm d}^2\bm{r}\, {\rm e}^{-{\rm i} \bm{k} \cdot \bm{r}}\,
\frac{\partial }{\partial Y} \,S(\bm{r}; Y)
\end{equation}
where the notation $\partial S(\bm{r}; Y)/\partial Y$ is merely a shortcut for
the r.h.s. of \eqref{BK} with $\bm{r}=\bm{x}-\bm{y}$. The appearance of the BK equation
was in fact to be expected: when $\xi\to 1$, Eqs.~\eqref{eq:nlosigma}--\eqref{Jvbare}
describe the emission of a soft primary gluon by the incoming quark,
in the eikonal approximation. By definition, the effect of this soft emission on the quark
multiplicity is equivalent to the first step in the BK evolution of the respective
LO result in \eqn{LO}. Note however that in general the $S$-matrices
implicit in the r.h.s. of \eqn{xi1} are meant to be computed {\em beyond} the LLA
and their rapidity argument $X(x,\pt)$ is a complicated function of the kinematics
of the emitted gluon.
\subsection{CGC factorization at NLO}
\label{sec:main}
After this preparation, we are now in a position to present our master formula for
the single-inclusive quark multiplicity valid through next-to-leading order
(i.e. which includes both the LO and the NLO contributions). The relation between
this formula and the factorization scheme proposed in Refs.~\cite{Chirilli:2011km,Chirilli:2012jd}
will be discussed later, in Sect.~\ref{sec:sub}. As already
mentioned, we systematically omit the NLO corrections proportional to the quark Casimir $C_{\rm F}$,
which play no special role for the high-energy evolution. These corrections
can be taken over from Refs.~\cite{Chirilli:2011km,Chirilli:2012jd} and simply added to
our master formula, which reads
\begin{align}
\label{NLO}\hspace*{-0.7cm}
\frac{{\rm d} N^{pA\to qX}}{{\rm d}^2\bm{k} \,{\rm d} \eta}\bigg|_{{\rm \scriptscriptstyle LO}+{\rm \scriptscriptstyle NLO}}\!
& = \frac{ x_p q(x_p)}{(2\pi)^2}\Big[{\mathcal{S}}_0(\bm{k})+\Delta{\mathcal{S}}(\bm{k},X_g)\Big]
+ \int\! \frac{{\rm d}^2 \bm{p}}{(2\pi)^2}
\int_{x_{\rm m}(\pt)}^{1} \!{\rm d} x \ \frac{\bar{\alpha}_s}{2\pi} \,
\frac{1+(1-x)^2}{2x}\nonumber\\*[0.2cm]
&\times\left\{
\frac{x_p}{1-x} q\left(\frac{x_p}{1-x}\right)\tilde \mathcal{J}\big(\bm{k} ,x; \bm{p}, X(x,\pt)\big)-
x_p q\left(x_p \right)\tilde \mathcal{J}_v\big(\bm{k} ,x; \bm{p}, X(x,\pt)\big) \right\}.
\end{align}
This formula is illustrated in Fig.~\ref{fig:fact}.
As before, ${\mathcal{S}}_0(\bm{k})$ denotes the tree-level contribution to the dipole $S$-matrix,
say as given by the MV model (see e.g.
\cite{Iancu:2002xk,Iancu:2003xm,Gelis:2010nm,Kovchegov:2012mbw} for an explicit expression).
The quantity $\Delta{\mathcal{S}}(\bm{k},X_g)$ within the square brackets
denotes a NLO correction to the dipole $S$-matrix, to be specified in
Sect.~\ref{sec:evol}. The last term in \eqn{NLO}, which
involves a double integration --- over the longitudinal fraction $x=1-\xi$
and the transverse momentum $\bm{p}$ of the primary gluon --- is the main term for our present purposes. It encodes the impact factor to NLO accuracy, the LO evolution of the dipole $S$-matrix
and also a part of the respective NLO contribution (namely, the part which is not included
in $\Delta{\mathcal{S}}(\bm{k},X_g)$; see Sect.~\ref{sec:evol} for details).
The new functions $\tilde\mathcal{J}\big(\bm{k} ,x; \bm{p}, X(x,\pt)\big)$
and $\tilde \mathcal{J}_v\big(\bm{k} ,x; \bm{p}, X(x,\pt)\big)$ are essentially the integrands in
Eqs.~\eqref{Jbare} and respectively \eqref{Jvbare}, in which we replaced $\xi\to 1-x$
and $\bm{q} \to \bm{p}+\bm{k}$. (As compared to \eqref{Jbare} and \eqref{Jvbare},
we now use $\bm{p}$ and $\bm{\ell}$ as integration variables; the integral over $\bm{p}$ is
explicit in \eqn{NLO}, while that over $\bm{\ell}$ is included in the definitions of
$\tilde\mathcal{J}$ and $\tilde\mathcal{J}_v$.) The notation emphasizes that the various dipole
$S$-matrices implicit in these functions should be evaluated at a target
rapidity $Y=\ln(1/X)$ with $X=X(x,\pt)$, cf. \eqn{Xx}.
The lower limit $x_{\rm m}(\pt)$ is shown in \eqn{xmM}. In the real term,
it is understood that the support of the quark distribution limits the integration
to $x < 1-x_p$.
\begin{figure}[t]
\begin{minipage}[b]{0.46\textwidth}
\begin{center}
\includegraphics[scale=0.45]{QuarkProd_MV
\end{center}
\end{minipage}\ \
\begin{minipage}[b]{0.46\textwidth}
\begin{center}
\includegraphics[scale=0.56]{QG_NLOfact2
\end{center}
\end{minipage}
\caption{\small Graphical illustration of the factorization in \eqn{NLO}.
Left: a diagram describing multiple scattering in the MV model;
this is representative for the tree-level term $\mathcal{S}_0$ in \eqn{NLO}.
Right: a diagram which exhibits the primary quark-gluon pair (the NLO impact
factor) and its multiple scattering off the gluon distribution of the target, itself
evolved to NLO; this is representative for the second term in \eqn{NLO}, which encodes all the
radiative corrections to the quark multiplicity through NLO.}
\label{fig:fact}
\end{figure}
For more clarity,
let us exhibit here the function $\tilde\mathcal{J}\big(\bm{k} ,x; \bm{p}, X(x,\pt)\big)$
which enters the `real contribution (the corresponding expression for $\tilde \mathcal{J}_v$ can
be similarly written):
\begin{align}\hspace*{-.5cm}
\label{Jreal}
\tilde\mathcal{J}\big(\bm{k} ,x; \bm{p}, X(x,\pt)\big)&=
\frac{2\bm{p} \cdot \big[(1-x)\bm{p}-x\bm{k}\big]}{\bm{p}^2 \big[(1-x)\bm{p}-x\bm{k}\big]^2}\,
\mathcal{S}\big(\bm{p}+\bm{k}, X(x,\pt)\big)\nonumber\\*[0.2cm]
&+\int\! \frac{ {\rm d}^2\bm{\ell}}{(2\pi)^2}
\frac{2(\bm{k}-\bm{\ell})\cdot \big[(1-x)\bm{p}-x\bm{k}\big]}{(\bm{k}-\bm{\ell})^2
\big[(1-x)\bm{p}-x\bm{k}\big]^2}\,
\mathcal{S}\big(\bm{p}+\bm{k}, X(x,\pt)\big)\,\mathcal{S}\big(\bm{\ell}, X(x,\pt)\big).
\end{align}
It is perhaps interesting to notice that the linear combination ${\bm P}\equiv
(1-x)\bm{p}-x\bm{k}$ which appears in the above integrand is the momentum
conjugated to the transverse separation $\bm{x}-\bm{z}$ between the quark and
the primary gluon. Similarly, the total momentum
${\bm q}\equiv\bm{k}+\bm{p}$ is conjugated to the center-of-mass $(1-x)\bm{x} + x\bm{z}$ of the quark-gluon pair.
(As in \eqn{BK}, $\bm{x}$ and $\bm{z}$ denote the transverse positions of the quark and the
primary gluon, respectively.)
To convincingly demonstrate the validity of \eqn{NLO} to the NLO accuracy of interest,
one still needs to describe the correction $\Delta{\mathcal{S}}(\bm{k},X_g)$ to the dipole $S$-matrix,
which we shall do in Sect.~\ref{sec:evol}. In the remaining part of this
subsection, we shall merely check that
\eqn{LO} properly encodes the LO result, cf. \eqn{LO}, together with
the NLO correction to the impact factor discussed
in Sect.~\ref{sec:CXY}, without any over-counting.
The LLA limit of \eqn{NLO} is obtained by making approximations appropriate
at small $x$, that is, by treating the emission of the primary gluon in the eikonal
approximation and by replacing
$\pt\sim \kt$ in the kinematical limits and the various rapidity variables;
that is, one approximates $x_{\rm m}(\pt)\simeq x_g= \kt^2/\hat s$ and $X(x,\pt)\simeq
X(x)=X_g/x$, with $X_g=x_g$ (cf. \eqn{XxLO}). Also, all dipoles $S$-matrices are now understood to
obey the LO BK evolution, from $X_0\simeq 1$ down to $X(x)$. Under these assumptions, one can
first commute the integrations over $\bm{p}$ and $x$ in \eqn{NLO} and then use the identity \eqref{xi1} to
rewrite the simplified version of this equation in the following, suggestive, form
\begin{align}
\label{NLOtoLO}
\frac{{\rm d} N^{pA\to qX}}{{\rm d}^2\bm{k} \,{\rm d} \eta}\bigg|_{{\rm \scriptscriptstyle LO}+{\rm \scriptscriptstyle NLO}}^{(0)}
& = x_p q(x_p)\,\frac{\mathcal{S}_0(\bm{k})}{(2\pi)^2}\,+\,x_p q(x_p)\,
\frac{1}{(2\pi)^2}\int_{x_g}^{1}
\frac{{\rm d} x}{x}\, \frac{\partial }{\partial Y} \,\mathcal{S}\big(\bm{k}; Y=\ln1/X(x)\big)\,.
\end{align}
The above integral runs over the longitudinal momentum fraction $x=p^+/k^+$
of the right-moving gluon, whereas the evolution of the dipole
$S$-matrix has been rather performed w.r.t. the longitudinal fraction $X$ of the gluons
in the target. However,
to LLA, $x$ and $X$ are related via $X(x)=X_g/x$, so one can
change the integration variable from $x$ to $Y\equiv \ln(1/X)$ and thus identify a total derivative
\begin{equation}
\int_{x_g}^{1}
\frac{{\rm d} x}{x}\, \frac{\partial }{\partial Y} \,\mathcal{S}\big(\bm{k}; Y=\ln1/X(x)\big)=
\int_0^{Y_g}{\rm d} Y\, \frac{\partial }{\partial Y} \,\mathcal{S}(\bm{k}; Y)\,=\,\mathcal{S}(\bm{k}; Y_g)-{\mathcal{S}}_0(\bm{k})\,.
\end{equation}
After also adding the initial condition in \eqn{NLOtoLO}, one recognizes the LO result \eqref{LO},
as anticipated.
The r.h.s. of \eqn{NLOtoLO} is recognized as the `integral' version of the LO BK equation
introduced in \eqn{BKint}.
Hence, the full result \eqref{NLO} can be viewed as the generalization of
that integral representation to NLO and to the exact kinematics for the primary
gluon emission.
(This will be confirmed by the discussion in Sect.~\ref{sec:approx}.)
The explicit separation of the first gluon emission
from the remaining evolution, as operated by this representation,
has allowed us to promote the calculation of the impact factor
to NLO accuracy, while at the same time avoiding over-counting.
At this point, it is important to more precisely specify the perturbative content of
\eqn{NLO}. As just explained, the integral term there fully encodes the LO evolution of
the dipole $S$-matrix, that is, it resums corrections of the type ${(\bar{\alpha}_s Y_g)^n}$ to all orders.
It obviously encodes NLO corrections due
to the fact that the emission of the primary gluon is treated exactly; that is, the integral
over $x$ also generates corrections of $\order{\bar{\alpha}_s}$ besides the dominant contribution
of $\order{\bar{\alpha}_s Y_g}$, which counts for the LO evolution. To ensure NLO accuracy, the
evolution of the various dipole $S$-matrices must be computed to NLO as well. Indeed,
the NLO BK kernel includes corrections of $\order{\bar{\alpha}_s}$; hence, the solution to the NLO
BK equation involves corrections proportional to ${\bar{\alpha}_s(\bar{\alpha}_s Y_g)^n}$, which count to NLO.
By a similar argument, one must include the (one-loop) running coupling corrections
within the QCD coupling $\bar{\alpha}_s$ associated with the primary gluon vertex.
This is why, in writing \eqn{NLO}, we have
inserted the factor $\alpha_s$ {\em inside} the double integral over $\bm{p}$ and $x$:
after including the running coupling corrections, this factor will depend upon the
transverse momenta $\bm{k}$ and $\bm{p}$ which enter the emission vertex
and possibly also upon $x$ (via the gluon kinematics). Specifying this dependence
requires a prescription, which is most conveniently formulated in the transverse coordinate
representation (since this is the representation in which the BK equation is generally
solved in practice). Such prescriptions will be discussed in Sect.~\ref{sec:evol}
(see e.g. \eqn{azero} there), together with the other NLO corrections to the dipole
evolution. Notice however that, in order to evaluate \eqn{NLO}, one also needs
a prescription for the running coupling which is directly formulated
in momentum space. In general such a prescription
will be different from the one in coordinate space. This mismatch could have consequences
for the fine-tinning issue to be discussed in Sect.~\ref{sec:sub}.
The above arguments show that, strictly speaking, the integral term in \eqn{NLO} also includes
terms of NNLO, as generated by the product between the NLO correction to the impact
factor and the NLO effects in the high-energy evolution, or in the running coupling.
As we shall explain in Sect.~\ref{sec:sub}, these various types of NLO effects can be disentangled
from each other via a reorganization of the perturbation theory which involves a
`rapidity subtraction', as in Refs.~\cite{Chirilli:2011km,Chirilli:2012jd}. Yet, this procedure has
some inconveniences, as we shall see (notably, it introduces the `fine-tuning' issue anticipated in
the Introduction). So it is important to stress here that, although going beyond a strict NLO
approximation, the result \eqref{NLO} is in fact the natural outcome of perturbative QCD --- that is,
the direct result of evaluating Feynman graphs at the loop-order of interest, before performing
additional manipulations like the `rapidity subtraction'.
We conclude this subsection with a discussion of potential difficulties
with using \eqn{NLO} in practice. All these issues
will be addressed in more detail in the next two subsections, where we shall provide
solutions to them, at least at the expense of further approximations.
First, \eqn{NLO} looks very cumbersome, notably due the intricacy of the
multiple integrations, over both transverse ($\bm{p}$, $\bm{\ell}$) and longitudinal ($x$) momenta,
which are entangled with each other. It is not clear to us whether these integrations
can be computed as such, not even numerically. The calculations
might be further complicated by the need to compute the Fourier transform
of the dipole $S$-matrix, as numerically obtained by solving the BK equation
in coordinate space.
Second, \eqn{NLO} is somewhat formal, in that the dipole $S$-matrices implicit
there are supposed to encode the evolution of the target gluon distribution at NLO.
However, the high-energy evolution of a dense nucleus has not been explicitly computed
beyond LO. (All NLO calculations to date refer to the evolution of a dilute projectile, like a dipole
\cite{Balitsky:2008zza,Balitsky:2013fea,Kovner:2013ona}.) Besides,
a purely NLO approximation to the high-energy evolution
is likely to become unstable (and thus require resummations)
in the `collinear' regime where the transverse momentum $\kt$ is relatively large ($\kt\gg Q_s$).
Finally, the NLO calculation based on \eqn{NLO} is formally sensitive
to the physics of the nuclear wavefunction at large values of $X$
(recall that $x\sim x_g$ corresponds to $X\sim 1$),
which is not really under control within the present, high-energy approximations.
This cannot be a serious difficulty in the physical context at hand: the NLO corrections
to the impact factor are controlled by relatively hard primary gluons
with $x\sim \order{1}$ and hence $X(x)\ll 1$; such a hard emission by the projectile
should be well separated from the valence structure of the target at $X\sim 1$. At the
end of the next subsection we shall describe an explicit procedure which implements this
separation.
\subsection{Simplifying the kinematics}
\label{sec:approx}
In this subsection, we shall propose strategies to deal with some of the problems alluded to
at the end of the previous subsection. First, we shall argue
that one can approximate $\pt\sim \kt$ within the rapidity variables for the high-energy evolution
and thus greatly simplify the structure of the transverse and longitudinal integrations in \eqn{NLO}.
Second, we shall discuss the prescription for the running of the coupling in the emission
of the primary gluon. Third, we shall reformulate the initial condition at low energy in such a
way to reduce the sensitivity to the large-$X$ region in the target wavefunction.
Concerning the first point --- the dependence of the kinematical constraints on the
high-energy evolution upon the transverse momenta
of the primary quark-gluon pair ---, we note that there are two interesting physical regimes:
\texttt{(i)} {\it Hard quark production and di-jet configurations:} $\kt\gg Q_s(X)$.
This is the regime which is primarily concerned by the negativity problem discussed in
Refs.~\cite{Stasto:2013cha,Stasto:2014sea,Kang:2014lha,Xiao:2014uba,Altinoluk:2014eka,Watanabe:2015tja,Ducloue:2016shw}. In this case, we have already argued in Sect.~\ref{sec:dijet}
that the transverse momentum of the primary gluon and that of the produced quark
must balance each other : $\pt\simeq\kt$.
\texttt{(ii)} {\it Semi-hard quark production:} $\kt\gtrsim Q_s(X)$. This regime includes the
`geometric scaling' window \cite{Stasto:2000er}, where the scattering is weak, but the $S$-matrix
is still influenced by non-linear effects, via the `saturation boundary' at $Q_s(X)$
\cite{Iancu:2002tr,Mueller:2002zm,Munier:2003vc}. In this regime, all the relevant transverse
momenta --- the $\kt$ of the produced quark, the $\pt$ of the primary gluon, and the $\qt$
transferred by the nuclear target --- take typical values of $\order{Q_s}$, since
this is the value naturally acquired via rescattering off the saturated gluon distribution of the
target.
We see that in both cases the quantities $\pt$ and $\kt$ cannot be very
different from each other, so one can approximate $\pt\simeq \kt$ when evaluating the rapidity
variables (cf. Eqs.~\eqref{Xx} and \eqref{xmM}) :
\begin{equation}\label{xmk}
x_{\rm m}(\pt)\,\to\,x_{\rm m}(\kt)=\frac{\kt^2}{\hat s}\equiv x_g\quad\mbox{and}\quad
X(x,\pt)\,\to\,X(x,\kt)=\frac{\kt^2}{x\hat s}\equiv X(x)\,.
\end{equation}
Remarkably, thanks to this approximation, we have returned to the same
expressions for the rapidity variables $x_g$ and $X(x)$ as at LO, cf. \eqn{Xx}.
This greatly simplifies \eqn{NLO} since the transverse momentum
integrations can now be performed prior to the integral over $x$. Then \eqn{NLO} takes
a form closer to that in \eqn{eq:nlosigma}, namely,
\begin{align}
\label{NLOkt
\frac{{\rm d} N^{pA\to qX}}{{\rm d}^2\bm{k} \,{\rm d} \eta}
& = \frac{ x_p q(x_p)}{(2\pi)^2}\Big[{\mathcal{S}}_0(\bm{k})+\Delta{\mathcal{S}}(\bm{k},X_g)\Big]
+\frac{\bar{\alpha}_s(\kt^2)}{2\pi} \int_{x_g}^{1} {\rm d} x \ \frac{1+(1-x)^2}{2x}\nonumber\\*[0.2cm]
&\qquad\qquad\times\left\{
\frac{x_p}{1-x} q\left(\frac{x_p}{1-x}\right) \mathcal{J}\big(\bm{k} ,x; X(x)\big)-
x_p q\left(x_p \right) \mathcal{J}_v\big(\bm{k} ,x; X(x)\big) \right\},
\end{align}
where the functions $\mathcal{J}\big(\bm{k} ,x; X(x)\big)$ and $\mathcal{J}_v\big(\bm{k} ,x; X(x)\big) $
have the same formal expressions as in Eqs.~\eqref{Jbare}--\eqref{Jvbare} [with $\xi=1\!-\! x$,
of course], except for the fact that the rapidity variable for the evolution of
the dipole $S$-matrices is now clearly identified, namely $Y=\ln \big(1/X(x)\big)$.
Once again, it is understood that the integral over $x$ in the real term
is restricted to $x < 1-x_p$.
In writing \eqn{NLOkt}, we have also identified the argument of the running coupling for the primary
vertex as $\kt^2$. This is unambiguous under the present assumptions, since $\kt\sim\pt$ is the
only hard scale involved in that splitting\footnote{Recall that the momentum
conjugated to the transverse separation between the quark and
the primary gluon is ${\bm P}=(1-x)\bm{p}-x\bm{k}$ ; when $\pt\sim\kt$, we have $P_\perp\sim \kt$
for any $x$, hence the transverse separation is of order $1/\kt$.}.
\eqn{NLOkt} also shows that the natural `LO approximation' for the high-energy evolution in
the problem at hand is the LO BK equation with running coupling (rcBK). Indeed, when evaluating
the second term in \eqn{NLOkt} within the eikonal approximation, as appropriate for $x\ll 1$,
the r.h.s. of this equation becomes proportional to the integral version of rcBK; that is,
this is tantamount to evaluating the LO formula \eqref{LO} with the solution $\mathcal{S}_{\rm \scriptscriptstyle rcBK}(\bm{k}, X_g)$
to rcBK.
The considerably simpler structure of \eqn{NLOkt} also allows us to reformulate the
initial condition at low-energy, in such a way to avoid the `dangerous' region at $X\sim 1$.
To that aim, let us introduce a `better' separation scale $X_0$ for the rapidity
evolution of the target, which is such that the high-energy approximations are indeed justified
for any $X\le X_0$. (For instance the value $X_0=0.01$
is often used in the fits to the HERA data for deep inelastic scattering; see e.g.
\cite{Albacete:2010sy,Iancu:2015joa}.) In particular, this scale should obey $x_g \ll X_0 \ll 1$.
We then separate the integral over $x$ into two regions, $x_g< x < x_{g}/X_0$
and $x_g/X_0 < x <1$, which in terms of $X=X_g/x$ correspond to $1 > X > X_0$ and respectively
$X_0 > X > X_g$. (We recall that $X_g=\kt^2/\hat s= x_g$.) Within the first region,
one has $x\ll 1$, hence one can replace
$x\to 0$ within the transverse kernels which enter the functions $\mathcal{J}$
and $\mathcal{J}_v$ [e.g., $\mathcal{J}\big(\bm{k} ,x; X(x)\big)\to \mathcal{J}\big(\bm{k} ,0; X(x)\big)$]
and also within the quark distribution.
Under these assumptions, the sum between the `initial condition' $\mathcal{S}_0$
in \eqn{NLOkt} and the part of the integral there which runs over the
small-$x$ interval at $x_g< x < x_{g}/X_0$ is formally the same as the
r.h.s. of the LO BK equation with running coupling (rcBK) integrated from $X=1$
down to $X_0$ (recall \eqn{xi1}). It might be tempting to interpret this sum
as the solution to rcBK evaluated at $X=X_0$, but we shall not adopt this point of
view: after all,
this `rcBK evolution' refers to the large-$X$ interval at $1 > X > X_0$, where
the high-energy approximations are not valid.
We shall rather {\em replace} the result of this fictitious `rcBK evolution'
with a {\em new} initial condition, denoted as $\mathcal{S}(\bm{k}, X_0)$,
which is formulated directly at $X_0$. That is, we replace
\eqn{NLOkt} by\footnote{Strictly speaking, the integral over $x$ implicit within
the NLO quantity $\Delta{\mathcal{S}}(\bm{k},X_g)$, as visible e.g. in \eqn{toyNLO2} below,
should be cutoff at $x_g/X_0$ as well when $\Delta{\mathcal{S}}(\bm{k},X_g)$ is inserted into \eqn{NLOX0}.}
\begin{align}
\label{NLOX0
\frac{{\rm d} N^{pA\to qX}}{{\rm d}^2\bm{k} \,{\rm d} \eta}
& = \frac{ x_p q(x_p)}{(2\pi)^2}\Big[{\mathcal{S}}(\bm{k},X_0)+\Delta{\mathcal{S}}(\bm{k},X_g)\Big]
+\frac{\bar{\alpha}_s(\kt^2)}{2\pi} \int_{x_g/X_0}^{1} {\rm d} x \ \frac{1+(1-x)^2}{2x}\nonumber\\*[0.2cm]
&\qquad\qquad \times\left\{
\frac{x_p}{1-x} q\left(\frac{x_p}{1-x}\right) \mathcal{J}\big(\bm{k} ,x; X(x)\big)-
x_p q\left(x_p \right) \mathcal{J}_v\big(\bm{k} ,x; X(x)\big) \right\},
\end{align}
where it is now understood that all the $S$-matrices implicit in the second term
in the r.h.s. are obtained by solving appropriate evolution equations with the initial
condition formulated at $X=X_0$. The evolution equations to be used in this
context will be discussed in the next subsection.
\subsection{The evolution of the color dipole beyond leading order}
\label{sec:evol}
In this subsection, we shall describe the NLO evolution of the dipole
$S$-matrices to be used in conjunction with the factorization scheme in \eqn{NLO} or
\eqref{NLOkt}. In particular, we shall present an explicit expression for
the NLO correction $\Delta{\mathcal{S}}$ which enters this scheme but so far has not
been specified. To simplify the arguments and the notations, we shall work
at the level of the kinematic approximations introduced in Sect.~\ref{sec:approx},
that is, we shall built on top of \eqn{NLOkt}.
There are two aspects which are rather subtle and that we shall try to clarify
in what follows. The first refers one to the proper inclusion of NLO corrections within
$\Delta{\mathcal{S}}$ without any over-counting: this quantity must contain only those
corrections to the high-energy evolution that are not already included in
the integral term in \eqn{NLOkt}. The second aspect refers to the relation between
the viewpoint of {\em target} evolution, that was explicitly used in our previous
arguments (for instance, when specifying the rapidity arguments of the various $S$-matrices),
and that of the evolution of the {\em projectile} (a color dipole),
for which the evolution equation is currently known at
NLO accuracy \cite{Balitsky:2008zza}, including the collinear improvement
\cite{Beuf:2014uia,Iancu:2015vea,Iancu:2015joa,Hatta:2016ujq}.
Indeed, the two evolutions refer to different variables (`evolution times')
--- $x=q^+/q_0^+$ for the right-moving projectile and respectively $X=q^-/P^-$
for the left-moving target ---, so the corresponding equations cannot be identical
beyond leading order, when the differences between various transverse momenta
and also the off-shell effects start to play a role.
\comment{
Our subsequent discussion will be rather schematic and not fully rigorous,
for several reasons.
First, this represents a slight detour from our main preoccupation in this paper,
which is the problem of the negativity of the NLO cross-section.
Indeed, as shown by the existing calculations in the literature and will be further
discussed in Sect.~\ref{sec:sub}, this problem already arises when using the
{\em leading-order} approximation to the high-energy evolution of the color dipoles.
There is of course the logical possibility that the negativity problem be related to such an
incomplete treatment of the NLO corrections, but this is unlikely, as we shall
argue in Sect.~\ref{sec:sub}.
Second, whereas the NLO approximation to the BK equation is explicitly known
\cite{Balitsky:2008zza}, the respective formulae are rather cumbersome
and not really needed for the present discussion. This is why, in what follows, we shall use
rather schematic notations and refer to the original literature for more explicit expressions.
Third, a direct application of the NLO version of the BK equation to the problem at hand
is hindered by at least two problems. \texttt{(i)} This equation has been obtained by
studying the high-energy evolution of the {\em dilute projectile} (the color dipole), whereas
our present factorization scheme explicitly refers to the evolution of the {\em dense target}
(the nucleus). Beyond leading order, the respective evolution equations cannot be identical
to each other, as they refer to different evolution variables: $x=q^+/q_0^+$ for the right-moving
dipole and respectively $X=q^-/P^-$ for the left-moving nucleus. So, strictly speaking,
we do not know the NLO version of the evolution equation obeyed by the $S$-matrices,
like $\mathcal{S}(\bm{q}, X)$, which enter Eq.~\eqref{NLOkt}.
\texttt{(ii)} The strict NLO approximation to the BK evolution is known
to be unstable \cite{Lappi:2015fma}, due to large NLO corrections which are
enhanced by double collinear logarithms. More precisely, such corrections appear
in the `hard-to-soft' evolution of the dilute projectile and in that context one has
devised techniques for their resummation to all orders
\cite{Beuf:2014uia,Iancu:2015vea,Iancu:2015joa,Hatta:2016ujq}.
This resummation yields a `collinearly improved' version of the BK equation,
which is both stable and suitable for the phenomenology
\cite{Iancu:2015vea,Iancu:2015joa,Albacete:2015xza,Lappi:2016fmu}.
However, it is not clear how to adapt this recent progress to the
evolution of the dense nucleus, while keeping the full NLO accuracy.
This being said, in what follows (and notably in Appendix \ref{sec:proj})
we shall provide a recipe in that sense which
is correct, at least, to the accuracy of the collinearly-improved
BK equation constructed in \cite{Iancu:2015vea,Iancu:2015joa}.
Third, the NLO version of the BK equation has been obtained by studying the high-energy
evolution of the {\em dilute projectile} (the color dipole): as it stands, this equation reflects the
evolution of the dipole wavefunction via the emission of soft gluons which carry smaller
and smaller fractions $x=q^+/q_0^+$ of the `plus' component of the light-cone momentum.
By contrast, our factorization scheme for the quark multiplicity at NLO is explicitly
formulated in terms of the high-energy evolution of the {\em dense target} (the nucleus)
--- that is, the evolution with decreasing $X=q^-/P^-$.
}
As mentioned after \eqn{NLO}, the unknown quantity $\Delta{\mathcal{S}}(\bm{k},X_g)$
represents a part of the NLO corrections to the evolution of the dipole $S$-matrix.
Since these corrections are fully known by now \cite{Balitsky:2008zza}, the simplest
way to obtain $\Delta{\mathcal{S}}(\bm{k},X_g)$ is by clarifying its relation to the NLO calculation in
Ref.~\cite{Balitsky:2008zza}. We first observe that, by definition, the full NLO
result for the quark multiplicity must involve two types of NLO corrections:
those associated with the impact factor and those related
to the high-energy evolution of the color dipoles. Hence, if one `switches off' the first type
of corrections (which one can do by computing the emission of the
primary gluon in the eikonal approximation), then the r.h.s. of Eq.~\eqref{NLOkt}
should be proportional to
${\mathcal{S}}(\bm{k}, X_g)$ --- the dipole $S$-matrix computed at NLO accuracy and for the
kinematics of interest. This argument implies
\begin{equation}\label{Sfull}
{\mathcal{S}}(\bm{k}, X_g)={\mathcal{S}}_0(\bm{k})+\Delta{\mathcal{S}}(\bm{k},X_g)
+2\pi\bar{\alpha}_s(\kt^2) \int_{x_g}^{1}\ \frac{{\rm d} x}{x}
\left\{\mathcal{J}\big(\bm{k} ,0; X(x)\big)-
\mathcal{J}_v\big(\bm{k} ,0; X(x)\big) \right\},
\end{equation}
where the function $\mathcal{J}\big(\bm{k} ,0; X(x)\big)$ is obtained from \eqn{Jbare} by letting
$x\equiv 1-\xi\to 0$ and by evaluating the $S$-matrices there to
NLO accuracy and for a rapidity argument $Y=\ln(1/X(x))$
[and similarly for the function $\mathcal{J}_v\big(\bm{k} ,0; X(x)\big)$]. Using
this condition together with the NLO result for the dipole $S$-matrix which emerges
from Ref.~\cite{Balitsky:2008zza}, it is possible to identify the quantity $\Delta{\mathcal{S}}(\bm{k},X_g)$.
By the `NLO result for ${\mathcal{S}}$', we mean the integral representation for
the NLO $S$-matrix, as obtained by formally solving the respective evolution equation
--- that is, the generalization of \eqn{BKint} to NLO.
\comment{This whole procedure may look a bit formal, for the reasons aforementioned --- the
NLO version of the BK equation is written for the projectile evolution in $x$, whereas
the $S$-matrix ${\mathcal{S}}(\bm{k}, X_g)$ depends upon the rapidity argument $X=X_g$
for the target. Yet, it should be possible to express at least the first step in this evolution
in terms of gluon emissions by the projectile. We already did so for the LO evolution
(cf. Sect.~\ref{sec:TvsP}). The only new ingredient which appears
when moving to NLO accuracy
is that the first evolution step can also involve the emission of a set of 2 gluons.
These 2 gluons have similar rapidities ($x_1\sim x_2$), but they are both soft as
compared to the parent quark ($x\equiv x_1+x_2\ll 1$).
}
Before we proceed, it is useful to, first, clarify what we precisely mean by the
`LO evolution' and, second, introduce some simplified notations, which focus on
the essential and help making the subsequent arguments more transparent.
As already mentioned after \eqn{NLOkt}, our LO approximation to the dipole
$S$-matrix is the solution $\mathcal{S}_{\rm \scriptscriptstyle rcBK}(\bm{k}, X)$ to the LO BK equation with running coupling
(rcBK) \cite{Kovchegov:2006wf,Kovchegov:2006vj,Balitsky:2006wa}. This choice deserves
some comment, in that it already includes a subset of the NLO (and
higher-order) corrections, via the running of the coupling. But precisely because of that,
rcBK offers a better starting point for a perturbative expansion than the {\em strict}
leading-order approximation --- the LO BK equation with fixed coupling. This is related
to the poor convergence of the perturbative expansion at high energy: the
LO BK equation with fixed coupling is well known to predict an unrealistically fast evolution
with increasing energy, meaning a too large value for the saturation exponent.
Hence, for sufficiently high energies, the strict LO estimate \eqref{LO} for the multiplicity
becomes {\em exponentially larger} (in the sense of an exponential in $Y=\ln(1/X)$)
than the actual result at NLO. This problem is considerably alleviated if one instead
uses rcBK as the `leading order' evolution: this approximation
predicts a significantly slower evolution \cite{Mueller:2002zm,Albacete:2007yr}
and offers a reasonably good description
for the phenomenology \cite{Albacete:2010bs,Albacete:2010sy,Albacete:2014fwa}.
The all-order resummation of running coupling corrections requires a prescription.
Here, we shall mention two popular such prescriptions, both built with
the one-loop approximation for the running coupling and which are roughly equivalent to
each other (see Ref.~\cite{Iancu:2015joa} for a recent discussion).
Consider the splitting of the parent dipole $(\bm{x},\bm{y})$ into two daughter dipoles
$(\bm{x},\bm{z})$ and $(\bm{z},\bm{y})$, as described by the LO BK equation \eqref{BK}.
The {\em smallest dipole prescription} consists in replacing\footnote{Clearly,
after such a replacement, the coupling $\bar{\alpha}_s(r_{\rm min})$ must be moved {\em inside}
the integral over $\bm{z}$ in \eqn{BK}.}
$\bar{\alpha}_s \to \bar{\alpha}_s(r_{\rm min})$, where
$r_{\rm min} \equiv \min\big\{|\bm{x} \!-\!\bm{y}|,|\bm{x} \!-\!\bm{z}|,|\bm{y} \!-\!\bm{z}|\big\}$ and
\begin{equation}\label{1Lrun}
\bar{\alpha}_s(r) =
\frac{1}{\bar b\ln\big[4/r\Lambda_{\rm QCD}^2\big]}\,,\qquad
\bar b=(11N_{\rm c}- 2 N_{\rm f})/12N_{\rm c}\,.
\end{equation}
The other prescription, known as {\it fastest apparent convergence (fac)},
amounts to $\bar{\alpha}_s \to \bar{\alpha}_{\rm fac}$, with
\begin{equation}
\label{azero}
\bar{\alpha}_{\rm fac}
\equiv \left[\frac{1}{\bar{\alpha}_s(|\bm{x} \!-\! \bm{y}|)}
+ \frac{(\bm{x} \!-\! \bm{z})^2 -
(\bm{y} \!-\! \bm{z})^2}{(\bm{x} \!-\! \bm{y})^2} \,
\frac{\bar{\alpha}_s(|\bm{x} \!-\! \bm{z}|) - \bar{\alpha}_s(|\bm{y} \!-\! \bm{z}|)}{\bar{\alpha}_s(|\bm{x} \!-\! \bm{z}|)\bar{\alpha}_s(|\bm{y} \!-\! \bm{z}|)}\right]^{-1}.
\end{equation}
This last prescription is particularly useful for what follows, in that it simplifies the
expression that we shall obtain for $\Delta{\mathcal{S}}$.
After a Fourier transform to momentum space, the solution to rcBK can be given
the following integral representation:
\begin{equation}
\label{rcBKint}
{\mathcal{S}}_{\rm \scriptscriptstyle rcBK}(\bm{k},X_g)={\mathcal{S}}_0(\bm{k})+2\pi\bar{\alpha}_s(\kt^2) \int_{x_g}^{1}\frac{{\rm d} x}{x}\,
\big[\mathcal{J}_{\rm \scriptscriptstyle rcBK}\big(\bm{k},0;X(x)\big)- \mathcal{J}_{v,{\rm \scriptscriptstyle rcBK}}\big(\bm{k},0;X(x)\big)\big]\,,\end{equation}
where the factorization of the running coupling $\bar{\alpha}_s(\kt^2)$ was possible
because of our assumption that $\kt$ is sufficiently hard (recall the discussion
after \eqn{NLOkt}). The functions $\mathcal{J}_{\rm \scriptscriptstyle rcBK}\big(\bm{k},0;X(x)\big)$ and
$\mathcal{J}_{v,{\rm \scriptscriptstyle rcBK}}\big(\bm{k},0;X(x)\big)$ are obtained from the respective functions
in \eqn{Sfull} after replacing $\mathcal{S}\to {\mathcal{S}}_{\rm \scriptscriptstyle rcBK}$.
\begin{figure}[t] \centerline
\includegraphics[width=.7\textwidth]{SLO.pdf}}
\caption{\small A graphical illustration of the integral version of the LO BK equation \eqref{toyLO}.
Running-coupling corrections and non-linear effects in $\mathcal{S}$ describing multiple scattering
are implicitly assumed, but not explicitly depicted.}
\label{fig:toyBK}
\end{figure}
We now introduce more schematic notations, as anticipated.
Specifically, let us ignore (just in our notations) the transverse momentum convolutions, the inessential numerical factors, and the non-linear structure of functions
like $\mathcal{J}$ w.r.t. the dipole $S$-matrix. Also, in writing the cross-section, we shall omit
the quark distribution function. That is, we shall rewrite \eqn{NLOkt} simply as
\begin{equation}\label{toyNLO}
\mathcal{N}_{{\rm \scriptscriptstyle LO}+{\rm \scriptscriptstyle NLO}}=\mathcal{S}_0 +\Delta{\mathcal{S}}(X_g)+ \bar{\alpha}_s(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x}
\,\mathcal{K}(x) \,\mathcal{S}\big(X(x)\big)
\,,\end{equation}
where the kernel $\mathcal{K}(x)$ encodes all the momentum space variables and convolutions
(but no dipole $S$-matrix) and $\mathcal{S}\big(X(x)\big)$ succinctly denotes all the factors involving
the dipole $S$-matrix, which could be either linear, or bi-linear, in $\mathcal{S}$. Both `real' and `virtual'
contributions are implicitly added in the above integral.
With these new notations, Eqs.~\eqref{Sfull} and \eqref{rcBKint} become
\begin{equation}\label{toySfull}
{\mathcal{S}}(X_g)={\mathcal{S}}_0+\Delta{\mathcal{S}}(X_g)
+ \bar{\alpha}_s(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x} \,\mathcal{K}(0) \,
\mathcal{S}\big(X(x)\big)\,,\end{equation}
and respectively
\begin{equation}\label{toyLO}
{\mathcal{S}}_{\rm \scriptscriptstyle rcBK}(X_g)=\mathcal{S}_0 + \bar{\alpha}_s(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x} \,\mathcal{K}(0) \,
\mathcal{S}_{\rm \scriptscriptstyle rcBK}\big(X(x)\big)\,,\end{equation}
where the kernel is now evaluated at $x=0$ (or $\xi=1$), that is, in the eikonal
approximation. Clearly, the compact
notation $\mathcal{K}(0)$ stays for the LO BK (or dipole) kernel. In particular, the rcBK equation
\eqref{toyLO} is graphically illustrated in Fig.~\ref{fig:toyBK}.
\begin{figure}[t] \centerline
\includegraphics[width=.9\textwidth]{SNLO.pdf}}
\caption{\small A graphical illustration of the integral version of the NLO BK equation
\eqref{toyNLOBK}. The middle term with kernel $\mathcal{K}(0)$ describes a soft gluon emission
and corresponds to the LO BK kernel, but with a running coupling. The last term
with kernel $\mathcal{K}_2(0)$ represents the NLO piece of the BK kernel, with the
running-coupling corrections excluded. The two $s$-channel gluons included within
$\mathcal{K}_2(0)$ can be close in rapidity.
The non-linear effects in $\mathcal{S}$ describing multiple scattering
are kept implicit.}
\label{fig:toyNLOBK}
\end{figure}
These simpler notations hopefully make clear that the quantity denoted as $\Delta{\mathcal{S}}$
must encode all NLO corrections to the BK kernel except for those expressing
the running of the coupling. These corrections should be
computed in the large-$N_c$ limit, for consistency with our previous approximations.
They can be inferred by inspection of the NLO
version of the BK equation shown in Eq. (5) of Ref.~\cite{Balitsky:2008zza}.
For convenience, we display the large-$N_c$ version of this equation
in Appendix \ref{sec:NLOBK}, where we also discuss its collinear improvement,
along the lines of Refs.~\cite{Iancu:2015vea,Iancu:2015joa}.
For simplicity, we shall stick here to our schematic notations and refer to
Appendix \ref{sec:NLOBK} for more explicit formulae.
The integral version of the NLO BK equation reads (in momentum space and adapted
to the kinematics at hand)
\begin{equation}\label{toyNLOBK}
{\mathcal{S}}(X_g)={\mathcal{S}}_0+ \bar{\alpha}_s(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x} \,\mathcal{K}(0) \,
\mathcal{S}\big(X(x)\big)+
\bar{\alpha}_s^2(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x} \,\mathcal{K}_2(0) \,
\mathcal{S}\big(X(x)\big)
\,,\end{equation}
where $\mathcal{K}_2(0)$ is a compact notation for the NLO piece of the BK kernel
alluded to above: its action on $\mathcal{S}$ generates
all the NLO terms visible in the r.h.s. of
\eqn{nlobkint} {\em except for the running coupling corrections},
which are explicitly included in the middle term of \eqn{toyNLOBK}.
More precisely, if one uses the prescription \eqref{azero} for the running coupling,
then in constructing $\mathcal{K}_2(0)$ one should omit the NLO terms in
\eqn{nlobkint} which are proportional to $\bar b$
(but keep all the other ones). Besides, the large NLO corrections enhanced by double or
single collinear logarithms require all-order resummations (`collinear improvement'),
to be shortly described (see also the discussion in Appendix \ref{sec:NLOBK}).
By comparing Eqs.~\eqref{toySfull} and \eqref{toyNLOBK}, it is now obvious that
$\Delta{\mathcal{S}}(X_g)$ must be identified with the third term in the r.h.s. of
\eqn{toyNLOBK}, that we now describe in some detail. As a general rule,
the NLO corrections to the BK kernel are obtained by evaluating all the 2-loop
graphs which yield a contribution of order $\alpha_s^2 Y$ and hence count for
a single step in the high-energy evolution (see Ref.~\cite{Balitsky:2008zza}
for an exhaustive list of diagrams and explicit calculations). A typical such a graph
involves a sequence of two gluon emissions,
whose longitudinal fractions $x_1$ and $x_2$ obey
$x_1\sim x_2\ll 1$; that is, the two gluons have similar rapidities, but they
are both soft compared to the projectile.
The integration variable $x$ in the last term in \eqn{toyNLOBK} can be
interpreted as $x\equiv x_1+x_2$. The other independent rapidity integration,
say over the variable $u\equiv x_2/(x_1+x_2)$ with $0 < u \le 1$, is implicitly included
in the structure of the NLO kernel $\mathcal{K}_2(0)$. This is possible since the scattering
between the 2-gluon system and the target is independent of $u$ to the accuracy of
interest\footnote{The double integral over $x$ and $u$ also generates contributions
of order $(\alpha_s Y)^2$, which count for 2 steps in the LO evolution. Such contributions
are exlicitly subtracted in Ref.~\cite{Balitsky:2008zza},
via the `plus' prescription, since already generated via 2 iterations
of the middle term in \eqn{toyNLOBK}.}. A rather schematic, but intuitive, graphical
illustration of \eqn{toyNLOBK} is shown in Fig.~\ref{fig:toyNLOBK}.
At large $N_c$, two successive gluon emissions
from the original $q\bar q$ dipole can generate up to three dipoles
in the fundamental representation. Accordingly, the third term in \eqn{toyNLOBK}
involves contributions which are cubic in the dipole
$S$-matrix, together with quadratic and linear terms.
All these contributions are visible in \eqn{nlobkint}.
To the accuracy of interest, the kinematics of the primary gluon
(with energy fraction $x=x_1+x_2$) can be treated in the LLA. This explains
why we were able to use the same lower limit $x_g$ in the integral over $x$
and also the same rapidity variable $X(x)$ for the relevant $S$-matrices as
in the middle term in \eqn{toyNLOBK}, which encodes the LO evolution.
\begin{figure}[t] \centerline
\includegraphics[width=.85\textwidth]{NTOT.pdf}}
\caption{\small A graphical illustration of the factorization of quark production
at NLO, as schematically encoded in \eqn{toyNLO2}. The primary gluon emission,
with kernel $\mathcal{K}(x)$, is included with exact kinematics. The evolution step depicted
as two-gluon emission with kernel $\mathcal{K}_2(0)$ succinctly represents the NLO
piece of the BK kernel (but without the running-coupling corrections, which were
already included in the middle term). Non-linear effects in $\mathcal{S}$,
corresponding to multiple scattering, are implicitly included but not explicitly shown.}
\label{fig:toyNLO}
\end{figure}
To conclude, the result for the quark multiplicity which is complete to NLO accuracy
can be compactly, but schematically, written as
\begin{equation}\label{toyNLO2}
\mathcal{N}_{{\rm \scriptscriptstyle LO}+{\rm \scriptscriptstyle NLO}}=\mathcal{S}_0 + \bar{\alpha}_s(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x}
\,\mathcal{K}(x) \,\mathcal{S}\big(X(x)\big)+
\bar{\alpha}_s^2(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x} \,\mathcal{K}_2(0) \,
\mathcal{S}\big(X(x)\big)
\,,\end{equation}
and is illustrated in Fig.~\ref{fig:toyNLO}. To deduce a more explicit expression,
one should use \eqn{NLOkt} together with the formula for $\Delta\mathcal{S}$ shown in
Appendix \ref{sec:NLOBK}.
Although the above factorization scheme is formally correct to NLO, it might be still
too hard, if not impossible, to achieve a full NLO accuracy in a practical calculation.
Indeed, as already mentioned, the target evolution is not known to NLO and
the relation between the function $\mathcal{S}\big(X(x)\big)$ which enters the above
factorization and the solution $\mathcal{S}(x)$ to the NLO BK equation for
the projectile is not known at the accuracy of interest. Besides, numerical calculations
might be hindered by the complexity of the NLO BK equation and of the transverse
convolutions implicit in the two integral terms in \eqn{toyNLO2}.
In view of that, we would like to propose two approximation schemes which we
believe have more chances to be transposed in practice. In both schemes,
the approximations refer to the dipole evolution, whereas the NLO impact factor,
as represented by the kernel $\mathcal{K}(x)$ in the middle term in \eqn{toyNLO2},
should be treated exactly.
The simplest approximation which is still physically meaningful is
the LO approximation (in the sense of rcBK) to the dipole evolution.
This is obtained by neglecting the third term in \eqn{toyNLO2} and replacing
$\mathcal{S}\simeq\mathcal{S}_{\rm \scriptscriptstyle rcBK}$ within the integrand of the middle term.
A similar strategy has been used \cite{Stasto:2013cha,Stasto:2014sea,Watanabe:2015tja}
in relation with the `plus' prescription \cite{Chirilli:2011km,Chirilli:2012jd} ; the respective
numerical calculations, albeit very complex, turned out to be tractable
\cite{Zaslavsky:2014asa}.
The second approximation, which is more ambitious, refers to the use of the collinearly-improved
version of the BK equation, as proposed in Refs.~\cite{Iancu:2015vea,Iancu:2015joa}.
We recall that, besides the running coupling corrections, this equation also resums
double-collinear logarithms together with a subset of the single collinear logarithms
(which includes the respective contribution at NLO). The corresponding approximation
to \eqn{toyNLO2} involves two aspects.
On the one hand, one must relate the function $\mathcal{S}\big(X(x)\big)$, which encodes the
evolution of the target, to the solution to the collinearly-improved BK equation (which
refers to the evolution of the projectile). This aspect is particularly important
for the middle term in \eqn{toyNLO2} and will be discussed in Appendix \ref{sec:proj}.
On the other hand, one must use a simplified
version of the NLO BK kernel $\mathcal{K}_2(0)$ which keeps only those corrections
to the LO kernel which refer to the collinear improvement.
This will be described in what follows.
To that aim, it is convenient to use the coordinate representation: the
last term in \eqn{toyNLO2}, which we recall corresponds to the piece
$\Delta{\mathcal{S}}(\bm{k},X_g)$ in \eqn{NLOkt}, will be written as
\begin{equation}\label{DeltaSF}
\Delta{\mathcal{S}}(\bm{k},X_g)=\int {\rm d}^2\bm{r}\, {\rm e}^{-{\rm i} \bm{k} \cdot \bm{r}} {\Delta S}(\bm{r},x_{\rm \scriptscriptstyle T})\,,
\end{equation}
where $\bm{r}=\bm{x}-\bm{y}$ and
\begin{align}\hspace*{-0.5cm}
\label{Scoll}
{\Delta S}(\bm{x},\bm{y}; x_{\rm \scriptscriptstyle T})\equiv \int_{x_{\rm \scriptscriptstyle T}}^{1}\frac{{\rm d} x}{x} \int \frac{{\rm d}^2\bm{z}}{2\pi}\,
\bar{\alpha}_{\rm fac}\,{\mathcal M}_{\bm{x}\bm{y}\bm{z}}\big(\mathcal{K}_{{\rm \scriptscriptstyle DLA}}
\mathcal{K}_{{\rm \scriptscriptstyle SL}}-1\big)
\Big[
S\Big(\bm{x},\bm{z}; \frac{x_{\rm \scriptscriptstyle T}}{x}\Big) S\Big(\bm{z}, \bm{y}; \frac{x_{\rm \scriptscriptstyle T}}{x}\Big)
-S\Big(\bm{x},\bm{y}; \frac{x_{\rm \scriptscriptstyle T}}{x}\Big)\Big].
\end{align}
In the above equation,
we recognize the dipole kernel ${\mathcal M}_{\bm{x}\bm{y}\bm{z}}$ and the running coupling
$\bar{\alpha}_{\rm fac}$ that were introduced before, together with two multiplicative
corrections to the
kernel, $\mathcal{K}_{{\rm \scriptscriptstyle DLA}}$ and $\mathcal{K}_{{\rm \scriptscriptstyle SL}}$, which encode the resummations
of double and respectively single collinear logarithms, as alluded to above. Physicswise,
$\mathcal{K}_{{\rm \scriptscriptstyle DLA}}$ implements the condition of time-ordering for the successive soft
gluon emissions by the projectile, whereas $\mathcal{K}_{{\rm \scriptscriptstyle SL}}$ resums a subset of the DGLAP
logarithms (see Refs.~\cite{Iancu:2015vea,Iancu:2015joa} for details).
The new rapidity argument $x_{\rm \scriptscriptstyle T}$ which appears too in \eqn{Scoll} will be later explained.
Specifically, $\mathcal{K}_{{\rm \scriptscriptstyle DLA}}$ is defined as the function
\begin{equation}\label{kdla}
\mathcal{K}_{{\rm \scriptscriptstyle DLA}}(\rho) = \frac{{\rm J}_1
\big(2\sqrt{\bar{\alpha}_s \rho^2}\big)}{\sqrt{\bar{\alpha}_s \rho^2}} =
1- \frac{\bar{\alpha}_s \rho^2}{2} + \frac{(\bar{\alpha}_s\rho^2)^2}{12} + \cdots,
\end{equation}
evaluated at $\rho^2={L_{\bm{x}\bm{z} r}L_{\bm{y} \bm{z} r}}$, with $L_{\bm{x}\bm{z} r} \equiv \ln[(\bm{x}-\bm{z})^2/r^2]$.
If the double logarithm $L_{\bm{x}\bm{z} r}L_{\bm{y} \bm{z} r}$ is negative, then one uses its absolute value
and the Bessel function ${\rm J}_1$ gets replaced by the modified Bessel function ${\rm I}_1$.
Note however that if, e.g., $(\bm{x}-\bm{z})^2 \ll r^2$, so that $L_{\bm{x}\bm{z} r} <0$, then
$(\bm{y}-\bm{z})^2 \simeq r^2$ and hence $L_{\bm{y} \bm{z} r}\simeq 0$. Accordingly, the relatively small
daughter dipoles bring no significant contributions to the difference
$\mathcal{K}_{{\rm \scriptscriptstyle DLA}}-1$. Furthermore,
\begin{equation}\label{ksl}
\mathcal{K}_{{\rm \scriptscriptstyle SL}}=\exp\left\{-\bar{\alpha}_s A_1\left|\ln
\frac{(\bm{x} \!-\! \bm{y})^2}{\min\{(\bm{x} \!-\! \bm{z})^2,(\bm{y} \!-\! \bm{z})^2\}}\right|\right\},\qquad
A_1 =\frac{11}{12} + \frac{N_{\rm f}}{6N_{\rm c}^3}\,,
\end{equation}
where $A_1$ is the `gluonic anomalous dimension' of the DGLAP evolution
(see e.g. \cite{Ciafaloni:1999yw} for details). Note that the difference
$\mathcal{K}_{{\rm \scriptscriptstyle DLA}}\mathcal{K}_{{\rm \scriptscriptstyle SL}}-1$ starts at $\order{\bar{\alpha}_s}$, as expected.
However, keeping only that lowest-order term in the expansion would artificially enhance
the importance of the `collinear' regions in phase-space --- the regions where
the successive gluon emissions
(or dipole splittings) are strongly ordered in transverse sizes, or momenta. This is the origin
of the instability of the strict NLO approximation to the high-energy evolution,
as previously mentioned.
Vice-versa, the all-order resummation of such corrections within the
factor $\mathcal{K}_{{\rm \scriptscriptstyle DLA}}\mathcal{K}_{{\rm \scriptscriptstyle SL}}$ suppresses the contributions from
the `collinear' regions and thus restores the convergence of perturbation theory.
This is rather obvious for the second factor $\mathcal{K}_{{\rm \scriptscriptstyle SL}}$, which exponentially
cuts off the configurations where the daughter dipoles are either much smaller, or much
larger, than the parent dipole. But this is also true for the other factor $\mathcal{K}_{{\rm \scriptscriptstyle DLA}}$,
which, as already mentioned, becomes important only when the daughter dipoles are
sufficiently large, such that $\bar{\alpha}_s \rho^2\gg 1$.
In that case, the Bessel function ${\rm J}_1\big(2\sqrt{\bar{\alpha}_s \rho^2}\big)$ is rapidly oscillating
when varying the position $\bm{z}$ of the emitted gluon, hence the integral over
the regions in space where `$\bm{z}$ is large' (in the sense
that $|\bm{z}-\bm{x}|\sim |\bm{z}-\bm{y}|\gg r$) averages to zero.
To conclude, for gluon emissions which are strongly ordered in transverse momenta,
we have $\mathcal{K}_{{\rm \scriptscriptstyle DLA}}\mathcal{K}_{{\rm \scriptscriptstyle SL}}\simeq 0$ and then the overall kernel
in \eqn{Scoll} reduces to {\em minus} the LO kernel $-\mathcal{K}(0)$. In turn, the latter
subtracts the soft and collinear contributions to the middle term in \eqn{toyNLO2},
that is, it implements the collinear improvement for the LO evolution, as it should.
There is one more aspect of \eqn{Scoll} which requires a few words of explanation: the rapidity
arguments of the various $S$-matrices and, related to them, the
lower limit $x_{\rm \scriptscriptstyle T}$ for the integral over $x$. If the evolution of these $S$-matrices is
computed to LO, i.e. according to rcBK, then one can identify
$x_{\rm \scriptscriptstyle T}\simeq x_g=\kt^2/\hat s$ and there is no distinction between projectile and target
evolutions. Albeit such an approximation would be formally justified
to NLO accuracy, it is still preferable to use the collinearly-improved
evolution, i.e. the BK equation with the resummed
kernel ${\mathcal M}_{\bm{x}\bm{y}\bm{z}}\mathcal{K}_{{\rm \scriptscriptstyle DLA}} \mathcal{K}_{{\rm \scriptscriptstyle SL}}$.
Indeed, this would ensure a smooth matching
with the middle term in \eqn{toyNLO2}, where the collinearly-improved version of the
evolution becomes compulsory. Since the latter has been formulated for projectile
evolution alone \cite{Iancu:2015vea,Iancu:2015joa}, one must understand what is
the longitudinal phase-space for the evolution in $x=q^+/q_0^+$ which corresponds
to the physical range for the evolution in $X=q^-/P^-$. In coordinate space
and to the accuracy of interest, one can write (see e.g. \cite{Iancu:2015vea})
\begin{equation}\label{SPST}
S_{\rm \scriptscriptstyle T}\big(\bm{r}, X\big) = S_{\rm \scriptscriptstyle P}\big(\bm{r}, x_{\rm \scriptscriptstyle T}\big)\,\qquad
\mbox{with}\qquad X=\frac{Q^2}{\hat s}\, \gg\, x_{\rm \scriptscriptstyle T}=\frac{Q^2_0}{\hat s}\,,
\end{equation}
where we have temporarily introduced the subscripts T (`target') and P (`projectile'),
to make the discussion more transparent and we recall that $\hat s=2q_0^+P^-$.
In this equation, $Q=1/r$ is the dipole
resolution scale in the transverse plane and can be also identified with the
transverse momentum $\kt$ of the produced quark, via the Fourier transform
\eqref{DeltaSF}; hence, $X\simeq X_g=x_g$. Furthermore,
$Q_0$ is the saturation scale in the nuclear target at low energy (say, within the MV model).
The interesting situation is such that $\kt \gtrsim Q_s(X_g)\gg Q_0$
and therefore $x_{\rm \scriptscriptstyle T} \ll x_g$, as indicated too in \eqn{SPST}.
The relation \eqref{SPST} can be understood as follows. The scattering between
the small dipole and the nuclear target probes the evolution of the latter
down to values of $X$ such that the longitudinal extent $\Delta x^+=1/(XP^-)$ of the
softest target fluctuations is still smaller than the lifetime $2q_0^+/Q^2$ of the dipole projectile.
This argument, which selects $X={Q^2}/{\hat s}$ as anticipated, is merely a variation
of our earlier derivation of \eqn{Xg} in Sect.~\ref{sec:kin}. If on the other hand the evolution
is encoded in the wavefunction of the projectile, then one should allow for the small-$x$
fluctuations with transverse momenta $\bm{q}$ within the range $Q^2 > \qt^2 > Q_0^2$
and with large enough lifetimes $2xq_0^+/\qt^2\gtrsim 1/P^-$ (indeed, $1/P^-$ is the
longitudinal extent of the un-evolved target). These conditions imply $x\gtrsim \qt^2/\hat s
\ge Q_0^2/\hat s$,
or $x \ge x_{\rm \scriptscriptstyle T}$, which explains the lower limit in the integral over $x$ in \eqn{Scoll}.
One may find surprising that the rapidity interval $\Delta y=\ln(1/x_{\rm \scriptscriptstyle T})$
available for the evolution of the projectile is much larger than that, $\Delta Y=\ln(1/X_g)$,
allowed for the evolution of the target. But one should keep in mind that the projectile
evolution with decreasing $x$ is strongly constrained by the condition of time-ordering,
which limits the corresponding transverse phase-space (via the multiplicative
correction $\mathcal{K}_{{\rm \scriptscriptstyle DLA}}$ to the kernel) and thus reduces the evolution speed.
By contrast, there is no similar constraint for the evolution of the target with
decreasing $X$, since in that case the physical condition of time-ordering is
automatically satisfied.
\subsection{Subtracting the leading order evolution: why is this subtle}
\label{sec:sub}
The factorization scheme that we have constructed in the previous sections,
cf. Eqs.~\eqref{NLOkt} or \eqref{toyNLO2}, does not involve any subtraction:
there is no over-counting of the relevant perturbative contributions and hence
no need for a subtraction. We therefore expect that calculations based
on this factorization should yield a positive result for the quark multiplicity.
More explicit arguments in that sense will be presented later in this section.
This represents a significant improvement over the previous proposal in
Refs.~\cite{Chirilli:2011km,Chirilli:2012jd}, so it is interesting to better understand
the relation between these two schemes.
From the discussion in Sect.~\ref{sec:CXY}, we recall that in the approach
\cite{Chirilli:2011km,Chirilli:2012jd} on explicitly subtracts the LO evolution of
the dipole $S$-matrix from the NLO correction to the impact factor. Such a
subtraction can also be performed within our present approach. To synthetically
describe that, we shall again use the schematic notations introduced in the previous
subsection. Using \eqn{toyNLOBK}, we can express the last term in the NLO cross-section
\eqref{toyNLO2} (a NLO correction to the dipole evolution)
as the difference between the dipole $S$-matrix at NLO and its LO evolution:
\begin{equation}\label{toyK2}
\bar{\alpha}_s^2(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x} \,\mathcal{K}_2(0)
\mathcal{S}\big(X(x)\big)={\mathcal{S}}(X_g) - \left[{\mathcal{S}}_0+ \bar{\alpha}_s(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x} \,\mathcal{K}(0) \,
\mathcal{S}\big(X(x)\big)\right]
\,.\end{equation}
The r.h.s. of this equation is not exactly the same as the difference $\mathcal{S}-\mathcal{S}_{\rm \scriptscriptstyle rcBK}$,
but it is very close to it.
(The expression between the square brackets would reduce to $\mathcal{S}_{\rm \scriptscriptstyle rcBK}$
if the $S$-matrix under the integral would be itself approximated by rcBK; recall
\eqn{toyrcBK}.) So, clearly, \eqn{toyK2} expresses a NLO correction which is {\em a priori}
small as the difference between two quantities which are individually large: each
of them includes the LO contribution $\mathcal{S}_{\rm \scriptscriptstyle rcBK}$.
Inserting \eqn{toyK2} into \eqn{toyNLO2}, one deduces an alternative expression for
the NLO cross-section,
\begin{equation}\label{toyNLO1}
\mathcal{N}_{{\rm \scriptscriptstyle LO}+{\rm \scriptscriptstyle NLO}}=\mathcal{S}(X_g)
\,+ \bar{\alpha}_s(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x} \,\big[\mathcal{K}(x)-\mathcal{K}(0)\big]\,\mathcal{S}\big(X(x)\big),
\end{equation}
in which the LO evolution (as represented by the kernel $\mathcal{K}(0)$) is explicitly subtracted
from the impact factor. Since the difference $\mathcal{K}(x)-\mathcal{K}(0)$
vanishes as $x\to 0$, the above integral is controlled by large values $x\sim 1$
and it does not generate a small-$x$ logarithm anymore.
That is, the second term in the r.h.s. of \eqn{toyNLO1} is a pure $\bar{\alpha}_s$ correction,
that can be viewed as the NLO contribution to the impact factor. The
NLO correction to the dipole evolution is now fully encoded in the first term
$\mathcal{S}(X_g)$. Indeed, to the NLO accuracy of interest, the $S$-matrix within the integral
can be evaluated by using the LO approximation, $\mathcal{S}\simeq\mathcal{S}_{\rm \scriptscriptstyle rcBK}$.
\eqn{toyNLO1} is very similar to the proposal in
Refs.~\cite{Chirilli:2011km,Chirilli:2012jd}, which in our schematic notations reads
\begin{align}\label{toyCXY
\mathcal{N}_{{\rm \scriptscriptstyle CXY}}=
\mathcal{S}(X_g)+ \bar{\alpha}_s(\kt^2) \int_{0}^{1} \frac{{\rm d} x}{x} \,\big[\mathcal{K}(x)-\mathcal{K}(0)\big]\,\mathcal{S}(X_g)\,.
\end{align}
The main difference w.r.t. \eqn{toyNLO1} is that
the $S$-matrix within the above integral over $x$
is not evaluated at the $x$-dependent rapidity argument $X(x)$,
but rather at its endpoint value $X_g=X(1)$. In other terms, \eqn{toyCXY}
is {\em local} in the target rapidity $X_g$, as standard for the $k_\perp$-factorization.
In spite of this difference, the results in Eqs.~\eqref{toyCXY} and \eqref{toyNLO1}
are perturbatively equivalent to NLO accuracy.
Indeed, since the integral over $x$ in \eqref{toyNLO1} is controlled by $x\sim 1$,
there is no loss of NLO accuracy if one replaces $\mathcal{S}\big(X(x)\big)\to \mathcal{S}(X_g)$.
This can be easily checked by using the dominant energy-behavior of the BK
solution in the weak scattering regime, namely $\mathcal{S}(X)\propto 1/X^\lambda$
with $\lambda=\order{\bar{\alpha}_s}$.
This being said, one should keep in mind that the limiting value $\mathcal{S}(X_g)$
is strictly larger than $\mathcal{S}\big(X(x)\big)$ for any $x<1$, because the function
$\mathcal{S}(X)$ increases when decreasing $X$.
Hence, albeit formally allowed to the accuracy of interest,
the replacement $\mathcal{S}\big(X(x)\big)\to \mathcal{S}(X_g)$ could still be troublesome,
in that it might result in an {\em over-subtraction}. Indeed,
as discussed in \cite{Ducloue:2016shw}, the difference
$\mathcal{K}(x)-\mathcal{K}(0)$ is strictly negative for sufficiently large $\kt\gtrsim Q_s$.
Hence, the negative correction to the impact factor at large $\kt$ is over included in
\eqref{toyCXY} as compared to \eqref{toyNLO1}, a feature which might
contribute to the `negativity' problem under consideration.
In order to discuss this problem --- the fact that the cross-section computed according
to \eqn{toyCXY} becomes negative at sufficiently large (but still semi-hard) $\kt$
---, it is useful to keep in mind that previous numerical evaluations of \eqn{toyCXY}
used approximate versions of the dipole $S$-matrix,
like $\mathcal{S}_{\rm \scriptscriptstyle rcBK}$ \cite{Stasto:2013cha,Stasto:2014sea,Watanabe:2015tja}.
Hence, it would be interesting to understand why \eqn{toyCXY} with
$\mathcal{S}\to\mathcal{S}_{\rm \scriptscriptstyle rcBK}$ can lead to a negative cross-section at large transverse momenta.
To that aim, let us consider the rcBK approximation to the `subtracted' version of our
factorization scheme, cf. \eqref{toyNLO1}. This reads
\begin{equation}\label{toyrcBK}
\mathcal{N}_{{\rm \scriptscriptstyle rcBK}}= \mathcal{S}_{{\rm \scriptscriptstyle rcBK}}+ \bar{\alpha}_s(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x} \,
\big[\mathcal{K}(x)-\mathcal{K}(0)\big]\,\mathcal{S}_{\rm \scriptscriptstyle rcBK}\big(X(x)\big)\,.
\end{equation}
Via the rcBK equation \eqref{toyLO}, this is equivalent to
\begin{equation}\label{toyrcBK2}
\mathcal{N}_{{\rm \scriptscriptstyle rcBK}} = \mathcal{S}_{0}+ \bar{\alpha}_s(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x} \,
\mathcal{K}(x)\,\mathcal{S}_{\rm \scriptscriptstyle rcBK}\big(X(x)\big),
\end{equation}
which is of course the same as our `un-subtracted' result in \eqn{toyNLO}
with $\Delta\mathcal{S}=0$ and $\mathcal{S}\to\mathcal{S}_{\rm \scriptscriptstyle rcBK}$.
We believe that the r.h.s. of \eqn{toyrcBK2} should be positive.
Indeed, when the primary gluon emission is treated in the eikonal approximation
($\mathcal{K}(x)\to \mathcal{K}(0)$), \eqn{toyrcBK2} is the same as the integral representation
of $\mathcal{S}_{\rm \scriptscriptstyle rcBK}$, which is well known to be positive definite (at least in coordinate space).
With the actual kernel $\mathcal{K}(x)$, the value of the integral term should be somewhat
reduced (since the correct phase-space for the primary gluon emission is smaller
than its eikonal estimate), but this should remain positive, as it describes the
growth of the dipole $S$-matrix via gluon emissions.
Hence, the `subtracted' result in \eqn{toyrcBK} should be positive as well.
In spite of that, we believe that explicit numerical calculations based on \eqn{toyrcBK}
may run into difficulties (in particular, yield a negative result)
because of the high degree of `fine-tuning' inherent in the subtraction method:
in going from \eqn{toyrcBK2} to \eqn{toyrcBK}, we have added and subtracted
the same quantity ($\mathcal{S}_{\rm \scriptscriptstyle rcBK}$), but we have done that in a very peculiar
way: we have added the l.h.s. of \eqn{toyLO}, that is, $\mathcal{S}_{\rm \scriptscriptstyle rcBK}$ itself, but we
have subtracted the r.h.s. of \eqn{toyLO}, that is, the integral representation of $\mathcal{S}_{\rm \scriptscriptstyle rcBK}$.
This procedure leaves the result unchanged if and only if the function
$\mathcal{S}_{\rm \scriptscriptstyle rcBK}$ is an exact solution to the integral equation \eqref{toyLO}.
But any numerical approximation in solving rcBK, or in the Fourier transform
$S(\bm{r},X)\to\mathcal{S}(\bm{k},X)$, may lead to an imbalance between
the terms that have been added and respectively subtracted.
For instance, such an imbalance
could be introduced by the treatment of the running coupling corrections,
which in general is not fully coherent between the coordinate-space
and the momentum-space representations.
In practice, the momentum-space version of the rcBK equation, as shown
in \eqref{toyLO}, is not the {\em exact} Fourier transform of the respective equation in
coordinate space: the Fourier transform is not also applied to the
running of the coupling. So, strictly speaking, there is some
mismatch between Eqs.~ \eqref{toyrcBK} and \eqref{toyrcBK2} even when
using the exact solution to the rcBK equation, as obtained in coordinate space.
When computing the cross-section according to \eqn{toyCXY},
this mismatch can be further enhanced by the fact that the function $\mathcal{S}_{\rm \scriptscriptstyle rcBK}\big(X(x)\big)$
gets replaced by its maximal value $\mathcal{S}_{\rm \scriptscriptstyle rcBK}(X_g)$ (cf. the discussion after
\eqn{toyCXY}).
From the above discussion, we see that the rcBK equation plays the role of a `self-consistency
condition' for rewriting the cross-section in a `subtracted' form.
To illustrate the importance of having `good' solutions to this equation, let us consider a
somewhat extreme example, where this condition is strongly violated.
(A similar discussion can be found in \cite{Ducloue:2016shw}.)
Namely, we consider the popular approximation in which the dipole $S$-matrix is taken from
the GBW model \cite{GolecBiernat:1998js}, which is a Gaussian:
\begin{equation}\label{GBW}
{\mathcal{S}}_{\rm \scriptscriptstyle GBW}(\bm{k}, X)\,=\,\frac{4\pi}{Q_s^2}\,{\rm e}^{-\frac{\kt^2}{Q_s^2}}\,,\end{equation}
where $Q_s^2(X)=Q_0^2(X_0/X)^\lambda$ with $\lambda\simeq 0.3$. Clearly, this is a very poor
approximation at high $\kt\gtrsim Q_s$, where it decays exponentially, in sharp contrast with
the power law tail $1/\kt^4$ predicted by pQCD.
Using this particular model within \eqn{toyNLO1}, one finds
\begin{equation}\label{toyrcGBW}
\mathcal{N}_{{\rm \scriptscriptstyle GBW}}= \mathcal{S}_{{\rm \scriptscriptstyle GBW}}(X_g)+ \bar{\alpha}_s \int_{x_g}^{1} \frac{{\rm d} x}{x} \,
\big[\mathcal{K}(x)-\mathcal{K}(0)\big]\,\mathcal{S}_{\rm \scriptscriptstyle GBW}\big(X(x)\big)\,.
\end{equation}
The `LO' piece $\mathcal{S}_{{\rm \scriptscriptstyle GBW}}$ exponentially vanishes at large $\kt\gg Q_s$, whereas
the `subtracted' piece shows a tail $\propto 1/\kt^4$, since obtained by iterating once the LO BK equation.
So, the `subtracted' piece is not only
larger than the `LO' one, but it is the only one to survive at large $\kt$. Hence,
at $\kt\gtrsim Q_s$ the overall result reduces to the second, `NLO', piece, which is
negative in that particular region of phase-space \cite{Ducloue:2016shw}, as already mentioned.
This is in agreement with the numerical findings in \cite{Watanabe:2015tja,Ducloue:2016shw}.
To summarize, albeit Eqs.~\eqref{toyrcBK} and \eqref{toyrcBK2} are in principle equivalent with each
other, the second equation is probably safer to use in practice. A similar discussion applies to the full NLO cross-section: the `unsubtracted' factorization \eqref{toyNLO2} should provide a meaningful result which is positive when evaluated with either the NLO $S$-matrix $\mathcal{S}$, or with its collinearly-improved version $\mathcal{S}_{{\rm \scriptscriptstyle collBK}}$. On the other hand, the `subtracted' version \eqref{toyNLO1}, which may look appealing since structurally simpler, is likely to be more tricky to use in practice.
\section{Summary and conclusions}
\label{sec:conc}
In this paper, we have established a factorization scheme allowing the computation of single-inclusive
particle production at forward rapidities in proton-nucleus collisions at next-to-leading order
in pQCD, in the presence of the non-linear effects associated with the high gluon
density in the nuclear target. The main difference with respect to the previous proposal in
\cite{Chirilli:2011km,Chirilli:2012jd} is that our result involves no rapidity subtraction,
meaning that it is free of the fine-tuning problem inherent
in any such a subtraction scheme. The `fine-tuning' refers to the fact that
the numerical solution to the evolution equation for the dipole $S$-matrix
must be precisely known to ensure that
the quantity which is included as `LO evolution' is properly subtracted from the
`NLO correction to the impact factor'. Further approximations
within the subtraction scheme, which are formally
allowed to NLO accuracy, or even small numerical errors in the
associated calculations, may spoil such a fine cancellation and lead to unphysical results.
In our opinion, this is the reason why
the cross-section computed within this scheme
appears to turn negative at sufficiently large transverse momenta
\cite{Stasto:2013cha,Stasto:2014sea,Watanabe:2015tja,Ducloue:2016shw}.
By contrast, our scheme is more robust and should converge to physical,
positive-definite, results with considerably less numerical efforts.
Our factorization scheme relies on the skeleton structure of perturbative QCD
up to two-loop order, which is the relevant loop-order for the NLO calculation.
The most general version of our result is exhibited in \eqn{NLO}.
This version however is probably too complicated to be used in practice.
Fortunately, important simplifications become possible in the interesting situation
where the transverse momentum $\kt$ of the
produced quark is relatively large, $\kt\gtrsim Q_s$. In that case, the primary gluon is hard as
well, $\pt\sim\kt$, and the general formula \eqref{NLO} can be then replaced by
\eqn{NLOkt} [or \eqref{NLOX0}], which is considerably simpler. The latter shows the same degree of
complexity, in terms of transverse integrations, like the formulae already used
in practice in relation with the subtraction method. So, we are confident that \eqn{NLOkt}
can indeed by explicitly evaluated, via the numerical tools developed in
\cite{Zaslavsky:2014asa,Stasto:2013cha,Stasto:2014sea,Watanabe:2015tja,Ducloue:2016shw}.
As discussed in Sect.~\ref{sec:dijet}, the fact that we can approximate $\pt\sim\kt$ is specific
to the problem of particle production. This would not apply, say, to deep inelastic scattering,
where the dipole evolution must be computed in transverse coordinate space.
In that case, if the parent dipole is sufficiently small ($r\ll 1/Q_s$) --- as appropriate for DIS at
relatively high virtuality $Q^2\gg Q_s^2$ ---, then the first gluon emission in the high-energy evolution
is considerably {\em softer} (it typically carries a transverse momentum $\pt \ll 1/r$) and the
dependence upon its transverse kinematics cannot be simplified at NLO. This is visible in the NLO
calculations of the DIS impact factor \cite{Balitsky:2010ze,Beuf:2011xd,Balitsky:2012bs,Beuf:2016wdz}.
To explicitly evaluate the quark multiplicity according to \eqn{NLOkt}, one also needs a suitable
approximation for the high-energy evolution of the dipole $S$-matrix.
The simplest such an approximation which is still meaningful for
phenomenology is the LO BK equation with running coupling
(rcBK) \cite{Kovchegov:2006wf,Kovchegov:2006vj,Balitsky:2006wa}. A more accurate
treatment of the evolution can be obtained by using the collinearly-improved BK equation
\cite{Beuf:2014uia,Iancu:2015vea,Iancu:2015joa,Albacete:2015xza,Lappi:2016fmu,Hatta:2016ujq},
which resums the double collinear logarithms inherent in the `hard-to-soft' evolution
of the projectile together with a subset of single collinear logs. The inclusion of the
collinear improvement in the problem at hand is quite subtle, due to the need to properly
identify the rapidity phase-space for the evolution of the dilute projectile
(the incoming quark, or the pair made with this quark and the
hard primary gluon). This is clarified in Sect.~\ref{sec:evol} and Appendix \ref{sec:proj}.
An important lesson emerging from our analysis is that one should not always insist in
writing the result of a NLO calculation in pQCD at high-energy in a `$\kt$-factorized form',
which is local in rapidity. This can be best appreciated by comparing our
results in Eqs.~\eqref{toyNLO2}, \eqref{toyNLO1}, and \eqref{toyCXY}. \eqn{toyNLO2} is obtained
via a direct evaluation of the relevant, one-loop and two-loop, diagrams. As such, it
does not involve any rapidity subtraction at one-loop level: the would-be NLO
correction to the impact factor and the first step in the high-energy evolution of the color dipole
are both encoded in the middle term in \eqn{toyNLO2}. In \eqn{toyNLO1}, the two
contributions that we just mentioned are explicitly separated from each other,
via a rapidity subtraction, but the result is still not `factorized': the
$\order{\alpha_s}$-correction to the impact factor and the dipole $S$-matrix are still entangled
by the rapidity integral over $x$ (besides the transverse momentum convolutions which are
implicit in our schematic notations). Finally, \eqn{toyCXY} expresses a genuine
$\kt$-factorization: the r.h.s. is local in the target rapidity $X_g$
and the NLO impact factor is explicitly factorized (in so far as the rapidity
dependence is concerned) from the dipole $S$-matrices,
which encode the high-energy evolution\footnote{Notice that in this high-density context where
gluon saturation and multiple scattering are important, the Fourier transform
of the dipole $S$-matrix plays the role of a generalized
unintegrated gluon distribution for the nuclear target; see e.g. \cite{Dominguez:2011wm}.}.
These three representations for the NLO cross-section, \eqref{toyNLO2}, \eqref{toyNLO1},
and \eqref{toyCXY}, are all consistent with each other to NLO accuracy. Yet, the two
representations involving a rapidity subtraction, Eqs.~\eqref{toyNLO1} and \eqref{toyCXY},
are potentially affected by the issue of fine-tuning that we have identified in this paper.
By contrast, the original result in \eqn{toyNLO2} is free
from this problem and therefore it should provide a positive-definite estimate for the cross-section.
We would like to conclude this section with some prospects for the extension of this
factorization program beyond NLO. Our factorization scheme suggests an interesting
pattern which is likely to survive beyond the present approximations. The distinguished
feature of this pattern is the fact that the higher-order corrections to the impact factor are
not explicitly separated from the relevant corrections to the high-energy evolution.
To illustrate this point, let us consider quark production at NNLO. Without any claim
to completeness and leaving aside the possibility of new contributions which go beyond
the dipole picture, let us here indicate the generic structure that we expect in view of
our previous analysis in this paper. Namely, we expect the NNLO cross-section to include
the following four terms (in schematic notations,
similar to those introduced in Sect.~\ref{sec:evol})
\begin{align}\label{toyNNLO}
\mathcal{N}_{{\rm \scriptscriptstyle LO}+{\rm \scriptscriptstyle NLO}+{\rm \scriptscriptstyle NNLO}}= \mathcal{S}_0 & + \bar{\alpha}_s(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x}
\,\mathcal{K}(x) \,\mathcal{S}\big(X(x)\big) \nonumber\\ & +
\bar{\alpha}_s^2(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x_1}{x_1} \int_{x_g}^{x_1} \frac{{\rm d} x_2}{x_2} \,
\big[\mathcal{K}_2(x_1,x_2/x_1)-\mathcal{K}_2(x_1,0)\big] \,
\mathcal{S}\big(X(x_1,x_2)\big)\nonumber\\ & +
\bar{\alpha}_s^3(\kt^2) \int_{x_g}^{1} \frac{{\rm d} x}{x} \,\mathcal{K}_3(0) \,
\mathcal{S}\big(X(x)\big)
\,,\end{align}
where it is now understood that the dipole $S$-matrices evolve according to the
(presently unknown) NNLO version of the BK equation. The argument of the running
couplings is taken as $\kt^2$, for illustration, but this should be the right choice
only in the limit where the transverse momentum of the produced quark is larger than
any other transverse scale in the problem. Also, the lower limit $x_g$ on the
integrals over $x$ is purely illustrative: the actual limit should depend upon the
exact kinematics. The first two terms in the r.h.s.
of \eqn{toyNNLO} have the same structure as the respective terms in \eqn{toyNLO2};
in particular, the second term encodes the NLO impact factor and the LO piece of the
BK kernel. The third term is a generalization of the last term in \eqn{toyNLO2}: it encodes
the NLO piece of the BK kernel together with the NNLO correction to the impact factor.
This term is generated by 2-loop graphs where the 2 emitted partons (say, gluons) can have
arbitrarily energy fractions $x_1$ and $x_2$ (with $x_2\le x_1$ for definiteness),
so their emissions must be computed with exact kinematics. The sum of all such graphs
is schematically represented by the kernel $\mathcal{K}_2(x_1,x_2/x_1)$. The subtraction of
$\mathcal{K}_2(x_1,0)$ is needed to avoid the over inclusion of the first two steps in the
high energy evolution, that were already included in the second term, with kernel
$\mathcal{K}(x)$. Finally, the fourth term in \eqn{toyNNLO} represents the NNLO correction
to the BK kernel, as generated by 3-loop graphs
in which the 3 gluons are all soft, but close in rapidity to each other:
$x_1\sim x_2\sim x_3\ll 1$. Clearly, in the limit where the very first emission by the projectile
is soft and computed in the eikonal approximation, the r.h.s. of \eqn{toyNNLO} becomes
proportional to the integral representation of the NNLO BK equation.
\section*{Acknowledgments}
\vspace*{-0.3cm}
We would like to thank Tuomas Lappi for inspiring discussions which triggered
our interest on this problem. We are grateful to Bertrand Duclou\'e, Dmitry Ivanov,
Tuomas Lappi, and Yan Zhu for useful comments on the manuscript.
A.H.M. would like to thank Bowen Xiao for useful and informative discussions.
D.N.T. would like to thank Guillaume Beuf for insightful remarks on the interplay
between evolution and factorization at next-to-leading.
A.H.M. and D.N.T. would like to acknowledge l'Institut de Physique Th\'eorique de Saclay for
hospitality during the early stages of this work.
The work of E.I. is supported in part by the European Research Council
under the Advanced Investigator Grant ERC-AD-267258. The work of A.H.M.
is supported in part by the U.S. Department of Energy Grant \# DE-FG02-92ER40699.
|
\section{Magnetic susceptibility}
\begin{figure}[tbp]
\linespread{1}
\par
\includegraphics[width=3.1in]{Magsus.pdf}
\par
\caption{Magnetic susceptibility data obtained for Er$_{2-x}$Y$_{x}$Ti$_2$O$_7$ with $x=$0 (pure), 0.2 (10$\%$-Y) and 0.4 (20$\%$-Y). Powder samples from crushed single crystals used in the neutron scattering measurements were employed. For each sample, both the zero field cooled (ZFC - closed symbols) and the field cooled (FC - open symbols) susceptibility are shown. The inset shows the Curie-Weiss fits performed for each sample.}
\label{MagSus}
\end{figure}
Single crystals of Er$_{2-x}$Y$_{x}$Ti$_2$O$_7$ with $x=0$ (pure), 0.2 (10$\%$-Y) and 0.4 (20$\%$-Y) were grown at McMaster University using a two-mirror floating zone image furnace. The growth procedure closely resembled the one used in Ref.~\cite{Niven2014,balakrishnan1998single,Bryan2012}. The single crystal of Er$_2$Ti$_2$O$_7$ used for this study is the same as the one studied by previous neutron scattering work~\cite{Ruff2008,PhysRevLett.109.167201,PhysRevLett.112.057201}. The susceptibility measurements were performed using a Quantum Design MPMS magnetometer equipped with a $^{3}$He insert and the results are shown in Fig.~\ref{MagSus}.
Magnetic susceptibility measurements were performed on a small portion of the crushed single crystal of undoped Er$_2$Ti$_2$O$_7$ and on crushed single crystal of the 10$\%$-Y and 20 $\%$-Y samples. The dc susceptibility was measured on warming using both a field cooled and a zero field cooled protocol with an external field of 0.01~T. In each case, the transition to long range magnetic order is marked by a cusp in the susceptibility as well as a bifurcation of the field cooled and zero field cooled susceptibilities. As the cusp and the bifurcation in the susceptibility do not occur concomitantly, we define $T_N$ as the cusp maxima, and the difference between the two as the error bar. The undoped Er$_2$Ti$_2$O$_7$ orders at $T_N = 1.1(1)$~K, consistent with previous studies~\cite{PhysRevB.68.020401}. The effect of diluting magnetic Er$^{3+}$ with non-magnetic Y$^{3+}$ is to suppress the magnetic ordering transition to lower temperature. The 10$\%$-Y and 20$\%$-Y samples have N\'eel temperatures of $T_N = 1.04(5)$~K and $T_N = 0.74(8)$~K, respectively. This suppression of $T_N$ is consistent with previous studies of yttrium-dilution in Er$_2$Ti$_2$O$_7$~\cite{Niven2014}.
At high temperatures, Curie-Weiss behavior is observed in both undoped and yttrium-doped Er$_2$Ti$_2$O$_7$. Curie-Weiss fits were performed for each sample between 50~K and 300~K, and the results of these fits are summarized in Table I of the main manuscript. The Curie-Weiss temperature ($\theta_{\text{CW}}$) does not change appreciably in these three materials, ranging from $-17.2(3)$~K to $-18.0(2)$~K. The paramagnetic moment extracted from the high-temperature Curie-Weiss fits, 9.555(4)$\mu_B$ for Er$^{3+}$ as well as prior bulk magnetization studies~\cite{Bramwell2000}. However, the paramagnetic moment per Er$^{3+}$ systematically decreases with yttrium doping, likely a consequence of differences in the high energy crystal electric field schemes in these three materials. We note, however, that our inelastic scattering study revealed no low-lying crystal electric field levels in either the 10$\%$-Y or 20$\%$-Y samples up to 3 meV.
\section{Long range ordered and diffuse contributions to the (220) Bragg peak in E\lowercase{r}$_{2-x}$Y$_x$T\lowercase{i}$_2$O$_7$}
\begin{figure*}[tbp]
\linespread{1}
\par
\includegraphics[width=7in]{FitBragg.pdf}
\par
\caption{Typical fits of the (220) magnetic Bragg peak performed in this work for Er$_{2-x}$Y$_x$Ti$_2$O$_7$ with $x=0$ (pure), 0.2 (10$\%$-Y) and 0.4 (20$\%$-Y)) at $T=0.1$~K in zero field. The yellow areas represent the elastic, resolution-limited (Gaussian) magnetic contribution to the scattering and the blue area represents the dynamic, quasi-elastic or frozen (Lorentzian) contributions to the scattering. The red area shows the structural scattering obtained via a fit of the 3T data set. The black lines shows the resulting fit which are obtained by the addition of the three curves (blue,yellow and red). }
\label{FitBragg}
\end{figure*}
To extract the long-range order as well as the dynamic, quasi-elastic and frozen contributions to the (220) magnetic Bragg peak, we performed an integration of the neutron scattering data set of $-0.2$ to 0.2 in the [00L] direction and $-0.1$ to 0.1 meV in energy transfer. The results of such integration for the three single crystals are shown in Fig.~\ref{FitBragg} for $T=0.1$~K and zero applied field. Each Bragg peak has been fit to a Gaussian function to account for the long-range order (blue shaded region in Fig.~\ref{FitBragg}) and a Lorentzian function for the dynamic, quasi-elastic and frozen contributions to the scattering (yellow shaded region in Fig.~\ref{FitBragg}). The width of the Gaussian function was fixed by the width of a strictly nuclear reflection in the 3T data set. An additional Gaussian contribution to the scattering was accounted for by the 3T data set itself, representing the nuclear contribution (red shaded region in Fig.~\ref{FitBragg}) and kept fix for the low field data sets. Finally, the width of the Lorentzian function was allowed to refine freely. The results for all samples show that the relative contribution of the Lorentzian scattering over the Gaussian (long-range order) contributions is strongly enhanced as a function of doping. As discussed in the main manuscript, this indicates that upon dilution, the spins do not order in the discrete states belonging to $\psi_2$ but instead freeze into a mosaic of $\psi_2$ and $\psi_3$ domains, whose characteristic size is approximated by the inverse of the $Q$-width of the Lorentzian component to the scattering, which gives 35 $\pm$ 7 \AA.
\section{Temperature dependence of the (220) Bragg peak for 20\%-Y E\lowercase{r}$_{2}$T\lowercase{i}$_2$O$_7$.}
\begin{figure}[tbp]
\linespread{1}
\par
\includegraphics[width=2.8in]{fig5.pdf}
\par
\caption{Elastic cut showing the (220) magnetic Bragg peak for varying temperature in Er$_{2-x}$Y$_x$Ti$_2$O$_7$ with $x$=0.4 (20$\%$-Y) in zero applied field. The black line represents the resulting fit for the $T=0.1$~K data set. The inset panel shows the temperature dependence of the Gaussian, long range ordered, and Lorentzian contributions to the (220) scattering, where the dashed line marks the N\'eel transition temperature, $T_N=0.74(8)$~K.}
\label{fig5}
\end{figure}
The temperature dependence of the (220) magnetic Bragg peak in Er$_{2-x}$Y$_x$Ti$_2$O$_7$ with $x=0.4$ (20$\%$-Y) has also been explored to investigate a possible phase transition between $\psi_2$ and $\psi_3$ states at temperatures between the base temperature of our neutron experiment ($T=0.1$~K) and the N\'eel temperature ($T=0.74(8)$~K) of this sample. Fig.~\ref{fig5} shows elastic cuts around the (220) Bragg position obtained by an integration of $-0.2$ to 0.2 in the [00L] direction and $-0.1$ to 0.1 meV in energy transfer. The same procedure followed in the previous section was again employed to fit the (220) magnetic Bragg peak at every temperature collected in our neutron experiment in zero applied field. In this figure, the structural contribution of the long-range order scattering at (220) has been removed using intensity obtained from the paramagnetic, $T=2$~K data set. The temperature dependence of the Lorentzian and the long-range order scattering are shown in the inset of Fig.~\ref{fig5}. The long-range order contribution shows typical order parameter behavior, correlated with $T_N$ and no anomaly is observed below that temperature. In contrast, the Lorentzian scattering has little temperature dependence but is enhanced near $T_N$, consistent with critical scattering near a continuous phase transition.
|
\section{Introduction}
\label{sec:intro}
Cosmological experiments indicate that about a quarter of the energy budget of the universe is sourced by dark matter (DM).
How did DM acquire this abundance?
The intimacy between thermodynamics and the history of the early universe suggests that it may have been set
by some thermal mechanism.
Usually, DM is assumed to have frozen out of equilibrium from a thermal bath of various particle species.
In the simplest models, the DM relic abundance $\Omega_\chi h^2$ is set by a Lee-Weinberg mechanism \cite{Lee:1977ua} involving the process $\chi \chi \rightarrow {\rm SM~SM'}$.
The key ingredient here is the thermally averaged annihilation cross-section, $\langle \sigma v\rangle_{\rm ann}$.
To satisfy $\Omega_\chi h^2$ measured by Planck \cite{Ade:2015xua}, one requires, at the stage of freeze-out,
\begin{equation}
\langle \sigma v\rangle_{\rm ann} = 3 \times 10^{-26}~{\rm cm}^3~{\rm s}^{-1} .
\label{eq:thermalxs}
\end{equation}
This value is obtainable in a minimal manner by parametrically taking $\langle \sigma v\rangle_{\rm ann} \sim \alpha^2/M^2$,
where $\alpha$ and $M$ are the DM-Standard Model (SM) coupling and DM mass respectively.
Strikingly, electroweak-size couplings and weak scale masses -- a combination generically required to understand the stability of the Fermi scale -- can lead to Eq.~(\ref{eq:thermalxs}), a coincidence denoted in the literature as the ``WIMP miracle".
However, there are well-known exceptions to this possibility.
These alternative mechanisms for obtaining the correct $\langle \sigma v\rangle_{\rm ann}$ have gained traction in recent decades in light of stringent limits placed on the minimal WIMP scenario.
Let us briefly review the mechanisms prominent in the literature:
\begin{itemize}
\item The combination $\alpha^2/M^2$ may be fixed by a choice of coupling strength and mass that need not be related to the weak scale, the so-called {\bf WIMPless miracle} \cite{Feng:2008ya}.
\item The usual Lee-Weinberg calculation does not apply to the ``{\bf semi-annihilation}" process $\chi_i \chi_j \rightarrow \chi_k {\rm S}$, where $S$ is a singlet under DM symmetries.
This process must be included in the computation of freezeout if allowed by the model \cite{D'Eramo:2010ep}.
\item $3 \to 2$ processes within the dark sector, as opposed to $2 \to 2$ DM annihilation to SM particles, may be the dominant DM number-changing process.
This occurs for {\bf SIMP DM}~\cite{Hochberg:2014dra, Hochberg:2014kqa}.
These scenarios require a DM-SM connection in order to be viable and, unlike the other exceptions listed here, DM in this case is automatically self-interacting.
\item An effective $\langle \sigma v\rangle_{\rm ann}$ that enters the computation of DM relic density may be set by multiple processes.
In the mechanism of {\bf co-annihilation}, all possible annihilation combinations between DM and nearly mass-degenerate states must be included \cite{Griest:1990kh,Gondolo:1990dk}.
\item In $s-$channel-mediated annihilations, the amplitude increases near a pole, enhancing $\langle \sigma v\rangle_{\rm ann} \sim |\mathcal{M}|^2$.
This is the so-called {\bf resonant annihilation} \cite{Griest:1990kh,Gondolo:1990dk}.
\item The thermal averaging of the annihilation cross-section may play a non-trivial role \cite{Griest:1990kh,Gondolo:1990dk}.
In particular, if the products of DM annihilation are heavier than DM, their limited phase space truncates the range of DM velocites over which the averaging is performed.
In such cases, the required value of $\langle \sigma v\rangle_{\rm ann}$ can be achieved if the non-thermal cross-section is large to begin with.
Since this type of annihilation cannot proceed at zero temperature, this class of models is referred to as ``{\bf forbidden dark matter}".
\end{itemize}
Currently, extensive DM searches through direct and indirect detection, and collider production, target the electroweak scale.
These searches are usually complementary and capable of probing the mechanism underlying DM freezeout.
While considerable attention has been paid to phenomenological implications of semi-annihilation, co-annihilation and resonant annihilation at the weak scale, the same cannot be said for forbidden annihilation.
This is surprising, given that DM annihilation to heavier states is a minimal deviation from standard WIMP freezeout.
In the quest of demystifying the nature of DM,
it is imperative to explore the full landscape of abundance-setting mechanisms.
The viability of the forbidden mechanism at the electroweak scale, a scenario that may be loosely described as forbidden WIMPs, is the primary concern of this paper.
The rest of the paper is set up as follows.
Sec.~\ref{sec:model-building} first reviews the literature of forbidden DM models in order to distinguish our work.
It then deals with aspects of building forbidden WIMP simplified models and shows that the top quark portal is the only viable model if we require $\Omega_\chi h^2$ to be set primarily by a forbidden mechanism.
Constraints on top portal models and future search sensitivities are studied in Sec.~\ref{sec:topportal}.
Sec.~\ref{sec:discs} provides some discussions and concludes the paper.
\section{Model-Building Aspects}
\label{sec:model-building}
\subsection{Literature Review}
\label{subsec:litreview}
{We begin by pointing out the principal difference between previous work on forbidden DM and our paper.
Whereas previous authors targeted indirect detection as the main means of probing their models,
our models cater exclusively to colliders and direct detection experiments.
In the following we will expand on this statement.
Forbidden DM was studied in \cite{Jackson:2009kg,Jackson:2013pjq,Jackson:2013rqp}
in the context of line signals in the sky.
In this series of papers, the SM gauge group was extended to include a $U(1)'$ sector, with its gauge boson $Z'$ mediating DM-SM interactions through interactions with the top.
For DM lighter than the $Z'$ and top quark, annihilation into the corresponding channels is suppressed, in turn quelling continuum photon emission.
This gives way to the domination of line signals produced by, e.g., $\gamma \gamma, \gamma h, \gamma Z$ final states from diagrams involving loops of the top quark and/or new fermion partners.
When the DM relic density was computed, it was by taking these final states into account; the contribution from forbidden final states was not included.
Forbidden DM was invoked again by the authors of \cite{Tulin:2012uq} to explain the (former) anomaly of the 130 GeV Fermi line.
A new vector-like fermion with mass $\geq 130$~GeV was introduced as the state to which DM annihilated, and a singlet pseudoscalar played the role of mediator.
The relic density was not considered.
More recently, Ref.~\cite{D'Agnolo:2015koa} re-introduced forbidden DM into the literature; here, the final state of annihilation is a massive dark photon, the gauge mediator of a hidden $U(1)'$ sector.
Emphasis was laid on a new calculation of the relic density.
Constraints on this model arise from possible gauge kinetic mixing, mainly from beam dump experiments and observations of the CMB and supernovae cooling.
Our approach to forbidden DM will differ from these models in several important respects:
\begin{itemize}[label={}]
\item(i) In keeping with minimality, we do not extend the SM gauge group.
Instead, we consider ``simplified models", effective low-energy theories that capture DM-SM interactions via dimension-4 Lagrangian terms
\cite{Chang:2013oia,
An:2013xka,
Bai:2013iqa,
DiFranzo:2013vra,
Papucci:2014iwa,
Garny:2014waa,
Abdallah:2014hon,
Buckley:2014fba,
Harris:2014hga,
Abdallah:2015ter,
Altmannshofer:2014cla,
Baker:2015qna}.
In these models the new physics sector usually comprises of no more than the DM state and a mediator.
The DM is charged odd under a $Z_2$ parity to ensure its stability.
Crucially, these models are highly sensitive to collider and direct detection experiments through the couplings of the mediator.
For a recent review of simplified models vis-a-vis LHC searches, see \cite{Abdallah:2015ter}.
We will study two different classes of simplified models, involving mediators either in the $s$- or $t$-channel.
These models are described in full detail in Sec.~\ref{sec:topportal}.
\item (ii) The DM abundance in our framework is set by forbidden annihilation to Standard Model states.
In contrast, the abundance in \cite{Jackson:2009kg,Jackson:2013pjq,Jackson:2013rqp} was set by allowed annihilations to (lighter-than-DM) SM states.
In the instances where forbidden annihilation was considered, the abundance was not set to the observed value.
In ~\cite{Tulin:2012uq} and \cite{D'Agnolo:2015koa}, the annihilation products were exotic particles.
Enforcing the condition that DM annihilate to SM has two virtues.
Since all the SM masses are known, we are able to obtain a fair indication of where the DM mass might lie.
Secondly, by not allowing DM to annihilate to mediator final states, we insure communication between the dark and luminous sectors, which is crucial for probing the forbidden mechanism under the lampposts of LHC and direct detection.
\item (iii)
Thermally averaged forbidden cross-sections are evaluated by integrating over the tail of the DM velocity distribution, usually assumed Maxwell-Boltzmann.
Parametrically, $\langle \sigma v\rangle_{\rm ann} \sim \sigma_{\rm ann} \exp(- k \Delta M/T)$, where $\sigma_{\rm ann}$ is the unaveraged cross-section, $\Delta M$ the mass deficit of DM w.r.t. the final state and $k$ a constant.
In order to obtain Eq.~(\ref{eq:thermalxs}), a large $\sigma_{\rm ann}$ is required to overcome the exponential Boltzmann suppression.
This was achieved in \cite{Jackson:2009kg,Jackson:2013pjq,Jackson:2013rqp,Tulin:2012uq} with an $s$-channel mediator so that the annihilation could transpire near a pole, bolstering $\sigma_{\rm ann}$.
In \cite{D'Agnolo:2015koa}, $\sigma_{\rm ann} \sim \alpha^2/M^2$ was enhanced by diminishing the mass scale of the dark system $M$.
A mediator lighter than 10 GeV was considered, with focus on the MeV scale.
In our approach, we will tackle this issue by two means.
First, in simplified models with an $s$-channel mediator, annihilation near a pole can overcome Boltzmann suppression {\em \`a la} \cite{Tulin:2012uq}.
Second, in simplified models with a $t$-channel mediator, where the interactions take the form $\mathcal{L} \supset \lambda \rm{(SM)(DM)(mediator)}$, the coupling $\lambda$ can be large ($>$ 1) as long as it is perturbative.
This is legitimate since simplified models are explicitly constructed as effective low-energy theories; in fact, large couplings were demonstrated to be a \textit{requisite} for obtaining the correct relic density \cite{Chang:2013oia,An:2013xka,Altmannshofer:2014cla} and for saturating collider bounds \cite{DiFranzo:2013vra}.
Sizeable couplings can then make the annihilation overcome Boltzmann suppression.
This amounts to saying that while \cite{D'Agnolo:2015koa} increased the ratio $\alpha^2/M^2$ by lowering $M$, it is here accomplished by raising $\alpha$.
\end{itemize}
Traditionally, $\langle \sigma v\rangle_{\rm ann}$ in the forbidden set-up is computed by weighting $\sigma_{\rm ann}$
with the Maxwell-Boltzmann distribution and carrying out a numerical integration over $v$ from the critical velocity to $\infty$.
As earlier mentioned, Ref.~\cite{D'Agnolo:2015koa} proposed a new algorithm, taking advantage of the principles of unitarity and detailed balance.
By this method, the forbidden cross-section may be obtained by computing the cross-section of the reverse process, SM~SM~$\rightarrow$~DM~DM.
We will adopt it for obtaining $\Omega_\chi h^2$ throughout our paper.
The details of the computation are enlarged in Appendix \ref{app:relabundcalc}.
\subsection{Simplified Forbidden WIMPs}
\label{sec:simpmods}
We now turn to building simplified models for forbidden WIMPs.
In principle, DM annihilation could proceed via some combination of forbidden and allowed processes.
That is, the sum of the final state masses could be greater or smaller than the sum of the DM masses.
In this paper, we will restrict ourselves to a \textit{principally} forbidden WIMP, a WIMP whose allowed annihilation channels are negligible.
We say ``principally", because the latter channels will always exist -- annihilation to light states through loops or three-body annihilations are usually unavoidable, though suppressed.
Our object here is construct a WIMP that is as forbidden as possible.
We do this to demonstrate that experiments are yet to rule out the forbidden mechanism as the chief means of providing the DM abundance\footnote{Forbidden WIMPs -- aren't.}.
The constraints one can impose on such a scenario then give a qualitative picture of a case where forbidden channels dominate the annihilation of DM.
Since we demand SM annihilation channels, we will now work through every massive SM particle and check whether it qualifies as the portal to a principally forbidden WIMP.
\begin{itemize}
\item
{\em Electroweak bosons}
It has been long known that a WIMP charged under the electroweak gauge group and interacting solely with the gauge bosons is ruled out by experiment unless it mixes with an SM singlet \cite{Cohen:2011ec}.
Even so, weak gauge boson final states are unsuitable for our purpose, on the following grounds.
Assuming renormalizable interactions, it is easily checked that any DM annihilation with $Z Z$ (and possible $W W$) final states must have annihilations to light fermions through an $s$-channel $Z$.\footnote{An even stronger argument can be given for non-renormalizable interactions.
For concreteness, assume a scalar DM field $\chi$ that couples to the $Z$ boson through the operator $c_{\chi Z} |\chi|^2 Z_{\mu \nu} Z^{\mu \nu}/\Lambda_{\rm new}^2$
and that, through some tuning, the analogous operator involving photons vanishes at the relevant scale.
Assuming the coefficient $c_{\chi Z}$ is $\mathcal{O}(1)$,
the $\chi$-$\chi$-$Z$-$Z$ coupling ($\equiv \lambda_{\chi Z}$) goes as $p_Z^2/\Lambda_{\rm new}^2$, where $p_Z$ is the $Z$ momentum.
Since the DM mass must be $m_{\rm DM} \lesssim m_Z$ for viable forbidden annihilations, the $Z$ velocity at freezeout should approximate the DM velocity, $v \sim 0.3$.
Thus, $p_Z \simeq m_Z v \simeq 30$ GeV.
For the validity of the contact interaction, $\Lambda_{\rm new}$ must be separated from $m_Z$ by at least an order of magnitude, thus we have
$\lambda_{\chi Z} < (30~{\rm GeV}/900~{\rm GeV})^2 \simeq 10^{-3}$.
From our usual intuition about WIMPs, that the correct abundance for weak scale particles is obtained with electroweak-size couplings, we may be certain that such a feeble coupling as $\lambda_{\chi Z} < 10^{-3}$ would, due to suppressed annihilation, overclose the universe.
If we demand forbidden annihilation, $\langle \sigma v\rangle_{\rm ann}$ is
only further suppressed.
These arguments hold also for fermion DM.
The only remedy here is to boost $\langle \sigma v\rangle_{\rm ann}$ by allowing for annihilations to lighter final states such as the photon.}
This makes electroweak bosons ill-suited to our ends.
\item{\em Higgs boson}
DM interacting with the SM via a Higgs portal are an attractive study for their phenomenological richness.
For DM DM $\rightarrow h h$ to proceed via a forbidden mechanism, we need the DM to be lighter than $m_h =$ 125 GeV by a few percent.
In this range, however, Higgs portal models that do not violate parity are in tension with limits placed by LUX and PandaX-II on DM-nucleon scattering \cite{LopezHonorez:2012kv,Craig:2014lda}.
This is the main reason why we will not pursue forbidden WIMP model-building via the higgs portal.
A second reason may be given: as in the case of annihilation to weak bosons, one finds that an annihilation to Higgs bosons is always accompanied by $s$-channel Higgs-mediated annihilations
to other SM states, in this case light fermions and (one or more off-shell) gauge bosons.
We remark however that these channels are still suppressed due to the smallness of the Higgs Yukawa couplings and phase space respectively.
\item{\em Light fermions}
If we expect forbidden DM to annihilate to the bottom quark and/or lighter quarks and leptons, we expect
the mediator(s) to be in a mass range of at most $\mathcal{O}(10~\rm GeV)$ --
heavier mediators would unduly suppress the annihilation.
Moreover, to enhance the forbidden cross-section the DM-SM couplings needed are usually large.
For these mediator masses and couplings, $s$-channel mediators would have been observed as resonances in di-fermion final states at colliders.
Similarly, $t$-channel mediators of mass $\leq$ 104 GeV would have been seen at LEP searches \cite{LEPcino,LEPslep} in missing energy signatures.
The lack of both these signals disfavors the possibility of light fermions being our annihilation products.
One potential exception to this argument is a DM particle coupling to the right-handed bottom quark via a colored mediator heavier than $104$~GeV, evading LEP searches.
We discuss this in more detail at the end of Sec.~\ref{subsec:tch}.
The parameter space where this scenario is viable is extremely small, and for that reason we do not study it in this work.
\item{\em Top quark}
At the moment, no searches have substantially excluded a top portal DM and its mediators.
Hence we can proceed to construct simplified models involving it, with the proviso that flavor violations are avoided.
Note that even now, as stated in the beginning of this section, we cannot expect the $t$-$\bar{t}$ final state to contribute $100 \%$ to the WIMP annihilation.
Processes involving loops that give the final states $b$-$\bar{b}, Z h, \gamma \gamma$, gluon-gluon, etc. are always present.
However, we find these to be highly sub-dominant to the the $t$-$\bar{t}$ channel due to loop factors and/or momentum-dependent vertices.
Three-body final states such as $tWb$ may also be at play \cite{Chen:1998dp,Hosotani:2009jk,Yaguna:2010hn,Jackson:2013rqp}, but are strongly phase-space suppressed for the DM-top mass splittings we consider.
We will illustrate this further in Sec.~\ref{subsec:3body}.
As we will discuss shortly, in models with $s$-channel mediation, the imposition of Minimal Flavor Violation (MFV) leads to light quark channels.
These are nevertheless negligible in comparison to the top quark channel due to their small Yukawa couplings.
\end{itemize}
\begin{figure*}
\begin{center}
\includegraphics[width=.7\textwidth]{Figures/Feyn}
\caption{
DM annihilation to top quarks can proceed via $s$-channel and $t$-channel processes.
For the $s$-channel the $\lambda_t$ vertex restricts $\widetilde{a}$ to be spin-0, whereas $\chi$ can be spin-0 or spin-1/2,
to avoid new sources of CP violation, we work with only spin-1/2 DM.
For the $t$-channel the spin of $\widetilde{T}$ (0 or 1/2) is fixed by that of $\chi$.
}
\label{fig:Feyn}
\end{center}
\end{figure*}
\section{Top Portal Forbidden WIMP}
\label{sec:topportal}
WIMPs can annihilate to top quark pairs via either an $s$-channel or $t$-channel mediator,
shown schematically in Fig.~\ref{fig:Feyn}.
The vertices in these diagrams capture all the interactions relevant to our phenomenology;
the exact Lagrangians are provided in Appendix~\ref{app:formulae}.
In the t-channel case the DM particle $\chi$ could either be spin-0 or spin-1/2, and the mediator $\widetilde{T}$ is correspondingly spin-1/2 or spin-0.
On the other hand, the $s$-channel mediator
$\widetilde{a}$ is always assumed spin-0 because it couples to SM fermion pairs.
We do not consider vector DM and mediators in this work, in keeping with our stance of not extending the SM gauge group.
A summary of the field content of our models is provided in Table~\ref{tab:fieldcontent}.
\begin{table*}[hbt]
\begin{center}
\begin{tabular}{c c c c c | c }
& & & \multicolumn{3}{c}{Spins} \\
\hline
Field | & $SU(3)_c \otimes SU(2)_W \otimes U(1)_Y$ & | $Z_2$ | & \multicolumn{2}{c}{$t$-channel}| &| $s$-channel
\\
\hline \hline
$\chi$ & ({\bf 1,1,0}) & {\bf-1} & {\bf1/2} & {\bf0} & {\bf1/2} \\
$\widetilde{T}$ & ({\bf 3,1,2/3}) & {\bf-1} & {\bf0} &{\bf1/2} &\\
$\widetilde{a}$ & ({\bf 1,1,0}) & {\bf+1} & & & {\bf0} \\ \hline
& & Free parameters | & \multicolumn{2}{c}{$\lambda, m_{\rm DM}, m_{\rm MED}$} | & | $\lambda_t, \lambda_\chi, m_{\rm DM}, m_{\rm MED}$ \\\end{tabular}
\end{center}
\caption{Summary of the field content of our models.}
\label{tab:fieldcontent}
\end{table*}
Of immediate concern at this point are constraints from flavor physics.
The fields introduced in these models can potentially induce dangerous flavor violating processes through loops.
To mitigate such effects, we invoke the principle of Minimal Flavor Violation (MFV) \cite{Hall:1990ac,Chivukula:1987py,Buras:2000dm,D'Ambrosio:2002ex} by which the flavor group is only broken by Yukawa spurions.
In the $s$-channel model, this means the singlet mediator couples to SM fermions proportional to their Higgs Yukawa strengths.
The
couplings to the down and lepton sectors
can be set to zero without offending MFV.
It was shown in Ref.~\cite{Dolan:2014ska} that even with MFV imposed, significant limits from flavor physics are incurred for DM lighter than 10 GeV.
This is not of concern for us, since our DM mass is a few GeV below the top quark.
In the $t$-channel model, one can imagine vertices analogous to those in Fig.~\ref{fig:Feyn} involving two other generations of mediators.
Then MFV implies these couplings are flavor-diagonal such that every quark flavor communicates to $\chi$ via a unique mediator.
The effects of the ``up'' and ``charm" mediator can then be decoupled by setting them heavy.
Alternatively, flavor violation may be avoided in both models by assuming that the new physics couplings are aligned with the Yukawa matrices such that, in the mass basis, we are only left with couplings to the top quark.
The interesting possibility of DM, rather than the mediator, charged under the top flavor has been considered in Ref.~\cite{Kilic:2015vka}.
More general simplified models with DM carrying flavor indices are explored in Refs.~\cite{Kile:2011mn,Agrawal:2011ze}.
The key difference between the mediators in our models is their $Z_2$ charge.
From their interaction vertices in Fig.~\ref{fig:Feyn}, it can be seen that the $s$-channel ($t$-channel) mediator must be charged $Z_2 = +1 (-1)$.
Hence $\widetilde{a}$ can potentially mix with the SM Higgs boson while $\widetilde{T}$ has no mixing with the SM.
Notice also the difference in the number of free model parameters.
In the first case, four are needed: the DM and mediator masses $m_{\rm DM}$ and $m_{\rm MED}$, and the DM-DM-mediator and top-antitop-mediator couplings $\lambda_\chi$ and $\lambda_t$.
The combination $\lambda \equiv \sqrt{\lambda_t \lambda_\chi}$ appears in all relevant cross-sections.
Moreover, the decay width of the mediator, $\Gamma_{\rm MED}$, usually plays a non-trivial role in $s$-channel annihilation and collider phenomenology.
Consequently the four free parameters are traded for $m_{\rm DM}$, $m_{\rm MED}$, $\lambda \equiv \sqrt{\lambda_\chi\lambda_t}$ and $\Gamma_{\rm MED}$.
In $t$-channel mediation, three parameters describe the model: the DM and mediator masses $m_{\rm DM}$ and $m_{\rm MED}$, and the DM-top-mediator coupling strength $\lambda$.
The final difference between the scenarios is how they achieve the right relic abundance:
as mentioned in Sec.~\ref{subsec:litreview}, forbidden annihilation is possible through the mediator $\widetilde{a}$ due to cross-section enhancement near a pole, and through the mediator $\widetilde{T}$ due to putatively large $\lambda$.
On account of the differences detailed above, these two scenarios will result in very different phenomenologies.
Hence we will consider each possibility separately in the following subsections.
Since our approach takes forbidden annihilation as the sole setter of DM abundance, we will fix $\lambda$ using the observed $\Omega_\chi h^2 = 0.1197$ when presenting our constraints.
This is done using the formulae presented in Appendices~\ref{app:relabundcalc} and \ref{app:formulae}, and comparing with {\tt MicrOmegas} 4.3 \cite{MuOmega} to double-check the result.
This method also allows all constraints to be shown in the space of mediator mass versus DM mass.
This is the treatment of simplified DM models adopted by \cite{Chang:2013oia}.
In the following two sub-sections we will consider in more detail the $s$- and $t$-channel models, and the relevant phenomenology.
\begin{figure}
\begin{center}
\includegraphics[width=.45\textwidth]{Figures/schFeyns}
\caption{
Example diagrams contributing to the monojet + MET signal in the $s$-channel model.
For a CP-odd $\widetilde{a}$, this cross-section is always higher than other channels like $t$-$\bar{t}$ and $b$-$\bar{b}$, yielding stricter bounds.}
\label{fig:schFeyns}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=.45\textwidth]{Figures/sch}
\caption{
Limits on the $s$-channel model.
The coupling $\lambda$ is fixed here with the observed DM abundance; the brown curves denote these contours of $\lambda$.
The blank region is where $\lambda > 1.12$ and is thus non-perturbative;
see the text for how this bound is estimated.
The area under the blue curve is excluded at 95\% C.L. by monojet searches at the 8 TeV LHC with $\mathcal{L} = 20~{\rm fb}^{-1}$.
The bands around the curve denote uncertainties due to renormalization and factorization scales.
The area under the magenta curve is coverable by the 95\% C.L. projection limits provided by ATLAS for $\sqrt{s}$ = 14 TeV and $\mathcal{L} = 3000~{\rm fb}^{-1}$.
On account of adopting an EFT treatment, these monojet bounds are more stringent than true limits.
See the text for more details.
In the region bounded by the dashed curves, the determination of $\lambda$ by Taylor-expanding $\langle \sigma v\rangle_{\rm ann}$ in $v$ may be inaccurate.}
\label{fig:sch}
\end{center}
\end{figure}
\subsection{$s$-channel mediator}
\label{subsec:sch}
Since we want $\chi$ to annihilate to SM, we must ensure that the channel $\chi \chi \rightarrow \widetilde{a} \widetilde{a}$ is closed.
Therefore we impose $m_{\rm MED} > m_t$.
The mediator $\widetilde{a}$ can generically have both scalar and pseudoscalar couplings to SM fermions, which would explicitly violate CP and confront strict constraints from neutron EDM measurements.
For this reason, $\widetilde{a}$ is usually taken to be a CP eigenstate (and all couplings taken real).
If we admit only CP-even couplings, spin-independent direct detection limits nearly exclude the space of our parameters.
The only region spared is in the neighborhood of $m_{\rm MED} \simeq 2 m_t$, where resonant DM annihilation allows
$\lambda$ to be small.
CP-odd couplings, on the other hand, are well-shielded from spin-independent direct detection searches \cite{Freytsis:2010ne,Dienes:2013xya}, leading to more interesting phenomenology.
Consequently, we will take $\widetilde{a}$ to be a pure pseudoscalar.
This choice restricts the spin of our DM.
If $\chi$ is a scalar, its couplings to $\widetilde{a}$ would introduce a new source of large CP violation.
Therefore, we take $\chi$ to be a Dirac fermion; the alternative choice of a Majorana fermion would lead to similar results.
We now turn to the constraints on this model.
The strongest limits are placed by collider searches for monojets and missing momentum \cite{Buckley:2014fba,
Haisch:2012kf,
Fox:2012ru,
Lin:2013sca,
Haisch:2013fla,
Haisch:2015ioa,
Arina:2016cqj}.
This signature is generated by the production of $\widetilde{a}$ and radiation of a gluon, such as in the Feynman diagrams in Fig.~\ref{fig:schFeyns}.
A potentially competitive probe is $t\bar{t} + \slashed{E}_T$ final states, but as pointed out in \cite{Haisch:2015ioa}, is always weaker than the monojet search when $\widetilde{a}$ is CP-odd.
A $t\bar{t}$ resonance via the one-loop process $gg \rightarrow \widetilde{a} \rightarrow t\bar{t}$ can modify top-pair production by $\mathcal{O}(1\%)$ \cite{Haisch:2013fla}, but is unlikely to be found at the LHC since the theoretical uncertainty on the cross-section is already $\sim 5 \%$~\cite{Czakon:2013goa}.
A dijet resonance via the two-loop process $gg \rightarrow \widetilde{a} \rightarrow gg$ is also possible, but again unresolvable due to loop suppression~\cite{Haisch:2013fla}.
The imposition of MFV implies that the couplings of $\widetilde{a}$ to the lighter quarks are suppressed with respect to $\lambda_t$ by factors of $m_q/m_t$.
Therefore, they are beyond the sensitivity of LHC searches.
See also Ref.~\cite{Dolan:2016qvg}, which performs a sensitivity study of the mediator in trying to determine its quantum spin and CP properties.
The introduction of the singlet mediator could have potentially given rise to mixing in the Higgs sector, invoking limits from Higgs coupling measurements at the LHC.
However, by choosing $\widetilde{a}$ to be CP-odd, mixing with the SM Higgs is pre-empted.
Finally, no limits from direct detection experiments apply to the region of parameters considered here
since both the spin-dependent and spin-independent cross-sections generated are too small to be probed.
The limits are sketched in Fig.~\ref{fig:sch} on the $m_{\rm DM}-m_{\rm MED}$ plane.
Here $\lambda$ is fixed by requiring $\Omega_\chi h^2 = 0.1197$.
Contours of $\lambda$ thus obtained are represented by brown curves.
There is a value of $\lambda$ above which the theory becomes non-perturbative, which we estimate as follows.
As mentioned in Appendix~\ref{app:formulae}, the $\lambda_t\widetilde{a}\bar{t}t$ vertex must arise from a term such as $c H \phi\bar{Q}_3 t^c/\Lambda$ in the unbroken electroweak phase, where $c$ is an $O(1)$ coefficient, $\Lambda$ is a new physics scale and $\phi$ is a complex scalar containing $\widetilde{a}$.
The coupling $\lambda_t$ is then identified with $c\cdot v/\Lambda$, where $v$ is the Higgs vacuum expectation value.
For the effective theory to be valid, we require $v/\Lambda$ to be at most $\sim 1/10$.
This fixes the maximum size of $\lambda_t$, namely 0.1.
No such constraint applies for $\lambda_\chi$, which can take on its maximum allowable perturbative size of $4\pi$.
Together, we obtain the upper bound $\lambda \leq \sqrt{(0.1)(4\pi)} \approx 1.12$.
The blank region in Fig.~\ref{fig:sch} is where $\lambda > 1.12$, becoming non-perturbative and invalidating the effective theory.
In the range of paramaters shown, we varied $\Gamma_{\rm MED}$ subject to $\lambda = \lambda_t \lambda_\chi$, and find no observable effect on $\Omega_\chi h^2$.
(Expressions for $\Gamma_{\rm MED}$ in terms of other model parameters can be found in \cite{Haisch:2015ioa,Abdallah:2015ter}.)
Therefore, we do not present constraints on $\Gamma_{\rm MED}$.
The blue curve denotes the 95\% C.L. exclusion from the CMS monojet + $\slashed{E}_T$ search at $\sqrt{s} = 8$ TeV with $\mathcal{L} = 20~{\rm fb}^{-1}$ \cite{CMS8TeVmonojet}.
We derive this curve by recasting the limits obtained in \cite{Haisch:2015ioa}, which had used an EFT description for setting bounds: the DM-top interaction was assumed a contact operator, with the mediator $\widetilde{a}$ integrated out.
The EFT suppression scale is given by
\begin{equation*}
\Lambda_{\rm EFT} = \left(\frac{\mtopm_{\rm MED}^2}{\lambda^2}\right)^{1/3}~.
\end{equation*}
The limit obtained was $\Lambda_{\rm EFT} = 170$ GeV.
The common renormalization and factorization scale in this analysis is $\mu_Q \equiv\left(\sqrt{m^2_{\chi \chi}+p^2_{T,j_1}} + p_{T,j_1}\right)/2$, where $m_{\chi \chi}$ is the $\chi$-$\bar{\chi}$ invariant mass and $p_{T,j_1}$ is the $p_T$ of the hardest jet in an event.
The bands around the blue curve corresponds to the theoretical uncertainty from varying this scale up and down by a factor of 2.
The range of uncertainty thus obtained in the limits is $\Lambda_{\rm EFT} \in [160,185]$ GeV.
These bounds are conservative on account of assuming a contact interaction between the DM and top.
This is best understood by inspecting the propagator in both the full theory and the EFT treatment.
If the full theory is used, the cross-section is suppressed when the propagator goes off-shell, which may happen for a tight cut on $\slashed{E}_T$.
On the contrary, in the EFT treatment, cutting hard on $\slashed{E}_T$ leads to a gross {\em overestimate} of the cross-section.
This happens because a tight cut on $\slashed{E}_T$ increases the energy of the final state, or the energy running through the loop, making it approach the suppression scale $\Lambda_{\rm EFT}$.
The EFT treatment then becomes inaccurate and does not capture the suppression of production rates.
See \cite{Fox:2012ru} and \cite{Haisch:2015ioa} for studies on the differences between an EFT approach and treating the mediator as an active degree of freedom.
The bound in Fig.~\ref{fig:sch} was obtained by imposing $\slashed{E}_T > 450~$GeV \cite{Haisch:2015ioa}.
Therefore the propagator would have been sent off-shell in the full theory and the cross-section diminished.
This has had the effect of rendering the EFT approach very crude.
Consequently, any bound in the region $m_{\rm MED} < 450$ GeV must be interpreted with extreme caution, and the true limit remembered to be weaker.
We also recast the future sensitivity of the monojet + $\slashed{E}_T$ search at $\sqrt{s}$ =14 TeV and $\mathcal{L} = 3000~{\rm fb}^{-1}$, keeping with the EFT approach.
This limit had been obtained from an ATLAS sensitivity study for this energy and luminosity \cite{ATLAS14TeVprospex}.
We denote it by the magenta curve, which corresponds to $\Lambda_{\rm EFT} = 250$ GeV.
The renormalization and factorization scale here is $2\mu_Q$, and the scale uncertainty was not presented in this case.
We notice from this plot that, even with the most optimistic reach of the LHC, the parameter space of this model is not probed well.
One concludes that much room is left for the $s$-channel forbidden WIMP scenario.
One possibly better means to probe the uncovered regions would be a future 100 TeV collider.
Whether monojet production or $t$-$\bar{t}$ production would be the better probe can only be answered with a detailed analysis involving well-chosen cuts.
Such an analysis is beyond our current scope.
We re-emphasize that, due to the EFT treatment, the reach denoted by the magenta curve is merely an order-of-magnitude estimate,
and must be interpreted as an upper limit on the actual reach.
Finally, a remark on the accuracy of the above limits is in order.
The Taylor expansion $\langle \sigma v\rangle_{\rm ann} \approx a + b~v^2$ may not be entirely valid close to the pole of the annihilation.
The discrepancy between naive analytical approximations and a full numerical treatment in computing $\langle \sigma v\rangle_{\rm ann}$ near the pole is described in \cite{Griest:1990kh} and \cite{Gondolo:1990dk}.
Here ``near the pole" means the region within about 10\% of the pole mass.
We indicate this region with dashed lines in Fig.~\ref{fig:sch}, which encompasses $m_{\rm MED} \in [310,390]$ GeV.\footnote{The resonant peak in the thermal cross-section with an $s$-channel mediator is usually near twice the DM mass.
In the case of forbidden DM, it is near twice the mass of the final state.
This is because in forbidden DM phase space, most of the annihilation comes from DM momenta near the annihilation threshold (which in our case is $2 m_t$).
In contrast, usual DM annihilation proceeds mostly near zero DM momentum.}
Our estimate of $\lambda$ in this range is unreliable and smaller than the actual value.
The discrepancy between the Taylor expansion and the full treatment grows starker with diminishing $\Gamma_{\rm MED}/m_{\rm MED}$ \cite{Griest:1990kh,Gondolo:1990dk}, which can happen in our scenario when $m_{\rm DM} > 155$ GeV and the decay mode $\widetilde{a} \rightarrow \chi \chi$ is absent. Fortunately, the area bounded by the 8 TeV monojet search does not fall within this region.
Taking into account the above limits and considerations, we find that the allowed range of parameters covers almost the entire region where the perturbativity bound $\lambda < 1.12$ is imposed.
Specifically, when perturbative, the model is allowed to reside in the mass ranges $m_{\rm DM} \in [145, 170]$ GeV and $m_{\rm MED} \in [175, 725]$ GeV.
As one expects, the allowed range of DM masses is at its largest in the ``funnel region", $m_{\rm MED} \sim 2 m_t$.
In the next sub-section we explore the viability and limits of the forbidden mechanism with a $t$-channel mediator.
\begin{figure*}
\begin{center}
\includegraphics[width=.45\textwidth]{Figures/tchMaj}
\quad \quad
\includegraphics[width=.45\textwidth]{Figures/tchReal}
\caption{
Limits on the $t$-channel models, with Majorana fermion (real scalar) DM on the left (right).
The coupling $\lambda$ is fixed with the observed DM abundance; the brown curves represent contours of this coupling.
The blank region corresponds to $\lambda > 3$.
The blue shaded regions are excluded at 90\% C.L. by LUX, the gray shaded regions excluded at 90\% C.L. by PandaX-II, and the magenta curve denotes the 90\% C.L. future sensitivity reach of XENON1T.
The red shaded regions are excluded at 95\% C.L. by top squark searches at the LHC;
the dashed green curve on the left denotes the 5$\sigma$ reach projected by ATLAS at $\sqrt{s} = 14$ TeV and $\mathcal{L} = 3000~{\rm fb}^{-1}$.
The black dashed curves on the left encompass the 5$\sigma$ reach at $\sqrt{s} = 13$ TeV and $\mathcal{L} = 3000~{\rm fb}^{-1}$ claimed by the strategy presented in \cite{An:2015uwa} for the ``compressed region" of stops.
For more details, see the text.}
\label{fig:tch}
\end{center}
\end{figure*}
\subsection{$t$-channel mediator}
\label{subsec:tch}
Here we consider two models, one with a Majorana fermion DM and a complex scalar mediator,
and another with a real scalar DM and Dirac fermion mediator.
To avoid confusion over terminology, we label these models ``{\tt tchFDM}" and ``{\tt tchSDM}" respectively.
In {\tt tchFDM}, the choice between Majorana and Dirac DM is slight.
As explained e.g. in \cite{Chang:2013oia}, Majorana DM annihilation picks up a chirality flip in the $s$-wave, so that $\langle \sigma v\rangle_{\rm Maj} \approx (m_f/m_{\rm DM})^2 \langle \sigma v\rangle_{\rm Dirac}$, where $m_f$ is the mass of the annihilation products.
Since in our case $m_f = m_t$ is close to $m_{\rm DM}$, the results for Majorana and Dirac are found to be qualitatively the same.
In {\tt tchSDM},
the dominant $s$-wave contribution is four times larger for a real scalar DM than for a complex scalar DM.
Since the annihilation cross-section scales as $\lambda^4$, the couplings are correspondingly only $4^{1/4} \simeq 1.4$ times smaller in the former case.
Therefore no qualitative difference exists between the use of a real and complex scalar for DM.
The constraints on these models come from collider and direct DM detection experiments.
Searches for top superpartners (stops) can be directly recast to our scenario.
The relevant search here is for pair-production of stops followed by decay to a pair of top quarks + $\slashed{E}_T$.
This signature applies to our model since at the LHC our mediator, like the stop, is QCD-produced and decays to top quark + DM.
The case of the scalar mediator of {\tt tchFDM} ~is particularly interesting.
In both {\tt tchFDM} ~and the MSSM, the top-flavored scalar is produced entirely through QCD and the BR = 100\%.\footnote{This is not the case had other mediator flavors been present, e.g., the production of the first two generations of mediators would involve processes with $t$-channel DM exchange, at variance with similar processes of squark production because of the difference in couplings.}
Hence, due to identicality of production and decay, the exclusion region for the scalar mediator of {\tt tchFDM} ~is congruent to that of the top superpartner (for LSP masses that range in our DM masses).
However, the rate of the fermion mediator production through QCD can be different in {\tt tchSDM}, as we explain below.
Therefore to set bounds, we first generate the mediator pair production process $pp \rightarrow \widetilde{T} \bar{\widetilde{T}}$ in {\tt MadGraph5} \cite{Alwall:2014hca} and obtain the production cross-sections at leading order.
Then we apply a $K$-factor of 1.5 (see \cite{Anandakrishnan:2015yfa}) to obtain the NLO cross-section.
Assuming similar acceptances for the production of a stop pair and $\widetilde{T} \bar{\widetilde{T}}$, we compare our rates with the 95\% C.L. exclusion cross-sections provided in \cite{CMS:2014wsa} and find the limits on $m_{\rm MED}$.
DM in {\tt tchSDM}, {\tt tchFDM}\, can scatter with nuclei through gluon loops
(see \cite{Hisano:2015bma} for all the attendant Feynman diagrams),
contributing to spin-independent direct detection rates.
These limits are shown on the $m_{\rm DM}-m_{\rm MED}$ plane in Fig.~\ref{fig:tch}, where the plot on the left (right) corresponds to {\tt tchFDM}~({\tt tchSDM}).
As in our treatment of the $s$-channel model, we fix $\lambda$ throughout these plots by the requirement $\Omega_\chi h^2 = 0.1197$ and denote these contours of $\lambda$ with brown curves.
The blank region in both plots is where $\lambda \geq 3$, violating perturbativity of the coupling.\footnote{Here we obtain this value from a conservative perturbativity condition: $\lambda^2/(8\pi^2) \lesssim 0.1$.
Furthermore, for $\lambda > 3$, its rapid RG running produces a Landau pole below a scale of 10 TeV, that we wish to avoid.}
In the plot we require $m_{\rm MED} > m_{\rm DM}$ since we require $\chi$ to be the lightest field charged odd under $Z_2$.
While nothing prevents $m_{\rm DM} < m_{\rm MED} < m_t$, we choose to present limits in the region $m_{\rm MED} > m_t$.
This is the range relevant to LHC searches in the $\widetilde{T} \rightarrow \chi t$ channel; in addition, the effects of co-annihilation between $\widetilde{T}$ and $\chi$ are mitigated by separating $m_{\rm MED}$ and $m_{\rm DM}$.
Although co-annihilations are in principle allowed in our models, we wish to focus solely on the effects of the forbidden mechanism.
We now describe the various curves in Fig.~\ref{fig:tch}. The red shaded region in both plots is ruled out at the 95\% C.L. limit by ATLAS stop searches at $\sqrt{s} = 8$~TeV with $\mathcal{L} = 20~{\rm fb}^{-1}$ \cite{Aad:2015pfx}.
CMS sets similar bounds \cite{Khachatryan:2015wza,CMS:2014wsa}.
We now show that the future prospects of these models are optimistic -- most of the parameter space in the perturbative region can be probed in upcoming LHC searches.
The region to the right of the dashed green curve in the left-hand plot of Fig.~\ref{fig:tch} denotes the $5\sigma$ discovery reach at $\sqrt{s} = 14$~TeV and $\mathcal{L} = 3000~{\rm fb}^{-1}$ as projected by ATLAS in \cite{ATLASfuture}.
This reach is provided in the projected search combining 0-lepton and 1-lepton channels.
The other end of this reach is at a mass of $m_{\rm MED}$ = 1400~GeV, well outside our perturbative region.
Projections in these channels have also been made by CMS \cite{CMSfuture} but at a lower luminosity of $\mathcal{L} = 300~{\rm fb}^{-1}$, due to which the reach does not appear as optimistic as the prospects presented by ATLAS.
Since no analogous projection studies have been provided for the fermionic mediator of {\tt tchSDM}, we do not provide their sensitivities.
Another interesting prospect for these models is to hunt for the mediator in the ``compressed" region near $m_{\rm MED} \approx m_{\rm DM} + m_t$.
A strategy dedicated to compressed stop searches was introduced in \cite{Hagiwara:2013tva,An:2015uwa}, exploiting the recoil of ISR jets against the $\slashed{E}_T$ seen in these parametric regions.
In \cite{An:2015uwa} the projected reach for a 5$\sigma$ discovery by this strategy for top squarks at $\sqrt{s} = 13~$TeV and $\mathcal{L} = 3~{\rm ab}^{-1}$ was presented.
We show this on the left-hand plot in the region flanked by the dashed black curves.
The technique can be seen to cover the compressed region all the way down to $m_{\rm MED} = m_{\rm DM} +m_t$.
Analogous sensitivities for {\tt tchSDM}~were not presented, however we may reasonably expect the same technique to cover the entire compressed region again.
Finally, the blue and gray shaded regions in the two plots are ruled out at the 90\% C.L. limit by spin-independent scattering cross-section bounds set by LUX \cite{Akerib:2013tjd} and PandaX-II \cite{Tan:2016zwf} respectively.
The magenta curve on the left-hand plot depicts the future sensitivity (at 90\% C.L. limit ) at the XENON1T experiment \cite{Aprile:xenon1T} proposed to go live in the year 2017.
On the right-hand plot, the corresponding magenta curve would fall outside the range of masses shown.
In the region left uncovered by PandaX-II, the smallest spin-independent scattering cross-section is $\sigma_{\rm SI}^{\rm scal} \simeq 2 \times 10^{-45}~{\rm cm}^2$.
This is about two orders of magnitude larger than the reach of XENON1T, and can thus be comfortably covered by the experiment.
When we compare across the two plots, we notice a difference:
while only a small region in the perturbative region of {\tt tchFDM} ~is seen to be excluded by LUX, a much wider range (upto $m_{\rm MED} \simeq$ 700 GeV) is covered for {\tt tchSDM}.
One understands this difference from the interplay between the annihilation and direct detection cross-sections, as follows.
By approximating $m_{\rm MED} \gg m_t$, the annihilation cross-section for both fermionic and scalar DM can be parametrized as $\sigma_{\rm ann} \sim \lambda^4 m_{\rm DM}^2/m_{\rm MED}^4$.
In the DM-nucleon scattering cross-section, on the other hand, there are two key differences between {\tt tchFDM} ~and {\tt tchSDM}.
First, consider the DM-DM-gluon-gluon effective operators in the two models:
\begin{eqnarray}
\nonumber f_N \frac{\alpha_s}{\pi}\bar{\chi}\chi \mathcal{G}_{\mu \nu}\mathcal{G}^{\mu \nu},~& &{\rm for~{\tt tchFDM},~and} \\
f_N \frac{\alpha_s}{\pi}\chi^2 \mathcal{G}_{\mu \nu}\mathcal{G}^{\mu \nu},~& &{\rm for~{\tt tchSDM}}.
\label{eq:ggDMDMops}
\end{eqnarray}
The Wilson co-efficients $f_N$ scale as
\begin{eqnarray}
\nonumber m_{\rm DM}/m_{\rm MED}^4,~& &{\rm for~{\tt tchFDM},~and} \\
\nonumber 1/m_{\rm MED}^2,~& &{\rm for~{\tt tchSDM}}.
\end{eqnarray}
The difference in the mass dimensions of the $f_N$'s reflects the fact that fermionic and scalar DM fields have different dimensions.
More crucially, a factor of $m_{\rm DM}$ appears in the numerator of the fermion $f_N$.
This factor can be understood by the fact that loop level effects (such as Eq.~(\ref{eq:ggDMDMops})) generated from a renomalizable theory must respect the same symmetries as the underlying theory.
Here, the operator in the {\tt tchFDM}\, case violates chiral symmetry and therefore must vanish as $m_{\rm DM} \to 0$ and chiral symmetry in the underlying theory is restored.
Second, the DM-nucleon scattering cross-section of scalar DM is suppressed in the denominator by a factor of $m_{\rm DM}^2$, which is absent for fermion DM.
This extra scaling arises from the ratio of the matrix elements
\begin{equation*}
\frac{\langle \chi| \chi^2 |\chi \rangle_{\rm scalar}}{\langle \chi | \bar{\chi} \chi | \chi \rangle_{\rm fermion}} \propto \frac{1}{m_{\rm DM}}~.
\end{equation*}
The resulting spin-independent direct detection cross-sections go as
\begin{eqnarray}
\nonumber \sigma_{\rm SI}^{\rm ferm} &\sim& \frac{\lambda^4 m_N^4m_{\rm DM}^2}{m_{\rm MED}^8}; \\
\nonumber \sigma_{\rm SI}^{\rm scal} &\sim& \frac{\lambda^4 m_N^4}{m_{\rm DM}^2m_{\rm MED}^4},
\end{eqnarray}
where $m_N$ is the nucleon mass.
These can be written as
\begin{equation}
\sigma_{\rm SI}^{\rm ferm} \sim \sigma_{\rm ann}m_N^4/m_{\rm MED}^4;~~ \sigma_{\rm SI}^{\rm scal} \sim \sigma_{\rm ann} m_N^4/m_{\rm DM}^4~.
\label{eq:DDannlinks}
\end{equation}
Since we fix the relic abundance to the observed value throughout our plots, we can now write a direct comparison between the fermion and scalar DM scattering cross-sections for a given combination of DM and mediator mass:
$\sigma_{\rm SI}^{\rm ferm} \sim \sigma_{\rm SI}^{\rm scal} (m_{\rm DM}/m_{\rm MED})^4$.
Hence {\tt tchFDM} ~direct detection rates are relatively suppressed, giving us weaker bounds and less optimistic projections.
Eq.~(\ref{eq:DDannlinks}) also helps one understand the shape of the curves in Fig.~\ref{fig:tch}.
We find that the {\tt tchFDM} ~bound falls quicker than {\tt tchSDM} ~with respect to $m_{\rm MED}$.
This is due to the $1/m_{\rm MED}^4$ scaling of $\sigma_{\rm SI}^{\rm ferm}$ that is not seen in $\sigma_{\rm SI}^{\rm scal}$.
At $m_{\rm MED} \gg m_t$, we find the {\tt tchSDM} ~bound to be insensitive to $m_{\rm MED}$.
This is because for a fixed $\sigma_{\rm ann}$, $\sigma_{\rm SI}^{\rm scal}$ does not scale with $m_{\rm MED}$.
}
The collider limits are stronger for fermion mediator pair production in {\tt tchSDM}.
This is because (a) fermions have more spin degrees of freedom than scalars, (b) scalar mediator production ({\em\`a la} stop production) in {\tt tchFDM}~is suppressed by a momentum-dependent coupling in production modes where a gluon mediates in the $s$-channel and $\widetilde{T}$ mediates in the $t$-channel; this is absent in the corresponding fermion mediator production process in {\tt tchSDM}.
The larger production rates of the fermion mediator results in two consequences: at the upper end of the mass bound, the reach is higher, while at the lower end, near the compressed region $m_{\rm MED} \simeq m_{\rm DM}$, the fermion mediator is able to better overcome the dwindling signal acceptance of the search.
Hence we find that while $m_{\rm MED} \in [380,690]$~GeV is excluded in {\tt tchFDM}, the corresponding range is $m_{\rm MED} \in [310,920]$~GeV in {\tt tchSDM}.
The latter is in good agreement with the recasting performed by the authors of \cite{Anandakrishnan:2015yfa} for fermionic top partners that are odd under a parity.
The Majorana DM model has a supersymmetric limit, where $\chi$ is a neutralino LSP of the MSSM, $\widetilde{T}$ is a top squark.
In this limit, the coupling $\lambda$ is of electroweak strength and hence feeble.
Therefore, it may naively preclude forbidden annihilation of neutralino DM as a viable option, but this can be circumvented with co-annihilation effects.
More discussion on co-annihilation and supersymmetry is relegated to the final section of the paper, Sec.~\ref{sec:discs}.
In Sec.~\ref{sec:simpmods} we mentioned the possibility of a ``bottom portal" forbidden WIMP with a mediator evading LEP bounds.
Such a theory can be built in a manner analogous to the models seen in this sub-section, with $t_R \rightarrow b_R$ and the $t$-channel mediator $\widetilde{T} \rightarrow \widetilde{B}$.
One then asks what couplings and DM masses are possible in this scenario.
The heavier $\widetilde{B}$ is, the more inefficient is the annihilation, and hence the stronger is the coupling and/or the more degenerate is $m_{\rm DM}$ with $m_{\rm bottom}$ = 4.2 GeV, if we need to achieve the correct DM abundance.
To maximize the forbidden annihilation cross-section, let us set the mediator mass $m_{\widetilde{B}} = 104$~GeV, which is the edge of the LEP limit~\cite{LEPcino,LEPslep}, and the coupling $\lambda = 3$, its maximum perturbative size.
If we now ask what value of $m_{\rm DM}$ yields the correct DM abundance, we find
\begin{eqnarray}
\nonumber m_{\rm DM} =& 4.0~{\rm GeV},\ &{\rm for\ Majorana\ fermion\ DM}, \\
\nonumber m_{\rm DM} =&~3.95~{\rm GeV},\ &{\rm for\ real\ scalar\ DM}.
\end{eqnarray}
The above DM masses are the minimum required, since larger $m_{\widetilde{B}}$ and/or smaller $\lambda$ values
reduce $\langle \sigma v\rangle_{\rm ann}$, compelling $m_{\rm DM}$ to move closer to $m_{\rm bottom}$ to overcome the Boltzmann suppression.
Thus the range of parameters where this scenario is viable is very limited, and we do not pursue it further.
\subsection{Indirect detection limits}
\label{subsec:indirect}
Constraints on the present-day annihilation cross-section of DM are set by Fermi-LAT \cite{Ackermann:2013yva} by the observation of 15 dwarf galaxies.
For a thermal cross-section $\langle \sigma v\rangle_{\rm ann} = 3 \times 10^{-26}~{\rm cm}^3~{\rm s}^{-1}$, the 95\% C.L. limit is $\lesssim$ 95 GeV in the $b$-$\bar{b}$ final state, which is the strongest constraint.
Our $m_{\rm DM} \supset [145, 170]$~GeV is above this limit.
Moreover the present-day forbidden annihilation to top quark pairs is exponentially suppressed as $\exp(-2(m_t-m_{\rm DM})/T_{\rm today})$, where $T_{\rm today} = 2.7$~K.
This cross-section is several orders of magnitude below $3 \times 10^{-26}~{\rm cm}^3 ~{\rm s}^{-1}$.
There are kinematically ``allowed" final states in our models from loops, which include $\gamma, g, Z$ and $h$.
Ref.~\cite{Jackson:2013rqp} considers these annihilation channels; their cross-sections are enhanced by assuming an $s$-channel vector mediator near resonance, coupled strongly to a vector-like fermion that mixes with the top quark.
In regions where the DM mass is close to the top quark's, the relic density was not calculated accounting for the forbidden mechanism.
This was because the focus of \cite{Jackson:2013rqp} was to demonstrate the possibility of non-trivial indirect detection signals in a forbidden set-up.
However, since our focus is on the relic density, these ``allowed" channels are always safe from indirect detection bounds.
One sees this by considering that our DM freezeout was dictated by forbidden annihilation, in comparison to which the light final state channels are sub-dominant due to loop-suppression.
This automatically insures that the latter have a thermal cross-section much smaller than $3 \times 10^{-26}~{\rm cm}^3 ~{\rm s}^{-1}$, and will therefore never saturate indirect detection bounds.
In a similar manner, the $b$-$\bar{b}$ final state from loops, as well as three-body annihilations to such states as $tWb$, are also expected to be unconstrained by line and diffuse photon searches.
\begin{figure}
\begin{center}
\includegraphics[width=.45\textwidth]{Figures/2vs3body}
\caption{
A comparison of the thermal cross-sections of 2-body ($\chi \chi \rightarrow t \bar{t}$) and the dominant 3-body ($\chi \chi \rightarrow t W b$) final states in our forbidden WIMP set-up.
For illustration, we have chosen the $t$-channel model with fermionic DM and $\lambda = 1$, $m_{\rm MED} = 180$ GeV.
The results are similar for other choices of models and parameters.
In spite of the Boltzmann exponential suppression of forbidden annihilation, the 2-body process dominates over the phase space-suppressed 3-body annihilation for $m_{\rm DM} > 145$ GeV.
This range happens to enclose the region where our $\lambda$ is perturbative (Figs.~\ref{fig:sch},~\ref{fig:tch}).}
\label{fig:2v3}
\end{center}
\end{figure}
\subsection{Three-body final states}
\label{subsec:3body}
In presenting the limits on our models above, we assumed that the thermal cross-section of the $t\bar{t}$ final state overwhelms that of a three-body final state $tWd_i$, where $d_i$ is a down-sector quark.
It is worthwhile to check the accuracy of this assumption.
It was found in \cite{Jackson:2013rqp,Chen:1998dp,Yaguna:2010hn} that for DM masses $m_{\rm DM} \lesssim m_t$, the $\chi\dm \rightarrow tWb$ process can have rates near thermal cross-sections for certain model parameters.
In these models the forbidden annihilation to $t\bar{t}$ was neglected.
In our case, at large $m_t$-$m_{\rm DM}$ splittings it is possible that the Boltzmann exponential suppression of annihilation to $t\bar{t}$ compares to the three-body phase suppression of annihilation to $tWb$.
Using analytic expressions provided in \cite{Chen:1998dp}, and setting $\lambda=1$ and $m_{\rm MED} = 180$ GeV, we compute $\langle \sigma v\rangle_{\rm ann}$ for $tWb$ production and plot it as a function of $m_{\rm DM}$ in Fig.~\ref{fig:2v3} (the orange curve).
For comparison we also plot the $t\bar{t}$ thermal cross-section in green with the same parameters.
We have chosen the $t$-channel model {\tt tchFDM}~ for illustration.
For $m_{\rm DM} > 145$~GeV, the two-body cross-section dominates.
For $m_{\rm DM} \geq 150$ GeV, the three-body rate is about 10\% or less, and has negligible impact on our phenomenology.
We have checked that this is true for all combinations of $\lambda$, $m_{\rm MED}$ and models considered here.
Let us now inspect the implications of this finding for our phenomenology.
We have illustrated before that $m_{\rm DM} > 145$ GeV encloses the region of our parameters where $\lambda$ is perturbative.
It follows that any significant modification the three-body annihilations may have on our experimental limits would be in the mass range $m_{\rm DM} \in [145, 150]$ GeV.
Since an extra annihilation channel is added, its effect is to decrease the $\lambda$ required for obtaining the right abundance, hence leading to weaker constraints.
In Fig.~\ref{fig:sch}, the region in the range $145~{\rm GeV} \leq m_{\rm DM} \leq 150~{\rm GeV}$ falls within the dashed curves, where our estimation of $\lambda$ is already unreliable due to the Taylor expansion.
In the plots of Fig.~\ref{fig:tch}, all the area in the region $145~{\rm GeV} \leq m_{\rm DM} \leq 150~{\rm GeV}$ is already in the non-perturbative region.
These observations indicate that in the region where our constraints apply, three-body final states may be neglected.
\section{Discussion}
\label{sec:discs}
In our work, after having gone through every SM state, we settled on the top quark as the only remaining suitor for a final state that contributes (almost) 100\% to DM annihilation via a forbidden mechanism at the weak scale and couples at the renormalizable level to the DM particle.
And by demanding that DM annihilation be predominantly forbidden, we have obtained the tightest
constraints possible for a forbidden WIMP.
Even before imposing experimental limits, we find the bound from requiring coupling constant perturbativity squeezing our parameter space into a tight region.
Specifically, in the $s$-channel model, we are confined to DM masses between 145 GeV and 170 GeV, and mediator masses between 175 GeV and 725 GeV, with the least restrictive region near the ``funnel" at $m_{\rm MED} \sim 2 m_t$; in the $t$-channel models, our DM masses could range between 150 GeV and 170 GeV, with the mediator masses between 175 GeV and 1000 GeV.
Applying experimental limits, we find that (i) the $s$-channel model is mostly unconstrained by monojet searches at the LHC (see Fig.~\ref{fig:sch}),
(ii) the $t$-channel models are constrained to different extents depending on the spin of dark matter.
The model with fermionic DM ({\tt tchFDM}) is poorly bounded by DM direct detection experiments, but excluded in the mediator mass range [380, 690] GeV by ATLAS stop searches.
The model with scalar DM ({\tt tchSDM}) is highly constrained by PandaX-II and the ATLAS search -- these probe the parameter space in complementary regions and collectively exclude most of it (see Fig.~\ref{fig:tch}).
Our bounds, however, may be relaxed if we also allowed for more annihilations to SM states lighter than $\chi$, which would diminish $\lambda$ and reflect in weaker limits from collider and direct detection searches.
There are several ways to arrange this.
For instance, one may allow for a more general coupling structure
that will allow $\chi$ to annihilate to lighter quarks of both up and down types, and to leptons.
In the $t$-channel models, one may still wish to keep annihilations restricted to the third generation
in order to safeguard against direct detection limits, but allow tree-level $b$-$\bar{b}$ final states
by charging the mediator $\widetilde{T}$ under $SU(2)_W$.
Forbidden final states involving weak boson and Higgs final states may be accommodated if $\chi$ also partially annihilated to light fermions.
All these alterations would expose these models to experimental probes not discussed here; we leave these avenues of study for future work.
The limits given here may also be weakened if one took into account the effects of co-annihilation.
Were $m_{\rm MED}$ close to $m_{\rm DM}$ in the $t$-channel models, the process $\widetilde{T} \chi \rightarrow {\rm gluon}+ {\rm top}$ would dominate the effective thermal cross-section.
New states that assist co-annihilation may also be present.
We may see this illustratively in a supersymmetric context, our analogue of which is the model {\tt tchFDM}.
Consider a bino LSP will all other electroweak-inos decoupled.
If one had a right-handed top squark $\widetilde{t}_R$ not much above in mass, and decoupled all other superpartners, the bino would annihilate predominantly to top quark pairs.
If its mass were just below $m_t$, only forbidden annihilation is possible.
Given the bino couples with hypercharge, this is an extremely inefficient process and would result in an unacceptably large relic abundance.
But if we imagine in the spectrum the presence of a gluino that is less than $\sim 10\%$ above the bino in mass, co-annihilation may play a crucial role.
The efficient self-annihilation of the gluino would now depopulate the bino abundance, and may set the effective thermal cross-section\footnote{As another example, if we take an LSP that has wino and/or higgsino content, the co-annihilation of the neutral components of the $SU(2)_W$ multiplets with their corresponding charged components may be balanced against the forbidden annihilation of the LSP to heavier SM states.}.
In this way, forbidden annihilation may be incorporated into supersymmetry models that provide DM candidates.
Here one might wonder if a gluino with a mass near $m_t$ is already ruled out by hadron collider experiments.
However, the ``compressed" nature of the gluino-bino spectrum, begetting soft jets and subdued MET, is unfriendly to the existing jets+$\slashed{E}_T$-based searches.
See e.g., Ref.~\cite{Alwall:2008ve}, where various cuts
were optimized to recast Tevatron searches for a simplified gluino-bino spectrum -- one finds the region of bino mass $\lesssim m_t$ uncovered.
To date, no dedicated search has been performed for the spectrum of interest.
An intriguing possibility (in both supersymmetric and non-supersymmetric contexts) is the forbidden annihilation of DM through an $s$-channel mediator tempered by co-annihilation with nearby states.
If the mediator mass were close to twice the DM mass, all three ``exceptions" outlined in \cite{Griest:1990kh} may be in action.
That Nature may have turned exceptions into a collective rule is an amusing scenario warranting further study.
It is worth emphasizing a crucial phenomenological difference between the forbidden scenarios considered here (annihilation to top quarks) and that of Ref.~\cite{D'Agnolo:2015koa} (annihilation to dark photons).
Forbidden WIMPs are only very weakly limited by DM indirect detection due to suppressed annihilation rates in the present day, whereas the requirement of sizeable couplings for forbidden annihilation makes them amenable to collider and direct detection probes.
On the other hand, indirect detection constraints play a decisive role in \cite{D'Agnolo:2015koa} -- when gauge kinetic mixing is introduced, annihilations to light SM states in the present day universe are possible.
At the same time, limits from direct detection and collider physics are weak or non-existent.
This is so because this model is an example of a secluded WIMP \cite{Pospelov:2007mp}.
The key concept is that direct DM-SM couplings are suppressed but the correct thermal cross-section is obtained by taking mediators light; therefore collider and direct detections limits are significantly weakened, whereas indirect detection signals become important.
In this way, forbidden WIMPs and secluded WIMPs are entirely contrasting ideas.
This is not to say that forbidden WIMPs may never yield signals from the sky.
Interesting prospects in indirect detection can be raised in our models if DM annihilates to new states that mix with the SM.
As \cite{Pospelov:2007mp} demonstrates, one may have annihilations to a $U(1)'$ boson (as in \cite{D'Agnolo:2015koa}), or to a spin-0 mediator mixing with the Higgs boson (possible in our $s$-channel model if $\widetilde{a}$ is CP-even) or even to a right-handed neutrino.
In such cases, one may also loosen the imposition of flavor structures discussed in this work, allowing more freedom in the way the mediators couple to multiple SM fields.
If the dark matter abundance is indeed set -- in full or part -- by forbidden annihilation to SM states, it might be argued that the DM mass appears to be tuned to fall just short of the SM mass.
Such spectral tuning is inevitable in forbidden WIMP models, and is somewhat akin to the mass degeneracy seen in scenarios involving co-annihilation, or to the careful arrangement of mass parameters that occasions resonant annihilation.
Our spectrum is perhaps the low energy manifestation of a flavor structure in the ultraviolet that is common to the dark and SM sectors.
In the $t$-channel models, the large sizes of both the effective couplings and the top Yukawas suggest the presence of some strong dynamics being responsible for the near-degeneracy of the $\chi$-$t$ system.
These hints encourage interesting model-building possibilities.
To conclude, forbidden annihilation is a freezeout scenario that we hope will garner more attention than priorly, from both theoretical and experimental communities wishing to constrain non-standard DM models.
We have shown that, even at the weak scale, such a mechanism may be principally responsible for making DM as abundant as observed.
This was arranged with minimal introduction of fields and parameters.
If this mechanism were in action at the weak scale, current collider and direct detection experiments are perfectly poised to probe it.
We hope to have pointed the direction to a significant stone to turn in the WIMP paradigm.
\section*{Acknowledgments}
\label{sec:ack}
Carlos Alvarado,
Joseph Bramante,
Spencer Chang,
Fatemeh Elahi
and
Roni Harnik
have variously improved this work with enlightening conversation.
This work was partially supported by the National
Science Foundation under Grants
No. PHY-1417118 and No. PHY-1520966.
|
\section{Introduction}
\subsection{Overview and main results}
Let $\field{T} := \field{R}^2/\field{Z}^2$ be the two-dimensional flat torus, and define $\Delta = \partial^2/\partial x_1^2+ \partial^2/\partial x_2^2$ to be the associated Laplace-Beltrami operator. Our aim in this paper is to characterize the high-energy behaviour of the zero set of complex-valued random eigenfunctions of $\Delta$, that is, of solutions $f$ of the {Helmholtz equation}
\begin{equation}\label{Hequation}
\Delta f + Ef = 0,
\end{equation}
for some adequate $E>0$. {{} In order to understand such a setting, recall} that the eigenvalues of $-\Delta$ are the positive reals of the form $E_n:=4\pi^2n$, where $n=a^2+b^2$ for some $a,b\in \field{Z}$ (that is, $n$ is an integer that can be represented as the sum of two squares). {Here, and throughout the paper}, we set
$$
S := \{ n \in \field{N} : a^2+b^2=n, \,\, \mbox{for some } \,\, a,b\in \field{Z}\},
$$
and for $n\in S$ we define
$$\Lambda_n:=\lbrace \lambda=(\lambda_1,\lambda_2)\in \field{Z}^2: \|\lambda\|^2:=\lambda_1^2+\lambda_2^2=n\rbrace$$
to be the set of {\bf energy levels} associated with $n$, while $\mathcal N_n:=|\Lambda_n|$ denotes its cardinality. An orthonormal basis (in $L^2(\mathbb T)$) for the eigenspace associated with $E_n$ is given by the set of complex exponentials $ \lbrace e_\lambda : \lambda\in \Lambda_n \rbrace$, defined as
$$
e_\lambda(x):={\rm e}^{i2\pi\langle \lambda,x\rangle}, \quad x\in \mathbb T,
$$
with $i = \sqrt{-1}$.
\smallskip
For every $n\in S$, the integer $\mathcal N_n =: r_2(n)$ counts the number of distinct ways of representing $n$ as the sum of two squares: it is a standard fact (proved e.g. by using Landau's theorem) that $\mathcal{N}_n$ grows on average as $\sqrt{\log n}$, and also that there exists an infinite sequence of prime numbers $p\in S$, $p\equiv 1\, {\rm mod}\, 4$, such that $\mathcal{N}_p=8$. A classical discussion of the properties of $S$ and $\mathcal N_n$ can be found e.g. in \cite[Section 16.9 and 16.10]{H-W}. In the present paper, we will systematically consider sequences $\{n_j\}\subset S$ such that $\mathcal{N}_{n_j}\to \infty$ (this is what we refer to as the {\bf high-energy limit}).
\medskip
The {complex waves} considered in this paper are natural generalizations of the real-valued arithmetic waves introduced by Rudnick and Wigman in \cite{RW}, and further studied in \cite{KKW, MPRW, ORW, RW2}; as such, they are close relatives of the complex fields considered in the physical literature --- see e.g. \cite{BD, Berry 2002, N-survey, NV}, {{} as well as the discussion provided below}. For every $n\in S$, we define the {\bf complex arithmetic random wave of order} $n$ to be the random field
\begin{equation}\label{e:defrf}
\Theta_n(x):=\frac{1}{\sqrt{\mathcal N_n}}\sum_{\lambda\in \Lambda_n} v_\lambda \, e_\lambda(x),\quad x\in \mathbb T,
\end{equation}
where the $v_\lambda,\, \lambda\in \Lambda_n$, are independent and identically distributed (i.i.d.) complex-valued Gaussian random variables such that, for every $\lambda \in \Lambda_n$, ${\rm Re}(v_\lambda)$ and ${\rm Im}(v_\lambda)$ are two independent centered Gaussian random variables with mean zero and variance one\footnote{{} Considering random variables $v_\lambda$ with variance 2 (instead of a more usual unit variance) will allow us to slightly simplify the discussion contained in Section \ref{ss:cr}.}.
The family $\{v_\lambda : \lambda\in \Lambda_n, \, n\in S\}$ is tacitly assumed to be defined on a common probability space $(\Omega, \mathscr{F}, {\mathbb{P}})$, with ${\mathbb{ E}}$ indicating expectation with respect to ${\mathbb{P}}$. It is immediately verified that $\Theta_n$ satisfies the equation \paref{Hequation}, {that is,} $\Delta \Theta_n + E_n \Theta = 0$, and also that $\Theta_n$ is {\bf stationary}, in the sense that, for every $y\in \field{T}$, the translated process $x\mapsto \Theta_n(y+x)$ has the same distribution as $\Theta_n$ (this follows from the fact that the distribution of $\{v_\lambda : \lambda\in \Lambda_n\}$ is invariant with respect to unitary transformations; see Section \ref{ss:cr} for further details on this {straightforward} but fundamental point).
\medskip
The principal focus of our investigation are the high-energy fluctuations of the {{} following} {\bf zero sets}:
\begin{eqnarray}
\mathscr{I}_n &:=& \{ x\in \field{T} : \Theta_n(x) = 0 \} \label{e:zeroset} \\ &=& \{ x\in \field{T} : {\rm Re} (\Theta_n(x)) = 0 \}\cap \{ x\in \field{T} : {\rm Im} (\Theta_n(x)) = 0 \}, \quad n\in S.\notag
\end{eqnarray}
We will show below (Part 1 of Theorem \ref{t:main}) that, with probability one, $\mathscr{I}_n$ is a finite {{} collection of isolated points} for every $n\in S$; throughout the paper, we will write
\begin{equation}\label{ienne}
I_n := |\mathscr{I}_n | = {\rm Card} (\mathscr{I}_n),\qquad n\in S.
\end{equation}
In accordance with the title of this work, the points of $\mathscr{I}_n$ are {called} {\bf phase singularities} for the field $\Theta_n$, in the sense that, for every $x\in \mathscr{I}_n$, the phase of $\Theta_n(x)$ (as a complex-valued random quantity) is not defined.
\medskip
As for nodal lines of real arithmetic waves \cite{KKW, MPRW}, our main results crucially involve the following collection of probability measures on the unit circle $S^1\subset \field{R}^2$:
\begin{equation}\label{e:spectral}
\mu_n (dz) := \frac{1}{\mathcal{N}_n} \sum_{\lambda\in \Lambda_n} \delta_{\lambda/\sqrt{n}} (dz), \quad n\in S,
\end{equation}
as well as the associated Fourier coefficients
\begin{equation}\label{e:fourier}
\widehat{\mu}_n (k ) := \int_{S^1} z^{-k} \mu_n (dz), \quad k\in \field{Z}.
\end{equation}
In view of the definition of $\Lambda_n$, the probability measure $\mu_n$ defined in \eqref{e:spectral} is trivially invariant with respect to the transformations $z\mapsto \overline{z}$ and $z\mapsto i \cdot z$. The somewhat erratic behaviours of such objects in the high-energy limit are studied in detail in \cite{KKW, KW}. Here, we only record the following statement, implying in particular that the sequences $\{\mu_n : n\in S\}$ and $\{\widehat{\mu}_n (4 ) : n\in S\}$ {\it do not} admit limits as $\mathcal{N}_n$ diverges to infinity within the set $S$.
Recall from \cite{KKW, KW} that a measure $\mu$ on $(S^1, \mathscr{B})$ (where $\mathscr{B}$ is the Borel $\sigma$-field) is said to be {\bf attainable} if there exists a sequence $\{n_j\}\subset S$ such that $\mathcal{N}_{n_j}\to \infty$ and $\mu_{n_j}$ converges to $\mu$ in the sense of the weak-$\star$ topology.
\begin{proposition}[\bf See \cite{KW, KKW}]\label{p:kw} The class of attainable measures is an infinite strict subset of the collection of all probability measures on $S^1$ that are invariant with respect to the transformations $z\mapsto \overline{z}$ and $z\mapsto i \cdot z$. Also, for every $\eta\in [0,1]$ there exists a sequence $\{n_j\}\subset S$ such that $\mathcal{N}_{n_j}\to \infty$ and $| \widehat{\mu}_{n_j}(4)| \to \eta$.
\end{proposition}
\medskip
Note that, if $\mu_{n_j}$ converges to $\mu_\infty$ in the weak-$\star$ topology, then $\widehat{\mu}_{n_j} (4) \to \widehat{\mu}_\infty(4)$. For instance, one knows from \cite{EH, KKW} that there exists a density one sequence $\{n_j\}\subset S$ such that $\mathcal{N}_{n_j}\to \infty$ and $\mu_{n_j}$ converges to the uniform measure on $S^1$, in which case $\widehat{\mu}_{n_j}(4)\to 0$.
\medskip
\noindent{\it Some conventions.} Given two sequences of positive numbers $\lbrace a_m\rbrace$ and $\lbrace b_m \rbrace$, we shall write $a_m\sim b_m$ if $a_m/b_m \to 1$,
and $a_m\ll b_m$ or (equivalently and depending on notational convenience) $a_m = O(b_m)$ if $a_m/b_m$ is asymptotically bounded. The notation $a_m=o(b_m)$ means as usual that $a_m/b_m\to 0$.
Convergence in distribution for random variables on $(\Omega, {\mathscr{F}}, {\mathbb{P}})$ will be denoted by $\overset{\rm law}{\Longrightarrow}$, whereas equality in distribution will be indicated by the symbol $\overset{\rm law}{=}$.
\medskip
The main result of the present work is the following exact characterization of the first and second order behaviours of $I_n$, {as defined by} \paref{ienne}, in the high-energy limit. As discussed below, it is a highly non-trivial extension of the results proved in \cite{KKW, MPRW}, as well as the first rigorous description of the {\bf Berry's cancellation phenomenon} \cite{Berry 2002} in the context of phase singularities of complex random waves.
\medskip
\begin{theorem} \label{t:main}
\begin{enumerate}
\item[\rm 1.] {\bf (Finiteness and mean)} With probability one, for every $n\in S$ the set $\mathscr{I}_n$ is composed of a finite collection of {isolated} points, and
\begin{equation}\label{e:exp}
{\mathbb{ E}}[I_n] = \frac{E_n}{4\pi} = \pi n.
\end{equation}
\item[\rm 2.] {\bf (Non-universal variance asymptotics)}
As $\mathcal{N}_n \to \infty$,
\begin{equation}\label{e:variance}
{\rm Var}(I_n) = d_n \times \frac{E_n^2}{\mathcal N_n^2} \, (1+o(1)),
\end{equation}
where
\begin{equation}\label{e:dn}
d_n := \frac{3\widehat \mu_n(4)^2 + 5}{128\pi^2}.
\end{equation}
\item[\rm 3.] {\bf (Universal law of large numbers)} Let $\lbrace n_j\rbrace\subset S$ be a subsequence such that $\mathcal N_{n_j}\to +\infty$. Then, for every sequence $\{\epsilon_{n_j}\}$ such that $\epsilon_{n_j}\mathcal{N}_{n_j}\to \infty$, one has that
\begin{equation}\label{e:wlln}
\mathbb{P}\left[ \left| \frac{I_{n_j}}{\pi n_j} - 1 \right| >\epsilon_{n_j} \right] \to 0.
\end{equation}
\item[\rm 4.] {\bf (Non-universal and non-central second order fluctuations)} Let $\lbrace n_j\rbrace\subset S$ be such that $\mathcal N_{n_j}\to +\infty$ and $|\widehat \mu_{n_j}(4)|\to \eta \in [0,1]$. Then,
\begin{eqnarray}\notag
\widetilde I_{n_j} &:=& \frac{I_{n_j} -{\mathbb{ E}}[I_{n_j}] }{\sqrt{{\rm Var}(I_{n_j}) }}\\ &\overset{\rm law}{\Longrightarrow}&
\frac{1}{2\sqrt{10+6\eta^2}}\left (\frac{1+\eta}{2}A
+\frac{1-\eta}{2}B
-2(C-2)\right) =: \mathcal{J}_\eta,\label{e:j}
\end{eqnarray}
with $A,B,C$ independent random variables such that $A\overset{\rm law}{=}B\overset{\rm law}{=}2X_1^2+2X_2^2-4X_3^2$
and $C \overset{\rm law}{=} X_1^2+X_2^2$, where
$(X_1,X_2,X_3)$ is a standard Gaussian vector of $\field{R}^3$.
\end{enumerate}
\end{theorem}
\begin{remark}\label{r:postmain}{\rm
\begin{enumerate}
\item The arguments leading to the proof of \eqref{e:exp} show also that, for every measurable $A\subset {\mathbb{T}}$,
\begin{equation}\label{e:locexp}
{\mathbb{ E}}[| \mathscr{I}_n\cap A |] = {\rm Leb}(A)\times \pi n,
\end{equation}
where `${\rm Leb}$' indicates the Lebesgue measure on the torus. The details are left to the reader.
\item For every $n\in S$, write $w(n) := {\rm Var}(I_n) \left( {E_n^2}/{\mathcal N_n^2}\right)^{-1}$. Standard arguments, based on compactness and on the fact that $\mu_n(4)\in [-1,1]$, yield that \eqref{e:variance} is equivalent to the following statement: for every $\{n_j\}\subset S$ such that $\mathcal{N}_{n_j}\to \infty$ and $|\mu_{n_j}(4)|\to \eta \in [0,1]$, one has that $w(n_j)\to d(\eta) := (3\eta^2+5)/128\pi^2$.
\item Relations \eqref{e:variance}--\eqref{e:dn} are {completely} new {and are {{} among} the main findings of the present paper; in particular, they} show that the asymptotic behaviour of the variance of $I_n$ is {\bf non-universal}. Indeed, when $\mathcal{N}_{n_j}\to \infty$, the fluctuations of the sequence $d_{n_j}$ depend on the chosen subsequence $\{n_j\}\subset S$, via the squared Fourier coefficients $\mu_{n_j}(4)^2$: in particular, the possible limit values of the sequence $\{d_{n_j}\}$ correspond to the whole interval $\left[\frac{5}{128\pi^2}, \frac{1}{16\pi^2}\right]$. As discussed in the sections to follow, such a non-universal behaviour {echoes} the findings of \cite{KKW}, in the framework of the length of nodal lines associated with arithmetic random waves. We will see that our derivation of the exact asymptotic relation \eqref{e:variance} follows a route that is different from the one exploited in \cite{KKW} --- as our techniques are based on chaos expansions, combinatorial cumulant formulae, as well as on a novel local Taylor expansion of the second order Kac-Rice kernel around the origin.
\item The support of the distribution of each variable $\mathcal{J}_\eta$ is the whole real line, but the distribution of $\mathcal{J}_\eta$ is {\it not} Gaussian (this follows e.g. by observing that law of $\mathcal{J}_\eta$ has exponential tails). As discussed in the forthcoming Section \ref{ss:cr}, similar non-central and non-universal second order fluctuations have been proved in \cite{MPRW} for the total nodal length of real arithmetic random waves. We will show below that this striking common feature originates from the same {\bf chaotic cancellation phenomenon} exploited in \cite{MPRW}, that is: {\it in the Wiener chaos expansion of the quantity $I_n$, the projection on the second chaos vanishes, and the limiting fluctuations of such a random variable are completely determined by its projection on the fourth Wiener chaos.} It will be clear from our analysis that, should the second chaotic projection of $I_n$ not disappear in the limit, then the order of ${\rm Var}(I_n)$ would be proportional to $E_n^2/\mathcal{N}_n$, as $\mathcal{N}_n\to \infty$.
\item Choosing $\epsilon_{n_j} \equiv \epsilon$ (constant sequence), one deduces from Point 3 of Theorem \ref{t:main} that the ratio $I_{n}/n$ converges in probability to $\pi$, whenever $\mathcal{N}_{n}\to \infty$. See e.g. \cite[p. 261]{D} for definitions.
\item It is easily checked (for instance, by computing the third moment) that the law of $\mathcal{J}_{\eta_0}$ differs from that of $\mathcal{J}_{\eta_1}$ for every $0\leq \eta_0< \eta_1\leq 1$. This fact implies that the sequence $\{| \mu_{n}(4)|\}$ dictates not only the asymptotic behaviour of the variance of $I_n$, but also the precise nature of its second order fluctuations.
\item Reasoning as in \cite[Theorem 1.2]{MPRW}, it is possible to suitably apply {\it Skorohod representation theorem} (see e.g. \cite[Chapter 11]{D}), in order to express relation \eqref{e:j} in a way that does not involve the choice of a subsequence $\{n_j\}$. We leave this routine exercise to the interested reader.
\end{enumerate}
}
\end{remark}
In the next section, we will discuss several explicit connections with the model of real arithmetic random waves studied in \cite{KKW, ORW, RW}.
\subsection{Complex zeros as nodal intersections}\label{ss:cr}
For simplicity, from now on we will write
\begin{equation}\label{e:tn}
T_n(x) := {\rm Re}(\Theta_n(x)), \quad \widehat{T}_n(x) := {\rm Im}(\Theta_n(x)),
\end{equation}
for every $x\in \field{T}$ and $n\in S$; in this way, one has that
\begin{eqnarray*}
\mathscr{I}_n &=&�T^{-1}_n(0)\cap \widehat{T}_n^{-1}(0) \quad \mbox{and} \quad I_n = |T^{-1}_n(0)\cap \widehat{T}_n^{-1}(0)|.
\end{eqnarray*}
We will also adopt the shorthand notation $${\bf T}_n := \{ {\bf T}_n(x) =(T_n(x), \widehat{T}_n(x)) : x\in \field{T}\}, \quad n\in S.$$
Our next statement yields a complete characterization of the distribution of the vector-valued process ${\bf T}_n$, as a two-dimensional field whose components are independent and identically distributed real arithmetic random waves, in the sense of \cite{KKW, MPRW, RW}.
\begin{proposition}\label{p:realim} Fix $n \in S$. Then, $T_n$ and $\widehat{T}_n$ are two real-valued independent centered Gaussian fields such that
\begin{equation}\label{e:cov}
{\mathbb{ E}}\Big[T_n(x)T_n(y)\Big] = {\mathbb{ E}}\Big[\widehat{T}_n(x)\widehat{T}_n(y)\Big] = \frac{1}{\mathcal{N}_n} \sum_{\lambda\in \Lambda_n} \cos(2\pi \langle \lambda, x-y\rangle )=:r_n(x-y).
\end{equation}
Also, there exist two collections of complex random variables
\begin{equation}\label{e:a}
{\bf A}(n) = \{ a_\lambda : \lambda \in \Lambda_n\} \quad \mbox{and} \quad \widehat{\bf A}(n)=\{ \widehat{a}_\lambda : \lambda \in \Lambda_n\},
\end{equation}
with the following properties:
\begin{itemize}
\item[\rm (i)] $ {\bf A}(n)$ and $\widehat{\bf A}(n)$ are stochastically independent and identically distributed as random vectors indexed by $\Lambda_n$;
\item[\rm (ii)] For every $\lambda\in \Lambda_n$, $a_\lambda$ is a complex-valued Gaussian random variable whose real and imaginary parts are independent Gaussian random variables with mean zero and variance $1/2$;
\item[\rm (iii)] If $\lambda \notin \{\sigma, -\sigma\}$, then $a_\lambda$ and $a_\sigma$ are stochastically independent;
\item[\rm (iv)] $a_\lambda = \overline{a_{-\lambda}}$;
\item[\rm (v)] For every $x\in \field{T}$,
\begin{equation}\label{e:realim}
T_n(x) = \frac{1}{\sqrt{\mathcal{N}_n}}\sum_{\lambda\in \Lambda_n} a_\lambda e_\lambda(x), \,\,\, \mbox{and} \,\,\,\,\, \widehat{T}_n(x) = \frac{1}{\sqrt{\mathcal{N}_n}}\sum_{\lambda\in \Lambda_n} \widehat{a}_\lambda e_\lambda(x).
\end{equation}
\end{itemize}
\end{proposition}
\noindent\begin{proof} One need only show that $T_n$ and $\widehat{T}_n$ are two centered independent Gaussian fields such that \eqref{e:cov} holds; the existence of the two families ${\bf A}(n)$ and $\widehat{\bf A}(n)$ can consequently be derived by first expanding $T_n$ and $\widehat{T}_n$ in the basis $\{e_\lambda : \lambda\in \Lambda_n\}$, and then by explicitly computing the covariance matrix of the resulting Fourier coefficients. Relation \eqref{e:cov} follows from a direct computation based on the symmetric structure of the set $\Lambda_n$, once it is observed that $T_n$ and $\widehat{T}_n$ can be written in terms of the complex Gaussian random variables $\{v_\lambda\}$ appearing in \eqref{e:defrf}, as follows:
\begin{eqnarray*}
T_n(x) &=& \frac{1}{\sqrt{\mathcal{N}_n}} \sum_{\lambda\in \Lambda_n} \Big\{{\rm Re}(v_\lambda)\cos(2\pi \langle \lambda, x\rangle)-{\rm Im}(v_\lambda)\sin(2\pi \langle \lambda, x\rangle)\Big\},\\
\widehat{T}_n(x) &=& \frac{1}{\sqrt{\mathcal{N}_n}} \sum_{\lambda\in \Lambda_n} \Big\{{\rm Im}(v_\lambda)\cos(2\pi \langle \lambda, x\rangle)+{\rm Re}(v_\lambda)\sin(2\pi \langle \lambda, x\rangle)\Big\}.
\end{eqnarray*}
\end{proof}
\medskip
The fact that $r_n$ only depends on the difference $x-y$ confirms in particular that ${\bf T}_n$ is a two-dimensional Gaussian stationary process.
\begin{assumption}\label{a:ass}{\rm Without loss of generality, for the rest of the paper we will assume that, for $n\neq m$, the two Gaussian families
$$
{\bf A}(n)\cup \widehat{\bf A}(n)\quad \mbox{and} \quad {\bf A}(m) \cup \widehat{\bf A}(m)
$$
are stochastically independent; this is the same as assuming that the two vector-valued fields ${\bf T}_n$ and ${\bf T}_m$ are stochastically independent.
}
\end{assumption}
As anticipated, relation \eqref{e:realim} implies that $T_n$ and $\widehat{T}_n$ are two independent and indentically distributed {\it real arithmetic random waves} of order $n$, such as the ones introduced in \cite{RW}, and then further studied in \cite{KKW, MPRW, ORW, RW2}. We recall that, according to \cite{C}, with probability one both $T_n^{-1}(0)$ and $\widehat{T}_n^{-1}(0)$ are unions of rectifiable curves, called {\bf nodal lines}, containing a finite set of isolated singular points. The following statement yields a further geometric characterisation of $\mathscr{I}_n$ and $I_n$: its proof is a direct by-product of the arguments involved in the proof of Part 1 of Theorem \ref{t:main}.
\begin{proposition}\label{p:inter} Fix $n\in S$. Then, with probability one the nodal lines of $T_n$ and $\widehat{T}_n$ have a finite number of {isolated intersection points}, whose collection coincides with the set $\mathscr{I}_n$; moreover, $\mathscr{I}_n$ does not contain any singular point for ${\bf T}_n$.
\end{proposition}
In view of Proposition \ref{p:inter}, it is eventually instructive to focus on the random {\bf nodal lengths}
$$
L_n := {\rm length}\, (T_n^{-1}(0)), \quad n\in S,
$$
for which we will present a statement collecting some of the most relevant findings from \cite{RW} (Point 1), \cite{KKW} (Point 2) and \cite{MPRW} (Point 3).
\begin{theorem}[\bf See \cite{RW, KKW, MPRW}]\label{t:lmain} \begin{enumerate}
\item[\rm 1.] For every $n\in S$
\begin{equation}\label{e:lexp}
{\mathbb{ E}}[L_n] = \frac{E_n}{2\sqrt{2}}.
\end{equation}
\item[\rm 2.]
As $\mathcal{N}_n \to \infty$,
\begin{equation}\label{e:lvariance}
{\rm Var}(L_n) = c_n \times \frac{E_n}{\mathcal N_n^2} \, (1+o(1)),
\end{equation}
where
\begin{equation}\label{e:cn}
c_n := \frac{1 + \widehat \mu_n(4)^2}{512}.
\end{equation}
\item[\rm 3.] Let $\lbrace n_j\rbrace\subset S$ be such that $\mathcal N_{n_j}\to +\infty$ and $|\widehat \mu_{n_j}(4)|\to \eta \in [0,1]$. Then,
\begin{eqnarray}\notag
\widetilde L_{n_j} &:=& \frac{L_{n_j} -{\mathbb{ E}}[L_{n_j}] }{\sqrt{{\rm Var}(L_{n_j}) }}\\ &\overset{\rm law}{\Longrightarrow}&
\mathcal{M}_\eta := \frac{1}{2\sqrt{1+\eta^2}} (2 - (1+\eta) X_1^2-(1-\eta) X_2^2),\label{e:m}
\end{eqnarray}
where $(X_1,X_2)$ is a standard Gaussian vector of $\field{R}^2$.
\end{enumerate}
\end{theorem}
As discussed e.g. in \cite{KKW,RW}, relations \eqref{e:lexp}--\eqref{e:lvariance} yield immediately a law of large numbers analogous to \eqref{e:wlln}. We stress that Theorem \ref{t:main} and Theorem \ref{t:lmain} share three common striking features (explained below in terms of a common chaotic cancellation phenomenon), namely: (a) an {inverse quadratic dependence} on $\mathcal{N}_n$, as displayed in formulae \eqref{e:variance} and \eqref{e:lvariance}, (b) non-universal variance fluctuations, determined by the quantities $d_n$ and $c_n$ defined in \eqref{e:dn} and \eqref{e:cn}, respectively, and (c) non-universal and non-central second order fluctuations (see \eqref{e:j} and \eqref{e:m}).
The estimate \eqref{e:lvariance} largely improves upon a conjecture formulated by Rudnick and Wigman in \cite{RW}, according to which one should have ${\rm Var}(L_n) = O(E_n/\mathcal{N}_n)$. The fact that the natural leading term $E_n/\mathcal{N}_n$ actually disappears in the high-energy limit, thus yielding \eqref{e:lvariance}, is connected to some striking discoveries by Berry \cite{Berry 2002}, discussed in the forthcoming section.
\subsection{More about relevant previous work}
\underline{\it Random waves and cancellation phenomena.} To the best of our knowledge, the first systematic analysis of phase singularities in wave physics appears in the seminal contribution by Nye and Berry \cite{NB}. Since then, zeros of complex waves have been the object of an intense study in a variety of branches of modern physics, often under different names, such as {\it nodal points}, {\it wavefront dislocations}, {\it screw dislocations}, {\it optical vortices} and {\it topological charges}. The reader is referred e.g. to \cite{D-survey, N-survey, UR-survey}, and the references therein, for detailed surveys on the topic, focussing in particular on optical physics, quantum chaos and quantum statistical physics.
One crucial reference for our analysis is Berry \cite{Berry 1978}, where the author studies several statistical quantities involving singularities of random waves on the plane. Such an object, usually called the (complex) {\bf Berry's random wave model} (RWM), is defined as a complex centered Gaussian field, whose real and imaginary parts are independent Gaussian functions on the plane, with covariance
\begin{equation}\label{e:rwm}
r_{\rm RWM}(x,y) := J_0\left(\sqrt{E}\, \| x-y\|\right), \quad x,y\in \field{R}^2,
\end{equation}
where $E>0$ is an energy parameter, and $J_0$ is the standard Bessel function (see also \cite{Berry 1977}). Formula \eqref{e:rwm} implies in particular that Berry's RWM is {\bf stationary and isotropic}, that is: its distribution is invariant both with respect to translations and rotations. As discussed e.g. in \cite[Section 1.6.1]{KKW}, if $\{n_j\} \subset S$ is a sequence such that $\mathcal{N}_{n_j}\to \infty$ and $\mu_{n_j}$ converges weakly to the uniform measure on the circle, then, for every $x\in {\mathbb{T}}$ and using the notation \eqref{e:cov},
\begin{equation}\label{e:scaling}
r_{n_j}\left(\sqrt{\frac{E}{n_j}}\cdot \frac{x}{2\pi}\right) \longrightarrow r_{\rm RWM}(x),
\end{equation}
showing that Berry's RWM is indeed the local scaling limit of the arithmetic random waves considered in the present paper.
Reference \cite{Berry 2002}, building upon previous findings of Berry and Dennis \cite{BD}, contains the following remarkable results: {\bf (a)} the expected nodal length per unit area of the real RWM equals $\sqrt{E} /(2\sqrt{2})$ \cite[Section 3.1]{Berry 2002}, {\bf (b)} as $E\to \infty$ the variance of the nodal length at Point {\bf (a)} is proportional to $\log E$ \cite[Section 3.2]{Berry 2002}, {\bf (c)} the expected number of phase singularities for unit area of the complex RWM is $E /(4\pi)$ \cite[Section 4.1]{Berry 2002}, and {\bf (d)} as $E\to \infty$ the variance of the number of singularities at Point {\bf (c)} is proportional to $E \log E$ \cite[Section 4.2]{Berry 2002}. Point {\bf (a)} and {\bf (c)} are perfectly consistent with \eqref{e:lexp} and \eqref{e:exp}, respectively. Following \cite{Berry 2002}, the estimates at Points {\bf (b)} and {\bf (d)} are due to an `obscure' cancellation phenomenon, according to which the natural leading term in variance (that should be of the order of $\sqrt{E}$ and $E^{3/2}$, respectively) cancels out in the high-energy limit. The content of Point {\bf (b)} has been rigorously confirmed by Wigman \cite{wig} in the related model of real {\it random spherical harmonics}, whose scaling limit is again the real RWM. See also \cite{ALW}.
As explained in \cite{KKW}, albeit improving conjectures from \cite{RW}, the order of the variance established in \eqref{e:lvariance} differs from that predicted in {\bf (b)}: this discrepancy is likely due to the fact that, differently from random spherical harmonics, the convergence in \eqref{e:scaling} does not take place uniformly over suitable regions. As already discussed, in \cite{MPRW} it was shown that the asymptotic relation \eqref{e:lvariance} is generated by a remarkable chaotic cancellation phenomenon, which also explains the non-central limit theorem stated in \eqref{e:m}.
The main result of the present paper (see Theorem \ref{t:main}) confirms that such a chaotic cancellation continues to hold for phase singularities of complex arithmetic waves, and that it generates non-universal and non-central second order fluctuations for such a random quantity. This fact lends further evidence to the natural conjecture that cancellation phenomena analogous to those described in {\cite{Berry 2002, wig, KKW, MPRW, Ro}} should hold for global quantities associated with the zero set of Laplace eigenfunctions on general manifolds, as long as such quantities can be expressed in terms of some area/co-area integral formula.
We stress that the fact that the order of the variance stated in \eqref{e:variance} differs from the one predicted at Point {\bf (d)} above, can once again be explained by the non-uniform nature of the scaling relation \eqref{e:scaling}.
\medskip
\noindent\underline{\it Leray measures and occupation densities}. While the present paper can be seen as a natural continuation of the analysis developed in \cite{KKW, MPRW}, the methods implemented below will substantially diverge from those in the existing literature. One fundamental difference stems from the following point: in order to deal with strong correlations between vectors of the type $(T_n(x), \partial/\partial_{1} T_n(x), \partial/\partial_{2} T_n(x))$ and $(T_n(y), \partial/\partial_{1} T_n(y), \partial/\partial_{2} T_n(y))$, $x\neq y$, the authors of \cite{KKW} extensively use results from \cite{ORW} (see in particular \cite[Section 4.1]{KKW}) about the fluctuations of the {\bf Leray measure}
$$
A_n := \int_{{\mathbb{T}}} \delta_0(T_n(x)) \, dx,
$$
which is defined as the limit in $L^2({\mathbb{P}})$ of the sequence $k\mapsto \int_{{\mathbb{T}}} \varphi_k(T_n(x)) \, dx$, with $\{\varphi_k\}$ a suitable approximation of the identity; on the other hand, following such a route in the framework of random phase singularities is impossible, since the formal quantity
$$
B_n:= \int_{{\mathbb{T}}} \delta_{(0,0)}(T_n(x), \widehat{T}_n(x)) \, dx
$$
{\it cannot} be defined as an element of $L^2({\mathbb{P}})$. The technical analysis of singular points developed in Section \ref{proof varianza sec} is indeed how we manage to circumvent this difficulty. We observe that, in the parlance of stochastic calculus, the quantity $A_n$ (resp. $B_n$) is the {\bf occupation density at zero} of the random field $T_n$ (resp. ${\bf T}_n$) --- in particular, the fact that $A_n$ is well-defined in $L^2({\mathbb{P}})$ and $B_n$ {is not} --- follows from the classical criterion stated in \cite[Theorem 22.1]{GH}, as well as from the relations
\begin{equation}\label{e:explosion}
\int_{\mathbb{T}} \frac{dx}{\sqrt{1-r^2_n(x)}} <\infty \quad \mbox{and} \quad \int_\field{T} \frac{dx}{1-r^2_n(x)} =\infty,
\end{equation}
where we have used the fact that, according e.g. to \cite[Lemma 5.3]{ORW}, the mapping $x\mapsto (1-r^2_n(x))^{-1}$ behaves like a multiple of $1/\|x-x_0\|^2$ around any point $x_0$ such that $r_n(x_0)=\pm1$.
\medskip
\noindent{\underline{\it Nodal intersections of arithmetic random waves with a fixed curve}.
{{} A natural problem related to the subject of our paper is that of studying the number of} nodal intersections with a fixed {{} deterministic} curve $\mathcal C\subset \mathbb T$ whose length equals $L$, i.e. number of zeroes of $T_n$ that lie on $\mathcal C$:
$$
\mathcal Z_n := T_n^{-1}(0)\cap \mathcal C.
$$
In \cite{RW2}, the case where $\mathcal C$ is a smooth curve with nowhere zero-curvature {has been investigated}. The expected number of nodal intersections is ${\mathbb{ E}}[|\mathcal Z_n|]=(\pi \sqrt 2)^{-1}\times E_n \times L$, hence proportional to the length $L$ of the curve times the wave number, independent of the geometry. The asymptotic behaviour of the nodal intersections variance in the high energy limit is a subtler matter: it {depends} on both the angular distribution of lattice points lying on the circle with radius corresponding to the given wavenumber, in particular on the sequence of measures $\lbrace \mu_n\rbrace$, and {on} the geometry of $\mathcal C$. The asymptotic distribution of $|\mathcal Z_n|$ is}{analyzed in \cite{RoW}}. {See \cite{Ma} for the case where $\mathcal C$ is a segment.}
\medskip
\noindent\underline{\it Zeros of random analytic functions/Systems of polynomials}. To the best of our expertise, our limit result \eqref{e:j} is the first non-central limit theorem for the number of zeros of random complex analytic functions defined on some manifold $\mathcal{M}$. As such, our findings should be contrasted with the works by Sodin and Tsirelson \cite{ST1, ST2}, where one can find central limit results for local statistics of zeros of analytic functions corresponding to three different models (elliptic, flat and hyperbolic). As argued in \cite[Section 1.6.4]{wig}, these results are roughly comparable to those one would obtain by studying zeros of complex random spherical harmonics, for which a central high-energy behaviour is therefore likely to be expected. References \cite{SZ1, SZ2}, by Shiffman and Zelditch, contain central limit result for the volume of the intersection of the zero sets of independent Gaussian sections of high powers of holomorphic line bundles on a K\"ahler manifold of a fixed dimension.
In view of Proposition \ref{p:inter}, our results have to be compared with works dealing with the number of roots of random system of polynomials.
{The first important result about the number of roots of random systems is due to Shub and Smale \cite{ss},
where the authors compute the expectation of the number of roots of a square system with independent
centered Gaussian coefficients with a particular choice of the variances that makes the distribution of the polynomials invariant under the action of the orthogonal group of the parameter space.
This model is called the {\bf Shub-Smale model}.
Later, Edelman and Kostlan, see \cite{k} and references therein, and Aza\"is and Wschebor \cite{aw-pol} extend these results to more general Gaussian distributions.
Wschebor \cite{w} studies the asymptotic for the variance of the number of roots of a Shub-Smale system in the case where the number of equations and variables tends to infinity.
Armentano and Wschebor \cite{ar-w} consider the expectation of non-centered (perturbed) systems.
McLennan \cite{mcl} studies the expected number of roots of multihomogeneous systems. Rojas \cite{ro} consider the expected number of roots of certain sparse systems. }
\subsection{Short plan of the paper}
{Section \ref{s:prelim} gathers some preliminary results that will be needed in the sequel. In Section \ref{s:approach} we explain the main ideas and steps of the proof of our main result (Theorem \ref{t:main}). Finally, the remaining sections are devoted to the detailed proofs. In particular, we collect in Section \ref{appendix} some technical computations and proofs of intermediate results.
}
\subsection{Acknowledgments}
This research was supported {by the grant F1R-MTH-PUL-15CONF (CONFLUENT) at the University of Luxembourg (Federico Dalmao and Ivan Nourdin)} and by
the grant F1R-MTH-PUL-15STAR (STARS)
at the University of Luxembourg (Giovanni Peccati and Maurizia Rossi).
The authors heartily thank Domenico Marinucci for suggesting the problem studied in this paper, and for several inspiring discussions.
\section{Some preliminary result}\label{s:prelim}
We will now present some useful notions and results connected to Wiener chaos, gradients, combinatorial moment formulae and arithmetic estimates.
\subsection{Wiener chaos }\label{ss:wiener}
Let $\{H_k : k=0,1,...\}$ be the sequence of {\bf Hermite polynomials} on $\mathbb{R}$, recursively defined as follows: $H_0 \equiv 1$, and, for $k\geq 1$,
$$H_{k}(t) = tH_{k-1}(t) - H'_{k-1}(t),\qquad t\in \field{R}.$$
It is a standard fact that the collection $\mathbb{H} := \{H_k/\sqrt{k!} : k\ge 0\}$ is a complete orthonormal system for $$L^2(\mathbb{R}, \mathscr{B}(\mathbb{R}), \gamma) :=L^2(\gamma),$$ where $\gamma(dt):= \phi(t)dt =\frac{ e^{-t^2/2}}{\sqrt{2\pi}} dt$ is the standard Gaussian measure on the real line. By construction, for every $k\geq 0$, one has that
\begin{equation}\label{e:sym}
H_{2k}(-t) = H_{2k}(t), \quad \mbox{and} \quad H_{2k+1}(-t) = -H_{2k+1}(t), \quad t\in \field{R}.
\end{equation}
\medskip
In view of Proposition \ref{p:realim} (recall also Assumption \ref{a:ass}), every random object considered in the present paper is a measurable functional of the family of complex-valued Gaussian random variables $$ \bigcup_{n\in S} \Big( {\bf A}(n) \cup \widehat{{\bf A}}(n) \Big),$$ where ${\bf A}(n) $ and $\widehat{{\bf A}}(n)$ are defined in \eqref{e:a}. Now define the space ${\bf A}$ to be the closure in $L^2(\mathbb{P})$ of all real finite linear combinations of random variables $\xi$ of the form $$\xi = c_1 (z \, a_\lambda + \overline{z} \, a_{-\lambda}) + c_2(u \, \widehat{a}_\tau + \overline{u} \, \widehat{a}_{-\tau})$$ where $\lambda, \tau \in \mathbb{Z}^2$, $z,u\in \mathbb{C}$ and $c_1, c_2\in \field{R}$. The space ${\bf A}$ is a real centered Gaussian Hilbert subspace
of $L^2(\mathbb{P})$.
\begin{defn}\label{d:chaos}{\rm For a given integer $q\ge 0$, the $q$-th {\bf Wiener chaos} associated with ${\bf A}$, denoted by $C_q$, is the closure in $L^2(\mathbb{P})$ of all real finite linear combinations of random variables of the type
$$
\prod_{j=1}^k H_{p_j}(\xi_j),
$$
with $k\ge 1$, where the integers $p_1,...,p_k \geq 0$ verify $p_1+\cdots+p_k = q$, and $(\xi_1,...,\xi_k)$ is a centered standard real Gaussian vector contained in ${\bf A}$ (so that $C_0 = \mathbb{R}$).}
\end{defn}
In view of the orthonormality and completeness of $\mathbb{H}$ in $L^2(\gamma)$, it is not difficult to show that $C_q \,\bot\, C_{q'}$ (where the orthogonality holds in $L^2(\mathbb{P})$) for every $q\neq q'$, and moreover
\begin{equation*}
L^2(\Omega, \sigma({\bf A}), \mathbb{P}) = \bigoplus_{q=0}^\infty C_q;
\end{equation*}
the previous relation simply indicates that every real-valued functional $F$ of ${\bf A}$ can be uniquely represented in the form
\begin{equation}\label{e:chaos2}
F = \sum_{q=0}^\infty {\rm proj}(F \, | \, C_q)=\sum_{q=0}^\infty F[q],
\end{equation}
where $F[q]:={\rm proj}(F \, | \, C_q)$ stands for the the projection of $F$ onto $C_q$, and the series converges in $L^2(\mathbb{P})$. By definition, one has $F[0]={\rm proj}(F \, | \, C_0) = {\mathbb{ E}} [F]$. See e.g. \cite[Theorem 2.2.4]{NP} for further details.
\subsection{About gradients}
Differentiating both terms in \eqref{e:realim} yields that, for $j=1,2$,
\begin{equation}\label{e:partial}
\partial_j T_n(x) = \frac{2\pi i}{\sqrt{\mathcal{N}_n} }\sum_{(\lambda_1,\lambda_2)\in \Lambda_n} \lambda_j a_\lambda e_\lambda(x),\,\, \mbox{and} \,\,\, \partial_j \widehat{T}_n(x) = \frac{2\pi i}{\sqrt{\mathcal{N}_n} }\sum_{(\lambda_1,\lambda_2)\in \Lambda_n} \lambda_j \widehat{a}_\lambda e_\lambda(x)
\end{equation}
(where we used the shorthand notation $\partial_j = \frac{\partial}{\partial x_j}$).
It follows that, for every $n \in S$ and every $x\in{\mathbb{T}}$,
\begin{equation}\label{e:six}
T_{n}(x),\, \partial_{1}T_{n}(x), \, \partial_{2}T_{n}(x), \widehat{T}_n(x),\, \partial_{1}\widehat{T}_n(x), \, \partial_{2}\widehat{T}_n(x) \in \bf A.
\end{equation}
Another important fact (that one can check by a direct computation) is that, for fixed $x\in \mathbb{T}$, the six random variables appearing in \eqref{e:six} are stochastically independent.
\medskip
Routine computations (see also \cite[Lemma 2.3]{RW}) yield that $${\rm Var}(\partial_j T_n(x)) ={\rm Var}(\partial_j \widehat{T}_n(x)) =\frac{E_n}{2},$$ for any $j=1,2$, any $n$ and any $x\in\mathbb{T}$. Accordingly, we will denote by $\widetilde{\partial}_j$ the normalised derivative
\[
\widetilde{\partial}_j:=\sqrt{\frac{2}{E_n}}\,\frac{\partial}{\partial x_j},
\]
and adopt the following (standard) notation for the gradient and its normalised version:
$$
\nabla :=\left( \begin{matrix} \partial_1 \\ \partial_2 \end{matrix}\right),\qquad \widetilde \nabla :=\left( \begin{matrix} \widetilde \partial_1 \\ \widetilde \partial_2 \end{matrix}\right).
$$
\subsection{Leonov-Shiryaev formulae}\label{ss:ls}
In the proof of our variance estimates, we will crucially use the following special case of the so-called {\bf Leonov-Shiryaev combinatorial formulae} for computing joint moments. It follows immediately e.g. from \cite[Proposition 3.2.1]{PT}, by taking into account the independence of $T_n$ and $\widehat{T}_n$, the independence of the six random variables appearing in \eqref{e:six}, as well as the specific covariance structure of Hermite polynomials (see e.g. \cite[Proposition 2.2.1]{NP}).
{\color{black}
\begin{proposition}\label{p:ls} Fix $n\in S$ and write
\begin{eqnarray*}
&& (X_0(x), X_1(x), X_2(x), Y_0(x), Y_1(x), Y_2(x))\\
&&\quad\quad\quad := (T_{n}(x),\, \widetilde{\partial}_{1}T_{n}(x), \, \widetilde{\partial}_{2}T_{n}(x), \widehat{T}_n(x),\, \widetilde{\partial}_{1}\widehat{T}_n(x), \, \widetilde{\partial}_{2}\widehat{T}_n(x)), \quad x\in {\mathbb{T}}.
\end{eqnarray*}
Consider integers $p_0,p_1,p_2, q_0, q_1, q_2\geq 0$ and $a_0,a_1,a_2, b_0, b_1, b_2\geq 0$, and write
\begin{eqnarray*}
(X^\star_1(x),...,X^{\star}_{p_0+p_1+p_2}(x))&:=& (\underbrace{X_0(x),...,X_0(x)}_{p_0\mbox{ times}} ,\underbrace{X_1(x),...,X_1(x)}_{p_1\mbox{ times}},\underbrace{ X_2(x),...,X_2(x)}_{p_2 \mbox{ times}} )\\
(X^{\star\star}_1(y),...,X^{\star\star}_{a_0+a_1+a_2}(y))&:=& (\underbrace{X_0(y),...,X_0(y)}_{a_0 \mbox{ times}} ,\underbrace{X_1(y),...,X_1(y)}_{a_1\mbox{ times}},\underbrace{ X_2(y),...,X_2(y)}_{a_2 \mbox{ times}} ) \\
(Y^\star_1(x),...,Y^{\star}_{q_0+q_1+q_2}(x))&:=& (\underbrace{Y_0(x),...,Y_0(x)}_{q_0\mbox{ times}} ,\underbrace{Y_1(x),...,Y_1(x)}_{q_1\mbox{ times}},\underbrace{ Y_2(x),...,Y_2(x)}_{q_2 \mbox{ times}} )\\
(Y^{\star\star}_1(y),...,Y^{\star\star}_{b_0+b_1+b_2}(y))&:=& (\underbrace{Y_0(y),...,Y_0(y)}_{b_0 \mbox{ times}} ,\underbrace{Y_1(y),...,Y_1(y)}_{b_1\mbox{ times}},\underbrace{ Y_2(y),...,Y_2(y)}_{b_2 \mbox{ times}} ).
\end{eqnarray*}
Then, for every $x,y \in {\mathbb{T}}$,
\begin{eqnarray}\label{e:ls}
&&{\mathbb{ E}}\left[\prod_{j=0}^2 H_{p_j}(X_j(x))H_{a_j}(X_j(y))\prod_{k=0}^2 H_{q_k}(Y_k(x))H_{b_k}(Y_k(y))\right] \\
&& \notag = {\bf 1}_{\{ p_0+p_1+p_2 = a_0+a_1+a_2\}} {\bf 1}_{\{ q_0+q_1+q_2 = b_0+b_1+b_2\}} \times\\
&&\quad\quad\times \sum_{\sigma, \pi} \left( \prod_{j=1}^{p_0+p_1+p_2} \, {\mathbb{ E}}[X^{\star}_j(x) X^{\star\star}_{\sigma(j)}(y)]\right)\left( \prod_{k=1}^{q_0+q_1+q_2} \, {\mathbb{ E}}[Y^{\star}_k (x) Y^{\star\star}_{\pi(k)}(y)]\right),\notag
\end{eqnarray}
where the sum runs over all permutations $\sigma, \pi$ of $\{1,..., p_0+p_1+p_2\}$ and of $\{1,...,q_0+q_1+q_2\}$, respectively.
\end{proposition}
}
\subsection{Arithmetic facts}\label{ss:arit}
We will now present three results having an arithmetic {flavour}, that will be extensively used in the proofs of our main findings. To this end, for every $n\in S$ we introduce the $4$- and $6$-{\bf correlation set of frequencies}
\begin{eqnarray*}
S_4(n) &:=& \{{\boldsymbol{\lambda}}=(\lambda,\lambda',\lambda'',\lambda''')\in\Lambda_n^4 : \lambda- \lambda'+\lambda'' -\lambda''' =0\},\\
S_6(n) &:=& \{{\boldsymbol{\lambda}}=(\lambda,\lambda',\lambda'',\lambda''', \lambda'''', \lambda^{ v})\in\Lambda_n^6 : \lambda- \lambda'+\lambda'' -\lambda''' +\lambda^{{} iv} - \lambda^{ v}=0\}.
\end{eqnarray*}
The first statement exploited in our proofs yields an exact value for $|S_4(n)|$; the proof is based on an elegant geometric argument due to Zygmund \cite{Zy}, and is included for the sake of completeness.
\begin{lemma}\label{l:zygmund}
For every $n\in S$, every element of $S_4(n)$ has necessarily one of the following four (exclusive) structures:
\begin{itemize}
\item[\rm a)] $\lambda=\lambda'$, and $\lambda''=\lambda'''$;
\item[\rm b)] $\lambda = -\lambda'=-\lambda'' = \lambda'''$;
\item[\rm c)] $\lambda\notin \{\lambda',-\lambda'\}$, $\lambda=-\lambda''$, and $\lambda'=-\lambda'''$;
\item[\rm d)] $\lambda\notin \{\lambda',-\lambda'\}$, $\lambda=\lambda'''$, and $\lambda'=\lambda''$.
\end{itemize}
In particular, $|S_4(n)| = 3\mathcal{N}_n(\mathcal{N}_n-1)$.
\end{lemma}
\noindent\begin{proof}
We {{} partition} the elements of $S_4(n)$ into three disjoint subset: $S_1 = \{\boldsymbol{\lambda}:\lambda=\lambda'\}\, {\cap \, S_4(n)}$, $S_2 = \{\boldsymbol{\lambda}:\lambda=-\lambda'\} \, {\cap \, S_4(n)}$, and $S_3 = \{\boldsymbol{\lambda}:\lambda\notin \{\lambda',-\lambda'\} \} \, {\cap \, S_4(n)}$. If ${\boldsymbol{\lambda}} \in S_1$, then necessarily $\lambda''=\lambda'''$, and consequently $|S_1| = \mathcal{N}_n^2$. If ${\boldsymbol{\lambda}}\in S_2$, then the relations $2\lambda+\lambda''=\lambda'''$ and $2\lambda-\lambda'''=-\lambda''$ show that both $-\lambda''$ and $\lambda'''$ must belong to the intersection of the circle $C_0$ of radius $\sqrt{n}$ centered at the origin (since they are both in $\Lambda_n$) with the circle $C'$ of radius $\sqrt{n}$ centered in $2\lambda$; since $C_0\cap C' = \{\lambda\}$, we conclude that necessarily $\lambda=-\lambda'=-\lambda''=\lambda'''$. Plainly, this argument also yields $|S_2| = \mathcal{N}_n$. If ${\boldsymbol{\lambda}} \in S_3$, then the relations $\lambda-\lambda'+\lambda'' = \lambda'''$ and $\lambda-\lambda'-\lambda''' = -\lambda''$ show that both $-\lambda''$ and $\lambda'''$ must belong to the intersection of the circle $C_0$ of radius $\sqrt{n}$ centered at the origin (again, since they are both in $\Lambda_n$) with the circle $C''$ of radius $\sqrt{n}$ centered in $\lambda-\lambda'$; since $C_0\cap C'' = \{\lambda, -\lambda'\}$, we conclude that necessarily either $-\lambda''=\lambda$ and $-\lambda'''=\lambda'$ or $\lambda'''=\lambda$ and $\lambda''=\lambda'$ (the configurations such that $-\lambda''=\lambda'''$ are excluded by the requirement that $\lambda\neq -\lambda'$). This yields $|S_3| = 2\mathcal{N}_n(\mathcal{N}_n-2)$. Summing up,
$$
|S_4(n)| = |S_1|+|S_2|+|S_3| = \mathcal{N}_n^2+ \mathcal{N}_n+2\mathcal{N}_n (\mathcal{N}_n-2) = 3\mathcal{N}_n(\mathcal{N}_n-1).
$$
\end{proof}
The second estimate involves $6$-correlations, and follows from a deep result by Bombieri and Bourgain \cite[Theorem 1]{B-B} --- see also \cite[Theorem 2.2]{KKW} for a slightly weaker statement.
\begin{lemma}[\bf See \cite{B-B}]\label{lemma BB}
As $\mathcal{N}_n\to \infty$,
$$
|S_6(n)| = O\left(\mathcal N_n^{7/2} \right).
$$
\end{lemma}
We will also need the following elementary fact about the behaviour of the correlation function $r_n$, as defined in \eqref{e:cov}, in a small square containing the origin.
\begin{proposition}\label{gio} Let $n\in S$, with $n\geq 1$, let $c_n= ( 1000 \sqrt {n} )^{-1} $, and $Q_n := \{(x,y)\in \field{R}^2 : |x|, |y| \leq c_n\}$. Assume that $z = (x,y) \in Q_n$ is such that $r_n(z) = \pm 1$; then, $z=0$.
\end{proposition}
\noindent\begin{proof} Assume first that $r_n(z) = 1$. Then, for every $(\lambda_1,\lambda_2) \in \Lambda_n$, one has necessarily that there exist $j,k,l\in \mathbb{Z}$ such that (i) $\lambda_1 x+\lambda_2 y = j$, (ii) $-\lambda_1 x+\lambda_2 y = k$, and (iii) $\lambda_1 y+\lambda_2 x = l$. Assume first that $\lambda_1 = 0$ (and therefore $|\lambda_2| = \sqrt{n} $): equation (i) implies that $|y| = |j|/\sqrt{n}$, and such an expression is $>c_n$ unless $j=y=0$; similarly, equation (iii) implies that $|x|>c_n$, unless $x=l=0$. The case where $\lambda_2 = 0$ is dealt with analogously. Now assume that $\lambda_1, \lambda_2\neq 0$: equations (i) and (ii) imply therefore that $y= (j+k)/2\lambda_2$ and $x = (j-k)/2\lambda_1$; since $|\lambda_i| \leq \sqrt{n}$, for $i=1,2$, we infer that $|x|, |y|\leq c_n$ if and only if $j+k = 0= j-k$, and therefore $x=y=j=k=0$.
If $r_n(z) = -1$, then, for every $(\lambda_1,\lambda_2) \in \Lambda_n$, one has necessarily that there exist $j,k,l\in \mathbb{Z}+\frac12$ such that equations (i), (ii), (iii) above are verified: reasoning exactly as in the first part of the proof, we conclude that $\max\{|x|, |y|\}>c_n$, and consequently $z$ cannot be an element of $Q_n$.
\end{proof}
{{}
Finally, we will make use of the following result, corresponding to a special case of \cite[Corollary 9, p. 80]{Kov}: it yields a local ersatz of B\'ezout theorem for systems of equations involving trigonometric polynomials. We recall that, given a smooth mapping $U : \field{R}^2 \to \field{R}^2$ and a point $x\in \field{R}^2$ such that $U(x) = (0,0)$, one says that $x$ is {\it nondegenerate} if the Jacobian matrix of $U$ at $x$ is invertible.
\begin{lemma}[See \cite{Kov}]\label{l:few} Fix $n\in S$, and consider two trigonometric polynomials on $\field{R}^2$:
$$
P(x) = c+ \sum_{\lambda\in \Lambda_n} a_\lambda e_\lambda(x), \quad \mbox{and} \quad Q(x) = c'+ \sum_{\lambda\in \Lambda_n} b_\lambda e_\lambda(x),
$$
where $c, c'\in \field{R}$ and the complex numbers $\{a_\lambda, b_\lambda\}$ verify the following:
\begin{itemize}
\item[--] for every $\lambda \in \Lambda_n$, one has that $a_\lambda = \overline{a_{-\lambda}}$ and $b_\lambda = \overline{b_{-\lambda}}$;
\item[--] every solution of the system $(P(x), Q(x)) = (0,0)$ such that $\|x\| < \pi/\sqrt{n}$ is nondegenerate.
\end{itemize}
Then, the number of solutions of the system $(P(x), Q(x)) = (0,0)$ contained in the open window $W := \{x\in \field{R}^2 : \|x\|< \pi/\sqrt{n}\}$ is bounded by a universal constant $\alpha(n) \in (0,\infty)$ depending uniquely on $\mathcal{N}_n = | \Lambda_n|$.
\end{lemma}
}
\medskip
The next section contains a precise description of the strategy we will adopt in order to attack the proof of Theorem \ref{t:main}
\section{Structure of the proof of Theorem \ref{t:main}}\label{s:approach}
\subsection{{Chaotic projections and cancellation phenomena}}
We will start by showing in Lemma \ref{lemma approx} that $I_n$ can be formally obtained in $L^2({\mathbb{P}})$ as
\begin{equation}\label{formula integrale1}
I_n = \int_{\mathbb T} \delta_{\bf 0}({\bf T}_n(x)) \left | J_{{\bf T}_n}(x)\right |\,dx,
\end{equation}
where $\delta_{\bf 0}$ denotes the Dirac mass in ${\bf 0}=(0,0)$, $J_{{\bf T}_n}$ is the Jacobian matrix
$$
J_{{\bf T}_n}=\left( \begin{matrix}
&\partial_1 T_n &\partial_2 T_n\\
&\partial_1 \widehat T_n &\partial_2 \widehat T_n
\end{matrix}\ \right )
$$
and $| J_{{\bf T}_n}|$ is shorthand for the absolute value of its determinant. Since $I_n$ is a square-integrable functional of a Gaussian field, according to the general decomposition \eqref{e:chaos2} one has that
\begin{equation}\label{series}
I_n = \sum_{q\ge 0} I_n[q],
\end{equation}
where $I_n[q]=\text{proj}(I_n|C_q)$ denotes the orthogonal projection of $I_n$ onto the $q$-th Wiener chaos $C_q$. Since $I_n[0]={\mathbb{ E}}[I_n]$, the computation of the $0$-order chaos projection will allow us to conclude the proof of Part 1 of Theorem \ref{t:main} in Section \ref{proof media sec}.
One crucial point in our analysis is that, as proved in Lemma \ref{berry's cancellation}, the projections of $I_n$ onto odd-order Wiener chaoses vanish and, more subtly, also the second chaotic component disappears. {Namely}, we will show that, for every $n\in S$, it holds
$$
I_n[q]=0 \qquad \text{for odd } q\ge 1
$$
and moreover
\begin{equation}\label{eq chaos 2}
I_n[2]=0.
\end{equation}
Our proof of \paref{eq chaos 2} is based on Green's identity and the properties of Laplacian eigenfunctions (see also \cite[Section 7.3 and p.134]{Ro}).
\subsection{Leading term: fourth chaotic projections}
The first non-trivial chaotic projection of $I_n$ to investigate is therefore $I_n[4]$. One of the main achievements of our paper is an explicit computation of its asymptotic variance, as well as a proof that it gives the dominant term in the asymptotic behaviour of the total variance ${\rm Var}(I_n)=\sum_{q\ge 2} {\rm Var}(I_n[2q])$. The forthcoming Propositions \ref{varianze chaos}, \ref{asymptotic} and \ref{limite 4}, that we will prove in Section \ref{sec chaos 4}, are the key steps in order to achieve our goals.
\begin{proposition}\label{varianze chaos}
Let $\lbrace n_j\rbrace_j\subset S$ be such that $\mathcal N_{n_j}\to +\infty$ and $|\widehat{\mu}_{n_j}(4)|\to \eta$. Then
\begin{equation*}
{\rm Var}(I_{n_j}[4]) = d(\eta)\, \frac{E_{n_j}^2}{\mathcal N_{n_j}^2}(1+o(1)),
\end{equation*}
where
\begin{equation*}
d(\eta) = \frac{3\eta^2 + 5}{128\pi^2}.
\end{equation*}
\end{proposition}
In view of Remark \ref{r:postmain}(2), Proposition \ref{varianze chaos} coincides with Part 2 of Theorem \ref{t:main}, once we replace $I_{n_j}[4]$ with $I_{n_j}$.
Let us now set, for $n\in S$,
\begin{eqnarray}\label{momento 4}
R_n(4) &:=& \int_{\mathbb T} r_n(x)^4\,dx= \frac{|S_n(4)|}{\mathcal {N}_n^4} =\frac{3\mathcal{N}_n(\mathcal{N}_n-1)}{\mathcal{N}_n^4} , \\
R_n(6)& :=& \int_{\mathbb T} r_n(x)^6\,dx= \frac{|S_n(6)|}{\mathcal N_n^6},\label{momento 6}
\end{eqnarray}
where $S_n(4), S_n(6)$ are the sets of $4$- and $6$-correlation coefficients defined in Section \ref{ss:arit}, and we have used Lemma \ref{l:zygmund} in \paref{momento 4}. The following result (Proposition \ref{asymptotic}), combined with Proposition \ref{varianze chaos} and Lemma \ref{lemma BB} allows us to conclude that, as $\mathcal{N}_n\to \infty$,
\begin{equation}\label{ciao}
{\rm Var}(I_{n}) \sim {\rm Var}(I_{n}[4]),
\end{equation}
thus achieving the proof of Part 2 of Theorem \ref{t:main}. Note that, by virtue of Lemma \ref{lemma BB} and \eqref{momento 6}, as $\mathcal{N}_n\to \infty$ one has that
$$
R_n(6) = o\left(\frac{1}{\mathcal{N}_n^2}\right) \quad \mbox{or, equivalently,} \quad R_n(6) = o\left(R_n(4)\right).
$$
\begin{proposition}\label{asymptotic}
As $\mathcal N_n\to +\infty$, we have
$$
\sum_{q\ge 3} {\rm Var}\left( I_n[2q] \right) = O\left(E_n^2 R_n(6) \right).
$$
\end{proposition}
Part 3 of Theorem \ref{t:main} follows immediately from the relation
$$
\mathbb{P}\left[ \left| \frac{I_{n}}{\pi n} - 1 \right| >\epsilon \right]\leq \frac{ {\rm Var}({I_{n}}/(\pi n) )^{1/2} }{\epsilon}
$$
(which is a consequence of the Markov inequality), as well as from Part 1 and Part 2 of the same Theorem. Finally, the proof of Part 4 of Theorem \ref{t:main} relies on a careful and technical investigation of $I_n[4]$, leading us to the following result, which indeed coincides with \eqref{e:j}, once replacing $\frac{I_{n_j}[4]}{\sqrt{{\rm Var}(I_{n_j}[4])}}$ with $\widetilde I_{n_j}$.
\begin{proposition}\label{limite 4}
Let $\lbrace n_j\rbrace_j\subset \lbrace n\rbrace$ be a subsequence such that $\mathcal N_{n_j}\to +\infty$ and $|\widehat \mu_{n_j}(4)|\to \eta$, then
$$
\frac{I_{n_j}[4]}{\sqrt{{\rm Var}(I_{n_j}[4])}}\, \stackrel{\rm law}{\Longrightarrow} \,
\mathcal{J}_\eta,
$$
where $\mathcal{J}_\eta$ is defined in \eqref{e:j}.
\end{proposition}
\subsubsection{Controlling the variance of higher-order chaoses}
In order to prove Proposition \ref{asymptotic}, we need to carefully control the remainder given by $\sum_{q\ge 3}{\rm Var}(I_n[2q])$; our argument (consisting in a substantial extension of techniques introduced in \cite[\S 6.1]{ORW} and \cite[\S 4.3]{RW2}) is the following.
We {{} partition} the torus into {a} union of disjoint squares $Q$ of side length $1/M$, where $M$ is proportional to $\sqrt{E_n}$. Of course
\begin{equation}\label{eq sum}
I_n=\sum_Q I_{n_{|_Q}},
\end{equation}
where $I_{n_{|_Q}}$ is the number of zeroes contained in $Q$. It holds that, for every $q\ge 0$, $
I_n[q]=\sum_Q I_{n_{|_Q}}[q]$ and hence
\begin{equation}\label{sum cov}
{\rm Var}\left( \sum_{q\ge 3} I_n[2q] \right) = \sum_{Q,Q'} {\rm Cov\,}\left(\text{proj} \left(I_{n_{|_Q}}|C_{\ge 6}\right) , \text{proj} \left(I_{n_{|_{Q'}}} |C_{\ge 6}\right)\right),
\end{equation}
where $\text{proj}\left(\cdot |C_{\ge 6} \right)$ denotes the orthogonal projection onto $\bigoplus_{q\ge 6} C_q$, that is, the orthogonal sum of Wiener chaoses of order larger or equal than six.
We now split the double sum on the RHS of \paref{sum cov} into two parts: namely, one over {\it singular} pairs of cubes and the other one over {\it non-singular} pairs of cubes. Loosely speaking, for a pair of non-singular cubes $(Q,Q')$, we have that for every $(z,w)\in Q\times Q'$, the covariance function $r_n$ of the field $T_n$ and all its normalized derivatives up to the order two $\widetilde \partial_i r_n, \widetilde \partial_{ij} r_n:=(E_n/2)^{-1}\partial/\partial_{x_i x_j} r_n$ for $i,j=1,2$ are bounded away from 1 and $-1$, once evaluated in $z-w$ (see Definition \ref{sing point} and Lemma \ref{lemma sing}).
\begin{lemma}[Contribution of the singular part]\label{varianza sing}
As $\mathcal N_n\to +\infty$,
\begin{equation}\label{eq sing}
\left | \sum_{(Q,Q')\text{ sing.}} {\rm Cov\,}\left (\text{proj}\left (I_{n_{ |_Q}}|C_{\ge 6}\right ), \text{proj}\left (I_{n_{ |_{Q'}}}|C_{\ge 6}\right )\right ) \right | \ll E_n^2 R_n(6).
\end{equation}
\end{lemma}
In order to show Lemma \ref{varianza sing} (see Section \ref{proof varianza sec}), we use the Cauchy-Schwarz inequality and the stationarity of ${\bf T}_n$, in order to reduce the problem to the investigation of nodal intersections in a small square $Q_0$ around the origin: for the LHS of \paref{eq sing} we have
\begin{equation*}
\begin{split}
&\left | \sum_{(Q,Q')\text{ sing.}} {\rm Cov\,}\left (\text{proj}\left (I_{n_{ |_Q}}|C_{\ge 6}\right ), \text{proj}\left (I_{n_{ |_{Q'}}}|C_{\ge 6}\right )\right ) \right |\cr
&\ll \sum_{(Q,Q')\text{ sing.}} {\mathbb{ E}}\left[I_{n_{ |_{Q_0}}}\left(I_{n_{ |_{Q_0}}}-1 \right) \right] + {\mathbb{ E}}\left[I_{n_{ |_{Q_0}}} \right].
\end{split}
\end{equation*}
Thus, we need to (i) count the number of singular pairs of cubes, (ii) compute the expected number of nodal intersections in $Q_0$ and finally (iii) calculate the second factorial moment of $I_{n_{ |_{Q_0}}}$. Issue (i) will be dealt with by exploiting the definition of singular pairs of cubes and the behavior of the moments of the derivatives of $r_n$ on the torus (see Lemma \ref{bello}), thus obtaining that
$$
| \lbrace (Q,Q')\text{ sing.} \rbrace | \ll E_n^2 R_n(6).
$$
Relations \eqref{e:locexp} and \paref{eq sum} yield immediately that ${\mathbb{ E}}\left[I_{n_{ |_{Q_0}}} \right]$ is bounded by a constant independent of $n$.
To deal with (iii) is much subtler matter. Indeed, we need first to check the assumptions for Kac-Rice formula (see \cite[Theorem 6.3]{AW} ) to hold in Proposition \ref{gio}. The latter allows us to write the second factorial moment $ {\mathbb{ E}}\left[I_{n_{ |_{Q_0}}}\left(I_{n_{ |_{Q_0}}}-1 \right) \right]$ as an integral on $Q_0\times Q_0$ of the so-called {\bf two-point correlation} function $K_2$, given by
$$
K_2(x,y) := p_{({\bf T}_n(x),{\bf T}_n(y))}({\bf 0},{\bf 0}){\mathbb{ E}}\left[ \left | J_{{\bf T}_n}(x)\right | \left | J_{{\bf T}_n}(y)\right |\Big |{\bf T}_n(x)={\bf T}_n( y)={\bf 0} \right],
$$
where {$x,y\in \mathbb T$ and} $p_{({\bf T}_n(x),{\bf T}_n(y))}$ is the density of $({\bf T}_n(x),{\bf T}_n(y))$.
The stationarity of ${\bf T}_n$ then reduces the problem to investigating $K_2(x):=K_2(x,0)$ around the origin. Hypercontractivity properties of Wiener chaoses and formulas for the volume of ellipsoids (see \cite{KZ}) yield the following estimation
\begin{equation}\label{psi n}
K_2(x)\ll \frac{|\Omega_n(x)|}{1-r_n(x)^2}=: \Psi_n(x),
\end{equation}
where $| \Omega_n(x)|$ stands for the absolute value of the determinant of the matrix $\Omega_n(x)$, defined as the covariance matrix of the vector $\nabla T_n(0)$, conditionally on $T_n(x)=T_n(0)=0$. An explicit Taylor expansion at $0$ for $\Psi_n$ (made particularly arduous by the diverging integral in \eqref{e:explosion} --- see Lemma \ref{lemma taylor}) will allow us to prove that $ {\mathbb{ E}}\left[I_{n_{ |_{Q_0}}}\left(I_{n_{ |_{Q_0}}}-1 \right) \right]$ is also bounded by a constant independent of $n$. This concludes the proof of Lemma \ref{varianza sing}.
To achieve the proof of Theorem \ref{t:main}, we will eventually show the following result, whose proof relies on Proposition \ref{p:ls}, on the definition of non-singular cubes, as well as on the behavior of even moments of derivatives of the covariance function $r_n$.
\begin{lemma}[Contribution of the non-singular part]\label{varianza nonsing}
As $\mathcal N_n\to +\infty$, we have
\begin{equation*}
\left| \sum_{(Q,Q')\text{ non sing.}} {\rm Cov\,}\left (\text{proj}\left(I_{n_{|_{Q}}}|C_{\ge 6}\right),\text{proj}\left(I_{n_{|_{Q'}}}|C_{\ge 6}\right) \right) \right|=O\left( E_n^2 R_n(6) \right).
\end{equation*}
\end{lemma}
The rest of the paper contains the formal proofs of all the statements discussed in the present section.
\section{Phase singularities and Wiener chaos}\label{wiener sec}
\subsection{Chaotic expansions for phase singularities}\label{chaos sec}
In this part we find the chaotic expansion \paref{series} for $I_n$. The first achievement in this direction is the following approximation result.
\subsubsection{An integral expression for the number of zeros}\label{sec approx}
For $\varepsilon>0$ and $n\in S$, we consider the $\varepsilon$-approximating random variable
\begin{equation}\label{formula integrale1eps}
I_n(\varepsilon):=\frac{1}{4\varepsilon^2}
\int_{\mathbb T}{\mathbf 1}_{[-\varepsilon,\varepsilon]^2}({\mathbf T}_n(x))
|J_{{\mathbf T}_n}(x)|\,dx,
\end{equation}
where ${\bf 1}_{[-\varepsilon, \varepsilon]^2}$ denotes the indicator function of the square $[-\varepsilon, \varepsilon]^2$. The following result makes the formal equality in \paref{formula integrale1} rigorous.
\begin{lemma}\label{lemma approx}
For $n\in S$, with probability one, $I_n$ {{} is composed of a finite number of isolated points} and, as $\varepsilon\to 0$,
\begin{equation}\label{e:cscs}
I_n(\varepsilon)\to I_n,
\end{equation}
both a.s. and in the {{} $L^p({\mathbb{P}})$-sense, for every $p\geq 1$. }
\end{lemma}
\noindent\begin{proof} {{} Fix $n\in S$. In order to directly apply some statements taken from \cite{AW}, we will canonically identify the random field $(x_1,x_2)\mapsto {\bf T}_n(x_1,x_2)$ with a random mapping from $\field{R}^2$ to $\field{R}^2$ that is 1-periodic in each of the coordinates $x_1, x_2$. In what follows, for $x\in \field{R}^2$ we will write ${\bf T}_n(x,\omega)$ to emphasize the dependence of ${\bf T}_n(x)$ on $\omega\in \Omega$. We subdivide the proof into several steps, numbered from (i) to (vi).
\begin{itemize}
\item[(i)] First of all, since ${\bf T}_n$ is an infinitely differentiable stationary Gaussian field such that, for every $x\in \field{R}^2$, the vector ${\bf T}_n(x)$ has a standard Gaussian distribution, one can directly apply \cite[Proposition 6.5]{AW} to infer that there exists a measurable set $\Omega_0\subset \Omega$ with the following properties: ${\mathbb{P}}(\Omega_0)=1$ and, for every $\omega \in \Omega_0$ and every $x\in \field{R}^2$ such that ${\bf T}_n(x, \omega) =0$, one has necessarily that the Jacobian matrix $J_{{\bf T}_n}(x, \omega)$ is invertible.
\item[(ii)] A standard application of the inverse function theorem (see e.g. \cite[p. 136]{AT}) implies that, for every $\omega \in \Omega_0$, any bounded set $B\subset \field{R}^2$ only contains a finite number of points $x$ such that ${\bf T}_n(x,\omega)=0$. This implies in particular that, with probability one, $\mathscr{I}_n$ (as defined in \eqref{e:zeroset}) is composed of a finite number of isolated points and $I_n <+\infty$.
\item[(iii)] Sard's Lemma yields that, for every $\omega \in \Omega_0$, there exists a set $ U_\omega \subset \field{R}^2$ such that $U_\omega^c$ has Lebesgue measure $0$ and, for every $u\in U_\omega$ there is no $x\in \field{R}^2$ such that ${\bf T}_n(x,\omega) = u$ and $J_{{\bf T}_n}(x,\omega)$ is not invertible. Note that, by definition, one has that $0\in U_\omega$ for every $\omega \in \Omega_0$.
{
\item[(iv)] Define $B := \{x = (x_1,x_2)\in \field{R}^2 : 0\leq x_i< 1/L\}, \,\, i=1,2\}$, where $L$ is any positive integer such that $L>\sqrt{n}$. For every $u\in \field{R}^2$, we set $I_{n,u}(B)$ to be the cardinality of the set composed of those $x\in B$ such that ${\bf T}_n(x) = u$; the quantity $I_{n,u}(\field{T})$ is similarly defined, in such a way that $I_{n,0}(\field{T})=I_n$. Two facts will be crucial in order to conclude the proof: (a) for every $\omega \in \Omega_0$ and every $u=(u_1,u_2) \in U_\omega$, by virtue of Lemma \ref{l:few} as applied to the pair $(P,Q)$ given by
$$
P(x) = T_n(x,\omega)-u_1 \quad \mbox{and} \quad Q(x) = \widehat{T}_n(x,\omega)-u_2,
$$
as well as of the fact that $B\subset W$, one has that $I_{n,u}(B)(\omega)\leq \alpha(n)$, and (b) as a consequence of the inverse function theorem, for every $\omega \in \Omega_0$ there exists $\eta_\omega \in (0, \infty)$ such that the equality $I_n(\omega) = I_{n,u}(\field{T})(\omega)$ holds for every $u$ such that $\| u \| \leq \eta_\omega$. Indeed, reasoning as in \cite[Proof of Theorem 11.2.3]{AT} if this was not the case, then there would exist a sequence $u_k \to 0$, $u_k \neq 0$, and a point $x\in \field{T}$ such that: (1) ${\bf T}_n(x,\omega) = 0$, and (2) for every neighborhood $V$ of $x$ (in the topology of $\field{T}$) there exist $k\geq 1$
and $x_0, x_1\in V$ such that $x_0\neq x_1$ and ${\bf T}_n(x_0) = {\bf T}_n(x_1) = u_k$ --- which is in contradiction with the inverse function theorem.
\item[(v)] By the area formula (see e.g. \cite[Proposition 6.1 and formula (6.2)]{AW}), one has that, for every $\omega \in \Omega_0$,
\begin{eqnarray}\label{e:asc}
&& \frac{1}{4\varepsilon^2}\int_{\field{T}}{\mathbf 1}_{[-\varepsilon, \varepsilon]^2}({\mathbf T}_n(x, \omega))|J_{{\mathbf T}_n}(x, \omega )|dx\\
&&\quad\quad\quad\quad =\frac{1}{4\varepsilon^2}\int_{[-\varepsilon,\varepsilon]^2}I_{n,u}(\field{T})(\omega) \,du
= \frac{1}{4\varepsilon^2}\int_{[-\varepsilon,\varepsilon]^2\cap U_\omega}I_{n,u}(\field{T})(\omega) \,du,\notag
\end{eqnarray}
where we used the property that the complement of $U_\omega$ has Lebesgue measure 0. Since the integral on the right-hand side of \eqref{e:asc} equals $ I_{n}$ whenever $\varepsilon \leq \eta_\omega/\sqrt{2}$, we conclude that \eqref{e:cscs} holds ${\mathbb{P}}$-a.s. .
\item[(vi)] According to the discussion at Point (iv)-(a) above and using stationarity, one has that
$$
{\mathbb{P}} [I_n \leq L^2\alpha(n)] = {\mathbb{P}}\left[ \frac{1}{4\varepsilon^2}\int_{[-\varepsilon,\varepsilon]^2}I_{n,u}(\field{T}) \,du \leq L^2\alpha(n)\right]= 1.
$$
The fact that \eqref{e:cscs} holds also in $L^p({\mathbb{P}})$ now follows from Point (v) and dominated convergence.}
\end{itemize}
}
\end{proof}
\subsubsection{Chaotic expansions}
Let us consider the collections of coefficients $\lbrace \beta_l : l\geq 0\rbrace$ and $\lbrace \alpha_{a, b, c, d} : a,b,c,d\geq 0\}$ defined as follows. For $l\ge 0$
\begin{equation}\label{e:beta}
\beta_{2l+1}:=0, \qquad \beta_{2l} := \frac{1}{\sqrt{2\pi}}H_{2l}(0),
\end{equation}
where (as before) $H_{2l}$ is the $2l$-th Hermite polynomial. For instance,
\begin{equation}\label{beta piccoli}
\beta_0=\frac{1}{{\sqrt{2\pi}}},\qquad \beta_2=-\frac{1}{{\sqrt{2\pi}}},\qquad \beta_4 = \frac{3}{\sqrt{2\pi}}.
\end{equation}
Also, we set
\begin{equation}\label{e:alpha}
\alpha_{a,b,c,d} := {\mathbb{ E}}[|XW-YV|H_a(X)H_b(Y)H_c(V)H_d(W)],
\end{equation}
with $(X,Y,V, W)$ a standard real four-dimensional Gaussian vector. Note that
on the right-hand side of \paref{e:alpha}, $|XW-YV|$ is indeed the absolute value of the determinant of
the matrix
$$
\left ( \begin{matrix}
&X &Y\\
&V &W
\end{matrix} \ \right).
$$
\begin{lemma}\label{nullity}
If $a,b,c,d$ do not have the same parity, then $$\alpha_{a,b,c,d}=0.$$
\end{lemma}
\noindent\begin{proof}
Let us assume without loss of generality (by symmetry) that $a$ is odd and that at least one integer among $b$, $c$ and $d$ is even. We will exploit \eqref{e:sym}. If $b$ is even, then, using that
$(X,Y,V, W)\overset{\rm law}{=}(-X,-Y,V, W)$, one can write that
\begin{eqnarray*}
\alpha_{a,b,c,d}&=&{\mathbb{ E}}[|XW-YV|H_a(X)H_b(Y)H_c(V)H_d(W)]\\
&=&{\mathbb{ E}}[|YV-XW|H_a(-X)H_b(-Y)H_c(V)H_d(W)]\\
&=&-{\mathbb{ E}}[|XW-YV|H_a(X)H_b(Y)H_c(V)H_d(W)]=-\alpha_{a,b,c,d},
\end{eqnarray*}
leading to $\alpha_{a,b,c,d}=0$.
If $c$ (resp. $d$) is even, the same reasoning based on
$(X,Y,V, W)\overset{\rm law}{=}(-X,Y,-V, W)$
(resp. $(X,Y,V, W)\overset{\rm law}{=}(-X,Y,V,-W)$)
leads to the same conclusion.
\end{proof}
We will not need the explicit values of $\alpha_{a,b,c,d}$, unless $a+b+c+d\in \lbrace{0,2,4\rbrace }$. The following technical result will be proved in Section \ref{appendix}.
\begin{lemma}\label{alpha piccoli} It holds that
\begin{equation*}
\begin{split}
\alpha_{0,0,0,0}&=1,\\
\alpha_{2,0,0,0}&= \alpha_{0,2,0,0}=\alpha_{0,0,2,0}=\alpha_{0,0,0,2}=\frac12,\\
\alpha_{4,0,0,0} &=\alpha_{0,4,0,0}=\alpha_{0,0,4,0}=\alpha_{0,0,0,4}=-\frac{3}{8},\\
\alpha_{2,2,0,0}&=\alpha_{0,0,2,2}=\alpha_{2,2,0,0}=-\frac{1}{8},\\
\alpha_{2,0,2,0}&=\alpha_{0,2,0,2}-\frac{1}{8},\\
\alpha_{2,0,0,2}&=\alpha_{0,2,2,0}=\frac{5}{8},\\
\alpha_{1,1,1,1}&=-\frac{3}{8}.
\end{split}
\end{equation*}
\end{lemma}
\begin{lemma}[Chaotic expansion of $I_n$]\label{berry's cancellation}
For $n\in S$ and $q\ge 0$, we have
\begin{equation}\label{chaos dispari}
I_n[2q+1]=0,
\end{equation}
and
\begin{equation}\label{proiezione q}
\begin{split}
I_n[2q]=&\frac{E_n}{2}\sum_{i_1+i_2+i_3+j_1+j_2+j_3=2q} \frac{\beta_{i_1}\beta_{j_1}}{i_1! j_1!}\frac{\alpha_{i_2 i_3 j_2 j_3}}{i_2! i_3! j_2! j_3!}\times\cr
&\!\!\!\!\times \int_{\mathbb T} H_{i_1}(T_n(x))H_{j_1}(\widehat T_n(x))H_{i_2}(\widetilde \partial_1 T_n(x))H_{i_3}(\widetilde \partial_2 T_n(x))H_{j_2}(\widetilde \partial_1 \widehat T_n(x))H_{j_3}(\widetilde \partial_2 \widehat T_n(x))\,dx,
\end{split}
\end{equation}
where the sum can be restricted to the set of those indices $(i_1,j_1,i_2,i_3,j_2,j_3)$ such that $i_1, j_1$ are even and $i_2, i_3, j_2, j_3$ have the same parity.
In particular,
\begin{equation}\label{chaos 2}
I_n[2]=0.
\end{equation}
The chaotic expansion for $I_n$ is hence
\begin{equation}\label{chaos exp}
\begin{split}
I_n=&I_n[0]+\sum_{q\ge 2}\frac{E_n}{2}\sum_{i_1+i_2+i_3+j_1+j_2+j_3=2q} \frac{\beta_{i_1}\beta_{j_1}}{i_1! j_1!}\frac{\alpha_{i_2 i_3 j_2 j_3}}{i_2! i_3! j_2! j_3!}\times\cr
&\times \int_{\mathbb T} H_{i_1}(T_n(x))H_{j_1}(\widehat T_n(x))H_{i_2}(\widetilde \partial_1 T_n(x))H_{i_3}(\widetilde \partial_2 T_n(x))H_{j_2}(\widetilde \partial_1 \widehat T_n(x))H_{j_3}(\widetilde \partial_2 \widehat T_n(x))\,dx,
\end{split}
\end{equation}
where the sum runs over the set of those indices $(i_1,j_1,i_2,i_3,j_2,j_3)$ such that $i_1, j_1$ are even and $i_2, i_3, j_2, j_3$ have the same parity.
\end{lemma}
\noindent\begin{proof} The main idea is to deduce the chaotic expansion for $I_n$ from the chaotic expansion for \paref{formula integrale1eps} and Lemma \ref{lemma approx}.
Let us first rewrite \paref{formula integrale1eps} as
\begin{equation}\label{areaformula}
I_n(\varepsilon)={\frac{E_n}{8\varepsilon^2}}\int_{\field{T}} 1_{[-\varepsilon,\varepsilon]^2}({\bf T}_n(x))
\big|
\widetilde{\partial}_1T_n(x)\widetilde{\partial}_2\widehat{T}_n(x)
-
\widetilde{\partial}_1\widehat{T}_n(x)\widetilde{\partial}_2T_n(x)
\big|dx.
\end{equation}
We recall the chaos decomposition of the indicator function (see e.g. \cite[Lemma 3.4]{MPRW}):
\begin{equation*}
\frac{1}{2\varepsilon}{1}_{[-\varepsilon ,\varepsilon ]}(\cdot
)=\sum_{l=0}^{+\infty }\frac{1}{l!}\beta_l^{\varepsilon}\,H_{l}(\cdot ),
\end{equation*}
where, for $l \geq 0$
\begin{equation}\label{e:beta eps}
\beta_0^{\varepsilon} = \frac{1}{2\varepsilon} \int_{-\varepsilon}^{\varepsilon} \phi(t)\,dt,\qquad \beta_{2l+1}^\varepsilon = 0,\qquad \beta^\varepsilon_{2l+2}= -\frac{1}{\varepsilon}
\phi \left (\varepsilon \right) H_{2l+1} \left (\varepsilon \right),
\end{equation}
and $\phi$ is still denoting the standard Gaussian density.
For the indicator function of $[-\varepsilon,\varepsilon]^2$ appearing in (\ref{areaformula}), we thus
have
\begin{equation}\label{delta0}
\frac{1}{4\varepsilon^2}1_{[-\varepsilon,\varepsilon]^2}(x,y) = \sum_{l=0}^\infty \sum_{q=0}^l \frac{\beta^\varepsilon_{2q}\beta^\varepsilon_{2l-2q}}{(2q)!(2l-2q)!}H_{2q}(x)H_{2l-2q}(y).
\end{equation}
The chaotic expansion for the absolute value of the Jacobian determinant appearing in \paref{areaformula} is, thanks to Lemma \ref{nullity},
\begin{equation}\label{determinant}
\begin{split}
&\big|
\widetilde{\partial}_1T_n(x)\widetilde{\partial}_2\widehat{T}_n(x)
-
\widetilde{\partial}_1\widehat{T}_n(x)\widetilde{\partial}_2T_n(x)
\big|\\
&=
\sum_{q\ge 0}\sum_{\substack{a+b+c+d=2q\\ (a,b,c,d\text{ the same parity}) }}^\infty \frac{\alpha_{a,b,c,d}}{a!b!c!d!}
H_a(\widetilde{\partial}_1 T_n(x))H_b(\widetilde{\partial}_2 T_n(x))H_c(\widetilde{\partial}_1 \widehat T_n(x))H_d(\widetilde{\partial}_2 \widehat T_n(x)),
\end{split}
\end{equation}
where $\alpha_{a,b,c,d}$ are given in \paref{e:alpha}. In particular, observe that Lemma \ref{nullity} ensures that the odd chaoses vanish in the chaotic expansion for the Jacobian.
It hence follows from \paref{delta0} and \paref{determinant} that the chaotic expansion for $I_n(\varepsilon)$ in \paref{areaformula} is (taking sums over even $i_1, j_1$ and $i_2, i_3, j_2, j_3$ with the same parity)
\begin{equation}\label{proiezione q eps}
\begin{split}
I_n(\varepsilon)=&\frac{E_n}{2}\sum_{q\ge 0}\sum_{i_1+i_2+i_3+j_1+j_2+j_3=2q} \frac{\beta^\varepsilon_{i_1}\beta^\varepsilon_{j_1}}{i_1! j_1!}\frac{\alpha_{i_2 i_3 j_2 j_3}}{i_2! i_3! j_2! j_3!}\times\cr
&\times \int_{\mathbb T} H_{i_1}(T_n(x))H_{j_1}(\widehat T_n(x))H_{i_2}(\widetilde \partial_1 T_n(x))H_{i_3}(\widetilde \partial_2 T_n(x))H_{j_2}(\widetilde \partial_1 \widehat T_n(x))H_{j_3}(\widetilde \partial_2 \widehat T_n(x))\,dx.
\end{split}
\end{equation}
Noting that, as $\varepsilon\to 0$,
$$
\beta_{l}^\varepsilon \to \beta_l,
$$
where $\beta_l$ are given in \paref{e:beta} and using Lemma \ref{lemma approx}, we prove both \paref{chaos dispari} and \paref{proiezione q}.
\
\noindent Let us now prove \paref{chaos 2} that allows to conclude the proof.
Equation \paref{proiezione q} with $q=1$ together with Equation \paref{beta piccoli} and Lemma \ref{alpha piccoli}, imply that the projection of $I_n$ on the second Wiener chaos equals the quantity
\begin{eqnarray*}
I_n[2]&:=& 2\pi^2 n \beta_0\beta_2 \alpha_{0,0,0,0} \int_{\field{T}} H_2(T_n(x)) dx
+2\pi^2 n \beta_2\beta_0 \alpha_{0,0,0,0} \int_{\field{T}} H_2(\widehat{T}_n(x)) dx \\
&& +2\pi^2 n \beta_0^2 \alpha_{2,0,0,0} \int_{\field{T}} H_2(\widetilde{\partial}_1 T_n(x))dx
+ 2\pi^2 n \beta_0^2 \alpha_{0,2,0,0} \int_{\field{T}} H_2(\widetilde{\partial}_2 T_n(x)) dx \\
&&+ 2\pi^2 n \beta_0^2 \alpha_{0,0,2,0} \int_{\field{T}} H_2(\widetilde{\partial}_1\widehat{T}_n(x))dx
+ 2\pi^2 n \beta_0^2 \alpha_{0,0,0,2} \int_{\field{T}} + H_2(\widetilde{\partial}_2\widehat{T}_n(x) ) dx \\
&= &\frac{\pi n}{2} \left\{ \int_{\field{T}} \big[H_2(\widetilde{\partial}_1 T_n(x)) + H_2(\widetilde{\partial}_2 T_n(x)) + H_2(\widetilde{\partial}_1\widehat{T}_n(x)) + H_2(\widetilde{\partial}_2\widehat{T}_n(x) ) \big] dx \right.\\
&& \left. \quad\quad\quad- 2\int_{\field{T}} \big[H_2(T_n(x)) + H_2(\widehat{T}_n(x))\big] dx\right\} .
\end{eqnarray*}
According to Green's first identity (see e.g. \cite[p. 44]{Lee}),
$$\int_{\field{T}}\nabla v\cdot\nabla w\ dx=-\int_{\field{T}}w\,\Delta v\ dx.$$
\noindent Using the facts that $H_2(t)=t^2-1$ {and that $T_n$ and $\widehat{T}_n$ are eigenfunctions of $\Delta$}, we eventually infer that
\begin{eqnarray*}
I_n[2]&= &\frac{1}{4\pi} \int_{\field{T}} \big[\|\nabla T_n(x)\|^2 + \|\nabla \widehat{T}_n(x)\|^2 \big] dx
-n\pi \int_{\field{T}} \big[T_n(x)^2 + \widehat{T}_n(x)^2\big] dx\\
&= &-\frac{1}{4\pi} \int_{\field{T}} \big[T_n(x)\, \Delta T_n(x) + \widehat{T}_n(x) \Delta \widehat{T}_n(x) \big] dx
-n\pi \int_{\field{T}} \big[T_n(x)^2 + \widehat{T}_n(x)^2\big] dx\\
&= &n\pi \int_{\field{T}} \big[T_n(x)^2 + \widehat{T}_n(x)^2 \big] dx
-n\pi \int_{\field{T}} \big[T_n(x)^2 + \widehat{T}_n(x)^2\big] dx=0.
\end{eqnarray*}
\end{proof}
\subsection{Proof of Part 1 of Theorem \ref{t:main}}\label{proof media sec}
According to Lemma \ref{alpha piccoli} and Equation \paref{beta piccoli}, for every $n\in S$ one has that
$$
I_n[0]=\mathbb{E}[I_n] = 2\pi^2 n\,\beta_0^2\,\alpha_{0,0,0,0} =\pi n = \frac{E_n}{4\pi},
$$
thus yielding the desired conclusion.
\section{Investigation of the fourth chaotic components}\label{sec chaos 4}
In this section we shall investigate fourth chaotic components. In particular, we shall prove Proposition \ref{varianze chaos} and Proposition \ref{limite 4}.
\subsection{Preliminary results}
For $n\in S$, from \paref{proiezione q} with $q=2$ we deduce that
\begin{equation}\label{eq 4 chaos}
\begin{split}
I_n[4]&=\frac{E_n}{2}\sum_{i_1+i_2+i_3+j_1+j_2+j_3=4} \frac{\beta_{i_1}\beta_{j_1}}{i_1! j_1!}\frac{\alpha_{i_2 i_3 j_2 j_3}}{i_2! i_3! j_2! j_3!}\times\cr
&\times \int_{\mathbb T} H_{i_1}(T_n(x))H_{j_1}(\widehat T_n(x))H_{i_2}(\widetilde \partial_1 T_n(x))H_{i_3}(\widetilde \partial_2 T_n(x))H_{j_2}(\widetilde \partial_1 \widehat T_n(x))H_{j_3}(\widetilde \partial_2 \widehat T_n(x))\,dx.
\end{split}
\end{equation}
where the sum only considers integers $i_1,j_1$ even and $i_2, i_3, j_2, j_3$ with the same parity. In order to compute an expression for $I_n[4]$ that is more amenable to analysis,
let us introduce, for $n\in S$, the following family of random variables:
\begin{eqnarray*}
W(n) &=& \frac{1}{\sqrt{\mathcal N_n}}\sum_{\lambda\in \Lambda_n}(|a_\lambda|^2-1),\\ \widehat W(n) &= &\frac{1}{\sqrt{\mathcal N_n}}\sum_{\lambda\in \Lambda_n}(|\widehat a_\lambda|^2-1)\\
W_j(n)&=&\frac{1}{n\sqrt{\mathcal N_n}}\sum_{\lambda\in \Lambda_n}\lambda_j^2(|a_\lambda|^2-1),\\ \widehat W_j(n)&=&\frac{1}{n\sqrt{\mathcal N_n}}\sum_{\lambda\in \Lambda_n}\lambda_j^2(|\widehat a_\lambda|^2-1),\quad j=1,2 \\
W_{1,2}(n)&=&\frac{1}{n\sqrt{\mathcal N_n}}\sum_{\lambda\in \Lambda_n}\lambda_1\lambda_2\,|a_\lambda|^2,\\ \widehat W_{1,2}(n)&=&\frac{1}{n\sqrt{\mathcal N_n}}\sum_{\lambda\in \Lambda_n}\lambda_1\lambda_2\,|\widehat a_\lambda|^2, \\
M(n)&=&\frac{1}{\sqrt{\mathcal N_n}} \sum_{\lambda\in \Lambda_n} a_\lambda \overline{\widehat{a}_\lambda} \\
M_j(n) &=& \frac{i}{\sqrt{n\mathcal N_n}} \sum_{\lambda\in \Lambda_n} \lambda_j a_\lambda \overline{\widehat{a}_\lambda},\quad j=1,2 \\
M_{\ell,j}(n)&=&\frac{1}{n\sqrt{\mathcal N_n}}\sum_{\lambda\in \Lambda_n} \lambda_\ell \lambda_j \ a_\lambda \overline{\widehat{a}_\lambda}\qquad j,\ell=1,2.
\end{eqnarray*}
Note that
$$
W_{1,2}(n)=\frac{1}{n\sqrt{\mathcal N_n}}\sum_{\lambda\in \Lambda_n}\lambda_1\lambda_2\,(|a_\lambda|^2-1),\quad \mbox{and}\quad \widehat W_{1,2}(n)=\frac{1}{n\sqrt{\mathcal N_n}}\sum_{\lambda\in \Lambda_n}\lambda_1\lambda_2\,(|\widehat a_\lambda|^2-1),
$$
since $\sum_{\lambda\in \Lambda_n} \lambda_1\lambda_2 = 0$, and also that $M_j$ is real-valued for $j=1,2$.
Now, let us express each summand appearing on the {}{right-hand side} of \paref{eq 4 chaos} in terms of $W{}{(n)}$, $W_1{}{(n)}$, $W_2{}{(n)}$, $W_{1,2}{}{(n)}$, $\widehat W{}{(n)}$, $\widehat W_1{}{(n)}$, $\widehat W_2{}{(n)}$, $\widehat W_{1,2}{}{(n)}$, $M{}{(n)}$, $M_1{}{(n)}$, $M_2{}{(n)}$, $M_{1,1}{}{(n)}$, $M_{2,2}{}{(n)}$ and/or $M_{1,2}{}{(n)}$. The proof of the following result will be given in Section \ref{appendix}. In what follows, the symbol $o_{{\mathbb{P}}}(1)$ indicates a sequence of random variables converging to zero in probability. In view of Remark \ref{r:postmain}-7, we will focus on sequences $\{n_j\}$ such that $\widehat{\mu}_{n_j}(4)$ converges to some number $\eta\in [-1,1]$.
\begin{lemma}\label{lemma infinito}
Let $\lbrace n_j\rbrace\subset S$ be such that $\mathcal N_{n_j}\to +\infty$ and $\widehat{\mu}_{n_j}(4)\to \eta\in[-1,1]$, then
\begin{enumerate}
\item[(i)] $\int_{\field{T}} H_4(T_{n_j}(x))\,dx=\frac{3}{\mathcal{N}_{n_j}}\big(W{}{(n_j)}^2-2+o_\mathbb{P}(1)\big)$;
\item[(ii)] $\int_{\field{T}} H_4(\widetilde{\partial}_k T_{n_j}(x))\,dx=\frac{3}{\mathcal{N}_{n_j}}\big(4W_k{}{(n_j)}^2-3-\eta+o_\mathbb{P}(1)\big)$, $k=1,2$;
\item[(iii)] $\int_{\field{T}} H_2(T_{n_j}(x))\big(H_2(\widetilde{\partial}_1 T_{n_j})(x))+H_2(\widetilde{\partial}_2 T_{n_j})(x))\big)\,dx=\frac{2}{\mathcal{N}_{n_j}}\big(W{}{(n_j)}^2-2+o_\mathbb{P}(1)\big)$;
\item[(iv)] $\int_{\field{T}} H_2(\widetilde{\partial}_1 T_{n_j}(x))H_2(\widetilde{\partial}_2 T_{n_j}(x))\,dx=\frac{1}{\mathcal{N}_{n_j}}\big(4W_{1}{}{(n_j)}W_2{}{(n_j)}+8W_{1,2}{}{(n_j)}^2-3+3\eta+o_\mathbb{P}(1)\big)$;
\item[(v)] $\int_{\field{T}} H_2(T_{n_j})(x))H_2(\widehat T_{n_j})(x))\,dx=\frac{1}{\mathcal{N}_{n_j}}\big(W{}{(n_j)}\widehat{W}{}{(n_j)}+2M{}{(n_j)}^2-2+o_\mathbb{P}(1)\big)$;
\item[(vi)] $\int_{\field{T}} H_2(T_{n_j})(x))\big(H_2(\widetilde{\partial}_1 \widehat T_{n_j}(x))+H_2(\widetilde{\partial}_2 \widehat T_{n_j})(x))\big)\,dx=\frac{2}{\mathcal{N}_{n_j}}\big(W{}{(n_j)}\widehat{W}{}{(n_j)}+M_1{}{(n_j)}^2+M_2{}{(n_j)}^2-1+o_\mathbb{P}(1)\big)$;
\item[(vii)] $\int_{\field{T}} H_2(\widetilde{\partial}_\ell T_{n_j}(x))H_2(\widetilde{\partial}_k \widehat T_{n_j}(x))\,dx=\frac{1}{\mathcal{N}_{n_j}}\big(4W_\ell{}{(n_j)} \widehat{W}_j{}{(n_j)}+8M_{\ell,k}{}{(n_j)}^2-(3+\eta){\bf 1}_{\{\ell= k\}}-(1-\eta){\bf 1}_{\{\ell\neq k\}}+o_\mathbb{P}(1)\big)$, $\ell,k=1,2$;
\item[(viii)] $\int_{\field{T}} \widetilde{\partial}_1 T_{n_j}(x)\widetilde{\partial}_2 T_{n_j}(x)\widetilde{\partial}_1 \widehat T_{n_j}(x)\widetilde{\partial}_2 \widehat T_{n_j}(x)\,dx=\frac{1}{\mathcal{N}_{n_j}}\big(4W_{1,2}{}{(n_j)}\widehat{W}_{1,2}{}{(n_j)}+4M_{1,1}{}{(n_j)}M_{2,2}{}{(n_j)}+4M_{1,2}{}{(n_j)}^2-1+\eta+o_\mathbb{P}(1)\big)$.
\end{enumerate}
\end{lemma}
We are now able to give an explicit expression for $I_n[4]$ in \paref{eq 4 chaos}.
\begin{lemma}\label{carino}
Let $\lbrace n_j\rbrace\subset S$ such that $\mathcal N_{n_j}\to +\infty$ and $\widehat{\mu}_{n_j}(4)\to \eta\in [-1,1]$, then
\begin{equation}\label{in4}
\begin{split}
I_{n_j}[4]=
\frac{n_j\pi}{8\,\mathcal N_{n_j}}\Big(&
\frac12W{}{(n_j)}^2+\frac12\widehat{W}{}{(n_j)}^2-3W{}{(n_j)}\widehat{W}{}{(n_j)}- W_1{}{(n_j)}^2 - W_2{}{(n_j)}^2 -\widehat{W}_1{}{(n_j)}^2\\
& - \widehat{W}_2{}{(n_j)}^2
+6W_1{}{(n_j)}\widehat{W}_2{}{(n_j)}+6\widehat{W}_1{}{(n_j)}W_2{}{(n_j)} -2W_{1,2}{}{(n_j)}^2
-2\widehat{W}_{1,2}{}{(n_j)}^2 \\
&- 12 W_{1,2}{}{(n_j)}\widehat{W}_{1,2}{}{(n_j)} - 4M_1{}{(n_j)}^2 -4M_2{}{(n_j)}^2 +4M{}{(n_j)}^2-2M_{1,1}{}{(n_j)}^2\\
&-2M_{2,2}{}{(n_j)}^2
-12M_{1,1}{}{(n_j)}M_{2,2}{}{(n_j)}+8M_{1,2}{}{(n_j)}^2+4+o_{\mathbb{P}}(1)
\Big).
\end{split}
\end{equation}
\end{lemma}
\noindent\begin{proof}
From Lemma \ref{alpha piccoli} and \paref{eq 4 chaos}, we find that
\begin{eqnarray}
I_n[4]&=&\frac{n\pi}{64}\,\Big(8\int_{\field{T}} H_4(T_n(x))\,dx -8\int_{\field{T}} H_2(T_n(x))H_2(\widetilde{\partial}_1 T_n(x))\,dx\notag\\
&&-8\int_{\field{T}} H_2(T_n(x))H_2(\widetilde{\partial}_2 T_n(x))\,dx- 2\int_{\field{T}} H_2(\widetilde{\partial}_1 T_n(x))H_2(\widetilde{\partial}_2 T_n(x))\,dx\notag\\
&&-\int_{\field{T}} H_4(\widetilde{\partial}_1 T_n(x))\,dx
-\int_{\field{T}} H_4(\widetilde{\partial}_2 T_n(x))\,dx \notag\\
&&+8\int_{\field{T}} H_4(\widehat T_n(x))\,dx -8\int_{\field{T}} H_2(\widehat T_n(x))H_2(\widetilde{\partial}_1 \widehat T_n(x))\,dx\notag
\end{eqnarray}
\begin{eqnarray}
&&-8\int _{\field{T}}H_2(\widehat T_n(x))H_2(\widetilde{\partial}_2 \widehat T_n(x))\,dx- 2\int_{\field{T}} H_2(\widetilde{\partial}_1\widehat T_n(x))H_2(\widetilde{\partial}_2 \widehat T_n(x))\,dx\notag\\
&&-\int_{\field{T}} H_4(\widetilde{\partial}_1\widehat T_n(x))\,dx -\int_{\field{T}} H_4(\widetilde{\partial}_2 \widehat T_n(x))\,dx \notag\\
&&+16\int_{\field{T}} H_2(T_n(x))H_2(\widehat T_n(x))\,dx
-8\int_{\field{T}} H_2(T_n(x))\big(H_2(\widetilde{\partial}_1 \widehat T_n(x))+H_2(\widetilde{\partial}_2 \widehat T_n(x))\big)\,dx\notag\\
&&
-8\int_{\field{T}} H_2(\widehat T_n(x))\big(H_2(\widetilde{\partial}_1 T_n(x))+H_2(\widetilde{\partial}_2 T_n(x))\big)dx\notag\\
&&
-2\int_{\field{T}} H_2(\widetilde{\partial}_1 T_n(x))H_2(\widetilde{\partial}_1 \widehat T_n(x))\,dx
-2\int_{\field{T}} H_2(\widetilde{\partial}_2 T_n(x))H_2(\widetilde{\partial}_2 \widehat T_n(x))\,dx\notag\\
&&+10\int_{\field{T}} H_2(\widetilde{\partial}_1 T_n(x))H_2(\widetilde{\partial}_2 \widehat T_n(x))\,dx
+10\int_{\field{T}} H_2(\widetilde{\partial}_2 T_n(x))H_2(\widetilde{\partial}_1 \widehat T_n(x))\,dx\notag\\
&&-24\int_{\field{T}} \widetilde{\partial}_1 T_n(x)\widetilde{\partial}_2 T_n(x)\widetilde{\partial}_1 \widehat T_n(x)\widetilde{\partial}_2 \widehat T_n(x)\,dx
\Big ).\label{eq infinita}
\end{eqnarray}
Using the previous identities (i)-(viii) in Lemma \ref{lemma infinito} in \paref{eq infinita}, and also using that
$W_1{}{(n_j)} + W_2{}{(n_j)} =W{}{(n_j)}$ and
$\widehat W_1{}{(n_j)} + \widehat W_2{}{(n_j)} = \widehat W{}{(n_j)}$, one concludes the proof.
\end{proof}
\subsection{Proofs of Proposition \ref{varianze chaos} and Proposition \ref{limite 4}}
Let us first study the asymptotic distribution of the centered random vector, defined for $n\in S$ as follows
\begin{equation*}
\begin{split}
{\bf W}(n):=(&W(n),W_1(n), W_2(n), W_{1,2}(n), \widehat W(n), \widehat W_1(n), \widehat W_2(n), \widehat W_{1,2}(n),\cr
&M(n),M_1(n),M_2(n),M_{1,1}(n),M_{2,2}(n),M_{1,2}(n))\in \mathbb R^{14}.
\end{split}
\end{equation*}
\begin{lemma}\label{CLT}
Let $\lbrace n_j\rbrace\subset S$ be such that $\mathcal N_{n_j}\to +\infty$ and $\widehat{\mu}_{n_j}(4)\to \eta\in [-1,1]$. Then, as $\mathcal N_{n_j}\to \infty$,
$$
{\bf W}(n_j) \stackrel{\rm law}{\Longrightarrow} {\bf G},
$$
where ${\bf G}=(G_1,\ldots,G_{14})$ denotes a Gaussian real centered vector with
covariance matrix given by
\begin{equation}\label{matrix M}
{\bf M}(\eta)=\left(\begin{matrix}
&{\bf A}{}{(\eta)} &0&0\\
&0&{\bf A}{}{(\eta)} &0\\
&0&0 &{\bf B}{}{(\eta)} \\
\end{matrix}\ \ \right ),
\end{equation}
where
$$
{\bf A}(\eta):=\left(\begin{matrix}
&2 &1 &1 &0 \\
&1 &\frac{3+\eta}{4} &\frac{1-\eta}{4} &0\\
&1 &\frac{1-\eta}{4} &\frac{3+\eta}{4} &0 \\
&0 &0 &0 &\frac{1-\eta}{4} \\
\end{matrix}\ \ \right ),
$$
and
$$
{\bf B}(\eta):=\left(\begin{matrix} &1 &0 &0 &\frac12 &\frac12 &0\\
&0 &\frac12 &0 &0 &0 &0 \\
&0 &0 &\frac12 &0 &0 &0\\
&\frac12 &0 &0 &\frac{3+\eta}{8} & \frac{1-\eta}{8} &0\\
&\frac12 &0 &0 &\frac{1-\eta}{8} &\frac{3+\eta}{8} &0\\
&0 &0 &0 &0 &0 &\frac{1-\eta}{8}
\end{matrix}\ \ \right ).
$$
\end{lemma}
\noindent\begin{proof}
First, for reasons related to independence it is easy to check that the covariance matrix of ${\bf W}(n)$ takes the form
\begin{equation}\label{sigma grande}
\Sigma_{{}{n}} =\left(\begin{matrix}
&{\bf A}_n &0&0\\
&0&{\bf A}_n&0\\
&0&0 &{\bf B}_n\\
\end{matrix}\ \ \right ),
\end{equation}
where ${\bf A}_n$ and ${\bf B}_n$ denote the covariance matrices of $(W(n),W_1(n),W_2(n),W_{1,2}(n))$ and
$(M(n),M_1(n),M_2(n),M_{1,1}(n),M_{2,2}(n),M_{1,2}(n))$ respectively. Let us first compute ${\bf A}_n$.
Since ${\mathbb{ E}}[(|a_\lambda|^2-1)(|a_{\lambda'}|^2-1)]=1$ if $\lambda=\pm \lambda'$ and is zero otherwise, one has
$$
{\mathbb{ E}}(W{}{(n)}^2) = \frac1{\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}{\mathbb{ E}}[(|a_\lambda|^2-1)(|a_{\lambda'}|^2-1)]
=2.
$$
Similarly,
$$
{\mathbb{ E}}(W{}{(n)}W_j{}{(n)}) = \frac1{n\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}\lambda_j^2{\mathbb{ E}}[(|a_\lambda|^2-1)(|a_{\lambda'}|^2-1)]
=\frac2{n\mathcal{N}_n}\sum_{\lambda\in\Lambda_n}\lambda_j^2=1,
$$
whereas
$$
{\mathbb{ E}}(W{}{(n)}W_{1,2}{}{(n)}) = \frac1{n\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}\lambda_1\lambda_2{\mathbb{ E}}[(|a_\lambda|^2-1)(|a_{\lambda'}|^2-1)]
=\frac2{n\mathcal{N}_n}\sum_{\lambda\in\Lambda_n}\lambda_1\lambda_2=0.
$$
We also have
$$
{\mathbb{ E}}(W_j{}{(n)}^2) = \frac1{n^2\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}\lambda_j^2{\lambda'_j}^2{\mathbb{ E}}[(|a_\lambda|^2-1)(|a_{\lambda'}|^2-1)]
=\frac2{n^2\mathcal{N}_n}\sum_{\lambda\in\Lambda_n}\lambda_j^4.$$
To express ${\mathbb{ E}}(W_j{}{(n)}^2)$ in a more suitable way, let us rely on $\widehat{\mu}_n(4)$:
\begin{eqnarray*}
\widehat{\mu}_n(4)&=&\int_{\mathcal{S}^1}z^4d\mu_n(z)=\frac{1}{n^2\mathcal{N}_n}\sum_{\lambda\in \Lambda_n}
(\lambda_1+i\lambda_2)^4
=\frac{1}{n^2\mathcal{N}_n}\sum_{\lambda\in \Lambda_n}
(\lambda_1^4-6\lambda_1^2\lambda_2^2+\lambda_2^4)\\
&=&\frac{1}{n^2\mathcal{N}_n}\sum_{\lambda\in \Lambda_n}(\lambda_1^2+\lambda_2^2)^2 -
\frac{8}{n^2\mathcal{N}_n}\sum_{\lambda\in \Lambda_n}\lambda_1^2\lambda_2^2
=1 -
\frac{8}{n^2\mathcal{N}_n}\sum_{\lambda\in \Lambda_n}\lambda_1^2\lambda_2^2.
\end{eqnarray*}
As a result,
$$
\sum_{\lambda\in \Lambda_n}\lambda_1^2\lambda_2^2 = \frac{n^2\mathcal{N}_n}{8}(1-\widehat{\mu}_n(4)),
$$
leading to
$$
{\mathbb{ E}}(W_j{}{(n)}^2) = \frac1{n^2\mathcal{N}_n}\sum_{\lambda\in\Lambda_n}(\lambda_1^4+\lambda_2^4)
=\widehat{\mu}_n(4)+\frac{6}{n^2\mathcal{N}_n}\sum_{\lambda\in \Lambda_n}\lambda_1^2\lambda_2^2
=\frac14(3+\widehat{\mu}_n(4)).
$$
Similarly,
\begin{eqnarray*}
{\mathbb{ E}}(W_{1,2}{}{(n)}^2) &= &\frac1{n^2\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}\lambda_1\lambda_2\lambda'_1\lambda'_2
{\mathbb{ E}}[|a_\lambda|^2|a_{\lambda'}|^2]=\frac2{n^2\mathcal{N}_n}\sum_{\lambda\in\Lambda_n}\lambda_1^2\lambda_2^2=\frac14(1-\widehat{\mu}_n(4)),
\end{eqnarray*}
as well as
\begin{eqnarray*}
{\mathbb{ E}}(W_1{}{(n)}W_2{}{(n)}) &= &\frac1{n^2\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}\lambda_1^2{\lambda'_2}^2
{\mathbb{ E}}[(|a_\lambda|^2-1)(|a_{\lambda'}|^2-1)]\\
&=&\frac2{n^2\mathcal{N}_n}\sum_{\lambda\in\Lambda_n}\lambda_1^2\lambda_2^2=\frac14(1-\widehat{\mu}_n(4)),
\end{eqnarray*}
and
\begin{eqnarray*}
{\mathbb{ E}}(W_j{}{(n)}W_{1,2}{}{(n)}) &= &\frac1{n^2\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}\lambda_j^2\lambda'_1\lambda'_2
{\mathbb{ E}}[(|a_\lambda|^2-1)(|a_{\lambda'}|^2-1)]\\
&=&\frac2{n^2\mathcal{N}_n}\sum_{\lambda\in\Lambda_n}\lambda_j^2\lambda_1\lambda_2=0.
\end{eqnarray*}
Taking all these facts into consideration, we deduce that
$${\bf A}_n=\left(\begin{matrix}
&2 &1 &1 &0 \\
&1 &\frac{3+\widehat \mu_n(4)}{4} &\frac{1-\widehat \mu_n(4)}{4} &0\\
&1 &\frac{1-\widehat \mu_n(4)}{4} &\frac{3+\widehat \mu_n(4)}{4} &0 \\
&0 &0 &0 &\frac{1-\widehat \mu_n(4)}{4} \\
\end{matrix}\ \ \right ).$$
Now, let us turn to the expression of ${\bf B}_n$.
Using that ${\mathbb{ E}}[a_\lambda a_{\lambda'}]=1$ if $\lambda'=-\lambda$ and is zero otherwise, we obtain
$$
{\mathbb{ E}}(M{}{(n)}^2)= \frac1{\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}{\mathbb{ E}}[a_\lambda a_{\lambda'}]{\mathbb{ E}}[ \overline{\widehat{a}_\lambda} \overline{\widehat{a}_{\lambda'}}]
=1.
$$
Similarly,
$$
{\mathbb{ E}}(M_j{}{(n)}^2)= -\frac1{n\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}\lambda_j\lambda'_j{\mathbb{ E}}[a_\lambda a_{\lambda'}]{\mathbb{ E}}[ \overline{\widehat{a}_\lambda} \overline{\widehat{a}_{\lambda'}}]
=\frac1{n\mathcal{N}_n}\sum_{\lambda\in\Lambda_n}\lambda_j^2=\frac12,
$$
as well as
\begin{eqnarray*}
{\mathbb{ E}}(M_{j,j}{}{(n)}^2)&=& \frac1{n^2\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}\lambda_j^2{\lambda'_j}^2{\mathbb{ E}}[a_\lambda a_{\lambda'}]{\mathbb{ E}}[ \overline{\widehat{a}_\lambda} \overline{\widehat{a}_{\lambda'}}]\\
&=&\frac1{n^2\mathcal{N}_n}\sum_{\lambda\in\Lambda_n}\lambda_j^4=\frac12{\mathbb{ E}}(W_j^2)=\frac18(3+\widehat{\mu}_n(4)),
\end{eqnarray*}
and
\begin{eqnarray*}
{\mathbb{ E}}(M_{1,2}{}{(n)}^2)&=& \frac1{n^2\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}\lambda_1\lambda_2{\lambda'_1}{\lambda'_2}{\mathbb{ E}}[a_\lambda a_{\lambda'}]{\mathbb{ E}}[ \overline{\widehat{a}_\lambda} \overline{\widehat{a}_{\lambda'}}]\\
&=&\frac1{n^2\mathcal{N}_n}\sum_{\lambda\in\Lambda_n}\lambda_1^2\lambda_2^2=\frac12{\mathbb{ E}}(W_{12}^2)=\frac18(1-\widehat{\mu}_n(4)).
\end{eqnarray*}
Besides, it is immediate to check that, for any $l,j$,
$$
{\mathbb{ E}}(M{}{(n)}M_j{}{(n)})={\mathbb{ E}}(M{}{(n)}M_{12}{}{(n)})={\mathbb{ E}}(M_j{}{(n)}M_{l,j}{}{(n)})={\mathbb{ E}}(M_{j,j}{}{(n)}M_{1,2}{}{(n)})=0.
$$
Finally,
$$
{\mathbb{ E}}(M{}{(n)}M_{j,j}{}{(n)})= \frac1{n\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}\lambda_j^2{\mathbb{ E}}[a_\lambda a_{\lambda'}]{\mathbb{ E}}[ \overline{\widehat{a}_\lambda} \overline{\widehat{a}_{\lambda'}}]
=\frac1{n^2\mathcal{N}_n}\sum_{\lambda\in\Lambda_n}\lambda_j^2=\frac12,
$$
whereas
\begin{eqnarray*}
{\mathbb{ E}}(M_{1,1}{}{(n)}M_{2,2}{}{(n)})&=& \frac1{n^2\mathcal{N}_n}\sum_{\lambda,\lambda'\in\Lambda_n}\lambda_1^2{\lambda'_2}^2\,
{\mathbb{ E}}[a_\lambda a_{\lambda'}]{\mathbb{ E}}[ \overline{\widehat{a}_\lambda} \overline{\widehat{a}_{\lambda'}}]\\
&=&\frac1{n^2\mathcal{N}_n}
\sum_{\lambda\in\Lambda_n}\lambda_1^2\lambda_2^2
=\frac12{\mathbb{ E}}(W_{12}{}{(n)}^2)=\frac18(1-\widehat{\mu}_n(4)).
\end{eqnarray*}
Putting everything together, we arrive at the following expression for ${\bf B}_n$
$$
{\bf B}_n=\left(\begin{matrix} &1 &0 &0 &\frac12 &\frac12 &0\\
&0 &\frac12 &0 &0 &0 &0 \\
&0 &0 &\frac12 &0 &0 &0\\
&\frac12 &0 &0 &\frac{3+\widehat \mu_n(4)}{8} & \frac{1-\widehat \mu_n(4)}{8} &0\\
&\frac12 &0 &0 &\frac{1-\widehat \mu_n(4)}{8} &\frac{3+\widehat \mu_n(4)}{8} &0\\
&0 &0 &0 &0 &0 &\frac{1-\widehat \mu_n(4)}{8}
\end{matrix}\ \ \right ).
$$
{
Now, let us prove that each component of ${\bf W}_{n_j}$ is asymptotically Gaussian as $\mathcal{N}_{n_j}\to +\infty$.
Since all components of ${\bf W}_{n_j}$ belong to the same {{} Wiener} chaos (the second one) and have a converging variance (see indeed the diagonal part of ${\bf B}_n$ just above), according to the Fourth Moment Theorem (see, e.g., \cite[Theorem 5.2.7]{NP}) it suffices to show that the fourth cumulant of each component of ${\bf W}_{n_j}$ goes to zero as $\mathcal{N}_{n_j}\to +\infty$. Since we are dealing with sum of independent random variables,
checking such a property is straightforward. For sake of illustration, let us only consider the case of $W_2(n_j)$ which is representative of the difficulty. {{} We recall that, given a real-valued random variable $Z$ with mean zero, the fourth cumulant of $Z$ is defined by $\kappa_4(Z) :={\mathbb{ E}}[Z^4] - 3{\mathbb{ E}}[Z^2]$. } Since the $a_\lambda$ are independent except for the relation $\overline{a_\lambda}=a_{-\lambda}$, we can write, setting $\Lambda_n^+=\{\lambda\in\Lambda_n:\,\lambda_2>0\}$,
\begin{eqnarray*}
\kappa_4(W_2(n))&=&
\kappa_4\bigg(\frac{2}{n\sqrt{\mathcal{N}_{n}}}
\sum_{\lambda\in\Lambda_n^+}\lambda_2^2(|a_\lambda|^2-1)\bigg)
=\frac{16\,\kappa_4(|N_\mathbb{C}(0,1)|^2)}{n^4\mathcal{N}_{n}^2}
\sum_{\lambda\in\Lambda_n^+}\lambda_2^8\\
&\leq&
\frac{8\,\kappa_4(|N_\mathbb{C}(0,1)|^2)}{\mathcal{N}_{n}};
\end{eqnarray*}
to obtain the last inequality, we have used that $\lambda_2^2\leq \lambda_1^2+\lambda_2^2= n$.
As a result, $\kappa_4(W_2(n_j))\to 0$ as $\mathcal{N}_{n_j}\to +\infty$
and it follows from the Fourth Moment Theorem that $W_2(n_j)$ is asymptotically Gaussian. It is not difficult to apply a similar strategy in order to prove that, actually, each component of ${\bf W}_{n_j}$ is asymptotically Gaussian as well; details are left to the reader.
}
{
Finally, we make use of
\cite[Theorem 6.2.3]{NP} to conclude the proof of Lemma \ref{CLT}.
}
Indeed, (i) all components of ${\bf W}_n$ belong to the same Wiener chaos (the second one), (ii) each component of ${\bf W}_{n_j}$ is asymptotically Gaussian (as $\mathcal{N}_{n_j}\to +\infty$), and finally (iii) $\Sigma_{k,l}(n_j)\to {\bf M}_{k,l}(\eta)$ for each pair of indices $(k,l)$.
\end{proof}
\bigskip
\noindent\begin{proof}[Proofs of Proposition \ref{varianze chaos} and Proposition \ref{limite 4}]
For each subsequence $\lbrace n'_j\rbrace\subset \lbrace n_j\rbrace$, there exists a subsubsequence $\lbrace n''_j\rbrace\subset \lbrace n'_j\rbrace$ such that it holds either (i) $\widehat{\mu}_{n''_j}(4)\to \eta$ or (ii) $\widehat{\mu}_{n''_j}(4)\to -\eta$.
Combining Lemma \ref{carino} with Lemma \ref{CLT}, we have, as $j\to +\infty$,
\begin{equation}\label{limit}
\begin{split}
\frac{8\mathcal{N}_{n''_j}}{n''_j\pi}I_{n''_j}[4]\Rightarrow &\frac12G_1^2+\frac12G_5^2-3G_1G_5- G_2^2 - G_3^2 -G_6^2 - G_7^2 +6G_2G_7+6G_6G_3-2G_4^2-2G_8^2\notag\\
&-12G_4G_8-4G_{10}^2-4G_{11}^2+4G_9^2-2G_{12}^2-2G_{13}^2+8G_{14}^2-12G_{12}G_{13},
\end{split}
\end{equation}
where $(G_1,\ldots,G_{14})$ denotes a Gaussian centered vector with covariance matrix \paref{matrix M}.
Since $\left \lbrace \frac{8N_{n''_j}}{n''_j\pi}I_{n''_j}[4] \right \rbrace$ is a sequence of random variables belonging to a fixed Wiener chaos and converging in distribution, by
standard arguments based on uniform integrability, we also have
\begin{equation*}
\begin{split}
{\rm Var}&\left(\frac{8N_{n''_j}}{n''_j\pi}I_{n''_j}[4]\right)\to {\rm Var}\Big(\frac12G_1^2+\frac12G_5^2-3G_1G_5- G_2^2 - G_3^2 -G_6^2 - G_7^2 +6G_2G_7+6G_6G_3\notag\\
&-2G_4^2-2G_8^2-12G_4G_8-4G_{10}^2-4G_{11}^2+4G_9^2-2G_{12}^2-2G_{13}^2+8G_{14}^2-12G_{12}G_{13}\Big);
\end{split}
\end{equation*}
the proof of Proposition \ref{varianze chaos} is then concluded, once computing
\begin{equation*}
\begin{split}
{\rm Var}\Big(&\frac12G_1^2+\frac12G_5^2-3G_1G_5- G_2^2 - G_3^2 -G_6^2 - G_7^2 +6G_2G_7+6G_6G_3-2G_4^2-2G_8^2\\
&-12G_4G_8-4G_{10}^2-4G_{11}^2+4G_9^2-2G_{12}^2-2G_{13}^2+8G_{14}^2-12G_{12}G_{13}\Big)= 8(3\eta^2 +5),
\end{split}
\end{equation*}
and noting that the latter variance is the same in both cases (i) and (ii).
Let us now prove Proposition \ref{limite 4}.
Let $(Z_1,\ldots,Z_{11})\sim N_{11}(0,I)$ be a standard Gaussian vector of $\field{R}^{11}$. Then one can check that the vector
$$
\left(
\begin{matrix}
\sqrt{2}\,Z_5\\
\frac{1}{\sqrt{2}}Z_5+\frac12\sqrt{\eta+1}\,Z_3\\
\frac{1}{\sqrt{2}}Z_5-\frac12\sqrt{\eta+1}\,Z_3\\
\frac12\sqrt{1-\eta}\,Z_8\\
\sqrt{2}\,Z_6\\
\frac{1}{\sqrt{2}}Z_6+\frac12\sqrt{\eta+1}\,Z_4\\
\frac{1}{\sqrt{2}}Z_6-\frac12\sqrt{\eta+1}\,Z_4\\
\frac12\sqrt{1-\eta}\,Z_9\\
Z_2\\
\frac{1}{\sqrt{2}}Z_{10}\\
\frac{1}{\sqrt{2}}Z_{11}\\
\frac12Z_2+\sqrt{\frac18(\eta+1)}\,Z_1\\
\frac12Z_2-\sqrt{\frac18(\eta+1)}\,Z_1\\
\sqrt{\frac18(1-\eta)}\,Z_7
\end{matrix}
\right)
$$
admits ${\bf M}{}{(\eta)}$ for covariance matrix as well.
Expressing (\ref{limit}) in terms of $(U_{1},\ldots,U_{11})$ leads to the fact that (\ref{limit}) has the same law
as $$
\frac{1+\eta}{2}A
+\frac{1-\eta}{2}B
-2(C-2),$$
with $A,B,C$ independent and
$A\overset{\rm law}{=}B\overset{\rm law}{=}2Z_1^2-Z_2^2-Z_3^2-6Z_2Z_3$ and
$C\overset{\rm law}{=}Z_1^2+Z_2^2$.
Finally, noting that the law of the random variable $\frac{1+\eta}{2}A
+\frac{1-\eta}{2}B
-2(C-2)$ is the same for case (i) and case (ii) and using that $(Z_1,Z_2,Z_3)\overset{\rm law}{=}(Z_1,\frac1{\sqrt{2}}(Z_2-Z_3),\frac1{\sqrt{2}}(Z_2+Z_3))$,
we get the desired conclusion.
\end{proof}
\section{The variance of higher order chaoses}\label{proof varianza sec}
In this section we shall prove Proposition \ref{asymptotic}.
Let us decompose the torus $\mathbb T$ as a disjoint union of squares $Q_k$ of side length $1/M$ (where $M\approx \sqrt{E_n}$ is a large integer\footnote{{$M= {{} \lceil d \sqrt{E_n} \rceil}$, $d\in \field{R}_{>0}$}}), obtained by translating along directions ${k}/M$, ${k}\in \mathbb Z^2$, the square $Q_0 := [0,1/M)\times [0,1/M)$ containing the origin. By construction, the south-west corner of each square is therefore situated at the point ${k}/M$.
\subsection{Singular points and cubes}
Let us first give some definitions, inspired by \cite[\S 6.1]{ORW} and \cite[\S 4.3]{RW2}. Let us denote by {\color{black} $0<\varepsilon_1<\frac{1}{10^{10} }$} a very small number\footnote{{Let us now choose $d$ such that $d\ge \frac{16\pi^2}{\varepsilon_1}$}.} that will be fixed until the end. From now on, we shall use the simpler notation $r_j:=\partial_j r_n$, and $r_{ij} := \partial_{ij} r_n$ for $i,j=1,2$.
\begin{definition}[Singular pairs of points and cubes]\label{sing point}
\
\noindent i) A pair of points $(x,y)\in \mathbb T\times \mathbb T$ is called singular if either $|r(x-y)|>\varepsilon_1$ or $|r_1(x-y)|>\varepsilon_1\, \sqrt{n}$ or $|r_2(x-y)|>\varepsilon_1\, \sqrt{n}$ or $|r_{12}(x-y)|>\varepsilon_1\, n$ or $|r_{11}(x-y)|>\varepsilon_1\, n$ or $|r_{22}(x-y)|>\varepsilon_1\, n$.
\noindent ii) A pair of cubes $(Q,Q')$ is called singular if the product $Q\times Q'$ contains a singular pair of points.
\end{definition}
For instance, $(0,0)$ is a singular pair of points and hence $(Q_0,Q_0)$ is a singular pair of cubes. In what follows we will often drop the dependence of $k$ from $Q_k$.
\begin{lemma}\label{lemma sing}
Let $(Q, Q')$ be a singular pair of cubes, then $|r(z-w)|>\frac12\varepsilon_1$ or $|r_1(z-w))|>\frac12\,\varepsilon_1\, \sqrt{n}$ or $|r_2(z-w))|>\frac12\varepsilon_1\, \sqrt{n}$ or $|r_{12}(z-w))|>\frac12\varepsilon_1\, n$ or $|r_{11}(z-w))|>\frac12\varepsilon_1\, n$ or $|r_{22}(z-w))|>\frac12\varepsilon_1\, n$, for every $(z,w)\in Q\times Q'$.
\end{lemma}
\noindent
\begin{proof} First note that the function $\mathbb T\ni s\mapsto r(s/\sqrt{n})$ and its derivatives up to the order two are Lipschitz with a universal Lipschitz constant $c=8\pi^3$ (in particular, independent of $n$).
Let us denote by $(x,y)$ the singular pair of points contained in $Q\times Q'$ and suppose that $r(x-y)>\varepsilon_1$. For every $(z,w)\in Q\times Q'$,
\begin{equation*}
\begin{split}
|r(z-w) - r(x-y)| &= \left |r\left( \frac{(z-w)\cdot \sqrt n}{\sqrt n} \right)-r\left( \frac{(x-y)\cdot \sqrt n}{\sqrt n} \right) \right |\cr
& \le c\sqrt n |(z-x) - (w-y)|\le 2c \sqrt{n} \frac{1}{M}.
\end{split}
\end{equation*}
{Since $d\ge \frac{16\pi^2}{\varepsilon_1}$ in $M= {{} \lceil d \sqrt{E_n} \rceil}$}, then
\begin{equation*}
\begin{split}
r(z-w) \ge r(x-y) - \varepsilon_1 /2 > \varepsilon_1 /2.
\end{split}
\end{equation*}
The case $r(x-y) < -\varepsilon_1$ in indeed analogous.
The rest of the proof for derivatives follows the same argument.
\end{proof}
Let us now denote by $B_Q$ the union of all squares $Q'$ such that $(Q,Q')$ is a singular pair. The number of such cubes $Q'$ is $M^2 {\rm Leb}(B_Q)$, the area of each cube being $1/M^2$.
\begin{lemma}\label{bello}
It holds that
$\text{Leb}(B_Q) \ll \int_{\mathbb T} r(x)^6\,dx$.
\end{lemma}
\noindent
\begin{proof}
Let us first note that
$$
B_Q \subset B_{Q}^0 \cup B_Q^1 \cup B_Q^2 \cup B_Q^{12} \cup B_Q^{11} \cup B_Q^{22},
$$
where $B_{Q}^0$ is the union of all cubes $Q'$ such that there exists $(x,y)\in Q\times Q'$ enjoying $|r(x-y)|>\varepsilon_1$ and for $i,j=1,2$, $B_Q^{i}$ is the union of all cubes $Q'$ such that there exists $(x,y)\in Q\times Q'$ enjoying $|r_i(x-y)|>\varepsilon_1\, \sqrt n$ and finally $B_Q^{ij}$ is the union of all cubes $Q'$ such that there exists $(x,y)\in Q\times Q'$ enjoying $|r_{ij}(x-y)|>\varepsilon_1\, n$. We can hence write
\begin{equation*}
\text{Leb}(B_Q) \le \text{Leb}(B_{Q}^0)+\text{Leb}(B_Q^1)+ \text{Leb}(B_Q^2)+ \text{Leb}(B_Q^{12})+ \text{Leb}(B_Q^{11})+ \text{Leb}(B_Q^{22}).\end{equation*}
Let us now fix $z\in Q$; then Lemma \ref{lemma sing} yields
\begin{equation*}
\text{Leb}(B_Q^0) = \int_{B_Q^0} \frac{|r(z-w)|^6}{|r(z-w)|^6}\,dw\le \varepsilon_1^{-6} \int_{B_Q^0} |r(z-w)|^6\,dw \le \varepsilon_1^{-6}\int_{\mathbb T} |r(x)|^6\,dx.
\end{equation*}
Moreover, for $i=1,2$,
\begin{equation*}
\text{Leb}(B_Q^i) = \int_{B_Q^i} \frac{|\widetilde r_i(z-w)|^6}{|\widetilde r_i(z-w)|^6}\,dw\le \varepsilon_1^{-6} \int_{B_Q^i} |\widetilde r_i(z-w)|^6\,dw \le \varepsilon_1^{-6}\int_{\mathbb T} |\widetilde r_i(x)|^6\,dx,
\end{equation*}
where $\widetilde r_i:=r_i/\sqrt{E_n}$ are the normalized derivatives.
Since
\begin{equation*}
\begin{split}
\int_{\mathbb T} \widetilde r_i(x)^6\,dx&= \frac{1}{\mathcal N_n^6}\sum_{\lambda, \lambda',\dots, \lambda^{v}} \frac{\lambda_i}{\sqrt n} \frac{\lambda_i'}{\sqrt n}\cdots
\frac{\lambda_i^v}{\sqrt n}\int_{\mathbb T} {\rm e}^{i2\pi\langle \lambda -\lambda'+\dots -\lambda^v,x \rangle}\,dx \cr
&= \frac{1}{\mathcal N_n^6}\sum_{\lambda-\lambda'+\dots-\lambda^{v}=0} \frac{\lambda_i}{\sqrt n} \frac{\lambda_i'}{\sqrt n}\cdots
\frac{\lambda_i^v}{\sqrt n}\cr
&\le \frac{|S_6(n)|}{\mathcal N_n^6}= \int_{\mathbb T} r(x)^6\,dx,
\end{split}
\end{equation*}
we have
$$
\text{Leb}(B_Q^i) \ll \int_{\mathbb T} r(x)^6\,dx.
$$
An analogous argument applied to $B_Q^{ij}$ for $i,j=1,2$ allows to conclude the proof.
\end{proof}
The number of cubes $Q'$ such that the pair $(Q,Q')$ is singular is hence {negligible {{} with respect to}} $E_n\, R_n(6)$.
\subsection{Variance and cubes}
We write the total number $I_n$ of nodal intersections as the sum of the number $I_{n_{ |_Q}}$ of nodal intersections restricted to each square $Q$, i.e.
\begin{equation*}
I_n = \sum_Q I_{n_{ |_Q}}.
\end{equation*}
We have
\begin{equation*}
\text{proj}\left(I_n|C_{\ge 6}\right )= \sum_Q \text{proj}\left (I_{n_{ |_Q}}|C_{\ge 6} \right),
\end{equation*}
so that
\begin{equation*}
{\rm Var}\left(\text{proj}(I_n|C_{\ge 6}) \right )= \sum_{Q,Q'} {\rm Cov\,}\left (\text{proj}\left (I_{n_{ |_Q}}|C_{\ge 6}\right ), \text{proj}\left (I_{n_{ |_{Q'}}}|C_{\ge 6}\right )\right ).
\end{equation*}
We are going to separately investigate the contribution of the singular pairs and the non-singular pairs of cubes:
\begin{equation*}
\begin{split}
{\rm Var}(\text{proj}(I_n|C_{\ge 6})) =& \sum_{(Q,Q')\text{ sing.}} {\rm Cov\,}(\text{proj}(I_{n_{ |_Q}}|C_{\ge 6}), \text{proj}(I_{n_{ |_{Q'}}}|C_{\ge 6})) \cr
&+ \sum_{(Q,Q')\text{ non sing.}} {\rm Cov\,}(\text{proj}(I_{n_{ |_Q}}|C_{\ge 6}), \text{proj}(I_{n_{ |_{Q'}}}|C_{\ge 6})).
\end{split}
\end{equation*}
\subsubsection{The contribution of singular pairs of cubes}
\begin{proof}[Proof of Lemma \ref{varianza sing}]
By Cauchy-Schwarz inequality and the stationarity of ${\bf T}_n$, recalling moreover
Lemma \ref{bello}, we have
\begin{equation*}
\begin{split}
&\left | \sum_{(Q,Q')\text{ sing.}} {\rm Cov\,}\left (\text{proj}\left (I_{n_{ |_Q}}|C_{\ge 6}\right ), \text{proj}\left (I_{n_{ |_{Q'}}}|C_{\ge 6}\right )\right ) \right | \cr
&\le \sum_{(Q,Q')\text{ sing.}} \left |{\rm Cov\,}\left (\text{proj}\left (I_{n_{ |_Q}}|C_{\ge 6}\right ), \text{proj}\left (I_{n_{ |_{Q'}}}|C_{\ge 6}\right )\right ) \right |\cr
& \le \sum_{(Q,Q')\text{ sing.}} \sqrt{{\rm Var}\left (\text{proj}\left (I_{n_{ |_Q}}|C_{\ge 6}\right )\right ) {\rm Var}\left (\text{proj}\left (I_{n_{ |_{Q'}}}|C_{\ge 6}\right )\right )}\cr &\ll E_n^2 R_n(6) {\rm Var}\left (\text{proj}\left (I_{n_{ |_{Q_0}}}|C_{\ge 6}\right )\right ),
\end{split}
\end{equation*}
where, from now on, $Q_0$ denotes the square containing the origin.
Now,
\begin{equation*}
\begin{split}
{\rm Var}\left (\text{proj}\left (I_{n_{ |_{Q_0}}}|C_{\ge 6}\right )\right )\le {\mathbb{ E}}\left [I_{n_{ |_{Q_0}}}^2\right ]= \underbrace{{\mathbb{ E}}\left [I_{n_{ |_{Q_0}}}^2\right ] - {\mathbb{ E}}\left [I_{n_{ |_{Q_0}}}\right ]}_{=:A}+{\mathbb{ E}}\left [I_{n_{ |_{Q_0}}}\right ].
\end{split}
\end{equation*}
It is immediate to check that
\begin{equation*}
{\mathbb{ E}}\left [ I_{n_{ |_{Q_0}}} \right ] = \frac{2\pi n}{M^2},
\end{equation*}
in particular ${\mathbb{ E}}\left [ I_{n_{ |_{Q_0}}} \right ] =O(1)$.
Note that $A$ is the $2$-th factorial moment of $I_{n_{ |_{Q_0}}}$:
\begin{equation*}
A = {\mathbb{ E}}\left[I_{n_{ |_{Q_0}}}\left(I_{n_{ |_{Q_0}}}-1 \right) \right].
\end{equation*}
Applying \cite[Theorem 6.3]{AW}, we can write
\begin{equation}\label{media}
\begin{split}
A={\mathbb{ E}}\left [ I_{n_{ |_{Q_0}}} \left( I_{n_{ |_{Q_0}}} -1 \right) \right ] = \int_{Q_0}\int_{Q_0} K_2(x,y)\,dx dy,
\end{split}
\end{equation}
where
\begin{equation*}\label{corr1}
K_2(x,y) := p_{({\bf T}_n(x),{\bf T}_n(y))}(0,0)\, {\mathbb{ E}}\left [\left | J_{{\bf T}_n}(x)\right |\cdot \left | J_{{\bf T}_n}(y)\right |\Big | {\bf T}_n(x) = {\bf T}_n(y) = 0 \right ]
\end{equation*}
is the so-called $2$-point correlation function. {Indeed, Proposition \ref{gio} ensures that for $(x,y)\in Q_0\times Q_0$, the vector $({\bf T}_n(x), {\bf T}_n(y))$ is non-degenerate {except} on the diagonal $x=y$}.
Note that, by stationarity of the model, we can write \paref{media} as
\begin{equation*}
\begin{split}
{\mathbb{ E}}\left [ I_{n_{ |_{Q_0}}} \left( I_{n_{ |_{Q_0}}} -1 \right) \right ] = \text{Leb}(Q_0)\int_{\widetilde Q_0} K_2(x)\,dx,
\end{split}
\end{equation*}
where $K_2(x):=K_2(x,0)$ and $\widetilde Q_0$ is $2 Q_0$.
Let us first check that the function $x\mapsto K_2(x)$ is integrable around the origin.
Note that, by Cauchy-Schwarz inequality,
\begin{equation*}
\begin{split}
K_2(x) &= \frac{1}{1-r^2(x)}\, {\mathbb{ E}}\left [\left | J_{{\bf T}_n}(x)\right |\cdot \left | J_{{\bf T}_n}(0)\right |\Big | {\bf T}_n(x) = {\bf T}_n(0) = 0 \right ]\cr
&\le \frac{1}{1-r^2(x)}\, {\mathbb{ E}}\left [\left | J_{{\bf T}_n}(0)\right |^2\Big | {\bf T}_n(x) = {\bf T}_n(0) = 0 \right ].
\end{split}
\end{equation*}
Hypercontractivity on Wiener chaoses \cite{NP} ensures that there exists $c>0$ such that
\begin{equation*}
K_2(x) \le c \frac{1}{1-r^2(x)}\, \left( {\mathbb{ E}}\left [\left | J_{{\bf T}_n}(0)\right |\Big | {\bf T}_n(x) = {\bf T}_n(0) = 0 \right ]\right)^2.
\end{equation*}
Now, thanks to \cite[(1.3)]{KZ}
\begin{equation*}
K_2(x) \le c \frac{1}{1-r^2(x)}\, \left( \sqrt{|\Omega_n(x)|}\right)^2
= c \frac{|\Omega_n(x)|}{1-r^2(x)},
\end{equation*}
where $\Omega_n(x)$ denotes the covariance matrix of $\nabla T_n(0)$ conditioned to ${T}_n(x) = { T}_n(0) = 0$ (see \paref{matrix omega} for a precise expression).
The Taylor expansion in Lemma \ref{lemma taylor} gives that, as $\|x\|\to 0$,
\begin{equation*}
\frac{|\Omega_n(x)|}{1-r^2(x)} = c E_n^2 + E_n^3O(\|x\|^2),
\end{equation*}
for some other constant $c>0$, where the constants involving in the `O' notation do not depend on $n$, so that
\begin{equation*}
\begin{split}
{\mathbb{ E}}\left [ I_{n_{ |_{Q_0}}} \left( I_{n_{ |_{Q_0}}} -1 \right) \right ] = \text{Leb}(Q_0)\int_{\widetilde Q_0} K_2(x)\,dx\ll \frac{E_n^2}{M^4},
\end{split}
\end{equation*}
which is the result we looked for.
\end{proof}
\subsubsection{The contribution of non-singular pairs of cubes}
\begin{proof}[Proof of Lemma \ref{varianza nonsing}]
For any square $Q$, we can write
\begin{eqnarray*}
&& \text{proj}\left(I_{n_{|_{Q}}}|C_{\ge 6}\right) \\
&& =\frac{E_n}{2} \sum_{q\ge 3}\sum_{i_1+i_2+i_3 +j_1+j_2+j_3=2q} \frac{\beta_{i_1}\beta_{j_1}\alpha_{i_2,i_3,j_2,j_3}}{i_1! i_2! i_3! j_1! j_2! j_3!}\times\\
&& \quad\times \int_{Q} H_{i_1}(T_n(x)) H_{i_2}(\partial_1 \widetilde T_n(x)) H_{i_3}(\partial_2 \widetilde T_n(x))H_{j_1}(\widehat T_n(x)) H_{j_2}(\partial_1 \widetilde {\widehat T}_n(x)) H_{j_3}(\partial_2 \widetilde {\widehat T}_n(x))\,dx,
\end{eqnarray*}
for even $i_1,j_1$ and $i_2,i_3,j_2,j_3$ with the same parity.
Recall that $\beta_l=0$ for odd $l$, and that $\beta_{2l}^2/{(2l)!}\approx 1/\sqrt{l}$, as $l\to \infty$. {\color{black}
We have
\begin{equation}\label{lungo1}
\begin{split}
& \left| \sum_{(Q,Q')\text{ non sing.}} {\rm Cov\,}\left (\text{proj}\left(I_{n_{|_{Q}}}|C_{\ge 6}\right),\text{proj}\left(I_{n_{|_{Q'}}}|C_{\ge 6}\right) \right) \right|\cr
&\le E_n^2 \sum_{q\ge 3} \sum_{i_1+i_2+i_3+j_1+j_2+j_3=2q}\sum_{a_1+a_2+a_3+b_1+b_2+b_3=2q} \left |\frac{\beta_{i_1}\beta_{j_1}\alpha_{i_2,i_3,j_2,j_3}}{i_1! i_2! i_3! j_1! j_2! j_3!}\right |\cdot \left |\frac{\beta_{a_1}\beta_{b_1}\alpha_{a_2,a_3,b_2,b_3}}{a_1! a_2! a_3! b_1! b_2! b_3!}\right |\cr
&\times\Big|\sum_{(Q,Q')\text{ non sing.}}\cr
& \int_Q \int_{Q'} {\mathbb{ E}}\Big [ H_{i_1}(T_n(x)) H_{i_2}(\widetilde{\partial_1} T_n(x)) H_{i_3}(\widetilde{\partial_2 }T_n(x))H_{j_1}(\widehat T_n(x)) H_{j_2}(\widetilde{\partial_1} {\widehat T}_n(x)) H_{j_3}(\widetilde{\partial_2} {\widehat T}_n(x))\cr
&\times H_{a_1}(T_n(y)) H_{a_2}(\widetilde{\partial_1} T_n(y)) H_{a_3}(\widetilde{\partial_2} T_n(y))H_{b_1}(\widehat T_n(y)) H_{b_2}(\widetilde{\partial_1} {\widehat T}_n(y)) H_{b_3}(\widetilde{\partial_2} {\widehat T}_n(y))
\Big ]\,dxdy\Big|.
\end{split}
\end{equation}
Let us now adopt the same notation as in Proposition \ref{p:ls}.
For $n\in S$ we set
\begin{eqnarray*}
&& (X_0(x), X_1(x), X_2(x), Y_0(x), Y_1(x), Y_2(x))\\
&&\quad\quad\quad := (T_{n}(x),\, \widetilde\partial_{1} T_{n}(x), \, \widetilde\partial_{2} T_{n}(x), \widehat{T}_n(x),\, \widetilde \partial_{1}\widehat{T}_n(x), \, \widetilde \partial_{2}\widehat{T}_n(x)), \quad x\in {\mathbb{T}}.
\end{eqnarray*}
From Proposition \ref{p:ls} and \paref{lungo1}, we have
\begin{eqnarray}\label{lungo2}
&& \left| \sum_{(Q,Q')\text{ non sing.}} {\rm Cov\,}\left (\text{proj}\left(I_{n_{|_{Q}}}|C_{\ge 6}\right),\text{proj}\left(I_{n_{|_{Q'}}}|C_{\ge 6}\right) \right) \right|\\
&&\le E_n^2 \sum_{q\ge 3} \sum_{i_1+i_2+i_3+j_1+j_2+j_3=2q}\sum_{a_1+a_2+a_3+b_1+b_2+b_3=2q} \left |\frac{\beta_{i_1}\beta_{j_1}\alpha_{i_2,i_3,j_2,j_3}}{i_1! i_2! i_3! j_1! j_2! j_3!}\right |\cdot \left |\frac{\beta_{a_1}\beta_{b_1}\alpha_{a_2,a_3,b_2,b_3}}{a_1! a_2! a_3! b_1! b_2! b_3!}\right | \notag \\
&&\quad\times {\bf 1}_{\{i_1+i_2+i_3 = a_1+a_2+a_3\}}{\bf 1}_{\{j_1+j_2+j_3 = b_1+b_2+b_3\}} \Big| V(i_1,i_2,i_3 ; j_1,j_2,j_3 ; a_1,a_2,a_3 ; b_1,b_2,b_3) \Big| ,\notag \\
&& :=E_n^2 \times Z,\label{e:rol}
\end{eqnarray}
where each of the terms $V = V(i_1,i_2,i_3 ; j_1,j_2,j_3 ; a_1,a_2,a_3 ; b_1,b_2,b_3)$ is the sum of no more than $(2q)!$ terms of the type
\begin{equation}\label{e:sergio}
v =\sum_{(Q,Q')\text{ non sing.}} \int_Q\int_{Q'} \prod_{u=1}^{2q} R_{l_u, k_u}({\color{black} x-y}) \, dxdy,
\end{equation}
where $k_u, l_u \in \{0,1,2\}$ and {\color{black} where}, for $l,k=0,1,2$ and $x,y\in \mathbb T$, we set
$$
{\color{black} R_{l,k}(x-y)}:= {\mathbb{ E}}\left[X_l(x) X_k(y) \right]= {\mathbb{ E}}\left[Y_l(x) Y_k(y) \right].
$$
Note that, for any {even $p\in \field{N}$}, we have
\begin{equation}\label{deriv}
\int_{\mathbb T} R_{l,k}(x)^{p}\,dx \leq \int_{\mathbb T} {\color{black} r_n}(x)^p\,dx{\color{black} =:R_n(p)}
\end{equation}
and recall moreover that, for $x,y\in \field{T}$, $|R_{l,k}(x-y)|\leq 1$, and, for $(x,y)\in Q\times Q'$,
\begin{equation}\label{epsuno}
|R_{l,k}(x-y)|<\varepsilon_1.
\end{equation}
Using the definition of a non-singular pair of cubes, as well as the fact that the sum defining $Z$ in \eqref{e:rol} involves indices $q\geq 3$, one deduces that, for $v$ as in \eqref{e:sergio},
\begin{eqnarray*}
|v| &\leq& \varepsilon_1^{2q-6} \sum_{(Q,Q')\text{ non sing.}} \int_Q\int_{Q'} \prod_{u=1}^{6} \left |R_{l_u, k_u}({\color{black} x-y})\right| dxdy\\
&\leq & \varepsilon_1^{2q-6} \int_\field{T} \prod_{u=1}^{6} \left |R_{l_u, k_u}(x)\right| dx \leq \varepsilon_1^{2q-6} R_n(6),
\end{eqnarray*}
where we have applied a generalized H\"older inequality together with \eqref{deriv} in order to deduce the last estimate. This bound implies that each of the terms $V$ contributing to $Z$ can be bounded as follows:
\begin{eqnarray*}
&&\Big| V(i_1,i_2,i_3 ; j_1,j_2,j_3 ; a_1,a_2,a_3 ; b_1,b_2,b_3) \Big|\\
&&\quad\quad \leq (2q)! \frac{R_n(6)}{\varepsilon_1^6} \varepsilon_1^{2q} = (2q)! \frac{R_n(6)}{\varepsilon_1^6} ( \sqrt{\varepsilon_1} )^{i_1+\cdots+ j_3}( \sqrt{\varepsilon_1} )^{a_1+\cdots+ b_3}.
\end{eqnarray*}
One therefore infers that
\begin{eqnarray*}
&& Z\leq \frac{R_n(6)}{\varepsilon_1^6} \sum_{q\geq 3} (2q)! \sum_{i_1+i_2+i_3+j_1+j_2+j_3=2q}\sum_{a_1+a_2+a_3+b_1+b_2+b_3=2q} \left |\frac{\beta_{i_1}\beta_{j_1}\alpha_{i_2,i_3,j_2,j_3}}{i_1! i_2! i_3! j_1! j_2! j_3!}\right | \times \notag \\
&&\quad\quad \quad\quad\quad \quad \left |\frac{\beta_{a_1}\beta_{b_1}\alpha_{a_2,a_3,b_2,b_3}}{a_1! a_2! a_3! b_1! b_2! b_3!}\right |\times ( \sqrt{\varepsilon_1} )^{i_1+\cdots+ j_3}( \sqrt{\varepsilon_1} )^{a_1+\cdots+ b_3} =:\frac{R_n(6)}{\varepsilon_1^6}\times S .\notag
\end{eqnarray*}
In order to show that $S$ is finite, we write
\begin{eqnarray*}
S& = & \sum_{q \geq 3} (2q)! \sum_{i_1+i_2+i_3+j_1+j_2+j_3=2q}\sum_{a_1+a_2+a_3+b_1+b_2+b_3=2q} \left |\frac{\beta_{i_1}\beta_{j_1}\alpha_{i_2,i_3,j_2,j_3}}{i_1! i_2! i_3! j_1! j_2! j_3!}\right | \times \notag \\
&&\quad\quad \quad\quad\quad \quad \left |\frac{\beta_{a_1}\beta_{b_1}\alpha_{a_2,a_3,b_2,b_3}}{a_1! a_2! a_3! b_1! b_2! b_3!}\right |\times ( \sqrt{\epsilon_1} )^{i_1+\cdots+ j_3}( \sqrt{\epsilon_1} )^{a_1+\cdots+ b_3}\\
&\leq & \sum_{q\geq 0} \sum_{i_1+i_2+i_3+j_1+j_2+j_3=2q}\sum_{a_1+a_2+a_3+b_1+b_2+b_3=2q} \left |\frac{\beta_{i_1}\beta_{j_1}\alpha_{i_2,i_3,j_2,j_3}}{i_1! i_2! i_3! j_1! j_2! j_3!}\right | \times \notag \\
&&\sqrt{(i_1+\cdots+j_3)!} \sqrt{(a_1+\cdots+b_3)!} \left| \frac{\beta_{a_1}\beta_{b_1}\alpha_{a_2,a_3,b_2,b_3}}{a_1! a_2! a_3! b_1! b_2! b_3!}\right |\times ( \sqrt{\epsilon_1} )^{i_1+\cdots+ j_3+a_1+\cdots+ b_3}\\
&\leq & \sum_{i_1,\ldots ,j_3, a_1\ldots, b_3} \left |\frac{\beta_{i_1}\beta_{j_1}\alpha_{i_2,i_3,j_2,j_3}}{i_1! i_2! i_3! j_1! j_2! j_3!}\right | \times \notag \\
&&\sqrt{(i_1+\cdots+j_3)!} \sqrt{(a_1+\cdots+b_3)!} \left| \frac{\beta_{a_1}\beta_{b_1}\alpha_{a_2,a_3,b_2,b_3}}{a_1! a_2! a_3! b_1! b_2! b_3!}\right |\times ( \sqrt{\epsilon_1} )^{i_1+\cdots+ j_3+a_1+\cdots+ b_3} \\
& \leq & \Bigg ( \sum_{i_1,\ldots ,j_3, a_1\ldots, b_3} \left |\frac{\beta_{i_1}\beta_{j_1}\alpha_{i_2,i_3,j_2,j_3}}{i_1! i_2! i_3! j_1! j_2! j_3!}\right |^2 (i_1+\cdots+j_3)! ( \sqrt{\epsilon_1} )^{i_1+\cdots+ j_3+a_1+\cdots+ b_3}\Bigg)^{1/2}\times \\
&& \times \Bigg( \sum_{i_1,\ldots ,j_3, a_1\ldots, b_3}\left| \frac{\beta_{a_1}\beta_{b_1}\alpha_{a_2,a_3,b_2,b_3}}{a_1! a_2! a_3! b_1! b_2! b_3!}\right |^2 (a_1+\cdots+b_3)!( \sqrt{\epsilon_1} )^{i_1+\cdots+ j_3+a_1+\cdots+ b_3}\Bigg)^{1/2}
\\ & =& \sum_{i_1,\ldots ,j_3, a_1\ldots, b_3} \left |\frac{\beta_{i_1}\beta_{j_1}\alpha_{i_2,i_3,j_2,j_3}}{i_1! i_2! i_3! j_1! j_2! j_3!}\right |^2 (i_1+\cdots+j_3)! ( \sqrt{\epsilon_1} )^{i_1+\cdots+ j_3+a_1+\cdots+ b_3} < \infty,
\end{eqnarray*}
where: (a) the third inequality follows by applying the Cauchy-Schwarz inequality to the symmetric finite measure $\mu$ on $\mathbb{N}^{12}$ such that
$$
\mu\{ (k_1,...,k_{12} ) \} = (\sqrt{\epsilon_1} )^{k_1+\cdots+ k_{12}},
$$
and, (b) writing $m = m(i_1,...,j_3) := i_1+\cdots +j_3$ for every $i_1,...,j_3$, the finiteness of the last sum is a consequence of the standard estimate
$$
\frac{ (i_1+\cdots +j_3)!}{i_1! i_2! i_3! j_1! j_2! j_3!} \leq \sum_{\substack{k_1,...,k_6\geq 0 \\ k_1+\cdots+k_6 = m} } \frac{ m!}{k_1! \cdots k_6!} = 6^m = 6^{i_1+\cdots +j_3},
$$
as well as of the fact that the mapping
$$
(i_1,...,j_3)\mapsto \frac{\beta^2_{i_1}\beta^2_{j_1}\alpha^2_{i_2,i_3,j_2,j_3}}{i_1! i_2! i_3! j_1! j_2! j_3!}
$$
is bounded, and $6\sqrt{\varepsilon_1}<1$ by assumption. This concludes the proof.
}
\end{proof}
\section{End of the Proof of Theorem \ref{t:main}}\label{proofs main sec}
\subsection{Proof of Part 2}
From Lemma \ref{berry's cancellation}, for $n\in S$ the chaotic expansion for $I_n$ is
$$
I_n = {\mathbb{ E}}[I_n] + \sum_{q\ge 2} I_n[2q],
$$
where $I_n[2q]$ is given in \paref{proiezione q}. Proposition \ref{varianze chaos}, Proposition \ref{asymptotic} together with Lemma \ref{lemma BB} immediately conclude the proof, once we recall that, by orthogonality of different Wiener chaoses
$$
{\rm Var}(I_{n}) = {\rm Var}(I_{n}[4]) + \sum_{q\ge 3} {\rm Var}(I_{n}[2q]).
$$
\subsection{Proof of Part 4}
Part 2 of Theorem \ref{t:main} yields that, as $\mathcal{N}_{n_j}\to +\infty$,
$$
\frac{I_{n_j}-{\mathbb{ E}}[I_{n_j}]}{\sqrt{{\rm Var}(I_{n_j})}} = \frac{I_{n_j}[4]}{\sqrt{{\rm Var}(I_{n_j}[4])}} + o_{\mathbb{P}}(1).
$$
Proposition \ref{limite 4} hence allows to conclude the proof.
\section{Some technical computations}\label{appendix}
\subsection{Technical proofs}
\begin{proof}[Proof of Lemma \ref{alpha piccoli}]
We have
\begin{eqnarray*}
\alpha_{0,0,0,0}&=&\mathbb{E} [|XW-YV|] \\ &=& \frac{1}{(2\pi)^2} \left( \int_0^\infty \rho^2 e^{-\rho^2/2} d\rho\right)^2 \int_0^{2\pi}\int_0^{2\pi} | \sin\theta \cos \theta' - \sin \theta'\cos\theta| d\theta d\theta'\\
&=& \frac{1}{(2\pi)^2} \left( \int_0^\infty \rho^2 e^{-\rho^2/2} d\rho\right)^2 \int_0^{2\pi}\int_0^{2\pi} | \sin( \theta - \theta') | d\theta d\theta' = 1.
\end{eqnarray*}
Setting $Z$ to be any of the variables $X,Y,V,W$ and $\varphi_Z(u)$ to be $\cos(u)$ if $Z=X,V$, or $\sin(u)$ if $Z = Y,W$, we have that
\begin{eqnarray*}
&&\mathbb{E} [|XW-YV| H_2(Z) ]\\
&=& \frac{1}{(2\pi)^2} \int_0^\infty \rho^2 e^{-\rho^2/2} d\rho \int_0^\infty \gamma^4 e^{-\gamma^2/2} d\gamma \int_0^{2\pi}\int_0^{2\pi} | \sin( \theta - \theta')| \varphi_Z(\theta)^2 d\theta d\theta' -1= \frac12.
\end{eqnarray*}
As a result, we deduce that
$$
\alpha_{2,0,0,0}= \alpha_{0,2,0,0}=\alpha_{0,0,2,0}=\alpha_{0,0,0,2}=\frac12.
$$
Let us now concentrate on $\alpha_{4,0,0,0} $. We have
\begin{eqnarray*}
\alpha_{4,0,0,0} &=&{\mathbb{ E}}[|XW-YV|H_4(X)]\\
&=&
{\mathbb{ E}}[|XW-YV|X^4] - 6 \underbrace{{\mathbb{ E}}[|XW-YV|X^2]}_{=\frac32\text{ from above}}+3\underbrace{{\mathbb{ E}}[|XW-YV|]}_{=1}.
\end{eqnarray*}
Thus, it remains to calculate
\begin{eqnarray*}
&&{\mathbb{ E}}[|XW-YV|X^4]\\
&=&\frac{1}{(2\pi)^2}\int_{\mathbb R^4}|xw-yv| x^4 {\rm e}^{-x^2/2}{\rm e}^{-y^2/2}{\rm e}^{-v^2/2}{\rm e}^{-w^2/2}\,dxdydvdw\\
&=&\frac{1}{2\pi}\underbrace{ \int_0^{2\pi}\cos^4\theta d\theta}_{=\frac{3\pi}{4}}\underbrace{\int_0^{2\pi} | \sin( \theta' - \theta)| d\theta'}_{=4}\underbrace{\frac{1}{\sqrt{2\pi}} \int_0^\infty(\rho')^2 {\rm e}^{-(\rho')^2/2}d\rho'}_{=\frac12}\underbrace{\frac{1}{\sqrt{2\pi}}\int_0^\infty \rho^6 {\rm e}^{-\rho^2/2}\, d\rho}_{=\frac{15}{2}}\\
&=&\frac{45}{8}.
\end{eqnarray*}
Plugging into the previous expression, we deduce
$$
\alpha_{4,0,0,0} =-\frac{3}{8}.
$$
Since
$\int_0^{2\pi}\cos^4\theta d\theta = \int_0^{2\pi}\sin^4\theta d\theta$,
it is immediate to check that
$$
\alpha_{4,0,0,0}=\alpha_{0,4,0,0}=\alpha_{0,0,4,0}=\alpha_{0,0,0,4}.
$$
Let us now compute $\alpha_{2,2,0,0} $. We have
\begin{eqnarray*}
&&\alpha_{2,2,0,0} = {\mathbb{ E}}[|XW-YV|H_2(X)H_2(Y)]\\
&=&{\mathbb{ E}}[|XW-YV|X^2 Y^2]
- \underbrace{{\mathbb{ E}}[|XW-YV|X^2]}_{=\frac32}
-\underbrace{{\mathbb{ E}}[|XW-YV|Y^2] }_{=\frac32}+\underbrace{{\mathbb{ E}}[
|XW-YV|]}_{=1},
\end{eqnarray*}
whereas
\begin{eqnarray*}
&&{\mathbb{ E}}[|XW-YV|X^2 Y^2] \\
&=&\frac{1}{(2\pi)^2}\int_{\mathbb R^4}|xw-yv| x^2y^2 {\rm e}^{-x^2/2}{\rm e}^{-y^2/2}{\rm e}^{-v^2/2}{\rm e}^{-w^2/2}\,dxdydvdw\\
&=&\frac{1}{2\pi}\underbrace{ \int_0^{2\pi} \!\!\!\cos^2\theta \, \sin^2\theta d\theta}_{=\frac{\pi}{4}}\underbrace{\int_0^{2\pi} \!\!\!| \sin( \theta' - \theta)| d\theta'}_{=4}\underbrace{\frac{1}{\sqrt{2\pi}} \int_0^\infty \!\!\!(\rho')^2 {\rm e}^{-(\rho')^2/2}d\rho'}_{=\frac12}\underbrace{\frac{1}{\sqrt{2\pi}}\int_0^\infty \!\!\!\rho^6 {\rm e}^{-\rho^2/2}\, d\rho}_{=\frac{15}{2}}\\
&=&\frac{15}{8}.
\end{eqnarray*}
Therefore
$$
\alpha_{2,2,0,0}=-\frac{1}{8}.
$$
Similarly,
$$
\alpha_{0,0,2,2}=\alpha_{2,2,0,0}=-\frac{1}{8}.
$$
Now, let us compute $\alpha_{2,0,2,0} $. We can write
\begin{eqnarray*}
\alpha_{2,0,2,0} &=& {\mathbb{ E}}[|XW-YV|H_2(X)H_2(V)]\\
&=&{\mathbb{ E}}[|XW-YV|X^2 V^2] - \underbrace{{\mathbb{ E}}[|XW-YV|X^2]}_{=\frac32}-\underbrace{{\mathbb{ E}}[|XW-YV|V^2] }_{=\frac32}+\underbrace{{\mathbb{ E}}[|XW-YV|]}_{=1},
\end{eqnarray*}
whereas
\begin{eqnarray*}
&&{\mathbb{ E}}[|XW-YV|X^2V^2] \\
&=&\frac{1}{(2\pi)^2}\int_{\mathbb R^4}|xw-yz| x^2v^2 {\rm e}^{-x^2/2}{\rm e}^{-y^2/2}{\rm e}^{-v^2/2}{\rm e}^{-w^2/2}\,dxdydvdw\\
&=&\frac{1}{2\pi}\underbrace{\int_0^{2\pi} \!\!\!\cos^2\theta \, d\theta\int_0^{2\pi} \!\!\!| \sin( \theta' - \theta)|\cos^2\theta' d\theta'}_{=\frac{5\pi}{3}}\underbrace{\frac{1}{\sqrt{2\pi}} \int_0^\infty\!\!\!(\rho')^4 {\rm e}^{-(\rho')^2/2}d\rho'}_{=\frac32}\underbrace{\frac{1}{\sqrt{2\pi}}\int_0^\infty \!\!\!\rho^4 {\rm e}^{-\rho^2/2}\, d\rho}_{=\frac{3}{2}}\\
&=&\frac{15}{8}.
\end{eqnarray*}
Then,
$$
\alpha_{2,0,2,0}=\alpha_{0,2,0,2}-\frac{1}{8}.
$$
We also compute
\begin{eqnarray*}
\alpha_{2,0,0,2} &=& {\mathbb{ E}}[|XW-YV|H_2(X)H_2(W)]\\
&=&{\mathbb{ E}}[|XW-YV|X^2W^2]
- \underbrace{{\mathbb{ E}}[|XW-YV|X^2 ]}_{=\frac32}
-\underbrace{{\mathbb{ E}}[|XW-YV|W^2] }_{=\frac32}
+\underbrace{{\mathbb{ E}}[|XW-YV|]}_{=1}.
\end{eqnarray*}
We have
\begin{eqnarray*}
&&{\mathbb{ E}}[|XW-YV|X^2W^2] \\
&=&\frac{1}{(2\pi)^2}\int_{\mathbb R^4}|xw-yv| x^2w^2 {\rm e}^{-x^2/2}{\rm e}^{-y^2/2}{\rm e}^{-v^2/2}{\rm e}^{-w^2/2}\,dxdydvdw\\
&=&\frac{1}{2\pi}\underbrace{\int_0^{2\pi} \!\!\!\cos^2\theta \, d\theta\int_0^{2\pi} | \sin( \theta' - \theta)|\sin^2\theta' d\theta'}_{=\frac{7\pi}{3}}\underbrace{\frac{1}{\sqrt{2\pi}} \int_0^\infty\!\!\!(\rho')^4 {\rm e}^{-(\rho')^2/2}d\rho'}_{=\frac32}\underbrace{\frac{1}{\sqrt{2\pi}}\int_0^\infty \!\!\!\rho^4 {\rm e}^{-\rho^2/2}\, d\rho}_{=\frac{3}{2}}\\
&=&\frac{21}{8},
\end{eqnarray*}
so that
$$
\alpha_{2,0,0,2}=\alpha_{0,2,2,0}=\frac{5}{8}.
$$
Finally, let us consider the case where $a=b=c=d=1$.
We have
\begin{eqnarray*}
&&\alpha_{1,1,1,1}={\mathbb{ E}}[|XW-YV|XYVW] \\
&=&\frac{1}{(2\pi)^2}\int_{\mathbb R^4}|xw-yv| xyvw {\rm e}^{-x^2/2}{\rm e}^{-y^2/2}{\rm e}^{-v^2/2}{\rm e}^{-w^2/2}\,dxdydvdw\\
&=&\frac{1}{2\pi}\underbrace{\int_{[0,2\pi]^2} | \sin( \theta' - \theta)|\cos\theta\cos\theta'\sin\theta\sin\theta' d\theta d\theta'}_{=-\frac{\pi}{3}}\underbrace{\frac{1}{\sqrt{2\pi}} \int_0^\infty(\rho')^4 {\rm e}^{-(\rho')^2/2}d\rho'}_{=\frac32}\\
&&\hskip8.5cm\times\underbrace{\frac{1}{\sqrt{2\pi}}\int_0^\infty \rho^4 {\rm e}^{-\rho^2/2}\, d\rho}_{=\frac{3}{2}}\\
&=&-\frac{3}{8}.
\end{eqnarray*}
\end{proof}
\subsection{Proof of Lemma \ref{lemma infinito}}
\noindent{\it Proof of (i)}. We have
\begin{eqnarray*}
&&\int_{\field{T}} H_4(T_n(x))\,dx = \int_{\field{T}} (T_n(x)^4 - 6T_n(x)^2+3)\,dx\\
&=&\frac{1}{\mathcal N_n^2}\!\sum_{\lambda, \lambda',\lambda'',\lambda'''\in\Lambda_n} \!\!\!\!\!\!a_\lambda \overline{a_{\lambda'}}\, a_{\lambda''} \overline{{a}_{\lambda'''}}\!\int \!{\rm e}_{\lambda - \lambda'+\lambda'' -\lambda'''}(x)\,dx -\frac{6}{\mathcal N_n}\sum_{\lambda, \lambda'\in\Lambda_n} \!\!\!a_\lambda \overline{a_{\lambda'}}\int {\rm e}_{\lambda - \lambda'}(x)\,dx +3\\
&=&\frac{1}{\mathcal N_n^2}\sum_{\lambda,\lambda''} |a_\lambda|^2
|a_{\lambda''}|^2
+\frac{1}{\mathcal N_n^2}\sum_{\lambda} |a_\lambda|^4
+\frac{2}{\mathcal N_n^2}\sum_{\lambda\neq \pm\lambda'} |a_\lambda|^2 |a_{\lambda'}|^2
-\frac{6}{\mathcal N_n}\sum_{\lambda} |a_\lambda|^2 +3\\
&=&\frac{3}{\mathcal N_n^2}\sum_{\lambda,\lambda''} (|a_\lambda|^2-1)
(|a_{\lambda''}|^2-1)
-\frac{3}{\mathcal N_n^2}\sum_{\lambda} |a_\lambda|^4
=\frac{3}{\mathcal N_n} W{}{(n)}^2 -\frac{3}{\mathcal N_n^2}\sum_{\lambda} |a_\lambda|^4.
\end{eqnarray*}
Since $\frac1{\mathcal N_n}\sum_{\lambda} |a_\lambda|^4\to 2$ by the law of large numbers, the claim $(i)$ follows.
\bigskip
\newpage
\noindent{\it Proof of (ii)}. We have
\begin{eqnarray*}
&&\int_{\field{T}} H_4(\widetilde{\partial}_j T_n(x))\,dx = \int_{\field{T}} (\widetilde{\partial}_j T_n(x)^4 - 6\,\widetilde{\partial}_j T_n(x)^2+3)\,dx\\
&=&\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda, \lambda',\lambda'',\lambda'''\in\Lambda_n} \lambda_j \lambda'_j\lambda''_j\lambda'''_j\,a_\lambda \overline{a_{\lambda'}}\, a_{\lambda''} \overline{{a}_{\lambda'''}}\int {\rm e}_{\lambda - \lambda'+\lambda'' -\lambda'''}(x)\,dx \\
&&-\frac{12}{n\mathcal N_n}\sum_{\lambda, \lambda'\in\Lambda_n} \lambda_j \lambda'_j\,a_\lambda \overline{a_{\lambda'}}\int {\rm e}_{\lambda - \lambda'}(x)\,dx +3\\
&=&\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda,\lambda''} \lambda_j^2{\lambda''_j}^2|a_\lambda|^2
|a_{\lambda''}|^2
+\frac{4}{n^2\mathcal\mathcal N_n^2}\sum_{\lambda} \lambda_j^4 |a_\lambda|^4
+\frac{8}{n^2\mathcal\mathcal N_n^2}\sum_{\lambda\neq \pm\lambda'} \lambda_j^2{\lambda'_j}^2|a_\lambda|^2 |a_{\lambda'}|^2
\\
&& -\frac{12}{n\mathcal N_n}\sum_{\lambda}\lambda_j^2 |a_\lambda|^2 +3
\end{eqnarray*}
\begin{eqnarray*}
&=&\frac{12}{n^2\mathcal N_n^2}\sum_{\lambda,\lambda''} \lambda_j^2{\lambda''_j}^2|a_\lambda|^2
|a_{\lambda''}|^2
-\frac{12}{n^2\mathcal\mathcal N_n^2}\sum_{\lambda} \lambda_j^4 |a_\lambda|^4
-\frac{12}{n\mathcal N_n}\sum_{\lambda} \lambda_j^2|a_\lambda|^2 +3\\
&=&\frac{12}{n^2\mathcal N_n^2}\sum_{\lambda,\lambda''} \lambda_j^2{\lambda''_j}^2(|a_\lambda|^2-1)
(|a_{\lambda''}|^2-1)
-\frac{12}{n^2\mathcal\mathcal N_n^2}\sum_{\lambda} \lambda_j^4 |a_\lambda|^4\\
&=&\frac{12}{\mathcal N_n} W_j{}{(n)}^2 -\frac{12}{n^2\mathcal\mathcal N_n^2}\sum_{\lambda} \lambda_j^4 |a_\lambda|^4 .
\end{eqnarray*}
Since $\frac1{n^2\mathcal N_n}\sum_{\lambda} \lambda_j^4 |a_\lambda|^4\to \frac14(3+\widehat{\mu}_\infty(4))$ by the law of large numbers, the claim $(ii)$ follows.
\bigskip
\noindent{\it Proof of (iii)}. We have
\begin{eqnarray*}
&&\int_{\field{T}} H_2(T_n(x))\big(H_2(\widetilde{\partial}_1 T_n(x)+H_2(\widetilde{\partial}_2 T_n(x))\big)\,dx
\\
&=&\int_{\field{T}} \big(T_n(x)^2\widetilde{\partial}_1 T_n(x)^2+T_n(x)^2\widetilde{\partial}_2 T_n(x)^2
-2T_n(x)^2-\widetilde{\partial}_1 T_n(x)^2-\widetilde{\partial}_2 T_n(x)^2+2
\big)\,dx
\\
&=&\frac{2}{n\mathcal N_n^2}\sum_{\lambda, \lambda',\lambda'',\lambda'''\in\Lambda_n} (\lambda''_1\lambda'''_1+\lambda''_2\lambda'''_2)\,a_\lambda \overline{a_{\lambda'}}\, a_{\lambda''} \overline{{a}_{\lambda'''}}\int {\rm e}_{\lambda - \lambda'+\lambda'' -\lambda'''}(x)\,dx \\
&&-\frac{2}{\mathcal N_n}\sum_{\lambda, \lambda'\in\Lambda_n} a_\lambda \overline{a_{\lambda'}}\int {\rm e}_{\lambda - \lambda'}(x)\,dx
-\frac{2}{n\mathcal N_n}\sum_{\lambda, \lambda'\in\Lambda_n} (\lambda_1\lambda'_1+\lambda_2\lambda'_2) a_\lambda \overline{a_{\lambda'}}\int {\rm e}_{\lambda - \lambda'}(x)\,dx
+2\\
&=&\frac{2}{\mathcal N_n^2}\sum_{\lambda, \lambda''} |a_\lambda|^2 | a_{\lambda''} |^2
- \frac{2}{\mathcal N_n^2}\sum_{\lambda} |a_\lambda|^4
+\frac{4}{n\mathcal N_n^2}\sum_{\lambda\neq \pm\lambda'} (\lambda_1\lambda'_1+\lambda_2\lambda'_2)|a_\lambda|^2 | a_{\lambda'} |^2\\
&&-\frac{4}{\mathcal N_n}\sum_{\lambda} |a_\lambda|^2 +2\\
&=&\frac{2}{\mathcal N_n^2}\sum_{\lambda, \lambda''} (|a_\lambda|^2 -1)(| a_{\lambda''} |^2 -1)
- \frac{2}{\mathcal N_n^2}\sum_{\lambda} |a_\lambda|^4
=\frac{2}{\mathcal N_n}\left\{W{}{(n)}^2 - \frac{1}{\mathcal N_n}\sum_{\lambda} |a_\lambda|^4 \right\}.
\end{eqnarray*}
Since $\frac1{\mathcal N_n}\sum_{\lambda} |a_\lambda|^4\to 2$ by the law of large numbers, the claim $(iii)$ follows.
\bigskip\newpage
\noindent{\it Proof of (iv)}. We have
\begin{eqnarray*}
&&\int_{\field{T}} H_2(\widetilde{\partial}_1 T_n(x))H_2(\widetilde{\partial}_2 T_n(x))\,dx
=\int_{\field{T}} \big(\widetilde{\partial}_1 T_n(x)^2\widetilde{\partial}_2 T_n(x)^2-\widetilde{\partial}_1 T_n(x)^2-\widetilde{\partial}_2 T_n(x)^2+1)\,dx
\\
&=&\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda, \lambda',\lambda'',\lambda'''\in\Lambda_n} \lambda_1\lambda'_1\lambda_2''\lambda'''_2\,a_\lambda \overline{a_{\lambda'}}\, a_{\lambda''} \overline{{a}_{\lambda'''}}\int {\rm e}_{\lambda - \lambda'+\lambda'' -\lambda'''}(x)\,dx \\
&&-\frac{2}{n\mathcal N_n}\sum_{\lambda, \lambda'\in\Lambda_n} \lambda_1\lambda'_1a_\lambda \overline{a_{\lambda'}}\int {\rm e}_{\lambda - \lambda'}(x)\,dx
-\frac{2}{n\mathcal N_n}\sum_{\lambda, \lambda'\in\Lambda_n} \lambda_2\lambda'_2a_\lambda \overline{a_{\lambda'}}\int {\rm e}_{\lambda - \lambda'}(x)\,dx
+1
\\
&=&\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda, \lambda''} \lambda_1^2{\lambda_2''}^2|a_\lambda|^2 | a_{\lambda''} |^2
+ \frac{4}{n^2\mathcal N_n^2}\sum_{\lambda} \lambda_1^2\lambda_2^2|a_\lambda|^4
+\frac{8}{n^2\mathcal N_n^2}\sum_{\lambda\neq \pm\lambda'} \lambda_1\lambda_2\lambda'_1\lambda'_2|a_\lambda|^2 | a_{\lambda'} |^2\\
&&-\frac{2}{\mathcal N_n}\sum_{\lambda} |a_\lambda|^2 +1\\
\end{eqnarray*}
\begin{eqnarray*}
&=&\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda, \lambda''} \lambda_1^2{\lambda_2''}^2(|a_\lambda|^2-1) (| a_{\lambda''} |^2-1)
- \frac{12}{n^2\mathcal N_n^2}\sum_{\lambda} \lambda_1^2\lambda_2^2|a_\lambda|^4
+\frac{8}{n^2\mathcal N_n^2}\left(\sum_{\lambda} \lambda_1\lambda_2|a_\lambda|^2 \right)^2\\
&=&\frac{4}{\mathcal N_n}\left\{W_1{}{(n)}W_2{}{(n)} +2W_{1,2}{}{(n)}^2 - \frac{3}{n^2\mathcal N_n}\sum_{\lambda} \lambda_1^2\lambda_2^2|a_\lambda|^4
\right\}.
\end{eqnarray*}
Since $\frac{1}{n^2\mathcal N_n}\sum_{\lambda} \lambda_1^2\lambda_2^2|a_\lambda|^4 \to \frac14(1-\widehat{\mu}_\infty(4))$ by the law of large numbers, the claim $(iv)$ follows.
\bigskip
\noindent{\it Proof of (v)}. We have
\begin{eqnarray*}
&&\int_{\field{T}} H_2(T_n(x))H_2(\widehat T_n(x))\,dx =\int_{\field{T}} \big(T_n(x)^2 \widehat T_n(x)^2- T_n(x)^2- \widehat T_n(x)^2 + 1\big)dx\\
&=&\frac{1}{\mathcal N_n^2}\sum_{\lambda, \lambda',\lambda'',\lambda'''\in\Lambda_n} a_\lambda \overline{a_{\lambda'}}\, \widehat{a}_{\lambda''} \overline{\widehat{a}_{\lambda'''}}\int {\rm e}_{\lambda - \lambda'+\lambda'' -\lambda'''}(x)\,dx
-\frac{1}{\mathcal N_n}\sum_{\lambda, \lambda'\in\Lambda_n} a_\lambda \overline{a_{\lambda'}}\int {\rm e}_{\lambda - \lambda'}(x)\,dx \\
&&-\frac{1}{\mathcal N_n}\sum_{\lambda, \lambda'\in\Lambda_n} \widehat{a}_\lambda \overline{\widehat{a}_{\lambda'}}\int {\rm e}_{\lambda - \lambda'}(x)\,dx
+1\\
&=&\frac{1}{\mathcal N_n^2}\sum_{\lambda,\lambda'} (|a_\lambda|^2-1) (|\widehat{a}_{\lambda'}|^2-1)
+\frac{1}{\mathcal N_n^2}\sum_{\lambda} a_\lambda^2\, \overline{\widehat{a}_{\lambda}}^2
+\frac{2}{\mathcal N_n^2}\sum_{\lambda\neq \pm \lambda'} a_\lambda\, \overline{a_{\lambda'}}\,\overline{\widehat{a}_{\lambda}}
\widehat{a}_{\lambda'}\\
&=&\frac{1}{\mathcal N_n^2}\sum_{\lambda,\lambda'} (|a_\lambda|^2-1) (|\widehat{a}_{\lambda'}|^2-1)
-\frac{1}{\mathcal N_n^2}\sum_{\lambda} a_\lambda^2\, \overline{\widehat{a}_{\lambda}}^2
+\frac{2}{\mathcal N_n^2}\left(\sum_{\lambda} a_\lambda\, \overline{\widehat{a}_{\lambda}}\right)^2 - \frac{2}{\mathcal N_n^2}\sum_{\lambda} |a_\lambda|^2|\widehat{a}_{\lambda} |^2\\
&=&\frac{1}{\mathcal N_n}\left\{W{}{(n)}\widehat{W}{}{(n)}+2M{}{(n)}^2-\frac{1}{\mathcal N_n}\sum_{\lambda} a_\lambda^2\, \overline{\widehat{a}_{\lambda}}^2 - \frac{2}{\mathcal N_n}\sum_{\lambda} |a_\lambda|^2|\widehat{a}_{\lambda} |^2\right\}.
\end{eqnarray*}
Since $\frac{1}{\mathcal N_n}\sum_{\lambda}a_\lambda^2\, \overline{\widehat{a}_{\lambda}}^2 \to 0$
and $\frac{1}{\mathcal N_n}\sum_{\lambda} |a_\lambda|^2|\widehat{a}_{\lambda} |^2 \to 1$ by the law of large numbers, the claim $(v)$ follows.
\bigskip
\newpage
\noindent{\it Proof of (vi)}. We have
\begin{eqnarray*}
&&\int_{\field{T}} H_2(T_n(x))H_2(\widetilde{\partial}_j \widehat T_n(x))\,dx=\int_{\field{T}} \big(T_n(x)^2 \widetilde{\partial}_j\widehat T_n(x)^2\,-T_n(x)^2\,- \widetilde{\partial}_j\widehat T_n(x)^2\, + 1\big)dx\\
&=&\frac{2}{n\mathcal N_n^2}\sum_{\lambda, \lambda',\lambda'',\lambda'''} \lambda_j'' \lambda_j''' a_\lambda \overline{a_{\lambda'}}\, \widehat{a}_{\lambda''} \overline{\widehat{a}_{\lambda'''}}\int {\rm e}_{\lambda - \lambda'+\lambda'' -\lambda'''}(x)\,dx -\frac{1}{\mathcal N_n}\sum_\lambda |a_\lambda|^2 \\
&&-\frac{2}{n\mathcal N_n}\sum_\lambda \lambda_j^2 |\widehat{a}_\lambda|^2 +1\\
&=&\frac{2}{n\mathcal N_n^2}\sum_{\lambda, \lambda''} {\lambda_j''}^2 |a_\lambda|^2 |\widehat{a}_{\lambda''}|^2-\frac{2}{n\mathcal N_n^2}\sum_{\lambda} \lambda_j^2 a_\lambda^2 {\widehat{a}_{\lambda}}^2
+ \frac{2}{n\mathcal N_n^2}\sum_{\lambda\neq \pm\lambda'} \lambda_j \lambda_j' a_\lambda \overline{a_{\lambda'}}\, \widehat{a}_{\lambda'} \overline{\widehat{a}_{\lambda}}\\
&&-\frac{1}{\mathcal N_n}\sum_\lambda |a_\lambda|^2 -\frac{2}{n\mathcal N_n}\sum_\lambda \lambda_j^2 |\widehat{a}_\lambda|^2 +1.
\end{eqnarray*}
Thus,
\begin{eqnarray*}
&&\int_{\field{T}} H_2(T_n(x))\big(H_2(\widetilde{\partial}_1 \widehat T_n(x))+H_2(\widetilde{\partial}_2 \widehat T_n(x))\big)\,dx\\
&=&\frac{2}{\mathcal N_n^2}\sum_{\lambda, \lambda''} (|a_\lambda|^2-1) (|\widehat{a}_{\lambda''}|^2-1)-\frac{2}{\mathcal N_n^2}\sum_{\lambda} a_\lambda^2 {\widehat{a}_{\lambda}}^2
+ \frac{2}{n\mathcal N_n^2}\sum_{\lambda\neq \pm\lambda'} (\lambda_1 \lambda_1' + \lambda_2 \lambda_2' )a_\lambda \overline{a_{\lambda'}}\, \widehat{a}_{\lambda'} \overline{\widehat{a}_{\lambda}}\\
&=&\frac{2}{\mathcal N_n^2}\sum_{\lambda, \lambda''} (|a_\lambda|^2-1) (|\widehat{a}_{\lambda''}|^2-1)
- \frac{2}{\mathcal N_n^2}\sum_{\lambda} |a_\lambda |^2\, |\widehat{a}_{\lambda}|^2
+ \frac{2}{n\mathcal N_n^2}\sum_{\lambda,\lambda'} (\lambda_1 \lambda_1' + \lambda_2 \lambda_2' )a_\lambda \overline{a_{\lambda'}}\, \widehat{a}_{\lambda'} \overline{\widehat{a}_{\lambda}}\\
&=&\frac{2}{\mathcal N_n}\left\{ W{}{(n)}\widehat{W}{}{(n)}+M_1{}{(n)}^2+M_2{}{(n)}^2-\frac{1}{\mathcal N_n}\sum_{\lambda} |a_\lambda |^2\, |\widehat{a}_{\lambda}|^2 \right\}.
\end{eqnarray*}
Since $\frac{1}{\mathcal N_n}\sum_{\lambda} |a_\lambda |^2\, |\widehat{a}_{\lambda}|^2 \to 1$
by the law of large numbers, the claim $(vi)$ follows.
\bigskip
\noindent{\it Proof of (vii)}. We have
\begin{eqnarray*}
&&\int_{\field{T}} H_2(\widetilde{\partial}_\ell T_n(x))H_2(\widetilde{\partial}_j \widehat T_n(x))\,dx\\
&=&\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda, \lambda',\lambda'',\lambda'''} \lambda_\ell \lambda_\ell'\lambda_j'' \lambda_j''' a_\lambda \overline{a_{\lambda'}}\, \widehat{a}_{\lambda''} \overline{\widehat{a}_{\lambda'''}}\int {\rm e}_{\lambda - \lambda'+\lambda'' -\lambda'''}(x)\,dx\\
&&-\frac{2}{n\mathcal N_n}\sum_\lambda \lambda_\ell^2|a_\lambda|^2 -\frac{2}{n\mathcal N_n}\sum_\lambda \lambda_j^2 |\widehat{a}_\lambda|^2 +1\\
&=&\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda,\lambda''} \lambda_\ell^2 {\lambda_j''}^2 |a_\lambda|^2\, |\widehat{a}_{\lambda''}|^2
+\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda} \lambda_\ell^2 \lambda_j^2
a_\lambda^2 \overline{\widehat{a}_{\lambda}}^2
+\frac{8}{n^2\mathcal N_n^2}\sum_{\lambda\neq \pm \lambda'} \lambda_\ell \lambda_\ell'\lambda_j \lambda_j' a_\lambda \overline{a_{\lambda'}}\, \overline{\widehat{a}_{\lambda}} \widehat{a}_{\lambda'}
\\
&&-\frac{2}{n\mathcal N_n}\sum_\lambda \lambda_\ell^2|a_\lambda|^2 -\frac{2}{n\mathcal N_n}\sum_\lambda \lambda_j^2 |\widehat{a}_\lambda|^2 +1\\
&=&\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda,\lambda''} \lambda_\ell^2 {\lambda_j''}^2 (|a_\lambda|^2-1)\, (|\widehat{a}_{\lambda''}|^2-1)
-\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda} \lambda_\ell^2 \lambda_j^2
a_\lambda^2 \overline{\widehat{a}_{\lambda}}^2\\
&&+\frac{8}{n^2\mathcal N_n^2}\sum_{\lambda, \lambda'} \lambda_\ell \lambda_\ell'\lambda_j \lambda_j' a_\lambda \overline{a_{\lambda'}}\, \overline{\widehat{a}_{\lambda}} \widehat{a}_{\lambda'}
-\frac{8}{n^2\mathcal N_n^2}\sum_{\lambda} \lambda_\ell^2\lambda_j^2 |a_\lambda|^2\,|\widehat{a}_{\lambda}|^2
\\
\end{eqnarray*}
\begin{eqnarray*}
&=&\frac{4}{\mathcal N_n}\left\{ W_l{}{(n)}\widehat{W}_j{}{(n)}+2M_{\ell,j}{}{(n)}^2+
\frac{1}{n^2\mathcal N_n}\sum_{\lambda} \lambda_{\ell}^2\lambda_j^2\,a_\lambda ^2\, \widehat{a}_{\lambda}^2-\frac{2}{n^2\mathcal N_n}\sum_{\lambda} \lambda_{\ell}^2\lambda_j^2|a_\lambda |^2\, |\widehat{a}_{\lambda}|^2 \right\}.
\end{eqnarray*}
Since $\frac{1}{n^2\mathcal N_n}\sum_{\lambda} \lambda_{\ell}^2\lambda_j^2\,a_\lambda ^2\, \widehat{a}_{\lambda}^2\to 0$ and $$\frac{1}{n^2\mathcal N_n}\sum_{\lambda} \lambda_{\ell}^2\lambda_j^2|a_\lambda |^2\, |\widehat{a}_{\lambda}|^2 \to \frac18(1-\widehat{\mu}_\infty(4)){\bf 1}_{\{l\neq j\}}+\frac18(3+\widehat{\mu}_\infty(4)){\bf 1}_{\{l= j\}}$$
by the law of large numbers, the claim $(vii)$ follows.
\bigskip
\noindent{\it Proof of (viii)}. We have
\begin{eqnarray*}
&&\int_{\field{T}} \widetilde{\partial}_1 T_n(x)\widetilde{\partial}_2 T_n(x)\widetilde{\partial}_1 \widehat T_n(x)\widetilde{\partial}_2 \widehat T_n(x)\,dx\\
&=&\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda, \lambda',\lambda'',\lambda'''} \lambda_1 \lambda_2'\lambda_1'' \lambda_2''' a_\lambda \overline{a_{\lambda'}}\, \widehat{a}_{\lambda''} \overline{\widehat{a}_{\lambda'''}}\int {\rm e}_{\lambda - \lambda'+\lambda'' -\lambda'''}(x)\,dx \\
&=&\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda, \lambda''} \lambda_1\lambda_2 {\lambda_1''}{\lambda_2''} |a_\lambda |^2|\widehat{a}_{\lambda''}|^2
+\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda} \lambda_1^2 \lambda_2^2 a_\lambda^2 \overline{\widehat{a}_{\lambda}}^2
+\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda\neq\pm \lambda'} \lambda_1^2 {\lambda_2'}^2 a_\lambda \overline{a_{\lambda'}}\, \widehat{a}_{\lambda'} \overline{\widehat{a}_{\lambda}}\\
&&
+\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda\neq\pm \lambda'} \lambda_1 \lambda_2\lambda_1' \lambda_2' a_\lambda \overline{a_{\lambda'}}\, \widehat{a}_{\lambda'} \overline{\widehat{a}_{\lambda}}\\
&=&\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda, \lambda''} \lambda_1\lambda_2 {\lambda_1''}{\lambda_2''} |a_\lambda |^2|\widehat{a}_{\lambda''}|^2
+\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda, \lambda'} \lambda_1^2 {\lambda_2'}^2 a_\lambda \overline{a_{\lambda'}}\, \widehat{a}_{\lambda'} \overline{\widehat{a}_{\lambda}}\\
&&+\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda, \lambda'} \lambda_1 \lambda_2\lambda_1' \lambda_2' a_\lambda \overline{a_{\lambda'}}\, \widehat{a}_{\lambda'} \overline{\widehat{a}_{\lambda}}
-\frac{8}{n^2\mathcal N_n^2}\sum_{\lambda} \lambda_1^2 \lambda_2^2 |a_\lambda|^2\, |\widehat{a}_{\lambda}|^2
-\frac{4}{n^2\mathcal N_n^2}\sum_{\lambda} \lambda_1^2 \lambda_2^2 a_\lambda^2\overline{\widehat{a}_{\lambda}}^2\\
&=&\frac{4}{\mathcal N_n}
\left\{
W_{1,2}{}{(n)}\widehat{W}_{1,2}{}{(n)}+M_{11}{}{(n)}M_{22}{}{(n)}+M_{12}{}{(n)}^2 - \frac{2}{n^2\mathcal N_n}\sum_{\lambda} \lambda_1^2 \lambda_2^2 |a_\lambda|^2\, |\widehat{a}_{\lambda}|^2\right.\\
&&\left. -\frac{1}{n^2\mathcal N_n}\sum_{\lambda} \lambda_1^2 \lambda_2^2 a_\lambda^2\overline{\widehat{a}_{\lambda}}^2
\right\}.
\end{eqnarray*}
Since $\frac{1}{n^2\mathcal N_n}\sum_{\lambda} \lambda_{1}^2\lambda_2^2\,a_\lambda ^2\, \widehat{a}_{\lambda}^2\to 0$ and $\frac{1}{n^2\mathcal N_n}\sum_{\lambda} \lambda_{1}^2\lambda_2^2|a_\lambda |^2\, |\widehat{a}_{\lambda}|^2 \to \frac18(1-\widehat{\mu}_\infty(4))$
by the law of large numbers, the claim $(viii)$ follows.
\subsection{Taylor expansions for the two-point correlation function}\label{taylor}
{The matrix
\begin{equation}\label{matrix omega}
\Omega_n({\bf x}) := \begin{pmatrix}&\frac{E_n}{2} - \frac{(\partial_1 r_n({\bf x}))^2}{1-r_n({\bf x})^2} & - \frac{\partial_1 r_n({\bf x}) \partial_2 r_n({\bf x})}{1-r_n({\bf x})^2}\\
&- \frac{\partial_1 r_n({\bf x}) \partial_2 r_n({\bf x})}{1-r_n({\bf x})^2} &\frac{E_n}{2} - \frac{(\partial_2 r_n({\bf x}))^2}{1-r_n({\bf x})^2}
\end{pmatrix}
\end{equation}
is the covariance matrix of the random vector $\nabla T_n(x)$ conditioned to $T_n(x)=T_n(0)=0$ (see \cite[Equation (24)]{KKW}).}
We will write $\Omega= \Omega_n({\bf x})$, $E=E_n$ and $r=r_n({\bf x})$ for brevity. The determinant of $\Omega$ is
$$
\det \Omega = \frac{E}{2}\left ( \frac{E}{2} - \frac{(\partial_1 r)^2 + (\partial_2 r)^2}{1-r^2} \right).
$$
It is easy to show that $$
\lim_{\|{\bf x}\|\to 0} \det \Omega({\bf x}) = 0.
$$
However, we need the speed of convergence to $0$ of $\det \Omega$. Hence, we will use a Taylor expansion argument around $0$. We denote ${\bf x}=(x,y)$.
\begin{lemma}\label{lemma taylor} As $\|x\|\to 0$, we have
$$\displaylines{
\det \Omega_n(x) = c E_n^3 \|{\bf x}\|^2 +E_n^4\,O( \|{\bf x}\|^4),
}$$
and hence
$$
\Psi_n(x) :=\frac{|\Omega(x)|}{1-r^2(x)}= cE_n^2 + E_n^3 O\left (\|x\|^2\right),
$$
where both $c>0$ and the constants in the `$O$' notation do not depend on $n$.
\end{lemma}
\noindent
\begin{proof}
Let us start with $r$.
$$\displaylines{
r({\bf x}) = r(0,0) + \frac{1}{2} \langle \text{Hess}_r(0,0) {\bf x}, {\bf x} \rangle +\cr +\frac{1}{4!} \partial_{1111} r(0,0) x^4
+ \frac{1}{4!} \partial_{2222} r(0,0) y^4 +\frac{1}{2! 2!} \partial_{1122} r(0,0) x^2 y^2 + o_n(\|{\bf x}\|^4).
}$$
We have
$$
\text{Hess}_r(0,0) =\begin{pmatrix} &-\frac{E}{2} &0\\
&0 &-\frac{E}{2}
\end{pmatrix}
$$
and moreover
$$\displaylines{
\partial_{1111} r(0,0) = (2\pi)^4 n^2 \psi_n\cr
\partial_{2222} r(0,0) = (2\pi)^4 n^2 \psi_n\cr
\partial_{1122}r(0,0) = (2\pi)^4 n^2 (1/2 - \psi_n),
}$$
where $$
\psi_n := \frac{1}{n^2 \mathcal N_n}\sum_{\lambda \in \Lambda_n} \lambda_1^4 =\frac{1}{n^2 \mathcal N_n}\sum_{\lambda \in \Lambda_n} \lambda_2^4.
$$
Therefore we can write, as $\|{\bf x}\|\to 0$,
$$\displaylines{
r({\bf x}) = 1 - \frac{E}{4}(x^2 + y^2) + \frac{1}{4!}(2\pi)^4 n^2 \psi_n x^4
+ \frac{1}{4!} (2\pi)^4 n^2 \psi_n y^4 +\frac{1}{2! 2!} (2\pi)^4 n^2 (1/2 - \psi_n) x^2 y^2 + R_n^r.
}$$
{More precisely, the remainder $R_n^r $ is of the form
$$
R_n^r = O \left(\sup \|\partial^6 r_n\| \| {\bf x }\|^6 \right ).
$$
It is easy to check that
$$
\Big |\partial^6 r_n \Big | \le E_n^3
$$
and hence
$$
R_n^r = E_n^3 O \left( {\bf x }\|^6 \right ),
$$
where the constants involved in the 'O' notation do not depend on $n$. }
Analogously, we find that
$$\displaylines{
\partial r_1 ({\bf x}) = -\frac{E}{2}x + \frac{1}{3!} (2\pi)^4 n^2 \psi_n (x^3+y^3) +\frac{1}{2} (2\pi)^4 n^2 (1/2 - \psi_n)xy^2 + R_n^1
}$$
and
$$\displaylines{
\partial r_2 ({\bf x}) = -\frac{E}{2}y + \frac{1}{3!} (2\pi)^4 n^2 \psi_n (x^3+y^3) +\frac{1}{2} (2\pi)^4 n^2 (1/2 - \psi_n)x^2 y+ R_n^2.
}$$
Also here, the remainders $R_n^1 $ and $R_n^2$ are both of the form
$$
R_n^j \le \sup \|\partial^6 r_n\| \cdot O(\| {\bf x }\|^5).
$$
where also here the constants involved in the 'O' notation do not depend on $n$. Squaring previous Taylor expansions we hence get
$$\displaylines{
r^2({\bf x})= 1 +\left( \frac{E}{4}\right)^2 (x^2+y^2)^2 - \frac{E}{2}(x^2 + y^2) +\cr
+2\left( \frac{1}{4!}(2\pi)^4 n^2 \psi_n x^4
+ \frac{1}{4!} (2\pi)^4 n^2 \psi_n y^4 +\frac{1}{2! 2!} (2\pi)^4 n^2 (1/2 - \psi_n) x^2 y^2\right) + o_n(\|{\bf x}\|^4)=\cr
=1 - \frac{E}{2}(x^2+y^2) + f_n(x,y) +E_n^3 \cdot O(\|{\bf x}\|^6),
}$$
where $f_n(x,y)$ is defined as
$$
\left( \frac{E}{4}\right)^2 (x^2+y^2)^2
+2\left( \frac{1}{4!}(2\pi)^4 n^2 \psi_n x^4
+ \frac{1}{4!} (2\pi)^4 n^2 \psi_n y^4 +\frac{1}{2! 2!} (2\pi)^4 n^2 (1/2 - \psi_n) x^2 y^2\right).
$$
{Therefore,
\begin{equation}\label{uno meno err quadro}
1-r^2 = \frac{E}{2}(x^2+y^2) - f_n(x,y) +E_n^3 \cdot O(\|{\bf x}\|^6).
\end{equation}}
Let us now investigate the derivatives {}{$(\partial_i r) ^2$, $i=1,2$. Firstly,}
$$\displaylines{
(\partial_1 r) ^2 =\left( \frac{E}{2}\right)^2 x^2 - E x \left( \frac{1}{3!} (2\pi)^4 n^2 \psi_n (x^3+y^3) +\frac{1}{2} (2\pi)^4 n^2 (1/2 - \psi_n)xy^2 \right) + E_n \cdot E_n^3 O(\|{\bf x}\|^6)
}$$
where the constants involved in the 'O' notation still do not depend on $n$.
{}{Secondly,}
$$\displaylines{
(\partial_2 r) ^2 =\left( \frac{E}{2}\right)^2 y^2 - E y \left( \frac{1}{3!} (2\pi)^4 n^2 \psi_n (x^3+y^3) +\frac{1}{2} (2\pi)^4 n^2 (1/2 - \psi_n)x^2y \right) + E_n \cdot E_n^3 O(\|{\bf x}\|^6).
}$$
For brevity, let us denote
$$
a_n(x,y):= - E x \left( \frac{1}{3!} (2\pi)^4 n^2 \psi_n (x^3+y^3) +\frac{1}{2} (2\pi)^4 n^2 (1/2 - \psi_n)xy^2 \right)
$$
and
$$
b_n(x,y):= - E y \left( \frac{1}{3!} (2\pi)^4 n^2 \psi_n (x^3+y^3) +\frac{1}{2} (2\pi)^4 n^2 (1/2 - \psi_n)x^2y \right),
$$
so that
$$\displaylines{
(\partial_1 r) ^2 =\left( \frac{E}{2}\right)^2 x^2 +a_n(x,y) + E_n \cdot E_n^3 O(\|{\bf x}\|^6),
}$$
and moreover
$$\displaylines{
(\partial_2 r) ^2 =\left( \frac{E}{2}\right)^2 y^2 + b_n(x,y) + E_n \cdot E_n^3 O(\|{\bf x}\|^6).
}$$
Thus we have, for fixed $n$, as $\|{\bf x}\|\to 0$, using also \paref{uno meno err quadro},
\begin{equation*}
\begin{split}
&\det \Omega = \frac{E}{2}\left ( \frac{E}{2} - \frac{(\partial_1 r)^2 + (\partial_2 r)^2}{1-r^2} \right)\cr
&= \frac{E}{2}\left ( \frac{E}{2} - \frac{\left( \frac{E}{2}\right)^2 (x^2+y^2) +a_n(x,y) + b_n(x,y) + E_n \cdot E_n^3 O(\|{\bf x}\|^6)}{\frac{E}{2}(x^2+y^2) - f_n(x,y) +E_n^3 O(\|{\bf x}\|^6)} \right)\cr
&= \frac{E}{2}\left ( \frac{E}{2} - \frac{E}{2} \frac{1 +\left(\frac{2}{E}\right)^2\left(\frac{a_n(x,y) + b_n(x,y)}{x^2 + y^2} + E_n \cdot E_n^3 O(\|{\bf x}\|^4)\right)}{1 - \frac{2}{E}\left(\frac{f_n(x,y)}{x^2+y^2} +E_n^3 O(\|{\bf x}\|^4)\right)} \right)\cr
&=\left(\frac{E}{2}\right)^2 \left( 1- \frac{1 +\left(\frac{2}{E}\right)^2\left(\frac{a_n(x,y) + b_n(x,y)}{x^2 + y^2} + E_n^4 O(\|{\bf x}\|^4)\right)}{1 - \frac{2}{E}\left(\frac{f_n(x,y)}{x^2+y^2} +E_n^3 O(\|{\bf x}\|^4)\right)} \right) \cr
&=\left(\frac{E}{2}\right)^2 \left[1- \left(1 +\left(\frac{2}{E}\right)^2\left(\frac{a_n(x,y) + b_n(x,y)}{x^2 + y^2} + E_n^4 O(\|{\bf x}\|^4)\right)\right)\right.\\
&\hskip6.3cm\times \left.\left(1 +\frac{2}{E}\left(\frac{f_n(x,y)}{x^2+y^2} +E_n^3 O(\|{\bf x}\|^4)\right)\right) \right]\cr
&=\left(\frac{E}{2}\right)^2 \left[1- \left(1 +\left(\frac{2}{E}\right)^2\frac{a_n(x,y) + b_n(x,y)}{x^2 + y^2} +\frac{2}{E}\frac{f_n(x,y)}{x^2+y^2} +E_n^2 O(\|{\bf x}\|^4)\right) \right]\cr
&=\left(\frac{E}{2}\right)^2 \left[ -\left(\frac{2}{E}\right)^2\frac{a_n(x,y) + b_n(x,y)}{x^2 + y^2} -\frac{2}{E}\frac{f_n(x,y)}{x^2+y^2} +E_n^2 O(\|{\bf x}\|^4)\right]\cr
&=c E_n^3 \|{\bf x}\|^2 +E_n^4\,O( \|{\bf x}\|^4).
\end{split}
\end{equation*}
\end{proof}
|
\section{Introduction}
\label{sec:intro}
Galaxies are assembled over cosmic time by the accumulation of stellar mass (M$_{*}$) through
star-formation (SF) processes.
This build-up is accompanied by an increase of metal content, typically measured through
the gas-phase oxygen abundance (O/H), the most abundant heavy element produced by
massive stars.
Stellar mass is a measure of the integrated SF activity over the history of the galaxy, while
the star-formation rate (SFR) indicates the current rate for conversion of gas into stars.
The gas-phase metallicity ($Z$) reflects not only the metal production from high-mass stars,
but also the level of galaxy interactions with environment through inflows and outflows
in the form of galactic winds.
Given the causal relation between star-formation processes and metal content
in galaxies, it is not surprising that M$_{*}$, SFR, and O/H are mutually
correlated.
The mass-metallicity relation \citep[MZR, e.g.,][]{tremonti04} is a
manifestation of the M$_{*}$\,$- Z$ correlation; the SF ``main sequence'' relates
M$_{*}$\ and SFR \citep[SFMS, e.g.,][]{brinchmann04,salim07,noeske07}.
The mutual relations among the three variables extend to specific SFR (sSFR\,$\equiv$\,SFR/M$_{*}$)
and metallicity which are also correlated \citep[e.g.,][]{salim14,yates14}.
These mutual correlations imply that residuals from the main relations (MZR, SFMS)
should be correlated with the third variable.
Indeed, from an analysis of data from the Sloan Digital Sky Survey (SDSS),
\citet{mannucci10} found an expression that connected the residuals in the
MZR to SFR; this was dubbed the ``Fundamental Metallicity Relation" (FMR) and
reduced the scatter in O/H over $\sim$80\,000 galaxies from $\sim$0.1\,dex
to 0.05$-$0.06\,dex.
In a similar vein, \citet{laralopez10} showed that
the 3D space of M$_{*}$, SFR, and O/H
for $\sim$33\,000 SDSS galaxies
could be expressed through a two-dimensional
(planar) surface (``Fundamental Plane", FP).
By fitting regressions to parameter pairs, they expressed the FP in terms of M$_{*}$\
and found a residual variation of $\sim$0.16\,dex, larger however than that
found for the FMR.
Given that reducing a three-dimensional (3D) parameter space to a (2D) plane is mathematically
equivalent to diagonalizing the 3D covariance matrix, a natural approach to this
problem is a Principal Component Analysis (PCA).
A PCA
was first applied to M$_{*}$, SFR, and O/H by \citet{hunt12} for $\sim$1000 galaxies
from $z\sim 0 - 3.5$ selected to span a range of
$\ga 10^5$ in SFR and two orders of magnitude in O/H\footnote{To avoid
problems with the curvature of the MZR at high metallicities, M$_{*}$\ was limited
to $\leq$10.5\,dex\,$M_\odot$, so those results are formally applicable
only to galaxies less massive than this limit.}.
The PCA showed that the principal component dominated by O/H was the
component {\it most dependent} on the other two,
as by itself it comprised only $\sim$2\% of the total variance.
The PCA resulted in a FP in metallicity (FPZ) with a spread of 0.17\,dex in O/H,
despite the vast range in the original parameters, including redshift.
This FPZ applied to the same SDSS samples used by \citet[][here mass
limited]{mannucci10} gave roughly the same residuals as the FMR,
0.06\,dex. Thus, \citet{hunt12} concluded that the FPZ could be used to estimate metallicities with
an accuracy of $\sim 40-50$\% over an extended range of M$_{*}$\ and SFR, and moreover was a good representation of O/H at $z\ga3$.
It is now well established that both the MZR and the SFMS extend to the highest redshifts examined so far,
but with differing normalizations relative to the Local Universe;
at a given M$_{*}$, SFR (and sSFR) increases with increasing redshift
\citep[e.g.,][]{noeske07,elbaz11,karim11,wuyts11,speagle14} while metallicity decreases
\citep[e.g.,][]{erb06a,maiolino08,mannucci09,cresci12,xia12,yabe12,henry13,cullen14,zahid14,troncoso14,steidel14,wuyts14,ly15,delosreyes15}.
Consequently, if we assume that the FPZ is redshift-invariant (an assumption
that we shall reassess below), the higher sSFRs found in high$-z$ galaxy populations must
be related, perhaps causally, to the lower metallicities observed at the same redshift.
This is the hypothesis we examine in this paper.
In order to observationally constrain the evolution of metallicity with
redshift, we have compiled a new dataset of $\sim 1000$ star-forming galaxies
from $z\simeq 0$ to $z\sim3.7$ with nebular oxygen abundance measurements;
we will refer this compilation as the ``MEGA'' dataset, corresponding to
{\it Metallicity Evolution and Galaxy Assembly}.
This compilation is a radical improvement over the dataset used by \citet{hunt12}
because of the inclusion of several more high$-z$ samples
and, more importantly, because of a common metallicity calibration.
Section \ref{sec:samples} describes the 19 individual samples which form
the MEGA compilation, together with our estimates of stellar masses and SFRs
for the samples at $z\simeq 0$.
The procedures for aligning the individual samples to a common O/H calibration
are outlined in Sect. \ref{sec:ohcalib}.
Sect. \ref{sec:scaling} describes the scaling relations for the MEGA dataset
and
re-evaluates the redshift invariance of the FPZ through
a linear analysis of the correlations of M$_{*}$, SFR, and O/H in the MEGA sample
and in $\sim 80\,000$ galaxies
at $z\sim0$ selected from the SDSS by \citet{mannucci10}.
The coevolution of SFR and O/H with redshift in the MEGA dataset
is presented in Section \ref{sec:coevo}, together with a comparison of
results with previous work.
We discuss our results and summarize our conclusions in Sect. \ref{sec:discussion}.
Throughout the paper we use a \citet{chabrier03} Initial-Mass Function (IMF)
and, when necessary, adopt the conversions for M$_{*}$\ and SFR given by \citet{speagle14}.
\section{Galaxy samples}
\label{sec:samples}
Because of the need to compare stellar mass, M$_{*}$,
SFR, and metal abundance [as defined by the
nebular oxygen abundance, 12$+$log(O/H)], we have selected only samples of galaxies
for which either these quantities are already available in the
literature, or can be derived from published data.
Here we discuss the estimates of M$_{*}$\ and SFR;
the metallicity determinations for the samples will be discussed
in Sect. \ref{sec:ohcalib}.
\subsection{Local Universe}
Four samples of galaxies in the Local Universe met these criteria:
the 11\,Mpc distance-limited sample of nearby galaxies or
Local Volume Legacy \citep[11HUGS, LVL:][]{kennicutt08,lee09,lee11};
the Key Insights into Nearby Galaxies: a Far-Infrared Survey with {\it Herschel}\ \citep[KINGFISH,][]{kennicutt11};
the starburst sample studied by \citet{engelbracht08}, and
the blue compact dwarf (BCD) sample by \citet{hunt10}.
There are 15 galaxies that appear both in the KINGFISH and LVL samples;
for these, we used the KINGFISH parameters from \citet{kennicutt11} because
of the uniform O/H calibration given by \citet{moustakas10}.
The starbursts from \citet{engelbracht08} were restricted
to only those galaxies (42) with
metallicities derived from the ``direct'' or ``T$_{\rm e}$" method based on electron temperatures
(see Table \ref{tab:samples} and Sect. \ref{sec:ohcalib});
the BCDs (23) all have T$_{\rm e}$-measured metallicities.
\subsubsection{Star-formation rates }
\label{sec:sfr}
In order to maximize consistency,
we have recalculated SFRs and M$_{*}$\ for the four local samples
starting from photometric fluxes reported in the literature.
SFRs were derived according to \citet{murphy11} using
for KINGFISH and LVL the hybrid method with far-ultraviolet (FUV)$+$total infrared luminosity ($L_{\rm TIR}$);
these data were available for 123 (of 138 non-KINGFISH) galaxies with O/H in the LVL
and for 50 (of 55) KINGFISH galaxies.
For the KINGFISH and LVL galaxies without these data, we adopted other SFR calibrations
given by \citet{murphy11} including TIR (5 galaxies in KINGFISH, 6 LVL),
UV (3 LVL), and H$\alpha$$+$24\,$\mu$m\ (4 LVL).
$L_{\rm TIR}$\ was calculated according to \citet{draine07} and fluxes
were taken from \citet{dale07,dale09} and \citet{lee09,lee11}.
For 20 LVL galaxies, the SFRs inferred from $L_{\rm TIR}$\ were larger than those from
FUV$+$$L_{\rm TIR}$\ using the prescriptions by \citet{murphy11};
in those cases, we adopted SFR($L_{\rm TIR}$).
For NGC\,253 and M\,82, SFR($L_{\rm TIR}$) is $\sim$2 times SFR(FUV$+$$L_{\rm TIR}$),
but for the other galaxies the two estimates agree to within 30\%.
We also compared for the LVL galaxies
the SFRs calculated with FUV$+$$L_{\rm TIR}$\ with those inferred
by combining H$\alpha$\ and 24-$\mu$m\ luminosities \citep[$L_{\rm{H}\alpha}$, $L_{24}$:][]{calzetti10,murphy11};
SFR(FUV$+$$L_{\rm TIR}$) tends to be
$\sim$1.6 times larger than SFR(H$\alpha$$+L_{24}$) with a scatter of $\sim$0.2\,dex.
This is consistent with the findings of \citet{leroy12} who found
a similar trend at low surface SFR densities such as those in the LVL galaxies.
Because FUV data are generally not available for the starbursts or the BCDs,
for these we adopted
the hybrid combination of H$\alpha$$+L_{24}$\,$\mu$m\ as prescribed by \citet{murphy11}.
Total H$\alpha$\ fluxes were taken from
\citet{dopita02,gildepaz03,james04,pustilnik04,cannon05,moustakas06,schmitt06,lopezsanchez08,kennicutt08,cairos10,james10}
and 24\,$\mu$m\ measurements from \citet{engelbracht08}.
When these data were unavailable (9 galaxies), we adopted the SFR(TIR) prescription
by \citet{murphy11} using the fluxes by \citet{engelbracht08}.
For SBS\,0335$-$052 and I\,Zw\,18, we adopted the SFRs from
radio free-free emission \citep{hunt04,hunt05,johnson09}, given
the superiority of such estimates over other methods \citep[e.g.,][]{murphy11}.
For one galaxy in the starburst sample (UM\,420), because of the lack of MIPS observations, the SFR was estimated
from H$\alpha$\ luminosities, and for one galaxy (ESO\,489$-$G56) there were no data available from which to infer SFR
(so it was not considered further).
For the (23) BCDs from \citet{hunt10}, total H$\alpha$\ fluxes were taken from \citet{gildepaz03,rosagonzalez07,perez11,lagos14},
and 24\,$\mu$m\ fluxes from Hunt et al. (2016, in prep.).
As for the starbursts, for the three galaxies without H$\alpha$\ data, we used SFR(TIR); and
for SBS\,1030$+$583 there were no MIPS data so we adopted SFR(H$\alpha$).
Considering the different SFR estimators discussed above,
and considering their varying degrees of applicability, for the local samples
the uncertainties on the SFRs are probably around a factor of 2 (0.3\,dex).
As mentioned above, the SFRs have been reported to a \citet{chabrier03} IMF.
\subsubsection{Stellar masses}
\label{sec:mstar}
We calculated the stellar masses according to \citet{wen13},
a method
based on WISE W1 (3.4\,$\mu$m) luminosities.
This approach exploits the approximately constant mass-to-light ratios
of stellar populations at near-infrared wavelengths,
independently of metallicity and age \citep{norris14,mcgaugh14}.
However, when W1 photometry was not available, we used
IRAC 3.6\,$\mu$m\ photometry instead.
In fact, the two bands are very similar;
using data from \citet{brown14}, \citet{grossi15} find for spirals a
mean flux ratio $F_{3.4}/F_{3.6}\,=\,1.02\pm0.035$.
Including also the data for dwarf irregulars from \citet{brown14} we find a
mean flux ratio $F_{3.4}/F_{3.6}\,=\,0.98\pm0.061$.
Thus, we conclude that the ratio of the W1 and IRAC 3.6\,$\mu$m\ bands is
unity, with 5-6\% scatter for galaxies like our targets.
For the starburst and BCD samples, we used the H{\sc ii}-galaxy formulation by \citet{wen13}, rather than what
they found for their full sample; the H{\sc ii}\ galaxies have the lowest mass-to-light
ratios in their compilation, corresponding roughly to the bluest regions of
the galaxies studied by \citet{zibetti09}.
To better take into account the weak trends with abundance found by \citet{wen13},
we also applied an approximate correction
for low metallicity \citep[by multiplying the mass-to-light ratio
by 0.8 when 12$+$log(O/H)$\leq$8.2:][see their Fig. 17]{wen13}.
Instead, for the LVL and KINGFISH samples, we adopted the
\citet{wen13} formulation based on morphological type, and considered
an ``early-type'' galaxy one with Hubble type $T<2$\footnote{The distinction used by \citet{wen13}
is based on colors which are not available for all our samples.}.
However, before applying the relations by \citet{wen13}, we first
subtracted nebular emission and emission from hot dust where possible.
In starbursts and BCDs,
such contamination can be very important in the near-infrared
and can contribute 50\% or more to the observed flux at these wavelengths
\citep{hunt01,hunt02,smith09,hunt12}.
The ionized gas continuum contribution to the 3.4-3.6\,$\mu$m\ flux was
estimated from the SFR using the emission coefficients from \citet{osterbrock06}.
When possible, we also subtracted the hot-dust component, with
the assumption that $H$-band emission is entirely stellar.
Because $H$-band photometry is available for some of our sample, we used the data
from \citet{brown14} to estimate the maximum possible IRAC 3.6\,$\mu$m/$H$-band
ratio in galaxies similar to our targets;
95\% of the spiral/dwarf irregular galaxies have a flux ratio $\leq$2.4.
This corresponds to a (Vega-based) [H-3.6] color of $\sim0.8$, consistent
with what is found for the pure stellar component in star-forming galaxies
\citep{hunt02}.
After subtraction of the nebular component, any excess over this ratio was
attributed to hot dust and subtracted; this subtraction was not possible
for 33 galaxies, including all the BCDs.
We compared the stellar masses obtained with the formulation of
\citet{wen13} to those calculated according to
\citet{lee06} based on IRAC 4.5\,$\mu$m\ luminosities \citep[used by][]{hunt12}.
For the BCDs, the masses based on \citet{wen13} are on average $-$0.2\,dex lower
than those based on \citet{lee06} with a scatter of 0.15\,dex;
this is not unexpected given the blue colors of these galaxies
and the results of \citet{zibetti09} who showed that the \citet{bell01}
calibration used by \citet{lee06} gives mass-to-light ratios
that are too high for such blue galaxies.
Instead for the starbursts the two estimates are in closer agreement,
with $-$0.06\,dex difference on average and a scatter of 0.20\,dex.
Moreover, for the galaxies having both W1 and IRAC data, the stellar masses
obtained from 3.6\,$\mu$m\ luminosities are within $\sim$5\% of those
from W1 as expected.
For LVL, we have compared our estimates with those from \citet{cook14}
who used a constant mass-to-light ratio and the IRAC 3.6\,$\mu$m\ luminosity.
The Wen-derived stellar masses are on average $-$0.25\,dex smaller than
the \citet{cook14} values, with a scatter of 0.09\,dex.
\citet{skibba11} derived stellar masses for the KINGFISH sample according
to the formulation of \citet{zibetti09} based on optical and $H$-band colors.
We have compared ours derived using \citet{wen13} to theirs and find the values
from \citet{skibba11} are smaller by $\sim$0.5\,dex on average, with a 0.3\,dex scatter.
Given the significant uncertainties inherent in the procedures to
derive stellar masses over a wide range of galaxy types,
and considering the offsets and scatters of our new M$_{*}$\
estimates,
the uncertainties on the stellar masses for the local samples
are at
most
a factor of 2 (0.3\,dex).
As above, these values are based on a \citet{chabrier03} IMF.
\subsection{SDSS10 $z\simeq 0$ galaxy sample}
\citet{mannucci10} analyzed a set of emission-line galaxies from the SDSS,
using the stellar masses from \citet{kauffmann03}, and SFRs measured
from H$\alpha$\ after correcting for extinction using the Balmer decrement;
they reported all values to a \citet{chabrier03} IMF.
The parameter range covered by this sample is much more limited
than the MEGA sample:
9.2\,$\la$\,dex(M$_{*}$)$\la$\,11.3\,$M_\odot$;
8.5\,$\la$\,12$+$log(O/H)$\la$\,9.1 (assuming the \citealt{kewley02} calibration, see below);
$-1.3\,\la$\,log(SFR)$\la$\,0.8\,$M_\odot$\,yr$^{-1}$.
Nevertheless, we include this sample, hereafter SDSS10, in our analysis because of its
superb statistics for
comparison both locally and at $z>0$.
\begin{figure}
\vspace{\baselineskip}
\includegraphics[height=0.4\textwidth]{mega_redshift_bins.png}
\caption{Redshift distribution of combined sample (without SDSS10).
The 7 redshift bins used throughout the paper for the MEGA dataset
are also shown as vertical dotted lines.
The colors of the portions of the histogram are arbitrary, with the aim
of illustrating ``low" redshift (blue), ``intermediate" redshifts (green),
and ``high" redshifts (red).
}
\label{fig:redshift}
\end{figure}
\begin{table*}
\caption{Characteristics of the individual samples in the MEGA dataset}
\begin{tabular}{lcrccll}
\hline
\multicolumn{1}{c}{Parent} &
\multicolumn{1}{c}{Redshift} &
\multicolumn{1}{c}{Number} &
\multicolumn{1}{c}{Selection} &
\multicolumn{1}{c}{Original} &
\multicolumn{1}{c}{SF method} &
\multicolumn{1}{c}{Reference} \\
\multicolumn{1}{c}{sample} &
\multicolumn{1}{c}{range} &&
\multicolumn{1}{c}{criterion} &
\multicolumn{1}{c}{O/H calibration} \\
\hline
\multicolumn{7}{c}{Local Universe} \\
KINGFISH & $-0.001-0.008$ & 55 & Representative & KK04 & FUV$+$TIR$^{\mathrm a}$ & \citet{kennicutt11} \\
LVL & $-0.001-0.003$ & 138 & Volume-limited & M91, KK04, Direct$^{\mathrm b}$ & FUV$+$TIR$^{\mathrm a}$ & \citet{kennicutt08} \\
Starburst & $ 0.0-0.058$ & 41 & Representative & Direct$^{\mathrm c}$ & H$\alpha$$+$24$\mu$m$^{\mathrm d}$ & \citet{engelbracht08} \\
BCD & $ 0.009-0.044$ & 23 & Primordial helium & Direct$^{\mathrm c}$ & H$\alpha$$+$24$\mu$m$^{\mathrm d}$ & \citet{hunt10} \\
\\
\hline
\multicolumn{7}{c}{$0.1 \leq z \la 0.9$} \\
COSMOS & $ 0.17-0.91$ & 334 & $I$ band & KD02 & H$\alpha$, H$\beta$ & \citet{cresci12} \\
COSMOS & $ 0.62-0.69$ & 26 & [O{\sc iii}]\,$\lambda\lambda$4959,5007 & KK04 & SED fitting & \citet{henry13} \\
DEEP2 & $ 0.71-0.91$ & 27 & [O{\sc iii}]\,$\lambda$4363 & Direct & H$\beta$ & \citet{ly15} \\
NewH$\alpha$\ & $ 0.79-0.82$ & 143$^{\mathrm e}$ & Narrow-band H$\alpha$\ & T04 & H$\alpha$ & \citet{delosreyes15} \\
{\it HST}-grism & $0.60-2.32$ & 11 & [O{\sc iii}]\,$\lambda\lambda$4959,5007, [O{\sc ii}]\,$\lambda$3727 & KK04 & H$\beta$ & \citet{xia12} \\
\\
\hline
\multicolumn{7}{c}{$z > 0.9$} \\
DEEP2 & $1.02-1.40$ & 9 & $R$ band & PP04N2 & H$\alpha$ & \citet{shapley05} \\
DEEP2 & $1.02-1.40$ & 7 & $R$ band & PP04N2 & H$\alpha$ & \citet{liu08} \\
VVDS & $1.27-1.53$ & 6 & [O{\sc ii}]\,$\lambda$3727 & T04 & H$\alpha$ & \citet{queyrel09} \\
BX & $2.11-2.43$ & 7 & $U_n-G$, $G-{\cal R}$ colors & PP04N2 & H$\alpha$ & \citet{shapley04} \\
KBSS & $2.02-2.55$ & 79 & $H$ band$^{\mathrm f}$ & PP04O3N2 & H$\alpha$$^{\mathrm g}$ & \citet{steidel14} \\
LSD & $2.93-3.41$ & 8 & Lyman-break dropout & KD02 & H$\alpha$ & \citet{mannucci09} \\
AMAZE & $3.04-4.87$ & 26 & Lyman-break dropout & KD02 & H$\alpha$ & \citet{troncoso14} \\
COSMOS & $2.97-3.69$ & 35 & Predicted H$\beta$\ & KD02 & UV & \citet{onodera16} \\
\\
\hline
\multicolumn{7}{c}{Stacked samples} \\
SXDS/UDS & $1.27-1.52$ & 5$^{\mathrm h}$ & $K$ band & PP04N2 & H$\alpha$ & \citet{yabe14} \\
COSMOS & $1.40-1.70$ & 10$^{\mathrm h}$ & sBzK & PP04N2 & H$\alpha$ & \citet{zahid14} \\
\\
\hline
\label{tab:samples}
\end{tabular}
\vspace{-\baselineskip}
\begin{flushleft}
$^{\mathrm a}$ ~If not available, then SFR(FUV), or as last choice SFR(TIR). \\
$^{\mathrm b}$ ~Taken from \cite{berg12} or \citet{marble10} when available, otherwise
from \citet{moustakas10} (KK04). \\
$^{\mathrm c}$ ~Taken from
\citet{berg12},
\citet{guseva03a},
\citet{guseva03b},
\citet{guseva11},
\citet{guseva12},
\citet{izotov04},
\citet{izotov06},
\citet{izotov07},
\citet{izotov09},
\citet{izotov12},
\citet{kobulnicky96},
\citet{kobulnicky97},
\citet{kniazev03},
\citet{kniazev04},
\citet{mattsson11},
\citet{perez05},
\citet{roennback95},
\citet{shi05},
\citet{thuan05},
\citet{vigroux87},
\citet{zhao10}. \\
$^{\mathrm d}$ ~If not available, then the maximum of SFR(TIR) and SFR(H$\alpha$). \\
$^{\mathrm e}$ ~AGN have been excluded.\\
$^{\mathrm f}$ ~This is only one of several ``layered'' criteria for selecting the galaxies for KBSS Keck-MOSFIRE observations.\\
$^{\mathrm g}$ ~SFRs are taken from \citet{steidel14a}. \\
$^{\mathrm h}$ ~These are from stacked spectra, but are treated here as individual measurements;
the redshifts are taken as the average given in the respective papers
\citep[$z\sim1.4$, and $z\sim1.6$, for][respectively]{yabe14,zahid14}.\\
\end{flushleft}
\end{table*}
\subsection{$z>0$ samples}
Because our analysis is focused on observationally constraining metal
content at high redshift, to construct the MEGA dataset
we have culled from the literature all available
samples at $z>0$ with measured M$_{*}$, SFR, and O/H.
Stacked analyses have been avoided
where possible, and are used only to increase statistics when
tabulations of observations for individual galaxies were not available
in the required redshift range.
We identified 14 samples at $z>0.1$ (see Table \ref{tab:samples}) for which
these three parameters were measured.
Unavoidably, this compilation is subject to a variety of selection
effects which change with sample and redshift.
Nevertheless, from the observational point of view, the MEGA dataset
constitutes a unique tool with which to assess basic
trends among M$_{*}$, SFR, and O/H, and establish how they vary with
redshift.
Table \ref{tab:samples} lists the samples that comprise the MEGA
dataset, together with their redshift range, selection technique,
and other information.
We postpone the important discussion of metallicity estimates to
Sect. \ref{sec:ohcalib}.
\subsubsection{$0.1 \leq z \leq 0.9$}
The most important representative samples in the redshift range $0.1\leq\,z\,\la\,0.9$ come from two surveys,
zCOSMOS \citep{lilly09,cresci12} and NewH$\alpha$\ \citep{delosreyes15}; these two
datasets alone comprise 477 galaxies.
The first, from COSMOS, is $I$-band selected and was first described by \citet{lilly09}.
Stellar masses were derived from fitting spectral energy distributions
(SEDs) of 12 photometric bands, including {\it Spitzer}/IRAC data
at 3.6$-$5.8\,$\mu$m.
SFRs were calculated from H$\alpha$\ and H$\beta$\ luminosities, after correcting
for extinction either via the Balmer decrement (for galaxies with $z\la0.49$)
or using the extinction estimated from the SED fitting with an appropriate
multiplicative factor.
The second large sample in this redshift range comes from the NewH$\alpha$\ survey,
selected from narrow-H$\alpha$\ band images designed to identify emission-line
galaxies around $z\approx 0.8$.
\citet{delosreyes15} calculate stellar masses through
SED fitting of eight photometric bands (up to observed frame $J$ band),
and estimate SFRs from
the H$\alpha$\ images after correcting for the contribution from [N{\sc ii}]\
and for extinction.
Unlike \citet{hunt12}, we do not include in the MEGA dataset the
Luminous Compact Galaxies (LCGs) by \citet{izotov11} and the ``Green Peas''
\citep{amorin10}; the latter galaxies are selected by bright [O{\sc iii}]\,$\lambda$5007
emission in the SDSS $r$ band \citep{cardamone09}.
LCGs, instead, are defined by requiring an [O{\sc iii}]\,$\lambda$4363
detection, large H$\beta$\ equivalent width (EW), and a flux limit in H$\beta$.
Thus the LCGs are young (because of the high H$\beta$\ EW), highly star forming (because of the H$\beta$\ flux limit)
and metal poor (because of the [O{\sc iii}]\,$\lambda$4363 detection).
Although highly interesting objects, the \citet{izotov11} selection criteria
favor young, metal-poor galaxies, and thus are not be representative of abundances
of typical galaxy populations at those redshifts.
There are $\sim$60 galaxies in the remaining three samples in this redshift range:
galaxies selected from a multi-slit narrowband spectroscopic survey with
[O{\sc iii}]\,$\lambda\lambda$4959,5007, [O{\sc ii}]\,$\lambda \lambda$3727,3729 at $z\sim0.6-0.7$ by \citet{henry13};
[O{\sc iii}]\,$\lambda$4363 DEEP2 selected objects at $z\sim0.7-0.9$ by \citet{ly15};
and galaxies selected from {\it HST}-grism observations ([O{\sc iii}], [O{\sc ii}]) by \citet{xia12}.
These three samples are very interesting because of their selection methods
which tend to favor less massive galaxies than typical broadband photometry
selections.
Stellar masses were derived from SED fitting of COSMOS imaging data including IRAC
bands \citep{henry13}; of 8-band photometry up to $z^\prime$ for the DEEP2
survey \citep{ly15}; and
of 10-band {\it HST} ACS/WFC3 photometry up to F160W \citep{xia12}.
With the exception of \citet{henry13} who used SED fitting to calculate SFRs,
H$\alpha$\ and H$\beta$\ corrected luminosities were used to infer SFRs.
\subsubsection{$z > 0.9$}
Most of the galaxies in this redshift range are color-selected
Lyman-Break Galaxies \citep[e.g.,][]{steidel99}.
However, the \citet{queyrel09} galaxies are
selected from the magnitude-limited Mass Assembly Survey with SINFONI
in VVDS \citep[MASSIV,][]{epinat09},
and the two $z\sim1$ samples
by \citet{shapley05} and \citet{liu08} are selected from
the DEEP2 Galaxy Redshift Survey \citep{davis03}.
Wavelength coverage for stellar-mass determinations varies,
with $U_n G {\cal R} K_s$ \citep{shapley04};
$BRI K_s$ \citep{shapley05,liu08};
$UBVRIZ_s JK$ \citep{queyrel09};
and 14 spectral bands from GOODS-MUSIC \citep{grazian06},
including IRAC 3.6, 4.5\,$\mu$m\ \citep{mannucci09,troncoso14}.
Stellar masses for the \citet{onodera16} COSMOS sample
are fit with $uBVrizYJHK$ and IRAC bands.
\citet{onodera16}
prefer SFRs inferred from extinction-corrected UV luminosities,
but all other SFRs in this redshift range
are determined from H$\alpha$\ suitably corrected for extinction.
To ensure better coverage of the redshift range $1.3 < z < 1.7$,
we have included also the two samples by
\citet{yabe14}
and \citet{zahid14}.
Neither group publishes data for individual galaxies,
so we have
adopted the parameters of their stacked spectra here
as individual galaxies, and used the average redshifts
of $z\sim1.4$ and $z\sim1.6$ for \citet{yabe14,zahid14}, respectively.
The redshift distribution of the MEGA dataset is shown in Fig. \ref{fig:redshift},
together with the 7 redshift bins that will be used throughout the paper.
As mentioned above,
Table \ref{tab:samples} gives the characteristics of the 19 individual samples comprising
the MEGA dataset; there is a total of 990 galaxies from $z\simeq 0$ to
$z \sim 3.7$ (and LnA1689$-$2 in the AMAZE sample at $z\,=\,4.87$).
\section{Metallicity calibrations}
\label{sec:ohcalib}
Oxygen abundance O/H is typically used as a proxy for metallicity in
emission-line galaxies.
Because the ionized gas in H{\sc ii}\ regions at lower metal abundance is hotter (as measured by electron temperature, T$_{\rm e}$),
the preferred technique to establish O/H is to measure T$_{\rm e}$\
and the physical conditions in the ionized plasma.
In this ``direct-temperature" or ``T$_{\rm e}$" method, the T$_{\rm e}$\
of the ionized gas is derived from the ratio of the [O{\sc iii}]\,$\lambda$4363 auroral line to lower-excitation lines
([O{\sc iii}]\,$\lambda$4959,\,5007);
such flux ratios are sensitive to temperature because the auroral and strong lines
originate from different excitation states (second and first excited states, respectively).
Because the oxygen transitions are collisionally excited, the relative population
of the excited states depends on T$_{\rm e}$.
Thus, the strengths of these forbidden lines,
combined with the measurement of T$_{\rm e}$\ and density in the nebula, can be converted to an
abundance, relative to hydrogen,
after
correcting for unseen phases of ionization \citep[e.g.,][]{osterbrock06}.
Although the T$_{\rm e}$\ method is more directly related to metallicity,
the auroral lines are weak and often difficult to detect, especially at high metallicity.
Thus, ``strong-line" methods are more generally used to estimate O/H,
especially in metal-rich objects and at high redshift.
It is necessary to calibrate these methods, either using
theoretical photoionization models \citep[e.g.,][hereafter KD02]{kewley02},
or measurements of T$_{\rm e}$\ \citep[e.g.,][hereafter PP04]{pettini04},
or a combination of the two
\citep[e.g.,][hereafter D02]{denicolo02}.
Despite the best efforts to correctly cross calibrate these methods over
a wide range of physical conditions,
there remain large discrepancies, as high as 0.6\,dex in log(O/H)
\citep[e.g.,][and references therein]{kewley08}.
Thus to correctly assess metal content and its evolution with redshift,
it is necessary to apply a common metallicity calibration to the
samples under discussion.
In nearby galaxies where spectra can be obtained with sufficient signal-to-noise,
the T$_{\rm e}$\ method is generally used.
As mentioned above,
the most widely used auroral T$_{\rm e}$\ diagnostic line is [O{\sc iii}]\,$\lambda$4363 because of
its relative ease of observation, high abundance of emitting ions,
and notable strength in the low- and intermediate-metallicity regime (i.e., below solar metallicity).
However, there are several potential problems with the T$_{\rm e}$\ method based on [O{\sc iii}]:
\begin{enumerate}[{\em i)}]
\item
Metallicities derived from collisionally-excited lines (CELs) such as [O{\sc iii}]\ can be underestimated when temperature fluctuations inside the nebula are present but neglected. The assumption of a single average CEL temperature for the whole nebula, usually higher than the temperature derived from the Balmer discontinuity, tends to lead to an underestimate of the abundances \citep[e.g.,][]{peimbert67,stasinska05,bresolin07,perez10,pena12}.
\item
Additional problems also plague the T$_{\rm e}$\ method including
possible non-Boltzmann electron distributions \citep[e.g.,][]{nicholls12,binette12,nicholls13};
depletion of oxygen onto dust grains \citep[e.g.,][]{peimbert10,pena12};
and potential shock waves within the nebulae \citep[e.g.,][]{binette12}.
\item
Finally, recent results suggest that metallicities derived from [O{\sc iii}]\ may
be more unreliable than those from other auroral lines such as [S{\sc iii}]\,$\lambda$6312
and [N{\sc ii}]\,$\lambda$5755 \citep[e.g.,][]{berg15}; however, these lines are even more
difficult than [O{\sc iii}]\ to measure in distant galaxies.
\end{enumerate}
An alternative to the use of T$_{\rm e}$-diagnostic lines can be the derivation of abundances from optical recombination lines (ORLs),
because of their reduced emissivity dependence on density and temperature.
Abundances derived from the ratio of the intensity of ORLs tend to be systematically higher than those from CELs
\citep[e.g.][]{peimbert93,liu95,liu01,tsamis04,ge07}. However, such differences may arise from the relation of the ORL abundances to
small H-deficient portions of the regions, while the CEL-based metallicities are more representative of the whole nebula \citep[see, e.g.,][]{liu00}.
Moreover, such lines are extremely faint, thus requiring very
high signal-to-noise spectra that are currently available only for the Galaxy
and the Local Group \citep[e.g.,][]{blanc15}.
There are also, perhaps more severe, problems with ``strong-line" methods,
and the simplifying assumptions made for photoionization model calibrations
\citep[e.g., photoionization structure, geometry, stellar age: see][for a thorough discussion]{moustakas10}.
As for the T$_{\rm e}$\ method,
there may also be systematic discrepancies due to the metallicity-dependent correction
for the depletion of oxygen onto dust grains \citep[e.g.,][]{peimbert10}.
Ultimately, the T$_{\rm e}$\ method (with [O{\sc iii}]) is generally considered to be the most viable, given
the limitations with other techniques.
Thus, to ensure the best possible comparison among different
samples that rely on different O/H calibrations,
it is advantageous to use the strong-line calibration method that
most closely resembles values inferred from oxygen-based T$_{\rm e}$-method estimations.
According to the results of \citet{andrews13},
who used a stacking technique to measure the oxygen abundances of
$\sim$200\,000 star-forming galaxies from the SDSS to enhance the
signal-to-noise ratio of the weak [O{\sc iii}]\,$\lambda$4363 line,
there are three such methods:
PP04 (both [N{\sc ii}]\ and [O{\sc iii}]$+$[N{\sc ii}]-based: hereafter PP04N2, PP04O3N2)
and D02.
Over the metallicity and M$_{*}$\ range covered by their calculations
of various strong-line methods \citep[see Fig. 10 of][]{andrews13},
the discrepancies between the T$_{\rm e}$\ method and these three methods
are $\la$0.1\,dex in 12$+$log(O/H).
Thus, in what follows, where there are no direct-T$_{\rm e}$\ estimates,
we have applied the transformations
given by \citet{kewley08} to convert the original strong-line O/H
calibrations for the MEGA dataset (and SDSS10 sample)
to the calibrations by D02 and PP04 (PP04N2, PP04O3N2).
As reported in Table \ref{tab:samples},
the original O/H calibrations include:
KD02 \citep[][]{kewley02,cresci12,mannucci10,troncoso14};
KK04 \citep[][]{kobulnicky04,kennicutt11,henry13,xia12};
M91 \citep[][]{mcgaugh91,marble10};
PP04N2, PP04O3N2 \citep[][]{pettini04,shapley04,shapley05,liu08,yabe12,zahid14,steidel14};
and T04 \citep{tremonti04,delosreyes15}.
\begin{figure*}
\hbox{
\includegraphics[width=\textwidth]{MZR_PP04N2_7allz.png}
}
\caption{Mass-metallicity relation over redshifts from $z\sim0$ to $z\ga3.3$,
binned as described in the text.
The solid (black) curve (labeled $z\sim0$) corresponds to the fit to SDSS10
(with stacked T$_{\rm e}$\ metallicity determinations) by
\citet{andrews13}, and the dotted (grey) curves to the polynomial
fits by \citet{maiolino08} with the KD02 calibration at
$z \approx 0.07$, $z \approx 0.7$, $z \approx 2.2$, and $z \approx 3.5$.
The O/H calibration for all galaxies is PP04N2 as described in the text.
Samples are labeled according to the legend in the lower rightmost panel,
except for $z\ \approx\ 0$ which are: LVL as small open circles
(colors correspond to Hubble types with late types (T$\geq$8) as cyan,
5$\leq$T$<$8 as blue, 3$\leq$T$<$5 as magenta, T$<$3 as red);
KINGFISH as (orange) $+$;
\citet{engelbracht08} as (purple) filled triangles;
\citet{hunt10} as (blue) filled squares.
The stacked samples \citep{yabe14,zahid14} at $z \approx 1$ are not plotted.
}
\label{fig:mzr}
\end{figure*}
\begin{figure*}
\hbox{
\includegraphics[width=\textwidth]{MS_SSFR_7allz.png}
}
\caption{Specific SFR vs. M$_{*}$\ (main sequence of star formation)
over redshifts from $z\sim0$ to $z\ga3.3$, binned as described in the text.
The solid (black) curve (labeled $z\sim0$) corresponds to the fit to the LVL$+$KINGFISH
samples, and
the dotted (grey) curves to the formulation of dependence with
M$_{*}$\ and $z$ by \citet{speagle14} for $z\approx 0.6$,
$z\approx 1.3$, $z\approx 2.3$, and $z\approx 3.3$.
Symbols are as in Fig. \ref{fig:mzr},
and the stacked samples \citep{yabe14,zahid14} are not plotted.
}
\label{fig:ms}
\end{figure*}
\section{Scaling relations and the fundamental plane}
\label{sec:scaling}
The MEGA dataset comprises
three parameters (pseudo-observables, as they are not directly observed):
nebular oxygen abundance (12$+$log(O/H)), stellar mass (M$_{*}$), and SFR.
As discussed in the Introduction, these three parameters are mutually correlated,
although O/H trends flatten at high M$_{*}$\ (and high O/H).
Here we discuss the scaling relations of the three parameters:
the mass-metallicity relation, MZR,
the ``main sequence'' of star formation, SFMS,
and the correlation (at least at $z\sim0$) between sSFR and metallicity.
\begin{figure*}
\hbox{
\includegraphics[width=0.95\textwidth]{SSFR_OH_PP04N2_7allz.png}
}
\caption{Specific SFR vs. 12$+$log(O/H)\
over redshifts from $z\sim0$ to $z\ga3.3$, binned as described in the text.
The solid (black) line
corresponds to the fit to the best-fit slope for the MEGA sample at $z\sim0$,
and
the dotted (grey) lines to the prediction of the redshift variation
assuming the increase in sSFR given by \citet{speagle14}
(from $z\sim0.6$, $z\sim1.3$, $z\sim2.3$, $z\sim3.3$)
of the Fundamental Plane for these two parameters described in Sect. \ref{sec:fp}.
As in Fig. \ref{fig:mzr},
the O/H calibration for all galaxies is PP04N2 as described in the text.
Symbols are as in Fig. \ref{fig:mzr},
and the stacked samples \citep{yabe14,zahid14} are not plotted.
}
\label{fig:ssfroh}
\end{figure*}
The MZR with the PP04N2 O/H calibration
for different redshift bins is shown in
Fig. \ref{fig:mzr}.
The solid curve shows the T$_{\rm e}$-method MZR derived by \citet{andrews13}
which well approximates the MEGA dataset at $z \approx 0$.
The dotted grey curves represent the polynomial fits given by \citet{maiolino08}
for the KD02 calibration;
at high M$_{*}$, these curves fail to capture the T$_{\rm e}$-derived (or PP04N2) metallicities
because of the different O/H calibration.
As virtually all previous work suggests,
the different panels illustrate that as $z$ increases, at a given M$_{*}$\
metallicity decreases.
However, at $z \approx 0$ for a given M$_{*}$,
the starburst and BCD samples tend to be more metal-poor than the LVL and KINGFISH galaxies;
they behave more like galaxies at $z \ga 1$ than like galaxies in the Local Universe,
presumably because of their higher sSFR.
The high sSFRs in the starburst and BCD $z \approx 0$ samples are more clearly seen in
Figure \ref{fig:ms}, which shows the SFMS, or sSFR plotted against M$_{*}$.
The solid line shows the SFMS calibrated with the LVL$+$KINGFISH samples,
having a slope of $-0.19\,\pm\,0.02$, roughly consistent with that ($-0.23$) found by \citet{elbaz07} for $z\sim0$ galaxies.
The dashed grey lines correspond to the \citet{speagle14} formulation for SFR as a function
of cosmic time (we have calculated cosmic age for representative redshifts and plotted the result).
The slope by \citet{speagle14} at $z\ga2$ is similar to what we find for the Local Universe,
which however is shallower (steeper in SFR-M$_{*}$\ space)
than their value for $z\sim0$.
Fig. \ref{fig:ms} illustrates that as redshift increases, for a given M$_{*}$,
sSFR also increases; galaxies that would be main-sequence galaxies at $z\ga1$ are starbursts if
found at $z\approx0$.
However, it is also seen from the figure that the individual high-$z$
samples do not clearly
follow the SFMS;
this is almost certainly due to selection effects and will be further discussed
in Sect. \ref{sec:massvariation}.
Because of the difficulty in measuring metallicities in high-$z$ emission-line galaxies,
flux limits for spectroscopy impose a commensurate limit in SFRs.
\begin{figure*}
\vspace{\baselineskip}
\hbox{
\includegraphics[width=0.95\textwidth]{SDSS_MZR_d02_ppn2_ppo3_new-crop}
}
\caption{SDSS10 galaxies: 12$+$log(O/H)\ plotted against (log of) M$_{*}$.
The three panels correspond to number densities of
12$+$log(O/H)\ (with the three O/H calibrations) and M$_{*}$.
In each panel, the solid curves
correspond to the mass-metallicity relation taken from \citet{andrews13}
for stacked O/H direct-temperature metallicity determinations as also shown
in Fig. \ref{fig:mzr}.
The dashed curves show instead the
relation of the form used by \citet{andrews13} but fit to the SDSS10 data used here.
}
\label{fig:mzrsdss}
\end{figure*}
The third correlation between sSFR and O/H is shown in Figure \ref{fig:ssfroh}.
As in previous figures, the solid line gives the local calibration on the LVL$+$KINGFISH
and the dotted grey lines show the redshift trend for O/H expected for the higher SFR
as predicted by the FPZ (see Sect. \ref{sec:fp}).
The SFRs of the local starbursts and BCDs are higher at a given O/H, relative to the
other local samples; again, they are more similar to galaxies at $z\ga1$ than to typical
local populations.
Similarly to the behavior of the MEGA dataset for the SFMS, the LVL$+$KINGFISH galaxies
show a well-defined correlation between sSFR and O/H, but the correlation disappears
for the higher-redshift samples.
The main point of this third correlation is that, at least locally, the three
psuedo-observables, 12$+$log(O/H), SFR, and M$_{*}$\ are mutually interdependent.
This makes it difficult to determine which is the primary parameter(s) driving
the relations, and it this point which we explore below in
Sect. \ref{sec:fp}.
\subsection{The SDSS10 relations}
\label{sec:sdss}
Similar correlations are found for the SDSS10 galaxies, although
the range in M$_{*}$, O/H, and sSFR is smaller than in the MEGA dataset.
Nevertheless, over the limited parameter range the sheer number statistics
afford precise determinations of scaling relations and fitting functions
which will be important for constraining our models.
Fig. \ref{fig:mzrsdss} gives the MZR for the three O/H calibrations of SDSS10 sample,
transformed from the original KD02.
The solid curves, also shown in Fig. \ref{fig:mzr},
give the MZR for the direct-method O/H as found by \citet{andrews13},
while the dashed ones are functions of the same form but fit to
the SDSS10 dataset itself.
The PP04 calibrations (middle and right panels) are, on average, the best approximation to the
direct-method O/H curve, although at 12$+$log(O/H)\,$\la$8.5, the D02 calibration
is superior.
In any case, the functional MZR form used by \citet{andrews13} does not well approximate the
SDSS10 data at low mass or low metallicities.
The low-mass, low-metallicity linear portion of the data has a slope of $\sim 0.38\,\pm\,0.003$,
similar to the MZR curve, but the latter is offset to higher masses.
The SFMS of the SDSS10 data is shown in Fig. \ref{fig:mssdss}; the solid line corresponds
to the linear regression for the LVL$+$KINGFISH galaxies (also shown in Fig. \ref{fig:ms}) and
the dashed curve to the Schechter-like functional form fitted by \citet{salim07}.
Although this last captures the low- and high-mass ends of the SDSS10 data, it does
not pass through the region with the highest density (blue colors in Fig. \ref{fig:mssdss}).
This could have something to do with the different ways that SFR is calculated;
\citet{salim07} used FUV while \citet{mannucci10} used extinction-corrected H$\alpha$.
Nevertheless, the LVL$+$KINGFISH regression well approximates this behavior, implying
that the SFR derivation is probably not the cause of the discrepancy.
The third correlation, between sSFR and 12$+$log(O/H)\ is not shown for the SDSS10 data;
it shows a similar behavior to the MEGA sample.
\begin{figure}
\vspace{\baselineskip}
\hbox{
\includegraphics[width=0.45\textwidth]{SDSS_SSFRMstar_crop}
}
\caption{SDSS10 galaxies: sSFR plotted against M$_{*}$.
The color scale corresponds to the number densities of the two parameters.
The solid curve gives the linear SF ``main sequence'' regression
to the LVL$+$KINGFISH samples (also shown in Fig. \ref{fig:ms}), and
the dashed curve to the Schechter function fitted by \citet{salim07}.
}
\label{fig:mssdss}
\end{figure}
\subsection{A planar approximation to scaling relations}
\label{sec:fp}
At high M$_{*}$\ and O/H, both the MZR and SFMS inflect and flatten
\citep[e.g.,][]{tremonti04,noeske07,whitaker14,lee15,gavazzi15}.
However, for M$_{*}$\ below a certain threshold, M$_{*}$$\leq3\times10^{10}$\,$M_\odot$, roughly the ``turn-over mass''
\citep{tremonti04,wyder07}, the relations among the variables
are approximately linear.
We propose that, at high M$_{*}$\ and O/H, the inflections the MZR and the SFMS compensate one another,
and hypothesize that even above this inflection threshold,
the trends in M$_{*}$, O/H, and SFR can be approximated by linear relations.
Consequently, as discussed above,
these observationally-defined variables could
define a {\it plane} which, given the relatively large scatters in the SFMS, the
MZR, and the SFR-O/H relation, is not viewed in the best projection.
Because the three parameters are mutually correlated, it is important to determine which
of the three is the most fundamental, and whether or not the planar
approximation is sufficient to describe the data.
This can be readily accomplished through a Principal Component Analysis
\citep[PCA, e.g.,][]{hunt12}.
The MEGA dataset is a significant improvement on the sample studied by \citet{hunt12},
and is particularly well suited for such an analysis.
In particular, the MEGA dataset
triples the number of galaxies at $z\ga2-3$ with respect to \citet{hunt12}.
It spans
almost two orders of magnitude in metallicity (12$+$log(O/H)\,=\,7.1 to $\sim9$),
a factor of $\sim10^6$ in SFR ($\sim10^{-4} \leq$ SFR $\leq\, \sim10^2$\,$M_\odot$\,yr$^{-1}$),
and a factor of $\sim10^5$ in stellar mass ($\sim10^6 \leq$ M$_{*}$\ $\leq\, \sim10^{11}$\,$M_\odot$);
moreover it includes galaxies at redshifts from $z \sim 0-3.8$ (see Fig. \ref{fig:redshift}).
Other samples previously analyzed to find scaling relations cover much smaller
parameter ranges: typically less than a decade in metallicity (12$+$log(O/H)$\geq$8.4),
a factor of $\sim$200 in SFR ($\sim$0.04$\la$ SFR$\la 6$\,$M_\odot$\,yr$^{-1}$),
and roughly 2 orders of magnitude in stellar mass (M$_{*}$$\ga 10^9$\,$M_\odot$)
\citep[e.g.,][]{tremonti04,mannucci10,laralopez10,yates12}.
Because more than 50\% of the MEGA dataset has $z > 0.5$,
and it includes galaxies at redshift $z \ga 3.5$,
we can test the assumption that the relations among the observationally-defined variables
are redshift invariant.
\begin{figure*}
\vspace{\baselineskip}
\includegraphics[width=\textwidth]{MEGAAllOnoderaALLMASS_FPZMEGA_d02_ppn2_ppo3_NEWNEW-crop.pdf}
\vspace{\baselineskip}
\includegraphics[width=\textwidth]{MEGAAllOnoderaALLMASS_Residuals0p05Bins_FPZMEGA_d02_ppn2_ppo3_NEWNEW-crop.pdf}
\caption{Top panel: FPZ projection of O/H for all O/H calibrations
of the MEGA dataset as described in the text.
All galaxies are shown (in gray intensity and contours), including those with M$_{*}$ $\geq 3\times10^{10}$\,$M_\odot$;
colored filled circles correspond to galaxies with $z\geq2$.
The dashed lines give the identity relation.
Bottom panel: histograms of FPZ residuals from the identity relation.
The grey histograms show all galaxies, while the (lower-amplitude) colored ones correspond
to galaxies with $z\geq2$.
The heavy dashed lines give the Gaussian fit of the residuals
(including the separate fit to the galaxies with $z\geq2$);
the $\sigma$ values in the upper left corner of each panel correspond to the mean residuals
for all galaxies.
The width of the best (PP04N2) residual distribution over all redshifts is $\sim$0.16\,dex (see middle panels).
}
\label{fig:fpzall}
\end{figure*}
\subsection{PCA of the MEGA and SDSS10 samples}
\label{sec:pca}
We have therefore performed a PCA for all three O/H calibrations
of the MEGA dataset
without imposing a limit in M$_{*}$\ \citep[c.f.,][]{hunt12}.
A PCA diagonalizes the 3D covariance matrix,
thus defining the orientation of the parameter space which minimizes the covariance.
The orientation is contained in the eigenvectors which are,
by definition, mutually orthogonal.
If the 3D space formed by the three psuedo-observables is truly planar,
we would expect most of the variance to be contained in the first
two eigenvectors (the orientation of the plane);
for the third eigenvector, perpendicular to the plane, the variance should be
very small.
Independently of the O/H calibration,
the PCA shows that the set of three observables truly defines a plane;
$\ga$98\% of the total variance is contained in the first two eigenvectors.
Most (87\%) of the variance is contained
in the first eigenvector alone (or Principal Component, PC), PC1;
it is dominated by SFR, with M$_{*}$\ contributing slightly
less, and O/H giving only a marginal contribution.
PC2, the second eigenvector, holds 10$-$11\% of the variance,
and is dominated by M$_{*}$, followed by SFR, and as in PC1,
with O/H again only marginal.
The smallest fraction of the variance ($\sim 1.5-1.8$\%) is contained in the third
eigenvector, PC3, which is dominated by O/H;
the implication is that O/H is the most {\it dependent} parameter,
governed almost completely by M$_{*}$\ and SFR.
Moreover, this means that the 3D space defined by O/H, SFR, and M$_{*}$\ is degenerate;
because of the mutual correlations of the psuedo-observables,
only two parameters are required to describe the properties of the galaxies.
We have also performed a PCA of the SDSS10 galaxies, and
obtained similar results: namely, the third eigenvector, PC3, the one
dominated by O/H, contains the smallest fraction of the variance.
The residuals of 0.05$-$0.06\,dex are comparable to that obtained
by the FMR formulation by \citet{mannucci10}.
That star-forming galaxies form a plane in O/H, SFR, and M$_{*}$\
is not a new result.
\citet{laralopez10} concluded that the 3D space of O/H, SFR, and M$_{*}$\ of
$\sim$ 33\,000 SDSS galaxies could be represented as a plane
but used a regression analysis rather than a PCA
\citep[although see][]{laralopez13}.
\citet{hunt12} derived a PCA for a dataset similar to ours, although dominated
by LCGs, and also concluded that a 2D plane was sufficient to describe the 3D dataset.
As in the Introduction,
we will refer to the resulting 2D plane as the FPZ (Fundamental
Plane in metallicity).
The FPZ for the MEGA dataset
is shown in the top panel of Figure \ref{fig:fpzall} where we have plotted 12$+$log(O/H)\ vs.
the equation that results from equating PC3 (PP04N2) to zero (see Table \ref{tab:fpzfmr}
for the other O/H calibrations):
\begin{equation}
12+\log(O/H) = -0.14\,{\rm log(SFR)} + 0.37\,{\rm log(M_*)} + 4.82 \ \
\label{eqn:fpzall}
\end{equation}
\noindent
The bottom panel of Fig. \ref{fig:fpzall} shows the residuals from the FPZ
for the different O/H calibrations.
For PP04N2, they are well approximated by a Gaussian with a $\sigma\,=\,0.16$,
corresponding to $\la$45\% uncertainty; the
other two O/H calibrations (D02, PP04O3N2) give similar results,
although slightly larger (see Table \ref{tab:fpzfmr}).
The residuals of the FPZ relation are independent of redshift to within 0.16\,dex,
the overall uncertainty;
nevertheless the different symbols plotted in the top panel
(and the different histograms in the bottom one) suggest some slight deviation with redshift
which we will explore in Sect. \ref{sec:fpzinvariance}.
Because of the turnover of the MZR and SFMS at high stellar masses,
our assumption of linearity in the FPZ could also produce
a residual correlation with M$_{*}$.
We have investigated this possibility and found that the FPZ residuals
and (log)M$_{*}$\ are uncorrelated;
the mean residuals of the regression are $\sim$0.16\,dex (for the PP04 calibration),
the same as those of the FPZ itself.
Moreover, the slope of the FPZ residuals vs. (log)M$_{*}$\ is zero to within
the uncertainties ($-0.015\,\pm\,0.01$).
Thus, our hypothesis that the inflections in the MZR and SFMS compensate
one another is apparently justified; the curvature in the MZR can be adequately accommodated by
the increasing SFRs at high M$_{*}$, at least to within the uncertainties of our data.
The FPZ dispersion of $\sim0.16-0.18$\,dex for the MEGA dataset is higher than that found by \citet{tremonti04} for the MZR
defined by 53\,000 galaxies from the SDSS (0.1\,dex), and also higher than the FMR (0.06\,dex)
found for the SDSS10 sample by \citet{mannucci10}.
However, as shown in Table \ref{tab:fpzfmr},
the SDSS10 data span limited ranges in O/H, SFR, stellar mass (and redshift) relative
to the parameter space covered by the MEGA data.
The mean and standard deviation of (log) stellar mass for the SDSS10 sample is
$10.26\pm0.41$\,$M_\odot$, while the comparable mean, standard deviation for MEGA is $9.44\pm0.93$\,$M_\odot$;
(log) SFR shows a similar pattern:
$-0.04\pm0.43$\,$M_\odot$\,yr$^{-1}$\ for SDSS10 compared to $0.16\pm1.12$\,$M_\odot$\,yr$^{-1}$\
for MEGA.
Thus,
the higher dispersion in the MEGA FPZ is not surprising despite the many more galaxies in SDSS.
The value of the FPZ dispersion is {\it lower} than the scatter of the MZR for $\sim$20\,000 VVDS galaxies
within individual redshift bins from $z \sim 0.3-0.9$ \citep[$\sim$0.20\,dex,][]{lamareille09}.
The FPZ dispersion for the $\sim$1000 galaxies studied here is only slightly higher than
that found for the MZR of 25 nearby dwarf galaxies \citep[0.12\,dex,][]{lee06},
a sample dominated by low-mass galaxies.
It is also only slightly higher than the rms scatter of 0.12\,dex found by \citet{henry13} for
18 galaxies at $z \sim 0.6$ in the mass range dex(8.5)$\leq$M$_{*}$$\leq$dex(9.0).
Because the dispersion in the MZR is found to increase with decreasing M$_{*}$\
\citep{tremonti04,mannucci11},
a $\sigma$ of 0.16\,dex is a reasonable value,
given the broad parameter space covered by our dataset.
\subsection{Comparison with the FMR}\label{sec:compfmr}
\begin{table*}
\setlength{\tabcolsep}{3pt}
\begin{center}
\caption {FPZ and FMR applied to MEGA and SDSS datasets$^{\mathrm a}$}
{\scriptsize
\begin{tabular}{lcccccccp{0.15\textwidth}}
\hline
\multicolumn{1}{c}{Sample} &
\multicolumn{1}{c}{Calibration} &
\multicolumn{1}{c}{$\sigma_{\rm fit}^{\mathrm b}$} &
\multicolumn{1}{c}{Offset(fit)$^{\mathrm b}$} &
\multicolumn{1}{c}{$\langle$12$+$log(O/H)$\rangle$$^{\mathrm c}$} &
\multicolumn{1}{c}{$\langle$Log(SFR)$\rangle$$^{\mathrm c}$} &
\multicolumn{1}{c}{$\langle$Log(M$_{*}$)$\rangle$$^{\mathrm c}$} &
\multicolumn{2}{c}{12$+$log(O/H)\,= } \\
\multicolumn{1}{c}{(1)} &
\multicolumn{1}{c}{(2)} &
\multicolumn{1}{c}{(3)} &
\multicolumn{1}{c}{(4)} &
\multicolumn{1}{c}{(5)} &
\multicolumn{1}{c}{(6)} &
\multicolumn{1}{c}{(7)} &
\multicolumn{2}{c}{(8)} \\
\hline
\multicolumn{9}{c}{MEGA FPZ applied to the MEGA and SDSS10 datasets} \\
\hline
\\
MEGA & D02 & 0.185 & 0.03 & $8.422\,\pm\,0.33$ & $0.159\pm1.12$ & $9.440\,\pm\,0.93$ &
\multicolumn{2}{c}{$-0.17 \ s + 0.44 \ m + 4.33$} \\
MEGA & PP04N2 & 0.159 & 0.02 & $8.332\,\pm\,0.30$ & $0.159\pm1.12$ & $9.440\,\pm\,0.93$ &
\multicolumn{2}{c}{$-0.14 \ s + 0.37 \ m + 4.82$} \\
MEGA & PP04O3N2 & 0.178 & 0.03 & $8.355\,\pm\,0.32$ & $0.159\pm1.12$ & $9.440\,\pm\,0.93$ &
\multicolumn{2}{c}{$-0.16 \ s + 0.41 \ m + 4.48$} \\
\\
\hline
\\
SDSS10 & D02 & 0.102 & $-0.04$ & $8.768\,\pm\,0.09$ & $-0.038\pm0.43$ & $10.263\,\pm\,0.41$ &
\multicolumn{2}{c}{As above for D02.} \\
SDSS10 & PP04N2 & 0.080 & $-0.004$ & $8.660\,\pm\,0.11$ & $-0.038\pm0.43$ & $10.263\,\pm\,0.41$ &
\multicolumn{2}{c}{As above for PP04N2.} \\
SDSS10 & PP04O3N2 & 0.088 & $-0.01$ & $8.709\,\pm\,0.12$ & $-0.038\pm0.43$ & $10.263\,\pm\,0.41$ &
\multicolumn{2}{c}{As above for PP04O3N2.} \\
\\
\hline
\multicolumn{9}{c}{FMR$^{\mathrm d}$ applied to the MEGA and SDSS10 datasets} \\
&&&&&&& \multicolumn{2}{c}{12$+$log(O/H)\,=} \\
&&&&&&&
\multicolumn{1}{c}{$\mu_{0.32}<$9.5\,$M_\odot$$^{\mathrm e}$} &
\multicolumn{1}{c}{$\mu_{0.32}\geq$9.5\,$M_\odot$$^{\mathrm e}$} \\
\multicolumn{1}{c}{(1)} &
\multicolumn{1}{c}{(2)} &
\multicolumn{1}{c}{(3)} &
\multicolumn{1}{c}{(4)} &
\multicolumn{1}{c}{(5)} &
\multicolumn{1}{c}{(6)} &
\multicolumn{1}{c}{(7)} &
\multicolumn{1}{c}{(8)} &
\multicolumn{1}{c}{(9)} \\
\hline
\\
MEGA & D02 & 0.175 & $-0.14$ & $8.422\,\pm\,0.33$ & $0.159\pm1.12$ & $9.440\,\pm\,0.93$ &
$-0.16 \ s + 0.51 \ m + 3.83 $ & $-13.8 + 4.17 m - 1.34 s + 0.12 m\,s - 0.19 m^2 - 0.054 s^2$ \\
MEGA & PP04N2 & 0.168 & $-0.25$ & $8.332\,\pm\,0.30$ & $0.159\pm1.12$ & $9.440\,\pm\,0.93$ &
$-0.16 \ s + 0.51 \ m + 3.83 $ & $-13.8 + 4.17 m - 1.34 s + 0.12 m\,s - 0.19 m^2 - 0.054 s^2$ \\
MEGA & PP04O3N2 & 0.181 & $-0.22$ & $8.355\,\pm\,0.32$ & $0.159\pm1.12$ & $9.440\,\pm\,0.93$ &
$-0.16 \ s + 0.51 \ m + 3.83 $ & $-13.8 + 4.17 m - 1.34 s + 0.12 m\,s - 0.19 m^2 - 0.054 s^2$ \\
\\
\hline
&&&&&&& \multicolumn{2}{c}{12$+$log(O/H)\,=} \\
&&&&&&&
\multicolumn{1}{c}{$\mu_{0.32}<$10.2\,$M_\odot$$^{\mathrm e}$} &
\multicolumn{1}{c}{$\mu_{0.32}>$10.5\,$M_\odot$$^{\mathrm e}$} \\
\multicolumn{1}{c}{(1)} &
\multicolumn{1}{c}{(2)} &
\multicolumn{1}{c}{(3)} &
\multicolumn{1}{c}{(4)} &
\multicolumn{1}{c}{(5)} &
\multicolumn{1}{c}{(6)} &
\multicolumn{1}{c}{(7)} &
\multicolumn{1}{c}{(8)} &
\multicolumn{1}{c}{(9)} \\
\hline
\\
SDSS10 & D02 & 0.05 & $-0.25$ & $8.768\,\pm\,0.09$ & $-0.038\pm0.43$ & $10.263\,\pm\,0.41$ &
$-0.15 \ s + 0.47 \ m + 4.20$ & \multicolumn{1}{c}{$9.07$} \\
SDSS10 & PP04N2 & 0.05 & $-0.34$ & $8.660\,\pm\,0.11$ & $-0.038\pm0.43$ & $10.263\,\pm\,0.41$ &
$-0.15 \ s + 0.47 \ m + 4.20$ & \multicolumn{1}{c}{$9.07$} \\
SDSS10 & PP04O3N2 & 0.06 & $-0.28$ & $8.709\,\pm\,0.12$ & $-0.038\pm0.43$ & $10.263\,\pm\,0.41$ &
$-0.15 \ s + 0.47 \ m + 4.20$ & \multicolumn{1}{c}{$9.07$} \\
\\
\hline
\label{tab:fpzfmr}
\end{tabular}
}
\vspace{-1.5\baselineskip}
\begin{flushleft}
$^{\mathrm a}$~Applying the FPZ and the FMR to the entire mass range according to Eqn. \ref{eqn:fpzall}.
In the equations of Cols. (8,9), $s$ corresponds to Log(SFR) and $m$ to Log(M$_{*}$).
The sense of the residuals is 12$+$log(O/H)(data) $-$ 12$+$log(O/H)(FPZ,FMR).\\
$^{\mathrm b}$~Standard deviation $\sigma$ and offset of the FPZ or FMR residuals (e.g., Fig. \ref{fig:fpzall}).\\
$^{\mathrm c}$~Means and standard deviations of the samples. \\
$^{\mathrm d}$~We took the FMR for the MEGA sample from the extension to lower M$_{*}$\ by \citet{mannucci11},
and for SDSS10 from \citet{mannucci10}; for both we have converted their recipes with $\mu_{0.32}$ to
the multiplicative formulation as for the FPZ. \\
$^{\mathrm e}$~The divisions for the FMR are in $\mu_{0.32}$,
where $\mu_{0.32}\,\equiv\,$Log(M$_{*}$) $-$ 0.32\ Log(SFR) \citep{mannucci10}. \\
\end{flushleft}
\end{center}
\end{table*}
Table \ref{tab:fpzfmr} gives the mean residuals and offsets of the FPZ applied to the MEGA and SDSS10 datasets,
as well as of the FMR from \citet{mannucci10} applied to SDSS10, and the FMR extended to lower M$_{*}$\
by \citet{mannucci11} applied to the MEGA dataset.
Despite the vastly different parameter ranges over which the FPZ and FMR are
calibrated,
results from the Table show that the FPZ and the FMR are roughly equivalent in terms of
the width of the residuals, i.e., the accuracy of the approximation.
The FPZ fits the SDSS10 dataset almost as well as the FMR itself ($\sigma\,\approx\,$0.08\,dex),
and the FMR is reasonably good at reproducing the metallicities of the MEGA dataset.
However, the salient difference between the FPZ and FMR formulations is the negative O/H offsets
of the FMR; Col. (4) of Table \ref{tab:fpzfmr} shows that the FMR predicts metallicities
both for the MEGA dataset and for the recalibrated SDSS10 that can be in excess by as much as $\sim -0.3$\,dex.
A similar result was found by \citet{hunt12} relative to the FMR,
although the offset was larger, $\sim -0.4$\,dex, presumably because of the original SDSS10 KD02 calibration.
Indeed, the over-large metallicities predicted by the FMR are almost certainly due to the different O/H
calibrations as discussed in Sect. \ref{sec:ohcalib}.
\citet{cullen14} found a similar discrepancy of their observations at $z \sim 2$
with respect to the FMR using the same (KD02) calibration as \citet{mannucci10}.
Some groups concluded that the FMR evolves with redshift because of
its failure to fit galaxies at $z \sim 2-3$ \citep[e.g.,][]{steidel14,troncoso14}.
However, we find that the FPZ, unlike the FMR, is apparently invariant with redshift;
our result almost certainly stems from the common O/H calibration and its
similarity to the T$_{\rm e}$\ method by which strong-line methods are calibrated
at low metallicity.
Thus, it is of extreme importance to compare galaxies at different redshifts
with a common O/H calibration that is as accurate as possible at low metallicities,
and that smoothly connects these with the difficult intermediate-metallicity regime
and with higher metallicities nearer to or exceeding Solar.
\section{Metallicity and SFR coevolution}
\label{sec:coevo}
Much work has been done to establish how metal abundance and SFR vary with
redshift.
As mentioned in the Introduction,
the picture that emerges from these studies is that
the shape of the MZR is relatively invariant while the metallicity
for a given M$_{*}$\ decreases with increasing redshift
\citep[e.g.,][]{shapley05,cowie08,henry13,yabe14,erb06a,zahid12,steidel14,maiolino08,mannucci09,troncoso14,onodera16}.
At the same time,
it is well known that the SFMS also remains relatively constant in shape,
but at a given M$_{*}$, SFR (and sSFR) increases with redshift \citep[e.g.,][]{noeske07,karim11,speagle14}.
In the context of the FPZ, the relatively small dispersion of the residuals suggests that
the FPZ formulation is apparently invariant with redshift to
$z\sim$3.7, even with the new MEGA sample that more than
triples the number of galaxies at $z\ga2-3$ with respect to \citet{hunt12}.
\textit{Thus, under the hypothesis that the FPZ is maintained even at high $z$,
the opposing redshift trends of O/H and SFR must somehow be mutually compensated. }
For typical galaxy populations, at fixed M$_{*}$,
the increase of SFR with redshift must be accompanied by a corresponding decrease in O/H.
We can quantify such trends with the MEGA dataset and the FPZ.
Table \ref{tab:medians} reports the median values of M$_{*}$, 12$+$log(O/H), and sSFR
for the MEGA dataset for 6 mass bins within the 7 redshift bins shown in Fig. \ref{fig:redshift}.
\begin{table*}
\setlength{\tabcolsep}{3pt}
\caption{Median stellar masses, (PP04N2) O/H, and sSFR in MEGA redshift bins$^{\mathrm a}$}
\begin{tabular}{llcccc}
\hline
\multicolumn{1}{c}{Redshift bin} &
\multicolumn{1}{c}{Mass bin} &
\multicolumn{1}{c}{Number} &
\multicolumn{1}{c}{Log(M$_{*}$)} &
\multicolumn{1}{c}{12$+$log(O/H)} &
\multicolumn{1}{c}{Log(sSFR)} \\
&&&
\multicolumn{1}{c}{($M_\odot$)} &
\multicolumn{1}{c}{(PP04N2)} &
\multicolumn{1}{c}{($M_\odot$\,yr$^{-1}$)} \\
\hline
$z \leq\ 0.1$ & $\log({\mathrm M}_*)<8.5$ & 128 & 7.76$^{+0.31}_{-0.58}$ & 7.92$^{+0.12}_{-0.22}$ & -9.61$^{+0.28}_{-0.23}$\\
& $8.5\leq \log({\mathrm M}_*) < 9$ & 24 & 8.69$^{+0.11}_{-0.09}$ & 8.16$^{+0.10}_{-0.10}$ & -9.78$^{+0.60}_{-0.18}$\\
& $9\leq\ \log({\mathrm M}_*) < 9.5$ & 34 & 9.18$^{+0.12}_{-0.11}$ & 8.39$^{+0.09}_{-0.13}$ & -9.84$^{+0.46}_{-0.18}$\\
& $9.5\leq\ \log({\mathrm M}_*) < 10$ & 18 & 9.64$^{+0.15}_{-0.05}$ & 8.39$^{+0.19}_{-0.21}$ & -9.94$^{+0.30}_{-0.11}$\\
& $10\leq\ \log({\mathrm M}_*) < 10.5$ & 26 & 10.28$^{+0.12}_{-0.13}$ & 8.69$^{+0.08}_{-0.14}$ & -9.99$^{+0.16}_{-0.15}$\\
& $\log({\mathrm M}_*)>10.5$ & 27 & 10.71$^{+0.12}_{-0.11}$ & 8.75$^{+0.04}_{-0.03}$ & -10.36$^{+0.30}_{-0.62}$\\
\\
$0.1 < z \leq\ 0.4$ & $\log({\mathrm M}_*)<8.5$ & 1 & 8.09& 8.16 & -8.91\\
& $8.5\leq \log({\mathrm M}_*) < 9$ & 16 & 8.83$^{+0.08}_{-0.16}$ & 8.22$^{+0.08}_{-0.04}$ & -9.22$^{+0.11}_{-0.32}$\\
& $9\leq\ \log({\mathrm M}_*) < 9.5$ & 39 & 9.35$^{+0.09}_{-0.23}$ & 8.34$^{+0.10}_{-0.07}$ & -9.52$^{+0.27}_{-0.33}$\\
& $9.5\leq\ \log({\mathrm M}_*) < 10$ & 84 & 9.71$^{+0.12}_{-0.11}$ & 8.53$^{+0.11}_{-0.10}$ & -9.69$^{+0.30}_{-0.24}$\\
& $10\leq\ \log({\mathrm M}_*) < 10.5$ & 18 & 10.11$^{+0.09}_{-0.05}$ & 8.62$^{+0.06}_{-0.11}$ & -9.78$^{+0.25}_{-0.24}$\\
& $\log({\mathrm M}_*)>10.5$ & 6 & 10.57$^{+0.10}_{-0.04}$ & 8.60$^{+0.04}_{-0.09}$ & -9.72$^{+0.08}_{-0.05}$\\
\\
$0.4 < z \leq\ 0.7$ & $\log({\mathrm M}_*)<8.5$ & 9 & 8.38$^{+0.02}_{-0.16}$ & 8.19$^{+0.04}_{-0.03}$ & -8.58$^{+0.30}_{-0.36}$\\
& $8.5\leq \log({\mathrm M}_*) < 9$ & 12 & 8.82$^{+0.15}_{-0.13}$ & 8.25$^{+0.05}_{-0.05}$ & -9.02$^{+0.36}_{-0.17}$\\
& $9\leq\ \log({\mathrm M}_*) < 9.5$ & 49 & 9.34$^{+0.09}_{-0.07}$ & 8.35$^{+0.13}_{-0.07}$ & -9.45$^{+0.25}_{-0.14}$\\
& $9.5\leq\ \log({\mathrm M}_*) < 10$ & 84 & 9.69$^{+0.09}_{-0.08}$ & 8.44$^{+0.09}_{-0.08}$ & -9.41$^{+0.18}_{-0.21}$\\
& $10\leq\ \log({\mathrm M}_*) < 10.5$ & 12 & 10.12$^{+0.10}_{-0.04}$ & 8.59$^{+0.02}_{-0.11}$ & -9.37$^{+0.22}_{-0.12}$\\
& $\log({\mathrm M}_*)>10.5$ & 3 & 10.53$^{+0.03}_{-0.00}$ & 8.64$^{+0.05}_{-0.01}$ & -9.57$^{+0.15}_{-0.05}$\\
\\
$0.7 < z \leq\ 0.9$ & $\log({\mathrm M}_*)<8.5$ & 6 & 8.10$^{+0.18}_{-0.18}$ & 8.04$^{+0.12}_{-0.03}$ & -7.58$^{+0.46}_{-0.26}$\\
& $8.5\leq \log({\mathrm M}_*) < 9$ & 16 & 8.73$^{+0.12}_{-0.05}$ & 8.17$^{+0.18}_{-0.20}$ & -8.07$^{+0.39}_{-0.38}$\\
& $9\leq\ \log({\mathrm M}_*) < 9.5$ & 45 & 9.34$^{+0.11}_{-0.07}$ & 8.33$^{+0.11}_{-0.07}$ & -9.13$^{+0.43}_{-0.28}$\\
& $9.5\leq\ \log({\mathrm M}_*) < 10$ & 91 & 9.73$^{+0.08}_{-0.11}$ & 8.44$^{+0.11}_{-0.09}$ & -9.21$^{+0.27}_{-0.29}$\\
& $10\leq\ \log({\mathrm M}_*) < 10.5$ & 22 & 10.21$^{+0.08}_{-0.12}$ & 8.60$^{+0.07}_{-0.08}$ & -9.23$^{+0.27}_{-0.24}$\\
& $\log({\mathrm M}_*)>10.5$ & 20 & 10.74$^{+0.24}_{-0.08}$ & 8.69$^{+0.04}_{-0.05}$ & -9.54$^{+0.34}_{-0.26}$\\
\\
$0.9 < z \leq\ 1.8$ & $\log({\mathrm M}_*)<8.5$ & 0 & \multicolumn{1}{c}{$-$} & \multicolumn{1}{c}{$-$} & \multicolumn{1}{c}{$-$}\\
& $8.5\leq \log({\mathrm M}_*) < 9$ & 1 & 8.74& 7.49 & -7.98\\
& $9\leq\ \log({\mathrm M}_*) < 9.5$ & 2 & 9.20$^{+0.02}_{-0.02}$ & 8.08$^{+0.05}_{-0.05}$ & -8.88$^{+0.08}_{-0.08}$\\
& $9.5\leq\ \log({\mathrm M}_*) < 10$ & 14 & 9.81$^{+0.11}_{-0.17}$ & 8.41$^{+0.04}_{-0.06}$ & -8.81$^{+0.28}_{-0.17}$\\
& $10\leq\ \log({\mathrm M}_*) < 10.5$ & 14 & 10.24$^{+0.05}_{-0.11}$ & 8.52$^{+0.02}_{-0.03}$ & -8.67$^{+0.06}_{-1.06}$\\
& $\log({\mathrm M}_*)>10.5$ & 12 & 10.68$^{+0.17}_{-0.13}$ & 8.54$^{+0.07}_{-0.03}$ & -9.12$^{+0.31}_{-0.81}$\\
\\
$1.8 < z \leq\ 2.8$ & $\log({\mathrm M}_*)<8.5$ & 0 & \multicolumn{1}{c}{$-$} & \multicolumn{1}{c}{$-$} & \multicolumn{1}{c}{$-$}\\
& $8.5\leq \log({\mathrm M}_*) < 9$ & 2 & 8.74$^{+0.07}_{-0.07}$ & 8.12$^{+0.01}_{-0.01}$ & -7.22$^{+0.07}_{-0.07}$\\
& $9\leq\ \log({\mathrm M}_*) < 9.5$ & 5 & 9.28$^{+0.03}_{-0.10}$ & 8.25$^{+0.03}_{-0.05}$ & -7.69$^{+0.05}_{-0.10}$\\
& $9.5\leq\ \log({\mathrm M}_*) < 10$ & 33 & 9.69$^{+0.16}_{-0.10}$ & 8.24$^{+0.08}_{-0.02}$ & -8.17$^{+0.16}_{-0.30}$\\
& $10\leq\ \log({\mathrm M}_*) < 10.5$ & 24 & 10.18$^{+0.12}_{-0.07}$ & 8.28$^{+0.09}_{-0.07}$ & -8.76$^{+0.26}_{-0.26}$\\
& $\log({\mathrm M}_*)>10.5$ & 24 & 10.89$^{+0.25}_{-0.26}$ & 8.46$^{+0.03}_{-0.14}$ & -9.21$^{+0.20}_{-0.12}$\\
\\
$2.8 < z \leq\ 3.8$ & $\log({\mathrm M}_*)<8.5$ & 2 & 8.41$^{+0.02}_{-0.02}$ & 7.93$^{+0.01}_{-0.01}$ & -7.90$^{+0.26}_{-0.26}$\\
& $8.5\leq \log({\mathrm M}_*) < 9$ & 6 & 8.82$^{+0.12}_{-0.05}$ & 7.88$^{+0.22}_{-0.06}$ & -7.61$^{+0.11}_{-0.16}$\\
& $9\leq\ \log({\mathrm M}_*) < 9.5$ & 12 & 9.35$^{+0.07}_{-0.06}$ & 8.12$^{+0.04}_{-0.17}$ & -8.09$^{+0.23}_{-0.20}$\\
& $9.5\leq\ \log({\mathrm M}_*) < 10$ & 24 & 9.71$^{+0.14}_{-0.05}$ & 8.16$^{+0.01}_{-0.15}$ & -8.19$^{+0.18}_{-0.42}$\\
& $10\leq\ \log({\mathrm M}_*) < 10.5$ & 16 & 10.12$^{+0.18}_{-0.05}$ & 8.21$^{+0.05}_{-0.04}$ & -8.77$^{+0.29}_{-0.19}$\\
& $\log({\mathrm M}_*)>10.5$ & 8 & 10.75$^{+0.06}_{-0.14}$ & 8.19$^{+0.09}_{-0.02}$ & -8.86$^{+0.38}_{-0.31}$\\
\\
\hline
\label{tab:medians}
\end{tabular}
}
\vspace{-\baselineskip}
\begin{flushleft}
$^{\mathrm a}$~Medians of values within each redshift bin;
the upper and lower values correspond to the 75\% and 25\% quantile levels,
respectively. We have not considered the AMAZE galaxy, LnA1689$-$2, at $z\,=\,4.87$.
\end{flushleft}
\end{table*}
\subsection{Redshift variation of O/H and sSFR}
\label{sec:redshiftvariation}
\begin{figure*}
\vspace{\baselineskip}
\hbox{
\includegraphics[width=0.475\textwidth]{MEGA_OHPP04N2_newbins_individualfit_OnoderaLiuStackedYabeZahid_vsz-crop}
\hspace{0.3cm}
\includegraphics[width=0.475\textwidth]{MEGA_SSFR_newbins_individualfit_OnoderaLiuStackedYabeZahid_vsz-crop}
}
\caption{Binned measurements of 12$+$log(O/H)\ and sSFR as a function of redshift
using the MEGA dataset.
As in previous figures, the O/H calibration is PP04N2;
the mass bins are shown in the lower right-hand corner.
Error bars correspond to the 25th percentile of the vertical
parameter within each mass bin.
The curves are an approximation to the observed
trends obtained by adopting the formulation of \citet{karim11} based on
separable functions of M$_{*}$\ and $z$: sSFR, O/H $\propto\ M_*^\beta\ (1+z)^n$.
The mean power-law indices, $n$, averaged over all mass bins (except the lowest one) are:
$\langle n \rangle$({\rm O/H})\,=\,$-0.57\,\pm\,0.17$ for 12$+$log(O/H)\,$\propto (1+z)^n$ (left panel), and
$\langle n \rangle$({\rm sSFR})\,=\,$2.79\,\pm\,0.52$ for sSFR\,$\propto (1+z)^n$ (right).
}
\label{fig:ohssfrvsz}
\end{figure*}
The trends with redshift of the MEGA dataset are shown in
Fig. \ref{fig:ohssfrvsz} where 12$+$log(O/H)\ and sSFR are plotted vs. redshift $z$.
Following \citet{karim11}, we have fit the redshift variation with
separable functions in M$_{*}$\ and $z$:
$$
\mathrm{sSFR}, \mathrm{O/H} ({\mathrm M}_*,z) \propto {\mathrm M}_*^\beta\ (1+z)^n.
$$
\noindent
The dependence on stellar mass is encompassed in the power-law index $\beta$,
and the $z$ dependence in the power-law index (slope in log space), $n$.
The symbols in Fig. \ref{fig:ohssfrvsz} correspond to data binned in redshift and in stellar mass as given in
the legend (see Fig. \ref{fig:redshift} for redshift intervals);
the error bars give the 25\% quantiles of the data within each bin.
For the trend of sSFR\,$\propto (1+z)^{n({\rm sSFR})}$, the mean power-law index, $\langle n \rangle$(sSFR),
averaged over all mass bins (except the lowest one,
because of the lack of low-mass galaxies at $z \ga 1$) is:
$\langle n \rangle$(sSFR)\,=\,$2.8\,\pm\,0.5$.
For the M$_{*}$\ bins between dex(8.5) and dex(10)\,$M_\odot$,
$n$ is relatively constant: $3.1\,\pm\,0.37$.
This value is roughly consistent with $n({\rm sSFR})\sim 3.4-3.5$
to $z \sim 2$ as reported by \citet{oliver10,karim11}.
For larger M$_{*}$\ (M$_{*}$\,$\geq10^{10}$\,$M_\odot$), the index decreases to $n({\rm sSFR})\,=\,2.2-2.4$.
Similar slopes and such a flattening are also seen in the highly star-forming sub-sample of COSMOS galaxies
described by \citet{karim11}; the MEGA dataset probably represents a similarly highly star-forming
sample, at least at the higher redshifts.
The redshift variation of O/H is represented by
the mean power-law index $\langle n \rangle$({\rm O/H}), averaged over all mass bins (except
the lowest one as above):
$\langle n \rangle$({\rm O/H})\,=\,$-0.57\,\pm\,0.17$ for 12$+$log(O/H)\,$\propto (1+z)^{n({\rm O/H})}$ (left panel of Fig. \ref{fig:ohssfrvsz}).
From the relatively small standard deviation, and visually evident in Fig. \ref{fig:ohssfrvsz},
it is apparent that the index $n({\rm O/H})$ is much more constant over variations in M$_{*}$\ than
the equivalent index $n({\rm sSFR})$ for sSFR.
Table \ref{tab:ohvszbins} gives the fitted coefficients for the (PP04N2) metallicity redshift variation
of the MEGA dataset;
these are the equations describing the dashed curves in the left panel of Fig. \ref{fig:ohssfrvsz}
and subsequent figures.
\begin{table}
\setlength{\tabcolsep}{3pt}
\caption {Fitted (PP04N2) O/H redshift variation of the MEGA dataset$^{\mathrm a}$}
{\scriptsize
\begin{tabular}{crcccc}
\hline
\multicolumn{1}{c}{Mass bin} &
\multicolumn{1}{c}{Number} &
\multicolumn{1}{c}{$a$} &
\multicolumn{1}{c}{$b$} &
\multicolumn{1}{c}{Median} &
\multicolumn{1}{c}{RMS} \\
\multicolumn{1}{c}{log($M_\odot$)} & \multicolumn{1}{c}{redshift} & & &
\multicolumn{1}{c}{residual} &
\multicolumn{1}{c}{$\sigma$} \\
& \multicolumn{1}{c}{points} \\
\hline
$8.5-9.0$ & 77 & $8.23\,\pm\,0.03$ & $-0.33\,\pm\,0.11$ & ~$0.010$ & $0.18$\\
$9.0-9.5$ & 186 & $8.43\,\pm\,0.02$ & $-0.45\,\pm\,0.07$ & ~$0.002$ & $0.15$\\
$9.5-10.0$ & 349 & $8.59\,\pm\,0.01$ & $-0.64\,\pm\,0.05$ & $-0.002$ & $0.14$\\
$10.5-10.5$ & 132 & $8.72\,\pm\,0.02$ & $-0.75\,\pm\,0.05$ & ~$0.012$ & $0.13$\\
$>10.5$ & 100 & $8.78\,\pm\,0.02$ & $-0.68\,\pm\,0.06$ & ~$0.005$ & $0.14$\\
\hline
\label{tab:ohvszbins}
\end{tabular}
}
\vspace{-\baselineskip}
\begin{flushleft}
$^{\mathrm a}$~Coefficients for robust fits to the individual data
points (within each mass bin) of the
form 12$+$log(O/H)\,=\,$a$\ + $b\,\log(1+z)$; $b$ corresponds to $n$(O/H) as
described in the text.\\
\end{flushleft}
\end{table}
\begin{figure*}
\vspace{\baselineskip}
\hbox{
\includegraphics[width=0.475\textwidth]{MEGA_OHPP04N2_newbins_individualfit_OnoderaLiuStackedYabeZahid_vsMstar-crop}
\hspace{0.3cm}
\includegraphics[width=0.475\textwidth]{MEGA_SSFR_newbins_individualfit_OnoderaLiuStackedYabeZahid_vsMstar-crop}
}
\caption{Binned measurements of 12$+$log(O/H)\ and sSFR as a function of M$_{*}$\
using the MEGA dataset.
As in previous figures, the representative O/H calibration is PP04N2;
the redshift bins are shown in the corners of the figures (see also Fig. \ref{fig:redshift}).
Error bars correspond to the 25th percentile of the vertical
parameter within each redshift bin.
As in Fig. \ref{fig:ohssfrvsz}, the curves are an approximation to the observed
trends obtained by adopting the formulation of \citet{karim11} based on
separable functions with the redshift trend given by $(1+z)^n$, and the
M$_{*}$\ variation by $M_*^\beta$.
The mean power-law indices, $\beta$, averaged over all redshift bins are:
$\langle \beta \rangle$({\rm O/H})\,=\,$0.21\,\pm\,0.05$ for 12$+$log(O/H)\,$\propto$\,M$_{*}$$^{\beta({\rm O/H})}$ (left panel), and
$\langle \beta \rangle$({\rm sSFR})\,=\,$-0.51\,\pm\,0.24$ for sSFR\,$\propto$\,M$_{*}$$^{\beta({\rm sSFR})}$ (right).
}
\label{fig:ohssfrvsmstar}
\end{figure*}
\subsection{Variation of O/H and sSFR with stellar mass}
\label{sec:massvariation}
We now examine the M$_{*}$\ dependence of the redshift variations of O/H and sSFR.
Fig. \ref{fig:ohssfrvsmstar} shows 12$+$log(O/H)\ vs. M$_{*}$\ (left panel) and
sSFR vs. M$_{*}$\ (right), binned into different redshift bins;
these are equivalent to the changes with redshift of the MZR (left panel) and the SFMS (right).
Here we assess $\beta$ in the separated formalism as above:
sSFR, O/H $\propto M_*^\beta\ (1+z)^n$.
$\beta$(O/H) corresponds roughly to the slope (power-law index) of the MZR,
and $\beta$(sSFR) to the slope of the SFMS.
For the M$_{*}$\ trend of O/H, $\langle \beta \rangle$\ (averaged over all mass bins) is:
$\langle \beta \rangle$(O/H)\,=\,$0.21\,\pm\,0.05$.
Both the normalization and the slope of the MZR are rather constant up to $z \sim 2$:
$\beta\,=\,0.24\,\pm\,0.03$,
12$+$log(O/H)(M$_{*}$\,=\,dex(9)\,$M_\odot$)\,=\,$8.28\,\pm\,0.04$
(averages and standard deviations over 5 equally-weighted redshift bins for $z < 1.8$).
However, for $z > 2$, both the normalization and the slope gradually decrease:
$\beta\,=\,0.15\,\pm\,0.03$ and 12$+$log(O/H)(M$_{*}$\,=\,dex(9)\,$M_\odot$)\,=\,$8.01\,\pm\,0.03$
within the highest ($z \sim 3$) redshift bin.
The slope $\beta$(O/H) for the MEGA dataset at $z \sim 1.4$ of $0.23\,\pm\,0.02$ is slightly
steeper than that by \citet{yabe14} who find $\beta$\,=\,0.15,
but shallower than the slope of $\sim 0.3$ found by \citet{liu08} using PP04N2 in a similar
redshift range (both samples are also included in the MEGA dataset).
The steeper slopes we find relative to \citet{yabe14} can be attributed to their use
of a Salpeter IMF, rather than the \citet{chabrier03} IMF used here.
In the redshift range $z \sim 0.5-0.9$, \citet{cowie08} find an MZR slope
of $0.13-0.17$ using the KK04 and T04 O/H calibrations and a Salpeter IMF;
this is also somewhat shallower than the PP04N2 $\beta$\,=\,$0.24\,\pm\,0.02$ in a similar redshift range,
but with the Chabrier IMF.
The steeper slopes we find for the MEGA dataset are consistent with those
found by \citet{zahid11} for DEEP2 galaxies at $z \sim 0.8$ using the KK04 O/H calibration
(and a Chabrier IMF).
As discussed by \citet{zahid11}, differences in fitting procedures are an
important consideration in comparing slopes of the MZR,
but the O/H calibration is also important.
\citet{kewley08} illustrate that both the slope (at the low-mass end) and
the absolute O/H determination depend strongly on the calibration.
Thus, the consistency of the MZR slopes $\beta$ relative to previous
work lends confidence our approach.
For the trends of sSFR with M$_{*}$, corresponding to the SFMS,
averaging over all redshifts gives $\langle \beta \rangle$(sSFR)\,=\,$-0.51\,\pm\,0.24$
(average and standard deviation over the individual redshift bins);
the mean slope is poorly determined because of the (possibly spurious, see below)
steepening toward high $z$.
At $z \simeq 0$,
we find $\beta$(sSFR)$\,=\,-0.21\,\pm\,0.025$\footnote{This slope is derived
from all data at $z \simeq 0$, while the slope of $-0.19\,\pm\,0.02$ in Sect. \ref{sec:scaling}
is found from the LVL$+$KINGFISH galaxies only; the two slopes
are in good agreement.}.
At $z\sim0.25$, we find a steeper slope, $\beta$(sSFR)$\,=\,-0.29\,\pm\,0.07$, roughly consistent
with the sSFR vs. M$_{*}$\ power-law index of $\sim -0.4$ estimated by \citet{karim11} and by \citet{speagle14}
for $z \approx 0.3$.
At $z \sim 3$, $\beta$(sSFR)$\,=\,-0.64\,\pm\,0.09$;
the observed steepening of $\beta$(sSFR) toward
higher redshift evident in Fig. \ref{fig:ohssfrvsmstar} is inconsistent with
the results of \citet{speagle14} who
find steeper slopes with increasing redshift in (log) SFR vs. M$_{*}$, corresponding to
shallower slopes in (log) sSFR.
Indeed, the MEGA dataset does not show clear evidence for a SFMS within
individual redshift bins (see Fig. \ref{fig:ms}); this could be because the galaxies at higher redshift are selected
basically for a constant SFR (see Table \ref{tab:samples}) rather than
a selection based on M$_{*}$.
Such a selection can result in a basically flat trend of SFR with M$_{*}$\
\citep[e.g.,][]{erb06b,lee13,renzini15},
which produces a steep dependence of sSFR with M$_{*}$, sSFR$\,\propto$\,M$_{*}$$^{-1}$,
similar to the behavior of the MEGA dataset at high $z$.
This is essentially a Malmquist bias since at low stellar masses, only
galaxies with relatively high SFR are selected.
Because the MEGA dataset requires emission lines in order to measure
metallicity spectroscopically, such an effect almost certainly
plays an important role \citep[e.g.,][]{juneau14}.
Ultimately, because of such selection effects,
the MEGA dataset may not be completely representative of the
SFR-M$_{*}$\ correlations at high redshift.
Nevertheless, it is the best dataset currently available for assessing
the evolution of the MZR.
\subsection{Redshift invariance of the FPZ?}
\label{sec:fpzinvariance}
One way to assess
the redshift invariance of the FPZ is by comparing the coefficients
for redshift variation discussed above in Sect. \ref{sec:redshiftvariation}
with those for the FPZ.
We thus performed multi-variable linear regressions on the MEGA dataset
for 12$+$log(O/H)\ as a function of M$_{*}$\ and redshift, and the same for sSFR.
Performing a robust fit\footnote{For all statistical calculations we use R,
a free software environment for statistical computing and graphics,
{\it https://www.r-project.org/}.},
we find (for the PP04N2 calibration):
\begin{equation}
12+\log({\rm O/H})\,=\,0.27\,\log({\rm M}_*) - 0.59\,\log(1+z) + 5.89
\label{eqn:ohvsz}
\end{equation}
\noindent
with a residual standard error of $\sim$0.15\,dex,
and
\begin{equation}
\log({\rm sSFR})\,=\,-0.29\,\log({\rm M}_*) + 2.88\,\log(1+z) - 7.16
\label{eqn:ssfrvsz}
\end{equation}
\noindent
with a residual standard error of $\sim$0.44\,dex,
The FPZ in Eqn. (\ref{eqn:fpzall}) can be expressed as a function of sSFR,
rather than SFR:
\begin{align}
12+\log({\rm O/H}) & = -0.14\,\log {\rm (sSFR)} + 0.23\,\log ({\rm M_*)} + 4.82
\label{eqn:fpzallssfr}
\end{align}
By inserting the redshift variation of sSFR given by Eqn. (\ref{eqn:ssfrvsz}) in
Eqn. (\ref{eqn:fpzallssfr}) for the FPZ, and comparing it with the redshift
variation of O/H given by Eqn. (\ref{eqn:ohvsz}), we can
compare the resulting difference equation term by term and assess the redshift invariance of the FPZ;
we would expect $\approx 0$ in such a case.
The resulting coefficient for the difference (FPZ $-$ $z$ fits) of the log(M$_{*}$) term is $0.0004$, consistent with 0.
For log($1+z$), we find a difference of $0.178$, roughly consistent with 0 to within the residual standard errors of the fits.
The resulting difference for the constant term is $-0.051$, again consistent with 0 within the standard errors.
Although this result pertains to PP04N2, similarly small difference coefficients are
obtained for the D02 and PP04O3N2 calibrations.
\textit{We thus conclude that the FPZ is approximately redshift invariant to within 0.15$-$0.16\,dex}
(see Sect. \ref{sec:fp}); for typical galaxy populations
the increase of sSFR with redshift is compensated by the decrease in O/H \citep[although see][]{wuyts14}.
Nevertheless, the log($1+z$) difference coefficient of $\sim 0.18$\,dex is slightly larger
than would be expected given the residuals of the (PP04N2) FPZ, $\sim 0.16$\,dex.
This implies that the FPZ is not a perfect formulation of the redshift evolution of metallicity.
Therefore, to investigate the amplitude of the residual trend with redshift, in
Fig. \ref{fig:fpresidualsvsz} we have plotted the FPZ O/H residuals vs. redshift.
The trend (shown by the solid line) is significant with:
\begin{align}
12+\log({\rm O/H}) - {\rm FPZ(PP04N2)} & = \nonumber \\
(-0.048\,\pm\,0.006)\,z + 0.047\,\pm\,0.007
\label{eqn:fpzresvsz}
\end{align}
\noindent
This implies that at $z\sim 3.5$, the FPZ predicts metallicities 12$+$log(O/H)\ that are
roughly $0.17$\,dex too large.
However, the residual standard error of the (PP04N2) fit in Eqn. (\ref{eqn:fpzresvsz}) is $0.16$\,dex,
equivalent to the residuals of the FPZ itself;
indeed, the spread of residuals at $z\sim0$ is as large or larger than the spread of residuals
at $z\ga3$.
Thus, while the current data suggest that the FPZ may not fully describe metallicity
evolution (or its lack thereof related to SFR), the discrepancies are within
the overall noise in the estimation.
The data at $z\ga 3$ are still relatively sparse, however, and more data
with accurate metallicity measurements should help in confirming (or refuting)
this conclusion.
\begin{figure}
\vspace{\baselineskip}
\hbox{
\includegraphics[width=0.47\textwidth]{MEGAAllResiduals_vsz_FPZMEGA_PPO4N2-crop.pdf}
}
\caption{FPZ residuals for the MEGA dataset with the PP04N2 calibration
as a function of redshift.
The color scale corresponds to density of data points, and the contours show
the full dataset.
Individual points are also plotted for $z\geq1$.
The horizontal dashed line guides the eye for zero residuals, while
the solid line shows the best-fit robust regression given by Eqn. \ref{eqn:fpzresvsz}.
}
\label{fig:fpresidualsvsz}
\end{figure}
\section{Discussion and summary}
\label{sec:discussion}
The FPZ presented in Sect. \ref{sec:fp} is based on the
hypothesis that the curvature in the MZR at high stellar masses is compensated by the inflection
in the SFMS.
Our results show that this hypothesis is reasonably good, at least to within $\sim$0.16\,dex in 12$+$log(O/H).
Comparison of the FPZ with the FMR shows that both formulations adequately represent
the mutual correlations of M$_{*}$, SFR, and O/H, but
also that the O/H calibration is crucial;
applying the FMR to arbitrary samples can result in metallicity offsets as large as $-0.2$ to $-0.3$\,dex,
compared with the three O/H calibrations considered here.
Some groups have concluded that at a given redshift
or over a narrow range of redshifts, that the MZR does not depend on SFR
\citep[e.g.,][]{wuyts14,sanders15}.
However,
the parameters of the MEGA sample in specific redshift intervals, given in Table \ref{tab:medians}, are
comparable to the other samples used to draw these conclusions.
It is likely that the broad parameter space spanned by the MEGA dataset
contribute to the differences in the outcome.
The PCA analysis presented here could also play a role;
indeed, if we fit the MEGA data with a simple multi-variable regression
of 12$+$log(O/H)\ with respect to
M$_{*}$, SFR, and redshift, we find very little dependence of O/H on SFR.
This is because of the strong dependence of SFR (and sSFR) on redshift
through the increasing normalization of the SFMS (e.g., Sect. \ref{sec:coevo}).
The mutual correlations of the variables underlying the FPZ must be
taken into account for any analysis considering the MZR and its dependence on SFR.
In conclusion, we have compiled a new MEGA dataset consisting of $\sim$1000 galaxies taken
from 19 individual samples spanning a wide range of stellar masses, SFRs, and
metallicities and covering redshifts from $z\simeq 0$ to $z\sim 3.7$. In
addition to larger numbers of high-$z$ galaxies, the main improvement of this
dataset over that of \citet{hunt12} is the common O/H calibrations derived for
the MEGA galaxies. The main results are as follows:
\noindent
\hangindent=0.05\linewidth
\hangafter=1
$\bullet$\ \ After examining the mutual correlations among these parameters,
a PCA of the MEGA dataset shows that the 3D parameter space can be described
by a plane, dubbed the FPZ.
\noindent
\hangindent=0.05\linewidth
\hangafter=1
$\bullet$\ \
The functional form of the (PP04N2) FPZ is given by $12+\log({\rm O/H}) = -0.14\,\log
{\rm (SFR)} + 0.37\,\log ({\rm M_*)} + 4.82$ over the entire mass and redshift range of the MEGA dataset.
\noindent
\hangindent=0.05\linewidth
\hangafter=1
$\bullet$\ \
The mean O/H residuals of the FPZ over the MEGA dataset are 0.16\,dex
(for the PP04N2 calibration, slightly larger for D02 and PP04O3N2);
such residuals are
smaller than those found previously, and consistent with trends found in
smaller galaxy samples with more limited ranges in M$_{*}$, SFR, and O/H.
\noindent
\hangindent=0.05\linewidth
\hangafter=1
$\bullet$\ \
The FPZ is also found to be roughly invariant with redshift enabling an estimation of
metallicity accurate to within 0.16\,dex over roughly 5 orders
of magnitude in M$_{*}$,
from $\ga 10^{6}$\,$M_\odot$\ to $\sim 10^{11}$\,$M_\odot$,
up to $z\sim 3.7$.
An additional correction for redshift may be employed to increase
slightly the accuracy of O/H estimates from the FPZ
for $z\ga 2$ (see Eqn. \ref{eqn:fpzresvsz}).
\section*{Acknowledgments}
We acknowledge the anonymous referee whose insightful comments greatly improved the paper.
We also thank G. Cresci for passing us the SDSS10 (KD02) data in electronic form, and are grateful to the DAVID network
({\it http://wiki.arcetri.astro.it/DAVID/WebHome}) for fostering a fruitful collaborative environment.
PD gladly acknowledges funding from the EU COFUND Rosalind Franklin program.
|
\section{Introduction}
Termites are protagonists of processes which are crucial to ecological functioning and whose understanding would unveil puzzles still persistent in evolutionary theory. Specifically, the fact that termites live symbiotically in familiar groups (``colonies'') comprising a highly integrated society, poses them at the upper end of the series of transitions characterizing organic evolution \cite{Maynard-Smith.Szathmary.1997.majortransitionsin} (Table~\ref{tab:transitions}).
Simultaneously, termites often establish symbioses with other organisms (even other termite species) which cohabit termite nests along with their building colony \cite{Florencio.etal.2013.Diet
. This makes a termite nest a hotspot of diversity, with important consequences for ecosystem functioning \cite{Pringle.etal.2010.Spatial}. It also makes a termite nest a hotspot of evolutionary processes, varying from inter-specific symbiotic relationships to full sociality (this latter being a form of symbiosis occuring at the intraspecific level).
Despite diverse in output, many of these processes are based on intra- or interspecific interactions occurring between individuals which cohabit the same termitarium. Being interindividual in nature, such processes may involve emergent phenomena and hence call for analytical solutions provided by computing modelling, as opposed to classical biological methods of analysis.
Here we provide selected examples on how can computing tools and modelling help better understanding of ecological and evolutionary processes through the study of termite societies and their associated cohabitants. Such examples will involve computer-aided tracking and path analyses, cellular automata simulations, the analyses of emergent properties in networks, and the use of genetic algorithms and artificial life simulations.
\subsection{What are termites?}
Termites are six-legged social animals, majorly tropical, but also present in many temperate countries. They belong to the same order as the coackroaches (Blattodea) and comprise the phylogenetically cohesive infraorder Isoptera \cite{Krishna.etal.2013}. Bees, ants and wasps, which are also social insects, belong to another order, the Hymenoptera.
Being social, termites live in colonies within nests built by themselves. Such colonies are a familiar group composed by both female and male parents plus their male and female offspring. Having a society composed by male and female members is the biggest difference between termites and other social insects such as ants and bees. In these latter, most members are females (males are short-lived and are produced sazonally). Other differences exist and we shall introduce them along the text, as needed.
\subsection{Why are termites important?}
Termites are quite important for both, practical and theoretical reasons. Although mostly known for their pest status, this is not their sole or their most important trait. Termites are among the rare animals which are able to digest celullosis and this means that they are ultimately responsible for transforming organic carbon into its inorganic form, which is the major source of carbon used by plants. By possessing their own endogenous cellulolitic enzimes in addition to cellulolitic gut symbionts, termites can oxidise nearly 99\% of the carbon they intake \cite{Slaytor2000Energy}, being then among the few animals to take active part in the carbon cycle in earth. Maybe more importantly, they prevent excessive accumulation of dead plant material \cite{DeSouza.etal.2009} and dung from large herbivores \cite{Freymann2008} on the Earth surface . It is worth mentioning that, if we copy termite’s diggestion process we can be more efficient in producing carbon-free fuel because we will be using, \textit{e.g.}, sugar but the whole of the sugarcane to produce ethanol \cite{ScharfBoucias2010Potential}.
The importance of termites to ecological process on Earth is stressed by their huge biomass: it is estimated that termites along with ants comprise a third of all Amazonia's animal biomass \cite{Fittkau.Klinge.1973}. A similar picture applies to termites in African savanna. Because termites are small their metabolism is faster than, e.g., that of a vertebrate. This would make their astonishing biomass even more significant, as per unit weight and time, termites are prone to process more materials than vertebrates.
In order to shelter their colonies, termites build nests which may vary from simple excavations in wood or soil to elaborate buildings on the soil surface or on trees. In doing so, termites modify the surrounding environment, allowing the estabilishment of other plant and animal species, either in the vicinity of the nest, on its external walls, or even inside the whole structure. This role as an ``ecosystem engineer" percolates to positively impact the animal and vegetal primary productivity of the region \cite{Pringle.etal.2010.Spatial}. In summary, termites are central do ecological processes on Earth.
The relevance of termites to theoretical insights into biology stems from the fact that these are the oldest social insect \cite{Grimaldi.Engel.2005.Evolution}, providing therefore, primordial views on the processes ruling the origin and maintenance of sociality. This is quite important, because sociality is considered one of the last transitions in evolution \cite{Maynard-Smith.Szathmary.1997.majortransitionsin}.
Sociality in termites (along with that in all ants, some bees, wasps, and a few other animals) is of a special kind because it can be posed at the highest level of all forms of colonial organisation. In such cases, sterile offspring cohabit with and cooperatively help their parents to raise fertile offspring, and this is called ``eusociality'' \cite{Wilson.1971.InsectSocieties}. This division of reproductive labour estabilishes within the colony distinct groups of individuals
which share similar behavioural and sometimes morphological traits, the so called ``castes''. Non-reproductive castes are, in termites, comprised by ``workers'' and ``soldiers''. The former are responsible for nest building, maintenance, foraging, etc, while the soldiers, as the name implies, take care of nest defense. In most termites, such sterile castes are irreversible, which means that the members of a termite society are truly dependent on each other for survival and reproductive success.
\section{The paradox of eusociality}
Eusociality presents an evolutionary paradox \cite{Miramontes.DeSouza.2014.Social}: if adaptive evolution is driven by differential reproduction of individuals, how can sterile members of a social insect colony evolve and persist? Darwin himself commented on these sterile individuals saying that [...] \textit{they represent one special difficulty, which at first appeared to me insuperable, and actually fatal to the whole theory}. As a solution, he proposed that queens which were able to produce cooperative sterile offspring in addition to ordinary fertile ones, would succeed better than those producing only non-cooperative fertile offspring, as these latter queens would not profit from the synergism inherent to cooperative work.
Nearly a hundred years after, in 1964, an elegant mathematical formalism aimed to solve the riddle was proposed by W. D. Hamilton. It consisted of the so-called ``kin selection'', sometimes also referred to as “inclusive fitness”. In essence, kin selection predicts that infertile individuals, by cooperating with fertile family members, warrant the transmission of their own genes indirectly, when their kin reproduce. This, of course, will only work if sterile helpers are closely related to their fertile nestmates.
Haplodiploidy in Hymenoptera (bees, ants, wasps), where males are haploid and females are diploid, seem a key to this puzzle because it fulfills the condition of high relatedness among nestmates. A hymenopteran female can share 75\% of its genes with her nestmate sisters. Haplodiploidy, therefore, warrants higher levels of kinship between females, which, even if abstaining reproduction and helping their mother to raise reproductive sisters, would transfer to the next generation more genes than they could do by direct reproduction.
The higher levels of kinship among haplodiploids can be easily spotted by a simple calculation. Consider a fully heterozygous haplodiploid cross:
\begin{center}
\begin{tabular}{l|r}
& B\\\hline
A & AB\\
a & aB\\\hline
\end{tabular}
\end{center}
Because in Hymenoptera diploids are females, all offspring produced from this cross is female (males are produced parthenogenetically). If we then inspect the degree of relatedness between these sisters we will get:
\begin{center}
\begin{tabular}{l|p{0.7cm}p{1cm}}
sisters & AB & aB\\\hline
AB & 1.0 & 0.5\\
aB & 0.5 & 1.0\\\hline
\end{tabular}
\end{center}
That is, in average, sisters are related to each other by: $(1.0+0.5+0.5+1.0)/4=0.75$%
Haplodiploidy, however, is not sufficient to explain the evolution of eusociality. On one hand, there are lots of solitary hymenopteran species which are also haplodiploid, which means that haplodiploidy does not lead necessarily to eusociality. On the other hand, and maybe more importantly, eusociality also occurs among diplo-diploid organisms (those in which both, males and females, are diploids). Among these we find all termite species (\textit{c.a.} 3,000) plus aphids, beetles, shrimps, and naked mole rats -- not to mention humans which may also be said eusocial under certain circumstances. Fully diploid organisms would profit more from their own reproduction than that of their relatives, since they are in average only 50\% akin, as detailed below. In fully heterozygous diplo-diploid cross the following offspring can be produced:
\begin{center}
\begin{tabular}{l|rp{0.2cm}r}
& B && b\\\hline
A & AB && Ab\\
a & aB && ab \\\hline
\end{tabular}
\end{center}
Siblings from such a cross will exhibit the following degree of relatedness:
\begin{center}
\begin{tabular}{l|p{0.7cm}p{0.7cm}p{0.7cm}p{0.7cm}}
siblings & AB & Ab & aB & ab \\\hline
AB & {1.0} & {0.5} & {0.5} & {0.0} \\%\hline
Ab & {0.5} & {1.0} & {0.0} & {0.5} \\%\hline
aB & {0.5} & {0.0} & {1.0} & {0.5} \\%\hline
ab & {0.0} & {0.5} & {0.5} & {1.0} \\\hline
\end{tabular}
\end{center}
In such case, therefore, the average relatedness between siblings is: $$((1*4)+(0.5*8))/16=0.50$$%
The reasons for the existence of steriles among diplo-diploids, therefore, remain unclear. In other words, we are still awaiting for a convincing explanation on why termites are eusocial and at the same time diploid \cite{Thorne.1997}.
\section{A fertile terrain for computing models and tools}
\subsection{Intraspecific interactions and social behaviour}
Termites are, thus, central to the theory of evolution because: (i) they are social and hence are posed at the pinnacle of evolutionary processes and (ii) their sociality is not easily explained. Why are them social, being not haplo-diploid?
To answer this question, computing models come in very handy. Being hemimetabolous, termites are highly mobile since birth, unlike bees, ants, or wasps, in which juveniles are grub-like and hence almost sessile. They are also highly interactive and excitable: they exchange food and care among themselves and are able to pass information from one to another member of the colony.
Termites, hence, seem autonomous interactive mobile agents and as such possess traits highly prone to generate group-level dynamics. An immediate hypothesis arises from this: eusociality can be an emergent phenomenon, arising form interindividual interactions among colony members \cite{Miramontes.DeSouza.2014.Social}. That is, eusociality could stem from group-level phenomena. If this is so, termites could profit from collective behaviour per se, offsetting small genetic advantages inherent to diplo-diploidy. Sterile castes would, then, profit from cooperation and group living, surviving longer and hence getting more time to pass their genes on indirectly.
Evidence points that this is likely: an experiment has shown that survival of termites depends on group size \cite{DeSouza.etal.2001} (Fig.~\ref{fig:survival-grouped}). This survival is ``social facilitated'' (Fig.~\ref{fig:survival-automata}), closely resembling the ``survival'' of mobile cellular automata also modelled in groups of varying size and governed by two simple rules only: (i) individual energy decays with time and (ii) when moving around, a automata can recover some energy by interacting to another \cite{Miramontes.DeSouza.1996}.
It has been also showed that interindividual interactivity at optimal density is key to this survival: under low density, termites are more mobile and survive longer and these experimental procedures reveal the presence of phase transitions in termite social behaviour \cite{DeSouza.Miramontes.2004,Miramontes.DeSouza.2008}.
If mobility is key to interactivity and this latter is adaptive, one should expect mobility patterns of termites to follow an optimized search strategy. This is precisely what we observed in movement patterns of a individual termites: a Lévy-like route was established \cite{Miramontes.etal.2014}. We described exploratory spatial behaviour in isolated termite workers kept in large containers, free from the constrained movements they experience within tunnels (Fig.~\ref{fig:termite-path-msd}). In this way we were able to assess individual free exploratory behaviour in clueless environments and away from social interactions. We concluded that their searching patterns are compatible with scale-free strategies based on a fractal exploration of space and that these are key to the efficient flow of information between nestmates, thereby providing expressive hints on how self-organization underlies social cohesion.
In conclusion, simple repeated interactions between individuals can produce complex adaptive patterns at the level of the group. Eusociality may, indeed, be an emergent phenomenon.
\subsection{Interspecific interactions and termitophily}
Because interactivity and awareness-of-the-other lie in the very heart of organic evolution, a deeper insight tempting arises. The awareness of the other has been proposed to be one of the traits helping organisms to cross the barrier to sociality and eusociality \cite{Wilson.2012}. In fact, one of the traits common to all transistions in evolution \cite{Maynard-Smith.Szathmary.1997.majortransitionsin} is that entities come about together to form other entities: molecules compose chromosomes which compose cells, and these compose multicellular organisms, which compose colonies. Moreover, many complex entities we see today are in fact a combination of distinct ones: our cells are in fact the product of the engulfing of once free-living prokaryotes by a primitive eukaryotic cell. This prokariotes are now what we know as ``mitochondria''\cite{Margulis.Sagan.1990.Origins}, which is now unable to live outside the the cell. A similar scenario applies to chloroplasts in plants.
Maybe not surprisingly, termites provide us with a macroscopic example of such strict integration: in the nest of many termite species, other termite species and other invertebrates coexist \cite{Cristaldo.etal.2012,Marins.etal.2016.Termitecohabitationrelative} and this is called ``termitophily''. Moreover, some do it in an obligatory way: they can not survive outside such nests \cite{Florencio.etal.2013.Diet}. Understanding such symbiosis may help us to unveil the process by which termites deal with strangers and, ultimately, the process by which two distinct entities come about to form single one.
Again, computing models may be good tools to reveal the underlying forces. One of such approaches is the use of network analysis to reveal hidden termite-termite interactions. We contrasted the emergent network properties of multispecies termite communities cohabting the same termitarium with the properties of detail-rich reference communities with known modes of interaction \cite{Campbell.etal.2016.Top} (Fig.~\ref{fig:network-termitophile}). The studied termite networks overlapped more closely with mutualistic plant–pollinator communities than to antagonistic host–parasitoid communities. The analysis raised the hypothesis that termite–termite cohabitation networks may be overall mutualistic. More broadly, it provided support for the argument that cryptic communities may be analyzed via comparison to well-characterized communities.
Another approach is to model, through genetic algorithms and artificial life simulations, how do inquilines in termites manage to cohabit the nest along with the building species? Why do host colonies tolerate intruders? What kind of evolutionary pressures can explain the origin of this phenomenon that apparently benefits only the inquilines? These issues are difficult to investigate using classic methods at disposal of biologists. Computer simulation models can help to test and generate new hypothesis that can be subsequently verified with targeted experiment in nature. A recent collaboration between Federal University of Viçosa and Aberystwyth University has started aiming to develop studies using such an approach.
\section{Conclusion}
Symbiotic processes occurring at the intra- and inter-specific level within termitaria area key to understand basal mechanisms leading to the major transitions in evolution. Because such processes involve interindividual interactions and may result and emergent phenomena, their full understanding may be better achieved by the use of computing tools and models rather than classical biological methods of analysis.
\section*{Acknowledgments}
This is an excerpt of a talk given by ODS at the 14th International Conference on the Simulation of Adaptive Behavior, at Aberystwyth, Wales, on 23-26 Aug 2016. Since it is intended as a review, this text stems from previous works by the authors and, in particular, it draws heavily from \cite{Campbell.etal.2016.Top,DeSouza.etal.2001,Florencio.etal.2013.Diet,Miramontes.DeSouza.1996,Miramontes.DeSouza.2014.Social,Miramontes.etal.2014}. The work described in this paper has been supported by The Brazilian National Council for Research Development (CNPq), the Coordination for the Improvement of Higher Personnel (CAPES), Minas Gerais State Foundation for Research Support (FAPEMIG) and the Newton Fund-CONFAP-FAPEMIG initiative (APQ-0811/15). ODS holds a CNPq Fellowship (CNPq PQ 305736/2013-2) This is contribution \#65 from the Lab of Termitology at Federal University of Viçosa (\url{http://www.isoptera.ufv.br}).
|
\section{INTRODUCTION}
With most of our used energy being lost as waste heat, there is a growing need for high-performance thermoelectric materials that can directly and reversibly convert heat into electricity. The efficiency of a thermoelectric material is determined by the dimensionless figure of merit $ZT=S^2\sigma T/(\kappa_e+\kappa_p)$, where $S$ is the Seebeck coefficient, $\sigma$ is the electrical conductivity, $T$ is the temperature, and $\kappa_e$ and $\kappa_p$ are the electronic and phonon thermal conductivities, respectively. To obtain a high $ZT$ value, one must try to increase the power factor (PF=$S^2\sigma$) and/or decrease the thermal conductivity. However, the transport coefficients ($S$, $\sigma$, $\kappa_e$ and $\kappa_p$) are not independently tunable because they all depend strongly on the details of the band structure and the scattering of the charge carriers. As a result, it is still a challenge to obtain high $ZT$ thermoelectric materials, hindering their wide applications.
Several successful concepts have been developed to increase the $ZT$ value, such as the use of low dimensionality,\cite{MS1,MS2} which could enhance the PF due to the sharper density of states (DOS) near the Fermi energy or reduce the thermal conductivity originating from the increased phonon scattering. Therefore, many efforts have been devoted to search for low-dimensional thermoelectric materials. Due to the weak van der Waals interactions between the neighboring layers, the layered materials serve as ideal candidates which can be readily exfoliated to two-dimensional (2D) films.\cite{synthesis1,synthesis2} On the other hand, a high valley degeneracy in the electronic structure also contributes to the enhancement of the thermoelectric performance.\cite{valley1,valley2,valley3} Such degeneracy can be engineered by tuning the doping and composition in a bulk material.\cite{valley-bulk1,valley-bulk2,valley-bulk3} If we could combine the above two concepts together, namely, introduce the high valley degeneracy into a low dimensional thermoelectric material, a large $ZT$ value may be achieved.
\begin{figure}[htpb]
\centering
\includegraphics[width=1.0\columnwidth]{Fig1}\caption{\label{fig1-structure and energy band}(a) Top and (b) side views of the HfS$_{2}$ monolayer. The black line in (a) denotes the primitive cell used in our calculations. (c) Band structures of the HfS$_{2}$ monolayer without (solid black line) and with (dashed red line) spin-orbit coupling (SOC). The inset is a magnified view of the region marked by the rectangular outline. (d) Absolute value of the Seebeck coefficient and (e) power factor (PF) as a function of the carrier concentration. The solid black and dashed red lines represent the results without and with SOC, respectively.}
\end{figure}
\begin{figure}[htpb]
\centering
\includegraphics[width=1.0\columnwidth]{Fig2}\caption{\label{fig2-band struct}Phonon spectra of the HfS$_{2}$ monolayer under different external strains of (a) 0\%, (b) 6\%, (c) 9\%, and (d) 10\%. The gap between acoustical and optical phonon bands is highlighted in yellow color.}
\end{figure}
Transition metal dichalcogenides (TMDCs), which have the formula M$X_2$ (M = transition metal; $X$ = S, Se, or Te), represent a large family of layered materials. TMDCs are presently being intensively researched due to their diverse and attractive properties.\cite{TMDC1,TMDC2,TMDC3} Some of them, TiS$_2$ for example,\cite{TiS2-bulk} were found to be good thermoelectric materials. When exfoliated into film, the 2D TiS$_2$ was predicted to have much improved thermoelectric performance compared with its bulk counterpart when the thickness fell below 14 layers.\cite{TiS2-layer} HfS$_2$ is another typical TMDC. It was theoretically predicted that the lattice thermal conductivity of the bulk HfS$_2$ was about 9 Wm$^{-1}$K$^{-1}$ at 300 K,\cite{HfS2-bulk} which is much smaller than those of the MoS$_2$\cite{MoS2-bulk-kappa} and WS$_2$,\cite{WS2-bulk-kappa} making the HfS$_2$ system a promising thermoelectric material. However, the room-temperature $ZT$ value of the bulk HfS$_2$ was about 0.06,\cite{HfS2-bulk}, still far away from the requirement of the thermoelectric application.\cite{ZT-require} If using the method of low dimensionality, the thermoelectric performance of the HfS$_2$ system may be further enhanced. In this work, we investigate the thermoelectric properties of the 2D HfS$_2$ monolayer. It is demonstrated that the HfS$_2$ monolayer is an excellent n-type thermoelectric material, with much improved thermoelectric performance compared with the bulk. Furthermore, the band valleys in the HfS$_2$ monolayer can be effectively tuned by the external strain, and the Seebeck coefficient and therefore the peak value of the PF are greatly increased. At the strain of 6\%, where the degeneracy of the valence band valleys reach the maximum, the $ZT$ value of the p-type doped system is dramatically increased.
\section{Computational details}
Our calculations were performed within the framework of the density functional theory (DFT), as implemented in the ABINIT code.\cite{ABINIT1,ABINIT2,ABINIT3} The exchange correlation energy was in the form of Perdew-Burke-Ernzerhof (PBE)\cite{PBE} with generalized gradient approximation (GGA). The Brillouin zone was sampled with a $12\times12\times1$ Monkhorst-Pack $k$ mesh. The cutoff energy for the plane-wave expansion was set to be 600 eV. Based on the electronic structure, the electronic transport coefficients are derived by using the semiclassical Boltzmann theory\cite{Boltzmann} within the relaxation-time approximation, and doping is treated by the rigid-band model.\cite{rigid-band} The electronic thermal conductivity $\kappa_e$ is calculated using the Wiedemann-Franz law $\kappa_e =L\sigma T$, where $L$ is the Lorenz number. In this work, we use a Lorenz number of $1.50\times10^{-8}$ W/K$^2$.\cite{Lorenz}
In the calculation of the phonon dispersion, the density functional perturbation theory (DFPT) as implemented in the VASP package\cite{vasp1,vasp2,vasp3} was used to calculate the force constant matrices. A $5\times5\times1$ supercell was used and the phonon frequencies were obtained by the PHONOPY code.\cite{phonopy} The lattice thermal conductivity was calculated by solving the phonon Boltzmann transport equation within the relaxation time approximation, as implemented in the ShengBTE code.\cite{ShengBTE} The second order harmonic and third order anharmonic interatomic force constants (IFCs) were calculated by using $5\times5\times1$ supercell with $2\times2\times1$ Monkhorst-Pack $k$ meshes and $4\times4\times1$ supercell with $\Gamma$ point, respectively. The interactions up to third-nearest neighbors were considered when calculating the third order IFCs.
\begin{figure*}[ht]
\centering
\includegraphics[width=1.5\columnwidth]{Fig3}\caption{\label{fig2-band struct}Band structures of the HfS$_{2}$ monolayer under different external strains of (a) 0\%, (b) 2\%, (c) 4\%, (d) 6\%, (e) 8\%, and (f) 9\%. The three valence band valleys are highlighted in different color, denoted by I, II, and III, respectively; the three conduction band valleys are denoted by IV, V, and VI, respectively. For the unstrained system, the VBM and CBM are determined by valleys II and IV located at $\Gamma$ and $M$ points, respectively, which are schematically drawn in the Brillouin zone in the inset of (a). The same are shown in the insets of (c), (d), and (f) for the systems under the strains of 4\%, 6\%, and 9\%, respectively. At the strain of 6\%, the three valence band valleys converge in energy and the degeneracy reaches the maximum, as shown in the inset of (d).}
\end{figure*}
\section{RESULTS AND DISCUSSION}
The bulk HfS$_{2}$ crystallizes in the 1$T$-CdI$_{2}$ structure with the space group of $P\bar{3}m$1.\cite{bulk-structure} Due to the weak van der Waals interaction between the layers, few-layer HfS$_{2}$ could be exfoliated from the bulk structure.\cite{few-layer HfS2} The top and side views of the HfS$_{2}$ monolayer are shown in Figs. 1(a) and (b), respectively. After full relaxation, the lattice parameters are calculated to be $a=b=3.639$ {\AA}, slightly larger than the in-plane parameters of the HfS$_{2}$ bulk.\cite{bulk-structure} The structural stability is investigated by calculating the phonon spectrum, which is shown in Fig. 2(a). There is no imaginary frequency in the phonon dispersion, indicating that the HfS$_{2}$ monolayer is stable. The result agrees well with the previous report,\cite{monolayer-phonon} confirming the reliability of our calculations. We next consider the electronic properties. The band structures without and with spin-orbit coupling (SOC) included are shown in Fig. 1(c), displayed by solid black and short dashed red lines, respectively. The HfS$_{2}$ monolayer is an indirect-band-gap semiconductor, with the valence band maximum (VBM) and conduction band minimum (CBM) located at the $\Gamma$ and $M$ points, respectively. The band splitting due to the SOC is observed. In particular, at $\Gamma$ point, the splitting energy $\Delta_{\mbox{so}}$ of the VBM is 0.13 eV, thus the degeneracy of the two valleys near the VBM changes from 2 without SOC to 1 when SOC is included. However, the degeneracy of the conduction band valley near the CBM is unchanged. The band gap is calculated to be 1.22 (1.29) eV with (without) SOC included. The electronic transport coefficients ($S$, $S^2\sigma/\tau$) are shown in Figs. 1(d) and (e), respectively. We can see that for the p-type doping, the Seebeck coefficient and therefore the PF (with relaxation time $\tau$ inserted) are decreased due to the SOC, while for the n-type doping, they keep unchanged. If we notice the change of the band structure due to the SOC, we can deduce that the decrease of the Seebeck coefficient may be caused by the decrease of the valley degeneracy. Therefore, a possible way to optimize the PF is to engineer the band valleys.
Strain is an effective way to tune the electronic structure of the 2D material. In the following, a biaxial tensile strain, which is defined as $\varepsilon=(a-a_0)/a_0\times100\%$, is applied to the HfS$_{2}$ monolayer. To see how large strain the HfS$_{2}$ monolayer can withstand, we calculate the phonon spectra when the strain is increased up to 10\%. The results for the strains of 6\%, 9\%, and 10\% are displayed in Figs. 2(b), (c) and (d), respectively. No imaginary frequencies are observed in the phonon spectra until the strain increases up to 9\%. When the strain reaches 10\%, small imaginary frequency appears along the $\Gamma-M$ direction, thus the system tends to be unstable. Therefore, the largest strain that the HfS$_{2}$ monolayer can withstand is 9\%. The HfS$_{2}$ monolayer can withstand a relatively large strain, mainly originating from the special sandwich structure. When the strain is applied, the outer S atom layers move inside, so the thickness of the HfS$_{2}$ monolayer is decreased, while the bond length between Hf and S atoms is slightly altered.
\begin{figure}[ht]
\centering
\includegraphics[width=0.8\columnwidth]{Fig4}\caption{\label{fig3-energy difference}Energy difference between the three valence band valleys ($\Delta_1$) and between the three conduction band valleys ($\Delta_2$) as a function of the applied strain.}
\end{figure}
\begin{figure*}[htpb]
\centering
\includegraphics[width=1.4\columnwidth]{Fig5}\caption{\label{fig5-band struct}Carrer concentration dependence of (a) absolute value of the Seebeck coefficient, (b) electrical conductivity, (c) electronic thermal conductivity, and (d) PF of the HfS$_{2}$ monolayer at 300 K under different external strains. The left and right panels are the p- and n-type doping, respectively.}
\end{figure*}
Next, the strain effect on the electronic properties is investigated. When the doping concentration is not very high, the electronic transport coefficients of the p- and n-type doped systems are determined by the bands near the VBM and CBM, respectively. For the valence bands, the three valleys highlighted in different color in Fig. 3(a) are denoted as I, II, and III, respectively, while the three conduction band valleys are denoted as IV, V, and VI. When no strain is applied, the energy of band valley II is the largest among the three valence band valleys and the energy difference between valleys II and I (III) is denoted by $\Delta_1$, as shown in Fig. 3(b). The energy of the valley IV is the smallest among the conduction band valleys and the energy difference between IV and V (VI) is denoted by $\Delta_2$. When the strain is applied, for the valence band valleys, the valleys I and III are elevated gradually when the strain is increased, while the valley II is gradually lowered, resulting in the decrease of $\Delta_1$, as displayed in Fig. 4. At the strain of 6\%, the three valleys converge in energy and $\Delta_1$ is decreased to zero. At this strain condition, the degeneracy of the valence band valley reaches the maximum, as displayed by the inset of Fig. 3(d). When further increasing the strain, the energy of valleys I and III becomes larger than that of the valley II, and the degeneracy is reduced, as demonstrated by the inset of Fig. 3(f). For the conduction band valleys, as the strain is increased, the valley IV is gradually elevated, while the valleys V and VI are lowered, therefore, the $\Delta_2$ decreases monotonically with the increase of the strain (see Fig. 4).
The increase of the band valley degeneracy may be beneficial to the thermoelectric performance. In the following, the electronic transport coefficients of the HfS$_{2}$ monolayer under different strains are calculated by using the semiclassical Boltzmann theory. Since the system becomes unstable when the strain reaches 10\%, we only consider the condition when the strain is not larger than 9\%. Figure 5(a) shows the calculated absolute values of the Seebeck coefficients at 300 K as a function of the doping concentration. We can see that for the p-type doping, the Seebeck coefficient first increases as the increase of the strain, reaches its maximum at the strain of 6\%, and then slightly decreases when further increasing the strain. For the n-type doping, however, the absolute value of the Seebeck coefficient increases monotonically as a function of the strain, thus it reaches the maximum at the strain of 9\%. If we notice the energy difference between the band valleys (see Fig. 4), we can find that the Seebeck coefficient increases as the absolute value of the valley energy difference is decreased. When $\Delta_1$ decreases to zero, the degeneracy of the valence band valleys becomes the largest, and thus the absolute value of the Seebeck coefficient of the p-type doping reaches the maximum.
The trend of the electrical conductivity as a function of the strain (see Fig. 5(b)) is just opposite to that of the Seebeck coefficient. Here the relaxation time $\tau$ is inserted as a parameter. The electrical conductivity $\sigma/\tau$ is generally decreased by the applied strain, which is detrimental to the PF. Whether the PF will be improved by the strain or not is determined by the balance between the Seebeck coefficient and electrical conductivity. In Fig. 5(d), we plot the PF (with $\tau$ inserted) at 300 K as a function of the doping concentration under different strains. For the p-type doping, the peak value of the PF first increases as increasing the strain, reaches the maximum at the strain of 6\%, and then decreases slightly when the strain is further increased. For the n-type doping, however, the PF increases monotonically with the increase of the strain. The tread of the PF as a function of the strain is the same as that of the Seebeck coefficient, indicating that the negative effect of the strain on the electrical conductivity is overweighed by the increase of the Seebeck coefficient. Therefore, the PF of the HfS$_{2}$ monolayer can be greatly improved by the valley engineering through the method of strain.
\begin{table}[h]
\small
\caption{\label{Table1-relaxation time}Effective mass ($m^*$), carrier mobility ($\mu$), relaxation time ($\tau$) at 300 K in the zigzag and armchair directions of the unstrained and 6\% strained HfS$_{2}$ monolayers.}
\begin{tabular}{ccccccc}
\hline
~&~&~& $m^*$ & $\mu$ & $\tau$ \tabularnewline
~&~&~& ($m_e$) & (cm$^2$V$^{-1}$s$^{-1}$) & ($10^{-13}$ s) \tabularnewline
\hline
Unstrained & Zigzag & $h$ & $-$0.26 & 1141.6 & 1.68 \tabularnewline
~&~& $e$ & 0.23 & 4774.3 & 6.35 \tabularnewline
~& Armchair & $h$ & $-$0.25 & 1219.6 & 1.76 \tabularnewline
~&~& $e$ & 2.26 & 502.9 &6.45 \tabularnewline
Strained & Zigzag & $h$ & $-$0.28 & 1801.1 & 2.88 \tabularnewline
~&~& $e$ & 0.31 & 1021.9 & 1.81 \tabularnewline
~& Armchair & $h$ & $-$0.28 & 1756.8 & 2.77 \tabularnewline
~&~& $e$ & 4.21 & 94.4 & 2.26 \tabularnewline
\hline
\end{tabular}
\end{table}
As for the electronic thermal conductivity, we can see from Fig. 5(c) that the topology of $\kappa_e/\tau$ as a function of the carrier concentration is the same as that of the electrical conductivity, since it is calculated based on $\kappa_e =L\sigma T$. The electronic thermal conductivity is reduced by the strain, which is another beneficial factor to the thermoelectric performance.
As mentioned above, within our method, the electrical conductivity and therefore the PF can only be calculated with the relaxation time $\tau$ inserted as a parameter. The relaxation time is determined by $\mu=e\tau/m^*$, where $\mu$ and $m^*$ are the carrier mobility and effective mass, respectively. Details of calculating the carrier mobility can be found in the ESI.$\dag$ Since at the strain of 6\%, the degeneracy of the valence band valleys reaches the maximum, in the following, we only focus on this strain condition. The calculated $m^*$ and room-temperature $\mu$ and $\tau$ of the unstrained and 6\% strained HfS$_{2}$ monolayer are summarized in Table 1. For the unstrained HfS$_{2}$ monolayer, the effective mass $m^*$ as well as the mobility of the hole are highly isotropic due to the isotropic band dispersion near the VBM. However, for the electron, the effective mass along the zigzag direction is much smaller than that along the armchair direction, since near the CBM, the band dispersion along the $M$-$K$ direction (zigzag direction in real space) is much steeper than that along the $M$-$\Gamma$ direction (armchair direction in real space). The calculated electron mobility along the zigzag direction is as high as 4774.3 cm$^2$V$^{-1}$s$^{-1}$, which is much larger than that of the MoS$_{2}$ monolayer.\cite{MoS2-mobility} The relaxation time of the hole is much smaller than that of the electron. However, for both hole and electron, the difference of the relaxation time $\tau$ between the zigzag and armchair directions is very small, so we will use the averaged $\tau$ along the two directions to evaluate the thermoelectric performance. When the strain of 6\% is applied, for both the zigzag and armchair directions, the effective mass $m^*$ of the hole are nearly unchanged, while the carrier mobility $\mu$ and relaxation time $\tau$ are slightly increased. The effective mass of electron is however increased, both along the zigzag and armchair directions, because the band dispersion near the CBM becomes flatter when the strain is applied (see Fig. 3). The carrier mobility and the relaxation time of electron are significantly decreased compared with those of the unstrained HfS$_{2}$ monolayer.
\begin{figure}[ht]
\centering
\includegraphics[width=0.9\columnwidth]{Fig6}\caption{\label{fig6-ZT value}(a) Phonon spectra and (b) ZT value at 300 K as a function of the carrier concentration for the unstrained and 6\% strained HfS$_{2}$ monolayers. The left and right panels of (b) represent the p- and n-type doping, respectively.}
\end{figure}
\begin{table*}
\centering
\small
\caption{\label{Table1-relaxation time}Optimal doping concentration ($n$) and the corresponding Seebeck coefficient ($S$), electrical conductivity ($\sigma$), electronic and lattice thermal conductivity ($\kappa_e$ and $\kappa_l$), and $ZT$ value of the unstrained and 6\% strained HfS$_{2}$ monolayers at 300 K for the p- and n-type doping.}
\begin{tabular}{ccccccccccccccc}
\hline
~&~&~&~& $n$ &~& $S$ &~& $\sigma$ &~& $\kappa_e$ &~& $\kappa_l$ &~& $ZT$ \tabularnewline
~&~&~&~& (10$^{12}$ cm$^{-2}$) &~& ($\mu$V/K) &~& (10$^5$ $\Omega$$^{-1}$m$^{-1}$) &~& (Wm$^{-1}$K$^{-1}$) &~& (Wm$^{-1}$K$^{-1}$) ~&~& \tabularnewline
\hline
Unstrained &~& p-type &~& 2.25 &~& 197.8 &~& 1.94 &~& 0.87 &~& 2.92 &~& 0.60 \tabularnewline
~&~& n-type &~& 2.75 &~& 360.5 &~& 6.98 &~& 3.14 &~& 2.92 &~& 4.48 \tabularnewline
Strained &~& p-type &~& 12.1 &~& 336.6 &~& 4.67 &~& 2.10 &~& 2.22 &~& 3.67 \tabularnewline
~&~& n-type &~& 7.07 &~& 316.1 &~& 4.24 &~& 1.91 &~& 2.22 &~& 3.08 \tabularnewline
\hline
\end{tabular}
\end{table*}
To evaluate the figure of merit, we have calculated the lattice thermal conductivity $\kappa_p$, which is 2.92 Wm$^{-1}$K$^{-1}$ at 300 K for the unstrained HfS$_{2}$ monolayer. The HfS$_{2}$ monolayer has much smaller lattice thermal conductivity than those of the Mo/W based TMDC monolayers,\cite{MoS2-monolayer-kappa,WS2-monolayer-kappa} the same as the case of the bulk systems.\cite{HfS2-bulk,MoS2-bulk-kappa,WS2-bulk-kappa} Moreover, when under the external strain of 6\%, the room-temperature $\kappa_p$ is further suppressed to be 2.22 Wm$^{-1}$K$^{-1}$, due to the phonon softening of transverse and longitudinal acoustic (TA and LA) modes, as well as the reduced gap between the acoustical and optical phonon bands (see Fig. 6(a)). The very small lattice thermal conductivity of the HfS$_{2}$ monolayer indicating that this system may have much improved thermoelectric performance.
Combining all the calculated coefficients together, we plot in Fig. 6(b) the room temperature $ZT$ value as a function of the carrier concentration. When no strain is applied, a $ZT$ value as high as 4.48 is achieved for the n-type doped system, which is significantly improved compared with the corresponding bulk.\cite{HfS2-bulk} The $ZT$ value of the n-type doped system is much larger than that of the p-type doped one. The HfS$_{2}$ monolayer can be used as an excellent n-type thermoelectric material. However, in the fabrication of thermoelectric modules, both the p- and n-legs are needed. Interestingly, we find that the $ZT$ value of the p-type doped system is significantly increased by the applied strain, from 0.60 for the unstrained HfS$_{2}$ monolayer to 3.67 for the 6\% strained one. However, for the n-type doping, in spite of the increased peak value of the power factor (with relaxation time inserted, see Fig. 5(d)), the $ZT$ value is decreased by the external strain, from 4.48 for the unstrained system to 3.08 for the strained one. This is because that when inserting the thermal conductivity to evaluate the $ZT$ value, the optimal doping concentration gets decreased, moving to the region where the PF of the strained system is slightly smaller than that of the unstrained one. Although slightly decreased, the room-temperature $ZT$ value is still above 3.0, reaching the requirement of the thermoelectric application.\cite{ZT-require}
In Table 2, we summarize the maximum $ZT$ values of the unstrained and 6\% strained HfS$_{2}$ monolayer, with the optimal doping concentration and the corresponding Seebeck coefficient, electrical conductivity, electronic and lattice thermal conductivities included. For the p-type doping, we can see that at the optimal doping concentration, where the maximum $ZT$ value is obtained, both the Seebeck coefficient and the electrical conductivity is increased by the strain, while the lattice thermal conductivity is decreased, which lead to the significantly increased $ZT$ value. For the n-type doping, at the optimal doping concentration, although the electronic and lattice thermal conductivities are decreased, the Seebeck coefficient and the electrical conductivity are also decreased, resulting in the decreased $ZT$ value for the 6\% strained system.
\section{Conclusion}
In conclusion, we have investigated the electronic, phonon, and thermoelectric properties of the HfS$_{2}$ monolayer. The band valleys of this 2D material can be effectively engineered by the external strain. At the strain of 6\%, the three valence band valleys converge in energy and the degeneracy of the valleys reaches the maximum, while the energy difference of the conduction band valleys decreases monotonically as the strain increases up to 9\%, the largest strain the HfS$_{2}$ monolayer can withstand. Although the electrical conductivity is decreased by the strain, the increase of the Seebeck coefficient overweighs the decrease of the electrical conductivity and thus the peak value of the power factor (with relaxation time inserted) increases monotonically with the decrease of the energy difference among the band valleys. At the strain condition of 6\%, the maximum room-temperature $ZT$ value of 3.67 can be achieved for the p-type doped system, which is five times larger than that of the unstrained one. Our results indicate that the thermoelectric performance of the HfS$_{2}$ monolayer can be greatly improved by the band valley engineering through the method of strain.
In experiments, we can introduce strain to a 2D material in various ways. Traditionally, strain can arise from the lattice mismatch between epitaxial thin films and substrates.\cite{mismatch} In recent years, with the progress in nanotechnology, one could transfer a 2D film to a soft supporting substrate and apply strain to the film by either stretching\cite{stretching1,stretching2} or bending\cite{bending1,bending2} the substrate. By using this method, extremely large strain could be achieved. For example, strain as high as 30\% has been realized in graphene.\cite{stretching1,bending2} These methods can be readily transferred to the HfS$_{2}$ monolayer, which deserves further study in the future experiments.
\section{Acknowledgement}
This work was supported by the National Key Research and Development Program under Contract No. 2016YFA0300404, National Natural Science Foundation of China under Contracts No. 11404340, 11274311, U1232139, 11674326 and 11574108, the Anhui Provincial Natural Science Foundation under Contract No. 1408085MA11, the China Postdoctoral Science Foundations (Grant No. 2014M550352 and 2015T80670). The calculation was partially performed at the Center for Computational Science, CASHIPS.
|
\section{Introduction}
The problem of generating frames by iterative
actions of operators \cite{ACMT, ACAMP, AP} has emerged within the research related to the dynamical sampling problem \cite{AADP13}-\cite{ACAMP}.
The
conditions under which a frame generated by iterative
actions of operators
exists for a finite-dimensional or a separable Hilbert space have been stated in \cite{ACMT} and \cite{ACAMP}. If we have a frame, then a linear combination of a dual frame with the dynamically sampled coefficients reproduce the original signal.
The natural follow-up questions to ask in this setup are: whether we can obtain a scalable frame under iterative actions, and if not, whether we can find a dual frame which preserves the dynamical structure.
Let $A $ be an operator on a separable Hilbert space $\mathbb{H}$. We consider a countable set of vectors $G $ in $\mathbb{H}$, and a function $L : G \rightarrow \mathbb Z_+$, where $\mathbb Z_+ = \mathbb N \cup \Set{0}$. Related to the iterated system of vectors \begin{equation}\label{oursystem}
\{A^j {\bf g} \; | \; {\bf g} \in G, \; 0 \leq j \leq L({\bf g}) \},\end{equation}
we answer the following two questions:
\begin{itemize}
\item[(Q1)]
What conditions on $A$, $ G$ and $L$ ensure that \eqref{oursystem}
is a scalable frame for $\mathbb{H}$?
\item[(Q2)] Assuming the system \eqref{oursystem}
is a frame for $\mathbb{H}$, can we obtain a dual frame for \eqref{oursystem}, perhaps by iterative actions of some operator?
\end{itemize}
The motivation for studying systems of type \eqref{oursystem} comes from the
{\it dynamical sampling problem }
(DSP): Find sampling locations
that allow the reconstruction of an unknown function ${\bf f} $
from the scarce
samples of ${\bf f}$,
and its
evolved states $A^n {\bf f}$.
In the DSP, $n$ represents time, and
$A^*$ is an evolution operator; for instance, $A^*$ can represent the heat evolution operator,
${\bf f}$ the temperature at time
$n
= 0$,
and
$(A^*)^n {\bf f}$ the temperature at time
$n$. The DSP for the heat evolution operator was studied in \cite{LV09, RCLV11}; generalizations of the DSP and related applications can be found in \cite{AADP13}-\cite{ACAMP}.
More precisely, the DSP is as follows: Let the initial state of a dynamical system be represented by an unknown element
${\bf f} \in \mathbb{H}$.
Say the initial state ${\bf f} $ is evolving under the action of an
operator $A^*$ to the states
${\bf f}_j = A^*{\bf f}_{j-1}$, where ${\bf f}_0 = {\bf f}$ and $j \in \mathbb Z_+$.
Given a set of vectors $G \subset \mathbb{H}$, one can find conditions on $ A$, $G $ and $ L = L({\bf g})$ which
allow the recovery of the initial state ${\bf f}$ from the set of samples
$\{ \langle A^{* j}
{\bf f} , {\bf g} \rangle \; | \; {\bf g} \in G\}_{ j =0}^{L({\bf g})}$.
In short, the problem of signal recovery via dynamical sampling is solvable if the
set of vectors $F_A^{L}(G) : =\{ A^{ j}
{\bf g} \; | \; {\bf g} \in G\}_{j=0}^{ L({\bf g})}$ is a frame for $\mathbb{H}$, \cite{ACMT}. In frame theory it is known that every frame has at least one dual frame; if $F_A^{L}(G)$ is a frame for $\mathbb{H}$, and its dual frame elements are ${\bf h}_{{\bf g}, j}$, then all ${\bf f} \in \mathbb{H}$ are reconstructed as
\begin{equation}\label{signalreconstructuonDS}
{\bf f} = \sum_{{\bf g} \in G} \sum_{j=0}^{L({\bf g})}\langle {\bf f}, A^j {\bf g} \rangle {\bf h}_{{\bf g}, j}.
\end{equation}If the frame $F_A^{L}(G)$ is {\textit{ scalable}}, then its dual frame elements are $w^2_{j,{\bf g}} A^j {\bf g}$ for some {\it scaling coefficients } $w_{j, {\bf g}}$, and the reconstruction formula \eqref{signalreconstructuonDS} is
\begin{equation}\label{signalreconDSscalable}
{\bf f} = \sum_{{\bf g} \in G} \sum_{j=0}^{L({\bf g})}w_{j,{\bf g}}^2 \langle {\bf f}, A^j {\bf g} \rangle A^j {\bf g}.
\end{equation}Notice that the frame coefficients in \eqref{signalreconstructuonDS} are exactly the samples
\begin{equation}
\label{sampleseq}
\langle A^{* j} {\bf f} , {\bf g} \rangle=\langle {\bf f} , A^j {\bf g} \rangle.
\end{equation} Thus the set of samples $\{ \langle A^{* j}
{\bf f} , {\bf g} \rangle \; | \; {\bf g} \in G\}_{ j =0}^{L({\bf g})}$ is sufficient for the recovery of ${\bf f}$.
Since \eqref{signalreconstructuonDS} requires that the dual frame of $F_A^{L}(G)$ is known, unless the frame is scalable as in \eqref{signalreconDSscalable}, it is significant to find the answers to questions (Q1) and (Q2).
\subsection{Contribution and organization} In Section \ref{prelim} we recall the notions of frames, scalable frames and, in particular, frames of iterative actions of operators, i.e., dynamical frames.
In Section \ref{allnewstuffbeyongAA}, we illustrate the dynamical nature of the canonical dual frame of \eqref{oursystem} in Theorem \ref{canondualframedysam}, and the fusion frame structure of dynamical frames (Corollary \ref{fusiondyn}). In Section \ref{mainresults} we give a characterization of scalability in Theorem \ref{multiscalablediagonalgen}, under the assumption that $A$ is normal. Section \ref{blockdiagOpsubsection} contains several generalized examples of frames and scalable frames in lower dimensions, and we characterize frame scalability in $\mathbb R^2$ and $\mathbb R^3$. In addition, we provide examples of operators which are not normal, yet generate scalable frames for $\mathbb R^2$ and $\mathbb R^3$.
Motivated by these results, we
study block-diagonal operators, which combine low-dimensional frames into higher-dimensional frames (Theorem \ref{blockresultbig}).
%
In Section \ref{compansection}, we also provide examples of dynamical scalable frames, generated using companion operators \cite{HJ85} and generalized companion operators. In section \ref{conclusion} we give initial answers to question (Q3), addressing frame scalability when multiple operators are involved.
\section{Preliminaries}\label{prelim}
Frames are a generalization of orthonormal bases.
For an orthonormal basis $ \{{\bf f}_i\}_{i \in I}$ of \( \mathbb{H} \), it holds
\begin{equation}\label{onmbrepr} {\bf f} = \sum_{i \in I} \inpro{{\bf f}, {\bf f}_i} {\bf f}_i \;\; \text{ for all } {\bf f} \in \mathbb{H}. \end{equation}
%
The uniqueness of representation \eqref{onmbrepr} is not always an advantage.
In applications such as image and signal processing, the loss of a single coefficient during data transmission will prevent the recovery of the original signal,
unless we ensure redundancy via frame spanning.
Since finding a dual frame can be computationally challenging, one significant direction of current research has been on the construction of tight frames in finite dimensions
\cite{ BM03, STDH07, CMKLT06, CFHWZ12, HKLW07}. A tight frame plays the role of its own dual, and provides a reconstruction formula as in \eqref{onmbrepr} up to a constant.
Recently, the theme of scalable frames has been developed as a method of constructing tight frames from general frames by manipulating the length of frame vectors.
Scalable frames maintain erasure resilience and sparse expansion properties of frames \cite{CC13, CKLMNPS14, KOF13, KOPT13, CKOPW15}.
First, let us review relevant definitions and known results. Throughout this paper $\mathbb{H}$ denotes a separable Hilbert space.
Given an index set $I$, a sequence $F = \{{\bf f}_i\}_{i \in I}$ of nonzero elements of $\mathbb{H}$ is a \textit{frame} for $\mathbb{H}$, if there exist $0<A \leq B < \infty$ such that
\begin{equation}
\label{frameineq}
A\Vert {\bf f} \Vert^2 \leq \sum_{i \in I} \vert \langle {\bf f} , {\bf f} _i \rangle \vert^2 \leq B\Vert {\bf f} \Vert^2 \;\; \text{ for all } {\bf f} \in \mathbb{H}.
\end{equation}
In finite dimensions,
we find it useful to express frames as matrices, so we abuse the notation of $F$ as follows: when $\dim \mathbb{H} = n$, a frame $F=\{{\bf f}_i\}_{i \in I}$ for $\mathbb{H}$ is often represented by a $n \times k$ matrix $F$, whose column vectors are ${\bf f}_i$, $i = 1, \ldots, k$.
The frame operator $S = FF^*$ is then positive, self-adjoint
and invertible.
For each frame $F$ there exists at least one \textit{dual} frame $ G= \{{\bf g}_i\}_{i \in I}$, satisfying
\begin{equation}
\label{framerepr}
{\bf f} = \sum_{i \in I} \langle {\bf f}, {\bf f}_i \rangle {\bf g}_i = \sum_{i \in I} \langle {\bf f}, {\bf g}_i \rangle {\bf f}_i \;\; \text{ for all } {\bf f} \in \mathbb{H}.
\end{equation}
The matrix equation $ F G^* = G F ^* = I$ is an equivalent expression to the frame representation \eqref{framerepr}.
The set $ \{{\bf g}_i=S^{-1}{\bf f}_i\}_{i \in I}$ is called the canonical dual frame.
Finding a dual frame can be computationally challenging; thus it is of interest to work with tight frames. We say that a frame is $A$-\textit{tight} if $A=B$ in \eqref{frameineq}. In this case, the function reconstruction is simplified since the frame operator is the identity operator up to scalar multiplication.
So, for an $A$-tight frame, we only need one frame for both analysis and reconstruction, as \eqref{framerepr} becomes
\begin{equation}
{\bf f} =\frac{1}{A} \sum_{i \in I} \langle {\bf f}, {\bf f}_i \rangle {\bf f}_i = \frac{1}{A}FF^*{\bf f} \;\; \text{ for all } {\bf f} \in \mathbb{H}.
\end{equation}
When $A=1$, we call $F$ a Parseval frame.
If a frame $F=\{ {\bf f}_i\}_{i \in I}$ is not tight, but we can find scaling coefficients $w_i \ge 0$, $i \in I$, such that the scaled frame $F_w=\{w_i{\bf f}_i\}_{i \in I}$ is tight, then we call the original frame $F$ a \textit{scalable} frame.
We note that the notion of scalability of a frame is defined for a unit-norm frame in \cite{CKLMNPS14}, but in this manuscript we do not require a scalable frame to be unit-norm.
For a scalable frame, the scaled frame representation becomes
\begin{equation}\label{defscalablerepr} {\bf f} = \sum_{i \in I} \langle {\bf f}, w_i {\bf f}_i \rangle w_i {\bf f}_i = F_wF_w^*{\bf f} = F D_{w^2}F ^*{\bf f}\;\; \text{ for all } {\bf f} \in \mathbb{H},\end{equation}
where $D_{w^2}$ denotes a diagonal operator with $w_i ^2$ as diagonal entries.
If the scaling coefficients $w_i $ are positive for all $i \in I$, then we call the original frame $F$ a \textit{strictly scalable} frame.
Let $I$ denote a finite or countable index set, let $G = \{ {\bf f}_s \}_{s\in I} \subset \mathbb{H}$ and let $A : \mathbb{H} \to \mathbb{H}$ be a bounded operator. We call the collection
\begin{equation}\label{orifdynfr}
F_{G}^{\bf L } (A) = \cup_{s \in I} \{A^j {{\bf f}}_{s} \,:\, j=0,1,\ldots,L_s \}
\end{equation}
a {\it dynamical system}, where $L_s \geq 0$ ($L_s$ may go to $\infty$) and ${\bf L}=(L_s)_{s \in I}$ is a sequence of iterations. The operator $A$, involved in generating the set \eqref{orifdynfr}, is sometimes referred to as a {\it dynamical operator}.
If $A$ is fixed, then we use the notation $F_{G} ^{\bf L}$, and if $G =\{ {\bf f} \}$ and ${\bf L} =\{ L \}$, then we label \eqref{orifdynfr} by $F_{{\bf f}} ^L$.
Note that in \cite{ACMT}, ${\bf f}_{s}$ are chosen to be the standard basis vectors, while in this manuscript, we allow the use of any nonzero vector ${\bf f}_{s} \in \mathbb{H}$. If \eqref{orifdynfr} is a frame for $\mathbb{H}$, then we call $F_{G}^{\bf L } (A)$ a \textit{dynamical frame}, generated by operator $A$, set $G$ and sequence of iterations ${\bf L}$.
\section{New results on dynamical frames}\label{allnewstuffbeyongAA}
As we are about to see in Theorem \ref{canondualframedysam}, the canonical dual frame of a dynamical frame preserves the dynamical structure, just like the canonical duals of wavelet or Gabor frames preserve the corresponding wavelet/Gabor structure \cite{Gro01}.
\begin{thm}\label{canondualframedysam}
Let $G = \{ {{\bf f}}_{s} \}_{s \in I} \subset\mathbb{H}$, where $I$ is a countable index set, and assume that $F_{G}^{\bf L} (A)$ is a frame for $\mathbb{H}$, with frame operator $S$.
The canonical dual frame of $F_{G}^{\bf L} (A)$ is the dynamical frame $F_{ G'} ^{\bf L} (B)$, generated by $B=S^{-1}AS$, $G' = \{ {{\bf g}}_s = S^{-1}{{\bf f}}_{s} \}_{ s \in I}$, and sequence of iterations $\bf L$.
That is, for every ${\bf f} \in \mathbb{H}$ the frame reconstruction formula is
\begin{equation}\label{frdynreprAB}
{\bf f} = \sum_{s\in I}\sum_{j=0} ^{L_s}\langle A^{*j}{\bf f}, {{\bf f}}_{s} \rangle B^j {{\bf g}}_{s}.
\end{equation}
\end{thm}
\begin{proof}
The elements of the canonical dual frame of $F_{G}^{\bf L} (A)$ are computed as $S^{-1} \left(A^j {{\bf f}}_{s}\right)$, $s \in I$, $j=0,1,\ldots,L_s$. Let ${{\bf g}}_s = S^{-1}{{\bf f}}_{s}$, $s \in I$, then for all $j \geq 0$ we have
$$B^j {{\bf g}}_{s} = (S^{-1}AS)(S^{-1}AS)\ldots(S^{-1}AS){{\bf g}}_s = S^{-1}A^j \left(S {{\bf g}}_{s}\right) = S^{-1}\left( A^j {{\bf f}}_{s}\right),$$
and \eqref{frdynreprAB} follows by \eqref{framerepr} and \eqref{sampleseq}.
\end{proof}
It is a known fact in frame theory that an invertible operator preserves the frame inequality. It follows from this that under the action of an invertible operator, the dynamical structure is preserved:
\begin{thm}\label{123}
Let $\mathbb{H}_1$ and $\mathbb{H}_2$ be two separable Hilbert spaces. Let $G = \{ {\bf f}_s \}_{s \in I} \subset \mathbb{H}_1$, where $I$ is a countable index set. Let ${\bf L}=(L_s)_{s \in I}$, $L_s \geq 0$.
Let $A$ be an operator on $\mathbb{H}_1$ and let
$B: \mathbb{H}_1 \rightarrow \mathbb{H}_2$ be an invertible operator. Set ${\bf g}_s = B {\bf f}_s \in \mathbb{H}_2$, $s \in I$, and $C = B AB^{-1}$. TFAE:
\begin{itemize}
\item[(i)] The set
$\displaystyle F=\cup_{s\in I} \{A^j {\bf f}_{s}\}_{j=0}^{L_s}$ is a frame for $\mathbb{H}_1$,
\item[(ii)] The set $\displaystyle BF= \cup_{s\in I} \{C^j {\bf g}_{s}\}_{j=0}^{L_s}$ is a frame for $\mathbb{H}_2$.
\end{itemize}
\end{thm}
\begin{proof}
Let ${\bf g}_s = B{\bf f}_s \in \mathbb{H}_2$, $s \in I$, and set $C = B AB^{-1}$. Note that $C^j = B A^jB^{-1}$, due to $B^{-1}B=I$. For all $A^j {\bf f}_s \in F\subset \mathbb{H}_1$, we have \begin{equation}
BA^j{\bf f}_s =BA^j B^{-1}B{\bf f}_s =BA^j B^{-1} {\bf g}_s = C^j{\bf g}_s\in BF\subset \mathbb{H}_2. \end{equation} The operator $B$ is invertible, thus $BF$ is a frame if and only if $F$ is a frame, so (i) and (ii) are equivalent.
\end{proof}
\begin{com}
If $\mathbb{H}=\mathbb{H}_1=\mathbb{H}_2$, then Theorem \ref{123} is a generalization of the change of basis result. Notice that under the action of an invertible operator $B: \mathbb{H} \rightarrow \mathbb{H} $, the elements of a dynamical frame $F$ for $\mathbb{H}$ preserve the dynamical structure, i.e., $BF$ is also a dynamical frame for $\mathbb{H}$.
\end{com}
Fusion frames \cite{CK04} are frames which decompose into
a union of frames for subspaces of a Hilbert space $\mathbb{H}$.
Given a countable index set $I$, let $\mathcal{W}: = \{W_i \, | \, i \in I \}$ be a family of closed subspaces in $\mathbb{H}$. Let the orthogonal projections
onto $W_i$ be denoted by by $P_i$. Then $\mathcal{W}$ is a \textit{fusion frame} for $\mathbb{H}$, if there exist $C, D >0$ such that
\[
C \Vert {\bf f} \Vert^2 \leq \sum_{i \in I} \Vert P_i({\bf f}) \Vert^2 \leq D\Vert {\bf f} \Vert^2 \;\; \text{ for all } {\bf f} \in \mathbb{H}. \]
Let $F_{i}=\{{\bf f}_{ij} \}_{ j \in J_i }$ be a frame for $W_i$, $i \in I$, with frame bounds $A_i$, $B_i$. If $0 < A = \inf_{i \in I} A_i \leq \sup_{i \in I} B_i = B < \infty$, then
\cite{CK04}:
\begin{equation}\label{cassazaff}
\hspace{-2mm} \cup_{i \in I} F_{i} \; \text{is a frame for $\mathbb{H}$ if and only if} \;
\{ W_i\}_{i \in I} \; \text{ is a fusion frame for $\mathbb{H}$.}
\end{equation}
If $F_i$ denotes the frame matrix formed by the frame vectors for each $W_i$, and $G_i$ contains the dual frame elements $\{ {\bf g}_{ij}\}_{j \in J_i}$, then the fusion frame operator $S$ positive and invertible on $\mathbb{H}$, and for all ${\bf f} \in \mathbb{H}$, we have
\begin{equation}
\label{fusionfrmatrix}
{\bf f} = \sum_{i \in I} F_iG_i ^* {\bf f} = \sum_{i \in I} G_iF_i ^* {\bf f}.
\end{equation}
By \eqref{cassazaff} and \eqref{fusionfrmatrix}, a dynamical frame induces a fusion frame:
\begin{corollary}\label{fusiondyn}
Let
$F= \cup_{s\in I} \{A^j {{\bf f}}_{s} \}_{ j=0}^{ L_s}$ be a frame for $\mathbb{H}$. We introduce subspaces of $\mathbb{H}$ by
\begin{equation}
W_s = \overline{span \{A^j {{\bf f}}_{ s} \,:\, 0 \leq j \leq L_s \}},\;\; \text{ for all } s \in I.
\end{equation}
Then $\{W_s \}_{s \in I}$ is a fusion frame of $\mathbb{H}$.
\end{corollary}
\section{ Scalable frames generated by dynamical operators}\label{mainresults}
Now, we
study the scalability of frames of type \eqref{oursystem}.
{A prior result on this topic (see Theorem 8 in \cite{AP}) has restrictive requirements, and delivers a tight frame if the involved operator $A$ is a contraction, i.e., $A^j {\bf f} \rightarrow 0$ for all elements ${\bf f}$ in the studied Hilbert space. }
Our research results illuminate the fact that - in finite dimensions - obtaining a tight or a scalable frame is possible in many cases.
If the operator $B$ occurring in Theorem \ref{123} is unitary, then the property of scalability is preserved, and we have:
\begin{corollary}\label{scalabilityinparalel}
Let $G = \{ {\bf f}_s \}_{s \in I} \subset \mathbb{H}$ and ${\bf L}=(L_s)_{s\in I} $, $L_s \geq 0$.
Let $A$ be a bounded operator on a separable Hilbert space $\mathbb{H}$.
If $B$ is a unitary operator on $\mathbb{H}$,
then
$\cup_{s \in I} \{A^j {\bf f}_{s}\}_{j=0}^{L_s}$ is a scalable frame if and only if
$\cup_{s \in I} \{C^j {\bf g}_{s}\}_{j=0}^{L_s}$ is a scalable frame, where $C =BA B^{*}$ and
${\bf g}_s = B{\bf f}_s$, $s \in I$. \end{corollary}
\begin{corollary}\label{generalSchurstatement}
Let $A, R$ be two operators on a separable Hilbert space $\mathbb{H}$, and let $U$ be a unitary operator on $\mathbb{H}$. Let $ {\bf f}_{s} \in \mathbb{H}$, and set ${\bf v}_s = U^* {\bf f}_{s}$ for all $s \in I$, where $I$ is a countable index set.
If $A=URU^*$, then TFAE:
\begin{itemize}
\item[(i)] $\displaystyle \cup_{s \in I} \{ A^j {\bf f}_{ s} \}_{j=0}^{L_s}$ is a scalable frame for $\mathbb{H}$,
\item[(ii)] $\displaystyle \cup_{s \in I} \{R^j {\bf v}_s \}_{j=0}^{L_s}$ { is a scalable frame for $\mathbb{H}$.}
\end{itemize}
\end{corollary}
Corollary \ref{generalSchurstatement} is relevant to the {\it Schur} decomposition: recall, any
operator $A$ on a finite-dimensional Hilbert spaces $\mathbb{H}$ has a non-unique Schur decomposition of type $A=URU^*$, where $U$ is a unitary $n \times n$ matrix, and $R$ is of Schur form.
When $A=A^*$, i.e., $A$ is normal, then the Schur decomposition becomes unique, and is reduced to the classical unitary diagonalization. In the next subsection, we exploit the simplicity of the unitary diagonalization of normal operators to give more explicit conditions on the normal operator $A$ in order to ensure scalability of a frame of type $F^{\bf L} _G(A)$.
\subsection{Normal operators }
Let $A$ be a normal operator on $\mathbb{H}$. By the spectral theorem, there exists a unitary operator $U$, and a diagonal operator $D$ such that
$A=UDU^*$; in fact, for each $j \in \mathbb Z_+$ $A^j = UD^jU^*$.
Now, let $\mathcal{G} = \{ {\bf f}_s \}_{s \in I}$ and set ${\bf v}_s=U^*{\bf f}_s$, $s \in I$.
Then for each $j \in \mathbb Z_+$,
\begin{equation}\label{connection}
A^j {\bf f}_s = UD^j U^* {\bf f}_s= UD^j {\bf v}_s = U(D^j {\bf v}_s )
\;\; \text{ for all } {\bf f}_s \in \mathcal{G}.
\end{equation}
Corollary \ref{generalSchurstatement} for normal operators reads as follows:
\begin{corollary}\label{connectSymDiagmulti}
Let $A$ be a normal operator on $\mathbb{H}$, and let $A=UDU^*$ be its unitary diagonalization.
Let $\{ {\bf f}_{s}\}_{ s \in I} \subset \mathbb{H}$, and set ${\bf v}_{s} = U^* {\bf f}_{s}$, $s \in I$.
TFAE
\begin{itemize}
\item[(i)] The set
$\displaystyle \cup_{ s \in I}\{ A^j {\bf f}_{s} | \; j=0,1,\ldots, L_s\}$ is a scalable frame for $\mathbb{H}$.
\item[(ii)] The set
$ \displaystyle \cup_{ s \in I} \{ D^j {\bf v}_{s} | \; j=0,1,\ldots, L_s\}$ is a scalable frame for $\mathbb{H}$.
\end{itemize}
\end{corollary}
We now restrict our attention to a finite dimensional Hilbert space $\mathbb{H} =\mathbb R^n$ or $\mathbb C^n$. Let us first point out that the frame scalability property is preserved under simple manipulations:
\begin{proposition} Let $F = \{{\bf f}_i\}_{i=1}^k $ be a scalable frame for $\mathbb{H}$, $\dim H = n$.
Then the following are also scalable frames:
\begin{itemize}
\item[(i)] any column or row permutation of $F $
\item[(ii)] \(\{ U {\bf f}_i\}_{i=1}^k\) for any unitary matrix $U$
\end{itemize}
\end{proposition}
Given a diagonal operator $D$ in a Hilbert space $\mathbb{H}$ with $\dim \mathbb{H} = n$, we first focus our attention on solving the \textit{one-vector problem}: we look for conditions on $D$, and an unknown vector ${\bf v} \in \mathbb{H}$, which generate a scalable frame for $\mathbb{H}$ of type \eqref{oursystem}.
Let $L\geq 0$, let $D$ denote a diagonal $n\times n$ matrix, with diagonal entries $a_1,\ldots,a_n$, and let ${\bf v} = (x(1), \ldots, x(n))^T \in \mathbb{H}$.
Let $ w_j \in \mathbb R_+$, $0\leq j \leq L$, be scaling coefficients such that $ F_W= \{ w_j D^j {\bf v} \}_{ j=0} ^{ L} $ is a Parseval frame for $\mathbb{H}$, i.e.,
\begin{equation}\label{wantedS}
F_W F_W^*= I.
\end{equation}
Note that \eqref{wantedS} is equivalent to the system of equations
\begin{eqnarray}\label{wanteddiagscalable}
|x(i)|^2 \sum_{k=0} ^L w_k ^2 | a_i |^{2k} & =& 1, \quad i=1,\ldots, n; \nonumber \\
\sum_{k=0} ^L w_k ^2 \left( a_i\bar{a_j } \right)^k& =& 0, \quad i \neq j.
\end{eqnarray}
There exist real solutions of \eqref{wanteddiagscalable} when $n \leq 2$. For instance, when $\mathbb{H} = \mathbb R^2$, the choice of ${\bf v} = (0.5, 0.5)^T$ and $D= diag(1, -1)$ generates the set $\{{\bf v}, D{\bf v}, D^2 {\bf v} , D^3{\bf v}\}$, which is a Parseval frame for $\mathbb R^2$ .
However, when $\mathbb{H} = \mathbb R^3$, the equation $\sum_{k=0} ^L w_k ^2 \left( a_ia_j \right)^k = 0, \; i \neq j$ implies that for the first three $a_i's$, we always have the relation $a_1a_2$, $ a_1a_3$, and $a_2a_3$ are all negative numbers assuming $w_i \neq 0, \, i =1, 2, 3$, which is not possible. Thus we have:
\begin{thm}
Let ${\bf v} \in \mathbb R^n$, and $a_1, \dots, a_n \in \mathbb R$.
If $n \ge 2$, then any normal operator for $\mathbb R^n$ can not generate a strictly scalable frame from ${\bf v}$.
\end{thm}
In contrast to the real case, there exists a solution to the one-vector problem in $\mathbb C^n$, involving the $k$-th root of unity
\begin{ex}
Let $\gamma = e^{2\pi i/ k}$, $k \ge n$.
Then the following dynamical operator $A$ and the vector ${\bf v}$
\begin{equation*}
A = \left( \begin{array}{ccc}
1& 0& 0 \\
0 & \ddots & 0 \\
0 & 0 & \gamma^{n-1}
\end{array}\right), \quad
{\bf v} = \frac{1}{\sqrt{k}} \left( \begin{array}{c}
1 \\
\vdots \\
1
\end{array}\right)
\end{equation*}
generate the Harmonic tight frame $F_{{\bf v}}^{k-1}$.
\end{ex}
Next, we consider the multi-generator case:
By \eqref{defscalablerepr}, the scaling coefficients $ w_{s,j}$ related to vectors $D^j {\bf v}_s$, $0\leq j \leq L(s)$, where ${\bf v}_s =(x_s(1), \ldots, x_s(n))^T$, $1\leq s \leq p$, need to be solutions to the following system of equations:
\begin{equation}\label{wantedScalDiagmultivrs}
\left\{ \begin{array}{ll}
\sum_{s=1} ^p |{x_{s}(i)}|^2 \left[w_{s,0}^2+ w_{s,1}^2 |{a_i}|^2 + \ldots + w_{s,L_s}^2 |{a_i}|^{2L_s} \right] = 1, \\
\sum_{s=1} ^p x_{s} (i) \bar{x_{s}} (j) \left[ w_{s,0}^2 + w_{s,1}^2 a_i \bar{a_j} + \ldots + w_{s,L_s}^2(a_i \bar{a_j})^{L_k} \right] = 0, \end{array}\right. \\
\end{equation}
for all $i,j=1,\ldots, n$, $i \neq j$.
\begin{proposition}\label{multiscalablediagonalgen}
Let $D$ be a diagonal $n\times n$ matrix with
diagonal entries $a_1,\ldots, a_n \in \mathbb C$, and let ${\bf v}_s = (x_{s}(1), \ldots, x_{s}(n))^T \in \mathbb C^n$, $s \in \{1,\cdots, p\}$, $p \geq 1$.
TFAE:
\begin{itemize}
\item[(i)] The set $\cup_{s=1} ^p \{ D^j {\bf v}_s \; | \; j =0,1,\ldots, L_s\}$ is a scalable frame for $\mathbb{H}$
\item[(ii)] There exist scaling coefficients $w_{s,0}, w_{s,1},\ldots, w_{s,L_s}$, $1\leq s \leq p$, which satisfy conditions \eqref{wantedScalDiagmultivrs}. \end{itemize}
\end{proposition}
By Corollary \ref{connectSymDiagmulti} and Proposition \ref{multiscalablediagonalgen}, the following result holds true for a finite dimensional Hilbert space $\mathbb{H}$:
\begin{thm}\label{symmetriccaseequivalence}
Let $A = UDU^*$ be a normal $n\times n$ matrix, where $U$ is unitary, and $D$ is diagonal, with diagonal entries $a_1, \ldots, a_n \in \mathbb C$.
Let ${\bf f}_s \in \mathbb{H}$, and set ${\bf v}_{s} = U^* {\bf f}_s = (x_s(1), \ldots, x_s(n))^T$, $1\leq s \leq p$.
The set $\cup_{s=1}^p \{ A^j {\bf f}_{s} \; | \; 0\leq j \leq L_s \}$ is a scalable frame of $\mathbb{H}$ if an only if there there exists a positive solution $w_{s,0}, w_{s,1},\ldots, w_{s,L_s}$, $1\leq s \leq p$ to the system of equations \eqref{wantedScalDiagmultivrs}, defined with respect to $a_1, \ldots, a_n$ and $x_s(1), \ldots, x_s(n)$, $1\leq s \leq p$.
\end{thm}
\begin{com}
The problem of finding specific conditions under which the set in item (ii) in Corollary \ref{generalSchurstatement} is a scalable frame for $\mathbb{H}$ is still open for operators which do not possess a unitary diagonalization.
For this reason, we further study several operators with special structures, such as block-diagonal operators (section \ref{blockdiagOpsubsection}) and companion operators (subsection \ref{compansection}).
\end{com}
\section{Block-diagonal operators } \label{blockdiagOpsubsection}
In this section,
we explore the case when the operator $A$ is of block-diagonal form. Block-diagonal operators give us a chance to offer a partial answer to (Q1) in the case when we don't have a unitary diagonalization.
Note that in subsection \ref{blocks} we give examples of
operators which generate scalable frames in Hilbert spaces of dimension $2$ and $3$. Since we can treat $\mathbb{H}$ with $\dim \mathbb{H} = n$ as a decomposition of several subspaces of dimensions 2 and 3, the examples in subsection \ref{blocks} provide infinite examples of block-diagonal operators which generate scalable frames for $\mathbb{H}$.
\begin{thm}\label{stackScale}
Let $F_s$ be a scalable frame for $\mathbb{H}_{s}$, with $\dim \mathbb{H}_s = n_s$, $s =1, \ldots p$, and let
\begin{equation}\label{scal333}
G = \left( \begin{array}{ccc}
F_1 & 0 & 0 \\
0 & \ddots & 0 \\
0 & 0 & F_p
\end{array}\right). \end{equation}
Then $G$ is a scalable frame for
$\mathbb{H}= \mathbb{H}_1 \oplus \ldots \oplus \mathbb{H}_p$.
\end{thm}
\begin{definition}\label{wellembededvr}
Let $A_s : \mathbb{H}_s \rightarrow \mathbb{H}_s$ be an operator on $\mathbb{H}_s$, with $\dim \mathbb{H}_s = n_s$, $1\leq s \leq p$.
Let $A : \mathbb{H}_s \rightarrow \mathbb{H}_s $ be a block-diagonal operator on $\displaystyle \mathbb{H}= \oplus_{s=1}^p \mathbb{H}_s$, constructed as follows:
\begin{equation}\label{blockdiagdynamoperator}
A = \left( \begin{array}{ccc}
A_1 & \ldots & 0\\
\vdots & \vdots & \vdots \\
0 & \ldots & A_p
\end{array} \right).
\end{equation}
Let ${\bf v} \in \mathbb{H}_s$ for some $1\leq s \leq p$. We say that ${\bf v}$ is {\it well-embeded } in ${\bf f} \in \mathbb{H}$ with respect to operator \eqref{blockdiagdynamoperator} if
\begin{equation}
\begin{cases} {\bf f}(j) = {\bf v}(i), &\mbox{if } j = n_1 +\ldots n_s +i\\
{\bf f}(j) = 0, & \mbox{otherwise.} \end{cases}
\end{equation}
\end{definition}
Whenever ${\bf v}$ is well-embedded in ${\bf f}$ with respect to \eqref{blockdiagdynamoperator}, we have
$$A{\bf f}=\left( \begin{array}{c}
0 \\
A_s {\bf v} \\
0
\end{array} \right). $$
\begin{thm}\label{blockresultbig} Let $A_s : \mathbb{H}_s \rightarrow \mathbb{H}_s$ be an operator on $\mathbb{H}_s$, with $\dim \mathbb{H}_s = n_s$, $1\leq s \leq p$.
Let $A : \mathbb{H}_s \rightarrow \mathbb{H}_s $ be a block-diagonal operator on $\displaystyle \mathbb{H}= \oplus_{s=1}^p \mathbb{H}_s$, constructed as in \eqref{blockdiagdynamoperator}.
Let ${\bf f}_{s,1}, \ldots, {\bf f}_{s,m_s} \in \mathbb{H}$, $1\leq s \leq p$ be well-embedded vectors ${\bf v}_{s, 1} \ldots, {\bf v}_{s,m_s} \in \mathbb{H}_s$, $1\leq s\leq p$.
\begin{equation}\label{bigguy}
\text{The set} \;\;\;\; \; \bigcup_{s=1}^p \{ A^j {\bf f}_{s, k} \;\; | \;\; 1\leq k \leq m_s \}_{j=0}^{ L_{s,k} } \;\;\;\; \; \;\;\;\; \; \;\;\;\; \;
\end{equation}
{ is a (scalable) frame of $\mathbb{H}$}
if and only if $\; \{ A_s^j {\bf v}_{s,k} \; | \; 1\leq k\leq m_s \}_{j=0}^{ L_{s,k} } $ are (scalable) frames of $\ \mathbb{H}_s $ for all $1\leq s \leq p$.
\end{thm}
\begin{proof}
We assume that all $m_s =1$, i.e., ${\bf f}_{s,k} = {\bf f}_s$, ${\bf v}_{s,k} = {\bf v}_s$, and $ L_{s,k} = L_s$, $1\leq s \leq p$, to simplify the presentation of the proof.
The matrix representation of $ \cup_{s=1}^p \{ A^j {\bf f}_s \}_{j=0}^{L_s}$ with scaling coefficients
$w_{s, j}$, $0\leq j \leq L_s$ for each $s=1,\ldots, p$ is of block-diagonal form:
\begin{equation*}
F=\left( \begin{array}{ccccccc}
w_{1,0} {\bf v}_{1} & \ldots & w_{1, L_1}A_1^{L_1} {\bf v}_{1}& &&& \\
&&&\ddots &&& \\
&&& & w_{p, 0}{\bf v}_{p} & \ldots & w_{p, L_p} A_p^{L_p}{\bf v}_{p}
\end{array}\right) . \end{equation*}
If $F$ is a tight frame, then row vectors of $F$ are orthogonal and have the same norm and so does
$ (w_{s, 0} {\bf v}_{s} \ldots w_{s, L_s} A_s^{L_s} {\bf v}_{s})$ for each
$s =1, \ldots, p$. This implies that the system $ \{ A_s^j {\bf v}_{s}\}_{j=0}^{L_k}$ is a scalable frame for $\mathbb{H}_s$ for all $1\leq s \leq p$.
Now, suppose that for each $1\leq s \leq p$, the system $ \{ A^j {\bf v}_{s}\}_{j=0}^{L_s}$ is a scalable frame for $\mathbb{H}_s$. Then, there exist some scaling coefficients $w_{s,j}$, $1\leq s \leq p$, $0\leq j\leq L_s$, such that $ \{ w_{s,j}A_s^{j} {\bf v}_{s} | 0\leq j \leq L_s\}$ is a Parseval frame for each $s=1, \ldots p$.
\end{proof}
\subsection{Scalable dynamical frames for $\mathbb R^2$ and $\mathbb R^3$}\label{blocks}
For the classification of a tight frame in this section, we use the notion of the {\it diagram vector}.
For any \({\bf f} \in\mathbb{R}^n\), we define the diagram vector associated with \({\bf f}\), denoted \(\tilde{{\bf f}}\), by
\begin{equation*}
\tilde{{\bf f}} =
\frac{1}{\sqrt{n-1}}
\left( \begin{array}{c}
{\bf f}(1)^2-{\bf f}(2)^2\\ \vdots \\ {\bf f}(n-1)^2 -{\bf f}(n)^2 \\
\sqrt{2n}{\bf f}(1){\bf f}(2) \\ \vdots \\ \sqrt{2n}{\bf f}(n-1){\bf f}(n)
\end{array} \right)
\in\mathbb{R}^{n(n-1)\times 1},
\end{equation*}
where the difference of squares
${\bf f}(i)^2- {\bf f}(j)^2$ and the
product \({\bf f}(i){\bf f}(j)\) occur exactly once for \(i < j, \ i = 1, 2, \cdots, n-1.\)
Analogously, for any vector \({\bf f}\in\mathbb{C}^n\), we define the diagram vector associated with \({\bf f}\), denoted \(\tilde{{\bf f}}\), by
\begin{equation*}
\tilde{{\bf f}} =
\frac{1}{\sqrt{n-1}}
\left( \begin{array}{c}{\bf f}(1) \overline{{\bf f}(1)}-{\bf f}(2)\overline{{\bf f}(2)} \\ \vdots \\ {\bf f}(n-1)\overline{{\bf f}(n-1)}-{\bf f}(n)\overline{{\bf f}(n)} \\
\sqrt{n}{\bf f}(1) \overline{{\bf f}(2)} \\ \sqrt{n} \overline{{\bf f}(1)} {\bf f}(2) \\ \vdots \\ \sqrt{n}{\bf f}(n-1)\overline{{\bf f}(n)}
\\ \sqrt{n} \overline{{\bf f}(n-1)} {\bf f}(n)
\end{array} \right) \in\mathbb{C}^{3n(n-1)/2},
\end{equation*}
where the difference of the form
${\bf f}(i) \overline{{\bf f}(i)} - {\bf f}(j) \overline{{\bf f}(j)}$ occurs exactly once for \(i < j, \ i = 1, 2, \cdots, n-1\) and the
product of the form \({\bf f}(i) \overline{{\bf f}(j)} \) occurs exactly once for \(i \neq j.\)
The diagram vectors give us the following characterizations of tight frames and scalable frames:
\begin{thm}
\label{charTight}\cite{ CKLMNS13, CKLMNPS14}
Let \(\{{\bf f}_i\}_{i=1}^k\) be a sequence of vectors in \( \mathbb{H} \), not all of which are zero. Then \(\{{\bf f}_i\}_{i=1}^k\) is a tight frame if and only if \(\sum_{i=1}^k\tilde{{\bf f}_i}=0\).
\end{thm}
\begin{thm}\label{charScale}\cite{CKLMNS13, CKLMNPS14}
Let \(\{{\bf f}_i\}_{i=1}^k\) be a unit-norm frame for $\mathbb{H}$ and $c_1, \cdots, c_k$ be nonnegative numbers, which are not all zero.
Let $\tilde{G}$ be the Gramian associated to the diagram vectors \(\{ \tilde{{\bf f}}_i\}_{i=1}^k\) .
Then $\{c_i {\bf f}_i\}_{i=1}^k $ is a tight frame for $\mathbb{H}$ if and only if
${\bf f} = \left( c_1^2 \ldots c^2_k \right)^T$
belongs to the null space of $\tilde{G}$.
\end{thm}
Let $\{{\bf e}_1, \ldots, {\bf e}_n\}$ be the standard orthonormal basis in $\mathbb R^n$ or $\mathbb C^n$.
\begin{proposition}\label{niceR2example}
Let $A= \left( \begin{array}{cc}
a & c \\
b & d
\end{array}\right)$ be an operator in $\mathbb R^2$, where $a, b, c, d$ are not all zeros.
If $a=0$ and $b \neq 0$, then $F_{{\bf e}_1}^1 $ is a scalable frame for $\mathbb R^2$.
\end{proposition}
\begin{proof}
If $a=0$ and $b\neq 0$, then $F_{{\bf e}_1}^1 = \{ (1,0)^T, (0,b)^T\}$. Since the two vectors in $F_{{\bf e}_1}^1 $ are orthogonal, $F_{{\bf e}_1}^1 $ is a strictly scalable frame for $\mathbb R^2$.
\end{proof}
We highlight that, when $b=d \neq 0$ and $c=-d/4$ in Proposition \ref{niceR2example}, the matrix $A$ is non-diagonalizable yet generates a scalable frame for $\mathbb R^2$.
\begin{proposition}\label{2tight}
Let $a, b, c, d $ be real numbers such that
$a \neq -d$, \[b= \frac{\pm 1}{a+d}\sqrt{\frac{a^2(a+d)^2 + (a+d)^2 +a^2}{1+(a+d)^2}}, \text{ and }\]
\[ c = \mp a(ad+a^2+1) \sqrt{\frac{1+(a+d)^2}{(a+d)^2 +a^2(a+d)^2 +a^2}}.\]
Then the operator $A= \left( \begin{array}{cc}
a & c \\
b & d
\end{array}\right)$ in $\mathbb R^2$ generates a tight frame
$$F_{{\bf e}_1}^2 =\left( \begin{array}{ccc}
1 & a & a^2 + bc \\
0 & b & ab+bd
\end{array}\right) . $$
\end{proposition}
\begin{thm}\label{2scale}
Let $a, b, c, d $ be real numbers such that
$a>0$ and $abcd \neq 0$. Then the following two statements are equivalent:
\begin{enumerate}
\item $0< -\frac{ac}{bd} <1$.
\item The system
$$ F= \left( \begin{array}{ccc}
1& a & c \\
0 & b & d
\end{array}\right)$$
is a strictly scalable frame for $\mathbb R^2$.
\end{enumerate}
\end{thm}
\begin{proof}
We first note that the condition
$0< -\frac{ac}{bd} <1$
is equivalent to
($a>0, \, -\frac{b}{c}> \frac{a}{d} >0$) or ($a>0, \, -\frac{d}{a}> \frac{c}{b} >0$). \\
$(1)\Rightarrow(2)$: \quad
The conditions $a>0, \, -\frac{b}{c}> \frac{a}{d} >0$ imply that
$$ d>0, \, ad-bc>0, \, \frac{ac}{bd} > -1$$
and the conditions $a>0, \, -\frac{d}{a}> \frac{c}{b} >0$
imply that $$ d<0, \, ad-bc < 0, \, \frac{ac}{bd} > -1. $$
Then
$$ x=\sqrt{ \frac{ac}{bd}+1}, \, y=\sqrt{ \frac{c}{-b(ad-bc)}}, \, z=\sqrt{ \frac{a}{d(ad-bc)}} $$
are positive numbers and
$$ F= \left( \begin{array}{ccc}
x & ya & zc \\
0 & yb & zd
\end{array}\right)$$
is a Parseval frame for $\mathbb R^2$. \\
$(1)\Leftarrow(2)$: \quad
It the system $F$ is strictly scalable, then the normalized system
$$F'=\left( \begin{array}{ccc}
1 & \frac{a}{\sqrt{a^2 + b^2}} & \frac{c}{\sqrt{c^2+d^2}} \\
0 & \frac{b}{\sqrt{a^2 + b^2}} & \frac{d}{\sqrt{c^2+d^2}}
\end{array}\right)$$
is a unit-norm scalable frame. By Theorem \ref{charScale}, the Gramian matrix of diagram vectors of $F'$ has positive scalings in its null space:
\begin{equation}\label{2x3e1}
\frac{a^2cd-abc^2+abd^2-b^2cd}{ab(c^2+d^2)}>0,
\end{equation}
\begin{equation}\label{2x3e2}
\frac{-cd(a^2+b^2)}{ab(c^2+d^2)}>0.
\end{equation}
Inequality (\ref{2x3e2}) implies that $ -\frac{ac}{bd}>0$.
Next we show that $-\frac{ac}{bd}<1$. \\
In case $b>0$, inequality (\ref{2x3e1}) implies that
$$ a^2cd+abd^2 > bc ( ac +bd).$$
If ($c>0$ and $ac +bd \ge 0$) or ($c<0$ and $ac +bd \le 0$), then $ a^2cd+abd^2 >0$, which implies $-\frac{ac}{bd}<1$.
If $c>0$ and $ac +bd < 0$, then $ ac < -bd$, which implies $1 < -\frac{bd}{ac}$ since $ac>0$.
Similarly, if $c<0$ and $ac +bd > 0$, then $ ac > -bd$, which implies $1 < -\frac{bd}{ac}$ since $ac<0$.
This is equivalent to $-\frac{ac}{bd}<1$. \\
In case $b<0$, suppose that $-\frac{ac}{bd} \ge 1$. Multiply both sides by the positive number $-abd^2$. On one hand we have
$ a^2cd \ge -abd^2 $ and on the other hand, from inequality (\ref{2x3e1}), we have
$a^2cd-abc^2 <-abd^2+b^2cd$. Since $ a^2cd \ge -abd^2 $, we have
$-abd^2-abc^2 <-abd^2+b^2cd$, which implies $-\frac{ac}{bd} < 1$. This contradicts our assumption.
\end{proof}
This observation provides us the conditions for a dynamical operator $A$ in $\mathbb R^2$ to generate a scalable frame $F_{{\bf e}_1}^2 $ for $\mathbb R^2$.
\begin{corollary}\label{2x3scale}
Let $a, b, c, d $ be real numbers such that
$a>0$ and $0< -\frac{a(a^2+bc)}{b^2(a+d)}<1$.
Then the operator $A= \left( \begin{array}{cc}
a & c \\
b & d
\end{array}\right)$ generates a strictly scalable frame
$$F_{{\bf e}_1}^2 =\left( \begin{array}{ccc}
1 & a & a^2 + bc \\
0 & b & ab+bd
\end{array}\right) .$$
\end{corollary}
If $2 \sin^2(\omega)-1 >0$, then the operator
$$A= \left( \begin{array}{cc}
\cos(\omega) & -\sin(\omega) \\
\sin(\omega) & \cos(\omega)
\end{array}\right)$$
satisfies the condition on Theorem \ref{2scale}. Consequently we have:
\begin{ex}
Let
$$A= \left( \begin{array}{cc}
\cos(\omega) & -\sin(\omega) \\
\sin(\omega) & \cos(\omega)
\end{array}\right),$$
where $2 \sin^2(\omega)-1 >0$.
Then the operator $A$ generates a strictly scalable frame
$$F_{{\bf e}_1}^2 = \left( \begin{array}{ccc}
1& \cos(\omega) & \cos(2\omega) \\
0& \sin(\omega) & \sin(2\omega)
\end{array}\right). $$
\end{ex}
\begin{proposition}\label{2x4scale}
Let $a, b, c, d $ be real numbers such that $abcd<0$.
Then the system
$$ F= \left( \begin{array}{cccc}
1 & 0& a & c \\
0 & 1 & b & d
\end{array}\right)$$
is a strictly scalable frame for $\mathbb R^2$.
\end{proposition}
\begin{proof}
We define
$$ p= \sqrt{ \left( \frac{acd}{b} -c^2 \right) s^2 + 1 }, \,q= \sqrt{ \left(\frac{bcd}{a} +d^2\right) s^2 + 1}, \, r =\sqrt{-\frac{cd}{ab}}. $$
For any $a, b, c, d$ such that $abcd<0$, one can select $s$ such that $p>0$ and $q>0$. Those choices of $p, q, r, s$ guarantee that
the system
$$ F= \left( \begin{array}{cccc}
p &0 & ra & sc \\
0 & q & rb & sd
\end{array}\right)$$
is a Parseval frame.
\end{proof}
\begin{corollary}\label{2x4scale}
Let $a, b$ be real numbers such that $a+b^2<0$.
Then the operator $A= \left( \begin{array}{cc}
0 & a \\
1 & b
\end{array}\right)$ generates a strictly scalable frame $F_{{\bf e}_1}^3 $ for $\mathbb R^2$.
\end{corollary}
We next explore when a dynamical operator $A$ generates a scalable frame $F_{{\bf e}_1}^3 $ in $\mathbb R^3$.
We first observe the following systems in $\mathbb R^3$ when $ab \neq 0$
\begin{equation}\label{twosystemsdynscal}
F1 = \left( \begin{array}{ccccc}
1 & 0 & 0 & x & y \\
0 & 1 & 0 & a& c \\
0 & 0 & 1& b & d \\
\end{array}\right), \quad
F2 = \left( \begin{array}{ccccc}
1 & 0 & x & y \\
0 & 1 & a& c \\
0 & 0 & b & d \\
\end{array}\right).
\end{equation}
If $F$ is a tight frame, by Theorem \ref{charTight}, we have
\begin{equation} \label{onlytwo}
\begin{array} {ccc}
ax + cy &=& 0\\
bx+dy &=&0 \\
ab+cd&=& 0,
\end{array}
\end{equation}
which implies that $x=y=0$.
That is, the last two vectors have only two nonzero elements in the same entries.
We note that if the first column of $A$ is ${{\bf e}}_1$, then the system $F_{{\bf e}_1}^3 $ can not be a frame for $\mathbb R^3$.
Let
\begin{equation}\label{genmatrR3}
A = \left( \begin{array}{ccc}
0 & a & x\\
1 & b & y \\
0 & c & z
\end{array}\right).
\end{equation}
Then the corresponding $F_4$ system has the following entries:
$$
F_{{\bf e}_1}^3 = \left( \begin{array}{cccc}
1 & 0 & a & ab+cx\\
0 & 1 & b & b^2+cy + a \\
0 & 0 & c & bc +cz
\end{array}\right).
$$
By (\ref{onlytwo}), for the system $F_{{\bf e}_1}^3 $ to be a strictly scalable frame, we need to assume
$ a=ab+cx=0$ or $b = b^2+cy + a =0$.
We first consider the case $ a=ab+cx=0$.
\begin{proposition}\label{prop7import}
Let $a, b, c, d $ be real
numbers such that
$a>0$ and $0< -\frac{a(a^2+bc)}{b^2(a+d)}<1$.
Then the operator
\begin{equation}\label{nonhermandherm}
A = \left( \begin{array}{ccc}
0 & 0 & 0\\
1 & a & c \\
0 & b & d
\end{array}\right)
\end{equation}
generates a strictly
scalable frame
\begin{equation}\label{26}
F_{{\bf e}_1}^3 = \left( \begin{array}{cccc}
1 & 0& 0 & 0\\
0 & 1 & a & a^2 + bc \\
0 & 0& b & ab+bd
\end{array}\right).
\end{equation}
\end{proposition}
\begin{proof}
This follows from Theorem \ref{stackScale} and Theorem \ref{2scale}.
\end{proof}
When $b = b^2+cy + a =0$, we have
$$ A = \left( \begin{array}{ccc}
0 & a & x\\
1 & 0 & -a/c \\
0 & c & cz
\end{array}\right).
$$
By applying row and column permutations, $F_{{\bf e}_1}^3 $ can be written in the same form as (\ref{26}).
Similarly, the following operator, with a suitable choice of the second and third column:
\begin{equation}\label{genmatrR3}
A = \left( \begin{array}{ccc}
0 & a & x\\
0 & b & y \\
1 & c & z
\end{array}\right)
\end{equation}
generates a scalable frame $F_{{\bf e}_1}^3 $, which also can be written in the same form as (\ref{26}).
We note that any tight or scalable frame in $\mathbb R^n$ with $n$ frame vectors is an orthogonal basis. A trivial example of a scalable dynamical frame is the following:
\begin{ex}\label{examplelemma} Let
\begin{equation}\label{companion}
A = \left( \begin{array}{cc}
0 & 1 \\
I_{n-1}& 0 \\
\end{array}\right) . \end{equation}
Then the sequence $F_{{\bf e}_1} ^ L $ is a scalable frame of $\mathbb R^n$ if and only if $L \geq n$.
\end{ex}
For instance, when $n= L=3$, the resulting frame is $F_{{\bf e}_1}^3 =\{ {\bf e}_1, {\bf e}_2,{\bf e}_3, {\bf e}_1\}$, and the scaled frame $ \{ {2}^{-1/2}{\bf e}_1, {\bf e}_2,{\bf e}_3,2^{-1/2}{\bf e}_1\}$ is a Parseval frame.
Notice that \eqref{companion} is an example of a companion \cite{HJ85} operator. It makes sense to explore the conditions under which a companion operator generates a scalable frame.
\section{Companion operators and generalizations}\label{compansection}
Let $a_1, \ldots, a_n \in \mathbb R$ which are not all zeros, then
\begin{equation}\label{companiondef}
A = \left( \begin{array}{c|c}
0 & a_1 \\
\hline
& a_2\\
I_{n-1}& \vdots \\
& a_n\\
\end{array}\right) \end{equation}
is called a companion operator \cite{HJ85}.
\begin{proposition}
Let the dynamical operator $A$ be a companion operator \eqref{companiondef} in $\mathbb R^n$, then we have
\begin{enumerate}
\item $ F_{{\bf e}_1}^{n-1} = I. $
\item for any orthogonal matrix $U$, the operator $UAU^{-1}$ generates an orthonormal basis $U$.
\end{enumerate}
\end{proposition}
It is known that the standard orthonormal basis $B$ can not be extended to a scalable frame by adding one vector ${\bf f} \in \mathbb{H} \setminus B$,
\cite{DKN15, KOF13}. Thus we explore when one can generate a dynamical frame by adding two vectors.
Although a companion operator $A$ does not generate a scalable frame $F_{{\bf e}_1}^{n} $, it can generate a scalable frame $F_{{\bf e}_1}^{n+1} $ under certain conditions.
Using the companion operator $A$, we have
\begin{equation}\label{thissystemisframe}
F_{{\bf e}_1}^{n} = ({\bf e}_1 \ldots {\bf e}_{n} \, \, {\bf f} ), \quad
F_{{\bf e}_1}^{n+1} = ({\bf e}_1 \ldots {\bf e}_{n} \, \, {\bf f} \,\, {\bf g}),
\end{equation}
where
\[ {\bf f}= \left( \begin{array}{c}
a_1 \\
a_2 \\
a_3\\
\vdots\\
a_{n-1} \\
a_n
\end{array}\right) \text{ and }
{\bf g}= \left( \begin{array}{c}
a_1a_n \\
a_1+ a_2a_n \\
a_2+ a_3a_n\\
\vdots\\
a_{n-2} + a_{n-1} a_n\\
a_{n-1} +a_n^2
\end{array}\right).
\]
Similar calculations as in observation (\ref{onlytwo}) produce the following result:
\begin{proposition}
\label{ext}
\cite{DKN15} Let $\{{{\bf e}}_1, \ldots {{\bf e}}_n\}$ be the standard orthonormal basis in $\mathbb R^n$ with $n \ge 2$. Let ${\bf f}$ and ${\bf g}$ be two unit-norm vectors in $\mathbb R^n$.
If either system $\{{{\bf e}}_1, \ldots {{\bf e}}_n, {\bf f}, {\bf g}\}$ or $\{{{\bf e}}_1, \ldots {{\bf e}}_{n-1}, {\bf f}, {\bf g}\}$ is scalable, then
${\bf f}$ and ${\bf g}$ have only two nonzero elements in the same entries.
\end{proposition}
We now assume that $F_{{\bf e}_1}^{n+1} $ is scalable. Then by Proposition \ref{ext},
$a_m=0$ implies that $a_{m-1}=0$ for $m\ge 2$. This implies that $a_1=\ldots = a_{n-2}=0$.
\begin{proposition}\label{companionstandardresult}
Let $a$ and $b$ be real numbers such that
$a>0$ and $0< -\frac{a^2}{a+b^2}<1$.
Then the companion operator $A$ in $\mathbb R^n$,
\begin{equation}\label{gennonherm}
A = \left( \begin{array}{cccccc}
0 & 0 & ...&0& 0 & 0 \\
1 & 0 & ...&0& 0& 0 \\
&.&.&.&.& \\
0 & 0 & ...&1 & 0 & a\\
0 & 0 & ...&0& 1 & b
\end{array}\right)
\end{equation}
generates a strictly scalable frame $F_{{\bf e}_1}^{n+1} $.
\end{proposition}
\begin{proof}
We have
\begin{equation}\label{nicegen}
F_{{\bf e}_1}^{n+1} =\left( \begin{array}{ccccc}
I_{n-2} &&&\\
& 1& 0 & a & ab \\
& 0& 1& b & a+b^2 \\
\end{array}\right).
\end{equation}
The strict scalability follows from Theorem \ref{2scale} and Theorem \ref{stackScale}.
\end{proof}
We note that the operator $A$ in (\ref{gennonherm}) is not diagonalizable. Next, we generalize the structure of $A$ while ensuring that the new matrix generates scalable frames by iterative actions.
\begin{ex}\label{g1}
Let $a$ and $b$ be real numbers such that $0< -\frac{a(a^2+bc)}{b^2(a+d)}<1$ and
$a>0$.
Then the operator
\begin{equation}\label{nicegen}
A =\left( \begin{array}{ccccc}
0 & 0 & 0 & \ldots & 0\\
1 & 0 & 0 &\ldots & 0\\
0 & 1 & 0 & \ldots & 0\\
\vdots & \vdots & \vdots &\vdots\\
0 & \ldots & 1 & a & c \\
0 & \ldots & 0 & b & d
\end{array}\right)
\end{equation}
generates a strictly scalable frame $F_{{\bf e}_1}^{n} $ for $\mathbb R^n$.
\end{ex}
\begin{proof}
We have
\begin{equation}\label{nicegen}
F_{{\bf e}_1}^{n} =\left( \begin{array}{cccc}
I_{n-2} &&&\\
& 1& a & a^2+bc \\
& 0& b & ab+bd \\
\end{array}\right).
\end{equation}
The strict scalability follows by Proposition \ref{2tight} and Proposition \ref{stackScale}.
\end{proof}
\begin{ex}
Let $2 \sin^2(\omega)-1 >0$. Then
\begin{equation}\label{nicegen}
A =\left( \begin{array}{ccccc}
0 & 0 & 0 & \ldots & 0\\
1 & 0 & 0 &\ldots & 0\\
0 & 1 & 0 & \ldots & 0\\
\vdots & \vdots & \vdots &\vdots\\
0 & \ldots & 1 & \cos(\omega) & -\sin(\omega) \\
0 & \ldots & 0 & \sin(\omega) & \cos(\omega)
\end{array}\right)
\end{equation}
generates a strictly scalable frame $F_{{\bf e}_1}^{n}$.
\end{ex}
\begin{ex}
Let $2 \sin^2(\phi)-1 >0$ and let
\begin{equation}\label{realschursimple}
A = \left( \begin{array}{cccccc}
\pm 1 & 0 & 0 & 0 & \ldots & 0 \\
0 & \pm 1 & 0 & 0 & \ldots & 0 \\
\vdots & \vdots & \vdots & \vdots & \vdots & \vdots \\
0& 0 & \ldots & \pm 1& 0 & 0 \\
0& 0 & \ldots & 0& \cos \phi & -\sin \phi \\
0& 0 & \ldots & 0& \sin \phi & \cos \phi
\end{array}\right).\end{equation}
The set
\begin{equation}
\{ {\bf e}_{n-1}, A{\bf e}_{n-1}, A^2{\bf e}_{n-1}\} \cup \bigcup_{l=1}^{n-2} \{ {\bf e}_l, A {\bf e}_l, \ldots, A^{L_l} {\bf e}_l \}
\end{equation}
is a strictly scalable frame of $\mathbb R^n$. \end{ex}
\section{Concluding remarks and generalizations}\label{conclusion}
We have studied the scalability of dynamical frames in a separable Hilbert space $\mathbb{H}$. Given an operator $A$ on $\mathbb{H}$ and a (at most countable) set $G \subset \mathbb{H}$,
we have explored the relations between $A$, $G$ and the number of iterations that make the system \eqref{oursystem} a scalable frame. When $\dim \mathbb{H}$ is finite, and $A$ is a normal operator, we have fully answered question (Q1).
Since we have not achieved a full answer for operators which are not unitary diagonalizable, we have offered a partial answer by studying block-diagonal operators, which are not necessarily normal. Note that the block-diagonal matrix $A$ in Theorem \ref{blockresultbig} cannot be normal if one of its blocks is not normal. Also, we have established the canonical dual frame for frames of type $F_G(A)$; in particular, we showed that the canonical dual frame has, as anticipated, an iterative set structure. This result holds true in any separable Hilbert space $\mathbb{H}$.
We now pose a new question, which is a generalization of (Q1):
\vspace{2.1mm}
(Q3) \textit{ Given multiple operators $A_s$, $s \in I$ on a separable Hilbert space $\mathbb{H}$, and one fixed vector ${\bf v} \in \mathbb{H}$, when is the system $\cup_{s \in I} \{ A_s ^j {\bf v} \}_{j=0}^{L_s}$ a (scalable) frame for $\mathbb{H}$?}
\vspace{2mm}
The next example shows how to generate a scalable frame for $\mathbb R^3$ using two dynamical operators.
\begin{ex}\label{multigenscalable} {Let } $\alpha =2\pi/3$,
\[ A_1 = \left( \begin{array}{ccc}
\cos{\alpha } & -\sin{\alpha} &0\\
\sin{\alpha} & \cos{\alpha } &0 \\
0&0&0
\end{array}\right), \; \text{and} \;
A_2 = \left( \begin{array}{ccc}
\cos{\alpha } & 0 &-\sin{\alpha } \\
0 &0&0 \\
\sin{\alpha } &0 & \cos{\alpha} \end{array}\right). \]
$$ \text{Then } \; \{ {\bf e}_1, A_1 {\bf e}_1, A_1^2 {\bf e}_1, A_2 {\bf e}_1, A_2^2 {\bf e}_1\} \; \text{is a strictly scalable frame for $\mathbb R^3$.}$$
\begin{center}
\begin{tikzpicture}[scale=1.75]
\draw [red] (-1,0) arc (180:360:1cm and 0.5cm);
\draw[red, dashed] (-1,0) arc (180:0:1cm and 0.5cm);
\draw [blue](0,1) arc (90:270:0.5cm and 1cm);
\draw[blue, dashed] (0,1) arc (90:-90:0.5cm and 1cm);
\draw (0,0) circle (1cm);
\shade[ball color=blue!10!white,opacity=0.20] (0,0) circle (1cm);
\draw[->] (-1.2,0)--(1.2, 0)node[right] {$y$};
\draw[->] (0,-1.2) -- (0,1.2) node[above] {$z$};
\draw[->] (0.5, 0.5) -- (-0.5, -0.5) node[left, below] {$x$};
\draw[ultra thick,->] (0, 0) -- (-0.45, -0.45) node[above] {${{\bf e}}_1$};
\draw[red, ultra thick, ->] (0, 0)--(0.93, 0.22 )node[right, above] {$A_1{{\bf e}}_1$};
\draw[red, ultra thick, ->] (0, 0)--(-0.67, 0.38 )node[right, above] {$A^2_1{{\bf e}}_1$};
\draw[blue, ultra thick, ->] (0, 0)--(0.28, -0.8 )node[right, above] {$A_2{{\bf e}}_1$};
\draw[blue, ultra thick, ->] (0, 0)--(0.27, 0.88 )node[right, above] {$A^2_2{{\bf e}}_1$};
\end{tikzpicture}
\end{center}
\end{ex}
The following proposition is a generalization of the principle introduced in Example \ref{multigenscalable}:
\begin{proposition}
Let $i, j, k,l \in \mathbb N$ be such that $p<k \le n, \, q <l \le n$, and let $N\in \mathbb N$.
For each $m=1, \ldots, N$, we define $A^{pq}_{kl} (m) =[ a_{ij} (m) ]_{ i, j =1}^n$ as
$$
a_{pq} (m) = a_m,\,
a_{pl} (m) = b_m,\,
a_{kq} (m) = c_m,\,
a_{kl} (m) = d_m.
$$
If for each $m =1, \ldots, N$, $a_m, b_m, c_m$ and $d_m$ satisfy the conditions of Corollary \ref{2scale}, and the system
\begin{equation} \label{nicesystm}
\{{{\bf e}}_1 \} \cup \cup_{m=1}^N \{ A^{pq}_{kl} (m) {{\bf e}}_1, (A^{pq}_{kl} (m))^2 {{\bf e}}_1 \}\end{equation} spans $\mathbb R^n$, then
\eqref{nicesystm} is a strictly scalable frame for $\mathbb R^n$.
\end{proposition}
\begin{proof}
By Corollary \ref{2scale}, the set
$ \{{{\bf e}}_1 \} \cup \{ A^{pq}_{kl} (m) {{\bf e}}_1, (A^{pq}_{kl} (m))^2 {{\bf e}}_1 \} $ is a scalable frame for a 2-dimensional subspace for each $m =1, \ldots, N$. Thus, there exist some suitable scaling coefficients $x(m), y(m), z(m)$, and by Theorem \ref{charTight},
$$ \widetilde{x(m){{\bf e}}_1} + \widetilde{y(m)A^{pq}_{kl} (m) {{\bf e}}_1} + \widetilde{z(m) A^{pq}_{kl} (m))^2 {{\bf e}}_1 }=0 .$$
This implies that the system \eqref{nicesystm}
is a scalable frame for $\mathbb R^n$.
\end{proof}
{For a frame generated by iterative actions of multiple operators, that is, a { \it multi-dynamical} frame, we find that its canonical dual frame is also multi-dynamical:}
\begin{thm}
Let $A_s$, $s \in I$, be operators on a separable Hilbert space $\mathbb{H}$, let $L_s \geq 0$, and fix a vector ${\bf v} \in \mathbb{H}$. If $\cup_{s \in I} \{ A_s ^j {\bf v} \}_{j=0}^{L_s}$ is a frame for $\mathbb{H}$, with frame operator $S$, then its canonical dual frame is
\begin{equation}
\cup_{s \in I} \{ B_s ^j {\bf f} \}_{j=0}^{L_s},
\end{equation}$\text{where} \; {\bf f} = S^{-1}{\bf v}, \; \text{and} \; B_s= S^{-1}A_s S, \, s \in I.$
\end{thm}
\begin{proof}
If $S$ denotes the frame operator of the frame $\cup_{s} \{ A_s ^j {\bf v} \}_{j=0}^{L_s}$ for $\mathbb{H}$, then its canonical dual frame elements are $S^{-1}A_s ^j {\bf v}$. Since $ B_s^j = S^{-1}A_s^j S$, we obtain that the dual frame elements are
$$S^{-1}A_s ^j {\bf v} =S^{-1}A_s ^j S S^{-1} {\bf v} =S^{-1}A_s ^j S {\bf f} = B_s^j {\bf f}.$$
\end{proof}
\section*{Acknowledgement}
We express our gratitude to Professor S. Narayan for many helpful conversations on this work.
Kim was supported by the Central Michigan University FRCE Research Type A Grant \#C48143. Aceska was supported by the BSU Aspire Research Grant ``Frame Theory and Modern Sampling Strategies''.
|
\section{Introduction}
It is well known \cite{Bek1} that the thermodynamic description of
mundane physical systems breaks down in the low-temperature regime
$T\sim \hbar/R$ \cite{Noter,Noteunit}, when the characteristic
thermal wavelengths $\lambda_{\text{thermal}}\sim \hbar/T$ are no
longer small on the scale $R$ set by the spatial size of the system.
The physical properties of these low-temperature systems are
dominated by quantum (rather than thermodynamic) effects. Thus, the
thermodynamic description of mundane physical systems is known to be
restricted to the high-temperature regime \cite{Bek1}
\begin{equation}\label{Eq1}
T\times R\gg \hbar\ .
\end{equation}
Interestingly, black holes are unique in this respect. It is well
known that the Bekenstein-Hawking temperature \cite{Bek2,Haw} of the
Schwarzschild black hole is given by $T_{\text{BH}}=\hbar/4\pi
r_{\text{H}}$, where $r_{\text{H}}=2M$ is the radius of the
black-hole horizon. Thus, Schwarzschild black holes are
characterized by the relation $T_{\text{BH}}\times
r_{\text{H}}\sim\hbar$.
Moreover, the Bekenstein-Hawking temperature of Kerr black holes is
given by
\begin{equation}\label{Eq2}
T_{\text{BH}}={{\hbar(r_+-r_-)}\over{4\pi r^2_+}}\ ,
\end{equation}
where $r_{\pm}=M+(M^2-a^2)^{1/2}$ are the black-hole (outer and
inner) horizon radii \cite{Notema}. Thus, rapidly-rotating Kerr
black holes in the near-extremal $r_+-r_-\ll r_+$ regime are
characterized by the strong inequality
\begin{equation}\label{Eq3}
T_{\text{BH}}\times r_+\ll \hbar\ .
\end{equation}
It is quite remarkable that black holes have a well defined
thermodynamic behavior in the low-temperature regime (\ref{Eq3}),
where mundane physical systems are governed by quantum effects and
no longer have a self-consistent thermodynamic description.
One naturally wonders whether the thermodynamic description of black
holes is valid all the way down to the zero temperature $T\times
r_+\to 0$ limit? In order to address this interesting question, we
shall analyze in this paper the regime of validity of the first law
of thermodynamics for black holes.
\section{Black holes and the first law of thermodynamics}
For a closed physical system with a well defined temperature $T$, a
change $\Delta E$ in the energy of the system results with a change
\cite{Lan,Notevol}
\begin{equation}\label{Eq4}
\Delta S={{\Delta E}\over{T}}\
\end{equation}
in the entropy of the system. This famous differential relation,
known as the first law of thermodynamics, is one of the most
important features of the thermodynamic description of mundane
physical systems.
It is often stated in the physical literature (see e.g.,
\cite{Mech}) that black holes, like ordinary thermodynamic systems,
also obey this law. However, the regime of validity of the standard
thermodynamic description of black holes has never been discussed in
the literature. In the present paper we would like to raise the
following intriguing question: Do black holes {\it always} obey the
first law of thermodynamics?
In order to address this interesting question, it proves useful to
examine carefully the assumptions made in the physical literature in
deriving the first law of black-hole thermodynamics. The
characteristic Bekenstein-Hawking entropy of a black hole is given
by a quarter of its horizon area \cite{Bek2,Haw}:
\begin{equation}\label{Eq5}
S_{\text{BH}}={{A}\over{4\hbar}}\ .
\end{equation}
Remembering that the surface area of Kerr black holes \cite{Notekr}
is given by $A=4\pi(r^2_++a^2)=8\pi Mr_+$, one finds \cite{Noteja}
\begin{equation}\label{Eq6}
S_{\text{BH}}={{2\pi M\{M+[M^2-(J/M)^2]^{1/2}\}}\over{\hbar}}\ .
\end{equation}
Consider now a physical process which changes the energy (mass) of
the black hole by a small amount $\Delta E\ll M$ \cite{Noteby}. From
Eq. (\ref{Eq6}) one finds that the resulting change in the
Bekenstein-Hawking entropy of the black hole is given by the rather
cumbersome expression
\begin{equation}\label{Eq7}
{{\hbar}\over{2\pi}}\Delta S_{\text{BH}}=(M+\Delta E) \Big\{M+\Delta
E+\Big[(M+\Delta E)^2-\Big({{J}\over{M+\Delta
E}}\Big)^2\Big]^{1/2}\Big\}-M\Big\{M+\Big[M^2-\Big({{J}\over{M}}\Big)^2\Big]^{1/2}\Big\}\
.
\end{equation}
A careful inspection of Eq. (\ref{Eq7}) reveals that {\it if} $M
\Delta E\ll (r_+-r_-)^2$ then, to leading-order in the small ratio
$M \Delta E/(r_+-r_-)^2$, the changes in the black-hole entropy and
energy are related to each other by the standard first law of
thermodynamics,
\begin{equation}\label{Eq8}
\Delta S_{\text{BH}}={{\Delta E}\over{T_{\text{BH}}}}\ ,
\end{equation}
where $T_{\text{BH}}$ as given by (\ref{Eq2}) is the familiar
Bekenstein-Hawking temperature of the black hole.
On the other hand, in the opposite regime $M \Delta E\gg
(r_+-r_-)^2$ one finds from (\ref{Eq7}) the non-standard relation
\begin{equation}\label{Eq9}
\Delta S_{\text{BH}}={{{\sqrt 8}\pi}\over{\hbar}}M^{3/2}\sqrt{\Delta
E}
\end{equation}
between the changes in the physical parameters (entropy and energy)
of the black hole.
One therefore concludes that, for black holes in Einstein gravity,
the validity of the standard first law of thermodynamics, Eq.
(\ref{Eq8}), is restricted to the regime
\begin{equation}\label{Eq10}
{{(r_+-r_-)^2}\over{r_+}}\gg \Delta E\geq\Delta E_{\text{min}}\ ,
\end{equation}
where $\Delta E_{\text{min}}$ is the smallest possible change in the
energy (mass) of the black hole. Thus, in order to determine the
regime of validity of the law in the context of black-hole physics,
one should first determine the value of the fundamental physical
parameter $\Delta E_{\text{min}}$.
How small can $\Delta E_{\text{min}}$, the minimal change in the
energy (mass) of a black hole, be made? The answer to this question
at the {\it classical} level was given by Christodoulou and Ruffini
\cite{Chr}: the capture of a point particle by a black hole is
characterized by the relation $\Delta E_{\text{min}}=0$ if the
particle is captured at the black-hole horizon from a radial turning
point of its motion. In this scenario the energy (as measured by
asymptotic observers) of the absorbed particle is completely
red-shifted. Substituting $\Delta E_{\text{min}}=0$ into
(\ref{Eq10}) one deduces that, at the classical level, the first law
of thermodynamics may be valid all the way down to the extremal
limit $T_{\text{BH}}\to 0$.
However, as emphasized by Bekenstein in his seminal work
\cite{Bek2}, the classical limit of a perfectly localized particle
(a point particle) is physically unacceptable in a self-consistent
{\it quantum} theory of relativity. In particular, due to the
quantum uncertainty principle \cite{Bor}, the particle cannot be
localized at the black-hole horizon without having a non-zero radial
momentum (kinetic energy). Specifically, the Heisenberg uncertainty
principle sets a lower bound on the smallest possible energy
delivered to the black hole by the captured particle \cite{Bek2}:
\begin{equation}\label{Eq11}
\Delta E_{\text{min}}=2\pi T_{\text{BH}}\ .
\end{equation}
Substituting (\ref{Eq11}) into (\ref{Eq10}) one finds that, within
the framework of a self-consistent quantum theory of gravity, the
validity of the first law of thermodynamics, Eq. (\ref{Eq8}), is
restricted to the regime \cite{Notecaf}
\begin{equation}\label{Eq12}
T_{\text{BH}}\times r_{\text{+}}\gg
\Big({{\hbar}\over{r_{\text{+}}}}\Big)^2\ .
\end{equation}
\section{Summary}
The seminal works of Bekenstein and Hawking have revealed that
gravity has a well-defined thermodynamic description. In particular,
it is often stated in the physical literature that black holes, like
mundane physical systems, obey the first law of thermodynamics:
$\Delta E=T_{\text{BH}}\Delta S$, where $T_{\text{BH}}$ is the
Bekenstein-Hawking temperature of the black hole. In the present
paper we have explored the regime of validity of this law. In
particular, we have shown that,
due to quantum effects, the first law of thermodynamics breaks down
in the low-temperature regime
\begin{equation}\label{Eq13}
T_{\text{BH}}\times r_{\text{H}}\lesssim
\Big({{\hbar}\over{r_{\text{H}}}}\Big)^2
\end{equation}
of near-extremal black holes.
\bigskip
\noindent {\bf ACKNOWLEDGMENTS}
\bigskip
This research is supported by the Carmel Science Foundation. I thank
Yael Oren, Arbel M. Ongo, Ayelet B. Lata, and Alona B. Tea for
stimulating discussions.
\newpage
|
\section{Introduction}
Special flows built over the rotation $R_\alpha$ of the circle by an irrational number $\alpha$ and under a piecewise $C^1$-function $f$ with non-zero sum of jumps were introduced and studied by J.~von~Neumann in \cite{vN32}.
He proved that such flows have continuous spectrum (i.e.\ are weakly mixing) for each irrational rotation.
The weak mixing result was generalised in two directions:
the $C^1$-condition on $f$ was replaced by absolute continuity in \cite{ILM99},
while A.~Katok, \cite{Ka01}, proved weak mixing replacing $R_\alpha$ by any ergodic interval exchange transformations.
Recall also that it follows from a result by A.~Ko\v{c}ergin, \cite{Ko72}, that von Neumann flows are never mixing.
As a matter of fact K.~Fr\c{a}czek and M.~Lema\'{n}czyk, \cite{FL04}, proved that von Neumann flows are spectrally disjoint from all mixing flows.
Moreover, whenever $\alpha$ has bounded type, von Neumann flows are mildly mixing, \cite{FL06}.
On the other hand, nothing is known about spectral multiplicity and rank properties of von Neumann flows. Recall that rank yields an upper bound on the spectral multiplicity \cite{Ki}.
In the paper we show that the rank of von Neumann flows for which $f$ has one discontinuity is infinite.
More precisely, let $\mathcal{T}=(T_t)_{t\in\mathbb{R}}$ be a von Neumann special flow over the rotation $R_\alpha\colon\mathbb{T}\to\mathbb{T}$ by an irrational $\alpha$ and under a roof function $f\colon \mathbb{T}\to\mathbb{R}_+$ of the form
\begin{equation}\label{roof}
f(x)=g(x)+A\{x\}+c,
\end{equation}
where $g\in C^1(\mathbb{T})$, $\int_\mathbb{T} g(x)dx=0$, $A\neq 0$, $c\in\mathbb{R}$ is such that $f>0$ and $\{x\}$ stands for the fractional part of $x$.
Our main result is the following:
\begin{tw}\label{mainth}
The flow $\mathcal{T}$ does not have local rank one.
\end{tw}
The key property which is used to prove Theorem \ref{mainth} is {\em slow divergence} of orbits of nearby points in the flow direction.
Such way of divergence of orbits is characteristic for {\em parabolic systems}.
It was first observed by M.~Ratner for the class of horocycle flows, \cite{Ra83}.
Later this property (called Ratner's property) was shown to hold for some von Neumann flows under the additional assumption that $\alpha$ has bounded type \cite{FL06} and for some mixing, smooth flows on surfaces \cite{FK}, \cite{KKU}.
In all these papers, Ratner's property was used to enhance mixing properties (weak mixing to mild mixing and mixing to multiple mixing).
In \cite{K16}, a property of Ratner's type, called in \cite{K16} {\em parabolic divergence}, was used to compute {\em slow entropy} of some mixing smooth surface flows with non-degenerate saddles.
A variant of the parabolic divergence property is the main ingredient in the proof of Theorem \ref{mainth}.
Let us shortly describe this property in our setting. Notice that the divergence of orbits of two close points is caused by two effects: slow (uniform) divergence by hitting the roof and fast (non-uniform) divergence by hitting the discontinuity. It is the first type of divergence which is characteristic for parabolic systems. It turns out that either going forward or backward in time, we can always avoid the discontinuity long enough to observe the effect of the uniform divergence (see Lemma~\ref{lin} and Proposition~\ref{mainprop}).
\section{Basic definitions}
We denote by $\mathbb{T}$ the circle group $\mathbb{R}/\mathbb{Z}$ which we will identify with the unit interval $[0,1)$.
For a real number $x$ denote by $\{x\}$ its fractional part, by $[x]=x-\{x\}$ its integer part and by $\|x\|$ its distance to the nearest integer.
Given $x,y\in\mathbb{T}$ with $\|x-y\|<\frac{1}{2}$, $[x,y]$ will denote the shortest interval in $\mathbb{T}$ connecting $x$ and $y$.
Lebesgue measure on $\mathbb{T}$ will be denoted by $m_\mathbb{T}$.
\subsection{Special flows}
Let $T$ be an ergodic automorphism of a standard Borel space $(X,\mathcal{B},\mu)$ (with $\mu(X)<+\infty$).
A measurable function $f\colon X\to \mathbb{R}$ defines a cocycle $\mathbb{Z}\times X \to \mathbb{R}$ given by
$$
f^{(n)}(x) =
\begin{cases}
f(x)+f(Tx)+\cdots+f(T^{n-1}x) &\mbox{if}\quad n>0 \\
\hfil 0 &\mbox{if}\quad n=0 \\
\hfil -(f(T^{n}x)+\cdots+f(T^{-1}x)) &\mbox{if}\quad n<0.
\end{cases}
$$
Assume that $f\in L^1(X,\mathcal{B},\mu)$ is a strictly positive function.
\begin{df}\label{specialflow}
The \emph{special flow} $\mathcal{T} = (T^f_t)_{t\in\mathbb{R}}$ over the \emph{base automorphism} $T$ under the \emph{roof function} $f$ is the flow acting on $(X^f,\mathcal{B}^f,\mu^f)$, where $X^f = \{ (x,s) \in X\times\mathbb{R} \mid 0 \leq s < f(x) \}$ and $\mathcal{B}^f$ and $\mu^f$ are the restrictions of $\mathcal{B} \otimes \mathcal{B}_\mathbb{R}$ and $\mu \otimes \lambda$ to $X^f$ respectively ($\lambda$ stands for Lebesgue measure on $\mathbb{R}$).
Under the action of the flow $\mathcal{T}$ each point in $X^f$ moves vertically upward with unit speed, and we identify the point $(x,f(x))$ with $(Tx,0)$.
More precisely, for $(x,s)\in X^f$ we have
\begin{equation}\label{spflow}
T^f_t(x,s) = (T^n x, s+t-f^{(n)}(x)),
\end{equation}
where $n\in\mathbb{Z}$ is the unique number such that $f^{(n)}(x) \leq s+t < f^{(n+1)}(x)$.
\end{df}
\noindent
We will identify the base $X$ with the subset $\{ (x,0) \mid x\in X\}\subset X^f$.
Notice that if $d$ is a
metric on $X$, then $d^f((x,t),(y,s)) = d(x,y) + |t-s|$ is a metric for $X^f$.
\subsection{Von Neumann flows}
We call a function $f\colon \mathbb{T} \to \mathbb{R}$ \emph{piecewise $C^1$} if there exist $\beta_1,\ldots,\beta_k\in\mathbb{T}$ such that $f|_{\mathbb{T}\setminus \{\beta_1,\ldots,\beta_k\}}$ is $C^1$ smooth and $f_\pm(\beta)=\lim_{x\to \beta\pm} f(x)$ is assumed to be finite.
Denote $d_i:=f_-(\beta_i)-f_+(\beta_i)$ the \emph{jump} of $f$ at point $\beta_i$.
The number $\sum_{i=1}^k d_i$ is the \emph{sum of jumps} of $f$.
\begin{df}
A \emph{von Neumann flow} is a special flow $\mathcal{T}$ over a rotation $R_\alpha\colon(\mathbb{T},m_\mathbb{T})\to(\mathbb{T},m_\mathbb{T})$ by an irrational $\alpha\in\mathbb{T}$ and under a piecewise $C^1$ roof function $f\colon \mathbb{T}\to\mathbb{R}_+$ with a non-zero sum of jumps.
\end{df}
We will consider the simplest case when $f$ has only one discontinuity.
Without loss of generality we may assume that $f$ is $C^1$ on $\mathbb{T}\setminus \{0\}$ with a jump $A=f_-(0)-f_+(0)\neq 0$.
Any such $f$ can be written in the form
$$
f(x)=g(x)+A\{x\}+c,
$$
where $g\in C^1(\mathbb{T})$, $\int_\mathbb{T} g(x)dx=0$, $A\neq 0$ ($A$ is called the \emph{slope}) and $c\in\mathbb{R}$ is such that $f>0$.
We will assume that $\int_\mathbb{T} f dm_\mathbb{T}=1$, that is we normalize the resulting measure to make it a probability measure.
\subsection{Finite rank systems and systems of local rank one}
In this section we recall the notion of finite rank and local rank one.
There are several equivalent ways to define a finite rank system (see \cite{Fe97}).
We will define rank properties in the language of special flows, \cite{Fa05}.
Let $\mathcal{T}=(T_t)_{t\in\mathbb{R}}$ be an ergodic flow on a standard probability space $(X,\mathcal{B},\mu)$.
Let $B\subset X$, $H\in\mathbb{R}_+$ and $0<\eta < 1$.
\begin{df}\label{tower}
A pair $(B,H)$ is an \emph{$\eta$-tower} (or simply a \emph{tower}) for $\mathcal{T}$ (with the \emph{base} $B$ and of \emph{height} $H$) if there exists an isomorphism $\varphi\colon (X,\mu) \to (Y^f,\nu^f)$ between the flow $\mathcal{T}$ and a special flow over an ergodic $S\colon (Y,\mathcal{C},\nu) \to (Y,\mathcal{C},\nu)$ under a roof function $f\colon Y \to \mathbb{R}_+$ such that
\begin{itemize}
\item $\varphi(B) \subset Y$ and $\nu(\varphi(B))>\eta \nu(Y)$;
\item $f(y)\leq H$ for every $y\in Y$;
\item $f(y)=H$ for every $y\in \varphi(B)$.
\end{itemize}
\end{df}
\noindent
By abuse of notation, we will identify the subset $\varphi^{-1}(Y)\subset X$ with the base $Y$ and the measure $\nu\circ \varphi^{-1}$ on $\varphi^{-1}(Y)$ with $\nu$, that is we will think of the base of the corresponding suspension flow as of a subset of $X$ equipped with a finite measure $\nu$.
Note that $\nu(Y)\geq\frac{1}{H}$ (since $1=\mu(X)=\nu^f(Y^f)\leq H\nu(Y)$).
The set $\bigsqcup_{t=0}^H T_t B$ will be also referred to as a \emph{tower}.
For $0\leq t<H$ the set $T_t B$ is called a {\em level} of the tower.
Fix a finite measurable partition $\mathcal{P}$ of $X$.
For $\varepsilon > 0$ we say that a level $T_t B$ is {\em $\varepsilon$- monochromatic}
(for $\mathcal{P}$) if a $1-\varepsilon$ proportion of it (with respect to the measure $\nu\circ T_{-t}$)
is contained in one atom of the partition $\mathcal{P}$.
A tower for $(T_t)_{t\in\mathbb{R}}$ is called {\em $\varepsilon$-monochromatic} (for $\mathcal{P}$) if a $1-\varepsilon$ proportion of its levels (with respect to the Lebesgue measure $\lambda$ on $[0,H)$) is $\varepsilon$-monochromatic.
\begin{df}\label{alpharankone}
Let $\beta\in(0,1]$.
An ergodic flow $\mathcal{T}=(T_t)_{t\in\mathbb{R}}$ has \emph{local rank one of order $\beta$} if for every finite partition $\mathcal{P}$ of $X$ and every $\varepsilon>0$ there exists a $(\beta-\varepsilon)$-tower for $\mathcal{T}$ which is $\varepsilon$-monochromatic for $\mathcal{P}$.
\end{df}
\begin{df}\label{localrankone}
An ergodic flow $\mathcal{T}$ has {\em local rank one} if it has local rank one of some order $\beta\in(0,1]$.
If $\beta=1$ then $\mathcal{T}$ is said to have {\em rank one}.
\end{df}
\begin{df}\label{finiterank}
An ergodic flow $\mathcal{T}=(T_t)_{t\in\mathbb{R}}$ has \emph{finite rank} if there exists $r\in\mathbb{N}$ such that for every finite partition $\mathcal{P}$ of $X$ and every $\varepsilon>0$ there are $r$ disjoint towers $(B_1,H_1),\ldots,(B_r,H_r)$ for $\mathcal{T}$ which are $\varepsilon$-monochromatic for $\mathcal{P}$ and such that $\mu(\bigsqcup_{i=1}^r \bigsqcup_{t=0}^H T_t B_i) > 1-\varepsilon$.
We get the definition of rank one flow if $r=1$.
\end{df}
\noindent
Clearly, all the above properties are measure theoretic invariants and the following implications hold:
$$
\mbox{rank one} \; \Rightarrow \; \mbox{finite rank} \; \Rightarrow \; \mbox{local rank one} \; \stackrel{\scriptsize\cite{Ki}}{\Rightarrow} \; \mbox{finite spectral multiplicity}.
$$
Let $\mathcal{P}\colon X \to \{1,\ldots,k\}$ be a finite partition of $X$.
Given $H>0$ we define the \emph{Hamming distance} between $x$ and $y\in X$ by
$$
d^\mathcal{P}_H(x,y) := \frac{1}{H} \lambda(\{ 0\leq t <H \mid \mathcal{P}(T_tx) \neq \mathcal{P}(T_ty) \}).
$$
\begin{lm}\label{hamlem1}
For an arbitrary $\varepsilon>0$, if $(B,H)$ is an $\frac{\varepsilon^2}{4}$-monochromatic tower for a flow $(T_t)_{t\in\mathbb{R}}$, then there exists $D\subset B$, $\nu(D)>(1-\varepsilon)\nu(B)$, such that $d^\mathcal{P}_{H}(x,y)<\varepsilon$ for all $x,y\in D$.
\end{lm}
\begin{proof}
For every $0\leq t<H$, the partition $\mathcal{P}$ induces a partition $\mathcal{Q}_t$ of $B$ by $\mathcal{Q}_t(x)=\mathcal{P}(T_t x)$.
Denote by $Q_t$ the atom of the partition $\mathcal{Q}_t$ with the largest $\nu$-measure.
If for some $t\in [0,H)$, $Q_t$ is not uniquely defined, take $Q_t$ to be any atom of $\mathcal{Q}_t$ for which $\nu$ obtains its maximum.
However, for $1-\frac{\varepsilon^2}{4}$ proportion of $t\in [0,H)$ (with respect to the Lebesgue measure) $Q_t$ is determined uniquely, since the tower is $\frac{\varepsilon^2}{4}$-monochromatic.
Denote by $P_t$ the atom of $\mathcal{P}$ such that $Q_t=B\cap T_{-t}(P_t)$.
Note that
$$
\frac{1}{H}\int_0^{H} \left( \int_B \mathbbm{1}_{B\setminus Q_t}(x) d\nu(x) \right) dt < \frac{\varepsilon^2}{2},
$$
since the tower is $\frac{\varepsilon^2}{4}$-monochromatic.
By Fubini's Theorem, we get
$$
\frac{1}{H}\int_B \left( \int_0^{H} \mathbbm{1}_{B\setminus Q_t}(x) dt \right) d\nu(x) < \frac{\varepsilon^2}{2}.
$$
Hence there exists a set $D\subset B$, $\nu(D) > (1-\varepsilon)\nu(B)$, such that for all $x\in D$,
$$
\frac{1}{H} \int_0^{H} \mathbbm{1}_{B\setminus Q_t}(x) dt < \frac{\varepsilon}{2}.
$$
This means that for every $x\in D$, $T_tx\in P_t$ for $1-\frac{\varepsilon}{2}$ proportion of $t\in[0,H)$.
So, for any $x,y\in D$ and for $1-\varepsilon$ proportion of $t\in [0,H)$ both $T_tx,T_ty\in P_t$, and therefore $d^\mathcal{P}_{H}(x,y)<\varepsilon$.
\end{proof}
Given a set $P$ in a metric space $(X,d)$, by its \emph{diameter} we mean $\diam P := \sup_{x,y\in P}d(x,y)$.
For a family of sets $\mathcal{P}$, $\diam \mathcal{P} := \sup_{P\in\mathcal{P}}\diam P$.
As we will see in the next section, von Neumann flows in consideration have a Ratner like property of slow divergence of nearby points (Proposition~\ref{mainprop}), which will be in contrast with the following property of local rank one special flows.
\begin{lm} \label{lem2}
Let $\mathcal{T}=(T_t)_{t\in\mathbb{R}}$ be a special flow over a circle rotation and let $\mathcal{P}$ be a finite partition of $X$.
If $\mathcal{T}$ has local rank one of order $\beta\in(0,1]$, then for any $\varepsilon > 0$ there exist $H=H(\varepsilon)>0$ arbitrary large and $(x_0,s_0),(y_0,s'_0)\in X$ such that
$$\|x_0-y_0\| > \frac{\beta}{10H}, \;
d^f((x_0,s_0),(y_0,s'_0)) \leq \diam \mathcal{P}, \;
d^\mathcal{P}_H((x_0,s_0),(y_0,s'_0))<\varepsilon,$$
and for $(x_1,s_1):=T_H(x_0,s_0)$, $(y_1,s'_1):=T_H(y_0,s'_0)$ we also have
$$\|x_1-y_1\| > \frac{\beta}{10H}, \quad
d^f((x_1,s_1),(y_1,s'_1)) \leq \diam \mathcal{P}.$$
\end{lm}
\begin{proof}
We assume that $\varepsilon<\frac{\beta}{20}$.
Since $\mathcal{T}$ has local rank one of order $\beta$, we can find an arbitrary high $(\beta-\varepsilon)$-tower $(B,h)$ which is $\frac{\varepsilon^2}{16}$-monochromatic for $\mathcal{P}$, $\nu(B)>(\beta-\varepsilon)\nu(Y)\geq \frac{\beta-\varepsilon}{h}$ (recall that $\nu(Y)\geq \frac{1}{h}$).
Fix $t_0\in[0, h/4)$ and $t_1\in[3h/4, h)$ for which the levels $T_{t_0} B$ and $T_{t_1} B$ are $\varepsilon$-monochromatic, i.e.\ $\nu(T_{-t_i}P_i\cap B)\geq (1-\varepsilon)\nu(B)$ for an atom $P_i$ of $\mathcal{P}$ ($i=0,1$).
We set $H:=t_1-t_0 > \frac{h}{2}$.
By Lemma~\ref{hamlem1} there exists a subset $D\subset B$ with $\nu(D)>(1-\varepsilon)\nu(B)$ and such that $d^\mathcal{P}_h((x,s),(y,s'))<\frac{\varepsilon}{2}$ for any $(x,s),(y,s')\in D$.
Let $E:=D \cap T_{-t_0}P_0 \cap T_{-t_1}P_1$, $\nu(E)>(1-3\varepsilon)\nu(B) > \frac{\beta-4\varepsilon}{h} > \frac{\beta}{2.5H}$.
Fix some $(x,s)\in E$ and set
$$(x_0,s_0):=T_{t_0}(x,s), \quad (x_1,s_1):=T_{t_1}(x,s) = T_H(x_0,s_0).$$
Similarly, for $(y,s')\in E$ denote $$(y_0,s'_0):=T_{t_0}(y,s'), \quad (y_1,s'_1):=T_{t_1}(y,s') = T_H(y_0,s'_0).$$
Consider two sets
\begin{align*}
C_0 &:=\{ (y,s')\in E \mid \|x_0-y_0\|\leq \tfrac{\beta}{10H} \} \quad\mbox{and} \\
C_1 &:=\{ (y,s')\in E \mid \|x_1-y_1\|\leq \tfrac{\beta}{10H} \}.
\end{align*}
It is clear that $\nu(C_0)=\nu(C_1)= \frac{\beta}{5H}$ (since $\nu|_B$ projects onto $m_\mathbb{T}|_{\pi(B)}$, where $\pi$ stands for the natural projection $X\to\mathbb{T}$), while $\nu(E)>\frac{\beta}{2.5H}$.
Therefore there exists $(y,s') \in E \setminus (C_0 \cup C_1)$.
The corresponding pair $(x_0,s_0)$ and $(y_0,s'_0)=T_{t_0}(y,s')$ satisfies the lemma.
Indeed, $\|x_0-y_0\|, \|x_1-y_1\| > \frac{\beta}{10H}$, because $(y,s') \notin C_0 \cup C_1$.
On the other hand,
$$
d^f((x_i,s_i),(y_i,s'_i)) \leq \diam \mathcal{P},
$$
because $(x_i,s_i),(y_i,s'_i)\in P_i$, $i=0,1$.
And finally,
\begin{align*}
d^\mathcal{P}_H((x_0,s_0),(y_0,s'_0))
&= H^{-1} \lambda(\{ t_0\leq t <t_1 \mid \mathcal{P}(T_t(x,s)) \neq \mathcal{P}(T_t(y,s')) \}) \\
&\leq H^{-1} \lambda(\{ 0\leq t <h \mid \mathcal{P}(T_t(x,s)) \neq \mathcal{P}(T_t(y,s')) \}) \\
&< 2h^{-1} \lambda(\{ 0\leq t <h \mid \mathcal{P}(T_t(x,s)) \neq \mathcal{P}(T_t(y,s')) \}) \\
&= 2d^\mathcal{P}_h((x,s),(y,s')) < \varepsilon,
\end{align*}
since $H>\frac{h}{2}$ and $(x,s),(y,s')\in D$.
The lemma is proved.
\end{proof}
\section{Proof of Theorem~\ref{mainth}}
Assume that an irrational $\alpha$ is fixed and $f$ is of the form \eqref{roof}.
We may assume without lost of generality that the slope in \eqref{roof} is $A=1$.
Denote $L := \min\{1,\inf_{x\in\mathbb{T}} f(x)\}$ and $M := \max\{1,\sup_{x\in\mathbb{T}} f(x)\}$, so that
$$
0<L\leq f(x) \leq M < \infty
$$
and $L\leq 1 \leq M$.
For an integer interval $J=[a,b]\subset \mathbb{Z}$ and $c>0$, by $cJ$ we denote the integer interval $[ca,cb]\cap\mathbb{Z}$.
The following proposition will be the main ingredient in the proof.
\begin{pr} \label{mainprop}
Let $\beta\in (0,1]$ be a fixed number.
There exists a constant $\delta_0>0$ such that for all $x,y\in \mathbb{T}$, $0<\|x-y\|<\delta_0$, there exists an interval $J_{x,y}\subset \mathbb{Z}$ such that the following holds:
\begin{enumerate}[(i)]
\item \label{prop1a} $0\in J_{x,y}\subset [-\frac{\beta}{100M\|x-y\|},
\frac{\beta}{100M\|x-y\|}]$;
\item \label{prop1b} for every $n\in \frac{10M}{L} J_{x,y}$, we have $|f^{(n)}(x)-f^{(n)}(y)|<\frac{50M}{L}$;
\item \label{prop1c} there exists an interval $U_{x,y}\subset J_{x,y}$ such that $|U_{x,y}|>\frac{1}{10}|J_{x,y}|$ and such that for every $n\in U_{x,y}$
$$
|f^{(n)}(x)-f^{(n)}(y)|>\frac{\beta L}{10^6M^2}.
$$
\end{enumerate}
\end{pr}
We will prove the proposition in the next section, now let us prove the main result.
\begin{proof}[Proof of Theorem~\ref{mainth}]
Assume, by contradiction, that $\mathcal{T}$ has local rank one of order $\beta\in(0,1]$.
Fix a positive
$$
\delta < \min\Bigl\{ \delta_0,\;
\frac12 \frac{\beta L}{10^6M^2},\;
\min_{\substack{k\in\mathbb{Z}\\ 0<k\leq \frac{60M}{L^2}}} \frac{\|k\alpha\|}{2} \Bigr\},
$$
where $\delta_0$ is as in Proposition~\ref{mainprop}.
Let $\mathcal{P}$ be a finite partition of $X$ of diameter less then $\delta$.
For an arbitrary $\varepsilon > 0$ we can find by Lemma~\ref{lem2} a positive $H>100M$
and a pair of points $(x_0,s_0),(y_0,s'_0)\in D$ such that, if we denote by $(x_t,s_t):=T_t(x_0,s_0)$ and $(y_t,s'_t):=T_t(y_0,s'_0)$, we have
\begin{equation}\label{eq1}
\|x_0-y_0\| > \frac{\beta}{10H}, \quad \|x_H-y_H\| > \frac{\beta}{10H},
\end{equation}
\begin{equation}\label{eq2}
d^f((x_0,s_0),(y_0,s'_0))<\delta, \quad d^f((x_H,s_H),(y_H,s'_H))<\delta,
\end{equation}
\begin{equation}\label{hamineq}
d^\mathcal{P}_{H}((x_0,s_0),(y_0,s'_0))<\varepsilon.
\end{equation}
Denote
\begin{align*}
C &:=\{ 0\leq t < H \mid d^f((x_t,s_t),(y_t,s'_t)) < \delta \} \quad\mbox{and} \\
F &:=\{ 0\leq t < H \mid d^f((x_t,s_t),(y_t,s'_t)) \geq \delta \}
\end{align*}
the times when the points $(x_t,s_t)$ and $(y_t,s'_t)$ are $\delta$-close and $\delta$-far respectively.
Notice that the mapping $\mathbb{R}\ni t\mapsto d((x_t,s_t),(y_t,s'_t))$ is piecewise constant by the definition of a special flow, whence both $C$ and $F$ are finite unions of (half-open) intervals.
It trivially follows from the definition of $F$ (since $\diam\mathcal{P}<\delta$) that $\lambda(F)/H \leq d^\mathcal{P}_{H}((x_0,s_0),(y_0,s'_0))$.
We will show at the end that $\lambda(F)>\varepsilon H$ if $\varepsilon$ is small to get a contradiction with (\ref{hamineq}).
For the reader's convenience, we split the proof into several steps.
Given $t\in C$, let $J_{x_t,y_t}$ be as in Proposition~\ref{mainprop}.
\smallskip
\textit{Claim~1.}
\textit{For every $t\in C$, $|J_{x_t,y_t}|\leq \frac{H}{5M}$.}
To prove the claim we will argue by contradiction.
Assume that there exists $t_0\in C$ such that $|J_{x_{t_0},y_{t_0}}| > \frac{H}{5M}$.
Then we derive from Proposition~\ref{mainprop}\eqref{prop1a} that
\begin{equation}\label{spk}
\|x_{t_0}-y_{t_0}\| < \frac{\beta}{10H}.
\end{equation}
Set $J:=\frac{10M}{L} J_{x_{t_0},y_{t_0}}$, so that $|J|>\frac{2H}L$ and for every $n\in J$
\begin{equation}\label{pb}
|f^{(n)}(x_{t_0})-f^{(n)}(y_{t_0})|< \frac{50M}{L}
\end{equation}
(by Proposition~\ref{mainprop}\eqref{prop1b}).
Consider the interval $I\subset \mathbb{R}$ given by
$$I:=[t_0-s_{t_0}+f^{(a)}(x_{t_0}), t_0-s_{t_0}+f^{(b)}(x_{t_0}) ),$$
where $a=\min J$, $b=\max J+1$.
Then $\lambda(I)>2H$ and it follows that either $0 \in I$ or $H \in I$.
Since the proof in both cases goes along the same lines, let us assume without loss of generality that $H \in I$.
By definition,
\begin{align*}
&(x_H,s_H) = T_{H-t_0}(x_{t_0},s_{t_0}) = (x_{t_0}+n\alpha, s_{t_0}+(H-t_0)-f^{(n)}(x_{t_0})), \\
&(y_H,s'_H) = T_{H-t_0}(y_{t_0},s'_{t_0}) = (y_{t_0}+m\alpha, s'_{t_0}+(H-t_0)-f^{(m)}(y_{t_0})),
\end{align*}
where $n,m\in\mathbb{Z}$ are such that $f^{(n)}(x_{t_0}) \leq s_{t_0}+(H-t_0) < f^{(n+1)}(x_{t_0})$, $f^{(m)}(y_{t_0}) \leq s'_{t_0}+(H-t_0) < f^{(m+1)}(y_{t_0})$.
Notice that $n\in J$ by the definition of $I$.
The distance between $(x_H,s_H)$ and $(y_H,s'_H)$ is smaller than $\delta$ by \eqref{eq2} and equals
\begin{equation}\label{dist}
\|x_{t_0}-y_{t_0}+(n-m)\alpha\| + |s_{t_0}-s'_{t_0}-f^{(n)}(x_{t_0})+f^{(m)}(y_{t_0})| < \delta.
\end{equation}
Since $f<M$, $|s_{t_0}-s'_{t_0}|<M$.
It follows from \eqref{pb}, \eqref{dist} and the cocycle equality $f^{(m)}(y_{t_0}) = f^{(n)}(y_{t_0}) + f^{(m-n)}(y_{t_0}+n\alpha)$ that
$$|f^{(m-n)}(y_{t_0}+n\alpha)|<\frac{50M}{L}+M+1<\frac{60M}{L}.$$
On the other hand, $f(x)\geq L$ and therefore $|f^{(k)}(x)| \geq |k|L$ for any $x\in\mathbb{T}$, $k\in\mathbb{Z}$.
Hence $|n-m|<\frac{60M}{L^2}$.
Since the first term in (\ref{dist}) is less then $\delta$ and $\|x_{t_0}-y_{t_0}\|<\delta$ (because $t_0\in C$), we have $\|(n-m)\alpha\|<2\delta$ with $|n-m|<\frac{60M}{L^2}$.
By the choice of $\delta$ this is only possible if $n=m$.
We conclude that $x_H=x_{t_0}+n\alpha$, $y_H=y_{t_0}+n\alpha$ and
$$
\|x_H-y_H\|=\|x_{t_0}-y_{t_0}\|\stackrel{\eqref{spk}}{<} \frac{\beta}{10H},
$$
which contradicts \eqref{eq1}.
Similarly, if $0\in I$ then one can show that $\|x_0-y_0\|=\|x_{t_0}-y_{t_0}\| < \frac{\beta}{10H}$.
Claim~1 is proved.
\smallskip
\textit{Claim~2.}
\textit{For any $t\in C\cap[\frac{H}{4},\frac{3H}{4}]$ there exists an interval $I_t\subset [0,H)$ containing $t$
such that $\lambda(I_t\cap F) > \gamma \lambda(I_t)$, where $\gamma = \frac{L}{10M}$.}
Fix $t\in C\cap[\frac{H}{4},\frac{3H}{4}]$.
Fix integer intervals $J_{x_t,y_t}$ and $U_{x_t,y_t}\subset \mathbb{Z}$ as in Proposition~\ref{mainprop}.
We set $I_t := [t-s_t+f^{(a)}(x_t), t-s_t+f^{(b)}(x_t) )$, where $a=\min J_{x_t,y_t}$, $b=\max J_{x_t,y_t}+1$.
Equivalently, by denoting by $t+I:=\{t+x \mid x\in I\}$ the sumset, we can write
$I_t = t + \bigsqcup_{n\in J_{x_t,y_t}} I_t^n$, where $I_t^n := [-s_t+f^{(n)}(x_t),-s_t+f^{(n+1)}(x_t))$.
Since $\frac{H}{4}\leq t < \frac{3H}{4}$, $|J_{x_t,y_t}|\leq \frac{H}{5M}$ (by Claim~1) and $f\leq M$, we have $I_t\subset [0,H)$.
Also $t\in I_t$, because $0\in J_{x_t,y_t}$.
For every $n\in J_{x_t,y_t}$ and $r\in I_t^n$ we have
\begin{align*}
(x_{t+r},s_{t+r})=T_{r}(x_t,s_t) &= (x_t+n\alpha, s_t+r-f^{(n)}(x_t)), \\
(y_{t+r},s'_{t+r})=T_{r}(y_t,s'_t) &= (y_t+m\alpha, s'_t+r-f^{(m)}(y_t)),
\end{align*}
where $f^{(m)}(y_t) \leq s'_t+r < f^{(m+1)}(y_t)$.
By estimating the distance between $(x_{t+r},s_{t+r})$ and $(y_{t+r},s'_{t+r})$ as in \eqref{dist} and
reasoning as in the proof of Claim~1 we can show that if $t+r\in C$, then nesessarily $m=n$.
Moreover, in this case $|s_t-s'_t|<\delta$ and $|s_{t+r}-s'_{t+r}|<\delta$, so that $|f^{(n)}(x_t)-f^{(n)}(y_t)|<2\delta< \frac{\beta L}{10^6M^2}$.
The latter inequality cannot hold if $n\in U_{x_t,y_t}$ (by Proposition~\ref{mainprop}\eqref{prop1c}).
In other words, if $n\in U_{x_t,y_t}$ then the corresponding interval $t+I_t^n$ is contained in $F$.
We can now estimate $\lambda(F\cap I_t) \geq L |U_{x_t,y_t}| > \frac{1}{10} L|J_{x_t,y_t}|$ by Proposition~\ref{mainprop}\eqref{prop1c}, while $\lambda(I_t) \leq M|J_{x_t,y_t}|$.
This proves Claim~2.
\smallskip
\textit{Final step.}
Let $F=\bigsqcup_{n} F_n$ where each $F_n$ is a (half-open) interval in $[0,H)$.
The family $\mathcal{I}=\{I_t\}_{t\in C\cap[\frac{H}{4},\frac{3H}{4}]} \cup \{F_n\}_n$ covers $[\frac{H}{4},\frac{3H}{4}]$.
First, we can enlarge each interval by 1\% of its length to get open intervals and by compactness select a finite subfamily (of enlarged intervals) that covers $[\frac{H}{4},\frac{3H}{4}]$.
Then, by Vitali covering lemma, we can find a finite disjoint subfamily of $\mathcal{I}$ of total measure at least $\frac{H}{7}$.
Since each interval in $\mathcal{I}$ contains more than $\gamma$ proportion of $F$, $\lambda(F) > \frac{\gamma}{7}H$.
This means that inequality (\ref{hamineq}) fails if $\varepsilon < \frac{\gamma}{7}$.
The obtained contradiction finishes the proof of Theorem~\ref{mainth}.
\end{proof}
\subsection{Proof of Proposition~\ref{mainprop}}
Before we prove Proposition~\ref{mainprop} let us give an outline of how to prove \eqref{prop1b} and \eqref{prop1c}.
They are both a consequence of Lemma~\ref{lin}.
For $x,y\in \mathbb{T}$ satisfying \eqref{distanc}, if \eqref{lem3b} or \eqref{lem3c} in Lemma~\ref{lin} holds (that is forward or backward orbit of $[x,y]$ avoids the discontinuity long enough),
we define $J_{x,y}$ to be respectively $[0, \frac{C}{\|x-y\|}]\cap \mathbb{Z}$ or $[-\frac{C}{\|x-y\|},0]\cap \mathbb{Z}$, $C=C(\beta,M,L)$.
By the definition of $J_{x,y}$ and Lemma~\ref{c1}, for $n\in J_{x,y}$, $|f^{(n)}(x)-f^{(n)}(y)|=n(\text{ slope }+o(1))\|x-y\|$.
This means that the divergence is linear and hence \eqref{prop1b} and \eqref{prop1c} follow.
If $x,y\in \mathbb{T}$ satisfy \eqref{lem3a}, we define $J_{x,y}$ as $[0,2q_n-1]\cap \mathbb{Z}$.
In this case the divergence is given by hitting the discontinuity, but the number of times the orbit of $[x,y]$ hits $0$ is bounded by \eqref{dmleq}.
This gives \eqref{prop1b}.
And once hitting the discontinuity, the two points cannot come close for some time, which leads to \eqref{prop1c}.
The following lemma is classical.
\begin{lm}\label{c1} Let $g\in C^1(\mathbb{T})$. Then
\begin{equation}\label{canc}
\lim_{|n|\to +\infty}\sup_{x,y\in\mathbb{T}, x\neq y}\frac{|g^{(n)}(x)-g^{(n)}(y)|}{|n|\|x-y\|}=0.
\end{equation}
\end{lm}
\begin{proof} Notice that for some $\theta^{(n)}_{x,y}\in\mathbb{T}$
$$
\frac{|g^{(n)}(x)-g^{(n)}(y)|}{|n|\|x-y\|}=
\left|\frac{1}{n}g'^{(n)}(\theta^{(n)}_{x,y})\right|
$$
and so the statement follows from the ergodic theorem and unique ergodicity of $R_\alpha$ since $g'\in C(\mathbb{T})$ and $\int_\mathbb{T} g'd m_\mathbb{T}=0$.
\end{proof}
Let $(q_n)_{n\in\mathbb{N}}$ denote the sequence of denominators of $\alpha$, that is
$$
\frac{1}{2 q_n q_{n+1}} < \left| \alpha - \frac{p_n}{q_n} \right| < \frac{1}{q_n q_{n+1}},
$$
where
\[
\begin{array}{lll}
q_0 = 1, & q_1 = a_1, & q_{n} = a_{n}q_{n-1} + q_{n-2},\\
p_0 = 0, & p_1 = 1, & p_{n} = a_{n}p_{n-1} + p_{n-2}
\end{array}
\]
and $[0;a_1,a_2,\ldots]$ stands for the continued fraction expansion of $\alpha$ (see \cite{Kh}).
It follows that
$$
\frac{1}{2 q_{n+1}} < \| q_n\alpha \| < \frac{1}{q_{n+1}}.
$$
The reader can easily check that the partition of $\mathbb{T}$ by $0,\alpha,...,(q_n-1)\alpha$ has the form
$\{ R_\alpha^k I_n \mid 0\leq k < q_n - q_{n-1} \} \cup \{ R_\alpha^k I'_n \mid 0\leq k < q_{n-1} \}$,
where $I_n = [0,q_{n-1}\alpha]$ and $I'_n = [(-q_{n-1}+q_n)\alpha,0]\subset\mathbb{T}$.
(We recall that for $x,y\in\mathbb{T}$, $[x,y]$ stands for the shortest interval in $\mathbb{T}$ connecting $x$ and $y$.)
In particular,
\begin{equation}\label{pdiam}
\min_{0 \leq i < j < q_n} \|i\alpha - j\alpha\| = \|q_{n-1}\alpha\| > \frac{1}{2q_n}.
\end{equation}
\begin{lm}\label{lin}
Fix $x,y\in \mathbb{T}$, $x\neq y$, and let $n\in \mathbb{N}$ be any integer such that
\begin{equation}\label{distanc}
\|x-y\|< \frac{1}{6q_n}.
\end{equation}
Then one of the following holds:
\begin{enumerate}[(a)]
\item \label{lem3a} $0\in \bigcup_{k=0}^{q_n-1}R^k_\alpha[x,y]$;
\item \label{lem3b} $0\notin \bigcup_{k=0}^{\left[\frac{q_{n+1}}{6}\right]}R^k_\alpha[x,y]$;
\item \label{lem3c} $0\notin \bigcup_{k=-\left[\frac{q_{n+1}}{6}\right]}^{0}R^k_\alpha[x,y]$.
\end{enumerate}
\end{lm}
\begin{proof} Assume that \eqref{lem3a} does not hold.
Consider the partition $\mathcal{P}_n$ of $\mathbb{T}$ by $0,-\alpha,...,(-q_n+1)\alpha$.
Let $i,j\in\{-q_n+1,...,0\}$ be unique such that
\begin{equation}\label{xysub}
[x,y]\subset [i\alpha,j\alpha]=I\in\mathcal{P}_n.
\end{equation}
By \eqref{distanc} and \eqref{pdiam} it follows that either $d(i\alpha,[x,y])$ or $d(j\alpha,[x,y])$ is greater then $\frac{1}{6q_n}$.
Assume without loss of generality that $d(i\alpha,[x,y])> \frac{1}{6q_n}$ (the proof in the other case is analogous).
If $q_n\alpha - p_n<0$ we will show \eqref{lem3b}, and if $q_n\alpha - p_n>0$ we will show \eqref{lem3c}.
Let us conduct the proof assuming that $q_n\alpha-p_n<0$, the proof in the other case follows the same lines.
We need to show that for every $k\in\{-\left[\frac{q_{n+1}}{6}\right],...,0\}$,
\begin{equation}\label{atm}
k\alpha\notin [x,y].
\end{equation}
Each such $k\alpha$ belongs to a unique atom of $\mathcal{P}_n$.
If $k\alpha\notin I$ then \eqref{atm} holds trivially by \eqref{xysub}.
All $k\in\{-\left[\frac{q_{n+1}}{6}\right],...,0\}$ for which $k\alpha\in I$ are of the form $k=i-m_kq_n$ for $m_k\leq \frac{q_{n+1}}{6q_n}$.
Therefore
$$
d(k\alpha,[x,y])\geq d(i\alpha,[x,y])-m_k\|q_n\alpha\|>\frac{1}{6q_n}-
\frac{q_{n+1}}{6q_n}\frac{1}{q_{n+1}}\geq 0.
$$
So \eqref{atm} also holds in this case. This finishes the proof of Lemma~\ref{lin}.
\end{proof}
Now we can prove Proposition~\ref{mainprop}.
\begin{proof}[Proof of Proposition~\ref{mainprop}]
Recall that $f=g+\{\cdot\}+c$, where $g\in C^1(\mathbb{T})$, $0<L\leq 1 \leq M$, $0<\beta<1$.
Fix $\epsilon< \frac{\beta L}{10^6 M^2}$.
Let $\delta_0$ be such that for every $x,y\in \mathbb{T}$, $\|x-y\|< \delta_0$, and every $n\in \mathbb{Z}$
\begin{equation}\label{c1part}
|g^{(n)}(x)-g^{(n)}(y)|<\epsilon\max\{ 1, |n| \|x-y\| \}.
\end{equation}
The existence of such $\delta_0$ follows from Lemma~\ref{c1}.
Indeed, by Lemma~\ref{c1} there exists $n_\epsilon>0$ such that for all $n$, $|n|\geq n_\epsilon$, we have $|g^{(n)}(x)-g^{(n)}(y)|<\epsilon |n| \|x-y\|$.
It is enough to define $\delta_0:=\frac{\epsilon}{n_\epsilon \sup_\mathbb{T} |g'|}$.
We will also assume that
\begin{equation}\label{delta0}
\delta_0 \leq \frac{\beta L}{10^5 M^2}.
\end{equation}
Fix $x,y\in \mathbb{T}$, $\|x-y\|< \delta_0$.
We want to define $J_{x,y}$.
We will use Lemma~\ref{lin}.
Let $n\in \mathbb{N}$ be unique such that
\begin{equation}\label{dxy}
\frac{\beta}{200Mq_{n+1}}<\|x-y\|\leq \frac{\beta}{200Mq_n}.
\end{equation}
Then $n$ satisfies \eqref{distanc}, since $M\geq 1$.
Define $J_{x,y}$ as follows:
\begin{enumerate}[(A)]
\item \label{p1} if $x,y$ satisfy \eqref{lem3a}, set $J_{x,y}:=[0, 2q_n-1]\cap \mathbb{Z}$;
\item \label{p2} if $x,y$ satisfy \eqref{lem3b}, set $J_{x,y}:=[0, \frac{\beta L}{10^5M^{2}\|x-y\|}]\cap \mathbb{Z}$;
\item \label{p3} if $x,y$ satisfy \eqref{lem3c}, set $J_{x,y}:=[-\frac{\beta L}{10^5M^{2}\|x-y\|},0]\cap \mathbb{Z}$.
\end{enumerate}
We will show that $J_{x,y}$ defined above satisfies the assertions of Proposition~\ref{mainprop}.
\smallskip
\textit{Case \eqref{p1}.}
Notice that \eqref{prop1a} in Proposition~\ref{mainprop} holds trivially by \eqref{dxy} (since $2q_n\leq \frac{\beta}{100M\|x-y\|}$).
Let us show \eqref{prop1b}.
For $m\geq 0$ let
$$
D_m:=\left|\{i\in\{0,...,m-1\}\;:\; 0\in R_\alpha^i[x,y]\}\right|
$$
(the number of times the orbit of $[x,y]$ hits the discontinuity up to time $m$).
Notice that for every $n\in \frac{10M}{L}J_{x,y}\subset [0,\frac{20M}{L}q_n]$ by \eqref{dxy} and \eqref{pdiam} we have
\begin{equation}\label{dmleq}
D_n\leq \frac{40M}{L}.
\end{equation}
Therefore for every such $n$,
using \eqref{c1part} and \eqref{dxy}, we obtain
\begin{align*}
|f^{(n)}(x)-f^{(n)}(y)|
&\leq |g^{(n)}(x)-g^{(n)}(y)|+|\sum_{i=0}^{n-1}(\{x+i\alpha\}-\{y+i\alpha\})|\\
&< \max\{ 1, n \|x-y\| \} + n\|x-y\| + D_n \\
&< \frac{M}{L}+ \frac{M}{L}+\frac{40M}{L}\\
&< \frac{50M}{L}
\end{align*}
and this finishes the proof of \eqref{prop1b}.
For \eqref{prop1c} let $U_{x,y}:=[q_n,2q_n-1]\cap \mathbb{Z}$.
Obviously $|U_{x,y}| >\frac{1}{10}|J_{x,y}|$.
Since $x,y$ satisfy \eqref{lem3a}, it follows that for every $n\in U_{x,y}$ we have
$$
D_n \geq 1.
$$
Therefore for every $n\in U_{x,y}$ by \eqref{c1part} and \eqref{dxy}
$$
|f^{(n)}(x)-f^{(n)}(y)|
\geq D_n-n\|x-y\|-|g^{(n)}(x)-g^{(n)}(y)|
> 1-1/2
= 1/2
$$
and this completes the proof of \eqref{prop1c}.
So if $\eqref{lem3a}$ holds then Proposition~\ref{mainprop} holds for $x,y$.
\smallskip
It remains to conduct the proof in cases \eqref{p2} and \eqref{p3}.
The proofs in both cases are completely symmetric, one just needs to switch the time direction from positive (in case \eqref{p2}) to negative (in case \eqref{p3}).
Therefore we will present the proof in case \eqref{p2}, the proof in case \eqref{p3} follows the same lines.
\smallskip
\textit{Case \eqref{p2}.}
Notice that \eqref{prop1a} follows automatically by the definition of $J_{x,y}$.
Let us show \eqref{prop1b}.
Notice that $\frac{10M}{L}J_{x,y}\subset[0,\frac{\beta}{10^4M\|x-y\|}]\subset [0,\frac{q_{n+1}}{6}]$.
Hence by \eqref{lem3b} (which gives $D_n=0$), \eqref{c1part} and \eqref{dxy} we get for every $n\in \frac{10M}{L}J_{x,y}$
\begin{align*}
|f^{(n)}(x)-f^{(n)}(y)|
&\leq n\|x-y\|+|g^{(n)}(x)-g^{(n)}(y)|\\
&\leq n\|x-y\| + \max\{ 1, |n| \|x-y\| \}\\
&\leq \frac{\beta}{10^4M} + 1\\
&<2.
\end{align*}
This gives \eqref{prop1b}.
To get \eqref{prop1c} define $U_{x,y}:=[\frac{1}{2}\frac{\beta L}{10^5M^{2}\|x-y\|}, \frac{\beta L}{10^5M^{2}\|x-y\|}]\cap \mathbb{Z}$.
Then trivially $|U_{x,y}|\geq \frac{1}{2}|J_{x,y}|$ (notice that $U_{x,y}\neq\emptyset$ by \eqref{delta0}).
Moreover, for every $n\in U_{x,y}$,
$$
|f^{(n)}(x)-f^{(n)}(y)|
\geq n\|x-y\|-|g^{(n)}(x)-g^{(n)}(y)|
\geq \frac{1}{2}\frac{\beta L}{10^5M^{2}}-\epsilon
> \frac{\beta L}{10^6M^{2}}
$$
by the choice of $\epsilon$.
This gives \eqref{prop1c} and finishes the proof of Proposition~\ref{mainprop}
\end{proof}
\section{Concluding remarks}
It follows by \cite{FL04} that the spectral type of von Neumann flows is purely singular.
Whether or not the maximal spectral multiplicity of von Neumann flows is finite remains however an open problem.
It follows from our result that the popular method of estimating spectral multiplicity by the rank fails in this case.
It seems that the methods used for one discontinuity can be carried out for many discontinuities (however the proof becomes more subtle).
In a forthcoming paper \cite{KaSo} we will show that the slope is an isomorphism invariant for the von Neumann flows under consideration (with one discontinuity), i.e.\ flows are non-isomorphic when the slopes of the roof functions are different.
This would mean, in turn, that even when $\alpha$ is fixed, we have considered an uncountable family of pairwise non-isomorphic von Neumann flows.
\section{Acknowledgements}
The authors would like to thank K.~Fr\c{a}czek and M.~Lema\'{n}czyk for several remarks on the subject.
The first author would like to thank A.~Katok for discussions on spectral multiplicity of von Neumann flows.
We thank M.~Lema\'{n}czyk and C.~Ulcigrai for their comments on the final version of this paper.
The second author would like to thank ICTP, Trieste, where part of this work was done, for hospitality.
The research leading to these results has received funding from the European Research Council under the European Union's Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement n.~335989.
|
\section{Introduction}
The Riemann surfaces are often used in the problems of string theory~\cite{DHP}, lower-dimensional
gravity~\cite{Lo,NR} and quantum geometry~\cite{Kash00}. Sometimes, chaotic behavior in nature
can be also related with non-trivial geometry \cite{Gutz,Naz}.
Here we study free geometrodynamics or kinematics of the surface in genus two associated with
hyperbolic octagon which is embedded into Poincar\'e disk and stable under rotation by $\pi/2$.
Identifying the opposite sides of such a domain, it is enough to use two real parameters in order to
describe the octagon geometry and the form of Fuchsian group generators. To generate a geometry
evolution, we operate by parameter space invariants instead of Riemannian metric which is unknown for us.
Note that the involution of similar surfaces and the associated generators were discussed in \cite{Sil,BS}.
Defining the Teichm\"uller space~\cite{IT} for a family of the given surfaces, we find the Fenchel--Nielsen
variables regarding as global coordinates on it and permitting us to endow the parameter space with the
Weil--Petersson (WP) symplectic two-form due to the Wolpert's theorem~\cite{Wo85}. Further, we demonstrate
that the parameter space can be densely covered by the closed curves of the constant octagon perimeter and
describe the diffeomorphism produced by isoperimetric constraint. We consider the set of isoperimetric
orbits as a tool for further parametrization and quantization~\cite{Hurt}.
We determine the canonically conjugate action--angle variables for isoperimetric orbits by identifying the
action with WP-area of domain bounded by the orbit in parameter space. This is a main point of our approach
using the integral characteristics in a contrast with the formalism based on local Fenchel--Nielsen
parameters of the surface \cite{Kash00}. In this way, we touch the problem of WP-area quantization, which
is similar to the one of the loop quantum gravity/cosmology~\cite{RS}.
In order to describe geometrodynamics and to perform area quantization, we extend the algebra of
action--angle variables up to generators of $\mathfrak{su}(1,1)$ associated with the Lorentz algebra in
(2+1)-dimensional space-time. Such an approach permits us to formulate the relativistic dynamics of Riemann
surface as a canonical transformation generated by the boost. It is realized in Section 4 and leads to
``big bounce'' scenario~\cite{APS,LM} in parameter space. We note that the quantization of the universe
represented by the Riemann surface in genus two was formulated algebraically in \cite{NR}.
\section{Model Octagon and Riemann Surface}
The Poincar\'e model of two-dimensional hyperbolic space is given by open disk
$\mathbb{D}=\left\{z=x+\rmi y||z|<1\right\}$ and the metric $\rmd s^2=4|\rmd z|^2/(1-|z|^2)^2$.
The geodesic in $(\mathbb{D},\rmd s^2)$ is an circle arc inside $\mathbb{D}$ with radius $R$ and
center at the point $z_0=\sqrt{1+R^2}\exp{(\rmi\phi)}$ lying beyond the unit disk.
In particular case, the geodesics emanating from the origin are the Euclidean straight
lines (diameters). All geodesics intersect the boundary $\partial\mathbb{D}$ orthogonally.
The group of all orientation-preserving isometries $\gamma$ of $(\mathbb{D},\rmd s^2)$,
denoted by ${\rm Isom^+(\mathbb{D})}$, acts via the M\"obius transformation:
\begin{equation}
z\mapsto \gamma[z]=\frac{uz+v}{\overline{v}z+\overline{u}},\quad
z\in\mathbb{D},
\end{equation}
where $u$ and $v$ satisfy relation $|u|^2-|v|^2=1$; $\bar u$, $\bar v$ are the complex
conjugates. Thus, it is convenient to identify a generator $\gamma$ with an element of group
\begin{equation}
{\rm SU(1,1)}=\left\{\left.
\left(\begin{array}{cc}
u&v\\
\overline{v}&\overline{u}
\end{array}\right)\right||u|^2-|v|^2=1
\right\}.
\end{equation}
The Riemann surface $S$ is understood here as a compact two-dimensional orientable manifold
with the metric of constant negative curvature. Such a surface is obtained from hyperbolic
simply connected octagon ${\cal F}$ in $\mathbb{D}$ via gluing opposite sides formed by
eight geodesic arcs, whose intersections serve as vertices.
\begin{figure}
\begin{center}
\includegraphics[width=4.2cm]{fig1a.eps}\qquad
\includegraphics[width=4.3cm]{fig1b.eps}
\end{center}
\vspace*{-3mm}
\caption{\small Octagon with $a=0.8$, $\alpha=\pi/3$ and generators $g_k$ of
Fuchsian group.}
\end{figure}
Let the vertices be at the points $a\exp{(\rmi k\pi/2)}$, $b\exp{[\rmi(\alpha+k\pi/2)]}$
(Fig.~1, left panel), where $0<\alpha<\pi/2$, $0<a,b<1$ and $k=\overline{0,3}$. We also
require that the sum of inner angles of ${\cal F}$ equals to $2\pi$ and hyperbolic
${\rm Area}({\cal F})=4\pi$ in the case of the surface in genus two.
Choosing two parameters $(a,\alpha)$ as independent real variables, we find that
$b$ and the inner angle $\beta$ by vertices $a\exp{(\rmi k\pi/2)}$ are
\begin{equation}\label{b}
b=\left(\sqrt{2}a\cos{\tilde\alpha}\right)^{-1}, \quad
\tan{\beta}=\frac{1-a^2}{1-b^2}; \quad
\tilde\alpha=\alpha-\frac{\pi}{4}.
\end{equation}
These allow us to determine the region ${\cal A}$ of variety of parameters $(a,\alpha)$:
\begin{equation}
-\pi/4<\tilde\alpha<\pi/4,\quad
\left(\sqrt{2}\cos{\tilde\alpha}\right)^{-1}<a<1,
\end{equation}
which is shown in Fig.~2 below.
To obtain the regular hyperbolic octagon, we should put $a=2^{-1/4}$, $\alpha=\pi/4$.
In the case at hand, the octagon boundary $\partial{\cal F}$ is formed by geodesics of two
kinds (labeled by ``$\pm$'' below), which are completely determined by the radii $R_\pm$ and
the angles $\phi_\pm+k\pi/2$, $k=\overline{0,3}$, defining the positions of the circle (arc)
centers. Geometrical conditions result in parametrization:
\begin{equation}\label{R-phi}
R_\pm=\frac{1}{2a}\sqrt{T^2_\pm+(1-a^2)^2},\qquad
\phi_\pm=\arctan{\left[\left(\frac{T_\pm}{1+a^2}\right)^{\pm1}\right]},
\end{equation}
here $T_\pm=a^2\pm\tan{\tilde\alpha}$ and $0<\phi_+<\alpha<\phi_-<\pi/2$.
Here, we connect the model octagon ${\cal F}$ with the corresponding Riemann surface $S$ and
the Fuchsian group $\Gamma\subset{\rm Isom^+(\mathbb{D})}$ isomorphic to fundamental group $\pi_1(S)$.
Since the opposite sides of ${\cal F}$ have the same lengths, we can define isometry
$g_k\in{\rm Isom}^+(\mathbb{D})$ mapping geodesic boundary segment $s_{k+4}$ onto $s_k$
for all $k=\overline{0,3}$ (see Fig.~1, right panel). Identifying any $z\in s_{k+4}$ with
$g_k[z]\in s_k$, we obtain a closed surface. Four isometries $g_k$ and their inverses
$g^{-1}_k$ generate Fuchsian group $\Gamma$ with a single relation:
\begin{equation}
g_0 g^{-1}_1 g_2 g^{-1}_3 g^{-1}_0 g_1 g^{-1}_2 g_3={\rm id},
\end{equation}
and we then define surface $S$ as a quotient $\mathbb{D}/\Gamma$.
Our calculations give the dependence of $g_0$ and $g_1$ on $(a,\alpha)$:
\begin{eqnarray}
&g_0=N(a,\tilde\alpha)\left(
\begin{array}{cc}
a(1-\tan{\tilde\alpha})& (a^2-\tan{\tilde\alpha})+\rmi(1-a^2)\label{g0}\\
(a^2-\tan{\tilde\alpha})-\rmi(1-a^2) & a(1-\tan{\tilde\alpha})
\end{array}
\right),\\
&g_1=N(a,\tilde\alpha)\left(
\begin{array}{cc}
a(1+\tan{\tilde\alpha})& (1-a^2)+\rmi(a^2+\tan{\tilde\alpha})\\
(1-a^2)-\rmi(a^2+\tan{\tilde\alpha}) & a(1+\tan{\tilde\alpha})
\end{array}
\right),
\end{eqnarray}
here
$$
N(a,\tilde\alpha)=\frac{-\cos{\tilde\alpha}}{\sqrt{(1-a^2)(2a^2\cos^2{\tilde\alpha}-1)}}.
$$
The remaining generators are simply obtained by rotations:
\begin{equation}\label{gn}
g_{2,3}=R_\frac{\pi}{2}g_{0,1}R^{-1}_\frac{\pi}{2}, \quad
g^{-1}_k=R_{\pi}g_kR^{-1}_{\pi};
\end{equation}
$$
R_\varphi=\left(
\begin{array}{cc}
\exp{\left(\rmi{\varphi}/{2}\right)}& 0\\
0 & \exp{\left(-\rmi{\varphi}/{2}\right)}
\end{array}
\right).
$$
Let us mark a surface by generators of $\Gamma$. Two marked surfaces $(S,\Gamma)$ and
$(S^\prime,\Gamma^\prime)$ are called marking equivalent if there exists an isometry
$\gamma: S\to S^\prime$ satisfying $g_k^\prime=\gamma g_k\gamma^{-1}$. Then all marking
equivalent surfaces form a marking equivalence class $[S,\Gamma]$ representing
the Riemann surface $S$.
The set of all marking equivalence classes of the closed and compact Riemann surfaces
in genus $g$ forms the Teichm\"uller space ${\cal T}_g$. The real dimension of ${\cal T}_g$
like vector space equals to $6g-6$ in accordance with the Riemann--Roch theorem.
In our case, the Riemann surfaces result in the subset of total ${\cal T}_2$.
\section{Structure of Parameter Space ${\cal A}$}
\subsection{Symplectic Two-Form}
A hyperbolic Riemann surface of genus $g$ without boundary always contains a system of
$3g-3$ simple closed geodesics that are neither homotopic to each other nor homotopically
trivial. The cut along these geodesics always decomposes surface into $2g-2$ pairs of pants
(three-holed spheres), playing a role of natural building blocks for Riemann surface~\cite{Bu}.
In the case at hand, surface $S$ is two-holed torus which can be decomposed into two pairs
of pants by a system of three closed geodesics. This surgery results in computing the
Fenchel--Nielsen (FN) parameters: lengths $\ell_k$ of these geodesics and the corresponding
twists $\tau_k$ (see \cite{IT,Bu}).
For one of possible pants decompositions, we have found these quantities~\cite{Naz2}:
\begin{eqnarray}
\ell_{1,2}=2~{\rm arccosh}~{\frac{a^2}{1-a^2}},&\quad
&\ell_3=2\ln{\frac{1+a}{1-a}},\\
\tau_{1,2}={\rm arccosh}\left[\frac{2a^2-1}{a^2(1-b^2)}-1\right],&\quad
&\tau_3=\ln{\frac{1+a}{1-a}}.
\end{eqnarray}
Since the Teichm\"uller space ${\cal T}_2$ is homeomorphic to $\mathbb{R}^6$, one can
identify the FN variables with global coordinates on it. Moreover, the Teichm\"uller space
carries additional structure, namely, the Weil--Petersson (WP) symplectic two-form. Due to
a theorem by Wolpert~\cite{IT,Wo08}, WP-form for compact closed Riemann surfaces of genus $g$
takes on a particularly simple form in terms of FN variables,
\begin{equation}\label{wwp}
\omega_{\rm WP}=\frac{1}{2}\sum\limits_{k=1}^{3g-3}\rmd\ell_k\wedge\rmd\tau_k,
\end{equation}
which is invariant with respect to any pants decomposition. Introducing $\theta_k=2\pi\tau_k/\ell_k$,
the simple Dehn twist $\theta_k\to\theta_k+2\pi$ gives us isometrically the same surface.
Substituting the expressions for $\ell_k$ and $\tau_k$ into (\ref{wwp}), WP-form becomes
\begin{equation}\label{wwp2}
\omega_{\rm WP}=\frac{8a}{(1-a^2)(2a^2\cos^2{\tilde\alpha}-1)}\rmd a\wedge \rmd\tilde\alpha.
\end{equation}
To verify the uniqueness of last formula, we can consider another pants decomposition.
In terms of our parameters, a new decomposition simply leads to replacement,
\begin{equation}\label{z2}
a\leftrightarrow b,\qquad \tilde\alpha\leftrightarrow-\tilde\alpha,
\end{equation}
in length and twist functions of previous decomposition. Although the set of new functions
is obtained, the resulting two-form $\omega_{\rm WP}$ remains the same.
Thus, domain ${\cal A}$ of admissible parameters $(a,\alpha)$ is symplectic manifold
$({\cal A},\omega_{\rm WP})$. We may also treat two-form (\ref{wwp2}) as an area
element of ${\cal A}$.
\subsection{Isoperimetric Orbits}
Since the hyperbolic area of admissible octagons is always equal to $4\pi$, a simplest
way to control the surface changes globally consists in consideration of octagon perimeter:
\begin{equation}\label{Per}
P=8~{\rm arccosh}~\frac{1-a^2b^2+\sqrt{(1-a^2)^2+(1-b^2)^2}}{(1-a^2)(1-b^2)}.
\end{equation}
Although perimeter $P$ is obviously invariant under automorphism and pants decomposition, it can also
take on the same value for different pairs $(a,\alpha)$. We are aiming to describe the corresponding orbits.
On the contrary to our case, two parameters of {\it flat} octagon with the same automorphism and
the homothety property are exactly defined by fixing its area and perimeter.
For further calculations it is useful to introduce two auxiliary quantities:
\begin{equation}\label{varconv}
T\equiv\tanh{(P/16)}, \qquad \varepsilon=\pm1,
\end{equation}
where the latter one reflects an existence of two sheets in ${\cal A}$ labeled by
${\rm sign}~\tilde\alpha$.
For a given $T(P)$, maximal and minimal values of $a$ at $\tilde\alpha=0$ are
\begin{equation}\label{apm}
a_\pm(T)=\frac{1}{2}\sqrt{2+T^2\pm\sqrt{(2+T^2)^2-8}}.
\end{equation}
Solution to equation $(2+T^2)^2=8$ is $T_{\rm reg}=\sqrt{2\sqrt{2}-2}$ that results in
quantities $P_{\rm reg}=8~{\rm arccosh}~(5+4\sqrt{2})$ and $a_{\rm reg}=2^{-1/4}$
corresponding to the regular octagon. Thus, $P_{\rm reg}\approx24.457$ is a minimal value of $P$
among possible ones and trajectory in ${\cal A}$ for $P_{\rm reg}$ is contracted to a point.
Isoperimetric orbits can then be parametrized as
\begin{eqnarray}
a(T,\varphi)&=&\frac{1}{2}\sqrt{2+T^2+\cos{\varphi}\sqrt{(2+T^2)^2-8}},
\label{asol}\\
\tilde\alpha(T,\varphi)&=&\arctan{\frac{\sqrt{2}\sqrt{(2+T^2)^2-8}\sin{\varphi}}
{2\sqrt{3T^2-2-\cos{\varphi}\sqrt{(2+T^2)^2-8}}}},
\label{alf}
\end{eqnarray}
where cyclic variable $\varphi\in[0,2\pi)$ is used.
\begin{figure}
\begin{center}
\includegraphics[width=6.2cm]{fig2a.eps}\ \
\includegraphics[width=5.9cm]{fig2b.eps}
\end{center}
\vspace*{-3mm}
\caption{\small Left panel: Parameter space ${\cal A}$ and orbits of constant perimeter.
Right panel: WP-Area of domain bounded by isoperimetric curve. Solid curve is computed
numerically; dashed curve is given by analytic approximation (\ref{Alog}).}
\end{figure}
These orbits are shown in Fig.~2 (left panel) for $P$ from $P=25$ to $P=41$ with step 2. Although
a role of octagon perimeter is unclear from physical point of view, it is important that the
isoperimetric constraint guarantees the dense covering of ${\cal A}$, which has an unique shape, by
the corresponding orbits. This fact allows us to quantize $({\cal A},\omega_{\rm WP})$ in a spirit
of \cite{Hurt}. To realize it, we should identify the Weil-Petersson area $A_{\rm WP}(P)$ of domain
in ${\cal A}$, bounded by curve for fixed $P$, with an action variable, that is, integral of
``motion''. Then, quantization has to give us the number of quantum ``cells'' inside of the domain.
It seems enough if additional physics (except quantization) is not used. However, we shall require
SO(2,1) symmetry of dynamics in accordance with Lorentzian (2+1)-dimensional gravity.
Using the symplectic form (\ref{wwp2}), let us define WP-area:
\begin{eqnarray}
A_{\rm WP}(P)&\equiv&\int_{P={\rm const}}\omega_{\rm WP}\nonumber\\
&=&\int_{x_-(T)}^{x_+(T)}
\frac{8\rmd x}{(1-x)\sqrt{2x-1}}~{\rm arccosh}~f(x,T),
\label{AA}
\end{eqnarray}
where $x\equiv a^2$; functions $x_\pm(T)\equiv a^2_\pm(T)$ are determined by (\ref{apm});
\begin{equation}
f(x,T)=\frac{1}{\sqrt{1-T^2}}\sqrt{\frac{2x-1}{x}}\sqrt{\frac{T^2-x}{T^2-2x+1}},
\end{equation}
and convention (\ref{varconv}) is applied.
Analytic estimation of WP-area is made in Appendix~A. We show in Fig.~2 (right panel)
that $A^{(0)}_{\rm WP}(P)$ tends to numerically computed $A_{\rm WP}(P)$ at large $P$, but
behavior at $P\to P_{\rm reg}$ looks incorrect. However, we shall see that the derivative
$\rmd A_{\rm WP}(P)/\rmd P$ can be calculated explicitly.
\subsection{Global Canonical Variables}
Finding the canonical variables for isoperimetric orbits, let us define
\begin{equation}
Q(x,T)=\frac{\sqrt{2}}{4}\frac{1-T^2}{\sqrt{T^2-x_-}}\left[
\frac{F(u,k)}{1-T^2}-\frac{\Pi(u,\nu_1,k)}{1-x_-}+\frac{\Pi(u,\nu_2,k)}{T^2-2x_-+1}\right],
\end{equation}
where $F$ and $\Pi$ are elliptic integrals of the first and third kind, respectively.
Amplitude $u$, module $k$, and parameters $\nu_1$ and $\nu_2$ are given by
\begin{eqnarray}
&&u=\sqrt{\frac{x-x_-}{x_+-x_-}},\qquad
k=\sqrt{\frac{x_+-x_-}{T^2-x_-}},\nonumber\\
&&\nu_1=\frac{x_+-x_-}{1-x_-},\qquad
\nu_2=2\frac{x_+-x_-}{T^2-2x_-+1}.
\end{eqnarray}
Then, symplectic WP-form becomes
\begin{equation}\label{WW}
\omega_{\rm WP}=\varepsilon\rmd Q(x,T)\wedge\rmd P.
\end{equation}
Integrating over $x$, we see that
\begin{eqnarray}\label{dAWP}
\frac{\rmd A_{\rm WP}}{\rmd P}&\equiv&\oint_{P={\rm const}}\varepsilon\rmd Q(x,T)
=\int_{x_-}^{x_+}\rmd Q(x,T)-\int_{x_+}^{x_-}\rmd Q(x,T)\nonumber\\
&=&2Q(x_+,T),
\end{eqnarray}
where amplitude $u=1$ results in the complete elliptic integrals.
Defining an action variable (or ``angular momentum'') as
\begin{equation}\label{AM}
J(P)=\frac{1}{4\pi}A_{\rm WP}(P),
\end{equation}
we find the angle variable $\Phi$ from equation $\omega_{\rm WP}=2\rmd\Phi\wedge\rmd J$;
the Poisson bracket is then $\{J,\Phi\}_{\rm WP}=1$. One first has that
\begin{equation}
\rmd\Phi=\frac{\pi\varepsilon}{Q(x_+,T)}\rmd Q(x,T),\qquad
\oint_{P={\rm const}}\rmd\Phi=2\pi,
\end{equation}
where the rule of calculation of integrals containing $\varepsilon$ (see (\ref{dAWP})) is applied.
At fixed $P$, we come to expression for the angle variable:
\begin{equation}\label{Phi}
\Phi=\pi\varepsilon\frac{Q(x,T)}{Q(x_+,T)},\quad
\Phi\in[-\pi,\pi].
\end{equation}
Now it seems trivially to quantize the system in terms of $J$ and $\Phi$ what leads
immediately to estimation (in the Planck units) for relatively large $n$:
\begin{equation}
A_{\rm WP}\sim4\pi n,\qquad n\in\mathbb{N}.
\end{equation}
We specify this formula below due to consideration of the relativity theory.
\section{Relativistic Kinematics}
At this stage, geometrodynamics of Riemann surface within the considered model is
ambiguous because of purely gauge nature. There are only the geometric constraints
defining the ``physical sector'' of parameters variety and no definitions of time and
the Hamiltonian function having the physical meaning and generating an evolution of
$J$ and $\Phi$. However, we appeal here to (2+1)-dimensional gravity
where ${\rm SO}(2,1)\sim{\rm SU}(1,1)$ plays a role of the Lorentz group. It allows
us to construct a dynamical model with the same symmetry as follows.
Combining $J$ and $\Phi$, we extend the set of observables up to
\begin{equation}\label{su}
J_0=J,\quad
J_\pm=\sqrt{J^2-C}\exp{(\mp\rmi\Phi)},
\end{equation}
where $C$ is a constant such that $J^2\geqslant C\geqslant 0$.
The Poisson algebra of new variables is $\mathfrak{su}(1,1)$ Lie algebra:
\begin{equation}
\{J_+,J_-\}_{\rm WP}=2\rmi J_0,\qquad
\{J_\pm,J_0\}_{\rm WP}=\pm\rmi J_\pm.
\end{equation}
More generally, generators $J_{0,\pm}$ may be replaced by an infinite
number of quantities $L_n=J\exp{(\rmi n\Phi)}$, $n\in{\mathbb Z}$,
generating the Witt algebra.
Since the evolution of conservative system is usually described by canonical transformations,
we find that indeed SU(1,1) transformations are canonical transformations of a given system.
Let us introduce the matrix
\begin{equation}
{\cal M}=\left(
\begin{array}{cc}
J_0&J_+\\
J_-&J_0
\end{array}
\right),
\end{equation}
whose determinant $C\equiv J^2_0-J_+J_-$ is the Casimir of $\mathfrak{su}(1,1)$ algebra.
Action of matrix $U\in{\rm SU}(1,1)$ looks like ${\cal M}\mapsto\tilde{\cal M}=U{\cal M}U^\dag$
and preserves the Casimir, $\det\tilde{\cal M}=C$.
\begin{figure}
\begin{center}
\includegraphics[width=6.0cm]{fig3a.eps}\ \
\includegraphics[width=6.3cm]{fig3b.eps}
\end{center}
\vspace*{-3mm}
\caption{\small Angular momentum $J$ and angle $\Phi$ as functions of
internal time $\tau$; $C=0$ and initial data are $a_0=0.8$, $\alpha_0=\pi/3$. The
surface starts with an infinite WP-area at $\tau\to-\infty$, contracts and bounces
to expand again to infinite area at $\tau\to\infty$.}
\end{figure}
Relativistic kinematics of the surface can be generated by the boost:
\begin{equation}
U_\tau=\left(\begin{array}{cc}
\cosh{(\tau/2)}&\sinh{(\tau/2)}\\
\sinh{(\tau/2)}&\cosh{(\tau/2)}
\end{array}\right).
\end{equation}
Computing ${\cal M}(\tau)=U_\tau{\cal M}(0)U^\dag_\tau$, one derives the trajectories
for $J_{0,\pm}$ or, equivalently, for $J$ and $\Phi$:
\begin{eqnarray}
J(\tau)&=&\breve{J}\cosh{\tau}+\sqrt{\breve{J}^2-C}\sinh{\tau}\cos{\breve{\Phi}},\label{Jt}\\
\Phi(\tau)&=&
\arccos{\frac{\breve{J}\sinh{\tau}+\sqrt{\breve{J}^2-C}\cosh{\tau}\cos{\breve{\Phi}}}{\sqrt{J^2(\tau)-C}}},
\label{Phit}
\end{eqnarray}
where $\breve{J}=J[P(a_0,\alpha_0)]$, $\breve{\Phi}=\Phi[a_0,P(a_0,\alpha_0)]$ are
the values of functions (\ref{AM}), (\ref{Phi}) at $\tau=0$.
On the other hand, the evolution can be described by the Hamiltonian equations:
\begin{eqnarray}
\partial_\tau J&\equiv&\{J,H\}_{\rm WP}=\sqrt{J^2-C}\cos{\Phi},\\
\partial_\tau\Phi&\equiv&\{\Phi,H\}_{\rm WP}=-\frac{J}{\sqrt{J^2-C}}\sin{\Phi},
\end{eqnarray}
where the Hamiltonian
\begin{equation}\label{H}
H(J,\Phi,C)=\sqrt{J^2-C}\sin{\Phi}
\end{equation}
belongs to $\mathfrak{so}(2,1)$ algebra.
Of course, the form of the Hamiltonian depends generally on the physical problem under
consideration. In relativistic cosmology, different scenarios lead to modifications
of (\ref{H}) which are discussed, for instance, in \cite{LM}.
In our case, time dependence of functions $J(\tau)$, $\Phi(\tau)$ is sketched in Fig.~3.
The figures demonstrate a ``bounce'' in parameter space. Finding the zeroes of derivatives
of (\ref{Jt}), (\ref{Phit}) with respect to $\tau$, the bounce characteristics are
\begin{eqnarray}
&&\tau_{\rm b}=-{\rm arctanh}\left(\frac{\sqrt{\breve{J}^2-C}}{\breve{J}}\cos{\breve{\Phi}}\right),
\nonumber\\
&&J_{\rm b}=\sqrt{\breve{J}^2\sin^2\breve{\Phi}+C\cos^2\breve{\Phi}},\quad
\Phi_{\rm b}=\varepsilon\frac{\pi}{2}.
\end{eqnarray}
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fig4.eps}
\end{center}
\vspace*{-3mm}
\caption{\small The trajectory of Riemann surface evolution for $a_0=0.8$,
$\alpha_0=\pi/3$ and $C=0$. Bold line corresponds to $\varepsilon=1$, dashed
line is for $\varepsilon=-1$. Closed curve is the isoperimetric orbit determined
by initial data.}
\end{figure}
In principal, we are interested here in the form of trajectory in parameter space ${\cal A}$ without
specification of evolution parameter or time. It can be found by means of constraint
\begin{equation}\label{Hc}
H\{J[P(a,\alpha)],\Phi[a,P(a,\alpha)],C\}=E,\qquad
(a,\alpha)\in{\cal A},
\end{equation}
where constant $E$ is the value of Hamiltonian function $H$ for initial data.
Although Eq.~(\ref{Hc}) says at first sight that the points of trajectory are simply determined from
an abstract equation $h(a,\alpha,C,E)=0$ with additional constants $C$ and $E$, we would like to
emphasize its structure reflecting the chain of our buildings.
In terms of $a$ and $\alpha$ the trajectory looks like in Fig.~4, where influence of $C$ is neglected
because of quantum nature assumed. It is interesting to note that the system geometry does not tend to
the regular octagon having a maximal information entropy. At the values $a_{\rm min}=1/\sqrt{2}$ and
$a_{\rm max}=1$ corresponding to the infinite past and the infinite future time, angle $\alpha$ of
hyperbolic octagon reaches the same value $\pi/4$. Thus, geometry with $\alpha=\pi/4$ is exceptional in
our model. We would like also to note that there are two configurations for $\varepsilon=\pm1$ (left
and right with respect to $\tilde\alpha=0$), which are not mixed during whole evolution generated by
the pure boost. The chosen preference is preserved from the origin to the end. However, we have already
seen that there is the diffeomorphism generated by conservation condition of hyperbolic octagon perimeter,
which allows us the transition between ``phases'' with $\varepsilon=\pm1$. Combining it with an action of
the boost, it is possible to construct a novel scenario of classical geometrodynamics, even in the case
when the Riemannian metric of the parameter (or moduli) space is undetermined.
Coming back to quantization problem, the generator $J_0$ and its spectrum describes WP-area.
For this reason its eigenvalues should be discrete and positive. We choose the irreducible representation
(\ref{su11}) with standard basis diagonalizing the Casimir and $J_0$ and with minimal positive spin
$j=1/2$. It leads straightforwardly to the spectrum (in the Planck units):
\begin{equation}
A_{\rm WP}=4\pi\left(n+\frac{1}{2}\right),\qquad
n\in\mathbb{N}.
\end{equation}
In classical picture, $A_{\rm WP}=0$ for $P_{\rm reg}$, while $A_{\rm WP}$ is always non-vanishing at
quantum level for the system with a given topology. It means that the regular octagonal configuration
is not achieved because of quantum effect.
Although the basis of $\mathfrak{su}(1,1)$ is enough to quantize $A_{\rm WP}$ as a global characteristic of
${\cal A}$, it looks insufficient to apply for finding the spectra of other geometric observables. Further
investigations are still needed.
\section{Conclusions}
Here, we pay great attention to the structure of admissible parameters space ${\cal A}$
determining the geometry of the Riemann surface in genus two with an order four automorphism.
First, using the Weil--Petersson (WP) geometry, two-dimensional space ${\cal A}$ is equipped
with the fundamental symplectic two-form. Further, we perform the dense covering of ${\cal A}$
by the orbits generated by the isoperimetric constraint which is imposed on fundamental domain
of the surface. It is argued that an existence of these orbits is due to the Riemann surface
definition. The canonically conjugate action--angle variables for isoperimetric orbits are
found by identifying WP-area of domain bounded by the orbit with the action variable.
As the result, we take on a possibility to construct relativistic model in terms of special
invariants without knowing the Riemannian metric of ${\cal A}$.
To build the physically meaningful model, we extend the set of global canonical variables up
to generators of $\mathfrak{su}(1,1)$ algebra. This trick leads to appearance of the
Casimir playing a role of additional parameter. We can only assume that its value should be
minimal and non-zero. However, it leads after quantization to non-vanishing discrete spectrum
of WP-area in a contrast with initial theory where WP-area becomes zero for the Riemann
surface associated with the regular hyperbolic octagon. In any case, we should remember that
WP-area determines the surface geometry up to diffeomorphism in classical theory.
We may need to use the infinite Witt and Virasoro algebras instead of $\mathfrak{su}(1,1)$
in order to describe the system spectrum.
We also consider relativistic kinematics or free geometrodynamics of the Riemann surface,
generated by the Lorentz boost which acts on the constructed generators of $\mathfrak{su}(1,1)$
algebra. The time dependence of global variables leads to ``big bounce'' scenario and is
similar for quantities of different origin. However, solving equations with respect to
the surface parameters, we have obtained the trajectory (independent on time definition) in
space ${\cal A}$. In this picture, the system does not tend to reach the regular octagon
configuration corresponding to the maximal information entropy~\cite{Naz} and preserves some
kind of $\mathbb{Z}_2$ symmetry related to admissibility range of angle variable during whole
evolution.
\section*{Acknowledgments}
A.N. is deeply indebted to A.M.~Gavrilik (BITP, Kiev) and I.V.~Mykytiuk (IAPMM, Lviv)
for fruitful discussions of mathematical aspects of the problem.
|
\section{Introduction and preliminaries}
For a net $(x_\alpha)$ in a vector lattice $X$, we write
$x_\alpha\goeso x$ if $(x_\alpha)$ \term{converges} to $x$ \term{in
order}. That is, there is a net $(u_\gamma)$, possibly over a
different index set, such that $u_\gamma\downarrow 0$ and for every
$\gamma$ there exists $\alpha_0$ such that
$\abs{x_\alpha-x}\le u_\gamma$ whenever $\alpha\ge\alpha_0$. We write
$x_\alpha\goesuo x$ and say that $(x_\alpha)$ \term{uo-converges} to
$x$ if $\abs{x_\alpha-x}\wedge u\goeso 0$ for every $u\in X_+$; ``uo''
stands for ``unbounded order''. For a net $(x_\alpha)$ in a normed
lattice $X$, we write $x_\alpha\goesnorm x$ if $(x_\alpha)$ converges
to $x$ in norm. We write $x_\alpha\goesun x$ and say that $(x_\alpha)$
\term{un-converges} to $x$ if $\abs{x_\alpha-x}\wedge u\goesnorm 0$
for every $u\in X_+$; ``un'' stands for ``unbounded norm''.
A variant of uo-convergence was originally introduced in
\cite{Nakano:48}, while the term ``uo-convergence'' was first coined
in \cite{DeMarr:64}. Relationships between uo, weak, and weak*
convergences were investigated in
\cite{Wickstead:77,GaoX:14,Gao:14}. Relationships between
uo-convergence and almost everywhere convergence were investigated and
applied in \cite{GaoX:14,Emelyanov:16,GTX}. We refer the
reader to \cite{GTX} for a further review of properties of
uo-convergence. Un-convergence was introduced in \cite{Troitsky:04}
and further investigated in \cite{DOT}. For unexplained terminology on
vector and Banach lattices we refer the reader to
\cite{Abramovich:02,Aliprantis:06}. All vector lattices are assumed to
be Archimedean.
Let us start by briefly going over some of the known properties of
these modes of convergence; we refer the reader to \cite{GTX,DOT} for
details. Both uo-convergence and un-convergence respect linear and
lattice operations; limits are unique. In particular,
$x_\alpha\goesuo x$ iff $\abs{x_\alpha-x}\goesuo 0$; similarly,
$x_\alpha\goesun x$ iff $\abs{x_\alpha-x}\goesun 0$. For order bounded
nets, uo-convergence agrees with order convergence while
un-convergence agrees with norm convergence. It follows that order
intervals are uo- and un-closed. For sequences in $L_p(\mu)$, where
$1\le p<\infty$ and $\mu$ is a finite measure, it is easy to see that
uo-convergence agrees with convergence almost everywhere, see, e.g.,
\cite[Example~2]{DeMarr:64}. Under the same assumptions,
un-convergence agrees with convergence in measure, see
\cite[Example~23]{Troitsky:04}. We write $L_p$ for $L_p[0,1]$.
Suppose that $X$ is a vector lattice. By \cite[Corollary~3.6]{GTX},
every disjoint sequence in $X$ is uo-null. Recall that a sublattice
$Y$ of $X$ is \term{regular} if the inclusion map preserves suprema
and infima of arbitrary subsets. It was shown
in~\cite[Theorem~3.2]{GTX} that uo-convergence is stable under passing
to and from regular sublattices. That is, if $(y_\alpha)$ is a net in
a regular sublattice $Y$ of $X$ then $y_\alpha\goesuo 0$ in $Y$ iff
$y_\alpha\goesuo 0$ in $X$ (in fact, this property characterizes
regular sublattices).
It is clear that if $X$ is an order continuous normed lattice then
uo-convergence implies un-convergence. Let $X$ be a Banach lattice and
$(x_n)$ a un-null sequence in $X$. Then $(x_n)$ has a uo-null
subsequence by Proposition~4.1 of~\cite{DOT}. A disjoint sequence need
not be un-null. For example, the standard unit sequence $(e_n)$ in
$\ell_\infty$ is not un-null. However, a un-null sequence has an
asymptotically disjoint subsequence. More precisely, we have the
following.
\begin{theorem}(\cite[Theorem~3.2]{DOT})\label{KP}
Let $(x_\alpha)$ be a un-null net. There is an increasing sequence
of indices $(\alpha_k)$ and a disjoint sequence $(d_k)$ such that
$x_{\alpha_k}-d_k\goesnorm 0$.
\end{theorem}
While uo-convergence need not be given by a topology, it was observed
in \cite{DOT} that un-convergence is topological. For every
$\varepsilon>0$ and non-zero $u\in X_+$, put
\begin{displaymath}
V_{\varepsilon,u}=\bigl\{x\in X\mid \bignorm{\abs{x}\wedge u}<\varepsilon\bigr\}.
\end{displaymath}
The collection of all sets of this form is a base of zero
neighborhoods for a topology, and the convergence in this topology
agrees with un-convergence. We will refer to this topology
as \emph{un-topology}.
Every time a new linear topology is discovered, one is expected to ask
several natural questions: is this topology metrizable? Is it
locally-convex? Complete? Can one characterize (relatively) compact
sets? Is this topology stronger or weaker than other known topologies?
In this paper, we study these and similar questions for
un-topology. In other words, our motivation for this paper is to
investigate topological properties of un-topology.
Throughout this paper, $X$ will be assumed to be a Banach
lattice, unless specified otherwise. We write $B_X$ for the closed
unit ball of $X$. It was observed in~\cite{DOT} that
$x_\alpha\goesun x$ implies $\norm{x}\le\liminf\norm{x_\alpha}$. This
yields that $B_X$ is un-closed.
The following facts will be used throughout the paper.
\begin{lemma}\label{monot}
\begin{enumerate}
\item\label{monot-uo} If $(x_\alpha)$ is an increasing net in a
vector lattice $X$ and $x_\alpha\goesuo x$ then
$x_\alpha\uparrow x$;
\item\label{monot-un} If $(x_\alpha)$ is an increasing net in a
normed lattice $X$ and $x_\alpha\goesun x$ then
$x_\alpha\uparrow x$ and $x_\alpha\goesnorm x$.
\end{enumerate}
\end{lemma}
\begin{proof}
Without loss of generality, $x_\alpha\ge 0$ for all $\alpha$;
otherwise, pick any index $\alpha_0$ and consider the net
$(x_\alpha-x_{\alpha_0})_{\alpha\ge\alpha_0}$, which converges to
$x-x_{\alpha_0}$. Since lattice operations are uo- and un-continuous,
we have $x\ge 0$.
\eqref{monot-uo} Take any $z\in X_+$. It follows from uo-continuity
of lattice operations that $x_\alpha\wedge z\goesuo x\wedge z$.
Since the net $(x_\alpha\wedge z)$ is order bounded and increasing,
this yields $x_\alpha\wedge z\goeso x\wedge z$ and, therefore
$x_\alpha\wedge z\uparrow x\wedge z$. It follows that
$x_\alpha\wedge z\le x$ for every $\alpha$ and every $z\in
X_+$. Applying this with $z=x_\alpha$ we get
$x_\alpha\le x$. Thus, the net $(x_\alpha)$ is order bounded and,
therefore, $x_\alpha\goeso x$, hence $x_\alpha\uparrow x$.
\eqref{monot-un} The proof is similar and uses the fact that every
monotone norm convergent net converges in order to the same
limit. We note that $x_\alpha\wedge z\goesnorm x\wedge z$ and,
therefore, $x_\alpha\wedge z\uparrow x\wedge z$ for every $z\in
X_+$. It follows that the net $(x_\alpha)$ is order bounded, which
yields $x_\alpha\goesnorm x$ and, therefore, $x_\alpha\uparrow x$.
\end{proof}
Recall that \cite[Question~2.14]{DOT} asks whether $x_\alpha\goesun 0$
implies that there exists an increasing sequence of indices
$(\alpha_k)$ such that $x_{\alpha_k}\goesun 0$. The following
counterexample was kindly provided to us by E.~Emelyanov.
\begin{example}\label{emelyanov}
Let $\Omega$ be an uncountable set; let $X$ be the closed sublattice
of $\ell_\infty(\Omega)$ consisting of all the functions with
countable support. For $\omega\in\Omega$, we write $e_\omega$ for
the characteristic function of $\{\omega\}$.
Let $\Lambda$ be the set of all countable subsets of $\Omega$,
ordered by inclusion. For each $\alpha\in\Lambda$, pick any
$\omega\notin\alpha$ and put $x_\alpha=e_\omega$. We claim that
$x_\alpha\goesun 0$. Indeed, let $u\in X_+$; let $\alpha_0$ be the
support of $u$. Then $x_\alpha\wedge u=0$ whenever
$\alpha\ge\alpha_0$.
On the other hand, let $(\omega_k)$ be any sequence in $\Omega$; we
claim that the sequence $(e_{\omega_k})$ is not un-null. Indeed, put
$\beta=\{\omega_k\mid k\in\mathbb N\}$ and let $u$ be the
characteristic function of $\beta$. Then
$e_{\omega_k}\wedge u=e_{\omega_k}$ for every $k$; hence it does not
converge in norm to zero.
In particular, if $(\alpha_k)$ is an increasing sequence of indices
in $\Lambda$ then $(x_{\alpha_k})$ is not un-null.
\end{example}
Let $e\in X_+$. Recall that the band $B_e$ generated by $e$ is norm
closed and contains the principal ideal $I_e$; hence
$I_e\subseteq\overline{I_e}\subseteq B_e$. Recall also that
\begin{itemize}
\item $e$ is a \term{strong unit} when $I_e=X$; equivalently, for
every $x\ge 0$ there exists $n\in\mathbb N$ such that $x\le ne$;
\item $e$ is a \term{quasi-interior point} if $\overline{I_e}=X$;
equivalently, $x\wedge ne\goesnorm x$ for every $x\in X_+$;
\item $e$ is a \term{weak unit} if $B_e=X$;
equivalently, $x\wedge ne\uparrow x$ for every $x\in X_+$.
\end{itemize}
In particular,
strong unit $\Rightarrow$ quasi-interior point
$\Rightarrow$ weak unit.
\section{Strong units}
It is easy to see that each $V_{\varepsilon,u}$ is solid. It is also
absorbing, that is, for every $x\in X$ there exists $\lambda>0$ such
that $\lambda x\in V_{\varepsilon,u}$. The following lemma is a
dichotomy: it says that $V_{\varepsilon,u}$ is either ``very small''
or ``very large''.
\begin{lemma}\label{dichotomy}
Let $\varepsilon>0$, and $0\ne u\in X_+$. Then $V_{\varepsilon,u}$
is either contained in $[-u,u]$ or contains a non-trivial ideal.
\end{lemma}
\begin{proof}
Suppose that $V_{\varepsilon,u}$ is not contained in $[-u,u]$. Then
there exists $x\in V_{\varepsilon,u}$ such that
$x\notin[-u,u]$. Replacing $x$ with $\abs{x}$, we may assume that
$x>0$. Let $y=(x-u)^+$; then $y>0$. It is an easy exercise to show
that
\begin{math}
(\lambda y)\wedge u\le x\wedge u
\end{math}
for every $\lambda\ge 0$; it follows that $\lambda y\in
V_{\varepsilon,u}$. Since $V_{\varepsilon,u}$ is solid, it contains
the principal ideal $I_y$.
\end{proof}
\begin{lemma}\label{V-bdd-su}
If $V_{\varepsilon,u}$ is contained in $[-u,u]$ then $u$ is a strong unit.
\end{lemma}
\begin{proof}
Let $x\in X_+$. There exists $\lambda>0$ such that $\lambda x\in
V_{\varepsilon,u}$, hence $\lambda x\in [-u,u]$. It follows that $u$
is a strong unit.
\end{proof}
Recall that if $e$ is a positive vector in $X$ then the principal
ideal $I_e$ equipped with the norm
\begin{displaymath}
\norm{x}_e=\inf\bigl\{\lambda>0\mid\abs{x}\le\lambda e\bigr\}
\end{displaymath}
is lattice isometric to $C(K)$ for some compact Hausdorff space $K$,
with $e$ corresponding to the constant one function $\one$; see, e.g.,
Theorems~3.4 and~3.6 in~\cite{Abramovich:02}. If $e$ is a strong unit
in $X$ then $I_e=X$; it is easy to see that in this case
$\norm{\cdot}_e$ is equivalent to the original norm; it follows that $X$ is
lattice and norm isomorphic to $C(K)$.
It is easy to see that if $x_\alpha\goesnorm x$ then
$x_\alpha\goesun x$, so norm topology generally is stronger than
un-topology.
\begin{theorem}\label{su}
Let $X$ be a Banach lattice. The following are equivalent.
\begin{enumerate}
\item Un-topology agrees with norm topology;
\item $X$ has a strong unit.
\end{enumerate}
\end{theorem}
\begin{proof}
Suppose that un-topology and norm topology agree. It follows that
$V_{\varepsilon,u}$ is contained in $B_X$ for some $\varepsilon>0$
and $u>0$. By \Cref{dichotomy}, we conclude that $V_{\varepsilon,u}$
is contained in $[-u,u]$; hence $u$ is a strong unit by \Cref{V-bdd-su}.
Suppose now that $X$ has a strong unit. Then $X$ is lattice and norm
isomorphic to $C(K)$ for some compact Hausdorff space $K$. Without
loss of generality, $X=C(K)$. It follows from $x_\alpha\goesun 0$
that $\abs{x_\alpha}\wedge\one\goesnorm 0$. Since the norm in $C(K)$
is the $\sup$-norm, it is easy to see that $x_\alpha\goesnorm 0$.
\end{proof}
\section{Quasi-Interior points and metrizability}
Given a net $(x_\alpha)$ in a vector lattice with a weak unit $e$,
then $x_\alpha\goesuo x$ iff $\abs{x_\alpha-x}\wedge e\goeso 0$; see,
e.g., \cite[Corollary~3.5]{GTX} (this was proved in~\cite{Kaplan:97}
in the special case when the lattice is order complete). That is, it
suffices to test uo-convergence on a weak unit. Lemma~2.11
in~\cite{DOT} provides a similar statement for un-convergence and
quasi-interior points. We now prove that this property actually
characterizes quasi-interior points.
\begin{theorem}\label{qip}
Let $e\in X_+$. The following are equivalent.
\begin{enumerate}
\item\label{qip-un} $e$ is a quasi-interior point;
\item\label{qip-net} For every net $(x_\alpha)$ in $X_+$, if $x_\alpha\wedge
e\goesnorm 0$ then $x_\alpha\goesun 0$;
\item\label{qip-seq} For every sequence $(x_n)$ in $X_+$, if $x_n\wedge
e\goesnorm 0$ then $x_n\goesun 0$.
\end{enumerate}
\end{theorem}
\begin{proof}
The implication \eqref{qip-un}$\Rightarrow$\eqref{qip-net} was
proved
in~\cite[Lemma~2.11]{DOT}.
\eqref{qip-net}$\Rightarrow$\eqref{qip-seq} is trivial. This leaves
\eqref{qip-seq}$\Rightarrow$\eqref{qip-un}.
Suppose \eqref{qip-seq}. Fix $x\in X_+$. We need to show that
$x\wedge ne\goesnorm x$ or, equivalently $(x-ne)^+\goesnorm 0$ as a
sequence of $n$. Put $u=x\vee e$. The ideal $I_u$ is lattice
isomorphic (as a vector lattice) to $C(K)$ for some compact space
$K$, with $u$ corresponding to $\one$. Since $x,e\in I_u$, we may
consider $x$ and $e$ as elements of $C(K)$. Note that $x\vee e=\one$
implies that $x$ and $e$ never vanish simultaneously.
For each $n\in\mathbb N$, we define
\begin{displaymath}
F_n=\bigl\{t\in K\mid x(t)\ge ne(t)\bigl\}
\text{ and }
O_n=\bigl\{t\in K\mid x(t)>ne(t)\bigl\}.
\end{displaymath}
Clearly, $O_n\subseteq F_n$, $O_n$ is open, and $F_n$ is closed.
\emph{Claim 1}: $F_{n+1}\subseteq O_n$. Indeed, let $t\in
F_{n+1}$. Then $x(t)\ge (n+1)e(t)$. If $e(t)>0$ then $x(t)>ne(t)$,
so that $t\in O_n$. If $e(t)=0$ then $x(t)>0$, hence $t\in O_n$.
By Urysohn's Lemma, we find $z_n\in C(K)$ such that $0\le z_n\le x$,
$z_n$ agrees with $x$ on $F_{n+1}$ and vanishes outside of $O_n$. We can
also view $ z_n$ as an element of $X$.
\emph{Claim 2}: $n(z_n\wedge e)\le x$. Let $t\in K$. If $t\in O_n$
then $n(z_n\wedge e)(t)\le ne(t)<x(t)$. If $t\notin O_n$ then
$z_n(t)=0$, so that the inequality is satisfied trivially.
\emph{Claim 3}: $\bigl(x-(n+1)e\bigr)^+\le z_n$. Again, let
$t\in K$. If $t\in F_{n+1}$ then
$\bigl(x-(n+1)e\bigr)^+\le x(t)=z_n(t)$. If $t\notin F_{n+1}$ then
$x(t)<(n+1)e(t)$, so that $\bigl(x-(n+1)e\bigr)^+(t)=0$ and the
inequality is satisfied trivially.
Now, Claim 2 yields $0\le z_n\wedge e\le\frac{1}{n}x\goesnorm 0$,
so that $z_n\wedge e\goesnorm 0$. By assumption, this yields
$z_n\goesun 0$. Since $0\le z_n\le x$ for every $n$, the sequence
$(z_n)$ is order bounded and, therefore, $z_n\goesnorm 0$. Now
Claim~3 yields $\bigl(x-(n+1)e\bigr)^+\goesnorm 0$, which concludes
the proof.
\end{proof}
\begin{theorem}\label{metriz}
Un-topology is metrizable iff $X$ has a quasi-interior point. If $e$
is a quasi-interior point then $d(x,y)=\bignorm{\abs{x-y}\wedge e}$
is a metric for un-topology.
\end{theorem}
\begin{proof}
Suppose that $e\in X_+$ is a quasi-interior point and put
$d(x,y)=\bignorm{\abs{x-y}\wedge e}$ for $x,y\in X$. It can be
easily verified that this defines a metric on $X$. Indeed,
$d(x,x)=0$ and $d(x,y)=d(y,x)$ for every $x,y\in X$. If $d(x,y)=0$
then $\abs{x-y}\wedge e=0$, hence $\abs{x-y}=0$ because $e$ is a
weak unit, so that $x=y$. The triangle inequality follows from the
fact that
\begin{displaymath}
\abs{x-z}\wedge e\le\abs{x-y}\wedge e+\abs{y-z}\wedge e.
\end{displaymath}
Note also that
$x_\alpha\goesun x$ iff $d(x_\alpha,x)\to 0$ for every net
$(x_\alpha)$ in $X$.
Conversely, suppose that un-topology is metrizable; let $d$ be a
metric for it. For each $n$, let $B_{\frac1n}$ be the ball of radius
$\frac1n$ centred at zero for the metric, that is,
\begin{displaymath}
B_{\frac1n}=\bigl\{x\in X\mid d(x,0)\le \tfrac1n\bigr\}.
\end{displaymath}
Since $B_{\frac1n}$ is a neighborhood of zero for the un-topology,
it contains $V_{\varepsilon_n,u_n}$ for some $\varepsilon_n>0$ and
$u_n>0$. Let $M_n=2^n\norm{u_n}+1$; then the series
$e=\sum_{n=1}^\infty\frac{u_n}{M_n}$ converges. Note that $M_n>1$
and $u_n\le M_ne$ for every $n$. We claim that $e$ is a
quasi-interior point.
It suffices that \Cref{qip}\eqref{qip-net} is satisfied.
Suppose that $x_\alpha\wedge e\goesnorm 0$ for some net $(x_\alpha)$
in $X_+$. Fix $n$. It follows from
\begin{displaymath}
x_\alpha\wedge u_n\le (M_nx_\alpha)\wedge(M_ne)=M_n(x_\alpha\wedge
e)\goesnorm 0
\end{displaymath}
that $x_\alpha\wedge u_n\goesnorm 0$. Then
there exists $\alpha_0$ such that $\norm{x_\alpha\wedge
u_n}<\varepsilon_n$ whenever $\alpha\ge\alpha_0$. Consequently,
$x_\alpha$ is in $V_{\varepsilon_n,u_n}$ and, therefore, in
$B_{\frac1n}$. It follows that $x_\alpha\to 0$ in the metric, hence
$x_\alpha\goesun 0$.
\end{proof}
Note that a linear Hausdorff topological space is metrizable iff it is
first countable, i.e., has a countable base of neighborhoods of zero,
see, e.g., \cite[pp.~49]{Kelley:63}. Therefore, Theorem~\ref{metriz}
implies, in particular, that un-topology is first countable iff
$X$ has a quasi-interior point. This should be compared with
Corollary~2.13 and Question~2.14 in~\cite{DOT} (we now know from
Example~\ref{emelyanov} that Question~2.14 has a negative answer).
\begin{proposition}\label{un-str-metr}
Un-topology is stronger than or equal to a metric topology iff $X$
has a weak unit.
\end{proposition}
\begin{proof}
Suppose that un-topology is stronger than or equal to a topology
given by a metric. Construct $e$ as in the second part of the proof
of \Cref{metriz}. We claim that $e$ is a weak unit. Suppose that
$x\wedge e=0$. It follows that $x\wedge u_n=0$ for every $n$ and,
therefore, $x\in V_{\varepsilon_n,u_n}$, hence $x\in
B_{\frac1n}$. It follows that $x=0$.
Conversely, let $e\in X_+$ be a weak unit. For $x,y\in X$, define
$d(x,y)=\bignorm{\abs{x-y}\wedge e}$. As in the first part of the
proof of \Cref{metriz}, this is a metric and $x_\alpha\goesun x$
implies $d(x_\alpha,x)\to 0$.
\end{proof}
\subsection*{When is every un-null sequence norm bounded?}
If $X$ has a strong unit then, by \Cref{su}, un-topology agrees with
norm topology, hence every un-null sequence is norm null and, in
particular, norm bounded. This justifies the following question:
\emph{If every un-null sequence in $X$ is norm bounded (or even norm
null), does this imply that $X$ has a strong unit?} The following
example shows that, in general, the answer in negative.
\begin{example}
Let $X$ be as in \Cref{emelyanov}. Clearly, $X$ does not have a
strong unit; it does not even have a weak unit. Yet, every un-null
sequence in $X$ is norm null. Indeed, suppose that $x_n\goesun
0$. Let $u$ be the characteristic function of
$\bigcup_{n=1}^\infty\supp x_n$. By assumption, $\abs{x_n}\wedge
u\goesnorm 0$. It follows that for every $\varepsilon\in(0,1)$ there
exists $n_0$ such that for every $n\ge n_0$ we have
$\bignorm{\abs{x_n}\wedge u}<\varepsilon$. It follows that
$\norm{x_n}<\varepsilon$.
\end{example}
However, we will see that the answer is affirmative under certain
additional assumptions.
Recall that every disjoint sequence is uo-null. Thus, if
$\dim X=\infty$, one can take any non-zero disjoint sequence, scale it
to make it norm unbounded, and thus produce a uo-null sequence which
is not norm bounded. However, this trick does not work for un-topology
because a disjoint sequence need not be un-null. Moreover, we have the
following.
\begin{proposition}\label{disj}
The following are equivalent.
\begin{enumerate}
\item\label{disj-oc} $X$ is order continuous;
\item\label{disj-seq} Every disjoint sequence in $X$ is un-null;
\item\label{disj-net} Every disjoint net in $X$ is un-null.
\end{enumerate}
\end{proposition}
\begin{proof}
\eqref{disj-oc}$\Rightarrow$\eqref{disj-seq} because every disjoint
sequence is uo-null and, therefore, un-null. To show that
\eqref{disj-seq}$\Rightarrow$\eqref{disj-oc}, note that every order
bounded disjoint sequence is norm null and apply
\cite[Theorem~4.14]{Aliprantis:06}.
\eqref{disj-net}$\Rightarrow$\eqref{disj-seq} is trivial. To show
that \eqref{disj-seq}$\Rightarrow$\eqref{disj-net}, suppose that
there exists a disjoint net $(x_\alpha)$ which is not un-null. Then
there exist $\varepsilon>0$ and $u\in X_+$ such that for every
$\alpha$ there exists $\beta>\alpha$ with
$\bignorm{\abs{x_\beta}\wedge u}>\varepsilon$. Inductively, we find
an increasing sequence $(\alpha_k)$ of indices such that
$\bignorm{\abs{x_{\alpha_k}}\wedge u}>\varepsilon$. Hence, the
sequence $(x_{\alpha_k})$ is disjoint but not un-null.
\end{proof}
\begin{corollary}
If $X$ is order continuous and every un-null sequence in $X$ is norm
bounded then $\dim X<\infty$ (and, therefore, $X$ has a strong
unit).
\end{corollary}
\begin{proof}
Suppose $\dim X=\infty$. Then there exists a non-zero disjoint
sequence in $X$. Scaling it if necessary, we may assume that
it is not norm bounded. Yet it is un-null. A contradiction.
\end{proof}
Note that Example~2.7 in~\cite{DOT} is an example of a disjoint but
non un-null sequence in an infinite-dimensional Banach lattice which
is not order continuous and lacks a strong unit.
\begin{proposition}
If $X$ has a quasi-interior point and every un-null sequence is norm
bounded then $X$ has a strong unit.
\end{proposition}
\begin{proof}
By \Cref{metriz}, the un-topology on $X$ is metrizable. Fix such a
metric. As before, for each $n$, let $B_{\frac1n}$ be the ball of
radius $\frac1n$ centred at zero for the metric. For each $n$,
$B_{\frac1n}$ contains $V_{\varepsilon_n,u_n}$ for some
$\varepsilon_n>0$ and $u_n>0$. If
$V_{\varepsilon_n,u_n}\subseteq[-u_n,u_n]$ for some $n$ then $u_n$
is a strong unit by \Cref{V-bdd-su}. Otherwise, by \Cref{dichotomy},
each $V_{\varepsilon_n,u_n}$ contains a non-trivial ideal. Pick any
$x_n$ in this ideal with $\norm{x_n}=n$. Then the sequence $(x_n)$
is norm unbounded; yet $x_n\in B_{\frac1n}$ for every $n$, so that
$x_n\goesun 0$; a contradiction.
\end{proof}
\section{Un-convergence in a sublattice}
Recall that if $(y_\alpha)$ is a net in a regular sublattice $Y$ of a
vector lattice $X$ then $y_\alpha\goesuo 0$ in $Y$ iff
$y_\alpha\goesuo 0$ in $X$. The situation is very different for
un-convergence. Let $Y$ be a sublattice of a normed lattice
$X$ and $(y_\alpha)$ a net in $Y$. If $y_\alpha\goesun 0$ in $X$
then, clearly, $y_\alpha\goesun 0$ in $Y$. However, the following
examples show that the converse fails even for closed ideals or bands.
\begin{example}\label{ex:c0}
The sequence of the standard unit vectors $(e_n)$ is un-null in
$c_0$ but not in $\ell_\infty$, even though $c_0$ is a closed ideal in
$\ell_\infty$.
\end{example}
\begin{example}
Let $X=C[-1,1]$ and $Y$ be the set of all $f\in X$ which vanish on
$[-1,0]$. It is easy to see that $Y$ is a band (though it is not a
projection band). Let $(f_n)$ be a sequence in $Y_+$ such
that $\norm{f_n}=1$ and $\supp
f_n\subseteq[\frac{1}{n+1},\frac1n]$.
Since $X$ has a strong unit, the un-topology on $X$ agrees with the
norm topology, hence $(f_n)$ is not un-null in $X$. However,
it is easy to see that $(f_n)$ is un-null in $Y$.
\end{example}
Nevertheless, there are some good news. Recall that a sublattice $Y$
of a vector lattice $X$ is \term{majorizing} if for every $x\in X_+$
there exists $y\in Y_+$ with $x\le y$.
\begin{theorem}\label{sublat}
Let $Y$ be a sublattice of a normed lattice $X$ and $(y_\alpha)$ a
net in $Y$ such that $y_\alpha\goesun 0$ in $Y$. Each of the
following conditions implies that $y_\alpha\goesun 0$ in $X$.
\begin{enumerate}
\item\label{sublat-maj} $Y$ is majorizing in $X$;
\item\label{sublat-dense} $Y$ is norm dense in $X$;
\item\label{sublat-projb} $Y$ is a projection band in $X$.
\end{enumerate}
\end{theorem}
\begin{proof}
Without loss of generality, $y_\alpha\ge 0$ for every $\alpha$.
\eqref{sublat-maj} is straightforward. To
prove~\eqref{sublat-dense}, take $u\in X_+$ and fix
$\varepsilon>0$. Find $v\in Y_+$ with $\norm{u-v}<\varepsilon$. By
assumption, $y_\alpha\wedge v\goesnorm 0$. We can find $\alpha_0$
such that $\norm{y_\alpha\wedge v}<\varepsilon$
whenever $\alpha\ge\alpha_0$. It follows from $u\le v+\abs{u-v}$
that
\begin{math}
y_\alpha\wedge u\le y_\alpha\wedge v+\abs{u-v},
\end{math}
so that
\begin{displaymath}
\norm{y_\alpha\wedge u}\le\norm{y_\alpha\wedge v}+\norm{u-v}
<2\varepsilon.
\end{displaymath}
It follows that $y_\alpha\wedge u\goesnorm 0$. Hence
$y_\alpha\goesun 0$ in $X$.
To prove~\eqref{sublat-projb}, let $u\in X_+$. Then $u=v+w$ for some
positive $v\in Y$ and $w\in Y^d$. It follows from $y_\alpha\perp w$ that
\begin{math}
y_\alpha\wedge u=y_\alpha\wedge v\goesnorm 0.
\end{math}
\end{proof}
Recall that every (Archimedean) vector lattice $X$ is majorizing in
its \term{order (or Dedekind) completion} $X^\delta$; see , e.g., \cite[p.~101]{Aliprantis:06}.
\begin{corollary}
If $X$ is a normed lattice and $x_\alpha\goesun x$ in $X$ then
$x_\alpha\goesun x$ in the order completion $X^\delta$ of $X$.
\end{corollary}
\begin{corollary}
If $X$ is a KB-space and $x_\alpha\goesun 0$ in $X$ then
$x_\alpha\goesun 0$ in $X^{**}$.
\end{corollary}
\begin{proof}
By \cite[Theorem~4.60]{Aliprantis:06}, $X$ is a projection band in
$X^{**}$. The conclusion now follows from
\Cref{sublat}\eqref{sublat-projb}.
\end{proof}
\Cref{ex:c0} shows that the assumption that $X$ is a KB-space cannot
be removed.
\begin{corollary}
Let $Y$ be a sublattice of an order continuous Banach lattice
$X$. If $y_\alpha\goesun 0$ in $Y$ then $y_\alpha\goesun 0$ in $X$.
\end{corollary}
\begin{proof}
Suppose that $y_\alpha\goesun 0$ in $Y$. By
\Cref{sublat}\eqref{sublat-maj}, $y_\alpha\goesun 0$ in the ideal
$I(Y)$ generated by $Y$ in $X$. By
\Cref{sublat}\eqref{sublat-dense}, $y_\alpha\goesun 0$ in the
closure $\overline{I(Y)}$ of the ideal. Since $X$ is order
continuous, $\overline{I(Y)}$ is a projection band in $X$.
It now follows from \Cref{sublat}\eqref{sublat-projb} that
$y_\alpha\goesun 0$ in $X$.
\end{proof}
\begin{question}
Let $B$ be a band in $X$. Suppose that every net in $B$ which is
un-null in $B$ is also un-null in $X$. Does this imply that $B$ is a
projection band?
\end{question}
\bigskip
\begin{proposition}\label{band-closed}
Every band in a normed lattice is un-closed.
\end{proposition}
\begin{proof}
Let $B$ be a band and $(x_\alpha)$ a net in $B$ such that
$x_\alpha\goesun x$. Fix $z\in B^d$. Then $\abs{x_\alpha}\wedge z=0$
for every $\alpha$. Since lattice operations are un-continuous, we
have $\abs{x}\wedge z=0$. It follows that $x\in B^{dd}=B$.
\end{proof}
\begin{remark}\label{pb}
Let $B$ be a projection band a normed lattice $X$. We write
$P_B$ for the corresponding band projection. It follows easily from
$0\le P_B\le I$ that if $x_\alpha\goesun x$ in $X$ then
$P_Bx_\alpha\goesun P_Bx$ both in $X$ and in $B$.
\end{remark}
\subsection*{Dense band decompositions.}
Let $X$ be a Banach lattice. By a \term{dense band
decomposition} of $X$ we mean a family $\mathcal B$ of pairwise
disjoint projection bands in $X$ such that the linear span of all of
the bands in $\mathcal B$ is norm dense in $X$.
\begin{lemma}
Let $\mathcal B$ be a family of pairwise disjoint projection bands
in a Banach lattice $X$. $\mathcal B$ is a dense band decomposition
of $X$ iff for every $x\in X$ and every $\varepsilon>0$ there exist
$B_1,\dots,B_n$ in $\mathcal B$ such that
$\bignorm{x-\sum_{i=1}^nP_{B_i}x}<\varepsilon$.
\end{lemma}
\begin{proof}
Suppose that $\mathcal B$ is a dense band decomposition
of $X$. Let $x\in X$ and $\varepsilon>0$. By assumption, we can find
distinct bands $B_1,\dots,B_n$ and vectors $x_1\in B_1,\dots,x_n\in
B_n$ such that
\begin{math}
\bignorm{x-\sum_{i=1}^nx_i}<\varepsilon.
\end{math}
Put $Q=I-\sum_{i=1}^nP_{B_i}$. Then $Q$ is also a band projection,
hence it is a lattice homomorphism and $0\le Q\le I$. Note also that
$Qx_i=0$ for $i=1,\dots,n$. We have
\begin{displaymath}
\bigabs{x-\sum_{i=1}^nx_i}
\ge Q\bigabs{x-\sum_{i=1}^nx_i}
=\bigabs{Qx-\sum_{i=1}^nQx_i}
=\bigabs{x-\sum_{i=1}^nP_{B_i}x}.
\end{displaymath}
It follows that $\bignorm{x-\sum_{i=1}^nP_{B_i}x}<\varepsilon$.
The converse implication is trivial.
\end{proof}
Our definition of a disjoint band decomposition is partially motivated
by following fact.
\begin{theorem}(\cite[Proposition~1.a.9]{Lindenstrauss:79})\label{wunit-decomp}
Every order continuous Banach lattice admits a dense band
decomposition $\mathcal B$ such that each band in $\mathcal B$ has a
weak unit.
\end{theorem}
It is easy to see that if $X$ is an order continuous Banach lattice
and $\mathcal B$ is a pairwise disjoint collection of bands such that
$x=\sup\{P_{B}x\mid B\in\mathcal B\}$ for every $x\in X_+$ then
$\mathcal B$ is a dense band decomposition.
\begin{theorem}\label{band-decomp}
Suppose that $\mathcal B$ is a dense band decomposition of a Banach
lattice $X$. Then $x_\alpha\goesun x$ in $X$ iff $P_Bx_\alpha\goesun
P_Bx$ in $B$ for each $B\in\mathcal B$.
\end{theorem}
\begin{proof}
Without loss of generality, $x=0$ and $x_\alpha\ge 0$ for every
$\alpha$. The forward implication follows immediately from
\Cref{pb}. To prove the converse, suppose that
$P_Bx_\alpha\goesun 0$ in $B$ for each $B\in\mathcal B$. Let
$u\in X_+$; it suffices to show that $x_\alpha\wedge u\goesnorm 0$.
Fix $\varepsilon>0$. Find $B_1,\dots,B_n\in\mathcal B$
such that
\begin{math}
\bignorm{u-\sum_{i=1}^nP_{B_i}u}<\varepsilon.
\end{math}
Since $P_{B_i}x_\alpha\goesun 0$ in $B_i$ as $i=1,\dots,n$, we can
find $\alpha_0$ such that
\begin{math}
\bignorm{P_{B_i}x_\alpha\wedge P_{B_i}u}<\frac{\varepsilon}{n}
\end{math}
for every $\alpha\ge\alpha_0$ and every $i=1,\dots,n$.
It follows from $x_\alpha\wedge P_{B_i}u\in B_i$ that
$x_\alpha\wedge P_{B_i}u=P_{B_i}x_\alpha\wedge P_{B_i}u$.
Therefore,
\begin{multline*}
\norm{x_\alpha\wedge u}\le
\Bignorm{x_\alpha\wedge\sum_{i=1}^nP_{B_i}u}+
\Bignorm{u-\sum_{i=1}^nP_{B_i}u}
\le\Bignorm{\sum_{i=1}^nx_\alpha\wedge P_{B_i}u}+\varepsilon\\
=\Bignorm{\sum_{i=1}^nP_{B_i}x_\alpha\wedge P_{B_i}u}+\varepsilon
\le n\cdot\frac{\varepsilon}{n}+\varepsilon
\le 2\varepsilon.
\end{multline*}
\end{proof}
\begin{remark}\label{atoms}
Recall that a positive non-zero vector $a$ in a vector lattice $X$
is an \term{atom} if the principal ideal $I_a$ generated by $a$
coincides with $\Span a$. In this case, $I_a$ is a projection band,
and the corresponding band projection $P_a$ has form $f_a\otimes a$
for some positive functional $f_a$, that is, $P_ax=f_a(x)a$. We say
that $X$ is \term{non-atomic} if it has no atoms. We say that $X$ is
\term{atomic} if $X$ is the band generated by all the atoms. In the
latter case,
\begin{math}
x=\sup\{f_a(x)a\mid a\text{ is an atom}\}
\end{math}
for every $x\in X_+$.
See, e.g., \cite[p.~143]{Schaefer:74}.
\end{remark}
It follows that if $X$ is an order continuous atomic Banach lattice, the
family $\{I_a\mid a\text{ is an atom}\}$ is a dense band decomposition
of $X$. Applying \Cref{band-decomp}, we conclude that in such spaces
un-convergence is exactly the ``coordinate-wise'' convergence:
\begin{corollary}\label{un-atomic}
Let $X$ be an atomic order continuous Banach lattice. Then
$x_\alpha\goesun x$ iff $f_a(x_\alpha)\to f_a(x)$ for every atom $a$.
\end{corollary}
\begin{remark}
The order continuity assumption cannot be removed. Indeed,
$\ell_\infty$ is atomic, the sequence $(e_n)$ converges to zero
coordinate-wise, yet it is not un-null.
\end{remark}
The following results extends \cite[Proposition 6.2]{DOT}.
\begin{proposition}\label{w-un}
The following are equivalent:
\begin{enumerate}
\item\label{w-un-net} $x_\alpha\goesw 0$ implies $x_\alpha\goesun 0$ for every
net $(x_\alpha)$ in $X$;
\item\label{w-un-seq} $x_n\goesw 0$ implies $x_n\goesun 0$ for every
sequence $(x_n)$ in $X$;
\item\label{w-un-aoc} $X$ is atomic and order continuous.
\end{enumerate}
\end{proposition}
\begin{proof}
\eqref{w-un-net}$\Rightarrow$\eqref{w-un-seq} is
trivial. The implication
\eqref{w-un-seq}$\Rightarrow$\eqref{w-un-aoc} is a part of
\cite[Proposition 6.2]{DOT}. The implication
\eqref{w-un-aoc}$\Rightarrow$\eqref{w-un-net} follows from \Cref{un-atomic}.
\end{proof}
\section{AL-representations and local convexity}
In this section, we will show that un-topology on an order continuous
Banach lattice $X$ is locally convex iff $X$ is atomic. Our main tool
is the relationship between un-convergence in $X$ and in an
AL-representation of $X$.
It was observed in \cite[Example~23]{Troitsky:04} that for a net
$(x_\alpha)$ in $L_p(\mu)$ where $\mu$ is a finite measure and
$1\le p<\infty$, one has $x_\alpha\goesun 0$ iff $x_\alpha\goesmu 0$
(i.e., the net converges to zero in measure). Note that this does not
extend to $\sigma$-finite measures. Indeed, let
$X=L_p(\mathbb R)$ and let $x_n$ be the characteristic function of
$[n,n+1]$. Then $x_n\goesun 0$ but $(x_n)$ does not converge to zero
in measure. On the other hand, let $(x_\alpha)$ be a net in $L_p(\mu)$
where $\mu$ is a $\sigma$-finite measure, let $(\Omega_n)$ be a
countable partition of $\Omega$ into sets of finite measure; it
follows from \Cref{band-decomp} that $x_\alpha\goesun 0$ iff the
restriction of $x_\alpha$ to $\Omega_n$ converges to zero in measure
for every $n$.
Suppose that $X$ is an order continuous Banach lattice with a weak
unit $e$. By \cite[Theorem~1.b.14]{Lindenstrauss:79}, $X$ can be
represented as an ideal of $L_1(\mu)$ for some probability measure
$\mu$. More precisely, there is a lattice isomorphism from $X$ onto a
norm-dense ideal of $L_1(\mu)$; with a slight abuse of notation we
will view $X$ itself as an ideal of $L_1(\mu)$. Moreover, this
representation may be chosen so that $e$ corresponds to $\one$,
$L_\infty(\mu)$ is a norm-dense ideal in $X$, and both inclusions in
$L_\infty(\mu)\subseteq X\subseteq L_1(\mu)$ are continuous. We call
$L_1(\mu)$ an \term{AL-representation} for $X$ and $e$. Let $(x_n)$ be
a sequence in $X$. It was shown in \cite[Remark~4.6]{GTX} that
$x_n\goesuo 0$ in $X$ iff $x_n\goesae 0$ in $L_1(\mu)$. It was shown
in \cite[Theorem~4.6]{DOT} that $x_n\goesun 0$ in $X$ iff
$x_n\goesmu 0$ in $L_1(\mu)$. Since un-topology and the topology of
convergence in measure are both metrizable on $X$ because $X$ has a
weak unit, it follows that these two topologies coincide on $X$. In
particular, $x_\alpha\goesun 0$ in $X$ iff $x_\alpha\goesmu 0$ in
$L_1(\mu)$ for every net $(x_\alpha)$ in $X$. This may also be deduced
from Amemiya's Theorem (see, e.g., Theorem~2.4.8 in
\cite{Meyer-Nieberg:91}) as follows:
\begin{displaymath}
x_\alpha\goesun 0\text{ in }X\quad\Leftrightarrow\quad
\norm{x_\alpha\wedge e}_X\to 0
\quad\overset{\text{Amemiya}}{\Leftrightarrow}\quad
\norm{x_\alpha\wedge\one}_{L_1}\to 0\quad\Leftrightarrow\quad
x_\alpha\goesmu 0\text{ in }L_1(\mu)
\end{displaymath}
for every net $(x_\alpha)$ in $X_+$.
\begin{proposition}\label{un-nonat-conv}
Let $X$ be a non-atomic order continuous Banach lattice and
$W$ a neighborhood of zero for un-topology. If $W$ is convex then $W=X$.
\end{proposition}
\begin{proof}
Fix $e\in X_+$; we will show that $e\in W$. We know that
$V_{\varepsilon,u}\subseteq W$ for some $\varepsilon>0$ and
$u>0$. Consider the principal band $B_e$. Since $X$ is order
continuous, $B_e$ is a projection band in $X$; let $P_e$ be the
corresponding band projection. Furthermore, $B_e$ is a non-atomic
order continuous Banach lattice with a weak unit. Let
$L_1(\Omega,\mathcal F,\mu)$ be an AL-representation for $B_e$ with
$e=\one$. Note that the measure $\mu$ is non-atomic because if a
measurable set $A$ were an atom for $\mu$ then its characteristic
function $\chi_A$ would be an atom in $X$. Fix $n\in\mathbb N$.
Using the non-atomicity of $\mu$, we find a measurable partition
$A_{n,1},\dots,A_{n,n}$ of $\Omega$ with $\mu(A_{n,i})=\frac1n$ as
$i=1,\dots,n$; see, e.g., Exercise~2 in~\cite[p.~174]{Halmos:70}.
Since $L_\infty(\mu)\subseteq B_e\subseteq L_1(\mu)$, we may view
the characteristic functions $\chi_{A_{n,i}}$ as elements of
$B_e$. Consider the vectors $(n\chi_{A_{n,i}})\wedge u$ as
$i=1,\dots,n$; they belong to $B_e$, so that we may view them as
functions in $L_1(\mu)$. Let $g_n$ be the function in this list
whose norm in $X$ is maximal; if there are more than one, pick any
one. Repeating this construction for every $n\in\mathbb N$, we
produce a sequence $(g_n)$ in $[0,u]\cap B_e$. It follows that
$g_n\le P_eu$ for every $n$. Since $P_eu$ may be viewed as an
element of $L_1(\mu)$ and the measure of the support of $g_n$ tends
to zero, it follows that $\norm{g_n}_{L_1}\to 0$. Amemiya's Theorem
yields $\norm{g_n}_X\to 0$. Fix $n$ such that
$\norm{g_n}_X<\varepsilon$. It follows from the definition of $g_n$
that $\bignorm{(n\chi_{A_{n,i}})\wedge u}_X<\varepsilon$ as
$i=1,\dots,n$, so that $n\chi_{A_{n,i}}$ is in $V_{\varepsilon,u}$
and, therefore, in $W$. Since $W$ is convex and
\begin{displaymath}
e=\one=\frac1n\sum_{i=1}^nn\chi_{A_{n,i}},
\end{displaymath}
we have $e\in W$. Therefore, $X_+\subseteq W$. Furthermore,
it follows from $n\chi_{A_{n,i}}\in V_{\varepsilon,u}$ that
$-n\chi_{A_{n,i}}\in V_{\varepsilon,u}$ for all $i=1,\dots,n$ and,
therefore, $-e\in W$. This yields
$X_-\subseteq W$. Finally, for every $x\in X$ we have
$x=\frac12\bigl(2x^++2(-x^-)\bigr)$, so that $x\in W$.
\end{proof}
\begin{theorem}\label{loc-conv}
Let $X$ be an order continuous Banach lattice. Un-topology on $X$ is
locally convex iff $X$ is atomic.
\end{theorem}
\begin{proof}
Suppose that $X$ is atomic. By \Cref{un-atomic}, un-topology is
determined by the family of seminorms $x\mapsto
\bigabs{f_a(x)}$ where $a$ is an atom of $X$; hence the topology is
locally convex.
Suppose that un-topology is locally convex but $X$ is not atomic. It
follows that there is $e\in X_+$ such that $B_e$ is non-atomic.
By \Cref{sublat}, un-topology on $B_e$ agrees
with the relative topology induced on $B_e$ by un-topology on
$X$; in particular, it is locally convex. On the other hand,
\Cref{un-nonat-conv} asserts that this topology on $B_e$ has no
proper convex neighborhoods; a contradiction.
\end{proof}
\subsection*{Un-continuous functionals}
\Cref{loc-conv} allows us to describe un-continuous linear
functionals. For a functional $\varphi\in X^*$, we say that $\varphi$
is \term{un-continuous} if it is continuous with respect to the
un-topology on $X$ or, equivalently, if $x_\alpha\goesun 0$ implies
$\varphi(x_\alpha)\to 0$.
\begin{proposition}\label{un-cont-ideal}
The set of all un-continuous functionals in $X^*$ is an ideal.
\end{proposition}
\begin{proof}
It is straightforward to verify that this set is a linear
subspace. Suppose that $\varphi$ in $X^*$ is un-continuous; we will
show that $\abs{\varphi}$ is also un-continuous. Fix $\delta>0$. One
can find $\varepsilon>0$ and $u>0$ such that
$\bigabs{\varphi(x)}<\delta$ whenever $x\in V_{\varepsilon,u}$. Fix
$x\in V_{\varepsilon,u}$. Since $V_{\varepsilon,u}$ is solid,
$\abs{y}\le\abs{x}$ implies $y\in V_{\varepsilon,u}$ and, therefore,
$\bigabs{\varphi(y)}<\delta$. By the Riesz-Kantorovich formula, we
get
\begin{displaymath}
\bigabs{\abs{\varphi}(x)}\le\abs{\varphi}\bigl(\abs{x}\bigr)
=\sup\bigl\{\bigabs{\varphi(y)}\mid\abs{y}\le\abs{x}\bigr\}
\le\delta.
\end{displaymath}
It follows that $\abs{\varphi}$ is un-continuous. Hence, the set of
all un-continuous functionals in $X^*$ forms a sublattice. It is
easy to see that if $\varphi\in X^*_+$ is un-continuous and
$0\le\psi\le\varphi$ then $\psi$ is also un-continuous; this
completes the proof.
\end{proof}
Recall that if $a$ is an atom then $f_a$
stands for the corresponding ``coordinate functional''.
\begin{corollary}\label{un-dual}
Suppose that $X$ is an order continuous Banach lattice and
$\varphi\in X^*$ is un-continuous.
\begin{enumerate}
\item\label{dual-atomic} If $X$ is atomic then
$\varphi=\lambda_1f_{a_1}+\dots+\lambda_nf_{a_n}$, where
$\lambda_1,\dots,\lambda_n\in\mathbb R$ and $a_1,\dots,a_n$ are
atoms;
\item If $X$ is non-atomic then $\varphi=0$.
\end{enumerate}
\end{corollary}
\begin{proof}
By \Cref{un-cont-ideal}, we may assume that $\varphi\ge 0$;
otherwise we consider $\varphi^+$ and $\varphi^-$.
Suppose $X$ is atomic; let $A$ be a maximal disjoint family of
atoms. We claim that the set $F:=\{a\in A\mid\varphi(a)\ne 0\}$ is
finite. Indeed, otherwise, take a sequence $(a_n)$ of distinct atoms
in $F$ and put $x_n=\frac{1}{\varphi(a_n)}a_n$. Then $x_n\goesun 0$
by \Cref{un-atomic}, yet $\varphi(x_n)=1$; a contradiction. This
proves the claim.
Since $X$ is order continuous, it follows from Remark~\ref{atoms}
that $X$ has a disjoint band decomposition
$X=B_F\oplus B_{A\setminus F}$. Since $\varphi(a)=0$ for all
$a\in A\setminus F$, $\varphi$ vanishes on the ideal
$I_{A\setminus F}$ and, therefore, on $B_{A\setminus F}$ because
$\varphi$ is order continuous. On the other hand, since $F$ is
finite, $B_F=\Span F$ and, therefore, is finite-dimensional. It
follows that $\varphi$ is a linear combination of
$\{f_a\mid a\in F\}$.
Suppose now that $X$ is non-atomic. Let $W=\varphi^{-1}(-1,1)$. Then
$W$ is a convex neighborhood of zero for the un-topology. By
\Cref{un-nonat-conv}, $W=X$. This easily implies $\varphi=0$.
\end{proof}
Case~\eqref{dual-atomic} of the preceding corollary essentially says
that every un-continuous functional on an atomic order continuous
space has finite support.
\begin{example}
Let $X=\ell_2$. By \Cref{un-dual}, the set of all un-continuous
functionals in $X^*$ may be identified with $c_{00}$, the linear
subspace of all sequences with finite support. Clearly, it is
neither norm closed nor order closed; it is not even
$\sigma$-order closed in $X^*$.
\end{example}
\begin{example}
Let $X=C_0(\Omega)$ where $\Omega$ is a locally compact Hausdorff
topological space. It was observed in \cite[Example~20]{Troitsky:04}
that the un-topology in $X$ agrees with the topology of uniform
convergence on compact subsets of $\Omega$.
Let $\varphi\in X^*_+$. By the Riesz Representation Theorem, there
exists a regular Borel measure $\mu$ such that $\varphi(f)=\int
f\,d\mu$ for every $f\in X$; see, e.g.,
\cite[Theorem~III.5.7]{Conway:90}. An argument similar to the proof
of \cite[Proposition~IV.4.1]{Conway:90} shows that $\varphi$ is
un-continuous iff $\mu$ has compact support.
\end{example}
\section{Un-completeness}
Throughout this section, $X$ is assumed to be an order continuous
Banach lattice. Since un-topology is linear, one can talk about
un-Cauchy nets. That is, a net $(x_\alpha)$ is un-Cauchy if for every
un-neighborhood $U$ of zero there exists $\alpha_0$ such that
$x_\alpha-x_\beta\in U$ whenever
$\alpha,\beta\ge\alpha_0$.
We investigate whether $X$
itself or some ``nice'' subset of $X$ is un-complete. First, we
observe that the entire space is un-complete only when $X$ is
finite-dimensional.
\begin{lemma}\label{disj-sum}
Let $(x_n)$ be a positive disjoint sequence in an order continuous
Banach lattice $X$ such that $(x_n)$ is not norm null. Put
$s_n=\sum_{i=1}^nx_i$. Then $(s_n)$ is un-Cauchy but not un-convergent.
\end{lemma}
\begin{proof}
The sequence $(s_n)$ is monotone increasing and does not converge in
norm; hence it is not un-convergent by
\Cref{monot}\eqref{monot-un}. To show that $(s_n)$ is un-Cauchy, fix
any $\varepsilon>0$ and a non-zero $u\in X_+$. Since $x_i$'s are
disjoint, we have $s_n\wedge u=\sum_{i=1}^n(x_i\wedge u)$. The
sequence $(s_n\wedge u)$ is increasing and order bounded, hence is
norm Cauchy by Nakano's Theorem; see
\cite[Theorem~4.9]{Aliprantis:06}. We can find $n_0$ such that
\begin{math}
\bignorm{s_m\wedge u-s_n\wedge u}<\varepsilon
\end{math}
whenever $m\ge n\ge n_0$. Observe that
\begin{displaymath}
s_m\wedge u-s_n\wedge u=
\sum_{i=n+1}^m(x_i\wedge u)=(s_m-s_n)\wedge u
=\abs{s_m-s_n}\wedge u.
\end{displaymath}
It follows that $\bignorm{\abs{s_m-s_n}\wedge u}<\varepsilon$, so
that $s_m-s_n\in V_{\varepsilon,u}$.
\end{proof}
\begin{proposition}
Let $X$ be an order continuous Banach lattice. $X$ is un-complete
iff $X$ is finite-dimensional.
\end{proposition}
\begin{proof}
If $X$ is finite-dimensional then it has a strong unit, so that
un-topology agrees with norm topology and is, therefore,
un-complete. Suppose now that $\dim X=\infty$. Then $X$ contains a
disjoint normalized positive sequence. By \Cref{disj-sum}, $X$ is
not un-complete.
\end{proof}
\begin{example}
Let $X=L_p$ with $1<p<\infty$. Pick $0\le x\in L_1\setminus L_p$ and
put $x_n=x\wedge(n\one)$. It is easy to see that $(x_n)$ is
un-Cauchy in $L_p$, yet it does not un-converge in $L_p$.
\end{example}
Even when the entire space is not un-complete, the closed unit ball
$B_X$ may still be un-complete; that is, complete in the topology
induced by un-topology on $X$. Since $B_X$ is
un-closed, it is un-complete iff every norm bounded
un-Cauchy net in $X$ is un-convergent. The following theorem
should be compared with \cite[Theorem~4.7]{GaoX:14}, where a similar
statement was proved for uo-convergence.
\begin{theorem}\label{BX-complete}
Let $X$ be an order continuous Banach lattice. Then $B_X$ is
un-complete iff $X$ is a KB-space.
\end{theorem}
\begin{proof}
Suppose $X$ is not KB. Then $X$ contains a lattice copy of
$c_0$. Let $(x_n)$ be the sequence in $X$ corresponding to the unit
basis of $c_0$. Let $s_n=\sum_{i=1}^nx_i$. Clearly, $(s_n)$ is norm
bounded. However, by \Cref{disj-sum}, $(s_n)$ is un-Cauchy but not
un-convergent.
Suppose now that $X$ is a KB-space. First, we consider the case when
$X$ has a weak unit. In this case, un-topology on $X$ and,
therefore, on $B_X$, is metrizable by Theorem~\ref{metriz}. Hence,
it suffices to prove that $B_X$ is sequentially un-complete. Let
$(x_n)$ be a sequence in $B_X$ which is un-Cauchy in $X$. Let
$L_1(\mu)$ be an AL-representation for $X$. It follows that $(x_n)$
is Cauchy with respect to convergence in measure in $L_1(\mu)$. By
\cite[Theorem~2.30]{Folland:99}, there is a subsequence $(x_{n_k})$
which converges a.e. It follows that $(x_{n_k})$ is uo-Cauchy in
$X$ by \cite[Remark~4.6]{GTX}. Then \cite[Theorem~4.7]{GaoX:14}
yields that $x_{n_k}\goesuo x$ for some $x\in X$. It follows that
$x_{n_k}\goesun x$.
Since $(x_n)$ is un-Cauchy, this yields that
$x_n\goesun x$.
Now consider the general case. Let $X$ be a KB-space and
$(x_\alpha)$ a net in $B_X$ such that $(x_\alpha)$ is un-Cauchy in
$X$; we need to prove that the net is un-convergent. We may assume
without loss of generality that $x_\alpha\ge 0$ for every $\alpha$;
otherwise, consider $(x_\alpha^+)$ and $(x_\alpha^-)$, which are
also un-Cauchy because
$\abs{x_\alpha^+-x_\beta^+}\le\abs{x_\alpha-x_\beta}$ and
$\abs{x_\alpha^--x_\beta^-}\le\abs{x_\alpha-x_\beta}$.
By \Cref{wunit-decomp}, there exists a dense band decomposition
$\mathcal B$ of $X$ such that each $B$ in $\mathcal B$ has a
weak unit. Put
\begin{displaymath}
\mathcal C=\bigl\{B_1\oplus\dots\oplus B_n\mid
B_1,\dots,B_n\in\mathcal B\bigr\}.
\end{displaymath}
Note that $\mathcal C$ is a family of bands with weak
units. Furthermore, $\mathcal C$ is a directed set when ordered by
inclusion, so the family of band projections
$(P_C)_{C\in\mathcal C}$ may be viewed as a net.
For every $C\in\mathcal C$, the net $(P_Cx_\alpha)$ is un-Cauchy by
\Cref{pb}. Since $C$ has a weak unit, the first part of the proof
yields that $(P_Cx_\alpha)$ un-converges to some positive vector $x_C$ in
$C$. This produces a net $(x_C)_{C\in\mathcal C}$. It is easy to
verify that $x_C=x_{B_1}+\dots+x_{B_n}$ whenever
$C=B_1\oplus\dots\oplus B_n$ for some $B_1,\dots,B_n\in\mathcal B$.
It follows that the net $(x_C)_{C\in\mathcal C}$ is increasing. On
the other hand,
\begin{math}
\norm{x_C}\le\liminf_\alpha\norm{P_Cx_\alpha}\le 1,
\end{math}
so that this net is norm bounded. Since $X$ is a KB-space, the net
$(x_C)_{C\in\mathcal C}$ converges in norm to some $x\in X$.
Fix $B\in\mathcal B$. On one hand, norm continuity of $P_B$ yields
$\lim_{C\in\mathcal C}P_Bx_C=P_Bx$. On the other hand, for every
$C\in\mathcal C$ with $B\subseteq C$ we have $P_Bx_C=x_B$, so that
$\lim_{C\in\mathcal C}P_Bx_C=x_B$. It follows that $P_Bx=x_B$, so
that $P_Bx_\alpha\goesun P_Bx$ for every $B\in\mathcal B$. Now
\Cref{band-decomp} yields $x_\alpha\goesun x$.
\end{proof}
The assumption that $X$ is order continuous cannot be removed: for
example, $\ell_\infty$ is not a KB-space, yet its closed unit ball is
un-complete (because the un and the norm topologies on $\ell_\infty$
agree).
\begin{example}
The following examples show that in general $B_X$ in
\Cref{BX-complete} cannot be replaced with an arbitrary convex closed
bounded set. Let $X=\ell_1$; let $C$ be the set of all vectors in $B_X$ whose
coordinates sum up to zero. Clearly, $C$ is convex, closed, and
bounded. Let $x_n=\frac12(e_1-e_n)$. Then $(x_n)$ is a sequence in
$C$ which un-converges to $\frac12e_1$ which is not in $C$. Thus,
$C$ is not un-closed in $X$; in particular, $C$ is not un-complete.
It is easy to construct a similar example in $X=L_1$; take
\begin{math}
C=\bigl\{x\in B_X\mid \int x=0\bigr\}
\end{math}
and put
$x_n=\chi_{[0,\frac12]}-\frac{n}{2}\chi_{[\frac12,\frac12+\frac{1}{n}]}$,
$n\ge 2$.
\end{example}
\begin{proposition}\label{sets-un-closed}
Suppose that $X^*$ is order continuous and $C$ is a norm closed
convex norm bounded subset of $X$. Then $C$ is un-closed.
\end{proposition}
\begin{proof}
Suppose that $x_\alpha\goesun x$ for a net $(x_\alpha)$ in $C$ and a
vector $x$ in $X$. Since $(x_\alpha)$ is norm bounded and $X^*$ is
order continuous, \cite[Theorem~6.4]{DOT} guarantees that
$(x_\alpha)$ converges to $x$ weakly. Since $C$ is convex and
closed, it is weakly closed, hence $x\in C$.
\end{proof}
\begin{corollary}
Let $X$ be a reflexive Banach lattice and $C$ a closed convex norm
bounded subset of $X$. Then $C$ is un-complete.
\end{corollary}
\begin{proof}
Since $X$ is reflexive, $X$ is a KB-space and $X^*$ is order
continuous. Let $(x_\alpha)$ be a un-Cauchy net in
$C$. \Cref{BX-complete} yields that
$x_\alpha\goesun x$ for some $x\in X$, while \Cref{sets-un-closed}
implies that $x\in C$.
\end{proof}
\section{Un-compact sets}
The main result of this section is \Cref{BX-compact}, which asserts
that $B_X$ is (sequentially) un-compact iff $X$ is an atomic
KB-space. We start with some auxiliary results. The following theorem
shows that, under certain assumptions, un-compactness is a ``local''
property.
\begin{theorem}\label{comp-local}
Let $X$ be a KB-space, $\mathcal B$ a dense band decomposition of
$X$, and $A$ a un-closed norm bounded subset of $X$. Then $A$ is
un-compact iff $P_B(A)$ is un-compact in $B$ for every $B\in\mathcal B$.
\end{theorem}
\begin{proof}
If $A$ is un-compact then $P_B(A)$ is un-compact in $B$ for every
$B\in\mathcal B$ because $P_B$ is un-continuous by \Cref{pb}. To
prove the converse, suppose that $P_B(A)$ is un-compact in $B$ for
every $B\in\mathcal B$. Let $H=\prod_{B\in\mathcal B}B$, the formal
product of all the bands in $\mathcal B$. That is, $H$ consists of
families $(x_B)_{B\in\mathcal B}$ indexed by $\mathcal B$, where
$x_B\in B$. We equip $H$ with the topology of coordinate-wise
un-convergence; this is the product of un-topologies on the bands
that make up $H$. This makes $H$ a topological vector space. Define
$\Phi\colon X\to H$ via $\Phi(x)=(P_Bx)_{B\in\mathcal B}$. Clearly,
$\Phi$ is linear. Since $\mathcal B$ is a dense band decomposition,
$\Phi$ is one-to-one. By \Cref{band-decomp}, $\Phi$ is a
homeomorphism from $X$ equipped with un-topology onto its range in
$H$.
Let $K$ be the subset of $H$ defined by $K=\prod_{B\in\mathcal
B}P_B(A)$. By Tikhonov's Theorem, $K$ is compact in $H$. It is
easy to see that $\Phi(A)\subseteq K$.
We claim that $\Phi(A)$ is closed in $H$. Indeed, suppose that
$\Phi(x_\alpha)\to h$ in $H$ for some net $(x_\alpha)$ in $A$. In
particular, the net $\bigl(\Phi(x_\alpha)\bigr)$ is Cauchy in
$H$. Since $\Phi$ is a homeomorphism, the net $(x_\alpha)$ is
un-Cauchy in $A$. Since $(x_\alpha)$ is bounded and $X$ is a
KB-space, $(x_\alpha)$ un-converges to some $x\in X$ by
\Cref{BX-complete}. Since $A$ is un-closed, we have $x\in A$. It
follows that $h=\Phi(x)$, so that $h\in\Phi(A)$.
Being a closed subset of a compact set, $\Phi(A)$ is itself
compact. Since $\Phi$ is a homeomorphism, we conclude that $A$ is
un-compact.
\end{proof}
Next, we discuss relationships between the sequential and the general
variants of un-closedness and un-compactness. Recall that for a set
$A$ in a topological space, we write $\overline{A}$ for the
closure of $A$; we write $\overline{A}^\sigma$ for the
\term{sequential closure} of $A$, i.e., $a\in\overline{A}^\sigma$ iff
$a$ is the limit of a sequence in $A$. We say that $A$ is
\term{sequentially closed} if $\overline{A}^\sigma=A$. It is well
known that for a metrizable topology, we always have
$\overline{A}^\sigma=\overline{A}$.
For a set $A$ in a Banach lattice, we write $\overline{A}^{\rm un}$
and $\overline{A}^{\sigma\text{-un}}$ for the un-closure and the
sequential un-closure of $A$, respectively. Obviously,
$\overline{A}^{\sigma\text{-un}}\subseteq\overline{A}^{\rm
un}$.
\begin{example}
In general, $\overline{A}^{\rm
un}\ne\overline{A}^{\sigma\text{-}\rm{un}}$. Indeed, in the
notation of \Cref{emelyanov}, let
$A=\{e_{\omega}\mid\omega\in\Omega\}$. It follows from
\Cref{emelyanov} that zero is in $\overline{A}^{\rm
un}$ but not in $\overline{A}^{\sigma\text{-un}}$.
\end{example}
\begin{proposition}\label{closures}
Let $A$ be a subset of a Banach lattice $X$. If $X$ has a
quasi-interior point \emph{or} $X$ is order continuous then
$\overline{A}^{\rm un}=\overline{A}^{\sigma\text{-}\rm{un}}$.
\end{proposition}
\begin{proof}
If $X$ has a quasi-interior point then its un-topology is metrizable
by \Cref{metriz}, hence $\overline{A}^{\rm
un}=\overline{A}^{\sigma\text{-}\rm{un}}$.
Suppose that $X$ is order continuous. Suppose that
$x\in\overline{A}^{\rm un}$; we need to show that
$x\in\overline{A}^{\sigma\text{-un}}$. Without loss of generality,
$x=0$. This means that $A$ contains a un-null net $(x_\alpha)$. By
\Cref{KP}, there exists an increasing sequence of indices
$(\alpha_k)$ and a disjoint sequence $(d_k)$ such that
$x_{\alpha_k}-d_k\goesnorm 0$. It follows that
$x_{\alpha_k}-d_k\goesun 0$. Since $(d_k)$ is disjoint, it is
uo-null and, since $X$ is order continuous, un-null. It follows that
$x_{\alpha_k}\goesun 0$ and, therefore,
$0\in\overline{A}^{\sigma\text{-un}}$.
\end{proof}
Recall that a topological space is said to be \term{sequentially
compact} if every sequence has a convergent subsequence. In a
Hausdorff topological vector space which is metrizable (or,
equivalently, first countable), sequential compactness is equivalent
to compactness, see, e.g., \cite[Theorem~7.21]{Royden:88}. We do not
know whether un-compactness and sequential un-compactness are
equivalent in general, yet we have the following partial result.
\begin{proposition}\label{scomp}
Let $A$ be a subset of a Banach lattice $X$.
\begin{enumerate}
\item\label{scomp-qip} If $X$ has a quasi-interior point, then $A$
is sequentially un-compact iff $A$ is un-compact.
\item\label{scomp-oc} Suppose that $X$ is order continuous. If $A$ is
un-compact then $A$ is sequentially un-compact.
\item\label{scomp-KB} Suppose that $X$ is a KB-space. If $A$ is
norm bounded and sequentially un-compact then $A$ is un-compact.
\end{enumerate}
\end{proposition}
\begin{proof}
\eqref{scomp-qip} follows immediately from \Cref{metriz}.
\eqref{scomp-oc} Let $(x_n)$ be a sequence in $A$. Find $e\in X_+$
such that $(x_n)$ is contained in $B_e$ (e.g., take
$e=\sum_{n=1}^\infty\frac{x_n}{2^n\norm{x_n}+1}$). Since $B_e$ is
un-closed, the set $A\cap B_e$ is un-compact in $B_e$. Since $e$ is
a quasi-interior point for $B_e$, the un-topology on $B_e$ is
metrizable, hence $A\cap B_e$ is sequentially un-compact. It follows
that there is a subsequence $(x_{n_k})$ which un-converges in
$B_e$ to some $x\in A\cap B_e$. By
\Cref{sublat}\eqref{sublat-projb}, $x_{n_k}\goesun x$ in $X$.
\eqref{scomp-KB} Clearly, $A$ is sequentially un-closed and,
therefore, un-closed by \Cref{closures}. Let $\mathcal B$ be as in
\Cref{wunit-decomp}. For each $B\in\mathcal B$, the band projection
$P_B$ is un-continuous by \Cref{pb}, so that $P_B(A)$ is
sequentially un-compact in $B$. Since $B$ has a weak unit, the
un-topology on $B$ is metrizable, so that $P_B(A)$ is un-compact in
$B$. The conclusion now follows from \Cref{comp-local}.
\end{proof}
\begin{theorem}\label{BX-compact}
For a Banach lattice $X$, TFAE:
\begin{enumerate}
\item\label{BX-compact-un} $B_X$ is un-compact;
\item\label{BX-compact-seq} $B_X$ is sequentially un-compact;
\item\label{BX-compact-KB} $X$ is an atomic KB-space.
\end{enumerate}
\end{theorem}
\begin{proof}
First, observe that both \eqref{BX-compact-un} and
\eqref{BX-compact-seq} imply that $X$ is order continuous and
atomic. Indeed, since order intervals are bounded and un-closed,
they are (sequentially) un-compact. But on order intervals, the
un-topology agrees with the norm topology, hence order intervals are
norm compact. This implies that $X$ is atomic and order
continuous; see, e.g., \cite[Theorem~6.1]{Wnuk:99}.
Suppose \eqref{BX-compact-un}. Since $X$ is order continuous,
\Cref{scomp}\eqref{scomp-oc} yields~\eqref{BX-compact-seq}.
Suppose \eqref{BX-compact-seq}. We already know that $X$ is
atomic. To show that $X$ is a KB-space, let $(x_n)$ be an increasing
norm bounded sequence in $X_+$. By assumption, it has a
un-convergent subsequence $(x_{n_k})$. By
\Cref{monot}\eqref{monot-un}, $(x_{n_k})$ converges in norm, hence
$(x_n)$ converges in norm. This yields~\eqref{BX-compact-KB}.
Suppose \eqref{BX-compact-KB}. Let $A$ be a maximal disjoint family
of atoms in $X$. Then $\bigl\{B_a\mid a\in A\bigr\}$ is a dense band
decomposition of $X$. For every $a\in A$, $P_a(B_X)$ is a closed
bounded subset of the one-dimensional band $B_a$, hence $P_a(B_X)$
is norm and un-compact in $B_a$. \Cref{comp-local} now implies that
$B_X$ is un-compact, which yields~\eqref{BX-compact-un}.
\end{proof}
\begin{example}
Let $X=c_0$ and $x_n=e_1+\dots+e_n$. Then $(x_n)$ is a sequence in
$B_X$ with no un-convergent subsequences.
\end{example}
\begin{proposition}\label{rcomp}
Let $A$ be a subset of an order continuous Banach lattice $X$. If
$A$ is relatively un-compact then $A$ is relatively sequentially
un-compact.
\end{proposition}
\begin{proof}
Let $(x_n)$ be a sequence in $A$. Find $e\in X_+$ such that $(x_n)$
is contained in $B_e$. Since $\overline{A}^{\rm un}$ is un-compact,
the set $\overline{A}^{\rm un}\cap B_e$ is un-compact in $B_e$ and,
therefore, sequentially un-compact in $B_e$ because the un-topology
on $B_e$ is metrizable. Hence, there is a subsequence $(x_{n_k})$
which un-converges in $B_e$ and, therefore, in $X$.
\end{proof}
\section{Un-convergence and weak*-convergence}
\subsection*{When does un-convergence imply weak*-convergence?}
It is easy to see that, in general, un-convergence does not imply
weak*-convergence. Indeed, let $X$ be an infinite-dimensional Banach
lattice with order continuous dual. Pick any unbounded disjoint
sequence $(f_n)$ in $X^*$. Being unbounded, $(f_n)$ cannot be
weak*-null. Yet it is un-null by \Cref{disj}. However, if we restrict
ourselves to norm bounded nets, the situation is more interesting. The
following result is analogous to \cite[Theorem~2.1]{Gao:14}. Recall
that for a net $(f_\alpha)$ in $X^*$, we write
$f_{\alpha}\xrightarrow{\abs{\sigma}(X^*,X)}0$ if
$\abs{f_\alpha}(x)\to 0$ for every $x\in X_+$.
\begin{theorem}\label{un-ws}
Let $X$ be a Banach lattice such that $X^*$ is order continuous. The
following are equivalent:
\begin{enumerate}
\item\label{un-ws-oc} $X$ is order continuous;
\item\label{un-ws-net} for any norm bounded net $(f_{\alpha})$ in $X^*$,
if $f_{\alpha}\goesun 0$, then $f_{\alpha}\goesws 0$;
\item\label{un-ws-net-abs} for any norm bounded net $(f_{\alpha})$ in $X^*$,
if $f_{\alpha}\goesun 0$, then
$f_{\alpha}\xrightarrow{\abs{\sigma}(X^*,X)}0$;
\item\label{un-ws-seq} for any norm bounded sequence $(f_n)$ in $X^*$,
if $f_n\goesun 0$, then $f_n\goesws 0$;
\item\label{un-ws-seq-abs} for any norm bounded sequence $(f_n)$ in $X^*$,
if $f_n\goesun 0$, then
$f_n\xrightarrow{\abs{\sigma}(X^*,X)}0$.
\end{enumerate}
\end{theorem}
The proof is similar to that of \cite[Theorem~2.1]{Gao:14} except that
in the proof of \eqref{un-ws-seq}$\Rightarrow$\eqref{un-ws-oc} we use
\Cref{disj}. Note that without the assumption that $X^*$ is
order continuous, we still get the following implications:
\begin{displaymath}
\eqref{un-ws-oc}\Rightarrow
\bigl[\eqref{un-ws-net}\Leftrightarrow\eqref{un-ws-net-abs}\bigr]
\Rightarrow
\bigl[\eqref{un-ws-seq}\Leftrightarrow\eqref{un-ws-seq-abs}\bigr].
\end{displaymath}
\subsection*{When does weak*-convergence imply un-convergence?}
Recall that for norm bounded nets, weak*-convergence
implies uo-convergence in $X^*$ iff $X$ is atomic and order continuous
by \cite[Theorem~3.4]{Gao:14}. Furthermore, \Cref{w-un} immediately
yields the following.
\begin{corollary}\label{ws-un-gen}
If $f_n\goesws 0$ implies $f_n\goesun 0$ for every sequence in $X^*$
then $X^*$ is atomic and order continuous.
\end{corollary}
The following example shows that the converse is false in general.
\begin{example}
Let $X=c$, the space of all convergent sequences. By \cite[Theorem
16.14]{Aliprantis:06a}, $X^*$ may be identified with
$\ell_1\oplus\mathbb R$ with the duality given by
\begin{displaymath}
\bigl\langle(f,r),x\bigr\rangle=r\cdot\lim_nx_n+\sum_{n=1}^\infty f_nx_n,
\end{displaymath}
where $x\in c$, $f\in\ell_1$, and $r\in\mathbb R$. It is easy to see
that $X^*$ is atomic and order continuous. Consider the sequence
$\bigl((e_n,0)\bigr)$ in $X^*$, where $e_n$ is the $n$-th standard
unit vector in $\ell_1$. It is easy to see that
$(e_n,0)\goesws(0,1)$ in $X^*$. On the other hand, this sequence is
disjoint and, therefore, un-null. Take $f_n=(e_n,-1)$; it follows
that $(f_n)$ is weak*-null but not un-null. Note that in this
example, $X^*$ is order continuous while $X$ is not.
\end{example}
Nevertheless, we will show that the converse implication is true under
the additional assumption that $X$ is order continuous.
\begin{theorem}\label{ws-un}
The following are equivalent:
\begin{enumerate}
\item\label{ws-un-net} For every net $(f_\alpha)$ in $X^*$, if
$f_\alpha\goesws 0$ then $f_\alpha\goesun 0$;
\item\label{ws-un-X} $X^*$ is atomic and both $X$ and $X^*$ are
order continuous.
\end{enumerate}
\end{theorem}
\begin{proof}
\eqref{ws-un-net}$\Rightarrow$\eqref{ws-un-X} By \Cref{ws-un-gen},
$X^*$ is atomic and order continuous. Suppose $X$ is
not order continuous. By \cite[Corollary 2.4.3]{Meyer-Nieberg:91}
there exists a disjoint norm-bounded sequence $(f_n)$ in $X^*$ which
is not weak*-null. One can then find a subsequence $(f_{n_k})$, a
vector $x_0\in X$ and a positive real $\varepsilon$ so that
$\bigabs{f_{n_k}(x_0)}>\varepsilon$ for every $k$. By
the Alaoglu-Bourbaki Theorem, there is a subnet $(g_\alpha)$ of
$(f_{n_k})$ such that $g_\alpha\goesws g$ for some $g\in X^*$. Since
$(f_{n_k})$ is disjoint and $X^*$ is order continuous, we have
$f_{n_k}\goesun 0$ and, therefore, $g_\alpha\goesun 0$. By
assumption, this yields $g=0$, so that $g_\alpha\goesws 0$. This
contradicts $\abs{g_\alpha(x_0)}>\varepsilon$ for every $\alpha$.
\eqref{ws-un-X}$\Rightarrow$\eqref{ws-un-net}
Let $f_\alpha\goesws 0$ in $X$. Let $A$ be a maximal disjoint
collection of atoms in $X^*$; for each atom $a\in A$ let $P_a$ and
$\varphi_a$ be the corresponding band projection and the coordinate
functional, respectively; $P_a$ and $\varphi_a$ are defined on
$X^*$. By \cite[Corollary~2.4.7]{Meyer-Nieberg:91}, $P_a$ and,
therefore, $\varphi_a$, is weak*-continuous. It follows that
$\varphi_a(f_\alpha)\to 0$ in $\alpha$. \Cref{un-atomic} yields that
$f_\alpha\goesun 0$.
\end{proof}
\begin{proposition}
Suppose that $X^*$ is atomic. The following are equivalent.
\begin{enumerate}
\item\label{aws-un-net} For every net $(f_\alpha)$ in $X^*$, if
\begin{math}
f_\alpha\xrightarrow{\abs{\sigma}(X^*,X)}0
\end{math}
then $f_\alpha\goesun 0$;
\item\label{aws-un-seq} For every sequence $(f_n)$ in $X^*$, if
\begin{math}
f_n\xrightarrow{\abs{\sigma}(X^*,X)}0
\end{math}
then $f_n\goesun 0$;
\item\label{aws-un-X} $X^*$ is order continuous.
\end{enumerate}
\end{proposition}
\begin{proof}
\eqref{aws-un-net}$\Rightarrow$\eqref{aws-un-seq} is trivial.
\eqref{aws-un-seq}$\Rightarrow$\eqref{aws-un-X} The proof is similar
to that of \Cref{w-un}. To show that $X^*$
is order continuous, suppose that $(f_n)$ is an order bounded
positive disjoint sequence in $X_+^*$. It follows that
\begin{math}
f_n\xrightarrow{\abs{\sigma}(X^*,X)}0
\end{math}
and, by assumption, $f_n\goesun 0$. Since the sequence is order
bounded, this yields $f_n\goesnorm 0$. Therefore, $X^*$ is order
continuous.
\eqref{aws-un-X}$\Rightarrow$\eqref{aws-un-net} By
\cite[Proposition~2.4.5]{Meyer-Nieberg:91}, band projections on
$X^*$ are $\abs{\sigma}(X^*,X)$-continuous. The proof is now
analogous to the implication
\eqref{ws-un-X}$\Rightarrow$\eqref{ws-un-net} in \Cref{ws-un}.
\end{proof}
\subsection*{Simultaneous weak* and un-convergence}
Section~4 of~\cite{Gao:14} contains several results that assert that
if a sequence or a net in $X^*$ converges in both weak* and
uo-topology then it also converges in some other topology. Several of
these results remain valid if uo-convergence is replaced with
un-convergence. In particular, this works for Proposition~4.1
in~\cite{Gao:14}. Propositions~4.3, 4.4, and~4.6 in~\cite{Gao:14}
remain valid under the additional assumption that $X^*$ is order
continuous (note that the dual positive Schur property already implies
that $X^*$ is order continuous by
\cite[Proposition~2.1]{Wnuk:13}). The proofs are analogous to the
corresponding proofs in~\cite{Gao:14}. Alternatively, the un-versions
of these may be deduced from the uo-versions using the following two
facts: first, every un-convergent sequence has a uo-convergent
subsequence and, second, a sequence $(x_n)$ converges to $x$ in a
topology $\tau$ iff every subsequence $(x_{n_k})$ has a further
subsequence $(x_{n_{k_i}})$ such that
$x_{n_{k_i}}\xrightarrow{\tau}x$.
\section{Un-compact operators}
Throughout this section, let $E$ be a Banach space, $X$ a Banach
lattice, and $T\in L(E,X)$. We say that $T$ is \term{(sequentially)
un-compact} if $TB_E$ is relatively (sequentially) un-compact in
$E$. Equivalently, for every bounded net $(x_\alpha)$ (respectively,
every bounded sequence $(x_n)$) its image has a subnet
(respectively, subsequence), which is un-convergent.
Clearly, if $T$ is compact then it is un-compact and sequentially
un-compact. \Cref{metriz,BX-compact} and \Cref{rcomp} yield the following.
\begin{proposition}\label{un-comp-op}
Let $T\in L(E,X)$.
\begin{enumerate}
\item If $X$ has a quasi-interior point then $T$ is un-compact iff
it is sequentially un-compact;
\item If $X$ is order continuous and $T$ is un-compact
then $T$ is sequentially un-compact;
\item\label{un-comp-op-atom} If $X$ is an atomic KB-space then $T$ is
un-compact and sequentially un-compact.
\end{enumerate}
\end{proposition}
\begin{proposition}
The set of all un-compact operators is a linear subspace of
$L(E,X)$. The set of all sequentially un-compact operators in
$L(E,X)$ is a closed subspace.
\end{proposition}
\begin{proof}
Linearity is straightforward. To prove closedness, suppose that
$(T_m)$ is a sequence of sequentially un-compact operators in $L(E,X)$
and $T_m\goesnorm T$. We will show that $T$ is sequentially
un-compact.
Let $(x_n)$ be a sequence in $B_E$. For every $m$, the sequence
$(T_mx_n)_n$ has a un-convergent subsequence. By a standard diagonal
argument, we can find a common subsequence for all these
sequences. Passing to this subsequence, we may assume without loss
of generality that for every $m$ we have $T_mx_n\goesun y_m$ for
some $y_m$. Note that
\begin{displaymath}
\norm{y_m-y_k}\le\liminf_n\norm{T_mx_n-T_kx_n}\le\norm{T_m-T_k}\to 0,
\end{displaymath}
so that the sequence $(y_m)$ is Cauchy and, therefore, $y_m\goesnorm
y$ for some $y\in X$.
Fix $u\in X_+$ and $\varepsilon>0$. Find $m_0$ such that
$\norm{T_{m_0}-T}<\varepsilon$ and
$\norm{y_{m_0}-y}<\varepsilon$. Find $n_0$ such that
\begin{math}
\bignorm{\abs{T_{m_0}x_n-y_{m_0}}\wedge u}<\varepsilon
\end{math}
whenever $n\ge n_0$. It follows from
\begin{displaymath}
\abs{Tx_n-y}\wedge u\le\abs{Tx_n-T_{m_0}x_n}+
\abs{T_{m_0}x_n-y_{m_0}}\wedge u+\abs{y_{m_0}-y}
\end{displaymath}
that
\begin{math}
\bignorm{\abs{Tx_n-y}\wedge u}<3\varepsilon,
\end{math}
so that $Tx_n\goesun y$.
\end{proof}
We do not know whether the set of all un-compact operators is closed.
It is easy to see that if we multiply a (sequentially) un-compact
operator by another bounded operator on the right, the product is
again (sequentially) un-compact. The following example shows that this
fails when we multiply on the left.
\begin{example}\label{Radem}
\emph{The class of all (sequentially) un-compact operators is not a
left ideal.} Let $T\colon\ell_1\to L_1$ be defined via
$Te_n=r_n^+$, where $(e_n)$ is the standard unit basis of $\ell_1$
and $(r_n)$ is the Rademacher sequence in $L_1$. Note that $T$ is
neither un-compact nor sequentially un-compact because the sequence
$(Te_n)$ has no un-convergent subsequences. On the other hand,
$T=TI_{\ell_1}$, where $I_{\ell_1}$ is the identity operator on
$\ell_1$. Observe that $I_{\ell_1}$ is un-compact by
\Cref{un-comp-op}\eqref{un-comp-op-atom}.
\end{example}
\begin{proposition}
In the diagram $E\xrightarrow{T}X\xrightarrow{S}Y$, suppose that $T$
is (sequentially) un-compact and $S$ is a lattice homomorphism. If the
ideal generated by $\Range S$ is dense in $Y$ then $ST$ is
(sequentially) un-compact.
\end{proposition}
\begin{proof}
We will prove the statement for the sequential case; the other case
is analogous. Let $(h_n)$ be a norm bounded sequence in $E$. By
assumption, there is a subsequence $(h_{n_k})$ such that
$Th_{n_k}\goesun x$ for some $x\in X$. Let $Z=\Range S$; it is a
sublattice of $Y$. Fix $u\in Z_+$. Then $u=Sv$ for some $v\in
X_+$, and $\abs{Th_{n_k}-x}\wedge v\goesnorm 0$. Applying $S$, we
get $\bigabs{STh_{n_k}-Sy}\wedge u\goesnorm 0$. Therefore,
$STh_{n_k}\goesun Sx$ in $Z$. It follows from
\Cref{sublat}\eqref{sublat-maj} and~\eqref{sublat-dense} that
$STh_{n_k}\goesun Sx$ in $Y$.
\end{proof}
\begin{example}
\emph{The set of all sequentially un-compact operators is not order
closed.} Let $T$ be as in \Cref{Radem}. Let $T_n=TP_n$, where
$P_n$ is the $n$-th basis projection on $\ell_1$, i.e.,
$T_nh=\sum_{i=1}^nh_ir_i^+$ for $h\in\ell_1$. It is easy to see that
each $T_n$ is finite rank and, therefore, sequentially
un-compact. Note that $T_n\uparrow T$, yet $T$ is not sequentially
un-compact.
\end{example}
\begin{proposition}
Suppose that for every sequence $(T_n)$ of sequentially un-compact
operators in $L(c_0,X)$, $T_n\uparrow T$ implies that $T$ is
sequentially un-compact. Then $X$ is a KB-space.
\end{proposition}
\begin{proof}
Suppose not. Then there is a lattice isomorphism $T\colon c_0\to
X$. Put $x_n=Te_n$, where $(e_n)$ is the standard unit basis of
$c_0$. Put $T_n=TP_n$, where $P_n$ is the $n$-th basis projection on
$c_0$, i.e., $T_nh=\sum_{i=1}^nh_ix_i$ for $h\in c_0$. It follows
that $T_nh\goesnorm Th$, so that $T_nh\uparrow Th$ for every $h\ge
0$ and, therefore, $T_n\uparrow T$. For each $n$, $T_n$ has finite
rank and, therefore, is sequentially un-compact.
We claim that, nevertheless, $T$ is not sequentially un-compact. Put
$w_n=e_1+\dots+e_n$ in $c_0$. Note that $(w_n)$ is norm bounded and
$Tw_n=x_1+\dots+x_n$. Since $T$ is an isomorphism, $(Tw_n)$ is not
norm-convergent. Since $(Tw_n)$ is increasing, it is not
un-convergent by \Cref{monot}\eqref{monot-un}. Similarly, no
subsequence of $(Tw_n)$ is un-convergent.
\end{proof}
We do not know whether the converse is true.
Next, we study whether un-compactness is inherited under
domination. The following example shows that, in general, the answer
is negative.
\begin{example}
Let $T$ be as in \Cref{Radem}. Let $S\colon\ell_1\to L_1$ be defined
via $Se_n=\one$. Then $S$ is a rank-one operator; hence it is
compact and un-compact. Clearly, $0\le T\le S$. Yet $T$ is not
un-compact.
\end{example}
\begin{proposition}
Suppose that $S,T\colon E\to X$, $0\le S\le T$, $X$ is a KB-space
and $T$ is a lattice homomorphism. If $T$ is (sequentially)
un-compact then so is $S$.
\end{proposition}
\begin{proof}
We will prove the sequential case; the other case is similar.
Let $(h_n)$ be a bounded sequence in $E$. Passing to a subsequence,
we may assume that $(Th_n)$ is un-convergent. In particular, it is
un-Cauchy. Fix $u\in X_+$. Note that
\begin{displaymath}
\abs{Sh_n-Sh_m}\wedge u
\le\bigl(S\abs{h_n-h_m}\bigr)\wedge u
\le\bigl(T\abs{h_n-h_m}\bigr)\wedge u
=\abs{Th_n-Th_m}\wedge u\goesnorm 0
\end{displaymath}
as $n,m\to\infty$. It follows that $(Sh_n)$ is un-Cauchy and,
therefore, un-converges by \Cref{BX-complete}.
\end{proof}
We would like to mention that the class of un-compact operators is
different from several other known classes of operators. We already
mentioned that every compact operator is un-compact. The converse is
false as the identity operator on any infinite-dimensional atomic
KB-space is un-compact but not compact.
Recall that an operator between Banach lattices is AM-compact if it
maps order intervals to relatively compact sets.
\begin{proposition}
Every order bounded un-compact operator is AM-compact.
\end{proposition}
\begin{proof}
Let $T\colon X\to Y$ be an order bounded un-compact operator between
Banach lattices. Fix an order interval $[a,b]$ in $X$. Since $T$ is
un-compact, $T[a,b]\subseteq C$ for some un-compact set $C$. Since
$T$ is order bounded, $T[a,b]\subseteq[c,d]$ for some $c,d\in
Y$. Note that $[c,d]$ is un-closed, hence $C\cap[c,d]$ is un-compact
and, being order bounded, is compact. It follows that $T[a,b]$ is
relatively compact.
\end{proof}
Note that the converse is false: the identity operator on $c_0$ is
AM-compact but not un-compact.
The identity operator on $\ell_1$ is un-compact, yet it is neither
L-weakly compact nor M-weakly compact.
Finally, we note that if $T$ is sequentially un-compact and
semi-compact then $T$ is compact. Indeed, let $(h_n)$ be a bounded
sequence in $E$. There is a subsequence $(h_{n_k})$ such that
$Th_{n_k}\goesun x$ for some $x\in X$. Since $T$ is semi-compact, the
sequence $(Th_{n_k})$ is almost order bounded and, therefore,
$Th_{n_k}\goesnorm x$ by \cite[Lemma~2.9]{DOT}.
Finally, we discuss when weakly compact operators are un-compact.
\begin{lemma}\label{w-un-lim}
If $x_n\goesw x$ and $x_n\goesun y$ then $x=y$.
\end{lemma}
\begin{proof}
Without loss of generality, $y=0$. By \Cref{KP}, there
exist a subsequence $(x_{n_k})$ and a disjoint sequence $(d_k)$
such that $x_{n_k}-d_k\goesnorm 0$. It follows that
$x_{n_k}-d_k\goesw 0$, so that $d_k\goesw x$. Now
\cite[Theorem~4.34]{Aliprantis:06} yields $x=0$.
\end{proof}
\begin{theorem}
A Banach lattice $X$ is atomic and order continuous iff $T$ is
sequentially un-compact for every Banach space $E$ and every weakly
compact operator $T\colon E\to X$.
\end{theorem}
\begin{proof}
The forward implication follows immediately from \Cref{w-un}. To
prove the converse, let $(x_n)$ be a weakly null sequence in $X$. By
\Cref{w-un}, it suffices to show that $x_n\goesun 0$. Define
$T\colon\ell_1\to X$ via $Te_n=x_n$. By
\cite[Theorem~5.26]{Aliprantis:06}, $T$ is weakly compact. By
assumption, $T$ is sequentially un-compact. It follows that $(Te_n)$
has a un-convergent subsequence, i.e., $x_{n_k}\goesun x$ for some
$x\in X$ and a subsequence $(x_{n_k})$. \Cref{w-un-lim} yields
$x=0$. By the same argument, every subsequence of $(x_n)$ has a
further subsequence which is un-null; since un-convergence is
topological, it follows that $x_n\goesun 0$.
\end{proof}
\begin{corollary}
Every operator from a reflexive Banach space to an atomic order
continuous Banach lattice is sequentially un-compact.
\end{corollary}
\bigskip
\subsection*{Acknowledgement and further remarks.}
Most of the work on this paper was done during a visit of the first
and the second author to the University of Alberta. After the work on
this paper was essentially completed, we learned of recent preprints
\cite{Zab,GLX}. In the former, the author studies \emph{unbounded
absolute weak} convergence; it is shown there that in certain
situations it agrees with un-convergence. In the latter, techniques
of unbounded convergence are used to study risk measures.
The authors would like to thank the reviewers for valuable comments
and improvements.
|
\section{Introduction}\label{sec:intro}
Many experimental measurements demonstrate the profound effect that initial state fluctuations of nuclear collisions have on the subsequent dynamics of the system, but there is growing awareness that fluctuations and their dissipation also occur throughout the evolution. In earlier work we suggested that viscous diffusion broadens the rapidity dependence of two-particle transverse momentum correlations \cite{Gavin:2006xd}. That work stimulated an experimental analysis by STAR \cite{Agakishiev:2011fs} and the discovery that the rapidity structure of these correlations not only broadens but also develops a distinctively non-Gaussian structure not predicted by ref. \cite{Gavin:2006xd}.
We employ second order hydrodynamics with stochastic noise to develop an evolution equation for two particle transverse momentum correlations \cite{Gavin:2016hmv}. In this work we will compare rapidity correlation structures evolving from first and second order diffusion. In sec. \ref{sec:theory} we briefly derive evolution equations for the transverse momentum correlation function as well as describe how this correlation function relates to experimental observables. In the interest of brevity, we leave the rigorous derivations to \cite{Gavin:2016hmv}. Instead we focus on the results in sec. \ref{sec:results}. We find that second order hydrodynamics is required to explain the measured rapidity correlation structure and consequently that these measurements can constrain a transport coefficient that is sensitive to the thermalization process.
\section{Fluctuations and Correlations}\label{sec:theory}
Nuclear collisions produce a high energy density fluid that flows outward with an average transverse velocity $v_r$. Small deviations of the flow occur in each event. These deviations perturb the transverse momentum current of the fluid by an amount $g_t = T_{0r} - \langle T_{0r}\rangle$, where $T^{\mu\nu}$ is the stress energy tensor. Viscous friction arises as neighboring fluid elements flow past one another. This friction reduces velocity fluctuations, driving the velocity toward $v_r$ and the momentum deviations $g_t$ to zero. The final size of the fluctuations depends on the magnitude of the viscosity and the lifetime of the fluid.
We consider fluctuations of a fluid at rest with energy density $e$ and pressure $p$. Small fluctuations produce a small velocity $\mathbf{v}$ corresponding to a momentum current $\mathbf{M} \approx (e+p)\mathbf{v}$. To linear order in the fluctuations, we write the conservation form of the relativistic Navier-Stokes equation:
\begin{equation}\label{eq:NavStokes2}
\frac{\partial}{\partial t} \mathbf{M} + \bm{\nabla} p = \frac{\zeta + \tfrac{1}{3}\eta}{w}\bm{\nabla}(\nabla\cdot \mathbf{M}) + \frac{\eta}{w}\nabla^2\mathbf{M}
\end{equation}
where $\eta$ and $\zeta$ are the shear and bulk viscosity coefficients and $w=e+p$ is the enthalpy density.
The momentum fluctuation density can be written in terms of curl free longitudinal modes $\mathbf{g}_l$ and divergence free shear modes $\mathbf{g}$ so we have $\mathbf{M} = \mathbf{g}_l+ \mathbf{g}$, where $\mathbf{\nabla} \times \mathbf{g}_l=0$ and $\mathbf{\nabla} \cdot \mathbf{g}=0$. By taking the curl of (\ref{eq:NavStokes2}) we find the shear modes satisfy
\begin{equation}\label{eq:DiffModes}
\frac{\partial }{\partial t}\mathbf{g} = \nu\nabla^2 \mathbf{g},
\end{equation}
where $\nu = \eta/w$ is the kinematic viscosity. Equation (\ref{eq:DiffModes}) shows that Fick's Law holds for the density of transverse momentum fluctuations from local equilibrium.
We look for the diffusion of \textit{correlations} of the fluctuations $\mathbf{g}$ at different points in the fluid
$r_g = \langle\mathbf{g}_1\mathbf{g}_2\rangle - \langle\mathbf{g}_1\rangle\langle\mathbf{g}_2\rangle $.
In equilibrium, thermodynamic noise will cause such correlations to occur so $r_g^{le}\neq 0$ and the interesting correlations are those that differ from the equilibrium value $\Delta r_g = r_g - r_g^{le}$. It turns out that for two particle momentum correlations (\ref{eq:DiffModes}) becomes $\left[\frac{\partial }{\partial t} - \nu\left(\nabla^2_1 + \nabla^2_2\right)\right] \Delta r_g = 0$. For a detailed derivation please see ref. \cite{Gavin:2016hmv}.
We assume the event-averaged flow velocity has the Bjorken form, $u^\mu=(t/\tau,0,0,z/\tau)$, where $\tau =(t^2 -z^2)^{1/2}$ is the proper time and $\eta = (1/2) \log ((t+z)/(t-z))$ is spatial rapidity.
In an expanding system, observe that the rapidity density of total momentum $G\equiv \int g \tau dx_{\bot}$, where the integral is over the transverse area of the two colliding nuclei. If one identifies spatial
rapidity $\eta$ with the momentum-space rapidity of particles, then $G$ is observable and we look for correlations of the form
\begin{equation}\label{eq:RGdef}
r_G = \langle G(\mathbf{x}_1) G(\mathbf{x}_2)\rangle - \langle G(\mathbf{x}_1)\rangle\langle G(\mathbf{x}_2)\rangle
\end{equation}
Additionally we switch coordinates to relative $\eta_r = \eta_1 -\eta_2$ and average $\eta_a = (\eta_1 + \eta_2)/2$ rapidity and finally we have our first order result
\begin{equation}\label{eq:DiffModes2}
\left[\frac{\partial }{\partial \tau} -
\frac{\nu}{\tau^2}\left( 2\frac{\partial^2}{\partial \eta_r^2} + \frac{1}{2}\frac{\partial^2}{\partial \eta_a^2}\right)
\right] \Delta r_G = 0.
\end{equation}
Second order hydrodynamics is especially important for diffusive phenomena, where it renders the theory causal. In first order diffusion (\ref{eq:DiffModes}), a delta function perturbation instantaneously spreads into a Gaussian, with tails extending to infinity. New transport coefficients at second order include relaxation times for shear and bulk stresses, among other terms. Linearized forms of the second order equations are discussed in \cite{Romatschke:2009im,Young:2014pka}. To linear order the shear modes satisfy a Maxwell-Cattaneo equation
\begin{equation}\label{eq:2DiffModes}
\left(\tau_\pi \frac{\partial^2}{\partial t^2} + \frac{\partial }{\partial t}\right)\mathbf{g} = \nu\nabla^2 \mathbf{g}
\end{equation}
where the transport coefficient $\tau_\pi$ is a relaxation time for the shear modes. As with the first order case, we look to solve (\ref{eq:2DiffModes}) for the correlations (\ref{eq:RGdef}). We find
\begin{equation}\label{eq:2VisDiff}
\left[ \frac{\tau_\pi}{2}\frac{\partial^2}{\partial \tau^2}
+\frac{\partial}{\partial \tau}
-\frac{\nu}{\tau^2}\left( 2\frac{\partial^2}{\partial \eta_r^2} + \frac{1}{2}\frac{\partial^2}{\partial \eta_a^2} \right) \right]
\Delta r_G = 0.
\end{equation}
For a detailed derivation please refer to ref. \cite{Gavin:2016hmv}. Observe that (\ref{eq:2VisDiff}) reduces to the first order result (\ref{eq:DiffModes2}) when $\tau_\pi=0$. Further observe that (\ref{eq:2VisDiff}) has features of both wave and diffusion equations. As we will discuss in sec. \ref{sec:results}, at early times, fluctuation signals propagate as waves while at later times diffusion fills in the space between the wave fronts. It is precisely this behavior that delays the growth of the rapidity correlation structure in comparison to the first order case.
The correlation function $\Delta r_G$ is observable by measuring the covariance
\begin{equation}\label{C0def}
{\cal C} = \langle N\rangle^{-2}\langle \sum_{a\neq b} p_{t,a}p_{t,b}\rangle -\langle p_t\rangle^2
= \langle N\rangle^{-2}\int \Delta r_G(\eta_r, \eta_a)d\eta_r d\eta_a,
\end{equation}
where $a$ and $b$ label particles from each event, and the brackets represent the event average. The average transverse momentum is $\langle p_t\rangle \equiv \langle \sum_a p_{t,a}\rangle/\langle
N\rangle$. In the absence of correlations ${\cal C} = 0$, as is the case for local equilibrium in an infinite system.
The STAR collaboration at RHIC reports a differential version of the quantity $\cal C$ as a function of relative pseudorapidity $\eta_r$ and azimuthal angle $\phi_r$ of pairs:
\begin{equation}\label{exptC}
{\cal C}(\eta_r, \phi_r) = \left({\langle N \rangle_1 \langle N \rangle_2}\right)^{-1}
\left\langle \sum\limits_{a\neq b} p_{_{t,a}}p_{_{t,b}}\right\rangle_{1,2}
- \langle p_t\rangle_1 \langle p_t\rangle_2,
\end{equation}
where the numbers $\langle N \rangle_k$ and $\langle p_t \rangle_k$ refer to the particle number and transverse momentum in $(\eta_k, \phi_k)$ bins for particles $k= 1,2$ \cite{Agakishiev:2011fs}. We will discuss this observable in detail in the next section.
\section{Results}\label{sec:results}
We will now explore the rapidity dependence of transverse momentum fluctuations. In an earlier work we show that the relative rapidity width, $\sigma$, of (\ref{exptC}) can be related to the shear viscosity of the system using first order diffusion \cite{Gavin:2006xd}. Here we discuss how the diffusion equation derived from second order hydrodynamics, (\ref{eq:2VisDiff}), modifies this result. In particular we demonstrate that the first order theory cannot simultaneously explain both the width and shape of (\ref{exptC}) in relative rapidity; where the first order theory fails, the second order theory succeeds. Interestingly, while the kinematic viscosity $\nu=\eta/Ts$ remains the dominant factor in determining the width, the relaxation time of the second order theory, $\tau_\pi$, is the primary influence on the shape.
In Au+Au collisions at the top RHIC energy, STAR has measured ${\cal C}$ in both its differential form (\ref{exptC}) and its relative rapidity width \cite{Agakishiev:2011fs}.
The measured differential shape of (\ref{exptC}) resembles that of similar measurements of the ``ridge'' without $p_t$ weights. In Ref.\ \cite{Agakishiev:2011fs} authors interpret the near side, $|\phi_r| < 1$, of (\ref{exptC}) as a peak sitting on a flat pedestal and fit the $\phi_r$ integrated profile ${\cal C}(\eta_r)$ to extract the magnitude of the pedestal.
The measured width of ${\cal C}(\eta_r)$ is shown in fig.\ \ref{fig:sigNp}. Solid circles represent the RMS width in $\eta_r$ of (\ref{exptC}) integrated over an azimuthal range of $|\phi_r| < 1$ radians.
The gray error band in fig.\ \ref{fig:sigNp} represents the uncertainty in their fit procedure.
Figure \ref{fig:C12} shows the ${\cal C}(\eta_r)$ profile from which these widths are calculated. Open stars are from Ref.\ \cite{Agakishiev:2011fs} and solid circles are from Ref.\ \cite{PrivComm}. In all cases we have subtracted the pedestal.
\begin{figure}
\includegraphics[width=0.58\textwidth]{fig1.pdf}%
\hspace{0.02\textwidth}%
\begin{minipage}[b]{0.4\textwidth}\caption{\label{fig:sigNp}Rapidity width of ${\cal C}(\eta_r)$ as a function of the number of participants for second order momentum diffusion calculations (solid curve) compared to first order results (dot-dashed curve). Data (solid circles) from STAR include shaded area to denote the systematic uncertainty in the fit procedure \cite{Agakishiev:2011fs}. }
\end{minipage}
\end{figure}
The flat pedestal might be interpreted as the long-range correlations usually associated with the ridge while shorter range correlations would influence the shape of the peak sitting on the pedestal. Our calculations here specifically address the short range correlations. As discussed in Ref.\ \cite{Gavin:2006xd}, the growth of the width of this peak from peripheral to central collisions can be used to extract the viscosity. To compute the width in the second order theory, we follow ref.\ \cite{Aziz:2004qu} and multiply (\ref{eq:2VisDiff}) by $\eta_r^n$. Next, we integrate over $\eta_r$ and $\eta_a$ and use $\int\eta_r^n\partial^2\Delta r_G/\partial\eta_r^2 = n(n-1)\int\eta_r^{n-2}\Delta r_G$, which is nonzero only for $n \ge 2$. We find
\begin{equation}\label{eq:moments}
\left( \frac{\tau_\pi}{2}\frac{d^2}{d\tau^2} + \frac{d}{d\tau}\right)
A\langle\eta_r^n\rangle =
\frac{2\nu}{\tau^2}n(n-1)A\langle\eta_r^{n-2}\rangle,
\end{equation}
where $\langle\eta_r^n\rangle = A^{-1}\int\eta_r^n\Delta r_G d\eta_r d\eta_a$ are the normalized moments of the rapidity correlation function. The amplitude $A$ and the mean $\langle \eta_r\rangle$ both satisfy (\ref{eq:moments}) with the right side equal to zero. The vanishing first moment $\langle\eta_r\rangle=0$ makes physical sense if we assume a symmetric system. Therefore, the width can then be described as $\sigma^2 = \langle\eta_r^2\rangle - \langle\eta_r\rangle^2 = \langle\eta_r^2\rangle$ and following from (\ref{eq:moments}) we obtain the differential equation for the width,
\begin{equation}\label{eq:moment2}
\left( \frac{\tau_\pi}{2}\frac{d^2}{d\tau^2} + \frac{d}{d\tau}\right)
\sigma^2 = \frac{4\nu}{\tau^2}.
\end{equation}
The solution of (\ref{eq:moment2}) yields the time dependence of the broadening of ${\cal C}(\eta_r)$. We can observe that we should expect that longer lived collisions have broader rapidity distribution by examining the first order case where $\tau_\pi=0$. The solution of (\ref{eq:moment2}) is then
\begin{equation}\label{eq:DeltaV}
\sigma^2 =\sigma_0^2+ \frac{4\nu}{\tau_0}\left(1-\frac{\tau_0}{\tau}\right),
\end{equation}
as found in Ref. \cite{Gavin:2006xd}. Here $\nu=\eta/Ts=const$ is the kinematic viscosity, $\tau_0$ is the formation time, and $\sigma_0$ is the width in peripheral collisions. STAR used this method to estimate the average shear viscosity to entropy density ratio to be $\eta/s=0.13\pm 0.03$ \cite{Agakishiev:2011fs}.
To find the rapidity width for second order diffusion using (\ref{eq:moment2}) we must specify an initial condition for $d\sigma^2/d\tau\equiv \theta_0^2$ at $\tau = \tau_0$, the value of which is unknown.
We examine two cases, one where $\theta_0^2=0$ which assumes that no modifications to the correlation function (\ref{eq:RGdef}) occur prior to the formation time, and one that attempts to account for the effects of initial state expansion. In this proceedings we focus on the latter case and direct the reader to Ref. \cite{Gavin:2016hmv} for a detailed comparison of these two conditions.
For (\ref{eq:2VisDiff}), we take the initial correlation function to satisfy
\begin{equation}\label{eq:CausalIC}
\frac{\partial \Delta r_G}{\partial \tau}\Big|_{\tau=\tau_0}
= \frac{\nu_0}{\tau_0^2}\left( 2\frac{\partial^2}{\partial \eta_r^2} + \frac{1}{2}\frac{\partial^2}{\partial \eta_a^2} \right)
\Delta r_G,
\end{equation}
which corresponds to $\theta_0^2=4\nu/\tau_0^2$ in (\ref{eq:moment2}). Using (\ref{eq:CausalIC}) and solving (\ref{eq:moment2}) we find
\begin{equation}\label{eq:SOwidth}
\sigma^2 = \sigma_0^2 +
\frac{\theta_0^2 \tau_\pi}{2} \left(1 - e^{-2(\tau-\tau_0)/\tau_\pi}\right)
+ \frac{8\nu}{\tau_\pi}\int\limits_{\tau_0}^\tau \! du \!
\int\limits_{\tau_0}^u \! \frac{ds}{s^2} e^{2(s-u)/\tau_\pi}.
\end{equation}
The solid black curve in fig.\ \ref{fig:sigNp} shows the value of (\ref{eq:SOwidth}) at the freeze out time parameterized as
\begin{equation}\label{eq:tauF}
\tau_F - \tau_0 = K(R(N_{part})-R_0)^2
\end{equation}
where $\tau_0$ is the formation time and $R_0$ is roughly the proton size. We compute $N_{part}$ and $R$ from a Glauber model and fix the constant $K$ so that the freeze out time in the most central collisions has a specified value $\tau_{Fc}$.
Again we take $\nu = \eta/Ts=const$ for $\eta/s = 1/4\pi$, but now with $T=150$~MeV for all centralities. We must now specify the second-order relaxation time $\tau_\pi = \beta \nu$, for which we take $\beta = 10$. The values $\tau_0 = 1.0$~fm and $\tau_{Fc} = 10$~fm then give superb agreement with data.
The dash-dot curve shows our best fit to this data using the first order result (\ref{eq:DeltaV}) evaluated at $\tau_F$, eq.\ (\ref{eq:tauF}). Again $\eta/s = 1/4\pi$ but we take the freeze out temperature to be $T = 140$~MeV for all centralities. Values of the space time parameters $\tau_0 = 0.65$~fm and $\tau_{Fc} = 12$~fm then specify the lifetime (\ref{eq:tauF}).
One might argue the rough agreement in fig.\ \ref{fig:sigNp} for the first order result is compelling were it not for the fact that it consistently overestimates the data in the region where the data grows the fastest.
It is a feature of first order diffusion that it occurs a-causally; correlations diffuse too rapidly. Hence, we see that collision systems that freeze out at earlier times disagree with the first order case while more central, longer lived collision systems match with the asymptotic limit of (\ref{eq:DeltaV}). For this reason the estimate of the shear viscosity in Ref. \cite{Agakishiev:2011fs} is still on strong footing since authors used only the difference in central and most peripheral widths.
\begin{figure}
\begin{minipage}{0.485\textwidth}
\includegraphics[width=\textwidth]{Ctau_constNu_hos008_tau065_tauF12_consTF140_noData.pdf}
\caption{\label{fig:Ctau1st}Time evolution of the rapidity structure of transverse momentum correlations from first order diffusion (\ref{eq:DiffModes2}) for $5-10\%$ central collisions.}
\end{minipage}\hspace{0.03\textwidth}%
\begin{minipage}{0.485\textwidth}
\includegraphics[width=\textwidth]{fig3.pdf}
\caption{\label{fig:Ctau2nd}Time evolution of the rapidity structure of transverse momentum correlations from second order diffusion (\ref{eq:2VisDiff}) for $5-10\%$ central collisions.}
\end{minipage}
\end{figure}
The asymptotic value of the first order solution (\ref{eq:DeltaV}) is
\begin{equation}\label{eq:DeltaVinf}
\sigma^2_\infty =\sigma_0^2+ \frac{4\nu}{\tau_0}.
\end{equation}
This saturation of the rapidity width to the value (\ref{eq:DeltaVinf}) is a straightforward consequence of Bjorken flow. In a stationary liquid, a spike in momentum diffuses over a range $\sim (2\nu t)^{1/2}$ that grows with time $t$. Bjorken expansion of the underlying fluid stretches the longitudinal scale $\propto t$, rapidly overtaking diffusion and ``freezing in'' the initial inhomogeneity. Correspondingly, one can observe this same behavior in (\ref{eq:2VisDiff}). At long times the rightmost term vanishes, meaning the rapidity of particles ceases to change due to shear forces.
Consequently, the rapidity width from second order diffusion, in the case of $\theta_0^2=0$, will reach the same asymptotic value. Importantly, the difference between the first and second order cases is that in the second order case the growth of the rapidity width, or equivalently the evolution of the correlation function (\ref{eq:RGdef}), is delayed according to the relaxation time $\tau_\pi$.
The freeze out time plays an important role in selecting the state of the correlation function measured by the experiment. For constant $\nu$ and $\tau_\pi$ all centralities follow the same trajectory in the solution of (\ref{eq:2VisDiff}) and consequently (\ref{eq:moment2}). If $\nu$ and $\tau_\pi$ change throughout the evolution, each collision would have a unique solution to (\ref{eq:2VisDiff}), but we leave that to future work. Figures \ref{fig:Ctau1st} and \ref{fig:Ctau2nd} show the time evolution of the correlation function for $5-10\%$ central collisions for the first and second order cases respectively.
In both cases we assume the initial transverse momentum correlation function to be $\Delta r_G(\eta_r,\eta_a,\tau_0) = A e^{-\eta_r^2/2\sigma_0^2}e^{-\eta_a^2/2\Sigma_0^2}$. This distribution is motivated by the rapidity dependence of measured correlation functions for multiplicity and net charge in pp collisions. We set the initial width in relative rapidity, $\sigma_0$ to fit the most peripheral distribution in fig.\ \ref{fig:C12}. Furthermore, we assume there is insufficient time for significant evolution in the three most peripheral cases in fig.\ \ref{fig:C12}. The data supports this claim and gives a consistent value of $\sigma_0 = 0.50$. The average pseudo-rapidity width $\Sigma_0\approx 5-6$ units is assumed to be a ``large'' value relative to the size of experimental acceptance. We will take $A$ to fit the peak value of the measured $\cal{C}$. This parameter has little impact on our current study, since we are only concerned with the shape of the function. We use (\ref{eq:CausalIC}) for the initial value of the first derivative.
\begin{figure}
\begin{center}
\includegraphics[width=0.75\textwidth]{fig2.pdf}%
\end{center}
\caption{\label{fig:C12} Second order momentum diffusion calculations (solid curves) compared to the rapidity dependence of the measured covariance (\ref{exptC}). First order calculations are also compared for best fit to $\sigma$ in fig.\ \ref{fig:sigNp} (dot-dashed curves) and the best fit to ${\cal C}(\eta_r)$ (dashed curves). Data (open stars) are from \cite{Agakishiev:2011fs} and (filled circles) from \cite{PrivComm}. Percentages of the cross section indicate centrality, with each panel corresponding to a width measurement in fig.\ \ref{fig:sigNp}. }
\end{figure}
The striking difference between figures \ref{fig:Ctau1st} and \ref{fig:Ctau2nd} is our primary result. In the first order case, diffusion is the only process taking place; starting from an initial Gaussian shape the rapidity distribution of correlations will always remain Gaussian. In the second order case, fluctuation signals propagate in opposite directions with a wave speed of $v=\sqrt{\nu/\tau_\pi}$ in Cartesian coordinates. In rapidity coordinates, as longitudinal expansion overtakes the signals, fluctuations are frozen into the final distribution. Depending on what time this takes place a flattened or even bimodal peak structure is visible. Indeed STAR's discovery of this type of rapidity structure in the measurement of (\ref{exptC}), shown in fig.\ \ref{fig:C12}, is the motivation for this work.
In fig.\ \ref{fig:C12} we show comparisons of first and second order cases to measured data (with the pedestal subtracted). The most noticeable feature is that in central collisions there is a flattening of the peak and even hints of a double bump structure. First order diffusion will never be able to achieve this shape regardless of choices of parameters. We illustrate this with two choices of parameter sets. For the dot-dashed curves we use the same parameters chosen to fit the width in fig. \ref{fig:sigNp}. Agreement is poor. Since the first order shape is always Gaussian, much of the value of the width will depend on how the tails are truncated to match the acceptance of the data. If we ignore the acceptance constraints and just try to match ${\cal C}(\eta_r)$ in fig. \ref{fig:C12} as best as possible we find the dashed curves for parameter values $\eta/s = 1/4\pi$, $T = 110$~MeV, $\tau_0 = 0.50$~fm, and $\tau_{Fc} = 10$~fm. Agreement with the measured shape is still quite poor. The solid black curves use the second order case with the same parameters as used in fig. \ref{fig:sigNp}. Agreement with data is much better than either of the first order cases and we see signs of the flattening of the peak in central collisions.
The flattening of the peak is definitively a second order effect that is modulated by the relaxation time $\tau_\pi=\beta\nu$. This suggests that the experiment has access to this relaxation time. In fig. \ref{fig:ChangeBeta} we examine the effect of $\tau_{\pi}$ on our $5-10\%$ centrality result. In all cases we use the same parameter choices as used in in figs. \ref{fig:sigNp}, \ref{fig:Ctau2nd}, and \ref{fig:C12} with the exception of $\beta$. Previously for all second order results we chose $\beta=10$ since it yielded the most consistently strong agreement with all centralities. This solution is represented as the thick black line in fig. \ref{fig:ChangeBeta}. Incidentally, kinetic theory of massless Boltzmann particles predicts a value of $\beta=5$. Larger values of $\beta$ correspond to slower propagation of fluctuation signals in the medium, a fact that could be used to characterize the nature of the medium. The slower signal propagation associated with increasing $\beta$ means that the system takes longer to reach the asymptotic diffusion limit and longitudinal expansion ``freezes'' the correlation structure earlier in its development. As we can see in fig. \ref{fig:Ctau2nd}, at earlier times the wave nature of the structure still dominates the correlation shape. So, for larger $\beta$ values we should see an emergence of a more well defined double bump structure, as we see in the dashed line in fig. \ref{fig:ChangeBeta} corresponding to $\beta=14$.
We remark that the resolution of the double peak structure also depends on the narrowness of the initial width $\sigma_0$. If the initial correlation structure is wide then we may never see a well defined double peak structure, but that structure will still be decidedly non-Gaussian. Conversely if the initial structure is very narrow we could still see a double peak structure with smaller $\beta$ values.
\begin{figure}
\includegraphics[width=0.58\textwidth]{fig5_newIC.pdf}%
\hspace{0.02\textwidth}%
\begin{minipage}[b]{0.4\textwidth}\caption{\label{fig:ChangeBeta} Rapidity structure of transverse momentum correlations at freeze out from second order diffusion (\ref{eq:2VisDiff}). Different values of $\tau_\pi=\beta\nu$ are chosen by changing the parameter $\beta$. Results are selected for $5-10\%$ central collisions and compared to data \cite{Agakishiev:2011fs,PrivComm}. }
\end{minipage}
\end{figure}
\section{Conclusions}\label{sec:conclusion}
In this paper we study the rapidity correlation structure of transverse momentum correlations in nuclear collisions (\ref{C0def}). In earlier work we suggested how these correlations could be used to study viscosity \cite{Gavin:2006xd}. In an effort to obtain an estimate of the shear viscosity to entropy density ratio, STAR measured a differential version of these correlations (\ref{exptC}) and, in doing so, discovered that the rapidity shape has a distinctive non-Gaussian flattening in the more central centrality classes. In ref. \cite{Gavin:2016hmv} we use second order hydrodynamics with stochastic noise to develop a second order diffusion equation (\ref{eq:2VisDiff}) for the evolution of transverse momentum fluctuations that lead to the correlations (\ref{C0def}).
Comparisons of first and second order diffusion results to data in figs. \ref{fig:sigNp} and \ref{fig:C12} suggest that first order diffusion cannot be used to simultaneously explain both the broadening of and the emergence of the non-Gaussian shape of the correlation structure from peripheral to central collisions. We attribute this failure to the a-causal nature of first order diffusion where fluctuation signals can propagate at infinite speeds throughout the medium. In second order diffusion fluctuation signals propagate as wave fronts with speed $\sqrt{\nu/\tau_\pi}$ where $\nu=\eta/Ts$ is the kinematic viscosity and $\tau_\pi=\beta\nu$ is the relaxation time related to the transport of the fluctuation signal through the medium. We solve the second order diffusion equation (\ref{eq:2VisDiff}) for constant $\nu$ and $\tau_\pi$ to find excellent agreement with data using a value of $\beta=10$. We further find that the kinematic viscosity is the primary factor in determining the broadening of the correlation structure and $\tau_\pi$ has significant influence over the flattening or even bimodal nature of the correlation peak.
The emergence of a flattened peak in the rapidity correlation structure and the natural way in which second order hydrodynamics accommodates this behavior suggests that experiments have the ability to measure the relaxation time $\tau_\pi$. Since $\tau_\pi$ is the characteristic time for the dissipation of fluctuations in the medium, studying this quantity can yield information on the thermalization process.
\ac
We thank Rajendra Pokharel for collaboration in the early stages of this work. Special thanks to Monika Sharma and Claude Pruneau for discussing the STAR data. We thank Victoria Drolshagen, Mauricio Martinez, Jaki Noronha-Hostler, Jorge Noronha, Scott Pratt, and Clint Young for discussions. This work was supported in part by the U.S. NSF grant PHY-1207687.
\section*{References}
\bibliographystyle{utphys}
|
\section{Semantics}
\label{sec:dsa}
In this section we formalize our technique
by explaining the language and defining representation of states.
\subsection{Language}
We formulate our analysis over a language modelled on C.
For simplicity of exposition we restrict our description to a subset of C which includes C style structures and 1-dimensional arrays.
Let \text{$\mathbb{C}$}\xspace, \text{\sffamily V$\!_s$}\xspace, and \text{$\mathbbm{E}$}\xspace be the sets of values computed by the program, variables appearing in the program, and expressions
appearing in the program respectively.
A value \text{$c \in \text{$\mathbb{C}$}\xspace$} can be an integer, floating-point or boolean value.
A variable \text{$v \in \text{\sffamily V$\!_s$}\xspace$}
can be a scalar variable, a structure variable, or an array variable.
We define our program to have only one array variable denoted as \text{$a$}\xspace.
However, in practice, we can handle multiple arrays in a program as explained in our technical report~\cite{Jana2016Scaling}.
We also define \text{$\text{$\mathbb{E}_A$}\xspace \subseteq \text{$\mathbbm{E}$}\xspace$}
as set of array expressions of the form \text{\text{$a$}\xspace[\text{\sf\em E}\xspace]}.
A lval \text{\sf\em L}\xspace can be an array access expression or a variable.
Let \text{$c \in \text{$\mathbb{C}$}\xspace$},
\text{$\text{$x$}\xspace,i \in (\text{\sffamily V$\!_s$}\xspace - \{\text{$a$}\xspace\})$}.
We consider assignment statements, conditional statement, loop statement, and assertion statements defined by the following grammar.
We define the grammar of our language using the following non-terminals:
Program \text{\sf\em P}\xspace consists of statements \text{\sf\em S}\xspace which may use
lvalues \text{\sf\em L}\xspace and expressions \text{\sf\em E}\xspace. We assume that programs are type correct as per C typing rules.
\begin{equation}
\renewcommand{\arraystretch}{1.2}
\begin{array}{rl}
\text{\sf\em P}\xspace \rightarrow & \; \,
\text{\sf\em S}\xspace
\\
\text{\sf\em S}\xspace \rightarrow & \; \,
\text{\bf if\/}\xspace\; (\text{\sf\em E}\xspace) \; \text{\sf\em S}\xspace \; \text{\bf else\/}\xspace \; \text{\sf\em S}\xspace
\;\; \big\lvert \;\; \text{\bf if\/}\xspace\; (\text{\sf\em E}\xspace) \; \text{\sf\em S}\xspace
\;\; \big\lvert \;\; \text{\bf for\/}\xspace \; (i = \text{\sf\em E}\xspace \,;\; \text{\sf\em E}\xspace \,;\; \text{\sf\em E}\xspace) \;\text{\sf\em S}\xspace
\;\; \big\lvert
\\
&
\;\; \text{\sf\em S}\xspace \; ; \; \text{\sf\em S}\xspace
\;\; \big\lvert \;\; \text{\sf\em L}\xspace = \text{\sf\em E}\xspace
\;\; \big\lvert \;\; \text{\bf assert\/}\xspace(\text{\sf\em E}\xspace)
\\
\text{\sf\em L}\xspace \rightarrow & \;\,
a[\text{\sf\em E}\xspace]
\;\; \big\lvert \;\; \text{$x$}\xspace
\\
\label{eq:grammar.org}
\text{\sf\em E}\xspace \rightarrow & \; \,
\text{\sf\em E}\xspace \oplus \text{\sf\em E}\xspace
\;\; \big\lvert \;\; \text{\sf\em L}\xspace
\;\; \big\lvert \;\; c
\end{array}
\end{equation}
In practice, we analyze \textsc{Ansi}-C language programs that includes functions, pointers, composite data-structures,
all kinds of definitions, and all control structures except multi-dimensional arrays.
\subsection{Representing Program States}
We define program states in terms of memory location and the value stored in the memory location.
We distinguish between \emph{atomic} variables (such as
scalar and structure variables) whose values can be copied atomically to a memory location, from
non-atomic variables such as arrays. Since we are considering 1-dimensional arrays, the
array elements are atomic locations.
Function \text{$\text{$\ell$}\xspace(a[i])$} returns the memory location corresponding to the $i^{th}$ index of array $a$.
The memory of an input program consists of all atomic locations:
\begin{align}
\text{$\mathbb{M}$}\xspace & = (\text{\sffamily V$\!_s$}\xspace - \{\text{$a$}\xspace\}) \cup \big\{ \text{$\ell$}\xspace(\text{$a$}\xspace[i]) \; \big\lvert \; 0 \leq i \leq \text{\sf\em lastof}\xspace(\text{$a$}\xspace) \big\}
\label{equation:mem.org}
\end{align}
The function \text{\sf\em lastof}\xspace(\text{$a$}\xspace) returns the highest index value for array \text{$a$}\xspace.
A \emph{program state} is a map $\text{$\sigma$}\xspace : \text{$\mathbb{M}$}\xspace \to \text{$\mathbb{C}$}\xspace$.
\text{\eval{e}{\ms}}\xspace denotes the value of expression \text{$e$}\xspace in the program state \text{$\sigma$}\xspace.
We transform a program by creating a pair \text{$\langle \text{$i_a$}\xspace,\text{$x_a$}\xspace \rangle$} for the array \text{$a$}\xspace where \text{$i_a$}\xspace is the witness index
and \text{$x_a$}\xspace is the witness variable.
The memory of a transformed program with additional variables is:
\begin{align}
\text{$\mathbb{M'}$}\xspace & = (\text{\sffamily V$\!_s$}\xspace - \{\text{$a$}\xspace\}) \cup \{\text{$x_a$}\xspace\} \cup \; \{\text{$i_a$}\xspace\}
\label{equation:mem.trans}
\end{align}
For a transformed program, a program state is denoted by \text{$\sigma'$}\xspace and is defined over \text{$\mathbb{M'}$}\xspace.
We explain the relation between states in original and transformed programs using an example.
Let a program \text{$P$}\xspace have an array variable $a$ and variable $k$ holding the size of the array $a$. Let
the array contain the values \text{$c_i \in \text{$\mathbb{C}$}\xspace$}, \text{$0 \leq i < n$}, where $n \in \text{$\mathbb{C}$}\xspace$ is the value of size of the array.
Then, a program state, $\text{$\sigma$}\xspace$ at any program point \emph{l} can be:
\begin{align}
\text{$\sigma$}\xspace &=
\big\{ \big(k,n\big),
\big(\text{$\ell$}\xspace \left(a[0]\right),c_0\big),
\big(\text{$\ell$}\xspace \left(a[1]\right),c_1\big),
\ldots,
\big(\text{$\ell$}\xspace \left(a[{n-1}]\right),c_{n-1} \big)
\big\}
\label{eq:state.p}
\end{align}
In the transformed program \text{$\mathbb{A}(P,A_n)$}\xspace, let \text{$x_a$}\xspace and \text{$i_a$}\xspace be the witness variable and witness index respectively.
Let $l'$ be the program point in \text{$\mathbb{A}(P,A_n)$}\xspace that corresponds to $l$ in \text{$P$}\xspace.
Then, all possible states in the transformed program at $l'$ are,
\begin{align}
\label{eq:set.state.pa}
\text{$\sigma'$}\xspace_{0} & = \{(k,n),(i_a,0),(x_a,c_{0})\}
\nonumber
\\
\text{$\sigma'$}\xspace_{1} & = \{(k,n),(i_a,1),(x_a,c_{1})\}
\nonumber
\\
\ldots
\nonumber
\\
\text{$\sigma'$}\xspace_{{n-1}} & = \{(k,n),(i_a,{n-1}),(x_a,c_{{n-1}})\}
\nonumber
\end{align}
We now formally define how a state at a program point in the transformed program represents a state at the
corresponding program point in the original program.
\begin{definition}
\label{def:abselement1}
Let \text{$\sigma$}\xspace be a state
at a program point in \text{$P$}\xspace and let \text{$\sigma'$}\xspace be a state at the corresponding program point in \text{$\mathbb{A}(P,A_n)$}\xspace.
Then, \text{$\sigma'$}\xspace represents \text{$\sigma$}\xspace, denoted as \text{$\text{$\sigma'$}\xspace \rightsquigarrow \text{$\sigma$}\xspace$} if
\[
\text{$\sigma'$}\xspace = \left\{
(\text{$i_a$}\xspace,c_1)
(\text{$x_a$}\xspace,c_2)
\right\}
\;\cup\;
\left\{
(y,c) \mid (y,c) \in \text{$\sigma$}\xspace,
y \in (\text{\sffamily V$\!_s$}\xspace-\{\text{$a$}\xspace\})
\right\}
\Rightarrow \; \big(\text{$\ell$}\xspace(a[c_1]),c_2\big) \in \text{$\sigma$}\xspace \;
\]
\end{definition}
Let \text{$A_n$}\xspace be the assertion at line n in program P. Let \text{$\sigma$}\xspace be a state reaching \text{$A_n$}\xspace in the original program with pair
\text{$\left(\text{$\ell$}\xspace\left(a\left[\eval{e_1}{\text{$\sigma$}\xspace}\right]\right),\eval{e_2}{\text{$\sigma$}\xspace}\right)$}.
Let \text{$\sigma'$}\xspace be the state in
transformed program, \text{$\sigma'$}\xspace represents \text{$\sigma$}\xspace. Thus, \text{$\sigma'$}\xspace has two pairs,
\text{$\left(\text{$i_a$}\xspace, \eval{e_3}{\text{$\sigma'$}\xspace}\right)$} and
\text{$\left(\text{$x_a$}\xspace,\eval{e_4}{\text{$\sigma'$}\xspace}\right)$}
such that
\text{$\eval{e_3}{\text{$\sigma'$}\xspace} = \eval{e_1}{\text{$\sigma$}\xspace}$} and \text{$\eval{e_4}{\text{$\sigma'$}\xspace} = \eval{e_2}{\text{$\sigma$}\xspace}$} .
Hence, if the assertion \text{$A_n$}\xspace holds in transformed program it holds in the original program too.
\section{Precision}
\label{sec::prec}
We characterize the assertions for which our transformation is precise -- an assertion will fail in
\text{$\mathbb{A}(P,A_n)$}\xspace if and only if it does so in \text{$P$}\xspace.
We denote such an assertion as \text{$A\!_n^{inv}$}\xspace. We focus on \text{$A\!_n^{inv}$}\xspace in a loop.
A program can have array accesses outside loops too. In such cases we do not claim precision; as per our experience such situations
are rare in programs with large-sized arrays.
Our transformations
replace array access expressions and loop statements while the statements involving
scalars alone outside the loop remain unmodified. Hence precision criteria need to focus on the statements within loops and not outside it.
Let assertion \text{$A\!_n^{inv}$}\xspace be in loop statement $S_\text{$A\!_n^{inv}$}\xspace$.
Let \text{\sf\em$\mathbb{V}_\text{\em imp}$}\xspace be the set of variables and \text{\sf\em$\mathbb{E}_\text{\em imp}$}\xspace be the set of array access expressions on which \text{$A\!_n^{inv}$}\xspace is
data or control dependent within the loop $S_\text{$A\!_n^{inv}$}\xspace$.
Let the set of loop statements from where definitions reach \text{$A\!_n^{inv}$}\xspace be denoted by \text{\sf\em$\mathbb{S}_\text{\em def}$}\xspace, note that this
set is a transitive closure for data dependence.
Our technique is precise when:
\begin{itemize}
\item $\text{\sf\em fullarrayaccess}\xspace(S)$ holds for each \text{$S \in \{ S_\text{$A\!_n^{inv}$}\xspace\} \cup \text{\sf\em$\mathbb{S}_\text{\em def}$}\xspace$} \hfill(rule $l_1$)
\item If $a[e] \in \text{\sf\em$\mathbb{E}_\text{\em imp}$}\xspace$ then
\begin{itemize}
\item the index expression $e = i$ where $i$ is the loop iterator of loop $S_\text{$A\!_n^{inv}$}\xspace$ \hfill
(rule $a_2$)
\item $\text{$a$}\xspace \notin \text{\sf\em loopdefs}\xspace(S)$ where \text{$S \in \{ S_\text{$A\!_n^{inv}$}\xspace\} \cup \text{\sf\em$\mathbb{S}_\text{\em def}$}\xspace$} \hfill (rule $a_3$)
\end{itemize}
\item If $\text{$x$}\xspace \in \text{\sf\em$\mathbb{V}_\text{\em imp}$}\xspace$ then
$\text{$x$}\xspace \notin \text{\sf\em loopdefs}\xspace(S)$ where \text{$S \in \{ S_\text{$A\!_n^{inv}$}\xspace\} \cup \text{\sf\em$\mathbb{S}_\text{\em def}$}\xspace$} \hfill (rule $s_4$)
\item For an assignment statement of the form $a[e_1] = e_2$ in loop $S$ where $S \in \text{\sf\em$\mathbb{S}_\text{\em def}$}\xspace$,
\begin{itemize}
\item if $e_2$ is an array access expression then it must be of the form $a[i]$ where $i$ is the loop iterator
of loop $S$ \hfill (rule $d_5$)
\item if $e_2$ is \text{$x$}\xspace then $\text{$x$}\xspace \notin \text{\sf\em loopdefs}\xspace(S)$ where \text{$S \in \text{\sf\em$\mathbb{S}_\text{\em def}$}\xspace$} \hfill
(rule $d_6$)
\end{itemize}
\end{itemize}
\begin{theorem}
\label{th:comp}
If the assertion \text{$A\!_n^{inv}$}\xspace; that satisfies above rules; holds in the original program \text{$P$}\xspace, then it will hold in the
transformed program \text{$\mathbb{A}(P,A_n)$}\xspace also.
\end{theorem}
\begin{proof}
The transformed program is over-approximative because our transformation rules (~\ref{transform.stmt.3},~\ref{transform.stmt.4},~\ref{transform.expr.2})
introduce non-deterministic values.
We prove this theorem by showing that if assertion is of the form \text{$A\!_n^{inv}$}\xspace then
none of these transformation rules introduce non-deterministic values in the transformed program.
\begin{itemize}
\item Since rule $l_1$ holds unconditionally, case~\ref{transform.stmt.4} will not apply.
Hence no extra paths are added in transformed program.
Also, since case~\ref{transform.stmt.3} applies, assignment $i=\text{$i_a$}\xspace$ will be added for
$S_\text{$A\!_n^{inv}$}\xspace$ and the loop statements in \text{\sf\em$\mathbb{S}_\text{\em def}$}\xspace.
\item When rule $a_2$ holds, since rule $l_1$ holds $a[e]$ get replaced by \text{$x_a$}\xspace always (case~\ref{transform.expr.2}).
\item When rule $a_3$ holds, assignment $\text{$x_a$}\xspace=nd()$ is not added (case~\ref{transform.stmt.3}).
\item When rule $s_4$ holds, assignment $\text{$x$}\xspace=nd()$ is not added (case~\ref{transform.stmt.3}).
\item When rule $d_5$ holds, since rule $l_1$ holds $a[e]$ in RHS gets replaced with \text{$x_a$}\xspace (case~\ref{transform.expr.2}).
\item When rule $d_6$ holds, scalars in RHS are not assigned with a non-deterministic value.
\end{itemize}
\end{proof}
Note that rule $s_4$ is a very strong condition to ensure that non-deterministic values do not reach \text{$A\!_n^{inv}$}\xspace.
We can relax this rule
when $\text{$x$}\xspace \in \text{\sf\em loopdefs}\xspace(S_\text{$A\!_n^{inv}$}\xspace)$ under these two conditions:
\begin{itemize}
\item definition of \text{$x$}\xspace appears before the assert statement in the loop
\item \text{$x$}\xspace is defined with a constant or using loop iterator $i$ only.
\end{itemize}
None of the transformation rules replace variable \text{$x$}\xspace. Definition of \text{$x$}\xspace to non-deterministic value ($\text{$x$}\xspace = nd()$)
gets re-defined by original assignment (retained in the transformed loop body) appearing before the assert statement.
Since \text{$x$}\xspace is defined with a constant or $i$ ($i=\text{$i_a$}\xspace$ is added for $S_\text{$A\!_n^{inv}$}\xspace$),
its value is not over-approximated.
\section{Conclusions and Future Work}
\label{conclusion}
Verification of programs with loops iterating over arrays is a challenging problem because of large sizes of arrays.
We have explored a middle ground between the two
extremes of relying completely on dynamic approaches of using model checkers on the one hand and using completely
static analysis involving complex domains and fix point computations on the other hand. Our experience shows that using static analysis
to transform the program and letting the model checkers do the rest is a sweet spot that enables verification of
properties of arrays using an automatic technique that is generic, sound, scalable, and reasonably precise.
Our experiments show that the effectiveness of our technique depends on the characteristics of programs and
properties sought to be verified.
We are able to eliminate 40-90\% of false warnings from
diverse applications.
This is a significant value addition to static analysis that try to find
defects and end up generating a large number of warnings which need to be resolved manually for safety critical applications.
Our effort grew out of our own experience of such manual reviews which showed
a large number of warnings to be false positives.
We plan to make our technique more precise by augmenting it with a refinement step to verify the programs that are
reported as unsafe by our current technique.
Finally, we wish to extend our technique on other data structures such as maps or lists.
\section{Informal Description}
\label{desc}
Given a program \text{$P$}\xspace containing loops iterating over an array \text{$a$}\xspace, we transform it
to a program \text{$\mathbb{A}(P,A_n)$}\xspace that has a pair \text{$\langle \text{$x_a$}\xspace, \text{$i_a$}\xspace \rangle$} of a
\emph{witness variable} and a \emph{witness index} for the array and the index
such that
$\text{$x_a$}\xspace$ represents the element $a[i_a]$ of the original program. Further, loops are replaced by
their customized bodies that operate only on $\text{$x_a$}\xspace$ instead of all elements of $a$.
To understand the intuition behind our transformation, consider a trace $t$ of \text{$P$}\xspace ending on the assertion
\text{$A_n$}\xspace. Consider the last occurrence of a statement \text{$s: a[e_1]=e_2$} in $t$.
We wish to transform \text{$P$}\xspace such that
there exists a trace $t'$ of \text{$\mathbb{A}(P,A_n)$}\xspace ending on \text{$A_n$}\xspace with value of \text{$i_a$}\xspace equal to that of $e_1$
and value of \text{$x_a$}\xspace equal to that of $e_2$.
We achieve this by transforming the program such that:
\begin{itemize}
\item \text{$i_a$}\xspace gets a non-deterministic value at the start of the program (this facilitates arbitrary choice of array element $a[\text{$i_a$}\xspace]$).
\item array writes and reads for $a[\text{$i_a$}\xspace]$ gets replaced with witness variable \text{$x_a$}\xspace.
\item array writes other than $a[\text{$i_a$}\xspace]$ gets eliminated and reads gets replaced with non-deterministic value.
\item loop body is executed only once either non-deterministically or unconditionally based on loop characteristics. During the
execution of the loop body,
\begin{itemize}
\item the loop iterator variable gets the value of \text{$i_a$}\xspace or a non-deterministic value (depending on loop characteristics), and
\item all other scalar variables whose values may be different in different iterations
gets non-deterministic values.
\end{itemize}
\end{itemize}
Figure~\ref{TEx1} shows the transformed program \text{$\mathbb{A}(P,A_n)$}\xspace for the program \text{$P$}\xspace of Figure~\ref{Mex1}.
Function \texttt{nd(l,u)} returns a non-deterministic value in the range $[l..u]$.
In \text{$\mathbb{A}(P,A_n)$}\xspace, the witness index $i\_a$ for array $a$ is globally assigned a
non-deterministic value within the range of array size (at line 9).
In a run of BMC, the assertion is checked for this non-deterministically chosen element $a[\text{$i_a$}\xspace]$.
To ensure that values for the same index $a[i_a]$ are written and read, we replace
array accesses by the witness variable \emph{\text{$x\_a$}} only when the value of index $i$ matches with $i\_a$ (lines 13, 14 and 17).
We remove loop header but retain loop body.
To over-approximate the effect of removal of loop iterations we add non-deterministic assignments to all variables modified in the loop body,
at the start of the transformed loop body and also
after the transformed loop body (lines 11 and 15).
Note that we retain the original assignment statements too (line 12).
Since the loops at line 8 and line 14 in the original program iterate over the entire array, we equate loop iterator variable $i$ to
$i\_a$ (line 11 and 16) and the transformed loop bodies
(lines 10--14 and lines 16--17) are executed unconditionally.
We explain the transformation rules formally in Section~\ref{transformation}.
The transformed program can be verified by an off-the-shelf BMC.
Note that each index will be considered in some run of the BMC since $i\_a$ is chosen non-deterministically.
Hence, if an assertion fails for any index in the original program, it fails in the transformed program too.
\section{Experimental Evaluation}
\label{exp}
We have implemented our transformation engine using static
analysis\footnote{PRISM, a static analyzer generator developed at TRDDC, Pune
~\cite{chimdyalwar2011effective,khare2011static}}. It supports
\textsc{Ansi}-C programs with 1-dimensional arrays. The experiments are
performed on a 64-bit Linux machine with 16 Intel Xeon processors running at
2.4GHz, and 20GB of RAM. More details of optimization and implementation, including handling of multiple arrays,
are provided in our
technical report~\cite{Jana2016Scaling}.
Our transformation engine outputs C programs. Although we could take
any off-the-shelf BMC for C program to verify the transformed code, we use CBMC
in our experiments as it is known to handle all the constructs of
\textsc{Ansi}-C. We discuss the results of our experiments on academic
benchmarks and industry codes. For want of space, we omit the results of
various BMCs on patterns from industry code; those results are shared in our
technical report~\cite{Jana2016Scaling}.
\subsection{Experiment 1 : SV-COMP Benchmarks} SV-COMP
benchmarks~\cite{svcompbench} contain an established set of programs under
various categories intended for comparing software verifiers.
Results for \emph{ArraysReach}\footnote{Programs in \emph{ArrayMemSafety}
access arrays without using index and cannot be transformed.} from the
\emph{array} category
for CBMC used in SV-COMP 2016 ($CBMC_\alpha$), CBMC 5.4 ($CBMC_\beta$) and CBMC
5.4 on transformed programs (Transformation+$CBMC_\beta$) are
consolidated\footnote{Case by case results available at \\
\htmladdnormallink{https://sites.google.com/site/datastructureabstraction/home/sv-comp-benchmark-evaluation-1}{https://sites.google.com/site/datastructureabstraction/home/sv-comp-benchmark-evaluation-1}}
in Table~\ref{svcomp_table}.
\emph{ArraysReach} has 118 programs. $CBMC_\alpha$, an unsound version of CBMC,
gave correct results for 53 programs.
However, $CBMC_\beta$ gave correct results for 14 programs.
We compare the results of Transformation+$CBMC_\beta$ on three criteria:
\begin{compactitem}
\item Scalability: it scaled up for all 118 programs.
\item Soundness: it gave sound results for all 118 programs. For the 6
program for which $CBMC_\alpha$ gave unsound results, our results are not only
sound but are also precise.
\item Precision: it gave precise results for 59 programs. Out of these
$CBMC_\alpha$ ran out of memory for 45 programs ($CBMC_\alpha$ ran out of
memory for 14 additional programs). On the other hand, 22 true
programs reported correctly by $CBMC_\alpha$ were verified as false by
Transformation+$CBMC_\beta$. Transformation+$CBMC_\beta$ verified 25 program
as true which did not include 8 of programs reported correctly as true by
$CBMC_\beta$.
\end{compactitem}
Our technique is imprecise for the other 59 of 118 programs as they do not comply with
the characterization of precision provided in Section~\ref{sec::prec}. As can
be seen, there is a trade-off between scalability and precision. From the view
point of reliability of results, soundness is the most desirable property of a
verifier. Our technique satisfies this requirement. Further, it not only scales up but is also precise implying its
practical usefulness.
\subsection{Experiment 2 : Real-life Applications}
\begin{table}[t]
\centering
\caption{Real-life Application Evaluation}
\label{RealApp_table}
\begin{tabular}{ccccccccccc}
\hline
\multicolumn{4}{|c|}{Application details} & \multicolumn{3}{c|}{Sliced+CBMC} & \multicolumn{3}{c|}{\begin{tabular}[c]{@{}c@{}}Sliced\\ +Transformation \\ +CBMC\end{tabular}} & \multicolumn{1}{c|}{\multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}\%\\ False\\ Positive \\ Reduction\end{tabular}}} \\ \cline{1-10}
\multicolumn{1}{|c|}{Name} & \multicolumn{1}{c|}{\begin{tabular}[c]{@{}c@{}}Size\\ (LoC)\end{tabular}} & \multicolumn{1}{c|}{ $\% loop^{full}$} & \multicolumn{1}{c|}{\#Asserts} & \multicolumn{1}{c|}{\#P} & \multicolumn{1}{c|}{\#F} & \multicolumn{1}{c|}{\#T} & \multicolumn{1}{c|}{\#P} & \multicolumn{1}{c|}{\#F} & \multicolumn{1}{c|}{\#T} & \multicolumn{1}{c|}{} \\ \hline \hline
\multicolumn{1}{|c|}{\it{navi1}} & \multicolumn{1}{c|}{1.54M} & \multicolumn{1}{c|}{100} & \multicolumn{1}{c|}{63} & \multicolumn{1}{c|}{0} & \multicolumn{1}{c|}{0} & \multicolumn{1}{c|}{63} & \multicolumn{1}{c|}{52} & \multicolumn{1}{c|}{1} & \multicolumn{1}{c|}{10} & \multicolumn{1}{c|}{82.5} \\ \hline
\multicolumn{1}{|c|}{\it{navi2}} & \multicolumn{1}{c|}{3.3M} & \multicolumn{1}{c|}{93.4} & \multicolumn{1}{c|}{103} & \multicolumn{1}{c|}{0} & \multicolumn{1}{c|}{0} & \multicolumn{1}{c|}{103} & \multicolumn{1}{c|}{95} & \multicolumn{1}{c|}{1} & \multicolumn{1}{c|}{7} & \multicolumn{1}{c|}{92.2} \\ \hline
\multicolumn{1}{|c|}{icecast\_2.3.1} & \multicolumn{1}{c|}{336K} & \multicolumn{1}{c|}{59.1} & \multicolumn{1}{c|}{114} & \multicolumn{1}{c|}{0} & \multicolumn{1}{c|}{0} & \multicolumn{1}{c|}{114} & \multicolumn{1}{c|}{53} & \multicolumn{1}{c|}{61} & \multicolumn{1}{c|}{0} & \multicolumn{1}{c|}{46.5} \\ \hline \hline
\multicolumn{11}{|c|}{\begin{tabular}[c]{@{}c@{}} $loop^{full}$ - loop \text{\sf\em S}\xspace where \text{\sf\em fullarrayaccess}\xspace(\text{\sf\em S}\xspace) holds, \\ P - Assertion Proved, F - Assertion Failed, T - Timeout\end{tabular}} \\ \hline
\end{tabular}
\end{table}
We applied our technique on 3 real-life applications -
{\it navi1} and {\it navi2} are industry codes implementing the navigation
system of an automobile and icecast\_2.3.1 is an open source project for
streaming media~\cite{icecast}. We appended assertions using \emph{null
pointer dereference} (NPD) warnings from a sound static analysis\footnote{TCS
Embedded Code Analyzer (TCS ECA)\\
\htmladdnormallink{http://www.tcs.com/offerings/engineering\_services/Pages/TCS-Embedded-Code-Analyzer.aspx}{http://www.tcs.com/offerings/engineering\_services/Pages/TCS-Embedded-Code-Analyzer.aspx}}
tool as follows. Lets say the dereference expression is $*a[i].p$. A
statement $assert(a[i].p!=null)$ is added in the code just before statement
containing dereference expression.
We ran CBMC on these applications with a time out of 30 minutes. CBMC did not
scale on the original as well as the sliced programs. We ran our transformation engine on
sliced programs. Table~\ref{RealApp_table} shows the consolidated results of
our experiments. Out of 280 assertions, sliced+transformation+CBMC proved 200
assertions taking 12 minutes on average for transformation+verification. This
is a much less in comparison to the time given to CBMC for sliced programs
(sliced+CBMC), which was 30 minutes.
To verify the correctness of our implementation, we analyzed the warnings
manually. We found that all 280 warnings were false, implying that all the
assertions should have been proved successfully.
\begin{compactitem}
\item CBMC could scale up for such large applications because there are no loops in transformed programs.
However, CBMC could not scale for 17 cases even after transformation because of the presence of a long recursive call chain of calls through function
pointers.
\item CBMC could not prove 63 of
the assertions since array definitions reaching at the assertion were from the
loops where $\text{\sf\em fullarrayaccess}\xspace(\text{\sf\em S}\xspace)$ did not hold.
Hence the witness variable takes over-approximated values.
\item CBMC proved 200 assertions, where
all the conditions for precision mentioned in Section~\ref{sec::prec} get fulfilled. In these experiments, we checked for the NPD property which is
value-independent. Moreover, we found that the assertions inserted by us are
not control-dependent on any scalar.
\end{compactitem}
Note that the number of false warnings
eliminated in an application is proportional to the number of loops for which
\text{\sf\em fullarrayaccess}\xspace(\text{\sf\em S}\xspace) hold.
Over a diverse set of applications, we found that our technique
could eliminate 40-90\% of false warnings. This is a significant value addition
to static analysis tools that try to find defects and end up generating a large
number of warnings. In fact, our own effort grew out of the need of handling
warnings that were generated by our proprietary static analysis tool, a large
fraction of which were false positives.
\section{Introduction}
\label{introduction}
Bounded Model Checking is one of the most successful techniques for
finding bugs~\cite{copty2001benefits} as evidenced by success
achieved by tools implementing this technique in verification
competitions~\cite{SVCOMP16,SVCOMP15}. Given a program P
and a property $\varphi$, Bounded Model Checkers (BMCs) unroll the loops in P a fixed
number of times and search for violations to $\varphi$ in the unrolled program.
However, for programs with loops of large or unknown bounds, bounded model
checking instances often exceed the limits of resources available. In our
experience, programs manipulating large-sized arrays invariably have such loops
iterating over indices of the array. Consequently, BMCs
routinely face the issue of scalability in proving properties on arrays. The
situation is not different even when the property is an {\it array invariant}
i.e., it holds for every element of the array, a characteristic which can
potentially be exploited for efficient bounded model checking.
Consider the example in Figure~\ref{Mex1} manipulating an array of structures
$a$. The structure has two fields, $p$ and $q$, whose values are assigned in
the first {\it for} loop (lines 8--13) such that $a[i].q$ is the square of
$a[i].p$ for every index $i$. The second {\it for} loop (lines 14--17) asserts
that this property indeed holds for each element in $a$. This is a {\it
safe} program i.e., none of the assertions admit a counterexample.
CBMC~\cite{cbmc}, a bounded model checker for C, in an attempt to unwind first
loop 100000 times, runs out of memory before it even reaches
the loop with assertion. In fact, we tried this example with several other
model checkers\footnote{Result for \emph{motivatingExample.c} at \htmladdnormallink{https://sites.google.com/site/datastructureabstraction/}{https://sites.google.com/site/datastructureabstraction/}}
and none of them
were able to prove this property because of large loop bounds.
\begin{figure*}[t]
\scriptsize
\begin{minipage}{0.45\textwidth}
\begin{lstlisting}[frame=single,language=c,escapeinside={(*@}{@*)},basicstyle=\ttfamily]
1. struct S {
2. unsigned int p;
3. unsigned int q;
4. } a[100000];
5. int i,k;
06. main()
07. {
08. for(i=0; i<100000; i++)
09. {
10. k = i;
11. a[i].p = k;
12. a[i].q = k * k ;
13. }
14. for (i=0; i<100000; i++)
15. {
16. (*@ \color[rgb]{0.75,0.164,0.164}{assert}@*)(a[i].q ==
a[i].p * a[i].p);
17. }
18. }
\end{lstlisting}
\caption{Motivating Example}
\label{Mex1}
\end{minipage}
\quad
\quad
\begin{minipage}{0.50\textwidth}
\begin{lstlisting}[frame=single,language=c,escapeinside={(*@}{@*)},basicstyle=\ttfamily]
1. struct S{
2. unsigned int p;
3. unsigned int q;
4. }x_a;
5. int i_a;
6. int i,k;
7. main()
8. {
9. i_a = nd(0,99999);
//first loop body
10. k = nd(0,100000);
11. i = i_a;
12. k = i;
13. (i == i_a)? x_a.p = k : k;
14. (i == i_a)? x_a.q = k * k : k*k ;
15. k = nd(0,100000);
//second loop body
16. i = i_a;
17. (*@ \color[rgb]{0.75,0.164,0.164}{assert}@*)(((i==i_a)?x_a.q:nd())
==((i==i_a)?x_a.p:nd())
*((i==i_a)?x_a.p:nd()));
18. }
\end{lstlisting}
\caption{Transformed Code}
\label{TEx1}
\end{minipage}
\end{figure*}
One of the ways of proving this example safe is to show that the property holds
for any arbitrary element of the array, say at index $i_c$. This allows us to
get rid of those parts of the program that do not update $a[i_c]$ which, in
turn, eliminates the loop iterating over all the array indices. This enables
CBMC to verify the assertion without getting stuck in the loop unrollings.
Moreover, since $i_c$ is chosen nondeterministically from the indices of $a$,
the property holds for every array element without loss of generality.
This paper presents the transformation sketched above with the aim that the
transformed program is easier for a BMC to verify as compared
to the original program. The transformation
is over-approximative i.e., it give more values than that by the original program.
This ensures that if the original program is safe with respect to the chosen
property, so is the transformed program. However, the over-approximation
raises two important questions spanning practical and intellectual considerations:
\begin{enumerate}
\item {\it Is the proposed approach practically useful? Does the
transformation enable a BMC to verify real-world programs,
and even academic benchmarks, fairly often?}
We provide an answer to this through an extensive experimental evaluation over
industry code as well as
examples in the array category of SV-COMP
2016 benchmarks. In all the cases, we show that our approach helps CBMC to scale.
We further demonstrate the applicability of
our technique to successfully identify a large number of false warnings (on an average 73\%)
reported by a static analyzer on arrays in large programs.
\item {\it Is it possible to characterize a class of properties for which it is
precise?}
In order to address this we provide a formal characterization of properties for
which the transformation is precise i.e., we state criteria under which
the transformed program is unsafe only when the original program is unsafe
(Section~\ref{sec::prec}).
\end{enumerate}
To summarize, this paper makes the following contributions:
\begin{itemize}
\item A new technique using the concept of \emph{witness index} that enables BMCs to verify array invariant
properties in programs with loops iterating over large-sized arrays.
\item A formal characterization of properties for which the technique is
precise.
\item A transformation engine implementing the technique.
\item An extensive experimental evaluation showing the applicability of our
technique to real-world code as well as to academic benchmarks.
\end{itemize}
The rest of the paper starts with an informal description of the transformation
(Section~\ref{desc}) before we define the semantics
(Section~\ref{sec:dsa}) and formally state the transformations rules
(Section~\ref{transformation}). Section~\ref{sec:proof} and ~\ref{sec::prec},
resp., describe the soundness and precision of our approach. Section~\ref{exp}
presents the experimental setup and results. We discuss the related work in
Section~\ref{relwork} before concluding in Section~\ref{conclusion}.
\subsection*{References}
\Urlmuskip= 0mu plus 2mu\relax
\bibliographystyle{abbrv}
\section{Soundness}
\label{sec:proof}
\externaldocument{transformation}
This section outlines the claim that the proposed transformation is sound, i.e. if the transformed
program is safe, then so is the original program. As discussed in Section~\ref{sec:dsa}, the
soundness is immediate if the abstract states ``represent'' the
original states. We, therefore, prove that the proposed transformations ensure
that the \emph{represents} relation, $\rightsquigarrow$, holds between abstract and
original states.
For the base case, we prove that $\rightsquigarrow$ holds in the beginning -
before applying any transformation (Lemma~\ref{lemm:program.start}).
In the inductive step, we prove
that if $\rightsquigarrow$ holds at some stage during the transformation, then the subsequent
transformation continues to preserve $\rightsquigarrow$ (Lemma~\ref{lemma:strucinduc.abs}). We prove this by structural
induction on program transformations. We prove that each transformed expression is over-approximated when $\rightsquigarrow$ holds in (Lemma~\ref{lemm:rhs1}).
Detailed proof is provided in our technical report~\cite{Jana2016Scaling}.
\newcommand{\noindent\text{\em Proof Outline.\ }}{\noindent\text{\em Proof Outline.\ }}
\begin{lemma}
\label{lemm:program.start}
Let the start of the original program (i.e. the program point just before the code derivable from non-terminal \text{\sf\em S}\xspace in
production \text{$\text{\sf\em P}\xspace \rightarrow \text{\sf\em S}\xspace$} in grammar defined in equation(~\ref{eq:grammar.org}) be denoted by $l$.
The corresponding program point in the
transformed program \text{$\mathbb{A}(P,A_n)$}\xspace, denoted by $l'$, is just after \text{\sf\em I}\xspace and just before the non-terminal $S$ in production \text{$\text{\sf\em P}\xspace \rightarrow \text{\sf\em I}\xspace \;; \; \text{\sf\em S}\xspace$}
(Grammar in equation~\ref{eq:grammar.trans}).
Let \text{$\sigma$}\xspace and \text{$\sigma'$}\xspace be the states at $l$ and $l'$ in \text{$P$}\xspace and \text{$\mathbb{A}(P,A_n)$}\xspace respectively.
Then,
\text{$\text{$\sigma'\!\!_{l'}$}\xspace \rightsquigarrow\text{$\sigma_l$}\xspace$}.
\end{lemma}
\noindent\text{\em Proof Outline.\ }
Since the initial values of non array variables are preserved, the initial value of the element of array $a[\text{$i_a$}\xspace]$ is assigned to \text{$x_a$}\xspace, and
\text{$i_a$}\xspace is non-deterministically chosen, the lemma holds.
\begin{lemma}
\label{lemm:rhs1}
Let $\text{$\sigma_l$}\xspace$ be a state at a program point $l$ in \text{$P$}\xspace
and $\text{$\sigma'\!\!_{l'}$}\xspace$ be a state at the corresponding program point $l'$ in transformed program \text{$\mathbb{A}(P,A_n)$}\xspace.
Consider an arbitrary expression \text{$e \in \text{$\mathbbm{E}$}\xspace$} just after $l$ in original program \text{$P$}\xspace.
Then,
\[
\text{$\text{$\sigma'\!\!_{l'}$}\xspace \rightsquigarrow\text{$\sigma_l$}\xspace \Rightarrow \eval{\text{\sf\em transform}\xspace(e)}{\text{$\sigma'\!\!_{l'}$}\xspace} \supseteq \eval{e}{\text{$\sigma_l$}\xspace}$}.
\]
\end{lemma}
\noindent\text{\em Proof Outline.\ }
Since $e$ is derived from \text{\sf\em E}\xspace (grammar~\ref{eq:grammar.org}),
the over-approximation of values can be proved by structural induction on the productions for \text{\sf\em E}\xspace.
\begin{lemma}
Let $l$ and $m$ be the program points just before and after a statement $s$ in \text{$P$}\xspace and let $\text{$\sigma_l$}\xspace$ and $\text{$\sigma_m$}\xspace$ be
the states at $l$ and $m$ respectively.
Let $l'$ and $m'$ be the program points just before and after the corresponding transformed statement $\text{\sf\em transform}\xspace(s)$ in \text{$\mathbb{A}(P,A_n)$}\xspace.
Let \text{$\sigma'\!\!_{l'}$}\xspace and \text{$\sigma'\!\!_{m'}$}\xspace be the states at $l'$ and $m'$ respectively.
Then, \text{$\text{$\sigma'\!\!_{l'}$}\xspace \rightsquigarrow \text{$\sigma_l$}\xspace \Rightarrow \text{$\sigma'\!\!_{m'}$}\xspace \rightsquigarrow \text{$\sigma_m$}\xspace$}.
\label{lemma:strucinduc.abs}
\end{lemma}
\noindent\text{\em Proof Outline.\ }
Since statement $s$ is derived from non-terminal \text{\sf\em S}\xspace in the grammar~\ref{eq:grammar.org}
the lemma can be proved by structural induction on \text{\sf\em S}\xspace.
\begin{theorem}
If the assertion \text{$A_n$}\xspace is violated in the original program \text{$P$}\xspace, then it will be violated in transformed program \text{$\mathbb{A}(P,A_n)$}\xspace also.
\end{theorem}
\begin{proof}
Let the assert get violated for some $a[c]$. Since $\text{$i_a$}\xspace$ is initialized
non-deterministically it can take the value $c$ and we have shown in Lemma~\ref{lemm:rhs1} that
all expressions in $\text{$\mathbb{A}(P,A_n)$}\xspace$ are over-approximated. Lemma~\ref{lemm:program.start} and Lemma~\ref{lemma:strucinduc.abs} ensure the premise for Lemma~\ref{lemm:rhs1}.
Hence the theorem follows.
\end{proof}
\section{Related Work}
\label{relwork}
The literature on automated reasoning about array-manipulating code can be broadly categorized into \emph{analysis} and \emph{verification}.
Most methods that analyze programs manipulating arrays~\cite{blanchet2002design,gopan2005framework,halbwachs2008discovering,cousot2011parametric,liu2014abstraction}
are based on abstract interpretation.
Cornish et al.~\cite{cornish2014analyzing} transform a program to remove arrays and discover
non-trivial universally quantified loop invariants by analyzing
the transformed program using off-the-shelf abstract scalar analysis.
Since they create additional blocks for each value
of \emph{summary variable}, the program size increases considerably raising concerns about scalability.
Similar to our approach, Monniaux et al.~\cite{monniaux2015simple} transform array programs by replacing array operations with a scalar. However they keep loops. These programs are then analyzed using methods producing invariants (\emph{back-ends}).
CBMC did not scale up
on the transformed ''array copy''
example (10000 loop bound) given in the paper, suggesting that scalability is a concern with this technique too.
However, using our technique CBMC scaled for the same program.
Dillig et al.~\cite{dillig2010fluid} introduced fluid updates of arrays in order
to do away with strong and weak updates. Their technique uses indexed locations
along with bracketing constraints, a pair of over- and under-approximative
constraints, to specify the concrete elements being updated.
In another
work~\cite{dillig2011precise}, they propose an automatic technique
to reason about contents of arrays (or containers, in general). However, they
introduce an abstraction to encode all values that (a subset of) elements may
have. In contrast, since our technique choses only one representative element
to work with, we can capture its value precisely.
Template-base methods~\cite{beyer2007invariant,Gulwani2008Lifting} have been
very useful in synthesizing invariants but these techniques are ultimately
limited by a large space of possible templates that must be searched to get a
good candidate template. This has also led to semi-automatic approaches, such
as~\cite{flanagan2002predicate}, where the predicates are usually suggested by
the user. Our approach, however, is fully automatic and proves safety by
solving a bounded model checking instance instead of computing an invariant
explicitly.
Verification tools based on CEGAR have been applied successfully to
certain classes of programs, e.g., device drivers~\cite{ball2002s}. However,
this technique is orthogonal to ours. In fact, a refinement framework in
addition to our abstraction would make our technique complete. Several other
techniques have been used to scale BMCs to tackle complex, real-world
programs such as acceleration~\cite{kroening2013under} and
loop-abstraction~\cite{darke2015over}. But these techniques are not shown to
be beneficial in abstracting complex data structures.
Booster~\cite{alberti2014booster}, a recent tool for verifying C-like programs
handling arrays, integrates acceleration and lazy abstraction with interpolants
for arrays~\cite{alberti2014decision,alberti2012lazy}. It exploits
acceleration techniques to compute an exact set of reachable states, whenever
possible, for programs with arrays. For instance, their technique works on
\emph{$simple_{\mathcal{A}}^0$} programs~\cite{alberti2014decision}. However,
there are syntactic restrictions that limit the applicability of acceleration
in general for programs handling arrays. Note that Booster uses acceleration,
instead of abstraction-based procedures, for want of a precise solution (not
involving over-approximations). Since our technique is also precise for a
characterizable class of programs, it is certainly possible to gainfully
combine the two techniques in order to handle a larger class of programs than
what either of them can handle in isolation.
\section{Transformation}
\label{transformation}
The transformation rules are given in Figure~\ref{fig:transform}. A transformed program
satisfies the following grammar
derived from that of the original program (grammar \ref{eq:grammar.org}).
Let \text{$x,\text{$x_a$}\xspace,\text{$i_a$}\xspace \in \text{\sffamily V$\!_s$}\xspace$} denote scalar variable, witness variable, and witness index, respectively.
Let \text{$c,l,u \in \text{$\mathbb{C}$}\xspace$} be values. Then,
\begin{equation}
\renewcommand{\arraystretch}{1.2}
\begin{array}{rl}
\text{\sf\em P}\xspace \rightarrow & \; \,
\text{\sf\em I}\xspace \;; \; \text{\sf\em S}\xspace
\\
\text{\sf\em I}\xspace \rightarrow & \; \,
\text{$i_a$}\xspace = nd (l,u)
\\
\text{\sf\em S}\xspace \rightarrow & \; \,
\text{\bf if\/}\xspace\; (\text{\sf\em E}\xspace) \; \text{\sf\em S}\xspace \; \text{\bf else\/}\xspace \; \text{\sf\em S}\xspace
\;\; \big\lvert \;\; \text{\bf if\/}\xspace\; (\text{\sf\em E}\xspace) \; \text{\sf\em S}\xspace
\;\; \big\lvert \;\; \text{\sf\em S}\xspace \; ; \; \text{\sf\em S}\xspace
\;\; \big\lvert \;\;
\text{\sf\em L}\xspace = \text{\sf\em E}\xspace
\;\; \big\lvert \;\; \text{\bf assert\/}\xspace(\text{\sf\em E}\xspace)
\\
\text{\sf\em L}\xspace \rightarrow & \;\,
\text{$x$}\xspace \;\; \big\lvert \;\; \text{$x_a$}\xspace \;\; \big\lvert \;\; \text{$i_a$}\xspace
\\
\text{\sf\em E}\xspace \rightarrow & \; \,
\text{\sf\em E}\xspace \oplus \text{\sf\em E}\xspace
\;\; \big\lvert \;\; \text{\sf\em L}\xspace
\;\; \big\lvert \;\; c
\;\; \big\lvert \;\; nd()
\;\; \big\lvert \;\; nd(l,u)
\end{array}
\label{eq:grammar.trans}
\end{equation}
The non-terminal \text{\sf\em I}\xspace represents the initialization statements for witness index.
Witness variable is initialized in the scope same as that in the original program.
We use the functions described below in the transformation rules.
\begin{compactitem}
\item Function $nd$ returns a non-deterministically chosen value from the given range $l,u$; $l$ and $u$ being the lower and upper limit respectively. When range is not provided, $nd$
returns a non-deterministic value based on the type of \text{\sf\em L}\xspace.
\item Function \text{\em transform\/}\xspace takes the text derived from a non-terminal and transforms it. Function \text{\sf\em emit}\xspace shows the actual
code that would be emitted. We ignore the details of number of parameters and the type of the parameters of \text{\sf\em emit}\xspace.
We assume that it takes the code emitted by \text{\em transform\/}\xspace and possibly some additional statements and outputs the combined code.
It has been used only to distinguish the transformation time activity and run time activity.
For example, the boolean conditions in cases~\ref{transform.expr.2} and~\ref{transform.stmt.1} are not evaluated by the
body of function \text{\em transform\/}\xspace but is a part of the transformed code and is evaluated at run time when the transformed program is executed.
Similar remarks apply to the \text{\bf if\/}\xspace statements and other operations inside the parenthesis of \text{\sf\em emit}\xspace function.
\item Function $\text{\sf\em fullarrayaccess}\xspace(\text{\sf\em S}\xspace)$ analyzes\footnote{Analysis can be over-approximated.} the characteristics of the loop \text{\sf\em S}\xspace.
\begin{itemize}
\item When the loop \text{\sf\em S}\xspace accesses array \text{$a$}\xspace completely, $\text{\sf\em fullarrayaccess}\xspace(\text{\sf\em S}\xspace)$ returns true.
This means that loop either reads or write all the indices of the array.
\item When the loop \text{\sf\em S}\xspace accesses array \text{$a$}\xspace partially, $\text{\sf\em fullarrayaccess}\xspace(\text{\sf\em S}\xspace)$ returns false. This means
that the loop may not access all the indices or some indices are being read while some other indices are being written.
\item When loop \text{\sf\em S}\xspace do not access an array, $\text{\sf\em fullarrayaccess}\xspace(\text{\sf\em S}\xspace)$ returns false.
\end{itemize}
\item Function $\text{\sf\em loopdefs}\xspace(\text{\sf\em S}\xspace)$ returns the over-approximated set of variables modified in the loop \text{\sf\em S}\xspace.
\begin{compactitem}
\item Scalar variables are included in this set if they appear on the left hand side of any assignment statement in \text{\sf\em S}\xspace
(except when the RHS is a constant).
\item Loop iterator variable $i$ of loop \text{\sf\em S}\xspace is not included in this set.
\item
Array variable \text{$a$}\xspace
is included in this set when the array access expression appears on the left hand side of an assignment
and the value of index expression is different from the current value of the loop iterator $i$.
\end{compactitem}
\item Function $\text{\sf\em lastof}\xspace(a)$ returns the highest index value for array \text{$a$}\xspace.
\end{compactitem}
With the above functions, the transformation rules are easy to understand. Here we explain non-trivial transformations.
\begin{compactitem}
\item To choose an array index for a run, witness index (\text{$i_a$}\xspace) is initialized at the start of the program
to a non-deterministically chosen value from the range of the indices of the array (case~\ref{transform.prog.1}).
This value determines the array element ($a[\text{$i_a$}\xspace]$) represented by the witness variable (\text{$x_a$}\xspace).
\item An array access expression in LHS or RHS is replaced by the witness variable (\text{$x_a$}\xspace) provided the
values of the witness index and index expression of the array access expression match.
If the values do not match, it implies that the element accessed
is not at the non-deterministically chosen index \text{$i_a$}\xspace.
Hence for any other index the assignment does not happen (case~\ref{transform.stmt.1}).
Similarly, when any other index is read in RHS, it is
replaced with a non-deterministic value (case~\ref{transform.expr.2}).
\item Loop iterations are eliminated by removing the loop header containing initialization, test, and increment expression for
loop iterator variable. The loop bodies are transformed as follows :
\begin{compactitem}
\item Each variable in the set returned by $\text{\sf\em loopdefs}\xspace(\text{\sf\em S}\xspace)$ is assigned a non-deterministic value
at the start of the loop body and also after the loop body.
These assignments ensure that values dependent on loop iterations are
over-approximated when used inside or outside the loop body.
\item The loop iterator $i$ is a special scalar variable.
A loop \text{\sf\em S}\xspace where \text{\sf\em fullarrayaccess}\xspace(\text{\sf\em S}\xspace) holds (case~\ref{transform.stmt.3}) essentially means that loop bound is same as the array size
and array is accessed using loop iterator as index.
Hence it is safe to replace array access with \text{$x_a$}\xspace where the
values of loop iterator and index expression match.
To ensure this we equate loop iterator with \text{$i_a$}\xspace.
This models the behaviour of the original program precisely.
However, when $\text{\sf\em fullarrayaccess}\xspace(\text{\sf\em S}\xspace)$ does not hold (case~\ref{transform.stmt.4}), we assign loop iterator $i$ to a
non-deterministically chosen value from the loop bound.
\item Each statement in the loop body is transformed
as per the transformation rules.
\item Finally, the entire loop body is made conditional using a non-deterministically chosen true/false value
when $\text{\sf\em fullarrayaccess}\xspace(\text{\sf\em S}\xspace)$ does not hold.
This models the partial
accesses of array indices which imply that some of the values defined before the loop may reach after the loop.
However, the transformed loop body is unconditionally executed when $\text{\sf\em fullarrayaccess}\xspace(\text{\sf\em S}\xspace)$ holds.
\end{compactitem}
\end{compactitem}
|
\section{Introduction}
Theories with scale symmetry \cite{W} may provide a solution to the
hierarchy and cosmological constant problems.
But scale symmetry is not a symmetry of the real world, therefore it must be broken.
In this work we discuss theories with scale invariance at the classical
and quantum level that is broken {\it only spontaneously}.
This is important since in a classical scale invariant
theory, quantum calculations usually break this symmetry {\it explicitly} due to the presence
of the subtraction (renormalization) {\it scale} ($\mu$).
This scale is introduced to regularize the loop integrals,
regardless of the regularization method: dimensional regularization (DR),
Pauli-Villars, etc, and its simple presence breaks explicitly this symmetry.
It is known however how to avoid this problem by using a subtraction scale that
is generated {\it spontaneously}, as the vacuum expectation value (vev) of a scalar
field $\sigma$ \cite{Englert}. This field is the Goldstone mode of scale
symmetry (dilaton) and then $\mu=z \langle\sigma\rangle$, where $z$ is a dimensionless
parameter. But before (spontaneous) scale symmetry breaking, with a field-dependent
subtraction {\it function} $\mu(\sigma)= z \sigma$, there is no scale in the theory.
One can use this idea to compute quantum corrections to
the scalar potential of a theory with a higgs-like scalar $\phi$
and dilaton $\sigma$ and obtain a scale invariant result at one-loop
\cite{S1,S2,S3,dmg,P} with a flat direction and spontaneous scale symmetry breaking.
Although the result is scale invariant at the quantum level, the couplings still run
with the momentum scale \cite{S3,dmg,tamarit}\footnote{After
spontaneous breaking of scale symmetry
$\langle\sigma\rangle\not=0$, the subtraction scale $\mu(\langle\sigma\rangle)$
and all other masses/vev's of the theory are generated, proportional to $\langle\sigma\rangle$.}.
To illustrate some of these ideas, consider a scale invariant theory in $d=4$
\medskip
\begin{eqnarray}
L=
\frac{1}{2}\partial_\mu \phi\partial^\mu\phi
+\frac{1}{2}\partial_\mu\sigma\partial^\mu\sigma
-V(\phi,\sigma)
\end{eqnarray}
\medskip\noindent
where $\phi$ is a higgs-like scalar and $\sigma$ is a dilaton.
In such a theory $V$ has a form
\begin{eqnarray}\label{form}
V(\phi,\sigma)=\sigma^4\, W(\phi/\sigma)
\end{eqnarray}
In this paper we assume
that we have spontaneous breaking of this symmetry, hence $\langle\sigma\rangle\!\not=\!0$.
We do not detail how $\sigma$ acquires a vev (expected to be large
$\langle\sigma\rangle\sim M_\text{Planck}$)
but search for solutions with $\langle\sigma\rangle\!\not=\!0$.
Then the two minimum conditions $\partial V/\partial\phi\!=\!\partial V/\partial\sigma\!=\!0$
become
\begin{eqnarray}\label{W}
W'(x_0)=W(x_0)=0,
\qquad
x_0\equiv\frac{\langle\phi\rangle}{\langle\sigma\rangle};\,
\quad \langle\sigma\rangle, \langle\phi\rangle\not=0.
\end{eqnarray}
\medskip\noindent
At a given order $n$ in perturbation theory,
one condition, say $W^\prime(x_0)=0$,
fixes the ratio $x_0\equiv \langle\phi\rangle/\langle\sigma\rangle$ in terms of the (dimensionless)
couplings of the theory. The second condition, $W(x_0)=0$, leads to
vanishing vacuum energy $V(\langle\phi\rangle,\langle\sigma\rangle)=0$ and
fixes a relation among
the couplings, corrected to that order ($n$) in perturbation theory from its version in
the lower perturbation order ($n-1$). If these two equations have
a solution $x_0$, then the system has a flat direction (Goldstone) in the plane
$(\phi,\sigma)$ with $\phi/\sigma=x_0$. Then a
massless state exists (dilaton) at this order.
This is true {\it provided that} quantum corrections
do not break {\it explicitly} the scale symmetry (otherwise, eq.(\ref{form}) is not valid
due to the presence of the ``usual'' DR scale $\mu$). With a scale invariant
regularization, it is possible to keep these properties
($V=0$, a flat direction, etc) and study spontaneously broken quantum scale invariance.
Why is this interesting? One reason is that this answers the question of Bardeen
\cite{Bardeen} on the mass hierarchy.
The Standard Model (SM) with a vanishing classical higgs mass term is scale invariant
and there is no mass hierarchy (ignoring gravity, as here\footnote{For related applications
that include gravity, see for example \cite{FHR,S4,O,KA}.}).
If quantum calculations preserve this symmetry, via a scale invariant regularization,
one can avoid a hierarchy problem and the fine-tuning of the higgs self-coupling and
keep it light relative to the high scale (physical mass of a new state)
generated by $\langle\sigma\rangle\not=0$.
One can arrange that $x_0=\langle\phi\rangle/\langle\sigma\rangle\!\ll\! 1$ by
a {\it single} classical tuning of the (ratio of the) couplings of the
theory \cite{GGR}. The hierarchy
$m_\text{higgs}^2\sim \langle\phi\rangle^2\!\ll\! \langle\sigma\rangle^2$
is maintained at one-loop \cite{S1,S2,S3,dmg,P,GGR,FK} and probably
beyond it, due to the spontaneous-only scale symmetry breaking.
The only difference from the usual SM is the presence of a massless
dilaton in addition to the SM spectrum. Also, the solution $x_0$ is related
to the (minimum) condition $V=0$. This suggests that in spontaneously broken
quantum scale invariant theories any fine tuning is related to vacuum energy
tuning at the same order of perturbation.
With this motivation, in this paper we extend the above results.
We consider a classically scale invariant theory of
$\phi$ and $\sigma$ and compute at two-loop the scalar potential and the
running of the couplings, in a scale invariant regularization.
We find that starting from two loops, the running of the couplings differs from that in
the same theory of $\phi$, $\sigma$ regularised with $\mu\!=$constant.
We show that effective {\it non-polynomial} operators like
$\phi^6/\sigma^2$, $\phi^8/\sigma^4$, are generated as two-loop counterterms.
If expanded about the ground state, these operators
generate an infinite series of polynomial terms, showing the non-renormalizability
of the theory. The Callan-Symanzik equation of the potential is verified at two loops.
The results are useful for phenomenology, {\it e.g.} to study a scale invariant version of the
SM (+dilaton).
\section{One-loop potential}
We first review the one-loop potential \cite{dmg,P}. Consider the classical potential\footnote{
In principle one can also include higher dimensional
terms like $\phi^6/\sigma^2$, $\phi^8/\sigma^4$, etc, ($\langle\sigma\rangle\not=0$), see later.}
\begin{eqnarray}
V = \frac{\lambda_\phi}{4!} \phi^4 + \frac{\lambda_m}{4} \phi^2 \sigma^2
+ \frac{\lambda_\sigma}{4!} \sigma^4.
\end{eqnarray}
Spontaneous scale symmetry breaking $\langle\sigma\rangle\not=0$
demands two conditions (eq.(\ref{W})) be met:
\begin{eqnarray}\label{minimum}
9 \lambda_m^2=\lambda_\phi\lambda_\sigma+\textsf{loops},\,\,(\lambda_m<0),
\qquad\text{and}\qquad
x_0^2\equiv \frac{\langle\phi\rangle^2}{\langle\sigma\rangle^2}=
-\frac{3 \lambda_m}{\lambda_\phi}+\textsf{loops.}
\end{eqnarray}
A massless (Goldstone) state exists corresponding to a flat
direction $\phi=x_0\,\sigma$ with $V_\textsf{min}=0$.
With $\phi$ being higgs-like, scale symmetry breaking implies electroweak symmetry breaking.
To compute quantum corrections in $d=4-2\epsilon$, the scalar potential is modified
to $\tilde V=\mu^{2\epsilon}V$ to ensure dimensionless quartic couplings, with $\mu$ the
``usual'' DR subtraction scale. General
principles\footnote{They demand quantum interactions between $\phi$ and $\sigma$ vanish
in their classically decoupling limit $\lambda_m\!=\!0$.} suggest
that the subtraction {\it function} $\mu(\sigma)$ depend on the dilaton only \cite{dmg} and
generate the subtraction scale $\mu(\langle\sigma\rangle)$
after spontaneous scale symmetry breaking;
$\mu(\sigma)$ is then identified on dimensional grounds (using $[\mu]=1$, $[\sigma]=(d-2)/2$).
Then the scale invariant potential in $d=4-2\epsilon$ and $\mu(\sigma)$ become
\begin{eqnarray}\label{tildeV}
\tilde V(\phi,\sigma)=\mu(\sigma)^{2\epsilon} V(\phi,\sigma),
\qquad \mu(\sigma)=z\,\sigma^{1/(1-\epsilon)},
\end{eqnarray}
where $z$ is an arbitrary dimensionless parameter\footnote{The parameter $z$ plays a
special role in the Callan Symanzik equation, see later.}.
The one-loop result is
\begin{eqnarray}
V_1&=&\tilde V -
\frac{i}{2}\,
\,\int \frac{d^d p}{(2\pi)^d}
\,{\rm Tr}\ln \big[ p^2-\tilde V_{\alpha\beta}+i\varepsilon\big]
\end{eqnarray}
\medskip\noindent
Here $\tilde V_{ij}=\partial^2 \tilde V/\partial s_i \partial s_j$, \, ($i,j=1,2$),
$s_1=\phi$, $s_2=\sigma$ and similar for $V_{ij}=\partial^2 V/\partial s_i\partial s_j$.
Also $\tilde V_{ij}
=\mu^{2\epsilon}\,\big[V_{ij} + 2\epsilon \,\mu^{-2}\,N_{ij}\big]+\cO(\epsilon^2)$,
where
\medskip
\begin{eqnarray}
N_{ij}\equiv \mu\, \Big\{
\,\frac{\partial\mu}{\partial s_i}\, \frac{\partial V}{\partial s_j}
+\frac{\partial\mu}{\partial s_j}\, \frac{\partial V}{\partial s_i}\Big\}
+\Big\{\mu\, \frac{\partial^2 \mu}{\partial s_i\partial s_j}
-\frac{\partial \mu}{\partial s_i}\frac{\partial \mu}{\partial s_j}
\,\Big\}\,V,\qquad i,j=1,2.
\label{m2prim}
\end{eqnarray}
Then
\begin{eqnarray}\label{lt}
V_1
=\mu(\sigma)^{2\epsilon}
\Big\{
V- \frac{1}{64 \pi^2}\,
\Big[
\sum_{s=\phi,\sigma}
M^4_s \, \Big(\, \frac{1}{\epsilon}
- \ln\frac{M_s^2(\phi,\sigma)}{c_0\,\mu^2(\sigma)} \Big)
+
\frac{4 \,(V_{ij}\,N_{ji})}{\mu^2(\sigma)}\Big]\Big\}
\end{eqnarray}
\medskip\noindent
with an implicit sum over $i,j$ and with $c_0=4\pi e^{3/2-\gamma_E}$. The one-loop Lagrangian is
\medskip
\begin{eqnarray}\label{l1}
L_1=\frac{1}{2} (\partial_\mu\phi)^2+\frac{1}{2}(\partial_\mu\sigma)^2-V_1.
\end{eqnarray}
\medskip\noindent
Above, $M_s^2$ denotes
the field-dependent eigenvalues of the matrix $V_{ij}$. The poles in $L_1$ are cancelled
by adding the counterterm Lagrangian $\delta L_1$ found using the expression of the $M_s^2$:
\medskip
\begin{eqnarray}\label{ZV}
\delta L_1\equiv -\delta V_1=
- \mu(\sigma)^{2\epsilon}
\Big\{ \frac{1}{4!} (Z_{\lambda_\phi}\!-1) \lambda_\phi\phi^4+
\frac{1}{4} (Z_{\lambda_m}\!-1)\lambda_m \phi^2\sigma^2
+\frac{1}{4!} (Z_{\lambda_\sigma}\!-1)\lambda_\sigma\sigma^4\big]
\end{eqnarray}
with
\begin{eqnarray}\label{Z}
Z_{\lambda_\phi}&=&1+\frac{3}{2 \kappa\,\epsilon} (\lambda_\phi+\lambda_m^2/\lambda_\phi),
\nonumber\\
Z_{\lambda_m}&=&1+\frac{1}{2 \kappa\, \epsilon} (\lambda_\phi+\lambda_\sigma+4\lambda_m),
\nonumber\\
Z_{\lambda_\sigma}&=&1+ \frac{3}{2 \kappa\, \epsilon} (\lambda_\sigma+\lambda_m^2/\lambda_\sigma),
\quad \kappa=(4\pi)^2.
\end{eqnarray}
\medskip\noindent
$Z_{\lambda}$'s are identical to their counterparts computed in the same theory
regularized with $\mu$=constant (when scale symmetry is broken explicitly).
The one-loop potential becomes
\begin{eqnarray}\label{U}
U_1&= &V+V^{(1)}+V^{(1,n)},
\\[6pt]
V^{(1)} &\equiv & \frac{1}{64\pi^2}\,
\sum_{s=\phi,\sigma} M^4_s(\phi,\sigma)\, \Big[\ln \frac{M^2_s(\phi,\sigma)}{\mu^2(\sigma)}
-\frac{3}{2}\,\Big],
\\
V^{(1,n)}
\!\!\!& \equiv &\!\!\! \frac{1}{48 \kappa}
\Big[ \lambda_\phi \lambda_m \frac{\phi^6}{\sigma^2}
\!-\! (16 \lambda_\phi \lambda_m + 18 \lambda_m^2
\!-\! \lambda_\phi \lambda_\sigma)\phi^4
\!-\! (48 \lambda_m\! +\! 25 \lambda_\sigma )\lambda_m\,
\phi^2 \sigma^2 \!-\! 7 \lambda_\sigma^2 \sigma^4 \Big].
\quad\,\,
\end{eqnarray}
\medskip\noindent
The potential simplifies further if we use the tree-level relation (\ref{minimum})
among $\lambda_s$ ($s=\phi,m, \sigma$)
that ensures the spontaneous scale symmetry breaking.
$U_1$ is scale symmetric and a flat direction exists also at the quantum level.
$V_1^{(1,n)}$ is a new, {\it finite} one-loop correction, independent of the parameter $z$;
it contains a {\it non-polynomial} term $\phi^6/\sigma^2$ that can be Taylor-expanded about
$\langle\phi\rangle$, $\langle\sigma\rangle\not=0$.
$V^{(1,\text{n})}\!\rightarrow\! 0$ in the classical decoupling limit $\lambda_m\!\rightarrow\! 0$.
The Coleman-Weinberg term is also present, with $\mu\rightarrow \mu(\sigma)$ and thus
depends on $z$.
This dependence replaces the ``traditional'' dependence of $V^{(1)}$
on the subtraction scale in theories regularized
with $\mu=$constant. But physics should be independent of this parameter,
which means that $U_1$ must respect the Callan-Symanzik equation: $dU_1/d\ln z=0$ \cite{tamarit}.
To check the Callan-Symanzik equation, we need the beta functions of
the couplings which run with the momentum, even in scale invariant theories \cite{S3,tamarit}.
These are computed from the condition
$d (\mu(\sigma)^{2\epsilon}\lambda_j Z_{\lambda_j})/d\ln z\!=\!0$ ($j$:\,fixed),
since the bare coupling is independent of $z$.
The result is identical to that in a theory regularized with $\mu$=constant:
\medskip
\begin{eqnarray}
\beta_{\lambda_\phi}^{(1)}&\equiv &\frac{d \lambda_\phi}{d\ln z}
=\frac{3}{\kappa} \,(\lambda_\phi^2+\lambda_m^2),
\\
\beta_{\lambda_m}^{(1)}& \equiv &\frac{d\lambda_m}{d\ln z}=
\frac{1}{\kappa} \,(\lambda_\phi+4\lambda_m+\lambda_\sigma)\lambda_m
\\
\beta_{\lambda_\sigma}^{(1)}& \equiv &\frac{d\lambda_\sigma}{d\ln z}=
\frac{3}{\kappa}\,(\lambda_m^2+\lambda_\sigma^2).
\end{eqnarray}
\medskip\noindent
The Callan Symanzik equation at one-loop is
\medskip
\begin{eqnarray}\label{uui}
\frac{d U_1 (\lambda_j,z)}{d\ln z}=\Big(\beta_{\lambda_j}^{(1)}\,\frac{\partial }{\partial \lambda_j}
+z\,\frac{\partial}{\partial z}\Big)\, U_1(\lambda_j,z)= \cO(\lambda_j^3),
\qquad(\rm{sum\,\,over}\,\, j=\phi,m,\sigma).
\end{eqnarray}
\medskip\noindent
Eq.(\ref{uui}) is easily verified with the above results for the beta functions.
The one-loop $U_1$ can be used for phenomenology of a scale invariant version of the
SM extended by the dilaton~\cite{dmg}.
\section{Two-loop potential}
\subsection{ New poles in the two-loop potential}
To compute the two-loop potential we use the background field expansion method about
$\phi,\sigma$. We Taylor-expand $\tilde V$ about these values
\begin{eqnarray}
\tilde V(\phi+\delta\phi,\sigma+\delta\sigma)=
V(\phi,\sigma)+ \tilde V_j\, s_j
+\frac12 \tilde V_{jk}\, s_j s_k
+\frac{1}{3!} \tilde V_{ijk}\, s_i s_j s_k
+\frac{1}{4!} \tilde V_{ijkl}\, s_i s_j s_k s_l
+\cdots
\end{eqnarray}
\medskip\noindent
where the subscripts $i, j, k, l$ of $\tilde V_{ij...}$
denote derivatives of $\tilde V$ wrt fields of the set $\{\phi,\sigma\}_j$;
with $i,j, k,l=1,2$. Also $s_1=\delta \phi$, $s_2=\delta\sigma$ are field fluctuations.
Notice
that there are new, evanescent interactions ($\propto\epsilon$)
in vertices $\tilde V_{ijk\cdots}$ generated by eq.(\ref{tildeV})
that impact on the loop corrections. The two-loop diagrams are presented below.
Let us first denote:
\begin{gather}
V_2=V_2^a+V_2^b+V_2^c.
\end{gather}
Then
\begin{eqnarray}\label{twoloop}
V_2^a &=& \frac{i}{12}\;
\parbox[c][2em][c]{4.0em}{\includegraphics[scale=0.14]{Va}}
=
\frac{i}{12}
\tilde V_{ijk} \tilde V_{lmn}
\int \frac{d^d p}{(2\pi)^d} \int \frac{d^d q}{(2\pi)^d}
\,(\tilde D_p^{-1})_{il}
\,(\tilde D_q^{-1})_{jm}
\,(\tilde D_{p+q}^{-1})_{kn}
\nonumber\\
&=&
\frac{\mu(\sigma)^{2\epsilon}}{\epsilon^2}\frac{1}{16\kappa^2}
\Big[\phi^4 \,(\lambda_\phi^3+\lambda_\phi\,\lambda_m^2+2\lambda_m^3)
+\sigma^4\,(2\lambda_m^3+\lambda_m^2\lambda_\sigma+\lambda_\sigma^3)
\nonumber\\
&&\hspace{2cm}
+\,\phi^2\,\sigma^2\,(\lambda_\phi^2\,\lambda_m+6 \lambda_\phi\,\lambda_m^2
+10\lambda_m^3+6\lambda_m^2\lambda_\sigma+\lambda_m\lambda_\sigma^2)\Big]
+\cO(1/\epsilon).\qquad
\end{eqnarray}
Also
\begin{eqnarray}
V_2^b &=&
\frac{i}{8}\;
\parbox[c][3em][c]{1.9em}{\includegraphics[scale=0.14]{Vb}}
=\frac{i}{8}
\tilde V_{ijkl}
\Big[\int \frac{d^d p}{(2\pi)^d} (\tilde D_p^{-1})_{ij}\Big]\,
\Big[ \int \frac{d^d q}{(2\pi)^d} (\tilde D_q^{-1})_{kl} \Big]
\hspace{3cm}
\nonumber\\
&=& \frac{\mu(\sigma)^{2\epsilon}}{\epsilon^2}\frac{1}{32\kappa^2}
\Big[
\phi^4\,(\lambda_\phi^3+2\lambda_\phi\lambda_m^2+\lambda_m^2\lambda_\sigma)
+\sigma^4\,(\lambda_\phi\lambda_m^2+2\lambda_m^2\lambda_\sigma+\lambda_\sigma^3)
\nonumber\\
& & \hspace{2cm}
+\,2\lambda_m \,\phi^2\sigma^2\,(\lambda_\phi^2+ 9 \lambda_m^2+\lambda_\phi\lambda_\sigma
+\lambda_\sigma^2)\Big]+
\cO(1/\epsilon),\qquad
\end{eqnarray}
\noindent
and finally
\begin{eqnarray}\label{dc}
V_2^c & = & \frac{i}{2}\;
\parbox[c][2em][c]{3.2em}{\includegraphics[scale=0.14]{Vc}}
=\frac{i}{2}
(\delta V_1)_{ij}\int \frac{d^d p}{(2\pi)^d} \,(\tilde D_p^{-1})_{ij}
\nonumber\\
&=&
\frac{\mu(\sigma)^{2\epsilon}}{\epsilon^2}\frac{(-1)}{16\kappa^2}
\Big[\phi^4\,
(3\lambda_\phi^3+4\lambda_\phi\lambda_m^2+4 \lambda_m^3+\lambda_m^2\lambda_\sigma)
+\sigma^4\,(\lambda_\phi\lambda_m^2+4\lambda_m^3+4\lambda_m^2\lambda_\sigma+3\lambda_\sigma^3)
\nonumber\\
&+&
\phi^2\sigma^2\,(4\lambda_\phi^2\lambda_m+12\lambda_\phi\lambda_m^2+38\lambda_m^3
+2\lambda_\phi\lambda_m\lambda_\sigma+12 \lambda_m^2\lambda_\sigma+4\lambda_m\lambda_\sigma^2)
\Big]+\cO(1/\epsilon).
\end{eqnarray}
\medskip\noindent
These diagrams are computed using \cite{DT}, see also \cite{JJ}.
The propagators are given by the inverse of the matrix
$(\tilde D_p)_{ij}=p^2\delta_{ij}-\tilde V_{ij}$. To simplify the calculation
they can be re-written as
$(\tilde D^{-1})_{ij}=\tilde a_{ij}/(p^2-\tilde V_p)+\tilde b_{ij}/(p^2-\tilde V_m)$,
with appropriate coefficients $\tilde a_{ij}$ and $\tilde b_{ij}$ and
where $\tilde V_p$, $\tilde V_m$ ($\tilde V_p>\tilde V_m$)
denote the field-dependent masses, eigenvalues of the matrix
$\tilde V_{ij}$, $i,j=\phi,\sigma$. Note that $\tilde V_{ij}$, $\tilde V_p$, $\tilde V_m$,
$\tilde a_{ij}$, $\tilde b_{ij}$ and also $\tilde V_{ijkl}$, $\tilde V_{ijk}$
contain positive powers of $\epsilon$;
this is relevant for the above calculation, since they contribute
to the finite and $1/\epsilon$ parts of the potential.
Their form is detailed in Appendix~\ref{appendixB} and \ref{appendixC}.
One notices that the poles $1/\epsilon^2$ in $V_2^{a,b,c}$ are identical to those in the
theory regularized with $\mu=$constant. This is expected for this
leading singularity, but this is not true for their sub-leading one ($1/\epsilon$) or
for their finite part (see later).
The long expressions $\cO(1/\epsilon)$ and $\cO(\epsilon^0)$ of each diagram $V_2^{a,b,c}$
are not shown here. The sum of these diagrams gives
\medskip
\begin{eqnarray}\label{V2}
\!\!\!\!V_2&=&
\frac{\mu(\sigma)^{2\epsilon}}{\epsilon^2}\frac{(-1)}{32 \kappa^2}
\,\Big[\phi^4 (3 \lambda_\phi^3
+\! 4 \lambda_\phi \lambda_m^2\! + 4\lambda_m^3 + \lambda_m^2 \lambda_\sigma)
+\!\sigma^4(3 \lambda_\sigma^3\! + {\lambda_\phi \lambda_m^2}
+\! 4 \lambda_m^3+ 4\lambda_m^2 \lambda_\sigma)
\nonumber \\
&& \hspace{1.7cm} +\,\phi^2 \sigma^2 (4 \lambda_\phi^2 \lambda_m
+ 12 \lambda_\phi \lambda_m^2 + 38 \lambda_m^3+
2 \lambda_\phi \lambda_m \lambda_\sigma + 12\lambda_m^2 \lambda_\sigma+ 4\lambda_m \lambda_\sigma^2)
\Big]
\nonumber\\
&+& \frac{\mu(\sigma)^{2\epsilon}}{\epsilon}\; \frac{1}{16 \kappa^2}
\,\Big[\phi^4 ( \lambda_\phi^3
+ \lambda_\phi \lambda_m^2 + 2\lambda_m^3 ) +\sigma^4(2 \lambda_m^3+ \lambda_m^2 \lambda_\sigma+\lambda_\sigma^3)
\\
&&\hspace{1.7cm} +\phi^2 \sigma^2 (\lambda_\phi^2 \lambda_m
+ 6 \lambda_\phi \lambda_m^2 + 10 \lambda_m^3 + 6 \lambda_m^2 \lambda_\sigma+ \lambda_m \lambda_\sigma^2)
\Big] +V_2^{1/\epsilon}\!\!+V^{(2)}+V^{(2,n)}.
\nonumber
\end{eqnarray}
\medskip\noindent
Here $V^{(2)}$ and $V^{(2,n)}$ are $\cO(\epsilon^0)$ i.e. finite quantum
corrections presented in Appendix~\ref{appendixB}.
$V_2^{1/\epsilon}=\cO(1/\epsilon)$ is a {\it new term} that contains $1/\epsilon$ poles
{\it not} present in the theory regularized with $\mu$=constant;
its origin is due to evanescent interactions ($\propto\epsilon$), which ``meet''
$1/\epsilon^2$ poles, thus giving $1/\epsilon$ terms.
One finds
\medskip
\begin{eqnarray}\label{opp}
\!\!\! V_2^{1/\epsilon}
\!\!&=&
\frac{\mu(\sigma)^{2\epsilon}}{16 \kappa^2\,\epsilon} \,
\Big[
\phi^4 \Big(\frac{20}{3}\lambda_\phi^2 \lambda_m \!+ \frac{7}{6}
\lambda_\phi \lambda_m^2\! - 2\lambda_m^3\! - \frac{1}{2} \lambda_\phi^2 \lambda_\sigma
\! - \frac{4}{3}\lambda_\phi \lambda_m \lambda_\sigma
\!+ \frac{7}{12} \lambda_m^2 \lambda_\sigma\! +\!
\frac{1}{4} \lambda_\phi \lambda_\sigma^2
\Big)
\nonumber\\
&+&\!\!
\phi^2 \sigma^2 \Big(\,8 \lambda_\phi \lambda_m^2
+\frac{41}{2} \lambda_m^3 + \lambda_\phi \lambda_m \lambda_\sigma
+ \frac{43}{3}\lambda_m^2\lambda_\sigma +\frac{1}{2} \lambda_m \lambda_\sigma^2 \Big)
+
\sigma^4\Big(4 \lambda_m^3\!+ \frac{1}{3}\lambda_m^2 \lambda_\sigma
\!+\frac{7}{4}\lambda_\sigma^3 \Big)
\nonumber\\
& +&
\frac{\phi^6}{\sigma^2}\, \Big( -\frac{7}{6} \lambda_\phi^2 \lambda_m
+ \frac{7}{3} \lambda_\phi \lambda_m^2 - \frac{1}{6} \lambda_\phi \lambda_m \lambda_\sigma \Big)
-
\frac{1}{4} \lambda_\phi \lambda_m^2 \,\,\frac{\phi^8}{\sigma^4}\, \Big].
\end{eqnarray}
\bigskip\noindent
In addition to usual counterterms ($\phi^4$, etc),
notice from eq.(\ref{opp}) the need for {\it non-polynomial}
counterterms $\phi^6/\sigma^2$ and $\phi^8/\sigma^4$
(see also $\phi^6/\sigma^2$ in eq.(\ref{U})).
The above two-loop results contribute to the Lagrangian
(below $\rho^{\phi},\rho^\sigma$ are wavefunction coefficients defined later)
\begin{eqnarray}\label{l2}
L_2=\frac{1}{2}\Big( \frac{\rho^\phi}{\epsilon}+\text{finite}\Big) (\partial_\mu\phi)^2
+
\frac{1}{2}\Big(\frac{\rho^\sigma}{\epsilon}+\text{finite}\Big) (\partial_\mu\sigma)^2
-V_2.
\end{eqnarray}
\medskip\noindent
A counterterm $\delta L_2$ cancels the poles in the sum $L_1+L_2$ of
eqs.(\ref{l1}), (\ref{l2}) up to two-loops
\medskip
\begin{eqnarray}\label{deltaL2}
\delta L_2&=&
\frac{1}{2}(Z_\phi-1)(\partial_\mu\phi)^2
+
\frac{1}{2}(Z_\sigma-1)(\partial_\mu\sigma)^2
-
\mu(\sigma)^{2\epsilon}\,
\Big\{\, (Z_{\lambda_\phi}-1)\frac{\lambda_\phi}{4!}\phi^4 +
\nonumber\\
&+&
\!\!\!\! (Z_{\lambda_m}-1)\frac{\lambda_m}{4}\phi^2\,\sigma^2
\,+\,(Z_{\lambda_\sigma}-1)\frac{\lambda_\sigma}{4!} \sigma^4
\!+\! (Z_{\lambda_6}-1)\frac{\lambda_6}{6} \frac{\phi^6}{\sigma^2}
\!+\! (Z_{\lambda_8}-1)\frac{\lambda_8}{8} \frac{\phi^8}{\sigma^4}
\,\Big\},\,\,\,
\end{eqnarray}
where
\begin{eqnarray}\label{Z1}
Z_{\lambda_\phi}&=&1+\frac{\delta_0^\phi}{\kappa\,\epsilon}+\frac{1}{\kappa^2}
\Big(\,\frac{\delta_1^\phi+\nu_1^\phi}{\epsilon}+\frac{\delta_2^\phi}{\epsilon^2}\,\Big),
\nonumber\\
Z_{\lambda_m}&=&1+\frac{\delta_0^m}{\kappa\,\epsilon}+
\frac{1}{\kappa^2}\Big(\,\frac{\delta_1^m+\nu_1^m}{\epsilon}+\frac{\delta_2^m}{\epsilon^2}\,\Big),
\nonumber\\
Z_{\lambda_\sigma}&=&1+
\frac{\delta_0^\sigma}{\kappa\,\epsilon}+
\frac{1}{\kappa^2}\Big(\,\frac{\delta_1^\sigma+\nu_1^\sigma}{\epsilon}
+\frac{\delta_2^\sigma}{\epsilon^2}\,\Big),
\nonumber\\
Z_{\lambda_6} &=& 1 + \frac{1}{\kappa^2} \frac{\nu_1^6}{\epsilon};
\qquad
Z_{\lambda_8}= 1 + \frac{1}{\kappa^2} \frac{\nu_1^8}{\epsilon},
\end{eqnarray}
\medskip\noindent
where one-loop $\delta_0^\phi$, $\delta_0^m$, $\delta_0^\sigma$ can be read from eq.(\ref{Z})
while the two-loop coefficients $\delta_{k}^{s}$, $k=1,2$, $s=\phi, m, \sigma$,
are shown in Appendix~\ref{appendixA}. They are obtained by comparing
$\delta L_2$ against $L_2$, using $V_2$ of eq.(\ref{V2}).
The coefficients $\delta_k^s$ are those of the theory regularized
with $\mu=$constant. However, there is an extra contribution from coefficients
$\nu_1^s$, $s=\phi,m,\sigma, 6,8$ (see Appendix~\ref{appendixA}),
that is generated by the new poles $1/\epsilon$ of
$V_2^{1/\epsilon}$. This new contribution brings a correction to the two-loop
beta functions of our theory, see later.
One can also show that the two-loop-corrected wavefunction coefficients have
expressions similar to those in the theory regularised with $\mu=$constant:
\medskip
\begin{eqnarray}
\label{Z2}
Z_\phi &=& 1 +\frac{\rho^\phi}{\kappa^2\epsilon},
\qquad\quad
\rho^\phi= - \frac{1}{24}\, (\lambda_\phi^2 + 3 \lambda_m^2),
\nonumber\\
Z_\sigma &=& 1 +\frac{\rho^\sigma}{\kappa^2\,\epsilon},
\qquad\quad
\rho^\sigma= - \frac{1}{24}\, (\lambda_\sigma^2 + 3 \lambda_m^2),
\end{eqnarray}
One often uses the notation $\gamma_\phi=-2\rho^\phi/\kappa^2$ and
$\gamma_\sigma=-2\rho^\sigma/\kappa^2$ for the anomalous dimensions.
\subsection{Two-loop beta functions}
With the above information, one obtains the two-loop beta functions. To this purpose, one
uses that the ``bare'' couplings $\lambda^B_j$ below are independent of the parameter $z$:
\begin{eqnarray}\label{b}
\lambda_\phi^B &=& \mu(\sigma)^{2\epsilon}\lambda_\phi\,Z_{\lambda_\phi}\,Z_\phi^{-2},
\nonumber\\
\lambda_m^B &=& \mu(\sigma)^{2\epsilon}\lambda_m\,Z_{\lambda_m}\,Z_\phi^{-1}\,Z_\sigma^{-1},
\nonumber\\
\lambda_\sigma^B &=& \mu(\sigma)^{2\epsilon}\lambda_\sigma\,Z_{\lambda_\sigma}\,Z_\sigma^{-2},
\nonumber\\
\lambda_6^B &=& \mu(\sigma)^{2\epsilon}\lambda_6\,Z_{\lambda_6}\,Z_\phi\,Z_\sigma^{-3}.
\end{eqnarray}
\medskip\noindent
We thus demand that $(d/d\ln z) \lambda_k^B=0$, $k=\phi,m,\sigma, 6, 8$ \footnote{
We also include the effect of
wavefunction renormalization of the subtraction function which demands replacing:
$\mu(\sigma)=z\,\sigma^{1/(1-\epsilon)}\rightarrow z\,(Z_\sigma^{1/2}\,\sigma)^{1/(1-\epsilon)}$; however,
this brings no correction in this order.}.
Taking the logarithm of the first expression in (\ref{b})
and then the derivative with respect to $\ln z$, one obtains
\medskip
\begin{eqnarray}
2\epsilon+\frac{\beta_{\lambda_\phi}}{\lambda_\phi}
+\sum_{j=\phi,m,\sigma}\,\beta_{\lambda_j}\,\frac{d}{d\ln z}
\ln \Big[Z_{\lambda_\phi} Z_\phi^{-2}\Big]=0
\end{eqnarray}
and similar expressions for the other couplings. Using the form of $Z'$s, one finds
\medskip
\begin{eqnarray}\label{bb}
\beta_{\lambda_\phi}
=
-2\epsilon\lambda_\phi
+
2\lambda_\phi \sum_{j=\phi,m,\sigma}
\lambda_j\,\frac{d}{d\lambda_j}
\Big(\frac{\delta_0^\phi}{\kappa}+\frac{\delta_1^\phi+\nu_1^\phi}{\kappa^2}
-\frac{2\,\rho^\phi}{\kappa^2}\Big).
\end{eqnarray}
\medskip\noindent
One easily obtains similar relations for $\beta_{\lambda_m}$ and $\beta_{\lambda_\sigma}$
(for $\beta_{\lambda_\sigma}$ just replace the sub-/super-script $\phi\rightarrow \sigma$).
The difference in these beta functions
from those in the same theory but regularized with $\mu=$constant is the presence
of a new contribution: $\nu_1^{\phi}$ ($\nu_1^m$, $\nu_1^\sigma$, respectively),
that we identified in eqs.(\ref{Z1}).
Eq.(\ref{bb}) is solved with particular attention to the
$\epsilon$-dependent terms, to find at two-loop:
\medskip
\begin{eqnarray}
\beta_{\lambda_\phi}&=&\frac{3}{\kappa}(\lambda_\phi^2+\lambda_m^2)
-\frac{1}{\kappa^2}(\frac{17}{3}\lambda_\phi^3 + 5\lambda_\phi\lambda_m^2 + 12 \lambda_m^3)
+\beta_{\lambda_\phi}^{(2,\text{n})},
\nonumber\\
\beta_{\lambda_m} &=&\frac{1}{\kappa}(\lambda_\phi + 4\lambda_m + \lambda_\sigma)\lambda_m
- \frac{\lambda_m}{6\kappa^2} (5\lambda_\phi^2 + 36 \lambda_\phi \lambda_m + 54 \lambda_m^2 + 36 \lambda_m \lambda_\sigma + 5\lambda_\sigma^2)
+ \beta_{\lambda_m}^{(2,\text{n})},
\nonumber\\
\beta_{\lambda_\sigma} &=& \frac{3}{\kappa}(\lambda_m^2+\lambda_\sigma^2)-
\frac{1}{\kappa^2}(12\lambda_m^3 + 5\lambda_m^2 \lambda_\sigma + \frac{17}{3}\lambda_\sigma^3)
+\beta_{\lambda_\sigma}^{(2,\text{n})}.
\end{eqnarray}
\medskip\noindent
The ``new'' terms $\beta_\lambda^{(2,n)}$ on the rhs are
\begin{eqnarray}\label{betas2}
\beta_{\lambda_\phi}^{(2,\text{n})} &=& \frac{1}{2 \kappa^2}
\left[ \lambda_m^2 (24\lambda_m - 7 \lambda_\sigma)+ \lambda_\phi(-14 \lambda_m^2 + 16 \lambda_m \lambda_\sigma -3 \lambda_\sigma^2)
+ \lambda_\phi^2(-80\lambda_m + 6 \lambda_\sigma) \right],
\nonumber\\
\beta_{\lambda_m}^{(2,\text{n})} &=& -\frac{\lambda_m}{6 \kappa^2} \,(48 \lambda_\phi \lambda_m + 6 \lambda_\phi \lambda_\sigma +
123 \lambda_m^2 + 86 \lambda_m \lambda_\sigma + 3 \lambda_\sigma^2),
\nonumber\\
\beta_{\lambda_\sigma}^{(2,\text{n})} &=& -\frac{1}{2 \kappa^2}\,(48\lambda_m^3 + 4 \lambda_m^2 \lambda_\sigma + 21 \lambda_\sigma^3),
\nonumber\\
\beta_{\lambda_6}^{(2,\text{n})} &=& \frac{1}{4\kappa^2}\, \lambda_\phi \lambda_m (7\lambda_\phi - 14 \lambda_m + \lambda_\sigma),
\nonumber\\
\beta_{\lambda_8}^{(2,\text{n})} &=& \frac{1}{2\kappa^2}\, \lambda_\phi \lambda_m^2.
\end{eqnarray}
\medskip\noindent
Here $\beta^{(2,\text{n})}_\lambda$ that appears for each $\lambda$ at two-loop is the
mentioned correction, that is missed if this theory is regularized
with $\mu=$constant, when one breaks explicitly the scale symmetry.
Notice that $\lambda_{6,8}$ also run in this order in the scale invariant theory.
We conclude that from the two-loop running of
the couplings, encoded by the beta functions,
one can distinguish between the theory with (spontaneously
broken) scale symmetry at quantum level and that in which this symmetry is broken
{\it explicitly} by quantum corrections (with $\mu=$constant).
There is a simple way to understand this difference:
the theory regularized with $\mu=$constant, and two fields $\phi$, $\sigma$ is renormalizable
while our model, scale invariant at quantum level, is non-renormalizable.
This is due to the scale-invariant non-polynomial terms of type $\phi^6/\sigma^2$,
$\phi^8/\sigma^4$ generated at one- and two-loop level\footnote{
This non-renormalizability argument is different
from that in \cite{S2} which does not apply here, see \cite{dmg}.}.
This justifies the different beta functions in the two approaches starting from
the two-loop level.
This is an interesting result of the paper.
\subsection{Two-loop potential after renormalization}
Finally, we present the two-loop potential $U$ after renormalization. It has the form
\medskip
\begin{eqnarray}\label{finalU}
U=\underbrace{V+V^{(1)}+V^{(1,\text{n})}}_{=U_1}+V^{(2)}+V^{(2,\text{n})}
\end{eqnarray}
\medskip\noindent
where $U_1$ is the one-loop result of (\ref{U}).
$V^{(2)}$ is a two-loop correction identical to that obtained in the theory regularized with
$\mu=$constant (up to replacing $\mu\rightarrow z\, \sigma)$,
while $V^{(2,\text{n})}$ are new two-loop terms that involve derivatives of $\mu(\sigma)$ wrt
$\sigma$ (similar to one-loop $V^{(1,\text{n})}$)\footnote{See \cite{EJ} for
further discussion on the Goldstone modes contributions to the potential.}
The long expressions of $V^{(2)}$, $V^{(2,n)}$ are given
in Appendix~\ref{appendixB}, eq.(\ref{B5}).
$U$ contains new, {\it non-polynomial} effective operators,
such as $\phi^6/\sigma^2$ and $\phi^8/\sigma^4$, etc:
\begin{eqnarray}\label{uu}
U=
\frac{7 \lambda_\phi^3}{576 \, \kappa^2}\frac{\phi^8}{\sigma^4}+
\frac{5\lambda_\phi^3}{24 \,\kappa^2}\frac{\phi^6}{\sigma^2}
+\cdots
\end{eqnarray}
\medskip\noindent
All non-polynomial terms present in the potential can be expanded about the ground state
\begin{eqnarray}
\phi=\langle\phi\rangle+\delta\phi,\qquad
\sigma=\langle\sigma\rangle+\delta\sigma
\end{eqnarray}
where
$\delta\phi$ and $\delta\sigma$ represent fluctuations about the ground state.
Then each non-polynomial operator becomes an infinite series expansion about the
point $\langle\phi\rangle/\langle\sigma\rangle$. For example
\medskip
\begin{eqnarray}\label{exp}
\frac{\phi^6}{\sigma^2}=(\langle\phi\rangle+\delta\phi)^4\,\,
\frac{\langle\phi\rangle^2}{\langle\sigma\rangle^2}\,\,
\Big(1+\frac{2\delta \phi}{\langle\phi\rangle}+\frac{\delta\phi^2}{\langle\phi\rangle^2}+\cdots\Big)
\,\,
\Big(1-\frac{2\delta \sigma}{\langle\sigma\rangle}+\frac{3\delta\sigma^2}{\langle\sigma\rangle^2}+\cdots\Big)
\end{eqnarray}
\medskip\noindent
and similarly for the operator $\phi^8/\sigma^4$ in $U$, etc.
Although we did not present the ground state of the one-loop potential, this
is known to satisfy the relation \cite{dmg}:
$\langle\phi\rangle^2/\langle\sigma\rangle^2=
-3\lambda_m/\lambda_\sigma (1+\text{loop-corrections})$ \cite{dmg}.
Using this information in eq.(\ref{exp}) and (\ref{uu}), one sees that in the classical
decoupling limit $\lambda_m\rightarrow 0$, the non-polynomial operators of (\ref{uu}) do vanish.
It is important to stress that only operators
of the form $\phi^{2n+4}/\sigma^{2n}$, $n\geq 1$ were generated in the two-loop potential,
but no operator like $\sigma^{2n+4}/\phi^{2n}$, $n\geq 1$ is present.
This is due to the way the subtraction function enters in the loop corrections,
via derivatives wrt $\sigma$ of $\mu(\sigma)^\epsilon$ which are suppressed by
positive powers of $\mu(\sigma)$. This means that all higher dimensional operators
are ultimately suppressed by (large) $\langle\sigma\rangle$ and not proportional to it.
This is welcome for the hierarchy problem, since such terms could otherwise lead
to corrections to the higgs mass of the type $\lambda_\phi^3 \langle\sigma\rangle^2$
requiring tuning the higgs self-coupling $\lambda_\phi$, and thus re-introducing the hierarchy problem.
This problem is avoided at least at one-loop \cite{S1,dmg}.
\section{Two-loop Callan-Symanzik for the potential}
A good check of our two-loop scale-invariant potential is
the Callan-Symanzik equation, in its version for scale invariant theories
\cite{tamarit}. This equation states the independence of the two-loop potential of
the subtraction (dimensionless) parameter $z$; this parameter fixes the subtraction scale to
$z\langle\sigma\rangle$, after spontaneous scale symmetry breaking.
The equation is
\medskip
\begin{eqnarray} \label{CS}
\frac{d\, U(\lambda,z)}{d \ln z}=
\Big(
z\frac{\partial}{\partial z}
+
\beta_{\lambda_j}\, \frac{\partial}{\partial \lambda_j}
-
\phi\gamma_\phi\,\frac{\partial}{\partial\phi}
-
\sigma\gamma_\sigma\,\frac{\partial}{\partial\sigma}\Big) U(\lambda,z)
=0\;,
\end{eqnarray}
where the $j$-summation runs over $\lambda_j=\lambda_\phi,\lambda_m,\lambda_\sigma,\lambda_6,\lambda_8$.
Eq.\eqref{CS} can be re-written as a
set of equations at a given order of $\lambda$'s (or number of loops).
To help one trace the difference between our scale-invariant result and that
of the same theory but with $\mu=$constant, below we use for $U$ the
decomposition given in eq.(\ref{finalU}) while for the beta functions
we write
\medskip
\begin{eqnarray}
\beta_{\lambda_j} &=& \beta^{(1)}_{\lambda_j} + \beta^{(2)}_{\lambda_j}+ \beta^{(2,\text{n})}_{\lambda_j}.
\end{eqnarray}
\medskip\noindent
The terms in the beta function
correspond to 1-loop ($\beta_\lambda^{(1)}$), 2-loop-only ($\beta_\lambda^{(2)}$)
and 2-loop-new parts ($\beta_\lambda^{(2,\text{n})}$).
Then, with the two-loop anomalous dimensions $\gamma_\phi$, $\gamma_\sigma$ defined
after eq.(\ref{Z2}), a careful analysis shows that
eq.\eqref{CS} splits into
\begin{eqnarray}
&& \frac{\partial \, V^{(1)}}{\partial \ln z} \,
+ \, \beta^{(1)}_{\lambda_j}\, \frac{\partial V}{\partial \lambda_j} =0 \label{CSOLD1}
\\
&& \frac{\partial \, V^{(1,\text{n})}}{\partial \ln z} = 0 \label{CSNEW1}
\\
&& \frac{\partial \, V^{(2)} }{\partial \ln z}
+ \left( \beta^{(2)}_{\lambda_j}
\frac{\partial \;}{\partial \lambda_j} - \gamma_\phi \phi
\frac{\partial}{\partial \phi}- \gamma_\sigma \sigma \frac{\partial}{\partial \sigma}\right)
V + \beta^{(1)}_{\lambda_j} \,\frac{\partial V^{(1)}}{\partial\lambda_j}=0 \label{CSOLD2}
\\
&&
\frac{\partial\, V^{(2,\text{n})}}{\partial \ln z} + \beta^{(2,\text{n})}_{\lambda_j}
\frac{\partial V}{\partial \lambda_j} + \beta^{(1)}_{\lambda_j}\,
\frac{\partial V^{(1,\text{n})}}{\partial \lambda_j}=0 \label{CSNEW2} \;.
\end{eqnarray}
\medskip\noindent
where $V$ includes the new terms
$(\lambda_6/6)\, \phi^6/\sigma^2+ (\lambda_8/8)\, \phi^8/\sigma^4$.
We checked that these equations are respected.
Eqs.\eqref{CSOLD1}, \eqref{CSOLD2} express the \textit{usual} Callan-Symanzik equation (of the theory
with $\mu=$constant), whereas \eqref{CSNEW1} and \eqref{CSNEW2} constitute a \textit{new} part,
which is nonzero only when $\mu=\mu(\sigma)$.
Eq.\eqref{CSNEW1} is obvious and hardly revealing. But
checking eq.\eqref{CSNEW2} is more difficult.
For this one also needs to take account of the ``new'' corrections to
two-loop beta functions of $\lambda_{6,8}$, see eq.(\ref{betas2}) and also
the $z$-dependent part of $V^{(2,\text{n})}$ which we write below
\smallskip
\begin{eqnarray}
V^{(2,\text{n})}& =& \frac{1}{192\,\kappa^2}\Big[
\overline{\ln}V_p + \overline{\ln}V_m\Big]\Big[
\big( -144 \lambda_\phi^2 \lambda_m
- 111 \lambda_\phi \lambda_m^2 - 168 \lambda_m^3 + 9 \lambda_\phi^2 \lambda_\sigma - 40 \lambda_m^2 \lambda_\sigma \big)\, \phi^4
\nonumber\\
&-& \big(192 \lambda_\phi \lambda_m^2 + 705 \lambda_m^3 +
37 \lambda_\phi \lambda_m \lambda_\sigma + 368 \lambda_m^2 \lambda_\sigma + 106 \lambda_m \lambda_\sigma^2\big)\,\phi^2 \sigma^2
\nonumber\\
& -& \big(48 \lambda_m^3 + 46 \lambda_m^2 \lambda_\sigma + 63 \lambda_\sigma^3\big)\,\sigma^4 +
\big( 18 \lambda_\phi^2 \lambda_m - 24 \lambda_\phi \lambda_m^2 + 3 \lambda_m^3 +
3 \lambda_\phi \lambda_m \lambda_\sigma \big)\, \frac{\phi^6}{\sigma^2}
\nonumber\\
& + & 3 \lambda_\phi \lambda_m^2\, \frac{\phi^8}{\sigma^4}
\Big] + \text{$z$-independent terms}, \quad
\text{where}\quad
\overline\ln A\equiv\ln\frac{A}{(z\sigma)^2 4\pi e^{-\gamma_E}}-1.
\end{eqnarray}
\smallskip\noindent
where $\gamma_E=0.5772....$.
Here $V_p$ and $V_m$ are field dependent eigenvalues of the matrix of second derivatives
$V_{ij}$ wrt $i,j=\phi,\sigma$ of the tree level potential. Given this, the
Callan-Symanzik equation of the potential is verified at the two-loop level.
\section{Conclusions}
Quantum scale invariance with spontaneous breaking may provide
a solution to the cosmological constant
and the hierarchy problem. The ``traditional'' method for loop calculations
breaks explicitly classical scale symmetry of a theory
due to the regularization which introduces a subtraction scale (DR scale, cut-off,
Pauli-Villars scale). However, it is known how
to perform quantum calculations in a manifestly scale invariant way: the
subtraction scale is replaced by a subtraction function of the field(s) (dilaton $\sigma$)
which when acquiring a vev spontaneously, generates this scale
$\mu(\langle\sigma\rangle)=z\langle\sigma\rangle$. The Goldstone mode of this
symmetry is the dilaton field which remains a flat direction of the
quantum scale-invariant potential.
Starting with a classically scale-invariant action,
we computed the two-loop scalar potential
of $\phi$ (higgs-like) and $\sigma$ in a scale invariant regularization.
The one- and two-loop potential are scale invariant and contain new terms
beyond the usual corrections obtained for $\mu=$constant (Coleman-Weinberg, etc),
due to field derivatives of $\mu(\sigma)$. They also contain
interesting effective {\it non-polynomial}
operators $\phi^6/\sigma^2$ and $\phi^8/\sigma^4$, etc, allowed by scale symmetry,
showing that such theories are non-renormalizable.
These operators can be expanded about the non-zero $\langle\phi\rangle$
and $\langle\sigma\rangle$,
to obtain an infinite series of effective polynomial ones,
suppressed by $\langle\sigma\rangle\gg\langle\phi\rangle$
(such hierarchy can be enforced by one initial, classical tuning of the couplings).
The non-polynomial operators emerge
from evanescent interactions ($\propto\epsilon$) between $\phi$
and $\sigma$ that vanish in $d=4$ but
are demanded by scale invariance in $d=4-2\epsilon$.
Previous works also showed that the higgs mass is stable against quantum corrections
at one-loop $m_\phi^2\ll \langle\sigma\rangle^2$,
which may remain true beyond it if only spontaneous scale symmetry breaking is present.
We checked the consistency of the two-loop scale invariant potential
by showing that it satisfies the Callan-Symanzik equation in its scale-invariant formulation.
To this purpose we computed the two-loop beta functions of the couplings of the theory.
While one-loop beta functions are exactly those of the same theory of $\phi$, $\sigma$
regularized with $\mu\!=$constant, the two-loop beta functions
differ from those of the theory regularized with explicit breaking of scale symmetry
($\mu=$constant). In conclusion, the running of the couplings enables one to distinguish
between spontaneous and explicit breaking of quantum scale symmetry of the action.
\begin{center}
--------------------------------
\end{center}
\section*{Appendix:}
\defC-\arabic{equation}}\defA{C{A-\arabic{equation}}
\defA{A}
\setcounter{equation}{0}
\subsection{Coefficients of the counterterms}\label{appendixA}
Assuming $\lambda_n=0$, $n \geqslant 6$ at tree-level,
the coefficients of eqs.(\ref{Z1}), (\ref{bb}) are:
\begin{eqnarray}
\delta_1^\phi &=& - \frac{3}{2} \left( \lambda_\phi^2 + \lambda_m^2 + 2 \frac{\lambda_m^3}{\lambda_\phi}\right) \\
\delta_1^m &=& - \frac{1}{4} \left( \lambda_\phi^2 + 6 \lambda_\phi\lambda_m + 10\lambda_m^2 + 6\lambda_m \lambda_\sigma + \lambda_\sigma^2 \right) \\
\delta_1^\sigma &=& -\frac{3}{2} \left( 2\frac{\lambda_m^2}{\lambda_\sigma} + \lambda_m^2 + \lambda_\sigma^2 \right) \\
\nu_1^\phi &=& \frac{1}{8} \left[ \frac{\lambda_m^2}{\lambda_\phi}(24 \lambda_m - 7 \lambda_\sigma)
-(14\lambda_m^2 - 16 \lambda_m \lambda_\sigma + 3 \lambda_\sigma^2)-\lambda_\phi(80\lambda_m - 6 \lambda_\sigma) \right] \\
\nu_1^m &=& - \frac{1}{24} \left[123 \lambda_m^2 + 86 \lambda_m \lambda_\sigma +3 \lambda_\sigma^2 + 6\lambda_\phi(8 \lambda_m + \lambda_\sigma)\right] \\
\nu_1^\sigma &=& -\frac{1}{8 \lambda_\sigma} \left( 48 \lambda_m^3 + 4 \lambda_m^2 \lambda_\sigma + 21 \lambda_\sigma^3 \right) \\
\nu_1^6 &=& \frac{1}{16} \frac{\lambda_\phi \lambda_m}{\lambda_6} \left(7 \lambda_\phi - 14 \lambda_m + \lambda_\sigma \right) \\
\nu_1^8 &=& \frac{1}{8} \frac{\lambda_\phi \lambda_m^2}{\lambda_8} \\
\delta_2^\phi &=& \frac{3}{4} \left[ 3\lambda_\phi^2 + 4\lambda_m^2 + \frac{\lambda_m^2}{\lambda_\phi} (4 \lambda_m + \lambda_\sigma)\right] \\
\delta_2^m &=& \frac{1}{4}\left(2\lambda_\phi^2 + 6\lambda_\phi \lambda_m + \lambda_\phi \lambda_\sigma + 19 \lambda_m^2 + 6\lambda_m \lambda_\sigma + 2 \lambda_\sigma^2 \right) \\
\delta_2^\sigma &=& \frac{3}{4}\left[ \frac{\lambda_m^2}{\lambda_\sigma} (\lambda_\phi + 4 \lambda_m) + 4\lambda_m^2 + 3 \lambda_\sigma^2 \right].
\end{eqnarray}
\defC-\arabic{equation}}\defA{C{B-\arabic{equation}}\defA{B}
\subsection{The finite part of the two-loop potential}
\setcounter{equation}{0}
\label{appendixB}
We provide here the finite part of the two-loop potential,
$V^{(2)}+V^{(2,n)}$ of eq.(\ref{V2}), (\ref{finalU}).
This is rather long, we thus use a simplified notation.
The propagators are found from:
$(\widetilde D_p)_{ij}=p^2\delta_{ij}-\widetilde V_{ij}$.
To simplify the calculation it helps to write them as
\begin{eqnarray}
&&(\widetilde D^{-1})_{ij} = \frac{\tilde a_{ij}}{p^2-\widetilde V_p}
+\frac{\tilde b_{ij}}{p^2-\widetilde V_m},
\qquad
\tilde b_{ij}=\delta_{ij}-\tilde a_{ij}
\nonumber\\
\tilde a_{11}\!\! &=&\!\!
\tilde b_{22} = \frac{\widetilde V_p-\widetilde V_{22}}{\widetilde V_p-\widetilde V_m},\quad
\tilde a_{22} = \tilde b_{11} = 1 - \tilde a_{11}
=\frac{\widetilde V_p-\widetilde V_{11}}{\widetilde V_p-\widetilde V_m},\quad
\tilde a_{12}=\tilde a_{21}=\frac{\widetilde V_{12}}{\widetilde V_p -\widetilde V_m},
\quad
\end{eqnarray}
\smallskip\noindent
where $\widetilde V_p$, $\widetilde V_m$ are the field-dependent eigenvalues of matrix
$\widetilde V_{ij}=\partial^2
\widetilde V/\partial s_i\partial s_j$, $i,j=1,2; \,\,s_1=\phi, s_2=\sigma$,
and $\widetilde V=\mu(\sigma)^{2\epsilon}\,V$ where $V$ is the tree level potential in $d=4$.
We introduce the following coefficients (without \, $\widetilde{}$\,\,)
of the Taylor expansions in $\epsilon$
(see Appendix~\ref{appendixC} for their values in terms of the couplings and fields):
\begin{eqnarray}\label{intro}
\tilde a_{ij}&=&a_{ij}+\epsilon\,a_{ij}^1+\epsilon^2\,a_{ij}^2+\cO(\epsilon^3),
\qquad
b_{ij}=\delta_{ij}-a_{ij}, \quad b_{ij}^1=-a_{ij}^1, \quad b_{ij}^2=-a_{ij}^2
\nonumber\\
\widetilde V_{ijk\ldots}\!\!\!\!&=&\!\!\!
\mu(\sigma)^{2\epsilon}\, \big[\,
v_{ijk\ldots}+\epsilon\,u_{ijk\ldots}+\epsilon^2\,w_{ijk\ldots}+\cO(\epsilon^3)\,\big],
\qquad \text{where:}
\nonumber\\
\widetilde V_{ijk\ldots}\!\!&=&\!\!
\frac{\partial^4 \widetilde V}{\partial s_i\partial s_j \partial s_k \cdots},
\quad
v_{ijk\ldots}=
\frac{\partial^4 V}{\partial s_i \partial s_j \partial s_k \cdots},
\quad
i,j,k,\ldots=1,2;
\quad
s_1\!=\!\phi,\, s_2\!=\!\sigma,
\nonumber\\
\widetilde V_p&=&\mu(\sigma)^{2\epsilon}\, V_p \,\big[
1+ c_p^1\,\epsilon +c_p^2\,\epsilon^2+\cO(\epsilon^3)\big]
\nonumber\\
\widetilde V_m&=&\mu(\sigma)^{2\epsilon}\, V_m \,\big[
1+ c_m^1\,\epsilon + c_m^2\, \epsilon^2+\cO(\epsilon^3)\big].
\end{eqnarray}
Here $V_p$ and $V_m$ are the field-dependent eigenvalues of
the matrix $V_{ij}$ of the-tree level $V$:
\begin{eqnarray}\label{Vp}
V_p=1/2 \Big[ V_{11}+V_{22}+\big[ (V_{11}-V_{22})^2+4 V_{12}^2\big]^{1/2}\Big]
\end{eqnarray}
with $V_m$ of similar expression but with $-$ in front of the square root.
$V_p$ and $V_m$ should not be confused
with derivatives $V_i$ of the potential.
We also use the notation
\begin{eqnarray}
\overline\ln A = \ln\frac{A}{t (z\sigma)^2}-1,
\qquad t=4\pi e^{-\gamma_E} \,.
\end{eqnarray}
\medskip\noindent
Then $V^{(2)}$ and $V^{(2,n)}$ of eq.(\ref{V2}), (\ref{finalU})
are shown below. $V^{(2)}$ is that of the
theory regularized with $\mu=$constant, while $V^{(2,n)}$ is a new correction.
They are sums of the diagrams of eq.(\ref{twoloop})
\begin{eqnarray}\label{B5}
V^{(2)} & = & V_{2, old}^a+V_{2, old}^b+V_{2, old}^c
\nonumber\\
V^{(2,\text{n})} & = & V_{2,n}^a+V_{2,n}^b+V_{2,n}^c
\end{eqnarray}
where $a,b, c,$ label the sunset ($a$), snowman ($b$), counterterm ($c$) diagrams,
respectively.
Then, in terms of the above coefficients, one finds for the sunset diagram ($a$):
\begin{eqnarray}
V^a_{2,n} & =& \!\frac{1}{4 \kappa^2} \, \Big\{
\Big[ v_{ij}w_{lmn} + \frac{1}{2} u_{ijk}u_{lmn} \Big]\big[ V_p a_{il} + V_m b_{il} \big]
\big[ a_{jm} + b_{jm} \big] \big[ a_{kn} + b_{kn} \big]
\nonumber\\
& +&\!\!\! v_{ijk} u_{lmn} \Big[
\Big( V_p\big[ a_{il}(1 \!-\! 2\overline\ln V_m \!+\! c_p^1)+ a^1_{il} \big]
+ V_m \big[b_{il}(1 \!-\! 2\overline\ln V_m \!+\! c_m^1) + b^1_{il} \big]
\Big)\big( \!a_{jm}\! +\! b_{jm}\! \big)
\nonumber\\
& +&
2\big[ V_p a_{il} + V_m b_{il} \big]\big[ a_{jm}^1+b_{jm}^1 \big] \Big]
\big[ a_{kn} + b_{kn} \big]
+
v_{ijk} v_{lmn} \,\,
\Big[
\frac{1}{2}\Big( V_p\big[a_{il}(c_p^2-c_p^1)+a^1_{il}c_p^1
\nonumber\\
&-& 2\overline\ln V_p(a_{il}c_p^1 + a_{il}^1) \big]
+
V_m\big[ b_{il}(c_m^2-c_m^1) + b_{il}^1 c_m^1 -2 \overline\ln V_m(b_{il}c_m^1 + b_{il}^1) \big] \Big)
\big(a_{jm} + b_{jm} \big)
\nonumber \\
& +&
\Big(V_p a_{il} \big[c_p^1-2\overline\ln V_p\big] + V_m b_{il}\big[c_m^1-2\overline\ln V_m\big] \Big)
\big(a_{jm}^1+b_{jm}^1\big)\Big] \big[ a_{kn} + b_{kn} \big]
\nonumber\\
&+&
v_{ijk} v_{lmn}
\Big[
V_p \Big( b_{il} \big[2 a_{jm}^1 a^1_{kn} + a^2_{jm} b_{kn} + a^1_{jm} b_{kn} + 2 a^1_{jm} b^1_{kn} \big]
+ a_{il} \big[3a_{jm}^1 a_{kn}^1 + 4a^2_{jn} b_{kn}
\nonumber\\
&+& 2 b_{jm} b_{kn}^1 + 4 a^1_{jm}(b_{kn} + b^1_{kn})
+ 2 b_{jm} b_{kn}^2 \big]
+
a_{il} a_{jm} \big[3(a_{kn}^1 + a^2_{kn}) + 2(b_{kn}^1 + b^2_{kn}) \big] \Big)
\nonumber\\
&+ & \!\! V_m \Big( a_{il} \big[2 b_{jm}^1 b^1_{kn} + b^2_{jm} a_{kn} + b^1_{jm} a_{kn}
+ 2 b^1_{jm} a^1_{kn} \big]
+ b_{il} \big[3b_{jm}^1 b_{kn}^1 + 4b^2_{jn} a_{kn} + 2 a_{jm} a_{kn}^1
\nonumber\\
& +& 4 b^1_{jm}(a_{kn}
+ a^1_{kn}) + 2 a_{jm} a_{kn}^2 \big]
+
b_{il} b_{jm} \big[3(b_{kn}^1 + b^2_{kn}) + 2(a_{kn}^1 + a^2_{kn}) \big] \Big)\,
\Big] \;
\Big\}
\end{eqnarray}
\vspace{-0.3cm}
\noindent
and
\vspace{-0.2cm}
\begin{eqnarray}
V_{2,old}^{a}
&=&
\frac{1}{4 \kappa^2} \, v_{ijk} v_{lmn}
\Big\{
\big[\, V_p a_{il} \overline\ln V_p + V_m b_{il}\,\overline\ln V_m \big] \big[\, a_{jm}a_{kn}
\overline\ln V_p + b_{jm}b_{kn} \overline\ln V_m \, \big]
\nonumber\\
&+&
2 \big[\, v_p a_{il} \overline{\ln}^2 V_p +
V_m b_{il} \overline{\ln}^2 V_m \big] a_{jm} b_{kn}
+ \frac{1}{2}
\big[ V_p b_{il} - V_m a_{il} \big] \big[ \overline{\ln}^2 V_p
- \overline{\ln}^2 V_m\, \big] a_{jm} b_{kn}
\nonumber\\
&- &
\big[V_p a_{il} \overline\ln V_p + V_m b_{il} \overline\ln V_m \big] \big[a_{jm} +
b_{jm} \big] \big[ a_{kn} + b_{kn} \big]
\nonumber \\
&+ & \big[V_p a_{il} + V_m b_{il} \big]\big[a_{jm} + b_{jm} \big] \big[ a_{kn}
+ b_{kn} \big] \Big[ \frac{3}{2} + \frac{\pi^2}{12} \Big]
\nonumber \\
&-& \Big[V_p \big(2a_{il}\Phi_{p,m} - \frac{1}{2} b_{il} \Phi_{m,p} \big)
+ V_m \big(2b_{il}\Phi_{m,p} - \frac{1}{2} a_{il} \Phi_{p,m} \big) \Big] a_{jm}b_{kn}
\nonumber\\
&-& \frac{1}{3}\, \big[\, V_p a_{il}a_{jm}a_{kn} + V_m b_{il}b_{jm}b_{kn}\, \big] C
\Big\},
\end{eqnarray}
with \cite{DT}
\begin{eqnarray}
\Phi_{p,m} &=&
\begin{cases}
\sqrt{\frac{y_{pm}}{y_{pm}-1}}
\Big[-4\, \mathrm{Li}_2\left( \frac{1-\eta_{pm}}{2}\right)
+ 2\ln^2 \frac{1-\eta_{pm}}{2} - \ln^2 4y_{pm} + \frac{\pi^2}{3} \Big],& y_{pm}>1
\\[4pt]
4\sqrt{\frac{y_{pm}}{1-y_{pm}}}
\mathrm{Cl}_2\left(\arcsin \sqrt{y_{pm}}\right),& \!\!\!\!\! 1> y_{pm} \geqslant 0
\end{cases}\;,
\\
y_{pm}& =& {V_m/V_p},
\qquad \eta_{pm} = (1- 1/y_{pm})^{1/2},
\quad C = -2 \sqrt{3} \mathrm{Cl}_2\left( \pi/3 \right)\cong 3.5\,\,\,
\\
\mathrm{Li}_2(\xi) &=& -\int_0^1 \mathrm{d}t \frac{\ln(1-\xi t)}{t},
\qquad \mathrm{Cl}_2(\theta) = -\int_0^\theta \mathrm{d} \theta \ln
\left| 2 \sin \frac{\theta}{2} \right|. \,\,
\end{eqnarray}
\medskip\noindent
Further, for the snowman diagram ($b$):
\medskip
\begin{eqnarray}
\!\!\! V_{2,n}^b\!\!\!
&=& \!\!\!
\frac{1}{8\,\kappa^2}
\Big\{
w_{ijkl} \,\,(a_{ij} V_p+b_{ij} \,V_m)\,(a_{kl} V_p+b_{kl} \,V_m)\,
- 2\,
u_{ijkl}\, \Big[
\,
(a_{ij} V_p+b_{ij} \, V_m)
\nonumber\\
&\times&
\Big(\,
a_{kl}\, V_p \,\overline \ln V_p+ b_{kl}\, V_m\, \overline \ln V_m
-
\big[
(a_{kl} c_{p1}+a_{kl}^1)\,V_p+
(b_{kl} c_{p1}+b_{kl}^1)\,V_m\big]\Big)
\Big]
\nonumber\\
&-& 2\,
v_{ijkl}
\,\Big[
2\, a_{ij} \,(a_{kl}\, c_{p1} +a_{kl}^1)\, V_p^2 \,\overline \ln V_p
+2 b_{ij}\, (b_{kl}\, c_{m1} +b_{kl}^1)\, V_m^2\,\overline \ln V_m
\nonumber\\
&+&
\big[a_{ij}\, (b_{kl}\,c_{m1}+b_{kl}^1)+ b_{ij}\, (a_{kl}\, c_{p1}+a_{kl}^1)\,\big]
\,V_p\, V_m\,(\overline\ln V_p+\overline\ln V_m)
-
(a_{ij} V_p+b_{ij}\,V_m)\,
\nonumber\\
&\times &
\Big(
\big[ \,a_{kl}\, (c_{p2}-c_{p1}) + a_{kl}^1 c_{p1}+a_{kl}^2\big]\, V_p+
\big[\, b_{kl}\, (c_{m2}-c_{m1}) + b_{kl}^1 c_{m1}+b_{kl}^2\big]\, V_m\Big)
\nonumber\\
&-&\!\!\!\frac{1}{2}
\big[ ( a_{ij} c_{p1} +a_{ij}^1) V_p+ (b_{ij} c_{m1}+ b_{ij}^1) V_m\big]
\big[ (a_{kl} c_p^1 +a_{kl}^1 ) V_p + (b_{kl} c_{m1}+ b_{kl}^1) V_m \big]
\Big]
\Big\},\qquad
\end{eqnarray}
\medskip\noindent
where $ i,j,k,l=1,2.$ Also
\begin{eqnarray}
V_{2,old}^b
& =&
\frac{1}{8 \kappa^2} v_{ijkl}
\Big\{ \big[ V_p a_{ij} + V_m b_{ij} \big]
\big[ V_p a_{kl} + V_m b_{kl} \big] \bigg[ 1+\frac{\pi^2}{6} \bigg]
+ V_p V_m a_{ij} b_{kl} \big[ \overline\ln V_p - \overline\ln V_m \big]^2
\nonumber\\
&+&
2\, \big[\, V_p a_{ij} \overline\ln V_p + V_m b_{ij}
\overline\ln V_m\, \big]\big[ \, V_p a_{kl} \overline\ln V_p + V_m b_{kl} \overline\ln V_m\, \big]
\Big\}.
\end{eqnarray}
\medskip\noindent
For the final ``counter-term'' diagram ($c$)
we need to introduce the coefficients
$\delta v_{ij}$, $\delta u_{ij}$, $\delta w_{ij}$ whose values will be presented shortly
(Appendix C). From eq.(\ref{ZV})
\medskip
\begin{eqnarray}\label{co}
\delta V_1 &=&
\frac{1}{\varepsilon\,\kappa}\,\,\mu^{2\epsilon}
\Big[ \delta^\phi_0\, \frac{\lambda_\phi}{4!} \phi^4 +
\delta^m_0 \,\frac{\lambda_m}{4} \phi^2 \sigma^2
+ \delta^\sigma_0\, \frac{\lambda_\sigma}{4!} \sigma^4\Big], \,\,\,\text{then}
\nonumber\\
(\delta V_1)_{ij} &=&
\frac{1}{\varepsilon\,\kappa}\,\, \mu^{2\epsilon} \Big[\,
\delta v_{ij} + \epsilon\, \delta u_{ij}
+ \epsilon^2\, \delta w_{ij}\, \Big],
\end{eqnarray}
\medskip\noindent
where $(\delta V_1)_{ij}=\partial^2(\delta V_1)/\partial s_i \partial s_j$, $i,j=1,2$,
$s_1,s_2=\phi,\sigma$.
With this notation, we find for diagram $(c)$:
\begin{eqnarray}
V_{2,n}^c & = &
\frac{1}{2\kappa^2}
\Big\{ \delta w_{ij} \Big[ a_{ij} V_p + b_{ij} V_m \Big]
+ \delta u_{ij} \Big[ V_p\big( a_{ij}[c_p^1 - \overline\ln V_p ] + a_{ij}^1 \big)
\nonumber\\
& +& V_m \big( b_{ij}[c_m^1 - \overline\ln V_m ] + b_{ij}^1 \big)\Big]
- \delta v_{ij} \Big[
V_p \overline\ln V_p \big(a_{ij} c_p^1 + a_{ij}^1 \big) + V_m \overline\ln V_m \big(b_{ij}c^1_m + b_{ij}^1 \big)
\nonumber\\
& - &
V_p \big(a_{ij} c_p^2 + [a^1_{ij} - a_{ij}]c_p^1 + a^2_{ij} \big)
- V_m \big(b_{ij} c_m^2 + [b^1_{ij} - b_{ij}] c^1_m + b_{ij}^2 \big)
\Big]
\Big\}
\end{eqnarray}
and
\begin{eqnarray}
V_{2,old}^c
= \frac{1}{4 \kappa^2}
\delta v_{ij}\Big[ V_p a_{ij}\Big(\, \overline{\ln}^2 V_p + 1+\frac{\pi^2}{6} \Big)
+ V_m b_{ij}\Big( \overline{\ln}^2 V_m + 1+\frac{\pi^2}{6} \,\Big) \Big].
\end{eqnarray}
\defC-\arabic{equation}}\defA{C{C-\arabic{equation}}\defA{C}
\subsection{Coefficients entering the two-loop potential}
\setcounter{equation}{0}
\label{appendixC}
\noindent
Below we provide the expressions of the various
coefficients introduced in the rhs of eq.(\ref{intro}), (\ref{co}) and
used in Appendix~\ref{appendixB}.
The coefficients $v_{ijkl}$, $v_{ijk}$ are functions
of $\lambda$'s and fields
\medskip
\begin{gather}
\left[
\begin{array}{ccc}
v_{1111} & u_{1111} & w_{1111} \\
v_{1112} & u_{1112} & w_{1112} \\
v_{1122} & u_{1122} & w_{1122} \\
v_{1222} & u_{1222} & w_{1222} \\
v_{2222} & u_{2222} & w_{2222} \\
\end{array}
\right]
\!=\!
\left[
\begin{array}{ccc}
\lambda _{\phi } & 0 & 0 \\
0 & 2\lambda_\phi \frac{\phi}{\sigma } & 2\lambda_\phi \frac{ \phi}{\sigma } \\
\lambda _m & 3 \lambda _m- \lambda_\phi \frac{\phi ^2 }{\sigma ^2} &
\lambda_\phi\frac{\phi ^2 }{\sigma ^2}+5 \lambda _m \\
0 & \frac{2}{3}\lambda_\phi\frac{\phi ^3 }{\sigma ^3}+
2\lambda_m\frac{ \phi }{\sigma } & 8 \lambda_m \frac{ \phi }{\sigma }
- \frac{4}{3} \lambda_\phi\frac{\phi ^3}{ \sigma ^3} \\
\lambda _{\sigma } & -\frac{1}{2}\lambda_\phi \frac{\phi ^4}{\sigma ^4}
+\frac{25}{6}\lambda_\sigma-\lambda_m\frac{\phi ^2}{\sigma ^2} &
\frac{4}{3} \lambda_\phi \frac{\phi ^4}{\sigma ^4}+10 \lambda _{\sigma }- 2\lambda_m\frac{ \phi ^2 }{\sigma ^2} \\
\end{array}
\right]
\end{gather}
and
\medskip
\begin{gather}
\left[
\begin{array}{ccc}
v_{111} & u_{111} & w_{111} \\
v_{112} & u_{112} & w_{112} \\
v_{122} & u_{122} & w_{122} \\
v_{222} & u_{222} & w_{222} \\
\end{array}
\right]
\!=\!
\left[
\begin{array}{ccc}
\lambda_\phi \phi & 0 & 0 \\
\lambda_m \sigma & \lambda_\phi\frac{\phi ^2}{\sigma }+\sigma \lambda _m & \lambda_\phi
\frac{\phi ^2 }{\sigma }+ \lambda_m \sigma \\
\lambda_m \phi & 3 \lambda_m \phi -\frac{1}{3} \lambda_\phi\frac{\phi ^3}{\sigma ^2} &
\frac{1}{3}\lambda_\phi\frac{\phi ^3 }{ \sigma ^2}+5 \lambda_m \phi
\\
\lambda_\sigma \sigma & \frac{1}{6} \lambda_\phi \frac{\phi ^4 }{\sigma ^3}+\frac{13}{6}
\lambda_\sigma \sigma + \lambda_m \frac{\phi ^2 }{\sigma } & -\frac{1}{3}\lambda_\phi \frac{\phi ^4 }{ \sigma ^3}
+\frac{11}{3} \lambda_\sigma \sigma +4 \lambda_m \frac{ \phi ^2}{\sigma }
\\
\end{array}
\right].
\end{gather}
\medskip\noindent
Further, coefficients $\delta v_{ij}$, $\delta u_{ij}$ and $\delta w_{ij}$ of eq.(\ref{co}) are
\medskip
\begin{eqnarray}
\delta v_{11} \!\!&=&\!\!
\frac{1}{4\kappa} \left[
3\left(\lambda_\phi^2 + \lambda_m^2 \right) \phi^2 + \lambda_m \left(\lambda_\phi + 4\lambda_m + \lambda_\sigma \right) \sigma^2
\right]
\;, \quad
\delta u_{11} = \delta w_{11} = 0
\\
\delta v_{12} \!\!&=&\!\! \frac{1}{2\kappa} \lambda_m \left( \lambda_\phi + 4 \lambda_m + \lambda_\sigma \right)\phi \sigma \\
\delta u_{12} \!\!&=&\!\! \delta w_{12} = \frac{1}{2 \kappa} \left[ \left( \lambda_\phi^2 + \lambda_m^2 \right) \phi^2 +
\lambda_m \left(\lambda_\phi + 4 \lambda_m + \lambda_\sigma \right) \sigma^2
\right] \frac{\phi}{\sigma}
\\
\delta v_{22} \!\!&=&\!\! \frac{1}{4\kappa^2} \left[
\left(4 \lambda_m^2 + \lambda_\phi \lambda_m + \lambda_m \lambda_\sigma \right) \phi^2 +
3\left(\lambda_m^2 + \lambda_\sigma^2 \right)\sigma^2 \right]
\\
\delta u_{22} \!\!&=&\!\!
\frac{1}{8 \kappa^2} \frac{1}{\sigma^2}\left[
-\left( \lambda_\phi^2 + \lambda_m^2 \right)\phi^4 +
6\lambda_m\left( \lambda_\phi + 4\lambda_m + \lambda_\sigma \right)\phi^2 \sigma^2 +
7\left( \lambda_\sigma^2 + \lambda_m^2 \right) \sigma^4
\right]
\\
\delta w_{22} \!\!&=&\!\!
\frac{1}{8 \kappa^2} \frac{1}{\sigma^2}\left[
\left( \lambda_\phi^2 + \lambda_m^2 \right)\phi^4 +
10\lambda_m\left( \lambda_\phi + 4\lambda_m + \lambda_\sigma \right)\phi^2 \sigma^2 +
9\left( \lambda_\sigma^2 + \lambda_m^2 \right) \sigma^4
\right]\qquad
\end{eqnarray}
\bigskip\noindent
The
coefficients $c_p^1$, $c_p^2$, $c_m^1$, $c_m^2$, also $V_p$ and $V_m$ introduced in
eq.(\ref{intro}) have the expressions:
\medskip
\begin{eqnarray}
\left[
\begin{array}{c}
V_p\, c_p^1 \\
V_m\, c_m^1 \\
\end{array}
\right]
\!=\!
\frac{1}{24}\frac{1}{\sigma^2}
\left[
\begin{array}{c}
-\lambda_\phi \phi ^4 + 18 \lambda_m \phi ^2\sigma^2+7 \lambda_\sigma \sigma ^4 + R_1 \\
-\lambda_\phi \phi ^4 + 18 \lambda_m \phi ^2\sigma^2+7 \lambda_\sigma \sigma ^4 - R_1 \\
\end{array}
\right]
\end{eqnarray}
\medskip\noindent
where
\begin{eqnarray}
R_1& =& \frac{1}{2 S} \Big[
\lambda_\phi\left(\lambda_\phi - \lambda_m \right)\phi^6 +
\left( 15 \lambda_\phi \lambda_m - \lambda_\phi \lambda_\sigma + 18 \lambda_m^2 \right)\phi^4 \sigma^2
\nonumber\\
&+&
\left( -7\lambda_\phi\lambda_\sigma + 78\lambda_m^2 + 25\lambda_m\lambda_\sigma \right)\phi^2 \sigma^4 +
7\lambda_\sigma\left( -\lambda_m + \lambda_\sigma \right)\sigma^6
\Big]
\end{eqnarray}
and
\medskip
\begin{equation}\label{S}
S^2 = \frac{1}{4} \Big[
\left( \lambda_\phi - \lambda_m \right)^2 \phi^4 +
2\left( \lambda_\phi \lambda_m + 7 \lambda_m^2 - \lambda_\phi \lambda_\sigma + \lambda_m \lambda_\sigma \right) \phi^2 \sigma^2 +
\left(\lambda_m -\lambda_\sigma \right)^2 \sigma^4
\Big],
\end{equation}
\medskip\noindent
while $V_m$ and $V_p$ (see also (\ref{Vp})) are
\medskip
\begin{eqnarray}
\left[
\begin{array}{c}
V_p \\
V_m \\
\end{array}
\right]
\!=\!
\frac{1}{4}
\left[
\begin{array}{c}
\left(\lambda _{\sigma }+\lambda _m\right) \sigma ^2
+ \left(\lambda _{\phi }+\lambda _m\right)\phi ^2 + 2S \\
\left(\lambda _{\sigma }+\lambda _m\right) \sigma ^2
+ \left(\lambda _{\phi }+\lambda _m\right)\phi ^2 - 2S
\end{array}
\right].
\end{eqnarray}
For $c_p^2$, $c_m^2$ (with the above $V_p$ and $V_m$):
\begin{eqnarray}
\left[
\begin{array}{c}
V_p \, c_p^2 \\
V_m \, c_m^2 \\
\end{array}
\right]
\!=\!
\frac{1}{24}\frac{1}{\sigma^2}
\left[
\begin{array}{c}
\lambda_\phi \phi ^4 + 30 \lambda_m \phi ^2\sigma^2 + 9 \lambda_\sigma \sigma ^4 + R_2\\
\lambda_\phi \phi ^4 + 30 \lambda_m \phi ^2\sigma^2 + 9 \lambda_\sigma \sigma ^4 - R_2
\end{array}
\right],
\end{eqnarray}
where
\begin{eqnarray}
\!\! R_2 &=& \!\!\frac{1}{\sigma^2 S^3}
\Big[
\sigma ^2 \phi ^{10} \lambda _{\phi } \big(13 \lambda _{\phi }^3
+27 \lambda _{\phi } \lambda _m^2-39 \lambda _{\phi }^2 \lambda _m+3 \lambda _m^3\big)
+ \sigma ^4 \phi ^8 \big(-23 \lambda _{\sigma } \lambda _{\phi }^3
\nonumber\\
& + &
3 \lambda _{\phi } \lambda _m^2 \left[3 \lambda _{\sigma }+16 \lambda _{\phi }\right]
+5 \lambda _{\phi }^2 \lambda _m \left[6 \lambda _{\sigma }
+25 \lambda _{\phi }\right]-423\lambda _{\phi } \lambda _m^3+90 \lambda _m^4\big)
\nonumber\\
&+&
\sigma ^6 \phi ^6 \big(\lambda _{\sigma } \lambda _{\phi }^2
\left[7 \lambda _{\sigma }-27 \lambda _{\phi}\right]+3 \lambda _m^3 \left[99 \lambda _{\sigma }+151 \lambda _{\phi }\right]
+\lambda _{\phi } \lambda _m^2 \left[553 \lambda _{\phi }-
615 \lambda _{\sigma }\right]
\nonumber\\
&+ &
\lambda _{\sigma } \lambda _{\phi } \lambda _m
\left[9 \lambda _{\sigma }+65\lambda _{\phi }\right]+2034 \lambda _m^4\big)
+
\sigma ^8 \phi ^4 \big(3 \lambda _{\sigma }^2
\lambda _{\phi } \left[\lambda _{\sigma }+27 \lambda _{\phi }\right]
+3 \lambda _m^3 \left[641 \lambda _{\sigma } + 261 \lambda _{\phi }\right]
\nonumber\\
&+ &
\lambda _{\sigma } \lambda _m^2 \left[351 \lambda _{\sigma }
-521\lambda _{\phi }\right]-\lambda_{\sigma } \lambda _{\phi }
\lambda _m \left[361 \lambda_{\sigma }+81 \lambda _{\phi }\right]+3438 \lambda_m^4\big)
+
\sigma ^{10} \phi ^2 \big(-81 \lambda _{\sigma }^3 \lambda_{\phi }
\nonumber\\
&+&
\lambda _{\sigma } \lambda _m^2 \left[292 \lambda _{\sigma }-81\lambda _{\phi }\right]
+
9 \lambda _{\sigma}^2 \lambda _m \left[19 \lambda _{\sigma }
+18 \lambda _{\phi }\right]-609 \lambda _{\sigma } \lambda _m^3+342 \lambda _m^4\big)
\nonumber\\
&-&
27 \sigma ^{12} \lambda _{\sigma } \left[\lambda _m-\lambda _{\sigma }\right]^3
\Big].
\end{eqnarray}
\medskip\noindent
Finally,
$a_{ij}$, $a_{ij}^1$, $a_{ij}^2$, $b_{ij}$, $b_{ij}^1$, $b_{ij}^2$
introduced in (\ref{intro}) and used in Appendix~\ref{appendixB} have the values
\medskip
\begin{eqnarray}
a_{11}& =& 1-a_{22} = b_{22}=1 - b_{11} =
\frac{1}{2} + \frac{1}{4S}\big[\lambda_\phi \phi^2 + \lambda_m(-\phi^2 + \sigma^2) - \lambda_\sigma \sigma^2 \big]
\\
a_{12}& =& a_{21} = - b_{12} = - b_{21} =\frac{\lambda_m \phi\, \sigma}{S}
\\
a_{11}^1 &=& \!\! -a_{22}^1 = - b_{11}^1 = b_{22}^1
=\frac{\lambda_m \phi^2}{6\, S^3}
\Big[
\lambda_\phi \big(-2\lambda_\phi + 3 \lambda_m \big) \phi^4
\nonumber\\
&& \hspace{2cm}+\, 2\big(\lambda_\phi\lambda_\sigma -4\lambda_\phi\lambda_m
-
6 \lambda_m^2 \big)
\phi^2 \sigma^2 -\big(6\lambda_m^2 + \lambda_m \lambda_\sigma \big) \sigma^4
\Big]
\end{eqnarray}
with $S$ of eq.(\ref{S}).
Also
\begin{eqnarray}
a_{12}^1 &=& a_{21}^1 = -b_{12}^1 = - b_{21}^1 = \frac{\phi}{24 \sigma S^3} \Big[
\phi ^6
\lambda _{\phi } \left(2 \lambda _{\phi }^2-5 \lambda
_{\phi } \lambda _m+3 \lambda _m^2\right)\qquad\qquad
\nonumber\\
&+&
\phi ^4 \sigma ^2
\left(-4 \lambda _{\sigma } \lambda _{\phi }^2+5 \lambda _{\phi } \lambda _m \left[\lambda _{\sigma }+2
\lambda _{\phi }\right]+\lambda _{\phi } \lambda _m^2-12 \lambda _m^3\right)
\nonumber \\
&+&
\phi ^2 \sigma ^4
\left(2 \lambda _{\sigma }^2 \lambda _{\phi }+\lambda _m^2 \left[14
\lambda _{\phi }-13 \lambda_{\sigma }\right]-9 \lambda _{\sigma } \lambda _{\phi } \lambda _m+6 \lambda _m^3\right)
\nonumber \\
&+&
\sigma ^6
\lambda _m \left(-\lambda _{\sigma }^2-5 \lambda _{\sigma } \lambda_m+6 \lambda _m^2\right)
\Big].
\end{eqnarray}
Further
\begin{eqnarray}
a_{11}^2 &=& -a_{22}^2=-b_{11}^2 = b_{22}^2
=
\frac{\phi^2}{288 \sigma^2 S^5}
\Big\{ \phi ^{10}\, \lambda _{\phi }^2 \,
\big(-4 \lambda _{\phi }^3
-31 \lambda _{\phi } \lambda _m^2
+20 \lambda _{\phi }^2 \lambda _m+15 \lambda _m^3\big)
\nonumber \\
&+&
\!\! \phi ^8 \sigma ^2 \lambda _{\phi }\, \Big[
12 \lambda _{\sigma } \lambda _{\phi }^3
+\lambda _{\phi } \lambda _m^2 \left[31 \lambda _{\sigma }+180 \lambda _{\phi }\right]
-
20 \lambda _{\phi }^2 \lambda _m \,\left[2 \lambda _{\sigma }
+3 \lambda _{\phi }\right] +5 \lambda _{\phi } \lambda _m^3-192 \lambda
_m^4\Big]
\nonumber \\
& +&
2 \phi ^6 \sigma ^4
\Big[ -6 \lambda _{\sigma }^2 \lambda _{\phi }^3
+\lambda _{\phi } \lambda _m^3 \left[44 \lambda _{\phi }
-163 \lambda _{\sigma }\right]-\lambda _{\phi }^2 \lambda _m^2
\left[19 \lambda _{\sigma }+132 \lambda _{\phi }\right]
\nonumber \\
& + &
2 \lambda _{\sigma } \lambda _{\phi }^2 \lambda _m
\left[5 \lambda _{\sigma }+28 \lambda _{\phi }\right]+450
\lambda _{\phi } \lambda _m^4+144 \lambda _m^5\Big]
-
2 \phi ^4 \sigma ^6
\Big[ -2 \lambda _{\sigma }^3 \lambda _{\phi }^2
+\lambda _m^4 \left[270 \lambda _{\phi }
\right.\nonumber\\
& -& \left.
96 \lambda _{\sigma }\right]+\lambda _{\phi } \lambda _m^3
\left[236 \lambda _{\phi }-463 \lambda _{\sigma }\right]
+
\lambda _{\sigma } \lambda _{\phi } \lambda _m^2
\left[71 \lambda _{\sigma }-118 \lambda _{\phi }\right]+22
\lambda _{\sigma }^2 \lambda _{\phi }^2 \lambda _m+1008 \lambda _m^5\Big]
\nonumber \\
&+&
\phi ^2 \sigma ^8
\lambda _m\, \Big[
-8 \lambda _{\sigma }^3\lambda _{\phi }+12 \lambda _m^3
\left[29 \lambda _{\sigma }-31 \lambda_{\phi }\right]
+
\lambda _{\sigma } \lambda _m^2 \left[184 \lambda _{\phi}
-117 \lambda _{\sigma }\right]+49 \lambda _{\sigma }^2 \lambda _{\phi }\lambda _m
\nonumber\\
& -& 468 \lambda _m^4\Big]
+ \sigma ^{10}
3\lambda _m^2 \big(-7 \lambda _{\sigma }^3+16 \lambda _{\sigma} \lambda _m^2
+27 \lambda _{\sigma }^2 \lambda _m-36 \lambda_m^3\big)
\Big\}.
\end{eqnarray}
Finally
\begin{eqnarray}
a_{12}^2 &=& a_{21}^2 = -b_{12}^2 = -b_{21}^2
=
-
\frac{\phi}{576\sigma^3 S^5} \Big[
- \phi ^{12} \lambda _{\phi }^2 \left(3 \lambda _m
-2 \lambda _{\phi }\right) \big(\lambda _m-\lambda _{\phi }\big)^2
\nonumber\\
&+&
2 \sigma ^2 \phi ^{10} \lambda _{\phi } \big(
-3 \lambda _{\phi }^3 \left[\lambda _{\sigma }+
2 \lambda _{\phi }\right]-\lambda _{\phi } \lambda _m^2 \left[4 \lambda_{\sigma }
+99 \lambda _{\phi }\right]
+\lambda _{\phi }^2 \lambda _m \left[7 \lambda _{\sigma }
+57 \lambda _{\phi }\right]
\nonumber\\
&+ &
22 \lambda _{\phi } \lambda _m^3+30 \lambda _m^4\big)
+
\sigma ^4 \phi ^8 \big(2 \lambda _{\sigma } \lambda _{\phi }^3
\left[3 \lambda _{\sigma }+17 \lambda _{\phi }\right]+\lambda _{\phi }
\lambda _m^3 \left[160 \lambda _{\sigma }+301 \lambda _{\phi }\right]
\nonumber\\
& +&
2 \lambda _{\phi }^2 \lambda _m^2 \left[28 \lambda _{\sigma }+237 \lambda
_{\phi }\right]
-\lambda _{\phi }^2 \lambda _m
\left[184 \lambda _{\sigma } \lambda _{\phi }+7 \lambda _{\sigma }^2
+84 \lambda _{\phi }^2\right]-972 \lambda _{\phi } \lambda _m^4-72 \lambda _m^5\big)
\nonumber\\
&+&
2 \sigma ^6 \phi ^6 \big(-\lambda_{\sigma }^2 \lambda _{\phi }^2
\left[\lambda _{\sigma }+15 \lambda_{\phi }\right]-6 \lambda _m^4 \left[7 \lambda _{\sigma }
-68 \lambda _{\phi }\right] +\lambda _{\phi } \lambda _m^3 \left[379 \lambda _{\phi }
-651 \lambda _{\sigma }\right]
\nonumber\\
&&+\lambda _{\phi } \lambda _m^2 \left[-5 \lambda _{\sigma }
\lambda _{\phi }+71 \lambda _{\sigma }^2-108 \lambda _{\phi}^2\right]
+\lambda _{\sigma } \lambda _{\phi }^2 \lambda _m \left[13
\lambda _{\sigma }+75 \lambda _{\phi }\right]+1116 \lambda _m^5\big)
\nonumber \\
&+&
\sigma ^8 \phi ^4 \big(6 \lambda _{\sigma }^3 \lambda _{\phi }^2+12
\lambda _m^4 \left[92 \lambda _{\sigma }+45 \lambda _{\phi }\right]
+\lambda _m^3 \left[540 \lambda_{\sigma } \lambda _{\phi }
+59 \lambda _{\sigma }^2-264 \lambda _{\phi }^2\right]
\nonumber \\
&&-6 \lambda _{\sigma } \lambda _{\phi } \lambda _m^2 \left[91
\lambda _{\sigma }-41 \lambda _{\phi }\right]+\lambda _{\sigma }^2 \lambda _{\phi } \lambda _m \left[44
\lambda _{\sigma }-37 \lambda _{\phi }\right]+324 \lambda _m^5\big)
\nonumber \\
&+&
2 \sigma ^{10} \phi ^2 \big(\lambda _{\sigma}^4 \lambda _{\phi }
+6\lambda _m^4 \left[31 \lambda _{\sigma }-13 \lambda _{\phi }\right]
+\lambda _{\sigma } \lambda _m^3 \left[89 \lambda _{\phi
}-103 \lambda _{\sigma }\right]
\nonumber \\&&+\lambda _{\sigma }^2 \lambda _m^2 \left[41
\lambda _{\sigma }+8 \lambda _{\phi }\right]-20 \lambda _{\sigma }^3 \lambda _{\phi } \lambda _m
+72 \lambda _m^5\big)
\nonumber \\
&+&
\sigma ^{12} \left(-\lambda _m\right) \left(\lambda _m
-\lambda_{\sigma}\right)^2 \left(-11 \lambda _{\sigma }^2
+24 \lambda _{\sigma } \lambda _m+ 36 \lambda _m^2\right)
\Big],
\end{eqnarray}
\medskip\noindent
which enter in the expression of the two-loop potential.
\bigskip\bigskip
\noindent
{\bf Acknowledgements: }
The work of D. Ghilencea was supported by a grant from Romanian National Authority for
Scientific Research (CNCS-UEFISCDI) under project number PN-II-ID-PCE-2011-3-0607.
The work of P. Olszewski and Z. Lalak was supported by the Polish NCN grants DEC-2012/04/A/ST2/00099
and 2014/13/N/ST2/02712.
|
\section{introduction}
The ability to efficiently generate and manage a combination of theoretical and experimental data is the foundation for data driven discovery of new materials and functions as well as methods to control manufacturing processes.\cite{MRS:7963345,curtarolo:nmat_review}
This formidable task requires a continuous feedback loop where descriptors\cite{curtarolo:nmat_review} of the functional properties are calculated for an enormous number of materials configurations, integrated in the databases,\cite{Curtarolo2012227,Curtarolo2012218,Jain2013,Pizzi2016218} compared with the available experiments, and exploited in the prediction cycle. In this paper we focus on a broad class of descriptors derived from the electronic structure calculations in order to provide easier integration with the experimental data. We introduce tight-binding methodologies for the calculation of electronic transport properties and the simulation of optical spectroscopies in the broadest energy range and with excellent accuracy as well as high computational efficiency.
The prerequisite for the simulation of both electron transport and optical properties is the accurate evaluation of the electronic structure of the system that is obtained by a fully self-consistent quantum-mechanical calculation either within density functional theory (DFT) or other first principles approaches. The electronic structure of solids is often described with Fourier basis functions that account naturally for the the periodicity and whose completeness is easily improvable up to any desirable accuracy. The delocalized character of plane waves, however, is not appropriate for the description of highly localized electronic systems unless a very large number of basis functions is used. The development of minimal-space solutions such as atomic orbital (AO) Bloch sums, which are capable of capturing with satisfactory accuracy the properties of solids and molecules on finite Hilbert spaces, has been central to methodological developments in quantum chemistry and solid state physics for many decades.
Atomic orbital basis sets provide the foundation for methods, such as tight-binding (TB), which combine
an intuitive physical representation of the interactions, low computational cost, and interesting alternatives for the study of the electronic structure of molecules and solids.\cite{Huckel:1931ft, Jones:1934go,Harrison:1989}
The TB Hamiltonian matrix is the central quantity that provides a compact real space representation of the many-body interactions, the accuracy of the electronic structure relies on the quality of such matrix.
Albeit computationally inexpensive and very intuitive for simple compounds, semiempirical TB implementations often fail in the prediction of electronic structure of complex materials and, in most cases, lack of predictive value when dealing with structural and chemical modifications as well as charge rearrangements.
In recent years, the reliability of the TB models has been largely improved with the introduction of \textit{ab initio} TB Hamiltonians derived from fully self-consistent quantum-mechanical calculations through a mapping into a much smaller space spanned by a set of atomic or atomic-like orbitals.
This procedure combine the accuracy and the predictive value of first principles approaches with the low computational cost of TB techniques. Furthermore, it is particularly useful for the evaluation of properties such as the electron conductivity and the optical absorption spectra that require a precise and ultrafine reciprocal space integration, typically very computationally expensive.
We recently developed a straightforward, non-iterative projection scheme that can exactly represent the first principles electronic structure of a periodic system on a finite AO-like basis.\cite{Agapito_2013_projectionsPRB, Agapito:2016jl, Agapito:2016en} By filtering the projections of Bloch states with high-kinetic-energy components and tuning the richness of the finite Hilbert space, we construct fully first principles TB Hamiltonians where the number of exactly reproduced bands with respect to the original DFT calculation can be selectively increased at a negligible computational cost. This provides cost effective solutions to design efficient algorithms for electronic structure simulations of realistic material systems and massive high-throughput investigations.
Our technique does not seek construction of (heavily customized, localized) basis functions. Its value resides on allowing non-iterative reproduction of a large number of energy bands using standard quantum-chemistry basis sets or the pseudo-atomic orbitals (PAO) of a standard pseudopotential calculation. Practically, the present methodology completely supersedes the need for engineered basis functions such as maximally localized Wannier functions\cite{Marzari:1997p231} or muffin-tin orbitals of arbitrary order (NMTO)\cite{Andersen:2000fc} in the context of the evaluation of transport \cite{Mostofi:2014hw,Calzolari:2004p39,Pizzi:2014ee} or optical properties.
Moreover, the knowledge of a localized orbital representation that maps seamlessly the electronic structure onto a localized AO basis set that is eventually fitted to a Gaussian basis set, opens the way to the fast (analytical) computation of two-electron integrals for solid-state applications and it is at the core of the development of the accurate and efficient ACBN0 functional.\cite{Agapito:2015iz}
The paper is organized as follows: in Sec. \ref{method} we will discuss the theoretical background and the practical implementation of the procedure with representative test cases. In Sec. \ref{results} we study, with our methodology, two significant materials problems in order to show the importance of fine reciprocal space sampling and the computational efficiency of the PAO projection to deal with very large systems. We chose to study the thermolectric properties of CoSb$_3$ and the transport and optical properties of core-shell nanowire of ZnO and ZnS. Finally in Sec. \ref{conclusions} we outline the conclusion of the present work.
\section{Methodology} \label{method}
\subsection{TB representation from PAO projections}
Accurate TB Hamiltonian matrices can be built from the direct projection of the Kohn-Sham (KS) Bloch states $\ket{\psi_{n\mathbf{k}}}$ onto a chosen basis set of fixed localized functions, as we discussed extensively in Ref.~\onlinecite{Agapito_2013_projectionsPRB,Agapito:2016jl,Agapito:2016en}.
There, we have shown that the real space Hamiltonians $ \hat{H}\left({\bf r}_{\alpha}\right)$ can be directly calculated using atomic orbitals or PAOs from the pseudopotential of any given element. The key in this procedure is in the mapping of the {\it ab initio} electronic structure (solved on a well converged and large plane waves basis set) into a model that precisely reproduces a selected number of bands of interest.\cite{Agapito_2013_projectionsPRB, Agapito:2016jl}
The crucial quantities that measure the accuracy of the basis set are the projectabilities $p_{n\mathbf{k}}=\bra{\psi_{n\mathbf{k}}} \hat{P} \ket{\psi_{n\mathbf{k}}} \ge 0$ ($\hat{P}$ is the operator that projects onto the space of the PAO basis set, as defined in
Ref.~\onlinecite{Agapito:2016jl}
) which indicate the representability of a Bloch state $\ket{\psi_{n\mathbf{k}}}$ on the chosen PAO set.
Maximum projectability, $p_{n\mathbf{k}}= 1$, indicates that the particular Bloch state can be perfectly represented in the chosen PAO set; contrarily, $p_{n\mathbf{k}} \approx 0$ indicates that the PAO set is insufficient and should be augmented.
Once the Bloch states with good projectabilities have been identified, the TB Hamiltonian is constructed as:
\begin{equation}
\hat{H}(\mathbf{k}) = AEA^\dagger + \kappa \left( I-A \left( A^{\dagger}A \right)^{-1}A^\dagger \right)
\label{eq:Hk}
\end{equation}
where $E$ is the diagonal matrix of KS eigenenergies and $A$ is the matrix of coefficients obtained from projecting the Bloch wavefunctions onto the PAO set (See Ref.~\onlinecite{Agapito:2016jl}.) Since the filtering procedure introduces a null space, the parameter $\kappa$ is used to shift all the unphysical solutions outside a given energy range of interest.
The procedure provides an accurate real space representation of the {\it ab initio} Hamiltonian $\hat{H}\left({\bf r}_{\alpha}\right)$ as a TB matrix of very small dimension, a crucial advantage for the accurate calculation of any physical properties that requires the precise integration in the reciprocal space.
By exploiting the PAO projection scheme described above we can easily Fourier transform the TB real space representation, interpolate to arbitrary precision, and perform derivatives in reciprocal space. For example,
the expectation value of the momentum operator, which is the main quantity in the definition of both and transport and optical descriptors described below, is computed as :
\begin{eqnarray}\label{momentum}
{\bf p}_{nm} ({\bf k})&=& \left < \psi_n ({\bf k}) \right |\hat{p}\left | \psi_m ({\bf k}) \right > = \\ \nonumber
&=& \left < \psi_n ({\bf k}) \right |\frac{m_0}{\hbar} \V{\nabla}_{\bf k} \hat{H}({\bf k})\left | \psi_m ({\bf k}) \right >
\end{eqnarray}
with
\begin{equation} \label{Hamiltonian.gradient}
\V{\nabla}_{{\bf k}} \hat{H}({\bf k}) =
\sum_{\alpha} i{\bf r}_{\alpha} \exp\left({i{\bf k}\cdot {\bf r}_{\alpha}}\right) \hat{H}\left({\bf r}_{\alpha}\right).
\end{equation}
$\hat{H}\left({\bf r}_{\alpha}\right)$ being the real space TB matrix and $\left | \psi_n ({\bf k}) \right >$ the eigenstates of the Hamiltonian $\hat{H}({\bf k})$.
\subsection{Boltzmann transport} \label{boltzmann}
Within the semiclassical theory, the electrical conductivity can be evaluated by solving the Boltzmann equation that describes
the evolution of the distribution function $f$ of an electron gas under external electric field and in presence of scattering mechanisms.\cite{Parravicini:2000ud,Singh2001125,Madsen200667}
In the so-called scattering-time approximation, the conductivity tensor $\sigma_{ij}$ can be expressed as an integral over the Brillouin Zone (BZ):
\begin{equation} \label{boltzmann.conductivity.equation.generalized}
\sigma_{ij}= \frac{e^2}{4\pi^3} \int_{BZ} \tau \sum_n v_n^i({\bf k})v_n^j({\bf k}) \left( -\frac{\partial f_0}{\partial \epsilon}\right)d {\bf k},
\end{equation}
where $\tau$ is the relaxation time, $v_n^i({\bf k})$ is the {\em i}-th component of the electron velocity (${\bf v}_n$) corresponding to the {\em n}-th band for each {\bf k}-point in the BZ, $f_0$ is the equilibrium distribution function, and $\epsilon$ is the electron energy.
Generalizing Eq. (\ref{boltzmann.conductivity.equation.generalized}) it is also possible to define analogue expressions for the Seebeck-coefficient $S$ and the electron contribution to thermal conductivity $\kappa_{el}$.
Following the notation of Ref. \onlinecite{Mecholsky:2014jv}, we introduce the { generating tensors} ${\mathcal L}_{\alpha}$ ($\alpha = 0, 1, 2$):
\begin{equation} \label{generating.tensors}
\mathcal L_{\alpha} = \frac{1}{4\pi^3} \int \tau \sum_n {\bf v}_n({\bf k})\left[{\bf v}_n({\bf k}) \cdot \hat{{\bf e}}\right] \left( -\frac{\partial f_0}{\partial \epsilon}\right)\left[\epsilon_n-\mu\right]^{\alpha}d {\bf k},
\end{equation}
where $\hat{{\bf e}}$ is the direction of the external electric field and $\mu$ is the chemical potential. The coefficients $\sigma$, $S$ and $\kappa_{el}$ can be expressed as follows:
\begin{eqnarray}
\sigma &=& e^2\mathcal L_0 \\ \nonumber
S &=& -\frac{1}{T e} \left[\mathcal L_0\right]^{-1} \cdot \mathcal L_1 \\ \nonumber
\kappa_{el} &=& \frac{1}{T} \left(\mathcal L_2 - \mathcal L_1 \cdot \left[\mathcal L_0\right]^{-1} \cdot \mathcal L_1\right),
\end{eqnarray}
where $T$ is the temperature.
From Eqs (\ref{boltzmann.conductivity.equation.generalized}-\ref{generating.tensors}) it is evident that the evaluation of the transport properties requires an accurate integration
over a fine grid of k-point in the BZ, especially for highly dispersive bands as in metal systems. This becomes a trivial task using the TB representation from the PAO projections and Eq. (\ref{Hamiltonian.gradient}).
As a validation of this approach we have calculated the transport coefficients ($\sigma$, $S$, $\kappa_{el}$) of Silicon and compared with the results of the code BoltzWann, where the interpolation of the real space Hamiltonian is done in using Maximally Localized Wannier Functions as basis functions.\cite{Pizzi:2014ee} The results are summarized in Fig. \ref{conductivity} and show excellent agreement between the two approaches.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.95\columnwidth]{Fig1.pdf}
\vspace{5mm}
\caption{\small (Color online) Boltzmann conductivity, Seebeck coefficient, and electron thermal coefficient for Silicon calculated with BoltzWann and with our approach.}
\label{conductivity}
\end{center}
\end{figure}
\subsection{Ballistic transport}\label{ballistic}
Calculations of the ballistic electrical conductance {\it \`a la} Landauer are naturally built on a local representation of the electronic structure like the one provided by TB hamiltonians. Our procedure reduces the problem of calculating electron transport \cite{Nardelli:1999p58,Calzolari:2004p39} to a computationally inexpensive post-processing maintaining the predictive power and the accuracy of first principles methods. Briefly, using the Landauer approach the conductance is determined via the transmission function that can be
written as: \cite{Fisher:1981gm,Nardelli:1999p58}
\[
\mathcal{T}_{el} = {\rm Tr}(\Gamma_L G_C^r \Gamma_R G_C^a),
\]
where $G_C^{\{r,a\}}$ are the retarded and advanced Green's functions
of the conductor, respectively, and $\Gamma_{\{L,R\}}$ are functions
that describe the coupling of the conductor to the leads. The Green's
function for the whole system can be explicitly written
as:\cite{Datta:1997tk}
\begin{equation}
G_C = (\epsilon -H_C -\Sigma_L -\Sigma_R)^{-1}
\label{gconduct}
\end{equation}
where $\Sigma_{L}$ and $\Sigma_R$ are the self-energy terms due to the
semi-infinite leads.
Once the self-energy functions are known, the coupling functions
$\Gamma_{\{L,R\}}$ can be easily obtained as\cite{Datta:1997tk}
\[
\Gamma_{\{L,R\}} = {\rm i}[\Sigma_{\{L,R\}}^r - \Sigma_{\{L,R\}}^a].
\]
The expression of the self-energies can be deduced along the lines of
Ref. \onlinecite{Nardelli:1999p58} using the formalism of principal layers in the
framework of the surface Green's function matching
theory. We obtain:
\begin{equation}
\begin{array}{ccl}
\Sigma_L & = & H_{LC}^\dagger (\epsilon -H_{00}^L-(H_{01}^L)^\dagger
\overline T_L)^{-1} H_{LC}\\ \Sigma_R & = & H_{CR} (\epsilon
-H_{00}^R-H_{01}^R T_R)^{-1} H_{CR}^\dagger,\\
\end{array}
\end{equation}
where $H_{nm}^{L,R}$ are the matrix elements of the Hamiltonian
between the layer orbitals of the left and right leads respectively,
and $T_{L,R}$ and $\overline{T}_{L,R}$ are the appropriate transfer
matrices. The latter are easily computed from the Hamiltonian matrix
elements via an iterative procedure.\cite{Nardelli:1999p58} This approach has been extensively validated and it is standard procedure in many electronic structure software packages such as WanT,\cite{Calzolari:2004p39} Wannier90,\cite{Mostofi:2014hw} and Smeagol.\cite{Rocha:2006fk}
\subsection{Dielectric function}
The optical properties of a material are generally described in semiclassical linear response theory by the dielectric tensor $\epsilon(\omega, {\bf q})$ that is a complex function describing the optical-response of the material in the presence of an external electromagnetic field at a given
frequency $\omega$ and momentum ${\bf q}.$\cite{Parravicini:2000ud,Onida:2002zz}
Quantities such as the refraction index and the absorption spectrum are easily derived from the real and imaginary part of the dielectric tensor.
In the limit of long wavelength (i.e. negligible momentum transfer ${\bf q}$),
the optical properties of the material depends only on the frequency of the field.
The dielectric tensor can then be expressed in terms of the dielectric
susceptibility $\chi_{ij}(\omega)$:
\begin{equation} \label{dielectric.function}
\epsilon_{ij}(\omega) = 1 + 4 \pi \chi_{ij}(\omega).
\end{equation}
Following the seminal work by M. Graf and P. Vogl,\cite{Graf:1995tl} the imaginary part of $\chi(\omega)$ in the single particle approximation can be written as:
\begin{eqnarray} \label{imaginary.dielectric.susceptibility}
\operatorname{Im} \chi_{ij}(\omega) &=& \frac{e^2 \pi}{\omega^2 \hbar m_0^2 \Omega} \sum\limits_{n,m,{\bf k}} \left[f_n({\bf k}) - f_m({\bf k}) \right] \cdot\\ \nonumber
& & p_{nm}^i({\bf k})p_{mn}^j({\bf k}) \delta(\omega - \omega_{mn}({\bf k})).
\end{eqnarray}
where $m_0$ is the bare electron mass, $\Omega$ the unit cell volume, $m,n$ the band indices, $f_{\ell}({\bf k})$ the Fermi-Dirac
distribution evaluated on the band with index ${\ell}$ at energy $E_{\ell}({\bf k})$, $p_{nm}^i({\bf k})$ are the matrix elements
of the momentum operator calculated over the states (both occupied and empty) with indices $m$ and $n$ and $\hbar \omega_{mn} = E_m({\bf k})-E_n({\bf k})$ is the energy of the optical transition.
The real part of the dielectric susceptibility can then be expressed as the Kramers-Kronig transformation of the imaginary part
\begin{equation} \label{KK.imaginary.dielectric.susceptibility}
\operatorname{Re}\chi(\omega) = \frac{2}{\pi} \int\limits_{0}^{\infty} z \frac{\operatorname{Im}\chi(\omega)}{z-\omega}dz.
\end{equation}
Eq.~(\ref{imaginary.dielectric.susceptibility}) implicitly contains both intra- ($n=m$) and inter-band ($n \ne m$)
transitions. Inter-band transitions are associated to the usual optical absorption processes in the UV-visible range,
while intra-band transitions are relevant in
the low frequency regime.
In the latter case, Eq.~(\ref{imaginary.dielectric.susceptibility}) naturally discriminates the different trends of quasi-static dielectric function ($\epsilon (\omega \rightarrow 0)$)
for insulating (finite behavior) and metallic (diverging behavior) systems.
In order to separate the intra- and inter-band contributions, we rewrite the expression $[f_n({\bf k}) - f_m({\bf k})]$ that appears in Eq.~(\ref{imaginary.dielectric.susceptibility}).
The presence of delta function selects the energy of the transitions $E_n({\bf k+q}) = E_n({\bf k}) + \hbar \omega$ and, consequently, it fixes the argument of the Fermi-Dirac distributions. Thus for $n = m$ we obtain:
\begin{eqnarray} \label{approximation.for.fermi.difference}
f_n({\bf k}) - f_n({\bf k + q}) &=& \\ \nonumber
f(E_n({\bf k})) -f(E_n({\bf k})+ \hbar \omega) &\approx& \\ \nonumber
- \hbar \omega \frac{\partial f(E)}{\partial E}|_{E=E_n(\bf k)}.
\end{eqnarray}
In the case of undoped semiconductors and insulators the intra-band transitions do not contribute to Eq.~(\ref{imaginary.dielectric.susceptibility}),
the imaginary part of the dielectric function vanishes, while the real part tends to static dielectric constant.
For metallic systems the presence of partially occupied bands at the Fermi energy
makes the derivative in Eq.~(\ref{approximation.for.fermi.difference}) converging to a finite quantity, while the term $\frac{1}{\omega^2}$ in Eq.~(\ref{imaginary.dielectric.susceptibility}) diverges. The divergency of the dielectric function gives rise to the well-known Drude-like {\em dc}-conductivity of metals, where electrons close to Fermi-level can undergo electron transitions
with negligible momentum transfer in the quasi-static regime ($\omega \rightarrow 0$).
The evaluation of the momentum matrix elements, $p_{nm}^i({\bf k})$, is a computational bottleneck since it requires an integration over all the pairs of occupied and empty electronic states across the whole BZ.
If the single particle wavefunctions are expanded in
large basis sets (e.g. planewaves), as in standard solid state implementations, this integration rapidly becomes a computational challenge. Once again, the momentum can be efficiently evaluated using the TB Hamiltonian projected on the PAOs, transforming the calculation of the frequency dependent dielectric function into a computationally trivial post-processing of the first principles calculation.
We have validated the method for calculating the dielectric function against well known experimental results for GaAs and Al. Results are summarized in Fig. \ref{optical} and the comparison confirms the validity of our approach.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.95\columnwidth]{Fig2.pdf}
\vspace{5mm}
\caption{\small (Color online) Real and Imaginary part of the dielectric function for GaAs (left panel) and Al (right panel). Results are validated against experimental data: Ref. \onlinecite{Ehrenreich:1963cq} for Al and Ref. \onlinecite{Aspnes:1983fy} for GaAs.}
\label{optical}
\end{center}
\end{figure}
\subsection{The ACBN0 functional} \label{acbn0}
The knowledge of first principles based localized orbital representation of the electronic structure combined with analytical expressions based on Gaussian basis set provides an efficient strategy for the fast computation of two-electron integrals for solid-state applications and the development of local exchange functionals (LEX). This is a critical advantage when dealing with the plethora of novel materials that are characterized by strong electron localization
and correlation and vigorously sought for their rich physical and chemical properties. For these materials the LDA+$U$ method, introduced by Liechtenstein and Anisimov, \cite{Anisimov1997,Liechtenstein1995} is the most practical choice to compensate for the simplified, nearly-homogeneous-electron-gas treatment of the electron density by LDA.
The success of LDA(GGA)+$U$ confirms that preserving the information of orbital localization from being averaged out is prevalent to the correct prediction of the electronic structure in compounds such as transition metal oxides.\cite{prbpriya:prep}
Our projection methodology allows the direct computation of two-electron integrals and, when combined with a density-matrix-based approach, the direct and self-consistent evaluation of the on-site Coulomb $U$ and exchange $J$
parameters needed in the treatment of correlated solid materials. This is at the core of the definition of the ACBN0 functional, recently introduced by some of us.\cite{Agapito:2015iz} ACBN0 satisfies the rather ambitious criteria outlined by Pickett {\it et al.}\cite{Pickett:1998tu} in one of the first seminal articles on LDA+$U$.
Due to the projection on AOs and the accurate TB representation, the evaluation of the \textit{U} and \textit{J} for atoms in different chemical environments or close to topological defects (surfaces, interfaces, impurities, etc.) or for closed-shell atoms (like Zn) becomes trivial, thus overcoming the limitations of traditional linear response techniques.\cite{reviewLDAU}
Results so far are striking: comparisons with available experimental and theoretical data show that the proposed computation of the on-site Coulomb and exchange parameters is a sound and high-throughput alternative to higher levels of theory such as hybrid functionals and the $GW$ approach that systematically yields results with outstanding accuracy.\cite{Agapito:2015iz,Gopal:2015bf}
ACBN0 only demands computational resources comparable to a regular (LDA)PBE calculation.
\section{Computational details.} \label{comp_details}
Ground state properties are obtained with density functional theory as implemented in the {\small Quantum ESPRESSO} package.\cite{Giannozzi:2009p1507} The starting electronic structure has been determined using ACBN0 and compared, when appropriate, with traditional PBE.
Ionic potentials are described by pseudopotentials with an extended basis set of pseudo-atomic orbitals for an improved TB mapping of the conduction bands.\cite{Agapito:2016en}
The workflow to perform all steps of the approach is part of the {\sf AFLOW$\mathlarger{\mathlarger{\mathlarger{{\pi}}}}$} software infrastructure.\cite{AFLOWpi:prep} {\sf AFLOW$\mathlarger{\mathlarger{\mathlarger{{\pi}}}}$}
has been designed for high-throughput first principles calculations and exploits the projections on PAO and the TB representation for the calculation of the band structure, density of states, transport coefficients, and frequency dependent dielectric constant. {\sf AFLOW$\mathlarger{\mathlarger{\mathlarger{{\pi}}}}$}
is integrated with AFLOW.\cite{Curtarolo2012218}
\section{Results} \label{results}
Our TB representation enables studies that involve extremely fine sampling of the full $E_n({\bf k})$ electronic structure and brings within reach nanostructured systems with hundreds of atoms without compromising first principles accuracy.
As examples to validate our methodology we will analyze with unprecedented details the conduction manifold of CoSb$_3$-based thermoelectric materials (Sec. \ref{cosb3}) and a core-shell nanowire with 588 atoms (Sec. \ref{coreshell}).
\subsection{CoSb$_3$}\label{cosb3}
Materials with the skutterudite structure are of great interest for their performance as thermoelectrics (TEs) since they well represent the paradigmatic case of a phonon-glass electron-crystal.\cite{Singh2001125,NolasBook,doi.org/10.1038/nmat2090,ANIE:ANIE200900598,doi:10.1179/095066003225010182}
They exhibit low thermal conductivity but also excellent electronic properties
and have been studied extensively both theoretically\cite{Singh:prb1994, PhysRevB.81.045204} and experimentally.\cite{doi:10.1146/annurev.matsci.29.1.89,Uher2001139}
The thermoeletric performances are characterized by the so-called figure of merit, $ZT=S^2\sigma T/\kappa$, where $S$ is the Seebeck coefficient, $\sigma$ is the electrical conductivity, $\kappa$ is the thermal conductivity, and
$T$ is the temperature at which the device operates.
The prototypical $n$-type compound within the skutterudites family is CoSb$_3$.\cite{doi:10.1146/annurev.matsci.29.1.89,Caillat:japphys1996, Tang:jmateriom2015}
Indeed, the figure of merit of R-filled CoSb$_3$ with R = Na, Yb, In, Ba, Ce
exceeds one in the temperature windows of practical interest.
Typical descriptors for enhanced TE electronic properties include reasonable large effective masses which favor large Seebeck coefficients and multi-valley character of the active bands to optimize the conductivity.\cite{Singh2001125,doi:/content/aip/journal/apl/74/24/10.1063/1.124220,doi:10.1021/acs.chemmater.5b04365} Doping and filling the semiconducting CoSb$_3$ with donor elements activates the bands at the bottom of the conduction manifold whose features critically contribute to optimize the value of electronic transport coefficients. The enhanced performance of RCoSb$_3$ can be rationalized with a detailed analysis of the full $E_n({\bf k})$ for all the bands in the proximity of the Fermi level. Various experimental and theoretical investigations have been performed in order to highlight the mechanisms responsible for its good thermoelectrical performance and the physics of the system is still under debate. In this paper we extend previous studies on the conduction band of doped CoSb$_3$ (Ref. \onlinecite{Tang:nmat2015}) by taking advantage the improved band structure provided by the ACBN0 functional discussed in Sec. \ref{acbn0} and the Boltzmann transport capabilities presented Sec. \ref{boltzmann}.
We have computed the starting electronic structure of CoSb$_3$ (prototype: A3B\_cI32\_204\_g\_c in Ref. \onlinecite{Mehl:2016wf}) using a {\bf k}-point grid of ($9 \times 9 \times 9$).\cite{Tang:nmat2015} The Boltzmann transport coefficients have been then calculated using a much finer grid with ($100 \times 100 \times 100$) {\bf k}-points.
In Fig. \ref{fig:cosb3}-(a) we compare the band structure of the system calculated with PBE and ACBN0. Qualitatively, the bands compare well with previous calculations\cite{Singh:prb1994}, however, in our ACBN0 calculations the energy gap between the occupied and the unoccupied manifold decreases from the PBE value of 0.23 eV to 0.16 eV. Both values are consistent with the experimental findings that range from 0.05 eV to 0.22 eV,\cite{Pei:advmat2012} but the reduction of the energy gap may be important to investigate effects associated with decreased performance due to bipolar transport.
The linearity of the dispersion\cite{Singh:prb1994} at the top of the valence band ($p$-derived band near $\Gamma$) slightly increases and the occupied Co-$d$ manifold moves to lower energy. This is consistent with the fact that the Sb Hubbard correction (U$^{Sb}$ = 0.648 eV) is small whereas the value of U for the Co cation (U$^{Co}$ = 4.375 eV) leads to a larger variation. This result is of particular importance when chemical substitution is used to reduce the energy separation between dispersive and flat bands in order to optimize the $p$-type transport coefficients.
The description of the conduction manifold also improves. In Fig. \ref{fig:cosb3}-(upper panel) the valleys at H and along the $\Gamma$-N direction in the BZ become quasi-degenerate ({$\Delta E_{ACBN0} = 10$ meV}) and flatten. It is remarkable that this quasi-degeneracy is not captured by standard PBE calculations ($\Delta E_{PBE} = 100$ meV).\cite{Tang:nmat2015} Phenomenologically these effects contribute to increase the number and the effective mass of charge carriers with positive consequences on the conductivity and Seebeck coefficient.
Within the rigid band approximation, we computed the Seebeck coefficient at values of the chemical potential corresponding to the experimental electron density\cite{Tang:nmat2015} and we compare our theoretical prediction with the experimental results as shown in Fig. \ref{fig:cosb3}-(lower panel).
In order to obtain the value of $\mu$ corresponding to a given experimental electron density $n_e$, we have used the free-electron 3D gas relationship $\mu = \frac{\hbar^2}{2 m^*}(3\pi^2n_e)^{2/3}$ where we have employed two effective masses calculated from the electronic structure at $\Gamma$ and at H assuming a parabolic dispersion.
Results are summarized in Fig. \ref{fig:cosb3}-(b) where we observe an excellent agreement between the calculated and experimental value of the Seebeck coefficient S.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.99\columnwidth]{Fig3.pdf}
\vspace{1mm}
\caption{\small (Color online) (Top panel) Electronic structure of CoSb$_3$. The black-dashed and the red-solid lines refers to PBE\cite{Tang:nmat2015} and ACBN0 calculations. (Botton panel) Seebeck coefficient of CoSb$_3$ measured experimentally (black filled circles)\cite{Tang:nmat2015} and computed theoretically (red filled squares) with a two-effective-masses model. The energy is measured from the bottom of the conduction band.}
\label{fig:cosb3}
\end{center}
\end{figure}
\subsection{Core-shell nanowire}\label{coreshell}
Quasi-one-dimensional, vertically aligned nanowires can be exploited to construct three-dimensional architectures with demonstrated advantages over conventional planar devices. Nanowires form building blocks for compact ultrafast electronics and optoelectronic devices (e.g. solar cells \cite{garnett10,shu13}, photodetectors\cite{soci10}, nanoscale lasers,\cite{duan03} and light emitting diodes\cite{bao06, sarwar15}).
One-dimensional (1D) component-modulated materials, such as coaxial core-shell heterojunctions offer the benefit of designing and fabricating nanodevices without further assembling and provide unique and tunable properties.\cite{lauhon02,boland15,Saha:2013fy}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.99\columnwidth]{Fig4.pdf}
\vspace{5mm}
\caption{\small (Color online) (a) Total and material-projected density of states plot; (b) Quantum conductance. Left inner panel: geometry of the CCS nanowire. Right inner panels: eigenchannels of the trasmission amplitude corresponding to the top of the valence and bottom of conduction bands. Zero energy reference is aligned to the Fermi level of the system.}
\label{wire}
\end{center}
\end{figure}
To highlight the capability of our methodology in evaluating the optical and transport properties of nanoscale systems, we have chosen as a prototypical example a coaxial core-shell (CCS) nanowire of ZnO and ZnS (see Fig. \ref{wire}(b), left inner panel) that has been recently synthesized\cite{Brayek:2014bt,Huang:2013kj,Nam:2011kl,Shuai:2011jb,Zhu:2008bi}.
ZnO has a wide band gap energy of 3.37 eV and it possesses unique optical and electronic properties that make it a promising candidate for UV lasers and detectors working in the 320-400 nm wavelength range. Additionally, it is transparent to visible light and can be made highly conductive by doping.\cite{wang04}
ZnS is also a wide band gap semiconductor, with bandgap energy of 3.66 eV widely exploited for optoelectronic devices and sensors. The electrical and the optical characterization of ZnS coated ZnO nanowires have been studied extensively both experimentally\cite{li06,wang10,meng10,Fang:2015kb,Fang:2016ep,Jeong:2014fr} and theoretically, although on much smaller diameter wires.\cite{Schrier:2007ds}
In this study, the core-shell nanowire is simulated using the ACBN0 functional in a large cell ($50.0 \times 50.0 \times 5.3$) A$^3$ with 588 atoms (i.e. ~5300 electrons). The core is made of ZnO and it has an internal radius $r_c$=1.1 nm, that is large enough to correctly reproduce a realistic ZnO wire\cite{PhysRevB.80.201304}. The shell is made of ZnS, the total radius of the heterostructure is $r_s$=1.9 nm, in agreement with experimental samples\cite{li06}. The wire is aligned along the polar c-axis of the wurtzite ZnO crystal. Due to the huge dimension of the system, we include two bilayers of ZnS-ZnO wurtzite material along the wire direction, \textit{i.e.} the minimum to obtain a periodic wire. The CCS structure has a hexagonal symmetry and exposes only non-polar (10$\bar{1}$0) faces.
The geometry is fully relaxed until forces on all atoms are lower than 0.03 eV/\AA. In the optimized structure the inner ZnO core almost maintains its ideal geometry, while the external ZnS shell undergoes to remarkable distortion due to the relaxation
of the mismatch at the interface. Nonetheless the outermost layer exhibits a buckled dimer arrangement, typical of the ZnS(10$\bar{1}$0) surface (see inset of Fig. \ref{wire}).
The resulting ACBN0 electronic structure is summarized in Fig. \ref{wire}a, where we plot the total (black line) and ZnO (red shaded area) and ZnS (blue thin line) projected density of states (DOS). Albeit ZnO and ZnS have similar band gap, the different ionization potential causes the formation of a staggered type-II band-alignment at the interface, with the top of ZnS valence band lying in the pristine gap of ZnO material. The ZnO core has a band gap E$_g\sim 3.1$ eV very similar to corresponding bulk, while the outer ZnS layer has a band gap E$_g\sim2.8$ eV. This gap reduction derives from the strong atomic deformation, which makes the final structure sensitively different from the ground state bulk one. The total band gap of the CCS nanowire is E$_g\sim 1.2 $ eV that lies in the near-IR range.
By using the approach described in Sec. \ref{ballistic}, we calculated the coherent electron transport along the nanowire. In the minimal TB-representation the solution of the Landauer problem reduces to matrix operations between (2940$\times$2940) on-site and hopping hamiltonians,\textit{i.e.} much smaller than the corresponding plane wave ones. This makes an otherwise unsolvable problem computationally feasible.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.95\columnwidth]{Fig5.pdf}
\vspace{5mm}
\caption{\small (Color online) Imaginary part of dielectric function of ZnS/ZnO CCS nanowire. Inset highlights band-alignment scheme and lowest energy vertical transitions.}
\label{wire_opt}
\end{center}
\end{figure}
For the nanowire we are considering here, the scattering contributions with the boundaries, due to extremely high surface-to-volume ratio, are the predominant effects that control the electron transport.\cite{rcsAvd} Thus, the coherent regime is a good first approximation for the description of transport, at least at low temperature.
The result for the CCS nanowire is shown in Figure \ref{wire}b.
The quantum transmittance $\mathcal{T}_{el}$ is proportional to the number of transmitting channels available for electron mobility, which are equal to the number of conducting bands at the same energy.
Within the scattering theory framework, the transmittance function $\mathcal{T}_{el}$ can be related to the transmittance amplitude $\mathrm{t}$ through the relation $\mathcal{T}_{el}=Tr[\mathrm{t}^{\dag}\mathrm{t}]$. The eigenvectors
of matrix $\mathrm{t}$ are called {\em eigenchannels} and
are defined as the linear combinations of the incoming modes in a lead that do not mix upon reflection on the scattering region. Thus, the spatial representation of the eigenchannels visually displays
the path travelled by charge carriers in their flow through the nanowire. The eigenchannels corresponding to the top of the valence and bottom of conduction bands are shown as inset in panel b: hole carriers flow in the external ZnS crown, while electrons move in the internal ZnO core, confirming the intrinsic charge separation observed experimentally.\cite{Jeong:2014fr}
Simulated optical properties are summarized in Fig \ref{wire_opt}. The peaks in the imaginary part of the dielectric function $\epsilon$ correspond to single particle valence-to-conduction transitions. As expected, it is easy to recognize the valence-to-conduction absorption edge in the external ZnS shell (2.8 eV, blue arrow) and in the inner ZnO (3.1 eV, red arrow),
which correspond to the E$_g$ values of each material, as discussed above. However, other lower-energy transitions are present in the range
1.2-2.8 eV, which correspond to inter-material ZnO-to-ZnS transition as depicted in the inset of the figure. Although the frontiers orbitals are mostly localized in the core (electrons) and in the shell (hole), the overlap and the symmetry of the wave functions
give low but not negligible oscillator strength to the first four transition. Except for excitonic effects (not included at this level of theory), this agrees well with dramatic red-shift of the absorption edge (i.e. from UV - to near-IR) observed experimentally
\cite{Fang:2015kb,Fang:2016ep}, also in agreement with previous theoretical calculations on smaller CCS wires\cite{Schrier:2007ds}. The facile photocharge injection and the intrinsic carrier separation make this system a very promising photo-conductor candidate for optoelectronic and photovoltaic devices.
\section{Conclusions} \label{conclusions}
We have extended the PAO projection technique to allow for the calculation of electronic transport and optical properties of materials with extreme accuracy and negligible computational cost. The exact tight binding representation of the first principles electronic structure allows on one hand to produce extremely dense band interpolations, essential requirement for the evaluation of Boltzmann transport or optical properties, and on the other, the local Green's function representation that is at the foundation of quantum conductance calculations.
We have demonstrated the potentiality of the method by studying the multi-valley band structure of CoSb$_3$ and a large ZnO-ZnS core-shell nanowire.
\acknowledgments
We want to thank the Texas Advanced Computing Center (TACC) at the University of Texas Austin and CINECA for providing computing facilities; the funding provided by DOD-ONR (N00014-13-1-0635, N00014-15-1-2266 and N00014-14-1-0526); the support from the University of Modena and Reggio Emilia through the grant ``Nano- and emerging materials and systems for sustainable technologies''; and the European Union Seventh Framework Program for the grant agreement 265073 (ITN-Nanowiring). The authors also acknowledge the Duke University Center for Materials Genomics and the CRAY Corporation for computational assistance.
|
\section{Basic 2-category theory in an \texorpdfstring{$\infty$}{infinity}-cosmos}\label{sec:2-cat}
\renewcommand\thesection{\arabic{section}}
The goal of these lectures is to give a very precise account of the foundations of $\infty$-category theory, paralleling the development of ordinary category theory.
To explain what we mean by the term ``$\infty$-categories,'' it is useful to distinguish between schematic definitions of infinite-dimensional categories and models of infinite-dimensional categories. For instance, $(\infty,1)$-\emph{category} is a schematic notion, describing a category that is weakly enriched in $\infty$-\emph{groupoids}, itself a schematic term for something like topological spaces. The main source of ambiguity in this definition is meaning of ``weakly enriched,'' which is not precisely defined.
Models of $(\infty,1)$-categories are precise mathematical objects meant to embody this schema. These include \emph{quasi-categories} (originally called \emph{weak Kan complexes}), \emph{complete Segal spaces}, \emph{Segal categories}, and \emph{naturally marked simplicial sets} (\emph{1-trivial saturated weak complicial sets}), among others.
To develop the theory of $(\infty,1)$-categories in a non-schematic way, it makes sense to choose a model. Andr\'e Joyal pioneered and Jacob Lurie extended a wildly successful project to extend basic category theory from ordinary categories to quasi-categories. A natural question is then: does this work extend to other models of $(\infty,1)$-categories? And to what extent are basic categorical notions invariant under change of models?
The general consensus is that the choice of model should not matter so much, but one obstacle to proving results of this kind is that, to a large extent, precise versions of the categorical definitions that have been established for quasi-categories had not been given for the other models.
Here we will use the term $\infty$-\emph{category} to refer to any of the models of $(\infty,1)$-categories listed above: quasi-categories, complete Segal spaces, Segal categories, or naturally marked simplicial sets.\footnote{Later, we will explain that the results we prove about $\infty$-\emph{categories} also apply to other varieties of infinite-dimensional categories, but let us not get into this now.} Note that, for us ``$\infty$-category'' will not function as a schematic term: several well-behaved models of $(\infty,1)$-categories will be included in the scope of its meaning, but others are excluded. With this interpretation in mind, let us describe a few features of our project to extend the basic theory of categories to $\infty$-categories.
\begin{itemize}
\item It is blind to which model of $\infty$-categories is being considered.
\end{itemize}
That is, our definitions of basic categorical notions will be stated and our theorems will be proven without reference to particular features of any model of $\infty$-categories and will apply, simultaneously, to all of them.
\begin{itemize}
\item It is compatible with the Joyal/Lurie theory of quasi-categories.
\end{itemize}
In the special case of quasi-categories, our presentation of the basic categorical notions will necessarily differ from existing one, as our formalism is unaware of the fact that quasi-categories are simplicial sets. Nonetheless the categorical concepts so-defined are precisely equivalent to the Joyal/Lurie theory, and so our approach can be mixed with the existing one. This is good news because, at present, the Joyal/Lurie theory of quasi-categories is considerably more expansive.\footnote{But we are not finished yet.}
\begin{itemize}
\item It is invariant under change of models between $(\infty,1)$-categories
\end{itemize}
Independent work of Bertrand T\"{o}en and of Clark Barwick and Chris Schommer-Pries proves that all models of $(\infty,1)$-categories ``have the same homotopy theory,'' in the sense of being connected by a zig-zag of Quillen equivalences of model categories \cite{Toen:2005vu} or having equivalent quasi-categories \cite{BSP:2011ot}. In lectures 3 and 4, we will see that the models of $(\infty,1)$-categories that fall within the scope of this theory all ``have the same basic category theory,'' in the sense that the answers to basic categorical questions are invariant under change of model. Finally:
\begin{itemize}
\item It is as simple as possible.
\end{itemize}
Our original aim in this project was to provide a streamlined, from the ground up, development of the foundations of quasi-category theory that more closely parallels classical category theory and is relatively easy to work with, at least for those with some affinity for abstract nonsense. With a bit of effort, these ideas could be extended to cover a wider variety of models of $(\infty,1)$-categories, but to do so would come at the cost of complicating the proofs.
The plan for the first lecture is to introduce the axiomatic framework that makes it possible to develop the basic theory of $\infty$-categories without knowing precisely which variety of infinite-dimensional categories we are working with. In the second installment, we will continue with a discussion of limits and colimits of diagrams valued inside an $\infty$-category, which we relate to the adjunctions between $\infty$-categories that we will define now. In the third lecture, we will introduce \emph{comma $\infty$-categories}, which will be the essential ingredient in our proofs that basic category theory is invariant under change of models. In the first two lectures, we will see how basic categorical notions can be developed ``model agnostically,'' i.e., simultaneously for all models. In lecture three, we will be able to prove that this theory is both preserved and reflected by certain change-of-model functors. Finally, in the final talk we will introduce \emph{modules} between $\infty$-categories, which we use to develop the theory of pointwise Kan extensions.
The work to date on this project can be found in a series of five papers:
\begin{enumerate}[label=\Roman*.]
\item The 2-category theory of quasi-categories \cite{RiehlVerity:2012tt}, \href{http://arxiv.org/abs/1306.5144}{ar\-Xiv:1306.5144}
\item Homotopy coherent adjunctions and the formal theory of monads \cite{RiehlVerity:2012hc},\\ \href{http://arxiv.org/abs/1310.8279}{arXiv:1310.8279}
\item Completeness results for quasi-categories of algebras, homotopy limits, and related general constructions \cite{RiehlVerity:2013cp}, \href{http://arxiv.org/abs/1401.6247}{arXiv:1401.6247}
\item Fibrations and Yoneda's lemma in an $\infty$-cosmos \cite{RiehlVerity:2015fy}, \href{http://arxiv.org/abs/1506.05500}{arXiv:1506.05500}
\item Kan extensions and the calculus of modules for $\infty$-categories \cite{RiehlVerity:2015ke}, \\ \href{http://arxiv.org/abs/1507.01460}{arXiv:1507.01460}
\end{enumerate}
references within which will have the form, e.g., \extRef{cohadj}{II}{thm:monadiccomparisonadj}.
\subsection{Formal category theory in a 2-category}
Recall the term $\infty$-\emph{categories} refers to the reader's choice of one of several listed models of infinite-dimensional categories. Each fixed model of $\infty$-categories has an accompanying notion of $\infty$-\emph{functors} between them. The reason we are able to develop the basic theory of $\infty$-categories without knowing exactly what the $\infty$-categories are is that
the $\infty$-categories and $\infty$-functors of each type are packaged inside an axiomatic framework that describes the surrounding category with these objects and morphisms. We will say more about this in a moment, but first we want to illustrate how closely our development of the theory of $\infty$-categories parallels the development of the theory of ordinary categories. One thing that is provided by the axiomatization that we will describe in a moment is a reasonable notion of $\infty$-\emph{natural transformations} between $\infty$-functors.
Categories, functors, and natural transformations naturally assemble into a \emph{2-category} $\category{Cat}_2$, a structure which records the various ways in which functors and natural transformations can be composed. Similarly, we will see that the $\infty$-categories of each fixed model, the $\infty$-functors between them, and the $\infty$-natural transformations between $\infty$-functors also assemble into a 2-category, which we call the \emph{homotopy 2-category}. Each variety of $\infty$-categories will have their own homotopy 2-category --- one example being $\category{Cat}_2$. A \emph{homotopy 2-category} is a strict 2-category whose:
\begin{itemize}
\item objects $A$, $B$, and $C$ are \emph{$\infty$-categories};
\item 1-cells $f \colon A \to B$, $g \colon B \to C$, are ($\infty$-)\emph{functors};
\item 2-cells $\xymatrix{ A \ar@/^2ex/[r]^f \ar@/_2ex/[r]_g \ar@{}[r]|{\Downarrow\alpha}& B,}$ written in line as $\alpha \colon f \Rightarrow g \colon A \to B$, are ($\infty$-)\emph{natural transformations}.
\end{itemize}
\begin{rec}[pasting diagrams in 2-categories]
The objects and 1-cells in a 2-category define its \emph{underlying 1-category}. In particular, the 1-cells have the familiar associative and unital composition law. The 2-cells can be composed in two ways:
\begin{itemize}
\item The \emph{vertical composite} of $\vcenter{\xymatrix@R=30pt{ A \ar@/^2ex/[r]^f \ar@/_2ex/[r]_g \ar@{}[r]|{\Downarrow\alpha}& B \\ A \ar@/^2ex/[r]^g \ar@/_2ex/[r]_h \ar@{}[r]|{\Downarrow\beta}& B}}$ defines a 2-cell $\xymatrix{ A \ar@/^2ex/[r]^f \ar@/_2ex/[r]_h \ar@{}[r]|{\Downarrow\beta\cdot\alpha}& B}$.
\item The \emph{horizontal composite} of $ \xymatrix{ A \ar@/^2ex/[r]^f \ar@/_2ex/[r]_g \ar@{}[r]|{\Downarrow\alpha}& B \ar@/^2ex/[r]^h \ar@/_2ex/[r]_k \ar@{}[r]|{\Downarrow\gamma} & C}$ defines a 2-cell $\xymatrix{ A \ar@/^2ex/[r]^{hf} \ar@/_2ex/[r]_{kg} \ar@{}[r]|{\Downarrow\gamma\alpha}& C}$.
\end{itemize}
A degenerate special case of horizontal composition, in which all but one of the 2-cells is an identity $\id_f$ on its boundary 1-cell $f$, is called \emph{whiskering}:
\[ \vcenter{\xymatrix{ A \ar@/^2ex/[r]^f \ar@/_2ex/[r]_f \ar@{}[r]|{\Downarrow\id_f}& B \ar@/^2ex/[r]^h \ar@/_2ex/[r]_k \ar@{}[r]|{\Downarrow\gamma} & C \ar@/^2ex/[r]^\ell \ar@/_2ex/[r]_\ell \ar@{}[r]|{\Downarrow\id_{\ell}} & D} } = \vcenter{\xymatrix{ A \ar[r]^f & B \ar@/^2ex/[r]^h \ar@/_2ex/[r]_k \ar@{}[r]|{\Downarrow\gamma} & C \ar[r]^\ell & D} } = \vcenter{\xymatrix{ A \ar@/^2ex/[r]^{\ell hf} \ar@/_2ex/[r]_{\ell kf} \ar@{}[r]|{\Downarrow\ell \gamma f} & D} }\]
Vertical composition is associative and unital with respect to identity 2-cells: there is a category of 1-cells and 2-cells between each fixed pair of objects. Horizontal composition is associative and unital with respect to those identity 2-cells $\id_{\id_A}$ on identity 1-cells. Moreover, the vertical and horizontal composition operations commute by the law of \emph{middle four interchange} which says that any \emph{pasting diagram}, examples of which are displayed below, has a unique composite 2-cell.
\[ \vcenter{\xymatrix{ A \ar@/^3.5ex/[r]^f_{\Downarrow\alpha} \ar[r]|g \ar@/_3.5ex/[r]_h^{\Downarrow\beta} & B \ar@/^3.5ex/[r]^j_{\Downarrow\gamma} \ar[r]|k \ar@/_3.5ex/[r]_\ell^{\Downarrow\delta} & C}}
\qquad\qquad \vcenter{\xymatrix@=15pt{ & B \ar[rr]^g \ar@{}[d]|(.6){\Downarrow\alpha} \ar[dr]|p & \ar@{}[d]|(.4){\Downarrow\beta} & C \ar@{}[dd]|{\Downarrow\gamma} \ar[dr]^h \\ A \ar[ur]^f \ar[dr]_\ell \ar[rr]|m & \ar@{}[dr]|{\Downarrow\delta} & G \ar[ur]|r \ar[dr]|s & & D \\ & F \ar[rr]_k & & E \ar[ur]_j}}
\]
\end{rec}
Now let us explore how such a framework can be used to develop the basic category theory of the $\infty$-categories that define the objects of the homotopy 2-category. For instance, the homotopy 2-category provides a convenient framework in which to define the notion of adjunction between $\infty$-categories.
\begin{defn}\label{defn:adjunction} An \emph{adjunction} between $\infty$-categories consists of:
\begin{itemize}
\item a pair of $\infty$-categories $A$ and $B$;
\item a pair of functors $f \colon B \to A$ and $u \colon A \to B$; and
\item a pair of natural transformations $\eta \colon \id_B \Rightarrow uf$ and $\epsilon \colon fu \Rightarrow \id_A$
\end{itemize}
so that the triangle identities hold:
\[\xymatrix@=1.5em{ & B \ar[dr]|f \ar@{}[d]|(.6){\Downarrow\epsilon} \ar@{=}[rr] & \ar@{}[d]|(.4){\Downarrow\eta} & B \ar@{}[d]^*+{=} & B && B \ar@{=}[rr] \ar[dr]_f & \ar@{}[d]|(.4){\Downarrow \eta} & B \ar[dr]^f \ar@{}[d]|(.6){\Downarrow\epsilon} & {\mkern40mu}\ar@{}[d]^*+{=} & B \ar@/^2ex/[d]^f \ar@/_2ex/[d]_f \ar@{}[d]|(.4){\id_f}|(.6){\Leftarrow} & \\A \ar[ur]^u \ar@{=}[rr] & & A \ar[ur]_u & {\mkern40mu} & A \ar@/^2ex/[u]^u \ar@/_2ex/[u]_u \ar@{}[u]|(.4){\Rightarrow}|(.6){\id_u} && &A \ar[ur]|u \ar@{=}[rr] & & A & A }\] The left-hand equality of pasting diagrams asserts that $u\epsilon \cdot \eta u = \id_u$, while the right-hand equality asserts that $\epsilon f \cdot f\eta = \id_f$.
\end{defn}
We write $f \dashv u$ to assert that the functor $f \colon B \to A$ is \emph{left adjoint} to the functor $u \colon A \to B$, its \emph{right adjoint}. From the standpoint of this definition, we can easily prove some basic properties of adjunctions. Note that in the special case of the homotopy 2-category $\category{Cat}_2$, these proofs are exactly the familiar ones.
\begin{prop}\label{prop:adjunctions-compose} Adjunctions compose: given adjoint functors
\[ \vcenter{\xymatrix{ C \ar@<1ex>[r]^{f'} \ar@{}[r]|\perp & B \ar@<1ex>[r]^f \ar@<1ex>[l]^{u'} \ar@{}[r]|\perp & A \ar@<1ex>[l]^u}} \qquad \rightsquigarrow\qquad \vcenter{\xymatrix{ C \ar@<1ex>[r]^{ff'} \ar@{}[r]|\perp & A \ar@<1ex>[l]^{u'u}}}\] the composite functors are adjoint.
\end{prop}
\begin{proof}
Writing $\eta \colon \id_B \Rightarrow uf$, $\epsilon \colon fu \Rightarrow \id_A$, $\eta' \colon \id_C \Rightarrow u'f'$, and $\epsilon' \colon fu \Rightarrow \id_B$ for the respective units and counits, the pasting diagrams
\[ \xymatrix@=1.5em{ C \ar[dr]_{f'} \ar@{=}[rrrr] & & \ar@{}[d]|{\Downarrow \eta'} & & C & & C \ar@{}[d]|(.6){\Downarrow\epsilon'} \ar[dr]^{f'} \\ & B \ar@{=}[rr] \ar[dr]_f & \ar@{}[d]|(.4){\Downarrow\eta} & B \ar[ur]_{u'} & & B \ar[ur]^{u'} \ar@{=}[rr] & \ar@{}[d]|{\Downarrow\epsilon} & B \ar[dr]^f \\ & & A \ar[ur]_u & & A \ar[ur]^u \ar@{=}[rrrr] & & & & A}\] define the unit and counit of $ff' \dashv u'u$ so that the triangle identities
\[ \xymatrix@=.70em{ C \ar[dr]_{f'} \ar@{=}[rrrr] & & \ar@{}[d]|{\Downarrow \eta'} & & C \ar@{}[d]|(.6){\Downarrow\epsilon'} \ar[dr]^{f'}& & && C \ar@/^3ex/[dd]^(.3){ff'} \ar@/_3ex/[dd]_(.3){ff'} \ar@{}[dd]|(.4){\id_{ff'}}|(.6){\Leftarrow} & && & C \ar@{}[d]|(.6){\Downarrow\epsilon'} \ar[dr]|{f'} \ar@{=}[rrrr] & & \ar@{}[d]|{\Downarrow \eta'} & & C & & C \\ & B \ar@{=}[rr] \ar[dr]_f & \ar@{}[d]|(.4){\Downarrow\eta} & B \ar[ur]|{u'} \ar@{=}[rr] & \ar@{}[d]|{\Downarrow\epsilon} & B \ar[dr]^f & & \ar@{}[l]|{\displaystyle =} & && & B \ar[ur]^{u'} \ar@{=}[rr] & \ar@{}[d]|{\Downarrow\epsilon} & B \ar[dr]|f \ar@{=}[rr] & \ar@{}[d]|(.4){\Downarrow\eta} & B \ar[ur]_{u'} & & \ar@{}[l]|{\displaystyle =} \\ & & A \ar[ur]|u \ar@{=}[rrrr] & & & & A & & A & & A \ar[ur]^u \ar@{=}[rrrr] & & & & A \ar[ur]_u & & & & A \ar@/^3ex/[uu]^(.3){u'u} \ar@/_3ex/[uu]_(.3){u'u} \ar@{}[uu]|(.6){\id_{u'u}}|(.4){\Rightarrow} & & } \]
hold.
\end{proof}
It is also straightforward to show that the existence of adjoint functors is equivalence-invariant, in the sense of the standard 2-categorical notion of equivalence.
\begin{defn}\label{defn:equivalence} An \emph{equivalence} between $\infty$-categories consists of:
\begin{itemize}
\item a pair of $\infty$-categories $A$ and $B$;
\item a pair of functors $f \colon B \to A$ and $g \colon A \to B$; and
\item a pair of natural isomorphisms $\eta \colon \id_B \cong gf$ and $\epsilon \colon fg \cong \id_A$.
\end{itemize}
An ($\infty$-)\emph{natural isomorphism} is a 2-cell in the homotopy 2-category that admits a vertical inverse 2-cell.
\end{defn}
We write $A \simeq B$ and say that $A$ and $B$ are \emph{equivalent} if there exists an equivalence between $A$ and $B$. The direction for the natural isomorphisms comprising an equivalence is immaterial. Our notation is chosen to suggest the connection with adjunctions conveyed by the following exercise.
\begin{exs}\label{exs:adjoint-equivalence} Show that an equivalence, defined in any 2-category, can always be promoted to an \emph{adjoint equivalence} by modifying one of the 2-cell isomorphisms. That is, show that the 2-cell isomorphisms in an equivalence can be chosen so as to satisfy the triangle identities.
\end{exs}
Combining Exercise \ref{exs:adjoint-equivalence} with the symmetry in the definition of an equivalence, we have:
\begin{cor}\label{cor:equivs-are-adjoints} Any functor $f \colon A \to B$ that defines an equivalence of $\infty$-categories admits both a left and a right adjoint.
\end{cor}
\begin{prop}\label{prop:equiv-invar-adjunction} If $\xymatrix{ B \ar@<1ex>[r]^f \ar@{}[r]|\perp & A \ar@<1ex>[l]^u}$ is an adjunction and $A \simeq A'$ and $B \simeq B'$ are any equivalences, then the equivalent functors $\xymatrix{ B' \ar@<1ex>[r]^{f'} \ar@{}[r]|\perp & A' \ar@<1ex>[l]^{u'}}$ are again adjoints.
\end{prop}
\begin{proof}
Promoting the equivalences to adjoint equivalences we have the composite adjunction
\[ \xymatrix{ B' \ar@<1ex>[r]^\sim \ar@{}[r]|\perp & B \ar@<1ex>[r]^f \ar@<1ex>[l]^\sim \ar@{}[r]|\perp & A \ar@<1ex>[r]^\sim \ar@{}[r]|\perp \ar@<1ex>[l]^u & A' \ar@<1ex>[l]^\sim,}\] defining $f' \dashv u'$.
\end{proof}
An analogous result to Proposition \ref{prop:equiv-invar-adjunction} holds for other notions of equivalent functors:
\begin{exs}\label{exs:iso-invar-adjunction} In any 2-category show that:
\begin{enumerate}[label=(\roman*)]
\item If $f' \cong f$ and $f \dashv u$ then $f' \dashv u$.
\item If $f \dashv u$ and $f' \dashv u$ then $f \cong f'$.
\item If $f' \cong f$ and $f$ is an equivalence then so is $f'$.
\end{enumerate}
\end{exs}
\subsection{\texorpdfstring{$\infty$}{infinity}-cosmoi}
The results in the previous section illustrate some of the basic formal category theory that can be developed internally to any 2-category. As our terminology suggests, Definitions \ref{defn:adjunction} and \ref{defn:equivalence} indeed define appropriate notions of adjunction and equivalence for many varieties of $(\infty,1)$-categories in such a way that \ref{prop:adjunctions-compose}, \ref{exs:adjoint-equivalence}, \ref{cor:equivs-are-adjoints}, \ref{prop:equiv-invar-adjunction}, and \ref{exs:iso-invar-adjunction} all hold. To understand this, we will now explain how to define homotopy 2-categories whose objects are, respectively, quasi-categories, complete Segal spaces, Segal categories, or naturally marked simplicial sets, among other examples. Each homotopy 2-category arises as a quotient of an $\infty$-\emph{cosmos}, which records the properties of the model category --- typically used to encode the ``homotopy theory'' of each type of $(\infty,1)$-category --- that will be required to develop their basic category theory.
It is not necessary to know anything about model categories in order to understand the definition of an $\infty$-cosmos, but for those who are familiar with them, we briefly say a few words to motivate this axiomatization. If our aim was to establish the basic homotopy theory of some type of objects, experience suggests that a good setting in which to work is a simplicial model category, that is in a simplicially enriched category equipped with a model structure that is enriched over the Kan/Quillen model structure on simplicial sets. In a simplicial model category, there is a formula for the homotopy limit or colimit of a diagram of any shape. In the subcategory of fibrant-cofibrant objects, the model-theoretic notion of ``weak equivalence'' coincides with a more symmetrically defined notion of ``homotopy equivalence.'' Also, mapping-spaces between objects that are fibrant and cofibrant have the ``correct homotopy type'' and are invariant under these notions of equivalence.
By analogy, we posit that a good setting in which to establish the basic category theory of some type of objects (the ``$\infty$-categories'') is in a simplicially enriched category equipped with a model structure that is enriched over the Quasi/Joyal model structure on simplicial sets. Each of the models of $(\infty,1)$-categories listed above arise as precisely the fibrant objects (which are also cofibrant) in a model structure of this kind. In fact, to do our work, we need only a portion of the axioms defining a Quasi/Joyal-enriched model structure, which define what we call an $\infty$-\emph{cosmos}.
\begin{defn}[$\infty$-cosmos]\label{qcat.ctxt.cof.def}
An $\infty$-\emph{cosmos} is a simplicially enriched category $\mathcal{K}$ whose
\begin{itemize}
\item objects we refer to as the \emph{$\infty$-categories} in the $\infty$-cosmos, whose
\item hom simplicial sets $\fun(A,B)$ are all quasi-categories\footnote{A quasi-category is a particular type of simplicial set. See \ref{rec:equiv-qcats}.},
\end{itemize} and that is equipped with a specified subcategory of \emph{isofibrations}, denoted by ``$\twoheadrightarrow$'',
satisfying the following axioms:
\begin{enumerate}[label=(\alph*)]
\item\label{qcat.ctxt.cof:a} (completeness) As a simplicially enriched category, $\mathcal{K}$ possesses a terminal object $1$, cotensors $A^U$ of objects $A$ by all\footnote{It suffices to require only cotensors with finitely presented simplicial sets (those with only finitely many non-degenerate simplices.} simplicial sets $U$, and pullbacks of isofibrations along any functor.\footnote{For the theory of homotopy coherent adjunctions and monads developed in \cite{RiehlVerity:2012hc}, retracts and limits of towers of isofibrations are also required, with the accompanying stability properties of \ref{qcat.ctxt.cof:b}. These limits are present in all of the $\infty$-cosmoi we are aware of, but will not be required for any results discussed here.}
\item\label{qcat.ctxt.cof:b} (isofibrations) The class of isofibrations contains the isomorphisms and all of the functors $!\colon A \twoheadrightarrow 1$ with codomain $1$; is stable under pullback along all functors; and if $p\colon E\twoheadrightarrow B$ is an isofibration in $\mathcal{K}$ and $i\colon U\hookrightarrow V$ is an inclusion of simplicial sets then the Leibniz cotensor $i\leib\pitchfork p\colon E^V\twoheadrightarrow E^U\times_{B^U} B^V$ is an isofibration. Moreover, for any object $X$ and isofibration $p \colon E \twoheadrightarrow B$, $\fun(X,p) \colon \fun(X,E) \twoheadrightarrow \fun(X,B)$ is an isofibration of quasi-categories.
\end{enumerate}
The underlying category of an $\infty$-cosmos $\mathcal{K}$ has a canonical subcategory of (representably-defined) \emph{equivalences}, denoted by ``$\xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu}$'', satisfying the 2-of-6 property. A functor $f \colon A \to B$ is an \emph{equivalence} just when the induced functor $\fun(X,f) \colon \fun(X,A) \to \fun(X,B)$ is an equivalence of quasi-categories for all objects $X \in \mathcal{K}$. The \emph{trivial fibrations}, denoted by ``$\stackrel{\smash{\mkern-2mu\mathlower{1.5}{\sim}}}\longtwoheadrightarrow$'', are those functors that are both equivalences and isofibrations.
\begin{enumerate}[label=(\alph*), resume]
\item\label{qcat.ctxt.cof:c} (cofibrancy) All objects are \emph{cofibrant}, in the sense that they enjoy the left lifting property with respect to all \emph{trivial fibrations} in $\mathcal{K}$.
\[ \xymatrix{ & E \ar@{->>}[d]^{\rotatebox{90}{$\displaystyle\sim$}} \\ A \ar[r] \ar@{-->}[ur]^{\exists} & B}\]
\end{enumerate}
It follows from \ref{qcat.ctxt.cof.def}\ref{qcat.ctxt.cof:a}-\ref{qcat.ctxt.cof:c} that:
\begin{enumerate}[label=(\alph*), resume]
\item\label{qcat.ctxt.cof:d} (trivial fibrations) The trivial fibrations define a subcategory containing the isomorphisms; are stable under pullback along all functors; and the Leibniz cotensor $i\leib\pitchfork p\colon E^V\stackrel{\smash{\mkern-2mu\mathlower{1.5}{\sim}}}\longtwoheadrightarrow E^U\times_{B^U} B^V$ of an isofibration $p\colon E\twoheadrightarrow B$ in $\mathcal{K}$ and a monomorphism $i\colon U\hookrightarrow V$ between presented simplicial sets is a trivial fibration when $p$ is a trivial fibration in $\mathcal{K}$ or $i$ is trivial cofibration in the Joyal model structure on $\category{sSet}$ (see \extRef{equipment}{V}{lem:triv.fib.stab}).
\item\label{qcat.ctxt.cof:e} (factorization) Any functor $f \colon A \to B$ may be factored as $f = p j$
\[ \xymatrix{ & Nf \ar@{->>}[dr]^p \ar@{->>}@/_3ex/[dl]_{q}^*-{\rotatebox{45}{$\labelstyle\sim$}} \\ A \ar[rr]_f \ar[ur]^*-{\rotatebox{45}{$\labelstyle\sim$}}_j & & B}\] where $p \colon Nf \twoheadrightarrow B$ is an isofibration and $j \colon A \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} N_f$ is right inverse to a trivial fibration $q \colon N_f \stackrel{\smash{\mkern-2mu\mathlower{1.5}{\sim}}}\longtwoheadrightarrow A$ (see \extRef{yoneda}{IV}{lem:Brown.fact}).
\end{enumerate}
An $\infty$-cosmos is \emph{cartesian closed} if it satisfies the extra axiom:
\begin{enumerate}[label=(\alph*), resume]
\item\label{qcat.ctxt.cof:g} (cartesian closure) The product bifunctor $-\times - \colon \mathcal{K} \times \mathcal{K} \to \mathcal{K}$ extends to a simplicially enriched two-variable adjunction
\[ \fun(A \times B,C) \cong \fun(A, C^B) \cong \fun(B,C^A).\]
\end{enumerate}
\end{defn}
\begin{rec}[quasi-categories]\label{rec:equiv-qcats}
A \emph{quasi-category} is a simplicial set satisfying the weak Kan condition, i.e., in which every inner horn $\Lambda^{n,k} \hookrightarrow \del@fn\parsedel@^n$, $0 < k< n$, has a filler. The quasi-categories define the fibrant objects in a model structure on simplicial sets due to Joyal in which the cofibrations are the monomorphisms and the fibrations between fibrant objects are called \emph{isofibrations}. An \emph{isofibration} is a map that has the right lifting property against the inner horn inclusions and also against the inclusion $\del@fn\parsedel@^0 \to {\mathbb{I}}$ of either endpoint into the ``interval'' ${\mathbb{I}}$, defined to be the nerve of the free category $\bullet \cong \bullet$ containing an isomorphism.
The weak equivalences between quasi-categories are precisely the \emph{equivalences between quasi-categories}, which can be understood as a type of ``simplicial homotopy equivalence'' with respect to the interval ${\mathbb{I}}$. That is, a map $f \colon A \to B$ of quasi-categories is an equivalence just when there exists a map $g \colon B \to A$ together with maps $A \to A^{\mathbb{I}}$ and $B \to B^{\mathbb{I}}$ that restrict along the vertices of ${\mathbb{I}}$ to the maps $\id_A$, $gf$, $fg$, and $\id_B$ respectively:
\[ f \colon A \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} B\quad \mathrm{iff}\quad \exists g \colon B \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} A\quad \mathrm{and} \quad \vcenter{\xymatrix{ & A \\ A \ar[r] \ar[ur]^{\id_A} \ar[dr]_{gf} & A^{\mathbb{I}} \ar@{->>}[u]_{p_0} \ar@{->>}[d]^{p_1} \\ & A}} \quad\mathrm{and}\quad \vcenter{\xymatrix{ & B \\ B \ar[r]\ar[ur]^{fg} \ar[dr]_{\id_B} & B^{{\mathbb{I}}} \ar@{->>}[u]_{p_0} \ar@{->>}[d]^{p_1} \\ & B}}\]
\end{rec}
The subcategory of fibrant objects in a model category that is enriched over the Joyal model structure on simplicial sets defines an $\infty$-cosmos --- assuming all fibrant objects are cofibrant.\footnote{This hypothesis is not essential; see \extRef{yoneda}{IV}{qcat.ctxt.def}.} In such examples, the convention will be to define the isofibrations to be the fibrations between fibrant objects. It follows that the equivalences are precisely the weak equivalences between fibrant objects in the model category. This is the source of each of the following examples of $\infty$-cosmoi.
\begin{ex} There exist $\infty$-cosmoi:
\begin{itemize}
\item $\category{Cat}$, whose objects are ordinary categories, with isofibrations and equivalences the usual categorical isofibrations and equivalences (\extRef{yoneda}{IV}{ex:cat-cosmos});
\item $\category{qCat}$, whose objects are quasi-categories (\extRef{yoneda}{IV}{ex:qcat-qcat-ctxt}), with isofibration and equivalences as in \ref{rec:equiv-qcats};
\item $\category{CSS}$, whose objects are complete Segal spaces (\extRef{yoneda}{IV}{ex:CSS-cosmos});
\item $\category{Segal}$, whose objects are Segal categories (\extRef{yoneda}{IV}{ex:segal-cosmos});
\item $\category{sSet}_+$, whose objects are naturally marked simplicial sets (\extRef{yoneda}{IV}{ex:marked-cosmos});
\end{itemize}
all of which are cartesian closed. Thus each of these varieties of $(\infty,1)$-categories are examples of $\infty$-\emph{categories}, in our sense; the associated $\infty$-\emph{functors} are just the usual functors, maps of simplicial sets, maps of bisimplicial sets, and maps of marked simplicial sets, respectively. For ordinary categories, the isofibrations and equivalences coincide with the usual categorical notions bearing these names. For the quasi-categories, complete Segal spaces, Segal categories, and marked simplicial sets, the equivalences of $\infty$-categories are exactly the weak equivalences between fibrant-cofibrant objects in the model structure that is used to present the basic homotopy theory of each variety of $(\infty,1)$-category.
\end{ex}
So, each model of $(\infty,1)$-categories mentioned in the introduction has an $\infty$-cosmos. These are not the only examples, however:
\begin{ex} There exists $\infty$-cosmoi:
\begin{itemize}
\item $\theta_n$-$\mathcal{S}\mathrm{p}$, whose objects are \emph{$\theta_n$-spaces}, a simplicial presheaf model of $(\infty,n)$-categories (\extRef{yoneda}{IV}{ex:theta-n-cosmos})
\item $\mathcal{R}\mathrm{ezk}_\mathcal{M}$, whose objects are \emph{Rezk objects} in a sufficiently nice model category $\mathcal{M}$.\footnote{Here, ``sufficiently nice'' means permitting left Bousfield localization. With the definition of $\infty$-cosmos presented in \ref{qcat.ctxt.cof.def} we also need to require that the resulting fibrant objects are all cofibrant.} Rezk objects are used to define iterated complete Segal spaces, another simplicial presheaf model of $(\infty,n)$-categories (\extRef{yoneda}{IV}{prop:rezk-cosmos}).
\end{itemize}
Moreover, if $\mathcal{K}$ is any $\infty$-cosmos and $B \in \mathcal{K}$, then there is a \emph{sliced $\infty$-cosmos} $\mathcal{K}/B$, whose objects are isofibrations with codomain $B$. Sliced $\infty$-cosmoi will play a big role in \ref{sec:modules}.
\end{ex}
In summary, $\infty$-categories, for us, are the objects in a universe called an $\infty$-cosmos that is suitable for the development of their basic category theory --- much like a simplicial model category is a suitable environment in which to develop the basic homotopy theory of its objects. In our definition of an $\infty$-cosmos, we are not seeking to axiomatize the universe surrounding any particular variety of infinite-dimensional category, in contrast to \cite{Toen:2005vu} or \cite{BSP:2011ot}. Rather, the axioms outline what is needed to prove our theorems. The axiomatization presented here could also be made more general --- indeed, \cite{RiehlVerity:2015fy} uses a weaker definition of $\infty$-cosmos than will be considered here and further weakenings are also possible. Our aim is to optimize for simplicity of presentation, while applying sufficiently broadly. A perpetual challenge in category theory, or in many areas of abstract mathematics, is to find the right level of generality, which is often not the maximal level of generality.
\subsection{The homotopy 2-category of an \texorpdfstring{$\infty$}{infinity}-cosmos}
In fact most of our work to develop the basic theory of $\infty$-categories takes place not in their ambient $\infty$-cosmos, but in a quotient of the $\infty$-cosmos that we call the \emph{homotopy 2-category}. Each $\infty$-cosmos has an underlying 1-category whose objects are the $\infty$-categories of that $\infty$-cosmos and whose morphisms, which we call $\infty$-\emph{functors} or more often simply \emph{functors}, are the vertices of the mapping quasi-categories.
\begin{defn}[the homotopy 2-category of $\infty$-cosmos] The \emph{homotopy 2-category} of an $\infty$-cosmos $\mathcal{K}$ is a strict 2-category $\mathcal{K}_2$ so that
\begin{itemize}
\item the objects of $\mathcal{K}_2$ are the objects of $\mathcal{K}$, i.e., the $\infty$-categories;
\item the 1-cells $f \colon A \to B$ of $\mathcal{K}_2$ are the vertices $f \in \fun(A,B)$ in the mapping quasi-categories of $\mathcal{K}$, i.e., the $\infty$-functors;
\item a 2-cell $\xymatrix{ A \ar@/^2ex/[r]^f \ar@/_2ex/[r]_g \ar@{}[r]|{\Downarrow\alpha}& B}$ in $\mathcal{K}_2$ is represented by a 1-simplex $\alpha \colon f \to g \in \fun(A,B)$, where a parallel pair of 1-simplices in $\fun(A,B)$ represent the same 2-cell if and only if they bound a 2-simplex whose remaining outer face is degenerate.
\end{itemize}
Put concisely, the homotopy 2-category is the 2-category $\mathcal{K}_2 \mathrel{:=} h_*\mathcal{K}$ defined by applying the homotopy category functor $h \colon \category{qCat} \to \category{Cat}$ to the mapping quasi-categories of the $\infty$-cosmos; the hom-categories in $\mathcal{K}_2$ are defined by the formula \[\hom(A,B)\mathrel{:=} h(\fun(A,B))\] to be the homotopy categories of the mapping quasi-categories in $\mathcal{K}$.
\end{defn}
The homotopy 2-category $\category{qCat}_2$ of the $\infty$-cosmos of quasi-categories was first introduced by Joyal in his work on the foundations of quasi-category theory.
\begin{obs}[functors representing (isomorphic) 2-cells]
We write ${\bigger{1.12}{\mathbbe{2}}}$ for the simplicial set $\del@fn\parsedel@^1$, which is the nerve of the walking arrow $\bullet \to \bullet$. A natural transformation $\xymatrix{ A \ar@/^2ex/[r]^f \ar@/_2ex/[r]_g \ar@{}[r]|{\Downarrow\alpha}& B}$ in the homotopy 2-category of an $\infty$-cosmos $\mathcal{K}$ is represented by a map of simplicial sets $\alpha \colon {\bigger{1.12}{\mathbbe{2}}} \to \fun(A,B)$, which transposes to define a functor in $\mathcal{K}$ that composes with the two projections to the maps $g$ and $f$ respectively.
\[ \xymatrix{ & B \\ A \ar[r]^-{\alpha} \ar[ur]^f \ar[dr]_g & B^{\bigger{1.12}{\mathbbe{2}}} \ar@{->>}[u]_{p_0} \ar@{->>}[d]^{p_1} \\ & B}\]
The 2-cell $\xymatrix{ A \ar@/^2ex/[r]^f \ar@/_2ex/[r]_g \ar@{}[r]|{\Downarrow\alpha}& B}$ is an isomorphism in the homotopy 2-category $\mathcal{K}_2$ if and only if the arrow in $\hom(A,B)$ represented by the map $h(\alpha) \colon h{\bigger{1.12}{\mathbbe{2}}} \to h\fun(A,B)$ is an isomorphism. This is the case if an only if the representing simplicial map $\alpha \colon {\bigger{1.12}{\mathbbe{2}}} \to \fun(A,B)$ extends to a simplicial map $\alpha' \colon {\mathbb{I}} \to \fun(A,B)$, which transposes to define a functor in $\mathcal{K}$
\[ \xymatrix{ & B \\ A \ar[r]^-{{\alpha}'} \ar[ur]^f \ar[dr]_g & B^{\mathbb{I}} \ar@{->>}[u]_{p_0} \ar@{->>}[d]^{p_1} \\ & B}\]
\end{obs}
A priori, it is a bit of a surprise that the homotopy 2-category remembers enough information from the $\infty$-cosmos to develop the basic category theory of its objects. The first result that shows why this might be the case is the following.
\begin{prop}[{\extRef{yoneda}{IV}{prop:equiv.are.weak.equiv}}]\label{prop:equiv.are.weak.equiv} A functor $ f\colon A \to B$ is an equivalence in the $\infty$-cosmos $\mathcal{K}$ if and only if it is an equivalence in the homotopy 2-category $\mathcal{K}_2$.
\end{prop}
\begin{proof}
By definition, any equivalence $f \colon A \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} B$ in the $\infty$-cosmos induces an equivalence $\fun(X,A) \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} \fun(X,B)$ of quasi-categories for any $X$, which becomes an equivalence of categories $\hom(X,A) \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} \hom(X,B)$ upon applying the homotopy category functor $h \colon \category{qCat} \to \category{Cat}$. Applying the Yoneda lemma in the homotopy 2-category $\mathcal{K}_2$, it follows easily that $f$ is an equivalence in the sense of Definition \ref{defn:equivalence}.
Conversely, as the map ${\mathbb{I}} \to \del@fn\parsedel@^0$ of simplicial sets is a weak equivalence in the Joyal model structure, the cotensor $ B^{\mathbb{I}}$ defines a path object for the $\infty$-category $B$.
\[ \xymatrix{ & B^{\mathbb{I}} \ar@{->>}[dr]^-{(p_1,p_0)} \\ B \ar[rr]_\Delta \ar[ur]_*-{\rotatebox{45}{$\labelstyle\sim$}}^-{\Delta} & & B \times B}\]
It follows from the 2-of-3 property that any functor that is isomorphic in the homotopy 2-category to an equivalence in the $\infty$-cosmos is again an equivalence in the $\infty$-cosmos. Now it follows immediately from the 2-of-6 property for equivalences in the $\infty$-cosmos and the fact that the class of equivalences includes the identities, that any 2-categorical equivalence in the sense of Definition \ref{defn:equivalence} is an equivalence in the $\infty$-cosmos.
\end{proof}
The upshot is that any categorical notion defined up to equivalence in the homotopy 2-category is also characterized up to equivalence in the $\infty$-cosmos.
Axioms \ref{qcat.ctxt.cof.def}\ref{qcat.ctxt.cof:a} and \ref{qcat.ctxt.cof:b} imply that an $\infty$-cosmos has finite products satisfying a simplicially enriched universal property. Consequently:
\begin{prop}\label{prop:cartesian-closure}
The homotopy 2-category of an $\infty$-cosmos has finite products, and if the $\infty$-cosmos is cartesian closed, then so is its homotopy 2-category.
\end{prop}
\begin{proof}
The homotopy category functor $h \colon \category{qCat} \to \category{Cat}$ preserves finite products. Applying this to the defining isomorphisms $\fun(X,1)\cong \del@fn\parsedel@^0$ and $\fun(X, A\times B)\cong\fun(X,A)\times\fun(X,B)$ for the simplicially enriched terminal object and binary products of $\mathcal{K}$ yields isomorphisms $\hom(X,1)\cong {\bigger{1.16}{\mathbbe{1}}}$ and $\hom(X, A\times B)\cong\hom(X,A)\times\hom(X,B)$. These demonstrate that $1$ and $A\times B$ are also the 2-categorical terminal object and binary products in $\mathcal{K}_2$.
In this case where $\mathcal{K}$ is cartesian closed, applying the homotopy category functor to the defining isomorphisms on mapping quasi-categories yields the required natural isomorphisms
\[ \hom(A \times B, C) \cong \hom(A,C^B) \cong \hom(B,C^A)\] of hom-categories.
\end{proof}
\begin{defn}\label{defn:qcat-ctxt-functor} A \emph{functor of $\infty$-cosmoi} $F \colon \mathcal{K} \to \mathcal{L}$ is a simplicial functor that preserves isofibrations and the limits listed in~\ref{qcat.ctxt.cof.def}\ref{qcat.ctxt.cof:a}. Simplicial functoriality implies that a functor of $\infty$-cosmoi preserves equivalences and hence also trivial fibrations.
\end{defn}
For any $\infty$-cosmoi that arise as the fibrant objects in a Joyal-enriched model category, a simplicially enriched right Quillen adjoint will define a functor of $\infty$-cosmoi. This is the source of many of the following examples.
\begin{ex}
The following define functors of $\infty$-cosmoi:
\begin{itemize}
\item $\fun(X,-) \colon\mathcal{K} \to \category{qCat}$ for any object $X \in \mathcal{K}$ (see Proposition \extRef{yoneda}{IV}{prop:representable-functors}).
\item As a special case, the \emph{underlying quasi-category functor} $\fun(1,-) \colon \mathcal{K} \to \category{qCat}$. Examples include the functors $\category{CSS} \to \category{qCat}$ and $\category{Segal} \to \category{qCat}$ that take a complete Segal space or Segal category to its 0th row (see \extRef{yoneda}{IV}{ex:CSS-cosmos} and \extRef{yoneda}{IV}{ex:segal-cosmos}) and the functor $\category{sSet}_+\to \category{qCat}$ that carries a naturally marked simplicial set to its underlying quasi-category (see \extRef{yoneda}{IV}{ex:marked-cosmos}).
\item $(-)^U \colon \mathcal{K} \to \mathcal{K}$ for any simplicial set $U$, by \ref{qcat.ctxt.cof.def}\ref{qcat.ctxt.cof:b} and the fact that simplicially enriched limits commute with each other.
\item The inclusion $\category{Cat}\to \category{qCat}$ of categories into quasi-categories that identifies a category with its nerve (see \extRef{yoneda}{IV}{ex:cat-cosmos}).
\item The functor $t^! \colon \category{qCat} \to \category{CSS}$ defined in example \extRef{yoneda}{IV}{ex:other-CSS-functor}.
\item The functor $\category{CSS} \to \category{Segal}$ that ``discretizes'' the 0th space of a complete Segal space.
\end{itemize}
\end{ex}
The appropriate notion of functor between 2-categories is called a \emph{2-functor}, preserving all of the structure on the nose. A functor $F \colon \mathcal{K} \to \mathcal{L}$ of $\infty$-cosmoi induces a 2-functor $F_2 \mathrel{:=} h_*F \colon \mathcal{K}_2 \to \mathcal{L}_2$ between their homotopy 2-categories. Because adjunctions and equivalences in a 2-category are defined equationally, they are preserved by any 2-functor; in particular, the 2-functor between homotopy 2-categories induced by a functor of $\infty$-cosmoi preserves adjunctions and equivalences. Hence:
\begin{prop}\label{prop:induced-adjunctions}
If $\xymatrix{ B \ar@<1ex>[r]^f \ar@{}[r]|\perp & A \ar@<1ex>[l]^u}$ is an adjunction between $\infty$-categories then \begin{enumerate}[label=(\roman*)]
\item For any $\infty$-category $X$, $\xymatrix{ \fun(X,B) \ar@<1ex>[r]^{\fun(X,f)} \ar@{}[r]|\perp & \fun(X,A) \ar@<1ex>[l]^{\fun(X,u)}}$ defines an adjunction of quasi-categories.
\item\label{itm:induced-adjunctions:ii} For any $\infty$-category $X$, $\xymatrix{ \hom(X,B) \ar@<1ex>[r]^{\hom(X,f)} \ar@{}[r]|\perp & \hom(X,A) \ar@<1ex>[l]^{\hom(X,u)}}$ defines an adjunction of categories.
\item For any simplicial set $U$, $\xymatrix{ B^U \ar@<1ex>[r]^{ f^U} \ar@{}[r]|\perp & A^U \ar@<1ex>[l]^{u^U}}$ defines an adjunction of $\infty$-categories.
\item If the $\infty$-cosmos is cartesian closed, then for any $\infty$-category $C$ the pre- and post-composition functors define adjunctions of $\infty$-categories:
\[ \xymatrix{ B^C \ar@<1ex>[r]^{f_*} \ar@{}[r]|\perp & A^C \ar@<1ex>[l]^{u_*} & C^B \ar@<1ex>[r]^{u^*} \ar@{}[r]|\perp & C^A \ar@<1ex>[l]^{f^*}}\]
\end{enumerate}
\end{prop}
Taking $X = 1$ in \ref{itm:induced-adjunctions:ii} yields an adjunction between the \emph{homotopy categories} associated to the $\infty$-categories $A$ and $B$.
\begin{proof}
The adjunction $f \dashv u$ in $\mathcal{K}_2$ is preserved by the 2-functors $\fun(X,-)\colon \mathcal{K}_2 \to \category{qCat}_2$, $\hom(X,-) \colon \mathcal{K}_2 \to \category{Cat}_2$, $(-)^U\colon \mathcal{K}_2 \to \mathcal{K}_2$, $(-)^C \colon \mathcal{K}_2 \to \mathcal{K}_2$, and $C^{(-)} \colon \mathcal{K}_2^{\mathord{\text{\rm op}}} \to \mathcal{K}_2$.
\end{proof}
Via the simplicial cotensor and the embedding $\category{Cat}\hookrightarrow\category{sSet}$, for any $\infty$-category $A$ in an $\infty$-cosmos $\mathcal{K}$, there is also a 2-functor $A^{(-)} \colon \category{Cat}_2^{\mathord{\text{\rm op}}}\to \mathcal{K}_2$, which is another source of adjunctions between $\infty$-categories:
\begin{ex}[{\extRef{found}{I}{ex:comp.ident.adj}}]
For any $\infty$-category $A$, there is an adjunction
\[\xymatrix@C=10em{
*+[l]{ A^{\bigger{1.12}{\mathbbe{2}}}\times_AA^{\bigger{1.12}{\mathbbe{2}}}}
\ar[r]|*+{\scriptstyle m} &
{A^{\bigger{1.12}{\mathbbe{2}}}}
\ar@/^2.5ex/[l]^{i_1}_{}="l" \ar@/_2.5ex/[l]_{i_0}^{}="u"
\ar@{} "u";"l" |(0.2){\bot} |(0.8){\bot}
}\] between the ``composition'' functor $m$ and the pair of functors that ``extend an arrow into a composable pair'' by using the identities at its domain and codomain.
To prove this, first note that there exists a pair of adjunctions
\begin{equation*}
\xymatrix@C=10em@R=1ex{
{{{\bigger{1.12}{\mathbbe{2}}}}}\ar[r]|*+{\scriptstyle \delta^1} & {{{\bigger{1.12}{\mathbbe{3}}}}}
\ar@/^2.5ex/[l]^{\sigma^0}_{}="l" \ar@/_2.5ex/[l]_{\sigma^1}^{}="u"
\ar@{} "u";"l" |(0.2){\bot} |(0.8){\bot}
}
\end{equation*}
between ordinal categories so that the counit of the top adjunction and unit of the bottom adjunction are identities. Applying $A^{(-)} \colon \category{Cat}_2^{\mathord{\text{\rm op}}}\to \mathcal{K}_2$ converts these into adjunctions
\begin{equation*}
\xymatrix@C=10em{
{A^{\bigger{1.12}{\mathbbe{3}}}}
\ar[r]|*+{\scriptstyle A^{\delta^1}} &
{A^{\bigger{1.12}{\mathbbe{2}}}}
\ar@/^2.5ex/[l]^{A^{\sigma^1}}_{}="l" \ar@/_2.5ex/[l]_{A^{\sigma^0}}^{}="u"
\ar@{} "u";"l" |(0.2){\bot} |(0.8){\bot}
}
\end{equation*}
in which the upper adjunction has identity unit and the lower adjunction has identity counit. We write $A^{\bigger{1.12}{\mathbbe{2}}} \mathrel{:=} A^{\del@fn\parsedel@^1}$ and refer to this as the \emph{arrow $\infty$-category} associated to $A$ on account of a weak 2-categorical universal property that we will describe in \ref{sec:comma}.
The horn inclusion $\Lambda^{2,1}\hookrightarrow\del@fn\parsedel@^2$ is a trivial cofibration in Joyal's model structure, inducing an equivalence $p \colon A^{\bigger{1.12}{\mathbbe{3}}} \stackrel{\smash{\mkern-2mu\mathlower{1.5}{\sim}}}\longtwoheadrightarrow A^{\bigger{1.12}{\mathbbe{2}}} \times_A A^{\bigger{1.12}{\mathbbe{2}}}$ of $\infty$-categories, whose codomain we identify from the left-hand pushout in simplicial sets, which induces the right-hand pullback in the $\infty$-cosmos:
\begin{equation*}
\xymatrix@=2em{ \Lambda^{2,1} \save[]+<5pt,-5pt>*{\pbcorner}\restore & \del@fn\parsedel@^1 \ar[l]_-{\delta^2} & &
{A^{\Lambda^{2,1}}}\save[]+<5pt,-5pt>*{\pbcorner}\restore
\ar[r]^-{\pi_0}\ar[d]_-{\pi_1} & {A^{\bigger{1.12}{\mathbbe{2}}}}\ar@{->>}[d]^-{p_1} \\ \del@fn\parsedel@^1 \ar[u]^{\delta^0} & \del@fn\parsedel@^0 \ar[u]_{\delta^0}\ar[l]^-{\delta^1} & &
{A^{\bigger{1.12}{\mathbbe{2}}}}\ar@{->>}[r]_-{p_0} & A
}
\end{equation*}
Proposition \ref{prop:equiv.are.weak.equiv} tells us that $p \colon A^{\bigger{1.12}{\mathbbe{3}}} \stackrel{\smash{\mkern-2mu\mathlower{1.5}{\sim}}}\longtwoheadrightarrow A^{\bigger{1.12}{\mathbbe{2}}} \times_A A^{\bigger{1.12}{\mathbbe{2}}}$ defines an equivalence in the homotopy 2-category. In particular, by Corollary \ref{cor:equivs-are-adjoints}, $p$ admits an equivalence inverse $p'$ that is simultaneously a left and a right adjoint.
Composing $p \dashv p' \dashv p$ with the displayed adjunction, we obtain the adjunctions $i_0 \dashv m \dashv i_1$.
In fact, these adjunctions can be defined so that the unit and counit 2-cells, and not just the functors, are fibered over the endpoint evaluation functors $A^{\bigger{1.12}{\mathbbe{2}}}\times_AA^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A \times A$ and $A^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A \times A$. The proof makes use of the fact that these maps are isofibrations; see \extRef{found}{I}{ex:comp.ident.adj}.
\end{ex}
\renewcommand\thesection{Lecture~\arabic{section}}
\section{Limits and colimits in \texorpdfstring{$\infty$}{infinity}-categories}\label{sec:limits}
\renewcommand\thesection{\arabic{section}}
Recall that we use the term $\infty$-\emph{category} to refer to any variety of infinite-dimensional category that inhabits an $\infty$-cosmos. An $\infty$-\emph{cosmos} is a simplicially enriched category $\mathcal{K}$, whose homs $\fun(A,B)$ are quasi-categories, that admits certain simplicially-enriched limit constructions and whose specified class of \emph{isofibrations} enjoy certain closure properties. The objects and morphisms of the underlying category of $\mathcal{K}$ define the $\infty$-\emph{categories} and $\infty$-\emph{functors} of the $\infty$-cosmos. There are $\infty$-cosmoi for quasi-categories, complete Segal spaces, Segal categories, and naturally marked simplicial sets, each of these being a model of $(\infty,1)$-categories.
Our development of the basic theory of adjunctions and equivalences between $\infty$-cat\-e\-gor\-ies takes place entirely within the \emph{homotopy 2-category} $\mathcal{K}_2$ of an $\infty$-cosmos $\mathcal{K}$. The objects and morphisms in the homotopy 2-category are again the $\infty$-categories and $\infty$-functors. A 2-cell $\xymatrix{ A \ar@/^2ex/[r]^r \ar@/_2ex/[r]_{a'} \ar@{}[r]|{\Downarrow \alpha} &B}$ between a parallel pair of functors between $\infty$-categories is represented by a 1-simplex $\alpha \colon f \to g \in \fun(A,B)$, where two parallel 1-simplices are equivalent if and only if they bound a 2-simplex whose third edge is degenerate. More concisely, the hom-category $\hom(A,B)$ between two objects in the homotopy 2-category is the homotopy category of the mapping quasi-category $\fun(A,B)$.
Proposition \ref{prop:cartesian-closure} demonstrates that the homotopy 2-category of any $\infty$-cosmos has finite products, satisfying a 2-dimensional universal property. For the terminal $\infty$-category 1, this says that $\hom(X,1) \cong {\bigger{1.16}{\mathbbe{1}}}$, that is there is a unique $\infty$-functor $! \colon X \to 1$ for any $\infty$-category $X$ and this functor admits no non-identity endomorphisms. Proposition \ref{prop:equiv.are.weak.equiv} demonstrates that the notions of equivalence between $\infty$-categories defined at the level of the $\infty$-cosmos coincide precisely with the notions of equivalence in the homotopy 2-category, the upshot being that equivalence-invariant 2-categorical constructions are appropriately homotopical.
Fix an ambient $\infty$-cosmos $\mathcal{K}$. We will work inside this universe for the remainder of this lecture. Our aim now is to define appropriate notions of limit and colimit of diagrams taking values in an $\infty$-category $A$ inside the $\infty$-cosmos. To define limits and colimits, we need a way to look inside the $\infty$-category $A$ without leaving the comfort of the $\infty$-cosmos axiomatization. For this, we make use of the terminal $\infty$-category $1$. A functor $a \colon 1\to A$ will be called an \emph{element}\footnote{Synonyms include \emph{point} or \emph{object}. The term ``element'' is perhaps less traditional but also less likely to be confused with other mathematical notions currently under consideration.} of $A$; by analogy, a functor $a \colon X \to A$ is a \emph{generalized element} of $A$. Elements of $A$ and the 2-cells between them define the \emph{homotopy category} of the $\infty$-category $A$, as we record in passing.
\begin{defn}\label{defn:homotopy-category}
The \emph{homotopy category} of an $\infty$-category $A$ is the category $\hom(1,A)$ whose objects are the elements of $A$ and whose morphisms $f \colon a \to a'$ are 2-cells $\xymatrix{ 1 \ar@/^2ex/[r]^a \ar@/_2ex/[r]_{a'} \ar@{}[r]|{\Downarrow f} &A}$.
\end{defn}
\subsection{Terminal elements}
Before introducing the general notion of limits, we will warm up with a special case of terminal objects, which we will call \emph{terminal elements}.
\begin{defn}\label{defn:terminal-element} A \emph{terminal element} in an $\infty$-category $A$ is a right adjoint $t \colon 1 \to A$ to the unique functor $! \colon A \to 1$. Explicitly, the data consists of:
\begin{itemize}
\item an element $t \colon 1 \to A$ and
\item a natural transformation $\eta \colon \id_A \Rightarrow t!$ whose component $\eta t$ at the element $t$ is an isomorphism.\footnote{If $\eta$ is the unit of the adjunction $! \dashv t$, then the triangle identities demand that $\eta t =\id_t$. However, by a 2-categorical trick, to show that such an adjunction exists, it suffices to find a 2-cell $\eta$ so that $\eta t$ is an isomorphism (see \extRef{found}{I}{lem:min-term-pres}).}
\end{itemize}
\end{defn}
Several basic facts about terminal elements can be deduced immediately from the general theory of adjunctions.
\begin{prop}$\quad$
\begin{enumerate}[label=(\roman*)]
\item\label{itm:term-i} An element $t \colon 1 \to A$ is terminal if and only if it is representably terminal, i.e., if for all $f \colon X \to A$ there exists a unique 2-cell $\vcenter{\xymatrix@=1em{ X \ar[rr]^f \ar[dr]_{!} & \ar@{}[d]|(.4){\Downarrow\exists!} & A \\ & 1 \ar[ur]_t}}$
\item\label{itm:term-ii} Terminal elements are preserved by right adjoints and by equivalences.
\item\label{itm:term-iii} If $A' \simeq A$ then $A$ has a terminal element if and only if $A'$ does.
\end{enumerate}
\end{prop}
\begin{proof}
For \ref{itm:term-i}, Proposition \ref{prop:induced-adjunctions}\ref{itm:induced-adjunctions:ii} proves that terminal elements are representably terminal; the converse follows from the Yoneda lemma. \ref{itm:term-ii} is a special case of Proposition \ref{prop:adjunctions-compose}, via Corollary \ref{cor:equivs-are-adjoints}; \ref{itm:term-iii} follows.
\end{proof}
\subsection{Limits}
Terminal elements are limits indexed by the empty set. We now turn to limits of generic diagrams whose indexing shapes are given by simplicial sets. We have a 2-category $\category{sSet}_2$ of simplicial sets, extending in the evident way the definition of the homotopy 2-category $\category{qCat}_2 \subset \category{sSet}_2$ of quasi-categories. The 2-category of categories sits as a full subcategory $\category{Cat}_2\subset\category{qCat}_2 \subset \category{sSet}_2$, with categories identified with the simplicial sets defining their nerves. In this way, diagrams indexed by categories are among the diagrams indexed by simplicial sets.
For any $\infty$-category $A$ in an $\infty$-cosmos $\mathcal{K}$, there is a simplicial functor $A^{(-)} \colon \category{sSet}^{\mathord{\text{\rm op}}} \to \mathcal{K}$, which descends to a 2-functor $A^{(-)} \colon \category{sSet}_2^{\mathord{\text{\rm op}}} \to \mathcal{K}_2$. These simplicial cotensors are used to define $\infty$-categories of diagrams.
\begin{defn}[diagram $\infty$-categories]\label{defn:diagram-cats} If $J$ is a simplicial set and $A$ is an $\infty$-category, then the $\infty$-category $A^J$ is the \emph{$\infty$-category of $J$-indexed diagrams in $A$}.
\end{defn}
\begin{rmk}
In the case where the $\infty$-cosmos is cartesian closed, in the sense of Definition \ref{qcat.ctxt.cof.def}\ref{qcat.ctxt.cof:g}, we could instead take the indexing shape $J$ to be an $\infty$-category, in which case the internal hom $A^J$ is the \emph{$\infty$-category of $J$-indexed diagrams in $A$}. The development of the theory of limits indexed by an $\infty$-category in a cartesian closed $\infty$-cosmos entirely parallels the development for limits indexed by a simplicial set. The conflated notation of \ref{qcat.ctxt.cof.def}\ref{qcat.ctxt.cof:a} and \ref{qcat.ctxt.cof.def}\ref{qcat.ctxt.cof:g} is intended to further highlight this parallelism.
\end{rmk}
In analogy with Definition \ref{defn:terminal-element}, we have:
\begin{defn}\label{defn:all-limits} An $\infty$-category $A$ \emph{admits all limits of shape $J$} if the constant diagram functor $\Delta \colon A \to A^J$, induced by the unique functor $!\colon J \to 1$, has a right adjoint:
\[ \xymatrix{ A \ar@<1ex>[r]^-\Delta \ar@{}[r]|-\perp & A^J \ar@<1ex>[l]^-{\lim}}\]
\end{defn}
From the vantage point of Definition \ref{defn:all-limits}, the following result is easy:
\begin{exs} Show, using \ref{prop:adjunctions-compose} and \ref{exs:iso-invar-adjunction}, that a right adjoint functor $u \colon A \to B$ between $\infty$-categories that admit all limits of shape $J$ necessarily preserves them, in the sense that the functors
\[ \xymatrix{ A^J \ar[d]_{\lim} \ar[r]^{u^J} & B^J \ar[d]^{\lim} \ar@{}[dl]|\cong \\ A \ar[r]_u & B}\] commute up to isomorphism.
\end{exs}
The problem with Definition \ref{defn:all-limits} is that it is insufficiently general: many $\infty$-categories will have certain, but not all, limits of diagrams of a particular indexing shape. With this aim in mind, we will now re-express Definition \ref{defn:all-limits} in a form that permits its extension to cover this sort of situation. For this, we make use of the following 2-categorical notion.
\begin{defn}[absolute right lifting] Given a cospan $C \xrightarrow{g} A \xleftarrow{f} B$, a functor $\ell \colon C \to B$ and a 2-cell \begin{equation}\label{eq:abs-right-lifting} \xymatrix{ \ar@{}[dr]|(.7){\Downarrow\lambda} & B \ar[d]^f \\ C \ar[ur]^\ell \ar[r]_g & A}\end{equation} define an \emph{absolute right lifting of $g$ through $f$} if any 2-cell as displayed below-left factors uniquely through $\lambda$ as displayed below-right
\[ \vcenter{\xymatrix{ X \ar[d]_c \ar[r]^b \ar@{}[dr]|{\Downarrow\chi} & B \ar[d]^f \\ C \ar[r]_g & A}} \mkern20mu = \mkern20mu \vcenter{\xymatrix{ X \ar[d]_c \ar[r]^b \ar@{}[dr]|(.3){\exists !\Downarrow}|(.7){\Downarrow\lambda} & B \ar[d]^f \\ C \ar[ur]|(.4)*+<2pt>{\scriptstyle \ell} \ar[r]_g & A}}
\]
\end{defn}
We refer to the 2-cell \eqref{eq:abs-right-lifting} as an \emph{absolute right lifting diagram}. In category theory, the term ``absolute'' typically means ``preserved by all functors.'' An absolute right lifting diagram is a right lifting diagram $\lambda \colon f\ell \Rightarrow g$ so that the restriction of $\lambda$ along any generalized element $c \colon X \to C$ again defines a right lifting diagram.
\begin{exs}\label{exs:adj-as-abs-lifting} Show that in any 2-category, a 2-cell $\epsilon \colon fu \Rightarrow \id_A$ defines the counit of an adjunction $f \dashv u$ if and only if
\[ \xymatrix{ \ar@{}[dr]|(.7){\Downarrow\epsilon} & B \ar[d]^f \\ A \ar[ur]^u \ar@{=}[r] & A}\] defines an absolute right lifting diagram.
\end{exs}
Applying Exercise \ref{exs:adj-as-abs-lifting}, Definition \ref{defn:all-limits} is equivalent to the assertion that the \emph{limit cone}, our term for the counit of $\Delta \dashv \lim$ defines an absolute right lifting diagram:
\begin{equation}\label{eq:all-limits-abs-lifting} \xymatrix{ \ar@{}[dr]|(.7){\Downarrow\epsilon} & A\ar[d]^\Delta \\ A^J \ar@{=}[r] \ar[ur]^{\lim} & A^J}\end{equation} This motivates the following definition.
\begin{defn}[limit]\label{defn:limit} A \emph{limit} of a $J$-indexed diagram in $A$ is an absolute right lifting of the diagram $d$ through the constant diagram functor $\Delta \colon A \to A^J$
\begin{equation}\label{eq:lim-diagram-defn} \xymatrix{ \ar@{}[dr]|(.7){\Downarrow\lambda} & A\ar[d]^\Delta \\ 1 \ar[r]_d \ar[ur]^{\lim d} & A^J}\end{equation} the 2-cell component of which defines the \emph{limit cone} $\lambda \colon \Delta \lim d \Rightarrow d$.
\end{defn}
If $A$ has all $J$-indexed limits, then the restriction of the absolute right lifting diagram \eqref{eq:all-limits-abs-lifting} along the element $d \colon 1\to A^J$ defines a limit for $d$. Interpolating between Definitions \ref{defn:limit} and \ref{defn:all-limits}, we can define a \emph{limit of a family of diagrams} to be an absolute right lifting of the family $d \colon K \to A^J$ through $\Delta \colon A \to A^J$. For instance:
\begin{thm}[{\extRef{found}{I}{thm:splitgeorealizations}}]\label{thm:totalization} For every cosimplicial object in an $\infty$-category that admits an coaugmentation and a splitting, the coaugmentation defines its limit. That is, for every $\infty$-category $A$, the functors
\[ \xymatrix{ \ar@{}[dr]|(.7){\Downarrow\lambda} & A\ar[d]^\Delta \\ A^{\del@fn\parsedel@[b]} \ar[r]_{\mathrm{res}} \ar[ur]^{\ev_{[-1]}} & A^{\del@fn\parsedel@}}\] define an absolute right lifting diagram.
\end{thm}
Here $\del@fn\parsedel@$ is the usual simplex category of finite non-empty ordinals and order-preserving maps. It defines a full subcategory of $\del@fn\parsedel@+$, which freely appends an initial object $[-1]$, and this in turn defines a wide subcategory\footnote{A \emph{wide} subcategory is a subcategory containing all of the objects.} of $\del@fn\parsedel@[b]$, which adds an ``extra degeneracy'' map between each pair of consecutive ordinals. Diagrams indexed by $\del@fn\parsedel@ \subset \del@fn\parsedel@+\subset \del@fn\parsedel@[b]$ are, respectively, called \emph{cosimplicial objects}, \emph{coaugmented cosimplicial objects}, and \emph{split cosimplicial objects}. The limit of a cosimplicial object is often called its \emph{totalization}.
\begin{proof}[Proof sketch]
In $\category{Cat}_2$, there is a canonical 2-cell
\[ \xymatrix{ \del@fn\parsedel@ \ar@{^(->}[r] \ar[d]_{!} & \del@fn\parsedel@[b] \\ {\bigger{1.16}{\mathbbe{1}}} \ar[ur]_{[-1]} & \ar@{}[ul]|(.7){\Uparrow\lambda}}\]
because $[-1] \in \del@fn\parsedel@[b]$ is initial. This data defines an absolute right extension diagram that is moreover preserved by any 2-functor, because the universal property of the functor $[-1] \colon {\bigger{1.16}{\mathbbe{1}}} \to \del@fn\parsedel@[b]$ and the 2-cell $\lambda$ is witnessed by a pair of adjunctions (see \extRef{found}{I}{lem:doms2catlemma}). The 2-functor $A^{(-)} \colon \category{Cat}_2^{\mathord{\text{\rm op}}} \to \mathcal{K}_2$ converts this into the absolute right lifting diagram of the statement.
\end{proof}
The most important result relating adjunctions and limits is of course:
\begin{thm}[{\extRef{found}{I}{prop:RAPL}}]\label{thm:RAPL} Right adjoints preserve limits.
\end{thm}
Our proof will closely follow the classical one. Given a diagram $d\colon 1 \to A^J$ and a right adjoint $u \colon A \to B$ to some functor $f$, a cone with summit $b \colon 1 \to B$ over $u^J d$ transposes to define a cone with summit $fb$ over $d$, which factors uniquely through the limit cone. This factorization transposes back across the adjunction to show that $u$ carries the limit cone over $d$ to a limit cone over $u^Jd$.
\begin{proof}
Suppose that $A$ admits limits of a diagram $d\colon 1\to A^J$ as witnessed by an absolute right lifting diagram \eqref{eq:lim-diagram-defn}. By Proposition \ref{prop:induced-adjunctions}, an adjunction $f \dashv u$ induces an adjunction $f^J \dashv u^J$. We must show that \[\xymatrix{ \ar@{}[dr]|(.7){\Downarrow\lambda} & A \ar[d]^-\Delta \ar[r]^u & B \ar[d]^-\Delta \\ 1 \ar[ur]^{\lim d} \ar[r]_d& A^J \ar[r]_{u^J} & B^J}\] is again an absolute right lifting diagram. Given a square
\[\xymatrix{ X \ar[d]_{!} \ar[rr]^b \ar@{}[drr]|{\Downarrow\chi} & & B \ar[d]^-{\Delta} \\ 1 \ar[r]_-{d} & A^J \ar[r]_{u^J} & B^J} \] we first ``transpose across the adjunction,'' by composing with $f$ and the counit.
\[\vcenter{\xymatrix{ X \ar[d]_-{!} \ar[rr]^b \ar@{}[drr]|{\Downarrow\chi} & & B \ar[d]^-\Delta \ar[r]^f & A \ar[d]^-\Delta \\ 1 \ar[r]_-{d} & A^J \ar@{=}@/_3.5ex/[rr]^{\Downarrow\epsilon^J} \ar[r]^{u^J} & B^J \ar[r]^{f^J} & A^J}} = \vcenter{\xymatrix{ X \ar@{}[drr]|(.3){\exists !\Downarrow\zeta}|(.7){\Downarrow\lambda} \ar[d]_-{!} \ar[r]^b & B \ar[r]^f & A \ar[d]^-\Delta \\ 1 \ar[urr]_(0.4){\lim d} \ar[rr]_{d} & & A^J}} \] The universal property of the absolute right lifting diagram $\lambda \colon \Delta \lim \Rightarrow d$ induces a unique factorisation $\zeta$, which may then be ``transposed back across the adjunction'' by composing with $u$ and the unit.
\[ \vcenter{\xymatrix{ X \ar@{}[drr]|(.3){\exists !\Downarrow\zeta}|(.7){\Downarrow\lambda} \ar[d]_-{!} \ar[r]^b & B \ar@{=}@/^3.5ex/[rr]_{\Downarrow\eta} \ar[r]|f & A \ar[d]^-\Delta \ar[r]_u & B \ar[d]^-\Delta \\ 1 \ar[urr]_(0.4){\lim d} \ar[rr]_-{d} & & A^J \ar[r]_{u^J} & B^J}}= \vcenter{\xymatrix{ X \ar[d]_-{!} \ar[rr]^b \ar@{}[drr]|{\Downarrow\chi} & & B \ar[d]^-\Delta \ar@{=}@/^3.5ex/[rr]_{\Downarrow\eta} \ar[r]_f & A \ar[d]^-\Delta \ar[r]_u & B \ar[d]^-\Delta \\ 1 \ar[r]_-{d} & A^J \ar@{=}@/_3.5ex/[rr]^{\Downarrow\epsilon^J} \ar[r]^{u^J} & B^J \ar[r]^{f^J} & A^J
\ar[r]_{u^J} & B^J}} \] \[ = \vcenter{\xymatrix{ X \ar[d]_-{!} \ar[rr]^b \ar@{}[drr]|{\Downarrow\chi} & & B \ar[d]^-\Delta \ar@{=}@/^3.5ex/[rr] & & B \ar[d]^-\Delta \\ 1 \ar[r]_-{d} & A^J \ar@{=}@/_3.5ex/[rr]^{\Downarrow\epsilon^J} \ar[r]^{u^J} & B^J \ar[r]|{f^J} \ar@{=}@/^3.5ex/[rr]_{\Downarrow\eta^J}& A^J \ar[r]_{u^J} & B^J}} = \vcenter{\xymatrix{ X \ar[d]_-{!} \ar[rr]^b \ar@{}[drr]|{\Downarrow\chi} & & B \ar[d]^-\Delta \\ 1 \ar[r]_-{d} & A^J \ar[r]_{u^J} & B^J}}\] Here the second equality is a consequence of the 2-functoriality of the simplicial cotensor, while the third is an application of a triangle identity for the adjunction $f^J \dashv u^J$. The pasted composite of $\zeta$ and $\eta$ is the desired factorisation of $\chi$ through $\lambda$.
The proof that this factorization is unique, which again parallels the classical argument, is left to the reader: the essential point is that the transposes defined via these pasting diagrams are unique.
\end{proof}
The same argument also shows that a right adjoint preserves the limit of a family of diagrams $d \colon K \to A^J$. On account of Exercise \ref{exs:adjoint-equivalence}, we have the immediate corollary:
\begin{cor}\label{cor:equiv-pres-lim} Equivalences preserve limits.
\end{cor}
Moreover, ``completeness'' of $\infty$-categories is transferred along equivalences.
\begin{prop}\label{prop:completeness-equivalence} If $A \simeq B$ then any family of diagrams in $A$ that admits limits in $B$ also admits limits in $A$.
\end{prop}
\begin{proof}
Consider a family of diagrams $d\colon K \to A^J$ of shape $J$ in $A$ that admits limits in $B$ after composing with the equivalence $A \simeq B$, given by an absolute right lifting diagram:
\[ \xymatrix{ \ar@{}[drr]|(.7){\Downarrow \lambda} & & B \ar[d]^\Delta \\ K \ar[urr]^\lim \ar[r]_d & A^J \ar[r]_{\sim} & B^J}\] We claim that the composite 2-cell
\[ \xymatrix{ \ar@{}[drr]|(.7){\Downarrow \lambda} & & B \ar[d]^\Delta \ar[r]^\sim & A \ar[d]^\Delta \\ K \ar[urr]^\lim \ar[r]_d & A^J\ar@{=}@/_3.5ex/[rr]^{\Downarrow\cong} \ar[r]_{\sim} & B^J \ar[r]^\sim & A^J }\] again defines an absolute right lifting diagram, proving that the original family $d \colon K \to A^J$ admits limits in $A$. Promoting the equivalence to an adjoint equivalence, as in Exercise \ref{exs:adjoint-equivalence}, the proof of this universal property is very similar to the proof of Theorem \ref{thm:RAPL}; the remaining details are left as an exercise.
\end{proof}
\subsection{Colimits}
The theory of colimits of $J$-indexed diagrams in an $\infty$-category $A$ is dual to the theory of limits by reversing the direction of the 2-cells but not the 1-cells in the ambient homotopy 2-category $\mathcal{K}_2$.
\begin{defn}[absolute left lifting] Given a cospan $C \xrightarrow{g} A \xleftarrow{f} B$, a functor $\ell \colon C \to B$ and a 2-cell \[\xymatrix{ \ar@{}[dr]|(.7){\Uparrow\lambda} & B \ar[d]^f \\ C \ar[ur]^\ell \ar[r]_g & A}\] define an \emph{absolute left lifting of $g$ through $f$} if any 2-cell as displayed below-left factors uniquely through $\lambda$ as displayed below-right
\[ \vcenter{\xymatrix{ X \ar[d]_c \ar[r]^b \ar@{}[dr]|{\Uparrow\chi} & B \ar[d]^f \\ C \ar[r]_g & A}} \mkern20mu = \mkern20mu \vcenter{\xymatrix{ X \ar[d]_c \ar[r]^b \ar@{}[dr]|(.3){\exists !\Uparrow}|(.7){\Uparrow\lambda} & B \ar[d]^f \\ C \ar[ur]|(.4)*+<2pt>{\scriptstyle \ell} \ar[r]_g & A}}
\]
\end{defn}
\begin{defn}[colimit]\label{defn:colimit} A \emph{colimit} of a $J$-indexed diagram in $A$ is an absolute left lifting of the diagram $d$ through the constant diagram functor $\Delta \colon A \to A^J$
\[ \xymatrix{ \ar@{}[dr]|(.7){\Uparrow\lambda} & A\ar[d]^\Delta \\ 1 \ar[r]_d \ar[ur]^{\colim d} & A^J}\] the 2-cell component of which defines the \emph{colimit cone} $\lambda \colon f \Rightarrow \Delta \colim f$.
\end{defn}
We leave the formulation of the evident duals of \ref{thm:totalization}, \ref{thm:RAPL}, \ref{cor:equiv-pres-lim}, and \ref{prop:completeness-equivalence} as an exercise.
\subsection{Arrow \texorpdfstring{$\infty$}{infinity}-categories}
By design, our definitions of limits and colimits of diagrams in an $\infty$-category and of adjunctions between $\infty$-categories are in a form that can be easily expressed internally to a 2-category. Consequently, we have not yet described the universal properties encoded by the $\infty$-functors represented by limit or colimit elements. Our aim in the next lecture will be to present equivalent definitions of limits, colimits, and adjunctions that can be given in these terms. These will require a more substantial use of the $\infty$-cosmos axiomatization than we have needed thus far and will also form the basis for our ``model independence'' results.
To prepare the way for this discussion, we now briefly introduce the $\infty$-categories that will serve as a vehicle to encode these sorts of representable universal properties, at least in the very simplest case: these are \emph{arrow $\infty$-categories}.
Here is the idea that motivates their importance. Recall that we use the terminal $\infty$-category $1$ to probe inside an $\infty$-category $A$, in the sense that an element of $A$ is defined to be a functor $a \colon 1 \to A$. We would like to be able to probe similarly for arrows in $A$. An approximate notion is given by a 2-cell of the form $\xymatrix{1 \ar@/^2ex/[r]^a \ar@/_2ex/[r]_{a'} \ar@{}[r]|{\Downarrow f} & A}$, but these correspond to arrows in the homotopy category of $A$, as in Definition \ref{defn:homotopy-category}, rather than arrows in $A$ itself. If we had a ``walking arrow'' $\infty$-category ${\bigger{1.12}{\mathbbe{2}}}$, then we could define an arrow in $A$ to be a functor $f\colon {\bigger{1.12}{\mathbbe{2}}} \to A$. The axioms of an $\infty$-cosmos do not guarantee this, but we can represent functors of this type in ``transposed'' form, as elements $f \colon 1 \to A^{\bigger{1.12}{\mathbbe{2}}}$ in the arrow $\infty$-category that we now define.
\begin{defn}[arrow $\infty$-categories] For any $\infty$-category $A$, the simplicial cotensor
\[ \xymatrix@C=30pt{ A^{\bigger{1.12}{\mathbbe{2}}} \mathrel{:=} A^{\del@fn\parsedel@^1} \ar@{->>}[r]^-{(p_1,p_0)} & {A^{\partial\Delta^1}} \cong A \times A}\] defines the \emph{arrow $\infty$-category} $A^{\bigger{1.12}{\mathbbe{2}}}$, equipped with an isofibration $(p_1,p_0)\colon A^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A \times A$, where $p_1 \colon A^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A$ denotes the codomain projection and $p_0 \colon A^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A$ denotes the domain projection.
\end{defn}
Using the notation ${\bigger{1.12}{\mathbbe{2}}} \mathrel{:=} \Delta^1$, the defining universal property of the simplicial cotensor asserts that the canonical map defines an isomorphism of quasi-categories
\[ \fun(X, A^{\bigger{1.12}{\mathbbe{2}}}) \stackrel{\cong}{\longrightarrow} \fun(X,A)^{{\bigger{1.12}{\mathbbe{2}}}}.\] In particular, taking $X=A^{\bigger{1.12}{\mathbbe{2}}}$, the identity functor $\id_{A^{\bigger{1.12}{\mathbbe{2}}}}$ transposes to define a vertex in $\fun(A^{\bigger{1.12}{\mathbbe{2}}},A)^{\bigger{1.12}{\mathbbe{2}}}$ which represents a 2-cell
\begin{equation}\label{eq:generic-arrow} \xymatrix{ A^{\bigger{1.12}{\mathbbe{2}}} \ar@{->>}@/^2ex/[r]^{p_0} \ar@{->>}@/_2ex/[r]_{p_1} \ar@{}[r]|-{\Downarrow\phi} & A}\end{equation} in the homotopy 2-category $\mathcal{K}_2$.
There is an analogous categorical notion of cotensor with ${\bigger{1.12}{\mathbbe{2}}}$: if $C$ is an ordinary 1-category, then $C^{\bigger{1.12}{\mathbbe{2}}}$ is the category whose objects are morphisms in $C$ and whose morphisms are commutative squares. In particular, for any pair of $\infty$-categories, we can form the arrow category $\hom(X,A)^{\bigger{1.12}{\mathbbe{2}}}$ of the hom-category between them. Now the arrow $\infty$-category $A^{\bigger{1.12}{\mathbbe{2}}}$ would be a strict ${\bigger{1.12}{\mathbbe{2}}}$-cotensor in the homotopy 2-category $\mathcal{K}_2$ if there were an isomorphism between $\hom(X,A^{\bigger{1.12}{\mathbbe{2}}})$ and the arrow category $\hom(X,A)^{\bigger{1.12}{\mathbbe{2}}}$, analogous to the isomorphism of mapping quasi-categories that demonstrates that $A^{\bigger{1.12}{\mathbbe{2}}}$ is a strict $\Delta^1$-cotensor in $\mathcal{K}$. Instead:
\begin{prop}[{\extRef{found}{I}{prop:weak-cotensors}}]\label{prop:weak-cotensors} For any $\infty$-categories $X$ and $A$, the canonical functor
\[ \hom(X, A^{\bigger{1.12}{\mathbbe{2}}}) \longrightarrow \hom(X,A)^{{\bigger{1.12}{\mathbbe{2}}}}\] is \emph{smothering}: that is, surjective on objects, full, and conservative.\footnote{A functor is \emph{conservative} when it reflects isomorphisms.}
\end{prop}
Surjectivity on objects asserts that any 2-cell $\alpha \colon f \Rightarrow g \colon X \to A$ is represented by a functor $\hat{\alpha} \colon X \to A^{\bigger{1.12}{\mathbbe{2}}}$ with the property that the whiskered composite of the representing functor with the canonical 2-cell $\phi \colon p_0 \Rightarrow p_1$ equals $\alpha$:
\[ \vcenter{\xymatrix{ X \ar[r]^-{\hat{\alpha}} & A^{\bigger{1.12}{\mathbbe{2}}} \ar@{->>}@/^2ex/[r]^{p_0} \ar@{->>}@/_2ex/[r]_{p_1} \ar@{}[r]|-{\Downarrow\phi} & A}} = \vcenter{ \xymatrix{X \ar@/^2ex/[r]^f \ar@/_2ex/[r]_g \ar@{}[r]|{\Downarrow\alpha} & A}}\]
These representatives are not unique:
\begin{exs} Show that any parallel pair of functors $\hat\alpha, \hat{\alpha}' \colon X \to A^{\bigger{1.12}{\mathbbe{2}}}$ over $A \times A$ that are connected by an invertible 2-cell that projects to the identity along $(p_1,p_0) \colon A^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A \times A$ will necessarily represent the same 2-cell $\alpha \colon f \Rightarrow g$.
\[ \xymatrix@=1.5em{ X \ar@/^2ex/[rr]^{\hat\alpha} \ar@/_2ex/[rr]_{{\hat\alpha}'} \ar@{}[rr]|{\Downarrow\cong} \ar[dr]_{(g,f)} & & A^{\bigger{1.12}{\mathbbe{2}}} \ar@{->>}[dl]^{(p_1,p_0)} \\ & A \times A}\] Then use Proposition \ref{prop:weak-cotensors} to prove the converse: that any parallel pair of functors that represent the same 2-cell $\alpha \colon f \Rightarrow g$ are connected by an isomorphism of this form.
Lemma \ref{lem:ess.unique.1-cell.ind}, proven below, generalizes this result.
\end{exs}
This non-uniqueness of representing functors implies, in particular, that the functor of Proposition \ref{prop:weak-cotensors} cannot define an isomorphism of categories. Nonetheless, we are able to make substantial use of the \emph{weak 2-dimensional universal property} of the arrow construction, as expressed by Proposition \ref{prop:weak-cotensors}, as we shall shortly discover.
\renewcommand\thesection{Lecture~\arabic{section}}
\section{Comma \texorpdfstring{$\infty$}{infinity}-categories and model independence}\label{sec:comma}
\renewcommand\thesection{\arabic{section}}
In the first two parts of this series, we have introduced notions of adjunctions between and limits and colimits of diagrams valued within $\infty$-categories, which are objects in some well-behaved universe we call an $\infty$-cosmos. In our treatment of adjunctions between $\infty$-categories and limits and colimits of diagrams valued in an $\infty$-category we have privileged definitions of these basic categorical notions that can be defined internally to any 2-category. Consequently, we have not yet seen the analogues of, for instance, the idea that an adjunction encodes a natural correspondence between certain arrows in a pair of $\infty$-categories, or the idea that a limit defines a terminal object in the $\infty$-category of cones. Our aim in this lecture is to present equivalent definitions in terms of these universal properties, utilizing additional structures in the homotopy 2-category guaranteed by the $\infty$-cosmos axioms. The axioms imply that the homotopy 2-category of an $\infty$-cosmos admits the construction of the \emph{comma $\infty$-category} for any cospan of functors $C \xrightarrow{g} A \xleftarrow{f} B$.
Immediately from their constructions, comma $\infty$-categories are preserved by functors of $\infty$-cosmoi. For a certain special class of functors, which we will call \emph{weak equivalences of $\infty$-cosmoi}, equivalences and comma $\infty$-categories are also created and reflected. Having encoded limits, colimits, and adjunctions as equivalences between comma $\infty$-categories, a corollary will be that these notions are preserved, reflected, and created by weak equivalences of $\infty$-cosmoi. In particular, these notions are invariant under change of models between quasi-categories, complete Segal spaces, Segal categories, and naturally marked simplicial sets.
\subsection{The weak 2-universal property of comma \texorpdfstring{$\infty$}{infinity}-categories}
Proposition \ref{prop:cartesian-closure} demonstrates that the homotopy 2-category of an $\infty$-cosmos has finite products, in the 2-categorical sense. Unlike the special case of $\category{Cat}_2$, a general homotopy category will admit few 2-dimensional limit notions. However, certain simplicially-enriched limits in the $\infty$-cosmos provided by the axioms \ref{qcat.ctxt.cof.def}\ref{qcat.ctxt.cof:a} and \ref{qcat.ctxt.cof:b} descend to a particular variety of weak 2-dimensional limits. Arrow $\infty$-categories define a special case of a more general \emph{comma construction} that we now introduce.
\begin{defn}[comma $\infty$-categories]\label{defn:comma} Any pair of functors $f\colon B\to A$ and $g\colon C\to A$ in an $\infty$-cosmos $\mathcal{K}$ has an associated \emph{comma $\infty$-category}, constructed by the following pullback, formed in $\mathcal{K}$:
\begin{equation}\label{eq:comma-as-simp-pullback}
\xymatrix@=2.5em{
{f\mathbin{\downarrow} g}\save[]+<5pt,-5pt>*{\pbcorner}\restore \ar[r]\ar@{->>}[d]_{(p_1,p_0)} &
{A^{\bigger{1.12}{\mathbbe{2}}}} \ar@{->>}[d]^{(p_1,p_0)} \\
{C\times B} \ar[r]_-{g\times f} & {A\times A}
}
\end{equation}
The top horizontal functor represents a 2-cell
\[
\xymatrix@=10pt{
& f \downarrow g \ar@{->>}[dl]_{p_1} \ar@{->>}[dr]^{p_0} \ar@{}[dd]|(.4){\phi}|{\Leftarrow} \\
C \ar[dr]_g & & B \ar[dl]^f \\
& A}
\]
in the homotopy 2-category $\mathcal{K}_2$, called the \emph{comma cone}, which is defined by composing with \eqref{eq:generic-arrow}. Note that, by construction, the map $(p_1,p_0) \colon f \mathbin{\downarrow} g \twoheadrightarrow C \times B$ is an isofibration.
\end{defn}
As a simplicially-enriched limit in $\mathcal{K}$, the $\infty$-category $f \mathbin{\downarrow} g$ has a universal property expressed via a natural isomorphism of quasi-categories
\[ \fun(X, f\mathbin{\downarrow} g)\stackrel{\cong}{\longrightarrow} \fun(X,f) \mathbin{\downarrow} \fun(X,g)\] for any $X \in \mathcal{K}$, where the right-hand side is computed by the analogous pullback to \eqref{eq:comma-as-simp-pullback}, formed in $\category{qCat}$.
There is an analogous 2-categorical notion of comma object. The comma $\infty$-category $f \mathbin{\downarrow} g$ would be a strict comma object in the homotopy 2-category $\mathcal{K}_2$ if there were an isomorphism between the hom-category $\hom(X,f \mathbin{\downarrow} g)$ and the comma category $\hom(X,f) \mathbin{\downarrow} \hom(X,g)$ constructed from the pair of functors
\[ \hom(X,C) \xrightarrow{\hom(X,g)} \hom(X,A) \xleftarrow{\hom(X,f)} \hom(X,B).\] Instead:
\begin{prop}[{\extRef{found}{I}{prop:weakcomma}}]\label{prop:comma-smothering} For any object $X$, the induced comparison functor of hom-categories
\begin{equation*}
\hom(X,f\mathbin{\downarrow} g)\longrightarrow\hom(X,f)\mathbin{\downarrow}\hom(X,g)
\end{equation*}
is \emph{smothering}: surjective on objects, locally surjective on arrows, and conservative.
\end{prop}
\begin{obs}[the weak universal property of commas]\label{obs:comma-UP}
Explicitly, the weak universal property expressed by Proposition \ref{prop:comma-smothering} supplies three operations in the homotopy 2-category:
\begin{enumerate}[label=(\roman*)]
\item (1-cell induction) Given a 2-cell $\alpha \colon fb \Rightarrow gc$
\[ \vcenter{ \xymatrix@=10pt{
& X \ar[dl]_{c} \ar[dr]^{b} \ar@{}[dd]|(.4){\alpha}|{\Leftarrow} \\
C \ar[dr]_g & & B \ar[dl]^f \\
& A}} \quad =\quad \vcenter{ \xymatrix@=10pt{ & X \ar@{-->}[d]^{a} \ar@/^2ex/[ddr]^b \ar@/_2ex/[ddl]_c \\
& f \downarrow g \ar[dl]_{p_1} \ar[dr]^{p_0} \ar@{}[dd]|(.4){\phi}|{\Leftarrow} \\
C \ar[dr]_g & & B \ar[dl]^f \\
& A}}
\]
over the pair of functors $f$ and $g$, there exists a 1-cell $a\colon X\to f\mathbin{\downarrow} g$, defined by \emph{1-cell induction}, so that $p_0a = b$, $p_1a = c$, and $\alpha = \phi a$.
\item (2-cell induction) Given a pair of functors $a,a'\colon X\to f\mathbin{\downarrow} g$ and a pair of 2-cells
\[
\vcenter{\xymatrix@=10pt{
& {X}\ar[dl]_{a'}\ar[dr]^{a}
\ar@{}[dd]|(.4){\tau_1}|{\Leftarrow} & \\
{f\mathbin{\downarrow} g}\ar[dr]_{p_1} & &
{f\mathbin{\downarrow} g}\ar[dl]^{p_1} \\
& C &
}}
\mkern30mu\text{and}\mkern30mu
\vcenter{\xymatrix@=10pt{
& {X}\ar[dl]_{a'}\ar[dr]^{a}
\ar@{}[dd]|(.4){\tau_0}|{\Leftarrow} & \\
{f\mathbin{\downarrow} g}\ar[dr]_{p_0} & &
{f\mathbin{\downarrow} g}\ar[dl]^{p_0} \\
& B &
}}
\]
with the property that
\[
\xymatrix@=10pt{ & X \ar[dl]_{a'} \ar[dr]^a \ar@{}[dd]|(.4){\tau_1}|{\Leftarrow} & & & \ar@{}[dd]|{\displaystyle =} & & & X \ar[dl]_{a'} \ar[dr]^a \ar@{}[dd]|(.4){\tau_0}|{\Leftarrow} \\ f \downarrow g \ar[dr]_{p_1} & & f \downarrow g \ar[dl]|{p_1} \ar[dr]^{p_0} \ar@{}[dd]|(.4){\phi}|{\Leftarrow} & & & & f \downarrow g \ar[dl]_{p_1} \ar[dr]|{p_0} \ar@{}[dd]|(.4){\phi}|{\Leftarrow} & & f \downarrow g \ar[dl]^{p_0} \\ & C \ar[dr]_g & & B \ar[dl]^f & & C \ar[dr]_g & & B \ar[dl]^f & & \\ & & A & & & & A}
\]
then there exists a 2-cell $\tau \colon a \Rightarrow a'$, defined by \emph{2-cell induction}, satisfying the equalities
\[ \vcenter{\xymatrix@=10pt{
& {X}\ar[dl]_{a'}\ar[dr]^{a}
\ar@{}[dd]|(.4){\tau_1}|{\Leftarrow} & \\
{f\mathbin{\downarrow} g}\ar[dr]_{p_1} & &
{f\mathbin{\downarrow} g}\ar[dl]^{p_1} \\
& C &
}} \mkern10mu = \mkern10mu
\vcenter{\xymatrix@=30pt{ X \ar@/^3ex/[d]^a \ar@/_3ex/[d]_{a'} \ar@{}[d]|(.4){\tau}|{\Leftarrow} \\ f \downarrow g \ar[d]^{p_1} \\ C}}.
\mkern20mu\text{and}\mkern20mu
\vcenter{\xymatrix@=10pt{
& {X}\ar[dl]_{a'}\ar[dr]^{a}
\ar@{}[dd]|(.4){\tau_0}|{\Leftarrow} & \\
{f\mathbin{\downarrow} g}\ar[dr]_{p_0} & &
{f\mathbin{\downarrow} g}\ar[dl]^{p_0} \\
& B &
}} \mkern10mu = \mkern10mu
\vcenter{\xymatrix@=30pt{ X \ar@/^3ex/[d]^a \ar@/_3ex/[d]_{a'} \ar@{}[d]|(.4){\tau}|{\Leftarrow} \\ f \downarrow g \ar[d]^{p_0} \\ B}}
\]
\item (conservativity) Any 2-cell $\xymatrix{ X \ar@/^2ex/[r]^{a} \ar@/_2ex/[r]_{a'} \ar@{}[r]|-{\Downarrow\tau} & { f \mathbin{\downarrow} g}}$ with the property that the whiskered 2-cells $p_0\tau$ and $p_1\tau$ are both isomorphisms is also an isomorphism.
\end{enumerate}
\end{obs}
\begin{lem}[{\extRef{found}{I}{lem:1cell-ind-uniqueness}}]\label{lem:ess.unique.1-cell.ind}
A parallel pair of functors over $C \times B$
\[ \xymatrix@=1.5em{ X \ar@<.5ex>[rr]^a \ar@<-.5ex>[rr]_{a'} \ar[dr]_{(c,b)} && f \mathbin{\downarrow} g \ar@{->>}[dl]^{(p_1,p_0)} \\ & C \times B}\]
are isomorphic over $C \times B$ if and only if $a$ and $a'$ both enjoy the same defining properties as 1-cells induced by the weak 2-universal property of $f\mathbin{\downarrow} g$, i.e., if these functors satisfy $\phi a = \phi a'$.\footnote{For $a$ and $a'$ to define a parallel pair of functors over $C \times B$ we must have $p_0a =p_0a'$ and $p_1 a = p_1a'$.} That is, 2-cells of the form displayed on the left
\[ \vcenter{\xymatrix@=10pt{
& X \ar[dl]_{c} \ar[dr]^{b} \ar@{}[dd]|(.4){\alpha}|{\Leftarrow} \\
C \ar[dr]_g & & B \ar[dl]^f \\
& A}} \qquad \leftrightsquigarrow \qquad \vcenter{ \xymatrix@=10pt{ & X \ar@{->}[dd]^{a} \ar[dr]^b \ar[dl]_c \\ C & & B \\ & f \mathbin{\downarrow} g \ar@{->>}[ul]^{p_1} \ar@{->>}[ur]_{p_0}}}\]
stand in bijection with isomorphism classes of maps of spans, as displayed on the right.
\end{lem}
\begin{proof} An isomorphism $\tau$ between a parallel pair $a,a' \colon X \to f \mathbin{\downarrow} g$ with the same defining conditions is induced by 2-cell induction and conservativity from the pair of identity 2-cells $p_1\tau = \id_c$ and $p_1 \tau = \id_b$. Conversely, by uniqueness of composites of pasting diagrams, if there is any isomorphism $\tau \colon a \cong a'$ over an identity on $C \times B$ then whiskered composites $\phi a$ and $\phi a'$ are equal.
\end{proof}
\begin{lem}[{\extRef{equipment}{V}{lem:equiv-to-comma}}]\label{lem:equiv-to-comma}
For any pair of functors $C \xrightarrow{g} A \xleftarrow{f} B$, the universal property of \ref{obs:comma-UP} characterizes a unique equivalence class of isofibrations $E \twoheadrightarrow C \times B$. That is:
\begin{enumerate}[label=(\roman*)]
\item\label{itm:equiv-to-comma-i} Any pair of isofibrations $E \twoheadrightarrow C \times B$ and $E' \twoheadrightarrow C \times B$ equipped with a comma cone that enjoys the weak universal property of \ref{obs:comma-UP} are equivalent, over $C \times B$.
\item\label{itm:equiv-to-comma-ii} Any isofibration $E \twoheadrightarrow C \times B$ that is equivalent over $C\times B$ to the comma $f \mathbin{\downarrow} g \twoheadrightarrow C \times B$, has the universal property of \ref{obs:comma-UP}.
\end{enumerate}
\end{lem}
\begin{proof}
The proof of \ref{itm:equiv-to-comma-i} is an elementary exercise in the application of the universal property of \ref{obs:comma-UP}, paralleling the standard proof that a strictly-defined limit is unique up to isomorphism. For \ref{itm:equiv-to-comma-ii}, an equivalence $E \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} f\mathbin{\downarrow} g$ defines a representable equivalence in the homotopy 2-category. Thus, for any object $X$, we have a composable pair of functors
\begin{equation}\label{eq:equiv-to-comma-smothering} \hom(X,E) \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} \hom(X,f \mathbin{\downarrow} g) \longrightarrow \hom(X,f) \mathbin{\downarrow} \hom(X,g),\end{equation} the first being an equivalence and the second being smothering, and so the composite is full, conservative, and essentially surjective on objects. To show that it is in fact surjective on objects, we make use of the fact that the equivalence $E \simeq f \mathbin{\downarrow} g$ is fibered over $C \times B$. Any object in the codomain of \eqref{eq:equiv-to-comma-smothering} lifts to a representing functor $X \to f \mathbin{\downarrow} g$, which is isomorphic over $C \times B$ to a functor of the form $X \xrightarrow{e} E \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} f \mathbin{\downarrow} g$. By Lemma \ref{lem:ess.unique.1-cell.ind}, $e$ defines the desired preimage in $\hom(X,E)$.
\end{proof}
\subsection{Hom-spaces and groupoidal objects}
The comma construction allows us to define hom-spaces between a pair of elements in an $\infty$-category.
\begin{defn}[hom-spaces]\label{defn:hom-space} Given a pair of elements $a,a' \colon 1\to A$ in an $\infty$-category $A$, their \emph{hom-space} is the comma $\infty$-category $a \mathbin{\downarrow} a'$ defined by the pullback
\[ \xymatrix{ a \mathbin{\downarrow} a' \ar@{->>}[d] \ar[r] \save[]+<5pt,-5pt>*{\pbcorner}\restore & A^{\bigger{1.12}{\mathbbe{2}}} \ar@{->>}[d]^{(p_1,p_0)} \\ 1 \ar[r]_-{(a',a)} & A \times A}\]
\end{defn}
We refer to the $\infty$-category $a \mathbin{\downarrow} a'$ as a hom-space because, as an easy corollary of the conservativity of its weak universal property, it defines a \emph{groupoidal object} in the $\infty$-cosmos.
\begin{defn}\label{defn:groupoidal-object} We say an object $E$ in an $\infty$-cosmos $\mathcal{K}$ is \emph{groupoidal} if the following equivalent conditions are satisfied:
\begin{enumerate}[label=(\roman*)]
\item\label{itm:groupoidal-object:i} $E$ is a groupoidal object in the homotopy 2-category $\mathcal{K}_2$, that is, every 2-cell with codomain $E$ is invertible.
\item\label{itm:groupoidal-object:ii} For each $X \in \mathcal{K}$, the hom-category $\hom(X,E)$ is a groupoid.
\item\label{itm:groupoidal-object:iii} For each $X \in \mathcal{K}$, the mapping quasi-category $\fun(X,E)$ is a Kan complex.
\item\label{itm:groupoidal-object:iv} The isofibration $E^{\mathbb{I}} \twoheadrightarrow E^{\bigger{1.12}{\mathbbe{2}}}$, induced by the inclusion of simplicial sets ${\bigger{1.12}{\mathbbe{2}}}\hookrightarrow{\mathbb{I}}$, is a trivial fibration.
\end{enumerate}
Here \ref{itm:groupoidal-object:ii} is an unpacking of \ref{itm:groupoidal-object:i}. The equivalence of \ref{itm:groupoidal-object:ii} and \ref{itm:groupoidal-object:iii} is a well-known result of Joyal \cite[1.4]{Joyal:2002:QuasiCategories}. Condition \ref{itm:groupoidal-object:iv} is equivalent to the assertion that $\fun(X,E)^{\mathbb{I}} \twoheadrightarrow \fun(X,E)^{\bigger{1.12}{\mathbbe{2}}}$ is a trivial fibration between quasi-categories for all $X$. If this is a trivial fibration, then surjectivity on vertices implies that every 1-simplex in $\fun(X,E)$ is an isomorphism, proving \ref{itm:groupoidal-object:iii}. As ${\bigger{1.12}{\mathbbe{2}}} \hookrightarrow{\mathbb{I}}$ is a weak homotopy equivalence, \ref{itm:groupoidal-object:iii} implies \ref{itm:groupoidal-object:iv}.
\end{defn}
\begin{rmk}
In the $\infty$-cosmoi whose objects model $(\infty,1)$-categories, we posit that the groupoidal objects are precisely the corresponding $\infty$-\emph{groupoids}. For instance, in the $\infty$-cosmos for quasi-categories, an object is groupoidal if and only if it is a Kan complex. In the $\infty$-cosmos for naturally marked simplicial sets, an object is groupoidal if and only if it is a Kan complex with every edge marked.
\end{rmk}
\subsection{Commas representing functors}
For any functors between $\infty$-categories, e.g., $f \colon B \to A$ and $u \colon A \to B$, the following pullbacks in $\mathcal{K}$
\[ \xymatrix@=1.5em{ f \mathbin{\downarrow} A \ar[d]_{(p_1,p_0)} \ar[r] \save[]+<5pt,-5pt>*{\pbcorner}\restore & A^{\bigger{1.12}{\mathbbe{2}}} \ar[d] & & B \mathbin{\downarrow} u \ar[d]_{(q_1,q_0)} \ar[r] \save[]+<5pt,-5pt>*{\pbcorner}\restore & B^{\bigger{1.12}{\mathbbe{2}}} \ar[d] \\ A \times B \ar[r]_{\id_A \times f} & A \times A & & A \times B \ar[r]_{u \times \id_B} & B \times B}\]
define the comma $\infty$-categories $f \mathbin{\downarrow} A$ and $B \mathbin{\downarrow} u$, each of which is equipped with an isofibration to $A \times B$. The horizontal functors represent 2-cells
\begin{equation}\label{eq:adj-comma-cone}
\xymatrix@=15pt{ & f \mathbin{\downarrow} A \ar@{->>}[dl]_{p_1} \ar@{->>}[dr]^{p_0} \ar@{}[d]|(.6){\Leftarrow\alpha} & && & B \mathbin{\downarrow} u \ar@{->>}[dl]_{q_1} \ar@{->>}[dr]^{q_0} \ar@{}[d]|(.6){\Leftarrow\beta} \\ A & & B \ar[ll]^f && B \ar[rr]_u & & A}
\end{equation}
in the homotopy 2-category which satisfy the following weak universal property derived from Lemma \ref{lem:ess.unique.1-cell.ind}:
\begin{itemize}
\item 2-cells $ \chi \colon fb \Rightarrow a$ stand in bijection with isomorphism classes of maps of spans:
\[\vcenter{\xymatrix@=15pt{
& X \ar[dl]_{a} \ar[dr]^{b} \ar@{}[d]|{\chi}|(.7){\Leftarrow} \\
A & & B \ar[ll]^f }} \qquad \leftrightsquigarrow \qquad \vcenter{\xymatrix@=10pt{ & X \ar@{->}[dd]^{x} \ar[dr]^b \ar[dl]_a \\ A & & B \\ & f \mathbin{\downarrow} A \ar@{->>}[ul]^{p_1} \ar@{->>}[ur]_{p_0}}}\]
\item 2-cells $\zeta \colon b \Rightarrow ua$ stand in bijection with isomorphism classes of maps of spans:
\[\vcenter{\xymatrix@=15pt{
& X \ar[dl]_{a} \ar[dr]^{b} \ar@{}[d]|{\zeta}|(.7){\Leftarrow} \\
A \ar[rr]_u & & B }}\qquad \leftrightsquigarrow \qquad \vcenter{\xymatrix@=10pt{ & X \ar@{->}[dd]^{y} \ar[dr]^b \ar[dl]_a \\ A & & B \\ & B \mathbin{\downarrow} u \ar@{->>}[ul]^{q_1} \ar@{->>}[ur]_{q_0}}}\]
\end{itemize}
Here the isomorphism classes are defined with respect to natural isomorphisms $x \cong x'$ projecting to an identity over $A \times B$. The bijection is implemented by whiskering a functor over $A \times B$ with the 2-cells of \eqref{eq:adj-comma-cone}.
If $f \dashv u$, then $\hom(X,f) \dashv \hom(X,u)$ defines an adjunction of categories for any $\infty$-category $X$. This tells us that there is a natural bijection between 2-cells
\[\vcenter{\xymatrix@=15pt{
& X \ar[dl]_{a} \ar[dr]^{b} \ar@{}[d]|{\chi}|(.7){\Leftarrow} \\
A & & B \ar[ll]^f }}\qquad \leftrightsquigarrow\qquad \vcenter{\xymatrix@=15pt{
& X \ar[dl]_{a} \ar[dr]^{b} \ar@{}[d]|{\zeta}|(.7){\Leftarrow} \\
A \ar[rr]_u & & B }}\] implemented by pasting along the bottom with the unit or with the counit. By a Yoneda-style argument, this yields the following result, which we instead prove directly.
\begin{prop}[{\extRef{found}{I}{prop:adjointequiv}}]\label{prop:adjoint-comma-equiv-I} If $\xymatrix{ B \ar@<1ex>[r]^f \ar@{}[r]|\perp & A \ar@<1ex>[l]^u}$ is an adjunction between $\infty$-categories, then there is a fibered equivalence between the comma $\infty$-categories:
\begin{equation*}
\xymatrix@=1em{
{f\mathbin{\downarrow} A}\ar@{->>}[dr]_(0.4){(p_1,p_0)}\ar[rr]^{w}_{\sim}
&& {B\mathbin{\downarrow} u}\ar@{->>}[dl]^(0.4){(q_1,q_0)} \\
& {A\times B}&
}
\end{equation*}
\end{prop}
The proof of this result mirrors the standard construction of the adjoint-transpose bijection using the unit and counit of an adjunction.
\begin{proof}
The composite 2-cells displayed on the left of the pasting equalities induce functors $w'\colon B \mathbin{\downarrow} u \to f \mathbin{\downarrow} A$ and $w\colon f \mathbin{\downarrow} A \to B \mathbin{\downarrow} u$
\begin{equation*}\xymatrix@C=10pt{ & B \mathbin{\downarrow} u \ar[dl]_{q_1} \ar[dr]^{q_0} \ar@{}[d]|(.6){\Leftarrow\beta} & & & B \mathbin{\downarrow} u \ar[d]^{w'} & && & f \mathbin{\downarrow} A \ar[dl]_{p_1} \ar[dr]^{p_0} \ar@{}[d]|(.6){\Leftarrow\alpha} & & & f \mathbin{\downarrow} A \ar[d]^{w} \\ A \ar@{=}[dr] \ar[rr]^u & \ar@{}[d]|{\Leftarrow\epsilon} & B \ar[dl]^f & = & f \mathbin{\downarrow} A \ar[dl]_{p_1} \ar[dr]^{p_0} \ar@{}[d]|(.6){\Leftarrow\alpha} & & & A \ar[dr]_u & \ar@{}[d]|{\Leftarrow\eta} & B \ar@{=}[dl] \ar[ll]_f & = & B \mathbin{\downarrow} u \ar[dl]_{q_1} \ar[dr]^{q_0} \ar@{}[d]|(.6){\Leftarrow\beta} \\ & A & & A & & B \ar[ll]^f & & & B & & A \ar[rr]_u & & B}\end{equation*}
that commute with the projections to $A \times B$.
\begin{equation*}
\xymatrix@=1em{ f\mathbin{\downarrow} A \ar@{->>}[dr]_{(p_1,p_0)} \ar@/^1ex/[rr]^{w} & & B \mathbin{\downarrow} u \ar@{->>}[dl]^{(q_1,q_0)} \ar@/^1ex/[ll]^{w'} \\ & A \times B}
\end{equation*}
Supposing $f \dashv u$, we have the following series of pasting equalities:
\begin{equation*}
\xymatrix@C=10pt{ & f \mathbin{\downarrow} A \ar[d]^{w} & & & f \mathbin{\downarrow} A \ar[d]^{w} & & & f \mathbin{\downarrow} A \ar[dl]_{p_1} \ar[dr]^{p_0}\ar@{}[d]|(.6){\Leftarrow\alpha} & \ar@{}[dr]|*+{=} & & f \mathbin{\downarrow} A \ar[dl]_{p_1} \ar[dr]^{p_0} \ar@{}[d]|(.6){\Leftarrow\alpha} \\ & B \mathbin{\downarrow} u \ar[d]^{w'} & \ar@{}[d]|*+{=} & & B \mathbin{\downarrow} u \ar[dl]_{q_1} \ar[dr]^{q_0} \ar@{}[d]|(.6){\Leftarrow\beta} & {=} & A \ar[drr]|u \ar@{=}[d] &\ar@{}[dr]|(.4){\Leftarrow\eta} & B \ar[ll]_f \ar@{=}[d] & A & & B \ar[ll]^f \\ & f \mathbin{\downarrow} A \ar[dl]_{p_1} \ar[dr]^{p_0} \ar@{}[d]|(.6){\Leftarrow\alpha} & & A \ar@{=}[dr] \ar[rr]^u & \ar@{}[d]|(0.4){\Leftarrow\epsilon} & B \ar[dl]^f & A \ar@{}[ur]|(0.4){\Leftarrow\epsilon} & & B \ar[ll]^f \\ A & & B \ar[ll]^f & & A & & & & }
\end{equation*}
in which the last step is an application of one of the triangle identities. This tells us that the endo-1-cells $w'w$ and $\id_{f\mathbin{\downarrow} A}$ on the object $(p_1,p_0)\colon f\mathbin{\downarrow} A\twoheadrightarrow A\times B$ both map to the same 2-cell $\alpha$ under the whiskering operation. By Lemma \ref{lem:ess.unique.1-cell.ind}, it follows that these are connected via a natural isomorphism $w'w$ and $\id_{f\mathbin{\downarrow} A}$ over $A \times B$. A dual argument provides a natural isomorphism $ww' \cong \id_{B \mathbin{\downarrow} f}$ over $A \times B$, defining a fibered equivalence $f \mathbin{\downarrow} A \simeq B \mathbin{\downarrow} u$ over $A\times B$.
\end{proof}
\begin{cor} If $f \colon B \to A$ and $u \colon A \to B$ are functors so that $f \dashv u$, then for any pair of elements $a \colon 1 \to A$ and $b \colon 1 \to B$, the hom-spaces $fa \mathbin{\downarrow} b$ and $a \mathbin{\downarrow} ub$ are equivalent.
\end{cor}
\begin{proof}
Fibered equivalences can be pulled back along any functor to define another fibered equivalence. Pulling back the equivalence of Proposition \ref{prop:adjoint-comma-equiv-I} along a pair of elements $1 \xrightarrow{(a,b)} A \times B$
\[ \begin{xy}
0;<5pc,0pc>:
*{\xybox{
\POS(0,0)*+{1}="A",
\POS(1.3,0)*+{A \times B}="B",
\POS(-0.3,1.3)*+{fb \mathbin{\downarrow} a}="F",
\POS(-0.25,1.25)*+{\pbcorner},
\POS(0.45,0.65)*+{\pbcorner},
\POS(0.4,0.7)*+{a \mathbin{\downarrow} ub}="F'",
\POS(1,1.3)*+{f \mathbin{\downarrow} A}="E",
\POS(1.6,0.7)*+{B \mathbin{\downarrow} u}="E'",
\ar"A";"B"_-{(a,b)}
\ar@{->>}@/_0.2pc/"F";"A"
\ar@{->>}@/^0.2pc/"F'";"A"
\ar@{->>}@/_0.2pc/"E";"B"_(0.7){(p_1,p_0)}|!{"F'";"E'"}\hole
\ar@{->>}@/^0.2pc/"E'";"B"^(0.4){(p_1,p_0)}
\ar"F";"E"
\ar"F'";"E'"
\ar"E";"E'"^*{\rotatebox{145}{$\labelstyle\sim$}}
\ar"F";"F'"^*{\rotatebox{145}{$\labelstyle\sim$}}
}}
\end{xy}\]
we obtain an equivalence of hom-spaces $fb \mathbin{\downarrow} a \simeq b \mathbin{\downarrow} ua$, the former in the $\infty$-category $A$ and the latter in the $\infty$-category $B$.
\end{proof}
The converse to Proposition \ref{prop:adjoint-comma-equiv-I} follows from a general result that will have other applications.
\subsection{Commas and absolute lifting diagrams}
A 2-cell $\lambda \colon f \ell \Rightarrow g$ induces a functor
\begin{equation}\label{eq:abs-R-lifting} \vcenter{\xymatrix{ \ar@{}[dr]|(.7){\Downarrow\lambda} & B \ar[d]^f \\ C \ar[ur]^\ell \ar[r]_g & A}} \qquad \rightsquigarrow\qquad \vcenter{\xymatrix@=1em{ B \mathbin{\downarrow} \ell \ar[rr]^w \ar@{->>}[dr]_(.4){(p_1,p_0)} & & f \mathbin{\downarrow} g\ar@{->>}[dl]^(.4){(p_1,p_0)} \\ & C \times B}}\end{equation} between comma $\infty$-categories defined by 1-cell induction for the comma $f \mathbin{\downarrow} g$ from the pasted composite of the comma cone for $B \mathbin{\downarrow} \ell$ and $\lambda$.
\[ \vcenter{\xymatrix@=1.2em{
& {B\mathbin{\downarrow}\ell}\ar[dl]_{p_1}\ar[dr]^{p_0} & \\
{C} \ar[dr]_{g}\ar[rr]|*+{\scriptstyle\ell} && {B}\ar[dl]^{f} \\
& A &
\ar@{} "1,2";"3,2" |(0.3){\Leftarrow\phi} |(0.7){\Leftarrow\lambda}
}}
\mkern 20mu = \mkern20mu
\vcenter{\xymatrix@=1.2em{
& {B\mathbin{\downarrow}\ell}\ar[d]^{w}\ar@/_1.5ex/[ddl]_{p_1}\ar@/^1.5ex/[ddr]^{p_0} & \\
& {f\mathbin{\downarrow} g}\ar[dl]^{p_1}\ar[dr]_{p_0} & \\
{C}\ar[dr]_{g} & & {B}\ar[dl]^{f} \\
& {A} &
\ar@{} "2,2";"4,2" |{\Leftarrow\phi}
}}\]
A Yoneda-style argument, making use of Lemmas \ref{lem:ess.unique.1-cell.ind} and \ref{lem:equiv-to-comma}, proves the following result:
\begin{prop}[{\extRef{found}{I}{prop:absliftingtranslation}, \extRef{found}{I}{prop:absliftingtranslation2}}]\label{prop:abs-lifting-via-commas} The data of \eqref{eq:abs-R-lifting} defines an absolute right lifting diagram in $\mathcal{K}_2$ if and only if the induced map $w\colon B \mathbin{\downarrow} \ell \to f \mathbin{\downarrow} g$ is an equivalence over $C \times B$. Conversely, an equivalence \[ \xymatrix@=1em{ B \mathbin{\downarrow} \ell \ar[rr]^w_\sim \ar@{->>}[dr]_(.4){(p_1,p_0)} & & f \mathbin{\downarrow} g\ar@{->>}[dl]^(.4){(p_1,p_0)} \\ & C \times B}\] over $C \times B$ induces a canonical 2-cell $\lambda \colon f\ell \Rightarrow g$ that defines an absolute right lifting of $g$ through $f$.
\end{prop}
\subsection{Adjunctions, limits, and colimits via commas}
Special cases of Proposition \ref{prop:abs-lifting-via-commas} provide characterizations of adjunctions and (co)limits as equivalences between comma $\infty$-categories.
\begin{prop}[{\extRef{found}{I}{prop:adjointequivconverse}}]\label{prop:adjoint-comma-equiv-II} If $f \colon B \to A$ and $u \colon A \to B$ are functors so that there exists a fibered equivalence between the comma $\infty$-categories:
\begin{equation*}
\xymatrix@=1em{
{f\mathbin{\downarrow} A}\ar@{->>}[dr]_(0.4){(p_1,p_0)}\ar[rr]^{w}_{\sim}
&& {B\mathbin{\downarrow} u}\ar@{->>}[dl]^(0.4){(q_1,q_0)} \\
& {A\times B}&
}
\end{equation*} then $f \dashv u$.
\end{prop}
\begin{proof}
By Proposition \ref{prop:abs-lifting-via-commas}, the fibered equivalence induces an absolute right lifting diagram \[ \xymatrix{ \ar@{}[dr]|(.7){\Downarrow\epsilon} & B \ar[d]^f \\ A \ar[ur]^u \ar@{=}[r] & A}\] Exercise \ref{exs:adj-as-abs-lifting} then implies that $f \dashv u$, with $\epsilon \colon fu \Rightarrow \id_A$ as the counit.
\end{proof}
\begin{defn}[the $\infty$-category of cones] Given a $J$-indexed diagram $d \colon 1 \to A^J$ in an $\infty$-category $A$, the \emph{$\infty$-category of cones over $d$} is the comma $\infty$-category $\Delta \mathbin{\downarrow} d$ formed over the cospan
\[ \xymatrix@=10pt{
&\Delta \mathbin{\downarrow} d \ar@{->>}[dl]_{p_1} \ar@{->>}[dr]^{p_0} \ar@{}[dd]|(.4){\phi}|{\Leftarrow} \\
1 \ar[dr]_d & & A \ar[dl]^\Delta \\
& A^J}\]
By the defining simplicial pullback \eqref{eq:comma-as-simp-pullback}, the data of an element in $\Delta \mathbin{\downarrow} d$ is comprised of an element $a \colon 1 \to A$ (the summit) together with an element of the hom-space from $\Delta a$ to $d$ in $A^J$ (the cone).
\end{defn}
Specializing Proposition \ref{prop:abs-lifting-via-commas}, we have:
\begin{prop}\label{prop:limit-comma-equiv} An element $\ell \colon 1 \to A$ defines a limit for a diagram $d \colon 1 \to A^J$ if and only if there is a fibered equivalence between the comma $\infty$-category represented by $\ell$ and the $\infty$-category of cones over $d$:
\begin{equation*}
\xymatrix@=1em{
{A \mathbin{\downarrow} \ell }\ar@{->>}[dr]_{p_0}\ar[rr]^{\sim}
&& {\Delta \mathbin{\downarrow} d}\ar@{->>}[dl]^{p_0} \\
& {A}&
}
\end{equation*}
\end{prop}
The conclusion of Proposition \ref{prop:limit-comma-equiv} asserts that the $\infty$-category of cones over $d$ is \emph{represented by} the element $\ell \colon 1 \to A$.
\begin{exs} Specializing to the case $J= \emptyset$, show that an element $t \colon 1 \to A$ is terminal if and only if the projection $p_0 \colon A \mathbin{\downarrow} t \stackrel{\smash{\mkern-2mu\mathlower{1.5}{\sim}}}\longtwoheadrightarrow A$ is a equivalence, and thus a trivial fibration.
\end{exs}
\begin{exs}\label{exs:terminal-in-representable} Use 1-cell induction, 2-cell induction, and 2-cell conservativity for the comma $A \mathbin{\downarrow} \ell$ associated to an element $\ell \colon 1 \to A$ to show that the identity at $\ell$ defines a terminal element ${\id_\ell} \colon 1 \to A \mathbin{\downarrow} \ell$, in the sense of Definition \ref{defn:terminal-element}.
\end{exs}
A fibered version of Exercise \ref{exs:terminal-in-representable} (\extRef{found}{I}{prop:right.liftings.as.fibred.terminal.objects}) proves the following:
\begin{prop}[{\extRef{found}{I}{prop:limits.as.terminal.objects}}]\label{prop:limits.as.terminal.objects} A limit of a diagram $d \colon 1 \to A^J$ defines a terminal element in the $\infty$-category $\Delta \mathbin{\downarrow} d \twoheadrightarrow A$ of cones over $d$. Conversely, a terminal element in the $\infty$-category of cones defines a limit for $d$.
\end{prop}
\begin{rmk} The $\infty$-category of \emph{cones in $A$ over any $J$-indexed diagram} is the comma $\Delta \mathbin{\downarrow} A^J \twoheadrightarrow A^J \times A$. Pulling back along an element $d \colon 1 \to A^J$ defines the $\infty$-category $\Delta \mathbin{\downarrow} d$ of cones over $d$. The defining simplicial pullback \eqref{eq:comma-as-simp-pullback} for $\Delta \mathbin{\downarrow} A^J$ reveals that it is isomorphic to the simplicial cotensor $A^{\del@fn\parsedel@^0 \diamond J}$, where ``$\diamond$'' is Joyal's ``fat join'' construction. For any pair of simplicial sets $I$ and $J$, there is a weak equivalence $I \diamond J \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} I \star J$ in the Joyal model structure under the disjoint union $I \coprod J$ from the \emph{fat join} $I \diamond J$ to the \emph{join} $I \star J$. Taking cotensors, this induces a fibered equivalence $A^{\del@fn\parsedel@^0 \star J} \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} \Delta \mathbin{\downarrow} A^J$ over $A^J \times A$, which pulls back to define an equivalence $\slicer{A}{d} \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} \Delta \mathbin{\downarrow} d$ between Joyal's \emph{slice $\infty$-category} and the $\infty$-category of cones over $d$; see \S\extRef{found}{I}{subsec:join}. This is the geometrical basis for the proof that characterization of the limit of a diagram valued in a quasi-category given in Definition \ref{defn:limit} and re-expressed by Proposition \ref{prop:limits.as.terminal.objects} agrees with the Joyal's original definition (\extRef{found}{I}{prop:limits.are.limits}).
\end{rmk}
\subsection{Model independence of basic \texorpdfstring{$\infty$}{infinity}-category theory}
We have seen that comma $\infty$-categories can be used to encode various universal properties including:
\begin{itemize}
\item the existence of an adjunction between a pair of functors $u \colon A \to B$ and $f \colon B \to A$
\item the property that an element $\ell \colon 1 \to A$ defines a limit for a diagram $d \colon 1 \to A^J$
\end{itemize}
We will now see that any categorical property that can be captured by the existence of a fibered equivalence between comma $\infty$-categories is ``model independent'' in the sense that it is preserved by any functor of $\infty$-cosmoi and reflected by those functors that define \emph{weak equivalences of $\infty$-cosmoi}.
Recall, a \emph{functor of $\infty$-cosmoi} $F \colon \mathcal{K} \to \mathcal{L}$ is a simplicial functor that preserves the limits listed in~\ref{qcat.ctxt.cof.def}\ref{qcat.ctxt.cof:a} and the class of isofibrations, and hence also the classes of equivalences and trivial fibrations. A functor $F \colon \mathcal{K} \to \mathcal{L}$ of $\infty$-cosmoi induces a 2-functor $F_2 \mathrel{:=} h_*F \colon \mathcal{K}_2 \to \mathcal{L}_2$ between their homotopy 2-categories.
\begin{prop}[{\extRef{equipment}{V}{prop:induced-2-functor}}]\label{prop:induced-2-functor} A functor $F \colon \mathcal{K} \to \mathcal{L}$ of $\infty$-cosmoi induces a 2-functor $F_2 \colon \mathcal{K}_2 \to \mathcal{L}_2$ between their homotopy 2-categories that preserves adjunctions, equivalences, isofibrations, trivial fibrations, groupoidal objects, products, and comma objects.
\end{prop}
\begin{proof}
Any 2-functor preserves adjunctions and equivalences. Preservation of isofibrations and products are direct consequences of the hypotheses in Definition \ref{defn:qcat-ctxt-functor}; recall that the class of trivial fibrations in this intersection of the classes of isofibrations and equivalences. Preservation of groupoidal objects is a consequence of the characterization \ref{defn:groupoidal-object}\ref{itm:groupoidal-object:iv}. Preservation of commas follows from the construction of \eqref{eq:comma-as-simp-pullback}, which is preserved by a functor of $\infty$-cosmoi, and Lemma \ref{lem:equiv-to-comma}\ref{itm:equiv-to-comma-i}, which says that all commas are equivalent an $\infty$-category constructed by the simplicial pullback formula.
\end{proof}
\begin{defn}[weak equivalences of $\infty$-cosmoi] A functor $F \colon \mathcal{K} \to \mathcal{L}$ of $\infty$-cosmoi is a \emph{weak equivalence} when it is:
\begin{enumerate}[label=(\alph*)]
\item surjective on objects up to equivalence: i.e., if for every $X \in \mathcal{L}$, there is some $A \in \mathcal{K}$ so that $FA\simeq X \in \mathcal{L}$.
\item a local equivalence of quasi-categories: i.e., if for every pair $A,B \in \mathcal{K}$, the map $\fun(A,B) \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} \fun(FA,FB)$ is an equivalence of quasi-categories.
\end{enumerate}
\end{defn}
\begin{ex}
The following define weak equivalences of $\infty$-cosmoi:
\begin{itemize}
\item The underlying quasi-category functor $\fun(1,-)\colon \category{CSS} \to \category{qCat}$ that takes a complete Segal space to its $0\th$ row (see example \extRef{yoneda}{IV}{ex:CSS-cosmos}).
\item The functor $t^! \colon \category{qCat} \to \category{CSS}$ defined in example \extRef{yoneda}{IV}{ex:other-CSS-functor}.
\item The underlying quasi-category functor $\fun(1,-)\colon \category{Segal} \to \category{qCat}$ that takes a Segal category to its $0\th$ row (see example \extRef{yoneda}{IV}{ex:segal-cosmos}).
\item The functor $d_* \colon \category{qCat} \to \category{Segal}$ defined by Joyal and Tierney \cite{Joyal:2007hb}.
\item The underlying quasi-category functor $\fun(1,-) \colon \category{sSet}_+\to \category{qCat}$ that carries a naturally marked simplicial set to its underlying quasi-category (see example \extRef{yoneda}{IV}{ex:marked-cosmos}).
\item The functor $(-)^\natural \colon \category{qCat} \to \category{sSet}_+$ that gives a quasi-category its ``natural'' marking.
\item The functor $\category{CSS} \to \category{Segal}$ that ``discretizes'' the 0th space of a complete Segal space. This commutes with the underlying quasi-category functors
\end{itemize}
\end{ex}
\begin{prop}\label{prop:we} If $F$ is a weak equivalence of $\infty$-cosmoi, then the induced 2-functor $F_2 \colon \mathcal{K}_2 \to \mathcal{L}_2$
\begin{enumerate}[label=(\roman*)]
\item\label{itm:we-bi-equiv} defines a biequivalence $F_2 \colon \mathcal{K}_2 \to \mathcal{L}_2$: i.e., the 2-functor $F_2$ is surjective on objects up to equivalence and defines a local equivalence of categories $\hom(A,B) \stackrel{\simeq}{\longrightarrow}\hom(FA,FB)$ for all $A,B \in \mathcal{K}$.
\item\label{itm:we-iso} induces a bijection between isomorphism classes of parallel functors: for all $A, B \in \mathcal{K}$, the functor $\hom(A,B) \stackrel{\simeq}{\longrightarrow} \hom(FA,FB)$ induces a bijection on isomorphism classes of objects.
\item\label{itm:we-groupoidal} preserves and reflects groupoidal objects: $A \in \mathcal{K}$ is groupoidal if and only if $FA \in \mathcal{L}$ is groupoidal.
\item\label{itm:we-equiv} preserves and reflects equivalence: $A \simeq B \in \mathcal{K}$ if and only if $FA \simeq FB \in \mathcal{L}$.
\item\label{itm:we-equivs} preserves and reflects equivalences: $f \colon A \to B \in \mathcal{K}$ is an equivalence if and only if $Ff \colon FA \to FB \in \mathcal{L}$ is an equivalence.
\item\label{itm:we-comma} preserves and reflects comma objects: given $E \twoheadrightarrow C \times B$ and $C \xrightarrow{g} A \xleftarrow{f} B$ in $\mathcal{K}$, then $E \simeq f \mathbin{\downarrow} g$ over $C \times B$ if and only if $FE \simeq Ff \mathbin{\downarrow} Fg \cong F(f \mathbin{\downarrow} g)$ over $FC \times FB$.
\end{enumerate}
\end{prop}
\begin{proof}
The homotopy category functor $h \colon \category{qCat} \to \category{Cat}$ carries equivalences of quasi-categories to equivalences of categories; thus, the local equivalence of mapping quasi-categories $\fun(A,B) \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} \fun(FA,FB)$ descends to an equivalence of hom-categories $\hom(A,B) \to \hom(FA,FB)$, proving \ref{itm:we-bi-equiv}. Any equivalence of categories induces a bijection between isomorphism classes of objects, proving \ref{itm:we-iso}. A category is a groupoid if and only if it is equivalent to a groupoid, so \ref{itm:we-groupoidal} follows similarly, via Definition \ref{defn:groupoidal-object}\ref{itm:groupoidal-object:ii}.
The preservation halves of \ref{itm:we-equiv}-\ref{itm:we-comma} holds for any functor of $\infty$-cosmoi, as observed in Proposition \ref{prop:induced-2-functor}. The reflection halves of \ref{itm:we-equiv} and \ref{itm:we-equivs} hold for any biequivalence, by a standard argument. The proof of the remaining half of \ref{itm:we-comma} is similar to the proof of \ref{itm:we-equiv}, using the fact that the local equivalence of mapping quasi-categories pulls back to define a local equivalence of fibered mapping quasi-categories
\[ \xymatrix@=1em{ \fun_B(E,E') \ar@{-->}[dr]^-{\rotatebox{155}{$\labelstyle\sim$}} \ar[rr] \save[]+<5pt,-5pt>*{\pbcorner}\restore \ar@{->>}[dd] & & \fun(E,E') \ar[dr]^-{\rotatebox{155}{$\labelstyle\sim$}} \ar@{->>}'[d][dd] \\ & \fun_{FB}(FE,FE') \save[]+<5pt,-5pt>*{\pbcorner}\restore \ar@{->>}[dd] \ar[rr] & & \fun(FE,FE') \ar@{->>}[dd] \\ \del@fn\parsedel@^0 \ar'[r][rr] \ar@{=}[dr] & & \fun(E,B) \ar[dr]^-{\rotatebox{155}{$\labelstyle\sim$}} \\ & \del@fn\parsedel@^0 \ar[rr] & & \fun(FE,FB)}\]
\end{proof}
The assertion made in \ref{prop:we}\ref{itm:we-comma} can be strengthened, using Lemma \ref{lem:equiv-to-comma}:
\begin{exs}\label{exs:we-comma-redux} Show that:
\begin{enumerate}[label=(\roman*)]
\item\label{itm:we-comma-redux:i} If $f \cong f' \colon B \to A$ and $g \cong g' \colon C \to A$, then $f \mathbin{\downarrow} g \simeq f' \mathbin{\downarrow} g'$ over $C \times B$.
\end{enumerate}
Combine this with Lemma \extRef{equipment}{V}{lem:equiv-invariance-of-commas}, which says that a commutative diagram
\[ \xymatrix{ C' \ar[r]^{g'} \ar[d]_{c}^{\rotatebox{90}{$\labelstyle\sim$}} & A' \ar[d]_a^{\rotatebox{90}{$\labelstyle\sim$}} & B' \ar[l]_{f'} \ar[d]^b_{\rotatebox{90}{$\labelstyle\sim$}} \\ C \ar[r]_g & A & B \ar[l]^f}\] induces an equivalence $f \mathbin{\downarrow} g \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} f' \mathbin{\downarrow} g'$ over $c \times b \colon C' \times B' \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} C \times B$, to conclude:
\begin{enumerate}[label=(\roman*),resume]
\item If $E \twoheadrightarrow C \times B$ is an isofibration in $\mathcal{K}$ whose image under a weak equivalence of $\infty$-cosmoi $F \colon \mathcal{K} \to \mathcal{L}$ defines a comma $\infty$-category for some pair of functors in $\mathcal{L}$, then $E \twoheadrightarrow C \times B$ defines a comma $\infty$-category in $\mathcal{K}$.
\end{enumerate}
\end{exs}
\begin{thm}[model independence of basic category theory I]\label{thm:model-independence} The following notions are preserved and reflected by any weak equivalence of $\infty$-cosmoi:
\begin{enumerate}[label=(\roman*)]
\item\label{itm:model-independence-i} The adjointness of a pair of $\infty$-functors $f \colon B \to A$ and $u \colon A \to B$.
\item\label{itm:model-independence-ii} The existence of a left or right adjoint to an $\infty$-functor $u \colon A \to B$.
\item\label{itm:model-independence-iii} The question of whether a given element $\ell \colon 1 \to A$ defines a limit or a colimit for a diagram $d \colon 1 \to A^J$.
\item\label{itm:model-independence-iv} The existence of a limit or a colimit for a $J$-indexed diagram $d \colon 1 \to A^J$ in an $\infty$-category $A$.
\end{enumerate}
\end{thm}
\begin{proof} \ref{itm:model-independence-i} and \ref{itm:model-independence-iii} follow directly from Proposition \ref{prop:we}\ref{itm:we-comma}, via the characterizations of Propositions \ref{prop:adjoint-comma-equiv-I}, \ref{prop:adjoint-comma-equiv-II}, and \ref{prop:limit-comma-equiv}. Then \ref{itm:model-independence-ii} and \ref{itm:model-independence-iv} follow, using \ref{prop:we}\ref{itm:we-iso} to lift an adjoint or limit element from $\mathcal{L}$ to $\mathcal{K}$ and Exercise \ref{exs:we-comma-redux}\ref{itm:we-comma-redux:i} to transport the universal property encoded by an equivalence of commas along isomorphic functors.
\end{proof}
\begin{rmk} By definition, if $F \colon \mathcal{K} \to \mathcal{L}$ is a functor of $\infty$-cosmoi, $A \in \mathcal{K}$, and $J \in \category{sSet}$, then $F(A^J) \cong FA^J \in \mathcal{L}$. If $\mathcal{K}$ and $\mathcal{L}$ are cartesian closed $\infty$-cosmoi and $J \in \mathcal{K}$, then there is a natural functor $F(A^J) \to FA^{FJ}$, the transpose of the image of the evaluation map $A^J \times J \to A$ under $F$. If $F$ is a weak equivalence of $\infty$-cosmoi, then this map induces a natural equivalence $\fun(X, F(A^J)) \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} \fun(X, FA^{FJ})$ for all $X\in\mathcal{L}$, and so $F(A^J) \simeq FA^{FJ}$. With more care, analogs of the assertions of Theorem \ref{thm:model-independence}\ref{itm:model-independence-iii} and \ref{itm:model-independence-iv} concerning limit and colimits of simplicial set-indexed diagrams can be proven for diagrams indexed by another $\infty$-category.
\end{rmk}
\renewcommand\thesection{Lecture~\arabic{section}}
\section{Fibrations, modules, and Kan extensions}\label{sec:modules}
\renewcommand\thesection{\arabic{section}}
A section of Saunders Mac Lane's \emph{Categories for the Working Mathematician} is famously entitled ``All concepts are Kan extensions.'' Our aim in this final lecture is to develop the theory of Kan extensions for functors between $\infty$-categories.
At first glance, this might seem easy. After all, any 2-category has an internally-defined notion of extension diagram.
\begin{defn}\label{defn:right-ext} Given a span $C \xleftarrow{f} A \xrightarrow{k} B$, a functor $r \colon B \to C$ and a 2-cell
\[ \xymatrix{ A \ar[r]^k \ar[d]_f & B \ar[dl]^r \\ C & \ar@{}[ul]|(.7){\Leftarrow\nu}}\]
define a \emph{right extension of $f$ along $k$} if any 2-cell as displayed below-left factors uniquely through $\nu$ as displayed below-right:
\[ \vcenter{\xymatrix{ A \ar[d]_f \ar[r]^k \ar@{}[dr]|(.3){\Leftarrow\chi} & B \ar[dl]^g \\ C & }} \mkern20mu = \mkern20mu \vcenter{ \xymatrix{ A \ar[r]^k \ar[d]_f & B \ar[dl]|r \ar@/^4ex/[dl]^g \\ C & \ar@{}[ul]|(.3){\Leftarrow\exists!}|(.7){\Leftarrow\nu}}}
\] Dually, a functor $\ell \colon B \to C$ and a 2-cell $\lambda \colon f \Rightarrow \ell k$ define a \emph{left extension of $f$ along $k$} if any 2-cell as displayed below-left factors uniquely through $\lambda$ as displayed below-right:
\[ \vcenter{\xymatrix{ A \ar[d]_f \ar[r]^k \ar@{}[dr]|(.3){\Rightarrow\chi} & B \ar[dl]^g \\ C & }} \mkern20mu = \mkern20mu \vcenter{ \xymatrix{ A \ar[r]^k \ar[d]_f & B \ar[dl]|\ell \ar@/^4ex/[dl]^g \\ C & \ar@{}[ul]|(.3){\Rightarrow\exists!}|(.7){\Rightarrow\lambda}}}
\]
\end{defn}
In the 2-category $\category{Cat}_2$, \ref{defn:right-ext} defines the usual right and left Kan extensions for functors between ordinary categories. Special cases of these, in turn, define adjunctions (by a dual to Exercise \ref{exs:adj-as-abs-lifting}) and limits and colimit (as right or left extensions of the diagram $f \colon A \to C$ along the functor $!\colon A \to 1$). However, in general it turns out that the universal property expressed by this naive notion of right extension is insufficiently strong. In particular, Definition \ref{defn:right-ext}, interpreted in the homotopy 2-category, does not define the correction notion of Kan extension for $\infty$-functors. Instead, the correction notion will be of \emph{pointwise right Kan extensions}.
In fact, we will give two equivalent definitions of pointwise Kan extensions, both of which make use of comma $\infty$-categories, our vehicle for encoding $\infty$-categorical universal properties. One of these could be stated immediately, but we instead delay it in order to first develop the prerequisite theory for the other. Specifically, our aim will be to describe the full universal property of the comma $\infty$-category construction: namely that an isofibration $(p_1,p_0) \colon f \mathbin{\downarrow} g \twoheadrightarrow C \times B$ constructed from functors $C \xrightarrow{g} A \xleftarrow{f} B$ encodes a \emph{module} from $C$ to $B$, with $C$ acting covariantly ``on the left'' and $B$ acting contravariantly ``on the right.'' The calculus of modules describes the 2-dimensional structure into which modules most naturally assemble, which turns out to be a familiar setting for formal category theory, the scope of which includes pointwise Kan extensions.
\subsection{Arrow \texorpdfstring{$\infty$}{infinity}-categories define modules}
The prototypical examples of modules are comma $\infty$-categories. For simplicity, we specialize to the case of the arrow $\infty$-category $(p_1,p_0)\colon A^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A \times A$. Lemma \ref{lem:ess.unique.1-cell.ind} describes its universal property: 2-cells $\xymatrix{X \ar@/^2ex/[r]^a \ar@/_2ex/[r]_b \ar@{}[r]|{\Downarrow g} & A}$ with codomain $A$ correspond to generalized elements $\hat{g} \colon X \to A^{\bigger{1.12}{\mathbbe{2}}}$, up to isomorphism over $A \times A$. By Lemma \ref{lem:equiv-to-comma}, this universal property characterizes the arrow $\infty$-category up to equivalence of isofibrations over $A \times A$. However, it does not capture the additional fact that 2-cells from $X$ to $A$ can be composed vertically
\[\vcenter{\xymatrix@R=30pt{ X \ar@/^5.5ex/[r]^x_{\Downarrow f} \ar@/_5.5ex/[r]_y^{\Downarrow h} \ar@/^2ex/[r]|a \ar@/_2ex/[r]|b \ar@{}[r]|{\Downarrow g}& A}} \qquad \qquad (h \cdot g) \cdot f = h \cdot g \cdot f = h \cdot (g \cdot f),\]
defining commuting contravariant and covariant actions on the domains and codomains of the 2-cell $g \colon a \Rightarrow b$.
Observe that the domain and codomain 1-cells for the 2-cell $g \colon a \Rightarrow b$ can be recovered as the composites of a representing functor $\hat{g} \colon X \to A^{\bigger{1.12}{\mathbbe{2}}}$ with the projection functors: $p_0 \hat{g} = a$ and $p_1 \hat{g} = b$.
One way to express these actions is to note that the domain-projection functor $p_0 \colon A^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A$ and the codomain-projection functor $p_1 \colon A^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A$ respectively define a \emph{cartesian fibration} and a \emph{cocartesian fibration}: any 2-cells as displayed on the left-hand side of the pasting equalities below admit lifts as displayed on the right-hand sides:
\[ \vcenter{\xymatrix{ X \ar[r]^{\hat{g}} \ar[dr]_x & A^{\bigger{1.12}{\mathbbe{2}}} \ar@{->>}[d]^{p_0} \ar@{}[dl]|(.3){\Uparrow f} \ar@{}[dr]|{\displaystyle ~=} & X \ar@/^2ex/[r]^{\hat{g}} \ar@/_2ex/[r]_{f^*(x)} \ar@{}[r]|{\Uparrow \chi_{f}} & A^{\bigger{1.12}{\mathbbe{2}}} \ar@{->>}[d]^{p_0} \\ & A & & A}} \qquad \vcenter{\xymatrix{ X \ar[r]^{\hat{g}} \ar[dr]_y & A^{\bigger{1.12}{\mathbbe{2}}} \ar@{->>}[d]^{p_1} \ar@{}[dl]|(.3){\Downarrow h} \ar@{}[dr]|{\displaystyle ~= } & X \ar@/^2ex/[r]^{\hat{g}} \ar@/_2ex/[r]_{h_*(y)} \ar@{}[r]|{\Downarrow \chi^{h}} & A^{\bigger{1.12}{\mathbbe{2}}} \ar@{->>}[d]^{p_1} \\ & A & & A}}
\]
Moreover, the lifted 2-cell $\chi_f$ can be chosen to project along $p_1 \colon A^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A$ to $\id_b$ and the lifted 2-cell $\chi^h$ can be chosen to project along $p_0 \colon A^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A$ to $\id_a$.
In summary, the arrow $\infty$-category $A^{\bigger{1.12}{\mathbbe{2}}}$ defines a \emph{module} from $A$ to $A$. The definition will make use of the fact that slices of $\infty$-cosmoi are again $\infty$-cosmoi.
\begin{defn}[{\extRef{equipment}{V}{defn:sliced-cosmoi}}]\label{defn:sliced-cosmoi}
If $\mathcal{K}$ is any $\infty$-cosmos and $B \in \mathcal{K}$ is any object, then there is an $\infty$-cosmos $\mathcal{K}/B$, the \emph{sliced $\infty$-cosmos of $\mathcal{K}$ over $B$}, whose:
\begin{itemize}
\item objects are isofibrations $p \colon E \twoheadrightarrow B$ with codomain $B$;
\item mapping quasi-category from $p \colon E \twoheadrightarrow B$ to $q \colon F \twoheadrightarrow B$ is defined by taking the pullback
\[
\xymatrix@=1.5em{
{\fun_B(p,q)}\save[]+<5pt,-5pt>*{\pbcorner}\restore\ar[r]\ar@{->>}[d] &
{\fun(E,F)}\ar@{->>}[d]^{\fun(E,q)} \\
{\del@fn\parsedel@^0}\ar[r]_-{p} & {\fun(E,B)}
}
\]
in simplicial sets;
\item isofibrations, equivalences, and trivial fibrations are created by the forgetful functor $\mathcal{K}/B \to \mathcal{K}$;
\end{itemize}
and in which the simplicial limits are defined in the usual way for sliced simplicial categories.
\end{defn}
\begin{defn}\label{defn:module} In an $\infty$-cosmos $\mathcal{K}$, a \emph{module $E$ from $A$ to $B$}, denoted by $\dmod{E}{A}{B}$, is given by an isofibration $(q,p) \colon E \twoheadrightarrow A \times B$ such that
\begin{enumerate}[label=(\roman*)]
\item\label{itm:cartesian-on-the-right} $\vcenter{\xymatrix@=1em{ E \ar@{->>}[dr]_q \ar[rr]^-{(q,p)} & & A \times B \ar@{->>}[dl]^{\pi_1} \\ & A}}$ is a \emph{cartesian fibration} in $\mathcal{K}/A$; informally, ``$B$ acts on the right of $E$, over $A$.''
\item\label{itm:cocartesian-on-the-left} $\vcenter{\xymatrix@=1em{ E \ar@{->>}[dr]_p \ar[rr]^-{(q,p)} & & A \times B \ar@{->>}[dl]^{\pi_0} \\ & B}}$ is a \emph{cocartesian fibration} in $\mathcal{K}/B$; informally, ``$A$ acts on the left of $E$, over $B$.''
\item\label{itm:groupoidal-fibers} $(q,p) \colon E \twoheadrightarrow A \times B$ is groupoidal as an object in $\mathcal{K}/A \times B$; this means that any 2-cell $\xymatrix{X \ar@/^2ex/[r]^e \ar@/_2ex/[r]_{e'} \ar@{}[r]|{\Downarrow} & E}$ over an identity in $A \times B$ is an isomorphism, which implies in particular that $(q,p) \colon E \twoheadrightarrow A \times B$ has groupoidal fibers.
\end{enumerate}
\end{defn}
\begin{ex}[{\extRef{equipment}{V}{prop:hom-is-a-module}}]\label{ex:hom-is-a-module} For any $\infty$-category $A$, the arrow $\infty$-category defines a module $\dmod{A^{\bigger{1.12}{\mathbbe{2}}}}{A}{A}$. The fact that $A^{\bigger{1.12}{\mathbbe{2}}} \twoheadrightarrow A \times A$ is groupoidal is related to but stronger than the fact that each fiber over a pair of elements in $A$, the hom-spaces of Definition \ref{defn:hom-space}, is a groupoidal $\infty$-category.
\end{ex}
\subsection{Cartesian and cocartesian fibrations}
To explain Definition \ref{defn:module}, we need to define what it means for a functor in an $\infty$-cosmos to be a \emph{cartesian fibration} or \emph{cocartesian fibration}. We will not actually require any of these details for out ultimate aim in this lecture, to initiate the theory of pointwise Kan extensions, but we include them because these fibration notions are of independent interest.
\begin{defn}[cartesian 2-cells]
A 2-cell $\chi \colon e' \Rightarrow e \colon A \to E$ in the homotopy 2-category of an $\infty$-cosmos is \emph{cartesian} for an isofibration $p \colon E \twoheadrightarrow B$ if and only if
\begin{enumerate}[label=(\roman*)]
\item\label{itm:weak.cart.i} (induction) for any pair of 2-cells $\tau \colon e'' \Rightarrow e$ and $\gamma\colon pe''\Rightarrow pe'$ with $p \tau = p\chi \cdot \gamma$ there is some $\overline{\gamma} \colon e'' \Rightarrow e'$ with $p\overline\gamma = \gamma$ ($\bar\gamma$ lies over $\gamma$) and the property that $\tau = \chi \cdot \bar\gamma$.
\item\label{itm:weak.cart.ii} (conservativity) for any 2-cell $\gamma \colon e' \Rightarrow e'$ if $\chi \cdot \gamma = \chi$ and $p\gamma$ is an identity then $\gamma$ is an isomorphism.
\end{enumerate}
\end{defn}
All isomorphisms with codomain $E$ are $p$-cartesian. The class of $p$-cartesian 2-cells is stable under composition and left cancelation (Lemmas~\extRef{yoneda}{IV}{lem:cart-arrows-compose} and~\extRef{yoneda}{IV}{lem:cart-arrows-cancel}).
\begin{defn}[cartesian fibration]\label{defn:cart-fib}
An isofibration $p\colon E\twoheadrightarrow B$ is a {\em cartesian fibration\/} if and only if:
\begin{enumerate}[label=(\roman*)]
\item Every 2-cell $\alpha \colon b \Rightarrow pe$ has a $p$-cartesian lift $\chi_\alpha \colon \alpha^*(e)\Rightarrow e$:
\[\vcenter{\xymatrix{ X \ar[r]^{e} \ar[dr]_b & E \ar@{->>}[d]^{p} \ar@{}[dl]|(.3){\Uparrow \alpha} \ar@{}[dr]|{\displaystyle ~=} & X \ar@/^2ex/[r]^{e} \ar@/_2ex/[r]_{\alpha^*(e)} \ar@{}[r]|{\Uparrow \chi_\alpha } &E \ar@{->>}[d]^{p} \\ & B & & B}} \]
\item The class of $p$-cartesian 2-cells for $p$ is closed under pre-composition by all 1-cells.
\end{enumerate}
\end{defn}
Importantly, there is a ``model independent'' characterization of cartesian fibrations given in terms of adjunctions between commas. Any functor $p \colon E \to B$ induces functors between comma $\infty$-categories
\begin{equation*}
\vcenter{\xymatrix@C=0.8em@R=1.2em{
& {E}\ar[d]^-{i} & \\
& {B\mathbin{\downarrow} p}\ar@{->>}[dl]_{p_1}\ar@{->>}[dr]^{p_0} & \\
{E}\ar[rr]_{p} && {B}
\ar@{} "2,2";"3,2" |(0.6){\Leftarrow\phi}
}} =
\vcenter{\xymatrix@C=0.8em@R=1.2em{
& {E}\ar@{=}[dl]\ar@{->>}[dr]^{p} & \\
{E}\ar[rr]_{p} && {B}
\ar@{} "1,2";"2,2" |(0.6){=}
}}
\mkern50mu
\vcenter{\xymatrix@C=0.8em@R=1.2em{
& {E^{\bigger{1.12}{\mathbbe{2}}}}\ar[d]^-{k} & \\
& {B\mathbin{\downarrow} p}\ar@{->>}[dl]_{p_1}\ar@{->>}[dr]^{p_0} & \\
{E}\ar[rr]_{p} && {B}
\ar@{} "2,2";"3,2" |(0.6){\Leftarrow\phi}
}} =
\vcenter{\xymatrix@C=0.8em@R=1.2em{
& {E^{\bigger{1.12}{\mathbbe{2}}}}\ar@{->>}_{q_1}[dl]\ar@{->>}[dr]^{pq_0} & \\
{E}\ar[rr]_{p} && {B}
\ar@{} "1,2";"2,2" |(0.6){\Leftarrow p\psi}
}}
\end{equation*}
that are well-defined up to isomorphism over $E \times B$.
\begin{thm}[{\extRef{yoneda}{IV}{thm:cart.fib.chars}}]\label{thm:cart.fib.chars} For an isofibration $p \colon E \twoheadrightarrow B$, the following are equivalent:
\begin{enumerate}[label=(\roman*)]
\item\label{itm:cart.fib.chars.i} $p$ is a cartesian fibration.
\item\label{itm:cart.fib.chars.ii} The functor $i\colon E\to B\mathbin{\downarrow} p$ admits a right adjoint which is fibered over $B$.
\[
\xymatrix@R=2em@C=3em{
{B\mathbin{\downarrow} p}\ar@{->>}[dr]_{p_0} \ar@{-->}@/_0.6pc/[]!R(0.5);[rr]_{r}^{}="a" & &
{E}\ar@{->>}[dl]^{p} \ar@/_0.6pc/[ll]!R(0.5)_{i}^{}="b"
\ar@{}"a";"b"|{\bot} \\
& B &
}
\]
\item\label{itm:cart.fib.chars.iii} The functor $k\colon E^{\bigger{1.12}{\mathbbe{2}}}\to B\mathbin{\downarrow} p$ admits a right adjoint right inverse, i.e., a right adjoint with invertible counit.
\[
\xymatrix@C=6em{
{B\mathbin{\downarrow} p}\ar@{-->}@/_0.8pc/[]!R(0.6);[r]!L(0.45)_{\bar{r}}^{}="u" &
{E^{\bigger{1.12}{\mathbbe{2}}}}\ar@/_0.8pc/[]!L(0.45);[l]!R(0.6)_{k}^{}="t"
\ar@{}"u";"t"|(0.6){\bot}
}
\]
\end{enumerate}
\end{thm}
\emph{Cocartesian fibrations} are defined dually, by reversing the direction of the 2-cells in Definition \ref{defn:cart-fib} and of the adjoints in Theorem \ref{thm:cart.fib.chars}. A cartesian or cocartesian fibration $p \colon E \twoheadrightarrow B$ that defines a groupoidal object in the sliced $\infty$-cosmos over $B$ is called a \emph{groupoidal cartesian fibration} or \emph{groupoidal cocartesian fibration}. A groupoidal cartesian fibration is a cartesian fibration with groupoidal fibers. Condition \ref{itm:gpd.cart.fib.chars.iii} of the following theorem provides a ``model independent'' characterization of groupoidal cartesian fibrations.
\begin{thm}[{\extRef{equipment}{V}{thm:gpd.cart.fib.chars}}]\label{thm:gpd.cart.fib.chars} For an isofibration $p \colon E \twoheadrightarrow B$, the following are equivalent:
\begin{enumerate}[label=(\roman*)]
\item\label{itm:gpd.cart.fib.chars.i} $p$ is a groupoidal cartesian fibration.
\item\label{itm:gpd.cart.fib.chars.ii} Every 2-cell $\alpha\colon b\Rightarrow pe\colon X\to B$ has an essentially unique lift $\chi\colon e'\Rightarrow e\colon X\to E$, where the essential uniqueness is up to composition into the domain of $\chi$ with an invertible 2-cell that projects along $p$ to an identity.
\item\label{itm:gpd.cart.fib.chars.iii} The functor $k\colon E^{\bigger{1.12}{\mathbbe{2}}}\to B\mathbin{\downarrow} p$ is an equivalence.
\end{enumerate}
\end{thm}
Recall that a functor $F \colon \mathcal{K} \to \mathcal{L}$ of $\infty$-cosmoi induces a 2-functor between the homotopy 2-categories that preserves equivalences, adjunctions, and commas, and if $F$ is a weak equivalence of $\infty$-cosmoi, then these notions are also reflected; see Proposition \ref{prop:induced-2-functor}, Proposition \ref{prop:we}, and Theorem \ref{thm:model-independence}. It is straightforward to extend Theorem \ref{thm:model-independence}\ref{itm:model-independence-ii} to see that the existence of a right adjoint right inverse is preserved and reflected as well. By Theorem \ref{thm:cart.fib.chars}\ref{itm:cart.fib.chars.iii} and Theorem \ref{thm:gpd.cart.fib.chars}\ref{itm:gpd.cart.fib.chars.iii} we conclude:
\begin{cor}\label{cor:model-indep-fib} $\quad$
\begin{enumerate}[label=(\roman*)]
\item Functors of $\infty$-cosmoi preserve cartesian fibrations, cocartesian fibrations, group\-oid\-al cartesian fibrations, groupoidal cocartesian fibrations, and modules.
\item Weak equivalences of $\infty$-cosmoi both preserve and reflect cartesian fibrations, cocartesian fibrations, groupoidal cartesian fibrations, groupoidal cocartesian fibrations, and modules: an isofibration in the domain is a functor of this type if and only if its image is.
\end{enumerate}
\end{cor}
\begin{proof}
The preservation and reflection of a module from $A$ to $B$ follows from the statements concerning fibrations together with Propositions \ref{prop:induced-2-functor} and \ref{prop:we}\ref{itm:we-groupoidal} applied to the sliced functor of $\infty$-cosmoi $\mathcal{K}/A\times B \to \mathcal{L}/FA \times FB$.
\end{proof}
\subsection{The calculus of modules}
The calculus of modules between $\infty$-categories bears a strong resemblance to the calculus of (bi)modules between unital rings, with functors between $\infty$-categories playing the role of ring homomorphisms.\footnote{In more detail, unital rings, ring homomorphisms, bimodules, and module maps define a \emph{proarrow equipment}, in the sense of Wood \cite{wood:proI}. This can be seen as a special case of the prototypical equipment comprised of $\mathcal{V}$-categories, $\mathcal{V}$-functors, $\mathcal{V}$-modules, and $\mathcal{V}$-natural transformations between then, for any closed symmetric monoidal category $\mathcal{V}$. The equipment for rings is obtained from the case where $\mathcal{V}$ is the category of abelian groups by restricting to abelian group enriched categories with a single object.}
\begin{center}
\fbox{\begin{tabular}{ccc}
unital rings & $A$ & $\infty$-categories \\
ring homomorphisms & $\vcenter{\xymatrix{ A \ar[d]_f \\ B}}$ & $\infty$-functors \\
bimodules between rings & $\xymatrix{ A \ar[r]|\mid^E & B}$ & modules between $\infty$-categories \\
module maps & $\raisebox{.75cm}{ \xymatrix{A' \ar[d]_a \ar@{}[dr]|{\Downarrow} \ar[r]|\mid^{E'} & B' \ar[d]^b \\ A \ar[r]|\mid_E & B}}$ & module maps
\end{tabular}}
\end{center}
Our first result, analogous to restriction of scalars, is that modules can be pulled back.
\begin{prop}[{\extRef{equipment}{V}{prop:two-sided-pullback}}]\label{prop:module-pullback} If $(q,p) \colon E \twoheadrightarrow A \times B$ defines a module from $A$ to $B$ and $a \colon A' \to A$ and $b \colon B' \to B$ are any functors then the pullback
\[ \xymatrix{ E(b,a) \ar@{->>}[d] \ar[r] \save[]+<5pt,-5pt>*{\pbcorner}\restore & E \ar@{->>}[d]^{(q,p)} \\ A' \times B' \ar[r]_{a \times b} & A \times B}\] defines a module $E(b,a)$ from $A'$ to $B'$.
\end{prop}
\begin{ex} Recall that comma $\infty$-categories are defined to be pullbacks of arrow $\infty$-categories. \[ \xymatrix@=2.5em{
{f\mathbin{\downarrow} g}\save[]+<5pt,-5pt>*{\pbcorner}\restore \ar[r]\ar@{->>}[d]_{(p_1,p_0)} &
{A^{\bigger{1.12}{\mathbbe{2}}}} \ar@{->>}[d]^{(p_1,p_0)} \\
{C\times B} \ar[r]_-{g\times f} & {A\times A}
}\]
Since we know from Example \ref{ex:hom-is-a-module} that $A^{\bigger{1.12}{\mathbbe{2}}}$ defines a module $\dmod{A^{\bigger{1.12}{\mathbbe{2}}}}{A}{A}$, Proposition \ref{prop:module-pullback} implies that $f\mathbin{\downarrow} g$ defines a module $\dmod{f \mathbin{\downarrow} g}{C}{B}$.
\end{ex}
As is the case for rings, horizontal composition of modules between $\infty$-categories is a complicated operation. For general modules $\dmod{E}{A}{B}$ and $\dmod{F}{B}{C}$, the pullback
\[ \xymatrix@!=5pt{ & & E \times_B F \save[]+<0pt,-5pt>*{\rotatebox{-45}{$\pbcorner$}}\restore \ar@{->>}[dl]_-{\pi_1} \ar@{->>}[dr]^-{\pi_0} \\ & E \ar@{->>}[dl]_q \ar@{->>}[dr]^p & & F \ar@{->>}[dl]_s \ar@{->>}[dr]^r \\ A & & B & & C}\] defines an isofibration $E \times_B F \twoheadrightarrow A \times C$ that is a groupoidal cartesian fibration in the slice over $A$ and a groupoidal cocartesian fibration in the slice over $C$. However, it fails to be a groupoidal object over $A \times C$, as can be seen by considering the case $E=F=A^{\bigger{1.12}{\mathbbe{2}}}$. We use the notation $E \times_B F$ to denote this pullback construction, reserving $E \otimes_B F$ for special cases in which there is a module from $A$ to $C$ that can be recognized as the horizontal composite. For particular types of $\infty$-categories, general composite modules $E \otimes_B F$ can be defined via a ``fiberwise coinverter'' construction, but to do so requires that we leave the $\infty$-cosmos axiomatization, which does not provide for any colimits.
Rather than leave the axiomatization in search of a general composition formula, it turns out to be simpler to do without it. There is a natural categorical framework into which modules assemble, even without general horizontal composites, that turns out to be sufficient for our real goal: developing a theory of pointwise Kan extensions. Rings, ring homomorphisms, modules, and module maps assemble into a 2-dimensional structure known as a \emph{double category}. By analogy, $\infty$-categories, $\infty$-functors, modules, and module maps, to be introduced assemble into a \emph{virtual double category}.
\begin{defn}[{\extRef{equipment}{V}{defn:modules-virtual}}] The \emph{virtual double category of modules} $\MMod{K}$ in an $\infty$-cosmos $\mathcal{K}$ consists of
\begin{itemize}
\item a category of \emph{objects} and \emph{vertical arrows}, here the $\infty$-categories and $\infty$-functors
\item for any pair of objects $A, B$, a class of \emph{horizontal arrows} $\dmod{~}{A}{B}$, here the modules from $A$ to $B$
\item \emph{cells}, with boundary depicted as follows
\begin{equation}\label{eq:generic-cell} \xymatrix{ A_0 \ar[d]_{f} \ar[r]|{\mid}^{E_1} & A_1 \ar[r]|{\mid}^{E_2} \ar@{}[dr]|{\Downarrow} & \cdots \ar[r]|{\mid}^{E_n} & A_n \ar[d]^g \\ B_0 \ar[rrr]|{\mid}_{F} & & & B_n}\end{equation} including those whose horizontal source has length zero, in the case $A_0 = A_n$. Here, a cell with boundary \eqref{eq:generic-cell} will be an isomorphism class of objects in the mapping quasi-category
\[ \xymatrix{ \fun_{f,g}(E_1 \pbtimes{A_1} \cdots \pbtimes{A_{n-1}} E_n, F) \ar[r] \ar[d] \save[]+<5pt,-5pt>*{\pbcorner}\restore & \fun(E_1 \pbtimes{A_1} \cdots \pbtimes{A_{n-1}} E_n, F) \ar[d] \\ \Delta^0 \ar[r] & \fun(E_1 \pbtimes{A_1} \cdots \pbtimes{A_{n-1}} E_n, B_0 \times B_1)}\]
\item a \emph{composite cell}, for any configuration
\[ \xymatrix@C=35pt{ A_0 \ar[d]_{f_0} \ar@{..>}[r]|{\mid}^{E_{11},\dots, E_{1n_1}} \ar@{}[dr]|{\Downarrow} & A_1 \ar@{..>}[r]|{\mid}^{E_{21},\ldots, E_{2n_2}} \ar[d]^{f_1} \ar@{}[dr]|{\Downarrow} & \cdots \ar@{..>}[r]|{\mid}^{E_{n1},\ldots, E_{nn_n}} \ar@{}[d]|\cdots \ar@{}[dr]|{\Downarrow} & A_n \ar[d]^{f_n} \\ B_0 \ar[d]_{g} \ar[r]|{\mid}^{F_1} & B_1 \ar[r]|{\mid}^{F_2} \ar@{}[dr]|{\Downarrow} & \cdots \ar[r]|{\mid}^{F_n} & B_n \ar[d]^h \\ C_0 \ar[rrr]|{\mid}_{G} & & & C_n}\]
\item an \emph{identity cell} for every horizontal arrow
\[ \xymatrix{ A \ar@{=}[d] \ar[r]|{\mid}^E & B \ar@{=}[d] \ar@{}[dl]|{\Downarrow \id_E} \\ A \ar[r]|{\mid}_E & B}\]
\end{itemize}
so that composition of cells is associative and unital in the usual multi-categorical sense.
\end{defn}
The following examples motivate our definition of \emph{module maps}, i.e., cells in the virtual double category $\MMod{K}$.
\begin{ex}\label{ex:cells-above-a-comma} Lemma \ref{lem:ess.unique.1-cell.ind}, which expresses 1-cell induction as a bijection between isomorphism classes of maps of spans whose codomain is a comma span and certain 2-cells in the homotopy 2-category, provides an alternate characterization of cells in the virtual double category of modules whose codomain is a comma module. Explicitly, for any cospan $C \xrightarrow{g} A \xleftarrow{f} B$, there is a bijection between cells in $\MMod{K}$ whose codomain is the comma module $\dmod{f \mathbin{\downarrow} g}{C}{B}$ and 2-cells in the homotopy 2-category $\mathcal{K}_2$ under the pullback of the spans encoding the domain modules and over the cospan defining the comma module $f \mathbin{\downarrow} g$.
\[ \vcenter{ \xymatrix{ A_0 \ar[d]_{c} \ar[r]|{\mid}^{E_1} & A_1 \ar[r]|{\mid}^{E_2} \ar@{}[dr]|{\Downarrow} & \cdots \ar[r]|{\mid}^{E_n} & A_n \ar[d]^b \\ C \ar[rrr]|{\mid}_{f \mathbin{\downarrow} g} & & & B}}\qquad \leftrightsquigarrow\qquad \vcenter{ \xymatrix@!0@C=40pt@R=30pt{ & E_1 \pbtimes{A_1} \cdots \pbtimes{A_{n-1}} E_n \ar[dl] \ar[dr] \ar@{}[ddd]|{\displaystyle\Leftarrow} \\ A_0 \ar[d]_c & & A_n \ar[d]^b \\ C \ar[dr]_g & & B \ar[dl]^f \\ & A}}\]
\end{ex}
\begin{exs}\label{exs:sample-cells} Use the correspondence described in Example \ref{ex:cells-above-a-comma} to define canonical nullary and unary cells in $\MMod{K}$ associated to any functor $f \colon A \to B$.
\[ \xymatrix { A \ar@{=}[r] \ar@{=}[d] \ar@{}[dr]|{\Downarrow\nu} & A \ar[d]^f & & A \ar[d]_f \ar[r]|\mid^{B \mathbin{\downarrow} f} \ar@{}[dr]|{\Downarrow\rho} & B \ar@{=}[d] & & A \ar@{=}[r] \ar[d]_f \ar@{}[dr]|{\Downarrow\nu} & A \ar@{=}[d] & & B \ar@{=}[d] \ar[r]|\mid^{f \mathbin{\downarrow} B} \ar@{}[dr]|{\Downarrow\rho} & A \ar[d]^f \\ A \ar[r]_{B \mathbin{\downarrow} f}|{\mid} & B & & B \ar[r]|\mid_{B^{\bigger{1.12}{\mathbbe{2}}}} & B & & B \ar[r]|\mid_{f \mathbin{\downarrow} B} & A & & B \ar[r]|\mid_{B^{\bigger{1.12}{\mathbbe{2}}}} & B}\]
\end{exs}
\begin{ex}[{\extRef{equipment}{V}{prop:two-sided-yoneda}}]\label{ex:yoneda}
Any functor $f \colon A \to B$ induces a map
\[ \vcenter{\xymatrix@C=0.8em@R=1.2em{
& {A}\ar[d]^-{t} & \\
& {B \mathbin{\downarrow} f}\ar@{->>}[dl]_{p_1}\ar@{->>}[dr]^{p_0} & \\
{A}\ar[rr]_{f} && {B}
\ar@{} "2,2";"3,2" |(0.6){\Leftarrow\phi}
}} =
\vcenter{\xymatrix@C=0.8em@R=1.2em{
& {A}\ar@{=}[dl]\ar[dr]^{f} & \\
{A}\ar[rr]_{f} && {B}
\ar@{} "1,2";"2,2" |(0.6){=}
}} \]
over $A \times B$. Then for any module $E$ from $A$ to $B$, pre-composition with $t \colon A \to B \mathbin{\downarrow} f$ induces an equivalence of quasi-categories
\[ \fun_{A \times B} ( B \mathbin{\downarrow} f , E) \stackrel{\simeq}{\longrightarrow} \fun_{A \times B} ( A , E ).\] This result is a direct application of the Yoneda lemma for groupoidal cartesian fibrations (\extRef{yoneda}{IV}{cor:groupoidal-yoneda}) in the slice $\infty$-cosmos $\mathcal{K}/A$. Passing to isomorphism classes of objects in the mapping quasi-categories, this result asserts that there is a bijection between cells in $\MMod{K}$
\[ \vcenter{\xymatrix{A \ar@{=}[d] \ar[r]|{\mid}^{B \mathbin{\downarrow} f} \ar@{}[dr]|{\Downarrow} & B \ar@{=}[d] \\ A \ar[r]|\mid_E & B}} \qquad \stackrel{\cong}{\mapsto}\qquad \vcenter{\xymatrix{A \ar@{=}[d]\ar@{=}[r] \ar@{}[dr]|{\Downarrow} & A \ar[d]^f \\ A \ar[r]|\mid_E & B}} \]
implemented by restricting along the nullary cell $\nu$, as defined in Exercise \ref{exs:sample-cells}, represented by the functor $t \colon A \to B \mathbin{\downarrow} f$.
\end{ex}
Proposition \ref{prop:module-pullback} tells us that modules in an $\infty$-cosmos can be pulled back. Given $\dmod{E}{A}{B}$ and functors $a \colon A' \to A$ and $b \colon B' \to B$, the horizontal functor $\rho \colon E(b,a) \to E$ in the diagram defining the pullback module
\begin{equation}\label{eq:simplicial-pullback-of-module} \xymatrix{ E(b,a) \save[]+<5pt,-5pt>*{\pbcorner}\restore \ar@{->>}[d]_{(q',p')} \ar[r]^-\rho & E \ar@{->>}[d]^{(q,p)} \\ A' \times B' \ar[r]_{a \times b} & A \times B}\end{equation} defines a unary cell in the virtual double category of modules with a universal property that we now describe.
\begin{prop}[{\extRef{equipment}{V}{prop:cartesian-pullback-cells}}]\label{prop:cartesian-pullback-cells} In $\MMod{K}$, the cell
\[ \xymatrix{ A' \ar[d]_a \ar[r]|{\mid}^{E(b,a)} \ar@{}[dr]|{\Downarrow\rho} & B' \ar[d]^b \\ A \ar[r]|{\mid}_E & B}\]
defined by pulling back a module $\dmod{E}{A}{B}$ along functors $a \colon A' \to A$ and $b \colon B' \to B$ has the property that any cell as displayed on the left
\[ \vcenter{ \xymatrix{ X_0 \ar[d]_{af} \ar[r]|{\mid}^{E_1} & X_1 \ar[r]|{\mid}^{E_2} \ar@{}[dr]|{\Downarrow} & \cdots \ar[r]|{\mid}^{E_n} & X_n \ar[d]^{bg} \\ A \ar[rrr]_E|{\mid} & & & B}} \quad = \quad \vcenter{ \xymatrix{ X_0 \ar[d]_{f} \ar[r]|{\mid}^{E_1} & X_1 \ar[r]|{\mid}^{E_2} \ar@{}[dr]|{\Downarrow \exists !} & \cdots \ar[r]|{\mid}^{E_n} & X_n \ar[d]^{g} \\ A' \ar[d]_a\ar[rrr]^{E(b,a)}|{\mid} & \ar@{}[dr]|{\Downarrow\rho} & & B' \ar[d]^b \\ A \ar[rrr]_E|{\mid} & & & B}} \]
factors uniquely as displayed on the right.
\end{prop}
Proposition \ref{prop:cartesian-pullback-cells} asserts that $\rho$ is a \emph{cartesian cell} in $\MMod{K}$.
\begin{proof}
The simplicial pullback \eqref{eq:simplicial-pullback-of-module}, induces an equivalence of hom quasi-categories
\[ \fun_{af,bg}(E_1 \times \cdots \times E_n, E) \simeq \fun_{f,g}(E_1 \times \cdots \times E_n, E(b,a)). \qedhere \]
\end{proof}
Each module $\dmod{A^{\bigger{1.12}{\mathbbe{2}}}}{A}{A}$ defined by the arrow construction comes with a canonical cell with nullary source. Under the identification of Example \ref{ex:cells-above-a-comma}, this cell corresponds via 1-cell induction to the isomorphism class of maps of spans representing the identity 2-cell at the identity 1-cell of the object $A$.
\[ \vcenter{\xymatrix { A \ar@{=}[d] \ar@{=}[r] \ar@{}[dr]|{\Downarrow\iota} & A \ar@{=}[d] \\ A \ar[r]_{A^{\bigger{1.12}{\mathbbe{2}}}}|{\mid} & A}} \quad \leftrightsquigarrow \quad \vcenter{
\xymatrix@R=2.5em@C=0.8em{
{A}\ar@/^3ex/[]!D(0.4);[d]!U(0.5)^{\id_A}
\ar@/_3ex/!D(0.4);[d]!U(0.5)_{\id_A}
\ar@{}[d]|{=} \\ {A}
}} \quad = \quad \vcenter{
\xymatrix@R=2.5em@C=0.8em{
{A}\ar[d]^-{j} \\
{A^{\bigger{1.12}{\mathbbe{2}}}}\ar@{->>}@/^3ex/[]!D(0.4);[d]!U(0.5)^{p_0}
\ar@{->>}@/_3ex/!D(0.4);[d]!U(0.5)_{p_1}
\ar@{}[d]|{\Leftarrow\phi} \\ {A}
}} \]
This cell also has a universal property in the virtual double category of modules.
\begin{prop}[{\extRef{equipment}{V}{prop:arrows-are-units}}]\label{prop:arrows-are-units} Any cell in the virtual double category of modules whose horizontal source includes the object $A$, as displayed on the left
\[ \vcenter{\xymatrix@C=15pt{ X \ar[d]_{f} \ar[r]|{\mid}^{E_1} & \cdots \ar[r]|-{\mid}^-{E_n} & A \ar@{}[d]|{\Downarrow} \ar[r]|-{\mid}^-{F_1} & \cdots \ar[r]|{\mid}^{F_m} & Y \ar[d]^g \\ B \ar[rrrr]|{\mid}_{G} & & & & C}} \quad=\quad \vcenter{ \xymatrix@C=20pt{ X \ar@{=}[d] \ar[r]|{\mid}^{E_1} \ar@{}[dr]|{\Downarrow\id_{E_1}{~}} & \cdots \ar@{=}@<-1.5ex>[d]^{\cdots} \ar@{=}@<1.5ex>[d] \ar[r]|-{\mid}^-{E_n} \ar@{}[dr]|{{~}{~}\Downarrow\id_{E_n}}& A \ar@{=}[r] \ar@{=}[d] \ar@{}[dr]|{\Downarrow\iota} & \ar@{=}[d]A \ar@{}[dr]|{\Downarrow\id_{F_1}{~}} \ar[r]|-{\mid}^-{F_1} & \cdots \ar@{=}@<-1.5ex>[d]^{\cdots} \ar@{=}@<1.5ex>[d] \ar[r]|{\mid}^{F_m} \ar@{}[dr]|{{~}{~}\Downarrow\id_{F_m}}& Y \ar@{=}[d] \\ X \ar[d]_{f} \ar[r]|{\mid}^{E_1} & \cdots \ar[r]|-{\mid}^-{E_n} & A \ar[r]|{\mid}^{A^{\bigger{1.12}{\mathbbe{2}}}} \ar@{}[dr]|{\Downarrow\exists !} & A \ar[r]|-{\mid}^-{F_1} & \cdots \ar[r]|{\mid}^{F_m} & Y \ar[d]^g \\ B \ar[rrrrr]|{\mid}_{G} & & & & & C}}\]
factors uniquely through $\iota$ as displayed on the right.
\end{prop}
Proposition \ref{prop:arrows-are-units} asserts that $\iota$ is a \emph{cocartesian cell} in $\MMod{K}$.
\begin{proof}
In the case where both of the sequences $E_i$ and $F_j$ are empty, the Yoneda lemma for modules, in the form described in Example \ref{ex:yoneda}, supplies an equivalence of quasi-categories
\[ \xymatrix{ \fun_{f,g}(A^{\bigger{1.12}{\mathbbe{2}}}, G) \simeq \fun_{A \times A}(A^{\bigger{1.12}{\mathbbe{2}}}, G(g,f)) \ar[r]^-{j^*}_{\simeq} & \fun_{A \times A} (A, G(g,f)) \simeq \fun_{f,g}(A,G) .}\] This equivalence descends to a bijection between isomorphism classes of objects, i.e., to a bijection between cells
\[ \vcenter{\xymatrix{A \ar[d]_f \ar[r]|{\mid}^{A^{\bigger{1.12}{\mathbbe{2}}}} \ar@{}[dr]|{\Downarrow} & A \ar[d]^g \\ B \ar[r]|\mid_G & C}} \qquad \stackrel{\cong}{\mapsto}\qquad \vcenter{\xymatrix{A \ar[d]_f \ar@{=}[r] \ar@{}[dr]|{\Downarrow} & A \ar[d]^g \\ B \ar[r]|\mid_G & C}} \]
implemented by restricting along the cocartesian cell $\iota$. See \extRef{equipment}{V}{prop:arrows-are-units} for the proof in the general case.
\end{proof}
Propositions \ref{prop:cartesian-pullback-cells} and \ref{prop:arrows-are-units} imply that the virtual double category of modules is a \emph{virtual equipment} in the sense introduced by Cruttwell and Shulman \cite[\S 7]{CruttwellShulman:2010au}.
\begin{thm}\label{thm:virtual-equipment} The virtual double category $\MMod{K}$ of modules in an $\infty$-cosmos $\mathcal{K}$ is a virtual equipment: i.e., $\MMod{K}$ is a virtual double category such that
\begin{enumerate}[label=(\roman*)]
\item For any module and pair of functors as displayed on the left, there exists a module and cartesian cell as displayed on the right satisfying the universal property of Proposition \ref{prop:cartesian-pullback-cells}. \[ \xymatrix{ A' \ar[d]_a & B' \ar[d]^b & \ar@{}[d]|{\displaystyle\stackrel{\exists}{\rightsquigarrow}} & A' \ar[d]_a \ar[r]|\mid^{E(b,a)} \ar@{}[dr]|{\Downarrow\rho} & B' \ar[d]^b \\ A \ar[r]|\mid_E & B & & A \ar[r]|\mid_E & B}\]
\item Every object $A$ admits a \emph{unit} module equipped with a nullary cocartesian cell satisfying the universal property of Proposition \ref{prop:arrows-are-units}.
\[ \xymatrix{ A \ar@{=}[r] \ar@{=}[d] \ar@{}[dr]|{\Downarrow\iota} & A \ar@{=}[d] \\ A \ar[r]|\mid_{A^{\bigger{1.12}{\mathbbe{2}}}} & A}\]
\end{enumerate}
\end{thm}
The \emph{virtual equipment of modules} in $\mathcal{K}$ has a lot of pleasant properties, which follow formally from the axiomatization of a virtual equipment. For instance, certain sequences of composable modules can be said to \emph{have composites}, witnessed by cocartesian cells as in Proposition \ref{prop:arrows-are-units} (see \extRef{equipment}{V}{lem:comp-with-unit}, \extRef{equipment}{V}{lem:comp-with-unit-cells}, \extRef{equipment}{V}{lem:one-sided-rep-comp}, and \extRef{equipment}{V}{cor:two-sided-rep-comp}). Also, for any functor $f \colon A \to B$, the modules $\dmod{B \mathbin{\downarrow} f}{A}{B}$ and $\dmod{f \mathbin{\downarrow} B}{B}{A}$ behave like adjoints is a sense suitable to a virtual double category; more precisely, the module $\dmod{B \mathbin{\downarrow} f}{A}{B}$ defines a \emph{companion} and the module $\dmod{f \mathbin{\downarrow} B}{B}{A}$ defines a \emph{conjoint} to $f \colon A \to B$ (see \extRef{equipment}{V}{thm:companion-conjoint} and \extRef{equipment}{V}{cor:companion-conjoint}). Another formal consequence of Theorem \ref{thm:virtual-equipment} is the following:
\begin{lem}[{\extRef{equipment}{V}{lem:yoneda-embedding}}]\label{lem:yoneda-embedding} For any pair of parallel functors there are natural bijections between 2-cells
$ \xymatrix{ A \ar@/^2ex/[r]^f \ar@/_2ex/[r]_g \ar@{}[r]|{\Downarrow} & B}$ in the homotopy 2-category and cells
\[ \xymatrix{ A \ar@{=}[d] \ar[r]|\mid^{B \mathbin{\downarrow} f} \ar@{}[dr]|{\Downarrow} & B \ar@{=}[d] & \ar@{}[d]|{\displaystyle\leftrightsquigarrow} & A \ar[r]|\mid^{A^{{\bigger{1.12}{\mathbbe{2}}}}} \ar[d]_g \ar@{}[dr]|{\Downarrow} & A \ar[d]^f & \ar@{}[d]|{\displaystyle\leftrightsquigarrow} & B \ar@{}[dr]|{\Downarrow} \ar@{=}[d] \ar[r]|\mid^{g \mathbin{\downarrow} B} & A \ar@{=}[d] \\ A \ar[r]|\mid_{B \mathbin{\downarrow} g} & B & & B \ar[r]|\mid_{B^{\bigger{1.12}{\mathbbe{2}}}} & B & & B \ar[r]|\mid_{f \mathbin{\downarrow} B} & A}\]
in the virtual equipment of modules.
\end{lem}
It follows from Lemma \ref{lem:yoneda-embedding} and the cited results about composition of modules that there are two locally-fully-faithful homomorphisms $\mathcal{K}_2 \hookrightarrow \MMod{K}$ and $\mathcal{K}_2^{\mathord{\text{\rm coop}}} \hookrightarrow\MMod{K}$ embedding the homotopy 2-category into the sub bicategory of $\MMod{K}$ comprised only of unary cells whose vertical boundaries are identities. The modules in the image of the first homomorphism are the covariant representables and the modules in the image of the second homomorphism are the contravariant representables. We refer to these as the \emph{covariant} and \emph{contravariant embeddings}, respectively.
\subsection{Pointwise Kan extensions}
We are now close to achieving our goal, definitions of pointwise Kan extensions between $\infty$-categories. By Proposition \ref{prop:cartesian-pullback-cells}, any cell in the virtual equipment of modules can be represented uniquely as a cell between parallel sequences of modules, which we display inline as $E_1 \times_{A_1} \cdots \times_{A_{n-1}} E_n \Rightarrow E$, where $E$ is a module from $A$ to $B$ and $E_1,\ldots, E_n$ is a composable sequence of modules starting at $A$ and ending at $B$.
\begin{defn}[right extension of modules]\label{defn:right-extension-mod}
In the virtual equipment $\MMod{K}$ of modules, a \emph{right extension} of a module $\dmod{F}{A}{C}$ along a module $\dmod{K}{A}{B}$ is given by a module $\dmod{R}{B}{C}$ together with a cell $\nu \colon K \times_B R \Rightarrow F$ so that for any composable sequence of modules $E_1,\ldots, E_n$ from $B$ to $C$, composition with $\nu$ defines a bijection
\[\vcenter{ \xymatrix@=1em{ A \ar[rr]|\mid^K \ar[dd]|{\rotatebox{90}{$\labelstyle\mid$}}_F & & B \ar[d]|(.4){\rotatebox{90}{$\labelstyle\mid$}}^{E_1} \\ & & A_1\ar@/^/@{..>}[dl] \\C & A_{n-1} \ar[l]|(.6)\mid^{E_n} \ar@{}[uu]|{\Leftarrow\chi}}} \qquad = \qquad \vcenter{ \xymatrix@=1em{ A \ar[rr]|\mid^K \ar@{}[ddrr]|(.3){\Leftarrow\nu}|(.7){\Leftarrow\exists!} \ar[dd]|{\rotatebox{90}{$\labelstyle\mid$}}_F & & B \ar[ddll]|{\rotatebox{45}{$\labelstyle\mid$}}^R \ar[d]|(.4){\rotatebox{90}{$\labelstyle\mid$}}^{E_1} \\ & & A_1\ar@/^/@{..>}[dl] \\C & A_{n-1} \ar[l]|(.6)\mid^{E_n} & }} \]
\end{defn}
In the case where the modules $\dmod{K}{A}{B}$, $\dmod{F}{A}{C}$, and $\dmod{R}{B}{C}$ are all covariant representables, the Yoneda lemma, in the form of Lemma \ref{lem:yoneda-embedding}, implies that the binary cell $\nu \colon K \times_B R \Rightarrow F$ arises from a 2-cell in the homotopy 2-category. The following lemma, whose proof is left as an exercise, asserts that this 2-cell defines a right extension in $\mathcal{K}_2$ in the sense of Definition \ref{defn:right-ext}.
\begin{lem}[{\extRef{equipment}{V}{lem:mod-extensions-in-2-cat}}]\label{lem:mod-extensions-in-2-cat}
If $\nu \colon B \mathbin{\downarrow} k \times_B C \mathbin{\downarrow} r \Rightarrow C \mathbin{\downarrow} f$ displays $\dmod{C \mathbin{\downarrow} r}{B}{C}$ as a right extension of $\dmod{C \mathbin{\downarrow} f}{A}{C}$ along $\dmod{B \mathbin{\downarrow} k}{A}{B}$ in $\MMod{K}$, then $\nu \colon rk \Rightarrow f$ displays $r$ as the right extension of $f$ along $k$ in $\mathcal{K}_2$.
\[ \xymatrix{ A \ar[r]|\mid^{B \mathbin{\downarrow} k} \ar[d]|{\rotatebox{90}{$\labelstyle\mid$}}_{C \mathbin{\downarrow} f} & B \ar[dl]|{\rotatebox{45}{$\labelstyle\mid$}}^{C \mathbin{\downarrow} r} & \ar@{}[d]|{\displaystyle \rightsquigarrow} & A \ar[r]^k \ar[d]_f & B \ar[dl]^r \\ C & \ar@{}[ul]|(.7){\Leftarrow\nu} & & C & \ar@{}[ul]|(.7){\Leftarrow\nu}}\]
\end{lem}
\begin{defn}[stability of extensions under pasting]
In any 2-category, a right extension diagram
\[ \xymatrix{ A \ar[r]^k \ar[d]_f & B \ar[dl]^r \\ C & \ar@{}[ul]|(.7){\Leftarrow\nu}}\]
is said to be \emph{stable under pasting with a square}
\[ \xymatrix{ D \ar[d]_g \ar[r]^h \ar@{}[dr]|{\Leftarrow\lambda} & E \ar[d]^b \\ A \ar[r]_k & B}\]
if the pasted diagram
\[ \xymatrix{ D \ar[r]^h \ar[d]_g \ar@{}[dr]|{\Leftarrow\lambda} & E \ar[d]^b\\
A \ar[r]^k \ar[d]_f & B \ar[dl]^r \\ C & \ar@{}[ul]|(.7){\Leftarrow\nu}}\]
displays $br$ as a right extension of $fg$ along $h$.
\end{defn}
The following proposition defines \emph{pointwise right Kan extensions} for $\infty$-categories:
\begin{prop}[{\extRef{equipment}{V}{prop:pointwise-kan}}]\label{prop:pointwise-kan} For a diagram
\[ \xymatrix{ A \ar[r]^k \ar[d]_f & B \ar[dl]^r \\ C & \ar@{}[ul]|(.7){\Leftarrow\nu}}\]
in the homotopy 2-category of an $\infty$-cosmos $\mathcal{K}$ the following are equivalent.
\begin{enumerate}[label=(\roman*)]
\item\label{itm:pointwise-comma-stable} $\nu \colon rk \Rightarrow f$ defines a right extension in $\mathcal{K}_2$ that is stable under pasting with commas squares.
\item\label{itm:pointwise-in-mod} The image $\nu \colon B \mathbin{\downarrow} k \times_B C \mathbin{\downarrow} r \Rightarrow C \mathbin{\downarrow} f$ of $\nu$ under the covariant embedding $\mathcal{K}_2\hookrightarrow\MMod{K}$ defines a right extension in $\MMod{K}$.
\end{enumerate}
\end{prop}
There are other equivalent conditions, one of which says that pointwise right Kan extensions are right extension diagrams that are stable under pasting with a larger class of \emph{exact squares} in $\mathcal{K}_2$. Exact squares, which are characterized using the virtual equipment $\MMod{K}$, can also be used to define fully faithful functors of $\infty$-categories (see \extRef{equipment}{V}{defn:fully-faithful}) and initial and final functors between $\infty$-categories (see \extRef{equipment}{V}{defn:initial-final}).
Definition \ref{defn:right-extension-mod} can be dualized, by turning around the 1-cells but not the 2-cells, to define a \emph{right lifting} $\dmod{R}{C}{B}$ of a module $\dmod{F}{C}{A}$ along a module $\dmod{K}{B}{A}$. In the case, where all three modules are contravariant representables, the Yoneda lemma, in the form of Lemma \ref{lem:yoneda-embedding}, implies that the binary cell $\nu \colon r \mathbin{\downarrow} C \times_B k \mathbin{\downarrow} B \Rightarrow f \mathbin{\downarrow} C$ arises from a 2-cell $\nu \colon f \Rightarrow rk$ in the homotopy 2-category, which defines a left extension in $\mathcal{K}_2$ in the sense of Definition \ref{defn:right-ext}.
The dual to Proposition \ref{prop:pointwise-kan} defines \emph{pointwise left Kan extensions} for $\infty$-categories:
\begin{prop} For a diagram
\[ \xymatrix{ A \ar[r]^k \ar[d]_f & B \ar[dl]^r \\ C & \ar@{}[ul]|(.7){\Rightarrow\nu}}\]
in the homotopy 2-category of an $\infty$-cosmos $\mathcal{K}$ the following are equivalent.
\begin{enumerate}[label=(\roman*)]
\item\label{itm:pointwise-comma-stable-L} $\nu \colon f \Rightarrow rk$ defines a left extension in $\mathcal{K}_2$ that is stable under pasting with commas squares.
\item\label{itm:pointwise-in-mod-L} The image $\nu \colon r \mathbin{\downarrow} C \times_B k \mathbin{\downarrow} B \Rightarrow f \mathbin{\downarrow} C$ of $\nu$ under the contravariant embedding $\mathcal{K}_2^{\mathord{\text{\rm coop}}}\hookrightarrow\MMod{K}$ defines a right lifting in $\MMod{K}$.
\end{enumerate}
\end{prop}
Recall Definition \ref{defn:limit}: in a cartesian closed $\infty$-cosmos, the \emph{limit} of a diagram $d \colon J \to A$ is a element $\ell \colon 1 \to A$ equipped with an absolute right lifting diagram
\begin{equation}\label{eq:limit-abs-lifting} \xymatrix{ \ar@{}[dr]|(.7){\Downarrow\nu} & A \ar[d]^{\Delta} \\ 1 \ar[r]_-{d} \ar[ur]^\ell & A^J}\end{equation} Here the 2-cell $\nu$ encodes the data of the limit cone.
\begin{prop}[{\extRef{equipment}{V}{prop:limits-as-pointwise-kan}}]\label{prop:limits-as-pointwise-kan} In a cartesian closed $\infty$-cosmos $\mathcal{K}$, any limit \eqref{eq:limit-abs-lifting} defines a pointwise right Kan extension
\[ \xymatrix{ J \ar[d]_d \ar[r]^{!} \ar@{}[dr]|(.3){\Leftarrow\nu} & 1 \ar[dl]^\ell \\ A& }\] Conversely, any pointwise right Kan extension of this form transposes to define a limit \eqref{eq:limit-abs-lifting} in $A$.
\end{prop}
Proposition \ref{prop:limits-as-pointwise-kan} suggests the way to extend the definition of limits and colimits of diagrams indexed by $\infty$-categories to non-cartesian closed $\infty$-cosmoi.
\subsection{Model independence revisited}
We conclude with some remarks concerning the model independence of these notions. Note that a functor of $\infty$-cosmoi $F \colon \mathcal{K} \to \mathcal{L}$ induces a functor of sliced $\infty$-cosmoi $F \colon \mathcal{K}/B \to \mathcal{L}/FB$ for any $B \in \mathcal{K}$.
\begin{prop}\label{prop:sliced-we} Suppose $F \colon \mathcal{K} \to \mathcal{L}$ is a weak equivalence of $\infty$-cosmoi. Then the induced functor
$F \colon \mathcal{K}/B \to \mathcal{L}/FB$ is also a weak equivalence of $\infty$-cosmoi.
\end{prop}
\begin{proof}
We first argue that the functor between slices defines a local equivalence of sliced mapping quasi-categories, as defined in \ref{defn:sliced-cosmoi}. Given a pair of isofibration $p \colon E \twoheadrightarrow B$ and $p' \colon E' \twoheadrightarrow B$ in $\mathcal{K}$, the induced functor on mapping quasi-categories is defined by
\[ \xymatrix@=1em{ \fun_B(p,p') \ar@{-->}[dr]^-{\rotatebox{155}{$\labelstyle\sim$}} \ar[rr] \save[]+<5pt,-5pt>*{\pbcorner}\restore \ar@{->>}[dd] & & \fun(E,E') \ar[dr]^-{\rotatebox{155}{$\labelstyle\sim$}} \ar@{->>}'[d][dd] \\ & \fun_{FB}(FE,FE') \save[]+<5pt,-5pt>*{\pbcorner}\restore \ar@{->>}[dd] \ar[rr] & & \fun(FE,FE') \ar@{->>}[dd] \\ \del@fn\parsedel@^0 \ar'[r][rr] \ar@{=}[dr] & & \fun(E,B) \ar[dr]^-{\rotatebox{155}{$\labelstyle\sim$}} \\ & \del@fn\parsedel@^0 \ar[rr] & & \fun(FE,FB)}\] As the maps between the cospans in $\category{qCat}$ are equivalences, so is the induced map between the pullbacks.
For surjectivity up to equivalence, consider an isofibration $q \colon L \twoheadrightarrow FB$ in $\mathcal{L}$. As $F$ is surjective on objects up to equivalence, there exists some $A \in \mathcal{K}$ together with an equivalence $i \colon FA \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} L \in \mathcal{L}$. As $F$ defines a local equivalence of mapping quasi-categories, there is moreover a functor $f \colon A \to B$ in $\mathcal{L}$ so that $Ff \colon FA \to FB$ is isomorphic to $qi$ in $\mathcal{L}_2$. The map $f$ need not be an isofibration, but \ref{qcat.ctxt.cof.def}\ref{qcat.ctxt.cof:e} allows us to factor $f$ as $A \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} K \xtfib{p} B$. Choosing an equivalence inverse $j \colon K \xrightarrow{\mkern10mu{\smash{\mathlower{0.6}{\sim}}}\mkern10mu} A$, the result defines a diagram in $\mathcal{L}_2$ that does not commute on the nose but which does commute up to isomorphism.
\[ \xymatrix{FK \ar@{->>}[dr]_{Fp} \ar[r]^{Fj}_{\sim} & FA \ar[d]|{Ff} \ar[r]^{i}_{\sim} & L \ar@{->>}[dl]^q \\ \ar@{}[ur]|(.7){\cong} & FB & \ar@{}[ul]|(.7){\cong} }\]
Now a basic fact about isofibrations in an $\infty$-cosmos that we have not had occasion to mention is that they define isofibrations in the homotopy 2-category. An \emph{isofibration} in a 2-category is a 1-cell that has a lifting property for isomorphisms with one chosen endpoint; see \extRef{yoneda}{IV}{rec:trivial-fibration} and \extRef{yoneda}{IV}{lem:isofib.are.representably.so}. In particular, as $q \colon L \twoheadrightarrow FB$ defines an isofibration in $\mathcal{L}_2$, and so we may lift the displayed isomorphism along $q$ to define a commutative triangle:
\[ \vcenter{\xymatrix{FK \ar@{->>}[dr]_{Fp} \ar[r]^{Fj}_{\sim} & FA \ar[d]|{Ff} \ar[r]^{i}_{\sim} & L \ar@{->>}[dl]^q \\ \ar@{}[ur]|(.7){\cong} & FB & \ar@{}[ul]|(.7){\cong} }} \qquad = \qquad \vcenter{ \xymatrix{FK \ar@{->>}[dr]_{Fp} \ar@/^2ex/[rr]^{i \cdot Fj} \ar@{}[rr]|\cong \ar@/_2ex/[rr]_e & & L \ar@{->>}[dl]^q \\ & FB & } }\] As $e$ is isomorphic to an equivalence $i \cdot Fj$, it must also define an equivalence, whence we have shown that the isofibration $p \colon K \twoheadrightarrow B$ maps under $F$ to an isofibration that is equivalent to our chosen $q \colon L \twoheadrightarrow FB$.
\end{proof}
By Proposition \ref{prop:induced-2-functor} and Corollary \ref{cor:model-indep-fib}, a functor $F \colon \mathcal{K} \to \mathcal{L}$ of $\infty$-cosmoi induces a functor of virtual equipments $F\colon \MMod{K} \to \MMod{L}$; the important point here is that modules and simplicial pullbacks are preserved.
\begin{prop}\label{prop:equipment-we} If $F \colon \mathcal{K} \to \mathcal{L}$ is a weak equivalence of $\infty$-cosmoi, then the induced functor $F \colon \MMod{K} \to \MMod{L}$ defines a biequivalence of virtual equipments: i.e., it is
\begin{enumerate}[label=(\roman*)]
\item\label{itm:equipment-we:i} surjective on objects up to equivalence;
\item\label{itm:equipment-we:ii} locally bijective on isomorphism classes of parallel vertical functors;
\item\label{itm:equipment-we:iii} locally bijective on equivalence classes of parallel modules;
\item\label{itm:equipment-we:iv} locally bijective on cells.
\end{enumerate}
\end{prop}
\begin{proof}
\ref{itm:equipment-we:i} and \ref{itm:equipment-we:ii} are restatements of Proposition \ref{prop:we}\ref{itm:we-bi-equiv} and \ref{itm:we-iso}. The proof of \ref{itm:equipment-we:iii} is more subtle. Corollary \ref{cor:model-indep-fib} tells us that the property of an isofibration $E \twoheadrightarrow A \times B$ in $\mathcal{K}$ defining a module is both preserved and reflected by $F$. An argument similar to that given in the proof of Proposition \ref{prop:sliced-we} shows that the weak equivalence of $\infty$-cosmoi $F \colon \mathcal{K}/A\times B \to \mathcal{L}/FA \times FB$ creates modules, and of course also preserves and reflects equivalences. Finally \ref{itm:equipment-we:iv} is an application of Proposition \ref{prop:we}\ref{itm:we-iso} to this weak equivalence of $\infty$-cosmoi.
\end{proof}
\begin{thm}[model independence of basic category theory II] Any basic $\infty$-categorical notion that can be encoded as an equivalence-invariant proposition in the virtual equipment of modules is model invariant: preserved and reflected by weak equivalences of $\infty$-cosmoi.
\end{thm}
We are only just beginning to explore the consequences of this result but they appear to be quite strong. For instance, we can prove that any (large) quasi-category $E$ that admits limits and colimits of every diagram indexed by a small category defines a \emph{derivator}
\[ \category{Cat}^{\mathord{\text{\rm op}}} \xrightarrow{E^{-}} \category{qCat} \xrightarrow{h} \category{Cat}\]
in the sense of \cite{Heller:1988ly}; see \extRef{equipment}{V}{rmk:qcat-derivator}. The proof makes use of two (non-formal) facts that we prove directly in $\category{qCat}$:
\begin{enumerate}[label=(\roman*)]
\item\label{itm:derivator-i} For any pair of modules $\dmod{G}{A}{C}$ and $\dmod{K}{A}{B}$ between quasi-categories, there exists a right extension $\dmod{R}{B}{C}$.
\item\label{itm:derivator-ii} A module $\dmod{R}{B}{C}$ is equivalent to a comma module $\dmod{C \mathbin{\downarrow} r}{B}{C}$ for some functor $r \colon B \to C$ between quasi-categories if and only if this property is true for the modules obtained by pulling back along each vertex $b \colon 1 \to B$.
\end{enumerate}
Now if $\mathcal{K} \to \category{qCat}$ is any weak equivalence of $\infty$-cosmoi, the biequivalence $\MMod{K} \to \MMod{\category{qCat}}$ implies that \ref{itm:derivator-i} also holds in $\mathcal{K}$. Given a pair of modules in $\mathcal{K}$ form the right extension of their images in $\MMod{\category{qCat}}$ and use \ref{prop:equipment-we}\ref{itm:equipment-we:iii} and \ref{itm:equipment-we:iv} to lift this module and the universal cell to a right extension diagram in $\MMod{K}$. Now if $\dmod{R}{B}{C}$ is a module in $\mathcal{K}$ that pulls back along all elements $b \colon 1 \to B$ to a represented module, then its image in $\category{qCat}$ also has this property, using \ref{prop:equipment-we}\ref{itm:equipment-we:ii} to see that elements in $B$ correspond to elements in $FB$ up to isomorphism. The representing functor for the module $\dmod{FR}{FB}{FC}$ between quasi-categories lifts to a representing functor for $\dmod{R}{B}{C}$.
As a consequence, we conclude that in an $\infty$-cosmos that is weakly equivalent to $\category{qCat}$, any $\infty$-category that admits limits and colimits of every small diagram defines a derivator.
By a similar argument, a related result --- \extRef{found}{I}{cor:pointwise} --- that says that universal properties in $\category{qCat}_2$ are determined ``pointwise'' can also be generalized to weakly equivalent $\infty$-cosmoi.
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction}
\label{sec::intro}
\subsection{Overview}
\label{sec::overview}
The unit area Brownian map (TBM) and the unit area Liouville quantum gravity (LQG) sphere (with parameter $\gamma = \sqrt{8/3}$) are two natural continuum models for ``random surfaces'' which are both homeomorphic to the Euclidean sphere ${\mathbf S}^2$. Both objects come naturally endowed with an area measure. However, an instance of TBM additionally comes endowed with a metric space structure (i.e., a two-point distance function) while an instance of the LQG sphere comes endowed with a conformal structure (i.e., it can be canonically parameterized by ${\mathbf S}^2$ up to M\"obius transformation).
This article is the third and last in a series of papers that explain how to endow each of these random objects with the {\it other's} structure in a canonical way and show that once this is done the two objects agree in law. In other words, once both TBM and the $\sqrt{8/3}$-LQG sphere are endowed with both structures, they encode exactly the same information and have exactly the same law. From this point of view, any theorem about TBM is henceforth also a theorem about the $\sqrt{8/3}$-LQG sphere and vice-versa.
The first two papers in the current series \cite{qlebm, qle_continuity} show how to equip an instance~${\mathcal S}$ of the unit area $\sqrt{8/3}$-LQG sphere (as defined in \cite{dms2014mating,quantum_spheres}) with a metric in such a way that the resulting metric measure space (mm-space) agrees in law with TBM.
To phrase this another way, suppose one first samples an instance ${\mathcal S}$ of the LQG sphere, and then uses it to construct an instance ${\mathcal S}_{{\mathrm {BM}}}$ of TBM (by endowing ${\mathcal S}$ with a metric structure and then forgetting about the conformal structure) together with the identity homeomorphism $\phi$ from ${\mathcal S}$ to $\mathcal {\mathcal S}_{{\mathrm {BM}}}$. Then the results of \cite{qlebm, qle_continuity} imply that the law of $({\mathcal S}, {\mathcal S}_{{\mathrm {BM}}})$ is a {\em coupling} of the $\sqrt{8/3}$-LQG sphere and TBM. Moreover, it is shown in \cite{qlebm,qle_continuity} that in this coupling the first object a.s.\ determines the second. The current paper shows that the second object a.s.\ determines the first; indeed, given only~${\mathcal S}_{{\mathrm {BM}}}$ one can a.s.\ reconstruct both~${\mathcal S}$ and~$\phi$. In particular, this implies that there is a.s.\ a canonical way to endow an instance of TBM with a conformal structure (namely, the conformal structure it inherits from $\mathcal S$).
Although the three papers belong to the same series, let us stress that the methods and results are extremely different. The current paper implements a subtle conformal removability analysis that has no parallel in either \cite{qlebm} or \cite{qle_continuity}, the first two papers in the series. As explained in Section~\ref{subsec::outline}, the current paper {\em does} have some strategic similarity to \cite[Section~10]{dms2014mating} but the conceptual details that appear here are more difficult and delicate. Let us also stress the practical importance of the current paper. Once one knows that an instance of TBM comes with a {\em canonical} conformal structure, one can construct Brownian motions, conformal loop ensembles, and other natural random geometric objects {\em on top of} an instance of TBM. This allows one to formulate many scaling limit conjectures (e.g., that simple random walk on random planar maps scales to Brownian motion on TBM) and may pave the way to a deeper understanding of the relationships between various discrete and continuous models. Moreover, conformal symmetries and conformal field theory have obviously been of great importance to the physics literature on random surfaces. Now that we have endowed TBM with a conformal structure, one can apply these techniques to the study of TBM.
Throughout most of this paper, we will actually work with an infinite volume variant of TBM called the {\em Brownian plane} (TBP), as defined in \cite{cl2012brownianplane}, which was shown in \cite{qle_continuity} to be equivalent to an infinite volume variant of the LQG-sphere, called a {\em quantum cone}, as defined in \cite{SHE_WELD, dms2014mating}. We will then use these results about infinite volume surfaces to deduce analogous results about finite volume spheres and disks
We remark that, in some sense, our conclusions are even more general. First we observe that in a certain sense both length space structure and conformal structure are ``local'' properties. To be more precise, let $U_1, U_2, \ldots, U_k$ be any finite cover of the sphere with open neighborhoods. It is not hard to see that if one knows the conformal structure on each $U_i$, one can recover the conformal structure of the whole sphere. Similarly, if one knows the ``internal metric'' structure of each $U_i$ (i.e., one can define the length of any path that lies strictly inside one of the $U_i$) then this determines the length of any path (and hence the overall metric) on the entire sphere, because if $\eta:[0,1] \to {\mathcal S}_{{\mathrm {BM}}}$ is such a path, then the connected components of the sets $\eta^{-1}(U_i)$ form an open cover of $[0,1]$ and hence must have a finite subcover since $[0,1]$ is compact.)
The method for producing a length space structure on LQG, as described in \cite{qlebm, qle_continuity}, was shown to be ``local'' in the sense that it is a.s.\ the case that the internal metric on any open subset $U_i$ of a $\sqrt{8/3}$-LQG sphere instance is determined (in a particular concrete way) by the measure and conformal structure {\em within} $U_i$. This paper will show that the method for recovering the conformal structure from the metric is local in a similar sense.
Informally, these statements imply that we can impose a canonical length space structure on {\em any} surface that looks like a $\sqrt{8/3}$-LQG sphere locally, and a canonical conformal structure on {\em any} surface that looks locally like an instance of TBM. This should allow one to address, for example, higher genus variants of TBM and LQG.
However, let us stress if one has a {\em specific} method for generating a higher genus form of TBM --- and another {\em specific} method for generating a higher genus LQG surface, one would still have to do some additional work to show that the first surface (endowed with its canonical conformal structure) and the second surface (endowed with its canonical length space structure) agree in law. There are many variants of both LQG and TBM (which may differ in genus, boundary structure, and number and type of special marked points). See for example the recently announced work in preparation by Bettinelli and Miermont (on the Brownian map side) \cite{bm_compact2} or the work by Guillarmou, Rhodes and Vargas (on the LQG side) in \cite{guillarmourhodesvargas}. It remains an open problem to show that these surfaces described in these papers are equivalent. We only note here that all of these variants can now be defined as random variables on the same space --- the space of (possibly marked) {\em conformal mm-spaces} --- and that comparing the laws of surfaces defined on the LQG side with the laws of surfaces defined on the Brownian map side is therefore at least possible in principle.
\subsection{Main results from previous papers}
\label{subsec::prior_main_results}
Before we proceed, let us recall briefly the approach taken in the first two papers in this series. The first paper \cite{qlebm} considers a sequence of points $(x_n)$ chosen i.i.d.\ from the measure on an instance ${\mathcal S}$ of the $\sqrt{8/3}$-LQG sphere, and shows how to define a metric~$d_\CQ$ on $(x_n)$. This is accomplished by letting $d_\CQ(x_i, x_j)$ denote the amount of time it takes for a certain growth process, started at $x_i$, to absorb the point $x_j$. This growth process is a form of the {\em quantum Loewner evolution with parameters} $(\gamma^2,\eta) = (8/3,0)$ (i.e.\ ${\rm QLE}(8/3,0)$) as introduced by the authors in \cite{ms2013qle}. We showed in \cite{qlebm} that~$d_\CQ$ is a.s.\ determined by ${\mathcal S}$ together with the set $(x_j)$, i.e., it does not depend on any additional randomness associated with the growth processes.
The second paper in the current series \cite{qle_continuity} shows that there is a.s.\ a unique way to continuously extend the metric~$d_\CQ$ on $(x_j)$ to a metric $\ol{d}_\CQ$ defined on the entire sphere~${\mathcal S}$. It is also shown that the identity map from~${\mathcal S}$ (endowed with the Euclidean metric $d$) to itself (endowed with the metric $\ol{d}_\CQ$) is a.s.\ H\"older continuous in both directions. It was further shown in \cite{qle_continuity} that as a random mm-space, the pair $({\mathcal S},\ol{d}_\CQ)$ agrees in law with TBM. Along the way to proving this result, we showed that $({\mathcal S},\ol{d}_\CQ)$ is geodesic, i.e.\ it is a.s.\ the case that every pair of points $x,y$ can be connected by a path whose length with respect to $\ol{d}_\CQ$ is equal to $\ol{d}_\CQ(x,y)$. We recall below three major statements obtained in \cite{qle_continuity}.
In making these statements, as just above, we use $d$ to denote the standard Euclidean path-length metric on ${\mathbf S}^2$ (when ${\mathbf S}^2$ is embedded in $\mathbf{R}^3$ in the standard way).
\begin{theorem}[\cite{qle_continuity}]
\label{thm::continuity}
Suppose that ${\mathcal S} = ({\mathbf S}^2,h)$ is a unit area $\sqrt{8/3}$-LQG sphere, $(x_n)$ is an i.i.d.\ sequence chosen from the quantum measure on ${\mathcal S}$, and $d_\CQ$ is the associated ${\rm QLE}(8/3,0)$ metric on $(x_n)$. Then $(x_i,x_j) \mapsto d_\CQ(x_i,x_j)$ is a.s.\ H\"older continuous with respect to the metric $d^2( (x_1, y_1), (x_2, y_2) ) = d(x_1, x_2) + d(x_2,y_2)$ that $d$ induces on the product space ${\mathbf S}^2 \times {\mathbf S}^2$. In particular, $d_\CQ$ uniquely extends to a H\"older continuous (w.r.t.\ $d^2$) function $\ol{d}_\CQ$ on ${\mathbf S}^2 \times {\mathbf S}^2$.
\end{theorem}
The following states that $\ol{d}_\CQ$ a.s.\ induces a metric on ${\mathbf S}^2$ which is isometric to the metric space completion of $d_\CQ$.
\begin{theorem}[\cite{qle_continuity}]
\label{thm::metric_completion}
Suppose that ${\mathcal S} = ({\mathbf S}^2,h)$ is a unit area $\sqrt{8/3}$-LQG sphere and that $\ol{d}_\CQ$ is as in Theorem~\ref{thm::continuity}. Then $\ol{d}_\CQ$ a.s.\ defines a metric on ${\mathbf S}^2$ which is isometric to the metric space completion of $d_\CQ$. Moreover, the identity map between $({\mathbf S}^2, d)$ and $({\mathbf S}^2, \overlined_\CQ)$ is a.s.\ H\"older continuous in both directions.
\end{theorem}
Throughout the rest of the paper, in order to lighten the notation, we will write $d_\CQ$ instead of $\ol{d}_\CQ$. We will also make use of the following notation. We will write $\ball{z}{\epsilon}$ for the open Euclidean ball centered at $z$ of radius $\epsilon$ and write $\qball{z}{\epsilon}$ for the ball with respect to $d_\CQ$. When we have fixed a given reference point $w$, we will write $\qhull{z}{\epsilon}$ for the hull of $\qball{z}{\epsilon}$ relative to $w$. That is, $\qhull{z}{\epsilon}$ is the set of points disconnected from $w$ by $\qball{z}{\epsilon}$. Typically, $w = \infty$. For a set $K$, we also let $\mathrm{diam}_\CQ(K)$ be the diameter of $K$ with respect to $d_\CQ$.
\begin{theorem}[\cite{qle_continuity}]
\label{thm::tbm_lqg}
Suppose that ${\mathcal S} = ({\mathbf S}^2,h)$ is a unit area $\sqrt{8/3}$-LQG sphere and that $d_\CQ$ is as in Theorem~\ref{thm::continuity}. Then the law of the random mm-space $({\mathbf S}^2, d_\CQ,\mu_h)$ is the same as the law of TBM. If ${\mathcal S} = (\mathbf{D},h)$ is instead a unit boundary length $\sqrt{8/3}$-LQG disk, then the law of $(\mathbf{D},d_\CQ,\mu_h)$ is the same as that of a Brownian disk with unit boundary length. Finally, if ${\mathcal S} = (\C,h)$ is a $\sqrt{8/3}$-LQG cone, then the law of $(\C,d_\CQ,\mu_h)$ is the same as the law of TBP.
\end{theorem}
It is also deduced in \cite{gm2016uihpq} that the mm-space structure associated with a weight-$2$ quantum wedge agrees in law with that of the so-called Brownian half-plane, in which it is also deduced from \cite{bettinelli_miermont} that uniformly random quadrangulations of the upper half-plane converge to the Brownian half-plane. (See also \cite{bmr2016classification} for more on the Brownian half-plane, as well a more general set of convergence results of this type.) The work \cite{gm2016uihpq} is part of a series of papers which also includes \cite{gm2016saw,gm2016gluing} and identifies the scaling limit of the self-avoiding walk on random quadrangulations with ${\rm SLE}_{8/3}$ on $\sqrt{8/3}$-LQG.
\subsection{Main result of current paper}
\label{subsec::main_results}
Theorem~\ref{thm::tbm_lqg} implies that there exists a coupling of the law of a $\sqrt{8/3}$-LQG unit area quantum sphere ${\mathcal S}$ and an instance ${\mathcal S}_{{\mathrm {BM}}} = (M,d,\mu)$ of TBM such that the mm-space $({\mathbf S}^2,d_\CQ,\mu_h)$ associated with ${\mathcal S}$ is a.s.\ isometric to $(M,d,\mu)$. Moreover, by the construction of $d_\CQ$ given in \cite{qlebm} we have that $({\mathbf S}^2,d_\CQ,\mu_h)$ and hence $(M,d,\mu)$ is a.s.\ determined by ${\mathcal S}$. That is, ${\mathcal S}_{{\mathrm {BM}}}$ is a measurable function of ${\mathcal S}$. Moreover, the analogous facts also hold in the setting of the Brownian disk, TBP, and the Brownian half-plane. As mentioned above, the main result of this paper is that the converse also holds.
\begin{theorem}
\label{thm::map_determines_embedding}
Suppose that ${\mathcal S} = ({\mathbf S}^2,h)$ is a $\sqrt{8/3}$-LQG unit area sphere, that $d_\CQ$ is as in Theorem~\ref{thm::continuity}, that ${\mathcal S}_{{\mathrm {BM}}}$ is the instance of TBM described by $({\mathbf S}^2, d_\CQ,\mu_h)$, and that $\phi$ is the identity homeomorphism from ${\mathcal S}$ to ${\mathcal S}_{{\mathrm {BM}}}$. Then ${\mathcal S}$ and $\phi$ are a.s.\ determined by ${\mathcal S}_{{\mathrm {BM}}}$. In particular, this means that ${\mathcal S}$ viewed as a random variable taking values in the space of quantum surfaces is a.s.\ determined by ${\mathcal S}_{{\mathrm {BM}}}$ viewed as a random variable taking values in the space of mm-spaces.
The analogous statement holds in the setting of the coupling of the Brownian disk and a quantum disk, TBP and a quantum cone, and the Brownian half-plane and a weight-$2$ quantum wedge.
\end{theorem}
\subsection{Proof strategy}
\label{subsec::outline}
Our proof of Theorem~\ref{thm::map_determines_embedding} is a variant of the proof given in
\cite[Section~10]{dms2014mating} for the fact that a certain pair of continuum random trees determines the conformal structure of the surface obtained by gluing those trees together. Indeed, the reader might wish to read \cite[Section~10]{dms2014mating} before reading the current paper. The goal of this paper is essentially the same as the goal in \cite[Section~10]{dms2014mating} except that the pair of trees in question is defined in a different way --- in terms of the coordinates of the head of a Brownian snake process, instead of the coordinates of a correlated two dimensional Brownian motion (see Section~\ref{sec::preliminaries}). In what follows, we will give an overview of the strategy to prove this result, which parallels the strategy used in \cite[Section~10]{dms2014mating}.
In both stories, we {\em first} give a procedure for constructing a quantum surface decorated by the pair of continuum trees (and a space-filling path forming an ``interface'' between the two trees) and {\em subsequently} show that in this construction, the pair of trees --- viewed simply as a coupled pair of random metric trees --- actually determine the conformal structure. In both stories, the basic idea is to show that if one {\em conditions} on the pair of trees and resamples the conformal structure (conditioned on the pair of trees) then it is a.s.\ the case that nothing changes, i.e., the conformal structure is a.s.\ the same as before.
In both \cite[Section~10]{dms2014mating} and the current paper, one begins by fixing a value of $\gamma \in (0,2)$ and considering a particular kind of infinite volume $\gamma$-LQG surface, namely the so-called \emph{$\gamma$-quantum cone}. We recall that in a $\gamma$-quantum cone, the origin looks like a ``typical'' point from the quantum measure (see Section~\ref{sec::preliminaries}).
In \cite[Section~10]{dms2014mating}, the value of $\gamma \in (0,2)$ is arbitrary and the curve $\eta'$ is an independent space-filling form of ${\rm SLE}_{16/\gamma^2}$ \cite{MS_IMAG4} that forms the interface between two continuum random trees; it is possible to make sense of those two trees as metric spaces, and the contour functions describing those trees are a pair of Brownian processes $(X, Y)$ that are correlated to an extent that depends on $\gamma$.
In the current setting, we fix $\gamma = \sqrt{8/3}$, and the curve is the interface between a continuum geodesic tree and its dual tree. As explained in \cite{qlebm, qle_continuity}, one may endow an instance of the $\sqrt{8/3}$-quantum cone (with parameter given by $\gamma = \sqrt{8/3}$) with a metric $d_\CQ$ such that the resulting metric space has the law of TBP, as defined in \cite{cl2012brownianplane}. We recall that it is a.s.\ the case that for almost all points in TBP, there is a unique geodesic from that point to the special point ``at $\infty$'' in TBP and that for almost all pairs of points, the geodesics to $\infty$ merge in finite time, so that one has a continuum geodesic tree. The law of this tree and its dual tree can be described directly. First, the dual tree ${\mathcal T}$ is an infinite continuum random tree (CRT) \cite{ald1991crt1,ald1991crt2,ald1993crt3} and has a contour function given by a Brownian motion $X$ indexed by $\mathbf{R}$. Let $\rho_{\mathcal T}$ be the associated map from $\mathbf{R}$ to ${\mathcal T}$. The geodesic tree then has a contour function given by a process $Y$, whose conditional law given $X$ is determined by the fact that $Y_{\rho_{\mathcal T}^{-1}(s)}$ is a Brownian motion indexed by $s \in {\mathcal T}$ (and its definition doesn't depend on which inverse of $\rho_{{\mathcal T}}$ one uses, i.e., $Y_a = Y_b$ whenever $\rho_{\mathcal T}(a) = \rho_{{\mathcal T}}(b)$). The map $\Gamma$ from a real time $t$ to the corresponding point in TBP is a space-filling curve, indexed by $t \in \mathbf{R}$, that fills all of TBP and traces a unit of area in a unit of time. (We will review the definition of TBP more carefully in Section~\ref{subsec::map_plane}.)
In both \cite[Section~10]{dms2014mating} and the current paper, we parameterize this quantum cone by the whole plane $\C$ using the ``smoothed canonical embedding'' described in \cite[Section~10]{dms2014mating}. We recall that in this embedding, the origin and $\infty$ correspond to distinguished marked points of the quantum cone, the rotation is uniformly random, and the scaling is chosen in such a way that $(h, \phi) = 0$, where $\phi$ is a particular smooth rotationally invariant function from $\C$ to $\mathbf{R}$.
We emphasize that the joint law of the pair $(h,\Gamma)$ of the present paper is very different from the joint law of the pair $(h,\eta')$ from \cite[Section~10]{dms2014mating}. Indeed, in the setting of \cite[Section~10]{dms2014mating} we have that $\eta'$ is a space-filling ${\rm SLE}_{16/\gamma^2}$ which is first sampled independently of $h$ and then reparameterized so that it fills one unit of quantum area in each unit of time. In contrast, the construction of $d_\CQ$ implies that the curve $\Gamma$ is a.s.\ determined by $h$. Moreover, $\Gamma$ serves to encode the metric structure associated with $h$ while $\eta'$ should be thought of as encoding the scaling limit of a statistical physics model (e.g., a UST instance) on a random planar map.
Let ${\mathcal U} = \Gamma([0,1])$ denote the unit area portion of TBP traced by $\Gamma$ during the time interval $[0,1]$. In light of the embedding described above, we may also view ${\mathcal U}$ as a subset of $\C$. Now fix some large $n \in \mathbf{N}$ and for $k \in \{1,\ldots,n\}$, let ${\mathcal U}_k = \Gamma([(k-1)/n,k/n])$ denote the portion traced during the interval $[(k-1)/n, k/n]$.
\begin{lemma}
\label{lem::brownianplaneradiusbound}
There exists a constant $c > 0$ such that the following is true. For each $n \in \mathbf{N}$ and $k \in \{1,\ldots,n\}$, the probability that the diameter of ${\mathcal U}_k$ with respect to the metric of TBP exceeds $a n^{-1/4}$ decays at least as fast as $e^{- c a^{4/3}}$.
\end{lemma}
Lemma~\ref{lem::brownianplaneradiusbound} (stated as Proposition~\ref{prop::path_chunk_diameter_bound} and proved below) is a corollary of a large deviations principle for the Brownian snake established in \cite{serlet_ldp} which states that if~$d^*$ is the diameter of TBM with unit area then $-\log {{\bf P}}[d^* \geq r]$ behaves like a constant times~$r^{4/3}$ as $r \to \infty$. Lemma~\ref{lem::brownianplaneradiusbound} also implies that for each $\epsilon > 0$ fixed, the probability that even one of the ${\mathcal U}_k$ has diameter larger than $n^{-1/4 + \epsilon}$ tends to zero superpolynomially as a function of $n$. For each $k,n$, let $\widetilde {\mathcal U}_k$ denote the filled metric ball of radius $n^{-1/4 + \epsilon}$ centered at $\Gamma(k/n)$.
\begin{lemma}
\label{lem::momentbound}
There exists constants $c > 0$, $\alpha > 4$, and a positive probability event ${\mathcal A}$ which depends only on the quantum surface parameterized $\C \setminus {\mathcal U}$ such that the following is true. For each $n \in \mathbf{N}$ and $k \in \{1,\ldots,n\}$, we have that
\[ {\bf E}[ \mathop{\mathrm{diam}}(\widetilde{{\mathcal U}}_k)^4 {\mathbf 1}_{{\mathcal A}} ] \leq c n^{-\alpha}.\]
\end{lemma}
Lemma~\ref{lem::momentbound} is stated carefully as Lemma~\ref{lem::metric_ball_stable_moment_bound} and proved below.
Suppose that $K \subseteq \C$ is a compact set. Recall that~$K$ is said to be \emph{conformally removable} if the following is true. Suppose that $U,V \subseteq \C$ are open subsets with $K \subseteq U$. If $\varphi \colon U \to V$ is a homeomorphism which is conformal on $U \setminus K$ then~$\varphi$ is conformal on~$U$. The notion of conformal removability is important in the context of LQG, see e.g.\ \cite{SHE_WELD}.
\begin{lemma}
\label{lem::removableunion}
For each $n \in \mathbf{N}$, the set $\cup_{k=1}^n \partial \widetilde {\mathcal U}_k$ is a.s.\ conformally removable.
\end{lemma}
Lemma~\ref{lem::removableunion} follows from Theorem~\ref{thm::finite_unions_removable} which is stated and proved in Appendix~\ref{app::removability}. We recall that a simply connected domain $U \subseteq \C$ is said to be a \emph{H\"older domain} if there exists a conformal map $\varphi \colon \mathbf{D} \to U$ which is H\"older continuous up to $\partial \mathbf{D}$. Establishing Lemma~\ref{lem::removableunion} will require us to rewrite, in a slightly more general way, some of the arguments for the conformal removability of H\"older domain boundaries that appear in \cite{js2000remove}. (See Appendix~\ref{app::removability}.)
\begin{lemma}
\label{lem::insideoutsideindependent}
Given the mm-space structure of TBP, the conformal structure associated with the inside and the outside of a filled metric ball of a fixed radius centered at the origin are conditionally independent.
\end{lemma}
Lemma~\ref{lem::insideoutsideindependent} is stated carefully as Lemma~\ref{lem::inside_outside_independent} below.
Once we have the four lemmas above, Theorem~\ref{thm::map_determines_embedding} will follow from the argument used to prove the rigidity of the conformal structure given the peanosphere structure in \cite[Section~10]{dms2014mating}.
\subsection{Outline}
We will review basic definitions in Section~\ref{sec::preliminaries} and provide the proof of Theorem~\ref{thm::map_determines_embedding} in Section~\ref{sec::map_determines_field}. Like the other papers in this series, this paper builds on several previous works by the current authors about LQG surfaces \cite{SHE_WELD, dms2014mating, quantum_spheres}, TBM \cite{map_making}, and imaginary geometry \cite{MS_IMAG,MS_IMAG2,MS_IMAG3,MS_IMAG4}. These in turn rely on sizable literatures on SLE, TBM, LQG and other topics, which are briefly reviewed in the first paper of the current series \cite{qlebm}.
\section{Preliminaries}
\label{sec::preliminaries}
In this section we will review a few preliminaries, including background on mm-spaces (Section~\ref{subsec::mm_spaces}), quantum surfaces (Section~\ref{subsec::quantum_surfaces}), a brief review of the Brownian map and plane (Section~\ref{subsec::map_plane}), and finally a few basic estimates that will be used in the proofs of our main theorems (Section~\ref{subsec::external_estimates}).
\subsection{Metric measure spaces}
\label{subsec::mm_spaces}
A \emph{metric measure space} (mm-space) is a metric space $(M,d)$ together with a measure $\mu$ on the Borel $\sigma$-algebra associated with $(M,d)$. In this article, we will be considering a number of \emph{random} mm-spaces and we will also want to consider conditional probabilities where the $\sigma$-algebra in question is generated by such a random mm-space. In order to make this precise, we need to specify a $\sigma$-algebra on mm-spaces. There are several possibilities that one could choose from here. In order to be consistent with \cite{map_making}, we will use the Borel $\sigma$-algebra generated by the so-called Gromov-weak topology. (See \cite[Section~2.4]{map_making} and the references therein for additional detail.)
Suppose that $(M,d,\mu)$ is a mm-space with $\mu$ a probability measure and let ${\bf E}_\mu$ denote the expectation associated with the matrix of distances $d_{ij} = d(x_i,x_j)$ where $(x_j)$ is an i.i.d.\ sequence sampled from $\mu$. The Gromov-weak topology is the weakest topology such that for each $k \in \mathbf{N}$ and each bounded, continuous function $\psi$ on $\mathbf{R}^{k^2}$ we have that the map
\[ (M,d,\mu) \mapsto {\bf E}_\mu[ \psi(M_k)]\]
is measurable where $M_k = (d_{ij})_{i,j=1}^k$.
More generally, the Gromov-weak topology on mm-spaces $(M,d,\mu,z_1,\ldots,z_n)$ with $n$ marked points is defined in the same way except we take the sequence $(x_j)$ so that $x_1=z_1,\ldots,x_n=z_n$ and the remaining elements of the sequence to be i.i.d.\ from $\mu$. We refer the reader to \cite{grevenpfaffelhuberwinter} for additional background on the Gromov-weak topology.
We let ${\mathcal M}$ be the Borel $\sigma$-algebra generated by the Gromov-weak topology. More generally, we let ${\mathcal M}^n$ be the Borel $\sigma$-algebra generated by the Gromov-weak topology with $n$ marked points.
We note that, using this topology, two mm-spaces $(M_i,d_i,\mu_i)$ for $i=1,2$ are equivalent if there exists a measure preserving map $\psi \colon M_1 \to M_2$ which restricts to an isometry on the support of $\mu_1$.
In \cite[Section~2.4]{map_making}, a number of properties of the Gromov-weak topology are recorded. For example, it is shown there that the following events are Gromov-weak measurable:
\begin{itemize}
\item $(M,d)$ is compact (more precisely this means that $(M,d,\nu)$ is equivalent to some $(\tilde M, \tilde d, \tilde \nu)$ for which $(\tilde M, \tilde d)$ is compact) \cite[Proposition~2.12]{map_making},
\item $(M,d)$ is compact and geodesic \cite[Proposition~2.14]{map_making}, and
\item $(M,d)$ is geodesic and homeomorphic to ${\mathbf S}^2$ \cite[Proposition~2.15]{map_making}.
\end{itemize}
We note that TBM with unit area can be viewed as a random variable taking values in the space of mm-spaces using the $\sigma$-algebra ${\mathcal M}$ as described just above.
We finish this subsection with two observations:
\begin{itemize}
\item If $(M,d,\mu)$ is an mm-space where $0 < \mu(M) < \infty$, then we can let $\mu^* = \mu/\mu(M)$ and encode $(M,d,\mu)$ by the pair consisting of $(M,d,\mu^*)$ and $\mu(M)$. Therefore such a space fits into the aforementioned framework, where we add an extra variable to keep track of the total mass $\mu(M)$ and the natural $\sigma$-algebra in this context is the product $\sigma$-algebra associated with ${\mathcal M}$ and the Borel $\sigma$-algebra on~$\mathbf{R}$. We note that TBM with random area can be viewed as a random variable taking values in the product of the space of mm-spaces as described above and~$\mathbf{R}$ with this $\sigma$-algebra.
\item Suppose that $(M,d,\mu,z)$ is a marked mm-space where $\mu(M) = \infty$ but $0 < \mu(B(z,r)) < \infty$ for all $r > 0$ and $M = \cup_{r > 0} B(z,r)$. For each $r > 0$, let $\mu_r^*$ be given by $\mu|_{B(z,r)} / \mu(B(z,r))$. For $r > 0$, each of the spaces $(B(z,r),d,\mu_r^*,z)$ together with $\mu(B(z,r))$ fits into the framework described just above and we can view such a space $(M,d,\mu,z)$ as a sequence $(B(z,n),d,\mu_n^*,z)$ together with $\mu(B(z,n))$ each of which fits into the framework from just above. This leads to a natural $\sigma$-algebra in this setting and this is the space in which we will view TBP as taking values.
\end{itemize}
\subsection{Quantum surfaces}
\label{subsec::quantum_surfaces}
Suppose that $h$ is an instance of some form of the Gaussian free field (GFF) on a planar domain $D$ and fix $\gamma \in (0,2)$. The $\gamma$-Liouville quantum gravity (LQG) measure associated with $h$ is formally given by $\mu_h = e^{\gamma h} dz$ where $dz$ denotes Lebesgue measure on $D$. Since $h$ is a distribution and not a function, this expression requires interpretation and can be made rigorous using a regularization procedure. Similarly, if $L$ is a linear segment of the boundary, then the $\gamma$-LQG boundary measure is given by $\nu_h = e^{\gamma h/2} dz$ where here $dz$ denotes Lebesgue measure on $L$. This expression is made rigorous using the same regularization procedure used to construct $\mu_h$.
Suppose that $\widetilde{D}$ is another planar domain and $\varphi \colon \widetilde{D} \to D$ is a conformal transformation. If one defines
\begin{equation}
\label{eqn::quantum_equivalence}
\widetilde{h} = h \circ \varphi + Q \log|\varphi'| \quad\text{where}\quad Q = \frac{2}{\gamma} + \frac{\gamma}{2}
\end{equation}
then $\mu_{\widetilde{h}}(A) = \mu_h(\varphi(A))$ (resp.\ $\nu_{\widetilde{h}}(A) = \nu_h(\varphi(A))$) for all $A \subseteq \widetilde{D}$ (resp.\ $A \subseteq \partial \widetilde{D}$) Borel. This in particular allows one to make sense of $\nu_h$ on segments of $\partial D$ which are not necessarily linear. A quantum surface is an equivalence class where a representative consists of a pair $(D,h)$ and two pairs $(D,h)$, $(\widetilde{D},\widetilde{h})$ are said to be equivalent if~$h$ and~$\widetilde{h}$ are related as in~\eqref{eqn::quantum_equivalence}. We refer to a representative $(D,h)$ of a quantum surface as an \emph{embedding}.
More a generally, a \emph{marked quantum surface} is an equivalence class where a representative consists of a triple $(D,h,\underline{z})$ where $\underline{z} = (z_1,\ldots,z_k)$ is a collection of points in $\overline{D}$ and two marked quantum surfaces $(D,h,\underline{z})$, $(\widetilde{D},\widetilde{h},\underline{\widetilde{z}})$ are said to be equivalent if $h$, $\widetilde{h}$ are related as in~\eqref{eqn::quantum_equivalence} and $\varphi(\widetilde{z}_j) = z_j$ for all $1 \leq j \leq k$.
In this work, we shall be primarily interested in several types of quantum surfaces: quantum wedges, cones, disks, and spheres. As the definitions of these surfaces were reviewed in earlier works in this series, we will not repeat them here and instead refer the reader to \cite{dms2014mating}.
\subsection{The Brownian map and plane}
\label{subsec::map_plane}
We now recall the definition of the Brownian map (TBM) \cite{lg2013uniqueness,mier2013bm} and the Brownian plane (TBP) \cite{cl2012brownianplane}. We first begin by reminding the reader of the construction of the so-called Brownian snake process $Y$ on $[0,1]$. One can sample from the law of the Brownian snake using the following two-step procedure:
\begin{enumerate}
\item Sample a Brownian excursion $X \colon [0,1] \to \mathbf{R}_+$ with $X(0) = X(1) = 0$ and let
\begin{equation}
\label{eqn::m_e_def}
m_X(s,t) = \inf_{u \in [s,t]} X(u).
\end{equation}
\item Given $X$, sample a mean-zero Gaussian process $Y$ with covariance given by $\cov{Y_s}{Y_t} = m_X(s,t)$.
\end{enumerate}
TBM (with unit area) is the random mm-space which is constructed from the Brownian snake as follows. For $s,t \in [0,1]$, we set
\[ d^\circ(s,t) = Y_s + Y_t - 2\max\left( \inf_{u \in [s,t]} Y_u,\inf_{u \in [t,s]} Y_u \right).\]
Here, we assume without loss of generality that $s < t$ and take $[t,s] = [0,s] \cup [t,1]$. Let ${\mathcal T}$ be the instance of the continuum random tree (CRT) associated with $X$ and let $\rho_{\mathcal T} \colon [0,1] \to {\mathcal T}$ be the projection map. For $a,b \in {\mathcal T}$, we then set
\begin{equation}
\label{eqn::d_circ_def}
d_{\mathcal T}^\circ(a,b) = \inf\{ d^\circ(s,t) : \rho_{\mathcal T}(s) = a, \rho_{\mathcal T}(t) = b\}.
\end{equation}
Finally, for $a,b \in {\mathcal T}$, we set
\begin{equation}
\label{eqn::d_metric_def}
d(a,b) = \inf\left\{ \sum_{j=1}^k d_{\mathcal T}^\circ(a_{j-1},a_j) \right\}
\end{equation}
where the infimum is taken over all $k \in \mathbf{N}$ and $a_0 = a,a_1,\ldots,a_k=b$ in ${\mathcal T}$. The unit-area Brownian map is the mm-space $(M,d,\mu)$ where $(M,d)$ is given by the metric quotient ${\mathcal T} / \cong$ where $a \cong b$ if and only if $d(a,b) = 0$ and $\mu$ is the pushforward of Lebesgue measure from $[0,1]$ to $(M,d)$ under the projection map $\rho \colon [0,1] \to M$. This construction naturally associates with $(M,d,\mu)$ a space-filling, non-crossing path $\Gamma$ where $\Gamma(t) = \rho(t)$ for $t \in [0,1]$. (In fact, $\Gamma$ is the peano curve which snakes between a tree of geodesics and its dual tree on $(M,d)$.) In particular, $(M,d,\mu)$ is naturally marked by $\Gamma(0)$.
Recalling that $Y_s$ represents the distance between $\rho(s)$ and the root of the geodesic tree in TBM, i.e., $\rho(s^*)$ where $s^*$ is the unique value in $[0,1]$ such that $Y_{s^*} = \inf_{s \in [0,s]} Y_s$, it follows that the diameter $d^*$ of $(M,d)$ satisfies the bounds
\[ \sup_{s \in [0,1]} Y_s - \inf_{s \in [0,1]} Y_s \leq d^* \leq 2\left(\sup_{s \in [0,1]} Y_s - \inf_{s \in [0,1]} Y_s\right).\]
The precise asymptotics for the tail of $d^*$ were determined by Serlet \cite[Proposition~14]{serlet_ldp}. We restate a variant of this result here because it will be important for our later arguments. (The result in \cite{serlet_ldp} in fact includes matching upper and lower bounds.)
\begin{proposition}[\cite{serlet_ldp}]
\label{prop::serlet}
There exists a constant $c_0 > 0$ such that, as $r \to \infty$, we have that
\begin{equation}
\label{eqn::serlet_ldp}
\log {{\bf P}}[ d^* \geq r] \leq -c_0 r^{4/3}.
\end{equation}
\end{proposition}
We observe that if we have a marked instance of TBM $(M,d,\mu,\Gamma(0))$, then for each $t \in [0,1]$ fixed we have that $(M,d,\mu,\Gamma(0)) \stackrel{d}{=} (M,d,\mu,\Gamma(t))$. That is, the law of TBM is invariant under re-rooting. The reason for this is that the CRT instance itself is invariant under re-rooting as is the corresponding Brownian snake process. In fact, the conditional law of $\Gamma(0)$ given $(M,d,\mu)$ is equal to $\mu$.
In this work, we will be interested in TBP \cite{cl2012brownianplane}. We will now remind the reader of the construction of TBP and explain why Serlet's bound can be used to deduce diameter bounds for the corresponding space-filling path.
Following \cite{cl2012brownianplane}, one can produce a sample from the law of TBP as follows:
\begin{enumerate}
\item Sample the process $X \colon \mathbf{R} \to \mathbf{R}_+$ by taking $X_t = R_t$ for $t \geq 0$ (resp.\ $X_t = R_{-t}'$ for $t \leq 0$) where $R,R'$ are independent $3$-dimensional Bessel processes starting from~$0$.
\item Given~$X$, let~$Y$ be the mean-zero Gaussian process with covariance $\cov{Y_s}{Y_t} = m_X(s,t)$ where $m_X$ is as in~\eqref{eqn::m_e_def} (with $X$ as in the previous item).
\end{enumerate}
We note that the two $3$-dimensional Bessel processes in the definition of $X$ correspond to zooming in near the start and end of the Brownian excursion in the definition of TBM.
The mm-space structure of TBP is constructed in a manner which is analogous to TBM. Also associated with it is a space-filling path $\Gamma$ which is given by projecting the identity map on $\mathbf{R}$. Moreover, the law of TBP is invariant under re-rooting according to $\Gamma$, as it is the infinite volume limit of TBM which enjoys the corresponding property. See, e.g., \cite[Proposition~4]{cl2012brownianplane}.
We are now going to explain why Serlet's tail bound (Proposition~\ref{prop::serlet}) extends to bound the diameter $d_{s,t}^*$ of the sets $\Gamma([s,t])$ for $s < t$ of the space-filling path on TBP.
\begin{proposition}
\label{prop::path_chunk_diameter_bound}
There exists a constant $c_0 > 0$ such that, as $r \to \infty$, we have that
\begin{equation}
\label{eqn::space_filling_chunk_size}
\log {{\bf P}}[ d_{0,1}^* \geq r] \leq -c_0 r^{4/3}.
\end{equation}
More generally, for each $s < t$ we have that
\begin{equation}
\label{eqn::space_filling_chunk_size_general}
\log {{\bf P}}[ d_{s,t}^* \geq r] \leq -c_0 (t-s)^{-1/3} r^{4/3}.
\end{equation}
\end{proposition}
\begin{proof}
We note that~\eqref{eqn::space_filling_chunk_size_general} follows from~\eqref{eqn::space_filling_chunk_size} and the scaling properties of TBP. In particular, we have that $d_{s,t}^* \stackrel{d}{=} (t-s)^{1/4} d_{0,1}^*$ for all $s < t$. This leaves us to prove~\eqref{eqn::space_filling_chunk_size}.
We will deduce~\eqref{eqn::space_filling_chunk_size} from Proposition~\ref{prop::serlet}. By the definition~\eqref{eqn::d_circ_def}, \eqref{eqn::d_metric_def} of the metric for TBP and the definition of the space-filling path $\Gamma$, we know that
\begin{equation}
\label{eqn::path_diam_bound}
d_{0,1}^* \leq 2 \left(\sup_{s \in [0,1]} Y_s - \inf_{s \in [0,1]} Y_s \right).
\end{equation}
Therefore to extract~\eqref{eqn::space_filling_chunk_size} from~\eqref{eqn::serlet_ldp}, we just need to control the Radon-Nikodym derivative of the law of a $3$-dimensional Bessel process run for time $1/2$ with respect to the law of a Brownian excursion run for time $1/2$. Indeed, this will imply the analog of~\eqref{eqn::path_diam_bound} with $[0,1/2]$ in place of $[0,1]$. The case with $[0,1]$ then follows from the case with $[0,1/2]$ by applying scaling.
We begin by observing that if $R$ is a $3$-dimensional Bessel process starting from $0$ and $X$ is a Brownian excursion from $0$ to $0$, then for each $x \geq 0$ we have that the conditional law of $R|_{[0,1/2]}$ given $R_{1/2} = x$ is equal to the conditional law of $X|_{[0,1/2]}$ given $X_{1/2} = x$. Therefore we just have to control the Radon-Nikodym derivative of the law of $R_{1/2}$ with respect to the law of $X_{1/2}$.
By \cite[Chapter~XII, Theorem~4.1]{RY04} and \cite[Chapter~XI]{RY04} the density of the laws of $X_{1/2}$ and $R_{1/2}$, respectively, with respect to Lebesgue measure on $\mathbf{R}_+$ are given by the following expressions evaluated at $t=1/2$:
\begin{equation}
\label{eqn::densities}
\frac{2 x^2}{(2\pi t^3(1-t)^3)^{1/2}} \exp\left( - \frac{x^2}{2t} - \frac{x^2}{2(1-t)} \right) \quad\text{and}\quad \frac{2 x^2}{(2 \pi t^3)^{1/2}} \exp\left(- \frac{x^2}{2t} \right).
\end{equation}
Therefore the claimed bounds follow by using the explicit form of the Radon-Nikodym derivative and applying H\"older's inequality. More precisely, if $r > 0$ and $A_r$ is the event that the right hand side of~\eqref{eqn::path_diam_bound} is at least $r$, with $[0,1/2]$ in place of $[0,1]$, and $\mu$ (resp.\ $\nu$) denotes the law of TBM (resp.\ TBP), then for conjugate exponents $p,q \geq 1$ we have that
\[ \nu[A_r] = \int Z {\mathbf 1}_{A_r} d \mu \leq \left( \int Z^p d\mu \right)^{1/p} \mu[A_r]^{1/q}\]
where $Z$ is given by the second expression in~\eqref{eqn::densities} divided by the first with $t=1/2$. Proposition~\ref{prop::serlet} gives us the rate of decay for $\mu[A_r]$ as $r \to \infty$; note that taking the $1/q^{\rm th}$ power only changes the constant in front of the term $r^{4/3}$. The result thus follows because by inspecting the explicit form of $Z$, we see that we can choose $p > 1$ sufficiently close to $1$ so that $\int Z^p d\mu < \infty$.
\end{proof}
The boundary of the range $\Gamma([s,t])$ of the space-filling path $\Gamma$ on the interval $[s,t]$ can be divided into four boundary segments: two branches of the geodesic tree (i.e., geodesics) and two branches of the dual tree. These sets are not known to be conformally removable, after embedding into $\C$ using the ${\rm QLE}(8/3,0)$ metric. For this reason, in our proof of Theorem~\ref{thm::map_determines_embedding} we will cover each such chunk by a filled metric ball and instead use the removability of filled metric ball boundaries \cite{ms2013qle}. (This is in contrast to \cite[Section~10]{dms2014mating}, in which it was known that the boundaries of the chunks of the space-filling path were conformally removable.)
\subsection{Variance and distortion estimates}
\label{subsec::external_estimates}
We now record a variance bound and a distortion estimate which will be important for the proof of Theorem~\ref{thm::map_determines_embedding}. The first estimate that we will state is a general bound, the so-called Efron-Stein inequality \cite{efronstein}, on the variance of a function of independent random variables.
\begin{lemma}
\label{lem::variance_bound}
Let $A = A(X_1,\ldots,X_n)$ be a function of $n$ independent random variables $X_1,\ldots,X_n$ such that $\ex{ A^2 } < \infty$. Then
\begin{equation}
\label{eqn::variance_bound}
\var{A} \leq \sum_{i=1}^n \ex{ \var{ A \,|\, {\mathcal F}_i}} \quad\text{where}\quad {\mathcal F}_i = \sigma(X_j : j \neq i).
\end{equation}
\end{lemma}
This result is stated in \cite[Lemma~10.4]{dms2014mating} together with an independent proof.
When we apply Lemma~\ref{lem::variance_bound} below, the role of the $X_i$'s will be played by quantum surfaces parameterized by hulls of metric balls and $A$ will be a function which determines the embedding of the quantum surface formed by the $X_i$'s and the surface formed by the complement of their union. We will compute the variance conditionally on the latter, so that $A$ is just a function of the $X_i$'s.
The other estimate that we will state is a general estimate for conformal maps and is stated in \cite[Lemma~10.5]{dms2014mating}. Before we recall this estimate, we first remind the reader of the following definition. Suppose that $K \subseteq \C$ is a compact hull and let $F \colon \C \setminus \overline{\mathbf{D}} \to \C \setminus K$ be the unique conformal map with $F(\infty) = \infty$ and $F'(\infty) > 0$. Then we can write
\[ F(z) = a_{-1} z + a_0 + \sum_{n=1}^\infty a_n z^{-n}.\]
We will refer to $a_0$ as the \emph{harmonic center} of $K$.
\begin{lemma}
\label{lem::conformal_map_bound}
There exist constants $c_1,c_2 > 0$ such that the following is true. Let $K_1$ be a hull of diameter at most $r$ and $K_2$ another hull such that there exists a conformal map $F \colon \C \setminus K_1 \to \C \setminus K_2$ of the form
\[ F(z) = z + \sum_{n=1}^\infty \alpha_n z^{-n}.\]
Then whenever $\mathop{\mathrm{dist}}(z,K_1) \geq c_1r$ we have that
\[ |F(z) - (z+b_2-b_1)| \leq c_2 r^2|z-b_1|^{-1}\]
where $b_i$ for $i=1,2$ is the harmonic center of $K_i$.
\end{lemma}
In the proof of Theorem~\ref{thm::map_determines_embedding}, Lemma~\ref{lem::conformal_map_bound} will be used to bound the summands in~\eqref{eqn::variance_bound} in terms of the fourth moment of the Euclidean diameters of a collection of quantum metric balls.
\section{Brownian surfaces determine their embedding}
\label{sec::map_determines_field}
The purpose of this section is to prove Theorem~\ref{thm::map_determines_embedding}. We begin in Section~\ref{subsec::quantum_cone_stable} by describing a type of embedding of a $\sqrt{8/3}$-quantum cone that will be useful when we complete the proof of Theorem~\ref{thm::map_determines_embedding} in Section~\ref{subsec::resampling_argument}.
\subsection{Smooth embeddings and stability}
\label{subsec::quantum_cone_stable}
Suppose that $(\C,h,0,\infty)$ is a $\sqrt{8/3}$-quantum cone and let $\mu_h$ be the associated LQG area measure. For each $z \in \C$ and $r >0$, let $h_r(z)$ be the average of $h$ on $\partial B(z,r)$. In many places, it is convenient to take the embedding of the quantum cone so that the supremum of $r > 0$ values at which the process $r \mapsto h_r(0) + Q \log r$, $\gamma=\sqrt{8/3}$ and $Q$ as in~\eqref{eqn::quantum_equivalence}, takes on the value $0$ is equal to $r=1$. This particular choice of embedding in some places is referred to as the circle-average embedding of $h$. It is useful because it is possible to give an explicit description of the law of $h$. When we complete the proof of Theorem~\ref{thm::map_determines_embedding} in Section~\ref{subsec::resampling_argument} below, however, we will need to consider a slightly different embedding of the quantum cone into $\C$. The reason for this is that we will be cutting out various parts of the surface and then gluing in new pieces and the aforementioned normalization is not compatible with this operation as the circles centered at the origin will typically not be invariant under the operation of cutting/gluing.
We are now going to describe an alternative choice of embedding of a quantum cone into $\C$ that we refer to as a \emph{smooth canonical description}. This type of embedding was introduced in \cite[Section~10]{dms2014mating} (and also made use of in \cite{ghms2015correlation}), but for completeness we will recall the definition now. For $t \geq 0$, let
\[ R_t = \frac{1}{\gamma} \log {\bf E}[ \mu_h(\ball{0}{e^{-t}}) \,|\, h_{e^{-t}}(0) ].\]
That is, $R_t$ is chosen so that $e^{\gamma R_t}$ is equal to the expected $\gamma$-LQG mass in $\ball{0}{e^{-t}}$ given $h_{e^{-t}}(0)$. Then $R_t = B_t + (\gamma-Q) t$ where $B_t$ for $t > 0$ is a standard Brownian motion \cite[Section~4]{DS08}; note that $\gamma-Q < 0$. Fix $\epsilon > 0$ very small. Now we will consider the quantity
\begin{equation}
\label{eqn::smooth_embedding_function}
\int_{-\epsilon}^0 R_t \phi(t) dt
\end{equation}
where $\phi \geq 0$ is a $C^\infty$ bump function supported on $(-\epsilon,0]$ with total integral one. We apply a coordinate change rescaling to~$h$ so that the function from~\eqref{eqn::smooth_embedding_function} achieves the value~$0$ for the first time at $0$. We refer to the $h$ with this type of scaling as the \emph{smooth canonical description} for the quantum cone.
When we perform the cutting/gluing operations in Section~\ref{subsec::resampling_argument} below, it will be important for us to work on a certain event on which the behavior of certain quantities is uniform. Moreover, it will be important that this event is determined by the quantum surface in question and is invariant under performing certain cutting/gluing operations. We will now make the definition of this event precise. Fix $s > 0$ and $r > 1$. Suppose that $(\C,h,0,\infty)$ is a $\sqrt{8/3}$-quantum cone and let $\Gamma$ be the space-filling path associated with TBP structure of $h$. We say that $h$ is \emph{$(r,s)$-stable} if the following are true.
\begin{itemize}
\item $\Gamma([-r^{-1},r^{-1}]) \subseteq \qhull{0}{s} \subseteq B(0,\tfrac{1}{2})$
\item For every $z \in \qhull{0}{s}$ it is the case that if we translate by $-z$, the scaling factor which puts the resulting field into a smooth canonical description is between $r^{-1}$ and $r$.
\item The previous items continue to hold if we cut out the surface parameterized by $\qhull{0}{s}$ and then weld in any surface which preserves the overall mm-space structure. In other words, if we consider any conformal map $\psi$ from $\C \setminus \qhull{0}{s}$ to $\C \setminus K$ for some hull $K \subseteq \C$ that is normalized so that
\begin{enumerate}[(i)]
\item $\psi$ fixes $\infty$,
\item $\psi$ has positive real derivative at $\infty$ (i.e., $\lim_{z \to \infty} \psi(z)/z > 0$),
\item $\psi$ is scaled in such a way that the pushforward of $h$ via the quantum coordinate change described by $\psi$ corresponds to a smooth canonical description
\end{enumerate}
then $K \subseteq B(0,\tfrac{1}{2})$. Moreover, for any $z \in K$, the scaling factor necessary to put the field into a smooth canonical description is between $r^{-1}$ and $r$.
\end{itemize}
A notion of stability is also introduced in \cite[Section~10.4.2]{dms2014mating}. The definition above differs from that in \cite{dms2014mating} in that the above is defined in terms of the mm-space structure of the surface while the version in \cite{dms2014mating} is focused on the behavior of the space-filling ${\rm SLE}$.
We are now going to show that by adjusting the parameters $r$ and $s$, the probability of the event that a $\sqrt{8/3}$-quantum cone is $(r,s)$-stable can be made arbitrarily close to $1$.
\begin{proposition}
\label{prop::stable_positive_probability}
Let $(\C,h,0,\infty)$ be a smooth canonical description of a $\sqrt{8/3}$-quantum cone. Fix $s > 0$, $r > 1$, and let ${\mathcal A}_{r,s}$ be the event that $h$ is $(r,s)$-stable. Then $\pr{{\mathcal A}_{r,s}} > 0$. Moreover, for every $\epsilon > 0$ there exists $s_0 > 0$ such that for all $s \in (0,s_0)$ there exists $r_0 > 1$ such that $r \geq r_0$ implies that $\pr{{\mathcal A}_{r,s}} \geq 1-\epsilon$.
\end{proposition}
\begin{proof}
The proof is similar to that of \cite[Proposition~10.16]{dms2014mating}, but we will include it here for completeness (though we will cite some of the results from \cite[Section~10.4.2]{dms2014mating} that we make use of verbatim).
We let $\Phi$ be the set of distorted bump functions of the form $|g'|^2 \phi \circ g$ for which $g^{-1}$ is conformal outside of $K$ for some $K \subseteq \ball{0}{\tfrac{1}{10}}$ and looks like the identity near $\infty$. Then we know that $\Phi$ is a sequentially compact subset of the space of test functions with respect to the topology of uniform convergence of all derivatives on compact sets. Indeed, this follows from the Arzel\'a-Ascoli theorem. Since this space is a Fr\'echet space (thus metrizable) $\Phi$ is also compact. In particular, for any distribution $h$, both of the quantities
\[ M_1(h) = \inf_{\widetilde{\phi} \in \Phi} (h, \phi - \widetilde{\phi}) \quad\text{and}\quad M_2(h) = \sup_{\widetilde{\phi} \in \Phi} (h,\phi-\widetilde{\phi})\]
are finite. By continuity, they are equal to the infimum and supremum taken over a countable dense subset of $\widetilde{\phi} \in \Phi$. In the case that $h$ is given by a GFF, the value of $(h,\phi-\widetilde{\phi})$ is a continuous function of $\widetilde{\phi}$ hence also has a maximum over $\Phi$.
Let $R_t^j = M_j(z \mapsto h(e^t z) + Qt)$. By \cite[Proposition~10.19]{dms2014mating}, both of the processes $R_t^1$ and $R_t^2$ tend to $\infty$ as $t \to \infty$ and to $-\infty$ as $t \to -\infty$. This implies that there a.s.\ exists $c \in (0,\infty)$ (random) such that we have $R_t^1, R_t^2 > 0$ for $t > c$ and $R_t^1, R_t^2 < 0$ for $t < -c$. Pick $c_1 \in (0,\infty)$ such that with
\begin{equation}
\label{eqn::p_c1_c}
E_1 = \{c \leq c_1\} \quad\text{we have}\quad \pr{E_1} \geq 1-\epsilon.
\end{equation}
If $K \subseteq \ball{0}{\tfrac{1}{2} e^{-c_1}}$ then we have that $a K \in \ball{0}{\tfrac{1}{2}}$ for any $a \in [0,e^{c_1}]$.
We know that $\qhull{0}{s} \subseteq \ball{0}{\tfrac{1}{2} e^{-c_1}}$ with positive probability for any value of $s > 0$. In fact, we have that for any $\epsilon > 0$ there exists $s_0 >0$ such that with
\begin{equation}
\label{eqn::p_c_c}
E_2 = \{ \qhull{0}{s_0} \subseteq \ball{0}{\tfrac{1}{2} e^{-c_1}} \} \quad\text{we have}\quad \pr{ E_2} \geq 1-\epsilon.
\end{equation}
On $E_1,E_2$, it is not hard to check that if one swaps out $\qhull{0}{s_0}$ with any other quantum surface, then the resulting appropriately scaled surface will still belong to $\ball{0}{\tfrac{1}{2}}$.
For any fixed value of $s \in (0,s_0)$, follows from Proposition~\ref{prop::path_chunk_diameter_bound} that there exists $r_0 > 0$ such that $r \geq r_0$ implies that with
\begin{equation}
\label{eqn::p_k_r}
E_3 = \{ \mathrm{diam}_\CQ(\Gamma([-r^{-1},r^{-1}])) \leq s\} \quad\text{we have}\quad \pr{E_3} \geq 1-\epsilon.
\end{equation}
On $E_1 \cap E_2 \cap E_3$, all of the properties for the configuration to be $(r,s)$-stable hold except for possibly the part of the event mentioned in the second bullet point. One can verify that this condition also holds with probability at least $1-\epsilon$ using the same argument which was used to establish~\eqref{eqn::p_c1_c}.
\end{proof}
\begin{lemma}
\label{lem::metric_ball_stable_moment_bound}
There exists a constant $c_0 > 0$ such that the following is true. Let $(\C,h,0,\infty)$ be a smooth canonical description of a $\sqrt{8/3}$-quantum cone. Fix $s \in (0,1)$, $r > 1$, and let ${\mathcal A}_{r,s}$ be the event that $h$ is $(r,s)$-stable. Fix $t \in [-r^{-1},r^{-1}]$ and let $y = \Gamma(t)$ where $\Gamma$ is the space-filling path associated with the Brownian plane structure of $h$. Let $a_1 > 4$ be the constant as in \cite[Equation~(4.3), Proposition~4.2]{qle_continuity}. Then we have that
\[ \ex{ \mathop{\mathrm{diam}}(\qhull{y}{\epsilon})^4 {\mathbf 1}_{{\mathcal A}_{r,s}}} \leq c_0 r^4 \epsilon^{a_1} \quad\text{for all}\quad \epsilon \in (0,1).\]
\end{lemma}
\begin{proof}
Let $X$ be the random scaling factor necessary to put $h$ translated by $-\Gamma(t)$ into a smooth canonical description. On ${\mathcal A}_{r,s}$, we have that $X \in [r^{-1},r]$. \cite[Proposition~4.2]{qle_continuity} is stated when $(\C,h,0,\infty)$ has the circle average embedding. When $h$ is instead a smooth canonical description with normalizing bump function supported in $\ball{0}{2} \setminus \ball{0}{1/2}$, it is easy to see that $h_{1/2}(0)$ is a normal random variable with mean zero and bounded variance. In particular, the probability that it is of order $c \sqrt{\log \epsilon^{-1}}$ decays like a power of $\epsilon$ that can be made arbitrarily large by making $c$ large. Thus it is not difficult to see that the result of \cite[Proposition~4.2]{qle_continuity} also applies in this setting, from which the result follows.
\end{proof}
\subsection{The resampling argument}
\label{subsec::resampling_argument}
We are now going to work towards completing the proof of Theorem~\ref{thm::map_determines_embedding}. The proof consists of two main steps.
\begin{enumerate}
\item Show that a $\sqrt{8/3}$-quantum cone is a.s.\ determined by the quantum surfaces which correspond to a finite number of metric hulls (which may overlap), their complement, and the mm-space structure of the overall surface (Lemma~\ref{lem::balls_embedding_determined}).
\item Show that, on the event ${\mathcal A}_{r,s}$ introduced in Section~\ref{subsec::quantum_cone_stable}, the metric hull $\qhull{0}{s}$ is a.s.\ determined by the mm-space structure up to a global rotation (Lemma~\ref{lem::stable_is_determined}).
\end{enumerate}
We then finish by deducing Theorem~\ref{thm::map_determines_embedding} from Lemma~\ref{lem::stable_is_determined}.
\begin{lemma}
\label{lem::balls_embedding_determined}
Suppose that $(\C,h,0,\infty)$, $(\C,\widetilde{h},0,\infty)$ are smooth canonical descriptions of $\sqrt{8/3}$-quantum cones and let $\Gamma, \widetilde{\Gamma}$ (resp.\ $d_\CQ,\wt{d}_\CQ$) be the space-filling curves (resp.\ metrics) associated with the corresponding Brownian plane structures of $h,\widetilde{h}$, respectively, with time normalized so that $\Gamma(0) = \widetilde{\Gamma}(0) = 0$. Suppose that $t_1,\ldots,t_k \in \mathbf{R}$. Let $X = \C \setminus \cup_{j=1}^k \partial \qhull{\Gamma(t_j)}{\epsilon}$ and $\widetilde{X} = \C \setminus \cup_{j=1}^k \partial \qhullT{\widetilde{\Gamma}(t_j)}{\epsilon}$. Suppose that there exists a conformal transformation $\psi \colon X \to \widetilde{X}$ such that $\wt{d}_\CQ(\psi(x),\psi(y)) = d_\CQ(x,y)$ for all $x,y \in X$, $\mu_{\widetilde{h}}(\psi(A)) = \mu_h(A)$ for all open $A \subseteq X$, and $\psi(z)-z \to 0$ as $z \to \infty$. Then $\psi$ extends continuously to $\C$ and is given by the identity map. In particular, $(\C,h,0,\infty)$ and $(\C,\widetilde{h},0,\infty)$ are a.s.\ equivalent as quantum surfaces.
\end{lemma}
\begin{proof}
We are first going to show that $\psi$ extends to a homeomorphism $\C \to \C$. To see this, we note that there exists $c_0,c_1,a_0,a_1 > 0$ (random) such that for each compact set $K \subseteq \C$ and all $x,y \in K \setminus X$ we have that
\begin{align*}
|\psi(x) - \psi(y)|
&\leq c_0 \wt{d}_\CQ(\psi(x),\psi(y))^{a_0} \quad\text{(Theorem~\ref{thm::metric_completion})}\\
&= c_0 d_\CQ(x,y)^{a_0} \quad\text{($\psi$ is an isometry)}\\
&\leq c_1 |x-y|^{a_1} \quad\text{(Theorem~\ref{thm::metric_completion})}.
\end{align*}
Therefore $\psi$ is H\"older continuous on $K \setminus X$ hence extends to be H\"older continuous on~$K$. Since $K \subseteq \C$ was an arbitrary compact set, we therefore have that~$\psi$ is a locally H\"older continuous map $\C \to \C$. The same argument implies that~$\psi^{-1}$ extends to be a locally H\"older continuous map $\C \to \C$. Therefore~$\psi$ is a homeomorphism $\C \to \C$ as desired. Theorem~\ref{thm::finite_unions_removable} combined with \cite[Proposition~2.1]{map_making} implies that $X$ is a.s.\ conformally removable since each $\partial \qhull{\Gamma(t_j)}{\epsilon}$ is a.s.\ the boundary of a simple H\"older domain. Therefore~$\psi$ is in fact a conformal transformation $\C \to \C$ with $\psi(z) - z \to 0$ as $z \to \infty$. This implies that~$\psi$ extends to the identity map on~$\C$, as desired.
\end{proof}
\begin{lemma}
\label{lem::inside_outside_independent}
Suppose that $(\C,h,0,\infty)$ is a $\sqrt{8/3}$-quantum cone and $s > 0$. Let~${\mathcal M}$ be the $\sigma$-algebra generated by the mm-space structure $(M,d,\mu)$ (as a random variable taking values in the space of marked mm-spaces using the $\sigma$-algebra from Section~\ref{subsec::mm_spaces}) associated with $h$. Then the quantum surfaces $(\qhull{0}{s},h)$ and $(\C \setminus \qhull{0}{s},h)$ are conditionally independent given ${\mathcal M}$.
\end{lemma}
\begin{proof}
Let ${\mathcal M}_I$ and ${\mathcal M}_O$ be the $\sigma$-algebras generated by the mm-space structures associated with the quantum surfaces $(\qhull{0}{s},h)$ and $(\C \setminus \qhull{0}{s},h)$, equipped with their internal metric, together with a marked point on $\partial \qhull{0}{s}$. Then we clearly have that ${\mathcal M}_I,{\mathcal M}_O \subseteq {\mathcal M}$. We claim that we in fact have that ${\mathcal M} = \sigma({\mathcal M}_I,{\mathcal M}_O)$. To show that this is the case, we first note that~${\mathcal M}_I$ and~${\mathcal M}_O$ together determine the geodesics from every point $z$ back to $0$ because if $z \in \qhull{0}{s}$ then such a geodesic is contained in $\qhull{0}{s}$ and if $z \notin \qhull{0}{s}$ then such a geodesic can be expressed as a concatenation of a shortest path from~$z$ to $\partial \qhull{0}{s}$ in $\C \setminus \qhull{0}{s}$ and then another shortest path in $\qhull{0}{s}$ from $\partial \qhull{0}{s}$ to $0$. As TBP is locally absolutely continuous with respect to TBM \cite{cl2012brownianplane} and in the case of TBM the tree of geodesics of this form determines the entire metric (recall its construction in Section~\ref{subsec::map_plane}), it follows that ${\mathcal M}_I$ and ${\mathcal M}_O$ in particular determine the length of paths contained in $\qhull{0}{2s}$. Therefore~${\mathcal M}_I$ and~${\mathcal M}_O$ together determine the length of any path because any path can be written as a concatenation of paths which stay outside of~$\qhull{0}{s}$ and which stay in~$\qhull{0}{2s}$. We conclude that ${\mathcal M} = \sigma({\mathcal M}_I,{\mathcal M}_O)$, as desired. Note also that the $\sigma$-algebra~${\mathcal B}$ generated by the quantum boundary length of $\partial \qhull{0}{s}$ is ${\mathcal M}$-measurable (see, e.g., \cite[Proposition~3.27]{map_making}).
Let~${\mathcal F}$ be the $\sigma$-algebra generated by the quantum surface $(\C \setminus \qhull{0}{s},h)$ and note that ${\mathcal M}_O \subseteq {\mathcal F}$. Suppose that $A$ is an event which is measurable with respect to the $\sigma$-algebra generated by the quantum surface $(\qhull{0}{s},h)$ and $B \in {\mathcal M}_I$ is a positive probability event. By the construction of ${\rm QLE}(8/3,0)$, we know that the quantum surface $(\qhull{0}{s},h)$ is conditionally independent of the quantum surface $(\C \setminus \qhull{0}{s},h)$ given its quantum boundary length. We also know that this quantum boundary length is ${\mathcal F}$-measurable and that~${\mathcal M}_I$ is contained in the $\sigma$-algebra generated by the quantum surface $(\qhull{0}{s},h)$. Consequently, we have that
\begin{equation}
\label{eqn::cond_ind_equation}
{{\bf P}}[ A \,|\, B, {\mathcal F}] = \frac{{{\bf P}}[ A,B \,|\, {\mathcal F}]}{{{\bf P}}[B \,|\, {\mathcal F}]} = \frac{{{\bf P}}[ A,B \,|\, {\mathcal B}]}{{{\bf P}}[B \,|\, {\mathcal B}]} = {{\bf P}}[A \,|\, B, {\mathcal B}].
\end{equation}
Since~\eqref{eqn::cond_ind_equation} holds for all such events $A,B$, it follows that for all such events $A$ we have that
\begin{equation}
\label{eqn::cond_ind_equation2}
{{\bf P}}[ A \,|\, {\mathcal M}_I, {\mathcal F}] = {{\bf P}}[ A \,|\, {\mathcal M}_I].
\end{equation}
Since~\eqref{eqn::cond_ind_equation2} holds for all such events $A$, ${\mathcal M} = \sigma({\mathcal M}_I,{\mathcal M}_O)$, ${\mathcal M}_I$ and ${\mathcal M}_O$ are conditionally independent given ${\mathcal B}$, and ${\mathcal B} \subseteq {\mathcal M}_I \cap {\mathcal M}_O$, it follows that
\begin{equation}
\label{eqn::cond_ind_equation3}
{{\bf P}}[ A \,|\, {\mathcal M}, {\mathcal F}] = {{\bf P}}[ A \,|\, {\mathcal M}].
\end{equation}
for all such events $A$. It follows from~\eqref{eqn::cond_ind_equation3} that the quantum surface $(\qhull{0}{s},h)$ is conditionally independent of the quantum surface $(\C \setminus \qhull{0}{s},h)$ given ${\mathcal M}$, as desired.
\end{proof}
\begin{lemma}
\label{lem::stable_is_determined}
Let $(\C,h,0,\infty)$ be a smooth canonical description of a $\sqrt{8/3}$-quantum cone. Fix $r,s > 0$, and let ${\mathcal A}_{r,s}$ be the event that $h$ is $(r,s)$-stable. Let ${\mathcal F}$ be the $\sigma$-algebra generated by the mm-space structure $(M,d,\mu)$ (as a random variable taking values in the space of marked mm-spaces using the $\sigma$-algebra from Section~\ref{subsec::mm_spaces}) associated with~$h$ and the quantum surface $(\C \setminus \qhull{0}{s},h)$. Let~$\widetilde{h}$ be the field which describes the quantum surface generated by starting with~$h$ and then resampling the quantum surface $(\qhull{0}{s},h)$ according to its conditional law given~${\mathcal F}$. Let $\Gamma,\widetilde{\Gamma}$ be the space-filling paths associated with the Brownian plane structures of $h,\widetilde{h}$, respectively, with time normalized so that $\Gamma(0) = \widetilde{\Gamma}(0) = 0$. Fix $-r^{-1} \leq a < b \leq r^{-1}$ and let $K_{a,b}$ (resp.\ $\widetilde{K}_{a,b}$) be the complement of the unbounded component of $\C \setminus \Gamma([a,b])$ (resp.\ $\C \setminus \widetilde{\Gamma}([a,b])$). We assume that the surface $(\C,\widetilde{h},0,\infty)$ is embedded so that there exists a unique conformal map $\varphi \colon \C \setminus K_{a,b} \to \C \setminus \widetilde{K}_{a,b}$ with $|\varphi(z)-z| \to 0$ as $z \to \infty$ and so that $\widetilde{h} = h \circ \varphi + Q \log|\varphi'|$ in $\C \setminus \Gamma([a,b])$. On ${\mathcal A}_{r,s}$, $\varphi(z) = z$ for all $z \in \C \setminus K_{a,b}$ a.s.
\end{lemma}
Note that Lemma~\ref{lem::stable_is_determined} implies that $K_{a,b}$ is ${\mathcal F}$-measurable for all $-r^{-1} \leq a < b \leq r^{-1}$, up to a global rotation. This implies that the same is true for $\Gamma([a,b])$, hence $\Gamma|_{[-r^{-1},r^{-1}]}$.
\begin{proof}[Proof of Lemma~\ref{lem::stable_is_determined}]
Fix $\epsilon > 0$, let $k = 2r^{-1} \epsilon^{-1}$, and assume that $r,\epsilon$ are such that $k$ is an integer. Let $-r^{-1} \leq t_0 < t_1 < \cdots < t_k \leq r^{-1}$ be $k+1$ equally spaced times.
Let ${\mathcal M}$ be the $\sigma$-algebra generated by the mm-space structure $(M,d,\mu)$ (as a random variable taking in the space of marked mm-spaces with the $\sigma$-algebra from Section~\ref{subsec::mm_spaces}) associated with~$h$.
Fix $u > 0$ small and let $\xi = 1/4 - u$. We let $h^0 = h$ (resp.\ $\widetilde{h}^0 = \widetilde{h}$) and let $\Gamma^0$ (resp.\ $\widetilde{\Gamma}^0$) denote the space-filling path associated with TBP structure of $h^0$ (resp.\ $\widetilde{h}^0$). Fix $j \geq 1$ and assume that we have defined fields $h^0,\ldots,h^{j-1}$ and $\widetilde{h}^0,\ldots,\widetilde{h}^{j-1}$ with associated space-filling paths $\Gamma^0,\ldots,\Gamma^{j-1}$ and $\widetilde{\Gamma}^0,\ldots,\widetilde{\Gamma}^{j-1}$. We then let~$h^j$ be the field which describes the quantum surface which is constructed by starting with~$h^{j-1}$ and then resampling the quantum surface $(\qhullj{\Gamma^{j-1}(t_{j-1})}{\epsilon^{\xi}}{j-1},h^{j-1})$, where $\qhullj{\Gamma^{j-1}(t_{j-1})}{\epsilon^{\xi}}{j-1}$ is the hull of the metric ball of radius~$\epsilon^{\xi}$ centered at~$\Gamma^{j-1}(t_{j-1})$ using the metric associated with~$h^{j-1}$, with respect to the field according to its conditional law given the quantum surface $(\C \setminus \qhullj{\Gamma^{j-1}(t_{j-1})}{\epsilon^{\xi}}{j-1},h)$ and ${\mathcal M}$. We take $h^j$ so that the embedding of the surface $(\C, h^j,0,\infty)$ is such that if $\varphi_j$ is the conformal map which takes $\C \setminus \qhullj{\Gamma^{j-1}(t_{j-1})}{\epsilon^{\xi}}{j-1}$ to $\C \setminus \qhullj{\Gamma^j(t_{j-1})}{\epsilon^{\xi}}{j}$ then $|\varphi_j(z) - z| \to 0$ as $z \to \infty$ and $h^j = h^{j-1} \circ \varphi_j^{-1} + Q \log |(\varphi_j^{-1})'|$. We let $\Gamma^j$ be the space-filling path associated with~$h^j$. We note that $h^j$ is not necessarily a smooth canonical description, however on ${\mathcal A}_{r,s}$ the scaling factor necessary to transform it into a smooth canonical description is between $r^{-1}$ and $r$.
For each $j \geq 1$, we construct $\widetilde{h}^j$ from $\widetilde{h}^{j-1}$ in the same way that we constructed $h^j$ from $h^{j-1}$. We also let $\widetilde{\Gamma}^j$ and $\qhullTj{x}{r}{j}$ be the space-filling path and filled metric ball associated with the metric of $\widetilde{h}^j$. We couple the two resampling procedures together so that $(\qhullj{\Gamma^j(t_j)}{\epsilon^{\xi}}{j},h^j)$ and $(\qhullTj{\widetilde{\Gamma}^j(t_j)}{\epsilon^{\xi}}{j},\widetilde{h}^j)$ are equivalent as quantum surfaces for each $j$. We also define the conformal maps $\widetilde{\varphi}_j$ in the same way as $\varphi_j$. We note that since we are working on the event that $h$ is $(r,s)$-stable, it follows from the definition of stability that $h^j$ is $(r,s)$-stable for each $j$. The same is likewise true for $\widetilde{h}^j$ for each $j$.
Proposition~\ref{prop::path_chunk_diameter_bound} implies there exists constants $c_0,c_1,a > 0$ such that with probability at least $1-c_0 \exp(-c_1 \epsilon^{-a})$ we have that $\qhull{0}{s} \subseteq \cup_{j=1}^k \qhull{\Gamma(t_j)}{\epsilon^\xi}$ as $\xi < 1/4$. This implies that, with probability at least $1-c_0 \exp(-c_1 \epsilon^{-a})$, the aforementioned procedure resamples the entire quantum surface corresponding to $(\qhull{0}{s},h)$. (We are using that the statement of Proposition~\ref{prop::path_chunk_diameter_bound} is a statement about the mm-space structure and does not depend on its specific embedding.) The same is likewise true with $\widetilde{h}$ in place of $h$.
Let $X = \C \setminus \cup_{j=1}^k \partial \qhullj{\Gamma^k(t_j)}{\epsilon^{\xi}}{k}$ and $\widetilde{X} = \C \setminus \cup_{j=1}^k \partial \qhullTj{\widetilde{\Gamma}^k(t_j)}{\epsilon^{\xi}}{k}$. On the aforementioned event, the resampling procedure yields a conformal map $\varphi \colon X \to \widetilde{X}$ which satisfies the hypotheses of Lemma~\ref{lem::balls_embedding_determined}. Therefore $\varphi$ is given by the identity map and the two quantum surfaces described by the fields $h^k$ and $\widetilde{h}^k$ are equivalent.
To finish proving the lemma, we will show that the maps $\psi = \varphi_k \circ \cdots \circ \varphi_1$ and $\widetilde{\psi} = \widetilde{\varphi}_k \circ \cdots \circ \widetilde{\varphi}_1$ both converge to the identity map as $\epsilon \to 0$. We will explain the argument in the former case, as the latter argument is analogous.
We note that $h^k$ is a function of the quantum surfaces $(\qhullj{\Gamma^j(t_j)}{\epsilon^{\xi}}{j},h^j)$ for $1 \leq j \leq k$ together with $(\C \setminus \qhull{0}{s},h)$. Indeed, this follows from the same removability argument used to prove Lemma~\ref{lem::balls_embedding_determined}. Moreover, these surfaces are by construction conditionally independent given ${\mathcal F}$. With $a_1$ as in Lemma~\ref{lem::metric_ball_stable_moment_bound}, we assume that $u > 0$ is chosen sufficiently small so that $a_1 \xi > 1$. Fix $z \in \C$ with $|z| > 2$. Let ${\mathcal G}_j$ be the $\sigma$-algebra generated by $(\qhullj{\Gamma^i(t_i)}{\epsilon^\xi}{i},h^i)$ for each $i \neq j$. Lemma~\ref{lem::variance_bound} implies that
\begin{equation}
\label{eqn::var_map_bound1}
\ex{\var{\psi(z) \,|\, {\mathcal F}} {\mathbf 1}_{{\mathcal A}_{r,s}}}
\leq \sum_{j=1}^k \ex{ \var{\psi(z) \,|\, {\mathcal F}, {\mathcal G}_j} {\mathbf 1}_{{\mathcal A}_{r,s}}}.
\end{equation}
By Lemma~\ref{lem::conformal_map_bound}, there exists a constant $c_0 > 0$ such that
\begin{equation}
\label{eqn::var_map_bound2}
\var{\psi(z) \,|\, {\mathcal F}, {\mathcal G}_j} \leq c_0 \mathop{\mathrm{diam}}(\qhullj{\Gamma^j(t_j)}{\epsilon^{\xi}}{j})^4.
\end{equation}
On ${\mathcal A}_{r,s}$, we note that the law of $h^j$ is the same as the law of $h^k$, up to a scaling factor which is between $r^{-1}$ and $r$. Consequently,
\begin{equation}
\label{eqn::diameter_scaling}
\ex{ \mathop{\mathrm{diam}}(\qhullj{\Gamma^j(t_j)}{\epsilon^{\xi}}{j})^4 {\mathbf 1}_{{\mathcal A}_{r,s}} } \leq r^4 \ex{ \mathop{\mathrm{diam}}(\qhullj{\Gamma^k(t_j)}{\epsilon^{\xi}}{k})^4 {\mathbf 1}_{{\mathcal A}_{r,s}}}
\end{equation}
Combining~\eqref{eqn::diameter_scaling} with~\eqref{eqn::var_map_bound1} and~\eqref{eqn::var_map_bound2} (and possibly increasing $c_0$) we see that
\begin{align*}
\ex{\var{\psi(z) \,|\, {\mathcal F}} {\mathbf 1}_{{\mathcal A}_{r,s}}}
&\leq \sum_{j=1}^k r^4 \ex{ \mathop{\mathrm{diam}}(\qhullj{\Gamma^k(t_j)}{\epsilon^{\xi}}{k})^4{\mathbf 1}_{{\mathcal A}_{r,s}} } \\
&\leq c_0 k r^8 \epsilon^{a_1 \xi} \quad\text{(Lemma~\ref{lem::metric_ball_stable_moment_bound})}\\
&= 2 c_0 r^9 \epsilon^{a_1 \xi-1} \to 0 \quad\text{as}\quad \epsilon \to 0 \quad\text{($a_1 \xi > 1$)}.
\end{align*}
The same argument also applies to $\widetilde{\psi}$. This implies that $\varphi$ is ${\mathcal F}$-measurable on ${\mathcal A}_{r,s}$, which implies the result.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm::map_determines_embedding}]
Let~${\mathcal F}$ be as in Lemma~\ref{lem::stable_is_determined}. As explained earlier, Lemma~\ref{lem::stable_is_determined} implies that on the event~${\mathcal A}_{r,s}$ that~$h$ is $(r,s)$-stable, we have that $\Gamma|_{[-r^{-1},r^{-1}]}$ is ${\mathcal F}$-measurable. We note that the intersection of the $\sigma$-algebras generated by the quantum surfaces $(\C \setminus \qhull{0}{s},h)$ for $s > 0$ is trivial. By applying scaling and using the scale invariance of the law of a $\sqrt{8/3}$-quantum cone, it therefore follows that the probability that~$\Gamma$, hence~$\mu_h$, hence~$h$ \cite{bss2014equivalence} is determined up to a global scaling and rotation by~${\mathcal M}$ is at least~$\pr{{\mathcal A}_{r,s}}$. The result follows because Proposition~\ref{prop::stable_positive_probability} implies that we can adjust $r, s$ so that~$\pr{{\mathcal A}_{r,s}}$ is arbitrarily close to~$1$.
\end{proof}
|
\section{Introduction}\label{sec:intro}
\begin{sloppypar}
In the very early days of the wireless industry,~\cite{Rap91a} predicted that wireless would be as pervasive as utility lines and house wiring by 2020. Now, that vision may be reached, as the fifth-generation (5G) of wireless standards are being developed for millimeter-wave (mmWave) frequency bands to provide tens of gigabits per second data rates, since today's frequencies below 6 GHz are too crowded to meet global traffic demand~\cite{Rap13a}. The 3rd Generation Partnership Project (3GPP), the cellular industry's global standards body, initiated a working group in September 2015 to develop channel models for spectrum above 6 GHz, as have other groups such as METIS~\cite{METIS2015}, MiWEBA~\cite{Miweba14a}, mmMagic~\cite{mmMagic}, ETSI~\cite{ETSI2015}, and IEEE 802.11ad. In just 10 months, 3GPP released TR 38.900 - Release 14 for channel models above 6 GHz in July 2016~\cite{3GPP.38.900}.
\end{sloppypar}
The development of 3GPP's channel models above 6 GHz was supported by numerous academic and industrial measurement campaigns and ray-tracing simulations for urban macrocell (UMa), urban microcell (UMi), and indoor hotspot (InH) scenarios~\cite{Rap15b,Nguyen16a,5GCM,Haneda16a,Haneda16b,Sun16b,Mac15b,Thomas16a,Samimi16a}. The channel models will be useful for the development of 5G waveforms, MAC, and PHY approaches, especially in light of the Federal Communications Commission's (FCC) goal for the USA to lead global 5G rollout via its recent Spectrum Frontiers ruling~\cite{FCC16-89}. The breakneck speed of 5G channel model development, however, increases the likelihood of adoption of models that are theoretically flawed, or unsubstantiated by empirical evidence.
While UMi, UMa, and InH scenarios were extensively studied~\cite{5GCM,Rap15b,Haneda16a,Haneda16b,Mac15b,Sun16b,Thomas16a,Nguyen16a,Samimi16a}, the rural macrocell (RMa) scenario was neglected and is not fully understood. The mmWave RMa model~\cite{3GPP.38.900} was hastily adopted from a cumbersome and two-decade-old propagation model meant for frequencies below 6 GHz, with very light validation from a very limited measurement campaign at 24 GHz~\cite{TDOC164975}. As shown here, the dual slope RMa model in~\cite{3GPP.38.900} is not valid, mathematically, above 9.1 GHz, meaning that a flawed and untested model currently exists in 3GPP. Also, we show that virtually no field measurements have been used to test the existing model. This paper offers a solution to the mathematical problem, and validates a much simpler path loss model for the RMa propagation scenario with field measurements at 73 GHz. This paper is organized as follows: Section~\ref{sec:3GPPRMa} describes the existing RMa path loss models in 3GPP~\cite{3GPP.38.900} and illuminates the mathematical problem and lack of evidence for the models, Section~\ref{sec:Meas} describes the 73 GHz RMa measurement campaign conducted in August 2016 in Riner, Virginia, Section~\ref{sec:PL} provides and discusses the empirical results and RMa path loss models for frequencies above 6 GHz, and conclusions are drawn in Section~\ref{sec:conc}.
\section{RMa Path Loss Model in 3GPP}\label{sec:3GPPRMa}
RMa path loss (PL) models enable engineers to predict signal strength as a function of propagation distance in rural environments from a tall tower (macrocell). The RMa path loss model equations provided in 3GPP~\cite{3GPP.38.900} are long and cumbersome with numerous input parameters specifying the base station height ($h_{BS}$), user terminal height ($h_{UT}$), average street width ($W$, a questionable variable for RMa scenarios), average building height ($h$, also questionable for rural settings), three-dimensional (3D) transmitter-receiver (T-R) separation distance ($d_{3D}$), and carrier frequency ($f_c$). The RMa line-of-sight (LOS) path loss model (for when a transmitter (TX) antenna can see the receiver (RX) antenna) is a dual slope model with a breakpoint given in~\eqref{eq:RMaLOS},\eqref{eq:dbp}~\cite{3GPP.38.900,ITU-RM.2135}:
\begin{align}\label{eq:RMaLOS}
\begin{split}
PL _1& = 20\log(40\pi \cdot d_{3D} \cdot f_c /3)+\min(0.03h^{1.72},10)\log_{10}(d_{3D}) \\
&-\min(0.044h^{1.72},14.77)+0.002\log_{10}(h)d_{3D}\\
PL_2 & = PL_1 (d_{BP})+40\log_{10}(d_{3D}/d_{BP})
\end{split}
\end{align}
where all heights and distances are in meters (m) and the shadow fading standard deviation is $\sigma_{SF}$ = 4 dB for $PL_1$ (before the breakpoint) and $\sigma_{SF}$ = 6 dB for $PL_2$ (after the breakpoint). (See~\cite{Rap02a} for treatment of large-scale path loss modeling). Eq.~\eqref{eq:RMaLOS} is adopted from ITU-R M.2135~\cite{ITU-RM.2135} as the LOS RMa path loss model. The breakpoint $d_{BP}$ in~\eqref{eq:RMaLOS} is the particular distance where the slope of the path loss changes, and is defined as:
\begin{equation}\label{eq:dbp}
d_{BP} = 2\pi \cdot h_{BS} \cdot h_{UT} \cdot f_c/c
\end{equation}
where $f_c$ is the carrier frequency in Hz and $c$ = $3.0\times10^8$ m/s (speed of light in air or free space).
The RMa non-LOS (NLOS) path loss model (for when buildings or foliage block the radio path) is given in~\eqref{eq:RMaNLOS} and has an odd physical imperfection such that it models close-in signals (say within 500 m) as being much stronger than physics would dictate, and thus requires a mathematical patch by requiring a lower bound equal to the RMa LOS path loss model~\cite{3GPP.38.900,Sun16b}:
\begin{align}\label{eq:RMaNLOS}
\begin{split}
PL & = \max(PL_{RMa-LOS},PL_{RMa-NLOS})\\
PL & _{RMa-NLOS} = 161.04-7.1\log_{10}(W)+7.5\log_{10}(h)\\
&-(24.37-3.7(h/h_{BS})^2)\log_{10}(h_{BS})\\
&(43.42-3.1\log_{10}(h_{BS}))(\log_{10}(d_{3D})-3)\\
&+20\log_{10}(f_c)-(3.2(\log_{10}(11.75h_{UT}))^2-4.97)
\end{split}
\end{align}
where all heights and distances are in meters and the shadow fading standard deviation $\sigma_{SF}$ = 8 dB. Table~\ref{tbl:appRange} provides the applicability range and default parameter values for the LOS and NLOS RMa path loss models.
\begin{table}
\centering
\caption{3GPP TR 38.900 RMa path loss model default values and applicability ranges~\cite{3GPP.38.900}.}\label{tbl:appRange}
\scalebox{0.83}{
\begin{tabu}{|l|}\hline
\textbf{RMa LOS Default Values Applicability Range} \\ \specialrule{1.5pt}{0pt}{0pt}
10 m $<d_{2D}<d_{BP}$, \\
$d_{BP} < d_{2D} <10\:000$ m,\\
$h_{BS} = 35$ m, $h_{UT}=1.5$ m, $W=20$ m, $h=5$ m\\
Applicability ranges: 5 m $<h<50$ m; 5 m $<W<50$ m; \\
10 m $<h_{BS}<150$ m; 1 m $<h_{UT}<10$ m \\ \hline
\textbf{RMa NLOS Default Values Applicability Range} \\ \specialrule{1.5pt}{0pt}{0pt}
10 m $<d_{2D}<5\:000$ m, \\
$h_{BS} = 35$ m, $h_{UT}=1.5$ m, $W=20$ m, $h=5$ m\\
Applicability ranges: 5 m $<h<50$ m; 5 m $<W<50$ m; \\
10 m $<h_{BS}<150$ m; 1 m $<h_{UT}<10$ m \\ \hline
\end{tabu}}
\end{table}
Similar to LOS, the RMa NLOS path loss model was adopted from~\cite{ITU-RM.2135}.
A footnote for the RMa path loss models in~\cite{3GPP.38.900} specifies that the applicable frequency range is $0.8$ GHz $ < f_c < f_H$, where $f_H$ is 30 GHz for RMa, but we found only one small measurement campaign (at 24 GHz) that tried to validate the RMa model~\cite{TDOC164975}. Surprisingly, the path loss models in~\cite{ITU-RM.2135} are for below 6 GHz, calling into question the validity of these models for mmWave.
\\
\\
\subsection{3GPP RMa LOS Path Loss Origin}
The LOS RMa path loss model in~\cite{3GPP.38.900}, adopted from~\cite{ITU-RM.2135}, originates from ITU 5D/88-E~\cite{ITU-5D/88-E}, which only shows a portion of the model in~\cite{3GPP.38.900}. The~\cite{ITU-5D/88-E} document cites work by the NTT Wireless Systems Laboratories as the original source of the RMa LOS path loss model, but measurements were only conducted at 2.6 GHz~\cite{Ichitsubo00a}. We can find no other publication or open-source document other than~\cite{TDOC164975} to support the 3GPP RMa path loss model above 6 GHz, yet major US carriers such as Verizon and AT\&T are eyeing rural mmWave service in their first trials of new mmWave spectrum. This lack of empirical support, and apparent misappropriation of the 3GPP RMa model have motivated the mmWave channel measurements and models herein.
Even more surprising is that the 3GPP LOS RMa dual slope path loss model~\eqref{eq:RMaLOS} is mathematically invalid for frequencies above 9.1 GHz, since the breakpoint distance~\eqref{eq:dbp} at 9.1 GHz or greater is farther than 10 km, the upper range specified for the model (See Table~\ref{tbl:appRange}). Figure~\ref{fig:dbp} displays a plot of the breakpoint distance vs. frequency for~\eqref{eq:dbp}, and shows that the RMa path loss model~\cite{3GPP.38.900} reverts to a single slope model for centimeter-waves above 9.1 GHz, and for all mmWave frequencies. An improved close-in reference distance single-slope model that avoids this problem is shown in Section~\ref{sec:PL}.
\begin{figure}
\centering
\includegraphics[width=0.40\textwidth]{RMA_dbp.eps}
\caption{LOS breakpoint distance vs. frequency in~\eqref{eq:dbp}.}
\label{fig:dbp}
\end{figure}
\subsection{3GPP RMa NLOS Path Loss Origin}
The NLOS RMa path loss model in~\cite{3GPP.38.900} and adopted from~\cite{ITU-RM.2135} can be traced back to a paper by Sakagami and Kuboi from 1991 that is based on empirical data from Tokyo at 813 MHz and 1443 MHz in a dense urban environment~\cite{Sakagami91a}, otherwise known as the extended Sakagami model~\cite{Ohta03a}. This explains the odd variables in~\eqref{eq:RMaLOS}--\eqref{eq:RMaNLOS} which are unneeded in rural settings. One difference between the legacy Sakagami model and models in~\cite{ITU-RM.2135} and~\cite{3GPP.38.900} is the first term in~\eqref{eq:RMaNLOS} which is 161.04 dB rather than 100 or 101 dB, since the Sakagami models had units of frequencies in MHz rather than GHz (the difference in free space path loss at 1 m between 1 MHz and 1 GHz is $\sim$ 60 dB).
The only effort~\cite{TDOC164975} to validate the 3GPP RMa model~\cite{3GPP.38.900} above 6 GHz described a limited measurement study at 24 GHz that combined LOS and NLOS scenarios, and which was never peer reviewed. The study was conducted over a very limited two-dimensional (2D) T-R separation distance range of 200 to 500 m, yet the published RMa model in~\eqref{eq:RMaLOS}--\eqref{eq:RMaNLOS}~\cite{3GPP.38.900} is specified over a 2D T-R distance from 10 m to 5 km or 10 km. Additionally,~\cite{TDOC164975} did not provide a best-fit indicator (e.g., RMSE) between the measured data and model. With such little evidence and questionable origins of the 3GPP RMa model, we set out to conduct a rural macrocell measurement and modeling study in LOS and NLOS beyond the 10 km distances stated in~\cite{3GPP.38.900}.
\section{73 GHz RMa Measurements}\label{sec:Meas}
A measurement campaign was conducted in Riner, Virginia, a rural town in the southwestern portion of the state using the 73 GHz mmWave frequency band. The TX was located at Prof. Rappaport's mountain home.
A narrowband CW tone was transmitted at a center frequency of 73.5 GHz with a maximum transmit power of 14.7 dBm (28 mW) with a 7$^\circ$ azimuth and elevation half-power beamwidth (HPBW) antenna having 27 dBi of gain, which resulted in 41.7 dBm effective isotropic radiated power (EIRP) (14.8 W EIRP), much lower power than traditional RMa cellular base stations. Figure~\ref{fig:TX_diag} displays the TX schematic where a 5.625 GHz CW tone is mixed with a 67.875 GHz signal (22.625 GHz x3 frequency multiplied inside the upconverter) to reach an RF center frequency of 73.5 GHz.
\begin{figure}
\centering
\includegraphics[width=0.40\textwidth]{TX_diag.eps}
\caption{73 GHz TX measurement equipment.}\label{fig:TX_diag}
\end{figure}
At the RX, an identical narrowbeam horn antenna with 7$^\circ$ azimuth and elevation HPBW and 27 dBi of gain was used to capture the RF signal which was downconverted (with 29.9 dB of gain) to bring the RF signal to an intermediate frequency (IF) of 5.625 GHz that was subsequently amplified with a low-noise amplifier (LNA) with 35 dB of gain (note the step attenuator in Figure~\ref{fig:RX_diag} to ensure linear operation). A Keysight E4407B spectrum analyzer in zero-span mode recorded received power levels, as depicted in Figure~\ref{fig:RX_diag} using a 15 kHz bandwidth setting. Occasional frequency tuning was required to account for system oscillator drift. Similar to the 73 GHz upconverter, the local oscillator (LO) of 22.625 GHz that enters the 73 GHz downconverter is x3 frequency multiplied to 67.875 GHz to demodulate the 73.5 GHz RF signal to the 5.625 GHz IF.
\begin{figure}
\centering
\includegraphics[width=0.43\textwidth]{RX_diag.eps}
\caption{73 GHz RX measurement equipment.}\label{fig:RX_diag}
\end{figure}
The maximum measurable path loss of the system was 190 dB, with local time averaging used to obtain received power at various RX locations.
\\
\subsection{Measurement Locations and Approach}
For the RMa measurements, 14 LOS locations and 17 NLOS locations were measured with detectable signal, and 5 additional locations resulted in outages where signal was not detectable. The 2D T-R separation distance ranged from 33 m (calibration distance) to 10.8 kilometers (km) for LOS scenarios and 3.4 km to 10.6 km for NLOS scenarios. The TX was located on a house porch on top of a mountain ridge, $\sim$ 110 m above surrounding terrain as shown in Figure~\ref{fig:TX_terrain}.
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{map.eps}
\caption{Map of TX and RX locations. The yellow star represents the TX, red pins indicate NLOS locations, and blue pins indicate LOS locations.}\label{fig:map}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{terrain_v5.eps}
\caption{Sketch of TX location and surroundings.}\label{fig:TX_terrain}
\end{figure*}
Figure~\ref{fig:TX_view} shows the northerly view from the TX to the surroundings below, where there is a 31 m distance from the house to the mountain drop edge.
\begin{figure}
\centering
\includegraphics[width=0.42\textwidth]{porch_1.eps}
\caption{Outward view from TX.}\label{fig:TX_view}
\end{figure}
During the measurements the TX antenna was set to a fixed downtilt of 2$^\circ$ and the azimuth was manually set to point in a direction that maximized received power at the various RX locations.
The measurement system was periodically calibrated at a distance of 33 meters to ensure theoretical free space path loss and accurate performance, with repeatability of 0.2 dB over the two-day campaign. The RX system was placed in a van and driven to measurement locations along highways and neighborhoods. At each RX location, the RX antenna was raised to an average height between 1.6 and 2 meters above ground while the best TX and RX antenna azimuth angles were manually determined based on the strongest received power.
Figure~\ref{fig:map} shows the Google Earth imagery of the TX and RX locations and the TX azimuth scanning window of $\pm10^\circ$ of true North so as to avoid a mountain that was west of the TX, and to avoid diffraction from the east side of the antenna due to a rising slope in the front yard of the house. RX locations that only had one or two small trees blocking the LOS path were considered LOS. Furthermore, two LOS locations (RX 31 and RX 32; top right corner of Figure~\ref{fig:map}) are shown but not used in the path loss model derivation due to the diffraction loss from the sloping yard (see Figure~\ref{fig:TX_view}).
\section{RMA Path Loss Model Results}\label{sec:PL}
\subsection{CI Path Loss Model}
As an alternative to the complicated RMa path loss model in~\eqref{eq:RMaLOS}--\eqref{eq:RMaNLOS} ~\cite{3GPP.38.900}, the close-in free space reference distance (CI) path loss model has a solid physical basis, a succinct form with only one modeling parameter, and is simultaneously applicable for frequencies both below and above 6 GHz~\cite{Rap15b,Mac15b,Sun16b,Mac13a,Haneda16a,Haneda16b}. Also, the CI path loss model exhibits excellent parameter stability and prediction accuracy for unanticipated use cases over distances outside of the original measurement range compared to other path loss models~\cite{Sun16b,Thomas16a}. The CI model has already been adopted as an optional path loss model for the UMi, UMa, and InH scenarios in~\cite{3GPP.38.900} based on works by the authors~\cite{5GCM,Haneda16a,Haneda16b,Sun16b,Thomas16a,Mac15b} due to its use of a frequency dependent free space path loss term in the first meter~\cite{Rap15b,Mac15b,Sun16b,Thomas16a}. The CI model, with its many virtues, is now shown to be well suited for the RMa scenario, as data confirm this model would be a sensible replacement for the current 3GPP RMa model. The general form of the CI path loss model is expressed as follows:
\begin{equation}\label{CI1}
\begin{split}
\PL^{\CI}(f_c,d)[\dB]=&\FSPL(f_c, d_0)[\dB]+10n\log_{10}\left(\frac{d}{d_0}\right)\\
&+\chi_{\sigma},~\text{where}~d\geq d_0
\end{split}
\end{equation}
where $\PL$ is the path loss measured in dB that is a function of T-R separation distance $d$ in m between the TX and RX, $f_c$ is the carrier frequency in GHz, and $d_0$ is the close-in free space reference distance in m. For distance $d$ between the TX and RX, 2D or 3D distances may be used, as the difference is de minimus for large separations (several km). In~\eqref{CI1}, $n$ represents the path loss exponent (PLE)~\cite{Rap15b,Mac15b,Sun16b,Rap02a,Haneda16a,Haneda16b}, and $\chi_{\sigma}$ denotes the shadow fading which is a zero-mean Gaussian random variable with standard deviation $\sigma$ in dB~\cite{Rap02a,Rap15b}. Note that in~\eqref{CI1}, $10\times n$, or $10\times \PLE$, is the coefficient in front of the log-distance term. The free space path loss in dB at a distance $d_0$ is given by Friis' free space path loss (FSPL)~\cite{Rap02a}:
\begin{equation}\label{FSPL1}
\FSPL(f_c,d_0)[\dB]=20\log_{10}\left(\frac{4\pi f_cd_0\times 10^9}{c}\right)
\end{equation}
where $c$ is the speed of light, $3\times 10^8$ m/s.
In the optional CI path loss model for UMi, UMa, and InH scenarios in 3GPP (July 2016)~\cite{3GPP.38.900}, the free space reference distance $d_0$ is set to 1 m, since there is clearly no obstruction in the first meter of transmission, and it simplifies the mathematical equation, provides a standardized modeling approach that may be used universally, and has been shown to yield superb model accuracy and parameter stability across a wide range of frequencies and distances~\cite{Sun16b,Rap15b,Thomas16a}. Based on the efficacy and stability of a 1 m reference distance~\cite{Rap15b,Mac15b,Sun16b,Thomas16a}, we use this in the RMa CI model. With $d_0$ = 1 m, the FSPL in Eq.~\eqref{FSPL1} can be reformulated as:
\begin{equation}\label{FSPL2}
\begin{split}
\FSPL(f_c,\text{1 m})[\dB]&=20\log_{10}\left(\frac{4\pi f_c\times 10^9}{c}\right)\\
&=20\log_{10}\left(\frac{4\pi \times 10^9}{c}\right)+20\log_{10}(f_c)\\
&=32.4+20\log_{10}(f_c)\;[\dB]\\
\end{split}
\end{equation}
Consequently, Eq.~\eqref{CI1} can be recast as:
\begin{equation}\label{CI2}
\begin{split}
\PL^{\CI}(f_c,d)[\dB]=&\FSPL(f_c, \text{1 m})[\dB]+10n\log_{10}(d)+\chi_{\sigma}\\
=&32.4+10n\log_{10}(d)+20\log_{10}(f_c)\\
&+\chi_{\sigma},~\text{where}~d\geq 1~\m\\
\end{split}
\end{equation}
\subsection{3GPP RMa Monte Carlo Simulation}
To test the efficacy of the CI model, we simulated the 3GPP RMa LOS and NLOS path loss models from~\eqref{eq:RMaLOS}--\eqref{eq:RMaNLOS} using default values in Table~\ref{tbl:appRange}. A Monte Carlo simulation was conducted for each environment at the following frequencies: 1, 2, 6, 15, 28, 38, 60, 73, and 100 GHz, with 50,000 instances each, for 2D T-R separation distances ranging between 10 m and 10 km in LOS and between 10 m and 5 km in NLOS. The CI model with a 1 m close-in reference distance~\eqref{CI2} was fit to the simulated 3GPP model sample points generated from random distances and normal (in dB) shadow fading sample values using~\eqref{eq:RMaLOS}--\eqref{eq:RMaNLOS} (frequencies above 9.1 GHz reverted to a single slope model and ignored the second slope portion of~\eqref{eq:RMaLOS}). From the simulated sample points, the equivalent CI path loss models were developed:
\begin{equation}\label{CILOS-3GPP}
\begin{split}
\PL^{\CI\text{-3GPP}}_{\LOS}(f_c,d)[\dB]=&32.4+23.1\log_{10}(d)+20\log_{10}(f_c)\\
&+\chi_{\sigma_{\LOS}},~\text{where}~d\geq 1~\m\\
\end{split}
\end{equation}
\begin{equation}\label{CINLOS-3GPP}
\begin{split}
\PL^{\CI\text{-3GPP}}_{\NLOS}(f_c,d)[\dB]=&32.4+30.4\log_{10}(d)+20\log_{10}(f_c)\\
&+\chi_{\sigma_{\NLOS}},~\text{where}~d\geq 1~\m\\
\end{split}
\end{equation}
\noindent where the large-scale shadow fading standard deviations ${\sigma_{\LOS}}$ and ${\sigma_{\NLOS}}$ were 5.9 dB and 8.3 dB, respectively. These simulated models were not based on measured data, but were formed by using the RMa model in~\cite{3GPP.38.900} to reproduce simulated data that could then be fit to the CI model. The resulting $\sigma$ values in the resulting CI models were quite reasonable but the PLE of 2.31 in LOS and 3.04 in the NLOS simulations show that the 3GPP RMa model predicts greater loss at larger distances compared to measured results (see Section~\ref{sec:OurRMa}), a phenomenon also found in other 3GPP models~\cite{Sun16b}. This exercise showed that the existing 3GPP RMa models could be recast in a much simpler and mathematically accurate form using the CI model with a single slope for all frequencies, but as measurements revealed, the 3GPP model is inaccurate. An elegant feature of the CI model is that 23.1 in~\eqref{CILOS-3GPP} corresponds to a PLE $n$ of 2.31 and is equivalent to $10n=23.1$, or 23.1 dB loss for each decade increase of distance.
\subsection{Proposed RMa Path Loss Model}\label{sec:OurRMa}
Using the measured data from the RMa measurement campaign at 73 GHz described in Section~\ref{sec:Meas}, we determined the minimum mean square error (MMSE) fit (the optimal PLE to minimize $\sigma$) using the CI model~\eqref{CI2} for LOS and NLOS RMa environments. Figure~\ref{fig:RMaPL} illustrates the scatter plot of path loss versus T-R separation distance, where blue circles represent measured LOS path loss data, red crosses denote measured NLOS data, and the two green diamonds indicate LOS data with partial diffraction from the edge of the yard near the TX. As shown by Figure~\ref{fig:RMaPL}, wireless communication links can be established using a very small TX power (and small bandwidth) to beyond 10 km in the RMa scenario at 73 GHz, implying a large coverage area for a very tall RMa base station. Moreover, the LOS PLE is 2.16, very close to the free space PLE of 2.0~\cite{Rap02a,Mac13a,Rap15b}; where the slightly higher PLE above free space may be caused by light foliage obstructions and misalignment between TX and RX antennas. The green diamonds in Figure~\ref{fig:RMaPL} indicate that diffraction by mountain edges may lead to a large amount of additional path loss in the LOS environment. The measured NLOS data render a NLOS PLE of 2.75, lower than the UMi and UMa cases published in~\cite{Sun16b,5GCM,3GPP.38.900} (where the PLE is between 2.9 and 3.2), which reveals a favorable propagation condition at mmWave frequencies in an RMa scenario when using such a tall TX antenna (110 m above the ground -- also called a ``boomer cell"). Fading variations over a few seconds were observed at some RX locations, with small fluctuations in received power that ranged from approximately a fraction of a dB about the average in LOS, and larger variations of $\sim \pm\:3-5\;\dB$ about the mean in NLOS due to foliage movement caused by wind.
\begin{figure}
\centering
\includegraphics[width=3.4in]{RMaPL_1km.eps}
\caption{73 GHz RMa path loss vs. T-R separation distance using a CI model with 1 m.}
\label{fig:RMaPL}
\end{figure}
From measured results depicted in Figure~\ref{fig:RMaPL}, the RMa CI path loss models for LOS and NLOS environments~\eqref{CI2} can be written as:
\begin{equation}\label{CILOS}
\begin{split}
\PL^{\CI}_{\LOS}(f_c,d)[\dB]=32.4+21.6\log_{10}(d)+20\log_{10}(f_c)\\
+\chi_{\sigma_{\LOS}},~\text{where}~d\geq 1~\m\text{, and }\sigma_{\LOS}=1.7\; \dB\\
\end{split}
\end{equation}
\begin{equation}\label{CINLOS}
\begin{split}
\PL^{\CI}_{\NLOS}(f_c,d)[\dB]=32.4+27.5\log_{10}(d)+20\log_{10}(f_c)\\
+\chi_{\sigma_{\NLOS}},~\text{where}~d\geq 1~\m\text{, and }\sigma_{\NLOS}=6.7\; \dB\\
\end{split}
\end{equation}
where the shadow fading standard deviations ${\sigma_{\LOS}}$ and ${\sigma_{\NLOS}}$ are 1.7 dB and 6.7 dB, respectively, according to the measured data. We note the UMa PLE in the CI model is not a function of carrier frequency when using a 1 m FSPL reference distance based on previous investigations~\cite{Sun16b,Haneda16a,Haneda16b}, thus the RMa PLE is also independent of frequency beyond 1 m. Furthermore,~\eqref{CILOS} and~\eqref{CINLOS} are based on measurements out to and beyond 10 km for both LOS and NLOS, whereas~\eqref{eq:RMaLOS} and~\eqref{eq:RMaNLOS} are limited to 10 km and 5 km, respectively. Measurements here show that RMa path loss~\eqref{CILOS},~\eqref{CINLOS} are valid for distances from 1 m to 11 km and frequencies from 500 MHz to 100 GHz. To match the existing 3GPP RMa model~\cite{3GPP.38.900}, one can increase the standard deviations in~\eqref{CILOS} and~\eqref{CINLOS} (e.g., to 4 dB and 8 dB for LOS and NLOS environments, respectively)~\cite{3GPP.38.900}. As this is the first in-depth empirical study to explore RMa path loss, more experiments would be valuable to verify the best RMa PLE and $\sigma$ model parameters. The CI models in~\eqref{CILOS} and~\eqref{CINLOS} have been validated by this 73 GHz measurement campaign, and prove that the CI model can accurately describe RMa path loss at mmWave bands, just as it has for UMi, UMa, and InH scenarios that are optional models in~\cite{3GPP.38.900}. These models are implemented in the NYUSIM open source 5G channel model simulator software~\cite{Samimi16a} as an optional model~\cite{3GPP.38.900}.
\section{Conclusions}\label{sec:conc}
This paper discussed the fundamentals of propagation path loss and demonstrated problems with the current RMa path loss model in 3GPP's standard above 6 GHz. We showed the rural macrocell (RMa) model used in~\cite{3GPP.38.900} is derived from urban models for below 6 GHz, and we presented an alternative model, the CI model, which has a solid physical basis, is simple, accurate over all frequencies, and is verified here. A first-of-its-kind RMa propagation measurement campaign at 73 GHz was conducted in a rural area to confirm the accuracy and validity of the proposed CI RMa model, while demonstrating the remarkable distances and coverage that may be obtained using mmWave communication beyond 10 km in an RMa scenario (the NLOS model in~\cite{3GPP.38.900} is limited to 5 km, and there was no effort to validate the 3GPP model, except for a few measurements to only 500 m). Using the measured results from the RMa campaign, the PLEs for LOS and NLOS environments were found to be 2.16 and 2.75, respectively, which were significantly different than the PLE values of 2.31 (LOS) and 3.04 (NLOS) found using simulated results based on the existing 3GPP RMa path loss model~\cite{3GPP.38.900}. This shows that the 3GPP model predicts greater path loss in LOS and NLOS compared to measured observations, as was also shown in~\cite{Sun16b,Thomas16a} for other 3GPP scenarios. The CI models in~\eqref{CILOS} and~\eqref{CINLOS} may be used for RMa from 500 MHz to 100 GHz, and users may wish to increase the standard deviations to 4 dB and 8 dB for LOS and NLOS, respectively, in order to match~\cite{3GPP.38.900}. It seems prudent for 3GPP to replace the current RMa models, which are shown to have originated from urban measurements, are unverified by rural measurements, fit poorly to measured data herein, and remain undefined above 9.1 GHz, with the RMa CI models given in~\eqref{CILOS} and~\eqref{CINLOS}. At the very least,~\eqref{CILOS} and~\eqref{CINLOS} should be optional in the 3GPP standard.
\section{Acknowledgments}
This work was supported by the NYU WIRELESS Industrial Affiliates Program, three National Science Foundation (NSF) Research Grants: 1320472, 1302336, and 1555332, and the GAANN Fellowship Program.
\bibliographystyle{abbrv}
|
\section{Related Work}\label{Section6}
Noise in networks has been widely studied since this phenomena influences the analysis results. Many papers ~\cite{borgatti2006robustness,yan2011finding, wang2014noise,kossinets2006effects} have studied how noise affects centrality measures. ~\cite{pfeiffer2011methods} outlined methods to measure centrality and cluster structures in networks with uncertain topologies. There have also been works to mine maximal cliques~\cite{mukherjee2013mining}as well as subgraph pattern matching~\cite{vesdapunt2014identifying} for uncertain graphs.~\cite{platig2013robustness} analyzed robustness of different network measures to different link errors. Attack vulnerability in presence of errors in networks is discussed in~\cite{Booker2012effects}. The entity resolution problem has also been tackled in the context of networks with limited information~\cite{moustafa2014subgraph,verroios2015entity}. Liu et al \cite{liu2015integrating} combines incomplete network discovery and community detection into a single framework. Authors in~\cite{kim2011network} have tried to complete a network by developing algorithms to infer missing nodes and links in a network.
\fi
\vspace{-.3cm}
\begin{figure}[!ht]
\centering
\begin{tabular}{c}
\includegraphics[scale=.2]{rank_football.eps}\\
\includegraphics[scale=.2]{railway_rank.eps} \\
\includegraphics[scale=.2]{lfr03_rank.eps}
\end{tabular}
\caption{The Jaccard Index between the top vertices of the original and the noisy networks for varying noise levels. We show the football (top), the railway (middle) and the LFR ($\mu=0.3$, bottom) networks respectively.}
\label{rank}
\vspace{-.5cm}
\end{figure}
\section{Conclusion}\label{Section7}
In this work, we have done rigorous experiments to understand the effect of noise in complex networks and compared different community scoring and centrality metrics in terms of sensitivity and reliability. A key observation is that in a majority of cases permanence worked better than all the other competing measures investigated. The central lesson is that while permanence is appropriately sensitive to different noise levels, the high permanence nodes are almost unaffected by the application of noise thus making the measure at the same time very reliable. \iffalse Our findings can be summarized as:
\begin{itemize}
\item We compared a large selection of community scoring and centrality metrics in noisy networks.
\item Permanence is appropriately sensitive to noise; this is immediately followed by closeness in many of the cases.
\item Permanence and in some cases closeness and betweenness are highly correlated to various structural features of a network.
\item Initiator selection based on permanence enables, in a majority of cases, a faster message broadcast compared to other metrics. This observation holds even when the underlying network is noisy. Betweenness and closeness based selections come immediately next to permanence based selection.
\item For the sensitivity as well as for a majority of the reliability experiments we obtain a partial ordering of the metrics that is relatively independent of the noise model and the type of the network. We believe that this finding shall be very helpful in making the appropriate choice of a metric in future applications dealing with noisy networks.
\end{itemize}
\fi
In future we would like to investigate the analytical reasons for the stability of high permanence nodes and, thereby, propose an algorithm to automatically identify the level of noise up to which this stability persists.
The data and the code are available in the public domain (\url{https://github.com/Sam131112/Noise_Models.git}).
\section{Permanence as a centrality metric}\label{definition}
We first provide a brief introduction to permanence. Permanence was introduced in ~\cite{chakraborty2014permanence} as a vertex-centric metric for evaluating the quality of communities in a network. If the community structure is known, permanence is defined for each vertex as follows;
$P(v) = [\frac{I(v)}{E_{max}(v)} \times \frac{1}{D(v)} - (1 - c_{in}(v)) ]$ \\
where $I(v)$ is the number of neighbours of $v$ in its community, $E_{max}(v)$ is the maximum number of connections of $v$ in its neighbouring community, $D(v)$ is the degree of $v$ and $c_{in}(v)$ is the internal clustering coefficient of the internal nodes of $v$ in the community. Permanence tries to formulate the ``pull'' experienced by a vertex $v$ from its external neighbors, i.e., neighbors residing in communities different from that of $v$. If this pull is low (low $E_{max}(v)$) and the internal neighbors in the community of $v$ are tightly connected (high $c_{in}(v)$) then there is a high probability that $v$ is very "`stable'' in its own community. In other words, the community assignment for $v$ is correct.
\begin{figure}[ht]
\centering
\begin{tabular}{c | c}
\includegraphics[width=.2\textwidth]{football_0} &
\includegraphics[width=.2\textwidth]{football_30}
\end{tabular}
\caption{Football network with ground truth community shown in palette of rainbow colors. On the left is the original network while on the right is the same network with 30\% of the edges removed uniformly at random.}
\label{foot}
\end{figure}
The value of permanence ranges from 1 (all neighbors form a cluster and are in the same community as the vertex) to -1 (all neighbors form a cluster and are in a different community than the vertex). Vertices in singleton communities have permanence zero. The permanence of a network is the average permanence of all its vertices. Therefore a network with more community-like structure will have higher permanence. For instance, the permanence of the node $v$ in Fig.~\ref{perm_toy} is equal to 0.12.
Since it is vertex-centric, permanence can also be considered as a centrality metric, where vertices in tightly bound clusters are marked to be important. Nodes with high permanence indicate that their neighbors form cliques or near-cliques. While clustering coefficient can also provide this information, it considers the entire set of neighbors around a node $v$. In contrast, permanence uses a restricted set, consisting only those neighbors that are part of the same community as $v$. This difference between ordinary clustering coefficient and permanence is illustrated in Fig.~\ref{perm_toy}; consider the two nodes $x$ and $y$ both of which have exactly same clustering co-efficient but different values of permanence. Therefore a high clustering coefficient does not indicate whether a vertex belongs to a community, whereas a high permanence means that the vertex belongs strongly within a community.
Permanence therefore provides a unique centrality value, that not only indicates the ranking of a vertex as per its importance, but also provides information about its position in a community. Because they belong to the core of communities, high permanence vertices are very effective as seeds for broadcasting messages and also as good bases for sampling. Moreover, because these vertices belong to tightly connected clusters their overall rankings are less affected by the varying levels of noise in the system. Therefore message spreading and sampling using high permanence vertices is more resistant to noise within the network
Interestingly, the mean permanence over {\em all} the vertices is very sensitive to the effect of noise. In Fig.~\ref{foot}, we illustrate this phenomena for the structure of the well-known football network~\cite{girvan2002community} (see Section~\ref{setup} for description). The left network shows the community structures present in the original network (each community is represented by a different color), while the right network shows the structure when 30\% of the edges have been removed uniformly at random. Note that the community labels for such moderate levels of perturbation to the network do not almost change. Further, while the standard quality metrics like modularity, conductance and cut-ratio cannot signal the differences between the left and the right network, permanence is the only metric that signals this difference. While from left to right network, average permanence significantly declines from 0.266 to 0.071, the other metrics hardly change -- modularity changes from 0.553 to 0.557, conductance changes from 0.402 to 0.399 and cut-ratio changes from 0.039 to 0.028\footnote{Note the scale-size is 2, 2, 1, 1 for permanence, modularity, conductance and cut-ratio respectively.}. Thus mean permanence over the network is very appropriately sensitive to noise as compared to the other metrics.
\section{Analysis of performance}\label{Section5}
\iffalse Our experiments show that among the tested parameters, permanence is the most sensitive to noise and also very effective at finding seed vertices for message passing in the majoirty of the cases.\fi In this section, we explore the characteristics of permanence that make it such a strong measure under noise.
\begin{figure}
\begin{tabular}{l}
\includegraphics[scale=.17]{football_perm_componts.eps} \\
\includegraphics[scale=.17]{football_mod_components.eps} \\
\end{tabular}
\caption{The variation in the different components of permanence and modularity when the noise levels are varied for the football network.}
\label{components1}
\vspace{-.5cm}
\end{figure}
\subsection{Sensitivity of permanence}
We compare the sensitivity of permanence with other community scoring metrics.
We break the permanence formula into two parts $PI=\frac{I(v)}{E_{max}(v)} \times \frac{1}{D(v)}$ and $c_{in}(v)$, and observe how they change for the different noise models. The results in Fig.~\ref{components1} show that $PI$ remains relatively constant, whereas the internal clustering coefficient is the major contributor to the change in permanence. When we contrast this result with the main factors in modularity (Fig.~\ref{components1}), namely the internal and external edges, we see that each factor remains relatively constant. A similar observation holds when we consider the other scoring metrics conductance and cut ratio.
\subsection {Rank of high permanence vertices under noise}
We compare the centrality metrics and permanence to check how their top ranking vertices alter under noise models. We identify the top 20 of the high valued vertices for each metric. Then for each noise level we compute the new top ranked vertices. We compute the Jaccard Index~\cite{gower1985measures} between the original vertex set and the new one obtained from the noisy network. A high Jaccard index (maximum value 1) indicates that most of the top ranked vertices are retained under noise, and a low value (minimum 0) indicates that the set has changed completely. As can be seen in Fig.~\ref{rank}, the Jaccard Index deteriorates much more slowly for permanence, than the other centrality metrics in most of the cases. This indicates that the ids of the high valued permanence vertices remain relatively constant under noise.
These experiments together provide a rationale of why permanence is effective both in evaluating noise through sensitivity and also reliable for message spreading.\iffalse The sensitivity experiments show that the internal clustering coefficient is most affected by noise can serve as an appropriate indicator of its level. The experiment related to ranking shows that the high permanence vertices retain their high ranks for low levels of noise. Thus, they still retain their relative centrality to the core of the community and thereby are effective for message broadcasting, even under noisy networks.\fi
\section{Experimental setup}\label{Section2}
\noindent {\bf Datasets.}
Here is a brief description of the different networks we used (see Table~\ref{tab:my_label} for properties of real-world networks).
\noindent{\bf LFR Benchmark:} We use the benchmark {\bf LFR model} \cite{lancichinetti2009benchmarks} that generates different networks and ground-truth communities of various quality. We use $n=1000$, $\mu=0.3$ keeping all other parameters to the default values of the original implementation~\footnote{\url{https://sites.google.com/site/santofortunato/inthepress2}}
\noindent{\bf Railway:} has been taken from Ghosh et al.~\cite{ghosh2011statistical}.
\noindent{\bf Football:} has been taken from Girvan et al.~\cite{girvan2002community}.
\begin{table}
\centering
\resizebox{6.5cm}{!}{
\begin{tabular}{|c | c | c | c | c | c | c | c |}
\hline
Network & \#Nodes & \#Edges & \textless k\textgreater & $k_{max}$ & $|c|$ & $n_{c}^{max}$ & $n_{c}^{min}$ \\ \hline
Football & 115 & 613 & 10.57 & 12 & 12 & 5 & 13 \\ \hline
Railway & 301 & 1224 & 6.36 & 48 & 21 & 1 & 46 \\ \hline
\end{tabular}}
\caption{Dataset statistics. $|c|$ denotes the number of communities in the ground-truth, $n_{c}^{min}$ and $n_{c}^{max}$ denote the number of nodes in the smallest and the largest size communities respectively.}
\label{tab:my_label}
\vspace{-.8cm}
\end{table}
\noindent{\bf Noise Models.}
We experiment with three noise models -- \textbf{uniform}, \textbf{crawled} and \textbf{censored} (see~\cite{yan2011finding} for detailed description), -- to simulate real-world sources of noise. We do not allow formation of disconnected components while introducing noise. We vary noise levels in steps of 2\% from $0$ to $30$ in all our experiments.
\iffalse \begin{enumerate}
\item {\bf Uniform noise model:} This models the uniform uncertainty of the presence/absence of an edges. To simulate this model we remove edges uniformly at random till a desired number of edges are removed. The number of edges removed is regulated by a user-defined parameter called the noise level.
\item {\bf Crawled noise model:} We simulate this noise model by executing a BFS (breadth-first search) from the node that has the highest closeness centrality and continue until the missing edges reach a pre-defined noise level.
\item {\bf Censored noise model:} This model limits the number of connections per node~\citep{zhu2012uncovering}. At each step, we randomly delete from the original network an edge from the highest degree node till the required noise level is reached.
\end{enumerate}
\fi
\noindent{\bf Metrics.} Our set of network parameters for evaluation include community-scoring metrics namely, modularity, cut-ratio, and conductance, and centrality metrics namely betweenness, closeness and Pagerank. For the definitions of these metrics the reader is referred to~\cite{newman2010networks}. We also include a recently introduced metric permanence~\cite{chakraborty2014permanence}, that serves both as a community scoring function as well as a measure of centrality.
\iffalse
\begin{figure}[ht]
\centering
\begin{tabular}{l}
\includegraphics[width=0.49\textwidth]{correl_plot2.eps}
\\
\includegraphics[width=0.49\textwidth]{new_correl1.eps}
\end{tabular}
\caption{Correlation of permanence, conductance, cut-ratio, modularity, closeness, pagerank and betweenness with the different network features. The top panel are the results for the real-world networks (football on the left and railway on the right) and the bottom panel is for LFR networks ($\mu=0.1$ and $\mu=0.3$ respectively).}
\label{correlation}
\vspace{-.5cm}
\end{figure}
\fi
\section{Introduction}
Network analysis has become an ubiquitous tool for understanding the behavior of various complex systems~\cite{Mitchell06complexsystems:}. The vertices in the network represent the entities of the complex system and the edges represent their pairwise interactions.
However, in the practical context, due to the limitations in data gathering, not all interactions can be observed. Consequently, the network can be potentially incomplete, as manifested by missing edges. It is therefore important to determine the effect of this incompleteness or noise on different network parameters and rank them according to how they behave under noise.
In this paper we study the effect of noise on two important classes of network analysis metrics -- (i) centrality measures and (ii) community scoring functions. Centrality measures are key to applications that rely on node ranking, and the community scoring functions determine the quality of clusters/communities that are used in many applications requiring unsupervised classification.
We evaluate these metrics based on two orthogonal qualities. The first is {\bf sensitivity}, that is whether the change in the value of the metric is commensurate with percentage of edges removed. If a metric is sensitive then it can serve as a good indicator of how much the network has been changed (see Section ~\ref{Section3}).
The second metric is {\bf reliability}, that is whether certain operations in the network can be performed efficiently in spite of the missing edges. If a metric is reliable then it guarantees good performance even under noise. Here we select message spreading as the candidate operation. The seed nodes that initiate message spreading are selected from the high valued entities of different metrics. A metric has high reliability if the time for message spreading under noise does not significantly degrade if its high valued entities are selected as seeds (see Section~\ref{Section4}). In contrast to previous work~\cite{borgatti2006robustness,yan2011finding,kossinets2006effects}, which focused on single noise models and primarily on centrality metrics, to the best of our knowledge this is the first comparative study encompassing several centrality and community-scoring parameters and different types of noise models.
{\bf Overview of Experiments (Section~\ref{Section2})} Among the centrality measures we consider closeness, betweenness and Pagerank and among the community scoring functions we consider modularity, cut-ratio and conductance. We also include a third type of metric, {\em permanence}~\cite{chakraborty2014permanence}. Although permanence is a community scoring metric, unlike the others it is vertex based. Therefore permanence can also be considered as a centrality measure.
We apply three different noise models on real-world and synthetic networks. We empirically evaluate the above metrics to estimate their sensitivity to varying levels of noise. We also measure their reliability by observing whether high valued vertices of these metrics can serve as effective seeds for message spreading.
In all our experiments, we ensure that in spite of the noise, the underlying community structure is not significantly disrupted from its original form and the giant component of the network remains connected. Nevertheless, as we shall see, even this constrained noise can significantly affect the analysis.
\noindent{\bf Key Results}
(i) For both the objectives -- sensitivity and reliability and for all the given noise models and networks, permanence proves to be the most sensitive and most reliable metric in majority of the cases. (ii) The other centrality metrics can be ranked in a partial order. The only other metric that exhibits sensitivity is closeness. For reliability, when a difference in the performance can be observed, closeness and betweenness also show high reliability. (iii) For all sensitivity experiments, and for most reliability experiments, the partial ordering of metrics is relatively independent of the noise model and type of network. Community scoring metrics,apart from permanence, are not sensitive.
\iffalse
\begin{enumerate}
\item For both the objectives -- sensitivity and reliability and for all the given noise models and networks, permanence proves to be the most sensitive and most reliable metric in majority of the cases.
\item The other centrality metrics can be ranked in a partial order. The only other metric that exhibits sensitivity is closeness. For reliability, when a difference in the performance can be observed, closeness and betweenness also show high reliability.
\item For all sensitivity experiments, and for most reliability experiments, the partial ordering of metrics is relatively independent of the noise model and type of network. Community scoring metrics are not sensitive.
\end{enumerate}
\fi
{\bf Rationale for the behavior of permanence (Section ~\ref{Section5}).} At a quick glance it would seem that sensitivity and reliability are mutually opposing properties. Sensitivity is used as an indicator of noise, whereas reliability is used to guarantee good performance in spite of noise. It is therefore surprising that permanence is both the most sensitive as well as the most reliable among all the metrics that we investigate.
We believe that this is because permanence encompasses both community-like and centrality-like properties. While the cumulative value of permanence is sensitive to the level of noise, satisfying the sensitivity criterion, its high rank vertices are stable under noise and therefore serve as effective seeds in the noisy versions of the networks. We compute the Jaccard Index (JI) for the high ranked vertices between the original and noisy networks. Permanence exhibits the highest JI and therefore the set of its high ranked vertices change the least.
\begin{figure}
\begin{tabular}{l}
\includegraphics[scale=.2]{Football_sensitivity.eps} \\ \includegraphics[scale=.2]{Railway_sensitivity.eps}
\\
\includegraphics[scale=.2]{lfr03_sensitivity.eps}
\end{tabular}
\caption{Sensitivity of the different quality metrics for varying levels of noise (in steps of 2\%) with x-axis as the noise level and y-axis as the metric values. The first and the second panels are for the football and railway the networks respectively. The third panel is for the LFR network ($\mu=0.3$).}
\label{sense}
\vspace{-.5cm}
\end{figure}
\section{Introduction}
\bibliographystyle{abbrv}
\section{Sensitivity of the metrics}\label{Section3}
A {\em sensitive} parameter is one whose change is commensurate with the amount of noise applied. For small amounts of noise, the change in the parameter values should be low, whereas, as the noise increases, the change should be much higher. A sensitive parameter can function as a good indicator of whether a network significantly changed from its original topology.\\
{\bf Our goal is to rank the network parameters by the extent to which they are sensitive to the noise level.}
{\bf Methodology.} We apply the three noise models on the one synthetic LFR network ($\mu = 0.3$), and two real-world, railway and football, networks. For each increasing level of noise we compute the value of the parameters. For the vertex-based metrics we take the average over all vertices. We compute the value of the community-scoring parameters based on the ground-truth community assignment from the original network. Our rationale is that because community detection is expensive, therefore, re-computing the community after each noise addition would defeat the purpose of quickly ascertaining the change in the network. Further, our selected noise level is low enough such that it does not significantly change the original ground-truth community. Our results are averaged over ten simulation runs.
{\bf Results.} The results in Fig.~\ref{sense} show that the change in permanance has the highest slope with respect to increasing noise. This indicates that permanence is most sensitive to noise as compared to the other parameters. However, there are some cases, e.g., the football network where the closeness centrality is also quite sensitive. While betweeness is slightly sensitive in the LFR networks, it shows an opposite trend, i.e. increase in value with noise for censored noise in the real-world networks. The rest of the metrics remain constant. We report the range of the average of each metric obtained for each noise model as a tuple -- (average metric value at 2\% noise level, average metric value at 30\% noise level) -- in Table~\ref{range}. In this table, permanence shows the largest separation.
\begin{table}
\centering
\resizebox{8.5cm}{!}{
\begin{tabular}{l}
\begin{tabular}{llll}
\hline
Metrics & Railway & Football & LFR (0.3) \\
Permanence & {\bf (0.12, -0.08)} & {\bf (0.261, 0.091)} & {\bf (0.121, 0.003)} \\
Closeness & (0.019, 0.0048) & (0.176, 0.041) & (0.336, 0.301) \\
Betweenness & (0.038, 0.043) & (0.394, 0.358) & (0.12, 0.1099) \\
Pagerank & (0.0033, 0.0034) & (0.008, 0.008) & (0.001, 0.001) \\
Modularity & (0.467, 0.464) & (0.555, 0.557) & (0.661, 0.658) \\
Conductance & (0.552, 0.551) & (0.407, 0.4003) & (0.303, 0.3) \\
CutRatio & (0.012, 0.0098) & (0.038, 0.027) & (0.004, 0.003)\\
\end{tabular}
\\
\hline
\begin{tabular}{llll}
Permanence & {\bf (0.116, -0.028)} & {\bf (0.355, 0.165)} & {\bf (0.109, -0.09)} \\
Closeness & (0.02, 0.004) & (0.176, 0.037) & (0.336, 0.293) \\
Betweenness & (0.041, 0.065) & (0.39, 0.40) & (0.162, 0.303) \\
Pagerank & (0.0033, 0.0034) & (0.008, 0.008) & (0.001, 0.001) \\
Modularity & (0.467, 0.502) & (0.555, 0.548) & (0.661, 0.659) \\
Conductance & (0.548, 0.513) & (0.406, 0.401) & (0.303, 0.301) \\
CutRatio & (0.012, 0.0098) & (0.038, 0.027) & (0.004, 0.003)
\end{tabular}
\\
\hline
\begin{tabular}{llll}
Permanence & {\bf (0.09, -0.016) } & {\bf(0.25, 0.066)} & {\bf (0.129, -0.097)} \\
Closeness & (0.02, 0.001) & (0.183, 0.047) & (0.336, 0.137) \\
Betweenneness & (0.038, 0.031) & (0.391, 0.198) & (0.106, 0.072) \\
Pagerank & (0.0033, 0.0033) & (0.008, 0.008) & (0.001, 0.001) \\
Modularity & (0.451, 0.4) & (0.549, 0.526) & (0.657, 0.644) \\
Conductance & (0.583, 0.576) & (0.407, 0.441) & (0.304, 0.327) \\
CutRatio & (0.012, 0.0098) & (0.038, 0.027) & (0.004, 0.003)
\\
\hline
\end{tabular}
\end{tabular}}
\caption{Range of metric values for uniform noise ($1^\textrm{st}$ row), censored noise ($2^\textrm{nd}$ row) and crawled noise ($3^\textrm{rd}$ row). The numbers in bold shows the range that has the largest separation.}
\label{range}
\vspace{-.8cm}
\end{table}
\iffalse
\subsection{Relation of sensitivity to network structure}
We study how structural properties of the network affects the sensitivity of the metric. To do this, we compute the correlation of the tested metrics with several structural features of the network. The definitions of standard metrics like clustering coeefficient, effective diameter (eff-dia), diameter (dia), degree disassortivity (dis-deg), average path length (avg path) can be obtained from~\cite{newman2010networks}. Non-standard features used are robustness, inter edges, intra edges, inter-intra edges. The definitions are as follows.
\begin{itemize}
\item {\em Robustness} is defined as average sum of inverse geodesics between all pairs of nodes in the noisy network divided by average sum of inverse geodesics in the original network \cite{sur2015attack,scellato2011understanding}.
\item {\em Inter edges} are defined as edges interlinking communities expressed as a fraction of all edges.
\item {\em Intra edges} are defined as edges internal to a community expressed as a fraction of all edges.
\item {\em Inter-intra} is defined as the ratio of inter over intra edges.
\end{itemize}
The correlation values are calculated as follows: Considering uniform noise model, for each network we apply varying noise levels (1\% to 30\%),
and create 10 networks for each noise level. The mean value of the metrics at each level are computed. We also compute the structural properties of the network and take their mean to represent that particular noise level. Thus each metric (closeness, permanence, etc.) and each structural feature (edge connectivity, robustness, etc.), have a vector associated with them. The values in the vector represent how each parameter changed with increasing noise level. We now do a pair-wise correlation between each element from the parameter list with each element for the structural list. The correlation between the different entities are given in Fig.~\ref{correlation}.
We see that all the three metrics that showed any sensitivity, i.e. permanence, betweenness and closeness, are highly (positively or negatively) correlated with intrinsic network features such as diameter and average path length. Betweenness which showed less sensitivity actually has lower correlation to these structural features.
In contrast, modularity and conductance are more correlated to features that are defined by the community structure, such as intra and inter edges.
We note that even though permanence is based on community assignment, it is not well correlated to any of the community dependent features. We believe that this is due to the local effect of permanence. Unlike modularity, it does not consider all edges within or outside a community, but only those to which it is tightly connected.
A question might arise that if permanence, betweenness and closeness are strongly correlated to network features why does permanence show the most sensitivity. This is because as shown in Table~\ref{range}, the range between the maximum and minimum value of permanence is much higher than the other two metrics. Also note that while closeness and betwenness are global metrics, permanence is localized within a community and its neighboring vertices.
\fi
\section{Reliability of the Metrics}\label{Section4}
In message spreading~\cite{chierichetti2010rumour}, a set of source vertices (seed nodes) start sending a message. At every time step, a vertex containing the message transfers the message uniformly at random to one of its neighbors who does not have the message. The algorithm terminates when
all vertices have received the message. The selection of the seed nodes is critical to how quickly the message spreads. A {\em reliable} metric is one whose high ranked nodes, if used as seeds, can spread the message quickly even under noise.
{\bf Methodology.} For each of the centrality metrics, closeness, betweeeness and Pagerank, and also for permanence we select a small fraction of the highest ranked nodes as the seed. We also select seeds (a) uniformly at random and (b) based on highest degree as baselines for the spreading experiments.
For different levels of noise, we compute the the number of iterations required to broadcast the message in the whole network and compare the values across the different centrality metrics, permanence and the two baselines.
\begin{figure}
\centering
\begin{tabular}{l}
\includegraphics[scale=.2]{lfr03_spread_new.eps} \\
\includegraphics[scale=.2]{railway_spread_new.eps} \\
\includegraphics[scale=.2]{football_spread_new.eps}
\end{tabular}
\caption{Time required to broadcast a message for different seed node selection mechanisms (permanence, closeness, betweenness, Pagerank) and varying noise levels (in steps of 2\%). LFR network ($\mu = 0.3$, first panel), railway network (second panel) and football network (third panel). The first column of figures represents results for uniform noise, the middle column represents results for censored noise and the last column represents results for the crawled noise.}
\label{spreading}
\vspace{-.5cm}
\end{figure}
{\bf Results:} In Fig.~\ref{spreading}, we plot the time required to broadcast for different levels of noise. For each noise level the results are averaged over ten different runs. The results can be divided into three groups.
{\em All metrics perform equally well:} Crawled noise in the LFR and the football network.
{\em Metrics perform differently based on the noise level, but no clear winner:} Uniform noise in the railway and football networks and censored noise in the football.
{\em One metric performs better in most of the noise levels:} For uniform and censored noise for the LFR network and censored noise for the railway network permanence takes the least time to spread messages. For crawled noise in the railway network, betweenness takes the least time.
We therefore see that for the . Therefore vertices with high betweenness centralities would be key connection points. This feature is exaggerated in the crawled noise since the network created using BFS-search has become tree-like.
|
\section{Introduction}
Consider a symmetric space $X=G/K$ of noncompact type, $G$ being a connected noncompact semisimple lie group with finite center and $K$
a maximal compact subgroup. Let $g=k+p$ be the corresponding Cartan decomposition, $p$ being the orthocomplement of $k$ relative to the
killing form of $g$. Let $a\subset p$ be a maximal abelian subspace. Let $G_{0}$ be the Cartan Motion group. This
group is defined as the semidirect product of $K$ and $p$ with respect to the adjoint action of $K$ on $p$. The $X_{0}=G_{0}/K$ is naturally
identified with the Euclidean space $p$. The element $g_{0}=(k,Y)$ actions on $p$ by
$g_{0}(Y^{'})=Ad(k)Y^{'}+Y \ k\in K,Y,Y^{'}\in p$,
So the algebra $\mathbb{D}(X_{0})$ of $G_{0}$-invariant differential operators on $X_{0}$ is identified with the algebra of $Ad(K)$-invariant
constant coefficient differential operators on $p$. The corresponding spherical functions on $X_{0}$ are given by
$\psi_{\lambda}(Y)=\int_{k}e^{i\lambda(Ad(k)Y)}dk \ \lambda\in a_{c}^{*}$
and $\psi_{\lambda}=\psi_{\mu}$ if and only if $\lambda$ and $\mu$ are $W$-conjugate. See e.g.[2],IV\S 4. Again, the maximal ideal space of
$L^{\natural}(G_{0})$ is up to $W$-invariance identified with the set of $\lambda$ in $a_{c}^{*}$ for which $\psi_{\lambda}$ is bounded. Since
$\rho$ is relative to the curvature of $G/K$ it is natural to expect the bounded $\psi_{\lambda}$ to come from replacing $c(\rho)$ by the origin,
where $c(\rho)$ is for the semisimple case also proved by Sigurdur Helgason [3]. In the words, $\psi_{\lambda}$ is would be expected to be bounded
if and only if $\lambda$ is real, that is $\lambda\in a^{*}$. In [1], Sigurdur Helgason proved when $G$ is complex, the spherical function $\psi_{\lambda}$ on $G_{0}$ is bounded if and only if $\lambda$ is real, i.e. $\lambda\in a^{*}$ mainly by using two results proved by Harish-Chandra
[4] and [5]. In this paper, I use a different way to prove when $G$ is real, the spherical function $\psi_{\lambda}$ on $G_{0}$ is bounded also
if and only if $\lambda$ is real. In this way, we generalize Sigurdur Helgason's results.
In Sigurdur Helgason's another paper [6], he considered Eigenspaces of the Laplacian on $\mathbb{R}^{n}$. Let $L$ denote the usual Laplacian on $\mathbb{R}^{n}$ and for each $\lambda \in \mathbb{C}$ let $\mathcal{E}_{\lambda}(\mathbb{R}^{n})$ denote the eigenspace
$\mathcal{E}_{\lambda}(\mathbb{R}^{n})=\{f\in \mathcal{E}(\mathbb{R}^{n})\mid Lf=-\lambda^{2}f\}$ with the topology induced by that of $\mathcal{E}(\mathbb{R}^{n})$. Let $G$ denote the group of all isometries of $\mathbb{R}^{n}$, and $K$ the group of rotations $O(n)$. Sigurdur Helgason mainly proved the natural action of $G$ on $\mathcal{E}_{\lambda}(\mathbb{R}^{n})$ is irreducible if and only if $\lambda\neq 0$. I will
prove when $K\subset O(n)$ is transitive on the spheres about 0 in $\mathbb{R}^{n}$,$n>1$ instead of $O(n)$ and $G=K\rtimes \mathbb{R}^{n}$, the same results holds. In this way, we do a further step of this kind of problem. Meanwhile, I will specific all the groups $K$ which is transitive on the spheres about 0 in $\mathbb{R}^{n}$.
Finally, according to [7], we know when $K\subset O(n)$ is transitively on the spheres about 0, the specific form of the spherical functions on
$K\rtimes \mathbb{R}^{n}/K\cong \mathbb{R}^{n}$. Then I will give a estimation for it when $r\rightarrow \infty$.
\section{The main theorem}
The notion of induced spherical function mirrors the notion of induced representation. Let $Q\subset G$ be a closed subgroup such that $K$ is transitive on $G/Q$, i.e.$G=KQ$,i.e.$G=QK$,i.e. $Q$ is transitive on $G/K$. Let $\zeta: Q\rightarrow \mathbb{C}$ be spherical for $(Q,Q\bigcap K)$.
The induced spherical function is
$[Ind_{Q}^{G}(\zeta)](g)=\int_{K}\widetilde{\zeta}(gk)d\mu_{K}(k)$ \ where $\widetilde{\zeta}(kq)=\zeta(q)\Delta_{G/Q}(q)^{-\frac{1}{2}}$
Here $\Delta_{G/Q}:Q\rightarrow \mathbb{R}^{n}$ is the quotient of modular functions, $\Delta_{G/Q}(q)=\Delta_{G}(q)/\Delta_{Q}(q)=\Delta_{Q}(q)^{-1}$
\begin{theorem}\label{equal}
Let $\lambda\in a_{c}^{*}$. Then $\psi_{\lambda}(Y)$ is the induced spherical function $Ind_{p}^{G}(\varphi_{\lambda})$, where
$\varphi_{\lambda}(Y)=e^{i\lambda(Y)}$ for every $Y\in a\subset p$.
\end{theorem}
\begin{proof}
Apply above formula to $\varphi_{\lambda}$ with $Q=P$. Since $G=K\rtimes p$ and $p$ are unimodular, it says that the induced spherical function is given by
$Ind_{p}^{G}(\varphi_{\lambda})(k,Y)=\int_{K}\varphi_{\lambda}(Ad(k_{0})Y)dk_{0}=\int_{K}e^{i\lambda(Ad(k_{0})Y)}dk_{0}=\int_{K}e^{i\lambda(Ad(k)Y)}dk
=\psi_{\lambda}(Y)=\psi_{\lambda}(k,Y)$ \ for\ $Y\in p, k\in K$. Note that $(k_{0},0)(k,Y)=(k_{0}k,Ad(k_{0})Y)$.
\end{proof}
We apply the Mackey little group method to $G$ relative to its normal subgroup $p$:If $\psi$ is an irreducible unitary representation of $G$,then it can be constructed (up to unitary equivalence) as follows: If $\varphi_{\lambda}(Y)=e^{i\lambda(Y)}$, where $\lambda \in p^{*}$, $Y\in p$, let
$K_{\varphi_{\lambda}}=\{k\in K\mid \varphi_{\lambda}(Ad(k)Y)=\varphi_{\lambda}(Y) \ \forall Y\in p\}$. $K_{\varphi_{\lambda}}$ is a closed subgroup of $K$. Let $G_{\varphi_{\lambda}}=K_{\varphi_{\lambda}}\rtimes p$. Write $\widetilde{\varphi_{\lambda}}$ for the extension of $\varphi_{\lambda}$ to
$G_{\varphi_{\lambda}}$ given by $\widetilde{\varphi_{\lambda}}((k,Y))=\varphi_{\lambda}(Y)$. If $\gamma$ is an irreducible unitary representation of $K_{\varphi_{\lambda}}$, let $\widetilde{\gamma}$ denote its extension of $G_{\varphi_{\lambda}}$ given by
$\widetilde{\gamma}((k,Y))=\gamma(k)$. Denote $\psi_{\varphi_{\lambda},\gamma}=Ind_{G_{\varphi_{\lambda}}}^{G}(\widetilde{\varphi_{\lambda}}\otimes \widetilde{\gamma})$, then there exist choices of $\varphi_{\lambda}$ and $\gamma$ such that $\psi=\psi_{\varphi_{\lambda},\gamma}$.
\begin{theorem}\label{equal}
In the notation above, $\psi_{\varphi_{\lambda},\gamma}$ has a $K$-fixed vector is given(up to scalar multiple by $u((k,Y))=e^{-i\lambda(Ad(k^{-1})Y)}$, if $\varphi_{\lambda}=e^{i\lambda(Y)}$.
\begin{proof}
The representation space $H_{\psi}$ of $\psi=\psi_{\varphi_{\lambda},\gamma}$ consists of all $L^{2}$ functions $f:G\rightarrow H_{\psi}$
such that $f(g^{'}(k^{'},x^{'}))=\gamma(k^{'})^{-1}\varphi_{\lambda}(x^{'})^{-1}f(g^{'})$ for $g^{'}\in G$, $x^{'}\in p$, $k^{'}\in K_{\varphi_{\lambda}}$, and $\psi$ acts by $(\psi(g)f)(g^{'})=f(g^{-1}g^{'})$.
Now suppose that $0\neq f\neq \in H_{\psi}$ is fixed under $\psi(K)$. If $k^{'}\in K_{\varphi_{\lambda}}$, then $\gamma(k^{'})f(1)=f(1)$. If $f(1)=0$, then $f(G_{\varphi_{\lambda}})=0$ and $K$-invariance says $f=0$, contrary to the assumption. Thus $f(1)\neq 0$ and irreducibility of $\gamma$ forces $\gamma$ to be trivial.
Conversely, if $\gamma$ is trivial, then $u((k,Y))=e^{-i\lambda(Ad(k^{-1})Y)}$ is a nonzero $K$-fixed vector in $H_{\psi}$. And it is the only one, up to scalar multiple, because any two $K$-fixed vectors must be proportional.
\end{proof}
\end{theorem}
\begin{lemma}\label{twist}
In the notation above, $Ind_{G_{\varphi_{\lambda}}}^{G}(\widetilde{\varphi_{\lambda}})$ is unitary equivalent to the subrepresentation of $Ind_{p}^{G}(\widetilde{\varphi_{\lambda}})$ generated by the $K$-fixed unit vector $u((k,Y))=e^{-i\lambda(Ad(k^{-1})Y)}$, if $\varphi_{\lambda}=e^{i\lambda(Y)}$.
\end{lemma}
\begin{theorem}\label{equal}
Let $\varphi$ be a $(K\rtimes p, K)$-spheirical function. Then $\varphi$ is positive definite if and only if it is of the form $\psi_{\lambda}$ for some $\lambda\in a^{*}$. Further, if $\lambda,\lambda^{'}\in a^{*}$, then $\psi_{\lambda}=\psi_{\lambda^{'}}$ if and only if $\lambda^{'}\in Ad(k)\lambda$.
\end{theorem}
\begin{proof}
Let $\lambda\in a^{*}$. $\sum_{i,j}e^{i\lambda(Ad(k)(-Y_{j}+Y_{i}))}\overline{c_{j}}c_{i}=\sum_{i,j}(e^{-i\lambda(Ad(k)Y_{j})}\overline{c_{j}})e^{i\lambda(Ad(k)Y_{i})}c_{i}
=(\sum_{i}e^{i\lambda(Ad(k)Y_{i})}c_{i})\overline{(\sum_{j}e^{i\lambda(Ad(k)Y_{j})}c_{j})}\geqslant 0$.
Since $\psi_{\lambda}$ is a limit of non-negative linear combinations of positive definite functions on $\mathbb{R}^{n}$, so it is positive definite.
Now let $\varphi$ be a positive definite $(G,K)$-spherical function. Let $\Pi_{\varphi}$ be the associated irreducible unitary representation, and $H_{\varphi}$ the representation space, such that there is a $K$-fixed unit vector $u_{\varphi}\in H_{\varphi}$ and let $\varphi(g)=<u_{\varphi},\prod_{\varphi}(g)u_{\varphi}>$ for all $g\in G$. Following the discussion of the Mackey little group method, and Theorem 2.2 , we have $\varphi_{\lambda}(Y)$ for some $\lambda\in a^{*}, Y\in p$, s.t. $\Pi_{\varphi}$ is unitarily equivalent to $Ind_{G_{\varphi_{\lambda}}}^{G}(\widetilde{\varphi_{\lambda}})$. Making the identification one, $K$-fixed unit vector in $H_{\varphi}$ is given by
$u((k,Y))=e^{-i\lambda(Ad(k^{-1})Y)}$. We have $\lambda\in a{*}$ s.t.$\varphi_{\lambda}=e^{i\lambda(Y)}$ and from above several Theorems and Lemma, we compute:
$\varphi(Y)=<u,\Pi_{\varphi}(Y)u >$
=$\varphi(Y)=<u,Ind_{G_{\varphi_{\lambda}}}^{G}(\widetilde{\varphi_{\lambda}})(Y)u >$
=$\varphi(Y)=<u,Ind_{G_{p}}^{G}(\varphi_{\lambda})(Y)u >$
=$Ind_{G_{p}}^{G}(\varphi_{\lambda})(Y)$
=$\psi_{\lambda}(Y)$
=$\int_{K}e^{i\lambda(Ad(k)Y)}dk$
For the second, if $\lambda^{'}=Ad(k_{0})\lambda$ for some $k_{0}\in K$, we have:
$\psi_{\lambda^{'}}(Y)=\int_{K}e^{i\lambda^{'}(Ad(k)Y)}dk=\int_{K}e^{iAd(k_{0})\lambda(Ad(k)Y)}dk$
=$\int_{K}e^{i\lambda(Ad(k_{0}^{-1})Ad(k)Y)}dk=\int_{K}e^{i\lambda(Ad(k)Y)}dk=\psi_{\lambda}(Y)$.
Conversely, suppose that $\lambda^{'},\lambda\in a^{*}$ with $\psi_{\lambda^{'}}=\psi_{\lambda}$. Then
(up to unitary equivalence)$Ind_{G_{\varphi_{\lambda}}}^{G}(\widetilde{\varphi_{\lambda}})$=
$Ind_{G_{\varphi_{\lambda^{'}}}}^{G}(\widetilde{\varphi_{\lambda^{'}}})$. That gives us direct integral decompositions
$\int_{K}^{\bigoplus}\psi_{Ad(k)\lambda}dk=Ind_{G_{\varphi_{\lambda}}}^{G}(\widetilde{\varphi_{\lambda}})\mid p$
=$Ind_{G_{\varphi_{\lambda^{'}}}}^{G}(\widetilde{\varphi_{\lambda^{'}}})\mid p=\int_{K}^{\bigoplus}\psi_{Ad(k)\lambda^{'}}dk$
\end{proof}
\begin{theorem}\label{equal}
If $N$ is an $n$-step group with $n\geq 3$, then there are no Gelfand pairs $(K,N)$, where $K\in Aut(N)$.
\end{theorem}
\begin{theorem}\label{equal}
We first consider $K$-spherical functions associated to a Gelfand pair $(K,N)$.
Suppose $\phi$ is a bounded $K$-spherical function on $N$. Then there is a $\pi\in \hat{N}$ and a unit vector $\xi\in H_{\pi}$ such that
\begin{equation}\label{l-invariant elements}
\phi(x)=\int_{K}<\pi(k.x)\xi,\xi>dk
\end{equation}
for each $x\in N$
\end{theorem}
\begin{proof}
Let $\lambda_{\phi}:L_{K}^{1}(N)\rightarrow \mathbb{C}$ be given by integration against $\phi$.
Since $L^{1}(N)$ is a symmetric Banach *-algebra,[8],there is a representation $\bar{\pi}$ of $L^{1}(N)$ and a one-dimensional subspace $H_{\phi}$ of $H_{\bar{\pi}}$ such that $(\bar{\pi}\mid_{L_{K}^{1}(N)},H_{\phi})$ is equivalent to $(\lambda_{\phi},\mathbb{C})$. As $\lambda_{\phi}$ is irreducible, the extension $\bar{\pi}$ is also irreducible(cf.[9]). Using approximate identities at each point of $N$, one can show that $\bar{\pi}$ is the integrated version of some $\pi\in \hat{N}$, with $H_{\pi}=H_{\bar{\pi}}$.
Choose $\xi\in H_{\phi}$ with $\left \|\xi \right \|=1$. Then for each $f\in L_{K}^{1}(N)$, $\pi(f)\xi=\lambda_{\phi}(f)\xi$, so that
\begin{equation}\label{l-invariant elements}
\begin{split}
&<\phi,f >=\lambda_{\phi}(f)=<\pi(f)\xi,\xi>\\
&=\int_{N}f(x)<\pi(x)\xi,\xi>dx\\
&=\int_{K}\int_{N}f(k^{-1}.x)<\pi(x)\xi,\xi>dxdk\\
\end{split}
\end{equation}
since $f$ is $K$-invariant
\begin{equation}\label{l-invariant elements}
=\int_{K}\int_{N}f(k.x)<\pi(x)\xi,\xi>dxdk
\end{equation}
Since $\phi$ is $K$-invariant, we change the order of integration and obtain
\begin{equation}\label{l-invariant elements}
\phi(x)=\int_{K}<\pi(x)\xi,\xi>dk
\end{equation}
\end{proof}
A complex-valued continous function $\phi$ on a locally compact group $G$ is called positive definite if
$\sum_{i,j=1}^{n}\phi(x_{i}^{-1}x_{j})\alpha_{i}\overline{\alpha_{j}}\geq 0$ for all finite sets
$x_{1},\ldots ,x_{n}$ of elements in $G$ and any complex numbers $\alpha_{1},\ldots ,\alpha_{n}$.
\begin{theorem}\label{equal}
For Gelfand pair $(K,N)$, where $N$ is at most 2-step nilpotent lie group, if $\phi$ is a bounded $K$-spherical function on $N$ is and only if $\phi$ is positive definite.
\end{theorem}
\begin{proof}
If $\phi$ is a bounded $K$-spherical function on $N$, for all finite sets
$x_{1},\ldots ,x_{n}$ of elements in $G$ and any complex numbers $\alpha_{1},\ldots ,\alpha_{n}$, we have:
$\sum_{i,j=1}^{n}\phi(x_{i}^{-1}x_{j})=\int_{K}<\pi(k(x_{i}^{-1}x_{j})\xi,\xi>dk$
=$\int_{K}<\pi(k(x_{i})^{-1}k(x_{j})\xi,\xi>dk=\int_{K}<\pi(k(x_{j})\xi,k(x_{i})\xi>dk$.
Therefore,
$\sum_{i,j=1}^{n}\phi(x_{i}^{-1}x_{j})\alpha_{i}\overline{\alpha_{j}}$
=$\sum_{i,j=1}^{n}\int_{K}<\overline{\alpha_{j}}\pi(k(x_{j}))\xi,\overline{\alpha_{i}}\pi(k(x_{i}))\xi>dk$
=$\int_{K}<\sum_{j}\overline{\alpha_{j}}\pi(k(x_{j}))\xi,\sum_{i}\overline{\alpha_{i}}\pi(k(x_{i}))\xi>dk\geq 0$.
Conversely, if $\phi$ is a bounded $K$-spherical function on $N$ then $\phi$ is positive definite. Let $\varphi$ be a positive definite $(G,K)$-spherical function. Let $\Pi_{\varphi}$ be the associated irreducible unitary representation, and $H_{\varphi}$ the representation space, such that there is a $K$-fixed unit vector $u_{\varphi}\in H_{\varphi}$ and let $\varphi(g)=<u_{\varphi},\prod_{\varphi}(g)u_{\varphi}>$ for all $g\in G$. Then we have
$\left |\varphi(g) \right |=\left |<u_{\varphi},\prod_{\varphi}(g)u_{\varphi}> \right |$
$\leq \left |<u_{\varphi},u_{\varphi}> \right |^{\frac{1}{2}}\times \left |<\prod_{\varphi}(g)u_{\varphi},\prod_{\varphi}(g)u_{\varphi}> \right |^{\frac{1}{2}}$
$\leq \left |<u_{\varphi},u_{\varphi}> \right |=1$
Therefore, $\varphi$ is bounded.
\end{proof}
\begin{theorem}\label{equal}
In the notation just above, assume the group $G$ real . The spherical function
$\psi_{\lambda}$ on $G_{0}$ is bounded if and only if $\lambda$ is real,i.e.$\lambda\in a^{*}$.
\end{theorem}
\begin{proof}
According to Theorem 2.4, we obtain The spherical function $\psi_{\lambda}$ on $G_{0}$ is positive definite if and only if $\lambda$ is real,i.e.$\lambda\in a^{*}$. According to Theorem 2.11, since $p$ is abelian, we know that $\psi_{\lambda}$ is a bounded $K$-spherical function on $p$ is and only if $\psi_{\lambda}$ is positive definite. Therefore, $\psi_{\lambda}$ on $G_{0}$ is bounded if and only if $\lambda$ is real,i.e.$\lambda\in a^{*}$.
\end{proof}
\section{Generalizations for the Eigenspaces of the Laplacian on $\mathbb{R}^{n}$}
\begin{lemma}\label{twist}
[7] Let $K$ be any closed subgroup of $O(n)$, if $K$ is transitive on the spheres about 0 in $\mathbb{R}^{n}$, then $\mathcal{D}(G/K)=\mathbb{C}[\bigtriangleup]$, algebra of polynomials in the Laplace-Beltrami operator $\bigtriangleup=-\sum \partial^{2}/\partial x_{i}^{2}.$
\end{lemma}
\begin{proof}
It is clear that $\mathbb{C}[\bigtriangleup]\subset \mathcal{D}(G/K)$. Now let $D\in \mathcal{D}(G/K)$ be of order m. Then the mth order symbol of $D$ is a pollynomial of pure degree m constant on spheres about 0 in $\mathbb{R}^{n}$, in other words a multiple $cr^{m}$ with $m$ even and $r^{2}=\sum x_{i}^{2}$. Now $D-(c(-\bigtriangleup)^{m/2})\in \mathcal{D}(G/K)$ and $D-(c(-\bigtriangleup)^{m/2})$ has order $<m$. By induction on the order, $D-(c(-\bigtriangleup)^{m/2})\in \mathcal{D}(G/K)$, so we have $D\in \mathcal{D}(G/K)$.
\end{proof}
Let $L$ denote the usual Laplacian on $\mathbb{R}^{n}$ and for each $\lambda\in \mathbb{C}$ let $\mathcal{E}_{\lambda}(\mathbb{R}^{n})$ denote the eigenspace
\begin{equation}\label{l-invariant elements}
\mathcal{E}_{\lambda}(\mathbb{R}^{n})=\{f\in \mathcal{E}(\mathbb{R}^{n})\mid Lf=-\lambda^{2}f\}
\end{equation}
with the topogy induced by that of $\mathcal{E}(\mathbb{R}^{n})$. Let $G=K\rtimes \mathbb{R}^{n}$ and $K$ is the closed subgroup of $O(n)$ as well as acting transitive on the spheres about 0 in $\mathbb{R}^{n}$.
\begin{theorem}\label{equal}
The natural action of $G$ on $\mathcal{E}_{\lambda}(\mathbb{R}^{n})$ is irreducible if and only if $\lambda\neq 0$.
\end{theorem}
\begin{proof}
It is clear that each function
\begin{equation}\label{l-invariant elements}
f(x)=\int_{S^{n-1}}e^{i\lambda(x,w)}F(w)dw, \ F\in L^{2}(S^{n-1}),
\end{equation}
lies in $\mathcal{E}_{\lambda}(\mathbb{R}^{n})$; here $(,)$ denotes the usual inner product on $\mathbb{R}^{n}$ and $dw$ the normalized volume element.
\end{proof}
\begin{lemma}\label{twist}
Let $\lambda\neq 0$. Then the mapping $F\rightarrow f$ defined by (3.4) is one-to-one.
\end{lemma}
\begin{proof}
Let $p(\zeta)=p(\zeta_{1},\cdots \zeta_{n})$ be a polynomial and $D$ the corresponding constant coefficient differential operator on $\mathbb{R}^{n}$ such that
\begin{equation}\label{l-invariant elements}
\int_{\mathbb{R}^{n}}e^{i(x,\zeta)}D_{x}(e^{-(\frac{1}{2})\left |x \right |^{2}})=p(\zeta)e^{-(\frac{1}{2})(\zeta_{1}^{2}+\cdots +\zeta_{n}^{2})}
\end{equation}
for$\zeta\in \mathbb{C}^{n}$. If $f\equiv=0$ in (1) we deduce from (3.6) that
\begin{equation}\label{l-invariant elements}
\int_{S^{n-1}}p(\lambda w_{1},\cdots ,\lambda w_{n})F(w)dw=0
\end{equation}
Since $\lambda \neq 0$, this implies $F\equiv 0$.
\end{proof}
\begin{lemma}\label{twist}
Let $\lambda\neq 0$. The $K$-finite solutions $f$ of the equation $Lf=-\lambda^{2}f$ are precisely
\begin{equation}\label{l-invariant elements}
f(x)=\int_{S^{n-1}}e^{i\lambda(x,w)}F(w)dw
\end{equation}
where $F$ is a $K$-finite function on $S^{n-1}$.
\end{lemma}
\begin{proof}
Let $\delta$ be an irreducible representation of $K$ and if $\Sigma$ is any sphere in $\mathbb{R}^{n}$ with center at 0 let $\mathcal{E}_{\delta}(\Sigma)$ denote the space of $K$-finite functions in $\mathcal{E}(\Sigma)$ of type $\delta$. We know from Lemma 1.5 p.134 in [10] that if $\Sigma$ is suitably chosen each function $f\mid \Sigma$ to$\Sigma$. With $F$ and $f$ as in (3.4) it follows that the maps
$F\rightarrow f\mid \Sigma, \ F\rightarrow f \ F\in L^{2}(S^{n-1})$
are one-to-one and commute with the action of $K$. For reasons of dimensionality, the first must therefore map $\mathcal{E}_{\delta}(S^{n-1})$ onto $\mathcal{E}_{\delta}(\Sigma)$. The lemma now follows.
\end{proof}
For $\lambda\neq 0$ let $\mathcal{H}_{\lambda}$ denote the space of functions $f$ as defined in (3.4); $\mathcal{H}_{\lambda}$ is a Hilbert space if the norm of $f$ is the $L^{2}$ norm of $F$ on $S^{n-1}$.
\begin{lemma}\label{twist}
Let $\lambda\neq 0$. Then the space $\mathcal{H}_{\lambda}$ is dense in $\mathcal{E}_{\lambda}(\mathbb{R}^{n})$.
\end{lemma}
\begin{proof}
Each eigenfunction of $L$ can be expanded in a convergent series of $K$-finite eigenfunctions (cf. Sect. 5 [6])
so the lemma follows from Lemma 3.8.
\end{proof}
We can now prove Theorem 3.3. We first prove that $G$ acts irreducibly on $\mathcal{H}_{\lambda}$. Let $V\neq 0$ be a closed invariant subspace of $\mathcal{H}_{\lambda}$. Then there exists an $h\in V$ such that $h(0)=1$. We write
\begin{equation}\label{l-invariant elements}
h(x)=\int_{S^{n-1}}e^{i\lambda(x,w)}H(w)dw
\end{equation}
and the average $h^{\natural}(x)=\int_{K}h(k.x)dk$ is then
\begin{equation}\label{l-invariant elements}
h^{\natural}(x)=\varphi_{\lambda}(x)=\int_{S^{n-1}}e^{i\lambda(x,w)}dw.
\end{equation}
If $f$ in (3.4) lies in the annihilator $V^{0}$ of $V$ the functions $F$ and $H$ are orthogonal on $S^{n-1}$. Since $V^{0}$ is $K$-invariant this remains true for $H$ replaced by its integral over $K$, in other words $\varphi_{\lambda}$ belongs to the double annihilator $(V^{0})^{0}=V$. Now, since $V$ is invariant under translations it follows that for each $t\in \mathbb{R}$ the function
\begin{equation}\label{l-invariant elements}
x\rightarrow \int_{S^{n-1}}e^{i\lambda(x,w)}e^{i\lambda(t,w)}dw
\end{equation}
belongs to $V$. But then Lemma 3.5 shows that the annihilator of $V$ in $\mathcal{H}_{\lambda}$ is ${0}$, whence the irreducibility of $G$ on $\mathcal{H}_{\lambda}$.
Passing now to $\mathcal{E}_{\lambda}$ let $V\subset \mathcal{E}_{\lambda}$ be a closed invariant subspace. Then $V\cap \mathcal{H}_{\lambda}$ is an invariant subspace of $\mathcal{H}_{\lambda}$; Schwartz' inequality shows easily that it is closed. Thus, by the above, $V\subset \mathcal{E}_{\lambda}$ is $\{0\}$ or $\mathcal{H}_{\lambda}$. In the second case $V=\mathcal{E}_{\lambda}$ by Lemma 3.10. In the first case consider for each $f\in V$ the convergent expansion
\begin{equation}\label{l-invariant elements}
f=\sum_{\delta\in \hat{K}}\alpha_{\delta}*f
\end{equation}
where $\alpha_{\delta}=d(\delta)\chi_{\delta}^{*}$ and
\begin{equation}\label{l-invariant elements}
(\alpha_{\delta}*f)(x)=\int_{K}\alpha_{\delta}(k)f(k^{-1}.x)dk
\end{equation}
$\chi_{\delta}$ being the character of $\delta$.Then $\alpha_{\delta}*f\in \mathcal{H}_{\lambda}$ by Lemma 3.8. Let $V^{0}\subset \mathcal{E}^{'}(\mathbb{R}^{n})$ be the annihilator of $V$. Then $V^{0}$ is $G$-invariant and if $T\in V^{0}$,
\begin{equation}\label{l-invariant elements}
\int_{\mathbb{R}^{n}}(\alpha_{\delta}*f)(x)dT(x)=\int_{K}\alpha_{\delta}(k)\int_{\mathbb{R}^{n}}f(x)dT(k.x)dk
\end{equation}
so $\alpha_{\delta}*f$ belongs to the double annihilator $(V^{0})^{0}=V$. Thus $\alpha_{\delta}*f\in V\cap \mathcal{H}_{\lambda}=\{0\}$ so, by (3.6),$f=0$. Thus $V=\{0\}$ so the proof is finished.
\section{A estimation for some spherical functions and the groups K}
From [7], we know if $K$ is transitive on the spheres about 0 in $\mathbb{R}^{n}$, then the spherical function on $K\rtimes \mathbb{R}^{n}\simeq \mathbb{R}^{n}$ is of the form:
$\varphi_{s}(r)=\varphi(r,s)=\int_{S^{n-1}}e^{s(\xi,x)}d\sigma(\xi)$
=$\frac{\Gamma(\frac{n}{2})}{\sqrt{\pi}\Gamma\frac{n-1}{2}}\int_{0}^{\pi}e^{sr\cos \theta}\sin^{n-2}\theta d\theta$.
Where $s$ is a complex number, and $r=\left \|x \right \|=\sqrt{x_{1}^{2}+\ldots +x_{n}^{2}}$, $S^{n-1}$ is the unit sphere in $\mathbb{R}^{n}$ and
$\sigma$ the normalized surface measure on $S^{n-1}$.
If Re$s$=0 then it follows from above equation that $\left |\varphi_{s}(x) \right |\leq 1$ for all $x\in G$.
Clearly,$\varphi_{s}=\varphi_{-s}$. We just need to consider the case Re$s\geq 0$.
\begin{theorem}\label{equal}
If Re$s>0$, we have
$\varphi(r,s)\sim \frac{\Gamma(\frac{n}{2})2^{\frac{n-3}{2}}}{\sqrt{\pi}}\frac{e^{sr}}{(sr)^{\frac{n-1}{2}}}$,
when $r\rightarrow \infty$.
\end{theorem}
\begin{proof}
From above equation, we know that, by an elementary substitution,
$\varphi_{s}(r)=\varphi(r,s)$
=$\frac{\Gamma(\frac{n}{2})}{\sqrt{\pi}\Gamma(\frac{n-1}{2})}\int_{-1}^{1}e^{srt}(1-t^{2})^{\frac{n-3}{2}}dt$
and, setting $t=1-\frac{u}{r}$, we obtain
$\varphi(r,s)=\frac{\Gamma(\frac{n}{2})}{\sqrt{\pi}(\frac{n-1}{2})}\frac{e^{sr}}{r^{\frac{n-1}{2}}}\int_{0}^{2r}e^{-su}u^{\frac{n-3}{2}}(2-\frac{u}{r})^{\frac{n-3}{2}}du$
For Re$s>0$, we get, using Lebesgue's dominated convergence theorem,
$\lim_{r\rightarrow \infty}\int_{0}^{2r}e^{-su}u^{\frac{n-3}{2}}(2-\frac{u}{r})^{\frac{n-3}{2}}du=\frac{\Gamma(\frac{n-1}{2})2^{\frac{n-3}{2}}}{s^{\frac{n-1}{2}}}$
\end{proof}
Finally, I will give all the possible $K$, which is transitive on the spheres about 0 on $\mathbb{R}^{n}$, $n>1$. [11]
When $K$ is transitive on the spheres about 0 in $\mathbb{R}^{n}$,$n>1$, its identity compoment $K^{0}$ is also transitive, and $K=K^{0}F$ where $F$ is a finite subgroup of the normalizer $N_{O(n)}(K^{0})$. The possibilities for $K^{0}$ are as follows:
(1)$n>1$ and $K^{0}=SO(n)$,
(2)$n=2m$ and (i)$K^{0}=SU(m)$ or (ii)$U(m)$,
(3)$n=4m$ and (i)$K^{0}=Sp(m)$ or (ii)$Sp(m).U(1)$ or $Sp(m).Sp(1)$,
(4)$n=7$ and $K^{0}$ is the exceptional group $G_{2}$,
(5)$n=8$ and $K^{0}=Spin(7)$, and
(6)$n=16$ and $K^{0}=Spin(9)$.
In case (1), $N_{O(n)}(K^{0})=O(n)$, so the relevant choices for $F$ are $\{I\}$ and $\{I,-I\}$, so $K$ is either $SO(n)$ or $O(n)$.
In case (2)(i),$N_{O(n)}(K^{0})=U(m)\bigcup \alpha U(m)$ where $\alpha$ is complex conjugation of $\mathbb{C}^{m}$ over $\mathbb{R}^{m}$. The relevant
choices for $F$ are the finite subgroups of $U(1)\bigcup \alpha U(1)$ where $U(1)$ consists of the unitary scalar matrices $e^{ix}I$, $x$ real. Those
are the cyclic groups $\mathbb{Z}_{l}=\{e^{2\pi ik/l}I\}$ of order $l\geq 1$ and the dihedral groups $\mathbb{D}_{l}=\mathbb{Z}_{l}\bigcup \alpha\mathbb{Z}_{l}$, so $K$ is a group $SU(m)\mathbb{Z}_{l}$ or $SU(m)\mathbb{D}_{l}$. In the case (2)(ii) the relevant possibilities for $F$ are $\{I\}$ and $\{\alpha,I\}$, so $K$ is either $U(m)$ or $U(m)\bigcup \alpha U(m)$.
In case (3){i},(3)(iii),(4),(5),(6), $K^{0}$ has no outer automorphism, so we may take $F$ in the centralizer $Z_{O(n)}(K^{0})$. Thus in the case (3)(i), $F$ can be any subgroup of $Sp(1)$, in the other words, a cyclic group $\mathbb{Z}_{l}$ of order $l$, a binary dihedral group $\mathbb{D}_{l}^{*}$ of order $4l$, a binary tetrahedral group $\mathbb{T}^{*}$ of order 24, a binary octahedral group $\mathbb{O}^{*}$ of order 48, or a binary icosahedral group $\mathbb{I}^{*}$ of order 60. Thus $K$ is a group $Sp(m)\mathbb{Z}_{l}$,$Sp(m)\mathbb{D}_{l}^{*}$,$Sp(m)\mathbb{T}^{*}$, $Sp(m)\mathbb{O}^{*}$ or $Sp(m)\mathbb{I}^{*}$. In case (3)(ii) the relevant possibilities for $F$ are $\{I\}$ and $\{\beta,I\}$, where the $U(1)$ factor of $K^{0}$ consists of all quaternion scalar multiplications by complex numbers $e^{ix}$, $x$ is real, as in the case (2), and $\beta$ is quaternion scalar multiplication by $j$.Thus $K$ is either $Sp(m)U(1)$ or $(Sp(m)U(1))\bigcup (Sp(m)U(1))\beta$. In case (3)(iii), $K^{0}$ is its own $O(n)$-centralizer so $F=\{I\}$ and $K=Sp(m)Sp(1)$.
In case (4),(5),(6),$K^{0}$ is absolutely irreducible on $\mathbb{R}^{n}$, so relevant $F$ would have to consist of real scalars. As $G_{2}$ does not contain $-I$ we see that the relevant $F$ for case (4) are $\{I\}$ and $\{I,-I\}$, resulting in $K=G_{2}$ and $K=G_{2}\bigcup (-I)G_{2}$. Both $Spin(7)$ and $Spin(9)$ do contain $-I$, so $F$ is trivial in case (5),(6).That gives $K=Spin(7)$ in case (5) and $K=Spin(9)$ for case (6).
|
\section{Introduction}
\label{introduction}
Since the early work of Boltzmann in classical statistical mechanics\cite{Boltzmann,Boltzmann_translation}
and Planck in quantum statistical mechanics\cite{Planck_1901},
the microcanonical ensemble,
in which the energy is held constant,
has been a fundamental starting point.
However, this use of
the microcanonical ensemble
has recently
come under renewed scrutiny
due to two fundamental problems it presents.
\begin{enumerate}
\item
As normally interpreted,
the quantum microcanonical ensemble
is defined by
the Boltzmann constant times
the logarithm of the degeneracy
at each discrete energy eigenvalue,
so that
the entropy is not a continuous function of
the
energy,
as required by thermodynamics\cite{Planck_1901,Callen,RHS_book,RHS_continuous}.
\item
In both quantum and classical statistical mechanics,
for first-order phase transitions
the microcanonical ensemble
predicts a range of positive second derivatives
of the entropy as a function of energy,
which violates a well-known stability criterion\cite{Callen,RHS_book,RHS_continuous}.
\end{enumerate}
It has been suggested that
the source of these violations of the thermodynamic postulates
lies in the neglect of
quantum states that are linear combinations of
energy eigenstates\cite{RHS_continuous}.
Such linear combinations are generated
whenever two macroscopic systems
exchange energy,
and give a continuous energy spectrum.
The canonical ensemble can then
provide an excellent description of
the statistical properties of a macroscopic system,
even in the absence of
contact with an infinite thermal reservoir.
The traditional derivation of the canonical probability distribution
involves the expansion of the joint probability distribution
for the energy of a system of interest
in thermal contact with a second, much larger
system,
designated a thermal reservoir.
In the limit that the reservoir is
infinitely larger than the system of interest,
it is then shown that the
thermal weight
(loosely termed the ``probability'')
of an eigenstate with energy $E_n$
is proportional to
$\exp( - \beta E_n )$,
where
$\beta = 1/k_B T$,
$k_B$ is Boltzmann's constant,
and
$T$ is the absolute temperature.
We are raising the question of whether the second system
must really be infinitely larger,
or whether it might even be smaller than
the primary system of interest.
If so,
this would be consistent with normal usage in thermodynamics,
in which a system may be put into thermal equilibrium with
any other macroscopic system regardless of its size
(Zeroth Law Of Thermodynamics).
Four important papers have recently dealt with
key aspects of this question\cite{Jin_2013_1,Jin_2013_2,Novotny_decoherence_2015,Novotny_2016_1}.
While we are primarily concerned with the equilibrium features
of the statistical mechanics of finite systems,
these papers explored the dynamical approach to equilibrium
under the deterministic time development
of Newton's equations for classical systems\cite{Jin_2013_1}
and Schr\"odinger's equation for quantum systems\cite{Jin_2013_2,Novotny_decoherence_2015,Novotny_2016_1}.
A consequence of their computations was that
for both classical and quantum systems
the projection of
the joint probability distribution of two finite systems
onto one of the systems
was increasingly well approximated by
the canonical distribution
as the system sizes increased.
The only limitation of these papers
was
the practical one
that the solutions of the dynamical equations
could only be carried out for relatively small systems.
In this paper,
we take a simpler approach to a more limited problem
than was treated in
Refs.~\cite{Jin_2013_1,Jin_2013_2,Novotny_decoherence_2015,Novotny_2016_1}.
We restrict ourselves to the equilibrium statistical mechanics
of finite systems,
so we ignore the time development of the
microscopic states.
This means that we only need the information
contained in
the degeneracies of the quantum energy levels, $ \omega ( E_n )$,
and their equilibrium occupation.
We will show that under very general conditions,
the distribution of energies is very nearly canonical.
This in turn has the consequence that the thermodynamic entropy
as a function of energy
is given by
the canonical entropy,
and not the microcanonical.
In the next section,
we give the basic definitions for analyzing
two interacting quantum systems,
including the joint probability distribution in equilbirum,
and the canonical distribution as an approximation
to that distribution.
In Section \ref{section_Massieu},
we recall the definition of the Massieu function,
which is a version of the Helmholtz free energy
that is applicable to both positive and negative temperatures.
Section \ref{section_SHO}
then defines
$\delta$,
which
was introduced in
Refs.~\cite{Jin_2013_1,Jin_2013_2,Novotny_decoherence_2015,Novotny_2016_1}
a measure of the difference between
the joint probability distribution
and the canonical approximation.
A system of simple harmonic oscillators
is used to illustrate how a finite reservoir
can generate an excellent approximation to the canonical distribution --
even if it is smaller than the system of interest.
In Section \ref{section_Ising}
we discuss the behavior of the two-dimensional Ising model, both at and away from the
second-order phase transition.
In Section \ref{section_first_order},
we discuss a corresponding analysis of the two-dimensional, twelve-state Potts model,
both at and away from the
first-order phase transition.
\section{Joint probability distribution of two finite quantum systems}
\label{section: quantum joint probability}
Although thermodynamics and statistical mechanics
also make predictions for other macroscopic variables,
we will limit consideration to energy distributions
to represent the general situation.
Similarly,
although a general treatment would include
the exchange of energy between an arbitrary number
$M \ge 2$ subsystems,
restrict ourselves to
$M=2$ subsystems,
since the generalization to more subsystems is obvious.
Label the systems with $S$ (for the system of interest)
and $E$ (for the environment).
The joint Hamiltonian is then denoted as
\begin{equation}
H_{\textrm{total}}
=
H_S + H_E + L_{S,E} ,
\end{equation}
where
$H_S$ is the Hamiltonian of the system of interest,
$H_E$ is the Hamiltonian of the environment,
and
$L_{S,E}$
denotes the interactions between
particles in different subsystems.
When the systems are separated,
each system has its own energy spectrum.
For simplicity,
we will assume that the separation between energy eigenvalues
is a multiple of $\epsilon$ in both the system and the environment.
Let us assume that the system and the environment together are separated from the rest of the universe,
so that the total energy is a constant $E_T$.
When the system and the environment are separated,
let the density of states of the system be $\omega_S(E_n)$
(the degeneracy of energy level $E_n$),
and
$\omega_E(E_T-E_n)$ for the environment,
where $n$ denotes an eigenstate of the system.
It is normally the case that the energies associated with
$L_{S,E}$
are negligible in comparison with the other energies in the problem.
Nevertheless, they allow the transfer of energy between the two systems,
by changing the energy spectrum from two separate spectra
to a single spectrum\cite{RHS_continuous}.
After the two systems have been brought together, equilibrated, and then separated, the energy distribution in each system is
not a delta function.
Each combination of states will be equally likely, so that the probability of two energy levels with total energy
$E_T$
will have a probability proportional to the product of the degeneracies
$\omega_S ( E_n ) \omega_E ( E_T - E_n )$.
Normalizing, this becomes the probability for energy level $n$ in the original system.
\begin{equation}\label{product probability}
P_S (E_n, E_T)
=
\frac{\omega_S ( E_n ) \omega_E ( E_T - E_n )}
{ \sum_{E_n} \omega_S ( E_n ) \omega_E ( E_T - E_n )}
\end{equation}
The equilibrium value of the energy in $S$
is given by
\begin{equation}\label{product equilibrium value}
\langle E_n \rangle_S
=
\sum_{E_n}
E_n
P_S (E_n, E_T)
\end{equation}
This should be compared to the probability for the energy level $n$
of the system in contact with the thermal reservoir.
This probability can be derived from
Eq.~(\ref{product probability})
by expanding the logarithm of $\omega_E$
about its equilibrium value.
\begin{eqnarray}
\ln P_S (E_n, E_T)
&=
\ln \omega_S ( E_n )
+
\ln \omega_E ( E_T - E_n ) \nonumber \\
&
- \ln \left[
\sum_{E_n} \omega_S ( E_n ) \omega_E ( E_T - E_n ) \right]
\end{eqnarray}
\begin{eqnarray}
\ln P_S (E_n, E_T)
&=
\ln \omega_S ( E_n )
+
\ln \omega_E \left( E_T - \langle E_n \rangle_S \right) \nonumber \\
&
-
(E_n - \langle E_n \rangle_S )
\frac{ \partial }{ \partial E_T }
\ln \left[ \omega_E ( E_T - \langle E_n \rangle_S ) \right] \nonumber \\
&
+ (\textrm{higher order terms}) \nonumber \\
&
- \ln \left[
\sum_{E_n} \omega_S ( E_n ) \omega_E ( E_T - E_n ) \right]
\end{eqnarray}
The factor
$ \beta = \frac{ \partial }{ \partial E_T }
\ln \left[ \omega_E ( E_T - \langle E_n \rangle_S ) \right]$
can be interpreted as the inverse temperature.
Collecting terms, and dropping the higher order terms, we have
\begin{equation}
\ln P_S (E_n, E_T)
\approx
\ln P_{S,C}(E_n, \beta) ,
\end{equation}
where
\begin{equation}
\ln P_{S,C}(E_n, \beta)
=
\ln \omega_S ( E_n )
- \beta (E_n - \langle E_n \rangle_S )
-\ln X ,
\end{equation}
or
\begin{equation}
P_{S,C} (E_n,\beta)
=
\frac{1}{X}
\omega_S ( E_n )
\exp [- \beta (E_n - \langle E_n \rangle_S ) ] .
\end{equation}
This is just the canonical distribution
\begin{equation}
P_{S,C} (E_n,\beta)
=
\frac{1}{Z}
\omega_S ( E_n )
\exp [- \beta E_n ]
\end{equation}
with different parameters
because it is expanded about
$\langle E_n \rangle_S$
instead of zero.
The constants are related by
\begin{equation}
\frac{1}{Z}
=
\frac{1}{X}
\exp [ \beta \langle E_n \rangle_S ) ]
\end{equation}
It is clear that
$ P_S (E_n, E_T)$
and
$P_{S,C} (E_n, E_T)$
differ,
but the question is whether this difference is large enough to matter.
The question is of fundamental importance.
If they differ measurably for macroscopic systems,
it is not sufficient to specify the temperature
of system $E$,
but the system size must also be known.
If the difference is not measurable,
the assumptions of thermodynamics
are valid.
Furthermore,
the entropy as a function of energy
(the fundamental relation)
must be given by the canonical entropy,
which we now derive.
\section{Massieu functions and the canonical entropy}
\label{section_Massieu}
It is well known that the canonical partition function
is related to the Helmholtz free energy,
$F=U-TS$,
by the equation
\begin{equation}
\ln Z ( \beta, V,N ) = - \beta F(T,V,N)
\end{equation}
The Helmholtz free energy is,
of course,
the Legendre transform of the
fundamental relation
$U=U(S,V,N)$
with respect to temperature,
which we will denote as
$F(T,V,N) = U[T]$,
indicating the Legendre transform
by the square brackets
around the new variable $T$\cite{Callen,RHS_book}.
For generality,
it is better the use a Massieu function,
which is the Legendre transform of the entropy\cite{Callen,RHS_book}.
The reason is that if
$S=S(U,V,N)$
is not monotonic in $U$,
the function cannot be inverted
to find
$U=U(S,V,N)$.
It will be particularly useful to define a dimensionless entropy,
$\tilde{S}=S/k_B$,
in forming Massieu functions.
From the differential form of the fundamental relation
for $dS$,
we can see that
\begin{equation}
d \tilde{S} = \beta \, dU + \beta P dV - \beta \mu \, dN,
\end{equation}
where
$P$ is the pressure, $V$ is the volume,
$\mu$ is the chemical potential, and $N$ is the number of particles.
The inverse temperature $\beta$ is found from the usual equation,
which can be written as
\begin{equation}\label{S[b]=S-bU}
\beta
=
\left(
\frac{ \partial \tilde{S} }{ \partial U }
\right)_{V,N} .
\end{equation}
The Legendre transform (Massieu function) is given by
\begin{equation}\label{s[b]= - b F}
\tilde{S}[\beta] = \tilde{S} - \beta U = - \beta \left( U - TS \right) = - \beta F ,
\end{equation}
so that
\begin{equation}\label{St b = ln Z}
\tilde{S}[\beta] =\ln Z ( \beta, V,N ) .
\end{equation}
The differential form
of the Massieu function $\tilde{S}[\beta]$
is then
\begin{equation}
d \tilde{S}[\beta] = - U d \beta + \beta P dV - \beta \mu dN .
\end{equation}
This immediately gives us
\begin{equation}\label{d tilde S / dbeta = - U}
\left(
\frac{ \partial \tilde{S}[ \beta ] }{ \partial \beta }
\right)_{V,N} =
- U
=
-
\left(
\frac{ \partial ( \beta F ) }{ \partial \beta }
\right)_{V,N} ,
\end{equation}
where the last equality is a well-known thermodynamic identity\cite{Callen,RHS_book}.
To carry out the inverse Legendre transform of
$\tilde{S}[\beta]$
to find
$S(U)$,
use
Eq.~(\ref{d tilde S / dbeta = - U})
to find
$U=U(\beta)$.
Since $U$ is a monotonic function of
$\beta$,
even for a non-monotonic density of states,
we can invert this equation to obtain
$\beta=\beta(U)$.
From
Eq.~(\ref{S[b]=S-bU}),
we can find
\begin{equation}\label{St = St b + b U}
\tilde{S} = \tilde{S}[\beta] +\beta(U) U .
\end{equation}
Finally, the entropy with the usual dimensions is given by
\begin{equation}
S
=
k_B \tilde{S} .
\end{equation}
\section{Comparison of canonical and joint distributions
for simple harmonic oscillators}
\label{section_SHO}
A set of simple harmonic oscillators (SHO)
provides a test for these ideas\cite{RHS_continuous}.
The energy spectrum is
\begin{equation}\label{SHO H}
E = \hbar \omega \sum_k^N \left( n_k + \frac{1}{2} \right) ,
\end{equation}
where $\hbar$ is Planck's constant,
and $\omega$ is the angular frequency,
and $n_k$ takes on the values of all nonnegative integers.
The canonical entropy is
\begin{equation}
S_{\textrm{SHO}}
=
N k_B
\left[
- x \ln x
+ ( 1 + x ) \ln (1+x) ,
\right]
\end{equation}
where
\begin{equation}
x
=
\frac{
U_{\textrm{SHO}} -\frac{1}{2}
\hbar \omega N
}{
N \hbar \omega
} .
\end{equation}
To provide a numerical measure for the
agreement between the canonical entropy
and the joint distribution,
we follow Novotny et al.\cite{Novotny_decoherence_2015,Novotny_2016_1}
in defining $\delta$.
\begin{equation}\label{delta_1}
\delta
=
\sqrt{
\sum_{E_n}
\left(
P_S(E_n, E_T)
-
P_{S,C}(E_n, \beta )
\right)^2
}
\end{equation}
where
\begin{equation}
P_{S,C}(E_n, \beta )
=
D_S(E_n)
e^{-\beta E_n} / Z ,
\end{equation}
with $D_S(E_n)$ to denote the density of states of the system $S$
(to avoid confusion between $\omega$ and $\omega_S(E_n)$ ),
and
\begin{equation}
Z = \sum_{E'_n}
D_S(E'_n)
e^{-\beta E'_n} .
\end{equation}
The value of
$\delta$ is zero when the two distributions are identical,
which is the case for
$N \rightarrow \infty$.
For finite systems,
it indicates how much the two distributions
differ.
Therefore,
a small value of delta means that the canonical
distribution provides a good description of the system.
We investigated the probability distributions for
two systems of simple harmonic oscillators
representing the system (S) and the environment (E).
We used sizes ranging from
$10$ to $10^6$.
The energy was fixed at $1.5\epsilon$ per oscillator
for $N_S$ and $N_E$
between $10$ and $10^6$.
Table~\ref{ table:_SHO_1} gives the values of
$\delta$.
\begin{table}[htb]
\caption{The values of $\delta$
for a system of $N_S$ simple harmonic oscillators
given in
Eq.~(\ref{SHO H})
from computations
with environment composed of $N_E$
simple harmonic oscillators.
The values of the system size, $N_S$,
are given as headings of the columns.
The average energy is $1.5$,
and the temperature is $1.44$.
}
\begin{center}
\begin{tabular}{c||c|c|c|c|c|c}
$N_E$ & $ 10 $ & $ 10^2 $ & $ 10^3 $ & $ 10^4 $ & $ 10^5 $ & $ 10^6 $ \\
\hline
$ 10 $ & 8.61E-2 & 2.19E-1
& 2.68E-1 & 2.61E-1
& 2.58E-1 & 2.57E-1 \\
$ 10^2 $ & 1.05E-2 & 4.61E-2
& 1.03E-1 & 1.33E-1
& 1.43E-1 & 1.42E-1 \\
$ 10^3 $ & 1.08E-3 & 5.90E-3
& 2.57E-2 & 5.68E-2
& 7.24E-2 & 7.79E-2 \\
$ 10^4 $ & 1.08E-4 & 6.09E-4
& 3.32E-3 & 1.45E-2
& 3.19E-2 & 4.06E-2 \\
$ 10^5 $ & 1.08E-5 & 6.11E-5
& 3.43E-4 & 1.86E-3
& 8.13E-3 & 1.79E-2 \\
$ 10^6 $ & 1.08E-6 & 6.11E-6
& 3.44E-5 & 1.93E-4
& 1.05E-4 & 4.57E-3
\end{tabular}
\end{center}
\label{ table:_SHO_1}
\end{table}
To understand the convergence of $\delta$,
first consider the dependence on the size of the environment, $N_E$.
Table~\ref{ table:_SHO_1} shows clearly that $\delta \rightarrow 1/N_E$,
which is due to the quadratic dependence
of the difference
on the energy
between $P_S(E_n, E_T)$
and
$P_{S,C}(E_n, \beta )$.
The size of the system, $N_S$,
enters in two places.
The width of the distribution is proportional to
$\sqrt{N_S}$,
which is a standard result for the width of a peak
in statistical mechanics.
This suggests scaling $\left( E_n-\langle E \rangle \right)/N_S$
by
$1/\sqrt{N_S}$.
To maintain the normalization of the probabilities,
we multiply them by $\sqrt{N_S}$.
The scaling behavior of the difference in the probabilities
is illustrated in
Fig.~\ref{SHO_diff_10000_10000}.
\begin{figure} [ht]
\includegraphics[width=\columnwidth]{diff_delta_10000_10000}
\caption{
The scaled difference in the probabilities
$(P_J-P_C)N_E/N_S^{1/4}$
is plotted against the
scaled system energies
$ N_S^{1/2}(E_n - \langle E \rangle ) / N_S\epsilon $.
The system contained $N_S=10^4$ SHO's,
while the environment contained $N_S=10^4$ SHO's
($R=N_S/N_E=1$).
The asymptotic behavior is shown,
and the plot will not change for larger systems.
The effective inverse temperature is $1.0$,
and the average energy is $1.44$.
}
\label{SHO_diff_10000_10000}
\end{figure}
The scaling dependence of $\delta$ is then given by
\begin{equation}\label{delta_scaled_1}
\delta = \tilde{\delta} (N_S/N_E) N_S^{3/4} N_E^{-1}
\end{equation}
where $ \tilde{\delta} (N_S/N_E)$ is a function of the ratio
$R=N_S/N_E$ for systems with either $N_S$ or $N_E$
larger than about $10^4$.
\begin{table}[hb]
\caption{The values of $\tilde{\delta}$
(the rescaled values of $\delta$)
for a system of $N_S$ simple harmonic oscillators
given in
Eq.~(\ref{SHO H})
from computations
with environment composed of $N_E$
simple harmonic oscillators.
The values of the system size, $N_S$,
are given as headings of the columns.
The average energy is $1.5$,
and the temperature is $1.44$.
}
\begin{center}
\begin{tabular}{c||c|c|c|c|c|c}
$N_E$ & $ 10 $ & $ 10^2 $ & $ 10^3 $ & $ 10^4 $ & $ 10^5 $ & $ 10^6 $ \\
\hline
$ 10 $ & 0.15314 & 0.06917 & 0.01507 & 0.00261 & 0.00046 &
0.00008 \\
$ 10^2 $ & 0.18610 & 0.14582 & 0.05795 & 0.01330 & 0.00255 & 0.00045 \\
$ 10^3 $ & 0.19155 & 0.18645 & 0.14472 & 0.05679 & 0.01287 & 0.00246 \\
$ 10^4 $ & 0.19215 & 0.19255 & 0.18649 & 0.14461 & 0.05666 & 0.01282 \\
$ 10^5 $ & 0.19221 & 0.19319 & 0.19266 & 0.18649 & 0.14459 & 0.05665 \\
$ 10^6 $ & 0.19221 & 0.19325 & 0.19331 & 0.19267 & 0.18649 & 0.14459
\end{tabular}
\end{center}
\label{ table:_SHO_1_delta_tilde}
\end{table}
As an illustration,
consider the following cases.
\begin{enumerate}
\item If the environment is much larger than the system
($N_S<<N_E$), then $ \tilde{\delta} (R)$ becomes
a constant
as $R \rightarrow 0$,
as seen in the bottom left side of
Table~\ref{ table:_SHO_1_delta_tilde}.
\item For a fixed ratio
$R=N_S/N_E$,
$\delta \rightarrow N^{-1/4}$,
as the sizes of the two systems increase
(where $N$ is either $N_S$ or $N_E$).
This can be seen by looking at any diagonal in
Table~\ref{ table:_SHO_1_delta_tilde},
where
the values of $\tilde{\delta}$ go to a constant.
that depends on $R$.
This decrease is rather slow,
but the deviation from canonical behavior
might still be hard to measure for a macroscopic system,
as seen in Table~\ref{table_fixed_ratio}.
\begin{table}[htb]
\caption{The values of $\delta$ for
$N_S=10^{20}$
as a function of the ratio
$R=N_S/N_E$.
}
\begin{center}
\begin{tabular}{c|c}
$R$ & $\delta$ \\
\hline
0.001 & 1.93E-9 \\
0.01 & 1.93E-8 \\
0.1 & 1.86E-7 \\
1 & 1.45E-6 \\
10 & 5.66E-6 \\
100 & 1.28E-5 \\
1000 & 2.46E-5
\end{tabular}
\end{center}
\label{table_fixed_ratio}
\end{table}
\item For the case that $N_E$ is large,
$\tilde{\delta}$
is independent of $N_S$,
and
$\delta$
goes as $N_S^{3/4}$,
as given by Eq.~(\ref{delta_scaled_1}),
and shown in Tables~\ref{ table:_SHO_1}
and \ref{ table:_SHO_1_delta_tilde}.
\end{enumerate}
The numerical calculations have confirmed the predictions of Eq.~(\ref{delta_scaled_1}).
\section{Comparison of canonical and joint distributions
for
the Ising model
(second-order phase transitions)}
\label{section_Ising}
The exchange of energy between an Ising model and
a system of simple harmonic oscillators
is different at the second-order phase transition
and away from it.
We will examine both cases.
\subsection{Away from the second-order phase transition}
We did calculations for an inverse temperature of $\beta=0.5$,
which is well away from the critical value of
$\beta_c=0.88137$.
The energy was set as close to the average energy
as possible,
given the discrete spectrum.
To obtain a quantitative measure of the agreement,
we compared a variety of sizes for both the system and the environment.
The canonical distribution
was used to obtain the best fit for the temperature.
The deviation of the temperature from the desired value
was small for all systems,
and got dramatically smaller as the systems got larger.
As can be seen from
Tables~\ref{ table:_Ising_b5_1_delta}
and \ref{ table:_Ising_b5_1_delta_tilde},
the data shows the same scaling given
in
Eq.~(\ref{delta_scaled_1}).
\begin{table}[htb]
\caption{The values of $\delta$
for a system of $L \times L$ Ising models
from computations
with environment composed of $N_E$
simple harmonic oscillators.
The values of the system size, $N_S = L^2$,
are given as headings of the columns.
The average energy is $-1.279$,
and the inverse temperature is $\beta=0.5$.}
\begin{center}
\begin{tabular}{c|c|c|c}
$N_E$ & $8 \times 8$ & $16 \times 16$ & $32 \times 32$ \\
\hline
$2^6$ & 2.05E-02 &5.04E-02 & 9.69E-02 \\
$2^8$ & 5.48E-03 & 1.46E-02& 3.56E-02\\
$2^{10}$ & 1.40E-03 & 3.84E-03 & 1.04E-02\\
$2^{12}$ & 3.51E-04 & 9.72E-04 & 2.72E-03\\
$2^{14}$ & 8.79E-05 & 2.44E-04 & 6.88E-04\\
$2^{16}$ & 2.20E-05 & 6.10E-05 & 1.72E-04
\end{tabular}
\end{center}
\label{ table:_Ising_b5_1_delta}
\end{table}
\begin{table}[htb]
\caption{The values of $\tilde{\delta}$
(the rescaled values of $\delta$)
for a system of $L \times L$ Ising models
from computations
with environment composed of $N_E$
simple harmonic oscillators.
The values of the system size, $N_S = L^2$,
are given as headings of the columns.
The average energy is $-1.279$,
and the inverse temperature is $\beta=0.5$.
}
\begin{center}
\begin{tabular}{c|c|c|c}
$N_E$ & $8 \times 8$ & $16 \times 16$ & $32 \times 32$ \\
\hline
$2^6$ & 0.05804& 0.05044 & 0.03424 \\
$2^8$ & 0.06197 & 0.05852 & 0.05033 \\
$2^{10}$ & 0.06322 & 0.06139 & 0.05866 \\
$2^{12}$ & 0.06359 & 0.06221 & 0.06146 \\
$2^{14}$ & 0.06368 & 0.06244 & 0.06223 \\
$2^{16}$ & 0.06369 & 0.06249 & 0.06227
\end{tabular}
\end{center}
\label{ table:_Ising_b5_1_delta_tilde}
\end{table}
\subsection{At the second-order phase transition}
The width of the peak in the energy is largest at the second-order phase transition
due the connection between the variance of the energy and the specific heat.
\begin{equation}
c_S = C_S/N_S=
\frac{1}{N_S k_B T^2}
\left\langle \left( E - \langle E \rangle \right)^2 \right\rangle
\end{equation}
The increased width of the energy peak
requires a larger environment to show a good approximation
to the canonical ensemble.
The improvement of the fit with increasing size is shown in
Fig.~\ref{Ising_256_SHO}.
\begin{figure}[htb]
\begin{subfigure} [htb] {0.3\textwidth}
\includegraphics[width=\columnwidth]{256ising_256osc-1}
\caption{
$2^{8}$ SHO's as environment.
}
\label{256ising_256osc}
\end{subfigure}
\begin{subfigure} [htb] {0.3\textwidth}
\includegraphics[width=\columnwidth]{256ising_1024osc-1}
\caption{
$2^{10}$ SHO's as environment.
}
\label{256ising_1024osc}
\end{subfigure}
%
\begin{subfigure} [htb] {0.3\textwidth}
\includegraphics[width=\columnwidth]{256ising_4096osc-1}
\caption{
$2^{12}$ SHO's as environment..
}
\label{256ising_4096osc}
\end{subfigure}
%
\caption{ Comparison of the canonical distribution (black dots)
with the true joint distribution (red dots) for a
Ising model
on a $16 \times 16$ lattice
at the transition energy
interacting with a system of simple harmonic oscillators.
The system is at the critical inverse temperature.}
\label{Ising_256_SHO}
\end{figure}
The increase of the width of the critical energy peak also
has the effect of increasing the value of $\delta$,
which is shown in
Table~\ref{ table:_Ising_critical_1_delta}.
The values should be compared to
Table~\ref{ table:_Ising_b5_1_delta}
for $\beta \ne \beta_c$,
which are all smaller.
\begin{table}[htb]
\caption{The values of $\delta$
for a system of $L \times L$ Ising models
from computations
with environment composed of $N_E$
simple harmonic oscillators.
The values of the system size, $N_S = L^2$,
are given as headings of the columns.
The temperature is $\beta_c=0.88137$.
}
\begin{center}
\begin{tabular}{c|c|c|c}
$N_E$ & $8 \times 8$ & $16 \times 16$ & $32 \times 32$ \\
\hline
$2^6$ & 1.00E-01 & 1.69E-01 & 2.14E-01 \\
$2^8$ & 3.09E-02 & 7.46E-02 & 1.24E-01 \\
$2^{10}$ & 8.18E-03 & 2.42E-02 & 5.73E-02 \\
$2^{12}$ & 2.08E-03 & 6.58E-03 & 1.95E-02 \\
$2^{14}$ & 5.21E-04 & 1.68E-03 & 5.42E-03 \\
$2^{16}$ & 1.30E-04 & 4.22E-04 & 1.39E-03
\end{tabular}
\end{center}
\label{ table:_Ising_critical_1_delta}
\end{table}
We should be able to rescale $\delta$
for the Ising critical point
to demonstrate the effect of the specific heat.
Unfortunately,
we have not been able to do this,
probably because the systems are not large enough
to show the asymptotic behavior.
\section{Comparison of canonical and joint distributions
for first-order transitions
($12$-state Potts model)}
\label{section_first_order}
The twelve-state Potts model in two dimensions
presents an interesting test of finite-canonical distributions.
Away from the first-order transition,
it shows the same
finite-canonical behavior distribution
as the system of simple harmonic oscillators
seen in the previous section.
At the first-order transition,
the behavior is different.
In each case, the energy
was chosen to correspond as closely as possible
to the temperature being investigated
($\beta=1.0$ and $\beta_c=1.496068$).
The value of the inverse temperature
used for the comparison with the canonical distribution
was then optimized to reduce the value of $\delta$.
This temperature correction was about $1\%$
for the smallest example ($8 \times 8$ Potts lattice and $64$ SHOs),
but went down to $7.6 \times 10^{-6}$
for a $32 \times 32$ lattice and $N_E=2^{18}$.
We first demonstrate the behavior away from the transition.
\subsection{Away from the first-order transition}
Values of $\delta$ were calculated for an inverse temperature of
$\beta=1.0$,
far from the first-order transition at
$\beta_c=1.49607$.
The scaled values for
$\tilde{\delta}$
are given in Table~\ref{table:_Potts_optimized_beta_1.0_delta_tilde}.
The patterns is the same as seen in
Table~\ref{ table:_SHO_1_delta_tilde},
although the values of $\tilde{\delta}$
are somewhat smaller.
The asymptotic scaling of $\delta$
is confirmed.
\begin{table}[htb]
\caption{The values of $\tilde{\delta}$
(the rescaled values of $\delta$)
for a two-dimensional, $12$-state Potts model
at $\beta \approx 1.0$ (away from the first-order transition).
The size of the environment
(a set of simple harmonic oscillators)
is shown in the left column.
Values of $\delta$
for the three sizes of the Potts model investigated
($8 \times 8$, $16 \times 16$, and $32 \times 32$)
are given in the next three columns.
}
\begin{center}
\begin{tabular}{r|r|r|r}
$N_E$ & $8 \times 8$ & $16 \times 16$ & $32 \times 32$ \\
\hline
$2^6$ & 0.10519 & 0.08274 & 0.04579 \\
$2^8$ & 0.11993 & 0.11032 & 0.08200\\
$2^{10}$ & 0.12432 & 0.12291 & 0.11048 \\
$2^{12}$ & 0.12556 & 0.12691 & 0.12289 \\
$2^{14}$ & 0.12592 & 0.12803 & 0.12671 \\
$2^{16}$ & 0.12599 & 0.13292 & 0.12769 \\
$2^{18}$ & 0.12605 & 0.12837 & 0.12794
\end{tabular}
\end{center}
\label{table:_Potts_optimized_beta_1.0_delta_tilde}
\end{table}
\subsection{At the first-order transition}
For a first-order transition,
the peak in the energy becomes a double peak
with a distance between the two maxima
that is
proportional to the size of the system.
The double peak structure at a first-order transition
makes the analysis of this case differ for all others.
For a small environment,
the width of the distribution is dramatically limited,
so that we see only a single peak,
as shown in
Fig.~\ref{Potts_256_SHO_4096}.
As $N_E$ increases in
Fig.~\ref{Potts_256_SHO_16384}
and \ref{Potts_256_SHO_65536},
the full canonical distribution,
with the double peak,
emerges.
\begin{figure}[htb]
\begin{center}
\begin{subfigure} [htb] {0.3\textwidth}
\includegraphics[width=\columnwidth]{256Potts_4096osc}
\caption{
$2^{12}$ simple harmonic oscillators.
}
\label{Potts_256_SHO_4096}
\end{subfigure}
\begin{subfigure} [htb] {0.3\textwidth}
\includegraphics[width=\columnwidth]{256Potts_16384osc}
\caption{
$2^{14}$ simple harmonic oscillators.
}
\label{Potts_256_SHO_16384}
\end{subfigure}
%
\begin{subfigure} [htb] {0.3\textwidth}
\includegraphics[width=\columnwidth]{256Potts_65536osc_2}
\caption{
$2^{16}$ simple harmonic oscillators.
}
\label{Potts_256_SHO_65536}
\end{subfigure}
\end{center}
%
\caption{ Comparison of the canonical distribution (black dots)
with the true joint distribution (red dots)
for a
$12$-states Potts model
on a $16 \times 16$ lattice
at the energy of the phase transition
interacting with a system of simple harmonic oscillators.}
\label{Potts_256_SHO}
\end{figure}
\begin{table}[H]
\caption{The values of $\delta$
for a two-dimensional, $12$-state Potts model
at $\beta_c$, the first-order transition.
The size of the environment
(a set of simple harmonic oscillators)
is shown in the left column.
Values of $\delta$
for the three sizes of the Potts model investigated
($8 \times 8$, $16 \times 16$, and $32 \times 32$)
are given in the next three columns.
}
\begin{center}
\begin{tabular}{r|r|r|r}
$N_E$ & $8 \times 8$ & $16 \times 16$ & $32 \times 32$ \\
\hline
$2^6$ & 2.367E-1 & 2.467E-1 & 2.474E-1 \\
$2^8$ & 1.485E-1 & 1.800E-1 & 1.770E-1 \\
$2^{10}$ & 6.106E-2 & 1.309E-1 & 1.280E-1 \\
$2^{12}$ & 1.837E-2 & 7.995E-2 & 9.415E-2 \\
$2^{14}$ & 4.868E-3 & 2.590E-2 & 6.801E-2 \\
$2^{16}$ & 1.234E-3 & 6.567E-3 & 3.435E-2 \\
$2^{18}$ & 3.096E-4 & 1.641E-3 & 9.576E-3 \\
$2^{20}$ & 7.748E-5 & 4.100E-4 & 2.412E-3
\end{tabular}
\end{center}
\label{table:_Potts_optimized_beta_c_delta}
\end{table}
Data for $\delta$
is shown in Table~\ref{table:_Potts_optimized_beta_c_delta}.
The scaling of $\delta$ with $N_E^{-1}$
is satisfied.
On the other hand,
the scaling with the size of the system, $N_S$,
is not even approximately satisfied.
This is expected, because the width of the peak in the energy
is not proportional to $N_S^{-1/2}$.
\section{Conclusions}
We have shown that the microcanonical ensemble,
which assumes that the system of interest is in an energy eigenstate,
does not give a correct description of an isolated system.
If the history of the system includes
any thermal contact with another system,
it has a probability distribution
that is spread over many eigenstates.
With certain reasonable conditions on the sizes
of the system of interest and the environment,
the probability distribution is that of the canonical ensemble.
This is true
even when the ``thermal reservoir'' is 1000 times smaller than the
system of interest.
At a second-order phase transition,
the energy peak is wider than at a point away from the transition
and
the approach to the canonical ensemble
with increased size is slower.
However, a macroscopic system is still well described by
the canonical distribution
when the
environment is substantially smaller than the system of interest.
For first-order transitions,
the double-peak probability distribution requires the environment to be
larger than the system,
although still not infinite.
In all cases,
the thermodynamic energy is a continuous variable,
and the validity of the canonical ensemble
does not require an infinite thermal reservoir.
\section*{Acknowledgement}
We would like to thank Lachlan Lancaster
for many helpful discussions.
One if us (RHS) would like to thank
Roberta Klatzky
for useful comments.
This research did not receive any specific grant from funding agencies
in the public, commercial, or
not-for-profit sectors.
\makeatletter
\renewcommand\@biblabel[1]{#1. }
\makeatother
|
\section{Introduction}
\label{intro}
In principle, there are two kinds of data collection designs for social network studies: either ``whole--network'' or ``egocentric'' designs \cite{RefMars}. Whole--network studies examine sets of interrelated actors linked by sets of relationships, the way they are extracted from certain relevant structural data sets \cite{RefWF}. The whole--network data collection designs gather, without exception, all available structural information about the concerned sets of actors and their relationships, as far as the examined social network object may be regarded as already constituting a distinctive and bounded social collectivity (with the important caveat, as Peter Marsden has apprised, that ``network boundaries are often permeable and/or ambiguous'' \cite{RefWF}).
On the other side, the egocentric data collection designs prioritize only particular focal actors and a limited range of their interactions by sampling and filtering out the pertinent information from a larger population of emprical structural data. Thus, an {\bf{ego--centered network}} (or {\bf{ego--net}} or {\bf{personal network}}) \cite{RefCBe} hinges on some actor in the network, called the {\bf{ego}}, and the set of actors surrounding the ego, called {\bf{alters}}. Usually, the network data sets for ego--centered networks originate from General Social Surveys \cite{RefSMH} or other measurements conducted in the context of social science surveys or/and questionnaires \cite{Marsden}.
Although whole--network and egocentric designs are conceptually and operationally distinct, they can be considered to be interrelated from a methodological point of view. At one end, given an extensive whole network, there is a trivial (but quite redundant) assemblage of a large number of ego--centered networks spanning the whole network, when each actor in the latter would have been treated as a separate ego \cite{RefMarsE}. On the other end, an egocentric design, in which egos are sampled as much densely as possible, may be reassembled to give an emerging whole--network construction \cite{RefKirke} (although as the outcome of a rather supernumerary data collection design).
Here, we are concentrating on the methodology of detecting an ego--centric decomposition of a whole network by following a rather parsimonious formal procedure. We are asking the following question: Given a whole network, how can one select a possibly minimal number of actors in such a way that, when these actors are considered as distinct egos, the total collection of the ego--centered networks, which are formed in this way, might reach all other actors and span the entire whole network? Of course, the answer to this question depends on what exactly one means by ``reaching''--``spanning'' and it is our intention to formulate this question in formal graph--theoretic terms by conceptualizing on the all--embracing structural patterns of adjacency, in which actors are embroiled in the network. In this way, a possible answer might be given by resorting to the formal graph--theoretic concept of the ``minimum dominating set,'' which is defined as the minimal set of actors whose members are adjacent to all other actors in the network. In other words, we propose a structurally conclusive solution to the problem of the efficient construction of a set of ego--centered networks, into which a given whole network might reduce, by introducing a decomposition of the whole network into smaller ego--centered subnetworks, which are built around the actors of a minimum dominating set of the whole network.\footnote{Of course, in such a decomposition, a minimum dominating set literally ``spans'' the vertex set. However, what is equally important to know is how the graph edge set is distributed among the constituent ego--centered subnetworks, a question that we intend to tackle at the end of section 3 (in Corollary 2).} Apparently, the problem at hand is an optimization problem in the sense that it focuses on finding the minimum dominating set of actors that achieves a fixed goal, i.e., in our case, the goal of reaching (``being adjacent'') to all other actors in the network. Note that, in the literature of social network analysis, one often encounters the reverse problem as well: to find a set of actors of fixed size that achieves a certain structural goal (as, for instance, the goal of optimally diffusing something through the network \cite{RefBor}).
Of course, it is well known that computationally the dominating set problem is $\mathcal{NP}$--complete \cite{RefJ}, although there are certain efficient algorithms (typically based on integer linear programming) for finding approximate solutions. Here, we are using the algorithms implemented in the Python--based SageMath software in order to find just the domination number of a given graph (network) and, subsequently, we are following our own methodology in order to detect all possible minimum dominating sets in the graph (network). The key point of our approach is based on a computational heuristics in partitioning the set of all actors (vertices) of a network (graph) into three sets, the always dominant, the possible dominant and the never dominant actors (vertices), when the domination number of the network is known. It turns out that knowing this partition simplifies the efficient computation of all minimum dominating sets (of course, always under the limitations imposed by the complexity of the problem on the size of the network). After defining the dominating ego--centered decomposition of a network in this way, we proceed in comparing the ensuing dominating ego--centered subnetworks through a number of structural measures that count the number of vertices and edges (relationships) inside and across the applied dominating decomposition.
So, in the second section of our investigation, we are discussing the fundamentals of the graph domination theory for social networks and sketch out our methodology for detecting all the minimum dominating sets of a given network (graph). In the third section, we are outlining the formal construction of the dominating ego--centered decomposition of a social network. Moreover, we are introducing the relevant structural measures and indices that arise in the context of this decomposition. Finally, in the last section, we are applying our techniques for the computation of all minimum dominating sets and the construction of the corresponding dominating ego--centered network decompositions to six examples of well known empirical social networks.
\section{Dominating Sets in Graphs}
\label{doms}
Let $G = (V, E)$ be a simple undirected graph with set of vertices $V$, where $|V| = n$,\footnote{For any set $X$, we denote by $|X|$ the {\bf{cardinality}} of $X$, i.e., the number of elements of $X$.} and set of edges $E = \{(u, v)\!: u \sim v, {\mbox{ for some }} u, v \in V, u \neq v\}$, where $u \sim v$ denotes that vertices $u$ and $v$ are adjacent (with the understanding that adjacency is a symmetric relationship, i.e., that edges $(u, v)$ and $(v, u)$ are identical). The open neighborhood of vertex $u$ is denoted as $N(u) = \{v \in V\!: (u, v) \in E\} = \{v \in V\!: v \sim u\}$, the closed neighborhood of $u$ is denoted as $N[u] = N(u) \cup \{u\}$, the degree of $u$ as $\deg(u) = |N(u)|$ and the (geodesic) distance between $u$ and $v$ as $d(u, v)$. For any $A \subset V$, we denote $N(A) = \bigcup_{v \in A} N(v), N[A] = N(A) \cup A$ and $d(u, A) = \min \{d(u, v)\!: v \in A\}$. Moreover, we assume that $G$ is the underlying graph of a social network and that the vertices of $G$ represent the actors of the social network, while the edges of $G$ represent the relationships/ties among actors in the social network. Thus, from now on, we are going to use interchangeably the terms network--graph, vertices--actors and edges--relationships/ties.
\begin{definition}
\cite{RefHHS} Let $D$ be a set of vertices of graph $G$ ($D \subset V$).
\begin{itemize}
\item[\small{$\bullet$}] $D$ is called {\bf{dominating set}} (or {\it{externally stable}} according to Claude Berge \cite{RefB}) if every vertex $v \in V$ is either an element of $D$ or is adjacent to an element of $D$. Moreover, $D$ is called {\it{minimal dominating set}} if no proper subset $D' \subset D$ is a dominating set.
\item[\small{$\bullet$}] $D$ is called a {\bf{minimum dominating set}} if the cardinality of $D$ is minimum among the cardinalities of any other dominating set.
\item[\small{$\bullet$}] The cardinality of a minimum dominating set $D$ is called the {\bf{domination number}} of graph $G$ and is denoted by $\gamma = \gamma(G)$.
\item[\small{$\bullet$}] Furthermore, $D$ is called {\bf{independent}} set (or {\it{internally stable}} set in the terminology of \cite{RefB}) whenever no two vertices of $D$ are adjacent. Note that an independent set $D$ is also a dominating set if and only if $D$ is a maximal independent set, in which case $D$ is called {\bf{independent dominating set}}.\footnote{The minimum cardinality of an independent dominating set of $G$ is called the {\bf{independent domination number}} of graph $G$ and is denoted by $i = i(G)$. Note that the minimum dominating set of $G$ is not necessarily independent in $G$ and, in general, $\gamma(G) \leq i(G)$.}
\end{itemize}
\end{definition}
The standard reference to graph domination theory is the book of T.W. Haynes, S.T. Hedetniemi and P.J. Slater, \textit{Fundamentals of Domination in Graphs} \cite{RefHHS}, where the proofs of the following basic results can be found.\\
A vertex $v \in D$ is called an {\it{enclave}} of $D$ if $N[v] \subseteq D$ and $v$ is called an {\it{isolate}} of $D$ if $N(v) \subseteq V \smallsetminus D$. A set is called {\it{enclaveless}} if it does not contain any enclaves.
Apparently, $D$ is is a dominating set if and only if one of the following holds (cf. \cite{RefHHS}):
\begin{itemize}
\item[\small{$\bullet$}] $N[D] = V$,
\item[\small{$\bullet$}] for every $v \in V \smallsetminus D$, $d(v, D) \leq 1$,
\item[\small{$\bullet$}] $V \smallsetminus D$ is enclaveless.
\end{itemize}
The first existence theorems about dominating sets in graphs were given by O. Ore in his 1962 book, \textit{Theory of Graphs} \cite{RefO}, and they are included in the next theorem and the corollary that follows:
\begin{theorem}
\label{Ore1}
\cite{RefO} A dominating set $D$ is a minimal dominating set if and only if, for each vertex $u \in D$, one of the following two conditions holds:
\begin{itemize}
\item[\small{$\bullet$}] either $u$ is an isolate of $D$,
\item[\small{$\bullet$}] or there exists a vertex $v \in V \smallsetminus D$ such that $N[v] \cap D = \{u\}$, in which case $v$ is called a {\bf{private neighbor}} of $u$.
\end{itemize}
\end{theorem}
\begin{corollary}
\label{Ore2}
\cite{RefO} Let $G$ be a graph with no isolated vertices. Then:
\begin{itemize}
\item[\small{$\bullet$}] $G$ has a dominating set $D$ and $D$'s complement $V \smallsetminus D$ is also a dominating set.
\item[\small{$\bullet$}] $\gamma(G) \leq \frac{n}{2}$.
\end{itemize}
\end{corollary}
Let us, from now on, assume that $G$ contains no isolated vertices. Furthermore, let us denote by $\mathcal{D} = \mathcal{D}(G)$ the set of all (minimum) dominating sets of graph $G$. Then we claim that we can partition the set of vertices $V$ in three distinct types of vertices, which are defined as follows:
\begin{definition} Let $v \in V$.
\begin{itemize}
\item[\small{$\bullet$}] Vertex $v$ is said to be {\bf{always dominant}} if, for every $D \in \mathcal{D}$, $v \in D$. We denote by $\frak{A} = \frak{A}(G)$ the set of all always dominant vertices of $G$.
\item[\small{$\bullet$}] Vertex $v$ is said to be {\bf{possibly dominant}} if there exists $D \in \mathcal{D}$ such that $v \in D$. We denote by $\frak{P} = \frak{P}(G)$ the set of all possibly dominant vertices of $G$.
\item[\small{$\bullet$}] Vertex $v$ is said to be {\bf{never dominant}} if, for every $D \in \mathcal{D}$, $v \notin D$. We denote by $\frak{N} = \frak{N}(G)$ the set of all never dominant vertices of $G$.
\end{itemize}
Thus, there exists the following partition of $V$, that we are going to call {\bf{domination partition}},
\[
V = \frak{A} \cup \frak{P} \cup \frak{N}.
\]
Note that at least one of the sets $\frak{A}, \frak{P}, \frak{N}$ must be nonempty.
\end{definition}
Below, we write $G - v$ for the induced subgraph $G(V \smallsetminus \{v\})$.
Our main result is the following:
\begin{lemma}
Let $\gamma$ be the domination number of $G$ and let $v$ be a vertex.
\begin{description}
\item[(a)] Vertex $v$ is always a dominant vertex if and only if $\gamma(G - v) > \gamma$, i.e.,
\[
\frak{A} = \{v \in V\!: \gamma(G - v) > \gamma\}.
\]
\item[(b)] If $|\frak{A}| = \gamma$, then
\begin{eqnarray*}
\frak{N} &=& V \smallsetminus \frak{A},\\
\frak{P} &=& \varnothing.
\end{eqnarray*}
In other words, if $|\frak{A}| = \gamma$, then there exists a unique (minimum) dominating set of $G$ (i.e., $|\mathcal{D}(G)| = 1$).
\item[(c)] If $|\frak{A}| := \gamma_1 < \gamma$, then there exists a set $\frak{R} \subset V \smallsetminus \frak{A}$, with $|\frak{R}| = \gamma - \gamma_1$ ($> 0$), such that
$\frak{A} \cup \frak{R} \in \mathcal{D}(G)$, and, therefore,
\begin{eqnarray*}
\frak{P} &=& \{\frak{R} \subset V \smallsetminus \frak{A}\!: |\frak{R}| = \gamma - \gamma_1 {\mbox{ and }} \frak{A} \cup \frak{R} \in \mathcal{D}(G)\},\\
\frak{N} &=& V \smallsetminus (\frak{A} \cup \frak{P}).
\end{eqnarray*}
\end{description}
\end{lemma}
\begin{proof} Part (a) follows from the ``uniqueness'' proof of Gunther, Hartnell, Markus and Rall \cite{RefGHM}. Part (b) is plain to see, while (c) is a consequence of Ore's fundamental results (Theorem 1 and Corollary 1).
\end{proof}
\begin{definition}
For a graph $G$, the {\bf{index of domination multiplicity}} $m = m(G)$ is defined as
\[
m = 1 - \frac{|\frak{A}|}{\gamma}.
\]
\end{definition}
Clearly, $m$ varies in $[0, 1]$ and it is $m = 0$, for any graph with unique dominating set, while $m = 1$, for a graph without always dominant vertices (for instance, this is the case with cycle graphs). In this way, when an applied analyst knows the domination number together with the index of domination multiplicity, she possesses comprehensive information in order to be able to infer the ``complexity'' of the network structure at the global level. Of course, in practical situations, this knowledge works like a new sort of centrality index, which, dissimilarly to the local computation for most centrality indices, now it is based on a combinatorial assessment of all possible structural circumstances enabling the constitution of the domination partition of an empirical network at hand. In this context, the number of always dominant vertices is inversely proportional to the size of the domination index and directly proportional to the graph domination number.
In our numerical computations, after using the SageMath software \cite{sage} to compute the domination number of a graph, we are computing all the minimum dominating sets of the graph through the following two algorithms that we have implemented in Python.
\begin{algorithm}[H]
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\underline{function AllwaysDom} $(G,\gamma)$\;
\Input{Graph $G$ and domination number $\gamma$ of $G$.}
\Output{The set $\frak{A}(G)$ of the always dominant vertices.}
\For{$\mathrm{node} \in G.\mathrm{nodes}$}
{\label{forins}
\If{$\gamma(G \smallsetminus \mathrm{node}) > \gamma$}
{
$\frak{A}(G) \leftarrow \mathrm{node}$\;
}}
\caption{Algorithm for finding the set $\frak{A}(G)$ of all always dominant vertices of graph $G$, when the domination number $\gamma = \gamma(G)$ is known.}
\end{algorithm}
\begin{algorithm}[H]
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\underline{function AllDoms} $(G,\frak{A}(G),\gamma)$\;
\Input{Graph $G$, the set $\frak{A}$ of all always dominant vertices and the domination number $\gamma$ of $G$.}
\Output{The collection $\mathcal{D}$ of all minimum dominating sets of $G$.}
$r \leftarrow \gamma - |\frak{A}|$\;
merger $\leftarrow$ the set all subsets of $G \smallsetminus \frak{A}$ of size $r$\;
\For{$S \in \mathrm{merger}$}
{\label{forins}
$S \leftarrow S \cup \frak{A}$\;
{
\If{$S \mathrm{ \ is \ Dominating \ Set}$}
{$\mathcal{D} \leftarrow S$}
}}
\caption{Algorithm for finding the collection $\mathcal{D} = \mathcal{D}(G)$ of all the minimum dominating sets of graph $G$, when the domination number $\gamma = \gamma(G)$ is known.}
\end{algorithm}
\section{Ego--Centered Decompositions in Graphs}
\label{egod}
\begin{definition}
Let $G = (V, E)$ be a simple undirected graph, $u \in U$ and $U \subset V$. The subgraph $G(U)$ induced by $U$ is called {\bf{ego--centered}} subgraph if
\[
U = N[u],
\]
in which case vertex $u$ is called the {\bf{ego}} of $G(N[u])$ and all vertices $w \in N(u)$ are called {\bf{alters}} of $G(N[u])$. Moreover, an alter $w \in N(u)$ is called:
\begin{itemize}
\item[\small{$\bullet$}] {\bf{private alter}} if $N(w) \subset N[u]$ and
\item[\small{$\bullet$}] {\bf{public alter}} if $N(w) \smallsetminus N[u] \neq \varnothing$.
\end{itemize}
\end{definition}
For $u_1, u_2$ two vertices in $G$, let us consider the two ego--centered subgraphs $G(N[u_1]), G(N[u_2])$.
\begin{itemize}
\item[\small{$\bullet$}] If $w \in N[u_1] \cap N[u_2]$ ($\neq \varnothing$), then vertex $w$ is called {\bf{shared alter}} by the two ego--centered subgraphs. Note that, as far as $u_1 \neq u_2$, a shared alter is necessarily a public alter.
\end{itemize}
Furthermore, let $w_1 \in N(u_1), w_2 \in N(u_2)$ be two alters such that $w_1 \neq w_2$ and $w_1 \sim w_2$. Note that, whenever $u_1 = u_2$, $w_1, w_2$ become alters in the same ego--centered subgraph. Then we have the following cases:
\begin{itemize}
\item[\small{$\bullet$}] If $u_1 \neq u_2$ and $u_1 \sim u_2$, then edge $(u_1, u_2)$ is called {\bf{bridge of egos}}.
\item[\small{$\bullet$}] If $w_1$ is a private alter in $G(N[u_1])$ and $w_2$ is a private alter in $G(N[u_2])$, then edge $(w_1, w_2)$ is called {\bf{bridge of private alters}}.
\item[\small{$\bullet$}] If $w_1$ is a public alter in $G(N[u_1])$ and $w_2$ is a public alter in $G(N[u_2])$, then edge $(w_1, w_2)$ is called {\bf{bridge of public alters}}.
\item[\small{$\bullet$}] If $w_1$ is a private alter in $G(N[u_1])$ and $w_2$ is a public alter in $G(N[u_2])$, then edge $(w_1, w_2)$ is called {\bf{bridge of private--to--public alters}}.
\end{itemize}
\begin{definition}
For any minimum dominating set $D \in \mathcal{D}(G)$, the family of ego-centered subgraphs $\{G(N[u])\}_{u \in D}$ is called {\bf{$D$--dominating ego--centered decomposition}} of graph $G$.
Moreover, this decomposition induces the following partitions of the vertex set $V$ and the edge set $E$:
\begin{eqnarray*}
V &=& D \cup V_{\mathrm{private}} \cup V_{\mathrm{public}},\\
E &=& E_{\mathrm{ego}} \cup E_{\mathrm{private}} \cup E_{\mathrm{public}} \cup E_{\mathrm{private}-\mathrm{public}},
\end{eqnarray*}
where the sets of private and (shared) public vertices, $V_{\mathrm{private}} = V_{\mathrm{private}}(D), V_{\mathrm{public}} = V_{\mathrm{public}}(D)$, and the sets of ego, private, public and private--public bridges, $E_{\mathrm{ego}} = E_{\mathrm{ego}}(D), E_{\mathrm{private}} = E_{\mathrm{private}}(D), E_{\mathrm{public}} = E_{\mathrm{public}}(D), E_{\mathrm{private}-\mathrm{public}} = E_{\mathrm{private}-\mathrm{public}}(D)$, are defined as follows:
\begin{eqnarray*}
V_{\mathrm{private}} &=& \{w \in V \smallsetminus D\!: \exists u \in D {\mbox{ such that }} w \in N(u) \smallsetminus D \\
& & \,\, {\mbox{ and }} N(w) \subset N[u]\},\\
V_{\mathrm{public}} &=& \{w \in V \smallsetminus D\!: \exists u_1, u_2 \in D {\mbox{ such that }} u_1 \neq u_2, \\
& & \,\,\,\, N(u_1) \cap N(u_2) \neq \varnothing {\mbox{ and }} w \in N(u_1) \cap N(u_2)\},\\
E_{\mathrm{ego}} &=& \{(u_1, u_2) \in E\!: u_1 \neq u_2 {\mbox{ and }} u_1, u_2 \in D\},\\
E_{\mathrm{private}} &=& \{(w_1, w_2) \in E\!: w_1 \neq w_2 {\mbox{ and }} w_1, w_2 \in V_{\mathrm{private}}\},\\
E_{\mathrm{public}} &=& \{(w_1, w_2) \in E\!: w_1 \neq w_2 {\mbox{ and }} w_1, w_2 \in V_{\mathrm{public}}\},\\
E_{\mathrm{private}-\mathrm{public}} &=& \{(w_1, w_2) \in E\!: w_1 \neq w_2 {\mbox{ and }} w_1 \in V_{\mathrm{private}}, w_2 \in V_{\mathrm{public}}\}.
\end{eqnarray*}
\end{definition}
\begin{proposition}
For any $D$--dominating ego--centered decomposition of graph $G$ with domination number $\gamma$,
\begin{eqnarray*}
|V| - |V_{\mathrm{private}}| - |V_{\mathrm{public}}| &=& \gamma ,\\
|E| - |E_{\mathrm{private}}| - |E_{\mathrm{public}}| &=& \sum_{u \in D} \deg(u) - |E_{\mathrm{ego}}| + |E_{\mathrm{private}-\mathrm{public}}|.
\end{eqnarray*}
\end{proposition}
\begin{proof}
This is a direct consequence of the previous definitions and the degree sum formula (handshaking lemma) for graph $G$.
\end{proof}
\begin{corollary}
Let $\{G(N[u])\}_{u \in D}$ be a $D$--dominating ego--centered decomposition of $G$. Then we have:
\begin{itemize}
\item[\small{$\bullet$}] $D$ is an independent dominating set if and only if $E_{\mathrm{ego}} = \varnothing$.
\item[\small{$\bullet$}] $G$ is the disjoint union of the ego-centered subgraphs $\{G(N[u])\}_{u \in D}$, i.e., $G = \sum_{u \in D} G(N[u])$, if and only if $V_{\mathrm{public}} = E_{\mathrm{ego}} = E_{\mathrm{public}} = E_{\mathrm{private}-\mathrm{public}} = \varnothing$, in which case $G$ is disconnected to a forest of ego--centered subgraphs.
\end{itemize}
\end{corollary}
Let us also remark that, for each graph $G$, one may possibly choose different minimum dominating sets $D$ and, thus, different $D$--dominating ego-centered decompositions. However, all these choices depend on the type of the graph $G$ through the value of the index of domination multiplicity $m$ for $G$. Apparently, when $m$ tends to $0$, then there are fewer choices of minimum dominating sets than when $m$ tends to $1$. Moreover, in case of multiple minimum dominating sets, Proposition 1 suggests that one may select a $D$--dominating ego-centered decomposition according to whether this decomposition maximizes or minimizes the following quantities:
\begin{itemize}
\item[(1)] $\sum_{u \in D} \deg(u)$, i.e., the density of connections between egos and alters;
\item[(2)] $|E_{\mathrm{ego}}|$, i.e., the density of connections among egos;
\item[(3)] $|E_{\mathrm{private}}|$, i.e., the density of connections among private alters;
\item[(4)] $|E_{\mathrm{public}}|$, i.e., the density of connections among public alters;
\item[(5)] $|E_{\mathrm{private}-\mathrm{public}}|$, i.e., the density of connections between private and public alters.
\end{itemize}
One may notice that criterion (1) represents a condition on the total valence of egos and criterion (5) focuses on the separability among the ego--centered subnetworks within a given decomposition. Moreover, the three criteria (2), (3) and (4) are based on the cohesiveness of the set of egos, private alters or public alters, respectively. As a matter of fact, (2), (3) and (4) measure the number of triangles with one of their vertices being an ego and the other two being private alters or public alters, respectively. Similarly, (5) measures the number of 3--paths between pairs of egos. Nevertheless, according to Proposition 1, the five criteria are interdependent in the sense that, for a given network $G$, they cannot be all maximized or minimized simultaneously, since $\sum_{u \in D} \deg(u) + |E_{\mathrm{private}}| + |E_{\mathrm{public}}| - |E_{\mathrm{ego}}| + |E_{\mathrm{private}-\mathrm{public}}| = |E|$.
\section{Applications}
\label{appls}
\subsection{Hage \& Harary's Voyaging Network}
\label{appls1}
The first example of our computation is on Hage and Harary's \cite{RefHH} voyaging network among the 14 western Carolines Islands: Satawal, Ulithi, Woleai, Puluwat, Faraulep, Fais, Pulusuk, Pulap, Elato, Ifaluk, Sorol, Namonuito, Eauripik and Lamotrek. For the corresponding graph $G$, the domination number is equal to 3, there are two always dominant vertices (Elato and Fais) and two possibly dominant vertices (Puluwat and Pulap). The domination partition of this network is plotted in Figure 1.
\begin{figure}[H]
\resizebox{1.4\columnwidth}{!}
\hspace*{-5cm}
\includegraphics{voyaging.png}
}
\caption{The Hage \& Harary's Voyaging Network.}
\label{fig:vn}
\end{figure}
In Harary's voyaging network, the index of domination multiplicity is $m = 0.333$ and the properties of the two minimum dominating sets are shown in Table 1. We observe that both minimum dominating sets [Elato, Fais, Puluwat] and [Elato, Fais, Pulap] are independent sets, their egos have sum of degrees equal to 11 and the number of bridges among their private--public alters is 4. In other words, this is an illustration of a relatively simple network (low domination number, low multiplicity number) with two minimum dominating sets that seem to be more or less equally desirable since the rows for them in Table 1 are identical.
\begin{table}[H]
\small
\caption{The Minimum Dominating Sets (MDSs) of the Hage \& Harary's voyaging network (E = Elato, F = Fais, Pulu = Puluwat, Pul = Pulap).}
\label{table:vn}
\begin{tabular}{lccccccc}
\bf{MDSs} & $|V_{\mathrm{private}}|$ & $|V_{\mathrm{public}}|$ & $\sum_{v \in S} \deg(v)$ & $|E_{\mathrm{private}}|$ & $|E_{\mathrm{public}}|$ & $|E_{\mathrm{private}-\mathrm{public}}|$ & $|E_{\mathrm{ego}}|$ \\
E, F, Pulu & 5 & 6 & 11 & 3 & 6 & 4 & 0 \\
E, F, Pul & 5 & 6 & 11 & 3 & 6 & 4 & 0 \\
\end{tabular}
\end{table}
For those who find examples helpful in understanding concepts, let us indicate the egos, private and public alters corresponding to the decompositions induced by the two minimum dominating sets of Harary's voyaging network.
\begin{figure}[H]
\resizebox{1\columnwidth}{!}
\includegraphics{pw.png}
}
\caption{Egos and alters in the Hage \& Harary's voyaging network.}
\label{fig:kk}
\end{figure}
For the first minimum dominating set, Elato, Fais and Puluwat are the three egos (colored red in Figure 2). Ego Elato has two alters, Haluk and Lamotrek, both being public alters (colored green). Ego Fais has five alters, two of them, Ulithi and Sorol, being private alters (colored cyan) and the other three, Eauripik, Woleai and Farauleo, being public alters. Ego Puluwat has four alters, one of them, Satwal, being public alter and three, Pulap, Namonuito and Pulusuk, being private alters. Thus, in total, the minimum dominating set of Elato, Fais and Puluwat has five private alters (Ulithi, Sorol, Pulap, Namonuito and Pulusuk) and six public alters (Haluk, Lamotrek, Eauripik, Woleai, Farauleo and Satwal).
The second minimum dominating set is composed of three egos, Elato, Fais and Pulap. The alters of egos Elato and Fais have been already identified. Ego Pulap has also four alters, one of them, Satwal, being public alter and three, Puluwat, Namonuito and Pulusuk, being private alters. Thus, in total, the minimum dominating set of Elato, Fais and Puluwat has five private alters (Ulithi, Sorol, Puluwat, Namonuito and Pulusuk) and six public alters (Haluk, Lamotrek, Eauripik, Woleai, Farauleo and Satwal).
\subsection{Krackhardt's Kite}
\label{appls2}
Next, we examine Krackhardt's Kite network \cite{RefKn} composed of 10 vertices. The corresponding graph $G$ has domination number equal to 2, but there is a unique minimum dominating set and, thus, there are no possibly dominant vertices. The domination partition of $G$ is plotted in Figure 3.
\begin{figure}[H]
\resizebox{1.1\columnwidth}{!}
\hspace*{-5cm}
\includegraphics{kite.png}
}
\caption{The Krackhardt kite.}
\label{fig:kk}
\end{figure}
The index of domination multiplicity of Krackhardt's Kite is $m = 0$ and the properties of the unique minimum dominating set are shown in Table 2: [3, 8] is an independent dominating set, the single ego has sum of degrees equal to 8 and the number of bridges among private--public alters is 4.
\begin{table}[H]
\small
\caption{The Minimum Dominating Sets (MDSs) of the Krackhardt kite network (with the traditional labelling of vertices: 1 = Andre, 2 = Beverley, 3 = Carol, 4 = Diane, 5 = Ed, 6 = Fernando, 7 = Garth, 8 = Heather, 9 = Ike, 10 = Jane).}
\begin{tabular}{lccccccc}
\bf{MDSs} & $|V_{\mathrm{private}}|$ & $|V_{\mathrm{public}}|$ & $\sum_{v \in S} \deg(v)$ & $|E_{\mathrm{private}}|$ & $|E_{\mathrm{public}}|$ & $|E_{\mathrm{private}-\mathrm{public}}|$ & $|E_{\mathrm{ego}}|$ \\
[3, 8] & 5 & 3 & 8 & 3 & 3 & 4 & 0 \\
\end{tabular}
\end{table}
\subsection{Florentine Families}
\label{appls3}
The third example of our computation is the Florentine Families network in the form that Breiger and Pattison have used \cite{RefBP} (extracted from a subset of data on the social relations among Renaissance Florentine families collected by John Padgett). The 15 vertices of this network are the Florentine families: Acciaiuoli, Albizzi, Barbadori, Bischeri, Castellani, Ginori, Guadagni, Lamberteschi, Medici, Pazzi, Peruzzi, Ridolfi, Salviati, Strozzi and Tornabuoni. The corresponding graph $G$ has domination number equal to 5, there is a single always dominant vertex (Medici) and 9 possibly dominant vertices, shown in the domination partition plot of Figure 4.
\begin{figure}[H]
\resizebox{1.25\columnwidth}{!}{%
\hspace*{-5cm}
\includegraphics{florentine.png}
}
\caption{The Florentine Families network.}
\label{fig:ff}
\end{figure}
Now, the index of domination multiplicity is $m = 0.8$ and the properties of the 20 minimum dominating sets are shown in Table 3. Notice that [Albizzi, Guadagni, Medici, Salviati, Strozzi] is the dominating set with maximum sum of egos' degrees (equal to 19), there are 5 dominating sets with zero number of bridges among egos (i.e., 5 independent dominating sets) and 12 dominating sets with zero number of bridges among private--public alters.
\begin{table}[H]
\small
\caption{The Minimum Dominating Sets (MDSs) of the Florentine Families network (0 = Albizzi, 1 = Castellani, 2 = Ginori, 3 = Guadagni, 4 = Lamberteschi, 5 = Medici, 6 = Pazzi, 7 = Peruzzi, 8 = Salviati, 9 = Strozzi).}
\begin{tabular}{lccccccc}
\bf{MDSs} & $|V_{\mathrm{private}}|$ & $|V_{\mathrm{public}}|$ & $\sum_{v \in S} \deg(v)$ & $|E_{\mathrm{private}}|$ & $|E_{\mathrm{public}}|$ & $|E_{\mathrm{private}-\mathrm{public}}|$ & $|E_{\mathrm{ego}}|$ \\
[1, 2, 3, 5, 6] & 2 & 8 & 15 & 0 & 5 & 0 & 0 \\
[0, 4, 5, 6, 7] & 2 & 8 & 14 & 0 & 7 & 0 & 1 \\
[2, 4, 5, 8, 9] & 3 & 7 & 14 & 0 & 5 & 2 & 1 \\
[0, 3, 5, 8, 9] & 5 & 5 & 19 & 0 & 2 & 2 & 3 \\
[2, 3, 5, 6, 7] & 2 & 8 & 15 & 0 & 5 & 0 & 0 \\
[2, 3, 5, 7, 8] & 3 & 7 & 16 & 0 & 5 & 0 & 1 \\
[0, 3, 5, 6, 7] & 3 & 7 & 17 & 0 & 5 & 0 & 2 \\
[2, 4, 5, 6, 7] & 1 & 9 & 12 & 0 & 8 & 0 & 0 \\
[2, 3, 5, 8, 9] & 4 & 6 & 17 & 0 & 2 & 2 & 1 \\
[0, 4, 5, 7, 8] & 3 & 7 & 15 & 0 & 7 & 0 & 2 \\
[2, 3, 5, 6, 9] & 3 & 7 & 16 & 0 & 2 & 2 & 0 \\
[0, 4, 5, 8, 9] & 4 & 6 & 16 & 0 & 4 & 2 & 2 \\
[0, 1, 3, 5, 6] & 3 & 7 & 17 & 0 & 5 & 0 & 2 \\
[2, 4, 5, 6, 9] & 2 & 8 & 13 & 0 & 5 & 2 & 0 \\
[0, 3, 5, 7, 8] & 4 & 6 & 18 & 0 & 5 & 0 & 3 \\
[2, 4, 5, 7, 8] & 2 & 8 & 13 & 0 & 8 & 0 & 1 \\
[0, 4, 5, 6, 9] & 3 & 7 & 15 & 0 & 4 & 2 & 1 \\
[0, 1, 3, 5, 8] & 4 & 6 & 18 & 0 & 5 & 0 & 3 \\
[0, 3, 5, 6, 9] & 4 & 6 & 18 & 0 & 2 & 2 & 2 \\
[1, 2, 3, 5, 8] & 3 & 7 & 16 & 0 & 5 & 0 & 1 \\
\end{tabular}
\end{table}
\subsection{Karate Club}
\label{appls4}
The fourth example is Zachary's Karate Club network \cite{RefZa} composed of 34 vertices. The corresponding graph $G$ has domination number equal to 4, there are two always dominant vertices (0 and 33) and 6 possibly dominant vertices (5, 6, 16, 24, 25, 31), as they are shown in the domination partition plot of Figure 5.
\begin{figure}[H]
\resizebox{1.3\columnwidth}{!}{%
\hspace*{-5cm}
\includegraphics{karate.png}
}
\caption{The Karate Club network.}
\label{fig:kc}
\end{figure}
In the Karate Club network, the index of domination multiplicity is $m = 0.5$ and the properties of the 9 minimum dominating sets are shown in Table 4. There are 2 dominating sets with maximum sum of egos' degrees (equal to 43), 2 dominating sets with zero number of bridges among egos (i.e., 2 independent dominating sets) and 3 dominating sets with minimum number of bridges among private--public alters (equal to 16).
\begin{table}[H]
\small
\caption{The Minimum Dominating Sets (MDSs) of the Karate Club Network.}
\begin{tabular}{lccccccc}
\bf{MDSs} & $|V_{\mathrm{private}}|$ & $|V_{\mathrm{public}}|$ & $\sum_{v \in S} \deg(v)$ & $|E_{\mathrm{private}}|$ & $|E_{\mathrm{public}}|$ & $|E_{\mathrm{private}-\mathrm{public}}|$ & $|E_{\mathrm{ego}}|$ \\
[0, 5, 31, 33] & 15 & 15 & 43 & 3 & 18 & 17 & 3 \\
[0, 16, 31, 33] & 15 & 15 & 41 & 4 & 19 & 16 & 2 \\
[0, 5, 24, 33] & 15 & 15 & 40 & 3 & 19 & 17 & 1 \\
[0, 5, 25, 33] & 15 & 15 & 40 & 3 & 19 & 17 & 1 \\
[0, 6, 31, 33] & 15 & 15 & 43 & 3 & 18 & 17 & 3 \\
[0, 6, 25, 33] & 15 & 15 & 40 & 3 & 19 & 17 & 1 \\
[0, 16, 24, 33] & 15 & 15 & 38 & 4 & 20 & 16 & 0 \\
[0, 6, 24, 33] & 15 & 15 & 40 & 3 & 19 & 17 & 1 \\
[0, 16, 25, 33] & 15 & 15 & 38 & 4 & 20 & 16 & 0 \\
\end{tabular}
\end{table}
\subsection{Southern Women}
\label{appls5}
The next example is the Southern Women network \cite{RefDavis}, which is a two-mode network composed of 18 southern women and 14 informal social events. The corresponding bipartite graph $G$ has domination number equal to 5, there are no always dominant vertices, but there are 18 possibly dominant vertices (8 women and 10 events), as they are shown in the domination partition plot of Figure 6.
\begin{figure}[H]
\resizebox{1.25\columnwidth}{!}{%
\hspace*{-5cm}
\includegraphics{southern.png}
}
\caption{The Southern Women network.}
\label{fig:lm}
\end{figure}
The index of domination multiplicity of the network of Southern Women is $m = 1$ (since $|\frak{A}| = 0$) and the properties of the 36 minimum dominating sets are shown in Table 5. There is a single dominating set with maximum sum of egos' degrees (equal to 52), all dominating sets have zero number of bridges among egos (because the graph is bipartite) and there are 12 dominating sets with minimum number of bridges among private--public alters (equal to 2).
\begin{table}[H]
\tiny
\caption{The Minimum Dominating Sets (MDSs) of the Southern Women Network.}
\begin{tabular}{lccccccc}
\bf{MDSs} & $|V_{\mathrm{private}}|$ & $|V_{\mathrm{public}}|$ & $\sum_{v \in S} \deg(v)$ & $|E_{\mathrm{private}}|$ & $|E_{\mathrm{public}}|$ & $|E_{\mathrm{private}-\mathrm{public}}|$ & $|E_{\mathrm{ego}}|$ \\
[0, 1, 6, E8, E11] & 0 & 27 & 38 & 0 & 53 & 0 & 2 \\
[1, 6, E3, E7, E9] & 0 & 27 & 44 & 0 & 49 & 0 & 4 \\
[1, 7, E7, E8, E11] & 0 & 27 & 43 & 0 & 49 & 0 & 3 \\
[1, 6, E7, E8, E11] & 0 & 27 & 44 & 0 & 48 & 0 & 3 \\
[5, 7, E4, E8, E11] & 0 & 27 & 36 & 0 & 55 & 0 & 2 \\
[0, 5, 6, E8, E11] & 0 & 27 & 37 & 0 & 54 & 0 & 2 \\
[1, 7, E4, E8, E11] & 0 & 27 & 37 & 0 & 55 & 0 & 3 \\
[0, 5, 6, E8, E9] & 0 & 27 & 45 & 0 & 46 & 0 & 2 \\
[1, 7, E5, E9, E11] & 0 & 27 & 39 & 0 & 53 & 0 & 3 \\
[1, 7, E3, E8, E11] & 0 & 27 & 39 & 0 & 53 & 0 & 3 \\
[0, 1, 4, E8, E11] & 0 & 27 & 36 & 0 & 55 & 0 & 2 \\
[1, 6, E5, E8, E11] & 0 & 27 & 42 & 0 & 50 & 0 & 3 \\
[0, 1, 7, E8, E11] & 0 & 27 & 37 & 0 & 54 & 0 & 2 \\
[1, 6, E7, E8, E9] & 0 & 27 & 52 & 0 & 41 & 0 & 4 \\
[1, 6, E3, E8, E11] & 0 & 27 & 40 & 0 & 52 & 0 & 3 \\
[1, 6, E4, E8, E11] & 0 & 27 & 38 & 0 & 54 & 0 & 3 \\
[1, 6, E3, E8, 'E9] & 0 & 27 & 48 & 0 & 45 & 0 & 4 \\
[1, 2, 6, E7, E9] & 0 & 27 & 42 & 0 & 50 & 0 & 3 \\
[1, 6, E5, E9, E10] & 0 & 27 & 41 & 0 & 52 & 0 & 4 \\
[1, 3, 4, E5, E9] & 0 & 27 & 39 & 0 & 53 & 0 & 3 \\
[1, 4, E7, E8, E11] & 0 & 27 & 42 & 0 & 49 & 0 & 2 \\
[1, 6, E5, E9, E12] & 0 & 27 & 42 & 0 & 51 & 0 & 4 \\
[0, 1, 6, E8, E9] & 0 & 27 & 46 & 0 & 46 & 0 & 3 \\
[5, 6, E4, E8, E11] & 0 & 27 & 37 & 0 & 54 & 0 & 2 \\
[1, 7, E5, E8, E11] & 0 & 27 & 41 & 0 & 51 & 0 & 3 \\
[4, 5, E4, E8, E11] & 0 & 27 & 35 & 0 & 56 & 0 & 2 \\
[1, 6, E5, E9, E11] & 0 & 27 & 40 & 0 & 53 & 0 & 4 \\
[5, 6, E4, E8, E9] & 0 & 27 & 45 & 0 & 46 & 0 & 2 \\
[1, 6, E6, E7, E9] & 0 & 27 & 46 & 0 & 48 & 0 & 5 \\
[1, 3, 7, E5, E9] & 0 & 27 & 40 & 0 & 52 & 0 & 3 \\
[0, 5, 7, E8, E11] & 0 & 27 & 36 & 0 & 55 & 0 & 2 \\
[1, 3, 6, E5, E9] & 0 & 27 & 41 & 0 & 51 & 0 & 3 \\
[1, 6, E5, E7, E9] & 0 & 27 & 46 & 0 & 47 & 0 & 4 \\
[1, 6, E4, E8, E9] & 0 & 27 & 46 & 0 & 47 & 0 & 4 \\
[1, 6, E5, E8, E9] & 0 & 27 & 50 & 0 & 43 & 0 & 4 \\
[0, 4, 5, E8, E11] & 0 & 27 & 35 & 0 & 56 & 0 & 2 \\
\end{tabular}
\end{table}
\subsection{Les Miserables}
\label{appls6}
The last example is the network of Les Miserables \cite{RefKLM}, which is the network of co-occurrences of 77 characters in Victor Hugo's novel ``{\it{Les Misérables}}.'' The corresponding graph $G$ has domination number equal to 10, there are 7 always dominant vertices and 6 possibly dominant vertices, as they are shown in the domination partition plot of Figure 7.
\begin{figure}[H]
\resizebox{1.3\columnwidth}{!}{%
\hspace*{-5cm}
\includegraphics{miserables.png}
}
\caption{The network of Les Miserables.}
\label{fig:lm}
\end{figure}
In the network of Les Miserables, the index of domination multiplicity is $m = 0.3$ and the properties of the 8 minimum dominating sets are shown in Table 6. There is a single dominating set with maximum sum of egos' degrees (equal to 130), a single dominating set with minimum number of bridges among egos (equal to 9) and 2 dominating sets with minimum number of bridges among private--public alters (equal to 39).
\begin{table}[H]
\tiny
\caption{The Minimum Dominating Sets (MDSs) of the network of Les Miserables.}
\begin{tabular}{lccccccc}
\bf{MDSs} & $|V_{\mathrm{private}}|$ & $|V_{\mathrm{public}}|$ & $\sum_{v \in S} \deg(v)$ & $|E_{\mathrm{private}}|$ & $|E_{\mathrm{public}}|$ & $|E_{\mathrm{private}-\mathrm{public}}|$ & $|E_{\mathrm{ego}}|$ \\
\tiny{[0, 11, 23, 25, 45, 46, 48, 49, 51, 67]} & 41 & 26 & 116 & 44 & 53 & 50 & 9 \\
\tiny{[0, 11, 23, 25, 28, 46, 48, 49, 51, 67]} & 41 & 26 & 119 & 44 & 52 & 49 & 10 \\
\tiny{[0, 11, 23, 25, 28, 47, 48, 49, 51, 57]} & 39 & 28 & 130 & 28 & 69 & 39 & 12 \\
\tiny{[0, 11, 23, 25, 28, 46, 48, 49, 51, 57]} & 38 & 29 & 129 & 28 & 69 & 39 & 11 \\
\tiny{[0, 11, 23, 25, 28, 47, 48, 49, 51, 67]} & 42 & 25 & 120 & 44 & 52 & 49 & 11 \\
\tiny{[0, 11, 23, 25, 45, 47, 48, 49, 51, 57]} & 39 & 28 & 127 & 28 & 70 & 40 & 11 \\
\tiny{[0, 11, 23, 25, 45, 47, 48, 49, 51, 67]} & 42 & 25 & 117 & 44 & 53 & 50 & 10 \\
\tiny{[0, 11, 23, 25, 45, 46, 48, 49, 51, 57]} & 38 & 29 & 126 & 28 & 70 & 40 & 10 \\
\end{tabular}
\end{table}
\section{Conclusion}
\label{concl}
The problem of decomposing a whole network into a minimal number of ego--centered subnetworks has been tackled here. In the graph--theoretic approach that we followed in order to solve this problem, the network egos were selected as the members of a minimum dominating set of the graph (assumed to be undirected here). However, since in general the solution to the graph domination problem may not be unique, we have developed an algorithm that allowed us to compute all minimum dominating sets of a graph, given (knowing) the graph domination number. Our algorithm was based on the partition of the set of vertices into three subsets, the always dominant vertices, the possible dominant vertices and the never dominant vertices. In this way, we managed to associate a dominating ego--centered decomposition to each one of the minimum dominating sets. Moreover, we introduced a number of structural measures through which we achieved comparisons and assessments of any such dominating ego--centered decompositions. In order to illustrate our methodology, we applied it to six empirical networks of various degrees of complexity in terms of graph order, size and multiplicity of minimum dominating sets.
In a subsequent work, we intend to generalize our approach to the case of directed graphs, for which the concepts of dominating sets have been already elaborated in the graph--theoretic literature. However, we need to do a number of modifications in our algorithm that will enable us to compute all the minimum dominating sets in a directed graph and assign to each one of them a corresponding graph decomposition into ego--centered subgraphs. Subsequently, after extending our methodology, we expect to be able to investigate the domination structure and ego-centered decompositions of such empirical networks as bibliometric citation networks or other cases of directed graphs extracted from social media mining (for instance, networks of RTs from Twitter data).
|
\section{Introduction}
Navigation has been a core component of the web since its inception:
users and scripts can follow hyperlinks, and can go back or forwards
through the navigation history. Users are exposed to this functionality
through following hyperlinks, and by the forward and back buttons.
Scripts have many ways of accessing session history, via the
navigation API~\cite[\S7.7]{whatwg} and the \verb|element.click()| method.
Prior formalizations of navigation history
include~\cite{HH:2006,Haydar:2004,HPS:2004,LHYT:2000,WP:2003}, which
predate and are not aligned with the \textsc{whatwg}
specification~\cite{whatwg}. The specification of the navigation API
is informal, and has complex dependencies on the rest of the HTML
specification. There is little discussion of the goals
of the API, and an unclear alignment with browser implementations.
In this paper, we present a formal model of navigation, aligned with
the HTML specification, and investigate its properties. The
starting point is that there is a total order of
\emph{documents}\footnote{%
We are eliding some of the technical details of the specification here,
in particular we are conflating a \emph{browsing context}
with the document it contains, and we are ignoring issues around
document loading and unloading, and around the current entry of the joint
session history.
}, one of which is \emph{active}, for example:
\[\begin{tikzpicture}
\node[doc](0) at (0,0){0};
\node[doc](1) at (1,0){1};
\node[doc,jshactive,fully](2) at (2,0){2};
\node[draw,dotted,fit=(0)(1)(2)] {};
\end{tikzpicture}\]
In diagrams, we use left-to-right order to indicate order,
and highlight the active document. The user can \emph{traverse}
the history which changes the active document, for example pressing
the back button:
\[\begin{tikzpicture}
\node[doc](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,0){1};
\node[doc](2) at (2,0){2};
\node[draw,dotted,fit=(0)(1)(2)] {};
\end{tikzpicture}\]
The user can also \emph{navigate}, which replaces any document
after the currently active document by a fresh active document:
\[\begin{tikzpicture}
\node[doc](0) at (0,0){0};
\node[doc](1) at (1,0){1};
\node[doc,jshactive,fully](3) at (3,0){3};
\node[draw,dotted,fit=(0)(1)(3)] {};
\end{tikzpicture}\]
Users can also traverse the history by more than one document
at a time, for example by using a pull-down menu from the back
or forwards button. This is called \emph{traversing by $\delta$},
for instance we can traverse our running example by $-2$
to get:
\[\begin{tikzpicture}
\node[doc,jshactive,fully](0) at (0,0){0};
\node[doc](1) at (1,0){1};
\node[doc](3) at (3,0){3};
\node[draw,dotted,fit=(0)(1)(3)] {};
\end{tikzpicture}\]
We formalize the notions of traversal and navigation in
\S\ref{sec:model}, and show the \emph{fundamental property of traversal}:
that traversing by $\delta$ then by $\delta'$
is the same as traversing by $\delta+\delta'$.
So far, the model is refreshingly simple, and corresponds well to
the specification and to browser implementations. Where the problems
arise is in the \emph{hierarchical} nature of documents. HTML
documents can contain \verb|iframe| elements, which
are independent documents in their own right, often
used to embed third party content such as advertisements.
We can treat each document as a tree, for example:
\[\begin{tikzpicture}
\node[doc,jshactive,fully](0) at (0,0){0};
\node[doc,active,fully](1) at (1,-1){1};
\node[doc,active,fully](2) at (2,-2){2};
\node[doc,active,fully](3) at (3,0){3};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)] {};
\node[draw,dotted,fit=(2)] {};
\node[draw,dotted,fit=(3)] {};
\draw[->](0)--(1);
\draw[->](1)--(2);
\draw[->](0)--(3);
\end{tikzpicture}\]
The problem comes from the ability of each document to
navigate separately and maintain its own session history,
but that traversal is a global operation that operates
on the \emph{joint session history}. For example
if document $2$ in the previous example navigates, the
resulting state is:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,active,fully](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc,active,fully](3) at (3,0){3};
\node[doc,jshactive,fully](4) at (4,-2){4};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)] {};
\node[draw,dotted,fit=(2)(4)] {};
\node[draw,dotted,fit=(3)] {};
\draw[->](0)--(1);
\draw[->](1)--(4);
\draw[->](0)--(3);
\end{tikzpicture}\]
and then if document $1$ navigates, the state is:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc,active,fully](3) at (3,0){3};
\node[doc,active](4) at (4,-2){4};
\node[doc,jshactive,fully](5) at (5,-1){5};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(5)] {};
\node[draw,dotted,fit=(2)(4)] {};
\node[draw,dotted,fit=(3)] {};
\draw[->](0)--(5);
\draw[->](1)--(4);
\draw[->](0)--(3);
\end{tikzpicture}\]
Note that node $4$ here is in an unusual state: it is active, but has
an inactive ancestor. The specification~\cite[\S7.7]{whatwg}
distinguishes between \emph{active} documents such as $4$, and
\emph{fully active} documents such as $0$, $3$ and $5$. Active
documents can become fully active by traversals involving their
ancestors. For example, after traversing by $-1$, document $4$ is
fully active:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc,active,fully](3) at (3,0){3};
\node[doc,active,fully](4) at (4,-2){4};
\node[doc](5) at (5,-1){5};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(5)] {};
\node[draw,dotted,fit=(2)(4)] {};
\node[draw,dotted,fit=(3)] {};
\draw[->](0)--(1);
\draw[->](1)--(4);
\draw[->](0)--(3);
\end{tikzpicture}\]
As even a simple example like this shows, the combination of features
quickly results in a complex mix of session history, ordering, and
document hierarchy, which leads to the problems:
\begin{itemize}
\item \emph{Formally} there is no simple model,
and the model provided by the specification does
not satisfy the traverse-then-traverse property.
\item \emph{Experimentally} the browsers disagree
with each other, and with the HTML specification,
about the semantics of navigation.
\end{itemize}
In this paper, we address these:
\begin{itemize}
\item \S\ref{sec:model} provides a formal model of navigation history,
which is intended to align with the specification. We show, through
a series of examples, that it does not satisfy the
fundamental property, and give patches to the model for
each example. The final model does satisfy the
fundamental property.
\item \S\ref{sec:experiments} investigates how well the patched
model aligns with existing browser implementations. We show
ways in which the browsers exhibit behaviours which are not
aligned with the specification, and discuss how our proposed
model matches these behaviours.
\end{itemize}
Finally, we propose changed wording to the specification, which
would bring it in line with our patched model.
\section{Model}
\label{sec:model}
In this section, we present our formal model of navigation history.
\S\ref{sec:preliminaries} contains definitions of concepts such as
tree and total order, and may be skipped by most readers. The model,
together with some examples, is given in \S\ref{sec:defns}. In
\S\ref{sec:properties} we define the fundamental property of
traversal, show that the model does \emph{not} satisfy
it, but can be patched to do so.
\subsection{Preliminaries}
\label{sec:preliminaries}
A \emph{directed graph} $G=(V,{\rightarrow})$ consists of:
\begin{itemize}
\item a set $V$ (the \emph{vertices}), and
\item a relation ${\rightarrow} \subseteq (V\timesV)$ (the \emph{edges}).
\end{itemize}
The \emph{transitive closure} of $\rightarrow$ is defined as $v\rightarrow^+v'$ whenever
there exists $v_0,\ldots,v_n$ such that:
\[
v=v_0\rightarrow\cdots\parentOfv_n=v'
\]
The \emph{reflexive transitive closure} of $\rightarrow$ is defined as $v\rightarrow^*v'$ whenever
$v\rightarrow^+v'$ or $v=v'$.
A \emph{forest} is a directed graph where:
\begin{itemize}
\item there is no $v$ such that $v\rightarrow^+v$ (\emph{acyclicity})
\item if $v\parentOfv'\childOfv''$ then $v=v''$ (\emph{at most one parent}).
\end{itemize}
A \emph{root vertex} of a forest is a vertex $v$ such that there is no $w\parentOfv$.
A \emph{tree} is a forest with a unique root vertex.
A \emph{preorder} is a directed graph $(V, {\le})$ such that:
\begin{itemize}
\item every $v$ has $v\lev$ (\emph{reflexivity}), and
\item if $v\lev'\lev''$ then $v\lev''$ (\emph{transitivity}).
\end{itemize}
A \emph{partial order} is a preorder such that:
\begin{itemize}
\item for every $v$ and $v'$, if $v\lev'$ and $v'\lev$ then $v=v'$
(\emph{antisymmetry}).
\end{itemize}
A \emph{total order} is a partial order such that:
\begin{itemize}
\item for every $v$ and $v'$, either $v\lev'$ or $v\gev'$ (\emph{totality}).
\end{itemize}
A \emph{equivalence} is a preorder $(V,{\sim})$ such that:
\begin{itemize}
\item if $v\simv'$ then $v'\simv$ (\emph{symmetry}).
\end{itemize}
\subsection{Definitions}
\label{sec:defns}
We can now formalize our model of navigation history, together with
the operations of navigation and traversal. This formalizes the
navigation history specification~\cite{whatwg}, and has a pleasant
diagrammatic presentation, but as we shall see in
\S\ref{sec:properties}, it has unexpected properties.
\begin{definition}[Navigation history]
A \emph{navigation history} $H=(D,A,{\rightarrow},{\le},{\sim})$ consists of:
\begin{itemize}
\item a finite set $D$ (the \emph{documents}),
\item a subset $A \subseteq D$ (the \emph{active} documents),
\item a forest $(D,{\rightarrow})$ (the \emph{child} relationship),
\item a total order $(D,{\le})$ (the \emph{chronological} order), and
\item an equivalence relation $(D,{\sim})$ (the \emph{same-session} equivalence).
\end{itemize}
such that:
\begin{itemize}
\item for every $d$ there is a unique $d'\inA$ such that $d \sim d'$,
\item for every $d \rightarrow e \sim e'$
we have $d \rightarrow e'$, and
\item for every $d \rightarrow e$, we have $d \le e$.
\hfill$\Box$
\end{itemize}
\end{definition}
We present such navigation histories ad diagrams, using
left-to-right position for chronological order, and grouping documents
in the same session. Since documents in the same session must have the
same parent, we only draw the document hierarchy for active children.
For example the diagram:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc,active,fully](3) at (3,0){3};
\node[doc,active,fully](4) at (4,-2){4};
\node[doc](5) at (5,-1){5};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(5)] {};
\node[draw,dotted,fit=(2)(4)] {};
\node[draw,dotted,fit=(3)] {};
\draw[->](0)--(1);
\draw[->](1)--(4);
\draw[->](0)--(3);
\end{tikzpicture}\]
represents:
\[\begin{array}{l}
D = \{ 0,1,2,3,4,5 \} \\[\jot]
A = \{ 0,1,3,4 \} \\[\jot]
0 \rightarrow 1 \quad 0 \rightarrow 3 \quad 0 \rightarrow 5 \quad 1 \rightarrow 2 \quad 1 \rightarrow 4 \\[\jot]
0 \le 1 \le 2 \le 3 \le 4 \le 5 \\[\jot]
1 \sim 5 \quad 2 \sim 4
\end{array}\]
In such a navigation history, we define:
\begin{itemize}
\item $d_0$ is the unique active root document,
\item $d \twoheadrightarrow e$ when $d \rightarrow e$ and $e \in A$
(the \emph{active child} relationship),
\item $F\!A = \{ d \mid d_0 \twoheadrightarrow^* d \}$
(the \emph{fully active} documents),
\item $d \lesssim e$ whenever $d \sim e$ and $d < e$,
\item the \emph{session future} of $d$ is $\{ e \mid d \lesssim e \}$,
\item the \emph{session past} of $d$ is $\{ e \mid d \gtrsim e \}$,
\item the \emph{joint session future} is $\{ e \mid \exists d \in F\!A \mathbin. d \lesssim e \}$,
\item the \emph{joint session past} is $\{ e \mid \exists d \in F\!A \mathbin. d \gtrsim e \}$,
\end{itemize}
These definitions are intended to align with the specification, for example
\cite[7.7.2]{whatwg} has the definition:
\begin{quote}
The \textbf{joint session history} of a top-level browsing context is the
union of all the session histories of all browsing contexts of all
the fully active \verb|Document| objects that share that top-level browsing
context, with all the entries that are current entries in their
respective session histories removed except for the current entry of
the joint session history.
\end{quote}
which (eliding the concept of ``current entry of the joint session history'')
corresponds to the above definitions of joint session future and past.
We now consider how to formalize operations on navigation histories.
staring with \emph{navigation}, which is triggered by following hyperlinks,
or other actions which trigger document loading.
\begin{definition}[Navigation]
Define \emph{deleting $d$ from $H$}, when $d$ is in the joint session future
to be $H'=(D',A,{\le},{\rightarrow},{\sim})$ where:
\begin{itemize}
\item $D' = D \setminus \{ e \mid d\rightarrow^* e \}$.
\end{itemize}
Define \emph{replacing $d$ by $d'$ in $H$}, where $d\inF\!A$ and
$d'\notinD$,
to be $H'=(D',A',{\le'},{\rightarrow'},{\sim'})$ where:
\begin{itemize}
\item $D' = D \cup \{d'\}$,
\item $e \in A'$ whenever
$e \in A$ and $e\ned$, or
$e=d'$,
\item $e \le' f$ whenever
$e \le f$, or $f = d'$,
\item $e \rightarrow' f$ whenever
$e \rightarrow f$, or
$e \rightarrow d$ and $f = d'$, and
\item $e \sim' f$ whenever
$e \sim f$, or
$e=f$, or
$e \sim d$ and $f = d'$, or
$d \sim f$ and $e = d'$.
\end{itemize}
Define \emph{navigating from $d$ to $d'$ in $H$}, where $d\inF\!A$ to be the result of:
\begin{itemize}
\item deleting the session future of $d$, and then
\item replacing $d$ by $d'$.
\hfill$\Box$
\end{itemize}
\end{definition}
There are two parts to navigation from $d$ to $d'$: deleting the session
future of $d$, followed by replacing $d$ by $d'$. For example,
navigating from $1$ to $6$ in:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc,active,fully](3) at (3,0){3};
\node[doc,active,fully](4) at (4,-2){4};
\node[doc](5) at (5,-1){5};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(5)] {};
\node[draw,dotted,fit=(2)(4)] {};
\node[draw,dotted,fit=(3)] {};
\draw[->](0)--(1);
\draw[->](1)--(4);
\draw[->](0)--(3);
\end{tikzpicture}\]
we first delete the session future of $1$ (which is $5$):
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc,active,fully](3) at (3,0){3};
\node[doc,active,fully](4) at (4,-2){4};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)] {};
\node[draw,dotted,fit=(2)(4)] {};
\node[draw,dotted,fit=(3)] {};
\draw[->](0)--(1);
\draw[->](1)--(4);
\draw[->](0)--(3);
\end{tikzpicture}\]
then replace $1$ by $6$:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc,active,fully](3) at (3,0){3};
\node[doc,active](4) at (4,-2){4};
\node[doc,jshactive,fully](6) at (6,-1){6};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(6)] {};
\node[draw,dotted,fit=(2)(4)] {};
\node[draw,dotted,fit=(3)] {};
\draw[->](0)--(6);
\draw[->](1)--(4);
\draw[->](0)--(3);
\end{tikzpicture}\]
We also define \emph{traversing the history}, which changes the active
documents.
\begin{definition}[Traversal]
Define \emph{traversing the history to $d$ in $H$}, where $d \in D$,
to be $H'=(D,A',{\le},{\rightarrow},{\sim})$ where:
\begin{itemize}
\item $e\inA'$ whenever $d\not\sime \in A$, or
$d=e$.
\end{itemize}
Define \emph{$H$ traverses the history by $+\delta$ to $H'$} when:
\begin{itemize}
\item the joint session future of $H$ is $d_1 < \cdots < d_\delta < \cdots$,
\item $H$ traverses the history to $d_\delta$ in $H'$
\end{itemize}
Define \emph{$H$ traverses the history by $-\delta$ to $H'$} when:
\begin{itemize}
\item the joint session past of $H$ is $d_1 > \cdots > d_\delta > \cdots$,
\item $H$ traverses the history to $d_\delta$ in $H'$
\end{itemize}
Define \emph{$H$ traverses the history by $0$ to $H'$} when $H=H'$.
\end{definition}
For example, to traverse the history by $-2$ in:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc,active,fully](3) at (3,0){3};
\node[doc,active](4) at (4,-2){4};
\node[doc,jshactive,fully](6) at (6,-1){6};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(6)] {};
\node[draw,dotted,fit=(2)(4)] {};
\node[draw,dotted,fit=(3)] {};
\draw[->](0)--(6);
\draw[->](1)--(4);
\draw[->](0)--(3);
\end{tikzpicture}\]
we find the joint session past (which is $2 > 1$)
and traverse the history to the second item (which is $1$)
to arrive at:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc,active,fully](3) at (3,0){3};
\node[doc,active,fully](4) at (4,-2){4};
\node[doc](6) at (6,-1){6};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(6)] {};
\node[draw,dotted,fit=(2)(4)] {};
\node[draw,dotted,fit=(3)] {};
\draw[->](0)--(1);
\draw[->](1)--(4);
\draw[->](0)--(3);
\end{tikzpicture}\]
These definitions are intended to formally capture the HTML
specification, for example \cite[\S7.7.2]{whatwg} includes:
\begin{quote}
To \textbf{traverse the history by a delta} $\delta$, the user agent
must append a task to this top-level browsing context's session
history traversal queue, the task consisting of running the
following steps:
\begin{enumerate}
\item If the index of the current entry of the joint session history
plus $\delta$ is less than zero or greater than or equal to the
number of items in the joint session history, then abort these
steps.
\item Let \emph{specified entry} be the entry in the joint session
history whose index is the sum of $\delta$ and the index of the
current entry of the joint session history.
\item Let \emph{specified browsing context} be the browsing context
of the specified entry.
\item If the specified browsing context's active document's unload a
document algorithm is currently running, abort these steps.
\item Queue a task that consists of running the following
substeps [\dots]
\begin{itemize}
\item[3.] Traverse the history of the specified browsing context
to the specified entry.
\end{itemize}
\end{enumerate}
\end{quote}
\subsection{Properties}
\label{sec:properties}
We now consider the fundamental property of navigation history:
\begin{definition}[Fundamental property]
\label{defn:fundamental}
$H$ satisfies the \emph{fundamental property of traversal} whenever
$H$ traverses the history by $\delta$ to $H'$
and $H'$ traverses the history by $\delta'$ to $H''$
implies $H$ traverses the history by $\delta+\delta'$ to $H''$.
\hfill$\Box$
\end{definition}
Unfortunately, navigation histories as specified do not always satisfy the fundamental property,
due to ways individual session histories are combined into the joint session history.
In this section, we give a series of counterexamples, and propose patches to
the model to address each counterexample.
\begin{counterexample}
\label{counterexample:intermediaries}
Let $H$ be:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,-2){1};
\node[doc](3) at (3,-2){3};
\node[doc,active,fully](2) at (2,-1){2};
\node[doc](4) at (4,-1){4};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(3)] {};
\node[draw,dotted,fit=(2)(4)] {};
\draw[->](0)--(1);
\draw[->](0)--(2);
\end{tikzpicture}\]
which traverses the history by $1$ to:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-2){1};
\node[doc,jshactive,fully](3) at (3,-2){3};
\node[doc,active,fully](2) at (2,-1){2};
\node[doc](4) at (4,-1){4};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(3)] {};
\node[draw,dotted,fit=(2)(4)] {};
\draw[->](0)to[out=300,in=180](3);
\draw[->](0)--(2);
\end{tikzpicture}\]
which traverses the history by $1$ to:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-2){1};
\node[doc,active,fully](3) at (3,-2){3};
\node[doc](2) at (2,-1){2};
\node[doc,jshactive,fully](4) at (4,-1){4};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(3)] {};
\node[draw,dotted,fit=(2)(4)] {};
\draw[->](0)to[out=300,in=180](3);
\draw[->](0)--(4);
\end{tikzpicture}\]
but $H$ traverses the history by $2$ to:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,active,fully](1) at (1,-2){1};
\node[doc](3) at (3,-2){3};
\node[doc](2) at (2,-1){2};
\node[doc,jshactive,fully](4) at (4,-1){4};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(3)] {};
\node[draw,dotted,fit=(2)(4)] {};
\draw[->](0)--(1);
\draw[->](0)--(4);
\end{tikzpicture}\]
\hfill$\Box$
\end{counterexample}
This counterexample is caused by the definition of `traverses the history by $\delta$' which
only traverses one document's session history. Instead, we should traverse
the history of all $\delta$ documents.
\begin{patch}[Traverse intermediaries]
Define \emph{$H$ traverses the history by $+\delta$ to $H'$} when:
\begin{itemize}
\item the joint session future of $H$ is $d_1 < \cdots < d_\delta < \cdots$,
\item there is some $H=H_0,\ldots,H_\delta=H'$, such that
\item $H_{i-1}$ traverses the history to $d_i$ in $H_i$ for each $1 \le i \le \delta$.
\end{itemize}
Define \emph{$H$ traverses the history by $-\delta$ to $H'$} when:
\begin{itemize}
\item the joint session past of $H$ is $d_1 > \cdots > d_\delta > \cdots$,
\item there is some $H=H_0,\ldots,H_\delta=H'$, such that
\item $H_{i-1}$ traverses the history to $d_i$ in $H_i$ for each $1 \le i \le \delta$.
\hfill$\Box$
\end{itemize}
\end{patch}
\begin{counterexample}
Let $H$ be:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,-1){1};
\node[doc](2) at (2,-1){2};
\node[doc,active](3) at (3,-2){3};
\node[doc](4) at (4,-2){4};
\node[doc](5) at (5,0){5};
\node[draw,dotted,fit=(0)(5)] {};
\node[draw,dotted,fit=(1)(2)] {};
\node[draw,dotted,fit=(3)(4)] {};
\draw[->](0)--(1);
\draw[->](2)--(3);
\end{tikzpicture}\]
which traverses the history by $1$ to:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc,jshactive,fully](2) at (2,-1){2};
\node[doc,active,fully](3) at (3,-2){3};
\node[doc](4) at (4,-2){4};
\node[doc](5) at (5,0){5};
\node[draw,dotted,fit=(0)(5)] {};
\node[draw,dotted,fit=(1)(2)] {};
\node[draw,dotted,fit=(3)(4)] {};
\draw[->](0)--(2);
\draw[->](2)--(3);
\end{tikzpicture}\]
which in turn traverses the history by $1$ to:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc,active,fully](2) at (2,-1){2};
\node[doc](3) at (3,-2){3};
\node[doc,jshactive,fully](4) at (4,-2){4};
\node[doc](5) at (5,0){5};
\node[draw,dotted,fit=(0)(5)] {};
\node[draw,dotted,fit=(1)(2)] {};
\node[draw,dotted,fit=(3)(4)] {};
\draw[->](0)--(2);
\draw[->](2)--(4);
\end{tikzpicture}\]
but $H$ traverses the history by $2$ to:
\[\begin{tikzpicture}
\node[doc](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc,active](2) at (2,-1){2};
\node[doc,active](3) at (3,-2){3};
\node[doc](4) at (4,-2){4};
\node[doc,jshactive,fully](5) at (5,0){5};
\node[draw,dotted,fit=(0)(5)] {};
\node[draw,dotted,fit=(1)(2)] {};
\node[draw,dotted,fit=(3)(4)] {};
\draw[->](0)--(2);
\draw[->](2)--(3);
\end{tikzpicture}\]
\hfill$\Box$
\end{counterexample}
The problem this time is that the definition of `joint session history' only includes
the fully active documents, not all active documents.
\begin{patch}[Active joint session history]
Define:
\begin{itemize}
\item the \emph{joint session future} is $\{ e \mid \exists d \in A \mathbin. d \lesssim e \}$, and
\item the \emph{joint session past} is $\{ e \mid \exists d \in A \mathbin. d \gtrsim e \}$.
\hfill$\Box$
\end{itemize}
\end{patch}
\begin{counterexample}
Let $H$ be:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc,jshactive,fully](2) at (2,-2){2};
\node[doc](3) at (3,-2){3};
\node[doc,active,fully](4) at (4,-1){4};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(4)]{};
\node[draw,dotted,fit=(2)(3)]{};
\draw[->](0)--(4);
\draw[->](0)to[out=-20,in=90](2);
\end{tikzpicture}\]
which traverses the history by $-1$ to:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,-1){1};
\node[doc,active,fully](2) at (2,-2){2};
\node[doc](3) at (3,-2){3};
\node[doc](4) at (4,-1){4};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(4)]{};
\node[draw,dotted,fit=(2)(3)]{};
\draw[->](0)--(1);
\draw[->](0)to[out=-20,in=90](2);
\end{tikzpicture}\]
which traverses the history by $1$ to:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,active,fully](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc,jshactive,fully](3) at (3,-2){3};
\node[doc](4) at (4,-1){4};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(4)]{};
\node[draw,dotted,fit=(2)(3)]{};
\draw[->](0)--(1);
\draw[->](0)to[out=-20,in=120](3);
\end{tikzpicture}\]
which is not the same as $H$.
\hfill$\Box$
\end{counterexample}
This counterexample is caused by an asymmetry in the definition
of traversal: it is defined in terms of navigating \emph{to} a document
$d$, and not navigating \emph{from} a document. We fix this
by making the definition symmetric:
\begin{patch}[Symmetric traversal]
Define \emph{$H$ traverses the history from $d'$} when there is some $d$ such that:
\begin{itemize}
\item $d\lesssimd'$,
\item for any $e\lesssimd'$ we have $e\led$, and
\item $H$ traverses the history to $d$.
\end{itemize}
Define \emph{$H$ traverses the history by $-\delta$ to $H'$} when:
\begin{itemize}
\item the joint session past and active documents of $H$ are $d_1 > \cdots > d_\delta > \cdots$,
\item there is some $H=H_0,\ldots,H_\delta=H'$, such that
\item $H_{i-1}$ traverses the history from $d_i$ in $H_i$ for each $1 \le i \le \delta$.
\hfill$\Box$
\end{itemize}
\end{patch}
For example, to traverse the history by $-1$ from:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc,jshactive,fully](2) at (2,-2){2};
\node[doc](3) at (3,-2){3};
\node[doc,active,fully](4) at (4,-1){4};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(4)]{};
\node[draw,dotted,fit=(2)(3)]{};
\draw[->](0)--(4);
\draw[->](0)to[out=-20,in=90](2);
\end{tikzpicture}\]
we find the joint session past and active documents (which is $4 > 2 > 1 > 0$)
and traverse the history from the first item (which is $4$)
which is the same as traversing the history to $1$:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,-1){1};
\node[doc,active,fully](2) at (2,-2){2};
\node[doc](3) at (3,-2){3};
\node[doc](4) at (4,-1){4};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(4)]{};
\node[draw,dotted,fit=(2)(3)]{};
\draw[->](0)--(1);
\draw[->](0)to[out=-20,in=90](2);
\end{tikzpicture}\]
\begin{counterexample}
\label{cex:not-well-formed}
Let $H$ be:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,active,fully](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc](3) at (3,-1){3};
\node[doc,jshactive,fully](4) at (4,-2){4};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(3)]{};
\node[draw,dotted,fit=(2)(4)]{};
\draw[->](0)--(1);
\draw[->](1)--(4);
\end{tikzpicture}\]
which traverses the history by $-1$ to:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,active,fully](1) at (1,-1){1};
\node[doc,jshactive,fully](2) at (2,-2){2};
\node[doc](3) at (3,-1){3};
\node[doc ](4) at (4,-2){4};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(3)]{};
\node[draw,dotted,fit=(2)(4)]{};
\draw[->](0)--(1);
\draw[->](1)--(2);
\end{tikzpicture}\]
which traverses the history by $1$ to:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc,active](2) at (2,-2){2};
\node[doc,jshactive,fully](3) at (3,-1){3};
\node[doc](4) at (4,-2){4};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(3)]{};
\node[draw,dotted,fit=(2)(4)]{};
\draw[->](0)--(3);
\draw[->](1)--(2);
\end{tikzpicture}\]
which is not the same as $H$.
\hfill$\Box$
\end{counterexample}
The problem here is not the definition of `traversing by $\delta$', but the definition
of navigation histories themselves. They allow for states such as $H$ from
Counterexample~\ref{cex:not-well-formed}, which includes the problematic documents:
\[\begin{tikzpicture}
\node[doc,active,fully](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc](3) at (3,-1){3};
\node[doc,jshactive,fully](4) at (4,-2){4};
\node[draw,dotted,fit=(1)(3)]{};
\node[draw,dotted,fit=(2)(4)]{};
\end{tikzpicture}\]
There are similar problems with documents:
\[\begin{tikzpicture}
\node[doc,active,fully](2) at (2,-1){2};
\node[doc](1) at (1,-2){1};
\node[doc](3) at (3,-1){3};
\node[doc,jshactive,fully](4) at (4,-2){4};
\node[draw,dotted,fit=(2)(3)]{};
\node[draw,dotted,fit=(1)(4)]{};
\end{tikzpicture}\]
and with documents:
\[\begin{tikzpicture}
\node[doc,active,fully](1) at (1,-1){1};
\node[doc](3) at (3,-2){3};
\node[doc](2) at (2,-1){2};
\node[doc,jshactive,fully](4) at (4,-2){4};
\node[draw,dotted,fit=(1)(2)]{};
\node[draw,dotted,fit=(3)(4)]{};
\end{tikzpicture}\]
It turns out that these are the only remaining cause of counterexamples,
and we will call examples like this not \emph{well-formed}.
\begin{definition}[Well-formed]
A navigation history is \emph{well formed} whenever
for any $a \lesssim b$ and $c \lesssim d$,
if $a \in A$ and $d \in A$ then $d \le b$.
\end{definition}
We have that traversal preserves being well-formed: if $H$ is well-formed, and $H$ traverses
by $\delta$ to $H'$, then $H'$ is well-formed. Unfortunately, this is not true for navigation,
because of the way it clears the session future.
\begin{counterexample}
\label{cex:wf-nav}
Let $H$ be the well-formed history:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,active,fully](1) at (1,-1){1};
\node[doc,jshactive,fully](2) at (2,-2){2};
\node[doc](3) at (3,-1){3};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(3)]{};
\node[draw,dotted,fit=(2)]{};
\draw[->](0)--(1);
\draw[->](1)--(2);
\end{tikzpicture}\]
which navigates from $2$ to:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,active,fully](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc](3) at (3,-1){3};
\node[doc,jshactive,fully](4) at (4,-2){4};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(3)]{};
\node[draw,dotted,fit=(2)(4)]{};
\draw[->](0)--(1);
\draw[->](1)--(4);
\end{tikzpicture}\]
which is not well-formed.
\hfill$\Box$
\end{counterexample}
Fortunately, we can patch navigation to address this, by requiring that
we clear the entire joint session future, not just the session future of the document
being navigated from.
\begin{patch}[Navigation deletes joint session future]
Define \emph{navigating from $d$ to $d'$ in $H$}, where $d\inF\!A$ to be the result of:
\begin{itemize}
\item deleting the joint session future, and then
\item replacing $d$ by $d'$.
\hfill$\Box$
\end{itemize}
\end{patch}
For example in Counterexample~\ref{cex:wf-nav}, navigation from 2 now results in the well-formed history:
\[\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,active,fully](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc,jshactive,fully](4) at (4,-2){4};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)]{};
\node[draw,dotted,fit=(2)(4)]{};
\draw[->](0)--(1);
\draw[->](1)--(4);
\end{tikzpicture}\]
With these patches, we can prove the fundamental property of traversal.
\begin{theorem}
\label{thm:fundamental}
For any well-formed navigation history $H$,
if $H$ traverses the history by $\delta$ to $H'$
and $H'$ traverses the history by $\delta'$ to $H''$
then $H$ traverses the history by $\delta+\delta'$ to $H''$.
\end{theorem}
\begin{proof}
In this paper, we give a proof sketch. The full details have been mechanically verified in Agda~\cite{AgdaProofs}.
Define:
\begin{itemize}
\item a document $d$ \emph{can go back} there is some $c \lesssim d$,
\item the \emph{back target} $b$ is the $\le$-largest active document which can go back, and
\item the \emph{forward target} $f$ is the $\le$-smallest document in the joint session future.
\end{itemize}
We then show some lemmas:
\begin{enumerate}
\item $H$ traverses by $+(\delta+1)$ to $H'$ if and only if
$H$ traverses to $f$, then traverses by $+\delta$ to $H'$.
\item $H$ traverses by $-(\delta+1)$ to $H'$ if and only if
$H$ traverses from $b$, then traverses by $-\delta$ to $H'$.
\item If $H$ is well-formed and $H$ traverses to $f$ with result $H'$,
then $f$ is the back target of $H'$, and $H'$ traverses from $f$ with result $H$.
\item If $H$ is well-formed and $H$ traverses from $b$ with result $H'$,
then $b$ is the forward target of $H'$, and $H'$ traverses to $b$ with result $H$.
\item If $H$ is well-formed and $H$ traverses to $f$ to $H'$, then $H'$ is well-formed.
\item If $H$ is well-formed and $H$ traverses from $b$ to $H'$, then $H'$ is well-formed.
\end{enumerate}
The result is then an induction on $\delta$.
\hfill$\Box$
\end{proof}
\section{Experiments}
\label{sec:experiments}
In this section, we summarize our experiments to validate the conformance of browser
implementations with respect to the \textsc{whatwg} specification, to our proposed
changes, and to each other.
We give details of how to recreate Counterexample~\ref{counterexample:intermediaries}
in detail, the other counterexamples are similar. We create an \textsc{html} page for the parent,
containing two \verb|iframe|s, both of which start at \verb|page1.html|, with a hyperlink
to \verb|page2.html|:
\begin{quote}
\raisebox{-.5\height}{
\includegraphics[width=.45\linewidth]{images/experiments/forwardback4/firefox/1.png}%
}~\raisebox{-.5\height}{
\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,-2){1};
\node[doc,active,fully](2) at (2,-1){2};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)] {};
\node[draw,dotted,fit=(2)] {};
\draw[->](0)--(1);
\draw[->](0)--(2);
\end{tikzpicture}
}
\end{quote}
Clicking on both hyperlinks loads both copies of \verb|page2.html|:
\begin{quote}
\raisebox{-.5\height}{
\includegraphics[width=.45\linewidth]{images/experiments/forwardback4/firefox/8.png}%
}~\raisebox{-.5\height}{
\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-2){1};
\node[doc,active,fully](3) at (3,-2){3};
\node[doc](2) at (2,-1){2};
\node[doc,jshactive,fully](4) at (4,-1){4};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(3)] {};
\node[draw,dotted,fit=(2)(4)] {};
\draw[->](0)to[out=300,in=180](3);
\draw[->](0)--(4);
\end{tikzpicture}
}
\end{quote}
Pressing the ``back'' button twice takes us to the initial state of Counterexample~\ref{counterexample:intermediaries}:
\begin{quote}
\raisebox{-.5\height}{
\includegraphics[width=.45\linewidth]{images/experiments/forwardback4/firefox/1.png}%
}~\raisebox{-.5\height}{
\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,jshactive,fully](1) at (1,-2){1};
\node[doc](3) at (3,-2){3};
\node[doc,active,fully](2) at (2,-1){2};
\node[doc](4) at (4,-1){4};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(3)] {};
\node[draw,dotted,fit=(2)(4)] {};
\draw[->](0)--(1);
\draw[->](0)--(2);
\end{tikzpicture}
}
\end{quote}
Now, the user can traverse the history by $+2$ (by holding down the ``forward'' button)
which results in state:
\begin{quote}
\raisebox{-.5\height}{
\includegraphics[width=.45\linewidth]{images/experiments/forwardback4/firefox/8.png}%
}~\raisebox{-.5\height}{
\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-2){1};
\node[doc,active,fully](3) at (3,-2){3};
\node[doc](2) at (2,-1){2};
\node[doc,jshactive,fully](4) at (4,-1){4};
\node[draw,dotted,fit=(0)] {};
\node[draw,dotted,fit=(1)(3)] {};
\node[draw,dotted,fit=(2)(4)] {};
\draw[->](0)to[out=300,in=180](3);
\draw[->](0)--(4);
\end{tikzpicture}
}
\end{quote}
Experimentally, this shows that Firefox is aligned with our patched model, rather than
with the unpatched model. We can set up similar experiments for the other counterexamples,
and execute them in other browsers, which gives results\footnote{%
Recall that Counterexample~4 depends on a non-well-formed navigation history,
and that the patch for it is to make such states unreachable, and so
experimentally unverifiable.
}:
\begin{center}
{\sffamily
\begin{tabular}{crrrr}
\rowcolor{black!50!blue}
&&&& \llap{\color{white} Counterexample} \\
\rowcolor{black!50!blue}
& \color{white}1 & \color{white}2 & \color{white}3 & \color{white}5 \\
\rowcolor{white!90!blue}
Firefox & P & P & P & P \\
Chrome & P & P & P & P \\
\rowcolor{white!90!blue}
Safari & P & P & P & P \\
Internet Explorer & U & U & P & P \\
\rowcolor{white!90!blue}
Servo & P & P & P & P \\
\end{tabular}
}
\quad
\begin{tabular}{ll}
\textsf{P}:& aligned with patched model \\
\textsf{U}:& aligned with unpatched model \\
\end{tabular}
\end{center}
Most browsers are compatible with the patched model rather than than
unpatched model, with the exception of Internet Explorer, which has mixed behaviour
(Edge is similar). Servo was designed from the patched model.
Moreover, performing these experiments shows some unexpected behaviours
in browser implementations. For example in Firefox, starting in state:
\begin{quote}
\raisebox{-.5\height}{
\includegraphics[width=.45\linewidth]{images/experiments/forwardback4/firefox/5.png}
}~\raisebox{-.5\height}{\rlap{
\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc](3) at (3,-1){3};
\node[doc,active,fully](4) at (4,-1){4};
\node[doc](5) at (5,-2){5};
\node[doc,jshactive,fully](6) at (6,-2){6};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(4)]{};
\node[draw,dotted,fit=(2)(6)]{};
\draw[->](0)to[out=0,in=140](4);
\draw[->](0)to[out=0,in=120](6);
\end{tikzpicture}
}}
\end{quote}
and traversing by $-4$ results in state:
\begin{quote}
\raisebox{-.5\height}{
\includegraphics[width=.45\linewidth]{images/experiments/forwardback4/firefox/6.png}
}~\raisebox{-.5\height}{\rlap{
\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc,jshactive,fully](2) at (2,-2){2};
\node[doc](3) at (3,-1){3};
\node[doc,active,fully](4) at (4,-1){4};
\node[doc](5) at (5,-2){5};
\node[doc](6) at (6,-2){6};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(4)]{};
\node[draw,dotted,fit=(2)(6)]{};
\draw[->](0)to[out=0,in=140](4);
\draw[->](0)to[out=-20,in=90](2);
\end{tikzpicture}
}}
\end{quote}
This state is unexpected, as document $4$ should have traversed to document $1$, and any state
showing \verb|page3.html| should be capable of going back.
In Safari, the use of \verb|pushState| and \verb|popState| for navigation has unexpected
results. We can use \verb|pushState| and \verb|popState| to construct state:
\begin{quote}
\raisebox{-.5\height}{
\includegraphics[width=.45\linewidth]{images/experiments/forwardback4state/safari/6.png}%
}~\raisebox{-.5\height}{\rlap{
\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc,active,fully](1) at (1,-1){1};
\node[doc,jshactive,fully](2) at (2,-2){2};
\node[doc](3) at (3,-1){3};
\node[doc](4) at (4,-1){4};
\node[doc](5) at (5,-2){5};
\node[doc](6) at (6,-2){6};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(4)]{};
\node[draw,dotted,fit=(2)(6)]{};
\draw[->](0)--(1);
\draw[->](0)to[out=-20,in=90](2);
\end{tikzpicture}
}}
\end{quote}
then traversing by $+4$ results in:
\begin{quote}
\raisebox{-.5\height}{
\includegraphics[width=.45\linewidth]{images/experiments/forwardback4state/safari/7.png}%
}~\raisebox{-.5\height}{\rlap{
\begin{tikzpicture}
\node[doc,active,fully](0) at (0,0){0};
\node[doc](1) at (1,-1){1};
\node[doc](2) at (2,-2){2};
\node[doc](3) at (3,-1){3};
\node[doc,jshactive,fully](4) at (4,-1){4};
\node[doc](5) at (5,-2){5};
\node[doc](6) at (6,-2){6};
\node[draw,dotted,fit=(0)]{};
\node[draw,dotted,fit=(1)(4)]{};
\node[draw,dotted,fit=(2)(6)]{};
\draw[->](0)to[out=0,in=140](4);
\end{tikzpicture}
}}
\end{quote}
After this traversal, we are unable to determine the active entry for one of the \verb|iframe|s
as its state is \verb|null|.
As these examples show, navigation history is difficult to implement: even major
browser implementations give unexpected behaviours when combining separate
\verb|iframe| session histories.
\section{Specification}
In this section, we discuss how the \textsc{whatwg}
specification~\cite[\S7.7.2]{whatwg} can be aligned with the model
from \S\ref{sec:model}. This is not a direct translation, due to some
of the features we elided in our model. In particular, we did not
discuss how documents are \emph{loaded} and \emph{unloaded}, which
includes downloading and allocating resources such as \textsc{html} or
\textsc{css}, and activating JavaScript content. Since
loading-then-unloading a document is wasteful, the specification
should be written to avoid loading intermediate pages when traversing
by a delta. This introduces complexity.
Our first proposed change is that the current specification is defined in terms
of the ``joint session history'' and makes use of the ``current entry
of the joint session history'', neither of which are used by our model.
We propose to remove the definition of ``joint session history''
and ``current entry of the joint session history'', and add the following:
\begin{quote}
The \textbf{session past} of a browsing context is the entries
of the session history added before the current entry
(and does not include the current entry).
The \textbf{session future} of a browsing context is the entries
of the session history added after the current entry
(and does not include the current entry).
If an entry has a next entry in the chronologically ordered session
history, it is its \textbf{successor}.
If an entry has a previous entry in the chronologically ordered session
history, it is its \textbf{predecessor}.
The \textbf{joint session past} of a top-level browsing context is the
union of all the session pasts of all browsing contexts
that share that top-level browsing context.
Entries in the joint session past are in decreasing chronological order of
the time they were added to their respective session histories.
The \textbf{joint session future} of a top-level browsing context is the
union of all the session futures of all browsing contexts
that share that top-level browsing context.
Entries in the joint session future are in increasing chronological order of
the time their predecessor were added to their respective session
histories.
\end{quote}
The second proposed change is to replace the definition of how a user agent
should``traverse the history by a delta'' by the following:
\begin{quote}
To \textbf{traverse the history by a delta} \emph{delta}, the user
agent must append a task to this top-level browsing context's session
history traversal queue, the task consisting of running the following
steps:
\begin{enumerate}
\item Define the \emph{entry sequence}
as follows:
\begin{enumerate}
\item If \emph{delta} is a positive integer $+n$, and if the length of the
joint session future is less than or equal to $n$ then let the \emph{entry sequence}
be the first $n$ entries of the joint session future.
\item If \emph{delta} is a negative integer $-n$, and if the length of the
joint session past is less than or equal to $n$ then let the \emph{entry sequence}
be the first $n$ entries of the joint session past.
\item Otherwise, abort traversing the history by a delta.
\end{enumerate}
\item A session entry is said to \textbf{become active} when
it is a member of the \emph{entry sequence}, and no
session entry after it in the \emph{entry sequence} has the same
browsing context.
\item A session entry is said to \textbf{stay active} when it it the
current entry of its browsing context, and there are no members of
the \emph{entry sequence} with the same browsing context.
\item A session entry is said to be \textbf{activating} when either
it will become active or stay active.
\textbf{Note:} the activating documents
will be active after traversal has finished.
\item A session entry is said to be \textbf{fully activating} if
is activating, and either its browsing context is a top-level
browsing context, or it has a parent browsing context
and the session entry through which it is nested is itself fully activating.
\textbf{Note:} the fully activating documents
will be fully active after traversal has finished.
\item Queue a task that consists of running the following
substeps. The relevant event loop is that of the specified
browsing context's active document. The task source for the queued
task is the history traversal task source.
\begin{enumerate}
\item For each \emph{specified entry} in the \emph{entry sequence},
run the following substeps.
\begin{enumerate}
\item Let \emph{specified browsing context} be the browsing context of the \emph{specified entry}.
\item If there is an ongoing attempt to navigate \emph{specified
browsing context} that has not yet matured (i.e. it has not
passed the point of making its \texttt{Document} the active
document), then cancel that attempt to navigate the browsing
context.
\item If the \emph{specified browsing context}'s active document
is not the same \texttt{Document} as the \texttt{Document} of
the specified entry, then run these substeps:
\begin{enumerate}
\item Prompt to unload the active document of the
\emph{specified browsing context}. If the user refused to
allow the document to be unloaded, then abort these steps.
\item Unload the active document of the \emph{specified
browsing context} with the recycle parameter set to false.
\end{enumerate}
\item If the \emph{specified entry} is activating but not fully activating,
then set the current entry of the session history of \emph{specified browsing context}
to be the \emph{specified entry}.
\textbf{Note:} in this case, the current entry of the session history should
be updated, but the document will not be fully active, so should not be loaded.
\item If the \emph{specified entry} is fully activating, then
traverse the history of the \emph{specified browsing context}
to the \emph{specified entry}.
\textbf{Note:} in this case, the document will be fully active, so should be loaded.
\end{enumerate}
\end{enumerate}
\end{enumerate}
\end{quote}
We believe that these changes bring the specification in line with our model, and
so satisfies the fundamental property of navigation.
\section{Conclusion}
We have proposed a model of web navigation compatible with the
\textsc{whatwg} specification, and investigated its ``fundamental
property'': that traversing by $\delta$ then by $\delta'$ is the same
as traversing by $\delta+\delta'$. Unfortunately, the specified model
does not satisfy this property, but we have shown that a patched model
does. Experimentally, it appears that the patched model is closer to
the behaviour of existing browser implementations.
\bibliographystyle{plain}
|
\section{Introduction}\label{s:intro}
It is well known that, in Einstein gravity, Minkowski space is stable against
arbitrarily small perturbations. The simplest way to understand this is by
noting that the formation of microscopic black holes exhibits critical
behaviour, usually referred to as Choptuik scaling \cite{Choptuik:1992jv}.
Specifically, as the amplitude of a small initial perturbation gradually
decrease, the dynamics undergoes a phase transition between black hole
formation (for large amplitudes) and dispersion to infinity (for small
amplitudes). Infinitesmally small perturbations invariably disperse. As with
all critical phenomena, the end state is highly sensitive to small variations
in the initial data near the transition region.
The properties of the phase transition depend both qualitatively and
quantitatively on the nature of the local dynamics. The transition is second
order in the absence of a fundamental scale in the problem, but can be either
first or second order when new scales are introduced. New scales can arise
either in the matter action or due to modifications to the gravitational
dynamics via, for example, the addition of higher curvature terms. It has been
shown \cite{Deppe:2012wk} that adding a Gauss-Bonnet (GB) curvature squared
term in
five and six spacetime dimensions radically affects the critical behaviour in
microscopic black hole formation.\footnote{Choptuik scaling in 5D
Einstein-Gauss-Bonnet gravity was
first considered in \cite{Golod:2012yt}.}
Naively one might expect Choptuik scaling, which was originally discovered
as a local phenomenon, to be insensitive to boundary conditions at infinity.
In particular, it was thought to be unaffected by the inclusion of a
cosmological constant. To the contrary, \cite{Bizon:2011gg} argued from
numerical
results that, in (global) anti-deSitter space (AdS), which exhibits reflecting
boundary conditions at the conformal boundary, black holes form from
arbitrarily small perturbations for massless scalar matter with
a large class of initial data.
The instability arises because subcritical matter that initially disperses
is able to return from the boundary in finite time to form a horizon near the
origin after additional gravitational focusing. Subsequent analysis by many
authors \cite{Garfinkle:2011hm,Jalmuzna:2011qw,Garfinkle:2011tc,Dias:2011ss,
Buchel:2012uh,Dias:2012tq,Buchel:2013uba,Maliborski:2013jca,Maliborski:2013ula,
Balasubramanian:2014cja,Bizon:2014bya,Maliborski:2014rma,Buchel:2014xwa,
Dimitrakopoulos:2014ada,
Craps2014,Craps2015,Balasubramanian:2015uua,Bizon:2015pfa,Fodor:2015eia,
Green:2015dsa,
Dimitrakopoulos:2015pwa,Craps:2015iia,Craps:2015jma,Craps:2015xya,Dias:2016ewl,
Dimitrakopoulos:2016tss,Deppe:2016gur}
has demonstrated the existence of ``islands of stability,'' i.e.
non-negligible regions of the initial data parameter space for which black
holes never form. In fact, some perturbative analysis suggests that
stability against horizon formation may be generic in the parameter space
of initial conditions, and it is still an open question whether
stability, instability, or both are technically generic at arbitrarily small
but finite amplitude. Other work has considered massive scalars
\cite{Buchel:2014dba,Okawa2015,Deppe:2015qsa}, a gauge field and charged
scalar \cite{Arias:2016aig}, and holographic models of confining theories
(related to the Poincar\'e patch rather than global AdS)
\cite{Craps:2014eba,Craps:2015upq}.
The stability of AdS spacetime is an interesting question in
mathematical physics in its own right, but the issue takes on particular
significance in the context of the AdS/CFT correspondence, in which gravity
in AdS spacetime is dual to a Yang-Mills theory on the conformal boundary.
Since black hole formation in the bulk AdS spacetime corresponds to
thermalization in the spatially compact boundary CFT, it is perhaps less
surprising to think that generic initial conditions lead to black holes.
Indeed, islands of stability are more surprising as they imply that
some low-energy perturbations of Yang-Mills theories on $S^3$ need not
thermalize. However, the high degree of symmetry in AdS (integrability of
the boundary theory) can lead to quasiperiodic behavior. It is clear that
AdS/CFT is a rich system with many lessons about nonperturbative dynamical
behavior.
\comment{\footnote{One caveat is that the small black holes formed at small
amplitudes in AdS suffer a rapid Gregory-Laflamme-like instability over
the extra dimensions in explicit examples of the AdS/CFT correspondence.
Nonetheless, it is not yet clear that the instability sets in substantially
before formation of the small horizon} }
The end state of gravitational collapse in AdS spacetime results from the
interplay of local (weak turbulence) and global (resonance) dynamics of the
spacetime. Quantum theory generically suggests the need for higher-derivative
terms in the gravitational and matter actions that necessarily alter the
short distance, high curvature dynamics near the final stages of gravitational
collapse, i.e. the local dynamics. In the AdS/CFT correspondence,
higher-curvature terms in the gravitational action correspond to finite $N$
and 't Hooft coupling effects in the dual theory, including differing
$a$ and $c$ central charges in 4D CFTs. Our focus in this paper is the
gravitational collapse of a massless scalar field in Einstein-Gauss-Bonnet
(EGB) gravity in the AdS$_5$/CFT$_4$ correspondence. We are motivated in
part by the possible relation of the boundary CFT to strong dynamics in
QCD.
While one expects a tower of higher-derivative couplings suppressed by powers
of the string scale, 5D EGB gravity has been an important model of
higher-curvature effects in the AdS/CFT correspondence because it is the first
example of Lovelock gravity \cite{Lanczos1938,Lovelock1971} beyond the
Einstein-Hilbert action.\footnote{There are no non-trivial Lovelock terms in
4D. The GB term is a total divergence.} The key feature of Lovelock terms in
the gravitational action
is that the equations of motion remain second order in derivatives of the
metric despite the fact that the action is higher order. Not only does this
imply that the theory is ghost-free when linearized around a flat background,
but it also makes the study of AdS stability tractable.
In 5D, only the lowest order Lovelock term (beyond Einstein), the Gauss-Bonnet
term, is relevant. As a result, it is the unique higher-curvature theory
of gravity with second-order equations of motion.
The present authors initiated a study of the stability of AdS in EGB gravity
in \cite{Deppe:2014oua}. The purpose of the current paper is to provide further
details of our calculations as well as new results with improved numerical
methods.
In particular, we present an additional discussion of structure in critical
behavior near transitions between collapse before and after reflection from
the conformal boundary, evidence for self-similar (that is, chaotic)
behaviour in the black hole formation time vs amplitude plots in transition
regions, and data hinting at the formation of naked singularities in
spacetimes with ADM mass below the algebraic bound for black hole formation.
The paper is organized as follows. In section \ref{s:egb gravity},
we review EGB gravity and
derive via Hamiltonian techniques the relevant equations of motion in
Schwarzschild coordinates. We also briefly describe our numerical methods
there. We discuss our results on the above topics in section
\ref{s:results}.
We close with a summary and prospects for future work. An appendix contains
the derivation of the equations of motion for the same system but in the
AdS analogue of flat slice coordinates for future reference.
\section{EGB Gravity and EoMs}\label{s:egb gravity}
In this section, we briefly review features of Einstein-Gauss-Bonnet gravity in
AdS$_5$ and the Hamiltonian derivation of both the mass function and scalar
equations of motion.
\subsection{Einstein-Gauss-Bonnet Gravity In AdS}
$5$-dimensional Einstein-Gauss-Bonnet gravity is a special case of
Lanczos-Lovelock gravity \cite{Lanczos1938,Lovelock1971}. The action is
\begin{equation}
I_{EGB}=\frac{1}{2\kappa_5{}^2}\int d^5x \sqrt{-g}\left({12}{\lambda}
+ {\mathcal{R}} + \frac{\lambda_3}{2} \left[\mathcal{R}^2 -
4 \mathcal{R}_{\mu\nu}\mathcal{R}^{\mu\nu} +
\mathcal{R}_{\mu\nu\rho\sigma}\mathcal{R}^{\mu\nu\rho\sigma}\right]
\right) \label{eq:action1}
\end{equation}
with $\lambda>0$ in AdS (we use $\mathcal{R}$ for the Riemann tensor and its
contractions).
The covariant equations of motion are \cite{Nozawa2008,Maeda2011}:
\begin{equation}
G_{\mu\nu} +\lambda_3 H_{\mu\nu}-\frac{\lambda}{24}g_{\mu\nu}=0\ ,
\label{eq:covariant}
\end{equation}
where
\begin{eqnarray}
G_{\mu\nu} &=&\mathcal{R}_{\mu\nu}-\frac{1}{2}g_{\mu\nu}\mathcal{R} \textnormal{ and}\\
H_{\mu\nu} &=& 2\left[\mathcal{R} \mathcal{R}_{\mu\nu} - 2\mathcal{R}_{\mu\alpha}\mathcal{R}^\alpha{}_\nu- 2
\mathcal{R}^{\alpha\beta}\mathcal{R}_{\mu\alpha\nu\beta}+ \mathcal{R}_\mu{}^{\alpha\beta\gamma}
\mathcal{R}_{\nu\alpha\beta\gamma}\right]-\frac{1}{2}g_{\mu\nu}L_{GB}\ .
\label{eq:H}
\end{eqnarray}
A key feature of spherically symmetric EGB is the existence of a generalized
mass function
\begin{equation}
\mathcal{M}= \frac{3}{2\kappa_5^2}R^{4}\left[\lambda + \frac{1}{R^2}\left(1-
|DR|^2\right) +\frac{\lambda_3}{R^4}\left(1-|DR|^2\right)^2\right]\ ,
\label{eq:M}
\end{equation}
where $R$ is the areal radius and
$|DR|^2 = g^{\mu\nu}R_{,\mu}R_{,\nu}$ \cite{Nozawa2008}.
In vacuum the mass function is constant on shell:
$\partial_\mu \mathcal{M} = 0 \, \rightarrow \, \mathcal{M} = M =$ constant.
The most general vacuum solution with compact (positive curvature) horizon,
given here in Schwarzschild-like coordinates, is
\begin{eqnarray}
ds^2 &=& -F^2(R;M)dt^2 + F^{-2}(R;M)dR^2 + R^2 d\Omega \nonumber\\
F^2(R;M) &=& 1+ \frac{R^2}{2\lambda_3}\left(1\mp
\sqrt{1+{4\lambda_3}\frac{2\kappa^2_5 M}{3R^{4}} -{4\lambda\lambda_3}} \right)
\ ,\label{eq:vacuum solution}
\end{eqnarray}
where $M$ is the on-shell value of the mass function and $F^2(R;M)=|DR|^2$ is
obtained by inverting (\ref{eq:M}). The minus sign in front of the square root
corresponds to the physical sector because it yields the
Schwarzschild-Tangherlini-AdS solution in the limit that $\lambda_3\to 0$.
Note that the GB term yields a modified effective cosmological constant
\begin{equation}
\lambda_{eff} \equiv \left[ \frac{1}{2\lambda_3}\left(1-
\sqrt{1 -{4\lambda\lambda_3}}\right)\right]^{-1}\ ,
\label{eq:effective lambda}
\end{equation}
as can be seen by taking either the $M\to 0$ or $R\to \infty$ limit in
(\ref{eq:vacuum solution}).
The vacuum solution describes a single horizon black hole spacetime. In terms
of the mass function, the horizon condition $|D R|^2_{R_H}=0$ is
\begin{equation}
\mathcal{M}(R_H)=\frac{1}{2}\left[\lambda R_H^{4} + R_H^{2}+{\lambda_3}\right]\ ,
\label{eq:massathorizon}
\end{equation}
which implies that $R_H\to 0$ as $\mathcal{M}(R_H)\to M_{crit}\equiv \lambda_3/2$
even in the dynamical context. This
suggests that it is impossible to form a black hole when the ADM mass is
less than this critical value. This is a special feature of 5D EGB,
as it depends critically on the exponent of $R_H$ in the third
term of the mass function. It is similar to the existence of a critical
mass for black holes in AdS$_3$ with Einstein gravity.
\subsection{Hamiltonian Analysis}
The total action describing the gravitational collapse of a massless scalar
field in EGB gravity is
\begin{equation}
I = I_{EGB} + I_M\ ,\ \textnormal{where}\ \
I_M = -\frac{1}{2}\int d^5x\sqrt{-g}\nabla_\mu\psi \nabla^\mu \psi\ .
\label{eq:TotalAction}
\end{equation}
The Hamiltonian analysis of spherically symmetric EGB without cosmological
constant was performed in \cite{Louko1997,Taves2012}
and extended to generic Einstein-Lanczos-Lovelock gravity in
\cite{Kunstatter2013}.
Following \cite{Louko1997,Taves2012} we use the ADM parametrization
\begin{equation}
ds^2 = - N^2dt^2 + \Lambda^2\left(dx+N_rdt\right)^2+ R^2d\Omega_{(3)}^2
\label{eq:adm metric}
\end{equation}
and integrate out the angular coordinates in (\ref{eq:TotalAction})
to obtain a two dimensional dimensionally reduced action.
With this metric, we define
\begin{equation} |DR|^2= -y^2 +\left(\frac{R'}{\Lambda}\right)^2\ \textnormal{where}\ \
y \equiv \frac{\dot{R}}{N}-\frac{N_r R'}{N} \end{equation}
for future convenience. Here and in the following, a dot is the partial
derivative with respect to $t$, and a prime is the partial with respect to
the radial coordinate $x$.
The dimensionally reduced Hamiltonian is a linear combination of the
Hamiltonian constraint $\mathcal{G}$ and diffeomorphism constraint
$\mathcal{F}$
\begin{equation}
H_{tot} = \int dx \left(N\mathcal{G} + N_r\mathcal{F}\right)\ ,
\label{eq:Htot}
\end{equation}
where we have dropped an overal factor equal to the integral over the
unit three sphere. As shown in \cite{Louko1997},
\begin{eqnarray}
\mathcal{G} &=& -\frac{6\lambda}{2\kappa_5^2} + yP_R
+ y^2\left[\Lambda R - \lambda_3 \left(\frac{R'}{\Lambda}\right)^\prime\right]
-\Lambda R\left[1-\left(\frac{R'}{\Lambda}\right)^2\right]\nonumber \\
& &+\left(\frac{R'}{\Lambda}\right)^\prime\left\{R^2 + \lambda_3\left[1-\left(\frac{R'}{\Lambda}\right)^2\right]\right\}
+ \frac{1}{2\Lambda}\left(\frac{P_\psi^2}{R^3}+ R^3 \psi'^2\right)\ ,\\
\mathcal{F} &=& R' P_R - \Lambda' P_\Lambda + \psi' P_\psi\ .
\end{eqnarray}
The momentum conjugate to $\Lambda$ is given by
\begin{equation}
P_\Lambda = - \frac{3}{2\kappa_5^2}\left[ R^2 y + 2\lambda_3 y\left(1-
\left(\frac{R^\prime}{\Lambda}\right)^2\right)+2\lambda_3 \frac{y^3}{3}\right]
\ ,\label{eq:PLambda}
\end{equation}
which determines $y=y(R,\Lambda,P_\Lambda)$ implicitly in terms of the other
gravitational phase space variables. Note that $y$ is independent of $P_R$
and that we do not require the expression for $P_R$ in the following.
By taking suitable linear combinations of the Hamiltonian and
diffeomorphism constraints, the total Hamiltonian (\ref{eq:Htot}) can,
up to boundary terms, be written \cite{Kunstatter2013} as
\begin{eqnarray}
H_{tot} &=& \int dx \left[\left(\frac{N\Lambda}{R^\prime}\right)
\left(-\mathcal{M}^\prime + \frac{R^\prime}{\Lambda^2}\rho_m-\frac{y}{\Lambda}P_\psi
\psi^\prime \right) + \left(N_r+\frac{Ny}{R^\prime}\right)\mathcal{F} \right]
\label{eq:Htotb}\ ,
\end{eqnarray}
where
\begin{eqnarray}
\rho_m &=& \frac{1}{2} \left(\frac{P_\psi^2}{R^3}+
R^3\left(\psi^\prime\right)^2\right)
\end{eqnarray}
and $\mathcal{M}$ is the mass function (\ref{eq:M}) expressed in terms of
phase space variables. It is important for the following that $P_R$ appears
only in the diffeomorphism constraint $\mathcal{F}$.
We choose as spatial coordinate $R= R(x)$ with
consistency condition $\dot{R}=0$, which requires
\begin{equation}
\frac{N_r}{N}=-\frac{y(R,\Lambda,P_\Lambda)}{R^\prime}\ .
\label{eq:Rdot}
\end{equation}
We can now set the diffeomorphism constraint, gauge fixing condition, and
consistency condition strongly to zero to obtain the partially reduced
Lagrangian
\begin{equation}
L_{sr}(t) = \int dx \left(P_\Lambda \dot{\Lambda} +P_\psi \dot{\psi}
- H_{sr}\right)\ ,
\label{eq:Lsr}
\end{equation}
where
\begin{equation}
H_{sr} = \int dx \left(\frac{N\Lambda}{R^\prime}\right)
\left(-\mathcal{M}^\prime + \frac{R^\prime}{\Lambda^2}\rho_m-\frac{y}{\Lambda}
P_\psi \psi^\prime \right)\ .
\end{equation}
The remaining coordinate freedom can be fixed in two distinct ways.
The first is to set the metric function $\Lambda=\Lambda(x)$ to be a
specific function of $x$. We outline this procedure in an appendix.
The more common choice, namely Schwarzschild-like coordinates, is used
for numerical studies in much of the current literature.
This class is obtained by the choice $y(R,\Lambda,P_\Lambda) = 0$,
which in turn implies that $P_\Lambda=0$. When $y=0$, (\ref{eq:M}) reduces to
\begin{equation}
\mathcal{M}= \frac{3}{2\kappa_n^2}R^{4}\left[\lambda + \frac{1}{R^2}
\left(1-\left(\frac{R^\prime}{\Lambda}\right)^2\right)
+\frac{\lambda_3}{R^4}\left(1-\left(\frac{R^\prime}{\Lambda}\right)^2\right)^2
\right] \label{eq:M2}
\end{equation}
The consistency condition, $\dot{y}=0$, for this gauge choice is
\begin{equation}
\left(\frac{N\Lambda}{R^\prime}\right)^\prime=-\frac{\kappa^2_5}{3}
\left(\frac{N\Lambda}{R^\prime}\right)\frac{R}{R^\prime}
\frac{\Pi^2+\Phi^2}{\left[\frac{1}{R^2}+ 2 \frac{\lambda_3}{R^4}
\left(1-\left(\frac{R^\prime}{\Lambda}\right)^2\right)\right]}\ ,
\label{eq:Nconsistency}
\end{equation}
where
\begin{equation}
\Pi\equiv \frac{P_\psi}{R^{n-2}}\ ,\qquad \Phi \equiv \psi^\prime\ .
\end{equation}
Using the Hamiltonian constraint,
\begin{equation}
\mathcal{M}^\prime = \frac{R^\prime}{\Lambda^2}\rho_m
= \frac{R^{n-2}R^\prime}{2\Lambda^2}(\Pi^2+\Phi^2)\ .
\label{eq:ham2}
\end{equation}
We note that in vacuum $(N\Lambda/ R')' =0$ and the constraint $\mathcal{M}'=0$
can be solved algebraically for $\Lambda$ and $N$ to give
(\ref{eq:vacuum solution}).
The dynamical equations can be obtained by varying the following fully
reduced Hamiltonian with respect to $\psi$ and $P_\psi$ ($\Pi$ and
$\Phi$ are not canonical variables):
\begin{equation}
H_{red}=\int dx \left(\frac{N\Lambda}{R'}\right)\left[-\mathcal{M}^\prime+
\frac{R^\prime}{2\Lambda^2}R^{n-2}\left(\Pi^2+ \Phi^2\right)\right] \ .
\label{eq:Hred}
\end{equation}
In the above, $N$ and $\Lambda$ are implicitly defined by the consistency
condition (\ref{eq:Nconsistency}) and Hamiltonian constraint (\ref{eq:ham2}),
respectively. They do not need to be varied, however, since the corresponding
variations of $H_{red}$ are proportional to the Hamiltonian constraint
and consistency condition.
The resulting evolution equations are
\begin{equation}
\dot{\Phi} = \left(\frac{N}{\Lambda}\Pi\right)^\prime\ \textnormal{and}\
\dot{\Pi} = \frac{1}{R^3}\left(\frac{ N}{\Lambda} R^{3}\Phi\right)^\prime\ .
\end{equation}
The above, along with (\ref{eq:M2}), (\ref{eq:Nconsistency}),
and (\ref{eq:ham2}) are the complete set of equations to solve.
We now put these equations into the form used in \cite{Deppe:2014oua}
by making the substitutions
\begin{equation}
\Lambda^2 = \frac{ R^\prime}{A}\ ,\ N^2 = R^\prime A e^{-2\delta}\
\Rightarrow\ \frac{N}{\Lambda}=Ae^{-\delta}\ .\end{equation}
We choose a compactified spatial coordinate $R(x) = l \tan(x/l)$
with $l= 1/\lambda_{eff}$.
The metric in terms of dimensionless coordinates $x\to x/l$, $t\to t/l$ is
\begin{equation}
ds^2 = \frac{1}{\cos^2(x)}\left(-Ae^{-2\delta} dt^2 + A^{-1}dx^2 + \sin^2(x)
d\Omega^{(n-2)}\right)\ ,
\end{equation}
while the Hamiltonian constraint in terms of the new metric functions becomes
\begin{equation}
\mathcal{M}
= \frac{3}{2\kappa_n^2}R^{4}\left[ \frac{\lambda_2}{l^2\sin^2(x)}(1-A) +
\frac{\lambda_3}{l^4\sin^4(x)}(1-A)^2\right]\ ,
\label{eq:M2b}
\end{equation}
where $\lambda_2= 1- 2\lambda_3/l^2$ and we have used the identity
\begin{equation}
\frac{1}{R^2}\left(1- {R^\prime} A\right)= -1 + \frac{1}{\sin^2(x)}(1-A)\ .
\end{equation}
We make the scalar field and its conjugate dimensionless by rescaling
$\Phi \to \kappa_5\Phi/\sqrt 3$ and $\Pi\to \kappa_5\Pi/\sqrt 3$.
Finally, we absorb $l^2$ into the mass function and $\lambda_3$ to make them
dimensionless as well.
In the end, we solve the following equations:
\begin{eqnarray}
\dot\Ph
&=& \left(Ae^{-\delta}\Pi\right)'\\
\dot\P
&=& \frac{3}{\sin(x)\cos(x)}Ae^{-\delta} \Phi + \left(Ae^{-\delta}\Phi
\right)'\\
\delta
&=& - \frac{\cos(x)\sin^3(x) (\Pi^2+\Phi^2)}{\left[\sin^2(x) -
2 \lambda_3\left(A-\cos^2(x)\right)\right]}
\label{eq:deltaPrime2}\\
\mathcal{M}
&=& \frac{A}{2}\tan^{3}(x) (\Pi^2+\Phi^2)
\label{eq:Mprime2}\\
A &=&1+ \frac{\sin^2(x) (1-2\lambda_3)}{2\lambda_3}\left[1 - \sqrt{1 +
\frac{8\mathcal{M}\lambda_3}{(1-2\lambda_3)^2\tan^{4}(x)}}\right]
\label{eq:Afinal}
\end{eqnarray}
Since nonlinear self-gravitation effects drop off sufficiently quickly at
large radius due to the dilution of energy density,
the scalar field satisfies the same asymptotic expansion as
in the linearized theory,
$\Phi = \rho^3\left(\Phi_0 + \Phi_2\rho^2+\cdots\right)$ and
$\Pi = \rho^2\left(\Pi_0 + \Pi_2 \rho^2 + \cdots\right)$, where
$\rho=\pi/2-x$. These are the boundary conditions of the
normalizable linear eigenmodes $e_k(x)$, which can be defined in terms of
Jacobi polynomials; the leading terms in these expansions correspond to
expectation values of operators in the dual field theory.\footnote{There are
also non-normalizable scalar modes (ignoring gravity) which lead to a
different asymptotic expansion and correspond to spacetime-varying operators
in the Hamiltonian of the dual theory.} At the origin, we require that
$\Pi$ be an even function of $x$ and $\Phi$ be odd for smoothness.
\subsection{Numerical Methods}\label{s:numerical methods}
We briefly outline our numerical methods and how these have changed
since our previous work \cite{Deppe:2014oua}. A detailed description is
provided in an appendix of \cite{Deppe:2015qsa}. The key improvement to
our code is that we now solve the spatial ordinary differential equations
using an adaptive fifth-order Dormand-Prince stepper. We set the desired
relative and absolute tolerances and the stepper will adjust the step size
over the spatial mesh so that the desired tolerances are met locally.
The adaptive method requires scalar field data in between grid points which we
supply using a cubic spline. We find that the stepper takes many small steps
near the origin and much larger steps further out.
\section{Results}\label{s:results}
\begin{figure}[t]
\centering
\includegraphics[width=0.70\textwidth]{GBscan.pdf}
\caption{Horizon formation times $t_H$ as a function of amplitude $\epsilon$.
Blue circles represent amplitudes which form a horizon; red triangles
represent lower limits for $t_H$ based on simulations at fixed resolution.}
\label{fig:GBscan}
\end{figure}
As in \cite{Deppe:2014oua}, we consider initial data of the form
\begin{equation}
\phi=\Phi=0\ ,\ \ \Pi = \frac{2}{\pi}\epsilon\exp\left(-\left(
\frac{2}{\pi}\frac{\tan(x)}{\sigma}\right)^2\right)\ ,\ \
\sigma=\frac{1}{16}
\end{equation}
ie, Gaussian in $\Pi$, and a GB parameter of $\lambda=0.002$.
Figure \ref{fig:GBscan} provides an overview of our results for the horizon
formation time $t_H$, which cover an amplitude range $\epsilon=27-48$.
In the figure, blue circles represent formation of a horizon, while red
triangles represent lower limits on $t_H$ for amplitudes which do not form
a horizon for $t<100$. For Einstein gravity ($\lambda=0$), $t_H$ would be
approximately piecewise constant appearing as ``steps'' with $t_H$ with
decreasing amplitude. Physically speaking, at large amplitude, gravitational
collapse can proceed immediately, but lower amplitude initial data disperses,
reflects from the conformal boundary one or more times, and finally collapses
after more gravitational focusing.
As in the earlier results of \cite{Deppe:2014oua}, gravitational collapse
in EGB gravity exhibits the same as well as additional features. First,
there is a transition from immediate collapse to collapse after one or more
reflections. There is critical behavior at these transitions, which has
been studied in some detail in \cite{Olivan:2015fmy,Santos-Olivan:2016djn}
for Einstein gravity in AdS. In the
following subsection, we study the first critical point at large amplitude,
when horizon formation stops occurring immediately, extending the analysis
of this region in \cite{Deppe:2014oua}.
Another key feature of figure \ref{fig:GBscan} is that $t_H$ appears to
demonstrate sensitivity to initial conditions in certain amplitude ranges.
That is, while there are some steps in horizon formation time where
$t_H$ remains approximately constant with $\epsilon$, $t_H$ varies wildly
in transition regions between the steps. At low enough amplitude, the
steps are apparently so narrow that they dissolve into the transition regions.
In subsection \ref{s:chaotic}, we explore in more detail whether the
transition regions exhibit chaotic behavior such as self-similarity.
Because horizon formation is apparently sensitive to initial conditions in
some regions of the amplitude, we have opted to keep all
the simulations of figure \ref{fig:GBscan} at a fixed
resolution of $2^{12}+1$ grid points, even when they begin to lose convergence
(as illustrated by a loss of conserved ADM mass). Otherwise, an increase
in resolution could act as a small shift in amplitude.
At this resolution, simulations lose up to 2.5\%
of the conserved ADM mass by $t=100$, so simulations that do not form a
horizon by this time are shown only as lower limits on $t_H$.
We have tested several amplitudes with $2^{13}+1$ grid points and found that
subcritical simulations remain subcritical while horizon formation times
in the step regions (which are stable vs change of initial conditions) have
a relative difference of $5\times 10^{-7}$.
Finally, at the lowest amplitudes shown, none of the simulations form a
horizon. As noted earlier, horizons cannot form below a critical conserved
mass $M_{crit}$ in EGB gravity.
In other words, all initial data must be stable at low amplitudes,
in apparent contrast to the case of Einstein gravity. For our choice of
initial data, the critical mass corresponds to an amplitude of approximately
$\epsilon_{crit}\sim 21.86$; figure \ref{fig:GBscan} hints that higher
amplitudes may also be dynamically stable against horizon formation.
It is also an interesting question whether evolution of initial data below
the critical amplitude is quasi-periodic or evolves toward a naked singularity.
In section \ref{s:naked}, we study the evolution for two amplitudes, one just
larger than and one just smaller than the critical amplitude, and present
evidence suggestive of naked singularity formation at finite origin time
below the critical amplitude.
\subsection{Critical Phenomenon}\label{s:critical phenomenon}
Critical phenomena in the gravitational collapse of a spherically
symmetric scalar field in Einstein gravity (for
4-dimensional asymptotically flat spacetime)
was first observed by Choptuik \cite{Choptuik:1992jv}. Choptuik found that
geometrical quantities such as the mass of the black hole obey the scaling law
\begin{align}
\label{eq:choptuikRelation}
M_{BH}\propto(p-p_\star)^\gamma
\end{align}
where $p$ is a parameter in the initial data profile, $p_\star$ is the
critical value of $p$,\footnote{For $p>p_\star$ the scalar field collapses to
a black hole, and for $p<p_\star$ the field disperses.}, and $\gamma$ is the
critical exponent. A detailed semi-analytic study by Gundlach in four
dimensions \cite{Gundlach:1996eg} found that $\gamma=0.374\pm 0.001$.
For small amplitude initial data in asymptotically AdS spacetime,
any horizon that forms will be small compared
to the AdS scale, so the critical behavior at any transition (ie, collapse
after $n$ reflections transitioning to collapse after $n+1$ reflections)
should have the same critical exponent as the asymptotically flat case.
In the case of Einstein gravity, \cite{Santos-Olivan:2016djn} confirms
the expectation, finding a critical exponent consistent with the Gundlach
value independent of the width of initial data or the number of reflections
before collapse.
The critical behavior of EGB gravity differs from Einstein gravity even
in asymptotically flat spacetime. For one, the Gauss-Bonnet term contributes
to the equations of motion only in 5 dimensions or more; in 5D Einstein
gravity, the critical exponent is $\gamma\approx 0.416$
\cite{Bland:2005kk,Taves:2011yt}.
Critical phenomena in 5D EGB gravity has been studied in
\cite{Golod:2012yt,Deppe:2012wk}, which found that the new, short distance
length scale alters the near-critical behavior
such that no black hole forms below a minimum horizon radius
\cite{Deppe:2012wk}. This is similar to the case of a massive scalar
field in asymptotically flat spacetime, which also has a dynamically
determined minimum horizon radius \cite{Brady:1997fj}.
\begin{figure}[t]
\centering
\includegraphics[width=0.50\textwidth]{GrZeroBounce}
\caption{Initial horizon radius vs amplitude for the immediate collapse
for Einstein gravity. ($\gamma\approx0.42$) }
\label{fig:GRScaling}
\end{figure}
Again, it is natural to ask which features of the critical behavior persist
or differ in asymptotically AdS$_5$ spacetime. As in 4D, we expect
critical behavior near each transition ($n$ to $n+1$ reflections) to
match that in asymptotically flat spacetime because the black holes formed
are initially much smaller than the AdS curvature scale.
In figure \ref{fig:GRScaling}, we show
$\log(r_{\rm{AH}})$ as a function of
$\log((\epsilon-\epsilon_\star)/\epsilon_\star)$, where $\epsilon_\star$
is the amplitude above which scalar field configurations collapse
immediately (the 0 to 1 reflection transition).
We find a critical exponent $\gamma\approx 0.42$ in agreement
with results in asymptotically flat spacetime.
\comment{
A natural question to ask is if the addition of a negative
cosmological constant alters this behavior, since now the matter
cannot disperse and is bounded. Near the critical solution
only some of the scalar
field collapses to form the black hole, with the rest
dispersing. However, since the field is now bounded, it can reflect
and fall into the black hole at a later time. If evolutions are run
sufficiently long
then this should be observable as a jump in the radius of the black
hole. However, in practice simulations are usually stopped
within a time of $t\approx0.7$ for the nearest to $\epsilon_\star$
evolutions. This is much less than the transit time for the scalar
field to reflect off the boundary and return to the origin, which is
approximately $\Delta t=\pi$. }
For EGB gravity, we expect a minimum horizon radius at each critical
point, as in asymptotically flat spacetime. Figure \ref{fig:GBchoptuik}
shows the scaling of the initial apparent horizon radius with
amplitude near the critical point for immediate collapse,
$\epsilon_\star\approx 45.3315$. It is initially apparent
from figure \ref{fig:RadiusGap}, which shows values of
$\epsilon$ far from $\epsilon_\star$ where the black holes form very quickly,
that there is in fact a radius gap as seen in \cite{Deppe:2014oua}.
Continuing to amplitudes with
$\epsilon-\epsilon_\star\lesssim 10^{-5}\epsilon_\star$ in figure
\ref{fig:StepsInScaling}, we observe persistence of the radius gap
$R_{min}\sim 10^{-1.9}$
along with sudden jumps, or steps, in the horizon radius.
\begin{figure}[t]
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{GbZeroBounceEarly.pdf}
\caption{Evidence for the radius gap}
\label{fig:RadiusGap}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{GbZeroBounceFull}
\caption{Step-like behavior from time dilation}
\label{fig:StepsInScaling}
\end{subfigure}
\caption{Initial apparent horizon radius vs amplitude near the critical
point for immediate collapse in 5D EGB gravity.}
\label{fig:GBchoptuik}
\end{figure}
An explanation for this behavior is apparent in animations of our
simulations. As the scalar field collapses, the initial profile fragments,
with the main portion of the mass remaining near the origin and driving
horizon formation while several pulses of mass disperse toward the
boundary. For $\epsilon-\epsilon_\star$ small, one or even two of these
subsidiary pulses can reflect from the boundary and return to the
neighborhood of the origin (possibly multiple times)
before $A(t,x)$ reaches the threshold for
approximate horizon formation. These subsidiary pulses are responsible
for the multiple local minima of $A(t,x)$ noted in \cite{Deppe:2014oua}.
Animations showing sub-pulses reflecting from the boundary once and twice are
available at \url{http://ion.uwinnipeg.ca/~afrey/AdSGB16.html}.
Although the horizon formation times $t_H$ for these amplitudes are small,
it is important to remember that $t$ is the proper time at the origin.
As it turns out, there is a significant redshift factor between this time
and the proper time outside $x\gtrsim r_{AH}$. To quantify the time
dilation factor, in figure \ref{fig:TimeDilation}
we plot the proper time of an observer at the AdS boundary, given by
\begin{align}
\label{eq:boundaryTime}
\tau=\int_0^t\exp\left[-\delta(t,x=\pi/2)\right]\,dt,
\end{align}
as a function of the proper time at the origin (in one particular
collapse). While
an insufficient amount of time apparently passes for the scalar field to
reflect off of the AdS boundary according to observers at the origin,
the relevant time is actually better approximated by the proper time at the
AdS boundary since $\delta(t,x)$ is roughly spatially constant outside the
main portion of matter. Specifically,
while only a time $\Delta t\approx 0.37$ passes, the corresponding
boundary time elapsed is $\Delta \tau\approx 4.4$, enough for the
subsidiary pulses to reflect off the boundary and interact with the forming
black hole. Interestingly, this effect should be observable in Einstein
gravity close enough to the critical amplitude (since infinite boundary
time passes before the metric function $A(t,x)\to 0$), but it appears
to be much more challenging to resolve. Some progress on the subject
has been made \cite{Allahyari:2014lta}, but a different gauge choice
and black hole excision may be needed to fully explore this
behavior. The GB term seems to enhance time dilation effects significantly.
\begin{figure}[t]
\centering
\includegraphics[width=0.70\textwidth]{ProperTime}
\caption{Plot of the proper time at the AdS boundary as function of
the proper time of an observer at the origin. The strong redshift
effects allows the dispersed matter to reflect of the AdS boundary
and interact with the black hole, causing an abrupt jump in the
radius of the black hole.}
\label{fig:TimeDilation}
\end{figure}
\subsection{Self-Similarity}\label{s:chaotic}
As we noted above, the initial horizon formation time $t_H$ exhibits a much
richer structure in EGB gravity than in pure Einstein gravity. By now it
is well-known that $t_H$ typically increases piecewise monotonically with
decreasing $\epsilon$ for massless scalar matter in Einstein gravity,
forming the ubiquitous step structure seen in many references.\footnote{There
are some widths $\sigma$ for the initial data on the ``coastlines'' of
islands of stability with non-monotonic $t_H$
\cite{Buchel:2013uba,Deppe:2015qsa}.}
In contrast, while figure \ref{fig:GBscan} also exhibits some steps in $t_H$,
the transitions from step to step exhibit a significant non-monotonic
scatter in $t_H$. For example, while initial data with $\epsilon\gtrsim 45.3$
collapses immediately and initial data with
$39.6\lesssim \epsilon\lesssim 44.0$ collapse after one reflection from the
boundary, amplitudes $44.0\lesssim\epsilon\lesssim 45.3$ vary wildly.
The appearance of smaller steps and apparent sensitivity
to initial conditions in the transition regions led \cite{Deppe:2014oua}
to speculate that the $t_H$ vs $\epsilon$ curve may have a fractal structure.
Here, we investigate the transition region $44.0\leq\epsilon\leq 45.3$
in more detail with the aim of uncovering signatures of chaotic behavior.
Since changing resolution amounts to a change in initial conditions, all
simulations discussed in this subsection are carried out at a fixed resolution
of $2^{14}+1$ grid points, following the same reasoning explained above.
\begin{figure}[t]
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{Chaos1.pdf}
\caption{$\epsilon=44.0$ to $45.3$}
\label{fig:chaosZoom1}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{Chaos2.pdf}
\caption{$\epsilon=44.3$ to $44.5$}
\label{fig:chaosZoom2}
\end{subfigure}
\begin{center}\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{Chaos3.pdf}
\caption{$\epsilon=44.40$ to $44.42$}
\label{fig:chaosZoom3}
\end{subfigure}\end{center}
\caption{$t_H$ vs amplitude in the transition region from immediate collapse
to collapse after $1$ reflection, $\epsilon\approx 44.0$ to $45.3$, over
several ranges.}
\label{fig:chaosPlots}
\end{figure}
Figure \ref{fig:chaosPlots} shows the transition region in detail for three
ranges; figure \ref{fig:chaosZoom1} shows the entire region, figure
\ref{fig:chaosZoom2} shows a small area surrounding the $t_H\approx 132$
point at $\epsilon=44.4$, and \ref{fig:chaosZoom3} shows
a smaller area to the right of that point.
Blue circles represent horizon formation, while red triangles represent
simulations that do not form a horizon within $t=305$, which can be taken
as a lower limit on $t_H$ for those amplitudes. These simulations
lose several percent of the ADM mass by that time, however, so a conservative
reader may prefer to read these as lower limits of $t_H\gtrsim 170$,
just greater than the largest values of $t_H$ for collapsing simulations.
Regardless, the plots for the three amplitude ranges show a similar structure
of rapidly varying horizon formation times with amplitude. This remains
suggestive of fractal-like, self-similar behavior, at least on the scales
shown.
To test the self-similarity of the $t_H$ vs $\epsilon$ curve quantitatively,
we use a variation on the box-counting-dimension estimate. Specifically,
we draw grid lines at each of the tick marks on figure \ref{fig:chaosZoom1}
and count the number of boxes so created that are occupied by data points.
For a first estimate, we include subcritical simulations as if they have
$t_H=300$. Data points lying on a grid line are counted as occupying the
box above or to the right as appropriate. In this case, a curve of dimension
$D$ should occupy $N=W/s^D$ boxes, where $W$ is the total horizontal range
($W=1.3$ in figure \ref{fig:chaosZoom1}) and $s$ is the length between
grid lines ($s=0.26$ in figure \ref{fig:chaosZoom1}); the box-counting
dimension is defined as $D=-\lim_{s\to 0} \ln(N/W)/\ln s$. To take an
approximate limit, we repeat the procedure for figures \ref{fig:chaosZoom2}
and \ref{fig:chaosZoom3}, keeping vertical grid lines at the tick marks shown
but scaling the vertical distance between horizontal grid lines with $s$.
We find $N=9,10,11$ and consequently $D_s=1.44,1.22,1.14$ respectively for
the three subfigures ($s=0.26,0.04,0.004$).
However, it is reasonable to argue that the
amplitudes that do not form a horizon may have different values of $t_H$
from each other (or be truly stable),
so we should not count them. If we repeat the box-counting test while
ignoring the apparently stable points and also subtracting from $W$ the
width of any boxes that contain no collapsing data points, we find
$D_s=1.35,1.15,1.13$ for the three subfigures. This provides weak but
suggestive evidence that the $t_H$ vs $\epsilon$ curve has a fractal dimension
of around 1.14, somewhat greater than unity.
Another characteristic of chaotic behavior is exponential growth of
some measure of distance between two systems with similar initial conditions,
$|\Delta|\sim \exp(\lambda t)$ for Lyapunov exponent $\lambda$.
We consider three neighboring amplitudes in figure \ref{fig:chaosZoom3},
$\epsilon_{1,2,3}=44.413,44.412,44.411$ and take as a measure of the distance
between them the difference in the upper envelope of the Ricci scalar at
the origin $\Delta_{12}=\bar\mathcal{R}_1(t)-\bar\mathcal{R}_2(t)$, etc. The bar
indicates the maximum (at the origin) over one full reflection from the
conformal boundary $\Delta t=\pi$. Figure \ref{fig:lyapRicci} shows the
$\bar\mathcal{R}$ for the three amplitudes (in green dashed, blue solid, and red dotted
curves); note that the three amplitudes lead to different values of $t_H$, so
the $\bar\mathcal{R}_{1,2}$ curves do not extend across the entire plot.
Figure \ref{fig:lyapExp} shows $\Delta_{12}$ (blue points)
and $\Delta_{23}$ (magenta squares) versus time along with best fit exponential
functions (blue solid and magenta dashed lines respectively). Both fits
are consistent with Lyapunov exponents $\lambda\sim 0.31$, or mild chaotic
behavior. The actual differences appear to have oscillation superimposed on
the exponential growth.
\begin{figure}[t]
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{lyapRicci.pdf}
\caption{Ricci scalar at origin}
\label{fig:lyapRicci}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.85\textwidth]{lyap1.pdf}
\caption{$\log|\Delta|$ vs $t$}
\label{fig:lyapExp}
\end{subfigure}
\caption{Left: Upper envelope of Ricci scalar for amplitudes
$\epsilon_1$ (green dashed), $\epsilon_2$ (blue solid), $\epsilon_3$
(red dotted). Right: Log of the absolute differences $\Delta_{12}$ (blue
points and solid line), $\Delta_{23}$ (magenta squares and dashed line)
and best fits to exponential growth.}
\label{fig:lyapunov}
\end{figure}
While \cite{Deppe:2014oua} first suggested that AdS gravitational collapse in
EGB gravity exhibits chaotic behavior, \cite{Brito:2016xvw} have also found
evidence for chaos in the gravitational collapse of two thin shells of
matter in AdS with Einstein gravity. As in our figure \ref{fig:chaosZoom1},
\cite{Brito:2016xvw} finds hints for self-similarity in the $t_H$ curve
as a function of initial conditions (in their case, the common initial radius
of the two shells of matter). In this system, energy transfers gravitationally
between the shells as they pass through each other; in the self-similar region,
the transfer back and forth leads to chaotic behavior in the horizon formation
time. The two shells are not both near the origin when the horizon forms;
instead, the horizon forms when one of the shells happens to have accumulated
a large enough density to form a horizon on its own.
In addition, \cite{Brito:2016xvw} also finds
a small but positive Lyapunov exponent for the deviations between nearby
initial conditions in the chaotic region of parameter space. Clearly this
is similar behavior, and there may be a deeper analogy between scalar collapse
in EGB gravity and the two-shell system. Specifically, at least for some
amplitudes, the GB term causes the initial scalar field pulse to break into
multiple pulses, each of which behaves as an independent shell of matter.
For shells with large radii, the GB term is negligible, so we are in fact
also studying the collapse of multiple transparent shells in Einstein gravity.
Examining one of our evolutions as an animation is instructive;
an animation of $\mathcal{M}'$ for $\epsilon=44.412$ is available at
\url{http://ion.uwinnipeg.ca/~afrey/AdSGB16.html}.
We see that the initial pulse slowly separates into two (groups of) pulses
of matter, which are approximately completely out of phase by $t\sim 15$
and each of which contains one tall, thin shell of matter.
Eventually, one of the pulses forms a horizon while the other is far away.
So, once the GB term separates the initial matter distribution into two
pulses, it seems that energy transfer between pulses may be responsible for
the chaotic behavior, as in the two-shell system. We have also examined
our simulations in the piecewise-constant regions of figure \ref{fig:GBscan}
for comparison;
while collapses that reflect from the boundary multiple times do exhibit
some pulse fragmentation, only the main pulse ever has a high, thin shell of
matter.
\subsection{Naked Singularity Formation}\label{s:naked}
In EGB gravity in AdS$_5$, horizons must contain at least a minimal mass
(even at zero radius); since the asymptotic value of the mass function is
conserved, this implies that horizons cannot form below a critical value of
the amplitude. For our initial data and choice of GB parameter,
$\epsilon_{crit}\sim 21.86$. We have already noted that we have failed to
find horizon formation for $t<100$ for any amplitude $\epsilon\leq 32$,
leaving several important questions. One, is there a dynamical mechanism
preventing gravitational collapse for small amplitudes that are nontheless
greater than $\epsilon_{crit}$? We can attempt to answer this by studying
these amplitudes with high-resolution simulations for long times.
For another, do amplitudes that do not form horizons lead to a stable,
quasiperiodic evolution, or can they form naked singularities? Is the
behavior the same or different for amplitudes above and below $\epsilon_{crit}$?
To address these questions, we have carried out simulations at
$\epsilon=20$ and $\epsilon=22$, increasing resolution as necessary to
carry the simulations to as long a time as possible.
We have been unable to find horizon formation in either case to times of
$t\sim 325,295$ and resolutions up to $n=18,19$ respectively for
$\epsilon=20,22$. The need for the high resolutions
is clear when we consider $\bar\mathcal{R}$, the upper envelope of the Ricci scalar
at the origin, which we show in figures \ref{fig:ricciA20},\ref{fig:ricciA22}.
In both cases, we find strong growth of the Ricci scalar to very large
values, eventually reaching values of order $\bar\mathcal{R}\sim 10^7$ while avoiding
formation of a horizon. From visual inspection of the simulations, the key
physics seems to be dispersal of the original matter pulse into two pulses,
which individually narrow, leading to very high curvatures, but which are
nonetheless not massive enough to form a horizon. Nonetheless, the extreme
growth of $\bar\mathcal{R}$ and pulse narrowing (which also drives the need for
increasingly higher resolution at late times) suggests the possibility that
these amplitudes will eventually form a naked singularity, rather than
behaving in a quasi-periodic fashion.
\begin{figure}[t]
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{ricciA20all.pdf}
\caption{Ricci scalar at origin}
\label{fig:ricciA20all}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{ricciA20late.pdf}
\caption{Detail of late time behavior}
\label{fig:ricciA20late}
\end{subfigure}
\caption{The upper envelope of Ricci scalar at the origin for amplitude
$\epsilon=20$. Different curves represent calculations at different
resolutions: $n=15$ (solid blue), $n=16$ (dashed green), $n=17$ (dotted orange),
$n=18$ (dot-dashed red). The right panel shows
detail for later times in the evolution.}
\label{fig:ricciA20}
\end{figure}
\begin{figure}[t]
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{ricciA22all.pdf}
\caption{Ricci scalar at origin}
\label{fig:ricciA22all}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{ricciA22late.pdf}
\caption{Detail of late time behavior}
\label{fig:ricciA22late}
\end{subfigure}
\caption{The upper envelope of Ricci scalar at the origin for amplitude
$\epsilon=22$. Different curves represent calculations at different
resolutions: $n=15$ (solid blue), $n=16$ (dashed green), $n=17$ (dotted orange),
$n=18$ (dot-dashed red), $n=19$ (thick black). The right panel shows
detail for later times in the evolution.}
\label{fig:ricciA22}
\end{figure}
As further suggestive evidence of singularity formation, we have studied the
late-time energy spectra of both evolutions. Figure \ref{fig:spectra} shows
the energy spectra (to the $j=1024$ eigenmode)
for both amplitudes at the latest time we were able to simulate in each case.
These show a slow power-law decay at large mode number, which is usually
characteristic of horizon formation.\footnote{To our knowledge the first
demonstration of a power law spectrum in gravitational collapse in AdS
was given in \cite{deOliveira:2012dt} for the Fourier modes of the Ricci
scalar at the origin near horizon formation.}
However, it is impossible for a
horizon to form for $\epsilon=20$, so the distribution of energy through
the higher modes suggests the possible development of a naked singularity.
Another point suggestive of singularity formation is that we find over 1\% of
the total energy lies in higher modes ($j>1024$) for times greater than
$t\sim 322.4$. Similarly, the $\epsilon=22$ evolution seems to be moving
rapidly toward either horizon or naked singularity formation for $t\sim 295$.
Over 1\% of the total energy is in higher modes for $t\gtrsim 287$,
and close to 3\% is in higher modes by the end of our simulation at
$t\sim 295$. This degree of energy in high eigenmodes allows an extreme
concentration of energy density near the origin, which could drive
the formation of a singularity. It is important to note that these spectra
differ from that at earlier times (see for example the supplemental
information of \cite{Deppe:2014oua}), which have an apparent exponential
cut-off for mode numbers less than $10^3$, indicating that the power has
continued to cascade into higher and higher modes as the evolution progressed.
This difference is one reason for a potentially different conclusion about
evolution at amplitudes near $\epsilon_{crit}$ in comparison to
\cite{Deppe:2014oua}.
\begin{figure}[t]
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{spectrumA20t325.pdf}
\caption{$\epsilon=20$}
\label{fig:spectrumA20}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.95\textwidth]{spectrumA22t295.pdf}
\caption{$\epsilon=22$}
\label{fig:spectrum22}
\end{subfigure}
\caption{Energy spectra as fraction of total energy per mode at the latest
times simulated. Spectra are derived from simulations with $n=18$
resolution in both cases for computational resource reasons.}
\label{fig:spectra}
\end{figure}
Assuming that a naked singularity does form at a finite time $t$, which is
the proper time at the origin, it is important to ask whether redshift
effects push the singularity formation to an infinite conformal time at the
boundary, which controls the physics in the dual CFT. Unlike the case
of horizon formation, however, time dilation effects seem to be unimportant
here; at the latest times probed by our simulations, the metric function
$\delta$ takes values of $-0.007$ and $-0.035$ at the boundary
for $\epsilon=20,22$ respectively, which are the minima over their evolution.
It is worth considering at this point what the formation of a naked singularity
at finite time would mean in the context of the AdS/CFT correspondence.
If the gravity side of the correspondence is described by the pure EGB
gravity with no additional higher-curvature terms, the translation of a
naked singularity to the dual field theory is unclear. It is tempting,
therefore, to postulate that a naked singularity is a sign of a pathology
in the theory; in fact, \cite{Camanho:2014apa} has already argued that
theories dual to pure EGB gravity in AdS suffer from acausalities. On the
other hand, if the Gauss-Bonnet term is just the first in a tower of
higher-curvature corrections, the extreme curvatures of figures
\ref{fig:ricciA20},\ref{fig:ricciA22} suggest that the additional corrections
will become significant (as seems to be the case in the self-similar
transition regions of section \ref{s:chaotic} as well). That would be
a signal that the effective gravity theory is breaking down and should be
replaced by the full string theory, and it is possible that the end state of
collapse is a gas of strings near the origin.\footnote{We thank A.~Buchel
for interesting discussions related to this point.}
Due to the very high resolutions necessary, it was not computationally
feasible to perform convergence testing for the entire simulations shown in
figures \ref{fig:ricciA20},\ref{fig:ricciA22}. However, convergence tests
for part of the $\epsilon=20,n=17$ calculations showed the expected 4th-order
convergence for the Ricci scalar at the origin, and the clear overlap of
the different resolutions in much of the figures argues for the reliability of
our results.
\section{Discussion}\label{s:conclusions}
In this manuscript, we have expanded on the analysis of black hole formation
in AdS Einstein-Gauss-Bonnet gravity first presented in
\cite{Deppe:2014oua}. In our examination of the horizon formation time $t_H$
as a function of amplitude (figure \ref{fig:GBscan}), we have considered
three particular physical phenomena in detail: critical behavior at
a transition in $t_H$, possibly chaotic behavior below the transition
amplitude, and long-time evolution at low amplitudes possibly hinting at
formation of naked singularities.
We first examined critical behavior at the transition from immediate collapse,
i.e. when a black hole forms without the matter first dispersing, to collapse
after one or more reflections from the conformal boundary
and confirmed the existence of a
dynamical radius gap of approximately $R_{min}\sim 10^{-1.9}$ (see
figure \ref{fig:GBchoptuik}). A number of questions remain about this
critical behavior. For example, do the transitions from other
step regions of figure \ref{fig:GBscan} exhibit the same radius gap as in
figure \ref{fig:GBchoptuik}, or does the value change (or even vanish)?
Is there a universal scaling law for the horizon radius for amplitudes
\comment{below the jumps in \ref{fig:StepsInScaling} or}
below any transitions in
figure \ref{fig:GBscan}, as was demonstrated in Einstein gravity in
\cite{Olivan:2015fmy,Santos-Olivan:2016djn}? We expect that the
radius gap is due to the existence of a massive critical
solution, in which case the local features of the critical behaviour
should be independent of the number of reflections from the boundary before
horizon formation. Thus, the radius gap should persist for higher numbers of
bounces. Some evidence in this direction was provided in figure 3 of
\cite{Deppe:2014oua}, which compares the scaling plots after one bounce
to those with no bounces.
Our simulations also shed light on a novel dynamical feature
of the critical behavior, first presented in \cite{Deppe:2014oua}. The
step-like behaviour in the scaling plot (figure \ref{fig:StepsInScaling})
a time dilation effect; part of the initial lump of matter disperses from
the origin rather than falling into the forming horizon. For amplitudes
close enough to the transition value, one or more of these sub-pulses have
enough time to reflect from the AdS boundary and return to the origin before
the simulation reaches our criterion for horizon formation. Although not
previously observed, this could in principle occur in any gravity theory
in AdS that exhibits critical behaviour, including Einstein
gravity. It would be interesting to check since this effect depends both
on global features of AdS and local dynamics.
The transitions from one piecewise-constant region of figure \ref{fig:GBscan}
to another also demonstrate significant scatter. We have presented
evidence that the $t_H$ vs $\epsilon$ curve is self-similar in the
region from $\epsilon\sim 44.0$ to 45.3. Furthermore, a positive Lyapunov
exponent between the evolutions of nearby amplitudes is a hallmark of
truly chaotic behavior. The chaotic behaviour appears to have as its
source the separation of the initial pulse into two (or possibly more)
pulses and the subsequent transfer of energy between them as they pass
through each other between reflections off infinity and the
origin. Here, too, questions remain: Is chaos present for any
gravitational system in which the matter forms multiple pulses, as
is the case in the two-shell system of \cite{Brito:2016xvw}? Does the
matter distribution fragment whenever the physics has a second scale
(other than the AdS scale), such as a mass for the scalar field?
Perhaps the most intriguing aspect of our analysis is the potential
evidence for naked singularity formation in the model. Below the
algebraic mass gap of $M_{crit}=\lambda_3/2$ ($\epsilon_{crit} = 21.86$
for our GB parameter and initial data),
there are only two possible end states: a
quasi-periodic steady state or naked singularity formation. While it
is impossible to provide definitive proof numerically, the observed
dramatic increase in curvature and concentration of energy into higher
modes in the absence of horizon formation suggest that the
end state will be a naked singularity for our two long evolutions near the
critical value (one just below, the other just above). In pure EGB
gravity, naked singularity formation may indicate a pathology of any dual
field theory. On the other hand, the extreme curvatures found may instead
indicate the excitation of other higher curvature terms and
string degrees of freedom, leading to the eventual production
of a gas of strings rather than the actual development of a singularity.
It is curious that horizon formation seems strongly
suppressed (takes a lot longer, or does not form at all) even above
the algebraic threshold $\epsilon_{crit}$. Our results allow us to speculate
as to the cause of this suppression, which seems to be a highly nonlinear
effect. In the case of our evolution just above threshold,
$\epsilon = 22$, the ADM mass is $M\sim 0.00101$, which is just barely
above the critical value (see figure \ref{fig:massAmp} for the conserved
mass as a function of amplitude).
Assuming the existence of a dynamical radius
gap of about $R_{min}\sim 10^{-1.9}$, equation (\ref{eq:massathorizon})
implies that the minimum amount of
mass dynamically required to form a horizon is actually close to
$\mathcal{M}(R_{min})\sim 0.00108$, so both evolutions discussed in
section \ref{s:naked} were below this dynamical limit. The
fact that we do not see a black hole form slightly above the critical
value is perhaps not a surprise. More surprising is the
apparent suppression of black hole formation for amplitudes near
$\epsilon = 32$ and below (see figure \ref{fig:GBscan}).
The mass at this amplitude
is close to 0.002, double $M_{crit}$. However, we have seen in chaotic
transition regions that the mass tends to split into at least
two pulses. In this case, the splitting can
produce smaller shells of matter that individually do not have enough energy to
form a horizon, so that horizon formation depends on the subsequent
energy transfer between pulses/shells. Indeed, this splitting occurs at
long times for $\epsilon=30.2$, as indicated in figure \ref{fig:twopulses},
so it seems that the lowest amplitudes shown in figure \ref{fig:GBscan}
are in a chaotic region. Since the two pulses appear to carry a substantial
fraction of the mass, horizon formation will require significant energy
transfer between pulses (more than in chaotic regions at higher amplitude).
It seems likely that this significant transfer is unlikely and will occur
only rarely, leading to very long horizon formation times. It is also
worth speculating if similar physics occurs for gravitational collapse in
AdS$_3$, which also has a critical black hole mass and apparent suppression
of horizon formation for amplitudes just above the critical mass
\cite{Bizon:2013xha,daSilva:2014zva}.
\begin{figure}[t]
\begin{subfigure}{0.48\textwidth}\ \\ \ \\
\includegraphics[width=0.88\textwidth]{massAmp.pdf}
\caption{ADM Mass vs amplitude}
\label{fig:massAmp}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=0.98\textwidth]{Mp_302_9399.pdf}
\caption{Mass in two pulses at low amplitude}
\label{fig:twopulses}
\end{subfigure}
\caption{Left: The conserved mass as a function of the amplitude
of initial data, including gravitational effects, as calculated from
simulations. The dashed horizontal lines indicate $M_{crit}$ (lower line)
and the dynamical limiting mass $\mathcal{M}(R_{min})$ (upper line). Right:
$M'$ vs $x$ at $t=93.99$ for the evolution of $\epsilon=30.2$ initial data.
There are two separated thin shells of matter.}
\label{fig:suppression}
\end{figure}
It is clear that gravitational collapse in EGB gravity in AdS is an
intricate system, showing first-order transitions, chaotic behavior, and
possible formation of naked singularities.
\acknowledgments
We gratefully acknowledge many helpful conversations with P.~Bizo\'n,
A.~Buchel, and S.~Green.
The work of ND is supported in part by a Natural Sciences and Engineering
Research Council of Canada PGS-D grant to ND, NSF Grants PHY-1306125
and AST-1333129 at Cornell University, and by a grant from the Sherman
Fairchild Foundation.
The work of AF and GK is supported by the Natural Sciences
and Engineering Research Council of Canada Discovery Grant program.
This research was enabled in part by support provided by WestGrid
(www.westgrid.ca) and Compute Canada Calcul Canada (www.computecanada.ca).
|
\section{Introduction}
The {\it Motzkin numbers} $M_n$ may be defined by
(cf.\ \cite[Ex.~6.37]{StanBI})
\begin{equation} \label{eq:Motzkin}
\sum_{n\ge0}M_n\,z^n=\frac {1-z-\sqrt{1-2z-3z^2}} {2z^2}.
\end{equation}
They have numerous combinatorial interpretations, see \cite[Ex.~6.38]{StanBI}.
The most basic one says that $M_n$ equals the number of
lattice paths from $(0,0)$ to $(n,0)$ consisting of steps taken from
the set
$\{(1,0),(1,1),(1,-1)\}$ never running below the $x$-axis.
Deutsch and Sagan \cite[Theorem~3.1]{DeSaAA} determined the parity of $M_n$.
Furthermore, they conjectured \cite[Conj.~5.5]{DeSaAA} (a conjecture
which is in part also due to Amdeberhan)
that $M_n$ is divisible by $4$ if, and only if,
$$
n=(4i + 1)4^{j+1} - 1\quad \text{or}\quad n = (4i + 3)4^{j+1} - 2,
$$
for some non-negative integers $i$ and $j$.
This conjecture prompted Eu, Liu and Yeh \cite{EuLYAB} to embark
on a more systematic study of Motzkin numbers modulo~$4$
and~$8$. They approached this problem by carefully analysing
a binomial sum representation for $M_n$ modulo~$4$ and ~$8$.
They succeeded in establishing the above conjecture
of Amdeberhan, Deutsch and Sagan. Moreover, they were able to characterise
the even congruence classes of $M_n$ modulo~$8$; see
\cite[Theorem~5.5]{EuLYAB}
and Corollary~\ref{thm:EuLYAB} at the end of the present paper.
One outcome of this characterisation is that $M_n$ is never divisible
by~$8$, thus proving another conjecture of Deutsch and Sagan
\cite[Conj.~5.5]{DeSaAA}.
As part of a general theory of diagonals of rational functions,
Rowland and Yassawi \cite[Sec.~3.2]{RoYaAA} are able to approach the problem
of determination of Motzkin numbers modulo powers of~$2$ from a
completely different point of view. Their general theory (and
algorithms) produce automata which output the congruence class of
diagonal coefficients of a given rational function modulo a
given power of a prime number.
Since the Motzkin numbers $M_n$
can be represented as the diagonal coefficients of a certain
rational function, this theory applies, and consequently,
for any prime power~$p^\alpha$, an automaton can be produced which outputs
the congruence class of $M_n$ modulo~$p^\alpha$.
Rowland and Yassawi present the automaton for the determination of
$M_n$ modulo~$8$ in \cite[Fig.~4]{RoYaAA}.
The main purpose of this paper is to present yet another approach
to the determination of Motzkin numbers $M_n$ modulo powers of~$2$,
namely by the use of generating functions. Our approach is in line
with our earlier studies \cite{KaKMAA,KrMuAE,KrMuAL}, in which we
developed a generating function method for determining the
mod-$p^\alpha$ behaviour of sequences whose generating functions
satisfy algebraic differential equations. Application of this
method to a concrete class of problems starts with the choice
of a (usually transcendental) basic series.
The goal then is
to show that the generating function of the sequence that we are
interested in can be expressed as a polynomial in this basic series
when the terms of the sequence are reduced modulo~$p^\alpha$.
Once this is achieved, one tries to extract a characterisation of the
congruence behaviour of the sequence under consideration modulo~$p^\alpha$
from this polynomial representation of the generating function.
However, the details --- such as the choice of basic series, or the
extraction of coefficients from its powers --- have to be adjusted
for each individual class of applications.
Here, we choose the series $\Om(z)$ defined by
\begin{equation} \label{eq:Omdef}
\Om(z)=\sum _{e\ge0} ^{}\sum _{f\ge0} ^{}
z^{4^e(2f+1)}
=\sum _{e\ge0} ^{} \frac {z^{4^e}} {1-z^{2\cdot 4^e}}
\end{equation}
as basic series.
In plain words, the series $\Om(z)$ is the sum of all monomials $z^m$,
where $m$ runs through all positive integers with an even number of
zeroes to the far right of their binary representation.
Our first main result, Theorem~\ref{thm:Motzkin}, states that the
generating function $\sum_{n\ge0}M_n\,z^n$
for the Motzkin numbers, when coefficients are reduced modulo a given
power of~$2$, can be expressed as a polynomial in $\Om(z^4)$
with coefficients that are Laurent polynomials in $z$ and $1-z$.
Theorem~\ref{thm:Motzkin} may be
implemented, so that these polynomial expressions can be found
automatically modulo any given power of $2$. Thus, our result in
principle opens up the possibility of deriving congruences modulo
arbitrary $2$-powers for the sequence of Motzkin numbers, provided one
is able to systematically extract coefficients from powers of the
series $\Om(z^4)$. However, for exponents greater than $3$, this task
turns out to be difficult and rather involved, though possible in
principle. For this reason, we proceed differently. We use our
algorithm to obtain a congruence modulo $16$ exhibiting the generating
function of the Motzkin numbers as a polynomial of degree $7$ in
$\Om(z^4)$, and then drastically simplify the resulting expression
modulo $2^3=8$ at the expense of introducing a certain
``error-series'' $E(z)$. The advantage is that, in this step, the
degree of $\Om(z^4)$ drops from $7$ to $1$. Extracting
coefficients of powers of~$z$ from the resulting expression, we are then able
to go beyond the result \cite[Theorem~5.5]{EuLYAB}
of Eu, Liu and Yeh by providing a formula for the congruence class
of $M_n$ also in the case that $M_n$ is odd; see
Theorem~\ref{thm:M8}. As demonstrated in Corollary~\ref{cor:Mn1},
this result may conveniently be used to precisely characterise those~$n$
for which $M_n$ lies in a specified congruence class modulo~$8$.
In a sense, our generating function method and the automaton method
of \cite{RoYaAA} are complementary to each other. It is through
our method that we were led to the explicit characterisation
of the congruence behaviour of Motzkin numbers modulo~8 in
Theorem~\ref{thm:M8} and Corollary~\ref{cor:Mn1}, and to
the explicit characterisations of the congruence behaviour
modulo~8 of Motzkin prefix numbers, of Riordan numbers,
of hex tree numbers, and of central trinomial coefficients
in Theorems~\ref{thm:MP8}--\ref{thm:T8}, together with proofs
in each case.
It seems hard to {\it extract\/} these results from the automata
that one obtains by the theory in \cite{RoYaAA}. On the other hand,
as pointed out by a referee, once these results are {\it known},
they can be {\it verified\/} using the theory of automata on the basis of the
automata that one finds by means of the automaton theory of \cite{RoYaAA},
since the functions $s_2(n)$, $s_2^2(n)$, $e_2(n)$, $n_4$, $n_{K+1}$
that appear in these results are $2$-regular sequences.
\medskip
Our paper is organised as follows.
In the next section, we derive some properties of our basic series
$\Om(z)$ defined in \eqref{eq:Omdef} that we shall need in the
sequel. Section~\ref{sec:method} outlines our generating function
approach based on the series $\Om(z)$.
Subsequently, in Section~\ref{sec:Motzkin}, we apply this
method to the generating function $M(z)=\sum_{n\ge0}M_n\,z^n$
for Motzkin numbers. This section contains our first main result,
Theorem~\ref{thm:Motzkin}, stating that $M(z)$, when coefficients
are reduced modulo a given power of~$2$, can be expressed as a
polynomial in $\Om(z)$. In Section~\ref{sec:M8}, we specialise
this result to the modulus $2^4=16$, see \eqref{eq:M16}.
This expression is then reduced modulo~$8$ and further simplified,
at the cost of introducing an ``error series" $E(z)$, defined
in \eqref{eq:Edef}. The final expression for the mod-$8$
reduction of $M(z)$ is presented in \eqref{eq:MM}.
The remaining task is to extract the coefficient of $z^n$ from
the right-hand side of \eqref{eq:MM}. In order to accomplish this,
we provide several auxiliary lemmas in Section~\ref{sec:aux}.
In Section~\ref{sec:Mn8}, making use of these lemmas, we are then able
to provide explicit formulae
for the congruence class of $M_n$ modulo~$8$ in terms of the
binary digits of~$n$. This leads to our second main result which
is presented in Theorem~\ref{thm:M8}. In Corollary~\ref{cor:Mn1},
we demonstrate by means of an example how to characterise those~$n$
for which $M_n$ lies in a given congruence class modulo~$8$.
Our final section then collects together results analogous to
Theorem~\ref{thm:M8} for sequences related to Motzkin numbers,
namely for the Motzkin prefix numbers, Riordan numbers, hex tree
numbers, and the central trinomial coefficients; see
Theorems~\ref{thm:MP8}--\ref{thm:T8}. These results generalise
earlier ones of Deutsch and Sagan \cite{DeSaAA} from modulus~$2$ to
modulus~$8$.
\begin{note}
This paper is accompanied by a {\sl Mathematica} file
and a {\sl Mathematica} notebook so that an interested reader is
able to redo (most of) the computations that are presented in
this article. File and notebook are available at the article's
website
{\tt http://www.mat.univie.ac.at/\lower0.5ex\hbox{\~{}}kratt/artikel/2motzkin.html}.
\end{note}
\section{The power series $\Om(z)$}
\label{sec:Om}
Here we consider the formal power series $\Om(z)$ defined in
\eqref{eq:Omdef}.
This series is the principal character in the method for determining
congruences of recursive sequences modulo $2$-powers
that we describe in Section~\ref{sec:method}.
This series is transcendental over
$\mathbb Z[z]$.\footnote{The quickest argument uses coefficient
asymptotics for algebraic functions. The coefficients of $\Om(z)$
do not approach any limit. Since these coefficients are $0$ or $1$,
there is no asymptotic formula for them.
On the other hand, by \cite[Theorem~VII.8]{FlSeAA}, the coefficients
of algebraic power series {\it do} have asymptotic formulae (whose form
can even be described very specifically). Consequently, the series
$\Om(z)$ cannot be algebraic.} However, if the
coefficients of $\Om(z)$ are considered modulo a $2$-power $2^\ga$,
then $\Om(z)$ obeys a polynomial relation with coefficients that are
Laurent polynomials in $z$ and $1-z$. This is shown in
Proposition~\ref{prop:minpol} below. In the second part of this
section, we show that the derivative of $\Om(z)$, when coefficients
are reduced modulo a given $2$-power, can be expressed as a
polynomial in $z$ and $(1-z)^{-1}$.
This is one of the crucial facts which
make the method described in Section~\ref{sec:method} work.
\medskip
Here and in the sequel,
given integral power series (or Laurent series) $f(z)$ and $g(z)$,
we write
$$f(z)=g(z)~\text {modulo}~2^\ga$$
to mean that the coefficients
of $z^i$ in $f(z)$ and $g(z)$ agree modulo~$2^\ga$ for all $i$.
\begin{proposition} \label{prop:minpol}
We have
\begin{equation} \label{eq:Om^2A}
\Om^2(z)+\Om(z)-\frac {z} {1-z}=0\quad
\text {\em modulo }2.
\end{equation}
\end{proposition}
\begin{proof}
It is straightforward to see that
$$
\Om^2(z)=\sum _{e\ge0} ^{}\sum _{f\ge0} ^{}
z^{2\cdot 4^e(2f+1)}\quad
\text{modulo }2.
$$
In plain words, this is the sum of all monomials $z^m$,
where $m$ runs through all positive integers with an odd number of
zeroes to the far right of their binary representation. Consequently,
we have
$$
\Om^2(z)+\Om(z)=\sum_{m\ge1}z^m=\frac {z} {1-z}\quad
\text{modulo }2.
$$
This is equivalent to the claim.
\end{proof}
\begin{remarknu}
Clearly, Proposition~\ref{prop:minpol} implies that $\Om(z)$
is algebraic modulo {\it any} power of~$2$. In particular, we have
\begin{equation} \label{eq:N}
\left(\Om^2(z)+\Om(z)-\frac {z} {1-z}\right)^N=0\quad
\text {modulo }2^N,
\end{equation}
a relation that we shall use later on.
\end{remarknu}
We now turn our attention to the derivative of the basic series
$\Om(z)$.
By straightforward differentiation, we obtain
\begin{equation} \label{eq:Om'}
\Om'(z)
= \sum_{e\ge0}^{}4 ^e \left(\frac {
z^{4^e-1}} {1-z^{2\cdot 4^e}}
+
\frac {2
z^{3\cdot 4^e-1}} {(1-z^{2\cdot 4^e})^2}\right).
\end{equation}
We would like to prove that $\Om'(z)$, when considered modulo a given
power of~$2$, can be expressed as a polynomial in $z$ and $(1-z)^{-1}$
with integer coefficients. The above relation does not quite achieve this,
since the denominators on the
right-hand side are not powers of $1-z$. However, the following is
true.
\begin{lemma} \label{lem:1+z}
For all non-negative integers $j$ and positive integers $\al$ and
$\be,$ we have
$$
\frac {1} {(1+z^{2^j})^\al}=\Pol^+_{j,\al,\be}\left(z,(1-z)^{-1}\right)
\quad \text {\em modulo }2^\be,
$$
where $\Pol^+_{j,\al,\be}\left(z,(1-z)^{-1}\right)$ is a polynomial in $z$ and
$(1-z)^{-1}$ with integer coefficients.
\end{lemma}
\begin{proof} We perform an induction with respect to $j+\be$.
For $j=0$ and arbitrary $\beta$, we have
$$
\frac {1} {(1+z)^\al}=
\frac {1} {(1-z+2z)^\al}=
(1-z)^{-\al}\sum_{k\ge0}\binom {\al+k-1}k 2^k\frac {z^k} {(1-z)^k}.
$$
When this is considered modulo~$2^\be$ for some fixed~$\be$,
the sum on the right-hand side becomes finite and is indeed
a polynomial in $z$ and $(1-z)^{-1}$.
For the induction step with $j\ge1$, we write
\begin{align*}
\frac {1} {(1+z^{2^j})^\al}&=
\frac {1} {(1+z^{2^{j-1}})^{2\al}}
+\frac {1} {(1+z^{2^j})^\al}
-\frac {1} {(1+z^{2^{j-1}})^{2\al}}\\
&=
\frac {1} {(1+z^{2^{j-1}})^{2\al}}
+\frac {1} {(1+z^{2^{j-1}})^{2\al}(1+z^{2^j})^\al}
\left((1+z^{2^{j-1}})^{2\al}-(1+z^{2^j})^\al\right)\\
&=
\frac {1} {(1+z^{2^{j-1}})^{2\al}}
+\frac {1} {(1+z^{2^{j-1}})^{2\al}(1+z^{2^j})^\al}\\
&\kern4cm
\times
\left((1+2z^{2^{j-1}}+z^{2^{j}})^{\al}
-(1+z^{2^j})^\al\right)\\
&=
\frac {1} {(1+z^{2^{j-1}})^{2\al}}
+\frac {1} {(1+z^{2^{j-1}})^{2\al}(1+z^{2^j})^\al}\\
&\kern4cm
\times
\sum _{\ell=1} ^{\al}\binom\al\ell
2^\ell\left(z^{2^{j-1}}\right)^{\ell}
\left(1+z^{2^j}\right)^{\al-\ell}\\
&=
\frac {1} {(1+z^{2^{j-1}})^{2\al}}
+
\sum _{\ell=1} ^{\al}\binom\al\ell
\frac {2^\ell z^{\ell \cdot 2^{j-1}}}
{(1+z^{2^{j-1}})^{2\al}(1+z^{2^j})^\ell}.
\end{align*}
Now the induction hypothesis can be applied, and yields
\begin{multline*}
\frac {1} {(1+z^{2^j})^\al}=
\Pol^+_{j-1,2\al,\be}\left(z,(1-z)^{-1}\right)\\
+
\sum _{\ell=1} ^{\al}\binom\al\ell
{2^\ell z^{\ell \cdot 2^{j-1}}}
(\Pol^+)^{2\al}_{j-1,2\al,\be-\ell}\left(z,(1-z)^{-1}\right)
(\Pol^+)^{\ell}_{j,\ell,\be-\ell}\left(z,(1-z)^{-1}\right)\\
\text {modulo }2^\be.
\end{multline*}
This completes the induction.
\end{proof}
\begin{lemma} \label{lem:1-z}
For all non-negative integers $j$ and positive integers $\al$ and
$\be,$ we have
$$
\frac {1} {(1-z^{2^j})^\al}=\Pol_{j,\al,\be}\left(z,(1-z)^{-1}\right)
\quad \text {\em modulo }2^\be,
$$
where $\Pol_{j,\al,\be}\left(z,(1-z)^{-1}\right)$ is a polynomial in $z$ and
$(1-z)^{-1}$ with integer coefficients.
\end{lemma}
\begin{proof}
We perform an induction on~$j$. Clearly, there is nothing to prove
for $j=0$.
For the induction step, we write
$$
\frac {1} {(1-z^{2^j})^\alpha}
=\frac {1} {(1+z^{2^{j-1}})^\alpha} \cdot \frac1 {(1-z^{2^{j-1}})^\alpha} .
$$
By Lemma~\ref{lem:1+z}, we know that the first factor on the right-hand
is a polynomial in $z$ and $(1-z)^{-1}$, while the induction
hypothesis guarantees that the second factor is such a polynomial as well.
\end{proof}
Identity~\eqref{eq:Om'} and Lemma~\ref{lem:1-z} together imply
our claim concerning $\Om'(z)$ (it should be observed that, modulo a
given $2$-power $2^\be$, the sum on the right-hand side of \eqref{eq:Om'}
is finite).
\begin{corollary} \label{cor:Om'}
The series $\Om'(z),$ when coefficients are reduced modulo
$2^\be$ for some fixed~$\be,$
can be expressed as a polynomial in $z$ and $(1-z)^{-1}$
with integer coefficients.
\end{corollary}
\section{The method}
\label{sec:method}
We consider a (formal) differential equation of the form
\begin{equation} \label{eq:diffeq}
\mathcal P(z;F(z),F'(z),F''(z),\dots,F^{(s)}(z))=0,
\end{equation}
where $\mathcal P$ is a polynomial with integer coefficients, which has a
unique power series solution $F(z)$ with integer coefficients when
the equation is considered modulo any fixed power of~$2$.
(In the literature, series obeying a polynomial relation of the
form \eqref{eq:diffeq} are known as {\it differentially algebraic}
series; see, for instance, \cite{BeReAA}.)
In this
situation, we propose the following algorithmic approach to
determining $F(z)$ modulo a $2$-power $2^{2^\al}$,
for some positive integer $\al$.
Let us fix a positive integer $\ga$.
We make the Ansatz
\begin{equation} \label{eq:Ansatz}
F(z)=\sum _{i=0} ^{2^{\al+1}-1}a_i(z)\Om^i( z^\ga)\quad
\text {modulo }2^{2^\al},
\end{equation}
where $\Om(z)$ is given by \eqref{eq:Omdef},
and where the $a_i(z)$'s
are (at this point) undetermined elements of
$\Z[z,z^{-1},(1- z^\ga)^{-1}]$.
Now we substitute \eqref{eq:Ansatz} into \eqref{eq:diffeq}, and
we shall gradually determine approximations $a_{i,\be}(z)$ to $a_i(z)$ such that
\eqref{eq:diffeq} holds modulo~$2^\be$, for $\be=1,2,\dots,2^\al$.
To initiate the procedure, we consider the differential equation
\eqref{eq:diffeq} modulo~$2$, with
\begin{equation} \label{eq:Ansatz1}
F(z)=\sum _{i=0} ^{2^{\al+1}-1}a_{i,1}(z)\Om^i( z^\ga)\quad \text {modulo
}2.
\end{equation}
By Corollary~\ref{cor:Om'}, we know that $\Om'(z^\ga)$, when
considered modulo~$2$, is in $\Z[z,(1-z^\ga)^{-1}]$.
Consequently, we see that
the left-hand side of \eqref{eq:diffeq} is a polynomial in
$\Om( z^\ga)$ with coefficients in
$\Z[z,z^{-1},(1- z^\ga)^{-1}]$.
We reduce powers $\Om^k( z^\ga)$ with $k\ge2^{\al+1}$ using
the relation
\begin{equation} \label{eq:OmRel}
\left(\Om^{2}( z^\ga)+\Om( z^\ga)
-\frac {z^\ga} {1- z^\ga}\right)^{2^\al}=0\quad \text {modulo
}2^{2^\al},
\end{equation}
which is the special case $N=2^\al$ of \eqref{eq:N}.
Since, at this point, we are only interested in finding a solution to
\eqref{eq:diffeq} modulo~$2$, Relation~\eqref{eq:OmRel} simplifies to
\begin{equation} \label{eq:OmRel2}
\Om^{2^{\al+1}}( z^\ga)+\Om^{2^{\al}}( z^\ga)
-\frac {z^{\ga\cdot 2^\al}} {(1- z^\ga)^{2^{\al}}}=0\quad \text {modulo
}2.
\end{equation}
Now we compare coefficients of powers $\Om^k( z^\ga)$,
$k=0,1,\dots,2^{\al+1}-1$. This yields a
system of $2^{\al+1}$ (differential) equations (modulo~$2$)
for unknown functions $a_{i,1}(z)$ in
$\Z[z,z^{-1},(1- z^\ga)^{-1}]$, for $i=0,1,\dots,
2^{\al+1}-1$,
which may or may not have a solution.
Provided we have already found
functions $a_{i,\be}(z)$ in
$\Z[z,z^{-1},(1- z^\ga)^{-1}]$, $i=0,1,\dots,\break
2^{\al+1}-1$,
for some $\be$ with $1\le \be\le 2^{\al}-1$, such that
\begin{equation} \label{eq:Ansatz2}
F(z)=\sum _{i=0} ^{2^{\al+1}-1}a_{i,\be}(z)\Om^i( z^\ga)
\end{equation}
solves \eqref{eq:diffeq} modulo~$2^\be$, we put
\begin{equation} \label{eq:Ansatz2a}
a_{i,\be+1}(z):=a_{i,\be}(z)+2^{\be}b_{i,\be+1}(z),\quad
i=0,1,\dots,2^{\al+1}-1,
\end{equation}
where the $b_{i,\be+1}(z)$'s are (at this point) undetermined
elements in $\Z[z,z^{-1},(1- z^\ga)^{-1}]$. Next we substitute
\begin{equation} \label{eq:Ansatz2b}
F(z)=\sum _{i=0} ^{2^{\al+1}-1}a_{i,\be+1}(z)\Om^i( z^\ga)
\end{equation}
in \eqref{eq:diffeq} and consider the result modulo~$2^{\be+1}$.
Using Corollary~\ref{cor:Om'} and Lemma~\ref{lem:1-z},
we see that derivatives of $\Om(z^\ga)$, when considered
modulo~$2^{\be+1}$, can be expressed as
polynomials in $z$ and $(1-z^\ga)^{-1}$. Consequently, we may
expand the left-hand side of \eqref{eq:diffeq}
as a polynomial in $\Om( z^\ga)$ with
coefficients in $\Z[z,z^{-1},(1- z^\ga)^{-1}]$. Subsequently,
we apply again the reduction
using Relation~\eqref{eq:OmRel}. By comparing coefficients of powers
$\Om^k( z^\ga)$, $k=0,1,\dots,2^{\al+1}-1$,
we obtain a
system of $2^{\al+1}$ (differential) equations (modulo~$2^{\be+1}$)
for the unknown functions $b_{i,\be+1}(z)$,
$i=0,1,\dots,2^{\al+1}-1$,
which may or may not have a solution. If we manage to push this
procedure through until $\be=2^\al-1$, then,
setting $a_i(z)=a_{i,2^\al}(z)$, $i=0,1,\dots,2^{\al+1}-1$,
the series $F(z)$ as given in \eqref{eq:Ansatz} is a solution to
\eqref{eq:diffeq} modulo~$2^{2^\al}$, as required.
We should point out that, in order for the method to be meaningful,
it is essential to assume that the differential equation \eqref{eq:diffeq},
when considered modulo an arbitrary $2$-power $2^e$,
determines the solution $F(z)$ {\it uniquely} modulo~$2^e$ (which we do).
Otherwise, our method could produce several different ``solutions,"
and it might be difficult to decide which of them actually
corresponds to the series $F(z)$ we are interested in.
\section{Motzkin numbers}
\label{sec:Motzkin}
Let $M_n$ be the $n$-th {\it Motzkin number}, that is, the number of
lattice paths from $(0,0)$ to $(n,0)$ consisting of steps taken from
the set
$\{(1,0),(1,1),(1,-1)\}$ never running below the $x$-axis.
It is well-known (cf.\ \cite[Ex.~6.37]{StanBI})
that the generating function $M(z)=\sum_{n\ge0}M_n\,z^n$ is given by
\eqref{eq:Motzkin}
and hence satisfies the functional equation
\begin{equation} \label{eq:MotzkinEF}
z^2M^2(z)+(z-1)M(z)+1=0.
\end{equation}
\begin{theorem} \label{thm:Motzkin}
Let $\Om(z)$ be given by \eqref{eq:Omdef}{\em ,}
and let $\al$ be some positive integer.
Then the generating function $M(z)=\sum_{n\ge0}M_n\,z^n$
for the Motzkin numbers,
reduced modulo~$2^{2^\al},$
can be expressed as a polynomial in $\Om(z^4)$ of the form
$$
M(z)=\sum _{i=0} ^{2\cdot 2^{\al}-1}a_{i}(z)\Om^{i}(z^4)\quad
\text {\em modulo }2^{2^\al},
$$
where the coefficients $a_{i}(z),$ $i=0,1,\dots,2^{\al+1}-1,$
are Laurent polynomials in $z$ and $1-z$.
\end{theorem}
\begin{proof}
We apply a slight variation of
the method from Section~\ref{sec:method}. We start by
making the Ansatz (compare with \eqref{eq:Ansatz1})
\begin{equation*}
M(z)=\sum _{i=0} ^{2^{\al+1}-1}a_i(z)\Om^i(z^4)\quad
\text {modulo }2^{2^\al},
\end{equation*}
where $\Om(z)$ is given by \eqref{eq:Omdef},
and where the $a_i(z)$'s
are (at this point) undetermined Laurent polynomials in $z$ and
$1-z$.
For the ``base step" (that is, for $\be=1$), we claim that
$$
M_1(z)=\sum_{k=1}^{\al+1}\frac {z^{2^k-2}} {(1-z)^{2^k-1}}+
\frac {1-z} {z^2}\Om^{2^\al}(z^4)
$$
solves \eqref{eq:MotzkinEF} modulo~2. Indeed, substitution of $M_1(z)$
in place of $M(z)$ on the left-hand side of \eqref{eq:MotzkinEF} yields
the expression
\begin{multline*}
{z^2}\left(
\sum_{k=1}^{\al+1}\frac {z^{2\cdot(2^k-2)}} {(1-z)^{2\cdot(2^k-1)}}+
\frac {(1-z)^2} {z^4}\Om^{2^{\al+1}}(z^4)
\right)\\
+(z-1)\left(\sum_{k=1}^{\al+1}\frac {z^{2^k-2}} {(1-z)^{2^k-1}}+
\frac {1-z} {z^2}\Om^{2^\al}(z^4)\right)
+1\quad \text{modulo }2.
\end{multline*}
Now one uses the relation \eqref{eq:OmRel2}
to ``get rid of" $\Om^{2^{\al+1}}(z^4)$. This leads
to the expression
\begin{multline*}
{z^2}\left(
\sum_{k=1}^{\al+1}\frac {z^{2\cdot(2^k-2)}} {(1-z)^{2\cdot(2^k-1)}}+
\frac {(1-z)^2} {z^4}\Om^{2^{\al}}( z^4)
+\frac {(1-z)^2} {z^4}\frac {z^{4\cdot 2^\al}} {(1- z^4)^{2^{\al}}}
\right)\\
+(z-1)\left(\sum_{k=1}^{\al+1}\frac {z^{2^k-2}} {(1-z)^{2^k-1}}+
\frac {1-z} {z^2}\Om^{2^\al}(z^4)\right)
+1\\
=
\sum_{k=1}^{\al+1}\frac {z^{2^{k+1}-2}} {(1-z)^{2^{k+1}-2}}+
\frac {(1-z)^2} {z^2}\Om^{2^{\al}}( z^4)
+\frac {z^{2^{\al+2}-2}} {(1- z)^{2^{\al+2}-2}}
\\
+\sum_{k=1}^{\al+1}\frac {z^{2^k-2}} {(1-z)^{2^k-2}}+
\frac {(1-z)^2} {z^2}\Om^{2^\al}(z^4)
+1
\quad \text{modulo }2,
\end{multline*}
which is indeed $0$ modulo~2.
\medskip
After we have completed the ``base step," we now proceed with the
iterative steps described in Section~\ref{sec:method}. We consider
the Ansatz \eqref{eq:Ansatz2}--\eqref{eq:Ansatz2b}, where the
coefficients $a_{i,\be}(z)$ are supposed to provide a solution
$M_{\be}(z)=\sum _{i=0} ^{2^{\al+1}-1}a_{i,\be}(z)\Om^i(z^4)$ to
\eqref{eq:MotzkinEF} modulo~$2^\be$. This Ansatz, substituted in
\eqref{eq:MotzkinEF}, produces the congruence
\begin{equation*}
z^2M_{\be}^2(z)+(z-1)M_{\be}(z)
+2^\be(z-1)\sum _{i=0} ^{2^{\al+1}-1}b_{i,\be+1}(z)\Om^i(z^4)
+1=0
\quad
\text {modulo }2^{\be+1}.
\end{equation*}
By our assumption on $M_{\be}(z)$, we may divide by $2^\be$.
Comparison of powers of $\Om(z^4)$ then yields a system of congruences
of the form
$$
(1-z)b_{i,\be+1}(z)+\text {Pol}_i(z)=0\quad
\text {modulo }2,\quad \quad
i=0,1,\dots,2^{\al+1}-1,
$$
where $\text {Pol}_i(z)$, $i=0,1,\dots,2^{\al+1}-1$, are in
$\Z[z,z^{-1},(1-z)^{-1}]$. This system being trivially
uniquely solvable, we have proved that, for an arbitrary positive
integer $\al$, our (variant of the) algorithm of
Section~\ref{sec:method} will produce a solution $M_{{2^{\al+1}}}(z)$
to\break \eqref{eq:MotzkinEF} modulo~$2^{2^\al}$ which is a
polynomial in $\Om(z^4)$ with coefficients that are
in\break $\Z[z,z^{-1},(1-z)^{-1}]$.
\end{proof}
\begin{comment}
We have implemented the algorithm explained in the above proof.
For $\al=1$, it produces the following result.
\begin{theorem} \label{thm:Motzkin-4}
Let $\Om(z)$ be given by \eqref{eq:Omdef}.
Then, we have
\begin{multline}
\label{eq:LoesM4}
\sum _{n\ge0} ^{}M_n\,z^n
=\frac {1} {1-z}+\frac {z^2} {(1-z)^3}+\frac {2z^4} {(1-z)^5}
+\frac {2z^2} {(1-z)^3}\Om(z^4)\\
+\left(z^{-2}+3z^{-1}+\frac {2} {z^2(1-z)}\right)\Om^2(z^4)
+2\left(z^{-2}+z^{-1}\right)\Om^3(z^4)
\quad
\text {\em modulo }4.
\end{multline}
\end{theorem}
As we show next, the above result can be drastically simplified.
Namely, Relation~\eqref{eq:OmRel} implies the congruence
$$
\Om^3(z^4)=\frac {1} {1-z^4}\Om(z^4)+\frac {z^4} {1-z^4}
=\frac {1} {(1-z)^4}\Om(z^4)+\frac {z^4} {(1-z)^4}
\quad \text{modulo }2.
$$
Use of this congruence and another application of
Relation~\eqref{eq:OmRel} allow one to transform \eqref{eq:LoesM4}
into the following, simpler form.
\begin{corollary} \label{cor:Motzkin-4}
Let $\Om(z)$ be given by \eqref{eq:Omdef}.
Then, we have
\begin{equation}
\label{eq:LoesM4A}
\sum _{n\ge0} ^{}M_n\,z^n
=\frac {1} {1-z}+\frac {z^2} {(1-z)^3}+\frac {2} {1-z}\Om(z^4)
+\frac {1-z} {z^2}\Om^2(z^4)
\quad
\text {\em modulo }4.
\end{equation}
\end{corollary}
We may now analyse the terms on the right-hand side of
\eqref{eq:LoesM4A} one by one. In the sequel, we shall employ the
notation $\coef{z^n}f(z)$ for the coefficient of $z^n$ in the Laurent
series $f(z)$. Our first auxiliary result concerns the second term on
the right-hand side of \eqref{eq:LoesM4A}. It is a simple consequence
of the binomial theorem.
\begin{lemma} \label{lem:(1-z)^3}
For all non-negative integers $n$, we have
$$
\coef{z^n}\frac {z^2} {(1-z)^3}\equiv
\begin{cases}
0\pmod4,&\text{if\/ }n\equiv0,1\pmod8,\\
1\pmod4,&\text{if\/ }n\equiv2,7\pmod8,\\
2\pmod4,&\text{if\/ }n\equiv4,5\pmod8,\\
3\pmod4,&\text{if\/ }n\equiv3,6\pmod8.\\
\end{cases}
$$
\end{lemma}
For the third term on the right-hand side of \eqref{eq:LoesM4A}, the
following result holds.
\begin{lemma} \label{lem:Om1-z}
Let $n=n_0+n_12+n_22^2+\dots+n_k2^k$ be the binary expansion of the
non-negative integer $n$. Then, we have
$$
\coef{z^n}\frac {1} {1-z}\Om(z^4)\equiv
n_2+n_3+\dots+n_k\pmod2.
$$
\end{lemma}
\begin{proof}
In order to determine the coefficient of $z^n$ in $\frac {1}
{1-z}\Om(z^4)$, we have to count all positive integers $m$
not exceeding $n$ whose
binary representation has an even number of zeroes on the far
right. Let
$$
m=2^{2e}+m_{2e+1}2^{2e+1}+m_{2e+1}2^{2e+1}+\dots+m_{k}2^{k},
$$
for appropriate positive integers $e$ and $m_i\in\{0,1\}$,
$i=2e+1,\dots,k$. If $n_{2e}=1$, then the number of such $m$'s equals
$$
1+(n_{2e+1}+n_{2e+2}2+n_{2e+3}2^2+\dots+n_k2^{k-2})\equiv 1+n_{2e+1}\pmod2.
$$
On the other hand, if $n_{2e}=0$, then all ``previous" $m$'s are still
allowed, with the exception of
$$
m=2^{2e}+n_{2e+1}2^{2e+1}+n_{2e+1}2^{2e+1}+\dots+n_{k}2^{k}.
$$
Consequently, the number of $m$'s which we want to determine here equals
$$
n_{2e+1}+n_{2e+2}2+n_{2e+3}2^2+\dots+n_k2^{k-2}\equiv n_{2e+1}\pmod2.
$$
To summarise: modulo~2, the result is $n_{2e}+n_{2e+1}$ in both cases.
If we now sum this over all $e\ge1$, we obtain the result.
\end{proof}
\begin{lemma} \label{lem:Om^2}
For all non-negative integers $n$, we have
\begin{equation} \label{eq:Om^2}
\coef{z^n}\Om^2(z)\equiv
\begin{cases}
2(n_2+n_3+\dots+n_k)\pmod4,\\
&\kern-3cm
\text{if\/ }n=1+n_12+n_22^2+\dots+n_k2^k,\\
2(1+n_2+n_3+\dots+n_k)\pmod4,\\
&\kern-3cm
\text{if\/
}n=2^{2e}(1+n_12+n_22^2+\dots+n_k2^k),\ e>0,\\
1+2n_1\pmod4,\\
&\kern-3cm
\text{if\/
}n=2^{2e+1}(1+n_12+n_22^2+\dots+n_k2^k).
\end{cases}
\end{equation}
\end{lemma}
\begin{proof}
Here, we have to count the number of pairs $(m_1,m_2)$ of positive
integers with $m_1+m_2=n$, both of whose
binary representations have an even number of zeroes on the far
right. Equivalently, we have to count the number of quadruples
$(e_1,f_1;e_2,f_2)$, where $e_1,e_2,f_1,f_2$
are non-negative integers, such that
\begin{equation*}
4^{e_1}(2f_1+1)+4^{e_2}(2f_2+1)=n.
\end{equation*}
Supposing that, without loss of generality, we have $e_1\le e_2$,
we may write
\begin{equation} \label{eq:ef1}
n=4^{e_1}(2f_1+1)+4^{e_2}(2f_2+1)=4^{e_1}\big(2f_1+1+4^{e_2-e_1}(2f_2+1)\big).
\end{equation}
Let first $n=1+n_12+n_22^2+\dots+n_k2^k$. Then,
quadruples $(e_1,f_1;e_2,f_2)$ satisfying \eqref{eq:ef1} must either satisfy
$e_1=0<e_2$ or $e_1>0=e_2$. We concentrate on the former case.
In view of \eqref{eq:ef1},
it is not difficult to see that the number of such quadruples equals
\begin{align}
\notag
\sum_{e_2>0}&\fl{\frac {1} {2}\left(\frac {n-1} {4^{e_2}}+1\right)}
=\sum_{e_2\ge1}
\fl{\frac {1} {2}\left(\frac {n_12+n_22^2+\dots+n_k2^k} {4^{e_2}}+1\right)}\\
\notag
&=\sum_{e_2\ge1}
\fl{\frac {n_1} {2^{2e_2}}
+\frac {n_2} {2^{2e_2-1}}+\dots
+\frac {n_{2e_2}+1} {2}+n_{2e_2+1}+2n_{2e_2+2}+\dots+2^{k-2e_2-1}n_k
}\\
\notag
&=\sum_{e_2\ge1}
(n_{2e_2}+n_{2e_2+1})\quad \text{modulo }2\\
&=n_2+n_3+\dots+n_k\quad \text{modulo }2.
\label{eq:n2n3}
\end{align}
This congruence has to be multiplied by $2$, in order to yield the
first case in \eqref{eq:Om^2}.
Next let $n=4^e(1+n_12+n_22^2+\dots+n_k2^k)$ with $e\ge1$. Then the above
arguments apply again (with $e_1=e$). Furthermore, there are now also
quadruples $(e_1,f_1;e_2,f_2)$ with $e_1=e_2<e$ satisfying
\eqref{eq:ef1}. Modulo~4, the only ones which are relevant in our count
are those with $e_1=e_2=e-1$ and
$$
2f_1+2f_2+2=4+n_18+\dots+n_k2^{k+2}.
$$
The number of pairs $(f_1,f_2)$ of non-negative integers satisfying
this equation equals
$$
2+n_14+\dots+n_k2^{k+1}\equiv 2\pmod 4.
$$
If this is added to twice the right-hand side of \eqref{eq:n2n3},
the claimed result in the second case of \eqref{eq:Om^2} is obtained.
The last remaining case is $n=2\cdot
4^e(1+n_12+n_22^2+\dots+n_k2^k)$. Here, quadruples $(e_1,f_1;e_2,f_2)$
satisfying \eqref{eq:ef1} must necessarily satisfy $e_1=e_2\le e$. It is
then easy to see that, modulo~$4$, the only quadruples relevant for
our count are those with $e_1=e_2=e$. Their number equals
$$
1+n_12+n_22^2+\dots+n_k2^k\equiv 1+2n_1\pmod 4,
$$
which is the claimed result corresponding to the last case in
\eqref{eq:Om^2}.
\end{proof}
If one carefully combines Corollary~\ref{cor:Motzkin-4} with
Lemmas~\ref{lem:(1-z)^3}--\ref{lem:Om^2}, then one obtains the
following result for the values of the Motzkin numbers modulo~4.
\begin{theorem} \label{thm:Motzkin4A}
Let $n=n_0+n_12+n_22^2+\dots+n_k2^k$ be the binary expansion of the
non-negative integer $n$.
The Motzkin number $M_n$, when reduced modulo~$4$, is congruent to
\begin{enumerate}
\item $1+2n_3+2n_4+\dots+2n_k$ {\em modulo 4}, if $n\equiv0,1,4,5$~{\em(mod~$8$)};
\item $2+2n_3$ {\em modulo 4}, if $n\equiv2$~{\em(mod~$8$)};
\item $2n_3$ {\em modulo 4}, if $n\equiv3$~{\em(mod~$8$)};
\item $2+2n_{e+1}$ {\em modulo 4}, if $n\equiv6$~{\em(mod~$8$)},
$n_3=\dots=n_{e-1}=1$, and $n_e=0$, for even $e$;
\item $3+2n_3+2n_4+\dots+2n_k+2n_{e+1}$ {\em modulo 4}, if\/
$n\equiv6,7$~{\em(mod~$8$)},
$n_3=\dots=n_{e-1}=1$, and $n_e=0$, for odd $e$;
\item $2n_{e+1}$ {\em modulo 4}, if $n\equiv7$~{\em(mod~$8$)},
$n_3=\dots=n_{e-1}=1$, and $n_e=0$, for even $e$.
\end{enumerate}
\end{theorem}
\end{comment}
\section{The generating function for Motzkin numbers modulo $8$}
\label{sec:M8}
We have implemented the algorithm described in the proof of
Theorem~\ref{thm:Motzkin}.\footnote{The {\sl Mathematica} input
files are freely available; see the Note at the end of the
Introduction.}
If we apply it with $\al=2$, then we obtain
\begin{multline} \label{eq:M16}
\sum_{n\ge0}M_n\,z^n
=
\frac{8 z^{15}+8 z^{11}+8 z^9+8 z^8+8 z^5}
{(1-z)^{19}}
+\frac{4 z^{10}+12 z^9+12 z^7}
{(1-z)^{13}}\\
+\frac{2 z^{10}+4 z^8+12 z^7+2 z^6+12 z^5+2 z^4}
{(1-z)^{11}}
+\frac{z^6}{(1-z)^7}
+\frac{z^2}{(1-z)^3}
+\frac{1}{1-z}\\
+ \frac{4 z^{10}
}{(1-z)^{11}}\Om(z)
+\left(\frac{8
z^{10}}{(1-z)^{13}}+\frac{4
z^8}{(1-z)^9}+\frac{2
z^6}{(1-z)^7}\right) \Om^2(z)\\
+\left(\frac{8
z^8}{(1-z)^9}+\frac{8
z^6}{(1-z)^9}+\frac{8
z^4}{(1-z)^9}+\frac{4
z^2}{(1-z)^7}\right) \Om^3(z)\\
+\left(
\frac{8 z^{10}}{(1-z)^{11}}
+\frac{8( z^9+ z^8+ z^6+ z^4+ z+1)}
{(1-z)^{11}}
+\frac{4 z^4+12 z+12 z^{-1}}
{(1-z)^5}\right.\\\left.
+\frac{2 z^2+12 z+6+12 z^{-1}+2 z^{-2}}
{(1-z)^3}
+\frac{15}{z}+\frac{1}{z^2}
\right) \Om^4(z)\\
+\left(
\frac{8 z^2}{(1-z)^3}
+\frac{4 z^{-2}}{(1-z)^3}\right) \Om^5(z)
+\left(
\frac{8 (z^3+1+ z^{-1})}{(1-z)^5}
+\frac{4 z^{-2}}{(1-z)}
+\frac{2}{z^2}
+\frac{14}{z}\right)
\Om^6(z)\\
+\left(\frac{4}{z^2}+\frac{12}{z}+\frac{8}{1-z}\right)
\Om^7(z)
\quad \quad \text{modulo }16.
\end{multline}
We are aiming at congruences for the Motzkin numbers
modulo~$8$, thus we need to reduce the above expression
modulo~$8$. In order to do this efficiently, we introduce the
error series $E(z)$ defined by the relation
\begin{equation} \label{eq:OmE}
\Om^2(z) + \Om(z) - \frac {z} {1 - z}-2E(z)=0.
\end{equation}
A straightforward computation shows that, explicitly, the series
$E(z)$ is given by
\begin{equation} \label{eq:Edef}
E(z)=\underset{(e_1,f_1)\prec(e_2,f_2)}{\sum_{e_1,f_1,e_2,f_2\ge0} }
z^{4^{e_1}(2f_1+1)+4^{e_2}(2f_2+1)}.
\end{equation}
Here, the symbol $\prec$ refers to the lexicographic order on pairs of
integers, i.e., $(e_1,f_1)\prec(e_2,f_2)$ if, and only if,
$e_1<e_2$, or $e_1=e_2$ and $f_1<f_2$.
We combine three reductions: (1) powers of $\Om(z^4)$ are reduced by
means of \eqref{eq:N} with $N=3$; (2) powers of $\Om(z^4)$ are further
reduced using Relation~\eqref{eq:OmE} (at the cost of introducing the
error series $E(z)$); (3) coefficients are reduced modulo~$8$.
The final result is
\begin{multline} \label{eq:MM}
M(z)= \left(\frac{4
(z+1)}{z^2}E(z^4)+\frac{7}{z^2}+\frac{1}{z}\right.\\
\left.+\frac{2 z^7+2
z^6+6 z^5+6 z^4+6 z^3+6 z^2+2
z+2}{1-z^8}\right)\Om(z^4)\\
+\frac{4
E(z^8) (z+1)}{z^2}-2 E(z^4)
\left(\frac{7}{z^2}+\frac{5}{z}+\frac{2}
{1-z}\right)\\
+\frac{\begin{matrix}
(z^{15}+5 z^{14}+7
z^{13}+3 z^{12}+3 z^{11}+3 z^{10}+5
z^9+5 z^8\kern2cm\\\kern3cm+z^7+5 z^6+3 z^5+7 z^4+3 z^3+3
z^2+z+1)\end{matrix}}{1-z^{16}}\\
\quad \quad \text{modulo }8.
\end{multline}
Here, we also used that $E^2(z^4)=E(z^8)$~modulo~$2$. (This explains
the occurrence of $E(z^8)$ in the above expression.)
In order to proceed, we need to know how to extract coefficients
of~$z^n$ from the series
$$
E(z^4)\Om(z^4),\
\Om(z^4),\
\frac {1} {1-z^8}\Om(z^4),\
E(z^4),\
\text{and }\frac {1} {1-z}E(z^4).
$$
The next section is devoted to solving these problems.
\section{Coefficient extraction}
\label{sec:aux}
We begin with coefficient extraction from the error series $E(z)$.
In all what follows, coefficients $n_i$ will be assumed to be
either 0 or~1.
\begin{lemma} \label{lem:E}
For all non-negative integers $n,$ we have
\begin{equation} \label{eq:E}
\coef{z^n}E(z)\equiv
\begin{cases}
\chi(e>0)(1+2n_1)+n_2+n_3+3n_4+n_5+3n_6+n_7+\cdots \pmod4,\\
&\kern-8cm
\text{if\/
}n=4^{e}(1+n_12+n_22^2+\dots+n_k2^k),\\
2\chi(e>0)+n_1+2n_2\pmod4,\\
&\kern-8cm
\text{if\/
}n=2\cdot4^{e}(1+n_12+n_22^2+\dots+n_k2^k),
\end{cases}
\end{equation}
where $\chi(\mathcal A)=1$ if the assertion $\mathcal{A}$ holds true,
and $\chi(\mathcal A)=0$ otherwise.
\end{lemma}
\begin{proof}
We have to count the number of quadruples
$(e_1,f_1;e_2,f_2)$, where $e_1,e_2,f_1,f_2$
are non-negative integers with $e_1<e_2$, or $e_1=e_2$ and $f_1<f_2$,
such that
\begin{equation*}
4^{e_1}(2f_1+1)+4^{e_2}(2f_2+1)=n.
\end{equation*}
We may write
\begin{equation} \label{eq:ef2}
n=4^{e_1}(2f_1+1)+4^{e_2}(2f_2+1)=4^{e_1}\big(2f_1+1+4^{e_2-e_1}(2f_2+1)\big).
\end{equation}
Let first $n=1+n_12+n_22^2+\dots+n_k2^k$. Then,
quadruples $(e_1,f_1;e_2,f_2)$ satisfying \eqref{eq:ef2} must satisfy
$e_1=0<e_2$. For given $e_2$, the possible numbers $f_2$ such that
\eqref{eq:ef2} is satisfied with a suitable $f_1$ are $0,1,\dots,F$,
where
$$F=\fl{\frac {1} {2}\left(\frac {n-1} {4^{e_2}}-1\right)}.$$
Thus, the number of above quadruples equals
\begin{align}
\notag
\sum_{e_2>0}&\fl{\frac {1} {2}\left(\frac {n-1} {4^{e_2}}+1\right)}
=\sum_{e_2\ge1}
\fl{\frac {1} {2}\left(\frac {n_12+n_22^2+\dots+n_k2^k} {4^{e_2}}+1\right)}\\
\notag
&=\sum_{e_2\ge1}
\fl{\frac {n_1} {2^{2e_2}}
+\frac {n_2} {2^{2e_2-1}}+\dots
+\frac {n_{2e_2}+1} {2}+n_{2e_2+1}+2n_{2e_2+2}+\dots+2^{k-2e_2-1}n_k
}\\
\notag
&=\sum_{e_2\ge1}
(n_{2e_2}+n_{2e_2+1}+2n_{2e_2+2})\pmod 4\\
&=n_2+n_3+3n_4+n_5+3n_6+n_7+3n_8+\cdots\pmod4.
\label{eq:n2n3B}
\end{align}
This is in agreement with the
first case in \eqref{eq:E} with $e=0$.
Next let $n=4^e(1+n_12+n_22^2+\dots+n_k2^k)$ with $e\ge1$. Then the above
arguments apply again (with $e_1=e<e_2$). Furthermore, there are now also
quadruples $(e_1,f_1;e_2,f_2)$ with $e_1=e_2<e$ satisfying
\eqref{eq:ef2}. Modulo~4, the only ones which are relevant in our count
are those with $e_1=e_2=e-1$ and
$$
2f_1+2f_2+2=4+n_18+\dots+n_k2^{k+2}.
$$
The number of pairs $(f_1,f_2)$ of non-negative integers with
$f_1<f_2$ satisfying this equation equals
$$
1+n_12+\dots+n_k2^{k}\equiv 1+2n_1\pmod 4.
$$
If this is added to the right-hand side of \eqref{eq:n2n3B},
the resulting expression is in agreement with the first case of
\eqref{eq:E} with $e>0$.
The last remaining case is $n=2\cdot
4^e(1+n_12+n_22^2+\dots+n_k2^k)$. Here, quadruples $(e_1,f_1;e_2,f_2)$
satisfying \eqref{eq:ef2} must necessarily satisfy $e_1=e_2\le e$. It is
then easy to see that, modulo~$4$, the only quadruples relevant for
our count are those with $e_1=e_2=e-1$ (which only exist if $e>0$),
whose number is congruent to
$2$~ modulo~$4$, and those with
$e_1=e_2=e$, whose number equals
$$
n_1+n_22+\dots+n_k2^{k-1}\equiv n_1+2n_2\pmod 4.
$$
Both cases together lead to
the claimed result corresponding to the second case in \eqref{eq:E}.
\end{proof}
Next, we turn to coefficient extraction from the series $E(z^4)/(1-z)$.
\begin{lemma} \label{lem:E1-z}
For all non-negative integers $n$ with binary representation
$n=n_0+n_12+n_22^2+\cdots,$ we have
\begin{multline} \label{eq:E1-z}
\coef{z^n}E(z^4)\frac {1} {1-z}\equiv
\sum_{e_2>0}
(n'_{2e_2+2}+n'_{2e_2+3})
(n_2+n_3+n_4+n_5+\cdots+n_{2e_2+1})\\
+\sum_{i\ge1}(n_{2i+1}+1)n_{2i+2}
\pmod2,
\end{multline}
where $n-4=n'_0+n'_12+n'_24+\cdots$.
\end{lemma}
\begin{proof}
We have to count the number of quintuples
$(e_1,f_1;e_2,f_2;g)$, where $e_1,e_2,f_1,f_2,g$
are non-negative integers with $e_1<e_2$, or $e_1=e_2$ and $f_1<f_2$,
such that
\begin{equation*}
4^{e_1+1}(2f_1+1)+4^{e_2+1}(2f_2+1)+g=n.
\end{equation*}
Given $e_2$, the range of possible $f_2$'s is $0,1,\dots,F_2$,
where
$$F_2=\fl{\frac {1} {2}\left(\frac {n-4} {4^{e_2+1}}-1\right)}.$$
Given $e_2$, $f_2$, and $e_1<e_2$, the range of possible $f_1$'s is
$0,1,\dots,F_1$, where
$$F_1=\fl{\frac {1} {2}\left(\frac {n-4^{e_2+1}(2f_2+1)}
{4^{e_1+1}}-1\right)}.$$
Thus, the number of above quintuples with $e_1<e_2$ equals
\begin{align}
\notag
\sum_{e_2>0}&
\sum_{f_2=0} ^{F_2}
\sum_{e_1=0} ^{e_2-1}
\fl{\frac {1} {2}\left(\frac {n-4^{e_2+1}(2f_2+1)} {4^{e_1+1}}+1\right)}\\
&=
\notag
\sum_{e_2>0}
\sum_{f_2=0} ^{F_2}
\sum_{e_1=0} ^{e_2-1}
\left(
\fl{\frac {1} {2}\left(\frac {n}
{4^{e_1+1}}+1\right)}
-2\cdot4^{e_2-e_1-1}(2f_2+1)
\right)
\\
&\equiv
\notag
\sum_{e_2>0}
\sum_{f_2=0} ^{F_2}
\sum_{e_1=0} ^{e_2-1}
(n_{2e_1+2}+n_{2e_1+3})
\pmod 2\\
&\equiv
\notag
\sum_{e_2>0}
\sum_{f_2=0} ^{F_2}
(n_2+n_3+n_4+n_5+\cdots+n_{2e_2+1})
\pmod 2\\
\notag
&\equiv
\sum_{e_2>0}
(F_2+1)\cdot
(n_2+n_3+n_4+n_5+\cdots+n_{2e_2+1})\pmod2\\
&\equiv
\sum_{e_2>0}
(n'_{2e_2+2}+n'_{2e_2+3})
(n_2+n_3+n_4+n_5+\cdots+n_{2e_2+1})
\pmod 2.
\label{eq:n2n3C}
\end{align}
where $n-4=n'_0+n'_12+n'_24+\cdots$.
On the other hand, if $e_1=e_2$, then we want to count all triples
$(e_1,f_1,f_2)$ of non-negative integers with $f_1<f_2$ such that
$$
2\cdot 4^{e_1+1}(f_1+f_2+1)\le n.
$$
Since the number of pairs $(x,y)$ of non-negative integers with $x<y$
such that $x+y=k$ equals $\fl{(k+1)/2}$, the number of above triples
equals
\begin{equation} \label{eq:e1F}
\sum_{e_1\ge0}\sum_{k=0}^{F} \fl{\frac {k+1} {2}}
\end{equation}
where
$$
F=\fl{\frac {n} {2\cdot4^{e_1+1}}}-1.
$$
If $m=m_0+m_12+m_24+\cdots$, then it is not difficult to see that
$\sum_{k=0}^{m} \fl{(k+1)/2}$ is odd if and only if
$m\equiv1$~(mod~4). In symbols,
\begin{equation} \label{eq:sumk}
\sum_{k=0}^{m} \fl{\frac {k+1}
{2}}=m_0(m_1+1)\pmod2.
\end{equation}
Thus, modulo~2, the sum in \eqref{eq:e1F} becomes
\begin{equation}
\sum_{e_1\ge0}\sum_{k=0}^{F} \fl{\frac {k+1} {2}}
\equiv (n_3+1)n_4+(n_5+1)n_6+\cdots\pmod2.
\label{eq:n2n3D}
\end{equation}
The sum of \eqref{eq:n2n3C} and \eqref{eq:n2n3D} then yields
\eqref{eq:E1-z}.
\end{proof}
Our next series to be considered is $\Om(z)/(1-z^2)$.
\begin{lemma} \label{lem:Omz^2}
For all non-negative integers $n$ with binary representation
$n=n_0+n_12+n_22^2+\cdots,$ we have
\begin{equation} \label{eq:Omz^2}
\coef{z^n} \frac {1} {1-z^2}\Om(z)\equiv
\begin{cases}
1+n_1+2n_2\pmod4,&\text{if $n$ is odd,}\\
n_2+n_3+3n_4+n_5+3n_6+n_7+\cdots\pmod4,&\text{if $n$ is even.}
\end{cases}
\end{equation}
\end{lemma}
\begin{proof}
We have to count the number of triples
$(e_1,f_1;g)$, where $e_1,f_1,g$
are non-negative integers such that
\begin{equation*}
4^{e_1}(2f_1+1)+2g=n.
\end{equation*}
Let first $n=1+n_12+n_22^2+\dots+n_k2^k$.
Then necessarily $e_1=0$, so that we want to count all non-negative
integers $f_1$ with $f_1\le (n-1)/2$. Modulo~4, the number $(n-1)/2$ is
congruent to $n_1+2n_2$. The number of possible $f_1$'s is by~1 larger, and
this observation explains the first case in \eqref{eq:Omz^2}.
Now let $n=n_12+n_22^2+\dots+n_k2^k$. Given $e_1>0$, the integer
$f_1$ may range from $0$ to $F$, where
$$F=\fl{\frac {1} {2}\left(\frac {n} {4^{e_1}}-1\right)}.$$
Thus, the number of possible triples $(e_1,f_1;g)$ is
\begin{align*}
\sum_{e_2>0}&\fl{\frac {1} {2}\left(\frac {n} {4^{e_1}}+1\right)}\\
&=\sum_{e_2>0}
\fl{\frac {n_1} {2^{2e_1}}
+\frac {n_2} {2^{2e_1-1}}+\dots
+\frac {n_{2e_1}+1} {2}+n_{2e_1+1}+2n_{2e_1+2}+\dots+2^{k-2e_1-1}n_k
}\\
&\equiv\sum_{e_2>0}(n_{2e_1}+n_{2e_1+1}+2e_{2e_1+2})\pmod4\\
&\equiv
n_2+n_3+3n_4+n_5+3n_6+n_7+\cdots\pmod4,
\end{align*}
as desired.
\end{proof}
Finally, we address coefficient extraction from the product
$E(z)\Om(z)$.
\begin{lemma} \label{lem:OmE}
Let $n$ be a non-negative integer.
\smallskip
{\em(1)} If $n=4^e(1+n_12+n_22^2+\dots+n_k2^k),$ then
\begin{multline} \label{eq:OmE1}
\coef{z^n} E(z)\,\Om(z)\equiv
(n_1+1)n_2+\sum_{i\ge1}n_{2i+1}(n_{2i+2}+1)
+\chi(e\ge1)\cdot(n_2+n_3+\cdots)\\
+\sum_{e_3>0}
(n'_{2e_3}+n'_{2e_3+1})
(n_2+n_3+n_4+n_5+\cdots+n_{2e_3-1})
\pmod2,
\end{multline}
where $n4^{-e}-4=1+n'_12+n'_24+\cdots$.
\smallskip
{\em(2)} If $n=2\cdot 4^e(1+n_12+n_22^2+\dots+n_k2^k),$ then
\begin{equation} \label{eq:OmE2}
\coef{z^n} E(z)\,\Om(z)\equiv
\chi(e\ge1)+(n_{1}+n_{2})
+(n_{1}+n_{2}+n_3+n_4+\cdots)(1+n_1)
\pmod2.
\end{equation}
\end{lemma}
\begin{proof}
We have to count the number of sextuples
$(e_1,f_1;e_2,f_2;e_3,f_3)$, where $e_1,e_2,f_1,f_2,\break e_3,f_3$
are non-negative integers with $(e_1,f_1)\prec (e_2,f_2)$,
such that
\begin{equation} \label{eq:ef3}
4^{e_1}(2f_1+1)+4^{e_2}(2f_2+1)+4^{e_3}(2f_3+1)=n.
\end{equation}
We need this number only modulo~2.
By pairing up sextuples $(e_1,f_1;e_2,f_2;e_3,f_3)$ with
$(e_1,f_1)\prec (e_3,f_3)\prec (e_2,f_2) $ with those
sextuples with $(e_3,f_3)\prec (e_1,f_1)\prec (e_2,f_2) $,
one sees that we may equivalently count all sextuples
$(e_1,f_1;e_2,f_2;e_3,f_3)$ with
$(e_1,f_1)\preceq (e_2,f_2)\preceq (e_3,f_3) $, but not all three
equal, satisfying \eqref{eq:ef3}. In the remainder of this proof,
we shall always assume these conditions.
\medskip
{\sc Case 1:}
$n=4^e(1+n_12+n_22^2+\dots+n_k2^k)$.
Here, we have to consider all possible cases for solutions to \eqref{eq:ef3}
to exist: (a) $e_1=e_2=e_3=e$; (b) $0\le e_1=e<e_2=e_3$; (c)
$0\le e_1=e<e_2<e_3$; (d) $0\le e_1=e_2<e_3$.
\smallskip
(a) Writing $n=4^eu$, we see that \eqref{eq:ef3} reduces to
\begin{equation} \label{eq:f1f2f3}
f_1+f_2+f_3=\frac {u-3} {2},
\end{equation}
with $f_1\le f_2\le f_3$ but not all three equal. Let us denote the
number of triples $(f_1,f_2,f_3)$ with $f_1+f_2+f_3=N$, $f_1\le f_2\le
f_3$, not all three equal, by $f(N)$. In order to compute $f(N)$,
we face a classical partition problem, namely the problem of counting
all integer partitions of $N$ with at most three parts, not all of
them equal. It is then folklore (cf.\ \cite[Sec.~3.2]{AndrAF}) that,
for the generating function we have
$$
\sum_{N=0}^\infty f(N)\,z^N=\frac {1} {(1-z)(1-z^2)(1-z^3)}-\frac {1} {1-z^3}
=\frac {z+z^2-z^3} {(1-z)(1-z^2)(1-z^3)}.
$$
Modulo~2, this reduces to
$$
\sum_{N=0}^\infty f(N)\,z^N=\frac {z} {1-z^4}
\quad \quad \text{modulo }2.
$$
In other words, $f(N)\equiv1$~(mod~2) if, and only if,
$N\equiv1$~(mod~4). Returning to our problem of counting triples
satisfying \eqref{eq:f1f2f3} modulo~2, this means that the number of the above
triples is odd if, and only if, $u\equiv5$~(mod~8). Hence, a slick way to
express this number of triples modulo~2 in terms of the binary digits of $u$ is
\begin{equation} \label{eq:(a)}
n_0 (n_1+1)n_2=(n_1+1)n_2\pmod2.
\end{equation}
\smallskip
(b) With the convention $n=4^eu$ from above, we see that we have to
count all triples $(e_2,f_2,f_3)$ with $e_2>e$ and $f_2\le f_3$ satisfying
$$
2\cdot4^{e_2-e}(f_2+f_3+1)\le u-1.
$$
Since the number of pairs $(x,y)$ of non-negative integers with $x\le y$
such that $x+y=k$ equals $\fl{(k+2)/2}$, the number of above triples
equals
\begin{equation} \label{eq:e1FA}
\sum_{e_2>e}\sum_{k=0}^{F} \fl{\frac {k+2} {2}}
\end{equation}
where
$$
F=\fl{\frac {u-1} {2\cdot4^{e_2-e}}}-1.
$$
Using \eqref{eq:sumk} again,
the sum in \eqref{eq:e1FA}, when reduced modulo~2, becomes
\begin{equation}
\sum_{e_2>e}\sum_{k=0}^{F} \fl{\frac {k+2} {2}}
\equiv n_3(n_4+1)+n_5(n_6+1)+\cdots\pmod2.
\label{eq:n2n3F}
\end{equation}
\smallskip
(c) With the same notation as above,
here we have to count all quadruples\break $(e_2,f_2;e_3,f_3)$ with
$e<e_2<e_3$ satisfying
$$
4^{e_2-e}(2f_2+1)+4^{e_3-e}(2f_3+1)\le u-1.
$$
Given $e_3$ with $e_3>e$, the range of possible $f_3$'s is $0,1,\dots,F_3$,
where
$$F_3=\fl{\frac {1} {2}\left(\frac {u-5} {4^{e_3-e}}-1\right)}.$$
Given $e_3$, $f_3$, and $e_2$ with $e<e_2<e_3$,
the range of possible $f_2$'s is $0,1,\dots,F_2$, where
$$F_2=\fl{\frac {1} {2}\left(\frac {u-4^{e_3-e}(2f_3+1)-1}
{4^{e_2-e}}-1\right)}.$$
Thus, the number of above quadruples with $e_2<e_3$ equals
\begin{align}
\notag
\sum_{e_3>e}&
\sum_{f_3=0} ^{F_3}
\sum_{e_2=e+1} ^{e_3-1}
\fl{\frac {1} {2}\left(\frac {u-4^{e_3-e}(2f_3+1)-1} {4^{e_2-e}}+1\right)}\\
&=
\notag
\sum_{e_3>e}
\sum_{f_3=0} ^{F_3}
\sum_{e_2=e+1} ^{e_3-1}
\left(
\fl{\frac {1} {2}\left(\frac {u-1}
{4^{e_2-e}}+1\right)}
-2\cdot4^{e_3-e_2-1}(2f_3+1)
\right)
\\
&\equiv
\notag
\sum_{e_3>e}
\sum_{f_3=0} ^{F_3}
\sum_{e_2=e+1} ^{e_3-1}
(n_{2e_2-2e}+n_{2e_2-2e+1})
\pmod 2\\
&\equiv
\notag
\sum_{e_3>e}
\sum_{f_3=0} ^{F_3}
(n_2+n_3+n_4+n_5+\cdots+n_{2e_3-2e-1})
\pmod 2\\
\notag
&\equiv
\sum_{e_3>e}
(F_3+1)\cdot
(n_2+n_3+n_4+n_5+\cdots+n_{2e_3-2e-1})\pmod2\\
&\equiv
\sum_{e_3>e}
(n'_{2e_3-2e}+n'_{2e_3-2e+1})
(n_2+n_3+n_4+n_5+\cdots+n_{2e_3-2e-1})
\pmod 2,
\label{eq:(c)}
\end{align}
where $u-4=1+n'_12+n'_24+\cdots$.
\smallskip
(d) Here, Equation \eqref{eq:ef3} reduces to
\begin{equation} \label{eq:2u}
(f_1+f_2+1)+2\cdot 4^{e_3-e_1-1}(2f_3+1)=2\cdot4^{e-e_1-1} u.
\end{equation}
From this equation it is obvious that $e_1\le e-1$.
Since the second term on the left-hand side is divisible by~$2$
because of $e_3>e_1$, the sum $f_1+f_2+1$ must also be divisible
by~$2$. We need not consider the case where $f_1+f_2+1\equiv0$~(mod~4)
since then the number of possible pairs $(f_1,f_2)$ with $f_1\le f_2$
is even.
Thus, we may assume that $f_1+f_2+1\equiv2$~(mod~4), in which case
the number of possible pairs $(f_1,f_2)$ with $f_1\le f_2$ is odd.
Inspecting
\eqref{eq:2u} under this assumption again, we conclude that $e_3\ge e+1$.
The counting problem that remains to be solved (modulo~2) hence is
to count all pairs $(e_3,f_3)$ with $e_3\ge e+1$ satisfying
\begin{equation} \label{eq:eu2}
4^{e_3}(2f_3+1)\le 4^{e} \cdot u.
\end{equation}
Given $e_3$, the integer $f_3$ ranges between $0$ and $F_3$, where
$$
F_3=\fl{\frac {1} {2}\left(\frac {u} {4^{e_3-e}}-1\right)}.
$$
The number of pairs $(e_3,f_3)$ satisfying
\eqref{eq:eu2} then is
\begin{align}
\notag
\sum_{e_3>e}(F_3+1)
&=
\sum_{e_3>e}\fl{\frac {1} {2}\left(\frac {u} {4^{e_3-e}}+1\right)}\\
\notag
&\equiv
\sum_{e_3>e}(n_{2e_3-2e}+n_{2e_3-2e+1})
\quad \quad \text{(mod 2)}\\
&\equiv
n_2+n_3+\cdots
\quad \quad \text{(mod 2)}.
\label{eq:(d)}
\end{align}
It should be noted that, since in the current case $0\le e_1\le e-1$,
we must have $e\ge1$ in order that \eqref{eq:(d)} can actually occur.
The contribution of Subcase~(d) to the final result therefore is
\begin{equation} \label{eq:(d2)}
\chi(e\ge1)\cdot(n_2+n_3+\cdots)
\quad \quad \text{(mod 2)}.
\end{equation}
\smallskip
Adding up the individual contributions \eqref{eq:(a)},
\eqref{eq:n2n3F}, \eqref{eq:(c)}, and \eqref{eq:(d2)}, we arrive
at the right-hand side of \eqref{eq:OmE1}.
\medskip
{\sc Case 2:}
$n=2\cdot 4^e(1+n_12+n_22^2+\dots+n_k2^k)$.
In order to have solutions to \eqref{eq:ef3} (with
$(e_1,f_1)\preceq (e_2,f_2)\preceq (e_3,f_3) $, but not all three
equal, as we assume throughout), we must have
$0\le e_1=e_2<e_3$. In that case, Equation~\eqref{eq:ef3} reduces to
\begin{equation} \label{eq:eu1}
2\cdot 4^{e_1}(f_1+f_2+1)+4^{e_3}(2f_3+1)=n=2\cdot 4^e\cdot u,
\end{equation}
where $u$ is some odd number.
Let us for the moment fix the sum of $f_1+f_2+1$ to equal~$k$, say.
If this number is divisible by~$4$, say $k=4k'$, then
the number of possible pairs $f_1,f_2$ with $f_1\le f_2$ and
$f_1+f_2+1=4k'$ is
$\fl{(4k'+1)/2}=2k'$, which is even. Since we are only interested in
numbers of solutions modulo~2, we may disregard these cases.
There are two possibilities which remain to be considered: either
\begin{equation} \label{eq:f1f2A}
f_1+f_2+1=2u_0+1
\end{equation}
for some $u_0\ge0$, or
\begin{equation} \label{eq:f1f2B}
f_1+f_2+1=2(2u_0+1)
\end{equation}
for some $u_0\ge0$.
\smallskip
In the first case, that is, if \eqref{eq:f1f2A} holds, then,
from \eqref{eq:ef3}, we obtain
$$
2\cdot4^{e_1}(2u_0+1)+4^{e_3}(2f_3+1)=n=2\cdot 4^e\cdot u,
$$
for some odd positive integer~$u$. Because of $e_1<e_3$, this implies
$e_1=e$.
Given $e_3$ with $e_3>e$, the range of possible $f_3$'s is $0,1,\dots,F_3$,
where
$$F_3=\fl{\frac {1} {2}\left(\frac {n} {4^{e_3}}-1\right)}.$$
Given $e_3>e$ and $f_3$, the numbers $e_1$ and $u_0$ are uniquely
determined (recall that $e_1=e$). Since
the number of pairs $(f_1,f_2)$ with $f_1\le f_2$
satisfying \eqref{eq:f1f2A} equals
$\fl{\frac {1} {2}\big((2u_0+1)+1\big)}=u_0+1$,
the number of solutions to \eqref{eq:ef3} which we have to consider in the
current case is
\begin{align}
\notag
\sum_{e_3>e}&
\sum_{f_3=0} ^{F_3}(u_0+1)
=
\sum_{e_3>e}
\sum_{f_3=0} ^{F_3}
\fl{\frac {1} {2}\left(\frac {n-4^{e_3}(2f_3+1)} {2\cdot
4^{e}}+1\right)}
\\
&=
\notag
\sum_{e_3>e}
\sum_{f_3=0} ^{F_3}
\left(
\fl{\frac {1} {2}\left(\frac {n}
{2\cdot4^{e}}+1\right)}
-4^{e_3-e-1}(2f_3+1)
\right)
\\
&\equiv
\notag
\sum_{e_3>e}
(F_3+1)
\left(
\fl{\frac {1} {2}\left(\frac {n}
{2\cdot4^{e}}+1\right)}
-\chi(e_3=e+1)
\right)
\pmod 2\\
&\equiv
\notag
\sum_{e_3>e}
(n_{2e_3-2e-1}+n_{2e_3-2e})
\left((1+n_1)
-\chi(e_3=e+1)
\right)
\pmod 2\\
&\equiv
(n_1+n_2+n_3+\cdots)(1+n_1)-(n_1+n_2)
\pmod 2.
\label{eq:n2n3E}
\end{align}
\smallskip
In the second case, that is, if \eqref{eq:f1f2B} holds, then the number
of possible pairs $(f_1,f_2)$ with $f_1\le f_2$ satisfying
\eqref{eq:f1f2B} equals $\fl{\frac {1} {2}\big(2(2u_0+1)+1\big)}=2u_0+1$,
which is odd. Since we are only interested in the number of solutions
to \eqref{eq:ef3} modulo~2, we may therefore simply continue with
$u_0$, and determine the number of quadruples $(e_1,u_0,e_3,f_3)$
for which
$$
4^{e_1+1}(2u_0+1)+4^{e_3}(2f_3+1)=2\cdot 4^e\cdot u,
$$
for some odd positive integer~$u$. Because of $e_1<e_3$, it follows
that $e_1+1=e_3$. Thus, the above relation becomes
$$
u_0+f_3+1=4^{e_3-e}\cdot u.
$$
There is no relation between $u_0$ and $f_3$, hence the number of
pairs $(u_0,f_3)$ satisfying the above relation equals the right-hand
side, which is $4^{e_3-e}u$. This quantity is odd if, and only if,
we have $e_3=e$. Since $e_3=e_1+1\ge1$,
the contribution of the current case is $\chi(e\ge1)$.
If we add this to the earlier contribution \eqref{eq:n2n3E}, then
we obtain the right-hand side of \eqref{eq:OmE2}.
\medskip
This completes the proof of the lemma.
\end{proof}
\section{Motzkin numbers modulo $8$}
\label{sec:Mn8}
Using the auxiliary results from Section~\ref{sec:aux} in
\eqref{eq:MM}, we are now able to provide explicit formulae for
the congruence class of~$M_n$ modulo~8 in terms of the binary digits
of~$n$.
\begin{theorem} \label{thm:M8}
Let $n$ be a positive integer with binary expansion
$$
n=n_0+n_1\cdot2+n_2\cdot2^2+\cdots.
$$
Furthermore, let $K$ denote the least integer $\ge4$ such that
$n_K=0$.
The Motzkin numbers $M_n$ satisfy the following congruences
modulo~$8${\em:}
$$M_n\equiv_8 \begin{cases}
2 s_2^2(n) + 4 e_2(n) + 1,&\text{if\/ }n\equiv0~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 3,
&\text{if\/ }n\equiv1~\text{\em(mod $16$),}\\
4 s_2(n) + 6,&\text{if\/ }n\equiv2~\text{\em(mod $16$),}\\
4,&\text{if\/ }n\equiv3~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 3,&\text{if\/ }n\equiv4~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 5,&\text{if\/ }n\equiv5~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 n_4\, s_2(n) + 4 e_2(n) + 2 n_4 + 7,
&\text{if\/ }n\equiv6~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 n_4\, s_2(n) + 4 e_2(n) + 2 n_4 + 5,
&\text{if\/ }n\equiv7~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 1,&\text{if\/ }n\equiv8~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 3,
&\text{if\/ }n\equiv9~\text{\em(mod $16$),}\\
4,&\text{if\/ }n\equiv10~\text{\em(mod $16$),}\\
4 s_2(n) + 2,&\text{if\/ }n\equiv11~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 3,
&\text{if\/ }n\equiv12~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 5,&\text{if\/ }n\equiv13~\text{\em(mod $16$),}\\
(n_{K+1} + 1) (4 s_2(n) + 6),
&\text{if\/ }n\equiv14~\text{\em(mod $16$) and $K$ is even,}\\
2 n_{K+1} + 4 n_{K+1} s_2(n) + 2 s_2^2(n) + 4 s_2(n) + 4 e_2(n) +
7,\hskip-3cm\\
&\text{if\/ }n\equiv14~\text{\em(mod $16$) and $K$ is odd,}\\
6 n_{K+1} + 4 n_{K+1} s_2(n) + 4,
&\text{if\/ }n\equiv15~\text{\em(mod $16$) and $K$ is even,}\\
2 n_{K+1} + 4 n_{K+1} s_2(n) + 2 s_2^2(n) + 4
e_2(n) + 5,\hskip-1cm\\
&\text{if\/ }n\equiv15~\text{\em(mod $16$) and $K$ is odd}.
\end{cases}
$$
Here, $s_2(n)$ denotes the sum of the binary digits of~$n,$
while $e_2(n)$ denotes the number of pairs of successive digits
in the binary expansion of~$n,$ with both digits equalling~$1$.
The symbol $x\equiv_8y$ is short for $x\equiv y$~{\em(mod $8$)}.
\end{theorem}
\begin{proof}[Sketch of proof]
We have to discuss 16 different cases depending on the congruence
class of $n$ modulo~$16$. Since the arguments in all these cases
are very similar, we content ourselves with presenting the most
complex case in detail, leaving the remaining cases to the reader.
Let $n\equiv15$~(mod~16), and
let $L$ be the least integer $\ge K+2$ with $n_L\ne0$.
In other words, the binary representation of the number $n$ is of the form
\begin{equation} \label{eq:n2}
(n)_2=
\dots\overset{L}{\overset{\downarrow}1}0\dots0n_{K+1}
\overset{K}{\overset{\downarrow}0}1\dots1
\raise-1pt\hbox to0pt{.\hss}
\raise2pt\hbox to0pt{.\hss}
\raise5pt\hbox to0pt{.\hss}
\raise8pt\hbox to0pt{.\hss}\hphantom{.}
1111.
\end{equation}
(The vertical dots separate significant digits from the four 1's on
the right making $n$ a number congruent to $15$
modulo~$16$. Otherwise, these dots may be safely ignored.)
Let first $K$ be even, say $K=2k$. For our specific~$n$, we extract the
coefficient of $z^n$ in \eqref{eq:MM}. The relevant terms are
\begin{multline} \label{eq:terms}
4z^{-1}E(z^4)\Om(z^4),\
z^{-1}\Om(z^4),\
\frac {6z^3} {1-z^8}\Om(z^4),\
\frac {2z^7} {1-z^8}\Om(z^4),\\
4z^{-1}E(z^8),\
-10z^{-1}E(z^4),\
-\frac {4} {1-z}E(z^4),\
\frac {z^{15}} {1-z^{16}}.
\end{multline}
Coefficient extraction for these terms is discussed in
Lemmas~\ref{lem:OmE},
\ref{lem:Omz^2},
\ref{lem:E},
and \ref{lem:E1-z},
respectively (ordered according to the order of the terms in \eqref{eq:terms}).
Adding all contributions, we obtain
\begin{multline} \label{eq:M15}
4(n_{2k+1}+1)n_{2k+2}+
4\sum_{i\ge k+1}n_{2i+1}(n_{2i+2}+1)+
4\sum_{i\ge 2k+2}n_i\\
+
4\sum_{e>k}(n'_{2e}+n'_{2e+1})(n_{2k+2}+n_{2k+3}+\dots+n_{2e-1})\\
+
1+6(2+2n_4)+2(n_4+n_5+3n_6+n_7+3n_8+n_9+\cdots)+
4(2+n_{2k+1}+2n_{2k+2})\\
+
6(1+2n_{2k+1}+n_{2k+2}+n_{2k+3}+3n_{2k+4}+n_{2k+5}+3n_{2k+6}+n_{2k+7}+\cdots)\\
+
4\sum_{e>k}(n_{2e+2}+n_{2e+3})(n_{2}+n_{3}+\dots+n_{2e+1})
+
4\sum_{i\ge1}(n_{2i+1}+1)n_{2i+2}+1,
\end{multline}
where the digits $n'_j$ are defined by
$$
(n+1)2^{-2k}-4=1+n'_{2k+1}2+n'_{2k+2}4+\cdots.
$$
In order to control the auxiliary digits $n'_j$,
we have to distinguish between $L$ (recall \eqref{eq:n2}) being
even or odd. If this is done, one sees that one may
simply drop the primes without changing the result.
Next, we observe that
$$
n_2+n_3+\dots+n_{2k+1}\equiv1+n_{2k+1}\pmod2.
$$
Using this, we may split the first sum in the last line of
\eqref{eq:M15} in the following way:
\begin{align} \notag
4\sum_{e>k}&(n_{2e+2}+n_{2e+3})(n_{2}+n_{3}+\dots+n_{2e+1})\\
\notag
&\equiv
4\sum_{e>k}(n_{2e+2}+n_{2e+3})(1+n_{2k+1})\\
\notag
&\kern2cm
+4\sum_{e>k}(n_{2e+2}+n_{2e+3})(n_{2k+2}+n_{2k+3}+\dots+n_{2e+1})
\pmod8\\
\notag
&\equiv
4(1+n_{2k+1})\sum_{i\ge 2k+4}n_i
+4\sum_{e>k}(n_{2e+2}+n_{2e+3})(n_{2k+2}+n_{2k+3}+\dots+n_{2e+1})\\
&\kern12cm
\pmod8.
\label{eq:n'n}
\end{align}
The last sum in this expression cancels with the sum in the
second line of \eqref{eq:M15}, when reduced modulo~$8$, except
for the term for $e=k+1$.
A similar partial cancellation, modulo~$8$, takes place between the
big terms in parentheses in the third and fourth lines of
\eqref{eq:M15}.
For the first sum in the first line and the last sum in the last
line, we observe that
\begin{align*}
4\sum_{i\ge k+1}&n_{2i+1}(n_{2i+2}+1)
+4\sum_{i\ge1}(n_{2i+1}+1)n_{2i+2}\\
&=
4\sum_{i\ge k+1}n_{2i+1}n_{2i+2}
+4\sum_{i\ge k+1}n_{2i+1}
+4\sum_{i\ge1}n_{2i+1}n_{2i+2}
+4\sum_{i\ge1}n_{2i+2}\\
&\equiv
4\sum_{i=1}^kn_{2i+1}n_{2i+2}
+4\sum_{i\ge 2k+2}n_{i}
+4\sum_{i=1}^kn_{2i+2}
\pmod8\\
&\equiv
4(k-2+n_{2k+1}n_{2k+2})
+4\sum_{i\ge 2k+2}n_{i}
+4(k-2+n_{2k+2})
\pmod8\\
&\equiv
4(n_{2k+1}+1)n_{2k+2}
+4\sum_{i\ge 2k+2}n_{i}
\pmod8.\\
\end{align*}
After further simplification, one obtains the claimed expression
$6 n_{K+1} + 4 n_{K+1} s_2(n) + 4$.
\medskip
Now let $K$ be odd, say $K=2k+1$. The relevant terms from which one
extracts the coefficient of $z^n$ are again the ones in
\eqref{eq:terms}. However, the relevant cases
in Lemmas~\ref{lem:OmE} and \ref{lem:E} are not the same.
More precisely, here we obtain
\begin{multline} \label{eq:M15B}
4+4n_{2k+2}+4n_{2k+3}+4(n_{2k+2}+1)\sum_{i\ge 2k+2}n_i\\
+0
+ 6(2+2n_4)+2(n_4+n_5+3n_6+n_7+3n_8+n_9+\cdots)\\
+
4(1+2n_{2k+2}+n_{2k+3}+n_{2k+4}+3n_{2k+5}+n_{2k+6}+3n_{2k+7}+n_{2k+8}+\cdots)\\
+
6(2+n_{2k+2}+2n_{2k+3})+
4\sum_{e>k}(n_{2e+2}+n_{2e+3})(n_{2}+n_{3}+\dots+n_{2e+1})\\
+
4\sum_{i\ge1}(n_{2i+1}+1)n_{2i+2}+1.
\end{multline}
This expression can be simplified in much the same way as
\eqref{eq:M15}; with one notable difference though. The sums
in the next-to-last and the last lines have no counterparts
(as opposed to the situation in \eqref{eq:M15}). Consequently,
they need a different treatment.
The sum in the next-to-last line of \eqref{eq:M15B} is first
transformed in the same way as before, see \eqref{eq:n'n}.
Then one observes that
\begin{align*}
4\sum_{e>k}&(n_{2e+2}+n_{2e+3})(n_{2k+2}+n_{2k+3}+\dots+n_{2e+1})\\
&=4\sum_{i>j\ge 2k+2}n_in_j-4\sum_{i\ge k+1}n_{2i}n_{2i+1}\\
&=2\bigg(\sum_{i\ge 2k+2}n_i\bigg)^2-
2\sum_{i\ge 2k+2}n_i-4\sum_{i\ge k+1}n_{2i}n_{2i+1}\\
&=2\bigg(s_2(n)-\sum_{i=0}^{2k+1}n_i\bigg)^2-
2\sum_{i\ge 2k+2}n_i
-4\sum_{i\ge k+1}n_{2i}n_{2i+1}\\
&\equiv2s_2^2(n)-4n_{2k+1}s_2(n)+2n_{2k+1}-
2\sum_{i\ge 2k+2}n_i
-4\sum_{i\ge k+1}n_{2i}n_{2i+1}
\pmod8.
\end{align*}
The last sum combines with the sum $4\sum_{i\ge1}n_{2i+1}n_{2i+2}$
appearing in the last line of \eqref{eq:M15B} into $4e_2(n)$, up to some
error that can be computed explicitly modulo~$8$.
We leave the remaining simplifications, leading to the claimed
expression
$$ 2 n_{2k+2} + 4 n_{2k+2} s_2(n) + 2 s_2^2(n) + 4
e_2(n) + 5,$$
to the reader.
\end{proof}
The following result of Eu, Liu and Yeh \cite{EuLYAB}, characterising
even congruence classes of~$M_n$ modulo~$8$, can now readily be
obtained through a straightforward case-by-case analysis using the
corresponding cases from Theorem~\ref{thm:M8}.
\begin{corollary} \label{thm:EuLYAB}
The Motzkin number $M_n$ is even if, and only if,
$n = (4i + \ep)4^{j+1} -\de$
for non-negative integers $i, j$, $\ep = 1, 3$, and
$\de = 1, 2$. Moreover, we have
$$
M_n\equiv_8\begin{cases}
4,&\text{if\/ }(\ep,\de)=(1,1)\text{ or }(3,2),\\
4y+2,&\text{if\/ }(\ep,\de)=(1,2)\text{ or }(3,1),
\end{cases}
$$
where $y$ is the number of\/ $1$'s in the binary expansion of\/
$4i+\ep-1$.
\end{corollary}
Moreover, Theorem~\ref{thm:M8} allows us to provide a characterisation
of all those~$n$ for which $M_n$ lies in {\it any} specified
congruence class modulo~$8$. As an example, we give the
characterisation for the congruence class~1.
\begin{corollary} \label{cor:Mn1}
The Motzkin number $M_n$ is congruent to~$1$ modulo~$8$ if, and only
if, $n$ satisfies one of the following conditions:
\begin{enumerate}
\item $n\equiv0$~{\em(mod~$16$)} and
$s_2(n)\equiv e_2(n)\equiv0$~{\em(mod~$2$);}
\item $n\equiv1$~{\em(mod~$16$),}
$s_2(n)\equiv1$~{\em(mod~$2$),} and $e_2(n)\equiv0$~{\em(mod~$2$);}
\item $n\equiv4$~{\em(mod~$16$),}
$s_2(n)\equiv1$~{\em(mod~$2$),} and $e_2(n)\equiv0$~{\em(mod~$2$);}
\item $n\equiv5$~{\em(mod~$16$),}
$s_2(n)\equiv0$~{\em(mod~$2$),} and $e_2(n)\equiv1$~{\em(mod~$2$);}
\item $n\equiv6$~{\em(mod~$16$),} $n_4=0,$
$s_2(n)\equiv e_2(n)\equiv1$~{\em(mod~$2$);}
\item $n\equiv6$~{\em(mod~$16$),} $n_4=1,$
$s_2(n)\equiv e_2(n)\equiv0$~{\em(mod~$2$);}
\item $n\equiv7$~{\em(mod~$16$),} $n_4=0,$
$s_2(n)\equiv0$~{\em(mod~$2$),} and $e_2(n)\equiv1$~{\em(mod~$2$);}
\item $n\equiv7$~{\em(mod~$16$),} $n_4=1,$
$s_2(n)\equiv e_2(n)\equiv1$~{\em(mod~$2$);}
\item $n\equiv8$~{\em(mod~$16$),}
$s_2(n)\equiv e_2(n)\equiv0$~{\em(mod~$2$);}
\item $n\equiv9$~{\em(mod~$16$),}
$s_2(n)\equiv1$~{\em(mod~$2$),} and $e_2(n)\equiv0$~{\em(mod~$2$);}
\item $n\equiv12$~{\em(mod~$16$),}
$s_2(n)\equiv1$~{\em(mod~$2$),} and $e_2(n)\equiv0$~{\em(mod~$2$);}
\item $n\equiv13$~{\em(mod~$16$),}
$s_2(n)\equiv0$~{\em(mod~$2$),} and $e_2(n)\equiv1$~{\em(mod~$2$);}
\item $n\equiv14$~{\em(mod~$16$),} $K$ odd, $n_{K+1}=0,$
$s_2(n)\equiv e_2(n)\equiv1$~{\em(mod~$2$);}
\item $n\equiv14$~{\em(mod~$16$),} $K$ odd, $n_{K+1}=1,$
$s_2(n)\equiv e_2(n)\equiv0$~{\em(mod~$2$);}
\item $n\equiv15$~{\em(mod~$16$),} $K$ odd, $n_{K+1}=0,$
$s_2(n)\equiv0$~{\em(mod~$2$),} and $e_2(n)\equiv1$~{\em(mod~$2$);}
\item $n\equiv15$~{\em(mod~$16$),} $K$ odd, $n_{K+1}=1,$
$s_2(n)\equiv e_2(n)\equiv1$~{\em(mod~$2$);}
\end{enumerate}
where $s_2(n),$ $e_2(n),$ and $K$ are defined as in Theorem~{\em\ref{thm:M8}}.
\end{corollary}
It should be observed that it is straightforward to generate all
possible~$n$ in any of the 16~cases in the characterisation of the
previous corollary.
Characterisations similar to the one in Corollary~\ref{cor:Mn1}
are available for all other congruence classes modulo~$8$.
We leave it as an open problem whether a description for the odd
congruence classes of~$M_n$ modulo~$8$ exists which is
comparably compact as the one of Eu, Liu and Yeh for the even ones,
given here in Corollary~\ref{thm:EuLYAB}. We do not hide
that we are very sceptical about this.
\section{Further applications}
\label{Sec:FurthApp}
In this final section, we show that the same approach that we
applied in Sections~\ref{sec:Motzkin}--\ref{sec:Mn8} to Motzkin numbers
also works for the sequences of Motzkin prefix numbers, of Riordan
numbers, of hex tree numbers, and of central trinomial coefficients.
Since the arguments are completely analogous, we content ourselves
with brief sketches of the main points and subsequent statements
of the corresponding results for congruences modulo~$8$.
\subsection{Motzkin prefix numbers modulo $8$}
Let $MP_n$ be the $n$-th {\it Motzkin prefix number}, that is, the number of
lattice paths from $(0,0)$ consisting of $n$ steps taken from
the set
$\{(1,0),(1,1),(1,-1)\}$ never running below the $x$-axis.
Gouyou-Beauchamps and Viennot \cite{GoViAA} have shown that
$MP_n$ also counts directed rooted animals with $n + 1$ vertices.
It is well-known (cf.\ e.g.\ \cite[Sec.~8]{KrMuAE})
that the generating function $MP(z)=\sum_{n\ge0}MP_n\,z^n$ satisfies the
functional equation
\begin{equation} \label{eq:MotzkinprefEF}
z(1-3z)MP^2(z)+(1-3z)MP(z)-1=0.
\end{equation}
When we apply our method from Section~\ref{sec:method},
the Ansatz for the base step is
\begin{equation} \label{eq:Ansatz-MP}
MP_1(z)=\frac {1} {z}\Om^{2^\al}(z^4)+\sum_{k=0}^{\al+1}\frac {z^{2^k-1}}
{(1-z)^{2^k}}.
\end{equation}
Subsequently, everything runs through in the same way as in
the proof of Theorem~\ref{thm:Motzkin}. Consequently,
the generating function $MP(z)$ for Motzkin prefix numbers
satisfies a completely analogous theorem.
If we now follow the line of argument in Section~\ref{sec:M8},
then we obtain that $MP(z)$ admits the following representation
modulo~$8$:
\begin{multline} \label{eq:MP}
MP(z)=\frac{4 E(z^4) \Om(z^4)}{z}+\left(\frac{1}{z}+\frac{4 z^7+2
z^6+2 z^4+2 z^2+4
z+2}{1-z^8}\right)
\Om(z^4)\\
+\frac{4 E(z^8)}{z}+
\left(\frac{6}{z}+\frac{4}{1-z^2}\right
)E(z^4)\\
+\frac{\begin{matrix}6 z^{15}+z^{14}+z^{12}+4
z^{11}+5 z^{10}+6 z^9+z^8\kern3cm\\\kern3cm +6 z^7+z^6+4
z^5+z^4+4 z^3+5 z^2+2 z+1\end{matrix}}{1-z^{16}}
\quad \quad \text{modulo }8.
\end{multline}
Coefficient extraction using Lemmas~\ref{lem:E}--\ref{lem:OmE}
leads to the following theorem generalising \cite[Cor.~3.2]{DeSaAA}.
\begin{theorem} \label{thm:MP8}
Let $n$ be a positive integer with binary expansion
$$
n=n_0+n_1\cdot2+n_2\cdot2^2+\cdots.
$$
The Motzkin prefix numbers $MP_n$ satisfy the following congruences
modulo~$8${\em :}
$$MP_n\equiv_8 \begin{cases}
2 s_2^2(n) + 4 e_2(n) + 1,&\text{if\/ }n\equiv0~\text{\em(mod $16$),}\\
4 s_2(n) + 6,
&\text{if\/ }n\equiv1~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 7,&\text{if\/ }n\equiv2~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 1,&\text{if\/ }n\equiv3~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 1,&\text{if\/ }n\equiv4~\text{\em(mod $16$),}\\
0,&\text{if\/ }n\equiv5~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 7,
&\text{if\/ }n\equiv6~\text{\em(mod $16$),}\\
4 s_2(n) + 2 n_4 + 4 n_4 s_2(n) + 2,
&\text{if\/ }n\equiv7~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 1,&\text{if\/ }n\equiv8~\text{\em(mod $16$),}\\
4 s_2(n) + 6,
&\text{if\/ }n\equiv9~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 7,&\text{if\/ }n\equiv10~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 7,&\text{if\/ }n\equiv11~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 1,
&\text{if\/ }n\equiv12~\text{\em(mod $16$),}\\
0,&\text{if\/ }n\equiv13~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 7,
&\text{if\/ }n\equiv14~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 6 n_{K+1} + 4 n_{K+1} s_2(n) + 1,\hskip-2cm\\
&\text{if\/ }n\equiv15~\text{\em(mod $16$) and $K$ is even,}\\
4 s_2(n) + 2 n_{K+1} + 4 n_{K+1} s_2(n) + 2,\hskip-1cm\\
&\text{if\/ }n\equiv15~\text{\em(mod $16$) and $K$ is odd},
\end{cases}
$$
where $s_2(n),$ $e_2(n),$ and $K$ are defined as in Theorem~{\em\ref{thm:M8}}.
\end{theorem}
\subsection{Riordan numbers modulo $8$}
Let $R_n$ be the $n$-th {\it Riordan number}, that is, the number of
lattice paths from $(0,0)$ to $(n,0)$ consisting of steps taken from
the set
$\{(1,0),(1,1),(1,-1)\}$ never running below the $x$-axis, and where
steps $(1,0)$ are not allowed on the $x$-axis.
It is well-known (cf.\ e.g.\ \cite[Sec.~9]{KrMuAE})
that the generating function $R(z)=\sum_{n\ge0}R_n\,z^n$ satisfies the
functional equation
\begin{equation} \label{eq:RiordanEF}
z(1+z)R^2(z)-(z+1)R(z)+1=0.
\end{equation}
When we apply our method from Section~\ref{sec:method},
the Ansatz for the base step is the same as the one in the case of
Motzkin prefix numbers, that is,
the right-hand side of \eqref{eq:Ansatz-MP}.
Subsequently, everything runs through in the same way as in
the proof of Theorem~\ref{thm:Motzkin}. Consequently,
the generating function $R(z)$ for Riordan numbers
satisfies a completely analogous theorem.
If we now follow the line of argument in Section~\ref{sec:M8},
then we obtain that $R(z)$ admits the following representation
modulo~$8$:
\begin{multline} \label{eq:R}
R(z)=
\frac{4 E(z^4)\Om(z^4)}{z} +\left(\frac{7}{z}+\frac{2 z^6+4
z^5+2 z^4+4 z^3+2
z^2+2}{1-z^8}\right)\Om(z^4) \\
+\frac{4 E(z^8)
}{z}+
\left(\frac{2}{z}+\frac{4}{1-z^2}\right)E(z^4)\\
+\frac{\begin{matrix}5 z^{14}+2 z^{13}+z^{12}+2
z^{11}+z^{10}+4 z^9+z^8\kern3cm\\\kern5cm +5 z^6+6
z^5+z^4+2 z^3+z^2+1\end{matrix}}{1-z^{16}}
\quad \quad \text{modulo }8.
\end{multline}
Coefficient extraction using Lemmas~\ref{lem:E}--\ref{lem:OmE}
leads to the following theorem generalising \cite[Cor.~3.3]{DeSaAA}.
\begin{theorem} \label{thm:R8}
Let $n$ be a positive integer with binary expansion
$$
n=n_0+n_1\cdot2+n_2\cdot2^2+\cdots.
$$
The Riordan numbers $R_n$ satisfy the following congruences
modulo~$8${\em :}
$$R_n\equiv_8 \begin{cases}
2 s_2^2(n) + 4 e_2(n) + 1,&\text{if\/ }n\equiv0~\text{\em(mod $16$),}\\
0,
&\text{if\/ }n\equiv1~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 3,&\text{if\/ }n\equiv2~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 5,&\text{if\/ }n\equiv3~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 1,&\text{if\/ }n\equiv4~\text{\em(mod $16$),}\\
4 s_2(n) + 6,&\text{if\/ }n\equiv5~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 3,
&\text{if\/ }n\equiv6~\text{\em(mod $16$),}\\
6 n_4 + 4 n_4 s_2(n) + 4,
&\text{if\/ }n\equiv7~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 1,&\text{if\/ }n\equiv8~\text{\em(mod $16$),}\\
0,
&\text{if\/ }n\equiv9~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 3,&\text{if\/ }n\equiv10~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 7,&\text{if\/ }n\equiv11~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 1,
&\text{if\/ }n\equiv12~\text{\em(mod $16$),}\\
4 s_2(n) + 6,&\text{if\/ }n\equiv13~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 3,
&\text{if\/ }n\equiv14~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 2 n_{K+1} + 4 n_{K+1} s_2(n) + 5,\hskip-2cm\\
&\text{if\/ }n\equiv15~\text{\em(mod $16$) and $K$ is even,}\\
6 n_{K+1} + 4 n_{K+1} s_2(n) + 4,
&\text{if\/ }n\equiv15~\text{\em(mod $16$) and $K$ is odd},
\end{cases}
$$
where $s_2(n),$ $e_2(n),$ and $K$ are defined as in Theorem~{\em\ref{thm:M8}}.
\end{theorem}
\subsection{Hex tree numbers modulo $8$}
Let $H_n$ be the $n$-th {\it hex tree number}, that is,
the number of planar rooted trees where each vertex may have a left, a
middle, or a right descendant, but never a left {\it and\/} middle
descendant, and never a middle {\it and\/} right descendant.
It is well-known (and easy to see)
that the generating function $H(z)=\sum_{n\ge0}H_n\,z^n$
satisfies the functional equation
\begin{equation} \label{eq:hex treeEF}
z^2H^2(z)+(3z-1)H(z)+1=0.
\end{equation}
When we apply our method from Section~\ref{sec:method},
the Ansatz for the base step is
\begin{equation*}
H_1(z)=\frac {1-z} {z^2}\Om^{2^\al}(z^4)+\sum_{k=1}^{\al+1}\frac {z^{2^k-2}}
{(1-z)^{2^k-1}}.
\end{equation*}
Subsequently, everything runs through in the same way as in
the proof of Theorem~\ref{thm:Motzkin}. Consequently,
the generating function $H(z)$ for hex tree numbers
satisfies a completely analogous theorem.
If we now follow the line of argument in Section~\ref{sec:M8},
then we obtain that $H(z)$ admits the following representation
modulo~$8$:
\begin{multline} \label{eq:H}
H(z)=
\frac{4 E(z^4)\Om(z^4)
(z+1)}{z^2}\\
+ \left(\frac{7}{z^2}+\frac{3}{z}+\frac{6 z^7+2
z^6+2 z^5+6 z^4+2 z^3+6 z^2+6
z+2}{1-z^8}\right)\Om(z^4)\\
+\frac{4
E(z^8) (z+1)}{z^2}
+
\left(\frac{2}{z^2}+\frac{2}{z}+\frac{4}
{1-z}\right)E(z^4)\\
+\frac{\begin{matrix}3 z^{15}+5 z^{14}+5
z^{13}+3 z^{12}+z^{11}+3 z^{10}+7 z^9+5
z^8\kern1cm\\\kern2.5cm +3 z^7+5 z^6+z^5+7 z^4+z^3+3 z^2+3
z+1\end{matrix}}{1-z^{16}}
\quad \quad \text{modulo }8.
\end{multline}
Coefficient extraction using Lemmas~\ref{lem:E}--\ref{lem:OmE}
leads to the following theorem generalising \cite[Cor.~3.4]{DeSaAA}.
\begin{theorem} \label{thm:H8}
Let $n$ be a positive integer with binary expansion
$$
n=n_0+n_1\cdot2+n_2\cdot2^2+\cdots.
$$
The hex tree numbers $H_n$ satisfy the following congruences
modulo~$8${\em :}
$$H_n\equiv_8 \begin{cases}
2 s_2^2(n) + 4 e_2(n) + 1,&\text{if\/ }n\equiv0~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 1,
&\text{if\/ }n\equiv1~\text{\em(mod $16$),}\\
4 s_2(n) + 6,&\text{if\/ }n\equiv2~\text{\em(mod $16$),}\\
4,&\text{if\/ }n\equiv3~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 3,&\text{if\/ }n\equiv4~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 7,&\text{if\/ }n\equiv5~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 2 n_4 + 4 n_4 s_2(n) + 7,
&\text{if\/ }n\equiv6~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 6 n_4 + 4 n_4\, s_2(n) + 7,
&\text{if\/ }n\equiv7~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 1,&\text{if\/ }n\equiv8~\text{\em(mod $16$),}\\
2 s_2^2(n) + 4 e_2(n) + 1,
&\text{if\/ }n\equiv9~\text{\em(mod $16$),}\\
4,&\text{if\/ }n\equiv10~\text{\em(mod $16$),}\\
4 s_2(n) + 6,&\text{if\/ }n\equiv11~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 3,
&\text{if\/ }n\equiv12~\text{\em(mod $16$),}\\
6 s_2^2(n) + 4 e_2(n) + 7,&\text{if\/ }n\equiv13~\text{\em(mod $16$),}\\
4 s_2(n) + 6 n_{K+1} + 4 n_{K+1} s_2(n) + 6,
&\text{if\/ }n\equiv14~\text{\em(mod $16$) and $K$ is even,}\\
6 s_2^2(n) + 4 e_2(n) + 2 n_{K+1} + 4 n_{K+1} s_2(n) + 7,\hskip-3cm\\
&\text{if\/ }n\equiv14~\text{\em(mod $16$) and $K$ is odd,}\\
2 n_{K+1} + 4 n_{K+1} s_2(n) + 4,
&\text{if\/ }n\equiv15~\text{\em(mod $16$) and $K$ is even,}\\
6 s_2^2(n) + 4 e_2(n) + 6 n_{K+1} + 4 n_{K+1} s_2(n) + 7,\hskip-1cm\\
&\text{if\/ }n\equiv15~\text{\em(mod $16$) and $K$ is odd},
\end{cases}
$$
where $s_2(n),$ $e_2(n),$ and $K$ are defined as in Theorem~{\em\ref{thm:M8}}.
\end{theorem}
Inspection of the above sixteen cases reveals the interesting
fact that hex tree numbers are never divisible by~$8$
(a fact that is also true for Motzkin numbers).
\subsection{Central trinomial coefficients modulo $8$}
Let $T_n$ be the $n$-th {\it central trinomial coefficient}, that is, the
coefficient of $t^n$ in $(1+t+t^2)^n$.
Already Euler knew
that the generating function $T(z)=\sum_{n\ge0}T_n\,z^n$
equals
\begin{equation} \label{eq:zentrin}
T(z)=(1-2z-3z^2)^{-1/2}
\end{equation}
(cf.\ \cite[solution to
Exercise~6.42]{StanBI}).
Here we proceed differently in order to express $T(z)$ in terms of our
basic series $\Om(z^4)$ (and the error series $E(z^4)$).
We are forced to do so since our method from
Section~\ref{sec:method} fails.\footnote{The base step solution
would be the one with $a_0(z)=1/(1-z)$ and $a_{2^\al}(z)=0$.
Subsequently, the iteration step modulo~$4$ would not succeed,
however. We believe that the reason for this phenomenon might be that
the minimal polynomial for $\Om$ (in the sense of
\cite{KaKMAA,KrMuAE,KrMuAL}) is of degree less than~$4$.}
Instead, we compare \eqref{eq:Motzkin} and
\eqref{eq:zentrin} to see that\footnote{A similar approach would also
have been possible for the Motzkin
prefix and the Riordan numbers since their generating functions
satisfy relations with the generating function $M(z)$ for Motzkin
numbers analogous to \eqref{eq:TM}. However, no such relation
exists between the generating function for hex tree numbers
and $M(z)$.}
\begin{equation} \label{eq:TM}
T(z)=\frac {1-z-2z^2M(z)} {1-2z-3z^3}.
\end{equation}
Relation \eqref{eq:TM} would allow us to establish a complete analogue
of Theorem~\ref{thm:Motzkin} for the generating
function $T(z)$ for central trinomial coefficients.
In order to determine these numbers modulo~$8$,
we substitute \eqref{eq:MM} in the above relation. Then
some simplification eventually leads to
\begin{multline} \label{eq:T}
T(z)=
\frac{6 z^3+6 z^2+2 z+2
}{1-z^4}\Om(z^4)+\frac{4 E(z^4)}{1-z}\\
+\frac{3
z^7+7 z^6+z^5+z^4+7 z^3+3
z^2+z+1}{1-z^8}
\quad \quad \text{modulo }8.
\end{multline}
Coefficient extraction using Lemmas~\ref{lem:E1-z} and \ref{lem:Omz^2}
leads to the following theorem for central trinomial coefficients modulo~$8$.
\begin{theorem} \label{thm:T8}
Let $n$ be a positive integer with binary expansion
$$
n=n_0+n_1\cdot2+n_2\cdot2^2+\cdots.
$$
Modulo~$8,$ the central trinomial coefficients $T_n$
satisfy the following congruences:
$$T_n\equiv_8 \begin{cases}
2 s_2^2(n) + 4 e_2(n) + 1,
&\text{if\/ }n\equiv0~\text{\em(mod $2$),}\\
6 s_2^2(n) + 4 e_2(n) + 3,
&\text{if\/ }n\equiv1~\text{\em(mod $2$),}\\
\end{cases}
$$
where $s_2(n)$ and $e_2(n)$ are defined as in Theorem~{\em\ref{thm:M8}}.
\end{theorem}
\section*{Note}
The result in Theorem~\ref{thm:M8} has in the meantime also been obtained
--- in slightly different form and using a different approach --- by Wang and Xin
in ``A Classification of Motzkin Numbers Modulo~$8$" [{Electron\@. J. Combin\@.}
{\bf 25}(1) (2018), Article~P1.54, 15~pp.].
\section*{Acknowledgement}
The authors are indebted to an anonymous referee for clarifying
the scope of the automaton method of Rowland and Yassawi in \cite{RoYaAA}.
|
\section*{}
\documentclass[11pt]{article}
\usepackage{lineno}
\usepackage[hidelinks]{hyperref}
\usepackage{color}
\usepackage{amssymb,amsmath}
\usepackage{subfigure, float, graphicx}
\usepackage[square,super,sort&compress]{natbib}
\usepackage[margin=1in]{geometry}
\graphicspath{{Figs/}}
\usepackage[title,titletoc,toc]{appendix}
\modulolinenumbers[5]
\renewcommand{\vec}[1]{\ensuremath\boldsymbol{#1}}
\newcommand{\textcolor{red}{NEED CITATION }}{\textcolor{red}{NEED CITATION }}
\newcommand{\ensuremath\textrm{Ca}}{\ensuremath\textrm{Ca}}
\newcommand{\ensuremath\textrm{Re}}{\ensuremath\textrm{Re}}
\newcommand{\ensuremath\textrm{Rm}}{\ensuremath\textrm{Rm}}
\newcommand{\ensuremath\textrm{Mn}}{\ensuremath\textrm{Mn}}
\bibliographystyle{elsarticle-num}
\begin{document}
\begin{center}
\huge
Vesicles in magnetic fields
\bigskip
\normalsize
David Salac \\
Department of Mechanical and Aerospace Engineering, \\ University at Buffalo SUNY, \\ 318 Jarvis Hall, Buffalo, NY 14260-4400, USA\\716-645-1460\\\href{mailto:<EMAIL>}{<EMAIL>}
\end{center}
\noindent\makebox[\linewidth]{\rule{\textwidth}{0.4pt}}
\bigskip
Liposome vesicles tend to align with an applied magnetic field. This is due to the directional magnetic susceptibility difference
of the lipids which form the membrane of these vesicles. In this work a model of liposome vesicles exposed to magnetic
field is presented. Starting from the base energy of a lipid membrane in a magnetic field, the force applied
to the surrounding fluids is derived. This force is then used to investigate the dynamics of vesicle in the presence
of magnetic fields.
\bigskip
\noindent\makebox[\linewidth]{\rule{\textwidth}{0.4pt}}
\section{Introduction}
Knowledge about the directed motion of biological and bio-compatible nano- and microstructures, particularly liposome vesicles,
is critically important for a number of biotechnologies.
For example,
in directed drug delivery it is critical that drug carriers be directed towards locations where they are to release an encapsulated
drug.~\cite{Kagan2010} The directed motion of biological cells could be used to sort cells~\cite{Toner2005,Xia2006,Chen208} and to form
larger structures.~\cite{Solovev2010}
Numerous techniques have been proposed as possible ways to control the motion of these soft-matter systems. Specially designed microfluidic
devices can use differences in size,~\cite{Tan2008} shape,~\cite{DuBose2014} and rigidity~\cite{Wang2013} to physically separate particles.
It is also possible to use light~\cite{MacDonald2003} and acoustic waves~\cite{Zhou2015} to induce particle motion.
Of particular interest is the use of externally controllable fields to direct the motion of soft particles. Electric fields
have been demonstrated to induce large deformation in liposome vesicles both experimentally~\cite{Dimova2009,dimova2007} and
theoretically.~\cite{Kolahdouz2015a,salipante2014,schwalbe2011,mcconnell2013}
Electric fields can be used to sort cells~\cite{Barrett2005,Cummings2003,Vahey2008} and to induce the formation of pores in
vesicle membranes.~\cite{Riske2005}
Magnetic fields also offer the opportunity to direct the motion of biologically compatible soft-matter.
Experiments have demonstrated that liposome vesicles tend to align with an externally applied magnetic field.~\cite{BOROSKE1978}
The phospholipids which compose the vesicle membrane are known to be diamagnetically anisotropic and tend to align
perpendicularly to an external field.~\cite{Rikken2014} Due to the nature of the liposome membrane this results in a rotational/alignment
force which aligns and stretches a vesicle parallel to the applied field, see Fig. \ref{fig:lipidAlignment}.
Due to the unique nature of vesicles this stretching is balanced by an increase in surface tension and bending energy.
Using this fact Helfrich developed a model for the deformation of a vesicle when exposed to a magnetic field to
determine the flexibility of the vesicle membrane.~\cite{Helfrich1973,Helfrich1973a}
\begin{figure}
\centering
\includegraphics[width=6cm]{lipidAlignment}
\caption{To minimize the total energy lipids will orient perpendicularly to an applied magnetic field. This results in a
force which aligns the vesicle to the applied field.}
\label{fig:lipidAlignment}
\end{figure}
Unlike electric field effects, the influence of magnetic fields on liposome vesicles has received much less attention. In addition to the
work of Helfrich mentioned above, it has been experimentally shown that liposomes made from dipalmitoylphosphatidyl choline (DPPC)
have temperature dependent deformation and permeability when exposed to magnetic fields.~\cite{Tenforde1988} In the same work a simple model
was developed to explore the observed behavior. The fusion of liposome vesicles for a range of applied magnetic fields has also been
experimentally demonstrated~\cite{Ozeki2000} while others have verified the alignment
of vesicles to the external magnetic field.~\cite{Qiu1993,Kiselev2008}
Theoretical and computational investigation of magnetohydrodynamics of liposome vesicles are much less common. Helfrich investigated
the birefringence of vesicles in magnetic fields~\cite{Helfrich1973} and very few investigations into the biomechanics of vesicles in magnetic field
have been performed.~\cite{Ye2015}
To the author's knowledge this is the first effort to model the general dynamics of vesicles when exposed to externally driven magnetic fields.
In the remainder of this work the governing equations and numerical methods used to model this system will be presented. Sample results of a
vesicle in the presence of magnetic fields will also be shown.
\section{Governing Equations}
Consider a vesicle suspended in a fluid and exposed to an externally driven magnetic field, Fig. \ref{fig:schematic}.
The density, electrical and magnetic properties between the inner and outer fluid are matched while the viscosity may vary. The magnetic
field in static in time but could be spatially varying. The size of the vesicle
is on the order of 10 $\mu$m while the thickness of the membrane is approximately 5 nm, which allows for the modeling of the membrane as a thin
interface separating two fluids.~\cite{Seifert1997} The vesicle membrane is impermeable to fluid molecules and
the number of lipids in a vesicle membrane does not change over time
while the surface density of lipids at room temperature is constant.~\cite{Seifert1997} These conditions result in an inextensible membrane with
constant enclosed volume and global surface area, in addition to local surface area incompressibility.
\begin{figure}
\centering
\includegraphics[width=6cm]{VesicleInMField}
\caption{Schematic of a vesicle exposed to an externally driven magnetic field. The magnetic field $\vec{B}$ is at an angle of $\theta^B$ from the $x$-axis
while the inclination angle of the vesicle is given by $\theta^v$. The inner, $\mu^-$, and outer, $\mu^+$, fluid viscosity may differ. The outward
unit normal vector to the interface is given by $\vec{n}$.
}
\label{fig:schematic}
\end{figure}
For any multiphase fluid system the time-scale associated with charges migrating towards the interface is given by the charge relaxation time,
$t_c=\epsilon/\sigma$, where $\epsilon$ and $\sigma$ are the the fluid permittivity and conductivity, respectively.~\cite{Saville1997,Melcher1969}
Typical values for these in experimental vesicles investigations
are $\epsilon\approx 10^{-9}$ F/m and $s\approx 10^{-3}$ S/m.~\cite{Riske2006,salipante2014,needham1989electro} This results in a charge relaxation time
of $t_c\approx 10^{-6}$ s, which is much faster than observed dynamics of vesicles when exposed to magnetic fields.~\cite{BOROSKE1978}
It is thus valid to assume that there are no free charges in the bulk fluids and thus the leaky-dielectric model can be assumed.~\cite{Saville1997}
This lack of free charges in the bulk fluids has implications when considering the forces on the fluid. If there are free-charges in the fluid then
the Lorentz force will drag the fluid into motion, which is common for magnetohydrodynamics using conducting fluids.~\cite{Guo2013,Ki2010,Tagawa2006}
In the absence of free charges the Lorentz force can be ignored in the bulk fluid and therefore fluid will be driven into motion only
by conditions at the membrane.
In general the applied magnetic field and induced electric field are coupled through Maxwell's equations. Under the assumption
that the magnetic field is static in time the only possible coupling between the electric field and the magnetic field is through the
electric current density, $\vec{j}=\sigma\left(-\nabla\Phi+\vec{u}\times\vec{B}\right)$, where $\Phi$ is the
induced electric potential, $\vec{u}$ is the fluid velocity and $\vec{B}$ is the applied magnetic field. As the induced electrical
current must divergence free, $\nabla\cdot\vec{j}=0$,
any induced electric potential will obey $\nabla\cdot\left(\sigma\nabla\Phi\right)=\nabla\cdot\left(\sigma\vec{u}\times\vec{B}\right)$.
Experimental investigations of vesicles in a 1.5 T magnetic field demonstrate that responses take on the order of 10 s.~\cite{BOROSKE1978}
Assuming that the distance traveled during this time is 20 $\mu$m this results in a velocity of $2\times10^{-6}$ m/s. In the
absence of an external electric field this results in an induced electric current density of $3\times10^{-9}$ A/m$^2$.
At such small induced current densities the induced electric field will be much smaller than those needed to induce vesicle
deformation.~\cite{Dimova2009,Kolahdouz2015a,salipante2014,vlahovska2009electrohydrodynamic}
From this analysis the induced electric field and it's contribution to the dynamics of the vesicle will be ignored.
\subsection{Forces Exerted by the Membrane}
\label{sec:membraneForces}
The motion of the fluid will be driven by the conditions and forces at the vesicle membrane. Let the lipid membrane be given by $\Gamma$.
The total energy of the membrane is composed of four components:
\begin{align}
E[\Gamma] &= E_k[\Gamma] + E_\gamma[\Gamma] + E_{m,bulk}[\Gamma] + E_{m,rot}[\Gamma], \label{eq:membraneEnergy} \\
E_k[\Gamma] &= \oint_\Gamma\left(\dfrac{1}{2}k_c H^2+k_g K\right)\;dA, \label{eq:bendingEnergy}\\
E_\gamma[\Gamma] &= \oint_\Gamma \gamma\;dA, \label{eq:tensionEnergy}\\
E_{m,bulk}[\Gamma] &= -\dfrac{\chi_\perp d}{2\mu_m}\oint_\Gamma B^2\;dA, \label{eq:magneticBulkEnergy} \\
E_{m,rot}[\Gamma] &= -\dfrac{\Delta\chi d}{2\mu_m}\oint_\Gamma\left(\vec{n}\cdot\vec{B}\right)^2\;dA. \label{eq:magneticRotEnergy}
\end{align}
The first integral, $E_k[\Gamma]$, provides the bending energy associated with the current membrane configuration where
$\kappa_{c}$ is the bending rigidity, $\kappa_{g}$ is the Gaussian bending rigidity, $K$ is the Gaussian curvature,
and $H$ is the total curvature, which equals the sum of the principle curvatures. In this work the Gaussian curvature energy
is ignored as the integral of the Gaussian curvature around any closed surface is a constant.~\cite{Carmo1976}
The second integral, $E_\gamma[\Gamma]$, is the energy associated with a non-uniform
tension, $\gamma$. These two energies are based on the Helfrich model in the absence of spontaneous curvature\cite{Helfrich1973a} and have been used extensively to
model liposome vesicles.~\cite{biben2003,salipante2014,vlahovska2007,Sohn2010}
The final two integrals, $E_{m,bulk}[\Gamma]$ and $E_{m,rot}[\Gamma]$, provide the energy of a lipid membrane in a magnetic field $\vec{B}$
when the outward unit normal to the interface is given by $\vec{n}$ and $\vec{B}\cdot\vec{B}=B^2$.~\cite{Rikken2014,SCHOLZ1984}
The first energy, $E_{m,bulk}$, provides the total (bulk) energy of a membrane with a magnetic susceptibility perpendicular to a
lipid axis given by $\chi_\perp$, a membrane thickness of $d$, and where $\mu_m$ is the magnetic permeability of the membrane.
As lipid molecules are diamagnetic materials, $\chi_\perp<0$.
The second, $E_{m,rot}$, is the magnetic alignment energy, with $\Delta \chi=\chi_\parallel-\chi_\perp$ being
the difference between the magnetic susceptibilities in the parallel and
perpendicular direction for lipid molecules.
For lipid molecules, the
perpendicular magnetic susceptibility is larger (less negative) than the parallel one and thus $\Delta\chi<0$,
although it is possible to change this by adding biphenyl moieties to
the phospholipids.~\cite{Tan2002} This magnetic susceptibility difference drives the lipid molecules to become perpendicular to an applied magnetic field,
which in turn causes the lipid vesicle itself to align with the field.~\cite{BOROSKE1978}
In order to minimize the energy, the membrane will exert a force on the surrounding fluid. These forces are calculated by taking the variation of the appropriate membrane
energy with respect to a change of membrane location.
For a vesicle and neglecting spontaneous curvature the ultimate forms of the bending and tension forces are found to be\cite{seifert1999fluid}
\begin{align}
\vec{\tau}_{k} &= -\kappa_c\left(\frac{1}{2}H^3-2HK + \nabla^2_{s}H\right)\vec{n},\\
\vec{\tau}_{\gamma} &=\gamma H\vec{n}-\nabla_s\gamma.
\end{align}
The surface gradient, $\nabla_s$, and surface Laplacian, $\nabla_s^2$, are defined using the projection operator
$\vec{P}=\vec{I}-\vec{n}\otimes\vec{n}$. More precisely, the surface gradient of a scalar field
is given by $\nabla_s f=\vec{P}\nabla f$ while the surface Laplacian is $\nabla_s^2 f=\nabla_s\cdot\nabla_s f$.
The magnetic force has not been presented in the literature and is derived here. In particular, the method outlined by
Napoli and Vergori is used to determine the variation of the energy.~\cite{Napoli2010} Let a generic energy functional be given by
$\oint_\Gamma w\;dA$, where $w$ is an arbitrary energy functional density per unit area which only a function of
the unit normal $\vec{n}$ and no other surface quantities. The first variation of the energy
with respect to a change of the interface is then given by $\nabla_s\cdot\left(w\vec{P}-\vec{n}\otimes\left(\vec{P}\vec{w}_n\right) \right)$,
where $\vec{w}_n=\partial w/\partial\vec{n}$ is the derivative of the energy density with respect to the unit normal.
Assume that a single lipid species is present. Therefore the material properties $\chi_\perp$, $\Delta \chi$, $d$, and $\mu_m$ are all
constants. First consider the contribution of the bulk energy of the lipid membrane in a magnetic field, $\oint_\Gamma B^2 dA$.
In this case $w=B^2$ and thus $\vec{w}_n=0$. Using the results shown in Appendix \ref{App:Geometric}, this results in
\begin{equation}
\nabla_s\cdot\left(B^2 \vec{P}\right) = \nabla_s B^2 - B^2 H \vec{n}.
\end{equation}
Next consider the rotational energy contribution, $\oint_\Gamma\left(\vec{n}\cdot\vec{B}\right)^2 dA$. From $w=\left(\vec{n}\cdot\vec{B}\right)^2=B_n^2$
the quantity $\vec{w}_n=2\left(\vec{n}\cdot\vec{B}\right)\vec{B}=2 B_n\vec{B}$ is obtained. Thus,
\begin{equation}
\nabla_s\cdot\left(B_n^2\vec{P}-\vec{n}\otimes\left(2B_n\vec{P}\vec{B}\right)\right)
= \nabla_s\cdot\left(B_n^2\vec{P}\right)-2\nabla_s\cdot\left(\vec{n}\otimes\left(B_n\vec{P}\vec{B}\right)\right)
\label{eq:magForceDerivation0}
\end{equation}
where $B_n=\vec{n}\cdot\vec{B}$ is the portion of the magnetic field in the normal direction.
Each of these components are considered in turn. The first component results in
\begin{equation}
\nabla_s\cdot\left(B_n^2\vec{P}\right) = \nabla_sB_n^2 - B_n^2 H \vec{n} = 2B_n\nabla_s B_n-B_n^2 H\vec{n}.
\label{eq:magForceDerivation1}
\end{equation}
The second component is
\begin{align}
\nabla_s\cdot\left(\vec{n}\otimes\left(B_n\vec{P}\vec{B}\right)\right)
&=B_n \left(\nabla_s\vec{n}\right)\vec{P}\vec{B}+\vec{n}\nabla_s\cdot\left(B_n\vec{P}\vec{B}\right) \nonumber \\
&=B_n \left(\nabla_s\vec{n}\right)\vec{B} + \vec{n}\left(\vec{B}\cdot\nabla_s B_n-B_n^2 H+B_n\nabla_s\cdot\vec{B}\right) \nonumber \\
&=B_n \vec{L}\vec{B} + \vec{n}\left(\vec{B}\cdot\nabla_s B_n-B_n^2 H+B_n\nabla_s\cdot\vec{B}\right),
\label{eq:magForceDerivation2}
\end{align}
where the surface gradient of the unit normal, $\nabla_s\vec{n}=\vec{L}$, is called the curvature tensor, or shape operator,
of the interface. It is a symmetric and real matrix which characterizes the curvature of the surface.~\cite{Fried2008, Napoli2010, Napoli2012}
One eigenvalue of $\vec{L}$ is zero and has a corresponding eigenvector in the direction of the unit normal, $\vec{n}$.
As $\vec{L}$ is symmetric and real, it can be decomposed as $\vec{L}=\kappa_t \vec{t}\otimes\vec{t}+\kappa_b \vec{b}\otimes\vec{b}$,
where $\kappa_t$ and $\kappa_b$ are the remaining eigenvalues of $\vec{L}$ with corresponding eigenvectors $\vec{t}$ and $\vec{b}$, respectively.
In this case the eigenvalues are principle curvatures of the interface while the eigenvectors are the principle tangent directions.
Thus, the first part of Eq. \eqref{eq:magForceDerivation2} can be written as
\begin{align}
B_n \vec{L}\vec{B}
= B_n\left(\kappa_t \vec{t}\otimes\vec{t}+\kappa_b \vec{b}\otimes\vec{b}\right)\vec{B}
= B_n\left(\kappa_t B_t \vec{t} + \kappa_b B_b \vec{b}\right),
\label{eq:magForceDerivation3}
\end{align}
where $B_t=\vec{t}\cdot\vec{B}$ and $B_b=\vec{b}\cdot\vec{B}$ are the components of the magnetic field in the principle directions.
Combining the results of Eqs. \eqref{eq:magForceDerivation0}-\eqref{eq:magForceDerivation3} and simplifying results in
\begin{align}
&\nabla_s\cdot\left(B_n^2\vec{P}-\vec{n}\otimes\left(2B_n\vec{P}\vec{B}\right)\right) = \nonumber \\
&2B_n \nabla_s B_n + B_n^2 H \vec{n} - 2 B_n\left(\kappa_t B_t \vec{t} + \kappa_b B_b \vec{b}\right) - 2\vec{n}\left(\vec{B}\cdot\nabla_s B_n\right)-2B_n \vec{n}\nabla_s\cdot\vec{B}.
\end{align}
Using these results the force due to the magnetic field is then
\begin{align}
\vec{\tau}_{m,bulk} = &-\dfrac{\delta E_{m,bulk}}{\delta \Gamma} = \dfrac{\chi_\perp d}{2\mu_m}\left(\nabla_s B^2-B^2 H\vec{n}\right), \label{eq:magBulkForce} \\
\vec{\tau}_{m,rot} = &-\dfrac{\delta E_{m,rot}}{\delta \Gamma} = \dfrac{\Delta\chi d}{2\mu_m}\left( 2B_n\nabla_s B_n + B_n^2 H \vec{n} \right.\nonumber \\
& \left.- 2 B_n\left(\kappa_t B_t \vec{t} + \kappa_b B_b \vec{b}\right) - 2\vec{n}\left(\vec{B}\cdot\nabla_s B_n\right)-2B_n \vec{n}\nabla_s\cdot\vec{B}\right), \label{eq:magRotForce}
\end{align}
recalling that $B^2=\vec{B}\cdot\vec{B}$, $B_n=\vec{n}\cdot\vec{B}$, $B_t=\vec{t}\cdot\vec{B}$, and $B_b=\vec{b}\cdot\vec{B}$.
For general situations, the above expressions works well. When the magnetic field is spatially constant simplifications can be made by
expanding the $\nabla_s B_n=\nabla_s\left(\vec{n}\cdot\vec{B}\right)$ terms:
\begin{align}
\nabla_s B_n=\nabla_s\left(\vec{n}\cdot\vec{B}\right)=\vec{B}\cdot\nabla_s\vec{n}+\vec{n}\cdot\nabla_s\vec{B}=\vec{B}\cdot\vec{L}+\vec{n}\cdot\nabla_s\vec{B}=\vec{B}\cdot\vec{L}=\vec{L}\vec{B},
\end{align}
as $\nabla_s\vec{B}=\left(\nabla\vec{B}\right)\vec{P}=\vec{0}$ and $\vec{L}=\vec{L}^T$.
Beginning with Eq. \eqref{eq:magRotForce} and using the $\vec{L}\vec{B}$ form this results in
\begin{align}
\vec{\tau}_{m,rot}
=& \dfrac{\Delta\chi d}{2\mu_m}\left( 2B_n\nabla_s B_n + B_n^2 H \vec{n} - 2 B_n \vec{L}\vec{B} - 2\vec{n}\left(\vec{B}\cdot\nabla_s B_n\right)-2B_n \vec{n}\nabla_s\cdot\vec{B}\right) \nonumber \\
=& \dfrac{\Delta\chi d}{2\mu_m}\left( 2B_n \vec{L}\vec{B} + B_n^2 H \vec{n} - 2 B_n \vec{L}\vec{B} - 2\vec{n}\left(\vec{B}\cdot\vec{L}\vec{B}\right)\right) \nonumber \\
=& \dfrac{\Delta\chi d}{2\mu_m}\left( B_n^2 H - 2\vec{B}\cdot\vec{L}\vec{B}\right) \vec{n} \nonumber \\
=& \dfrac{\Delta\chi d}{2\mu_m}\left( B_n^2 H - 2\vec{B}\cdot\left(\kappa_t B_t\vec{t}+\kappa_b B_b\vec{b}\right)\right) \vec{n} \nonumber \\
=& \dfrac{\Delta\chi d}{2\mu_m}\left( B_n^2 H - 2 \kappa_t B_t^2-2\kappa_b B_b^2\right) \vec{n},
\end{align}
due to the fact that $\nabla_s\cdot\vec{B}=\vec{P}:\nabla\vec{B}=0$ when $\vec{B}$ is spatially constant.
\subsection{Interface Description}
In this work a level-set formulation is used to describe the location of the vesicle membrane. Let the evolving interface be given as the set of points where
a level-set function $\phi(\vec{x},t)$ is zero: $\Gamma(t)=\left\{\vec{x}:\phi(\vec{x},t)=0\right\}$, where $\vec{x}$ is a position in space and $t$ is time.
Instead of explicitly tracking the location of the interface $\Gamma$ through time the position is implicitly tracked by evolving $\phi$.
Following convention the inner fluid, $\Omega^-$, is given by the region $\phi<0$ while the outer fluid, $\Omega^+$, is given by $\phi>0$.
The entire domain is denoted as $\Omega=\Omega^-\cup\Omega^+$. Using the
level-set geometric quantities are easily obtained. For example, the outward facing unit normal vector and the total curvature (sum of principle
curvatures) is given by
\begin{align}
\vec{n}=&\dfrac{\nabla\phi}{\|\nabla\phi\|}, \\
H =& \nabla\cdot\vec{n}.
\end{align}
It is also possible to use the level set function to define varying material parameters using a single relation.
Consider the determination of the viscosity at any location in the domain. Letting $\mu^-$ be the inner viscosity and
$\mu^+$ be the outer viscosity the viscosity at a point $\vec{x}$ is given by $\mu(\vec{x})=\mu^{-}+(\mu^{+}-\mu^{-})\mathcal{H}(\phi(\vec{x}))$,
where $\mathcal{H}$ is the Heaviside function. Similar expressions hold for other material quantities.
In practice a smoothed version of the Heaviside function is used to ensure numerical stability.~\cite{Towers2009}
Finally, motion of the interface is obtained by advecting the level set function,
\begin{equation}
\dfrac{\partial \phi}{\partial t}+\vec{u}\cdot\nabla\phi=0.
\label{eq:levelsetadvection}
\end{equation}
Details of the numerical implementation will be presented later.
\subsection{Fluid Flow Equations}
Define the bulk fluid hydrodynamic stress tensor in each fluid as
\begin{equation}
\vec{T}_{hd}^{\pm}=-p^{\pm}\vec{I}+\mu^{\pm}(\nabla\vec{u}^{\pm}+\nabla^T\vec{u}^{\pm}) \textnormal{ in } \Omega^{\pm}.
\label{eq:hydrodynamic_stress}
\end{equation}
The forces derived in Section \ref{sec:membraneForces} are balanced by a jump in the fluid stress tensor,
\begin{equation}
\vec{n}\cdot(\vec{T}_{hd}^+-\vec{T}_{hd}^-) = \vec{\tau}_{k} + \vec{\tau}_{\gamma} + \vec{\tau}_{m,bulk} + \vec{\tau}_{m,rot}.
\label{eq:interface_force}
\end{equation}
Note that in general there is a contribution from a jump in the Maxwell stress tensor acting on the interface. In the absence of electric fields and with
matched magnetic fluid properties this contribution is zero and thus is not included.
Using the level set formulation it is possible to write the fluid momentum equations and the interface force balance as a single equation valid over the
entire domain:\cite{Chang1996,Kolahdouz2015b}
\begin{align}
\rho(\phi)\frac{D \vec{u}}{Dt}=&-\nabla p + \nabla\cdot\left(\mu(\phi)\left(\nabla\vec{u}+\nabla^{T}\vec{u}\right)\right) \nonumber \\
&+\delta(\phi)\kappa_{c}\left(\frac{H^3}{2} -2KH+\nabla^2_s H \right)\nabla\phi \nonumber \\
&+\delta(\phi)\|\nabla\phi\|\left(\nabla_s\gamma-\gamma H\vec{n}\right) \nonumber \\
&-\delta(\phi)\|\nabla\phi\|\dfrac{\chi_\perp d}{2\mu_m}\left(\nabla_s B^2-B^2 H\vec{n}\right) \nonumber \\
&-\delta(\phi)\|\nabla\phi\|\dfrac{\Delta\chi d}{2\mu_m}\left( 2B_n\nabla_s B_n + B_n^2 H \vec{n} \right.\nonumber \\
&\left.- 2 B_n\left(\kappa_t B_t \vec{t} + \kappa_b B_b \vec{b}\right)
- 2\vec{n}\left(\vec{B}\cdot\nabla_s B_n\right)-2B_n \vec{n}\nabla_s\cdot\vec{B}\right),
\label{eq:CSF_model}
\end{align}
where the full form of the force, Eqs. \eqref{eq:magBulkForce} and \eqref{eq:magRotForce}, have been used.
The use of the Dirac function $\delta(\phi)$ localizes the contributions from the membrane forces near the $\phi=0$ contour, which is the location of the interface.\cite{Towers2008}
Volume and surface area conservation are provided by ensuring that
\begin{align}
\nabla\cdot\vec{u}&=0 \textrm{ in } \Omega,\\
\nabla_s\cdot\vec{u}&=0 \textrm{ on } \Gamma.
\end{align}
Note that in the fluid formulation, Eq. (\ref{eq:CSF_model}), the tension is used to enforce surface area constraint $\nabla_s\cdot\vec{u}=0$ and is computed
as part of the problem alongside the pressure.
\subsection{Nondimensional Parameters and Equations}
Assume that the density is matched between the inner and outer fluids while the viscosity has a ratio of $\eta=\mu^-/\mu^+$. Each of the forces
acting on the vesicle membrane have an associated time-scale which depends on the material properties. The time-scale associated
with the bending of the membrane is\cite{schwalbe2011}
\begin{equation}
t_k=\dfrac{\mu^+(1+\eta) l_0^3}{\kappa_c},
\end{equation}
where $l_0$ is the characteristic length scale.
The magnetic field introduces two times scales, one associated with each component of the magnetic
field energy, Eqs. \eqref{eq:magneticBulkEnergy} and \eqref{eq:magneticRotEnergy}.
In both case the magnetic forces are compared to the viscous forces.
The first is the time scale of the bulk energy,
\begin{equation}
t_{m,bulk}=\dfrac{\mu^+(1+\eta)l_0 \mu_m}{|\chi_\perp| d B_0^2},
\end{equation}
while the magnetic rotation time scale is
\begin{equation}
t_{m,rot}=\dfrac{\mu^+(1+\eta)l_0 \mu_m}{|\Delta\chi| d B_0^2},
\end{equation}
where $B_0$ is the characteristic magnetic field strength and recalling
that $\mu_m$ is the magnetic permeability of the membrane.
Let the characteristic time be given by $t_0$, which allows for the definition of the characteristic velocity: $u_0=l_0/t_0$.
Define a capillary-like number providing the relative strength of the bending, $\ensuremath\textrm{Ca}=t_k/t_0$,
a magnetic Mason number indicating the strength of the bulk motion, $\ensuremath\textrm{Mn}=\textrm{sgn}(\chi_\perp)t_{m,bulk}/t_0$,
and a magnetic field induced rotational force number, $\ensuremath\textrm{Rm}=\textrm{sgn}(\Delta \chi)t_{m,rot}/t_0$,
while the Reynolds number is given by $\ensuremath\textrm{Re}=\rho u_0 l_0/\mu^+=\rho\;l_0^2/(\mu^+ t_0)$.
The use of $\textrm{sgn}(\chi_\perp)$ and $\textrm{sgn}(\Delta \chi)$
takes into account the fact that the values of $\chi_\perp$ and $\Delta \chi$ can be either positive or negative. Therefore,
the dimensionless parameters $\ensuremath\textrm{Mn}$ and $\ensuremath\textrm{Rm}$ can either be positive or negative, depending on if the membrane is a paramagnetic
or diamagnetic material.
It is then possible to write
the dimensionless fluid equations as
\begin{align}
\frac{D \vec{u}}{Dt}=&-\nabla p + \dfrac{1}{Re}\nabla\cdot\left(\mu(\phi)\left(\nabla\vec{u}+\nabla^{T}\vec{u}\right)\right) \nonumber \\
&+\dfrac{1}{\ensuremath\textrm{Ca}\;\ensuremath\textrm{Re}}\delta(\phi)\left(\frac{H^3}{2} -2KH+\nabla^2_s H \right)\nabla\phi \nonumber \\
&+\delta(\phi)\|\nabla\phi\|\left(\nabla_s\gamma-\gamma H\vec{n}\right)\nonumber \\
&-\dfrac{1}{2\;\ensuremath\textrm{Re}\;\ensuremath\textrm{Mn}} \delta(\phi)\|\nabla\phi\|\left(\nabla_s B^2-B^2 H\vec{n}\right) \nonumber \\
&-\dfrac{1}{2\;\ensuremath\textrm{Re}\;\ensuremath\textrm{Rm}} \delta(\phi)\|\nabla\phi\|\left(2 B_n \nabla_s B_n + B_n^2 H \vec{n} - 2 B_n\left(\kappa_t B_t \vec{t} + \kappa_b B_b \vec{b}\right) \right.\nonumber \\
& \hspace{1cm} \left.- 2\vec{n}\left(\vec{B}\cdot\nabla_s B_n\right)-2B_n \vec{n}\nabla_s\cdot\vec{B}\right),
\label{eq:CSF_nondim_model}
\end{align}
where all quantities are now dimensionless and the viscosity at a point can be calculated using $\mu(\phi)=\eta+(1-\eta)\mathcal{H}(\phi)$.
The energy is normalized by using the bending rigidity, $\kappa_{c}$, as the characteristic energy scale. When writing the normalized total energy,
the contribution from the tension, Eq. \eqref{eq:tensionEnergy}, is not included as it is a co-dimension one parameter used to enforce surface incompressibility.
Therefore, the normalized energy of the system is then
\begin{equation}
E[\Gamma] = \dfrac{1}{2} \oint_\Gamma H^2 dA - \dfrac{1}{2}\dfrac{\ensuremath\textrm{Ca}}{\ensuremath\textrm{Mn}}\oint_\Gamma B^2 dA - \dfrac{1}{2}\dfrac{\ensuremath\textrm{Ca}}{\ensuremath\textrm{Rm}}\oint_\Gamma\left(\vec{n}\cdot\vec{B}\right)^2 dA.
\end{equation}
It is useful to compare the time-scales associated with a typical experiment. Boroske and Helfrich reported the alignment of vesicles in
magnetic fields with a strength of 1.5 T and reported that the magnetic susceptibility difference to be $-3.52\times10^{-8}$ in SI units,
assuming a membrane thickness of 6 nm.~\cite{BOROSKE1978}
Assuming the viscosity is matched and equal to water, $\mu^+=10^{-3}$ Pa s, the membrane magnetic permeability is that of free-space, $\mu_m=4\pi\times10^{-7}$ H/m,
and a characteristic length of 10 $\mu$m (as estimated by the figures in Boroske and Helfrich) the membrane rotation time is $t_{m,rot}\approx 53$ s. It was
reported by Boroske and Helfrich that rotation through an angle of $\pi/2$ took approximately 100 s, which matches well with the characteristic time calculated here.
Using the same values and given that the bending rigidity for the system is approximately $10^{-19}$ J,\cite{Beblik1985} the time-scale associated with bending is 20 s, which
agrees with the experimental results as no deformation of the membrane was observed during rotation.
\section{Numerical Methods}
\label{sec:numericalMethods}
Two numerical methods must be discussed. The first is the advection of the level set field. In this work a new semi-implicit level set Jet scheme is used.~\cite{Velmurugan2016}
In addition to the level set function the gradient of the level set are also tracked, which increases the accuracy of the method.~\cite{Seibold2012}
The extension allows the original level set Jet scheme to be used for stiff advection problems. It is composed of three main steps. First, the level set field
is advanced using a second-order, semi-implicit, semi-Lagrangian update,
\begin{equation}
\dfrac{3\phi^{n+1}-2\phi_d^n+\phi_d^{n-1}}{2\Delta t}=\beta\nabla^2\phi^{n+1}-\beta\nabla^2\hat{\phi},
\end{equation}
where $\beta=0.5$ is a constant, $\phi_d^n$ is the departure value of the level set at time $t^n$,
$\phi_d^{n-1}$ is the departure value of the level set at time $t^{n-1}$, and $\hat{\phi}=2\phi^n-\phi^{n-1}$ is an approximation of the level set value at time
$t^{n+1}$. Once the smooth level set field is obtained the
effect of smoothing is captured by defining a source term,
\begin{equation}
S_{\phi}=\beta\nabla^2\left(\phi^{n+1}-\hat{\phi}\right).
\end{equation}
This advection source term is used to update the level set values on a sub-grid which surrounds all grid points,
\begin{equation}
\dfrac{3\phi^{s,n+1}-2\phi_d^{s,n}+\phi_d^{s,n-1}}{2\Delta t}=S_{\phi}.
\end{equation}
Using these updated sub-grid level set values, $\phi^{s,n+1}$, finite difference approximations are used to calculate the updated gradient field.
It was shown that this method results in an accurate
and stable scheme for the modeling of moving interfaces under stiff advection fields.~\cite{Velmurugan2016}
The fluid field is obtained using a projection-based method.~\cite{Kolahdouz2015b}
The first step is to calculate a tentative field using a semi-implicit, semi-Lagrangian
method:
\begin{align}
\dfrac{3\vec{u}^{\ast}-2\vec{u}_d^n+\vec{u}_d^{n-1}}{2\Delta t}=&-\nabla \hat{p} +\dfrac{1}{\ensuremath\textrm{Re}}\nabla\cdot\left(\mu\left(\nabla \vec{u}^{\ast} +\nabla^T\hat{\vec{u}}\right)\right)\nonumber \\
&+\vec{f}_k^n+\vec{f}_\gamma^n+\vec{f}_{m,bulk}^n+\vec{f}_{m,rot}^n,
\end{align}
where $\vec{f}_k^n$, $\vec{f}_\gamma^n$ and $\vec{f}_{m,bulk}^n+\vec{f}_{m,rot}^n$ are the bending, tension, and magnetic forces
while $\vec{u}_d^n$ and $\vec{u}_d^{n-1}$ are the departure velocities and $\hat{p}=2p^n-p^{n-1}$ is an extrapolation of the pressure
to time $t^{n+1}$.
The next step is to calculate the corrections to the pressure and tension to enforce volume and surface area conservation,
\begin{equation}
\dfrac{3}{2}\dfrac{\vec{u}^{n+1}-\vec{u}^{\ast}}{\Delta t}=-\nabla q +\delta(\phi)\|\nabla\phi\|\left(\nabla_s \xi-\xi H\nabla\phi \right),
\label{eq:projection}
\end{equation}
where $q$ and $\xi$ are the corrections needed for the pressure and tension, respectively. The pressure and tension are computed simultaneously
to enforce both local and global conservation of the enclosed volume and surface area. Complete details of the method are provided in
Kolahdouz and Salac.~\cite{Kolahdouz2015b}
\section{Results}
The experimental results most applicable are those of Boroske and Helfrich.~\cite{BOROSKE1978} As such the results presented here
will be modeled after those experiments. The characteristic time is chosen to be $t_0=1$ s. Assuming a bending rigidity of
$\kappa_c\approx 25 k_B T \approx 10^{-19}$ J,\cite{Vitkova2013} a vesicle radius of $10\;\mu$m, fluid density of 1000 kg/m$^3$, and outer fluid viscosity of $10^{-3}$ Pa s
with matched viscosity ($\eta=1$) the capillary-like number becomes $\ensuremath\textrm{Ca}=20$ while the Reynolds is $\ensuremath\textrm{Re}=10^{-4}$.
Using a magnetic susceptibility difference on the order of $\Delta \chi=-3\times10^{-8}$ in SI units, with a membrane
magnetic permeability of $\mu_m=4\pi\times 10^{-7}$ H/m and membrane thickness of 6nm, the magnetic rotation constant
scales as $\ensuremath\textrm{Rm}\sim -140/B_0^2$. Assuming a magnetic field strength between $B_0=1$ T and $B_0=10$ T, this results in a range of
$-150\lesssim \ensuremath\textrm{Rm} \lesssim -1$.
The magnetic susceptibility perpendicular to the lipid axis, $\chi_\perp$, is not
readily available in the literature, and thus it is not clear what the magnitude should be.
In spatially constant magnetic field the bulk magnetic energy contribution, Eq. \eqref{eq:magneticBulkEnergy}, is constant
for incompressible membranes and thus the bulk magnetic force, Eq. \eqref{eq:magBulkForce}, does not need to be included.
For spatially varying magnetic fields it is assumed that the magnetic susceptibility perpendicular is of the same order as the
magnetic susceptibility difference and therefore the magnetic Mason number will be taken to be $-100\lesssim \ensuremath\textrm{Mn} \lesssim -1$
Due to the long simulation times, and to facilitate a larger number of trials,
the results will be in the two-dimensional regime, resulting in zero Gaussian curvature: $K=0$.
Unless otherwise stated, the computational domain is a square spanning $[-6.4,6.4]^2$ using a $257^2$ grid and
periodic boundary conditions while the time step is fixed at $\Delta t=0.1 h$, where $h=12.8/256=0.05$ is the grid spacing.
The choice of this domain size and time step is justified in Sec. \ref{sec:domainParameters}.
Vesicles are be characterized by several parameters. Specifically, the viscosity ratio $\eta=\mu^-/\mu^+$, the
inclination angle $\theta^v$, the deformation parameter $D$, and the reduced area $\nu$.
Inclination angles are determined by calculating the eigenvalues and eigenvectors of
the vesicle's inertia tensor about its center of mass. The eigenvector corresponding to
the larger of the two eigenvalues provides the direction of the long axis of a vesicle. The angle
between the eigenvector associated with the long axis of the vesicle and the $x$-axis is denoted as the inclination angle.
The deformation parameter is given by $D=(a-b)/(a+b)$, where $a$ and $b$ are the long and short axes of an ellipse with
the same inertia tensor as the vesicle.~\cite{Salac2012,RAMANUJAN1998}
The vesicle reduced area indicates how deflated a vesicle is compared to a circle with the same interface length, and is given by $\nu=4 A\pi/L^2$, where
$A$ and $L$ are the enclosed area and interface length, respectively. A value of $\nu=0.5$ indicates that the enclosed area is one-half of a circle with the same
interface length while $\nu=1$ denotes a circle.
All simulations begin with an ellipse having an interfacial length of $2\pi$ and a reduced area of $\nu=0.71$.
This reduced area was estimated from Boroske and Helfrich, Fig. 1.~\cite{BOROSKE1978}
The vesicle is then allowed to evolve in the absence of a magnetic field to obtain a shape near the bending energy minimum.~\cite{Salac2011} This
shape is then used as the initial condition for the magnetically driven results.
The initial orientation of all vesicles is vertical, which is denoted as having an inclination angle of $\theta^v=\pi/2$, see Fig. \ref{fig:schematic}.
It is also assumed throughout that the viscosity is matched, $\eta=1$.
\subsection{Direct comparison with Boroske and Helfrich}
To provide the reader a better understanding of the vesicle dynamics, a direct comparison with the results of Boroske and Helfrich is performed.~\cite{BOROSKE1978}
Using Fig. 1 from that manuscript, it was estimated that the angle between the vesicle and the applied magnetic
field is $0.455\pi$. In the simulation,
the vesicle is initially aligned with the vertical axis and the magnetic field has an angle of $\theta^B=0.045\pi$, which matches
the conditions of the experiment. As the magnetic field is spatially constant the bulk magnetic energy is ignored while the dimensionless magnetic
rotation constant is set to $\ensuremath\textrm{Rm}=-7.5$. The inclination angle up to a time of 200 is shown in Fig. \ref{fig:boroskeAngle} while the computationally
derived vesicle shapes are compared to the experimental result in Fig. \ref{fig:boroske}.
\begin{figure}[]
\centering
\includegraphics[width=8cm]{Rm0075}
\caption{The inclination angle as a function of time for a vesicle with reduced area of $\nu=0.71$ in a magnetic field at an angle of $0.045\pi$ with
a rotation constant of $\ensuremath\textrm{Rm}=-7.5$. The dots indicate the angles determined from Fig. 1 of Boroske and Helfrich,\cite{BOROSKE1978} after an appropriate rotation is done
to take into account the different initial angles.}
\label{fig:boroskeAngle}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=8cm]{BoroskeHelfrich_Fig1}
\caption{Comparison between experimental results of Boroske and Helfrich and the simulation. Due to the different initial orientation, the simulation
results are first flipped about the horizontal axis and then rotated $0.045\pi$ counter-clockwise.
Reprinted from Biophysical Journal, Vol 24 (3), Boroske and Helfrich, ``Magnetic anisotropy of egg lecithin membranes", Pages 863-868., December 1978, with permission from Elsevier.}
\label{fig:boroske}
\end{figure}
The computational results match very well with the experimental results. Assuming a membrane thickness of $d=6$ nm, the properties of water, and an
applied magnetic field strength of $B_0=1.5$ T, and using the value of $\ensuremath\textrm{Rm}=-7.5$, the magnetic susceptibility difference is calculated
to be $\Delta \chi=-2.48\times 10^{-7}$. While this value is larger than that estimated by Boroske and Helfrich, it is within other experimentally
determined values.~\cite{BRAGANZA1984}
\subsection{Verification of domain parameters}
\label{sec:domainParameters}
To verify the choice of domain size, grid size, and time step a systematic investigation is performed by varying each simulation parameter
individually.
The magnetic field is spatially constant and fixed at an angle of $\theta^B=0.045\pi$. Three magnetic rotation constants used are $\ensuremath\textrm{Rm}=-1$,
$\ensuremath\textrm{Rm}=-10$, and $\ensuremath\textrm{Rm}=-100$. As the magnetic field is spatially constant, the bulk magnetic field contribution is neglected.
First consider the influence of the grid size on the results. Using a $[-6.4,6.4]^2$ domain, grid sizes ranging from $129^2$ to $513^2$
are used. In all cases the time step is set to $\Delta t = 0.1h$, where $h$ is the grid spacing. The results shown in Fig. \ref{fig:domainN}
indicate that a grid size of $129^2$ is not sufficient. This shouldn't be surprising as with this grid spacing only approximately 8 grid
points are used to describe the vesicle at it's narrowest point. The difference in the results using more than a grid size of $257^2$ are not noticeable
for any of the three magnetic rotation constants, which justifies that particular choice.
\begin{figure}
\centering
\includegraphics[width=8cm]{N}
\caption{The inclination angle versus time for magnetic rotation strengths of $\ensuremath\textrm{Rm}=-1$, $\ensuremath\textrm{Rm}=-10$, and $\ensuremath\textrm{Rm}=-100$ and grid sizes ranging from $129^2$ to $513^3$.
All results use a domain size of $[-6.4,6.4]^2$ with a time step of $\Delta t=0.1h$, where $h$ is the grid spacing.
No change in the results are seen past a grid size of $257^2$.}
\label{fig:domainN}
\end{figure}
Next consider the influence of domain size on the rotation dynamics. Using a constant grid spacing of $h=0.05$ and time step of $\Delta t=0.1h$,
various domain sizes from $[-2.4,2.4]^2$ to $[-8.0,8.0]^2$ are considered, see Fig. \ref{fig:domainL}. Clearly, boundary effects are
present in the smallest domains, particularly when $\ensuremath\textrm{Rm}=-10$. Once the domain size reaches $[-6.4, 6.4]^2$, only small differences are observed.
\begin{figure}
\centering
\includegraphics[width=8cm]{L}
\caption{The inclination angle versus time for magnetic rotation strengths of $\ensuremath\textrm{Rm}=-1$, $\ensuremath\textrm{Rm}=-10$, and $\ensuremath\textrm{Rm}=-100$ and domain sizes ranging from
$[-2.4,2.4]^2$ to $[-8.0,8.0]^2$. The number of grid points is adjusted so that a constant grid spacing of $h=0.05$ and
constant time step $\Delta =0.1h$ is used for each simulation.
No change in the results are seen past a domain size of $[-6.4,6.4]^2$.}
\label{fig:domainL}
\end{figure}
Finally consider the influence of the time step on the rotation dynamics. Using a $[-6.4,6.4]^2$ domain with $257^2$ grid points, various time
steps from $\Delta t=0.02h$ to $\Delta t=0.5h$ are considered. Note that using time steps of $\Delta t=h$ proved unstable.
There are almost no differences using time steps smaller than $\Delta t=0.1h$, and thus that is the time step chosen for further results.
\begin{figure}
\centering
\includegraphics[width=8cm]{DT}
\caption{
The inclination angle versus time for magnetic rotation strengths of $\ensuremath\textrm{Rm}=-1$, $\ensuremath\textrm{Rm}=-10$, and $\ensuremath\textrm{Rm}=-100$ and time steps ranging from
$\Delta t=0.02 h$ to $\Delta t=0.5h$. The domain is fixed at $[-6.4,6.4]^2$ while the size of the domain is $257^2$.
No change in the results are seen past a time step of $\Delta t=0.1h$.
}
\label{fig:domainDT}
\end{figure}
\subsection{Influence of $\ensuremath\textrm{Rm}$}
\label{sec:RM}
The influence of the magnetic rotation force is explored by varying $\ensuremath\textrm{Rm}$ within the range from 1 to 100
up to a time of $t=200$. The resulting inclination angle over time is shown in Fig. \ref{fig:RmResults_Angle},
while the amount of time needed to rotate through an angle of $0.05\pi$, $0.25\pi$, and $0.4\pi$
is shown in Fig. \ref{fig:RmResults_TOC}. There are several points to be made. First,
the equilibrium angle of the vesicle, given enough time, will match that of the applied magnetic field.
Second, the amount of time that is required to rotate through a particular angle is linearly dependent on the
$\ensuremath\textrm{Rm}$ value. It should be noted that due to the definition of $\ensuremath\textrm{Rm}$, this is related to the quadratic of the magnetic
field strength, i.e. a 2-fold increase in the magnetic field results in a 4-fold decrease of the $\ensuremath\textrm{Rm}$ parameter.
Therefore, increasing the magnetic field strength by a factor of two reduces the amount of time needed to rotate by a factor
of four.
\begin{figure}
\centering
\subfigure[Inclination angle over time for various values of $\ensuremath\textrm{Rm}$, indicated by the numbers.]{
\includegraphics[width=8cm]{RM_Angle}
\label{fig:RmResults_Angle}
}
\subfigure[Time it takes to rotate through an angle of $0.05\pi$, $0.25\pi$, and $0.4\pi$.]{
\includegraphics[width=8cm]{RM_TimeToCross}
\label{fig:RmResults_TOC}
}
\caption{The influence of the magnetic field-induced rotational force, $\ensuremath\textrm{Rm}$, on the inclination angle.
As $\ensuremath\textrm{Rm}$ increases it takes additional time to align with the magnetic field.}
\label{fig:RmResults}
\end{figure}
An investigation of the energy for three characteristic rotation strengths, $\ensuremath\textrm{Rm}=-1$, $\ensuremath\textrm{Rm}=-10$, and $\ensuremath\textrm{Rm}=-100$, is shown in Fig. \ref{fig:RM_energy}.
As expected, the total energy decreases as the vesicle becomes aligned with the magnetic field. The overall rotation rate is directly correlated
to the initial rotation energy, as a higher initial magnetic rotation energy correlates to faster rotation time. It is interesting to note that when
the magnetic rotation strength is strong, $\ensuremath\textrm{Rm}=-1$, the vesicle membrane can not respond quickly to changes in bending energy and thus the
bending energy contribution increases, Fig. \ref{fig:RM_energy_A}. This is in contrast to the weaker rotation forces shown in
Figs. \ref{fig:RM_energy_B} and \ref{fig:RM_energy_C}, where both the rotation and bending energy are strictly decreasing.
\begin{figure}
\centering
\subfigure[$\ensuremath\textrm{Rm}=-1$]{\label{fig:RM_energy_A} \includegraphics[width=8cm]{Rm001_Energy}}
\subfigure[$\ensuremath\textrm{Rm}=-10$]{\label{fig:RM_energy_B} \includegraphics[width=8cm]{Rm010_Energy}}
\subfigure[$\ensuremath\textrm{Rm}=-100$]{\label{fig:RM_energy_C} \includegraphics[width=8cm]{Rm100_Energy}}
\caption{The bending, rotation, and total energy for rotation strengths of $\ensuremath\textrm{Rm}=-1$, $\ensuremath\textrm{Rm}=-10$, and $\ensuremath\textrm{Rm}=-100$.}
\label{fig:RM_energy}
\end{figure}
To further explore the influence of $\ensuremath\textrm{Rm}$ on the vesicle shape, the deformation parameter for the three characteristic
$\ensuremath\textrm{Rm}$ values is shown in Fig. \ref{fig:RM_DeformationParameter}. It is clearly observed that the strong rotation force
given by $\ensuremath\textrm{Rm}=-1$ causes larger deformations than the $\ensuremath\textrm{Rm}=-10$ and $\ensuremath\textrm{Rm}=-100$ cases. The shape of the vesicle using $\ensuremath\textrm{Rm}=-1$, as shown in
Fig. \ref{fig:RM001_Shapes}, can be compared to that shown in Fig. \ref{fig:boroske}, and it is clear that larger deformation are observed
before the vesicle flattens out.
\begin{figure}
\centering
\includegraphics[width=6cm]{Rm_DeformationParameter}
\caption{
The deformation parameter for $\ensuremath\textrm{Rm}=-1$, $\ensuremath\textrm{Rm}=-10$, and $\ensuremath\textrm{Rm}=-100$. Strong magnetic field effects induce larger shape deformations.
The circles on the $\ensuremath\textrm{Rm}=-1$ correspond to the interfaces shown in Fig. \ref{fig:RM001_Shapes}. The shapes for $\ensuremath\textrm{Rm}=-10$ and $\ensuremath\textrm{Rm}=-100$
do not look qualitatively different from that shown in Fig. \ref{fig:boroske}.
}
\label{fig:RM_DeformationParameter}
\end{figure}
\begin{figure}
\centering
\subfigure[$t=0$]{\includegraphics[width=4cm]{Rm001_00}}
\vspace{0.5cm}
\subfigure[$t=3$]{\includegraphics[width=4cm]{Rm001_03}} \\
\subfigure[$t=6$]{\includegraphics[width=4cm]{Rm001_06}}
\vspace{0.5cm}
\subfigure[$t=9$]{\includegraphics[width=4cm]{Rm001_09}}
\caption{The shape of the vesicle at various times using $\ensuremath\textrm{Rm}=-1$.}
\label{fig:RM001_Shapes}
\end{figure}
It should be noted that the initial angle between the long-axis of the vesicle and the applied magnetic field is less than $\pi/4$. If the
angle is equal to $\pi/4$, then the mechanism of alignment is no longer rotation, but large-scale deformation of the interface.
This can be seen in Fig. \ref{fig:Horizontal}, where the vesicle is placed in a magnetic field aligned with the $x$-axis and
the rotation/alignment strength is $\ensuremath\textrm{Rm}=-1$. Up until approximately $t=30$, the major axis of the vesicle is aligned with the $y$-axis.
The vesicle undergoes large deformations, as demonstrated by the decrease in the deformation parameter. After $t=30$, the major axis is
aligned with the magnetic field along the $x$-axis and the vesicle begins to elongate to reduce both the bending and magnetic energies.
It should be noted that the final deformation parameter for this case, $D\approx 0.46$, is similar to that shown in Fig. \ref{fig:RM_DeformationParameter},
despite the difference in the magnetic angle.
\begin{figure}
\begin{center}
\subfigure[Interface Location at the times indicated]{\label{fig:Horizontal_Interface} \includegraphics[width=6cm]{Horizontal}}
\subfigure[Angle and Deformation Parameter]{\label{fig:Horizontal_AD} \includegraphics[width=8cm]{Horizontal_Angle_Deformation}}
\end{center}
\caption{Sample interface locations, the inclination angle, and deformation parameter for a vesicle in a magnetic field aligned
along the $x$-direction. In this case $\ensuremath\textrm{Rm}=-1$.}
\label{fig:Horizontal}
\end{figure}
\subsection{Spatially varying magnetic field}
\label{sec:VaryingMagneticField}
Next consider the influence of a spatially varying magnetic field. In this case, the full magnetic energy contribution must be considered, and thus both
$\ensuremath\textrm{Rm}$ and $\ensuremath\textrm{Mn}$ will be varied. To construct the variable magnetic field, the vesicle is placed inside a domain spanning $[-3.2,3.2]^2$ using a
grid size of $129^2$ so that $h=0.05$. In this case wall boundary conditions are assumed. To induce the magnetic field, two infinitely long
wires are placed at the locations $(-3.2,0)$ and $(3.2,0)$. Each of these wires has a current of magnitude $I_0$ in the vertical $z-$direction
and produces a magnetic field given by
\begin{equation}
\vec{B} = \dfrac{B_0}{2\pi}\left(\dfrac{x-a}{((x-a)^2+(y-b)^2},-\dfrac{y-b}{((x-a)^2+(y-b)^2} \right)
\end{equation}
where the $B_0=\mu_0 I_0$ is the strength of the induced magnetic field surrounding a wire at $(a,b)$.~\cite{Tzirtzilakis2005} Due to the linearity of the magnetic field,
the total magnetic field is simply the summation of that induced by both wires.
A vesicle with a reduced area of $\nu=0.71$ and matched viscosity, $\eta=1$, is then centered at $(-2,0)$. It is expected that the vesicle will
migrate towards the center of the domain, which is the location of lowest magnetic field strength.
An example of the magnetic field and initial vesicle location is given in Fig. \ref{fig:magneticField}, which shows both the magnetic field lines and the intensity of the magnetic field.
\begin{figure}
\centering
\includegraphics[width=6cm]{MagneticField}
\caption{The magnetic field lines, initial location of the vesicle, and representation of the magnetic field strength (color online). The magnetic field is strongest
at the center of the two current carrying wires located at $(-3.2,0)$ and $(3.2,0)$, indicated by the color red, and quickly decays towards the center of the domain, indicated
by blue.}
\label{fig:magneticField}
\end{figure}
The location of the interface at times of $t=0$, $t=100$, and $t=200$ for $\ensuremath\textrm{Mn}$ and $\ensuremath\textrm{Rm}$ values between 1 and 100 is shown in Fig. \ref{fig:variableInterfaces}.
In all cases the interface migrates towards the center of the domain. The rate of this migration and the overall deformation of the interface strongly depends
on both the $\ensuremath\textrm{Mn}$ and $\ensuremath\textrm{Rm}$ parameters. In general, as the strength of the alignment and bulk magnetic effects increases, the rate of of migration also
increases. It should also be observed that for stronger rotational strengths, denoted by lower $\ensuremath\textrm{Rm}$ values, the vesicle tends to align with the local magnetic
field. As the underlying local magnetic field is close to circular, the interface adopts this configuration.
\begin{figure*}
\centering
\includegraphics[width=12cm]{Variable}
\caption{The location of the interface at times $t=0$, $t=100$, and $t=200$ for a spatially variable magnetic field using various values of $\ensuremath\textrm{Mn}$ and $\ensuremath\textrm{Rm}$.
The results show a portion of the entire domain, which spans $[-3.2,3.2]^2$. The magnetic field arises due to current carrying wires embedded at locations
$(-3.2,0)$ and $(3.2,0)$.}
\label{fig:variableInterfaces}
\end{figure*}
The location of the $x$-centroid and the deformation parameter of the vesicle when exposed to this spatially varying magnetic field is shown in
Fig. \ref{fig:Variable_Results}. It is clear that the fastest migration is achieved with small values of $\ensuremath\textrm{Rm}$ and $\ensuremath\textrm{Mn}$. Even in situations
where the bulk-magnetic field effects are small, such as when $\ensuremath\textrm{Mn}=-100$, migration can occur due to the alignment energy. This is due to the fact
that Eq. \eqref{eq:magneticRotEnergy} can be decreased by not only aligning the interface with the magnetic field, but also by
pushing the interface towards regions of lower magnetic field strength.
The deformation parameter results mimic those seen in the spatially constant results. As the alignment strength increases, the vesicle becomes more deformed.
As the value of $\ensuremath\textrm{Mn}$ increases, this deformed state persists longer. This is due to the fact that it takes longer for the vesicle to migrate towards the
center of the domain when $\ensuremath\textrm{Mn}$ is large. This results in the vesicle remaining closer to the stronger and more compact magnetic field centered
at $(-3.2,0)$.
\begin{figure}
\centering
\subfigure[$x$-centroid]{
\includegraphics[width=8cm]{Variable_CenterX}
}
\subfigure[Deformation Parameter]{
\includegraphics[width=8cm]{Variable_DeformationParameter}
}
\caption{The evolution of the $x$-centroid and deformation parameter of the vesicle over time for various values of $\ensuremath\textrm{Mn}$ and $\ensuremath\textrm{Rm}$.
The combinations shown here correspond to those shown in Fig. \ref{fig:variableInterfaces}. The legend is common to both figures.}
\label{fig:Variable_Results}
\end{figure}
\begin{figure}
\centering
\subfigure[$\ensuremath\textrm{Rm}=-1$]{\includegraphics[width=8cm]{Mn010_Rm001_Energy}}
\subfigure[$\ensuremath\textrm{Rm}=-10$]{\includegraphics[width=8cm]{Mn010_Rm010_Energy}}
\subfigure[$\ensuremath\textrm{Rm}=-100$]{\includegraphics[width=8cm]{Mn010_Rm100_Energy}}
\caption{The bending, bulk, rotation, and total energy for three different rotation strengths. The results assume $\ensuremath\textrm{Mn}=-10$.}
\label{fig:Variable_energy}
\end{figure}
Finally, the energy of the system over time for a bulk constant of $\ensuremath\textrm{Mn}=-10$ and three alignment strengths, $\ensuremath\textrm{Rm}=-1$, $\ensuremath\textrm{Rm}=-10$, and $\ensuremath\textrm{Rm}=-100$ is shown
in Fig. \ref{fig:Variable_energy}. As in Sec. \ref{sec:RM}, the magnetic energies are strictly decreasing over time. For the cases
of strong magnetic field effects, particularly for $\ensuremath\textrm{Mn}=-10$ and $\ensuremath\textrm{Rm}=-1$, the bending energy increases above the initial value, and remains
elevated throughout the simulation. It should be expected that as the vesicle moves towards the center of the domain, where the magnetic
field is weakest, the bending energy should have a larger influence.
\section{Conclusion}
In this work a numerical model of vesicles in magnetic fields is presented. Based on the energy of the membrane, the interface forces
due to magnetic rotation/alignment and the bulk magnetic energy are derived. These magnetic interface forces, in addition to the bending
and tension forces of a vesicle, are used in conjunction with a level set description of the interface and a projection method for the fluid field
to investigate the dynamics of a two-dimensional vesicle. The simulation is compared to the experimental results of Boroske and Helfrich,
and good agreement is achieved. A systematic investigation of the influence of the rotation/alignment parameter, $\ensuremath\textrm{Rm}$,
on the vesicle membrane is performed for spatially constant magnetic field. In general, there is a linear relationship between
$\ensuremath\textrm{Rm}$ and the amount of time it takes a vesicle to rotate through a particular angle. It was also demonstrated that if the angle
between a vesicle and the magnetic field is $\pi/4$, then the alignment is not done through rotation, but by bulk deformation of the membrane.
The movement of a vesicle in a spatially-varying magnetic field was also considered by placing a vesicle between two current-carrying wires.
This magnetic field induced linear motion of the vesicle, with the rate of migration dependent on both the alignment parameter $\ensuremath\textrm{Rm}$
and the bulk magnetic field parameter $\ensuremath\textrm{Mn}$. The particular nature of the underlying magnetic field induced deformations of the membrane,
with the magnitude of these deformations depending on the particular parameter set.
The use of magnetic fields opens up new possibilities for characterization and processing of not only liposome vesicle, but also other soft-matter multiphase systems such
as polymer vesicles or biological cells. For example, it is imagined that using the experimental equivalent to the simulations shown here
it could be possible to determine material properties
such as the magnetic susceptibilities of the membrane molecules. This knowledge could then be used to design processing techniques, possibly in conjuncture with electric fields,
to precisely control the dynamics of vesicles. Future work will explore these possibilities.
\section*{Acknowledgments}
This work has been supported by the National Science Foundation through the Division of Chemical, Bioengineering, Environmental, and Transport Systems Grant \#1253739.
\begin{appendices}
\section{Calculus on Surfaces}
\label{App:Geometric}
One issue with derivatives on surfaces is that operations require information of not only how
a function varies on the interface, but also how the interface itself varies. For this reason,
some standard vector calculus identities may not hold. In this section the surface vector calculus
identities used to calculate the magnetic field force are derived.
Let the interface be orientable with an outward unit normal $\vec{n}$.
Without loss of generality, it is assumed that the interface is described as the zero contour of
a function $\Psi$ such that $\Psi$ is the solution to the Eikonal equation, $|\nabla\Psi|=1$
within a distance of $r$ to the interface, where $r$ depends on the curvature of the interface.
With this assumption the normal is simply $\vec{n}=\nabla\Psi$. As the normal is now defined in a small region
surrounding the interface, quantities such as the gradient of the unit normal, $\nabla\vec{n}$, are
well-defined near the interface.
The projection operator is
given by $\vec{P}=\vec{I}-\vec{n}\otimes\vec{n}$, or in component form $P_{ij}=\delta_{ij}-n_i n_j$, where $\delta_{ij}$
is the Kronecker delta function. In this work, indices $i$ and $j$ are free indices and
while $p$, $q$, and $r$ are dummy indices. The projection operator is symmetric, $\vec{P}=\vec{P}^T$, and idempotent,
\begin{align}
\left[\vec{P}\vec{P}\right]_{ij}
& =P_{ip}P_{pj} \nonumber \\
& =\left(\delta_{ip}-n_i n_p\right)\left(\delta_{pj}-n_p n_j\right) \nonumber \\
& =\delta_{ip}\delta_{pj}-n_i n_p \delta_{pj} - n_p n_j \delta_{ip} + n_i n_p n_p n_j \nonumber \\
& = \delta_{ij} - n_i n_j - n_i n_j + n_i n_j \nonumber \\
& = \delta_{ij} - n_i n_j = \left[\vec{P}\right]_{ij},
\end{align}
where $\left[\vec{v}\right]_i$ is the $i^{th}$ component of a vector $\vec{v}$, $\left[\vec{A}\right]_{ij}$ is the
$i,j$ compnent of a tensor $\vec{A}$, and
repeated indices indicate summation.
The generalized surface gradient function can be written as
$\nabla_s\vec{A}=\left(\nabla\vec{A}\right)\vec{P}$, where $\vec{A}$ can be either a scalar, vector, or tensor field.~\cite{Fried2008,Gurtin1975,Napoli2012}
For example, the surface gradient of a scalar field $a$ in component form would be written as
\begin{align}
\left[\nabla_s a\right]_i
=\left[\left(\nabla a\right)\vec{P}\right]_i
=\dfrac{\partial a}{\partial x_p}P_{pi},
\end{align}
The surface gradient of a scalar field $a$ squared is
\begin{align}
\left[\nabla_s a^2\right]_i
&=\left[\left(\nabla a^2\right)\vec{P}\right]_i=\dfrac{\partial a^2}{\partial x_p}P_{pi} \nonumber \\
&=2a\dfrac{\partial a}{\partial x_p}P_{pi}=\left[2a \nabla_s a\right]_i.
\end{align}
For a vector field $\vec{v}$ the surface gradient would be
\begin{align}
\left[\nabla_s \vec{v}\right]_{ij}
=\left[\left(\nabla \vec{v}\right)\vec{P}\right]_{ij}
=\dfrac{\partial v_i}{\partial x_p}P_{pj}.
\end{align}
The surface gradient of a vector dot product is
\begin{align}
\left[\nabla_s\left(\vec{v}\cdot\vec{w}\right)\right]_i&=\dfrac{\partial\left(v_p w_p\right)}{\partial x_q}P_{qi}\nonumber \\
&= w_p \dfrac{\partial v_p}{\partial x_q}P_{qi} + v_p\dfrac{\partial w_p}{x_q}P_{qi} \nonumber \\
&= \left[\vec{w}\cdot\nabla_s\vec{v}+\vec{v}\cdot\nabla_s\vec{w}\right]_i.
\end{align}
The surface divergence of any vector $\vec{v}$ can be written as
$\nabla_s\cdot\vec{v}=\textnormal{tr}\nabla_s\vec{v}=\vec{P}:\nabla\vec{v}$.~\cite{Fried2008}
In component form this is written as
\begin{align}
\left[\nabla_s\cdot\vec{v}\right]
=\left[\vec{P}:\nabla\vec{v}\right]
=P_{pq}\dfrac{\partial v_p}{\partial x_q}.
\end{align}
The surface divergence of a tensor field $\vec{A}$ is defined as\cite{Fried2008}
\begin{align}
\left[\nabla_s\cdot\vec{A}\right]_{i}=\left[\left(\nabla \vec{A}\right)\vec{P}\right]_i=\dfrac{\partial A_{ip}}{\partial x_q}P_{qp}.
\end{align}
The surface divergence of the projection operator is
given by
\begin{align}
\left[\nabla_s\cdot\vec{P}\right]_i
& =\left[\left(\nabla \vec{P}\right)\vec{P}\right]_i
=\dfrac{\partial P_{ip}}{\partial x_q}P_{qp} \nonumber \\
&= \dfrac{\partial}{\partial x_q}\left(\delta_{ip}-n_i n_p\right)P_{qp} \nonumber \\
& =-\dfrac{\partial n_i}{\partial x_q}P_{qp}n_p -n_i\dfrac{\partial n_p}{\partial x_q}P_{qp} \nonumber \\
& =\left[-\left(\nabla\vec{n}\right)\vec{P}\vec{n}-\vec{n}\nabla_s\cdot\vec{n}\right]_i
=\left[-H \vec{n}\right]_i
\end{align}
due to the definition of total curvature, $H=\nabla_s\cdot\vec{n}$, and the fact that $\vec{P}\vec{n}=0$:
\begin{align}
\left[\vec{P}\vec{n}\right]_i
& =P_{ip}n_p
=\left(\delta_{ip}-n_i n_p\right)n_p \nonumber \\
& =\delta_{ip}n_p - n_i n_p n_p=n_i-n_i
=\left[0\right]_i.
\end{align}
Let $a$ be a scalar field. The surface divergence of this scalar field times the projection operator is
\begin{align}
\left[\nabla_s\cdot\left(a\vec{P}\right)\right]_i
&=\left[\left(\nabla\left(a\vec{P}\right)\right)\vec{P}\right]_i
=\dfrac{\partial \left(a P_{ip}\right)}{\partial x_q}P_{qp} \nonumber \\
& =\dfrac{\partial a}{\partial x_q}P_{ip}P_{qp}+a\dfrac{\partial{P_{ip}}}{\partial x_q}P_{qp} \nonumber \\
&=\dfrac{\partial a}{\partial x_q}P_{qp}P_{pi}+a\dfrac{\partial{P_{ip}}}{\partial x_q}P_{qp} \nonumber \\
&=\dfrac{\partial a}{\partial x_q}P_{qi}+a\dfrac{\partial{P_{ip}}}{\partial x_q}P_{qp}
=\left[\nabla_s a-a H\vec{n}\right]_i.
\end{align}
Next, consider the surface divergence of the tensor (outer) product of the unit normal $\vec{n}$ and any vector $\vec{v}$:
\begin{align}
\left[\nabla_s\cdot\left(\vec{n}\otimes\vec{v}\right)\right]_i
& =\left[\left(\nabla\left(\vec{n}\otimes\vec{v}\right)\right)\vec{P}\right]_i
=\dfrac{\partial\left(n_i v_p\right)}{\partial x_q}P_{qp} \nonumber \\
& = \dfrac{\partial n_i}{\partial x_q}P_{qp}v_p +n_i \dfrac{\partial v_p}{\partial x_q}P_{pq}
= \left[\left(\nabla_s\vec{n}\right)\vec{v} + \vec{n}\nabla_s\cdot\vec{v}\right]_i
\end{align}
Finally, consider the surface divergence of a scalar, the projection operator, and a vector,
\begin{align}
\left[\nabla_s\cdot\left(a\vec{P}\vec{v}\right)\right]
&=\left[\vec{P}:\nabla\left(a\vec{P}\vec{v}\right)\right]
=P_{pq}\dfrac{\partial\left(a P_{pr} v_r\right)}{\partial x_q} \nonumber \\
& =P_{pq}\left(\dfrac{\partial a}{\partial x_q}P_{pr}v_r + a\dfrac{\partial P_{pr}}{\partial x_q}v_r +a P_{pr}\dfrac{\partial v_r}{\partial x_q}\right) \nonumber \\
&=\dfrac{\partial a}{\partial x_q}P_{qr}v_r + a \dfrac{\partial P_{pr}}{\partial x_q} P_{pq} v_r + a \dfrac{\partial v_r}{\partial x_q} P_{qr} \nonumber \\
& =\dfrac{\partial a}{\partial x_q}P_{qr}v_r + a \dfrac{\partial P_{rp}}{\partial x_q} P_{qp} v_r + a \dfrac{\partial v_r}{\partial x_q} P_{rq} \nonumber \\
&=\left[\vec{v}\cdot\nabla_s a+a\left(\nabla_s\cdot\vec{P}\right)\cdot\vec{v}+a\nabla_s\cdot\vec{v}\right] \nonumber \\
& =\left[\vec{v}\cdot\nabla_s a-aH \vec{n}\cdot\vec{v}+a\nabla_s\cdot\vec{v}\right]
\end{align}
\end{appendices}
\bibliographystyle{rsc}
|
\section{Introduction}\label{sec:Intro}
Let $X$ be a random vector on $\mathbb{R}^d$ with density $f$. The entropy associated to $X$ is defined by
\begin{align}
h(X) = -\int_{\mathbb{R}^d} f \log f,
\end{align}
provided the integral exists. The non-Gaussianness of $X$, denoted by $D(X)$, is given by
\begin{align}
D(X) = h(G^X) - h(X),
\end{align}
where $G^X$ denotes a Gaussian random vector with the same covariance as $X$. Evidently, $D(X)$ is the relative entropy of $X$ with respect to $G^X$, and is therefore nonnegative. Moreover, $D(X)=0$ if and only if $X$ is Gaussian.
Our main result may be informally stated as follows: Let $X, X_*$ be IID radially symmetric random vectors on $\mathbb{R}^d$, $d\geq 2$, with sufficiently regular density $f$. For any $\varepsilon > 0$
\begin{align}
h(\tfrac{1}{\sqrt{2}}(X+X_*) ) - h(X)
&\geq C_{\varepsilon}(X) D(X) ^{1+\varepsilon},\label{mainBound}
\end{align}
where $C_{\varepsilon}(X)$ is an explicit function depending on $\varepsilon$, the regularity of $f$, and a finite number of moments of $X$. In particular, if $\mathbb{E}|X|^{4+\delta}<\infty$ for some $\delta>0$, then $C_{2/\delta}(X)>0$. A precise statement can be found in Section \ref{sec:MainResults}, along with an analogous result for Fisher information and a related estimate that imposes no regularity conditions. In interpreting the result, it is important to note that, although a radially symmetric density $f:\mathbb{R}^d \to [0,\infty)$ has a one-dimensional parameterization, the convolution $f \ast f$ is inherently a $d$-dimensional operation unless $f$ is Gaussian. Thus, it does not appear that \eqref{mainBound} can be easily reduced to a one-dimensional problem.
The quantity $h(\tfrac{1}{\sqrt{2}}(X+X_*) ) - h(X)$ characterizes the \emph{entropy production} (or, entropy jump) associated to $X$ under rescaled convolution. Similarly, letting $J(X) = 4 \int_{\mathbb{R}^d} \left|\nabla \sqrt f \right|^2$ denote the Fisher information of $X$, the quantity $J(X) - J(\tfrac{1}{\sqrt{2}}(X+X_*) )$ characterizes the \emph{dissipation of Fisher information}. By the convolution inequalities of Shannon \cite{shannon48}, Blachman \cite{blachman1965convolution} and Stam \cite{stam1959some} for entropy power and Fisher information, it follows that both the production of entropy and the dissipation of Fisher information under rescaled convolution are nonnegative. Moreover, both quantities are identically zero if and only if $X$ is Gaussian.
The fundamental problem of bounding entropy production (and dissipation of Fisher information) has received considerable attention, yet quantitative bounds are few. In particular, the entropy power inequality establishes that entropy production is strictly greater than zero unless $X$ is Gaussian, but gives no indication of how entropy production behaves as, say, a function of $D(X)$ when $X$ is non-Gaussian. As a consequence, basic stability properties of the entropy power inequality remain elusive, despite it being a central inequality in information theory. For instance, a satisfactory answer to the following question is still out of reach: \emph{If the entropy power inequality is nearly saturated, are the random summands involved quantifiably close to Gaussian?}
Perhaps the first major result to address the question of entropy production in this setting is due to Carlen and Soffer \cite{carlen1991entropy}, who showed for each random vector $X$ with $J(X)<\infty$ and $\operatorname{Cov}(X)=\mathrm{I}$, there exists a nonnegative function $\Theta_{\psi,\chi}$ on $[0,\infty)$, strictly increasing from 0, and depending only on the two auxiliary functions
\begin{align}
\psi(R) = \mathbb{E}[|X|^2 \mathbf{1}_{\{|x|\geq R\}}],
\end{align}
and
\begin{align}
\chi(t) = h\left(X_t \right)- h(X), ~~~~X_t :=e^{-t}X + (1-e^{-2t})^{1/2}G^X,
\end{align}
such that
\begin{align}
h(\tfrac{1}{\sqrt{2}}(X+X_*) ) - h(X) \geq \Theta_{\psi,\chi}( D(X) ).
\end{align}
Moreover, the function $ \Theta_{\psi_0,\chi_0}$ will do for any density $f$ with $\psi_f \leq \psi_0$ and $\chi_f \leq \chi_0$. Hence, this provides a nonlinear estimate of entropy production in terms of $D(X)$ that holds uniformly over all probability densities that exhibit the same decay and smoothness properties (appropriately defined). Unfortunately, the proof establishing existence of $\Theta_{\psi,\chi}$ relies on a compactness argument, and therefore falls short of giving satisfactory quantitative bounds.
A random vector $X$ with density $f$ has spectral gap $c>0$ (equivalently, finite Poincar\'e constant $1/c$) if, for all smooth functions $g$ with $\int_{\mathbb{R}^d} fg = 0$,
\begin{align}
c \int_{\mathbb{R}^d} f g^2 \leq \int_{\mathbb{R}^d} f |\nabla g|^2.
\end{align}
Generally speaking, a non-zero spectral gap is a very strong regularity condition on the density $f$ (e.g., it implies $X$ has finite moments of all orders).
In dimension $d=1$, if $X$ has spectral gap $c>0$, then \cite{ball2003entropy, johnson2004fisher} established the linear bound\footnote{Technically speaking, Barron and Johnson establish a slightly different inequality. However, a modification of their argument gives the same result as Ball, Barthe and Naor. See the discussion surrounding \cite[Theorem 2.4]{johnson2004information}.}
\begin{align}
h(\tfrac{1}{\sqrt{2}}(X+X_*) ) - h(X) \geq \tfrac{c}{2+2c} D(X).\label{specGap}
\end{align}
In dimension $d\geq 2$, Ball and Nguyen \cite{ball2012entropy} recently established an essentially identical result under the additional assumption that $X$ is isotropic (i.e., $\operatorname{Cov}(X)=\mathrm{I}$) with log-concave density $f$. Along these lines, Toscani has a strengthened EPI for log-concave densities \cite{toscani2015strengthened}, but the deficit is qualitative in nature in contrast to the quantitative estimate obtained by Ball and Nguyen. %
Clearly, entropy production and Fisher information dissipation is closely related to convergence rates in the entropic and Fisher information central limit theorems. Generally speaking though, bounds in the spirit of \eqref{specGap} are unnecessarily strong for establishing entropic central limit theorems of the form $D(S_n) = O(1/n)$, where $S_n := \tfrac{1}{\sqrt{n}}\sum_{i=1}^n X_i$ denotes the normalized sum of $n$ IID copies of $X$. Indeed, it was long conjectured that $D(S_n) = O(1/n)$ under moment conditions. This was positively resolved by Bobkov, Chistyakov and G\"otze~\cite{bobkov2013rate, bobkov2014berry} using Edgeworth expansions and local limit theorems. By Pinsker's inequality, we know that $D(X)$ dominates squared total-variation distance, so $D(S_n) = O(1/n)$ is interpreted as a version of the Berry-Esseen theorem for the entropic CLT with the optimal convergence rate. However, while the results of Bobkov et al. give good long-range estimates of the form $D(S_n) = O(1/n)$ (with explicit constants depending on the cumulants of $X$), the smaller-order terms propagate from local limit theorems for Edgeworth expansions and are non-explicit. Thus, explicit bounds for the initial entropy jump $h(\tfrac{1}{\sqrt{2}}(X+X_*) ) - h(X)$ cannot be readily obtained.
Along these lines, we remark that Ledoux, Nourdin and Peccati \cite{ledoux2016stein} recently established the weaker convergence rate $D(S_n) = O(\frac{\log n}{n})$ via the explicit bound
\begin{align}
D(S_n) \leq \frac{S^2(X|G) }{2 n} \log\left( 1 + n \frac{I(X|G)}{S^2(X|G)}\right) , \label{steinDeficit}
\end{align}
where $I(X|G)$ is the relative Fisher information of $X$ with respect to the standard normal $G$, and $S(X|G)$ denotes the \emph{Stein discrepancy} of $X$ with respect to $G$, which is defined when a Stein kernel exists (see \cite{ledoux2016stein} for definitions). In principle, this has potential to give an explicit bound on the entropy production $h(\tfrac{1}{\sqrt{2}}(X+X_*) ) - h(X)$ in terms of $S^2(X|G)$ by considering $n=2$.
Unfortunately, a Stein kernel may not always exist; even if it does, further relationships between $S^2(X|G)$ and $I(X|G)$ or $D(X)$ would need to be developed to ultimately yield a bound like \eqref{mainBound}.
Another related line of work in statistical physics considers quantitative bounds on entropy production in the Boltzmann equation (see the review \cite{villani2002review} for an overview). The two problems are not the same, but there is a strong analogy between entropy production in the Boltzmann equation and entropy jumps associated to rescaled convolution as can be seen by comparing \cite{carlen1992strict} to \cite{carlen1991entropy}. The details of this rich subject are tangential to the present discussion, but we remark that a major milestone in this area was achieved when the entropy production in the Boltzmann equation was bounded from below by an explicit function of $D(X)$ (and various norms of $X$), where $X$ models the velocity of a particle in a rarified gas\cite{toscani1999sharp,villani2003cercignani}. A key ingredient used to prove this bound was an earlier result by Desvillettes and Villani that controls relative Fisher information $I(X|G)$ via entropy production in the Landau equation:
\begin{lemma}\label{lem:DV} \cite{desvillettes2000spatially}
Let $X$ be a random vector on $\mathbb{R}^d$, satisfying $\mathbb{E}|X|^2 = d\geq 2$ and having density $f$. Then,
\begin{align}
\frac{1}{2} \iint \left| x-x_*\right|^2f(x) f(x_*) \left| \Pi(x-x_*)\left[ \frac{\nabla f}{f}(x) -\frac{\nabla f}{f}(x_*) \right] \right|^2 \mathrm{d}x \, \mathrm{d}x_*\geq \lambda \, (d-1) I(X|G), \label{landauBound}
\end{align}
where $\lambda$ is the minimum eigenvalue of the covariance matrix associated to $X$, and $\Pi(v)$ is the orthogonal projection onto the subspace orthogonal to $v\in \mathbb{R}^d$.
\end{lemma}
Our proof of \eqref{mainBound} follows a program similar to \cite{toscani1999sharp,villani2003cercignani}, and is conceptually straightforward after the correct ingredients are assembled. In particular, we begin by recognizing that the LHS of \eqref{landauBound} resembles dissipation of Fisher information when written in the context of $L^2$ projections (cf. \cite[Lemma 3.1]{johnson2004fisher}). Using the radial symmetry assumption, we are able to bound the Fisher information dissipation from below by error terms plus entropy production in the Landau equation, which is subsequently bounded by relative Fisher information using Lemma \ref{lem:DV}. Care must be exercised in order to control error terms (this is where our regularity assumptions enter), but the final result \eqref{mainBound} closely parallels that proved in \cite{toscani1999sharp} for the Boltzmann equation. We remark that the assumption of a non-vanishing Boltzmann collision kernel in \cite{toscani1999sharp} has a symmetrizing effect on the particle density functions involved; the rough analog in the present paper is the radial symmetry assumption.
\subsection*{Organization}
The rest of this paper is organized as follows. Section \ref{sec:Notation} briefly introduces notation and definitions that are used throughout. Main results are stated and proved in Section \ref{sec:MainResults}, followed by a brief discussion on potential extensions to non-symmetric distributions. Section \ref{sec:LSI} gives an application of the results to bounding the deficit in the Gaussian logarithmic Sobolev inequality.
\section{Notation and Definitions}\label{sec:Notation}
For a vector $v\in \mathbb{R}^d$, we let $|v| :=(\sum_{i=1}^d v_i^2 )^{1/2}$ denote its Euclidean norm. For a random variable $X$ on $\mathbb{R}$ and $p\geq 1$, we write $\|X\|_p := \left( \mathbb{E} |X|^p \right)^{1/p}$ for the usual $L^p$-norm of $X$. It will be convenient to use the same notation for $0 < p <1$, with the understanding that $\| \cdot \|_p$ is not a norm in this case.
Throughout, $G\sim N(0,\mathrm{I})$ denotes a standard Gaussian random vector on $\mathbb{R}^d$; the dimension will be clear from context. For a random vector $X$ on $\mathbb{R}^d$, we let $G^X = \sqrt{d^{-1}\mathbb{E}|X|^2} G$ be a normalized Gaussian vector, so that $\mathbb{E}|X|^2 = \mathbb{E} |G^X|^2$. For $d\geq 2$, we denote the coordinates of a random vector $X$ on $\mathbb{R}^d$ as $(X_1,X_2, \dots, X_d)$. Thus, for example, $G_1^X$ is a zero-mean Gaussian random variable with variance $d^{-1}\mathbb{E}|X|^2$.
For a random vector $X$ with smooth density\footnote{All densities are with respect to Lebesgue measure.} $f$, we define the Fisher information
\begin{align}
J(X) = 4 \int \left|\nabla \sqrt f \right|^2 = \int_{f>0} \frac{|\nabla f|^2}{f} \label{FIidentity}
\end{align}
and the entropy
\begin{align}
h(X) = -\int f \log f,
\end{align}
where `$\log$' denotes the natural logarithm throughout. For random vectors $X,Q$ with respective densities $f,g$, the relative Fisher information is defined by
\begin{align}
I(X|Q) = 4 \int g \left|\nabla \sqrt {f/g} \right|^2
\end{align}
and the relative entropy is defined by
\begin{align}
D(X|Q) = \int f \log \frac{f}{g}.%
\end{align}
Evidently, both quantities are nonnegative and
\begin{align}
&I(X) := I(X|G^X) = J(X) - J(G^X) & D(X):=D(X|G^X) = h(G^X) -h(X).
\end{align}
Finally, we recall two basic inequalities that will be taken for granted several times without explicit reference: for real-valued $a,b$ we have $(a+b)^2 \leq 2 a^2 +2 b^2$, and for random variables $X,Y$, we have Minkowski's inequality: $\|X+Y\|_p \leq \|X\|_p + \|Y\|_p$ when $p\geq 1$.
\begin{definition}
A random vector $X$ with density $f$ is radially symmetric if $f(x) = \phi(|x|)$ for some function $\phi : \mathbb{R}\to [0,\infty)$.
\end{definition}
We primarily concern ourselves with random vectors that satisfy certain mild regularity conditions. In particular, it is sufficient to control $|\nabla \log f (x) |$ pointwise in terms of $|x|$.
\begin{definition}
A random vector $X$ on $\mathbb{R}^d$ with smooth density $f$ is $c$-regular if, for all $x\in \mathbb{R}^d$,
\begin{align}
|\nabla \log f (x) | \leq c \left( |x|+ \mathbb{E}|X| \right). \label{cRegIneq}
\end{align}
\end{definition}
We remark that the smoothness requirement of $f$ in the definition of $c$-regularity is stronger than generally required for our purposes. However, it allows us to avoid further qualifications; for instance, the identities \eqref{FIidentity} hold for any $c$-regular function. Moreover, since $\nabla \log f = \tfrac{\nabla f}{f}$ for smooth $f$, we have $J(X)<\infty$ for any $c$-regular $X$ with $\mathbb{E}|X|^2<\infty$.
Evidently, $c$-regularity quantifies the smoothness of a density function. The following important example shows that any density can be mollified to make it $c$-regular.
\begin{proposition}\label{prop:addNoise} \cite{polyWuWasserstein2016}
Let $X$ and $Z$ be independent, where $Z\sim N(0,\sigma^2 \mathrm{I})$. Then $Y = X+Z$ is $c$-regular with $c = 4 /\sigma^2$.
\end{proposition}
Observe that, in the notation of the above proposition, if $X$ is radially symmetric then so is $Y$. Therefore, Proposition \ref{prop:addNoise} provides a convenient means to construct radially symmetric random vectors that are $c$-regular. Indeed, we have the following useful corollaries (proofs are found in the appendix).
\begin{proposition}\label{prop:StartRegular}
Let $X$ be a random vector on $\mathbb{R}^d$, and let $X_t = e^{-t} X + (1-e^{-2t})^{1/2}G$ for $t\geq0$. If $X$ is $c$-regular, then $X_t$ is $(5c\,e^{2t})$-regular.
\end{proposition}
\begin{proposition}\label{prop:approxR}
Let $R_0$ be a non-negative random variable with $\mathbb{E} R_0^2 = 1$ and distribution function $F_{R_0}$. For any $d\geq 2$ and $t, \varepsilon>0$, there exists a $(4/\varepsilon)$-regular radially symmetric random vector $X$ on $\mathbb{R}^d$ with $\mathbb{E}|X|^2 = d$ and $R = \frac{1}{\sqrt{d}}|X|$ satisfying
\begin{align}
F_{R_0}\left( \tfrac{ r - \sqrt{(t+1)\varepsilon}}{\sqrt{1-\varepsilon}} \right) - e^{-d t^2 /8} \leq F_{R}(r) & \leq F_{R_0}\left( \tfrac{ r + \sqrt{(t+1)\varepsilon}}{\sqrt{1-\varepsilon}} \right) + e^{-d t^2 /8},
\end{align}
where $F_{R}$ is the distribution function of $R$.
\end{proposition}
\section{Main Results } \label{sec:MainResults}
In this section, we establish quantitative estimates on entropy production and Fisher information dissipation under rescaled convolution. As can be expected, we begin with an inequality for Fisher information, and then obtain a corresponding entropy jump inequality by integrating along the Ornstein-Uhlenbeck semigroup.
\subsection{Dissipation of Fisher Information under Rescaled Convolution}
\begin{theorem}\label{thm:FIjumps}
Let $X, X_*$ be IID radially symmetric random vectors on $\mathbb{R}^d$, $d\geq 2$, with $c$-regular density $f$. For any $\varepsilon > 0 $
\begin{align}
J(X)-J(\tfrac{1}{\sqrt{2}}(X+X_*)) &\geq K_{\varepsilon}(X) I(X)^{1+\varepsilon}, \label{FIIjump}
\end{align}
where
\begin{align}
K_{\varepsilon}(X) =
\frac{(\varepsilon / 8)^{\varepsilon} }{ \left( 8 (1+\varepsilon) \right)^{1+\varepsilon} }
\cdot
\frac{ \||X|^2\|_1^{1+\varepsilon} }{c^{{2 \varepsilon}} \left\| |X|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} }
. \label{KpX_FI}
\end{align}
\end{theorem}
\begin{remark}
We have made no attempt to optimize the constant $K_{\varepsilon}(X)$. %
\end{remark}
A few comments are in order. First, we note that inequality \eqref{FIIjump} is invariant to scaling $t: X \mapsto tX$ for $t>0$. Indeed, if $X$ is $c$-regular, then a change of variables shows that $tX$ is $(c /t^2)$-regular. So, using homogeneity of the norms, we find that
\begin{align}
K_{\varepsilon}(t X) = t^{2\varepsilon} K_{\varepsilon}(X).
\end{align}
Combined with the property that $t^2 J(t X) = J(X)$, we have
\begin{align}
K_{\varepsilon}(tX) I(tX|G^{tX})^{1+\varepsilon} = t^{-2} K_{\varepsilon}(X)I(X|G^{X})^{1+\varepsilon},
\end{align}
which has the same scaling behavior as the LHS of \eqref{FIIjump}. That is,
\begin{align}
J(tX)-J(\tfrac{1}{\sqrt{2}}(tX+tX_*)) = t^{-2}\left( J(X)-J(\tfrac{1}{\sqrt{2}}(X+X_*))\right).
\end{align}
Second, inequality \eqref{FIIjump} does not contain any terms that explicitly depend on dimension. However, it is impossible to say that inequality \eqref{FIIjump} is dimension-free in the usual sense that both sides scale linearly in dimension when considering product distributions. Indeed, the product of two identical radially symmetric densities is again radially symmetric if and only if the original densities were Gaussian themselves, which corresponds to the degenerate case when the dissipation of Fisher information is identically zero. However, inequality \eqref{FIIjump} does exhibit dimension-free behavior in the following sense: Suppose for simplicity that $X$ is normalized so that $\mathbb{E}|X|^2=d$. Since $X$ is radially symmetric, it can be expressed as the product of independent random variables $X =\sqrt{d} R U$, where $U$ is uniform on the $(d-1)$-dimensional sphere $\mathbb{S}^{d-1}$ and $R=\frac{1}{\sqrt{d}}|X|$ is a nonnegative real-valued random variable satisfying $\mathbb{E} R^2 = 1$. Now, by the log Sobolev inequality and Talagrand's inequality, we have
\begin{align}
I(X|G) \geq 2 D(X|G) \geq W_2^2 \Big( (\sqrt{d} R U) , G \Big) = d \, W_2^2 \Big( R , \tfrac{1}{\sqrt{d}} |G| \Big)
\end{align}
The equality follows since, for any vectors $v,w$ we have $|v-w|^2 \geq \big| |v|-|w|\big|^2$. However, this can be achieved with equality by the coupling $G = |G| U$. Thus, we have
\begin{align}
K_{\varepsilon}(X) I(X|G)^{1+\varepsilon} &\geq \frac{(\varepsilon / 8)^{\varepsilon} \, d^{1+2 \varepsilon}W_2^{2\varepsilon} \Big( R , \tfrac{1}{\sqrt{d}} |G| \Big) }{c^{{2 \varepsilon}} \left( 8 (1+\varepsilon) \right)^{1+\varepsilon} \, \left\| |X|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} } I(X|G)\\
&= \frac{(\varepsilon / 8)^{\varepsilon} \, W_2^{2\varepsilon} \Big( R , \tfrac{1}{\sqrt{d}} |G| \Big) }{c^{{2 \varepsilon}} \left( 8 (1+\varepsilon) \right)^{1+\varepsilon} \, \left\| R \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} } I(X|G).
\end{align}
Now, we note that $\mathbb{E} R^2 = \mathbb{E} (\tfrac{1}{\sqrt d} |G|)^2 = 1$, so we have a bound of the form
\begin{align}
J(X)-J(\tfrac{1}{\sqrt{2}}(X+X_*)) &\geq \widetilde{K}_{\varepsilon}(X) I(X|G^X),
\end{align}
where the function $\widetilde{K}_{\varepsilon}(X)$ is effectively dimension-free in that it only depends on the (one-dimensional) quadratic Wasserstein distance between $R$ and $\tfrac{1}{\sqrt{d}} |G|$. For $d\to \infty$, the law of large numbers implies that $\tfrac{1}{\sqrt{d}} |G|\to 1$ a.s. Therefore, $ W_2^2 \Big( R , \tfrac{1}{\sqrt{d}} |G| \Big)$ behaves similarly to $\mathbb{E}|R-1|^2 \geq \operatorname{Var}(R)$ high dimensions. Indeed, by the triangle inequality applied to $W_2$,
\begin{align}
\left| W_2 \Big( R , \tfrac{1}{\sqrt{d}} |G| \Big) - \|R-1\|_2 \right| \leq \left\|\tfrac{1}{\sqrt{d}} |G| -1\right\|_2 = O\left(\tfrac{1}{\sqrt{d}}\right). %
\end{align}
So, we see that \eqref{FIIjump} depends very weakly on $d$ when the marginal distribution of $|X|$ is preserved and dimension varies.
One important question remains: As dimension $d\to\infty$, do there exist random vectors $X$ on $\mathbb{R}^d$ with sufficient regularity for which the associated random variable $R$ is not necessarily concentrated around $1$? The answer to this is affirmative in the sense of Proposition \ref{prop:approxR}: we may approximate any distribution function $F_{R_0}$ to within arbitrary accuracy, at the (potential) expense of increasing the regularity parameter $c$.
\begin{proof}[Proof of Theorem \ref{thm:FIjumps}]
As remarked above, inequality \eqref{FIIjump} is invariant to scaling. Hence, there is no loss of generality in assuming that $X$ is normalized according to $\mathbb{E}|X|^2 = d$. Also, since $X$ is radially symmetric, $X-X_*$ is equal to $X+X_*$ in distribution, therefore we seek to lower bound the quantity
\begin{align}
J(X) - J(\tfrac{1}{\sqrt{2}}(X+X_*))= J(X) - 2 J(X-X_*).
\end{align}
Toward this end, define $W = X-X_*$, and denote its density by $f_W$. By the projection property of the score function of sums of independent random variables, the following identity holds (e.g., \cite[Lemma 3.4]{johnson2004information}):
\begin{align}
2 \left( J(X) - 2 J(X-X_*)\right) &= \mathbb{E}\left| 2 \rho_W(W) - \left( \rho(X) - \rho(X_*)\right) \right|^2,
\end{align}
where $\rho = \nabla \log f$ is the score function of $X$ and $\rho_W = \nabla \log f_W$ is the score function of $W$.
For $v\in \mathbb{R}^d$, let $\Pi(v)$ denote the orthogonal projection onto the subspace orthogonal to $v$. Now, we have
\begin{align}
2 J(X) - 4 J(X-X_*) &= \mathbb{E}\left| 2 \rho_W(W) - \left( \rho(X) - \rho(X_*)\right) \right|^2\\
&\geq \mathbb{E}\left| 2 \Pi(W) \rho_W(W) - \Pi(X-X_*)\left( \rho(X) - \rho(X_*)\right) \right|^2\label{eq1symm} \\
&= \mathbb{E}\left| \Pi(X-X_*)\left( \rho(X) - \rho(X_*)\right) \right|^2.\label{eq2symm}
\end{align}
The inequality follows since $\Pi(w) = \Pi(x-x_*)$ by definition, and $|v| \geq |\Pi(w) v|$ since $\Pi(w)$ is an orthogonal projection. The last equality follows since $\Pi(w) \rho_W(w) =0$ due to the fact that $\Pi(w) \nabla f_W(w)$ is the tangential gradient of $f_W$, which is identically zero due to radial symmetry of $f_W$.
Next, for any $R>0$, use the inequality
\begin{align}
1 \geq \frac{|x-x_*|^2}{ R^2} - \frac{|x-x_*|^2}{ R^2} \mathbf{1}_{\{|x-x_*|> R\}}
\end{align}
to conclude that
\begin{align}
2 J(X) - 4 J(X-X_*) &\geq \frac{1}{R^2} \mathbb{E}\left[\left| X-X_*\right|^2\left| \Pi(X-X_*)\left( \rho(X) - \rho(X_*)\right) \right|^2\right]\notag \\
&-\frac{1}{R^2} \mathbb{E}\left[\left| X-X_*\right|^2\left| \Pi(X-X_*)\left( \rho(X) - \rho(X_*)\right) \right|^2 \mathbf{1}_{\{|X-X_*|> R\}}\right].
\end{align}
We bound the second term first. By $c$-regularity and the triangle inequality, we have
\begin{align}
\left| \Pi(x-x_*)\left( \rho(x) - \rho(x_*)\right)\right| \leq \left| \rho(x) - \rho(x_*)\right| \leq c(|x|+|x_*|)+2c\mathbb{E}|X|. \end{align}
So, noting
the inclusion
\begin{align}
\{|x-x_*|> R\} \supseteq \{|x|\geq R/2, |x_*|\leq |x| \} \cup\{|x_*|\geq R/2, |x|\leq |x_*| \},
\end{align} we have the pointwise inequality
\begin{align}
&\mathbf{1}_{\{|x-x_*|> R\}} |x-x_*|^2 \left| \Pi(x-x_*) \left( \rho(x) - \rho(x_*)\right) \right|^2 \\
&\leq \mathbf{1}_{\{|x|\geq R/2, |x_*|\leq |x| \} \cup\{|x_*|\geq R/2, |x|\leq |x_*| \} } |x-x_*|^2 \left| \rho(x) - \rho(x_*) \right|^2\\
&\leq \mathbf{1}_{\{|x|\geq R/2 \} }4 |x|^2 \left( 2c|x|+2 c\mathbb{E}|X| \right)^2 + \mathbf{1}_{\{|x_*|\geq R/2 \} }4 |x_*|^2 \left( 2c|x_*|+2 c\mathbb{E}|X| \right)^2.
\end{align}
Taking expectations and using the fact that $X,X_*$ are IID, we have for any conjugate exponents $p,q\geq 1$ and $\beta>0$,
\begin{align}
&\mathbb{E}\left[\left| X-X_*\right|^2\left| \Pi(X-X_*)\left( \rho(X) - \rho(X_*)\right) \right|^2 \mathbf{1}_{\{|X-X_*|> R\}}\right]\notag\\
&\leq 16 \,c^2\, \mathbb{E}\left[\left| X \right|^2\left( |X| + \mathbb{E}|X| \right)^2 \mathbf{1}_{\{|X|\geq R/2\}}\right]\\
&\leq 32 \, c^2 \mathbb{E}\left[\left| X \right|^4 \mathbf{1}_{\{|X|> R/2\}}\right] +
32 \, c^2 \, \left(\mathbb{E}|X|\right)^2 \, \mathbb{E}\left[\left| X \right|^2 \mathbf{1}_{\{|X|\geq R/2\}}\right]\\
&\leq 32 \, c^2 \,\left\| |X|^2 \right\|_{ 2p}^2\left( \Pr\left\{ |X|\geq R/2\right\}\right)^{1/q}
+ 32 \, c^2 \, \left(\mathbb{E}|X|\right)^2\, \left\| |X|^2 \right\|_{ p} \left( \Pr\left\{ |X|\geq R/2\right\}\right)^{1/q}\\
&\leq 32 \, c^2 \left\| |X|^2 \right\|_{ 2p}^2 \left( \frac{\mathbb{E}|X|^{\beta}}{(R/2)^{\beta}}\right)^{1/q}
+ 32 \, c^2 \left(\mathbb{E}|X|\right)^2\, \left\| |X|^2 \right\|_{ p}\left( \frac{\mathbb{E}|X|^{\beta}}{(R/2)^{\beta}}\right)^{1/q}\\
&= \frac{32\cdot 2^{\beta/q} \, c^2}{R^{\beta /q}} \left( \left\| |X|^2 \right\|^2_{ 2p}
+ \left\| |X|^2 \right\|_{1/2}\, \left\| |X|^2 \right\|_{ p} \right) \left(\mathbb{E}|X|^{\beta}\right)^{1/q}\\
&\leq \frac{64\cdot 2^{\beta/q} \, c^2}{R^{\beta /q}} \left\| |X|^2 \right\|^2_{ 2p} \left\||X|^{2}\right\|_{\beta/2}^{\beta/(2q)} .
%
\end{align}
Since $\mathbb{E}|X|^2 = d$, radial symmetry implies $\operatorname{Cov}(X) = \mathrm{I}$. Therefore, by Lemma \ref{lem:DV}, we have
\begin{align}
\mathbb{E}\left[\left| X-X_*\right|^2\left| \Pi(X-X_*)\left( \rho(X) - \rho(X_*)\right) \right|^2\right] \geq 2(d-1) I(X|G).
\end{align}
Continuing from above, we have proved that
\begin{align}
J(X) - 2 J(X-X_*)&\geq \frac{d-1 }{R^2} I(X|G) -\frac{32\cdot 2^{\beta /q} \, c^2 }{R^{2+\beta/q}} \left\| |X|^2 \right\|^2_{ 2p} \left\||X|^{2}\right\|_{\beta/2}^{\beta/(2q)}.
\end{align}
For any $s > 0$, Taking $R = \left( \frac{b(2+s)}{2a}\right)^{1/s}$ yields the identity
\begin{align}
\frac{a}{R^2}-\frac{b}{R^{2+s}} = \frac{1}{1+2/s}\left( \frac{2/s}{b(1+2/s)}\right)^{2/s} a^{1+2/s}.
\end{align}
So, putting $\varepsilon = 2q/\beta$, $p = 1+1/(2\varepsilon)$ and simplifying, we obtain
\begin{align}
J(X) - 2 J(X-X_*)&\geq \frac{(\varepsilon /8)^{\varepsilon}}{\left( 4 (1+\varepsilon) \right)^{1+\varepsilon}}\left(\frac{ 1 }{ c \left\| |X|^2 \right\|_{ 2p} }\right)^{2 \varepsilon} \left( \frac{1}{ \left\||X|^{2}\right\|_{q/\varepsilon}}\right)\left( (d-1) I(X|G)\right) ^{1+\varepsilon}\\
&\geq \frac{(\varepsilon /8)^{\varepsilon}}{\left( 8 (1+\varepsilon) \right)^{1+\varepsilon}}\left(\frac{ 1 }{ c \left\| |X|^2 \right\|_{ 2p} }\right)^{2 \varepsilon} \left( \frac{1}{ \left\||X|^{2}\right\|_{q/\varepsilon}}\right)\left( d\, I(X|G)\right) ^{1+\varepsilon}\\
&= \frac{(\varepsilon / 8)^{\varepsilon}}{c^{2\varepsilon }\left( 8 (1+\varepsilon) \right)^{1+\varepsilon}}\left(\frac{ 1 }{ \left\| |X|^2 \right\|_{ 2+\varepsilon^{-1}} }\right)^{1+2 \varepsilon} \left( d \, I(X|G)\right) ^{1+\varepsilon}\\
&= \frac{(\varepsilon / 8)^{\varepsilon} \, \||X|^2\|_1^{1+\varepsilon} }{c^{{2 \varepsilon}} \left( 8 (1+\varepsilon) \right)^{1+\varepsilon} \, \left\| |X|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} } \, I(X|G) ^{1+\varepsilon},\label{FI_jump_beforeSimplify}
\end{align}
where we have made use of the crude bound $(d-1)/d\geq 1/2$ and substituted $d = \mathbb{E}|X|^2 = \| |X|^2\|_1$.
\end{proof}
\subsection{Entropy Production under Rescaled Convolution }
As one would expect, we may `integrate up' in Theorem \ref{thm:FIjumps} to obtain an entropic version. A precise version of the result stated in Section \ref{sec:Intro} is given as follows:
\begin{theorem}\label{thm:Hjump}
Let $X, X_*$ be IID radially symmetric random vectors on $\mathbb{R}^d$, $d\geq 2$, with $c$-regular density $f$. For any $\varepsilon > 0 $
\begin{align}
h(\tfrac{1}{\sqrt{2}}(X+X_*))-h(X)
&\geq C_{\varepsilon}(X) \,D(X)^{1+\varepsilon} ,\label{EJineq}
\end{align}
where
\begin{align}
C_{\varepsilon}(X) = \left( \frac{d\varepsilon}{1+( d+2)\varepsilon}\right)^{1+2\varepsilon}
\frac{ 2^4 \, (d/100)^{\varepsilon} }
{
%
\left( 2^8 (1+\varepsilon) (1+2\varepsilon) \right)^{1+\varepsilon}
%
} \cdot \frac{ \||X|^2\|_1 }{ c^{2\varepsilon}\, \left\| |X|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} } .
\end{align}
\end{theorem}
\begin{remark}
Although the constant $C_{\varepsilon}(X)$ appears to grow favorably with dimension $d$, this dimension-dependent growth can cancel to give a bound that is effectively dimension-free. An illustrative example follows the proof.
\end{remark}
\begin{proof}
Similar to before, the inequality \eqref{EJineq} is scale-invariant. Indeed, all relative entropy terms are invariant to scaling $t : X\mapsto t X$, and we also have $C_{\varepsilon}(tX) =C_{\varepsilon}(X) $ due to $tX$ being $(c/t^2)$-regular if $X$ is $c$-regular and homogeneity of the norms. Thus, we may assume without loss of generality that $X$ is normalized so that $\mathbb{E}|X|^2 = d$.
Next, define $W = \tfrac{1}{\sqrt{2}}(X+X_*)$, and let $X_t, W_t$ denote the Ornstein-Uhlenbeck evolutes of $X$ and $W$, respectively. That is, for $t\geq 0$
\begin{align}
&X_t = e^{-t} X + (1-e^{-2t})^{1/2}G, & W_t = e^{-t} W + (1-e^{-2t})^{1/2}G.
\end{align}
By Proposition \ref{prop:StartRegular}, $X_t$ is $(5 c e^{2t})$-regular for all $t\geq 0$. Noting that $\mathbb{E}|X_t|^2 = \mathbb{E}|X|^2$, an application of Theorem \ref{thm:FIjumps} gives
\begin{align}
I(X_t|G) - I(W_t|G) &\geq \frac{(\varepsilon / 8)^{\varepsilon} \, \||X_t|^2\|_1^{1+\varepsilon} }{(5c)^{{2 \varepsilon}} \left( 8 (1+\varepsilon) \right)^{1+\varepsilon} \, \left\| |X_t|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} } \,e^{-4 \varepsilon t} I(X_t|G) ^{1+\varepsilon} \\
&\geq \frac{(\varepsilon / 8)^{\varepsilon} \, \||X|^2\|_1^{1+\varepsilon} }
{ (5 c)^{{2 \varepsilon}}
%
\left( 8 (1+\varepsilon) \right)^{1+\varepsilon} \,
%
(2(\tfrac{1+(2+d)\varepsilon}{d\varepsilon} ))^{1+2\varepsilon}
%
\left\| |X|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} }
%
\,e^{-4 \varepsilon t} I(X_t|G) ^{1+\varepsilon} \label{ineqToJustify} \\
&=
\left( \frac{d\varepsilon}{2+(2d+4)\varepsilon}\right)^{1+2\varepsilon}
\frac{ (\varepsilon / 8)^{\varepsilon} \, \||X|^2\|_1^{1+\varepsilon} }
{ (5 c)^{{2 \varepsilon}}
%
\left( 8 (1+\varepsilon) \right)^{1+\varepsilon} \,
%
\left\| |X|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} }
%
\,e^{-4 \varepsilon t} I(X_t|G) ^{1+\varepsilon} ,\notag
\end{align}
where \eqref{ineqToJustify} holds since, for $p\geq 1$,
\begin{align}
\left\| |X_t|^2 \right\|_{p} = \left( \mathbb{E} |X_t|^{2 \cdot p} \right)^{1/p} &\leq 2 \left( \mathbb{E} (e^{-2t}|X|^2 +(1-e^{-2t}) |G|^2 )^{p} \right)^{1/p}\\
&= 2 \left\| e^{-2t}|X|^2 +(1-e^{-2t}) |G|^2 \right\|_{p}\\
&\leq 2\left( e^{-2t} \left\| |X|^2 \right\| _{p}+ (1-e^{-2t}) \left\| |G|^2 \right\|_{p} \right)\\
&\leq 2 (1+ \tfrac{p}{d}) \left\| |X|^2 \right\| _{p} . \label{chiMoment}
%
%
\end{align}
The bound \eqref{chiMoment} uses the fact that $|G|^2$ is a chi-squared random variable with $d$ degrees of freedom, and hence (using $\mathbb{E}|X|^2 = d$):
\begin{align}
\left\| |G|^2 \right\|_{p} = \left( 2^p \frac{\Gamma(p+\tfrac{d}{2})}{\Gamma(\tfrac{d}{2})} \right)^{1/p} &=\mathbb{E}|X|^2 \left( \frac{\Gamma(p+\tfrac{d}{2})}{\Gamma(\tfrac{d}{2}) \left( \tfrac{d}{2}\right)^p} \right)^{1/p}\\
&\leq \mathbb{E}|X|^2 \left( 1+\tfrac{p}{d} \right) \label{chiSquareBound} \\
&\leq \left\| |X|^2 \right\| _{p}(1+ \tfrac{p}{d}) .
\end{align}
Now, the claim will follow by integrating both sides. Indeed, by the classical de Bruijn identity, we have
\begin{align}
\int_{0}^{\infty}\left( I(X_t|G) - I(W_t|G) \right) dt = D(X|G)- D(W|G) = h(\tfrac{1}{\sqrt{2}}(X+X_*))-h(X) .
\end{align}
By Jensen's inequality,
\begin{align}
\int_{0}^{\infty}e^{-4\varepsilon t } I(X_t|G)^{1+\varepsilon} dt &\geq
\frac{1}{(4\varepsilon)^{\varepsilon}}\left( \int_{0}^{\infty}e^{-4\varepsilon t } I(X_t|G) dt \right) ^{1+\varepsilon} \\
&\geq \frac{1}{(4\varepsilon)^{\varepsilon}} \left( \int_{0}^{\infty} I(X_{t+2\varepsilon t }|G) dt \right) ^{1+\varepsilon} \\
&= \frac{1}{(4\varepsilon)^{\varepsilon}(1+2\varepsilon)^{1+\varepsilon}}D(X|G)^{1+\varepsilon} ,
\end{align}
where we used the bound $I(X_{t+s}|G)\leq e^{- 2 s}I(X_t|G)$ due to exponential decay of information along the semigroup (e.g., \cite{bakry2013analysis}), a change of variables, and the identity $\int_{0}^{\infty}I(X_t|G) dt = D(X|G)$.
Thus, we have proved
\begin{align}
&h(\tfrac{1}{\sqrt{2}}(X+X_*))-h(X) \notag \\
&\geq
\left( \frac{d\varepsilon}{2+(2d+4)\varepsilon}\right)^{1+2\varepsilon}
\frac{ (\varepsilon / 8)^{\varepsilon} \, \||X|^2\|_1^{1+\varepsilon} }
{ (5 c)^{{2 \varepsilon}}
%
\left( 8 (1+\varepsilon) \right)^{1+\varepsilon} \,
%
\left\| |X|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} }
\cdot \frac{1}{(4\varepsilon)^{\varepsilon}(1+2\varepsilon)^{1+\varepsilon}}D(X|G)^{1+\varepsilon} \\
%
&= \left( \frac{d\varepsilon}{1+( d+2)\varepsilon}\right)^{1+2\varepsilon}
\frac{ 2^4 \, (d/100)^{\varepsilon} }
{
%
\left( 2^8 (1+\varepsilon) (1+2\varepsilon) \right)^{1+\varepsilon}
%
} \cdot \frac{ \||X|^2\|_1 }{ c^{2\varepsilon}\, \left\| |X|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} }
D(X|G)^{1+\varepsilon} .
%
\end{align}
\end{proof}
\begin{example}[Centered Gaussian Mixtures]
Define $f_i$ to be the density associated to the centered Gaussian distribution with covariance $\sigma_i^2\, \mathrm{I}$. Let $X$ be a random vector on $\mathbb{R}^d$ with density $f = \sum_{i=1}^n p_i f_i$. For convenience, assume $\sigma_1^2 =\min_i \sigma_i^2$ and that the $\sigma_i$'s are normalized so that $\sum_{i=1}^n p_i \sigma_i^2=1$ ($\{p_1, \dots, p_n\}$ is a probability vector). Then, for $d$ large,
\begin{align}
h(\tfrac{1}{\sqrt{2}}(X+X_*))-h(X) \gtrsim \frac{ 1 }
{
%
2^{15}
%
} \cdot \frac{ \sum_{i=1}^n p_i (\sigma_i-1)^2 }{ \sigma_1^2 \, \sum_{i=1}^n p_i \left( \sigma^2_i/ \sigma_1^2\right)^{3} }\,
D(X).
\end{align}
\end{example}
It is easy to verify that $f$ is $(1/ \sigma_1^2 )$-regular. Moreover, $\mathbb{E}|X|^2 = d$, so using the bound \eqref{chiSquareBound}, we have
\begin{align}
\| |X|^2\|_{2+1/\varepsilon}\leq \left(\sum_{i=1}^n p_i \sigma_i^{2(2+1/\varepsilon)} \right)^{1/(2+1/\varepsilon)} (d + 2+1/\varepsilon)
\end{align}
Now, by Talagrand's inequality, we may lower bound $D(X) \geq \frac{d}{2} W_2^2 \left( \tfrac{1}{\sqrt{d}} |X| , \tfrac{1}{\sqrt{d}} |G| \right)$ as we did in the discussion following Theorem \ref{thm:FIjumps}.
Putting everything together and simplifying, we obtain:
\begin{align}
&h(\tfrac{1}{\sqrt{2}}(X+X_*))-h(X) \\
&\geq \left( \frac{d\varepsilon}{1+( d+2)\varepsilon}\right)^{2(1+2\varepsilon)}
\frac{ 2^4 \, (1/200)^{\varepsilon} }
{
%
\left( 2^8 (1+\varepsilon) (1+2\varepsilon) \right)^{1+\varepsilon}
%
} \cdot \frac{ W_2^{2\varepsilon } \left( \tfrac{1}{\sqrt{d}} |X| , \tfrac{1}{\sqrt{d}} |G| \right) }{ \sigma_1^2 \, \left(\sum_{i=1}^n p_i \left( \sigma^2_i/ \sigma_1^2\right)^{2+1/\varepsilon} \right)^{\varepsilon} }\,
D(X).\notag
\end{align}
An easy consequence of the LLN is the limit
\begin{align}
W_2^{2 } \left( \tfrac{1}{\sqrt{d}} |X| , \tfrac{1}{\sqrt{d}} |G| \right) \to \sum_{i=1}^n p_i (\sigma_i-1)^2~~~\text{as $d\to\infty$},
\end{align}
so the claim follows by putting $\varepsilon=1$ and crudely bounding.
\medskip
It is straightforward to remove the explicit requirement in Theorem \ref{thm:Hjump} for $c$-regularity:
\begin{corollary}\label{cor:HjumpNoRegularity}
Let $X, X_*$ be IID radially symmetric random vectors on $\mathbb{R}^d$, $d\geq 2$, with finite Fisher information. For any $\varepsilon > 0 $
\begin{align}
h(\tfrac{1}{\sqrt{2}}(X+X_*))-h(X)
\geq \widetilde{C}_{\varepsilon}(X)
\frac{D(X)^{1+3\varepsilon}}{I(X)^{2\varepsilon} } ,\label{noC_Hjumps}
\end{align}
where
\begin{align}
\widetilde{C}_{\varepsilon}(X) = \left( \frac{d\varepsilon}{1+( d+2)\varepsilon}\right)^{2+4\varepsilon}
\frac{ 2^{12} \, (d/100)^{\varepsilon} }
{
%
\left( 2^{17} (1+\varepsilon) (1+2\varepsilon) \right)^{1+\varepsilon}
%
} \cdot \frac{ \||X|^2\|_1 }{ %
\left\| |X|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} }.\label{CtildeDefn}
\end{align}
\end{corollary}
\begin{remark}
Although the requirement of $c$-regularity is eliminated in \eqref{noC_Hjumps}, we see that the statement of \eqref{noC_Hjumps} is effectively the same as \eqref{EJineq}, with $c$-regularity being replaced by another measure of regularity of $X$, i.e., the relative Fisher information $I(X)$.
\end{remark}
\begin{proof}
Observe that inequality \eqref{noC_Hjumps} is invariant to scaling $X$. Indeed, for $t>0$ we have $t^{4\varepsilon} \widetilde{C}_{\varepsilon}(t X) = \widetilde{C}_{\varepsilon}(X)$ and $t^2 I(tX | G^{tX}) = I(X | G^{X})$. Therefore,
\begin{align}
\widetilde{C}_{\varepsilon}(t X) \frac{1}{I(tX|G^{tX})^{2\varepsilon} } =\widetilde{C}_{\varepsilon}(X) \frac{1}{I(X|G^{X})^{2\varepsilon} }.
\end{align}
Since $D(tX | G^{tX})=D(X|G^X)$ and $h(\tfrac{1}{\sqrt{2}}(tX+tX_*))-h(tX) =h(\tfrac{1}{\sqrt{2}}(X+X_*))-h(X)$, the claim follows. Thus, as before, we will assume without loss of generality that $X$ is normalized so that $\mathbb{E}|X|^2 = d$.
Next, define $W = \tfrac{1}{\sqrt{2}}(X+X_*)$, and let $X_t, W_t$ denote the Ornstein-Uhlenbeck evolutes of $X$ and $W$, respectively. That is, for $t\geq 0$
\begin{align}
&X_t = e^{-t} X + (1-e^{-2t})^{1/2}G, & W_t = e^{-t} W + (1-e^{-2t})^{1/2}G.
\end{align}
By Proposition \ref{prop:addNoise}, $X_t$ is $4\,(1-e^{-2t})^{-1}$-regular for all $t\geq 0$. Noting that $\mathbb{E}|X_t|^2 = \mathbb{E}|X|^2$, an application of Theorem \ref{thm:Hjump} gives
\begin{align}
&h(W_t) - h(X_t) \notag \\
&\geq \left( \frac{d\varepsilon}{1+( d+2)\varepsilon}\right)^{1+2\varepsilon}
\frac{ 2^8 \, (d/100)^{\varepsilon} }
{
%
\left( 2^{12} (1+\varepsilon) (1+2\varepsilon) \right)^{1+\varepsilon}
%
} \cdot \frac{(1-e^{-2t})^{2\varepsilon} \||X_t|^2\|_1 }{ \left\| |X_t|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} }
D(X_t|G)^{1+\varepsilon} \\
&\geq \left( \frac{d\varepsilon}{1+( d+2)\varepsilon}\right)^{1+2\varepsilon}
\frac{ 2^8 \, (d/100)^{\varepsilon} }
{
%
\left( 2^{12} (1+\varepsilon) (1+2\varepsilon) \right)^{1+\varepsilon}
%
} \cdot \frac{(1-e^{-2t})^{2\varepsilon} \||X|^2\|_1 }{ (2(\tfrac{1+(2+d)\varepsilon}{d\varepsilon} ))^{1+2\varepsilon}
%
\left\| |X|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} }
D(X_t|G)^{1+\varepsilon} \label{ineqToJustify2} \\
&= \left( \frac{d\varepsilon}{1+( d+2)\varepsilon}\right)^{2+4\varepsilon}
\frac{ 2^9 \, (d/100)^{\varepsilon} }
{
%
\left( 2^{14} (1+\varepsilon) (1+2\varepsilon) \right)^{1+\varepsilon}
%
} \cdot \frac{(1-e^{-2t})^{2\varepsilon} \||X|^2\|_1 }{ %
\left\| |X|^2 \right\|_{ 2+1/\varepsilon}^{1+2 \varepsilon} }
D(X_t|G)^{1+\varepsilon} ,
\end{align}
where \eqref{ineqToJustify2} follows by the same logic as \eqref{ineqToJustify}.
Now, by de Bruijn's identity and the convolution inequality for Fisher information, it follows that
\begin{align}
\frac{d}{dt}\left( h(W_t) - h(X_t)\right) = J(W_t)-J(X_t) \leq 0.
\end{align}
Thus, $ h(W) - h(X)\geq h(W_t) - h(X_t)$. The map $t \mapsto D(X_t |G)$ is continuous on $t\in[0,\infty)$ (e.g., \cite{carlen1991entropy}). Hence, using the fact that $\frac{d}{dt} D(X_t |G) = -I(X_t|G)$ (de Bruijn's identity) and that $t\mapsto I(X_t|G)$ is monotone-decreasing on $t\in[0,\infty)$ (convolution inequality for Fisher information), we have the inequality $
D(X_t |G) \geq D(X|G) - t I(X|G)$.
Thus, to finish the proof, put $t = \frac{D(X|G)}{2 I(X|G)}$ and note that $1-e^{-x}\geq\tfrac{1}{\sqrt{2}}x$ for $x\in [0,1/4]$, which applies for our choice of $t$ due to the log Sobolev inequality, i.e., $\tfrac{1}{2}I(X|G)\geq D(X|G)$.
\end{proof}
\subsection{Beyond radial symmetry}
Theorem \ref{thm:Hjump} can be immediately extended to distributions that are radially symmetric, modulo an affine transformation. Indeed, in this case, we can apply the appropriate linear transformation $X \mapsto A (X - \mu)$, and then invoke the formulae for the behavior of entropy under such transformations to bound the desired entropy jump. A similar statement may be made for Theorem \ref{thm:FIjumps}.
Less immediately, there is potential to bound the entropy jumps associated to a general non-symmetric random vector $X$ by considering its symmetric decreasing rearrangement $X^{\star}$ (see, e.g., \cite{wang2014beyond} for definition). In this case, it is known that $h(X) = h(X^{\star})$ and $h(X+X_* ) \geq h( X^{\star}+X^{\star}_*)$ for $X,X_*$ independent \cite{wang2014beyond}. Thus, we might expect to be able to bound the entropy jump $h(\tfrac{1}{\sqrt{2}}(X+X_*) ) - h(X)$ from below by $D(X^{\star})$. There are two issues here that must be considered in order to proceed along this route. The first issue is fundamental: we cannot expect to bound $D(X^{\star})$ by $D(X)$ in general (e.g., consider $X$ to be a rearrangement of $G$). The second issue is technical and arises from the assumed regularity conditions. In particular, one would need to establish conditions under which the regularity of $X^{\star}$ can be appropriately controlled by the regularity of $X$ (either in the sense of $c$-regularity or relative Fisher information).
In any case, it may be possible to apply the general proof idea to non-symmetric $X$. In particular, radial symmetry of $X$ was only critically used in passing from \eqref{eq1symm} to \eqref{eq2symm}. In general, by the projection property of the score function, we have
\begin{align}
&\mathbb{E}\left| 2 \Pi(W) \rho_W(W) - \Pi(X-X_*)\left( \rho(X) - \rho(X_*)\right) \right|^2 \notag\\
&=
\mathbb{E}\left| \Pi(X-X_*) \left( \rho(X) - \rho(X_*)\right) \right|^2 - 4 \mathbb{E}\left| \Pi(W) \rho_W(W) \right|^2,
\end{align}
so it would be sufficient to bound
\begin{align}
4 \mathbb{E}\left| \Pi(W) \rho_W(W) \right|^2 < \alpha \, \mathbb{E}\left| \Pi(X-X_*) \left( \rho(X) - \rho(X_*)\right) \right|^2
\end{align}
for some $\alpha<1$ (possibly depending on the distribution of $X$) in order for the proof to carry over to general $X$. Note that the distribution of $W = X-X_*$ is symmetric in general, but not radially symmetric, which may be useful toward establishing such a bound.
\section{Estimate of the Deficit in the Log Sobolev Inequality }\label{sec:LSI}
The problem of quantitatively bounding the deficit in the logarithmic Sobolev in equality has received considerable attention lately (e.g., \cite{fathi2014quantitative, bobkov2014bounds, indrei2013quantitative}), but the problem of obtaining a satisfactory estimate that is dimension-free remains open in general. In this section, we apply our previous results to bound the deficit in the LSI for radially symmetric functions.
For a random vector $X$ on $\mathbb{R}^d$, define the entropy power of $X$ as
\begin{align}
N(X) = \frac{1}{2\pi e} \exp\Big( {\tfrac{2}{d}h(X)}\Big).
\end{align}
The following entropy power inequality was proved by the author in \cite{courtade2016strengthening}:
\begin{theorem}
Let $X,X_*,Z$ be independent random vectors on $\mathbb{R}^d$, with $Z$ being Gaussian. Then
\begin{align}
N(X+X_*+Z)N(Z) + N(X)N(X_*) \leq N(X+Z)N(X_*+Z).
\end{align}
\end{theorem}
Particularizing to the case where $Z\sim N(0,\mathrm{I})$, we have the immediate corollary:
\begin{align}
\frac{N(X+\sqrt{t}Z)N(X_*+\sqrt{t}Z) - N(X)N(X_*)}{t}\geq N(X+X_*+ \sqrt{t} Z) \geq N(X+X_*).
\end{align}
Letting $t\to 0$ and applying de Bruin's identity gives the inequality:
\begin{align}
d N(X+X_*) \leq N(X)N(X_*)\left(J(X) + J(X_*) \right).
\end{align}
Supposing $X,X_*$ are identically distributed, we obtain the following improvement of Stam's inequality \cite{stam1959some}, which states that $N(X)J(X) \geq d$:
\begin{align}
\tfrac{1}{d}N(X) J(X) \geq \exp\left\{\frac{2}{d}\left( h(\tfrac{1}{\sqrt{2}}(X+X_*))-h(X) \right) \right\} . \label{improvedStam}
\end{align}
By work of Carlen \cite{carlen1991superadditivity}, it is well-known that Stam's inequality is equivalent to Gross' log Sobolev inequality for the Gaussian measure \cite{gross1975logarithmic}; i.e., $\tfrac{1}{2}I(X|G) \geq D(X|G)$. Using the inequality $\log x\leq x-1$, it is straightforward to convert \eqref{improvedStam} into the following improved log Sobolev inequality:
\begin{align}
\delta_{\mathsf{LSI}}(X) \geq \left( h(\tfrac{1}{\sqrt{2}}(X+X_*))-h(X) \right), \label{improvedLSI}
\end{align}
where $\delta_{\mathsf{LSI}}(X) := \frac{1}{2}I(X|G) - D(X|G)$ denotes the deficit in the log Sobolev inequality.
In view of Corollary \ref{cor:HjumpNoRegularity}, we have established the following logarithmic Sobolev inequality in quantitative form for radially symmetric $X$:
\begin{theorem}\label{thm:stabilityNoReg}
Let $X$ be a radially symmetric random vector on $\mathbb{R}^d$, $d\geq 2$, with $D(X)<\infty$. It holds that
\begin{align}
\delta_{\mathsf{LSI}}(X) \geq \sup_{\varepsilon>0 }\widetilde{C}_{\varepsilon}(X)
\frac{ D(X)^{1+3\varepsilon}}{I(X)^{2\varepsilon}},
\end{align}
where $\widetilde{C}_{\varepsilon}(X)$ is defined as in \eqref{CtildeDefn}.
\end{theorem}
Theorem \ref{thm:stabilityNoReg} yields a quantitative stability result for the LSI whenever $\mathbb{E}|X|^{p}<\infty$ for some $p>4$. As an illustrative example, suppose $X$ is normalized such that $\mathbb{E}|X|^2=d$ (so that $D(X)=D(X|G)$ and $I(X)=I(X|G)$) and $\mathbb{E}|X|^8<\infty$. Then, a clean bound is obtained by setting $\varepsilon=1/2$, rearranging, solving a quadratic inequality and bounding constant terms:
\begin{align}
\frac{1}{2} I(X) %
&\geq D(X) \frac{1}{2}\left(1+
\sqrt{1+ \frac{1}{10^8} \frac{ \,\, \||X|^2\|^{5/2}_1 }{ %
\left\| |X|^2 \right\|^2_{ 4} } \sqrt{ D(X)} }\right) .
\end{align}
Note that this is a very strong stability result.
Indeed, it is a straightforward exercise (identify $\||X|^2\|_1 = d$, apply Minkowski's inequality to $\left\| |X|^2 \right\|_{ 4}$ and simplify) to show that
\begin{align}
D(X) \leq 10^8 \max\left\{ \delta_{\mathsf{LSI}}(X), \delta^{1/2}_{\mathsf{LSI}}(X) \sqrt{\frac{\mathbb{E}|X_1|^8}{d}}\right\}.
\end{align}
We emphasize that $\mathbb{E}|X_1|^8$ is the eighth moment of the \emph{one-dimensional} random variable $X_1$, the first coordinate of $X$. Hence, we would generally expect that $\mathbb{E}|X_1|^8 \ll d$ in high dimension.
If $X$ is known to be $c$-regular for some $c$, then we can establish the following stability estimate, which has no explicit dependence on dimension:
\begin{theorem}\label{thmLSIdeficit}
Let $X$ be a radially symmetric random vector on $\mathbb{R}^d$, $d\geq 2$, with $c$-regular density $f$. For any $\varepsilon > 0$
\begin{align}
\delta_{\mathsf{LSI}}(X) &\geq \frac{1}{4} K_{\varepsilon}(X) I(X)^{1+\varepsilon},
\end{align}
where $K_{\varepsilon}(X)$ is as defined in \eqref{KpX_FI}.
\end{theorem}
\begin{proof}
All quantities in \eqref{improvedLSI} are invariant to translations of $X$, so we assume without loss of generality that $\mathbb{E} X=0$ for the moment. In this case, we may write
\begin{align}
\delta_{\mathsf{LSI}}(X) &\geq D(X|G) - D( \tfrac{1}{\sqrt{2}}(X+X_*) |G) \\
&= -\delta_{\mathsf{LSI}}(X) + \frac{1}{2}\left( I(X|G) - I( \tfrac{1}{\sqrt{2}}(X+X_*) |G) \right) + \delta_{\mathsf{LSI}}( \tfrac{1}{\sqrt{2}}(X+X_*))\\
&\geq -\delta_{\mathsf{LSI}}(X) + \frac{1}{2}\left( J(X) - J( \tfrac{1}{\sqrt{2}}(X+X_*) ) \right) .
\end{align}
Thus, in view of Theorem \ref{thm:FIjumps}, we have proved the claim.
\end{proof}
\begin{remark}
Although Theorems \ref{thm:stabilityNoReg} and \ref{thmLSIdeficit} consider radially symmetric densities, we see from \eqref{improvedLSI} and the following inequalities that $\delta_{\mathsf{LSI}}(X)$ dominates entropy production and dissipation of Fisher information in general. Stated another way, any quantitative lower bound on entropy production (or Fisher information dissipation) will provide a lower bound on $\delta_{\mathsf{LSI}}(X)$.
\end{remark}
In closing, we mention here that a quantitative form of the sharp Sobolev inequality was established by Cianchi, Fusco, Maggi and Pratelli \cite{cianchi2009sharp} through a reduction of the general inequality to the setting of radially symmetric functions by considering spherically symmetric rearrangements. On this note, from \cite[Corollary 8.7]{wang2014beyond}, it is easy to verify that $\delta_{\mathsf{LSI}}(X) \geq \delta_{\mathsf{LSI}}(X^{\star})$, where as before, $X^{\star}$ is the symmetric decreasing rearrangement of $X$. Provided $X^{\star}$ has regular density, we obtain
\begin{align}
\delta_{\mathsf{LSI}}(X) &\geq \frac{1}{4} K_{\varepsilon}(X^{\star}) I(X^{\star} )^{1+\varepsilon}.
\end{align}
This should be compared to a result by Bobkov, Gozlan, Roberto and Samson \cite{bobkov2014bounds} (see also \cite{indrei2013quantitative}): If $X$ is a random vector on $\mathbb{R}^d$ with smooth density $p = e^{-V}$ satisfying $V''\geq \varepsilon \mathrm{I}$ for some $\varepsilon>0$, then
\begin{align}
\delta_{\mathsf{LSI}}(X) &\geq c_{\varepsilon} W_2^2(\bar{X},G),
\end{align}
where $c_{\varepsilon}$ is a constant depending on $\varepsilon$, only, and $\bar{X}$ corresponds to a rearrangement of the vector $X$ such that its one-dimensional marginals $\bar{X}_1, \bar{X}_2, \dots, \bar{X}_d$ form a martingale \cite{bobkov2014bounds}. We remark that the log-concavity assumption can be more restrictive than our regularity assumptions, but the symmetry of $\bar{X}$ is less restrictive than radial symmetry. We also mention that Fathi, Indrei and Ledoux \cite{fathi2014quantitative} have established another quantitative estimate on $\delta_{\mathsf{LSI}}(X)$, under the assumption that $X$ satisfies a Poincar\'e inequality. Namely, if $X$ is a centered random vector that has spectral gap $\varepsilon$, then
\begin{align}
\delta_{\mathsf{LSI}}(X) &\geq c_{\varepsilon} I(X),
\end{align}
where $c_{\varepsilon}$ is a constant depending on $\varepsilon$, only. It is interesting to recall from our above discussion that stability of entropy jumps (in dimension one) has previously been established under a spectral gap condition \cite{ball2003entropy}. We have now seen in \eqref{improvedLSI} that the two stability problems are closely connected.
\section*{Appendix: Proof of Propositions \ref{prop:StartRegular} and \ref{prop:approxR}}
\begin{proof}[Proof of Proposition \ref{prop:StartRegular}]
First, suppose $X$ is $c$-regular, and let $V = X+Z$, $Z\sim N(0,\sigma^2 I)$. We claim that $V$ is $(5c)$-regular. Toward this end, let $f_X$, $f_Z$ and $f_V$ denote the densities of $X,Z,V$, respectively. Now,
\begin{align}
\left| \nabla f_V(v) \right| &=\left| \int (\nabla f_X(v-z)) f_Z(z) dz \right|\\
&\leq \int \left| \nabla f_X(v-z)\right| f_Z(z) dz \\
&\leq c \int f_X(v-z) \left( \left|v-z\right| +\mathbb{E}|X| \right) f_Z(z) dz \\
&\leq c\, f_V(v) \left(|v| + \mathbb{E}|X| \right) + c\, \int \left| z\right| f_X(v-z) f_Z(z) dz \\
&= c\, f_V(v) \left(|v| + \mathbb{E}|X| \right) + c\, f_V(v)\, \mathbb{E}\left[ \,|X-v| \, |V=v \right] \\
&\leq c\, f_V(v) \left(|v| + \mathbb{E}|X| \right) + c\, f_V(v)\, \left( 3 |v| + 4 \mathbb{E}|X| \right) \\
&\leq 5 c\, f_V(v) \left(|v| + \mathbb{E}|V| \right).
\end{align}
A proof of the only nontrivial inequality $\mathbb{E}\left[ \,|X-v| \, |V=v \right] \leq 3 |v| + 4 \mathbb{E}|X|$ can be found in \cite[Proposition 2]{polyWuWasserstein2016}. Since $V$ is smooth with nonvanishing density, $\nabla \log f_V = \frac{\nabla f_V }{f_V}$; this completes the proof of the claim.
By a change of variables, we observe that $e^{-t} X$ is $(c\,e^{2t})$-regular. Combining with the previous claim finishes the proof.
\end{proof}
\begin{remark}
The claim of Proposition \ref{prop:StartRegular} may be strengthened. Indeed, if $X$ is $c$-regular, then $X_t$ is $(5c + 4 )$-regular for all $t\geq 0$. To see this, note that Proposition \ref{prop:addNoise} establishes that $X_t$ is $(4 /
(1-e^{-2t}))$-regular. However, Proposition \ref{prop:StartRegular} shows that $X_t$ is $(5c\,e^{2t})$-regular; maximizing the minimum of these two quantities over $t\geq 0$ establishes the strengthened claim. The weaker claim is more convenient for our purposes, so that regularity of $X_t$ is a multiple of $c$. This ensures that the inequalities are invariant to scaling of $X$.
\end{remark}
\begin{proof}[Proof of Proposition \ref{prop:approxR}]
The proof is elementary, but provided here for completeness. Define $X_0 = \sqrt{d}R_0 U$, where as before $U$ is uniform on $\mathbb{S}^{d-1}$ and independent of $R_0$. Now, let
\begin{align}
X = \sqrt{1-\varepsilon}X_{0} + \sqrt{\varepsilon} G,
\end{align}
and set $R = \frac{1}{\sqrt{d}} |X|$. Clearly, $\mathbb{E} R^2 = 1$ and, by Proposition \ref{prop:addNoise}, $X$ is $(4/\varepsilon)$-regular. Now, observe that
\begin{align}
F_{R}(r) &= \Pr\left\{\tfrac{1}{\sqrt{d}} | \sqrt{1-\varepsilon}X_{0} + \sqrt{\varepsilon} G | \leq r \right\}\\
&\geq \Pr\left\{ \tfrac{\sqrt{1-\varepsilon}}{\sqrt{d}}|X_{0}| + \tfrac{\sqrt{\varepsilon}}{\sqrt{d}}|G| \leq r \right\}\\
&\geq \Pr\left\{ \tfrac{\sqrt{1-\varepsilon}}{\sqrt{d}} |X_{0}| \leq r -\sqrt{(t+1)\varepsilon} \right\}\Pr\left\{ \tfrac{\sqrt{\varepsilon}}{\sqrt{d}}|G| \leq \sqrt{(t+1)\varepsilon} \right\}\\
&= F_{R_0}\left( \tfrac{ r -\sqrt{(t+1)\varepsilon}}{\sqrt{1-\varepsilon}} \right) \left(1- \Pr\left\{ \tfrac{1}{{d}} |G|^2 -1 > t\right\}\right)\\
&\geq F_{R_0}\left( \tfrac{ r - \sqrt{(t+1)\varepsilon}}{\sqrt{1-\varepsilon}} \right) - e^{-d t^2 /8} ,
\end{align}
where the final inequality is due to the tail bound for a chi-squared random variable with $d$ degrees of freedom. The other direction is similar:
\begin{align}
F_{R}(r) &= \Pr\left\{\tfrac{1}{\sqrt{d}} | \sqrt{1-\varepsilon}X_{0} + \sqrt{\varepsilon} G | \leq r \right\}\\
&\leq \Pr\left\{ \tfrac{\sqrt{1-\varepsilon}}{\sqrt{d}}|X_{0}| - \tfrac{\sqrt{\varepsilon}}{\sqrt{d}}|G| \leq r \right\}\\
&\leq \Pr\left\{ \tfrac{\sqrt{1-\varepsilon}}{\sqrt{d}} |X_{0}| \leq r +\sqrt{(t+1)\varepsilon} \right\} + \Pr\left\{ \tfrac{\sqrt{\varepsilon}}{\sqrt{d}}|G| > \sqrt{(t+1)\varepsilon} \right\}\\
&= F_{R_0}\left( \tfrac{ r + \sqrt{(t+1)\varepsilon}}{\sqrt{1-\varepsilon}} \right) + e^{-d t^2 /8}.
\end{align}
\end{proof}
\subsection*{Acknowledgment}
This work was supported in part by NSF grants CCF-1528132 and CCF-0939370 (Center for Science of Information).
\bibliographystyle{unsrt}
|
\section{Abstract}
Dynamic behavior of an isotropic Ashkin-Teller model in the presence of a
periodically
oscillating magnetic field has been analyzed by means of the mean field
approximation.
The dynamic equation of motion has been constructed with the
help of a Glauber type stochastic process and solved for a square lattice.
After defining the possible dynamical phases of the system, phase diagrams have been
given and the behavior of the hysteresis
loops has been investigated in detail. The hysteresis loop for specific order
parameter of isotropic
Ashkin-Teller model has been defined and characteristics of this loop in
different dynamical phases have been given.
Keywords: \textbf{Dynamic isotropic Ashkin-Teller model;
hysteresis loops; hysteresis loop area}
\section{Introduction}\label{introduction}
Ashkin-Teller Model (ATM) has been introduced for description of the cooperative
phenomena of quaternary
alloys \cite{ref1}. It has four states per site and may be useful to describe
magnetic systems with two easy axes.
The ATM is a ‘staggered’ version of the eight-vertex model \cite{ref2}. In two
dimension, the ATM can be mapped onto
a staggered eight-vertex model at the critical point. The non-universal
critical behavior along a self-dual line, where the exponents vary continuously
\cite{ref3},
is one of the interesting critical property of the model. On the other hand it
has been shown that,
three dimensional model has much richer phase diagrams than the ATM in two
dimension \cite{ref4}.
There appear some first-order phase transitions and continuous phase
transitions,
even an XY -like transition and a Heisenberg-like multicritical point.
It has been shown that ATM could be described in Hamiltonian form appropriate
for spin systems \cite{ref5}. In this form, the model can be viewed as two
coupled Ising models which is named as 2-color ATM. If two of these Ising models
are identical then the model named as isotropic Ashkin-Teller model (IATM),
otherwise the model is anisotropic Ashkin-Teller model (AATM). In a similar
manner, ATM that formed by $N$ coupled Isig model entitled as N-color ATM as
introduced in \cite{ref6}.
One of the well known physical realizations for this model is the compound of
Selenium adsorbed on a Ni
surface \cite{ref7}. ATM can be used to describe chemical interactions in
metallic alloys \cite{ref8},
thermodynamic properties in superconducting cuprates (ATM represents the
interactions between orbital current loops
in $CuO_2$ -plaquettes) \cite{ref9} and elastic response of DNA molecule to
external force and torque \cite{ref10}. Besides, oxygen ordering in $YBa_2 Cu_3
O_z$
may also be understood in analogy with the two-dimensional IATM
\cite{ref11,ref12,ref13}. Also, ATM has many interesting applications in neural
networks \cite{ref14} and cosmology \cite{ref15}. Besides, some mappings between
the ATM and some other models are possible. This makes the ATM valuable in a
theoretical manner. For instance, the random N-color quantum ATM can be
described by an $O(N)$ Gross-Neveu model with random mass \cite{ref16}.
Similarly the relation between the two-dimensional N-component Landau-Ginzburg
Hamiltonian with cubic anisotropy and N-color ATM has been discussed in
\cite{ref17,ref18}.
Critical properties of the ATM has been widely investigated in literature.
IATM has been investigated by mean field renormalization group technique
\cite{ref19,ref20}, Monte Carlo Simulation (MC)
\cite{ref21,ref22,ref23,ref24,ref25,ref26,ref27}, effective field theory (EFT)
\cite{ref28}, transfer-matrix finite-size-scaling method \cite{ref29}, high and
low temperature series expansion and MC \cite{ref30}, MC and renormalization
group technique \cite{ref31,ref32} and damage spreading simulation
\cite{ref33,ref34,ref35}.
On the other hand, the AATM has been investigated within several techniques such
as, mean field approximation (MFA) \cite{ref36}, MC \cite{ref37,ref38}, real-space
renormalization-group approach \cite{ref39} and Monte Carlo renormalization
group technique \cite{ref40}.
Theoretical works devoted to ATM is not limited to regular (translationally
invariant lattices) lattices. For instance, the model has been solved on Bethe
Lattice \cite{ref41}, diamond-like hierarchical lattice \cite{ref42,ref43} and
Cayley tree \cite{ref44}.
Also, some extensions and variants of the model have been studied. Extended ATM
(ATM with reduced degeneracy) has been introduced \cite{ref46} and solved within
the MFA \cite{ref47}, mean field renormalization group technique
\cite{ref48} and MC \cite{ref49}. Some other extensions like ATM with
Dzyaloshinskii –Moriya interaction \cite{ref50}, ATM with spin-1
\cite{ref51,ref52}, as well as mixed spin ATM \cite{ref45} can be found in the
literature.
Likewise, N-color ATM introduced in \cite{ref6} has been investigated within
the transfer matrix analysis \cite{ref53}, renormalization group technique
\cite{ref54}, Monte Carlo renormalization group technique \cite{ref55}, MFA
and MC \cite{ref56}. This model has also been solved exactly
for large N in 2D \cite{ref57,ref58}.
Although, magnetic systems under a time dependent external magnetic field has
been attracted much interest from both theoretical and experimental points of
view; to the best of our knowledge, ATM under the time dependent magnetic field
has not yet been studied. It will be interesting to investigate the dynamic
character of ATM, which has rich critical properties in the static case, and
obtain the corresponding dynamic phases of the system.
Dynamic phase transition (DPT) in magnetic systems comes from the competition
between the relaxation time of the system and period of the driving periodic
external magnetic field \cite{ref59}. The time average of the magnetization
over a full period of the oscillating magnetic field can be used as
dynamic order parameter (DOP) of the system. On the other hand appeared
hysteresis behavior
from the delay of the response of the system to the driving cyclic force
includes important clues of the dynamic character of the system.
From the experimental point of view,
DPTs and hysteresis behaviors can be observed
experimentally in different types of magnetic systems.
Experiments on ultrathin Co films \cite{ref60}, Fe/Au(001)
films \cite{ref61}, epitaxial Fe/GaAs(001) thin
films \cite{ref62}, fcc Co(001), and fcc NiFe/Cu/Co(001) layers \cite{ref63}
Fe/InAs(001) ultrathin films \cite{ref64} are among them.
On the theoretical side, there has been growing interest in the DPT which was
first observed within the MFA \cite{ref65} for the s-$1/2$ Ising model. Since that time,
DPT and hysteresis behaviors of the s-$1/2$ Ising model have been widely studied
within the
several techniques such as MFA \cite{ref66}, Monte Carlo
simulation (MC) \cite{ref67},
effective field theory (EFT) \cite{ref68}.
The aim of this work
is to investigate the IATM under a magnetic field
oscillating in time within the MFA formulation
and Glauber type of stochastic process \cite{ref71}. For this aim the paper is
organized
as follows: In Sec. \ref{formulation} we briefly present the
model and formulation. The results and discussions are
presented in Sec. \ref{results}, and finally Sec. \ref{conclusion} contains our
conclusions.
\section{Model and Formulation}\label{formulation}
The Hamiltonian of the dynamical IATM is given by
\eq{denk1}{\mathcal{H}=-J_2\summ{<i,j>}{}{\paran{\sigma_i \sigma_j+s_i s_j}}
-J_4\summ{<i,j>}{}{\sigma_i \sigma_j s_i s_j}-
H(t)\summ{i}{}{\paran{\sigma_i+s_i}},}
where the first two summations are over the nearest neighbors of the lattice,
while the
last one is over all the lattice sites. Here, $\sigma_i$ and $s_i$ are the $z$
components of the spin variables at a site $i$, $J_2$ is the bilinear exchange
interactions of both of the Ising models. $J_4$ is the four spin interaction,
which couples two Ising models to produce the IATM. Dynamical character of the
model comes from the time dependent external longitudinal magnetic field $H(t)$
and it is given by
\eq{denk2}{H(t)=H_0\sin{\paran{\omega t}},}
where $H_0$ is the amplitude and $\omega$ is the angular frequency of
the periodic magnetic field and $t$ stands for the time.
In the static case (i.e. $H_0=0$) the model turns to the IATM, which is
equivalent to the four component Potts model for $J_2=J_4$ \cite{ref72}. When
$J_4=0$ the model reduces to two independent Ising models.
We use a Glauber-type stochastic process \cite{ref71} to
investigate dynamic properties of the considered system. In general, in the
Glauber type of stochastic process (as done in Ref. \cite{ref65} for the
Ising model) the thermal average (denoted with $\sandd{}$) of a spin variable
$\tau_i$, which can take values $\pm 1$
can be given as
\eq{denk3}{
\theta \tur{\sandd{\tau_i}}{t}=-\sandd{\tau_i}+\sandd{\frac{Tr_i \tau_i
\exp{\paran{-\beta \mathcal{H}_i}}}{Tr_i \exp{\paran{-\beta \mathcal{H}_i}}}}
}
in this type of process. Here, $\theta$ is the transition rate per
unit time, $\beta=1/(k_BT)$, $k_B$ and $T$ denote the Boltzmann constant and
temperature, respectively. $Tr_i$ stands for the trace operation over the site
$i$. Also $\mathcal{H}_i$ denotes the part of the Hamiltonian of the system
related to the site $i$, which is given by,
\eq{denk4}{
\mathcal{H}_i=-J_2\sigma_i\summ{j}{}{ \sigma_j}-J_2s_i\summ{j}{}{ s_j}
-J_4\sigma_i s_i\summ{j}{}{ \sigma_j s_j}-
H(t)\paran{\sigma_i+s_i},
} where all summations are carried over the nearest neighbor sites of the site
$i$.
In order to handle these spin-spin interactions, usual approximation can be
adopted, that is; all spin-spin interactions of these variables are represented by
local fields as
\eq{denk5}{
h_1=J_2\summ{j=1}{z}{\sigma_j},\quad h_2=J_2\summ{j=1}{z}{s_j},\quad
h_4=J_4\summ{j=1}{z}{\sigma_j s_j},
} where $z$ is the coordination number (i.e. number of nearest neighbor sites of
any site $i$) of the lattice. Then one spin cluster Hamiltoian gets the form
\eq{denk6}{
\mathcal{H}_i=-\sigma_i h_1-s_i h_2
-\sigma_i s_i h_4-\paran{\sigma_i +s_i}H(t)
}
By writing Eq. \re{denk6} in Eq. \re{denk3} and performing the $Tr_i$ operations
we can get equations for $\tau_i=\sigma_i,s_i$ and $\sigma_i s_i$ as
$$
\theta
\tur{\sandd{\sigma_i}}{t}=-\sandd{\sigma_i}+\sandd{F\paran{x,y,z}|_{x=h_1+H(t),
y=h_2+H(t),z=h_4}}
$$
\eq{denk7}{
\theta
\tur{\sandd{s_i}}{t}=-\sandd{s_i}+\sandd{F\paran{x,y,z}|_{x=h_2+H(t),y=h_1+H(t),
z=h_4}}
}
$$
\theta \tur{\sandd{\sigma_i s_i}}{t}=-\sandd{\sigma_i
s_i}+\sandd{F\paran{x,y,z}|_{x=h_4,y=h_1+H(t),z=h_2+H(t)}}
$$ where
\eq{denk8}{
F\paran{x,y,z}=\frac{\tanh{(\beta x)}+\tanh{(\beta y)}\tanh{(\beta
z)}}{1+\tanh{(\beta x)}\tanh{(\beta y)}\tanh{(\beta z)}}
}
The simplest way to one calculate Eq. \re{denk7} is adopt MFA. Within this approximation, all local fields defined in Eq.
\re{denk5} written in terms of the thermal averages of the spin variables,
\eq{denk9}{
h_1=zJ_2m_\sigma,\quad h_2=zJ_2m_s,\quad h_4=zJ_4m_{\sigma s},
} where
\eq{denk10}{
m_\sigma=\sandd{\sigma_j}, m_s=\sandd{s_j}, m_{\sigma s}=\sandd{\sigma_j s_j}.
} Here, the translationally invariance property of the lattice is adopted, i.e.
all lattice sites are equivalent. Then Eq. \re{denk7} gets the form
$$
\theta \tur{m_\sigma}{t}=-m_\sigma+F\paran{x,y,z}|_{x=h_1+H(t),y=h_2+H(t),z=h_4}
$$
\eq{denk11}{
\theta \tur{m_s}{t}=-m_s+F\paran{x,y,z}|_{x=h_2+H(t),y=h_1+H(t),z=h_4}
}
$$
\theta \tur{m_{\sigma s}}{t}=-m_{\sigma
s}+F\paran{x,y,z}|_{x=h_4,y=h_1+H(t),z=h_2+H(t)}
$$
We can obtain explicit form of the
dynamic equation of motion Eq. \re{denk11}, by writing local fields defined in
Eq. \re{denk9} into the function Eq. \re{denk8} given with order in Eq.
\re{denk11}. Eq. \re{denk11} is a coupled first order differential equation
system and can be solved with standart methods such as Runge-Kutta method
\cite{ref73}. This iterative method starts by some initial values of the order
parameters ($m_\nu(0)$, $\nu=\sigma,s,\sigma s$) and results in the desired solution after the convergency criteria
$m_\nu\paran{t}=m_\nu\paran{t+2\pi/\omega}$ is satisfied.
By this way we can obtain DOP as
\eq{denk12}{Q_\nu=\frac{\omega}{2\pi}\intego{}{}{m_\nu\paran{t}dt}} where
$m_\nu\paran{t}$
is a stable and periodic function. Dynamical critical points can be determined
by obtaining the variation of the $Q_\nu$
with temperature for given set of Hamiltonian parameters.
We can construct the hysteresis loops which are nothing but the variation
of the $m_\nu(t)$ with $H(t)$ in one period of the periodic magnetic field.
Hereafter, once the hysteresis loop is determined, some quantities about
it can be calculated. One of them is dynamical hysteresis loop area (HLA) and
can be calculated
via integration over a complete cycle of the magnetic field,
\eq{denk13}{A_\nu=\intego{}{}{}m_\nu\paran{t}dH} and corresponds to the energy
loss due to the hysteresis.
\section{Results and Discussion}\label{results}
In order to determine DPT and hysteresis characteristics of the system, let us
scale all Hamiltonian parameters with $J_2$, i.e. the unit of energy is $J_2$,
\eq{denk21}{K_4=\frac{J_4}{J_2}, \tau=\frac{k_BT}{J_2},
h_0=\frac{H_0}{J_2},h(t)=\frac{H(t)}{J_2}.
}
Our investigation will be restricted to square ($z=4$) lattice.
We set $\theta=1$ throughout our numerical calculations.
We note that, under the transformation $\sigma_i\rightarrow s_i$, Hamiltonian
and the formulation used here do not change, then it will be enough to
investigate only one of the $m_\sigma$ or $m_s$, due to the fact that
$m_\sigma=m_s$.
Different phases of the static model is well known:
\begin{itemize}
\item Ferromagnetic (Baxter) phase: All magnetizations are different from zero.
\item Paramagnetic phase: All magnetizations are zero.
\item $\sandd{\sigma s}, \sandd{\sigma}, \sandd{ s}$ phases : All magnetizations
are zero except the magnetization that have index which is the name of the
phase. For instance in $\sandd{\sigma}$ phase only $m_\sigma$ is different from
zero, while other two are equal to zero.
\end{itemize}
We investigate corresponding dynamical phases in this work. Three different
dynamical phases (corresponding to the phases of the static IATM) will manifest
themselves. Also, as discussed in Ref. \cite{ref65} and successive works related to
the DPT in Ising model, some regions may appear in the phase diagram, which is
overlap region of the dynamically ordered and disordered phases. This overlap
region mostly occur for relatively large values of $h_0$ for the Ising model.
Let us enumerate these phases with prefix DP (which stands for dynamical phase).
\begin{itemize}
\item DP1: All DOPs are different from zero (corresponds to the Baxter phase in
static the case )
\item DP2: All DOPs are equal to zero (corresponds to the paramagnetic phase in
static the case )
\item DP3: Only $m_{\sigma s}$ is different from zero while $m_\sigma=m_s=0$
(corresponds to the $\sandd{\sigma s}$ phase in static the case )
\item DP4: Overlap of the DP1 and DP2 (or DP3) phases.
\end{itemize}
These phases and borders between them can be determined by calculating
magnetizations defined in Eq. \re{denk10} by using formulation presented in this
work. Typical time series corresponding to these four different phases can be seen
in Fig. \ref{sek1}.
\begin{figure}[h]\begin{center}
\epsfig{file=man_seri_ornek-3.7-hh.ps, width=7.2cm}
\epsfig{file=man_seri_ornek-4.1-hh.ps, width=7.2cm}
\epsfig{file=man_seri_ornek-4.2-hh.ps, width=7.2cm}
\epsfig{file=man_seri_ornek-15.0-hh.ps, width=7.2cm}
\end{center}
\caption{Time series of the order parameters of the IATM which corresponding to
(a) DP1, (b) DP4, (c) DP3 and (d) DP2. All time series are calculated for
the Hamiltonian parameters $K_4=0.5, \omega=1.0, h_0=0.5$. Initial values of the
order parameters and temperature are shown in each figure. In each figure,
the thicker lines correspond to $m_\sigma=m_s$ and the thinner lines correspond to
$m_{\sigma s}$. $\nu$ stands for $\sigma, s, \sigma s $. }
\label{sek1}\end{figure}
Although the values of the magnetizations in the DP1, DP2 and DP3 phases are not
dependent on the initial values of the magnetizations, as one can see from to Fig.
\ref{sek1} (b) that, this does not hold for the DP4 phase. The value of the
$m_\sigma=m_s$ converges zero or specific nonzero values depending on the
initial value. In other words, formulation gives two different values (such that
one of them is zero) for $m_\sigma=m_s$ depending on the choice of the
initial values.
The phase diagram for static IATM in $(K_4,\tau)$ plane can be seen in Fig.
\re{sek2}. This diagram can be obtained by numerical solutions of the Eq.
\re{denk11} in the static limit. The same phase diagram obtained within the RG
\cite{ref19,ref20} and MC \cite{ref25} can be found in the literature. In the static IATM,
for the higher values of the $K_4$, $\sandd{\sigma s}$ phase may appear when
the temperature rises, as seen in Fig. \ref{sek2}.
\begin{figure}[h]\begin{center}
\epsfig{file=phase_diag-omega-2.0-hh-0.0.ps, width=10.0cm}
\end{center}
\caption{Phase diagram of the static IATM in the $(K_4,\tau)$ plane. }
\label{sek2}\end{figure}
\subsection{Dynamic Phase Boundaries}
Dynamic Phase Boundaries (DPB) of the model separate the different dynamic
phases of the model. The relation of the dynamic behavior of the system with
driving dynamical magnetic field is well known \cite{ref59}. The relation
between the exchange interaction, temperature, amplitude and the frequency of
the magnetic field determines the dynamic phase of the system. Ferromagnetic
exchange interaction of the model tends to to keep the nearest neighbors of the spins
of the system parallel to each other. Rising temperature gives rise to the thermal thermal
fluctuations and drive the system to the disordered phases. Besides, in the dynamical
case, the amplitude and the frequency of the magnetic field are also decisive.
Higher amplitude, dictates the spins to align parallel to the field. But when the
frequency is high, spins cannot follow to magnetic field due to the difference
between the period of the magnetic field and relaxation time of the spins.
In order to compare with the static case, let us depict the phase diagrams in
$(K_4,\tau)$ plane (as in Fig. \ref{sek2}) for some selected values of $\omega$
and $h_0$. This can be seen in Fig. \ref{sek3} for low frequency ($\omega=0.1$)
and in Fig. \ref{sek4} for high frequency ($\omega=5.0$), with selected values of
$h_0=0.1,0.2,0.5,1.0$ shown in figures labeled as (a),(b),(c) and (d)
respectively.
\begin{figure}[h]\begin{center}
\epsfig{file=phase_diag-omega-0.1-hh-0.1.ps, width=7.2cm}
\epsfig{file=phase_diag-omega-0.1-hh-0.2.ps, width=7.2cm}
\epsfig{file=phase_diag-omega-0.1-hh-0.5.ps, width=7.2cm}
\epsfig{file=phase_diag-omega-0.1-hh-1.0.ps, width=7.2cm}
\end{center}
\caption{Phase diagrams of the dynamical IATM in a $(K_4,\tau)$ plane for selected
values of $\omega=0.1$ and (a) $h_0=0.1$, (b) $h_0=0.2$, (c) $h_0=0.5$, (d)
$h_0=1.0$. } \label{sek3}\end{figure}
\begin{figure}[h]\begin{center}
\epsfig{file=phase_diag-omega-5.0-hh-0.1.ps, width=7.2cm}
\epsfig{file=phase_diag-omega-5.0-hh-0.2.ps, width=7.2cm}
\epsfig{file=phase_diag-omega-5.0-hh-0.5.ps, width=7.2cm}
\epsfig{file=phase_diag-omega-5.0-hh-1.0.ps, width=7.2cm}
\end{center}
\caption{Phase diagrams of the dynamical IATM in a $(K_4,\tau)$ plane for selected
values of $\omega=5.0$ and (a) $h_0=0.1$, (b) $h_0=0.2$, (c) $h_0=0.5$, (d)
$h_0=1.0$. }
\label{sek4}\end{figure}
As seen in Fig. \ref{sek3}, at low frequencies, rising amplitude shrinks the
$DP_1$ region and shifts the $DP_2$ region towards the higher temperature
region of the $(K_4,\tau)$ plane. Also disappearing of the $DP_4$ phase with
rising $h_0$ draws attention. The same reasoning holds for the high frequency
regime, which can be seen in Fig. \ref{sek4}. For higher frequency the $DP_4$
phase survives in a larger interval of $h_0$, e.g. while for the value of
$h_0=1.0$, high frequency has $DP_4$ phase in $(K_4,\tau)$ plane (Fig.
\ref{sek4} (d)), the same phase don't occur for lower frequency (Fig. \ref{sek3}
(d)). This behavior is similar to the behavior of the dynamical Ising model as
shown in \cite{ref65}.
\subsection{Hysteresis characteristics}
The hysteric response of the system to the varying temperature, frequency and
amplitude of the driven field is well known from the investigations on the Ising
model \cite{ref59}. Thus, in this section we especially want to discuss the
effect of the coupling constant $K_4$ on these hysteresis characteristics.
\begin{figure}[h]\begin{center}
\epsfig{file=hyst_ornek_xx.ps, width=7.4cm}
\epsfig{file=hyst_ornek_om.ps, width=7.4cm}
\epsfig{file=hyst_ornek_hh.ps, width=7.4cm}
\epsfig{file=hyst_ornek_JJ.ps, width=7.4cm}
\end{center}
\caption{Hysteresis behavior of the IATM on a square lattice for changing (a)
temperature, (b) frequency of the field, (c) amplitude of the field and (d) $K_4$,
for selected values of Hamiltonian parameters.} \label{sek5}\end{figure}
Typical behaviors of the hysteresis loops with changing Hamiltonian parameters
can be seen in Fig. \ref{sek5}. If we group hysteresis loops (as for the Ising
model) as paramagnetic and ferromagnetic loops, we observe from Fig. \ref{sek5} that, in the manner of
transition between two kind of loops
\begin{itemize}
\item The effect of the rising temperature and amplitude of the field are
similar, they give rise to transition from ferromagnetic loops to
paramagnetic loops (see Fig. \ref{sek5} (a) and (c))
\item The effect of the rising frequency of the field and $K_4$ are similar,
they give rise to transition from paramagnetic loops to ferromagnetic loops (see Fig. \ref{sek5} (b) and (d)).
\end{itemize}
Rising temperature causes enhanced thermal fluctuations and the spins can
follow the driving periodic magnetic field. The same reasoning holds for the rising
amplitude of the field, since rising amplitude of the field means that more
energy is supplied to the system from the magnetic field. These two mechanisms
cause a transition from the ordered phase to the disordered phase. The reverse
transition can be seen for the rising frequency of the field. After a specific
frequency (which depends on all other Hamiltonian parameters), spins cannot
follow the magnetic field, then ferromagnetic phase occurs. These facts have already
been observed and explained for the dynamical Ising model \cite{ref59}. For the
IATM we observe from Fig. \ref{sek5} (d) that rising $K_4$ can give rise to
transition from the paramagnetic hysteresis loops to the ferromagnetic loops
(compare loops labeled by A,B,C in Fig. \ref{sek5} (d)). This behavior is
consistent with the dynamic phase diagram which was depicted in Fig. \ref{sek3}
(d). As seen in Fig. \ref{sek3} (d) while system is in the $DP_3$ phase for the
temperature value of $\tau=3.8$, rising coupling constant $K_4$ gives rise to a
transition to the $DP_1$ phase which has ferromagnetic hysteresis
characteristics.
Note that there is not any significant difference between the observed hysteresis loops belonging to $DP_2$
and $DP_3$ phases. This is because both of the
phases have $m_\sigma=0$. Thus the phases $DP_2$ and $DP_3$ could not be
distinguished by hysteresis behaviors of order parameter $m_{\sigma}$. But the variation of the $m_{\sigma s} $
with magnetic field in one period may distinguish between these two phases. This loop may be
called as hysteresis loop of $m_{\sigma s} $. In order to determine the response
of this new order parameter (which is absent in the Ising model), we depict
typical loops representing the phases $DP_2, DP_3$ and $DP_1$ in Fig. \ref{sek6}.
The other Hamiltonian parameter values are chosen as $\omega=0.1,
h_0=0.1,K_4=0.2$. We can see from Fig. \ref{sek3} (a) that, the curve
labeled A in Fig. \ref{sek6} represents the hysteresis loop of the phase
$DP_2$, B represents the phase $DP_3$ and C and D correspond to the phase
$DP_1$. First of all, the loop corresponds to the phase $DP_2$ (curve labeled by
A) indeed a $m_{\sigma s}=0$ line, means that the order parameter $m_{\sigma s}=0$ does not
respond to the varying magnetic field. As seen in Fig. \ref{sek6}, the order
parameter $m_{\sigma s}$ starts to reply the changing magnetic field within
the phase $DP_3$ (curve labeled B) and similar loop is valid for the phase
$DP_1$ (curve labeled C). But one difference draws our attention: while the
loop belonging to the phase $DP_3$ has one $m_{\sigma s}$ value for the zero
field, the loop about the $DP_1$ phase has two very close but slightly different
values of $m_{\sigma s}$ for the $h=0$ case. When the temperature is lowered, the loop
like knot becomes dissolved and evolve into the loops like labeled D in Fig. \ref{sek6}. As
a result, we can distinguish the phases $DP_1,DP_2$ and $DP_3$ by looking at the
hysteresis loops of the $m_{\sigma s}$ as explained above.
\begin{figure}[h]\begin{center}
\epsfig{file=plot_hyst_mss_ornek.ps, width=12cm}
\end{center}
\caption{Variation of $m_{\sigma s} $ within the one period of the magnetic
field for selected temperature values as $\tau=8.00,4.50,3.69,3.65$. Other
values of the Hamiltonian parameters are fixed as $\omega=0.1,h_0=0.1, K_4=0.2$. }
\label{sek6}\end{figure}
Lastly we want to elaborate the HLA properties of the system. For this aim we depict the
contour plots of the HLA in ($\omega,K_4$) plane for selected values of
$(h_0,\tau)$ pair, which can be seen in Fig. \ref{sek7}. HLA of the dynamical Ising
model has been widely inspected in the literature and the variation of the HLA with the
frequency is well known. We want to determine the relation between $K_4$ and
the HLA. This can be seen in contour plots for several cases. At first sight
one difference is take our attention between the contour plots given in Fig.
\ref{sek7} (a)-(d). Rising $K_4$ with low frequencies result in lowering of the
HLA in Fig. \ref{sek7} (a) while the reverse relation holds in Figs. \ref{sek7}
(b)-(d). As seen in Fig. \ref{sek3} (d), rising $K_4$ give rise to transition
from the phase $DP_3$ to $DP_1$ and this transition causes depressed HLA. We
cannot face this situation in Fig. \ref{sek7} (b). For low frequencies, rising
$K_4$ enhances the HLA. Again from Fig. \ref{sek3} (d), we can see that, this
variation cannot change the phase of the system, the system lies in the $DP_3$
phase and rising coupling $K_4$ between the two Ising models give rise to
rising HLA. Same situation holds for Figs. \ref{sek7} (c) and (d). The only
difference is wider region of the $(K_4,\omega)$ plane has big HLA due to the
rising $h_0$ enlarges the region that have $DP_2,DP_3$ regions.
\begin{figure}[h]\begin{center}
\epsfig{file=plot_hla-hh-1.0-xx-3.0.ps, width=7.4cm}
\epsfig{file=plot_hla-hh-1.0-xx-5.0.ps, width=7.4cm}
\epsfig{file=plot_hla-hh-2.0-xx-3.0.ps, width=7.4cm}
\epsfig{file=plot_hla-hh-2.0-xx-5.0.ps, width=7.4cm}
\end{center}
\caption{Contour plots of the HLA in ($\omega,K_4$) plane for selected values of
$(h_0,\tau)$ pair as (a) ($1.0,3.0$) (b) ($1.0,5.0$)
(c) ($2.0,3.0$) (d) ($2.0,5.0$). Also contour lines added as starting with the
value of $0.5$ and increment of $0.5$. } \label{sek7}\end{figure}
\section{Conclusion}\label{conclusion}
The phase diagrams and the hysteresis characteristics of the IATM has been
investigated within the MFA with Glauber type of stochastic process.
First, the phase diagrams of the model have been obtained in the $(\tau,K_4)$ plane, by
defining the possible dynamical phases of the system obtainable within the used
approximation. The effect of the frequency and the amplitude of the periodic
time dependent magnetic field on these diagrams has been investigated in
detail.
To the hysteresis part of the work, since the effect of the Hamiltonian
parameters on the hysteresis characteristics on the Isng model is well known and
the effect of the Hamiltonian parameter $K_4$ on the hysteresis characteristics
is mostly investigated. Dynamical phase transitions induced by changing $K_4$
shows itself on the hysteresis behaviors. These behaviors have been discussed in detail.
Besides, the behavior of the order parameter of the IATM namely $m_{\sigma s}$
with the magnetic field has also been discussed. The relation between these hysteresis loops
and possible phase transitions explored.
We hope that the results obtained in this work may be beneficial
form both theoretical and experimental point of view.
\newpage
|
\section{Introduction}
Type Ia supernovae (SNe~Ia) are thermonuclear explosions of white dwarfs (WDs) in binary systems. Nevertheless, the specific
progenitor systems of the SNe Ia have not been identified (Hillebrandt \& Niemeyer 2000; Maoz et al. 2014). Two broad classes
of progenitor binary systems have been suggested: the single-degenerate (SD) scenario (Whelan \& Iben 1973; Nomoto 1982),
in which the WD accretes mass from a non-degenerate stellar companion (e.g., main sequence or red giant star) and explodes
when it exceeds the Chandrasekhar mass; and the double-degenerate (DD) scenario (Iben \& Tutukov 1984; Webbink 1984),
involving the merger of two WDs.
The supernova of 1572 (SN~1572), also widely known as ``Tycho's supernova", is a well-established Type Ia supernova,
verified from its light echoes (Rest et al. 2008; Krause et al. 2008). As one of the few historical SNRs in the Milky Way, Tycho's
SNR has been widely observed over the entire electromagnetic spectrum (see, e.g., Warren et al. 2005; Gomez et al. 2012).
These observations have found in the remnant a complete shell-like structure with a diameter of approximately 8\,arcmin,
produced by the shocks from the SN explosion (see Warren et al. 2005). The distance of Tycho's SNR has been estimated
to be between 2 and 5\,kpc, but recent studies suggest a value closer to 2.5 and 3\,kpc (Tian \& Leahy 2011; Zhang et al. 2013).
Multi-wavelength observations suggested the existence of circumstellar material (CSM), including gas (Reynoso et al. 1999;
Lee et al. 2004; Cai et al. 2009; Xu et al. 2011) and dust (Ishihara et al. 2010; Gomez et al. 2012), in the northeast and east
front of Tycho's SNR shell. Nevertheless, Tian \& Leahy (2011) suggest that the atomic hydrogen (HI) gas at velocities
of $-$47 to $-$53\,km\,s$^{-1}$, which was once considered to be interacting with the shock waves from the explosion (Reynoso
et al. 1999), is located in front of Tycho's SNR. These authors also argued that the CO molecular gas at a velocity
of $\sim$\,$-$\,64\,km\,s$^{-1}$ (Lee et al. 2004; Cai et al. 2009; Xu et al. 2011) is not interacting with the remnant either,
because the derived gas density from the CO observations ($\sim$\,200\,cm$^{-3}$) is much larger than the density referred
from the X-ray observations ($\sim$\,0.2\,cm$^{-3}$; Katsuda et al. 2010). However, the re-analysis of the high-energy observations
(including X- and $\gamma$-ray data) of Tycho's SNR suggested a denser ambient medium ($\sim$\,4-12 cm$^{-3}$
on average; Zhang et al. 2013).
Recently, Zhou et al. (2016) found that the CO gas around the shell is expanding and that there is an enhanced $^{12}$CO\,(2--1)/(1--0)
intensity ratio ($\sim$\,1.6) in the northeast front of the shell (hereafter we refer to the circumstellar CO gas located at the edges
of the 8$'$ diameter shell as the rim). These new observational results provide evidence for the interaction between the supernova
shocks and the molecular gas near the remnant. In addition, this interaction is believed to be responsible for the acceleration of
cosmic-ray protons detected in Tycho's SNR (see, e.g., Zhang et al. 2013).
It should be noted that most of the previous observations toward Tycho's SNR focused on only a limited region (less than 30
arcmin) around the shell-like structure. In this paper, we present large-field (3$^\circ$\,$\times$\,2$^\circ$) and high-sensitivity
CO\,(1--0) molecular line observations toward Tycho's SNR, using the 13.7-meter millimeter-wavelength telescope of the
Purple Mountain Observatory (PMO), which is part of the Milky Way Imaging Scroll Painting (MWISP) project for investigating
the nature of the molecular gas along the northern Galactic Plane. In Section 2 we describe the observations and data reduction.
Observational results are presented in Section 3, discussed in Section 4, and summarized in Section 5.
\section{Observations and data reduction}
\subsection{PMO 13.7-meter CO observations}
The CO\,(1--0) observations toward Tycho's SNR is part of the MWISP CO line survey project operated by the PMO, which were made
from 2011 November to 2016 February with the 13.7-meter millimeter-wavelength telescope of the Qinghai station of PMO at Delingha in
China. The nine-beam Superconducting Spectroscopic Array Receiver (SSAR) was working as the front end in sideband separation mode
(see Shan et al. 2012). Three CO\,($J$ = 1--0) lines were simultaneously observed, $^{12}$CO at the upper sideband (USB) and two other
lines, $^{13}$CO and C$^{18}$O, at the lower sideband (LSB). Typical system temperatures were around 210\,K for the USB and around
130\,K for the LSB, and the variations among different beams are less than 15\%.
The total of pointing and tracking errors is about 5$''$, while the half-power beam width (HPBW) is $\sim$\,52$''$. The main-beam efficiencies
during the observations were $\sim$\,44\% for USB with the differences among the beams less than 16\%, and $\sim$\,48\% for LSB with the
differences less than 6\%. We mapped a 3$^\circ$\,$\times$\,2$^\circ$ area around Tycho's SNR via on-the-fly (OTF) observing mode, and
the data were meshed with the grid spacing of 30$''$. A Fast Fourier Transform (FFT) spectrometer with a total bandwidth of 1000\,MHz and
16,384 channels was used as the back end. The corresponding velocity resolutions were $\sim$\,0.16\,km\,s$^{-1}$ for the $^{12}$CO line
and $\sim$\,0.17\,km\,s$^{-1}$ for both the $^{13}$CO and C$^{18}$O lines. The average rms noises of all final spectra are about 0.5\,K for
$^{12}$CO and about 0.3\,K for $^{13}$CO and C$^{18}$O. All data were reduced using the GILDAS package (see
https://www.iram.fr/IRAMFR/GILDAS/).
\subsection{Complementary CO data}
A number of large CO line surveys along the Galactic Plane have been carried out in the past (see a review by Heyer \& Dame 2015), such
as the $^{12}$CO\,(1--0) surveys at the Center-for-Astrophysics (CfA) 1.2-meter telescope (Dame et al. 2001) and the Five College Radio
Astronomy Observatory (FCRAO) 14-meter telescope (Heyer et al. 1998). These previous surveys also obtained large-field $^{12}$CO images
around Tycho's SNR, which provided complementary CO data in this work.
\section{Results}
Figure~1 shows the velocity-integrated intensity image of the MWISP $^{12}$CO\,(1--0) emission around Tycho's SNR.
The integrated velocity range is between $-$68 and $-$57\,km\,s$^{-1}$, a velocity range suggested by previous CO line
observations toward the shell-like structure in the remnant (Lee et al. 2004; Cai et al. 2009; Xu et al. 2011; Zhou et al. 2016).
The large-field MWISP $^{12}$CO image reveals three large molecular clouds, located in the southeast, northeast, and west
of the remnant, respectively. Interestingly, the three clouds individually show an arc toward the remnant center (see solid
yellow lines in Fig.\,1). The three arcs could be fitted by one complete ellipse, which outlines a previously undiscovered cavity
around Tycho's SNR (see Fig.\,1). The radii of the cavity, depending on the azimuths, range from $\sim$\,0.3 degrees (to
the south of the remnant) to $\sim$\,0.6 degrees (to the west of the remnant), corresponding to $\sim$\,13-27\,pc (adopting a
distance of 2.5\,kpc).
The MWISP $^{12}$CO image also reveals three stream-like structures in the cavity (see Fig.\,1). One is to the northeast of the
remnant (named stream$_{\rm NE}$), the other, showing a wiggle morphology, to the southeast (named stream$_{\rm SE}$).
These two stream-like structures spatially connect the east edge of the cavity and the inner rim, which was found in the
previous CO line observations (with small fields of view) toward Tycho's SNR (e.g., Lee et al. 2004; Zhou et al. 2016).
Another faint stream-like structure can be roughly distinguished to the northwest of the remnant (named stream$_{\rm NW}$).
The three ``streams", joining at the center of the remnant, appear to radiate from the center.
Figure~2 shows the velocity-integrated intensity image of the MWISP $^{13}$CO\,(1--0) emission around Tycho's SNR. The
$^{13}$CO emission also shows the three clouds, but the emission is faint and the arc-like structures seen in the $^{12}$CO
emission are not clear in the $^{13}$CO intensity image. The $^{13}$CO emission is also detected from the north rim around the
shell-like structure. Figure~3 shows the CO spectra sampled from six positions in the surrounding clouds (see Figure~2). The
CO spectra do not show the Gaussian-like shape which is generally seen in quiescent molecular clouds, but present line
broadenings (in a velocity range between $-$70 and $-$55\,km\,s$^{-1}$) and asymmetries (see Figure~3). We note that no C$^{18}$O
line emission is detected around Tycho's SNR in the MWISP observations. The MWISP $^{12}$CO velocity channel map is
shown in Fig.\,A.1, where detailed kinematic information of the molecular gas around Tycho's SNR can be found.
The physical properties of the clouds, stream-like structures, and inner rim are listed in Table~1. Here two methods have been
used in the derivation of the H$_{2}$ gas column densities and masses. In the first method, on the assumption of local thermodynamic
equilibrium (LTE) and the $^{12}$CO\,(1--0) line being optically thick, we can derive the excitation temperature from the peak
radiation temperature of the $^{12}$CO\,(1--0). The $^{13}$CO\,(1--0) emission is optically thin and the $^{13}$CO column density
is converted to the H$_2$ column density using N(H$_2$)/N($^{13}$CO) $\approx$ 7\,$\times$\,10$^5$ (Frerking et al. 1982).
In the second method, the H$_2$ column density is estimated by adopting the mean CO-to-H$_2$ mass conversion factor
1.8\,$\times$\,10$^{20}$\,cm$^{-2}$\,K$^{-1}$\,km$^{-1}$\,s (Dame et al. 2001). The difference between the H$_2$ column
densities derived by the two methods are mainly caused by the small filling factors of the $^{13}$CO emission, since there is much
less $^{13}$CO emission than $^{12}$CO emission in the region (see Figures~1 and 2).
\section{Discussion}
\subsection{Scenario of a cavity produced by bright massive stars}
The large-field CO\,(1--0) observations show a large cavity around Tycho's SNR. It is well-known that bright (O- or early B-type)
stars are able to produce large bubbles in the ISM through strong stellar winds and UV radiation (see, e.g., Churchwell et al. 2006;
Deharveng et al. 2010). Therefore, the first question is whether the cavity around Tycho's SNR was produced by the bright stars
in the region.
As found by Chen et al. (2013), there is a linear relationship between the radius of a main-sequence bubble in a molecular environment
($R_{\rm bubble}$) and the initial mass of the energy source star ($M_{\rm star}$):
$R_{\rm bubble}$ (pc) $\approx$ 1.22$M_{\rm star}$/$M_\odot$ $-$ 9.16\,pc, assuming a constant interclump pressure (see Chen
et al. 2013 for more details). For the large cavity found in the MWISP observations (major radius of $\sim$\,0.6$^\circ$ or 27\,pc), a
massive star with a mass of $\sim$\,30\,$M_\odot$ (O7 or earlier types) is required. However, no such massive early stars are found
in the center of the cavity in the SIMBAD Astronomical Database\footnote{http://simbad.u-strasbg.fr/simbad/}. In addition, there is no
large HII region around the remnant either. Therefore, we can exclude the scenario of massive star bubble in the case of the large
cavity around Tycho's SNR.
\subsection{Possibility of the random distribution of a cavity}
The progenitor of Tycho's supernova was an evolved WD system (the time delays of most SNe~Ia range from $\sim$\,275\,Myr
to 1.25\,Gyr, with a median of $\sim$\,650\,Myr; see Schawinski 2009), and would not be expected to remain associated with its
natal molecular cloud. Therefore, the second question is whether the large cavity is randomly distributed around the remnant.
After checking much larger CO images from previous surveys (e.g., the FCRAO data), we find that the cavity-like structure is commonly
seen in the field and the estimated probability for a `cavity' to be detected by chance is higher than 0.06 (see Appendix~B). On the other
hand, we note that line broadenings and asymmetries are detected in the CO spectra of the surrounding clouds (see Figure 3). For instance,
the measured $^{12}$CO FWHM linewidths in the southeast cloud are typically $\sim$\,6-7\,km\,s$^{-1}$, which are much broader than the
linewidths found in other molecular clouds, such as the L1157 cloud in Taurus ($\sim$\,0.5-1.0\,km\,s$^{-1}$; see Hacar et al. 2016), for example.
This kind of spectra, suggesting the shock effect on the surrounding gas, is frequently found in the molecular clouds interacting with SNRs
(see, e.g., Jiang et al. 2010, Zhou et al. 2014). Furthermore, the observed kinematics in the CO gas suggests that the whole cavity is
expanding at a velocity of $\sim$\,3-4\,km\,s$^{-1}$\ (see discussion below in $\S$\,4.4.1). These results, in concert with the enhanced
$^{12}$CO(2--1)/(1--0) line ratio (about 1.6) found in the rim in the previous studies (Zhou et al. 2016; see discussion below in $\S$\,4.4.3),
suggest the interaction of the large cavity with a strong wind in the region.
Nevertheless, it must be noted that we cannot rule out the possibility of the random distribution of a large cavity-like structure around
Tycho's SNR thus far. Further observations, for example, searching for 1720\,MHz OH maser around the remnant (see discussion in
Chen et al. 2014 and Dubner \& Giacani 2015), are needed to verify the physical association between the large cavity and remnant.
\subsection{Potential cavity opened by accretion wind}
According to theoretical studies (e.g., Hachisu et al. 1996; 1999), accreting WDs in a binary system with a non-degenerate
companion would blow substantial outflows (also known as ``accretion winds''), excavating low-density cavities in the surrounding
interstellar medium (ISM) in the few 10$^6$\,yr prior to explosion (see Badenes et al. 2007 and references therein). On the other
hand, in the DD scenario, there would be no such winds and cavities. Therefore, one direct and effective way to distinguish between
progenitor scenarios is to search for large cavities (10-30\,pc) in the ISM, centered on the explosion sites of the SNe Ia. However,
to our knowledge, no such cavities have been observationally found yet (see, e.g., Badenes et al. 2007). A promising candidate
could be RCW~86, an SNR located in a 12-pc-radius cavity (see Broersen et al. 2014 and references therein), though the Type Ia
origin of this remnant is still uncertain.
The dimension of the large cavity around Tycho's SNR is far larger than the shock front from the supernova could produce.
In this context, this large cavity could be explained by the accretion wind from the progenitor system of Tycho's supernova.
As mentioned above, the progenitor system of Tycho's supernova was an evolved system. The molecular clouds found
around Tycho's progenitor system are likely newly formed in the Galactic Plane, where the typical lifetime of molecular
clouds is $\sim$\,10-100\,Myr (see Heyer \& Dame 2015). These newly-formed clouds, affected by generations of supernovae,
stellar winds, colliding flows, and turbulence (see, e.g., Dobbs et al. 2014), are irregularly distributed around Tycho's progenitor
system and then impacted by its accretion wind. This may explain why the north and northwest part of the cavity is open, that is,
there are no molecular clouds initially formed/distributed in these directions. Interestingly, a similar situation is also seen in the RCW\,86
case, where CO clouds are only seen in the east, south, and northwest of the cavity, leaving most of the circumference open
(see Sano et al. 2016).
Nevertheless, the CO observations (see Figs.\,1 and 2) find inner rim and stream-like structures within the cavity. This raises
another question about how the progenitor system can excavate a large cavity through a strong wind whilst leaving a significant
amount of residual material in the center.
It should be noted that the cavity models discussed in Badenes et al. (2007) make three important approximations: the ISM
around the progenitor system is homogeneous, and accretion wind from the system is spherically symmetric in space and
continuous in time. As we discussed above, the interstellar gas around the progenitor systems could be very complicated.
For accretion winds, in fact, numerical studies have shown that (1) the winds could be bipolar, which may lead to a bipolar
CSM structure similar to planetary nebulae (Balick \& Frank 2002), and (2) the winds are normally not steady but instead
episodic (e.g., Hachisu \& Kato 2003a, 2003b). Furthermore, during the mass growth of the WD, the mass transfer between
the donor star and WD is unstable, where nova-like outbursts are expected (see Maoz et al. 2014 and references therein)
and will lead to a complex CSM structure (even with several shells). Therefore, a totally empty cavity is only an ideal situation.
\subsection{Kinematics of the gas around the remnant}
\subsubsection{The expansion of the surrounding clouds}
Figure~4 shows the $^{12}$CO intensity-weighted velocity field around Tycho's SNR. In this velocity-field image, the mean
local standard of rest velocities ($V_{\rm LSR}$) of the molecular gas around the remnant can been found. The velocity field
shows that the CO emission from the most part of the inner rim, as well as the stream-like structures, is relatively blueshifted (i.e.,
moving toward us), compared with the nearby clouds. The measured mean velocity of the inner rim is $-$65.0\,$\pm$\,0.5\,km\,s$^{-1}$,
while the velocity measured along the cavity edge is roughly $-$62\,$\pm$\,1\,km\,s$^{-1}$\ (see Fig.\,4).
Figure~5 shows the position-velocity (PV) diagrams along the three large clouds (see the routes of the PV diagrams in Fig.\,4).
Interestingly, the PV diagrams of the three clouds all show curve-shaped morphologies, which could be fitted by ellipses with
velocity radii of $\sim$\,3--4\,km\,s$^{-1}$\ (see Fig.\,5). These results imply the expansion of the surrounding gas, which is likely pushed
by the accretion wind. Furthermore, we find that the expansion directions of the northeast cloud (blueshifted) and west cloud
(redshifted) appear to be opposite each other (see Fig.\,6). Therefore, the observed kinematics could by explained by a
complete view that the large cavity is expanding.
Using the standard method (see Weaver et al. 1977), the value of the mechanical luminosity of the wind ($L_{\rm wind}$) can
be estimated by
$L_{\rm wind}$\,$\approx$\,$\frac{1}{3}$$\frac{n_{\rm gas}}{{\rm cm^{-3}}}$($\frac{R_{\rm c}}{{\rm pc}}$)$^2$($\frac{V_{\rm c}}{{\rm km\,s^{-1}}}$)$^3$\,$\times$\,
10$^{30}$~ergs\,s$^{-1}$, in order to excavate a cavity with a radius of $R_{\rm c}$ and expansion velocity of $V_{\rm c}$ in
a molecular cloud with a density of $n_{\rm gas}$. For the large cavity around Tycho's SNR, the expansion velocity $V_{\rm c}$
is estimated to be $\sim$\,3.5\,km\,s$^{-1}$, while the density $n_{\rm gas}$ of the clouds is measured at $\sim$\,30\,cm$^{-3}$
from the CO observations. Adopting the major radius of the cavity ($R_{\rm c}$\,=\,27\,pc), the estimated
$L_{\rm wind}$ ($\sim$\,3\,$\times$\,10$^{35}$\,ergs\,s$^{-1}$) could be fed by a wind with a mass-loss rate of the order of
10$^{-6}$\,$M_{\odot}$\,yr$^{-1}$ at a velocity of $\sim$\,800\,km\,s$^{-1}$, while the timescale of the wind needed for opening such
a large cavity ($\frac{16}{27}$$\frac{R_{\rm c}}{\rm pc}$$\frac{\rm km\,s^{-1}}{V_{\rm c}}$\,$\times$\,10$^6$\,yr; Weaver et al.
1977) is $\sim$\,4-5\,$\times$\,10$^6$\,yr. The estimated wind velocity and timescale are both consistent with the predictions
from the wind-regulated accretion models, in which the wind velocity ranges from $\sim$\,200\,km\,s$^{-1}$\ to $\sim$\,1000\,km\,s$^{-1}$\
and timescale is $\sim$\,10$^{6}$ yr (see Badenes et al. 2007 and references therein).
\subsubsection{The stream-like structures}
The morphologies of the stream-like structures seen in the CO images resemble the jets/outflows driven by young stellar objects
(YSOs; see, e.g., Reipurth \& Bally 2001; Arce et al. 2007). In the optical and infrared observations, those YSO jets can move
more than 10\,pc from their driving sources, with a typical jet velocity of 100\,km\,s$^{-1}$ (see Reipurth \& Bally 2001). If the
`streams' found in the cavity were indeed the YSO jets, there should be a small group of YSOs in the center of the remnant. However,
we did not find this YSO group in the infrared observations (X.~Chen et al., in preparation).
In the wind-regulated accretion models (e.g., Hachisu et al. 1996; 1999), the accretion wind could last for a few million years, and
the WD may explode as a Type Ia supernova while the accretion wind is still active. Therefore, one possible explanation is that
these stream-like structures actually record the accretion winds from the progenitor system. In the theoretical models, the velocities
of the accretion winds range from $\sim$\,200\,km\,s$^{-1}$\ up to $\sim$\,1000\,km\,s$^{-1}$\ (see discussion in $\S$\,4.4.1), which is high enough
to shock interstellar clouds with large velocity dispersions. Indeed, in the CO observations (see PV diagrams in Figs.\,5a \& 5b),
the stream-like structures show large velocity ranges ($\sim$\,10\,km\,s$^{-1}$). The observed result suggests shocked emission along the
streams, which could be caused by the underlying wind.
\subsubsection{The inner rim}
The rim around the shell-like structure in Tycho's SNR was observed for decades (Lee et al. 2004; Cai et al. 2009; Xu
et al. 2011), and was recently suggested to represent a wind-blown cavity ($\sim$\,8$'$ in diameter) from the progenitor
system of the supernova, based on the detection of the expansion of the rim (see Zhou et al. 2016, and also Chiotellis et al.
2013). The expansion velocity of the rim is $\sim$\,4.5\,km\,s$^{-1}$, while the estimated wind velocity and timescale for this
putative cavity are $\sim$\,140\,km\,s$^{-1}$ and 3.9\,$\times$\,10$^5$\,yr, respectively (Zhou et al. 2016).
In the MWISP CO intensity images (see Fig.\,7a), there is extended CO emission within the shell region, and a small cavity-like
structure is found around the shell (as seen at the edges of the rim in Fig.\,7a, and in the green contours in Fig.\,2). Figure~7
shows the PV diagrams across the rim but in various directions. The PV diagrams along the northeast rim (see Figs.\,5b \& 7c)
show ring-like morphologies with small radii (about 0.1$^\circ$), which suggests the expansion of the gas in this direction with
a velocity of $\sim$\,3.5--4.0\,km\,s$^{-1}$. This is consistent with the result found by Zhou et al. (2016).
As discussed in $\S$\,4.3, during the accretion of the WD, nova-like outbursts could happen due to the instability in the
mass transfer between the donor star and WD. For instance, Langer et al. (2000) even found a long ($\sim$\,10$^6$\,yr)
switch-on phase of the mass transfer in their binary models. Therefore, there is a possibility that the rim (found in previous
observations) and the large cavity (found in this work) represent two independent cavities resulting from two different
accretion/outburst epochs.
On the other hand, if the rim indeed represents a small cavity, we may expect to find the isotropic expansion of this small
cavity. Nevertheless, we find that the expansion is only seen in the northeast direction of the rim, but not distinct in the other
directions (see Figs.\,5a \& 7b). Alternatively, we suggest that the rim may trace a large amount of CSM swept-up by the
(asymmetric) accretion wind in the northeast direction (see discussion above) for the following reasons: (1) The `expansion',
as well as the most circumstellar gas, is seen in the northeast direction of the rim; (2) the rim is connected with the
stream-like structures (stream$_{\rm NE}$ and stream$_{\rm SE}$; see Fig.\,1) radiating from the remnant center toward the
edges of the large cavity.
\section{Summary}
We present large-field CO\,(1--0) molecular line observations toward Tycho's SNR, using the PMO 13.7-meter telescope.
The CO images reveal, from the outside in, large molecular clouds, stream-like structures, and an inner rim around the remnant.
We derived the basic properties (column density, mass, and kinematics) of these objects based on the CO observations.
The large molecular clouds individually show an arc toward the remnant center, outlining a large cavity with radii of
$\sim$\,0.3$^\circ$\,$\times$\,0.6$^\circ$ (or 13\,pc\,$\times$\,27\,pc at a distance of 2.5\,kpc) around the remnant.
The observed CO line broadenings and asymmetries in the surrounding clouds, together with the enhanced $^{12}$CO\,(2--1)/(1--0)
intensity ratio detected in previous observations, suggest the interaction of the large cavity with a wind in the region.
After excluding the scenario of a large bubble produced by bright massive stars, we suggest that the large cavity could
be explained by the accretion wind from the progenitor system of Tycho's supernova.
The observed CO gas kinematics suggests that the large cavity is expanding at a velocity of $\sim$\,3-4\,km\,s$^{-1}$.
The estimated velocity ($\sim$\,800\,km\,s$^{-1}$, with a mass-loss rate of $\sim$\,10$^{-6}$\,$M_{\odot}$\,yr$^{-1}$) and timescale
($\sim$\,4-5\,$\times$\,10$^6$\,yr) of the wind needed for creating such a cavity are consistent with the predictions from
the wind-regulated accretion model.
Nevertheless, we note that the possibility of the random distribution of a large cavity around Tycho's SNR cannot be ruled
out thus far. Further observations are needed to confirm the physical association of the large cavity with the remnant, as well
as to comprehensively understand the nature of the streams and inner rim found therein. If the large cavity is really associated
with the remnant, our result may imply that Tycho's supernova, the prototypical Type~Ia supernova in the Milky Way, arose from
accretion onto a white dwarf.
\section*{ACKNOWLEDGMENTS}
We thank the anonymous referee for providing insightful suggestions and comments, which helped us to improve this work.
We are grateful to all the members of the Milky Way Imaging Scroll Painting CO line survey group, especially the staff of Qinghai
Radio Station of PMO at Delingha for support during the observations. This work was supported by the National Natural Science
Foundation of China (grants Nos. 11473069, 11233007, and U1431231) and the Strategic Priority Research Program of the Chinese
Academy of Sciences (grant No. XDB09000000). X.C. acknowledges the support of the Thousand Young Talents Program of China.
\clearpage
|
\section{Motivation}
\label{sec:motivation}
Testing is an important step to get confidence
in the functionality of a program.
The advantage of testing compared to program verification
is its potential for automation.
If we do not execute test cases only manually for some inputs
but encode them as input to test frameworks,
we can automatically run and repeat them when the software
is further developed, which is also known as regression testing.
A difficulty in testing is to find appropriate inputs for the
individual tests.
For this purpose, property-based testing has been proposed,
well known in the functional language Haskell with the
QuickCheck tool \cite{ClaessenHughes00}.
Basically, properties are predicates parameterized over one
or more arguments.
QuickCheck automates the test execution by applying properties
to randomly generated test inputs.
Since this idea is particularly reasonable for declarative
languages, it is been adapted in different forms
to functional and logic programming languages.
For instance, SmallCheck \cite{RuncimanNaylorLindblad08}
and GAST \cite{KoopmanAlimarineTretmansPlasmeijer03}
focus on a systematic enumeration of test inputs
for functional programs,
PropEr \cite{PapadakisSagonas11} adapts ideas of QuickCheck
to the concurrent functional language Erlang,
PrologCheck \cite{AmaralFloridoSantosCosta14}
transfers and extends ideas of QuickCheck to Prolog,
and EasyCheck \cite{ChristiansenFischer08FLOPS}
exploits functional logic programming features
to property-based testing of Curry programs.
CurryCheck follows the same ideas.
Actually, it is based on EasyCheck to define properties.
However, CurryCheck is intended as a comprehensive tool to simplify
the automation of test execution.
To use CurryCheck, properties are interspersed
into the program as top-level definitions.
Thus, properties are used to document the intended
semantics of the source code, which also supports
test-driven program development known as ``extreme programming.''
When CurryCheck is applied to a (set of) Curry modules,
it extracts all properties, generates a program to test these
properties, executes this generated program, and reports any errors.
Furthermore, CurryCheck also analyzes possible contracts
\cite{AntoyHanus12PADL} provided in source programs
and generates properties to test these contracts.
Thanks to this automation, CurryCheck is a useful tool
for continuous integration and deployment processes.
Actually, it is used for this purpose in the Curry implementations
PAKCS \cite{Hanus16PAKCS} and KiCS2 \cite{BrasselHanusPeemoellerReck11}.
In this paper we present the ideas and usage of CurryCheck.
After a review of the main features of Curry in the next section,
we introduce properties in Sect.~\ref{sec:properties}
and explain how they are tested in Sect.~\ref{sec:testproperties}.
The support of CurryCheck to define test inputs is presented
in Sect.~\ref{sec:gentestdata}.
CurryCheck's support for contract checking is described in
Sect.~\ref{sec:contracts}.
Some initial features of CurryCheck to combine testing and verification
are sketched in Sect.~\ref{sec:verify}.
We report about our practical experience with CurryCheck
in Sect.~\ref{sec:impl}
before we compare CurryCheck to some related tools and conclude.
\section{Functional Logic Programming and Curry}
\label{sec:flp}
Functional logic languages \cite{AntoyHanus10CACM,Hanus13}
integrate the most important features
of functional and logic languages in order to provide a variety
of programming concepts.
They support functional concepts like
higher-order functions and lazy evaluation
as well as logic programming concepts like
non-deterministic search and computing with partial information.
This combination led to new design patterns \cite{AntoyHanus11WFLP}
as well as better abstractions for application programming.
The declarative multi-paradigm language Curry \cite{Hanus97POPL}
is a modern functional logic language with advanced programming concepts.
In the following, we briefly review some features of Curry
relevant for this paper.
More details can be found in recent surveys on
functional logic programming \cite{AntoyHanus10CACM,Hanus13}
and in the language report \cite{Hanus16Curry}.
The syntax of Curry is close to Haskell \cite{PeytonJones03Haskell}.
In addition to Haskell, Curry allows \emph{free} (\emph{logic}) \emph{variables}
in rules and initial expressions.
Function calls with free variables are evaluated by a possibly
non-deterministic instantiation of demanded arguments.
\begin{example}\label{ex:concdup}
The following simple program shows the functional and logic features
of Curry. It defines the well-known list concatenation
and an operation that returns
some element of a list having at least two occurrences:
\begin{curry}
(++) :: [a] -> [a] -> [a] someDup :: [a] -> a
[] ++ ys = ys someDup xs | xs == _$\,$++$\,$[x]$\,$++$\,$_$\,$++$\,$[x]$\,$++$\,$_
(x:xs) ++ ys = x : (xs ++ ys) = x where$\;$x$\;$free
\end{curry}
The (optional) type declaration (\ccode{::}) of the operation \ccode{++}
specifies that \ccode{++} takes two lists as input and produces
an output list, where all list elements are of the same (unspecified) type.
Since \ccode{++} can be called with free variables in arguments,
the condition in the rule of \code{someDup}
is solved by instantiating \code{x} and
the anonymous free variables \ccode{\us} to appropriate values
before reducing the function calls.
This corresponds to narrowing \cite{Slagle74,Reddy85},
but Curry narrows with possibly non-most-general unifiers
to ensure the optimality of computations \cite{AntoyEchahedHanus00JACM}.
\end{example}
Note that \code{someDup} is a \emph{non-deterministic operation}
since it might deliver more than one result for a given argument,
e.g., the evaluation of \code{someDup$\,$[1,2,2,1]} yields the values
\code{1} and \code{2}.
Non-deterministic operations, which can formally be
interpreted as mappings from values into sets of values \cite{GonzalezEtAl99},
are an important feature
of contemporary functional logic languages.
Hence, Curry has also a predefined \emph{choice} operation:
\begin{curry}
x ? _ = x
_ ? y = y
\end{curry}
Thus, the expression \ccode{0$~$?$~$1} evaluates to \code{0} and \code{1}
with the value non-deterministically chosen.
\emph{Functional patterns} \cite{AntoyHanus05LOPSTR} are useful
to define some operations more easily.
A functional pattern is a pattern occurring in an argument
of the left-hand side of a rule containing defined operations
(and not only data constructors and variables).
Such a pattern abbreviates the set of all standard patterns to which the
functional pattern can be evaluated (by narrowing).
For instance, we can rewrite the definition of \code{someDup} as
\begin{curry}
someDup (_++[x]++_++[x]++_) = x
\end{curry}
Functional patterns are a powerful feature to express arbitrary selections
in tree structures, e.g., in XML documents \cite{Hanus11ICLP}.
Details about their semantics and a constructive implementation
of functional patterns by a demand-driven unification procedure
can be found in \cite{AntoyHanus05LOPSTR}.
Curry has also features which are useful for application programming,
like \emph{set functions} \cite{AntoyHanus09} to encapsulate
non-deterministic computations,
\emph{default rules} \cite{AntoyHanus16PADL} to deal
with partially specified operations and negation,
and standard features from functional programming,
like modules or monadic I/O \cite{Wadler97}.
Other features are explained when they are used in the following.
\section{Properties}
\label{sec:properties}
In this section we briefly discuss which kind of program properties
to be tested are supported by CurryCheck.
Since CurryCheck extends the functionality of
EasyCheck \cite{ChristiansenFischer08FLOPS},
it supports all kinds of EasyCheck's properties
which we review first.
Properties are defined top-level entities with a distinct type (see below).
Thus, their syntax and type-correctness
can be checked by the standard front end of any Curry system.
Properties do not require a specific naming convention
but CurryCheck recognizes them by their type.
Moreover, the name and position of the property in the source file
are used by CurryCheck to identify properties when errors are reported.
For instance, consider the list concatenation operation \ccode{++}
defined in Example~\ref{ex:concdup}.
Before discussing general properties, we define some
unit tests for fixed arguments, like
\begin{curry}
concNull12 = [] ++ [1,2] -=- [1,2]
concCurry = "Cu" ++ "rry" -=- "Curry"
\end{curry}
The infix operator \ccode{-=-} specifies a test which is
successful if both sides have single values which are identical
(we will later see tests for non-deterministic operations).
Since the expressions can be of any type (of course, the two
arguments must be of the same type),
the operator is polymorphic and has the type
\begin{curry}
(-=-) :: a -> a -> Prop
\end{curry}
Hence, all entities defined above have type \code{Prop}.
The power of CurryCheck and similar property-based test frameworks
comes from the fact that we can also test properties which are
parameterized over some input data.
For instance, we can check whether the concatenation operation
is associative by:
\begin{curry}
concIsAssociative xs ys zs = (xs++ys)++zs -=- xs++(ys++zs)
\end{curry}
This property is parameterized over three input values
\code{xs}, \code{ys}, and \code{zs}.
To test this property, CurryCheck guesses values for these parameters
(see below for more details) and tests the property for these values:
\begin{curry}
concIsAssociative_ON_BASETYPE (module ConcDup, line 18):
OK, passed 100 tests.
\end{curry}
Indicated by the suffix \code{\us{}ON\us{}BASETYPE},
we see another feature of CurryCheck.
If properties are polymorphic in their input values
(the above property has type
\code{[a]$\;\to\;$[a]$\;\to\;$[a]$\;\to\;$Prop}),
CurryCheck specializes the type to some base type,
since there is no concrete value of a polymorphic type
(and EasyCheck would fail on such properties).
As a default, CurryCheck uses the predefined type \code{Ordering}
having the three values \code{LT}, \code{EQ}, \code{GT}
(another more involved method to
instantiate polymorphic types in purely functional programs
can be found in \cite{BernardyJanssonClaessen10}).
This default type can be changed to other
base types, like \code{Bool}, \code{Int}, or \code{Char},
with a command-line option.
One could also provide an explicit type declaration for the
property. For instance, we can test the commutativity
of the list concatenation on lists of integers by the property
\begin{curry}
concIsCommutative :: [Int] -> [Int] -> Prop
concIsCommutative xs ys = (xs ++ ys) -=- (ys ++ xs)
\end{curry}
Of course, this property does not hold so that CurryCheck
reports an error together with a counter-example:
\begin{curry}
$\ldots$
concIsCommutative (module ConcDup, line 20) failed
Falsified by 8th test.
Arguments: [-1] [-3]
Results: ([-1,-3],[-3,-1])
\end{curry}
Note that the arguments of a test are ordinary expressions
so that one can use any defined operation in the tests.
For instance, we can (sucessfully) check whether the list concatenation
is the addition on their lengths:
\begin{curry}
concAddLengths xs ys = length xs + length ys -=- length (xs++ys)
\end{curry}
Since Curry covers also logic programming features,
CurryCheck supports the testing of non-deterministic properties.
For instance, one can check whether an expression reduces to some
given value with the operator is \ccode{\char126>}:
\begin{curry}
someDup1 = someDup [1,2,1,2] ~> 1
\end{curry}
Another important operator is \ccode{<\char126>} which
denotes a test which succeeds if both arguments
have the same set of values.
We can write unit tests by enumerating all expected values
with the choice operator \ccode{?}:
\begin{curry}
someDup12 = someDup [1,2,1,2,1] <~> (1$\,$?$\,$2)
\end{curry}
It should be noted that the operator \ccode{<\char126>}
really compares sets and not multi-sets:
Although the evaluation of \code{someDup$\;$[1,2,1,2,1]}
returns the value \code{1} three times in a typical Curry system,
the property \code{someDup12} holds.
This is intended since CurryCheck tests declarative properties which
are independent of specific compiler optimizations
(this is in contrast to PrologCheck which tests operational
properties like multiplicity of answers and modes
\cite{AmaralFloridoSantosCosta14}).
As another example, consider the following definition
of a permutation of a list by exploiting a functional pattern
to select some element in the argument list:
\begin{curry}
perm (xs++[x]++ys) = x : perm (xs++ys)
perm [] = []
\end{curry}
An important property of a permutation is that the length of
the list is not changed. Hence, we check it by the property
\begin{curry}
permLength xs = length (perm xs) <~> length xs
\end{curry}
Note that the use of \ccode{<\char126>} (instead of \ccode{-=-})
is relevant since non-deterministic values are compared.
Actually, the left argument evaluates to many (identical) values.
We might also want to check whether our definition of \code{perm} computes the
correct number of solutions. Since we know that a list of length $n$
has $n!$ permutations, we write the following property,
where \code{fac} is the factorial function and
the property \code{$x$ \# $n$} is satisfied if $x$ has $n$ different values:
\begin{curry}
permCount :: [Int] -> Prop
permCount xs = perm xs # fac (length xs)
\end{curry}
However, this test will be falsified with the test input \code{[1,1]},
since this list has only one permuted value (actually, both computed
values are identical).
We can obtain a correct property if we add the condition
that all elements in the input list \code{xs} are pairwise different.
For this purpose, we use a \emph{conditional property}:
the property \code{$b$ ==> $p$} is satisfied if $p$
is satisfied for all values where $b$ evaluates to \code{True}.
If the predicate \code{allDifferent} is satisfied iff its argument list
does not contain duplicated elements,
then we can reformulate our property as follows:
\begin{curry}
permCount xs = allDifferent xs ==> perm xs # fac (length xs)
\end{curry}
Furthermore, we want to check the existence of distinguished permutations.
For this purpose, consider a predicate to check whether a list is sorted:
\begin{curry}
sorted :: [Int] -> Bool
sorted [] = True
sorted [_] = True
sorted (x:y:zs) = x<=y && sorted (y:zs)
\end{curry}
Then we can check whether there are sorted permutations
(the property \ccode{eventually$\;x$} is satisfied if some value
of $x$ is \code{True}):
\begin{curry}
permIsEventuallySorted :: [Int] -> Prop
permIsEventuallySorted xs = eventually (sorted (perm xs))
\end{curry}
Property-based testing is appropriate for declarative languages
since the absence of side effects allows the execution
of tests on any number of test data without influencing
the individual tests.
Nevertheless, real programming languages have to deal with the
real world so that they support also I/O operations.
Clearly, such operations should also be tested.
Although there are methods to test monadic code \cite{ClaessenHughes02},
Curry supports only I/O monadic operations where testing with
arbitrary data seems not reasonable.
Therefore, CurryCheck supports only non-parameterized
unit tests for I/O operations.
For instance, the test \code{($a$ `returns` $x$)} is satisfied
if the I/O action $a$ returns the value $x$.
For instance, we can test whether writing a file and reading it
yields the same contents:
\begin{curry}
writeReadFile = (writeFile "TEST" "Hello" >> readFile "TEST")
`returns` "Hello"
\end{curry}
Since CurryCheck executes the tests written in a source program
in their textual order, one can write also several I/O tests
whose side effects depend on each other.
For instance, we can split the previous I/O test into two
consecutive tests:
\begin{curry}
writeTestFile = (writeFile "TEST" "Hello") `returns` ()
readTestFile = (readFile "TEST") `returns` "Hello"
\end{curry}
\section{Testing Properties}
\label{sec:testproperties}
After having seen several methods to define properties,
we sketch in this section how they are actually tested.
Our motivation for the development of CurryCheck is manifold:
\begin{enumerate}
\item
Properties are an executable documentation for the
intended semantics of operations.
\item
Properties increase the confidence in the quality of the developed
software.
\item
Properties can be used for software verification
by proving their validity for all possible input data.
\end{enumerate}
The first point is supported by interspersing properties
into the source code of the program instead of putting them
into separate files.
Thus, properties play the same role as comments or type annotations:
they document the intended semantics.
Hence, they can be extracted and put into the program
documentation by automatic documentation tools
\cite{Hanus02WFLP,Hermenegildo00}.
In order to avoid that properties influence the interface
of a module, they do not need to be exported.
As an example, consider the following simple module defining
the classical list reverse operation (the imported module
\code{Test.EasyCheck} contains the definitions of the property combinators
introduced in Sect.~\ref{sec:properties}):
\begin{curry}
module Rev(rev) where$\vrule width0pt depth1.25ex$
import Test.EasyCheck$\vrule width0pt depth1.25ex$
rev :: [a] -> [a]
rev [] = []
rev (x:xs) = rev xs ++ [x]$\vrule width0pt depth1.25ex$
revLength xs = length (rev xs) -=- length xs
revRevIsId xs = rev (rev xs) -=- xs
\end{curry}
We can run all tests of this module by invoking the CurryCheck executable
with the name of the module:\footnote{One can also provide several
module names so that they are tested at once.
Furthermore, CurryCheck has various options to influence
the number of test cases, default types for polymorphic properties, etc.}
\begin{curry}
> currycheck Rev
Analyzing module 'Rev'...
...
Executing all tests...
revLength_ON_BASETYPE (module Rev, line 9):
OK, passed 100 tests.
revRevIsId_ON_BASETYPE (module Rev, line 10):
OK, passed 100 tests.
\end{curry}
Although module \code{Rev} only exports the operation \code{rev},
all properties defined in the top-level of \code{Rev}
are passed to the underlying EasyCheck library for testing.
For this purpose, CurryCheck creates a copy of this module
where all entities are exported (note that the original module
cannot be modified since it might be imported to other modules
to be tested).
For each property a corresponding call to
an operation of EasyCheck is generated which actually
performs the generation of test data, runs the test,
and collects all results which are passed back to CurryCheck.
Furthermore, polymorphic properties are not checked
but a corresponding new property on the default base type
is generated which calls the polymorphic property.
For instance, if the default base type is \code{Int},
CurryCheck generates the new property
\begin{curry}
revRevIsId_ON_BASETYPE :: [Int] -> Prop
revRevIsId_ON_BASETYPE = revRevIsId
\end{curry}
which is actually checked instead of \code{revRevIsId}.
Note that it might lead to a failure if the type of \code{revRevIsId}
is directly specialized, since the polymorphic property \code{revRevIsId}
might be used in other property definitions with a different specialized type.
After these preparations, EasyCheck tests the properties
by generating test data as described in
\cite{ChristiansenFischer08FLOPS}.
EasyCheck does not use random generators
like QuickCheck or PrologCheck,
but it exploits functional logic programming features
to enumerate possible input values.
Since logic variables are equivalent to non-deterministic
generators \cite{AntoyHanus06ICLP},
one can evaluate a logic variable of a particular type
in order to get all values of this type in a non-deterministic manner.
For instance, if we evaluate the Boolean variable
\code{b::Bool}, we obtain the values \code{False} and \code{True}
as results. Similarly, for \code{bs::[Bool]}
we obtain values like \code{[]}, \code{[False]}, \code{[True]},
\code{[False,False]}, etc.
In order to select a finite amount of these infinite values,
one can use Curry's feature for encapsulated search to
collect all non-deterministic results in a tree structure,
traverse the tree with some strategy and return the
result of the traversal into a list.
If one selects only a finite amount of this list,
the lazy evaluation strategy of Curry ensures a finite computation
even if the tree is infinite.
Based on these features, the EasyCheck library contains an operation
\label{sec:valuesOf}
\begin{curry}
valuesOf :: a -> [a]
\end{curry}
which computes the list of all values of the given argument
according to a fixed strategy (in the current implementation by
randomized level diagonalization \cite{ChristiansenFischer08FLOPS}).
Hence, we can get 20 values for a list of integers by
\begin{curry}
$\ldots$> take 20 (valuesOf (_::[Int]))
[[],[-1],[-3],[0],[1],[-1,0],[-2],[0,0],[3],[-1,1],[-3,0],[0,1],
[2],[-1,-1],[-5],[0,-1],[5],[-1,2],[-9],[0,2]]
\end{curry}
It should be noted that \code{valuesOf} enumerates all values
of the given type completely and without duplicates.\footnote{%
In order to get an idea of the distribution of the generated test data,
CurryCheck also provides property combinators
\fcode{collect} and \fcode{classify} known from QuickCheck.}
Hence, if the set of possible input values is finite,
it is ensured that all of them are tested if sufficiently many tests
are performed. In this case, the property is also verified
(where QuickCheck or PrologCheck does not give such guarantees).
For instance, consider the De Morgan law from Boolean algebra:
\begin{curry}
negOr b1 b2 = not (b1 || b2) -=- not b1 && not b2
\end{curry}
This property is proved by CurryCheck after four tests
with all possible input values, and the output of CurryCheck
indicates that the testing was exhaustive:
\begin{curry}
negOr (module BoolTest, line 4):
Passed all available tests: 4 tests.
\end{curry}
\section{User-Defined Test Data}
\label{sec:gentestdata}
Due to the use of functional logic features to generate test data,
one can write properties not only on predefined data types
but also on user-defined data types.
For instance, consider the following definition
of general polymorphic trees:
\begin{curry}
data Tree a = Leaf a | Node [Tree a]
\end{curry}
We define operations to compute the leaves of a tree and mirroring a tree:
\begin{curry}
leaves (Leaf x) = [x]
leaves (Node ts) = concatMap leaves ts$\vrule width0pt depth1.25ex$
mirror (Leaf x) = Leaf x
mirror (Node ts) = Node (reverse (map mirror ts))
\end{curry}
The following properties should increase our confidence in the
correctness of the implementation:
\begin{curry}
doubleMirrorIsId t = mirror (mirror t) -=- t$\vrule width0pt depth1.25ex$
leavesOfMirrorAreReversed t = leaves t -=- reverse (leaves (mirror t))
\end{curry}
CurryCheck successfully tests these properties without providing
any further information about how to generate test data.
However, in some cases it might be desirable to define our own test data
since the generated structures are not appropriate for testing.
For instance, if we test algorithms working on balanced search trees,
we need correctly balanced search trees as test data.
As a naive approach, we can limit the tests to correct test inputs
by using conditional properties.
As a simple example, consider the following operation that adds all numbers
from 1 to a given limit:
\begin{curry}
sumUp n = if n==1 then 1 else n + sumUp (n-1)
\end{curry}
Since there is also a simple formula to compute this sum,
we can check it:
\begin{curry}
sumUpIsCorrect n = n>0 ==> sumUp n -=- n * (n+1) `div` 2
\end{curry}
Note that the condition is important since \code{sumUp}
diverges on non-positive numbers.
As a result, Curry\-Check tests this property by enumerating integers
and dropping tests with non-positive numbers.
While this works well, since CurryCheck performs a fairly good
distribution between positive and negative numbers,
this approach might have a serious drawback if the proportion
of correct test data is small.
In the case of balanced search trees,
there are many more unbalanced trees than balanced search trees.
This has the effect that CurryCheck generates many test data
and drops it so that it does not make much progress.
Actually, CurryCheck has an upper limit for dropping test data
in the conditional operator in order to avoid spending too much
work on generating unusable test data.
For instance, if we want to test the above property \code{revRevIsId}
on long input lists, we could define it as follows:
\begin{curry}
revRevIsIdLong :: [Int] -> Prop
revRevIsIdLong xs = length xs > 100 ==> rev (rev xs) -=- xs
\end{curry}
Since there are a huge number of integer lists with a length smaller than 100,
CurryCheck does not find any test case
(with a default limit of dropping at most 10,000 incorrect test data
values):
\begin{curry}
revRevIsIdLong (module Rev, line 13):
Arguments exhausted after 0 test.
\end{curry}
This shows that the fully automatic generation of test data
is not always appropriate.
Therefore, CurryCheck provides some combinators to
explicitly define test data (more advanced enumeration combinators
in the context of Scala are discussed in \cite{KurajKuncakJackson15}).
To show the method for test data generation in more detail,
we have to review Curry's methods to encapsulate non-deterministic
computations. For this purpose, Curry defines the following structure
to represent the results of a non-deterministic computation
\cite{BrasselHanusHuch04JFLP}:
\begin{curry}
data SearchTree a = Value a | Fail | Or (SearchTree a) (SearchTree a)
\end{curry}
\code{(Value v)} and \code{Fail} represent a single value
or a failure (i.e., no value), respectively,
and \code{(Or t1 t2)} represents a non-deterministic choice
between two search trees \code{t1} and \code{t2}.
Furthermore, there is a primitive search operator
\begin{curry}
someSearchTree :: a -> SearchTree a
\end{curry}
which yields a search tree for an expression.
For instance, \code{someSearchTree (0?1)} evaluates to the
search tree
\begin{curry}
Or (Value 0) (Value 1)
\end{curry}
The expression
\begin{curry}[mathescape=false]
someSearchTree (id $## (_::[Bool]))
\end{curry}
(where \ccode{\$\#\#} is an infix application operator which
evaluates the right argument to ground normal form before applying
the left argument to it) yields an (infinite)
search tree of all Boolean lists:
\begin{curry}
(Or (Value []) (Or (Or (Or (Value [False]) ...) (Or ...)) ...))
\end{curry}
Basically, EasyCheck defines various strategies to traverse
such search trees (see \cite{ChristiansenFischer08FLOPS} for details)
in order to enumerate test data.
Hence, if we want to define our own test data,
we have to define an operation that generates a search tree
containing the test data in \code{Value} leaves.
Although this is not difficult for simple data types,
it could be demanding for polymorphic types where
generators for the polymorphic arguments must be weaved
with the generators for the main data structure.
To simplify this task, CurryCheck offers a family of
combinators \code{genCons$n$} where each combinator
takes an $n$-ary constructor function
and $n$ generators as arguments and produces a search tree
for this constructor and all combinations of generated arguments.
Hence, these combinators have the type
\begin{curry}
genCons$n$ :: ($a_1 \to \cdots \to a_n \to a$) -> SearchTree $a_1$ -> $\cdots$ -> SearchTree $a_n$
-> SearchTree $a$
\end{curry}
Furthermore, there is an infix combinator \ccode{|||} to
combine two search trees.
For instance, consider the straightforward definition of
Peano numbers:
\begin{curry}
data Nat = Z | S Nat
\end{curry}
Then we can define a search tree generator for this type as follows:
\begin{curry}
genNat :: SearchTree Nat
genNat = genCons0 Z ||| genCons1 S genNat
\end{curry}
Similarly, we can define a search tree generator for polymorphic trees
which takes a search tree for the argument type as a parameter
(where \code{genList} denotes the corresponding generator for list values):
\begin{curry}
genTree :: SearchTree a -> SearchTree (Tree a)
genTree ta = genCons1 Leaf ta ||| genCons1 Node (genList (genTree ta))
\end{curry}
The explicit definition of value generators is reasonable
when only a subset of all values should be used for testing.
For instance, \code{sumUpIsCorrect} should be testest with
positive numbers only.
Hence, we define a generator for positive numbers:
\begin{curry}
genPos = genCons0 1 ||| genCons1 (+1) genPos
\end{curry}
Since these numbers are slowly increasing, i.e., the search tree
is actually degenerated to a list, we can also use
the following definition to obtain a more balanced search tree:
\begin{curry}
genPos = genCons0 1 ||| genCons1 (\n -> 2*(n+1)) genPos
||| genCons1 (\n -> 2*n+1) genPos
\end{curry}
In order to test properties with user-defined data,
CurryCheck provides the property combinator
\begin{curry}
forAll :: [a] -> (a -> Prop) -> Prop
\end{curry}
which is satisfied if the parameterized property given as the second
argument is satisfied for all values of the first argument list.
Since there is also a library operation
\begin{curry}
valuesOfSearchTree :: SearchTree a -> [a]
\end{curry}
(actually, the operation \code{valuesOf} introduced in
Sect.~\ref{sec:valuesOf} is defined via this operation) to enumerate
all values of a search tree, we can redefine the property
\code{sumUpIsCorrect} as follows:
\begin{curry}
sumUpIsCorrect = forAll (valuesOfSearchTree genPos)
(\n -> sumUp n -=- n*(n+1) `div` 2)
\end{curry}
Using this technique, we could also define finite tests
for potentially infinite structures, e.g., one can easily define
tree generators that generate all trees up to a particular depth.
Finally, we show the implementation of the
combinators to generate search trees.
The definition of \ccode{|||} and \code{genCons0} is straightforward:
\begin{curry}
x ||| y = Or x y$\vrule width0pt depth1.25ex$
genCons0 v = Value v
\end{curry}
To define the further combinators like \code{genCons1},
we have to replace in a given search tree (for the argument)
the \code{Value} nodes by new nodes where the constructor operation
is applied to the given value.
This task is done by the following auxiliary operation:\footnote{%
This operation is similar to the monadic bind operation
in Haskell's \code{MonadPlus}, but we use this definition
due to the lack of type classes in the current language definiton
of Curry.}
\begin{curry}
updateValues :: SearchTree a -> (a -> SearchTree b) -> SearchTree b
updateValues (Value a) f = f a
updateValues Fail f = Fail
updateValues (Or t1 t2) f = Or (updateValues t1 f) (updateValues t2 f)
\end{curry}
The definition of the remaining combinators is now easy
(we only show the first two ones):
\begin{curry}
genCons1 c gena = updateValues gena (\a -> Value (c a))$\vrule width0pt depth1.25ex$
genCons2 c gena1 gena2 =
updateValues gena1 (\a1$\;$->$\;$updateValues gena2 (\a2$\;$->$\;$Value (c a1 a2)))
\end{curry}
\section{Contract and Specification Testing}
\label{sec:contracts}
As discussed in detail in \cite{AntoyHanus12PADL},
the distinctive features of Curry (e.g., non-deterministic operations,
demand-driven evaluation, functional patterns, set functions)
support writing high-level specifications
as well as efficient implementations for a given problem
in the same programming language.
When applying this idea, Curry can be used
as a wide-spectrum language \cite{BauerEtAl78} for software development.
If a specification or contract is provided for some function,
one can exploit this information to support
run-time assertion checking with these specifications and contracts.
As an additional use of this information,
CurryCheck automatically generates properties
to test the given specifications and contracts,
which is described in the following.
According to the notation proposed in \cite{AntoyHanus12PADL},
a \emph{specification}\index{specification}
for an operation $f$ is an operation \code{$f$'spec}
of the same type as $f$.
A \emph{contract}\index{constract} consists
of a pre- and a postcondition, where the precondition could be omitted.
When provided, a \emph{precondition}\index{precondition} for an operation $f$
of type $\tau \to \tau'$ is an operation
\begin{curry}
$f$'pre :: $\tau$ ->$~$Bool
\end{curry}
putting demands on allowed arguments, whereas
a \emph{postcondition}\index{postcondition} for $f$
is an operation
\begin{curry}
$f$'post :: $\tau$ ->$~\tau'$ ->$~$Bool
\end{curry}
which relates input and output values
(the generalization to operations with more than one argument
is straightforward).
A specification should precisely describe the meaning of an operation,
i.e., the declarative meaning of the specification and the implementation
of an operation should be equivalent.
In contrast, a contract is a partial specification, e.g.,
all results computed by the implementation should satisfy the
postcondition.
As a concrete example, consider the problem of sorting a list.
The specification defines a sorted version of a given list
as a permutation of the input which is sorted.
Exploiting the operations introduced in Sect.\ref{sec:properties},
we define the following specification for the operation \code{sort}:
\begin{curry}
sort'spec :: [Int] -> [Int]
sort'spec xs | ys == perm xs && sorted ys = ys where ys free
\end{curry}
A postcondition, which is easier to check, states that the
input and output lists should have the same length:
\begin{curry}
sort'post :: [Int] -> [Int] -> Bool
sort'post xs ys = length xs == length ys
\end{curry}
To provide a concrete implementation, we implement the quicksort algorithm
as follows:
\begin{curry}
sort :: [Int] -> [Int]
sort [] = []
sort (x:xs) = sort (filter (<x) xs) ++ [x] ++ sort (filter (>x) xs)
\end{curry}
Note that specifications and contracts are optional.
However, if they are included in a module processed with CurryCheck,
CurryCheck automatically generates and checks properties that
relate the specification and contract to the implementation.
For instance, an implementation satisfies a specification
if both yield the same values, and a postcondition is satisfied
if each value computed for some input satisfies the postcondition
relation between input and output. For our example, CurryCheck generates
the following properties (if there are also
preconditions for some operation, these preconditions are used
to restrict the test cases via the condition operater \ccode{==>}):\footnote{%
The property ``{\fcode{always$\;x$}}'' is satisfied
if all values of $x$ are \fcode{True}.}
\begin{curry}
sortSatisfiesSpecification :: [Int] -> Prop
sortSatisfiesSpecification x = sort x <~> sort'spec x$\vrule width0pt depth1.25ex$
sortSatisfiesPostCondition :: [Int] -> Prop
sortSatisfiesPostCondition x = always (sort'post x (sort x))
\end{curry}
With CurryCheck, the framework of \cite{AntoyHanus12PADL}
becomes more useful since contracts are not only used
as run-time assertions in concrete computations,
but many possible computations are checked with various test data.
For instance, CurryCheck reports that the above implementation of
\code{sort} is incorrect for the example input \code{[1,1]}
(as the careful reader might have already noticed).
When reporting the error, the module and source code line number
of the erroneous operation is shown so that the programmer
can easily spot the problem.
Another kind of contracts taken into account by CurryCheck
are determinism annotations.
An operation that yields always identical results
(maybe multiple times) on identical argument values
can be annotated as ``deterministic'' by adding \code{DET} to the
result type of its type signature.
For instance, the following operation tests whether a list
represents a set, i.e., has no duplicate elements (the definition
exploits functional patterns \cite{AntoyHanus05LOPSTR} as
well as default rules \cite{AntoyHanus16PADL}):
\begin{curry}
isSet :: [a] ->DET Bool
isSet (_++[x]++_++[x]++_) = False
isSet'default _ = True
\end{curry}
The determinism annotation \ccode{$\to$DET} has the effect that
at most one result is computed for a given input, e.g., a single value
\code{False} is returned for the call
\code{isSet$\;$[1,3,1,3,1]}, although the first rule can be applied
multiple times to this call. Thus, after computing a first value,
all attempts to compute further values are ignored.
In order to ensure that this does not destroy completeness,
i.e., it behaves like ``green cuts'' in Prolog,
such operations must be deterministic from a semantical point of view.
CurryCheck tests this property by generating a property
for each \code{DET}-annotated operation
that expresses that there is at most one value for each input.
For instance, for \code{isSet}, the \code{DET} annotation is removed
and the property
\begin{curry}
isSetIsDeterministic x1 = isSet x1 #< 2
\end{curry}
is added by CurryCheck, where \ccode{$e\;$\#<$\;n$} is satisfied
if the \emph{set} of all values of $e$ contains less than $n$ elements.
\section{Combining Testing and Verification}
\label{sec:verify}
The objective of CurryCheck is to increase the confidence
in the reliability of Curry programs.
Testing with a lot of input data is one important step
but, in case of input data types with infinite values,
it can only show possible errors but not the absence of them.
In order to support the latter,
CurryCheck has also some (preliminary) support
to include the verification of program properties.
For this purpose, a programmer might prove properties
stated in a source program.
Since there are many different possibilities to prove such properties,
ranging from manual proofs to interactive proof assistants
and fully automatic provers, CurryCheck does not enforce
a particular proof technique.
Instead, CurryCheck trusts the programmer and uses a naming
convention for files containing proofs:
if there is a file with name \code{proof-$t$.*},
CurryCheck assumes that this file contains
a valid proof for property $t$.
For instance, the following property states that
sorting a list does not change its length:
\begin{curry}
sortlength xs = length (sort xs) <~> length xs
\end{curry}
If there is a file \code{proof-sortlength.agda},
containing a proof for the above property
(\cite{AntoyHanusLibby16} addresses techniques how to prove such
properties in the dependently typed language Agda),
CurryCheck considers this property as valid and does not check it.
Moreover, it uses it to simplify other properties to be tested.
In our case, the property \code{sortSatisfiesPostCondition}
of the previous section can be simplified to
\code{always$\;$True} so that it does not need to be tested.
Similarly, a determinism annotation for operation $f$ is not tested
if there is a proof file \code{$f$IsDeterministic.*}.
Since program verification is a notoriously difficult task,
a mixture of different techniques is required.
For instance, \cite{JohanssonEtAl14} discusses techniques
to use the Isabelle/HOL proof assistant to verify purely functional
properties inspired by QuickCheck.
\cite{AntoyHanusLibby16} describes a method to prove
non-deterministic computations by translating Curry programs
into Agda programs. Since these proofs can be verified by the Agda compiler,
CurryCheck can test the validity of a given proof file by simply
invoking the Agda compiler.
Some purely functional properties can be proved
in a fully automatic way. For instance, the properties
\begin{curry}
concLength xs = length (xs ++ ys) -=- length xs + length ys
revLength xs = length (rev xs) -=- length xs
\end{curry}
can be proved by the SMT solver Alt-Ergo.
To support the use of such solvers,
we have started the development of tools to automatically
translate Curry programs into the syntax of Agda and other
proof systems. We omit more details since this is outside the
scope of this paper.
\section{Practical Experience}
\label{sec:impl}
The implementation of CurryCheck is available with the
(Prolog-based) Curry implementation PAKCS \cite{Hanus16PAKCS} (since version
1.14.0) and the
(Haskell-based) Curry implementation KiCS2 \cite{BrasselHanusPeemoellerReck11}
(since version 0.5.0).
The implementation exploits meta-programming features
available in these systems to parse programs and transform
them into new programs as described in the previous sections.
Although we could show in this paper only simple examples,
we would like to remark that CurryCheck is successfully applied
in a larger context.
CurryCheck is regularly used for automatic regression testing
during continuous integration and nightly builds of PAKCS and KiCS2.
Currently, approximately 500 properties (the number is continuously growing)
are regularly used to test the libraries and tools of these systems.
Our practical experience is quite promising.
After the development and use of CurryCheck,
we found a bug in the implementation of the prelude operations
\code{quot} and \code{rem} w.r.t.\ negative numbers and free variables
which was undetected for a couple of years.
Although the bug was easy to fix, the definition of a general property
relating both operations and testing it
with all smaller values was essential for its detection.
The run time of CurryCheck mainly depends on the specific
properties to be tested.
The initial translation phase, which extracts properties,
contracts, and specifications from a given module and transforms
them into executable tests, is a straightforward compilation
process.
The run time of the subsequent test execution phase
depends on the number of test cases and the time needed to evaluate
each property.
The functional logic programming technique to generate
test data described in Sect.~\ref{sec:testproperties}
(i.e,. collecting all non-deterministic results of evaluating
a logic variable)
is reasonable in practice.
For instance, KiCS2 needs 0.6 seconds to test a trivial property
on a list of integers with 10,000 test cases
computed by the randomized level diagonalization
strategy described in \cite{ChristiansenFischer08FLOPS}
(on a Linux machine with Intel Core i7-4790/3.60Ghz and 8GiB of memory).
\section{Related Work}
\label{sec:related}
Since testing is an important part of the software development process,
there is a vast literature on this topic. In the following,
we compare our approach to testing, in particular, property-based testing,
in declarative languages. We already mentioned
QuickCheck \cite{ClaessenHughes00} which was influential
in this area and followed by other property-testing systems
for functional languages, like
GAST \cite{KoopmanAlimarineTretmansPlasmeijer03}
or SmallCheck \cite{RuncimanNaylorLindblad08}.
The same idea has also been transferred to other languages
like PropEr \cite{PapadakisSagonas11} for Erlang and
PrologCheck \cite{AmaralFloridoSantosCosta14} for Prolog.
In contrast to CurryCheck, most of these systems (except for SmallCheck)
are based on randomly
generating test data so that they do not provide guarantees for a complete
enumeration if the sets of input values are finite, i.e., they cannot verify
properties.
PropEr also supports contract checking but these function contracts are limited
to type specifications.
PrologCheck could also check operational aspects likes modes
or multiplicity of answers, whereas our properties
are oriented towards declarative aspects, i.e., the input/output
relation of values.
Closely related to CurryCheck is EasyCheck \cite{ChristiansenFischer08FLOPS}
since it can be seen as our back end.
EasyCheck is the only property-based test tool
covering functional and logic aspects but it is more limited than CurryCheck.
EasyCheck does not support polymorphic properties, I/O properties, or
combinators for user-defined generation of test data.
This has been added in CurryCheck together with a full automation
of the test process in order to obtain an easily usable tool.
Moreover, CurryCheck expands the use of automatic
testing by using it for contract and specification checking,
where functional logic programming has been shown to be an appropriate
framework \cite{AntoyHanus12PADL},
and combines it with static verification techniques.
\section{Conclusion}
\label{sec:conclusion}
We have presented CurryCheck, the first fully automatic tool
to test functional as well as non-deterministic properties
of Curry programs. CurryCheck supports unit tests and tests of
I/O operations with fixed inputs as well as property tests
which are parameterized over some arguments.
In the latter case, they are executed with test inputs
which are systematically generated for the given argument types.
Moreover, CurryCheck also supports specification and contract testing
if such information is present in the source program.
To simplify and, thus, enhance the use property testing,
properties can be interspersed in the source program
and are automatically extracted by CurryCheck.
Hence, CurryCheck supports test-driven program development methods
like extreme programming.
Properties are not only useful to obtain more reliable programs,
but they can also be used by automated documentation tools
to describe the intended meaning of operations,
a feature which has been recently added to the
CurryDoc \cite{Hanus02WFLP} documentation tool.\footnote{%
See \url{www.informatik.uni-kiel.de/~pakcs/lib/Combinatorial.html}
for an example.}
For future work we plan to extend the functionality
of CurryCheck (the current version does not support
the generation of floating point numbers and functional values).
Furthermore, we intend to integrate into CurryCheck
more features that can help to improve the reliability of the
source code, like abstract interpretation to approximate
specific run-time properties
\cite{FaehndrichLogozzo11,StulovaMoralesHermenegildo16},
or program covering to show
whether the test data was sufficient to reach all parts
of a source program.
\paragraph{Acknowledgements.}
The author is grateful to Jan-Patrick Baye
for implementing an initial version of CurryCheck.
|
\section{Introduction} \label{Introduction}
The Modularity Theorem\cite{CON01} and Eichler-Shimura\cite{EICH57}\cite{SHI59} theory enable weight $2$ newforms for $\Gamma_0(N)$ with integral coefficients to be uniquely associated to every isogeny class of elliptic curves over $\mathbb{Q}$. Within these results are explicit methods for constructing the $2$-dimensional complex lattice $\Lambda_E$ associated to an elliptic curve $E$ given a weight $2$ newform.
Throughout the paper, fix the weight $2$ newform associated to an elliptic curve $E/\mathbb{Q}$ as
\begin{equation}\label{e:modform}
f_E(z) = \sum_{n = 1}^{\infty} a_E(n) q^n; \quad q = \exp(2\pi i z); \qquad z \in \mathcal{H}, \end{equation}
where $\mathcal{H}$ is the upper half-plane.
Here we denote the complex analytic realization of an elliptic curve as $\mathbb{C}/\Lambda_E$, and its modular parametrization as $\phi_E: X_0(N) \rightarrow \mathbb{C}/\Lambda_E$. The modular degree is the degree of the map $\phi_E$. We consider in this paper only elliptic curves $E$ of modular degree $1$ with conductor $N$ such that $\text{genus}(X_0(N)) = 1$, which restricts $N$ to the finite set $\{11, 14, 15, 17, 19, 21, 27, 32, 36, 49\}$.
Our expressions make use of the Weierstrass mock modular form associated to $E$ (a more detailed exposition can be found in Section~\ref{mockmodular}). These functions were recently introduced by Alfes, Bringmann, Griffin, Guerzhoy, Ono, and Rolen in \cite{ALF15,BRI13} and arise from the Weierstrass $\zeta$-function. Recall that the classical description of an elliptic curve, $E \simeq \mathbb{C}/\Lambda_E$, for some $2$-dimensional complex lattice $\Lambda_E$. Then the \emph{Weierstrass $\zeta$-function} is defined by
\begin{equation}\label{d:zeta}
\zeta(\Lambda_E; z) = \frac{1}{z} + \sum_{w \in \Lambda_E \setminus \{0\}} \left( \frac{1}{z-w} + \frac{1}{w} + \frac{z}{w^2} \right).
\end{equation}
It is related to the Weierstrass $\wp$-function by differentiation
\[\frac{d}{dz}\zeta(\Lambda_E; z) = -\wp(\Lambda_E; z). \]
We consider $\mathcal{E}_{f_E}(z)$, the \emph{Eichler integral} of $f_E$:
\begin{equation}\label{e:eichler}\mathcal{E}_{f_E}(z) = \sum_{n = 1}^\infty \frac{a_E(n)}{n}q^n. \end{equation}
Note that this is essentially the antiderivative of the newform $f_E$ associated to the elliptic curve $E$.
Although the Weierstrass $\wp$-function is doubly periodic with respect to $\Lambda_E$, the $\zeta$-function $\zeta(\Lambda_E;z)$ is not. Motivated by this, Eisenstein \cite{Ono09} constructed a modification of $\zeta(\Lambda_E;z)$, a lattice invariant, non-holomorphic function $\mathfrak{Z}_E(z)$, defined by
\begin{equation}\label{e:zedE} \mathfrak{Z}_E(z) = \zeta(\Lambda_E; z) - S(\Lambda_E)z - \frac{\pi}{\vol(\Lambda_E)} \overline{z}, \end{equation}
where
\begin{equation} \label{d:slambda} S(\Lambda_E) = \lim_{s \rightarrow 0} \sum_{w \in \Lambda_E \setminus \{0\}} \frac{1}{w^2|w|^{2s}}. \end{equation}
This value $S(\Lambda_E)$ is essentially the weight $2$ Eisenstein series for the lattice $\Lambda_E$. Using this, we define the function $\widehat{\mathfrak{Z}}_E(z)$ as the evaluation of $\mathfrak{Z}_E(z)$ at the holomorphic Eichler integral $\mathcal{E}_{f_E}(z)$,
\begin{equation}\label{e:zed}
\widehat{\mathfrak{Z}}_E(z) = \mathfrak{Z}_E(\mathcal{E}_{f_E}(z)).
\end{equation}
Then $\widehat{\mathfrak{Z}}_E(z)$ is a Harmonic Maa{\ss} form (see Section~\ref{mockmodular}) and can be written as a sum of a holomorphic and non-holomorphic part
\[\widehat{\mathfrak{Z}}_E(z) = \widehat{\mathfrak{Z}}^+_E(z) + \widehat{\mathfrak{Z}}^-_E(z).\] We refer to the holomorphic function $\widehat{\mathfrak{Z}}_E^+(z)$ as the \textit{Weierstrass mock modular form} associated to $E$. This follows from the proof that $\widehat{\mathfrak{Z}}_E(z)$ is a weight $0$ harmonic Maa{\ss} form since $\widehat{\mathfrak{Z}}_E^+(z)$ is holomorphic when the modular degree is $1$ (see the work of Alfes, Griffin, Ono and Rolen in \cite{ALF15}).
The Hasse-Weil $L$-function $L(E,s) = L(f_E, s)$ plays a central role in the arithmetic of $E$. Indeed, the Birch and Swinnerton-Dyer Conjecture is the assertion that the arithmetic invariants of $E/\mathbb{Q}$ are encoded by the analytic behavior of $L(E,s)$ at $s=1$. There are further $L$-function associated to $E$. Here, we consider the \emph{shifted convolution} $L$-functions evaluated at $s=1$ defined by
\begin{equation}\label{e: Dhat} D_{f_E}( h; s) = \sum_{n = 1}^{\infty} a_E(n + h) \overline{a_E(n)} \left( \frac{1}{(n+h)^s} - \frac{1}{n^s} \right).\end{equation} For convenience, we denote the generating function of these values in $h$-aspect by \begin{equation}\label{e:lser}\mathbb{L}_{f_E}(z) = \sum_{h = 1}^{\infty} D_{f_E}(h; 1)q^h.\end{equation}
Motivated by the problem of numerically computing these values and understanding the explicit construction of the Weierstrass mock modular form, we offer a closed formula for these generating functions $\mathbb{L}_{f_E}(z)$ for newforms associated to elliptic curves $E$ of modular degree $1$ and conductor $N$ such that $\text{genus}(X_0(N)) = 1$. To make this precise, let $F^{\rho(i)}_{N,2}$ be the Eisenstein series for $\Gamma_0(N)$ nonvanishing and normalized to be $1$ only at the cusp $\rho(i)$ and vanishing at all other cusps (as in Section~\ref{eisenstein}).
Throughout this paper, for any function $f$, we let $f[h]$ denote the coefficent of $q^h$ in the Fourier expansion of $f$, and recall that we let $f_E = \sum_{n = 1}^{\infty} a_E(n) q^n$ denote the weight $2$ newform associated to the elliptic curve $E$.
For the first result, we restrict to the case where $N$ is squarefree, restricting to the case where $N$ is in the set $\{11, 14, 15, 17, 19, 21\} $.
\begin{thm} \label{t:thm1}
Assume the notation and hypotheses above. Then, we have that
$$\mathbb{L}_{f_E}(z) = \frac{\vol(\Lambda_E)}{\pi} \left( (f_E(z) \cdot \widehat{\mathfrak{Z}}_E^+(z)) - \alpha f_E(z) - F^{\infty}_{N,2}(z) \right),$$ where $$\alpha = (f_E \cdot \widehat{\mathfrak{Z}}_E^+)[1] - \frac{\pi}{\vol{\Lambda_E}}\mathbb{L}_{f_E}(1; 1) - F^{\infty}_{N,2}[1].$$
\end{thm}
If we drop the condition that $N$ is squarefree and instead suppose that $E$ has complex multiplication (restricting $N$ to be in the set $\{27, 32, 36, 49\}$), we can prove a stronger result.
\begin{thm}\label{t:thm1.2}
Let $E$ be an elliptic curve as above with complex multiplication. With $\mathbb{L}_{f_E}(z)$ defined as above, we have
\[\mathbb{L}_{f_E}(z) = \frac{\vol(\Lambda_E)}{\pi}\big((f_E(z) \cdot \widehat{\mathfrak{Z}}_E^+)(z) - F^\infty_{N,2}(z) \big). \]
\end{thm}
These identities reveal a number of relations between weight $2$ newforms and shifted convolution $L$-values that can be interpreted in many ways. Notably, these identities can be used to compute the shifted convolution $L$-values to arbitrary precision.
\section{Preliminaries}\label{prelim}
\subsection{Harmonic Maa{\ss} Forms and Mock Modular Forms}
We begin with a review of harmonic Maa{\ss} forms. These real-analytic modular forms were first introduced by Bruinier and Funke in \cite{BRU04}. Among many other important roles these forms play in number theory, work by Zwegers (see \cite{ZW03}) shows that Ramanujan's examples of mock $\theta$-functions arise from holomorphic parts of harmonic Maa{\ss} forms. These forms are also intimately connected to our study of the holomorphic projections associated to certain elliptic curves.
Let $\mathcal{H} = \{x + iy : x, y \in \mathbb{R},\, y > 0\}$ denote the upper half-plane. Consider $z = x+iy \in \mathcal{H}$ and as noted in the introduction, let $q = e^{2\pi i z}$. For any $\gamma = \begin{bmatrix}
a & b \\
c & d \\
\end{bmatrix} \in \SL_2(\mathbb{Z})$ we abbreviate the action of $\gamma$ on a function $f$ by the Petersson slash operator of weight $k$, $$(f |_k \gamma)(z) = (cz + d)^{-k} f \left(\frac{az + b}{cz + d} \right).$$
Consider the congruence subgroup of level $N$, defined as follows: $$\Gamma_0(N) = \left\{ \begin{bmatrix}
a & b \\
c & d \\
\end{bmatrix} \in \SL_2(\mathbb{Z}) \,|\, c \equiv 0 \hspace{-3pt} \mod N
\right \}.$$
\begin{defn}(\hspace{-4pt} \cite{BRU04}) \label{d:harmonicmaass}
A \textit{weak harmonic Maa{\ss} form} of weight $k \in \mathbb{Z}$ on $\Gamma_0(N)$ is a smooth function $f$ on $\mathcal{H}$ that satisfies the following three conditions:
\begin{enumerate}
\item $(f |_k \gamma)(z) = f(z)$ for all $\gamma \in \Gamma_0(N)$ (i.e. $f$ transforms like a modular form on $\Gamma_0(N)$),
\item $\Delta_k f(z) \equiv 0$, where if $z = x + iy$, $\Delta_k = -y^2 \left( \frac{\partial^2}{\partial x^2} + \frac{\partial^2}{\partial y^2} \right) + 2 i k y \left( \frac{\partial}{\partial x} + i \frac{\partial}{\partial y} \right)$
\item $f(z)$ has poles at most at the cusps of $\Gamma_0(N)$ (i.e. there exists some polynomial $P(z)$, such that $f(z) - P(q^{-1}) = O(e^{- \alpha y})$ as $y \rightarrow \infty$ and similar is true for the other cusps of $\Gamma_0(N)$).
\end{enumerate}
\end{defn}
Note that hereafter, we will simply term these weight $k$ harmonic Maa{\ss} forms for $\Gamma_0(N)$. Following immediately from the definition above, we see that a harmonic Maa{\ss} form admits a Fourier expansion.
\begin{prop}
A weight $k$ harmonic Maa{\ss} form $f(z)$ admits a Fourier expansion of the form,
$$f(z) = f^+(z) + f^-(z);$$
$$f^+(z) = \sum_{n \gg -\infty} c_f^+(n) q^n; \qquad f^-(z) = \sum_{n = 1}^{\infty} c_f^-(n) q^n \Gamma(k-1, 4 \pi n y).$$
Here, $\Gamma(\alpha, \beta) = \int_\beta^\infty e^{-t}t^{\alpha - 1}dt$ denotes the incomplete Gamma function, and $f^+(z)$ and $f^-(z)$ are the holomorphic and non-holomorphic parts of $f(z)$ respectively.
\end{prop}
This construction also gives rise to a characterization of mock modular forms in terms of harmonic Maa{\ss} forms:
\begin{defn}
Consider $f(z) = f^+(z) + f^-(z)$ as defined above. If $f^-(z) \neq 0$, then $f^+(z)$ is called a \emph{mock modular form}.
\end{defn}
We define a differential operator (as in \cite{Ono09}) to help characterize $f^-(z)$ when it is nontrivial. Denote by $H_{k}(\Gamma)$ the space of weight $k$ harmonic Maa{\ss} forms for some congruence subgroup $\Gamma \le \SL_2(\mathbb{Z})$.
\begin{prop}\label{p:xi}
Define a differential operator $\xi_{2-k}: H_{2-k}(\Gamma) \rightarrow S_k(\Gamma)$ where $\Gamma \le \SL_2(\mathbb{Z})$ as $$\xi_{2-k}(f(z)) = 2 i y^k \frac{\overline{\partial f}}{\partial \overline{z}}.$$
Then $\xi_{2-k}$ is a well-defined, surjective, antilinear map with kernel $M_{2-k}^!(\Gamma)$, the space of weakly holomorphic weight $2-k$ modular forms for $\Gamma$.
Further, $$\xi_{2-k} f(z) = -(4\pi)^{k-1} \sum_{n = 1}^{\infty} c_f^-(n) q^n.$$
\end{prop}
We term $\sum_{n = 1}^{\infty} c_f^-(n) q^n$ the \emph{shadow} of $f(z)$ or $f^+(z)$. When the shadow is non-trivial, we retreive some conditions on the cusp behavior of $f(z)$.
\begin{prop} (Lemma 2.3 of~\cite{BR12}) \label{p:nonconstant}
If $f(z) \in H_{2-k}(\Gamma_0(N))$ has the property that $\xi_{2-k}(f) \neq 0$, then the principal part of $f(z)$ is nonconstant for at least one cusp.
\end{prop}
\subsection{Weierstrass Mock Modular Forms}\label{mockmodular}
We begin the construction of the Weierstrass mock modular form associated with an elliptic curve $E$ of conductor $N$, which will prove very useful to our analysis. This Weierstrass mock modular form was introduced in \cite{BRI13} as a mechanism by which to understand properties of elliptic curves and their associated newforms through the language of harmonic Maa{\ss} forms. The Weierstrass mock modular form has been one of the primary objects of study by many recently, as in \cite{ALF15,CL16}.
The theory of elliptic curves gives rise to a notable example of a weight $0$ harmonic Maa{\ss} form. Recall that $E \simeq \mathbb{C}/ \Lambda_E$ where $\Lambda_E$ is a $2$ dimensional complex lattice. Recall the Weierstrass $\zeta$-function defined in~\eqref{d:zeta}. Although it is not elliptic, its derivative is negative the Weierstrass $\wp$-function. This relation gives the Laurent expansion of $\zeta$:
\begin{prop}\cite{DIA06}
The Laurent expansion of $\zeta$ is
\[ \zeta(\Lambda_E; z) = \frac{1}{z} - \sum_{k = 1}^\infty G_{k+2}(\Lambda_E) z^{2k+1},\]
where $G_k(\Lambda_E)$ is the Eisenstein series of weight $k$ associated to a lattice $\Lambda_E$.
\end{prop}
Again, recall the construction of the Weierstrass mock modular form given in \eqref{e:zed}:
\begin{align*}
\mathfrak{Z}^+_E(\mathcal{E}_{f_E}(z)) &= \zeta(\Lambda_E;\mathcal{E}_{f_E}(z)) - S(\Lambda_E)\mathcal{E}_{f_E}(z).
\end{align*}
The following theorem outlines some important properties about the Weierstrass mock modular form.
\begin{thm}(\hspace{-4pt} \cite{BRI13})
Assume the notation and hypotheses above. Then
\begin{enumerate}
\item The holomorphic part $\widehat{\mathfrak{Z}}^+_E(z) = \zeta(\Lambda_E;\mathcal{E}_f(z)) - S(\Lambda_E)\mathcal{E}_f(z)$ has poles exactly when $\mathcal{E}_{f_E}(z)$ is a lattice point.
\item If $\widehat{\mathfrak{Z}}^+_E(z)$ has poles in the upper half plane, there is a canonical meromorphic modular function $M_E(z)$ such that $\widehat{\mathfrak{Z}}^+_E(z) - M_E(z)$ is holomorphic on $\mathcal{H}$.
\item $\widehat{\mathfrak{Z}}_E(z) - M_E(z)$ is a harmonic Maa{\ss} form of weight $0$ on $\Gamma_0(N)$ and $\xi_0(\widehat{\mathfrak{Z}}^+_E(z) - M_E(z)) = -(4\pi)\cdot f_E(z)$. In particular, $\widehat{\mathfrak{Z}}^+_E(z) - M_E(z)$ is a weight $0$ mock modular form.
\end{enumerate}
\end{thm}
In particular, $\widehat{\mathfrak{Z}}_E^+(z) - M_E(z)$ is called the Weierstrass mock modular form of $E$.
The kernel of the surjective $\xi_{2-k}$ operator defined above is infinite dimensional. Selecting a suitable class of harmonic Maa{\ss} forms to serve as preimages under this $\xi$ operator depends intimately on the following notion:
\begin{defn}
A harmonic weak Maa{\ss} form $F(z) \in H_{2-k}(\Gamma_0(N))$ is \textit{good} for $f(z) \in S_k(\Gamma_0(N))$ if it satisfies the following $3$ conditions:
\begin{enumerate}
\item The principal part of $F(z)$ at the cusp $\infty$ is in $\mathbb{F}_f[q^{-1}]$, where $\mathbb{F}_f$ is the field obtained by adjoining the Fourier coefficients of $f$ to $\mathbb{Q}$.
\item The principal part of $F(z)$ at all other inequivalent cusps of $\Gamma_0(N)$ is constant
\item $\xi_{2-k}(F(z)) = \| f(z)\|^{-2} f(z)$ where $\xi_{2-k}$ is the differential operator defined in Proposition~\ref{p:xi}.
\end{enumerate}
\end{defn}
This definition enables us to state the following result:
\begin{prop}[\hspace{-4pt} \cite{MER16}]\label{p:good}
The weak harmonic Maa{\ss} form $\widehat{\mathfrak{Z}}_E(z) - M_E(z)$, as defined above, is good for $f_E(z)$, where $f_E(z)$ is the weight $2$ newform associated to $E$.
\end{prop}
In particular, we will be interested in the case where $\widehat{\mathfrak{Z}}^+_E(z)$ itself is a mock modular form of weight $0$. This arises when one can choose $M_E(z)$ to be identically zero. The following proposition gives sufficient conditions for this to occur.
\begin{lemma}\label{p:pole}
Let $E$ be an elliptic curve of conductor $N$ with modular parametrization $\phi_E: X_0(N) \rightarrow E$. If the modular degree $\deg(\phi_E)$ is $1$, then $\widehat{\mathfrak{Z}}^+_E(z)$ does not have poles in the upper half-plane.
\end{lemma}
\begin{proof}
Recall the mock modular form $\widehat{\mathfrak{Z}}_E^+(z)$ is given by the Fourier expansion in~\ref{e:zed}:
\begin{align*} \widehat{\mathfrak{Z}}_E^+(z) &= \widehat{\mathfrak{Z}}_E(\mathcal{E}_f(z)) \\
&= \frac{1}{\mathcal{E}_f(z)} - \sum_{k = 1}^\infty G_{2k+2}(\Lambda_E)\mathcal{E}_f(z)^{2k+1} - S(\Lambda_E)\mathcal{E}_f(z).
\end{align*}
The Eichler integral $\sum_{n = 1}^\infty \frac{a_E(n)}{n}q^n$ is holomorphic on the upper half-plane, so it suffices to show that $\mathcal{E}_f(z)$ does not vanish for any $z \in \mathcal{H}$.
The modular parametrization $\phi: X_0(N) \rightarrow \mathbb{C}/\Lambda_E$ is induced from the map $\phi_1: \mathcal{H} \rightarrow \mathbb{C}$ that can be given by
\begin{align*} \phi_1(z) &= - 2\pi i \int_z^{i\infty} f(\tau)d\tau \\
&= \sum_{n = 1}^\infty \frac{a_E(n)}{n}q^n = \mathcal{E}_f(z). \end{align*}
If the modular degree $\deg(\phi)$ is $1$ (requiring that the genus of $X_0(N)$ is $1$, since $\deg(\phi_E) \geq \text{genus}(X_0(N))$, then the map $\phi$ is an isomorphism. Thus, $\mathcal{E}_f(z)$ does not vanish for any $z \in \mathcal{H}$.
\end{proof}
Moreover, $\widehat{\mathfrak{Z}}^+_E(z)$ has rational Fourier coefficients if $E$ has complex multiplication (see Theorem 1.3 in \cite{BRU08}).
\subsection{Maa{\ss}-Poincar\'e Series}
The modular parametrization of an elliptic curve $E$ is given by a map $\phi_E: X_0(N) \rightarrow E$ where $X_0(N)$ is the compactification of the curve $\Gamma_0(N) \setminus \mathcal{H}$. Let $f_E$ be the weight $2$ newform associated to this parametrization. The Petersson norm of $f_E$ is then
\[\|f_E \|^2 = \langle f_E, f_E \rangle = \int_{z \in \Gamma_0(N) \setminus \mathcal{H}} |f(z)|^2dx \wedge dy. \]
Using Petersson norms, we can relate the degree of $\phi_E$ with the area of the fundamental parallelogram of the period lattice $\Lambda_E$, which is the volume of the elliptic curve, $\vol(\Lambda_E)$.
\begin{prop}[\hspace{-4pt} \cite{ZAG85}]
The volume $\vol(\Lambda_E)$ of an elliptic curve $E$ is
\[\vol(\Lambda_E) = \frac{4\pi^2 \| f_E\|^2}{\deg(\phi_E)}. \]
\end{prop}
The Petersson inner product can also be used to extract Fourier coefficients of cusp forms through Poincar\'e series.
A generic index $m$ Poincar\'e series is given by
\begin{equation}
\mathbb{P}(m,k,\phi_m,N;z) = \sum_{\gamma \in \Gamma_\infty \setminus \Gamma_0(N)} (\phi_m^* |_k \gamma)(z),
\end{equation}
where $\Gamma_\infty = \{ [\begin{smallmatrix} 1 & n \\ 0 & 1 \end{smallmatrix}] : n \in \mathbb{Z} \}$ and $\phi_m^*(z) = \phi_m(y)\exp(2\pi i m x)$ for a function $\phi_m: \mathbb{R}_{> 0} \rightarrow \mathbb{C}$ which satisfies $\phi_m(y) = O(y^\alpha)$ as $y \rightarrow 0$ for some $\alpha \in \mathbb{R}$.
Using this, the classical index $m$ Poincar\'e series $P(m, k, N; z)$ and the Maa{\ss}-Poincar\'e series $Q(-m,k,N;z)$ are defined as
\begin{align}
P(m,k,N;z) &= \mathbb{P}(m,k,\exp(-my), N;z) \\
Q(-m, k, N; z) &= \mathbb{P}(-m, 2-k, N, \mathcal{M}_{1-\frac{k}{2}}(-4\pi m y); z)
\end{align}
where $\mathcal{M}_s(y)$ is defined in terms of the $M$-Whittaker function
\[\mathcal{M}_s(y) = |y|^{-\frac{k}{2}}M_{\frac{k}{2}\text{sgn}(y),s-\frac{1}{2}}(|y|). \]
We can characterize a set of Maa{\ss}-Poincar\'e series in terms of the above functions, with a Fourier expansion given using Bessel functions and the Kloosterman sum $K(m, n; c)$ we recall below:
$$K(m, n;c) = \sum_{d \textnormal{ mod }c; (c, d) = 1} \exp\left( 2\pi i \frac{m\overline{d} + nd}{c} \right),$$ where $\overline{d}$ is the multiplicative inverse of $d$ modulo $c$.
The Fourier expansion and behavior of the Poincar\'e series at the cusps of $\Gamma_0(N)$ is explained by the following proposition.
\begin{prop}(6.2 in \cite{BRU08}, 3.3 in \cite{RH12})\label{p:poincusp}
If $k \in 2\mathbb{N}$, and $m, N \geq 1$, then $Q(-m, k, N;z) \in H_{2-k}(\Gamma_0(N))$, and has Fourier expansion of the form
\[Q(-m,k,N;z) = Q^+(-m,k,N;z) + Q^-(-m, k, N; z),\]
where
\[Q^+(-m, k, N;z) = q^{-m} + \sum_{n = 0}^\infty b_Q(-m,k,N;n)q^n \]
and for integers $n \geq 0$ we have
\begin{align*}
b_Q(-m, k, N; n) &= -2\pi (-1)^{k/2} \cdot \sum_{\substack{c > 0 \\ c \equiv 0 \pmod N}} \left( \frac{m}{n}\right)^{\frac{k-1}{2}} \frac{K(-m, n, c)}{c} \cdot I_{k-1}\left(\frac{4\pi \sqrt{|mn|}}{c} \right), \\
b_Q(-m, k, N; 0) &= -\frac{2^k \pi^k (-1)^\frac{k}{2} m^{k-1}}{(k-1)!} \cdot \sum_{\substack{c > 0 \\ c \equiv 0 \pmod N}} \frac{K(-m, 0, c)}{c^k}.
\end{align*}
Additionally, the principal part at all other cusps is zero.
\end{prop}
The classical Poincar\'e series $P(m, k, N;z)$ and the Maa{\ss}-Poincar\'e series are also related by the differential operator $\xi_{2-k}$ in the following proposition.
\begin{prop}(2.6 in \cite{MER16})
If $k \geq 2$ is even and $m, N \geq 1$, then
\[\xi_{2-k}(Q(-m, k, N; z) = (4\pi)^{k-1} m^{k-1} (k-1) \cdot P(m, k, N; z) \in S_k(\Gamma_0(N)). \]
\end{prop}
We can also associate a Fourier expansion to these Poincar\'e series that uses $J$-Bessel functions and Kloosterman sums.
\begin{prop}[\S3 \cite{LIU06}]\label{p:petFourier}
Consider the weight $k$ Poincar\'e series of index $m$ and level $N$, $P(m, k, N; z)$. Then this Poincar\'e series has a Fourier expansion as $P(m, k, N; z) = \sum_{n = 0}^{\infty} b_P(m, k, N;n) q^n$ where $b_P(m,k,N;n)$ can be defined as follows (when $k \equiv 0 \pmod{2}$):
$$b_P(m,k,N;n) = \left(\frac{n}{m}\right)^{(k-1)/2} \left(\delta_{m, n} + 2\pi i^{-k} \sum_{c > 0; N | c} J_{k -1} \left( \frac{2\pi \sqrt{mn}}{c} \right) \frac{K(m, n; c)}{c} \right).$$
\end{prop}
\subsection{Holomorphic Projection and Shifted Convolution Dirichlet Series}\label{holproj}
The study of holomorphic projection is motivated by a desire to transform a general modular form into a holomorphic modular form such that linear functionals and the Petersson inner product are respected. The holomorphic projection was first introduced in \cite{STU80} by Sturm and further developed in the work of Gross and Zagier in \cite{GRO86}. We will use it here to give a closed-form algebraic characterization of the shifted convolution $L$-series values for some modular forms associated to elliptic curves. Our approach follows previous work as in \cite{ALF15, BRI13, MER16, BRI16}.
We define the holomorphic projection for continuous functions $f: \mathcal{H} \rightarrow \mathbb{C}$, where $f(z) = \sum_{n \in \mathbb{Z}} a(n , y)q^n$ that transform like a modular form of weight $k \ge 2$ for $\Gamma_0(N)$ and has moderate growth at the cusps. We can make this idea more precise. Suppose the cusps of $\Gamma_0(N)$ are $\rho(i)$ (where $\rho(1)$ is chosen to be the infinite cusp), and consider $\sigma_i \in \Gamma_0(N)$ so that $\sigma_i \infty = \rho(i)$. Then $f$ has \emph{moderate growth at the cusps} if for $n > 0$ $$a(n, y) = O(y^{2-k}), \quad y \rightarrow 0,$$ and $$f |_k\, \sigma_i = c_0^{(i)} + O(e^{-\alpha y}), \quad y \rightarrow \infty.$$
\begin{defn}[\hspace{-4pt} \cite{STU80}]
Consider a continuous function $f(z) = \sum_{n \in \mathbb{Z}} a(n , y)q^n$ as above (where $z = x + iy$) that transforms like a modular form of weight $k \ge 2$ for $\Gamma_0(N)$ and has moderate growth at the cusps. Then, the \textit{holomorphic projection} of $f(z)$, denoted $\pi_{\text{hol}}(f)(z)$ is constructed as follows: $$\pi_{\text{hol}}(f)(z) = c_0^{(1)} + \sum_{n = 1}^{\infty} c(n) q^n;$$ $$c(n) = \frac{(4 \pi n)^{k-1}}{(k-2)!} \int_{0}^{\infty} a(n, y) e^{-4 \pi n y}y^{k - 2} dy.$$
\end{defn}
The holomorphic projection of a function satisfies several natural properties that arise from its construction:
\begin{prop}[\hspace{-4pt} \cite{GRO86}]\label{p:holprops}
Consider $f$ as defined above. Then the holomorphic projection of $f$ satisfies the following three properties.
\begin{enumerate}
\item If $f$ is a holomorphic modular form, $\pi_{\text{hol}}(f) = f$
\item If $k > 2$, $\pi_{\text{hol}}(f) \in M_k(\Gamma_0(N))$, the space of weight $k$ modular forms for $\Gamma_0(N)$. If $k = 2$, $\pi_{\text{hol}}(f) \in M_2(\Gamma_0(N)) \oplus \mathbb{C} E_2 = \tilde{M_2}(\Gamma_0(N))$, the space of weight $2$ quasi-holomorphic modular forms for $\Gamma_0(N)$.
\item $\langle g, f \rangle = \langle g, \pi_{\text{hol}}(f) \rangle$ for any $g \in S_k(\Gamma_0(N)).$
\end{enumerate}
\end{prop}
We can understand the holomorphic projection more explicitly in specific cases, such as the following product of a harmonic Maa{\ss} form and a cusp form.
\begin{prop}[\hspace{-4pt} \cite{MER16}] \label{p:hol}
Let $M_{f_1}$ be the weight $2-k$ harmonic Maa{\ss} form whose shadow is $f_1 \in S_k(\Gamma_0(N))$ where $f_1(z) = \sum_{n = 1}^{\infty} a_1(n)q^n$, so $\xi_{2 - k} M_{f_1} = -(4 \pi)^{k-1} f_1$. Consider also the weight $2$ cusp form for $\Gamma_0(N)$ $f_2(z) = \sum_{n = 1}^{\infty} a_2(n)q^n$. Then, \begin{equation}\label{e:piholMff} \begin{split} \pi_{\text{hol}} (M_{f_1} \cdot f_2)(z) & = M_{f_1}^+ (z)\cdot f_2(z) \\ & - (k - 2)! \sum_{h = 1}^{\infty} \left[ \sum_{n = 1}^{\infty} a_2(n+h) \overline{a_1(n)} \left( \frac{1}{(n + h)^{k-1}} - \frac{1}{n^{k-1}} \right) \right] q^h. \end{split}\end{equation}
\end{prop}
Now consider a strong Weil curve $E$ with associated weight $2$ newform $f_E$ for $\Gamma_0(N)$, where the genus of the modular curve $X_0(N)$ is $1$, and let $\widehat{\mathfrak{Z}}_E$ be defined as in~\eqref{e:zed}. Then, we can compute the holomorphic projection of $f_E \cdot \widehat{\mathfrak{Z}}_E$ as follows:
\begin{lemma}\label{p:picomputation}
Let $E$ be a strong Weil curve with associated weight $2$ newform $f_E$ for $\Gamma_0(N)$, where the $\text{genus}(X_0(N)) = 1$ and $\deg(\phi_E) = 1$, and let $\widehat{\mathfrak{Z}}_E$ be defined as in~\eqref{e:zed}. Then, we have the following:
$$\pi_{\text{hol}}(f_E \cdot \widehat{\mathfrak{Z}}_E)(z) = \frac{\vol(\Lambda_E)}{\pi} f_E(z) \widehat{\mathfrak{Z}}_E^+(z) - \sum_{h = 1}^{\infty} D_{f_E}(h; 1) q^h.$$
\end{lemma}
\begin{proof}
Since $\text{dim}(S_2(\Gamma_0(N))) = \text{genus}(X_0(N)) = 1$, the modular form $f_E$ is a scalar multiple of the Poincaré series $f_E = \frac{1}{\beta} P(1, 2, N;z)$, where $P(m, k, N;z)$ is the Poincaré series described in Proposition~\ref{p:petFourier}. Then the Petersson coefficient formula (\S3 \cite{LIU06}) and Proposition~\ref{d:slambda} yields $\beta = \frac{\pi}{\vol(\Lambda_E)}$. Following the computation in Corollary 1.2 of~\cite{MER16}, we obtain the holomorphic projection in terms of the Poincaré series
\[\pi_{\text{hol}}(f_E \cdot \widehat{\mathfrak{Z}}_E)(z) = P(1, 2, N; z) \widehat{\mathfrak{Z}}_E^+(z) - \sum_{h = 1}^{\infty} \mathbb{L}_{f_E}[h] q^h. \]
We can apply Proposition~\ref{p:hol} with $P(1,2,N; z) = \frac{1}{\beta} f_E$, which yields the Lemma.
\end{proof}
\begin{rem}
For the remainder of the paper, we will define a new function $\widehat{\pi}_{\text{hol}}$, a scalar multiple of the holomorphic projection by the constant $\frac{\pi}{\vol(\Lambda_E)}$ for ease of algebraic computation and numerical characterization. Thus, we will say $$\widehat{\pi}_{\text{hol}}(f_E \cdot \widehat{\mathfrak{Z}}_E)(z) = f_E(z) \widehat{\mathfrak{Z}}_E^+(z) - \frac{\pi}{\vol(\Lambda_E)} \sum_{h = 1}^{\infty} D_{f_E}(h; 1) q^h.$$
\end{rem}
\subsection{Eisenstein Series}\label{eisenstein}
In order to understand $\widehat{\pi}_{\text{hol}}(f_E \cdot \widehat{\mathfrak{Z}}_E)$, we define a basis for the space of \emph{weight $2$ quasimodular forms} for $\Gamma_0(N)$, the space $\mathcal{E}_2(\Gamma_0(N)) \oplus \mathbb{C} E_2$, denoted $\widetilde{\mathcal{E}_2}(\Gamma_0(N))$. To do this, we follow the construction given in \S2 of Chapter VII in \cite{SCH12} and arrive at a set of forms $F_{N, 2}^{-a_2/a_1}$ described below:
\begin{prop}
There is a set of weight $2$ quasimodular Eisenstein forms for $\Gamma_0(N)$, say $F_{N,2}^{-a_2/a_1}$, such that if $(a_1, a_2) = 1$, then $F_{N, 2}^{-a_2/a_1}$ is $1$ at the cusps $\Gamma_0(N)$-equivalent to $-a_2/a_1$ and $0$ at the other cusps. The set of linearly independent $F_{N,2}^{-a_2/a_1}$ forms a basis for $\widetilde{\mathcal{E}}_2 (\Gamma_0(N))$.
\end{prop}
Using these Eisenstein series, we can give an alternate representation of the holomorphic projection described in the previous section.
\begin{lemma}\label{l:holprojRHS}
Consider a strong Weil curve $E$ with associated modular form $f_E$ and $\widehat{\mathfrak{Z}}_E$ as defined in~\ref{e:zed}. Then, \begin{equation}\label{e:basis}\widehat{\pi}_{\text{hol}}(f_E \cdot \widehat{\mathfrak{Z}}_E) = \alpha f_E + \sum_i \beta_i F_{N, 2}^{\rho(i)}, \quad \beta_i \in \mathbb{C} \end{equation} where $F_{N, 2}^{\rho(i)}$ is the weight $2$ quasimodular form for $\Gamma_0(N)$ that takes the value $1$ at the cusp $\rho(i)$ and vanishes at all other inequivalent cusps.
\end{lemma}
\section{Proofs of Theorems}
Throughout this sections, let $E$ be a strong Weil elliptic curve with conductor $N_E$ and associated weight $2$ newform $f_E = \sum_{n = 1}^{\infty} a_E(n)q^n$ of level $N = N_E$, where $\text{genus}(X_0(N)) = 1$.
\begin{lemma}\label{l:zvanish}
Consider $E$ as defined above. Then $\widehat{\mathfrak{Z}}_E^+$ vanishes at all cusps not equivalent to the $\infty$ cusp of $\Gamma_0(N)$.
\end{lemma}
\begin{proof}
If $E$ satisfies the above conditions, then $\widehat{\mathfrak{Z}}_E$ is good for $f_E$. In particular, $\widehat{\mathfrak{Z}}_E$ has a pole at $\infty$ and constant principal part at the cusps of $\Gamma_0(N)$ by Proposition~\ref{p:good}.
On the other hand, the index $-1$ Maa{\ss}-Poincar\'e series $Q(-1, 2,N;z)$ has a pole at $\infty$ and zero principal part at the cusps (\ref{p:poincusp}).
The difference $\widehat{\mathfrak{Z}}_E(z) - Q(-1, 2, N;z)$ is a weight $0$ harmonic Maass form with no poles and constant value at each of the cusps.
The differential operator $\xi_0$ maps harmonic Maa{\ss} forms of weight $2-k$ to cusp forms of weight $k$. Then since the dimension of $S_2(\Gamma_0(N))$ is $1$, we have
\[\xi_0(\widehat{\mathfrak{Z}}_E(z) - Q(-1, 2, N;z)) = c \cdot f_E(z) \]
for some constant $c$. However, by Lemma~\ref{p:nonconstant}, since $\widehat{\mathfrak{Z}}_E(z) - Q(-1, 2, N;z)$ has constant principal part at all cusps, $c = 0$. This implies that the difference is holomorphic, and so is constant. Then $\widehat{\mathfrak{Z}}_E(z)$ and $Q(-1, 2, N;z)$ are equal up to an additive constant since both have leading term $q^{-1}$, as are their holomorphic parts $\widehat{\mathfrak{Z}}_E^+(z)$ and $Q^+(-1, 2, N;z)$. Thus, as in \cite{RH12}, since $Q^+(-1, 2, N; z)$ vanishes at all non-infinity cusps by Lemma~\ref{p:poincusp}, $\widehat{\mathfrak{Z}}_E^+(z)$ vanishes at all non-infinite cusps.
\end{proof}
Using this Lemma, we are able to prove the first theorem.
\begin{proof}[Proof of Theorem~\ref{t:thm1}]
From Lemma~\ref{p:picomputation} and Lemma~\ref{l:holprojRHS}, we obtain that $$\widehat{\pi}_{\text{hol}}(\widehat{\mathfrak{Z}}_E \cdot f_E)(z) = f_E(z) \cdot \widehat{\mathfrak{Z}}_E^+(z) - \frac{\pi}{\vol(\Lambda_E)} \sum_{h = 1}^{\infty} D_{f_E}(h; 1) q^h = \alpha f_E + \sum_i \beta_i F_{N, 2}^{\rho(i)}(z).$$
We can compute the holomorphic projection at each cusp $\rho(i)$ gives the value of $\beta_i$. Note that the $L$-series generating function $\mathbb{L}_{f_E}(z)$ and $f_E$ both vanish at all cusps $\rho(i)$. Applying Lemma~\ref{l:zvanish}, $\widehat{\mathfrak{Z}}_E^+$ vanishes at all cusps inequivalent to the $\infty$ cusp of $\Gamma_0(N)$. Consequently the holomorphic projection vanishes at all cusps not $\Gamma_0(N)$-equivalent to $\infty$ and is $1$ at the cusp $\infty$. If we let $\rho(1)$ denote the $\infty$ cusp, $\beta_i = 0$ for $i \neq 1$ and $\beta_1 = 1$. Using this, we can compute $\alpha$ by equating the first Fourier coefficient of both expressions for $\widehat{\pi}_{\text{hol}}$. Rearranging gives the desired expression for $\mathbb{L}_{f_E}(z)$.
\end{proof}
We also consider the case where $E$ as defined in the beginning has complex multiplication, which will not only give an analogous result for the shifted convolution $L$-series values, but it will also give a recursion on the coefficients of the modular form $f_E(z)$ that can be written in terms of $G_4$ and $G_6$, an explicit example of the Modularity Theorem.
Now, we can prove Theorem~\ref{t:thm1.2}.
\begin{proof}[Proof of Theorem~\ref{t:thm1.2}]
Following the proof for Theorem~\ref{t:thm1}, we obtain that $\beta_1 = 1$, and $\beta_i = 0$ for $i \neq 1$ in Equation~\eqref{e:basis}. Since $f_E$ is only supported at coefficients $1$ mod $p$, where the left hand side is not supported (see Lemma~\ref{l:piholvanish}), $\alpha = 0$.
Thus $$\widehat{\pi}_{\text{hol}}(\widehat{\mathfrak{Z}}_E \cdot f_E)(z) = F_{N, 2}^{\infty}(z).$$
Then, we can compute the desired closed form expression for the shifted-convolution $L$-series values: \[\mathbb{L}_{f_E}(z) = \frac{\vol(\Lambda_E)}{\pi}\big((f_E(z) \cdot \widehat{\mathfrak{Z}}_E^+(z)) - F^\infty_{N,2}(z) \big). \]
\end{proof}
\begin{lemma}\label{p:lservanish}
Consider $E$ as defined above which also has complex multiplication and thus conductor $N \neq 49$ with associated modular form $f_E(z)$. Then $ D_{f_E}(h; 1) = 0$ if $h \not \equiv 0 \pmod{n_0}$.
\end{lemma}
\begin{proof}
Suppose that $E$ has complex multiplication. Then if $p$ is a prime inert in the CM field, $a_E(p) = 0$. Since $f_E$ is a weight $2$ newform, its coefficients are multiplicative. Thus, as in \S5 of~\cite{BRU08}, $a_E(n) = 0$ for all $n \not \equiv 1 \pmod{n_0}$ where $n_0 | N$ is a curve-dependent value always at least $3$. For example, in the case of the $\Gamma_0(27)$-optimal elliptic curve with complex multiplication, $n_0 = 3$. Now recall that $D_{f_E}(h; 1)$ is defined by
\[ D_{f_E}(h; 1) = \sum_{n = 1}^\infty a_E(n+h)a_E(n) \left(\frac{1}{n+h}-\frac{1}{n} \right). \]
Suppose that $ D_{f_E}(h; 1)$ is nonvanishing. Then both $a_E(n+h)$ and $a_E(n)$ must be nonvanishing and thus $n+h, n \equiv 1 \pmod{n_0}$. This yields $h \equiv 0 \pmod{n_0}$.
\end{proof}
\begin{rem}
For $N = 49$, the support of the Hecke eigenvalues $a_E(n)$ are at $n \equiv 1, 2, 4 \pmod{7}$. In this case, the proof does not hold, and the $L$-series has support everywhere.
\end{rem}
\begin{lemma}\label{l:piholvanish}
Consider $E$ with complex multiplication as defined at the beginning of the section with conductor $N \neq 49$ and associated modular form $f_E$. Then there exists some $n_0 \ge 3$ with $n_0 | N$ such that $\widehat{\pi}_{\text{hol}}(f_E \cdot \widehat{\mathfrak{Z}}_E)[h] = 0$ if $h \not \equiv 0 \pmod{n_0}$.
\end{lemma}
\begin{proof}
The derivative of the Weierstrass mock modular form $\widehat{\mathfrak{Z}}_E^+$ for the three strong Weil curves of conductor $N = 27, 32,$ and $36$ are given as eta-quotients in the following table (see \cite{CL16}).
\begin{center}
\begin{tabular}{ |c||c|c| }
\hline
$N$ & $\widehat{\mathfrak{Z}}_E^+( \cdot )$ & $q \frac{d}{dq} (\widehat{\mathfrak{Z}}_E^+)$ \\ \hline
$27$ & $q^{-1} + \dfrac12 q^2 + \dfrac15 q^5 + \dfrac34 q^8 + \cdots$ & $- \dfrac{ \eta(3\tau)\eta(9\tau)^6}{\eta(27\tau)^3}$ \\ \hline
$32$ & $q^{-1} + \dfrac23 q^3 + \dfrac17 q^7 - \dfrac{2}{11} q^{11} + \cdots$ & $- \dfrac{\eta^2(4\tau) \eta^6(16\tau)}{ \eta^{4}(32\tau)}$ \\ \hline
$36$ & $q^{-1} + \dfrac35 q^5 + \dfrac{1}{11} q^{11} + \cdots$ & $-\dfrac{\eta^3(6\tau) \eta(12\tau) \eta(18\tau)}{\eta^3(36\tau)}$ \\ \hline
\end{tabular}
\end{center}
The support of the derivative $q \frac{d}{dq}(\widehat{\mathfrak{Z}}_E^+)$ is the same as the support of the Weierstrass mock modular form. The form of the eta-quotients indicates that the support for $\widehat{\mathfrak{Z}}_E^+$ for $N = 27$ is $-1 \pmod{3}$, for $N = 32$ is $-1 \pmod{4}$, and for $N = 36$ is $-1 \pmod{6}$. Let $n_0 = 3, 4, 6$ for $N = 27, 32, 36$ respectively.
Each of these curves has complex multiplication, and their associated modular forms $f_E$ have support at $1 \pmod{n_0}$. Thus the product of the Weierstrass mock modular form $\widehat{\mathfrak{Z}}_E^+$ and $f_E$ has support at $0 \pmod{n_0}$.
By Lemma~\ref{p:lservanish}, the $L$-series is also onlylt supported at $0 \pmod{n_0}$, which yields that the holomorphic projection
\[\pi_{\text{hol}}(f_E \cdot \widehat{\mathfrak{Z}}_E^+) = f_E \cdot \widehat{\mathfrak{Z}}_E^+ - \frac{\pi}{\vol(\Lambda_E)} \mathbb{L}_{f_E}(z) \]
has support only at $0 \pmod{n_0}$.
\end{proof}
\begin{rem}
Note that results such as the above theorems may be able to be transformed to give recurrence relations on the modular form coefficients, similar to those obtained in \cite{Ono09,CHO10}.
\end{rem}
\section{Examples}\label{examples}
\subsection{Conductor $N = 11$}
Consider the modular curve $X_0(11)$ of dimension $1$. There is a single isogeny class of elliptic curves, and the strong Weil curve is given by the Weierstrass equation
\[E: y^2 + y = x^{3} - x^{2} - 10 x - 20. \]
Numerically, we find that $S(\Lambda_E) = 0.38124\dots $. Using this and the Fourier expansion of the Weierstrass $\zeta$-function, the corresponding weight $0$ mock modular form $\widehat{\mathfrak{Z}}_E^+(z)$ is given by
\[ q^{-1} + 1 + 0.9520\dots q + 1.547\dots q^2 + 0.3493\dots q^3 + 1.976\dots q^4 -2.609\dots q^5 + O(q^6).\]
Using the formula given in Proposition~\ref{p:picomputation}, one can compute the $L$-series numerically. Using $100000$ coefficients of $f_E$, we have the Fourier expansion $\mathbb{L}_{f_E}(z) = \sum_{n = 1}^\infty D_{f_E}( h; 1)q^h$:
\begin{equation}\label{e:11L} \mathbb{L}_{f_E}(z) = -0.7063\dots q - 1.562\dots q^2 - 0.0944\dots q^3 - 1.237\dots q^4 + 2.026\dots q^5 + O(q^6) \end{equation}
The space of weight $2$ quasimodular Eisenstein series for $\Gamma_0(11)$ is $2$ dimensional, and a basis is given by $F_{11,2}^0$ and $F_{11,2}^\infty$, where $F_{11,2}^\infty$ is
\begin{align*}
F_{11,2}^\infty &= 1 + \frac15q + \frac35 q^2 + \frac45 q^3 + \frac75 q^4 + \frac65q^5 + \frac{12}{5}q^6 + O(q^7)
\end{align*}
Then, if we take $\alpha = .00159$, we obtain another way to retreive the $L$-series values in Equation~\eqref{e:11L}:
\begin{equation*} \begin{split}
\frac{\vol(\Lambda_E)}{\pi}\left((f_E \cdot \widehat{\mathfrak{Z}}_E^+) - \alpha f_E - \sum \beta_{\rho(i)} F^{\rho(i)}_{11,2} \right) &=
\\ -0.706\dots q - 1.562\dots q^2 - 0.0930\dots q^3 &- 1.234\dots q^4 + 2.024\dots q^5 + O(q^6)
\end{split}\end{equation*}
\subsection{Conductor $N = 27$}
Recall the strong Weil curve of conductor $27$ given by the Weierstrass equation $E_{27}: y^2 + y = x^3 - 7$ (Cremona label 27a1).
The weight $2$ modular form associated with $E_{27}$ is given by
\[f_{E_{27}} = q - 2q^{4} - q^{7} + 5q^{13} + 4q^{16} - 7q^{19} + O(q^{20}) .\]
Using the Fourier expansion of the Weierstrass $\zeta$-function, the weight $0$ mock modular form associated to $\widehat{\mathfrak{Z}}_{E_{27}}^+(z)$ is given by
\begin{equation} \label{e:zed27} \widehat{\mathfrak{Z}}_{E_{27}}^+(z) = q^{-1} + \frac{1}{2}q^2 + \frac{1}{5}q^5 + \frac{3}{4}q^8 - \frac{6}{11}q^{11} - \frac{1}{2}q^{14} + O(q^{17}) .\end{equation}
The holomorphic projection is given by
\[\widehat{\pi}_{\text{hol}}(f_E \cdot \widehat{\mathfrak{Z}}_{E_{27}}^+)(z) = 1 + 3q^9+ 9q^{18} - 12q^{27} \dots. \]
On the other hand, we find that the normalized element of the weight $2$ quasimodular forms for $\Gamma_0(27)$ that vanishes at all cusps but $\infty$ is given by
\[F_{27,2}^\infty = 1 + 3q^9+ 9q^{18} - 12q^{27} \dots \]
This agrees with the computation for the holomorphic projection.
Then the generating function $\mathbb{L}_f(z)$ can also be computed by Theorem~\ref{t:thm1.2}, giving arbitrary precision computations for these slowly convergent shifted convolution $L$-series values that could only previously be computed term-by-term.
\bibliographystyle{abbrv}
|
\section{Introduction}
\label{sec:introduction}
\textbf{Motivation:} Reusing intermediates in databases to speed-up analytical query processing has been studied in the past ~\cite{monetdb2009,cod2001,vvreuse_13,qpipe2005,qpipe2006,mrshare,mrmqo}.
These solutions typically require intermediate results of individual operators be materialized into temporary tables to be considered for reuse in subsequent queries.
However, these approaches are fundamentally ill-suited for use in modern main memory databases.
The reason is that modern main memory DBMSs are typically limited by the bandwidth of the memory bus and query execution is thus heavily optimized to keep tuples in the CPU caches and registers ~\cite{hyper2011,DBLP:reference/db/Ross09,llvm}.
To that end, adding additional materialization operations into a query plan not only add additional traffic to the memory bus but more importantly prevent the important cache- and register-locality, which results in high performance penalties.
Consequently, the benefits of materia\-lization-based reuse techniques heavily depends on the characteristics of the workload: i.e., how much overlap between queries of a given workload exists.
In the worst case, if the overlap is low the extra cost caused by materialization operations might even result in an overall performance degradation for analytical workloads.
The goal of this paper is to revisit "reuse" in the context of modern main memory databases~\cite{hyper2011,hana2013,tupleware2015}.
The main idea is to leverage internal data structures for reuse that are materialized anyway by pipeline breakers during query execution.
This way, reuse comes for free without any additional materialization costs.
Moreover, as a result reuse becomes more robust towards workloads with different reuse potentials and provides benefits for a wider range of workloads even if the overlap between queries is not that high.
In this paper, we present our system called \emph{HashStash}{} that implements reuse of internal data structures.
The focus of this work is on the most common internal data structure, hash tables (HTs), as found in hash-join and hash-aggregate operations.
We leave other operators and data structures (e.g., trees for sorting) for future work.\\
\vspace{-1.5ex}
\textbf{Contributions:} To the best of our knowledge this is the first paper that studies the reuse of internal data structures for query processing.
As a major contribution, we present a new system called \emph{HashStash}{} that extends a classical DBMS architecture to support the reuse of internal hash tables.
The architecture of \emph{HashStash}{} supports two reuse models:
\emph{(1) Single-query Reuse}: In this re-use model, users or applications submit a single query to a \emph{HashStash}{}-based DBMS just as in normal DBMSs.
However, different from a classical DBMS, a \emph{HashStash}{}-based DBMS identifies the best {\em reuse-aware plan} that leverages existing intermediate hash tables.
To support this model, we extend the DBMS architecture by three components: (a) {\em a cache for hash tables} that keeps lineage and statistics information, (b) {\em a reuse-aware optimizer} that uses new operator cost models and enumeration strategies to determine which hash tables should be reused by which operators in order to minimize the total query runtime, and finally (c) {\em a garbage collector} that evicts hash tables from the cache as needed using eviction strategies.
\emph{(2) Multi-query Reuse}: Many analytical applications today execute multiple queries at the same time to show different aspects over the same data set.
In order to support multiple queries that are submitted at the same time, we leverage the concept of shared plans as introduced in \cite{sharedb2014, sharedb2012} and extend them in the following directions:
(a) we develop {\em shared reuse-aware plans}, i.e., shared plans can also re-use the hash tables in~\emph{HashStash}{} and
(b) we extend the optimizer in~\emph{HashStash}{} to create optimal reuse-aware shared plans for a given batch of queries.
Finally, we evaluate the performance of the \emph{HashStash}{} system under workloads with different reuse potentials.
Our experiments show that \emph{HashStash}{} outperforms materia\-lization-based reuse for any of these workloads independent of the reuse potential.\\
\vspace{-1.5ex}
\textbf{Outline:} The rest of this paper is structured as follows.
Section \ref{sec:arch} gives an overview of our suggested \emph{HashStash}{}-based architecture to support single-query and multi-query reuse.
Section \ref{sec:single} and Section \ref{sec:multi} then present the details for each of these reuse cases and discuss novel optimization strategies to support them.
Afterwards, Section \ref{sec:garbage} discusses how garbage collection works in \emph{HashStash}{}.
Section \ref{sec:eval} presents our evaluation of our \emph{HashStash}{} prototype.
Finally, Section \ref{sec:related} discusses related work and Section \ref{sec:concl} concludes with a summary and outlines potential future work.
\section{HashStash Overview}
\label{sec:arch}
The main goal of \emph{HashStash}{} is to leverage internal hash tables for reuse that are materialized during query execution.
To achieve this, in \emph{HashStash}{} we add the following components to a classical DBMS architecture (see Figure \ref{fig:arch}):
(1) a \emph{Reuse-aware Query Optimizer} (RQO) that replaces the traditional (non-reuse-aware) optimizer, and
(2) a \emph{Hash Table Manager} (HTM) that consists of a cache of hash tables and a garbage collector.
In the following, we discuss each component individually and then present an example to illustrate the main ideas of \emph{HashStash}{}.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{figures/architecture.eps}
\caption{Additional \emph{HashStash}{} Components}
\vspace{-3.5ex}
\label{fig:arch}
\end{figure}
\subsection{Reuse-Aware Query Optimizer}
The Reuse-Aware Query Optimizer (RQO) offers two interfaces for compiling and optimizing queries: a query-at-a-time interface for single-query reuse and a query-batch interface to support multi-query reuse. \\
\vspace{-1.5ex}
\textbf{Query-at-a-time Interface:} This interface accepts a single query, optimizes this query, and returns a reuse-aware execution plan.
The main goal of the reuse-aware optimizer is to decide which hash tables in the cache to reuse such that the overall query execution time is minimized.
In order to select the best reuse-aware execution plan, the reuse-aware \emph{HashStash}{} optimizer enumerates different join orders and decides for each plan which is the best reuse case based on the hash tables in the cache.
Different from a traditional query optimizer, our reuse-aware optimizer therefore implements two important extensions:
(1) In order to decide which hash table to reuse, the optimizer leverages so called \emph{reuse-aware cost models}.
Different from normal cost models, reuse-aware cost models additionally take statistics of a candidate hash table into account in order to estimate the execution costs for the different reuse cases discussed before.
(2) The reuse-aware optimizer implements benefit-oriented optimizations.
The main intuition is that a plan is preferred over another if it creates hash tables that promise more benefits for future reuse even if the initial execution is a ``little'' more expensive.
Furthermore, the \emph{HashStash}{} optimizer supports four different cases for reuse-aware operators: \emph{exact}-, \emph{subsuming}-, \emph{partial}-, and \emph{overlapping}-reuse.
This is different from the existing approaches in \cite{monetdb2009,cod2001,vvreuse_13}, which only support the \emph{exact}-reuse, and the \emph{subsuming}-reuse cases.
The \emph{exact} case enables a join or aggregation operator to reuse a cached hash table which contains exactly the tuples required by the query.
In that case complete sub-plans might be eliminated (e.g., the one which build the hash table of a hash-join).
Compared to the case before, the \emph{subsumption} case is possible when the reused hash table contains more tuples than needed. This might lead to false-positives, which need to be post-filtered by an additional selection.
The \emph{overlapping} and the \emph{partial} case are different.
Both cases allow the reuse of a hash table where some tuples are ``missing''. These tuples thus need to be added during query execution.
To support all these different reuse cases the optimizer also applies different rewrites rules during optimization.
Figure \ref{fig:arch} shows how the reuse-aware optimizer is integrated into the \emph{HashStash}{} architecture.
First, the optimizer enumerates different join orders and retrieves candidate hash tables for reuse.
Once the "optimal" reuse-aware execution-plan is found, the optimizer sends the information of which hash tables will be reused to the HTM for book-keeping as shown in step 2 (Figure \ref{fig:arch}).
Finally, the optimizer sends the reuse-aware execution plan to the executor as shown in step 3 (Figure \ref{fig:arch}).
Once the plan execution is finished, the DBMS runtime informs the HTM to release all used reused hash tables as shown in step 4 (Figure \ref{fig:arch}), which make them available for garbage collection for instance.
Details about the query-at-a-time interface are described in Section \ref{sec:single}. \\
\vspace{-1.5ex}
\textbf{Query-Batch Interface:} The query-batch interface is different from the query-at-a-time interface since it accepts multiple queries submitted as a batch.
Different from the query-at-a-time interface, subsets of queries submitted in the same batch can share the same execution plan; called \emph{reuse-aware shared plan} in the sequel.
Different from the approach presented in \cite{sharedb2014}, the main contribution of \emph{HashStash}{} is that it integrates the before-mentioned reuse techniques into shared plans.
In order to find the best reuse-aware shared plan, we developed a novel reuse-aware multi-query optimizer that merges individual reuse-aware plans using a dynamic programming based approach.
Details about the query-batch interface are described in Section \ref{sec:multi}.
\subsection{Hash Table Manager}
The two components of the Hash Table Manager (HTM) are the hash table cache and the garbage collector.\\
\vspace{-1.5ex}
\textbf{Hash Table Cache:}
The hash table cache manages hash tables for reuse; it stores pointers to cached hash tables, as well as lineage information about how each one of them was created.
It also stores statistics to enable the cost-based hash table selection by the optimizer.
For our initial prototype, we allow only one query to reuse a hash-table in the cache at a time (except for the query-batch interface).
However, for future work, we plan to look into sharing the same hash table between concurrent queries. \\
\vspace{-1.5ex}
\textbf{Garbage Collector:}
The main goal of the garbage collector is to decide which hash tables to evict as necessary.
The garbage collector is triggered by the cache if no more memory is available to admit new hash tables.
Therefore, the garbage collector maintains usage information and implements eviction strategies to decide which hash tables can be removed.
\subsection{Reuse Example}
Figure \ref{fig:example} illustrates a reuse example for a sequence of three queries from a data exploration session.
The initial query $Q1$ executes an aggregation over a 3-way join of the tables \texttt{Customer}, \texttt{Orders}, and \texttt{Lineitem} for all lineitems shipped after \texttt{2015-02-01}.
For this query no reuse is possible (since it is the initial one).
However it materializes all three hash tables HT1-HT3 in the cache of \emph{HashStash}{}.
The follow-up query $Q2$ then executes a query that differs from $Q1$ only in the the filter predicate; i.e., it selects lineitems that shipped after \texttt{2015-01-01}.
In order to execute $Q2$, \emph{HashStash}{} can reuse hash table $HT2$ (exact-reuse) and thus avoids to recompute the join of \texttt{Customer} and \texttt{Orders}.
Moreover, the hash table $HT3$ produced by the aggregation of $Q1$ can also be reused to compute the aggregation operator in $Q2$ well. However, since $HT3$ does not aggregate all required lineitems (due to partial-reuse), the base table \texttt{Lineitem} needs to be re-scanned for the ``missing'' tuples between \texttt{2015-01-01} and \texttt{2015-02-01}. These tuples are added to $HT3$ by the reuse-aware plan of $Q2$.
Finally, the last query $Q3$ is similar to $Q2$.
The only difference is that it removes the group-by attribute \texttt{c.age}.
For executing $Q3$, \emph{HashStash}{} can directly reuse the hash table $HT3$ (exact-reuse).
However, due to the removed group-by attribute a post-aggregation operator needs to be added.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figures/example.eps}
\caption{Reuse Example}
\vspace{-3.5ex}
\label{fig:example}
\end{figure}
\section{Single-query Reuse}
\label{sec:single}
In this section, we describe how to find the best reuse-aware execution plan for the query-at-a time interface.
As discussed before, finding the best reuse-aware plan is implemented by the optimizer of \emph{HashStash}{}.
Therefore, we first give an overview of how the plan enumeration procedure of our optimizer works and then discuss the cost models of our reuse-aware hash-join and hash-aggregate operator.
Afterwards, we present the details on how the matching and rewriting procedures in \emph{HashStash}{} work to enable the different reuse-cases (exact, subsuming, partial, and overlapping).
Finally, we discuss some benefit-oriented optimizations that increase the effect of reuse by spending initially a little more execution cost to create "better" hash tables.
\subsection{Reuse-aware Plan Enumeration}
\label{sec:single:pe}
The plan enumeration algorithm in~\emph{HashStash}{} can be applied to complex nested SQL queries.
In order to simplify the presentation, we first show the basic procedure that only enumerates different join plans for a given SPJ (select-project-join) query .
More complex queries including aggregations and nesting are discussed at the end of this section. \\
\vspace{-1.5ex}
\textbf{Basic Procedure:} Algorithm \ref{alg:planenum} shows the basic recursive procedure for enumerating different reuse-aware plans for SPJ queries based on a top-down partitioning search strategy.
Compared to existing top-down partitioning search strategies such as \cite{topdown2007}, Algorithm \ref{alg:planenum} additionally implements the following ideas to support reuse-aware plans:
(1) when partitioning the join graph into a left and right partition of $G$, the algorithm enumerates the different candidate hash tables (including a new empty hash table) for the right partition $G_r$ and the left partition $G_l$ that can be reused for building the hash table of a hash join (line 8 to 16 and 19 to 27).
(2) Another difference from existing top down enumeration algorithms is to rewrite the respective sub-plan that would reuse a given candidate hash table (line 9 and 20).
This rewrite possibly eliminates the complete sub-plan (i.e., in the best case $G'_r$ (line 9) and $G'_l$ (line 20) might become an identity operation over the reused hash table if an exact-reuse is possible).
We discuss details of the rewrite procedure for all four different reuse cases (exact-, subsuming-, overlapping-, and partial-reuse) later in this section.
(3) The last difference is that the cost estimation (line 13 and 24) uses the cost models for the reuse-aware join and aggregation operator to estimate the runtime costs when reusing a given candidate hash table.
We also discuss the details of these cost models later in this section.
\begin{algorithm}
\scriptsize
\SetAlgoLined
\SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output}
\SetKwFunction{optimize}{getBestReusePlan}
\SetKwFunction{genPlan}{generatePlans}
\SetKwFunction{buildTree}{buildTree}
\SetKwFunction{getGraph}{getJoinGraph}
\SetKwFunction{createTree}{createTree}
\SetKwFunction{rewriteTree}{rewriteTree}
\SetKwFunction{checkHash}{getCandHTs}
\SetKwFunction{createHashSpec}{createHTSpec}
\SetKwFunction{newHash}{createNewHT}
\SetKwFunction{cost}{cost}
\SetKwData{Left}{left}
\Input{Join Graph $G$ of SPJ Query $Q$}
\Output{Reuse-Aware Execution Plan $P$}
\BlankLine
\textbf{Algorithm} {\optimize{$G$}}\\
{
\eIf{$bestPlans[G] \neq NULL$}{
\KwRet{$bestPlans[G]$}\;
}
{
\ForEach{$partition$ $(G_l,G_r)$ in $G$}{
$candHTs \leftarrow \checkHash(G_r)$\;
$candHTs \leftarrow candHTs \cup \newHash(G_r)$\;
\BlankLine
\ForEach{$candHT \in candHTs$}{
$G_r' \leftarrow \rewriteTree(G_r,candHT)$\;
$P_l \leftarrow \optimize(G_l)$\;
$P_r' \leftarrow \optimize(G_r')$\;
$curTree \leftarrow \createTree(P_l, P_r', candHT)$\;
\If{$\cost(curTree) \leq \cost(bestPlans[G])$}{
$bestPlans[G] = curTree$\;
}
}
\BlankLine
$candHTs \leftarrow \checkHash(G_l)$\;
$candHTs \leftarrow candHTs \cup \newHash(G_l)$\;
\BlankLine
\ForEach{$candHT \in candHTs$}{
$G_l' \leftarrow \rewriteTree(G_l,candHT)$\;
$P_l' \leftarrow \optimize(G_l')$\;
$P_r \leftarrow \optimize(G_r)$\;
$curTree \leftarrow \createTree(P_r, P_l', candHT)$\;
\If{$\cost(curTree) \leq \cost(bestPlans[G])$}{
$bestPlans[G] = curTree$\;
}
}
\BlankLine
}
}
}
\caption{Plan Enumeration in \emph{HashStash}{}}
\label{alg:planenum}
\end{algorithm}
\textbf{Complex Queries:}
The general idea to support more complex queries is similar to exiting optimizers.
First, nested queries are unnested using joins if possible.
Second, if unnesting is not possible for the complete query, the enumeration procedure shown before is applied to each query block individually. In \emph{HashStash}{}, query blocks can be in the form of either SPJ (select-project-join) or SPJA (select-project-join-aggregation) queries.
In order to extend Algorithm \ref{alg:planenum} for SPJA queries, we only need to iterate over all candidate hash tables as well as an empty (new) hash table for the additional aggregation operator in the SPJA block.
After selecting the most optimal hash table for reuse, we then need to apply the rewrite rules for the aggregation operator once the procedure shown in Algorithm \ref{alg:planenum} returns its result.
\subsection{Reuse-Aware Operators and Cost Models}
\label{sec:single:rop}
In the following, we discuss the reuse-aware operators (join and aggregation) and the cost models to select hash tables for reuse.
\vspace{-0.75ex}
\subsubsection{Reuse-Aware Hash-Join}
\label{sec:single:rhj}
A reuse-aware hash-join (RHJ) works similarly to a traditional hash-join; i.e., the join first builds a hash table from one of its inputs and then probes into the hash table using each tuple of the other input.
However, an RHJ has two major differences:
(1) in the build phase, the RHJ operator might need to add the ``missing'' tuples into the reused hash table (for overlapping- and partial-reuse), and
(2) in the probe phase, the RHJ operator might need to post-filter false-positives (for overlapping- and subsuming-reuse); i.e., tuples that are stored in a reused hash table but not required to execute the current join operator.
Running the join without post-filtering would produce false-positives during the probing phase.
For each candidate hash table $HT$ that can be reused to compute a given join, the optimizer in \emph{HashStash}{} needs to estimate the total runtime costs.
In the following, we explain the details of our cost model.\\
\vspace{-1.5ex}
\textbf{Cost Model:} The main components that determine the cost of an RHJ are the resize cost $c_{resize}$, the build cost $c_{build}$ and the probe cost $c_{probe}$.
\vspace{-1.5ex}
\begin{scriptsize}
\begin{equation*}
c_{RHJ} = c_{resize}(HT) + c_{build}(HT) + c_{probe}(HT)
\end{equation*}
\end{scriptsize}
\vspace{-1.5ex}
Our cost model for the RHJ explicitly considers the cost for resizing the hash table, $c_{resize}$.
In order to minimize the cost of resizing in \emph{HashStash}{}, we use a hash table that implements extendible hashing using linked lists for collision handling.
Thus, instead of re-hashing all entries, only the bucket array needs to get resized and entries can be assigned to the new buckets lazily.
The costs for building and probing of an RHJ are different from a traditional hash-join and depend additionally on two parameters:
(1) the contribution-ratio $contr$ and
(2) the overhead-ratio $overh$ of a candidate hash table $HT$.
The first parameter, the contribution-ratio $contr$, defines how much of the data in the candidate hash table $HT$ already contributes to the operator if that operator reuses this hash-table; i.e., this data does not need to be added to the hash table anymore and makes the build phase faster.
For example, if $contr(HT)=0.5$ then only $50\%$ of missing tuples need to be added to the hash table $HT$ during the build phase.
The second parameter, the overhead-ratio $overh$, defines how much unnecessary data is stored in the hash table; i.e., this data contributes to the total memory footprint of the hash table and makes the building and probing phases slower since the hash table might spill out of the CPU caches.
The overhead-ratio also determines the additional cost needed to post-filter false positives.
For example, if $overh(HT)=0.7$ then $70\%$ of tuples in the hash table are not required by the RHJ.
In the sequel, we discuss how to use both parameters ($contr$ and $overh$) to model all of the four reuse-cases (exact, subsuming, partial, and overlapping).
In the following equations, we show how \emph{HashStash}{} estimates the costs of the build phase $c_{build}(HT)$ and the probe phase $c_{probe}(HT)$ of an RHJ using these two parameters.
\vspace{-1.5ex}
\begin{scriptsize}
\begin{equation*}
\begin{split}
c_{build}(HT) &= \underbrace{|Builder| \cdot (1-contr(HT))}_\text{\#tuples to insert} \cdot \underbrace{ c_i(htSize,tWidth)}_\text{cost of a single insert} \\
c_{probe}(HT) &= \underbrace{|Prober|}_\text{\#tuples to probe}\cdot \underbrace{c_r(htSize,tWidth)}_\text{cost of a single lookup}
\end{split}
\end{equation*}
\end{scriptsize}
\vspace{-1.5ex}
The build cost $c_{build}(HT)$ is determined by the number of missing tuples that need to be inserted times the cost of a single insertion $c_i$ into the resized hash table.
The probe cost $c_{probe}(HT)$ is determined by the number of tuples that need to probe into the hash table times the lookup cost $c_l$ for a single probe into the hash table.
The cost of a single insert/lookup $c_i$ and $c_l$ depend on two parameters:
(1) the memory footprint of the resized hash table $htSize$ (shown in the following equation) and
(2) the width of a tuple $tWidth$ stored in the cached hash table $HT$.
While the memory footprint $htSize$ determines if a hash table fits into the CPU caches or not and thus influences the insert/lookup costs, the tuple width $tWidth$ determines the number of I/O operations required to transfer a tuple between main memory and CPU caches.
Since the hash table might contain more attributes than needed by the query (e.g., for post-filtering), the tuple-width $tWidth$ might actually be bigger as for a hash table that we create individually for this query.
\begin{figure*}
\begin{minipage}[b]{.25\textwidth}
\centering
\includegraphics[width=\textwidth,trim=5mm 0mm 4mm 0mm]{experiments/cost_model/singleinsert.eps}
\vspace{-3.5ex}\subcaption{Cost of a single insert}
\label{fig:single_insert}
\end{minipage}
\hfill
\begin{minipage}[b]{.25\textwidth}
\centering
\includegraphics[width=\textwidth,trim=5mm 0mm 4mm 0mm]{experiments/cost_model/singleprobe.eps}
\vspace{-3.5ex}\subcaption{Cost of a single probe}
\label{fig:single_probe}
\end{minipage}
\hfill
\begin{minipage}[b]{.25\textwidth}
\centering
\includegraphics[width=\textwidth,trim=5mm 0mm 4mm 0mm]{experiments/cost_model/singleupdate.eps}
\vspace{-3.5ex}\subcaption{Cost of a single update}
\label{fig:single_update}
\end{minipage}
\hfill
\\
\vspace{-1.5ex}
\caption{Reuse-aware Cost Parameters}
\vspace{-3.5ex}
\label{fig:cost_model_param}
\end{figure*}
Moreover, for estimating the build cost $c_{build}$ and the probe cost $c_{probe}$, we need to be able to estimate the cost of a single insert/lookup ($c_i$ and $c_l$).
However, these costs depend on the specific hash table implementations and other hardware-dependent parameters; e.g., how prefetching into CPU caches is implemented.
Therefore, these costs need to be determined by a set of micro benchmarks which calibrate the cost model for varying hash table sizes and tuple widths.
We implemented such a micro benchmark using C++ and used GCC 4.9.2 as compiler.
Figure \ref{fig:single_insert} and Figure \ref{fig:single_probe} show the results of our micro benchmark for the hash table implementation used in \emph{HashStash}{} on a machine with an Intel Xeon E5-2660 v2 processor using $128$GB RAM running Ubuntu Server 14.04.1.
The cache sizes of the processor are: $32$KB data + $32$KB instruction L1 cache (private), $256$KB L2 cache (private) and $25$MB L3 cache (shared).
For both, the insert and lookup operations, we can clearly see the effect of different hash table sizes ($1KB$ to $1GB$) and cache boundaries on the insertion/lookup costs.
The effect of the tuple-width ($8B$ to $256B$) is also visible but needs some more explanation.
For insertion, the cost does not change as long as a tuple fits into one cache line, which is $64$B in our processor.
Once the tuple-width exceeds the cache line size, the cost increases as shown for $128$B and $256$B in Figure \ref{fig:single_insert}.
For lookup, the behavior is slightly different: due to the prefetching of one cache line by the CPU, the cost to lookup one tuple increases only when the tuple-width exceeds $128$B.
\vspace{3.5ex}
\subsubsection{Reuse-Aware Hash-Aggregation}
\label{sec:single:rha}
Similar to the reuse-aware hash-join (RHJ), the reuse-aware hash-aggregate (RHA) can reuse an existing cached hash table.
Similar as for the RHJ, the RHA might also need to add ``missing'' tuples (for overlapping- and partial-reuse) and post-filter tuples (for overlapping- and subsuming-reuse).
In the following, we discuss the cost model, that estimates the runtime cost of an RHA for a given hash table.
\\
\vspace{-1.5ex}
\textbf{Cost Model:}
For a given candidate hash table, the optimizer estimates the total runtime costs of a reuse-aware hash aggregate as shown by the following equation:
\vspace{-1.5ex}
\begin{scriptsize}
\begin{equation*}
c_{RHA} = c_{resize}(HT) + c_{insert}(HT) + c_{update}(HT)
\end{equation*}
\end{scriptsize}
\vspace{-1.5ex}
The cost of an RHA consists of three components:
(1) the resize cost $c_{resize}$,
(2) the cost $c_{insert}$ to insert the initial tuple for each distinct missing group-by key, and
(3) the update costs $c_{update}$ of the aggregated value for the other input tuples.
For example, assume an RHA has $100$ missing input tuples with $10$ distinct missing group-by keys.
In that case, the RHA needs to pay $10$ times the insert cost and $90$ times the update cost to reuse the given hash table.
Similar to the RHJ, the contribution-ratio $contr$ and the overhead-ratio $overh$ have an influence on the insert/update costs.
In the following, we take a closer look into defining different cost components for a given candidate hash table $HT$.
The cost component $c_{resize}$ represents the cost to resize the hash table for the distinct missing group-by keys.
Again, these costs are dependent on the implementation details of the hash table.
For RHA, we use the same hash table implementation as for the RHJ operator (i.e., we use the same cost estimates for resizing).
The idea to estimate the other two cost components is that the insertion cost only need to be paid for the input tuples that represent the missing distinct group-by keys (i.e., the first tuple that is inserted for a missing key). All other tuples need to pay only the update cost.
The functions to estimate the insertion cost $c_{insert}$ and the update cost $c_{update}$ are shown in the following equations.
\vspace{-1.5ex}
\begin{scriptsize}
\begin{equation*}
\begin{split}
c_{insert}(HT) &= \underbrace{|distinct(Input.key)| \cdot (1-contr(HT))}_\text{\#tuples to insert} \cdot \\
& \underbrace{ c_i(htSize,tWidth)}_\text{cost of a single insert}
\end{split}
\end{equation*}
\end{scriptsize}
\begin{scriptsize}
\begin{equation*}
\begin{split}
c_{update}(HT) &= \underbrace{(|Input|-|distinct(Input.key)|) \cdot (1-contr(HT))}_\text{\#tuples to update} \cdot \\
&\underbrace{c_u(htSize,tWidth)}_\text{cost of a single update}
\end{split}
\end{equation*}
\end{scriptsize}
\vspace{-1.5ex}
The equations above need an estimate for the insert/update cost ($c_i$, $c_u$) for a input single tuple.
Same as for the RHJ, these costs depend on the total size of the hash table $htSize$ and the width of the tuples $tWidth$ stored in the candidate hash table $HT$ as well as the other hardware-dependent factors (i.e., the CPU cache sizes).
Different from the RHJ, however, the size of the hash table $htSize$ only depends on the number of distinct keys in the input since only one aggregated tuple per distinct key need to be stored.
In order to calibrate our costs for our hash-table implementation and the underlying hardware, we again executed a set of micro-benchmarks.
The insertion costs $c_i$ are the same as the ones shown for the RHJ operator in Section \ref{sec:single:rhj}.
Figure \ref{fig:single_update} shows the results for the update costs $c_u$ (single-threaded) for the same setup (CPU, hash table implementation and compiler) used for the RHJ operator.
\subsection{Matching and Rewriting}
\label{sec:single:rw}
The goal of the matching procedure is to find those candidate hash tables in the hash table cache which qualify for reuse for a given operator $r$ in the enumerated plan.
As discussed before, in \emph{HashStash}{} we support reuse for hash-join and hash-aggregations only.
We use the following notation: $C$ (cached plan) represents the plan that produced a \emph{cached} hash table and $R$ (requesting plan) represents the plan rooted by $r$.
For a hash-join $r$, $R$ represents the sub-plan below the join input that builds the hash-table (including the join itself but excluding the probing branch).
For a hash-aggregate $r$, the sub-plan $R$ represents the operator tree below the aggregation operator including the aggregation itself.
For finding a matching hash table that can be reused for $r$, our hash table manager stores lineage information in a similar way as described in \cite{vvreuse_13} using a so called the recycle graph $G_C$.
The graph $G_C$ merges the lineage of all hash tables in one graph $G_c$.
Figure \ref{fig:recycler} shows an example of a recycler graph that resulted from merging two query plans; the first query that produces $HT1$ contains an aggregation operator over the \texttt{Customer} table and the second query plan contains a join operator that builds the hash table $HT2$ over the \texttt{Customer} table on the join key \texttt{cid}.
Different from the recycler graph in \cite{vvreuse_13}, for each node $n_c \in G_C$ we additionally store if it refers to a cached hash table or not (since not all operators materialize a hash table).
The problem of matching is now to find a subtree in $G_C$ that matches a given plan $R$ rooted by the requesting operator $r$.
Matching $R$ to the graph $G_C$ builds on the notion of bisimilarity; i.e., all nodes $n_c$ in the graph $G_C$ are compared to the root node $r$ of $R$.
In our case, we can prune the search space and only need to compare $r$ to those nodes $n_c$ that actually refer to a cached hash-table (i.e., all operators that represent joins and aggregates in $G_c$).
The matching procedure used in \cite{vvreuse_13} to check if a node $n_c$ exactly matches $r$ is: (i) $n_c$ and $r$ represent the same type of operation (e.g., both are selections); (ii) the parameters of the two nodes are equal (e.g., they evaluate the same selection predicate); (iii) they have the same number of children and there is an exact match for all children of $n_c$ and $r$.
This procedure, however, tests only for the exact-reuse case.
\cite{vvreuse_13} also discusses an extension to test for subsuming-reuse and not only for exact-reuse.
In the following, we explain, how we extend the above matching procedure to check for the all four reuse-cases in the given order and discuss which rewrites need to be applied in each reuse case.
\\
\vspace{-1.5ex}
\textbf{Exact-reuse:} The matching procedure detects an exact-reuse of the hash-table referred by $n_c$, if the predicates of all selection operators in $R$ match exactly one predicate in the query tree rooted by $n_c$; i.e., for each predicate $r_i$ in $R$ there exists a predicate $c_i$ in the query tree rooted by $n_c$ for which $r_i = c_j$ holds.
In that case $R$ can be replaced by the operator over the cached hash table that is referenced by $n_c$.
Figure \ref{fig:recycler} (top right) shows an example where the selection predicate $\sigma_{c.age \geq 30}$ of $R$ matches the sub-plan rooted by the aggregation operator in the recycler graph $G_c$.
The rewrite rule thus replaces the original plan $R$ directly by $HT1$.
In case that the root node $r$ of $R$ represents a hash-aggregation, we additionally allow that the group-by attributes $G_r$ of $r$ are only a subset of the group-by attributes $G_c$ of $n_c$ if all aggregation functions are additive (e.g., sum, count, min, max).
In that case, $R$ can be replaced by a sub-plan that consists of the aggregation operator $r$ over the reused hash-table.
Figure \ref{fig:example} (right hand side) already has shown an example for this reuse case.\\
\vspace{-1.5ex}
\textbf{Subsuming-reuse:} We test if there exists a plan $C$ in $G_c$ where $r_i \subseteq c_j$ holds for all selection predicates $r_i$ in the plan rooted by $r$.
In that case $R$ can be replaced by a selection operator $\sigma_{post}$ over the reused hash table.
The predicate $post$ represents the conjunction of all predicated $r_i$ of $R$.
If the hash table does not contain the attributes needed to test $post$, it does not qualify for reuse.
Figure \ref{fig:recycler} (bottom left) shows this case. Since the hash table $HT2$ contains all customers for $age \geq 20$ and the join $r$ only requires only customer with $age \geq 30$, all false positives must be post filtered after probing using the filter predicate $\sigma_{age \geq 30}$.
\\
\vspace{-1.5ex}
\textbf{Partial-reuse:} In order to support this reuse case, we test if there exists a plan $C$ in $G_c$ where $c_j \subseteq r_i$ holds for all selection predicates $r_i$ in the plan rooted by $r$.
This means, that the reused hash table does not contain all necessary tuples.
To the end, $R$ is rewritten to a plan $R'$ which adds the missing tuples from the base tables to the reused hash table by replacing $r_i$ in $R$ by $r_i \wedge !c_j$.
Figure \ref{fig:recycler} (bottom right) shows an example where $HT2$ can be partially reused; i.e., all customers $20 \geq age < 30$ must be added to $HT2$ from the base table \texttt{Customer}.
\\
\vspace{-1.5ex}
\textbf{Overlapping-reuse:} We test if there exists a plan $C$ in $G_c$ where $c_j \wedge r_i \ne \emptyset$ holds for all selection predicates $r_i$ in the plan rooted by $r$.
For rewrite, we apply both rewrites that we have discussed for the partial-reuse and the subsuming-reuse case before.
\begin{figure}
\centering
\includegraphics[width=0.30\textwidth]{figures/recycler.eps}
\vspace{1.5ex}
\caption{Match and Rewrite}
\vspace{-3.5ex}
\label{fig:recycler}
\end{figure}
\subsection{Benefit-oriented Optimizations}
\label{sec:single:benefit}
\emph{HashStash}{} additionally implements the following benefit-oriented optimizations.
The main intuition behind these optimizations is that one plan $A$ is preferred over another plan $B$ if the plan $A$ creates hash tables that promise higher benefits for future reuse even if the plan execution is a ``little'' more expensive then for the optimal plan.\\
\vspace{-1.5ex}
\textbf{Aggregate Rewrite:} AVG is always rewritten to SUM and COUNT to support the partial- and overlapping-reuse at the cost of initially creating a slightly bigger hash-table. \\
\vspace{-1.5ex}
\textbf{Additional Attributes:} For join hash tables, all attributes used in selection operations in the sub-plan of the input which build the hash-table are added to the cache as well to enable post-filtering of false positives to increase the reuse potential of a hash table. \\
\vspace{-1.5ex}
\textbf{Join Order:} Typically hash tables are always built over the smaller join input. However, if the hash table is reused in future it might be also beneficial to build the hash table over the bigger input.
We therefore integrated a simple heuristic approach into our optimizer that is similar to the techniques presented in \cite{vvreuse_13} to determine which intermediate will provide more benefit for future queries based on the history of queries executed.\\
\section{Multi-query Reuse}
\label{sec:multi}
In this section, we describe the techniques in \emph{HashStash}{} that enable shared plans to reuse cached hash tables.
We call these plans \emph{reuse-aware shared plans}.
In the following, we first discuss the details of reuse-aware shared plans.
Afterwards, we present how we extend our optimizer in \emph{HashStash}{} to find an optimal reuse-aware shared plan for a given query-batch and a set of cached hash tables.
\subsection{Reuse-Aware Shared Plans}
The basic idea of shared plans is shown in Figure \ref{fig:shared}.
Instead of compiling each query into a separate plan, multiple queries are compiled into one shared plan that reuses hash tables.
The idea of shared plans has been presented in \cite{sharedb2014} already.
In \emph{HashStash}{}, we extend shared plans to allow them to reuse cached hash tables.
In the following, we first reiterate over the idea of shared plans and then discuss the relevant modifications for our reuse-aware operators to work correctly in shared plans.
Different from a normal plan, in a shared plan individual operators execute the logic of multiple queries.
The most common shared operator is the shar\-ed scan operator that evaluates the filter predicates of multiple queries in one scan.
In order to keep track of which tuples qualify for which query, shared operators in \cite{sharedb2014} use a Data-Query Model where each tuple is tagged by the IDs of those queries it qualifies for.
For example, if a tuple produced by a shared scan satisfies the predicates of query $Q_1$ and $Q_3$ but not of query $Q_2$, this tuple will be tagged using $Q_1$ and $Q_3$ (or $101$ if a bitlist is used to represent query IDs).
Moreover, other operators such as joins and aggregation operators can be shared as well.
Figure \ref{fig:shared} shows an example of a shared plan where the selection operators and the hash-join are shared by three queries ($Q_1$ to $Q_3$) while the aggregation is not shared (i.e., there exists one separate operator for each query).
For the hash-join, we see that tuples tagged with query IDs (qids) are stored in its hash table.
The query IDs are used in the probing phase to produce the output of the join.
In the following, we describe our extensions for the reuse-aware hash-join and hash-aggregate such that they can execute multiple queries at a time.\\
\vspace{-1.5ex}
\textbf{Shared Reuse-Aware Hash-Joins:}
In general, the shar\-ed reuse-aware hash-join (SRHJ) operator works similarly to the non-shared reuse-aware hash-join (RHJ) presented in Section \ref{sec:single:rhj}:
Instead of recomputing the hash table in the build phase from scratch, a cached hash table is reused to avoid re-computation.
However, there are some important differences between an SRHJ that has to support query-batches and a non-shared RHJ that only supports a single query.
First, the SRHJ can only reuse hash tables that include query IDs for tagging (as shown in Figure \ref{fig:shared}).
A hash table that does not include query IDs can not be reused for a shared operator.
Second, before the SRHJ operator starts to execute it has to re-tag all tuples stored in the reused hash table using the predicates of current query-batch.
Otherwise, if it does not re-tag all tuples in the reused hash table, these tuples will be tagged with obsolete query IDs from a previous (non-active) query-batch, which might lead to wrong query results if query IDs are recycled.
To that end, re-tagging represents an overhead that has to be considered in the cost model of an SRHJ.\\
\vspace{-1.5ex}
\textbf{Shared Reuse-Aware Hash-Aggregates:}
Shared aggregates are different from normal aggregation operators since they split the execution into two phases:
a first phase that groups the input tuples by keys and a subsequent aggregation phase.
While the grouping phase is shared for all queries, the subsequent aggregation phase is carried out for each query separately (i.e., the output of the grouping phase is split based on query IDs).
In this paper, we focus on shared hash-aggregates that store the output of the grouping phase in a hash table before applying the aggregation functions on the individual tuples stored in the hash-table.
The goal of a shared reuse-aware hash-aggregate (SRHA) is to reuse hash tables to avoid the re-computation of the grouping phase.
This is very different than reusing hash tables for a non-shared RHA operator since hash tables of an SRHA store individual tuples and not aggregates.
Another difference is that the SRHA operator also needs to re-tag all the tuples stored in the reused hash table (just as for the SRHJ operator) before the operator is executed.
Both these aspects (i.e., storing individual tuples and the need for re-tagging) influence the cost of an SRHA and must be included in the cost model.
Finally, SRHA and the RHA operators also differ in how they select candidate hash tables from the cache.
While an RHA must find hash tables with the same aggregation functions, an SRHA is more flexible since it can recompute any arbitrary aggregate function on the grouped data.
For example, a hash table which was built for an SRHA operator that computes one aggregation function (e.g., SUM) can be reused by another SRHA operator, which computes a different aggregation function (e.g., MIN).
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figures/sharedplan.eps}
\caption{Shared Plans using a Data-Query Model}
\vspace{-4.5ex}
\label{fig:shared}
\end{figure}
\subsection{Plan Enumeration}
In the following, we discuss the plan enumeration implemented in \emph{HashStash}{} to support query-batches.
The goal of the optimizer is to find a set of reuse-aware shared plans $\{S_1, S_2,$ $\dots, S_n\}$ for a given query-batch $\{Q_1,$ $Q_2, \dots, Q_m\}$ with $n \leq m$ that minimizes the total runtime to execute all queries in the given batch by reusing cached hash tables.
In order to find the optimal set of reuse-aware shared plans $\{S_1, S_2,$ $\dots, S_n\}$, \emph{HashStash}{} uses a {\em dynamic programming approach} to merge query plans incrementally into reuse-aware shared plans.
Figure \ref{fig:sharedplanenum} shows the dynamic programming process for three queries (e.g., such as those in Figure \ref{fig:shared}).
Each node in the dynamic programming graph in \ref{fig:sharedplanenum} represents a so-called merge configuration that describes which queries should be merged together into a shared reuse-aware plan and which should be executed using a separate non-shared reuse-aware plan.
In terms of notation, $\{Q_1,Q_{2+3}\}$ represents a merge configuration, which defines that two separate plans should be generated: one non-shared reuse-aware plan for query $Q_1$ and one shared reuse-aware plan for queries $Q_2$ and $Q_3$.
In \emph{HashStash}{}, it depends on two aspects if two queries are merged or not.
First, two queries are merged if the total runtime of the shared plan is less than the sum of executing two individual plans.
Second, not all queries are mergeable.
In order to guarantee a correct plan execution, two queries $Q_1$ and $Q_2$ can only be merged if they have the same join graph.
Otherwise, these queries cannot be merged and the plans must be kept separate.
If two queries $Q_1$ and $Q_2$ are mergeable, the result of merging in \emph{HashStash}{} is a shared reuse-aware plan where (1) all join operations are shared (i.e, SRHJ operators are used for joins) and (2) all aggregation operators that use the same group-by keys are shared (i.e, SRHA operators are used for aggregations).
In order to find the merge configuration that results in the minimal total runtime (i.e., the total sum over all plans), \emph{HashStash}{} starts the dynamic programming process with merge configurations of size $1$ (called level $1$).
On level $2$, the optimizer then continues to find the merge configurations for all possible combinations of two queries which has the minimal total runtime by extending the merge configurations from the level below until the process reaches level $m$.
For example, in order to compute the merge configuration on level $3$ in Figure \ref{fig:sharedplanenum}, the dynamic programming process merges query $Q_3$ into the merge configuration $\{Q_1,Q_2\}$ of level $2$ amongst the other possible combinations (e.g., merging $Q_2$ into $\{Q_1,Q_3\}$ or merging $Q_1$ into $\{Q_{2+3}\}$.
In order to merge $Q_3$ with the merge configuration $\{Q_1,Q_2\}$, the dynamic programming process enumerates all three possible merge configurations $\{Q_{1+3},Q_2\}$, $\{Q_{1},Q_{2+3}\}$, and $\{Q_1,Q_2, Q_3\}$ and keeps only the one with the minimal total runtime.
Moreover, in order to avoid to analyze the same merge configuration twice, \emph{HashStash}{} memoizes merge configurations and their estimated runtime.
Finally, to estimate the total runtime of a merge configuration, the optimizer computes the optimal reuse-aware (shared) plan associated with each entry of the given merge configuration.
In order to find the best reuse-aware (shared) plan associated with entry in a merge configuration, the optimizer applies a variant of the enumeration process presented in Section \ref{sec:single:pe} that supports query graphs and not only query trees.
For example, given the merge configuration $\{Q_{1+3},Q_2\}$ the optimizer applies the enumeration procedure to find the best (shared) plan separately for $Q_{1+3}$ and $Q_2$.
In order to find the best reuse-aware shared plans (e.g., for $Q_{1+3}$), the the plan enumeration in Section \ref{sec:single:pe} uses reuse-aware shared operators (i.e., SRHJ and SRHA) instead of using non-shared reuse-aware operators.
\begin{figure}
\centering
\includegraphics[width=0.25\textwidth]{figures/merging.eps}
\vspace{1.5ex}
\caption{Dynamic Programming based Plan Merging}
\vspace{-4.5ex}
\label{fig:sharedplanenum}
\end{figure}
\section{Garbage Collection}
\label{sec:garbage}
In this section, we provide the details of how garbage collection is implemented in \emph{HashStash}{}.
The main goal of garbage collection is to evict hash tables from the cache that are most likely not to be reused by other queries in future.
In Section \ref{sec:arch}, we already described that the hash table manager monitors the hash table cache and starts an eviction process, whenever the total memory footprint of the cached hash tables exceeds a threshold (i.e., no more memory is available to store new hash tables).
To decide on which hash tables to discard is the crucial part of the eviction process.
Different from the eviction process used in database buffers, the garbage collection in \emph{HashStash}{} does not work on the granularity of pages.
Instead it can either work in a coarse-grained mode on the granularity of complete hash tables or in a more fine-grained mode on the granularity of individual hash table entries.
While a coarse-grained mode needs less storage space for book keeping and requires less monitoring overhead than a fine-grained mode, it tends to keep ``old'' entries in a hash table even if other entries in the hash table are only used.
Moreover, evicting individual entries from a hash table in a fine-grained mode requires a scan of individual bucket of the hash table.
Finally, in a fine-grained mode, concurrent access of the eviction process and queries to the same hash table need to be synchronized.
In \emph{HashStash}{}, we have implemented this fine-grained mode.
However, initial results showed that this mode results in a high additional load that reduces the efficiency of \emph{HashStash}{}.
Therefore, we have decided to integrate only a \emph{least recently used (LRU)} policy that evicts complete hash tables instead of evicting individual entries of hash tables (i.e., garbage collection is working in a coarse-grained mode) in \emph{HashStash}{}.
In order to implement the LRU policy, the Garbage Collector of \emph{HashStash}{} keeps a \emph{timestamp} of the last access for each hash table in its usage information.
Based on this timestamp, the eviction process picks the hash table with the oldest timestamp and evicts it from the cache.
The garbage collection process repeats the eviction until the memory footprint drops below the memory threshold.
In our experiments we see that this policy is able to efficiently deal with different workloads where queries build on recent results. Moreover, the coarse-grained mode introduces only a minimal overhead for book keeping and for executing the eviction process.
\section{Experimental Evaluation}
\label{sec:eval}
In this section, we report the results of our experimental evaluation of the techniques presented in this paper.
The main goal of this evaluation is to:
(1) compare the efficiency of reusing internal hash-tables in \emph{HashStash}{} to materialization-based reuse,
(2) present the performance gains for both interfaces: the query-at-a-time and the query-batch interface,
(3) show the efficiency and the accuracy of our optimizer and the cost models,
(4) analyze the overhead of applying garbage collection in \emph{HashStash}{}.
In the following, we explain the details of the experimental setup that are commonly used for all experiments.\\
\vspace{-1.5ex}
\textbf{Workload and Data:} In order to analyze the efficiency of different re-use strategies we are using three different types of analytical workloads with (1) low-, (2) medium-, and (3) high-reuse potential.
Each of the workload consists of $64$ different queries over the TPC-H database schema.
For the workload with the low-reuse potential, the average overlap of common data read from base tables by two subsequent queries is $1\%$.
This simulates the fact that users often look at different parts of a data set.
For the medium-reuse case, the overlap is $10\%$ and $50\%$ for the high-reuse case.
The idea is that the spatial locality increases in these workloads to simulate users; i.e., in the medium- and high-reuse case users typically explore data in a common region by several queries before changing focus to other parts of the data.
The queries in each workload have the following characteristics: The initial query in each workload is TPC-H query $Q3$ that has a three-way join over the tables \texttt{Lineitem}, \texttt{Orders}, and \texttt{Customer} with an aggregation operator on top.
We used this query since it represents a medium-complex query with three join and one aggregation operator that initially creates three hash tables / intermediate results for potential reuse.
All other queries in our workloads resulted from applying different modifications to simulate different user interactions that are commonly used operations in analytical frontends such as Tableau \cite{tableau2012} or Spotfire \cite{spotfire}.
The user interactions simulated by different queries are: zooming-in/-out, shifting as well as drill-down and roll-up operations.
While zooming-in/-out and shifting only change the selection predicate of the previous query, drill-down and roll-up add/remove TPC-H tables that use a join and add/remove group-by attributes respectively.
The resulting queries are all SPJ or SPJA queries.
By using \emph{Drill-Down} operations, we introduce new joins into the workload that include \texttt{Part} and \texttt{Supplier} tables to achieve more complex queries (i.e., to form five-way join queries).
Finally, as the main data set in all our experiments, we use a TPC-H database of $SF=10$ with secondary indexes on all selection attributes used in our query workloads.
We do not use other scaling factors since the relative performance gains of \emph{HashStash}{} compared to materialization-based reuse is the same.
However, for some of our experiments that contain micro benchmarks, we use synthetic data sets (e.g., to show the effects of our cost models).
We describe these synthetic data sets further in the corresponding section.
\\
\vspace{-1.5ex}
\textbf{Implementation and Hardware:}
We implemented our \emph{HashStash}{} prototype using C++ and GCC 4.9.2 as the compiler.
For execution in our \emph{HashStash}{} prototype, we generate C++ execution plans for all SQL queries following the ideas described in \cite{hyper2011}.
In order to show the pure effects of reuse only, our prototype system implements a single-threaded execution model.
For running all experiments, we used one machine with an Intel Xeon E5-2660 v2 processor and $128$GB RAM running Ubuntu Server 14.04.1.
The cache sizes of the processor are: $32$KB data + $32$KB instruction L1 cache, $256$KB L2 cache and $25$MB L3 cache.
\begin{figure}
\begin{minipage}[b]{.22\textwidth}
\centering
\includegraphics[width=\textwidth, height=0.6\textwidth, trim=5mm 0mm 4mm 0mm]{experiments/enumeration/fig7_a.eps}
\vspace{-3.5ex}
\subcaption{Speed-up of Runtime}
\label{fig:exp:runtime}
\end{minipage}
\hfill
\begin{minipage}[b]{.23\textwidth}
\centering
\scriptsize
\resizebox{\columnwidth}{!}{
\begin{tabular}{|l l |c|c|}
\hline
& & & \\
\multicolumn{2}{|l|}{\bf Workload / Stats} & {\bf Mat. } & {\bf HashStash} \\
& & & \\\hline
\multirow{ 3}{*}{\bf Low} & Mem Size & $1052$ MB & $1160$ MB \\
& Hit Ratio & $0$ & $0.01$ \\
& Time &
164.22 (sec) &
151.66 (sec) \\\hline
\multirow{ 3}{*}{\bf Med} & Mem Size & $858$ MB & $643$ MB \\
& Hit Ratio & $1.27$ & $1.45$ \\
& Time &
117.58 (sec) &
98.85 (sec)
\\\hline
\multirow{ 3}{*}{\bf High} & Mem Size & $817$ MB & $520$ MB \\
& Hit Ratio & $1.37$ & $1.58$ \\
& Time &
103.04 (sec) &
79.99 (sec) \\\hline
\end{tabular}
}
\subcaption{Workload Statistics}
\label{fig:exp:stats}
\end{minipage}
\vspace{0.5ex}
\caption{Single-Query Reuse}
\vspace{-4.5ex}
\label{fig:exp:single}
\end{figure}
\subsection{Exp. 1: Single-Query Reuse}
\label{sec:eval:exp0}
In this experiment, we analyze the benefits of \emph{HashStash}{} for the single-query interface using the different workloads mentioned before.
In order to show the efficiency of \emph{HashStash}{}, we first executed each workload using the no-reuse strategy, which does not recycle any intermediates but also has no cost for materialization.
Afterwards, we executed the two reuse strategies: (1) materialization-based reuse where intermediate results are spilled out to a temporary table in memory, and
(2) \emph{HashStash}{} which reuses internal hash-tables.
We implemented all approaches (1) and (2) as well as the no-reuse strategy in \emph{HashStash}{}.
In order to compare (1) with our reuse strategy (2), we materialize the same intermediates as \emph{HashStash}{} does (i.e., the input of joins for which we build a hash table and outputs of aggregations).
Moreover, as described in \cite{vvreuse_13} (1) only supports exact and subsuming-reuse but neither partial nor overlapping-reuse.
The results of this experiment are shown in Figure \ref{fig:exp:single}, where we first present the resulting total runtime with some statistics (e.g., cache sizes, hit ratio) for executing the different workloads.
In this experiment, we turned the garbage collection (GC) off.
The effects of GC are analyzed in Section \ref{sec:eval:exp4}.
Figure \ref{fig:exp:runtime} shows the overall speed-up of both reuse strategies over the no-reuse strategy when running under different workloads.
We see that our strategy in \emph{HashStash}{} shows the highest speed-up for all workloads (low-, medium-, and high-reuse).
For the workload with high reuse potential \emph{HashStash}{} achieves a speed-up of $90\%$ over the no-reuse strategy, while the materialization-based reuse strategy only achieves $40\%$.
For the workload with low-reuse potential which simulates a user randomly browsing the data, \emph{HashStash}{} has a performance comparable to the no-reuse strategy; i.e., it does not introduce additional overhead even if there is (almost) no reuse potential.
This is different for the materialized-reuse which introduces a slow down of $10\%$ caused by the additional materialization costs.
Figure \ref{fig:exp:stats} shows additional statistics for each workload: memory footprint, hit ratio, and the total runtime.
For the materialized-reuse strategy, we report the memory footprint for all temporary tables as well as the hit ratio per temporary table (i.e., how often a temporary table was reused).
For \emph{HashStash}{}, we report the footprint for all cached intermediate hash-tables tables as well as the hit ratio per hash table.
The hit ratio is given as the average ratio of how often each element in the cache was re-used by a query.
For the medium- and high-reuse case, we see that \emph{HashStash}{} requires less memory in total while providing a higher hit ratio per cached element than the materialized-reuse strategy.
The main reason for this is that the materialized-reuse strategy only supports two out of four reuse-cases supported by \emph{HashStash}{}.
To that end, less intermediates are reused and more new ones are added to the cache.
For the low-reuse case, we see that the hit ratio of the cache is almost $0$ in both strategies.
In this case, the memory footprint is the highest since queries just register new elements to the cache without actually reusing them.
Moreover, the memory footprint of \emph{HashStash}{} is slightly higher than the materialized-reuse case.
The reason is that hash tables have an additional overhead (e.g., pointers for linked lists of extends) when compared to a temporary table which is essentially an array in memory without any overhead.
However, it is interesting to note that this does not have an effect on the runtime of \emph{HashStash}{} since caching the internal hash tables does not cause any additional memory I/O compared to the no-reuse strategy.
This is different from the materialized-reuse strategy, which requires additional I/O to persist the output of operators to the memory in order to support reuse.
\subsection{Exp. 2: Efficiency of Query Optimizer}
\label{sec:eval:exp1}
In this experiment, we show the benefits of our reuse-aware optimizer.
We therefore study the runtime of
(a) reuse on the query-level as well as
(b) reuse on the operator-level (i.e., for reuse-aware joins and aggregations).
The main goal is to compare the performance of our cost-model based strategy with two baselines: the first baseline is \emph{never-share}, where we turn reuse in our system completely off. The second baseline is \emph{always-share}, where all operators use a greedy-heuristic to reuse the matching hash table in the cache with the highest reuse ratio. We include this strategy in order to show that greedily reusing hash tables can result in a performance that is even worse than the performance of the \textit{Never Share} strategy and to emphasize the need for a cost model that decides whether to reuse a hash table or not.\\
\vspace{-1.5ex}
\textbf{Exp. 2a - Reuse on the Query-level:}
In this experiment, we selected a subset of seven queries from the workload with high-reuse potential.
We selected these queries, since each query represents a different type of user interaction and thus provides different reuse potentials for join and aggregation operators.
We selected the high-reuse case in order to show that the always-share baseline might result in non-optimal plans and showing that our cost-based approach finds better reuse-aware plans.
The first query of the sequence we picked is a 5-way SPJA query over the tables \texttt{Lineitem}, \texttt{Orders}, \texttt{Part}, \texttt{Customer}, and \texttt{Supplier}.
The details of the six follow-up queries are summarized in Table \ref{exp:table:states}.
The first column of this table lists the type of user interaction that was applied .
The second column shows the difference of each query to its predecessor:
The first four follow-up queries modify the selection predicate on the attribute \texttt{o\_orderdate}.
The last two queries modify the group-by keys.
\begin{figure}
\begin{minipage}[b]{.23\textwidth}
\centering
\includegraphics[width=\textwidth, height=0.6\textwidth, trim=5mm 0mm 4mm 0mm]{experiments/enumeration/fig8_a.eps}
\vspace{-3.5ex}
\subcaption{Runtime}
\label{fig:exp:states}
\end{minipage}
\hfill
\begin{minipage}[b]{.23\textwidth}
\centering
\scriptsize
\resizebox{\columnwidth}{!}{
\begin{tabular}{|l|c|c|}
\hline
{\bf Mutation} & {{\bf Update}} & {\bf \begin{tabular}[c]{@{}c@{}}Selected Reuse Scheme\\ (O, P, C, S, Agg)\end{tabular}} \\ \hline
{\bf Zoom In} & \begin{tabular}[c]{@{}c@{}}1996-06-01\\ 1996-09-01\end{tabular} & NSSSN \\ \hline
{\bf Zoom Out} & \begin{tabular}[c]{@{}c@{}}1992-01-01 \\ 1998-01-01\end{tabular} & NSSSS \\ \hline
{\bf Shift Much} & \begin{tabular}[c]{@{}c@{}}1996-09-01\\ 1998-01-01\end{tabular} & NSSSN \\ \hline
{\bf Shift Less} & \begin{tabular}[c]{@{}c@{}}1994-01-01\\ 1998-01-01\end{tabular} & NSSSS \\ \hline
{\bf Drill-down} & \begin{tabular}[c]{@{}c@{}}add\\ p\_brand\end{tabular} & SNSSN \\ \hline
{\bf Roll-up} & \begin{tabular}[c]{@{}c@{}}remove\\ p\_mfgr\end{tabular} & XXXXS \\ \hline
\end{tabular}
}
\subcaption{User Interactions}
\label{exp:table:states}
\end{minipage}
\\
\vspace{-1.5ex}
\caption{Reuse on the Query-Level}
\vspace{-6.5ex}
\end{figure}
For running this experiment, we executed all seven queries sequentially over the TPC-H database using our reuse strategy as well as using the two baselines (never-share and always-share).
The first query populates the cache with five hash tables in total: four resulting from the joins and one from the aggregation.
The results for the six follow-up queries (that are candidates for reuse) are shown in Figure \ref{fig:exp:states}.
In this figure, we see that the \emph{Cost Model} strategy, which is based on our optimizer, outperforms the two other baselines since it always picks the optimal reuse strategy.
In the best case (i.e., the \emph{RollUp} follow-up query), the speed up factor is about two orders of magnitude better than \emph{never-share}.
The reason is that the cached aggregation hash table is sufficient to execute the \emph{RollUp} query (i.e., no missing tuples need to be added and thus no joins need to be executed at all).
For the \emph{Drill Down} query, we could not execute the \textit{Always Share} strategy since the \emph{p\_brand} attribute was never included in the corresponding hash table in previous executions and thus that hash table is not reusable.
The last column of Table \ref{exp:table:states} shows the detailed decisions of our optimizer (i.e., for the \textit{Cost Model} strategy) for all operators of the six follow-up queries, which explain our performance results in Figure \ref{fig:exp:states}.
The string in this column uses one character to encode the decision for each operator (join and aggregation).
The operators from left to right are shown in the header of the last column:
For example, the $O$ character represents the hash table created by the build phase of a join that scanned the \texttt{Orders} relation.
The other characters represent the hash tables created by the build phase that scanned the \texttt{Part}, \texttt{Customer}, and \texttt{Supplier} tables.
\emph{Agg} represents the aggregation operator that is executed on top of all joins.
The characters encode the following decision:
$N\ (Not\ Shared)$ states that a new hash table was created for the operator whereas $S\ (Shared)$ states that the existing hash table was reused.
Moreover, $X$ defines that this operator was not need to be executed at all for the given query.
For instance, this case occurs for the \textit{Roll Up} operation, where the new query just needs to read the cached aggregation hash table.
\\
\vspace{-1.5ex}
\textbf{Exp. 2b - Reuse on the Operator-level (RHJ):} In this experiment, we show the efficiency of our optimizer for the reuse-aware hash-join (RHJ).
For showing the efficiency, we directly execute the reuse-aware hash-join operators on two input synthetic tables.
The table for the building phase was $16MB$ in size and the table for the probing phase had $10\times$ the size of the table for the build phase.
In order to show the efficiency of our optimizer for RHJ, after adding a candidate hash table of $16MB$ to the hash table cache we executed multiple runs with different contribution ratios from $100\%$ to $0\%$.
$100\%$ contribution-ratio means that the RHJ can reuse all tuples in the cached hash table and does not need to post-filter any tuples after probing; whereas $0\%$ contribution-ratio means that the RHJ can not reuse any tuples in the cached hash table.
Moreover, $0\%$ contribution-ratio means there is $100\%$ overhead in the reused hash table (i.e., all tuples must be post-filtered) due to the fact that for all contribution-ratios we keep the size of the cached hash-table the same.
Same as in the previous experiment, here we compare our \textit{Cost Model} based strategy against the \textit{Never Share} (i.e., a traditional hash-join) and the \textit{Always Share} strategy which always picks the cached hash table for reuse.
Figure \ref{fig:exp:join_16_16} shows the results.
We see that the \textit{Never Share} strategy pays a constant price since it never reuses the hash table.
Moreover, the costs for the \text{Always Share} strategy are constantly increasing since more and more missing tuples need to be added to the reused hash table (if the contribution-ratio decreases).
At approx. $70\%$ contribution-ratio, the \text{Always Share} gets more expensive than the \textit{Never Share} strategy due to the overhead incurred in the cached hash table (i.e., tuples in the hash table that are not required by the RHJ).
As an important result, we see that our \text{Cost Model} always picks the best strategy with the minimal cost: for a contribution-ratio from $100\%$ to $70\%$ it reuses the cached hash table and below $70\%$ it decides to create a new hash table since the total runtime costs are cheaper when not reusing the candidate hash table in the cache.\\
\vspace{-1.5ex}
\textbf{Exp. 2c - Reuse on the Operator-level (RHA):}
In this experiment, we show the effect of reusing hash tables for reuse-aware hash-aggregates (RHAs).
We again varied the contribution-ratio of the cached hash table as in the experiment before.
Figure \ref{fig:exp:agg} shows that our cost model still picks the best strategy with the minimal cost.
\begin{figure}
\begin{minipage}[b]{.23\textwidth}
\centering
\includegraphics[width=\textwidth, trim=5mm 0mm 4mm 0mm]{experiments/single_operator/join/16_16.eps}
\vspace{-2.5ex}
\subcaption{ Reuse-Aware Hash-Join }
\label{fig:exp:join_16_16}
\end{minipage}
\hfill
\begin{minipage}[b]{.23\textwidth}
\centering
\includegraphics[width=\textwidth, trim=5mm 0mm 4mm 0mm]{experiments/single_operator/aggregation/agg_case1.eps}
\vspace{-2.5ex}
\subcaption{ Reuse-Aware Hash-Agg. }
\label{fig:exp:agg}
\end{minipage}
\\
\vspace{-1.5ex}
\caption{Reuse on the Operator-Level}
\label{fig:join_exp}
\vspace{-4.5ex}
\end{figure}
\subsection{Exp. 3: Accuracy of Query Optimizer}
\label{sec:eval:exp2}
\begin{figure}
\centering
\includegraphics[width=.4\textwidth]{experiments/enumeration/plan_enum.eps}
\vspace{0.5ex}
\caption{Accuracy of Cost Models}
\vspace{-6.5ex}
\label{fig:exp:enum}
\end{figure}
As described in Section \ref{sec:single:pe}, the plan enumeration algorithm is one of the core elements of \emph{HashStash}{} to select a reuse-aware plan with minimal runtime for a given set of cached hash tables.
In this experiment, we validate the accuracy of the cost estimation component of our optimzer (i.e., the \texttt{cost} function used in Algorithm \ref{alg:planenum}).
For this experiment, we execute the workload described in Section \ref{sec:eval:exp0} with medium-reuse potential.
In order to analyze the accuracy of our cost estimation, we select one of the $5$-way join queries over the tables \texttt{Lineitem}, \texttt{Orders}, \texttt{Part}, \texttt{Customer}, and \texttt{Supplier} from this workload and analyze the estimated and actual cost of the optimizer.
We selected this query since it is a complex query with multiple joins and the optimal reuse-aware plan contains both cases: operators that reuse a cached hash table and other operators that create a new hash table.
In order to analyze the accuracy of our cost estimates, we compare the estimated and the actual cost for each enumerated sub-plan of this query.
Figure \ref{fig:exp:enum} shows the results.
As a general observation, we can see that our cost models are accurate since the actual and estimated costs follow the same trend.
To better understand the results, we clustered the costs into groups that represent equivalent sub-plans (i.e., one group represents sub-plans over the same partition of the join graph).
For example, the group \texttt{CO} represents the enumerated join plans over the two tables \texttt{Customer} and \texttt{Orders} for all hash tables in the cache.
Moreover, to better compare the actual and estimated costs, we are using normalized costs (called time units).
For normalization, the lowest cost per group uses the cost of $1$.
As discussed in Section \ref{sec:single:pe}, plan enumeration works incrementally and picks the cheapest sub-plan per group and composes the complete plan based on these optimal sub-plans.
Thus, the quality of the optimizer depends only on the fact that whether or not it finds the cheapest plan per group.
In order to see this, if the optimizer finds the cheapest plan per group, the normalized costs are sufficient (i.e., the absolute costs do not matter for this decision).
Figure \ref{fig:exp:enum} orders the sub-plans per group by their actual costs.
As shown in Figure \ref{fig:exp:enum}, the estimated costs follow the same trend as the actual costs.
Even more important is that the first plan per group, which has the lowest actual cost resulting from the ordering, always has the lowest estimated cost as well.
To that end, our optimizer is able the find the most optimal sub-plan per group for the query.
\subsection{Exp. 4: Multi-Query Reuse}
\label{sec:eval:exp3}
In this section, we present the evaluation results for the query-batch interface as explained in Section \ref{sec:multi}.
In order to generate the batches of queries, we group the query trace of $64$ queries of the workload with medium-reuse potential of the experiment in Section \ref{sec:eval:exp0} into smaller sets of $4$, $8$, and $16$ queries.
In order to populate the cache, we first executed one batch of the given size (e.g., $4$ queries) and afterwards executed $10$ randomly selected batches of the same size and report the average run time for one batch as a result.
Moreover, in order to show the effect of reuse in shared plans, we executed the same sequence of batches using different modes:
The first mode (\emph{single-query plan, wo reuse}) is the traditional database execution mode where queries are executed individually that do not reuse any cached hash tables.
The second mode (\emph{single-query plan, w reuse}) executes all queries individually as well, but in a set up where reuse of intermediate hash tables using our cost-model is enabled.
The last mode (\emph{shared plan, w reuse}) represents the mode where we use our reuse-aware shared plans as introduced in Section \ref{sec:multi} to execute the batch.
Figure \ref{fig:exp:batch} shows the total runtime summed up for all queries in a batch of a given size.
As expected, the first mode (\emph{single-query plan, wo reuse}) has the highest total runtime for all batch sizes.
The second mode (\emph{single-query plan, w reuse}) reduces the total runtime on average by approx. $20\%$ to run all queries in the batch, which purely results from reusing intermediates.
Finally, the execution of the (\emph{shared plan, w reuse}) mode results in the lowest total runtime, which is on average approximately $40\%$ lower than for the first mode (\emph{single-query plan, wo reuse}).
The additional reduction in runtime compared to the \emph{single-query plan without reuse} mode result from the effect of shared reuse-aware plans. Our plan enumeration algorithm of the query-batch interface in \emph{HashStash}{} creates on average, $2$ shared reuse-aware plan to execute the batch of size $4$, $3$ plans for the batch of size $8$, and $4$ plans for the batch of size $16$.
\begin{figure}
\centering
\includegraphics[width=.3\textwidth, trim=4mm 0mm 5mm 0mm]{experiments/batches/batch.eps}
\caption{Batch Execution in \emph{HashStash}{}}
\vspace{-4.5ex}
\label{fig:exp:batch}
\end{figure}
\subsection{Exp. 5: Effects of Garbage Collection}
\label{sec:eval:exp4}
In this experiment we show the effect of garbage collection on the performance of \emph{HashStash}{}.
We again used the workloads as described in Section \ref{sec:eval:exp0} and executed the complete trace using two modes:
The first mode (\emph{wo GC}) represents the case where we execute all queries using the query-at-a-time interface with reuse, however, no garbage collector was active; i.e., \emph{HashStash}{} used as much memory as needed to cache all hash tables.
For the second mode (\emph{with GC}), we additionally activated the garbage collector (GC).
For the hash table cache, we used $20\%$ of the memory that would be required to store all hash tables.
As a result, we measured the additional runtime overhead that was caused by the effects of the garbage collector (i.e., monitoring the size of all caches hash tables, evicting and reloading evicted hash tables).
Compared to \emph{HashStash}{} without GC, our experiment shows that \emph{HashStash}{} with GC introduces approximately only a $10\%$ higher overhead for the medium- and high-reuse case.
For the high-reuse case this is negligible when looking at the performance gains of \emph{HashStash}{} over a DBMS without any reuse (as we have shown in the experiments before).
For the medium-reuse case, \emph{HashStash}{} can still achieve a performance speed-up of $10\%$ over the no-reuse case.
Note, however, that when increasing the cache size to $50\%$ of the total memory required to cache all hash tables, the overhead of garbage collection drops down to $5\%$.
Most interestingly, for the low-reuse workload, GC causes almost no overhead since intermediate hash-tables are anyway almost never reused.
\section{Related Work}
\label{sec:related}
\textbf{Reuse of Intermediates:}
In order to better support user sessions in DBMSs, various techniques have been developed in the past to reuse intermediates \cite{cod2001,monetdb2009,vvreuse_13}.
All these techniques typically require that results of individual operators are materialized into temporary tables.
This is very different from \emph{HashStash}{}, which revisits "reuse" in the context of modern main memory DBMSs and promotes to leverage internal hash tables that are materialized by pipeline breakers and thus does not add any additional materialization cost to query execution.
In the following, we discuss further differences when comparing these techniques to the ideas of \emph{HashStash}{}.
\cite{cod2001} introduces an optimizer to select which intermediates should be reused.
Different from \emph{HashStash}{}, the cost models are rather coarse-grained and centered around the I/O benefits in disk-based DBMS. To that end, their cost models do not take the peculiarities of hash tables as well as hardware-dependent parameters such CPU caches into account.
In \cite{monetdb2009}, the authors integrate reuse techniques into MonetDB, that implements an operator-at-a-time execution model which anyway relies on full materialization of all intermediate results and thus does not need to tackle the issues that result form additional materialization cost as in pipelined databases.
\cite{vvreuse_13} extends the work of \cite{monetdb2009} for pipelined databases and integrates the ideas into Vectorwise.
In this paper, the authors introduce a cache with lineage which is similar to the ideas of the hash table manager in \emph{HashStash}{}.
A major difference is, however, that in both cases intermediate results of operators are reused and not internal data structures of operators as we suggest in \emph{HashStash}{}.
Moreover, compared to all the approaches mentioned before \cite{cod2001,monetdb2009,vvreuse_13}, our work also supports reuse-cases for partial- and overlapping reuse and most importantly introduces a reuse-aware optimizer.
Another area where reuse of intermediates was analyzed is in the context of Hadoop.
ReStore \cite{redoop2014} is able to reuse the output of whole MapReduce jobs that are part of a workflow implemented in PigLatin.
Moreover, it additionally materializes the output individual operators that are executed within a MapReduce job.
Since Restore is based on Hadoop and not tailored towards reuse in main memory systems, it makes their reuse techniques fundamentally different from those presented in \emph{HashStash}{}.
Finally, buffer pools and query caches in database systems \cite{querycache1994,semcache1996} serve as a cache for frequently accessed data.
However, the main purpose of buffer pools and query caches is to speed-up the access to base data (in case of the database buffer) or the final query result (in case of query caches) but not to reuse intermediates.\\
\vspace{-1.5ex}
\textbf{Materialized Views:}
Reusing results has also been the main motivation of materialized views \cite{matview2001}.
Again, the main difference of materialized views and \emph{HashStash}{} is that reuse is for materialized views is externalized as an additional "table" rather than leveraging internal data structures that are produced by query processing.
Moreover, our reuse-aware optimizer implements cost-models that target the reuse of internal data structures and introduces benefit-oriented optimizations, both aspects that are not covered by traditional optimizers that rewrite queries for materialized views.
\\
\vspace{-1.5ex}
\textbf{Automatic Physical Schema Design:}
Another line related to our work are techniques for online physical schema tuning \cite{online_tuning07}.
The main goal of this work is to create additional database objects such as indexes or materialized views (discussed before) without involving a database administrator.
Adaptive indexing \cite{adaptive_idx11,partial_idx89} also falls into this category and suggests to create indexes partially as a side effect of query processing.
However, again these techniques do not consider internal data structures for reuse but externalize their decision by creating additional (partial) indexes, views, etc.
\\
\vspace{-1.5ex}
\textbf{Multi-Query-Optimization:} Another area of related work is Multi-Query-Optimization (MQO) \cite{mqo}.
The main idea of MQO is to identify common sub-expressions of a set of queries that are active in a DBMS at the same time.
In order to save resources, common sub-expressions are only executed once.
One problem of MQO is that in most workloads, common sub-expressions are a rather rare case.
Therefore, MQO is typically used to optimize OLAP workloads over a star schema where the chance of common sub-expressions is higher since most queries join the dimension tables with the same fact table.
All ideas in MQO are orthogonal to the reuse ideas presented in this paper; i.e., reuse of hash tables can be integrated into plans created by MQO techniques.
\\
\vspace{-1.5ex}
\textbf{Work-Sharing:} Work-sharing systems \cite{shredscanibm2008,sharedmonet2007,crescando2009,qpipe2006,cjoin2009,sharedb2014} have similar goals as MQO since they also process multiple queries at a time by sharing work.
However, different from MQO they do not require to identify the very same sub-expression to share work.
One of the techniques for work-sharing is the shared (or cooperative) scan operator \cite{shredscanibm2008,crescando2009,sharedmonet2007}.
The idea of shared scans is that the scan operation can be shared by queries even if queries use different selection predicates.
Other systems such as QPipe \cite{qpipe2006}, CJoin \cite{cjoin2009}, SharedDB \cite{sharedb2014} extend the idea of work-sharing to other operators such as joins and aggregations.
All these ideas for work-sharing are again orthogonal to the reuse ideas presented in this paper.
In this paper, we actually extended the ideas of \cite{sharedb2014} to integrate reuse into shared-plans.
\section{Conclusions and Future Work}
\label{sec:concl}
Salient characteristics of modern main memory DBMSs and interactive analytical workloads require a critical rethinking of reuse in query processing. Our solution, called \emph{HashStash}{}, focuses on the reuse of hash tables populated with intermediate query results. We avoid additional materialization costs by leveraging hash tables that are already materialized at pipeline breakers. We also do not incur the overhead of casting hash tables to relations and vice versa by treating hash tables as native units of reuse. Our reuse-aware optimizer can accurately model hash table usage and its impact in query performance, leading to highly profitable reuse choices that offer up to $100\times$ performance improvement over the no-reuse baseline for realistic workloads.
We plan to extend the ideas presented in this paper to other data structures (e.g., trees) as well as concurrent reuse.
|
\section{Introduction}
\label{sect:introduction}
The celebrated sensitivity of cluster abundance to the growth rate of structure means that the mass density of the Universe, $\Omega_m$, and the amplitude of the power spectrum, $\sigma_8$ should be particularly accurately derived from cluster surveys, \citep{Peebles1989, Oukbir1992, Eke1996, Viana1996, Oukbir1997, Carlberg1997, Bahcall1997, Fan1997, Henry1997, Voit2005, Brodwin2007, Mantz2010, Rozo2010, Clerc2012, Benson2013, Huterer2015} providing a welcomed consistency check of the current cosmology. Even the presence of a few massive clusters at $z>0.5$ has been enough to overcome the long desired $\Omega_m=1$ consensus \citep{Bahcall1997}, favouring a sub-critical mean mass density for the Universe \citep{Bahcall1998}. Empirically, $\Omega_m\simeq 0.2$--$0.3$ has long been argued by simply extrapolating galaxy mass-to-light ratios (M/L) to large scales \citep{Ostriker1974} and clarified with dynamical measurements on larger scales \citep{Bahcall1995, Peacock2001} and of course independently confirmed with increasingly accurate claims from CMB acoustics \citep{Spergel2003, Planck2015XIII}.
New underway surveys to find large samples of clusters above $z \geqslant 0.5$ with lensing based masses are very exciting in this respect, so growth can be tracked as a function of cluster mass with unprecedented precision, including subtle modification by cosmological neutrinos \citep{Younger2005, Lahav2010, Carbone2012, Weinberg2013, Leistedt2014}, with little complication anticipated from detailed cosmological simulations that include gas physics \citep{Bocquet2016}. Using clusters, the current best estimates of the $\sigma_8\,\Omega_m^{\simeq 0.5}$ combination that principally determines growth \citep{Kaiser1984, Rozo2010, Mandelbaum2013} has hitherto been limited to the local volumes where cosmic variance and relatively small samples means it is rather unclear how to assess differences with the CMB based $\Omega_m h^2$ combination, fixed principally by the first peak of the CMB, and $\sigma_8$, where the uncertain level of electron scattering, $\tau$, smooths the amplitude of CMB fluctuations. The {\em Planck} weighted values of these observationally interdependent parameters, $\sigma_8$, $\Omega_m$, $h$, $\tau$, are now claimed to be in significant tension with the constantly high value of $H_0$ derived locally from the distance ladder \citep{Riess2016}.
Undermining the use of clusters in such comparisons is the indirectness of cluster mass estimates for which empirical scaling relations have to be relied on for converting observables to mass. Cluster richness seems to provide a robust connection as it is close to being linearly related to mass \citep{Rozo2009a, Rozo2009b, Rykoff2012}, with a slope of $d\log M_{200}/d\log N \simeq 1.1$, and a $\simeq 20\%$ inherent scatter inferred \citep{Andreon2016b} and with little evidence of evolution \citep{Younger2005, Andreon2014, Hennig2016}. Power-law scalings to convert X-ray and SZ measurements to total cluster masses are complicated by compressed gas and shocks from cluster interactions, so that the selection function and its evolution is challenging \citep{Suto2000, Allen2011, Mroczkowski2012, Molnar2015, Sereno2015a, Andreon2016a}. Lensing based scaling relations are now feasible for limited samples but for which the initial X-ray or SZ selection complicates matters \citep{Rozo2014b, Sereno2015c}.
Deeper, higher resolution, wide-field imaging surveys that exceed the SDSS, should soon rectify this lack of clusters with masses based on lensing and finally provide the long hoped statistically large sample of ``mass selected'' clusters to greater depth, in particular the Subaru/Hyper Suprime-Cam (HSC) \citep{Takada2010, Miyazaki2015}, J-PAS Northern Sky \citep{Benitez2014} and DES \citep{DES2016} surveys now underway. Broad-band surveys like HSC will require careful avoidance of foreground/member dilution of the background lensing signal that is feasible by excluding degenerate colour space \citep{Broadhurst2005, Medezinski2007, Medezinski2010, Umetsu2008}, with the fullest wavelength coverage to maximise the numbers of galaxies redder than the cluster and also dropouts. In the case of J-PAS the many narrow bands will provide unambiguous redshifts based on resolved spectral features, allowing clusters to be identified cleanly and to relatively low mass \citep{Ascaso2016}. Such data will improve future and past SZ, X-ray and radio ``relic'' based cluster searches \citep{Rottgering2011}, which will benefit enormously from lensing masses and the definition of cluster membership \citep{Hennig2016}, and can enhance the utility of the SDSS in the North.
Beyond cluster abundances, higher order moments of the density field, including the correlation function of galaxy clusters, also relate directly to the growth of structure \citep{Brodwin2007}. The clustering of clusters is in this respect far more useful than for galaxies where the strong dependence on Hubble-type implies a complex ``astrophysical bias'' \citep{Peacock2001, Lahav2002}. A major advantage of using clusters is their clear relation to the mass distribution, especially if direct lensing masses can be obtained for statistically large samples of clusters. This is unlike galaxies where ellipticals are measured to be much more spatially correlated than disk galaxies, implying as may be expected that the creation of galaxies from the underlying mass distribution is not simply related to the local density of dark matter. In the case of clusters the bias is more simply mass-density related and is not expected to be significantly influenced by gas physics, allowing relatively clean comparisons between theory and observation. For clusters a nearly linear relation is established between the measured richness and mass with a modest scatter, so richness can be reliably transformed statistically when examining the clustering of clusters. Previous clustering work with the SDSS has been either with relatively small local samples with an uncertain mass--richness relation, or relies on corrections for the wide smoothing by photometric redshifts \citep{Sereno2015c}, or on the angular clustering \citep{Baxter2016}. Here we establish the first spectroscopically complete analysis of cluster statistics using the depth of the SDSS survey, beyond the local Universe.
The careful redMaPPer work has been a big advance in identifying clusters by their red sequence of member galaxies and deriving reliable richnesses using the SDSS/BOSS survey data which has sufficient depth to detect clusters to $z~\simeq 0.3$ with high completeness \citep{Rykoff2014}. Currently 70\% of the brightest cluster galaxies in this clusters have redshift measurements with the SDSS/DR12 release. Here we augment these BCG measurements with additional cluster member redshift measurements, by correlating redMaPPer identified red sequence galaxies with the enlarged DR12 redshift sample from SDSS/BOSS, which we show here provides spectroscopic completeness to 93\% overall, and $>97$\% for the redMaPPer richness complete redshift range $z<0.325$ that we focus on in this paper.
In tandem with this observational progress, advances in the N-body simulations of $\Lambda{\rm CDM}$ have extended to volumes several times that of the observable Universe \citep{Angulo2012}. Large simulated volumes are necessary to accurately predict the number of massive clusters, given their rarity. The cosmological parameters chosen for these simulations follow the tradition set by \cite{Springel2005a} for such ground breaking simulations allowing consistency checks between these generations of simulations. The former consensus values adopted for these simulations \citep{Seljak2005} differ significantly from the present {\em Planck} weighted values of $\sigma_8$ and $\Omega_m$ that principally influence cluster predictions. The cause of this may be traced mainly to the relatively large $\tau$ estimated by WMAP \citep{Spergel2003}, raising $\sigma_8=0.9$, and lower $\Omega_m=0.25$. The amplitude of the CMB fluctuations on large scales scale as $A_s\,exp(-2\tau)$ where $A_s$ is the amplitude of the matter power spectrum; hence a higher $\tau$ implies a higher $A_s$ and consequently a higher $\sigma_8 \propto A_s$. The {\em Planck} weighted values today are significantly ``reversed'' for these key parameters mainly because of the much lower inferred $\tau$.
The structure of this paper is as follows. In Sect.~\ref{sect:data} we describe the data, namely, the redMaPPer cluster catalogue and the SDSS/BOSS spectroscopic sample. In Sec.~\ref{sect:mxxl} we describe the MXXL simulations that we use for comparison, and in Sec.~\ref{sect:correlation} and Sec.~\ref{sect:abundances} we describe the results obtained from the clustering and abundances analyses, respectively. Then, in Sec.~\ref{sect:likelihood} we perform a likelihood analysis to obtain the cosmologically favoured mass--richness relation, and explore some dependence on cosmology and its consistency with the clustering and the abundances results. In Sec.~\ref{sect:redshift_enhancement} we present the latest measurements of the redshift enhancement effect. Finally, we present our conclusions in Sec.~\ref{sect:conclusions}. Throughout we compare both the \cite{Planck2015XIII} weighted cosmological parameters and the consensus parameters set in 2003 used for the largest available simulations that differ significantly in terms of $\sigma_8$, $\Omega_m$ and $H_0$.
\section{Data}
\label{sect:data}
All our observational data comes from the Sloan Digital Sky Survey (SDSS), the most successful photometric and spectroscopic survey conducted to date \citep{Gunn2006}. Since it began in 2000, the SDSS has mapped the largest portion of the Universe to date, and provides high-precision data that has proved very useful for several kinds of large-scale structure analyses. The telescope has scanned more than 14,000 deg$^2$ of the sky with a mosaic CCD with five color-bands, {\it u, g, r, i } and {\it z} \citep{Fukugita1996}. It has also obtained the spectra of more than 1,600,000 unique objects during the SDSS-I/II progammes \citep{York2000}, and more than 1,500,000 unique spectra with a more advanced 1,000-fiber spectrograph \citep{Smee2013} during SDSS-III \citep{Eisenstein2011}, completed on 2014. The information obtained in this survey has been made public to the scientific community on a series of different Data Releases (DR), being the latest DR12 \citep{Alam2015}, which contains, among other information, 14,500 deg$^2$ of imaging that includes photometric information of 208,478,448 galaxies, and the optical spectra of 2,401,952 unique galaxies to $z=0.7$ in $\sim$10,500 deg$^2$ of the sky.
\subsection{redMaPPer Cluster Catalogue}
The red-sequence Matched-filter Probabilistic Percolation (redMaPPer, \citealp{Rykoff2014}) is the most robust and complete cluster catalogue based on SDSS data that has been produced to date. The redMaPPer cluster finder algorithm relies on a self-training procedure that calibrates the red-sequence as a function of redshift from a sample of galaxy clusters with known red spectroscopic galaxies. This red-sequence pattern is then used in the photometric data to find potential clusters and find, through an iterative process, the central galaxy (CG), the redshift $z_\lambda$, and the richness $\lambda$ of each cluster. This richness estimator is a multi-color evolution of the previous estimators described in \cite{Rozo2009b} and \cite{Rykoff2012}, which where employed in maxBCG \citep{Koester2007a} and GMBCG \citep{Hao2010} catalogues.
The richness of a cluster is defined as:
\begin{equation}
\lambda=\sum p_i\,\theta_i^L\,\theta_i^R\,,
\end{equation}
where $p_i$ is the probability that each galaxy found near the cluster is actually a cluster member, and $\theta_i^L$ and $\theta_i^R$ are the luminosity and radius-dependant optimized weights:
\begin{equation}
\theta_i^L=\frac{1}{2}\left[1+{\rm erf} \left( \frac{m_{\rm max}-m_i}{\sigma_i}\right) \right]\,,
\end{equation}
\begin{equation}
\theta_i^R=\frac{1}{2}\left[1+{\rm erf} \left( \frac{R(\lambda)-R}{\sigma_R}\right) \right]\,,
\end{equation}
with $m_{\rm max}$ the magnitude that corresponds to the $0.2 L_*$ luminosity threshold, $\sigma_i$ the photometric error of galaxy $i$, $\sigma_R=0.05 \,h^{-1}{\rm Mpc}$, and $R(\lambda)$ the richness-dependant aperture:
\begin{equation}
R(\lambda)=\left( \frac{\lambda}{100}\right)^{0.2}\,h^{-1}{\rm Mpc}\,.
\end{equation}
We refer the reader to \cite{Rykoff2014} and \cite{Rozo2015b} for an in-depth explanation of the algorithm features.
This cluster finder algorithm was designed to process future large photometric surveys like DES and Large Synoptic Survey Telescope (LSST), but it has been already run on the SDSS DR8 photometric and SDSS DR9 spectroscopic data in order to check the possible systematics associated to the algorithm itself, and compare its results with other clusters catalogues, like the SZ {\em Planck} cluster catalogue \citep{Rozo2015a}, or the X-CLASS X-ray cluster \citep{Sadibekova2014}. These analyses led to an updated 5.10 version of the algorithm described in \cite{Rozo2015b}, which we will augment here with the latest DR12 release that we show below provides a substantial enhancement in the number of the clusters with reliable spectroscopic redshifts.
The resulting redMaPPer catalogue covers an effective area of $\sim$10,400 deg$^2$, and contains 26,350 clusters in the $0.08 \leqslant z_{\rm photo} \leqslant 0.55$ redshift range.
There are several reasons why we select this catalogue among all the others described in the literature:
\begin{itemize}
\item High purity, where purity here is not understood as false detections (like is usually defined in SZ and X-ray cluster samples), but as the fraction of clusters that is not affected by projection effects that may lead to overestimated richness measurements. This purity is claimed to be $>95\%$ in \cite{Rykoff2014}, although a more recent study \citep{Simet2016} finds a higher rate of projection effects of the order of $12\%\pm4\%$.
\item A conservative low richness cutoff, or detection threshold of $\lambda/S(z)>20$, that enhances the performance of the resulting cluster catalogue. The ``scale factor'' $S(z)$ is introduced in order to take into account the limited depth of the sample, so a cluster of richness $\lambda$ has $\lambda/S(z)$ galaxies above the magnitude limit of the survey, and $S=1$ at $z<0.35$, where the DR8 is volume limited. This richness cut corresponds to a mass limit of approximately $M_{200c}\geqslant1.4\times10^{14}\,M_\odot\,h^{-1}$.
\item Together with the photometric redshift estimates of each cluster, an SDSS spectroscopic redshift measurement is provided for the CG when available. Of the original 26,350 clusters contained in the catalogue, 16,259 contain this information.
\item Instead of providing a unique CG candidate for the cluster, the algorithm indicates the centering probability $p_{\rm cen}$ of the five most probable CGs, together with their position in the cluster.
\item As previous studies \citep{Jimeno2015} that checked and compared different cluster catalogues have shown, redMaPPer is the catalogue that offered the most consistent results in terms of a cross comparison of the redshift enhancement and the gravitational redshift effects associated to clusters.
\item Provides supplementary information of the member galaxies that have been considered in the richness estimation, with their membership probability. There is information for 1,736,221 member galaxies, of which 72,642 also contain spectroscopic redshift measurements.
\end{itemize}
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/redMaPPer_Sky_Distribution}}
\caption{redMaPPer catalogue and BOSS footprint. The region inside the red line represents the area of the sky covered by the BOSS that overlaps with the redMaPPer catalogue. As can be seen from the figure, there is a region in the bottom left corner with redMaPPer clusters outside the BOSS region. We exclude those clusters from the final sample so that we have uniformly high completeness between the redMaPPer and the BOSS samples required in our analysis.}
\label{fig:redMaPPer_Sky_Distribution}
\end{figure}
\subsection{BOSS spectroscopic sample}
The Baryon Oscillations Spectroscopic Survey (BOSS, \citealp{Dawson2013}), designed to measure the baryon acoustic oscillations (BAO) scale at $z=0.3$ and $z=0.57$ to a 1.0\% accuracy, is the largest of the four surveys that comprise the SDSS-III progamme. Now completed, it uniformly targeted and obtained the spectra of galaxies in two redshift ranges: $0.15<z<0.4$, which lead to the color-selected ``LOWZ'' sample, composed of the brightest and the reddest of the low redshift galaxies, and $0.4<z<0.8$, designed to obtain through a series of photometric colour cuts a volume-limited sample of galaxies with approximately constant stellar mass, the so called ``CMASS'' sample. A total of about a half and a million unique galaxy spectra were measured over 10,400 deg$^2$ of the sky in the Northern and Southern Galactic Caps.
We update the publicly available catalogue based on DR8 and DR9 data using the more recent DR12 BOSS spectroscopic data. In order to work with the most reliable data, from all the BOSS galaxy spectra we select those that satisfy the following quality flag conditions: \texttt{PLATEQUALITY}$\,=\,$\texttt{good}, \texttt{Z\_ERR\_NOQSO}$\,<0.001$ and \texttt{ZWARNING\_NOQSO}$\,=0$ or \texttt{ZWARNING\_NOQSO}$\,=16$. The footprint of the final considered sample, which contains the spectra of 1,339,107 galaxies, is shown in Fig.~\ref{fig:redMaPPer_Sky_Distribution}. We cross correlate the angular position on the sky of the five most probable CGs associated to each cluster with the whole BOSS spectroscopic sample, and identify those objects that are closer than 0.50 arcseconds, finding 3,772 matches for the most probable CGs that did not have an spectroscopic redshift measurement before. More than 99\% of all the identifications are done for pairs that are closer than 0.02 arcseconds, but we actually find a gap between 0.40 and 1.50 arcseconds where no identification at all is made, making clear that we are safe from any possible misidentification due to close galaxy pairs. We repeat this identification process with the second, third, fourth and fifth most probable CGs in each cluster. A sizeable fraction of CGs do not have measured redshifts yet whereas one or more of the highly probably CGs often does. This fuller comparison of members with redshifts provides now for the first time a highly spectroscopically complete sample of clusters with which make several precise calculations below described.
Finally, we exclude those clusters that, as can be seen in Fig.~\ref{fig:redMaPPer_Sky_Distribution}, are outside the considered BOSS area. This leaves us a final sample of 19,473 clusters with spectroscopic redshift measurement of their CG, and a total of 23,135 clusters with spectroscopic redshift measurement of one of their most likely centrals. In comparison, in the original photometric redshift catalogue there were 24,869 clusters in the same BOSS region, meaning that we have now spectroscopic information of more than the 93 percent of the clusters.
To evaluate the final redshift of those clusters with more than one CG with a spectroscopic measurement, we make an average of their redshifts $z_{{\rm CG},i}$, weighted by the centering probability $p_{{\rm cen},i}$ associated to each of the five potential CGs:
\begin{equation}\label{eq:z_cluster}
z_{\rm cluster}=\frac{\sum\limits_{i=1}^5 z_{{\rm CG},i}\,p_{{\rm cen},i}}{\sum\limits_{i=1}^5 p_{{\rm cen},i}}
\end{equation}
The final redshift distribution of the clusters inside the BOSS region is shown in Fig.~\ref{fig:redMaPPer_Redshift_Distribution}. The difference between our definition of $z_{\rm cluster}$, and $z_{\rm photo}$, the original photometric redshift provided by the catalogue is shown in Fig.~\ref{fig:Spec_vs_Photo}.
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/redMaPPer_Redshift_Distribution}}
\caption{
Original photometric redshift (dot-dashed red line), DR9 spectroscopic redshift (dashed blue line) and updated DR12 spectroscopic redshift (solid green line) redMaPPer clusters redshift distribution. As the grey shaded line, we plot for comparison the proportional volume in each redshift bin. From the original 24,869 clusters with photometric redshift estimates, 15,936 clusters also have DR9 spectroscopic redshift information of their most likely CG, and 23,135 clusters have DR12 spectroscopic redshift measurements of at least one of their most likely CG. The clusters that fall outside the BOSS region have been excluded from these samples. Note that there is near full completeness in spectroscopic redshift to the peak of the redMaPPer selection function at $z=0.325$, and the bump feature at $z\simeq 0.25$ in the photometric redshift distribution, in contrast to the spectroscopic sample where the numbers smoothly rise with the increasing volume, as expected for a complete sample.}
\label{fig:redMaPPer_Redshift_Distribution}
\end{figure}
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/Spec_vs_Photo}}
\caption{
Relation between our definition of $z_{\rm cluster}$ (Eq.~\ref{eq:z_cluster}), based on DR12 data and the centering probabilities $p_{{\rm cen},i}$ of all the CG candidates, and $z_{\rm photo}$ for each of the clusters in the redMaPPer catalogue. The residual wave behaviour is understood to be related to the ``quantization'' of photometric redshift estimates from the discrete wavelength coverage of the photometric bands.}
\label{fig:Spec_vs_Photo}
\end{figure}
\section{MXXL simulations}
\label{sect:mxxl}
The ``Millennium-XXL'' (MXXL, \citealp{Angulo2012}) simulation is one of the largest dark matter N-body simulations performed to date. It follows the nonlinear creation and growth of dark matter structures within a cube of 3,000 Mpc $h^{-1}$ on a side, which contained 6720$^3$ particles of mass $m_p = 8.456\times 10^9\,M_\odot$.
Compared to its predecessor the ``Millennium Simulation'' (MS, \citealp{Springel2005a}), the MXXL simulation volume is 200 times bigger, comprising the equivalent volume of the Universe to $z=0.72$, or 7 times the volume of the BOSS survey. Although the particle resolution is 7 times lower in the MXXL simulation than in the MS, it is 300 times higher than the ``Hubble Volume Simulation'' \citep{Evrard2002} and has ample resolution for our massive clusters related purposes. Note that these large simulations deliberately share the same pre-{\em Planck} cosmology set by WMAP in 2003 to allow consistency checks. This is not in practice a limita-
tion for our work given the realization by \cite{Angulo2010} that a simple rescaling of size and redshifts can effectively provide predictions for other cosmologies using the same simulation, within the context of $\Lambda{\rm CDM}$.
This simulation was designed to provide enough statistical power as to study and interpret some of the problems related to the observation of galaxy clusters, like the scaling between the real mass of a cluster with the associated cluster observables, i.e., richness, lensing mass, X-ray luminosity and thermal Sunyaev-Zel'dovich signal. The code employed in the simulation is a extremely memory-efficient version of \texttt{GADGET-3}, which is itself a more sophisticated and efficient version of \texttt{GADGET-2} \citep{Springel2005b}, the code used for the MS.
This code is also designed to carry out halo and subhalo finding procedures during its execution with a friends-of-friends algorithm (FoF, \citealt{Knebe2011}), combined with the \texttt{SUBFIND} algorithm \citep{Springel2001}, which identifies locally overdense regions within the parent haloes found by the FoF. The halo catalogues produced during this search provide information about masses, velocity dispersions, halo shapes, velocities, etc...
The original cosmology employed in the MXXL simulation is $\Lambda{\rm CDM}$, with the same cosmological parameters that were employed in the MS: $\Omega_m=0.25$, $\Omega_\Lambda=0.75$, $\sigma_8=0.9$ and $H_0=73$ km s$^{-1}$ Mpc$^{-1}$. We update this cosmology and the resulting halo catalogue using the prescription given in \cite{Angulo2010}, with this new set of cosmological parameters taken from the combined analysis of {\em Planck} CMB data, BAO surveys and the JLA sample of Type Ia SNe \citep{Planck2015XIII}: $\Omega_m=0.3089$, $\Omega_\Lambda=0.6911$, $\sigma_8=0.8159$ and $H_0=67.74$ km s$^{-1}$ Mpc$^{-1}$. This algorithm to update simulations relies basically on the reassignments of masses, velocities and lengths, and the rescaling of the time steps, i.e., the redshifts of the snapshots, to match the shape of the smoothed linear matter power spectra of the desired cosmology, and thus the growth of structure. The modification of the long wavelength modes relies on the Zel'dovich approximation so the difference on large scales is also taken into account. We use the code \texttt{CAMB} \citep{Lewis2000} to compute the linear matter power spectra required for comparison. From now on, we will refer to this updated version as the MXXL simulations.
\subsection{MXXL synthetic cluster catalogue}
We now proceed to create a synthetic optical cluster catalogue from the MXXL simulations, so we can compare it with the redMaPPer catalogue. We use the data coming from 5 adjacent snapshots of the MXXL simulations, corresponding to the redshifts $z=0.027,\,0.128,\,0.242,\,0.393$ and $0.486$.
In each of these snapshots, we select of the order of $\sim7$ million DM haloes with masses above $M_{200c}=10^{12}\,M_\odot\,h^{-1}$, where here $M_{200c}$ is defined as the mass that is enclosed in a sphere centered in the potential minimum of the halo, that has a mean density 200 times the critical density of the Universe.
Now, in order to build a synthetic cluster catalogue from the DM halo catalogue, where the ``true'' mass, the position and peculiar velocity of the halo is known, we need to assign a richness value and artificially place a CG in each of these haloes, which is, at the end, what is provided by the redMaPPer catalogue, as described in the previous section.
\subsubsection{Richness}
First of all, we need to associate an estimate of the richness to all the haloes in the MXXL simulations. We employ the usual mass--richness $\left<M|\lambda\right>$ relation, which adopts the form:
\begin{equation}\label{eq:MR_relation}
\ln \left( \frac{M_{200c}(\lambda)}{10^{14}\,M_\odot\,h^{-1}} \right) = \ln \left( \frac{M_{200c}(\lambda_0)}{10^{14}\,M_\odot\,h^{-1}} \right) + \alpha_{M|\lambda}\,\ln\left(\frac{\lambda}{\lambda_0}\right)\,,
\end{equation}
where $M_{200c}(\lambda_0)$ is a reference mass at a given value of $\lambda=\lambda_0$, and $\alpha_{M|\lambda}$ is the slope of the mass--richness relation.
This relation has been the focus of intense research in the past decade, as an accurate scaling between the mass of a halo and the cluster observables is mandatory to constrain cosmological parameters from cluster counting techniques \citep{Rozo2007,Tinker2012}. If one wants to obtain a calibration which does not rely on matching a certain cosmology, a relation between the richness $\lambda$ and the lensing mass $M_{\rm lens}$, the X-ray luminosity $L_X$, or the thermal Sunyaev-Zel'dovich signal $Y_{SZ}$ must be found. \cite{Johnston2007}, using weak lensing measurements, find a slope of $\alpha_{M|\lambda}=1.28 \pm 0.04$ for a reference mass of $M_{200}(\lambda_0=20) = 0.88 \times10^{14} \,M_\odot\,h^{-1}$. \cite{Wen2010}, using a sample of 24 nearby clusters with weak lensing and X-ray mass measurements, found a steeper slope $\alpha_{M|\lambda}=1.55$, although it should be noted that their richness definition differs from the one given by maxBCG and redMaPPer algorithms. More recently, using SDSS weak lensing data and the redMaPPer catalogue in the $0.1\leqslant z \leqslant 0.33$ redshift region, \cite{Simet2016} find a slope of $\alpha_{M|\lambda}=1.33^{+0.09}_{-0.10}$ for a reference mass of $M_{200m}(\lambda_0=40) \sim 2.22 \times10^{14} \,M_\odot\,h^{-1}$. On the other hand, one can assume a determined cosmology and use matching abundances techniques to constrain this scaling relation. \cite{Rykoff2012} provide a tentative mass--richness relation with an slope of $\alpha_{M|\lambda}=1.06$ for a reference mass of $M_{200c}(\lambda_0=60) = 4.4\times10^{14}\,M_\odot\,h^{-1}$. For a fixed cosmology and using the photometric redshift estimates from DR8, \cite{Baxter2016} measured the angular clustering of redMaPPer clusters in order to calibrate the mass--richness and mass--bias relations, and obtained an slope of $\alpha_{M|\lambda}=1.18\pm0.16$ and no indication of any evolution of the mass--richness relation with redshift.
Because these cluster observables are intrinsically noisy, for each mass-observable relation there is a scatter associated to the distribution of the values of such observable given a true mass. The uncertainty in the value of this scatter alters the effectivity of any cosmological analysis as much as the uncertainty in the mean mass--observable relation. The fractional scatter $\sigma_{M|\lambda}$ on the halo mass at fixed richness can be described by a log-normal distribution:
\begin{equation}
\Delta \ln \left( \frac{M_{200c}(\lambda)}{10^{14}\,M_\odot\,h^{-1}} \right) = \sigma_{M|\lambda}\,.
\end{equation}
There have been several studies in the past years trying to evaluate $\sigma_{M|\lambda}$, using either simulations or data, considering mass proxies like X-ray luminosity.
\cite{Rozo2009a}, using maxBCG catalogue and weak lensing and X-ray data, found $\sigma_{M|\lambda} = 0.45^{+0.20}_{-0.18}$ at $\lambda \approx 40$.
Introducing modifications on the maxBCG richness estimator (modifications that would later be implemented in the redMaPPer algorithm), \cite{Rykoff2012} obtained a value of the scatter at fixed mass between $\sigma_{M|\lambda} \approx 0.2-0.3$.
Taking a different approach, \cite{Angulo2012} first populated the MXXL simulations haloes with galaxies using a halo occupation distribution (HOD) model, identified clusters and measured their richness using a procedure similar to the one employed by the maxBCG algorithm, and obtained a predicted scatter of $\sigma_{M|\lambda} = 0.36$.
\cite{Becker2007} found relatively higher values for the scatter, $\sigma_{M|\lambda} \sim 0.6$, using the velocity dispersion in maxBCG clusters to obtain a mass estimation.
In Sec.~\ref{sect:likelihood}, we will find, through a likelihood analysis, the values of $\kappa_{M|\lambda}$, $\alpha_{M|\lambda}$ and $\sigma_{M|\lambda}$ that best describe the observations, where for clarity we have defined:
\begin{equation}
\kappa_{M|\lambda} \equiv \ln \left( \frac{M_{200c}(\lambda_0=60)}{10^{14}\,M_\odot\,h^{-1}} \right)\,,
\end{equation}
and use these values in Secs.~\ref{sect:correlation} and \ref{sect:abundances} when comparing any data obtained from the redMaPPer with the ``model'' produced by the MXXL simulations. We will follow the results obtained by \cite{Angulo2012} and \cite{Rozo2010}, and consider the value of the scatter independent of the richness, and ignore any possible redshift evolution of the scatter or of the slope of the mass--richness relation. Also, because in the simulations we know the value of the true mass, rather than the value of the observable, we need to convert $\sigma_{M|\lambda}$ into $\sigma_{\lambda|M}$ inverting Eq.~\ref{eq:MR_relation}, so $\sigma_{M|\lambda}= \alpha_{M|\lambda}\,\sigma_{\lambda|M}$. In Fig.~\ref{fig:MR_relation} we show the distribution of the masses of the MXXL clusters as a function of one realization of the richness associated through a mass--richness relation with, e.g., a pivot mass of $\kappa_{M|\lambda}=1.35$, a slope of $\alpha_{M|\lambda} = 1.10$, and a scatter of $\sigma_{M|\lambda} = 0.20$, in comparison to the ideal mass--richness relation without scatter. The upscattering of low mass clusters into high richness regions increases with higher values of the slope or larger scatter.
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/MR_relation}}
\caption{Mass distribution as a function of richness of the MXXL DM haloes found in snapshot 54, that corresponds to $z=0.242$, close to the mean redshift of the complete cluster sample. The ``true'' distribution, shown as the black dashed line, is the theoretical mass--richness relation with $\kappa_{M|\lambda}=1.35$ and $\alpha_{M|\lambda} = 1.10$. When the scatter is not considered, there are $\sim$116,000 clusters above the richness threshold. If a scatter of $\sigma_{M|\lambda}=0.20$ is introduced, we obtain the cluster distribution shown in the figure, with around $\sim$128,000 clusters with richness, $\lambda>22$ due to the upscatter of low mass clusters.}
\label{fig:MR_relation}
\end{figure}
\subsubsection{CG miscentering}
Recent studies have shown that the assumption that central galaxies (CGs) lie basically at rest in the deepest part of the potential well is not accurate. Although there is evidence of an special correlation between the mass of a cluster halo and the properties (mass, morphology, star formation rates, stellar population, colour, etc...) of its brightest cluster galaxy (BCG, \citealp{vonderLinden2007}), we may expect a large proportion of clusters are non-relaxed dynamical systems that are still evolving, with BCGs following an evolving orbit that need not be located at the time varying minimum of the cluster potential.
Miscentering is one of the main sources of error on stacked measurements on clusters, including weak-lensing mass determinations \citep{Umetsu2011b, Umetsu2011a}, analyses of the power spectrum \citep{Reid2010}, gravitational redshift measurements \citep{Jimeno2015}, or velocity dispersion calculations \citep{Becker2007}. That is why in the last few years many authors have tried to determine the level of miscentering statistically.
There are two reasons why this miscentering may occur: the galaxy identified as the CG not being the galaxy with the lowest specific potential energy (i.e., the one that could be considered as the CG), or the real CG not being in the center of mass of the DM halo.
In cluster catalogues like the maxBCG \citep{Koester2007b}, the GMBCG \citep{Hao2010} or the WHL12 \citep{Wen2012}, the center of a cluster is identified with the position of the BCG. Although this assumption seems to improve the overall centering performance of cluster-finder algorithms \citep{Rozo2014a}, \cite{Skibba2011} claimed that between the 25\% and the 40\% of cases the BCG is not the real CG, but a satellite galaxy, and \cite{vonderLinden2007} found in a sample of 625 clusters that in more than a half of them the BCG was not located in the center of the cluster.
Trying to quantify the level of miscentering, and using mock catalogues, \cite{Soares-Santos2011} found an offset distribution that could be fitted by a 2D gaussian with a standard deviation of $\sigma=0.47\,h^{-1}{\rm Mpc}$, meanwhile \cite{Johnston2007} found $\sigma=0.42\,h^{-1}{\rm Mpc}$ for the BCGs that were not accurately centered, which ranged from 40\% to 20\% as a function of cluster richness.
In order to address this problem, the redMaPPer iterative self-training centering algorithm used BCGs as the seeds for the centering process, but in successive calibrating iterations all the galaxies found in the cluster that were consistent with the red-sequence were considered as potential CGs, and had a centering probability assigned. When convergence is obtained, the galaxy with the highest probability of being in the center of the cluster is tagged as the CG. It should be noted that the miscentering introduced by the red-sequence prior that does not allow blue galaxies to be selected as CGs is expected to affect less than 2\% of the clusters. In any case, when the cluster centers found on high resolution X-ray data were compared to the CGs found in the redMaPPer catalogue, the redMaPPer algorithm was claimed to have a centering success rate of $\approx 86\%$ \citep{Rozo2014a}.
To include this effect into our synthetic catalogue, we follow the results of \cite{Johnston2007} and introduce a miscentering probability as a function of the halo richness. The probability $p_{mc}$ of a CG being displaced from the center of its host halo follows: $p_{mc}(\lambda) = (2.13+0.046\,\lambda)^{-1}$, and such displacement is given by a 2D Gaussian of width $\sigma_{mc}=0.42\,h^{-1}{\rm Mpc}$.
\subsubsection{CG peculiar motion}
There are good reasons to believe that, compared to other member galaxies, CGs are a cold population of galaxies due to dynamical friction or possible central gas cooling. How much so is still a matter of discussion. Hierarchical merging of clusters means we must expect some level of dispersion periodically as CGs respond to a rapidly evolving potential and merge with each other. In any case, it is clear that their peculiar velocity cannot be ignored and in some cases is very anomalous \citep{Sharples1988} and with a small net gravitational redshift expected \citep{Broadhurst2000} and measured \citep{Wojtak2011, Jimeno2015, Sadeh2015}. Puzzlingly, a significant fraction of CGs with peculiar motions are located in the peak of the X-ray emission. Given the hydrodynamical forces relevant for the cluster, it is not expected that the gas should move together with the CG galaxy during cluster encounters \citep{Ricker2001, Molnar2015}.
The relation between the value of $\sigma_{\rm CG}$, the distribution of CG motions, and $\sigma_{\rm gal}$, the value of the dispersion associated to the cluster satellite galaxies, is still quite unknown. \cite{Oegerle2001} found that from a sample of 25 clusters almost all CGs showed peculiar velocities relative to the mean velocity of the clusters studied, with $\sigma_{\rm BCG}\approx175\,{\rm km\,s}^{-1}$. \cite{Coziol2009}, studying a much larger sample of clusters, found that CGs having peculiar motions within the cluster was a general phenomenon, with less than the 30 percent of them having velocities compatible with zero, and more than half of them having velocities higher than $0.3\,\sigma_{\rm gal}$, depending this value slightly on cluster richness. \cite{Skibba2011}, studying the miscentering of CGs, also found a relatively high value for their velocities, with $\sigma_{\rm CG}\simeq0.5\,\sigma_{\rm gal}$, that had also little dependence with the mass of the host cluster.
In any case, to mimic this motion within the cluster of our already placed CGs, we will assign a peculiar velocity to them given by $\sigma_{\rm CG} \simeq 0.4 \, \sigma_{\rm vir}$, where $\sigma_{\rm vir}$ is the virial velocity associated to the mass of the cluster.
Once this CG peculiar motion has been added to the motion of the cluster, we move CGs from real-space positions $\bm{x}$ to redshift-space positions $\bm{s}$:
\begin{equation}
\bm{s} = \bm{x} + \left( 1 + z \right) \frac{\bm{v} \cdot \bm{\hat{x}_i}}{H(z)}\,\bm{\hat{x}_i}\,,
\end{equation}
where $\bm{\hat{x}_i}$ is an arbitrarily line-of-sight chosen direction, and $\bm{v}$ is the final velocity of the CG.
\section{Correlation Function}
\label{sect:correlation}
We compute the two-point redshift-space correlation function $\xi(s,\mu)$ using the \cite{Landy1993} estimator, where $s$ is the redshift-space distance in $\,h^{-1}{\rm Mpc}$ units, and $\mu = \cos\theta$, where $\theta$ is the angle of the pair with respect to the line-of-sight:
\begin{equation}
\xi(s,\mu)=\frac{DD(s,\mu)-2DR(s,\mu)+RR(s,\mu)}{RR(s,\mu)}\,,
\end{equation}
where $DD(s,\mu)$, $DR(s,\mu)$ and $RR(s,\mu)$ are the normalized number of pairs found at a distance of $s$ and an angular separation of $\mu$ in the data-data, data-random and random-random samples.
We want to investigate the multipoles of the correlation function $\xi(s,\mu)$, which are:
\begin{equation}
\xi_\ell(s)=\left(\frac{2\ell+1}{2}\right)\int_{-1}^{+1}\xi(s,\mu)L_\ell(\mu)d\mu\,,
\end{equation}
where $L_\ell(\mu)$ is the $\ell$th Legendre polynomial. We weight $DD$, $DR$ and $RR$ pairs with the associated value of $L_\ell(\mu)$ for the monopole ($\ell=0$) and the quadrupole ($\ell=2$):
\begin{equation}
\xi_\ell(s)=\left(\frac{2\ell+1}{2}\right)\frac{DD_\ell(s)-2DR_\ell(s)+RR_\ell(s)}{RR_0(s)}
\end{equation}
To optimally weight regions with different number densities, we apply FKP weighting \citep{Feldman1994} to each cluster:
\begin{equation}
w_P=\frac{1}{1+n(z)P_{_{\rm FKP}}}\,,
\end{equation}
where $n(z)$ is the mean cluster density at redshift $z$, and $P_{_{\rm FKP}}=$ 20,000$ \,h^3\,{\rm Mpc}^{-3}$.
For the data sample, i.e., redMaPPer clusters, instead of considering only the most probable CGs, we use all the five CG candidates provided in the redMaPPer catalogue for each cluster, taking their centering probabilities $p$ to weight their contribution to the final pair count, so that $\sum_{\alpha=1}^5 p_{i,\alpha} = 1$ for cluster $i$. The final data-data, data-random and random-random pair counts can be expressed as:
\begin{equation}
DD_\ell(s)=\sum_{i=1}^N \sum_{j=i+1}^N \sum_{\alpha=1}^5 \sum_{\beta=1}^5 w_{ P,i}\,w_{P,j}\,p_{i,\alpha}\,p_{j,\beta}\,L_\ell(\mu)\,,
\end{equation}
\begin{equation}
DR_\ell(s)=\sum_{i=1}^N \sum_{j=1}^{\widetilde{N}} \sum_{\alpha=1}^5 w_{ P,i}\,w_{P,j}\,p_{i,\alpha}\,L_\ell(\mu)\,,
\end{equation}
\begin{equation}
RR_\ell(s)=\sum_{i=1}^{\widetilde{N}} \sum_{j=i+1}^{\widetilde{N}} w_{ P,i}\,w_{P,j}\,L_\ell(\mu)\,,
\end{equation}
where $N$ is the number of clusters in the data sample, and $\widetilde{N}$ is the number of objects in the synthetic random catalogue.
We are also interested in the projected correlation function $\Xi(r_\perp)$, which provides information of the real-space clustering so that we do not need to worry about the complications of peculiar motions \citep{Davis1983}. It is obtained integrating the 2D correlation function $\xi(r_{los},\,r_{\perp})$ along the line-of-sight:
\begin{equation}
\Xi(r_\perp)=2\int_0^\infty \xi(r_{los},\,r_{\perp})\,dr_{los}\,,
\end{equation}
where $r_{los}=\mu\,s$, and $r_\perp=\sqrt{s^2-r_{los}^2}$, are the parallel and perpendicular directions to the line-of-sight. To compute it from $\Xi(r_\perp)$, we use the estimator:
\begin{equation}
\Xi(r_\perp)=2\sum_i^{r_{los\,\max}} \xi(r_{los},\,r_{\perp i})\,\Delta r_{los}\,,
\end{equation}
where we bin the 2D correlation function into linearly spaced bins of constant size $\Delta r_{los} = 5\,h^{-1}{\rm Mpc}$, and select a maximum summation distance of $r_{los\,\max}=30\,h^{-1}{\rm Mpc}$.
We use the jackknife method to compute the covariance matrices of the correlation function. For each cluster sample we randomly create 80 cluster subsamples that comprise 1/80th part of the total, and then compute 80 times the monopole, dipole and projected correlation function of the total cluster sample with one of those cluster subsamples removed. The covariance matrix associated to this sample is then:
\begin{equation}
C_{i,j}=\frac{N-1}{N}\sum_{k=1}^N \left(\overline{\chi}_i - \chi_i^k\right) \left(\overline{\chi}_j - \chi_j^k\right)\,,
\end{equation}
where $\chi_i$ corresponds to $\xi_0$, $\xi_2$ or $\Xi$ at the $i^{th}$ bin, and $\overline{\chi}_i$ is the mean value of the $N=80$ calculations at the $i^{th}$ bin.
\subsection{Results}
\subsubsection{Redshift-space two-point correlation function}
First of all, and as a check of the power of both the spectroscopic sample based on the redMaPPer catalogue that we have constructed, and our probability-weighted estimator $\xi(s)$, we compute the redshift-space two-point correlation function up to $s=80\,h^{-1}{\rm Mpc}$. For that, we use a test sample containing all the clusters in the $0.080 \leqslant z \leqslant 0.325$ redshift range, and richness $\lambda>22$, which contains 7,143 clusters. The values obtained are shown in Fig.~\ref{fig:Comparison_CF}, where we also compare our results with previous measurements found in the literature \citep{Bahcall2003, Estrada2009, Sereno2015b}. Notice in Fig.~\ref{fig:Comparison_CF} the correlation function rises continuously to small radius with a slope similar to previous work but with much higher precision because of the larger numbers of clusters sampled to higher redshift. The increasingly shallower slope of \cite{Sereno2015b} at smaller scales is due to smoothing by the relatively large proportion of photometric redshifts in their analysis of the GMBCG cluster catalogue.
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/Comparison_CF}}
\caption{Our measurement of the redshift space correlation function (black circles) shows the huge improvement in precision now possible with the latest SDSS release, for the spectroscopically complete sample of 7,143 redMaPPer clusters contained in the full range $0.08 \leqslant z \leqslant 0.325$, with a lower richness limit of $\lambda>22$. For comparison, we plot previous measurements found in the literature, using other SDSS cluster samples: Bahcall et al. (2003) as red squares, Estrada et al. (2009) as green diamonds, and Sereno et al. (2015b) as blue triangles. Notice the correlation function slope and amplitude is similar to previous work but with much higher precision because of the larger number of clusters sampled to higher redshift, and without the photometric redshifts that smooth the Sereno et al. (2015b) correlations (blue triangles) at small scales.}
\label{fig:Comparison_CF}
\end{figure}
\subsubsection{Measured monopole, quadrupole, 2D, and projected correlation function}
Here we divide the redMaPPer catalogue into three different richness bins, $\lambda_1\in[22, 30)$, $\lambda_2\in[30, 45)$, and $\lambda_3\in[45, 200)$, with similar numbers of clusters and each of the resulting subsamples is again divided into two redshift regions in order to have a ``low-z'' and a ``high-z'' sample so we can examine evolution. The redshift regions for each of these six subsamples, together with the mean redshift, mean richness and number of clusters contained, are listed in Table~\ref{tab:redMaPPer_Subsamples}.
\begin{table}
\center
\caption{
\label{tab:redMaPPer_Subsamples}
Richness range $\lambda$, redshift region $z_{\rm clu}$, number of clusters $N$, mean richness $\left<\lambda\right>$, and mean redshift $\left<z_{\rm clu}\right>$ of the six redMaPPer cluster subsamples considered in the measurement of the correlation function.
}
\begin{tabular}{l c c c c c}
\hline
\hline
Subsample &$\lambda$ & $z_{\rm clu}$ & $N$ & $\left<\lambda\right>$ & $\left<z_{\rm clu}\right>$\\
\hline
$\lambda_1$ low-z &$[22, 30)$ & $[0.080, 0.250)$ &1770 &25.3 &0.189 \\
$\lambda_1$ high-z &$[22, 30)$ & $[0.250, 0.400]$ &4493 &25.6 &0.334 \\
\vspace{-5pt}\\
$\lambda_2$ low-z &$[30, 45)$ & $[0.080, 0.275)$ &1527 &36.0 &0.205 \\
$\lambda_2$ high-z &$[30, 45)$ & $[0.275, 0.425]$ &4008 &35.9 &0.363 \\
\vspace{-5pt}\\
$\lambda_3$ low-z &$[45, 200)$ & $[0.080, 0.300)$ &1024 &63.1 &0.221 \\
$\lambda_3$ high-z &$[45, 200)$ & $[0.300, 0.450]$ &2384 &62.11 &0.388 \\
\hline
\end{tabular}
\end{table}
We now compute the monopole $\xi_0(s)$, the quadrupole $\xi_2(s)$, the 2D correlation function $\xi(r_{los},\,r_{\perp})$, and the projected correlation function $\Xi(r_\perp)$ following the procedure described before. We bin s in 8 logarithmic distributed bins between 5$\,h^{-1}{\rm Mpc}$ and 35$\,h^{-1}{\rm Mpc}$, and bin $r_{los}$ and $r_{\perp}$ in linearly spaced bins of a size equal to 0.5$\,h^{-1}{\rm Mpc}$. The obtained values of $\xi_0$, $\xi_2$ and $\Xi$ for the six subsamples are shown in Fig~\ref{fig:1D_CF}, together with the MXXL realization model that best describes the real-space correlation function (see Sec.~\ref{sect:likelihood}). In the same way, the redMaPPer and the MXXL 2D correlation functions are shown in Fig.~\ref{fig:2D_CF}. We see a clear trend towards higher correlation amplitude with richness, and little dependence on redshift.
Later in Sec.~\ref{sect:likelihood} we discuss these measurements in comparison with the simulations predictions for the correlation functions after first defining the mass--richness relation between the mass function of the simulations with the observed cluster abundances.
\begin{figure*}
\centering
\includegraphics[width=0.98\textwidth]{Images/1D_CF}
\caption{Monopole $\xi_0(s)$ (left column), quadrupole $\xi_2(s)$ (central column) and projected correlation function $\Xi(r_\perp)$ (right column) for two redshift samples: low-z (red squares for redMaPPer and red dashed lines for the model) and high-z (blue circles for redMaPPer and blue dot-dashed lines for the model), and three richness ranges: $\lambda \in [22, 30)$ (upper panels), $[30, 45)$ (central panels), and $[45, 200)$ (lower panels). Notice the clear increase of the amplitude of the correlation functions with richness, and the near independence with redshift, in excellent agreement with the model predictions, which match very well with radius and with richness. The model curve here is derived in Sec.~\ref{sect:likelihood} from our likelihood analysis based on the MXXL simulations, and the best fitting mass--richness relation.}
\label{fig:1D_CF}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.49\textwidth]{Images/2D_CF_s0}
\includegraphics[width=0.453\textwidth]{Images/2D_CF_s1}
\caption{2D correlation function $\xi(r_{los},\,r_{\perp})$ for two redshift samples: low-z (left panels) and high-z (right panels), and three richness bins: $\lambda \in [22, 30)$ (bottom), $[30, 45)$ (middle), and $[45, 200)$ (top). The model results derived in Sec.~\ref{sect:likelihood} from the MXXL simulations are shown for comparison. Dashed contours correspond to values $\xi(r_{los},\,r_{\perp}) = (0.0, 1.5, 3.0, 4.5)$. For clarity, a 2D gaussian smoothing with a kernel of width $5 \,h^{-1}{\rm Mpc}$ has been applied to the images. Some significant differences are apparent here in the redshift direction given peculiar motions of the member galaxies that define the observed cluster redshifts.}
\label{fig:2D_CF}
\end{figure*}
Although the measurements of the quadrupole are too noisy to obtain any information from them, as we can see there is a clear increase in the amplitude of both the monopole and the projected correlation function for higher richness bins, following the behaviour found in the MXXL simulations. In each of these three subsamples, there is no clear evidence of any evolution between the two redshift bins, which, as we will see below is in good agreement with the MXXL simulations for the relatively small redshift range of the data. Some differences are apparent here in the redshift direction in Fig.~\ref{fig:2D_CF} with enhancement along the line of of sight compared to MXXL, which could be due to the higher than expected peculiar motion of the CGs used to define the observed cluster redshifts.
\subsubsection{Correlation length}
In order to fit the real-space correlation function $\xi(r)$, we approximate it by a power law:
\begin{equation}
\xi(r)= \left(\frac{r}{r_0}\right)^{-\gamma}\,,
\end{equation}
where both the correlation length $r_0$ and the slope $\gamma$ are left as free parameters. Considering this power law, the projected correlation function is:
\begin{equation}
\begin{split}
\Xi(r_\perp) & =2\int_0^\infty \xi\left[\left(r_\perp^2+r_{los}^2\right)^{1/2}\right]\,dr_{los} \\
& = 2\int_{r_\perp}^\infty \xi(r)\left(r^2-r_\perp^2\right)^{-1/2}\,dr \\
& = \sqrt{\pi}\,\frac{\Gamma\left(\left(\gamma-1\right)/\,2\right)}{\Gamma\left(\gamma\,/\,2\right)}\,r_0^\gamma\,r_\perp^{1-\gamma}\,.
\end{split}
\end{equation}
The values found for both $r_0$ and $\gamma$ for each of the six redMaPPer subsamples considered before are given in Table~\ref{tab:redMaPPer_CF}. The redshift and richness dependency of these results is shown in Fig.~\ref{fig:Projected_CF_Fit_vs_z}, in comparison with the MXXL model that adopts the ``clustering'' mass--richness relation parameters described later in Sec.~\ref{sect:likelihood}.
\begin{table}
\center
\caption{
\label{tab:redMaPPer_CF}
Values of the correlation length $r_0$ and the real-space correlation function slope $\gamma$ obtained for the six redMaPPer cluster subsamples considered in Table~\ref{tab:redMaPPer_Subsamples}.
}
\begin{tabular}{l c c}
\hline
\hline
Sample &$r_0\,[\,h^{-1}{\rm Mpc}]$ &$\gamma$ \\
\hline
$\lambda_1$ low-z &$14.53 \pm 1.20$ &$2.04 \pm 0.18$\\
$\lambda_1$ high-z &$15.58 \pm 0.61$ &$2.16 \pm 0.11$\\
\vspace{-5pt}\\
$\lambda_2$ low-z &$17.32 \pm 0.90$ &$2.26 \pm 0.19$\\
$\lambda_2$ high-z &$17.74 \pm 0.51$ &$2.34 \pm 0.10$\\
\vspace{-5pt}\\
$\lambda_3$ low-z &$23.05 \pm 1.08$ &$2.55 \pm 0.20$\\
$\lambda_3$ high-z &$22.19 \pm 0.65$ &$2.52 \pm 0.13$\\
\hline
\end{tabular}
\end{table}
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/Projected_CF_Fit_vs_z}}
\caption{Values of the correlation length $r_0$ obtained for the six redMaPPer subsamples listed in Table~\ref{tab:redMaPPer_Subsamples} as a function of the average redshift of the sample. Within two redshift regions, three richness ranges have been considered: $\lambda \in [22, 30)$ (red squares), $[30, 45)$ (blue triangles), and $[45, 200)$ (green circles). Shown as dotted lines, the model values of $r_0$ that we obtain from the MXXL simulations for these three richness ranges (from bottom to top, respectively) and the snapshots available. The slow increasing trend with redshift that is apparent here for all three richness ranges is well matched by MXXL and corresponds to the increasing bias at fixed mass with redshift, corresponding to rarer more biased peaks in the density field (Kaiser 1986). The mass--richness relation needed to obtain the MXXL model curves is obtained through a likelihood analysis, as described later in Sec.~\ref{sect:likelihood}.}
\label{fig:Projected_CF_Fit_vs_z}
\end{figure}
The values of $r_0$ for the cluster subsamples considered agree, within the noise, with the MXXL simulations expected values, with the only exception of the $\lambda_3$ high-z subsample, which is slightly more than 1$\sigma$ below the expect value. This could be an indication of the limitations of the redMaPPer algorithm above $z>0.35$, where it may be overestimating the richness of some clusters and thus diluting the amplitude of the correlation function.
Now we make the same measurement for these three richness subsamples for the
conservative redshift region $0.080\leq z_{\rm clu}\leq0.325$ where the upper redshift limit is
defined by the careful redMaPPer analysis as the limit of their volume complete region for clusters with richness $\lambda > 20$. The results obtained are shown in Fig.~\ref{fig:Projected_CF_Fit_vs_R}, where the relation between correlation length and richness, that we already noticed in Fig.~\ref{fig:1D_CF}, with an obvious rising trend that is very well fitted by MXXL, reflecting the enhanced bias expected for more massive clusters formed in a Gaussian random field \citep{Kaiser1986}.
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/Projected_CF_Fit_vs_R}}
\caption{The correlation length $r_0$ obtained from the projected correlation function (red squares) for three redMaPPer richness subsamples, $\lambda \in [22, 30)$, $[30, 45)$, and $[45, 200)$, in the redshift range $0.080\leq z_{\rm clu}\leq0.325$. The dotted black line is the equivalent MXXL correlation length points in the same richness ranges, once the mass--richness relation obtained later in Sec.~\ref{sect:likelihood} is applied, showing consistency between the data and the MXXL simulations of $\Lambda{\rm CDM}$.}
\label{fig:Projected_CF_Fit_vs_R}
\end{figure}
From these measured values of the correlation length $r_0(\left<\lambda_1\right>)$, $r_0(\left<\lambda_2\right>)$ and $r_0(\left<\lambda_3\right>)$, where $\left<\lambda_1\right>$, $\left<\lambda_2\right>$, and $\left<\lambda_3\right>$ represent the average cluster richness of the three richness subsamples considered, we observe the following linear relation between the value of the richness $\lambda$ and the value of the correlation length $r_0$:
\begin{equation}
r_0(\lambda) = 9.87 \pm 0.17 + \left( 0.198 \pm 0.004 \right)\lambda\,.
\end{equation}
The data showing such a clear trend with richness rising at the rate that is consistent with the MXXL simulations is very clear evidence for the standard physical understanding of the formation of structure from a Gaussian random field under gravity.
\section{Cluster Abundances}
\label{sect:abundances}
We proceed now to study the comoving density of clusters as a function of their richness. For this analysis, we limit our sample to the redshift range $0.080\leq z_{\rm clu}\leq0.325$, where the redMaPPer team has established volume completeness. We can also see in Fig.~\ref{fig:redMaPPer_Redshift_Distribution} that this claim is supported by the way the numbers of clusters scale in proportion to the cosmological volume. Above $z=0.35$, the richness calculated by the redMaPPer algorithm is increasingly limited to a diminishing proportion of relatively luminous galaxy members so that an uncertain estimate has to be made to take into account undetected galaxies below the survey magnitude limit, so that richness estimates become more noisy above this redshift; thus, we expect to obtain more robust results working with a cluster redshift cutoff of $z\leq0.325$.
The number of clusters $n_i$ that one may expect to find in the redshift range $[z_{\rm min}, z_{\rm max}]$, and within a richness range $[\lambda_i, \lambda_i+\Delta\lambda]$ is given by:
\begin{equation}
n_i = \Delta\Omega\int^{z_{\rm max}}_{z_{\rm min}}\int^{\lambda_i+\Delta\lambda}_{\lambda_i} dz\,d\lambda\,\frac{dV}{dz}\frac{dN(\lambda)}{dV\,d\lambda}\,,
\end{equation}
where $\Delta\Omega$ is the fraction of the sky covered by the survey, $dV/dz$ is the comoving volume per unit redshift, and $dN/dV/d\lambda$ is the theoretical cluster richness function.
\subsection{Results}
\subsubsection{Comoving densities}
To compute the cluster abundances, we average between several realizations of the redMaPPer richness distribution to take into account the effect on the number of clusters contained in each richness bin that the error on the algorithm richness estimate of each cluster (provided by the redMaPPer catalogue) can introduce. We take the mean values when convergence is obtained, and consider the standard deviation obtained from all these realizations an additional source of systematics, adding it in quadrature to the Poisson noise in each bin.
The comoving density of clusters $n(z)$ found within the redshift region previously mentioned is shown in Fig.~\ref{fig:redMaPPer_vs_MXXL_Comoving_Density} for three richness samples, where the previously mentioned $\lambda_1\in[22, 30)$, $\lambda_2\in[30, 45)$, and $\lambda_3\in[45, 200)$ binning has been applied, together with the MXXL model abundances.
\begin{figure*}
\centering
\includegraphics[width=0.98\textwidth]{Images/redMaPPer_vs_MXXL_Comoving_Density}
\caption{Cluster comoving densities $n(z)$ in the $0.080\leq z_{\rm clu}\leq0.325$ redshift range of three richness subsamples, $\lambda \in [22, 30)$ (left), $[30, 45)$ (middle), and $[45, 200)$ (right). The shaded region represents Poisson noise and errors introduced by the uncertainty in the richness measurement of each cluster, as given by the redMaPPer catalogue. Dotted lines represent the MXXL model density distributions in the same three richness ranges, once the optimal mass--richness relation obtained in Sec.~\ref{sect:likelihood} is used to obtain the synthetic cluster catalogue. A decline is apparent in the data, similar to the predictions of MXXL for this redshift range, corresponding to the expected growth of massive clusters over the past 3 Gyrs.}
\label{fig:redMaPPer_vs_MXXL_Comoving_Density}
\end{figure*}
The agreement between the data and the simulations is within the noise in this redshift range, showing a systematic decline of about 20\%. This is similar to the predictions of MXXL for this redshift range, corresponding to the expected growth of massive clusters over the past 3 Gyrs in the context of $\Lambda{\rm CDM}$.
\subsubsection{Richness function}
To compute the cluster richness function, we restrict ourselves to all the clusters in the redshift range $0.080\leq z_{\rm clu}\leq0.325$, and then we divide them in two redshift bins such that there is approximately equal numbers in each redshift bin. This results in two subsamples with $0.080\leq z_1<0.246$, and $0.246\leq z_2\leq0.325$, with mean redshifts $\left<z_1\right> = 0.186$ and $\left<z_2\right> = 0.287$, respectively. We then consider 10 log-spaced richness bins in the range $\lambda \in [22,\,200]$. The results obtained are shown in Fig.~\ref{fig:redMaPPer_Richness_Function_Evolution}. For comparison we also show the MXXL model results, interpolated to the mean redshifts of the data, $z=0.186$ and $z=0.287$.
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/redMaPPer_Richness_Function_Evolution}}
\caption{Cluster richness function of the redMaPPer clusters for two redshift subsamples: $0.080\leq z_1<0.246$ with $\left<z_1\right> = 0.186$ (red squares) and $0.246\leq z_2\leq0.325$ with $\left<z_2\right> = 0.287$ (blue circles). These two subsamples are divided into 10 log-spaced richness bins in the range $\lambda \in [22,\,200]$. Error bars include both Poisson noise and the errors on the measurement of the richness of each cluster provided by redMaPPer. A zoomed region $\lambda \in [22, 50]$ shows better the clear evolution with redshift relative to the small errors. The dashed red and dotted blue lines represent the richness functions predicted by the model, that accounts very well for the measured reduction in abundance with redshift. This model is derived in Sec.~\ref{sect:likelihood} from our likelihood analysis based on the MXXL simulations, and the best fit mass--richness relation obtained.
}
\label{fig:redMaPPer_Richness_Function_Evolution}
\end{figure}
It should be noted that, as described in the following section, no evolutionary information was used at all to calibrate the mass--richness relation applied to the MXXL simulations. Thus, the accurate agreement here in Fig.~\ref{fig:redMaPPer_Richness_Function_Evolution} of the data with the evolution predicted by MXXL shows the degree of consistency with the predictions of $\Lambda{\rm CDM}$. There was no guarantee that this comparison would reveal the same evolutionary trend in the cluster richness function.
\section{Likelihood Analysis}
\label{sect:likelihood}
We now obtain the mass--richness relation parameter values that best describe the observations through a likelihood analysis. We will be comparing the redMaPPer results with MXXL, covering a wide range of values for a power-law mass--richness relation, where, as we described in Sec.~\ref{sect:mxxl}, $\kappa_{M|\lambda}$ is the pivot mass normalization, $\alpha_{M|\lambda}$ is the slope, and the dispersion is given by $\sigma_{M|\lambda}$. As the assignment of richness given a value of the mass is an stochastic process, we generate several realizations of the synthetic cluster catalogues produced using the MXXL simulations (as we did above in Sec.~\ref{sect:abundances} with the redMaPPer catalogue) averaging the results until convergence is obtained. Note that in this analysis we restrain ourselves to the volume complete region, $0.080\leq z_{\rm clu}\leq0.325$.
For the pivot mass and for the slope we will consider uniform flat priors: $\kappa_{M|\lambda}\,\in\left[1.000,\,2.000 \right]$, and $\alpha_{M|\lambda}\,\in\left[0.500,\,1.600 \right]$. For the case of the scatter, a hard cutoff $\sigma_{M|\lambda}>0$ could bias the resulting posterior distribution, so we adopt the inverse-gamma distribution prior IG$(\epsilon, \epsilon)$ for $\sigma_{M|\lambda}^2$ \citep{Andreon2010, Sereno2015c, Umetsu2016} with $\epsilon$ a very small number (in our case, we take $\epsilon=10^{-3}$). We sample the posterior distribution using a large enough three-dimensional regular fine grid.
\subsection{Clustering}
\label{subsect:clustering_likelihood}
We first obtain an independent mass--richness relation through a likelihood analysis between the clustering data and MXXL. In this case, to obtain the values for $\kappa_{M|\lambda}$, $\alpha_{M|\lambda}$ and $\sigma_{M|\lambda}$ that best describe the clustering results, we rely on the projected correlation function $\Xi(r_\perp)$ of the whole sample, without any redshift nor richness binning. To make the joint analysis more consistent, we limit the sample to the volume complete region $0.080\leq z_{\rm clu}\leq0.325$, as we did in Sec.~\ref{sect:abundances}. We follow the procedure described in Sec.~\ref{sect:correlation} to obtain the points and the covariance matrix required for the likelihood analysis. The likelihood employed has the form:
\begin{equation}
\mathcal{L}(\bm{\Xi}|\bm{\theta},\,\mathcal{C})\propto \exp\left[-\frac{1}{2}\chi^2\left(\bm{\Xi},\,\bm{\theta},\,\mathcal{C}\right)\right]\,,
\end{equation}
where $\bm{\Xi}$ is the 8 dimensional data vector obtained from the cluster sample, $\bm{\theta}$ is the mass--richness relation vector $\bm{\theta}$ = ($\kappa_{M|\lambda}$, $\alpha_{M|\lambda}$, $\sigma_{M|\lambda}$), $\mathcal{C}$ is the $8\times8$ covariance matrix, and
\begin{equation}
\chi^2\left(\bm{\Xi},\,\bm{\theta},\,\mathcal{C}\right) = \left(\bm{\Xi}-\bm{\mu}(\bm{\theta})\right) \mathcal{C}^{-1} \left(\bm{\Xi}-\bm{\mu}(\bm{\theta})\right)^T\,,
\end{equation}
with $\bm{\mu}(\bm{\theta})$ the projected correlation function vector obtained from the MXXL simulations when the mass--richness with $\bm{\theta}$ is considered.
The projected correlation function obtained is shown in Fig.~\ref{fig:Best_Fit_Correlation_Function}, together with the MXXL realization with the mass--richness relation parameters given in Table~\ref{tab:Likelihood_Results}.
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/Best_Fit_Correlation_Function}}
\caption{Projected correlation function $\Xi(r_\perp)$ of the redMaPPer clusters for the full $0.080\leq z_{\rm clu}\leq0.325$ redshift range. The dashed green line is the expected correlation function derived for MXXL when we convert between mass and richness using the best fit mass--richness relation from the clustering data alone, clearly there is good self consistency. The blue dotted line represents an independent check using the abundance based mass--richness relation which lies a little below the data at the $2.5\sigma$ level.}
\label{fig:Best_Fit_Correlation_Function}
\end{figure}
Finally, we mention that these results are compatible with the ones obtained when we perform the same analysis independently in three richness ranges, the previously mentioned $\lambda_1\in[22, 30)$, $\lambda_2\in[30, 45)$, and $\lambda_3\in[45, 200)$ richness bins.
\subsection{Abundances}
\label{subsect:abundances_likelihood}
We now obtain the optimal mass--richness relation from the richness function. To compute it, we consider again all the clusters in the redshift range $0.080\leq z_{\rm clu}\leq0.325$, without any redshift binning, and distribute them in 10 log-spaced richness bins in the range $\lambda \in [22,\,200]$ (the maximum richness in this redshift range is found near $\sim200$). The median redshift of the sample considered, which comprises 8,152 clusters, is equal to $\left< z_{\rm clu} \right>=0.24$, and we select the same redshift bin corresponding to the appropriate time step outputted in the MXXL simulations.
To obtain the best fit and the associated confidence intervals, we follow the procedure derived by \cite{Cash1979} for Poisson statistics. To do so, we define the quantity $\mathcal{C}(\bm{n}|\bm{\theta}) = -2\ln \mathcal{L}(\bm{n}|\bm{\theta})$, where $\mathcal{L}(\bm{n}|\bm{\theta})$ is the likelihood function that depends on both the data vector $\bm{n}$, and the mass--richness relation $\bm{\theta}$. The deviations of $\mathcal{C}$ from the minimum follow a $\chi^2$ distribution, and in our case it is equal to:
\begin{equation}
\mathcal{C}(\bm{n}|\bm{\theta}) = -2\ln \mathcal{L}(\bm{n}|\bm{\theta}) = 2 \left(E(\bm{\theta}) - \sum_{i=1}^N n_i \ln e_i(\bm{\theta}) \right)\,,
\end{equation}
where $E(\bm{\theta})$ is the total number of clusters expected in all the $N$ bins once the mass--richness relation with $\bm{\theta}$ = ($\kappa_{M|\lambda}$, $\alpha_{M|\lambda}$, $\sigma_{M|\lambda}$) has been applied to the MXXL simulations, and $n_i$ and $e_i (\bm{\theta})$ are the observed and the expected number of clusters in the bin $i$, respectively.
The distribution we obtain, together with the MXXL richness function with the mass--richness relation given in Table~\ref{tab:Likelihood_Results}, is shown in Fig.~\ref{fig:Best_Fit_Richness_Function}. This measurement differs from the one described in Sec.~\ref{sect:abundances} and shown in Fig.~\ref{fig:redMaPPer_Richness_Function_Evolution} in that we are only considering one single redshift bin to improve our statistics. A very accurate fit to the data is found for MXXL with the mass--richness relation defined as a power-law, as described and shown in Fig.~\ref{fig:Best_Fit_Richness_Function}. Note that this fit is inherently more accurate than for the correlation function above for which the link between mass and richness is less direct than for abundances as discussed below in Sec.~\ref{sect:conclusions}.
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/Best_Fit_Richness_Function}}
\caption{Cluster richness function of the redMaPPer clusters for the full $0.080\leq z_{\rm clu}\leq0.325$ redshift range, obtained dividing the sample in 10 log-spaced richness bins in the range $\lambda \in [22,\,200]$. Error bars combine Poisson errors with the errors on the measurement of the richness of each cluster. The dotted blue and green dashed lines correspond to the richness function of the MXXL realization with the mass--richness relation parameters obtained from the abundances and clustering analysis (Table~\ref{tab:Likelihood_Results}), respectively.}
\label{fig:Best_Fit_Richness_Function}
\end{figure}
Again, the results obtained through this procedure are compatible with the ones obtained when we perform the same analysis in two redshift subsets of the catalogue with equal number of clusters contained in them, like we did in Sec.~\ref{sect:abundances}, with $0.080\leq z_1<0.246$ and $0.246\leq z_2\leq0.325$.
\subsection{Results and combined analysis}
The $1\sigma$ and the $2\sigma$ confidence regions of the $\kappa_{M|\lambda}$, $\alpha_{M|\lambda}$ and $\sigma_{M|\lambda}$ values, obtained through the abundance analysis and the joint analysis are shown in Fig.~\ref{fig:Likelihood_Contours}, and the marginalized posterior probabilities of each parameter are given in Fig.~\ref{fig:Likelihood_Marg_Probabilities}.
To compute the center (mean) and the scale (dispersion) of the marginalized 1D posterior distribution, we use the robust estimators described in \cite{Beers1990}. The results derived from the abundances analysis, the clustering analysis, and the combined analysis are listed in Table~\ref{tab:Likelihood_Results}.
\begin{table*}
\center
\caption{
\label{tab:Likelihood_Results}
The confidence values of the marginalized posteriors of the mass--richness relation parameters when the abundances, the clustering, and the combined abundances + clustering analyses are performed.
}
\begin{tabular}{c c c c}
\hline
Parameter &{\bf Abundances} &{\bf Clustering} &{\bf Abundances + Clustering}\\
\hline
\hline
$\kappa_{M|\lambda}$ &$1.351 \pm 0.039$ &$1.548 \pm 0.205$ &$1.341 \pm 0.031$\\
$\alpha_{M|\lambda}$ &$1.127 \pm 0.021$ &$1.102 \pm 0.197$ &$1.120 \pm 0.017$\\
$\sigma_{M|\lambda}$ &$0.194 \pm 0.053$ &$0.226 \pm 0.089$ &$0.164 \pm 0.039$\\
\hline
\end{tabular}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=0.98\textwidth]{Images/Likelihood_Contours}
\caption{Constraints on the mass--richness relation parameters $\kappa_{M|\lambda}$, $\alpha_{M|\lambda}$ and $\sigma_{M|\lambda}$ at $1\sigma$ and $2\sigma$ confidence levels when the abundances (blue) and the abundances + clustering (red) analysis are performed.}
\label{fig:Likelihood_Contours}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.98\textwidth]{Images/Likelihood_Marg_Probabilities}
\caption{Posterior probability distribution of the mass--richness relation parameters $\kappa_{M|\lambda}$, $\alpha_{M|\lambda}$ and $\sigma_{M|\lambda}$ marginalized over the other 2 parameters when the abundances (dashed blue) data only and the abundances + clustering (solid red) combined data are considered.}
\label{fig:Likelihood_Marg_Probabilities}
\end{figure*}
From our results, it is clear that the abundance analysis is much powerful in means of constraining the mass--richness relation, as the range of parameters that satisfy the clustering observations is much larger in comparison.
It should be noted, though, that both measurements provide very different mass--richness relations. If we compare how the MXXL synthetic catalogues drawn from both mass--richness relation perform in terms of agreeing with the data, we find that the abundances mass--richness relation (i.e., the one obtained in Sec.~\ref{subsect:abundances_likelihood}) is discrepant with the correlation function results at the $2.4\sigma$ level (Fig.~\ref{fig:Best_Fit_Correlation_Function}). On the other hand, when we apply the clustering mass--richness relation (i.e., the one obtained in Sec.~\ref{subsect:clustering_likelihood}) to the MXXL and measure the richness function, we find that there is a deviation of $8.1\sigma$ with respect to the best fit, making clear that the clustering mass--richness relation is not suitable to describe the cluster abundances found in the data.
In any case, when we combine both analyses and perform a joint likelihood, we obtain a concordance mass--richness relation that describes both measurements, with a pivot mass of:
\begin{equation}
\kappa_{M|\lambda} \equiv \ln \left( \frac{M_{200c}(\lambda_0=60)}{10^{14}\,M_\odot\,h^{-1}} \right) = 1.341 \pm 0.031\,,
\nonumber
\end{equation}
an slope equal to:
\begin{equation}
\alpha_{M|\lambda} = 1.120 \pm 0.017\,,
\nonumber
\end{equation}
and a scatter:
\begin{equation}
\sigma_{M|\lambda} \equiv \Delta \ln \left( \frac{M_{200c}(\lambda)}{10^{14}\,M_\odot\,h^{-1}} \right) = 0.164 \pm 0.039\,.
\nonumber
\end{equation}
For comparison with \cite{Simet2016} results, and using \cite{Hu2003} recipe to convert between mass definitions, we can compute the mass obtained through this mass--richness relation at $\lambda=40$:
\begin{equation}
\log_{10} \left(M_{200m}(\lambda=40)\right)=(14.534 \pm 0.015) \,M_\odot\,h^{-1}\,.
\end{equation}
\section{Redshift Enhancement}
\label{sect:redshift_enhancement}
Finally, as an independent consistency check of the mass--richness relation obtained, we measure the redshift enhancement effect produced by these redMaPPer clusters as a function of their richness, and compare it with the model drawn from such mass--richness relation. This effect is caused by a cluster acting as a gravitational lens with shear $\gamma(r)$ and convergence $\kappa(r)=\Sigma(r)/\Sigma_{\rm crit}$, where $\Sigma(r)$ is the projected mass of the cluster as a function of distance from the center, and $\Sigma_{\rm crit}$ is the critical surface density defined as:
\begin{equation}
\Sigma_{\rm crit} = \frac{c^2}{4\pi G}\frac{D_s}{D_l\,D_{ls}}\,,
\end{equation}
with $D_s$, $D_l$ and $D_{ls}$ the angular-diameter distances between the observer and the source, the observer and the lens, and the source and the lens, respectively. So, for a lens with a given shear and a convergence, the magnification produced by it is equal to:
\begin{equation}
\mu = \frac{1}{\left(1-\kappa \right)^2-\left|\gamma\right|^2}\,.
\end{equation}
This magnification leads to the so called {\it magnification bias} \citep{Broadhurst1995}, as the number of background lensed sources is modified by a combination of two effects: the luminosity limit of the survey increasing from $L_{\rm lim}$ to $L_{\rm lim}/\mu$ in the lens region, as the fluxes from the sources are amplified, and the number of objects behind the lens decreasing, as sky area behind it is expanded or dilated by a factor $1/\mu$. Thus, the original number of objects observed, $n_0$, is now modified to:
\begin{equation}
n_{\rm obs}\left[L_{\rm lim}(z)\right]=\frac{1}{\mu}n_0\left[>\frac{L_{\rm lim}(z)}{\mu}\right]\,.
\end{equation}
So, this modification on the number of observed objects, $n_{\rm obs}$, results in an enhancement of the average redshift of these same background objects, given by:
\begin{equation}
\overline{z}_{\rm back}=\frac{\int n_{\rm obs} (z)\,z\,dz}{\int n_{\rm obs} (z)\,dz}\,.
\end{equation}
So, we define the redshift enhancement $\delta_z$ as:
\begin{equation}\label{eq:redshift_enhancement}
\delta_z(r)\equiv\frac{\overline{z}(r)-\overline{z}_{\rm total}}{\overline{z}_{\rm total}}\,,
\end{equation}
where $\overline{z}(r)$ is the measured average redshift of the lensed $n(r)$ background objects inside a radial bin at a distance $r$ from the center of the lens, and $\overline{z}_{\rm total}$ is the average redshift of the unlensed background objects, as would be seen without a lens. We refer the reader to \cite{Coupon2013} and \cite{Jimeno2015} to a more in-depth description of this measurement. To model this effect, we will also adopt the same assumptions that were considered by them, that is, a NFW profile \cite{Navarro1996}, the mass-concentration relation of \cite{Bhattacharya2013}, and the Schechter parametrization of the V-band luminosity function as given by \cite{Ilbert2005}. Once the redshift enhancement model as a function of cluster mass is obtained, we use the mass--richness relation obtained in Sec.~\ref{sect:likelihood} to convert it into a function of cluster richness.
\subsection{Results}
We proceed now to measure the integrated redshift enhancement for three different richness cluster subsamples $\lambda_1\in[22, 30)$, $\lambda_2\in[30, 45)$, and $\lambda_3\in[45, 200)$ at redshifts below $z<0.400$. To do so, we first measure the average redshift $\overline{z}_{\rm total}$ of all the CMASS galaxies above $z>0.425$, to ensure there is a gap between the clusters and them, and then measure the average redshift of all those background galaxies that lie in the $r_\perp < 0.3 \,h^{-1}{\rm Mpc}$ region, where $r_\perp$ is the transverse distance to the CG of the redMaPPer cluster. The redshift enhancement signal $\delta_z$ is then obtained from Eq.~\ref{eq:redshift_enhancement}.
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/Redshift_Enhancement}}
\caption{Integrated redshift enhancement signal in the $r_\perp < 0.3 \,h^{-1}{\rm Mpc}$ region for three richness redMaPPer subsamples with $\lambda\in[22, 30)$, $[30, 45)$, and $[45, 200)$, in the redshift range $z_{\rm clu}<0.400$. The model prediction as a function of richness is obtained through the mass--richness relation of the joint analysis, as given by Table~\ref{tab:Likelihood_Results}.}
\label{fig:Redshift_Enhancement}
\end{figure}
Finally, for each richness subsample we generate 200 random samples with the same number of clusters, distributed in the same volume and following their same redshift distribution. Then, we measure the integrated redshift enhancement in each of them, and then compute the standard deviation of these 200 measurements to have an estimation of the errors associated to this measurement.
The redshift enhancement signal $\delta_z$ obtained for these three richness subsamples is shown in Fig.~\ref{fig:Redshift_Enhancement}, together with the lensing model whose dependency with richness is obtained from converting mass into richness using the mass--richness relation of the combined analysis, as given in Table~\ref{tab:Likelihood_Results}.
These results are a significant improvement with respect to previous measurements \citep{Coupon2013, Jimeno2015} because of the increase in the numbers of clusters and background galaxies with measured redshifts. The redshift enhancement effect is detected with a significance of $6.6\sigma$ for the whole sample. Also, we observe a clearer trend of higher values of $\delta_z$ for larger values of the average richness of the sample considered, as expected. It should be noted, though, that all the values of $\delta_z$ fall below the model, with a deviation of $\sim1\sigma$ for the richness subsamples $\lambda_1$ and $\lambda_3$. This could be an indication of the level of miscentering present in the position of the CGs with respect to the deepest part of the gravitational potential well, which would be the main cause of the dilution of the redshift enhancement signal.
\section{Discussion and Conclusions}
\label{sect:conclusions}
From the original redMaPPer catalogue and the latest DR12 spectroscopy, we have created a sample of $\sim$ 25,000 clusters with spectroscopic redshifts, and studied the cluster correlation function and the cluster abundances as a function of their optical richness, in the volume complete region $0.080\leqslant z \leqslant 0.350$.
For the correlation function calculation, we have used the centering probabilities of the candidate central galaxies, that the redMaPPer catalogue provides, as an additional weight to obtain a clearer signal. We detect a significant increase of the amplitude of both the monopole and the projected correlation function for higher average richness subsamples, but we do not notice any redshift dependency when those richness subsamples are split into different redshift bins.
On the other hand, when we measure the cluster comoving density on the range $0.080 \leqslant z \leqslant 0.325$, and measure the richness function for two redshift slices, $z_1=0.186$ and $z_2=0.287$, we do detect a clear continuous evolution of the abundances of clusters, which is in excellent agreement with the MXXL simulations.
We have compared our measurements with a synthetic cluster catalogue that we have created from the MXXL simulation, one of the largest cosmological simulations available, that allows us to mimic the distribution, miscentering and peculiar velocities of central galaxies within clusters. To do so, we have assumed the usual power-law mass--richness relation to convert the masses of the DM haloes found in the simulations into richness. Assuming a $\Lambda{\rm CDM}$ {\em Planck} cosmology, we find that the best agreement with the data is obtained for a mass--richness relation with a pivot mass $\ln \left( M_{200c}(\lambda_0=60)/10^{14}\,M_\odot\,h^{-1} \right) = 1.333\pm0.034$, a slope $\alpha_{M|\lambda} = 1.118\pm0.017$, and a scatter $\sigma_{M|\lambda}=0.158\pm0.044$.
The MXXL fits to the data are remarkably good in all respects as a function of richness, redshift and separation with no clear discrepancy visible in each of these plots, providing very strong confirmation of the detailed viability of the $\Lambda{\rm CDM}$ model. However, we do notice that the amplitude of the correlation function is slightly higher than expected when only the abundances are used to constrain the mass--richness relation, with a deviation of $\simeq 2.5\sigma$. The clustering measurements are described by a large range of mass--richness relation parameters, many of them agreeing with the abundance analysis, but the inherent precision is lower for defining the mass--richness relation via clustering. This possible tension between the abundance defined mass--richness relation and the amplitude of the observed correlation function has motivated us to make the same consistency check with the former MXXL and Millennium Simulation WMAP consensus values of $\sigma_8$, $\Omega_m$ and $H_0$ because they were until recently significantly different from the most recent {\em Planck} weighted values and these parameters are the most important for predicting the abundances of clusters and their correlation function. With these former values we find a completely acceptable consistency between the correlation function and the abundances. In Fig.~\ref{fig:Cosmo_Comparison} we show the comparison between the results obtained from both cosmologies when we use the mass--richness relation obtained from the abundance matching technique in both cosmologies. This may prove very interesting because of the increased support for $H_0=73$ km s$^{-1}$ from the independent local distance ladder measurements \citep{Riess2016} that adds support to previous claims for this value at somewhat lower significance \citep{Freedman2001} and from independent lensing time delay estimates of $H_0$ \citep{Bonvin2016}.
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/Cosmo_Comparison}}
\caption{The projected correlation function compared with MXXL derived self consistently for two different cosmologies when the mass--richness obtained from the abundance matching method is used: as the green dotted line, the former WMAP cosmology with $\Omega_m=0.25$ and $\sigma_8=0.9$, and as the blue dashed line, the {\em Planck} cosmology with $\Omega_m=0.309$ and $\sigma_8=0.816$. The WMAP cosmology results are much more in agreement with the observations than the values obtained when the {\em Planck} cosmology is considered.}
\label{fig:Cosmo_Comparison}
\end{figure}
Finally, we measured the redshift enhancement effect produced by the gravitational magnification of the redMaPPer clusters on the background CMASS subsample of BOSS galaxies. We found once again a clear relation between the richness and the amplitude of the gravitational lens magnification effect, with consistency at the $1.1\sigma$ level, a little below the model prediction, which could be produced by the intrinsic miscentering of central galaxies with respect to the center of the projected gravitational potential well and will be examined more thoroughly in the future.
The J-PAS survey of the North now coming on line \citep{Benitez2014} will provide lensing masses and redshifts for all SDSS clusters and will go beyond in redshift, measuring unprecedentedly accurate growth as a function of cluster mass in the range $z \lesssim 1.0$ (Fig.~\ref{fig:MXXL_Richness_Function_Evolution}), with the prospect of constraining the total relativistic species contribution to $\Omega_m$, and obtaining a clearer insight into the emerging tensions between the parameters describing the standard $\Lambda{\rm CDM}$ model.
\begin{figure}
\resizebox{84mm}{!}{\includegraphics{Images/MXXL_Richness_Function_Evolution}}
\caption{The model curves show the MXXL richness function for snapshots with redshift (from top to bottom): 0.027, 0.083, 0.154, 0.242, 0.351, 0.486, 0.652, 0.857, 1.110, 1.424. The mass--richness relation of the joint analysis, as given by Table~\ref{tab:Likelihood_Results}, has been used to convert between mass and richness. For comparison, the cluster richness function of the redMaPPer clusters in the range $22 < \lambda < 100$ and for two redshift subsamples are shown: $0.080\leq z_1<0.246$ with $\left<z_1\right> = 0.186$ (red squares) and $0.246\leq z_2\leq0.325$ with $\left<z_2\right> = 0.287$ (blue circles), illustrating the enormous potential of the upcoming deeper imaging surveys for evaluating the growth of structure.}
\label{fig:MXXL_Richness_Function_Evolution}
\end{figure}
\hspace{0pt}\\ \textbf{Acknowledgements}\\
TJB is supported by IKERBASQUE, the Basque Foundation for Science.
RL is supported by the Spanish Ministry of Economy and Competitiveness through research projects FIS2010-15492 and Consolider EPI CSD2010-00064, and the University of the Basque Country UPV/EHU under program UFI 11/55.
PJ acknowledges financial support from the Basque Government grant BFI-2012-349.
TJB, RL and PJ are also supported by the Basque Government grant for the GIC IT956-16 research group.
REA acknowledges support from AYA2015-66211-C2-2.
J.M.D acknowledges support of the projects AYA2015-64508-P (MINECO/FEDER, UE), AYA2012-39475-C02-01 and the consolider project CSD2010-00064 funded by the Ministerio de Economia y Competitividad.
KU acknowledges partial support from the Ministry of Science and Technology of Taiwan (grants MOST 103-2112-M-001-030-MY3 and MOST 103-2112-M-001-003-MY3).
Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/. SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III Collaboration including the
University of Arizona,
the Brazilian Participation Group,
Brookhaven National Laboratory,
University of Cambridge,
Carnegie Mellon University,
University of Florida,
the French Participation Group,
the German Participation Group,
Harvard University,
the Instituto de Astrofisica de Canarias,
the Michigan State/Notre Dame/JINA Participation Group,
Johns Hopkins University,
Lawrence Berkeley National Laboratory,
Max Planck Institute for Astrophysics,
Max Planck Institute for Extraterrestrial Physics,
New Mexico State University,
New York University,
Ohio State University,
Pennsylvania State University,
University of Portsmouth,
Princeton University,
the Spanish Participation Group,
University of Tokyo,
University of Utah,
Vanderbilt University,
University of Virginia,
University of Washington,
and Yale University.
|
\chapter{Acknowledgements}
First of all, I would like to thank my thesis advisor, Tam\'as Hausel, who made all this possible by pointing out the project to me, supervising my activity through these years and forming a lively geometry group in which I grew a lot as a person and as a mathematician.
I would also like to thank our beloved Pierrette, that helped and took care of me even before my official start at EPFL. She always says that it is her job, but one can notice the passion, kindness and care for others that she provides every day.
I have a huge debt of gratitude towards Daniele Boccalini, Alexandre Peyrot, Dimitri Wyss and Martina Rovelli, my closest "academic relatives" and good friends, for providing countless hours of fun, company and mathematical discussions as well.
A big thank you goes to all present and former members of the Geometry Group: Michael Groechenig, Michael Lennox Wong, Ben Davison, Martin Mereb, Zongbin Chen, Alexander Noll. I was very lucky to share this time with all of you and thank you for your useful suggestions.
Many visitors came to Lausanne to exchange thoughts on research for brief periods of time. Among those, I would like to thank Luca Migliorini and Jochen Heinloth for taking an interest in my work and answering many of my questions. Another big thank you for accepting to review my final work. I would like to thank Kathryn Hess and Donna Testerman as well for being so kind to spend their time being part of the jury for my PhD defense.
Last but not least I would like to thank my friends and family whose moral support was fundamental through these years.
During these years, my research was supported by the Project Funding number 144119: \emph{Arithmetic harmonic analysis on Higgs, character and quiver varieties} issued by the Swiss National Science Foundation.
\chapter{Bradlow-Higgs triples}
\section{Definitions, properties and basic results}
Consider a smooth projective curve $C$ over $\mathbb{C}$ of genus $g\geq 2$ and the total space $X=T^*C$ of $K$, the canonical bundle of $C$. Let us also denote by $\P X$ the projectivization of $X$ obtained by adding the divisor at infinity i.e. $\P X= \P(\mathcal{O} \oplus K)$.
We want to study moduli spaces of objects that are a variant of both Higgs bundles and Bradlow pairs.
\begin{defn}[Bradlow-Higgs triples and $\sigma$-stability]
\label{sigmast}
A triple $(E,\phi,s)$ is said to be a \emph{Bradlow-Higgs triple} if $(E,\phi)$ is a Higgs bundle and $s \in H^0(C,E)$ is a nonzero section of the underlying vector bundle $E$.
Let $\sigma$ be a positive real number. We say $(E,\phi,s)$ is \emph{$\sigma$-(semi)stable} if, for all proper $\phi$-invariant subbundles $F \subset E$, we have:
\begin{align*}
\frac{\deg(F)}{\rk(F)} \operatornamewithlimits{<}_{(=)} \frac{\deg(E)+\sigma}{\rk(E)} \qquad & \text{if } s \notin H^0(C, F)\\
\frac{\deg(F)+\sigma}{\rk(F)} \operatornamewithlimits{<}_{(=)} \frac{\deg(E)+\sigma}{\rk(E)} \qquad & \text{if } s \in H^0(C, F)
\end{align*}
A morphism of Bradlow-Higgs triples is a morphism of the underlying Higgs bundles that sends the section of the source to a scalar multiple of the section of the target.
\end{defn}
The idea is to introduce a section of the underlying vector bundle so that the stability condition is allowed to vary together with the positive real parameter $\sigma$. This will produce more complicated moduli spaces but will also allow to study the relation among them as $\sigma$ varies. As we will see later (and as already seen about Bradlow pairs) there are two interesting extremal stability conditions, one of which relates the moduli space of Bradlow-Higgs triples with the moduli space of semistable Higgs bundles, and the other one relates the moduli space of Bradlow-Higgs triples with a relative Hilbert scheme of points.
Here is a basic but fundamental result on Bradlow-Higgs triples.
\begin{prop}
Let $\sigma>0$ and $(E,\phi,s)$ be a $\sigma$-stable triple. Let $f \in H^0(\End E)$ such that $[f, \phi]=0$ and $f\cdot s=0$. Then $f=0$.\\
In particular, the only endomorphisms of a $\sigma$-stable triple $(E,\phi,s)$ are scalar multiples of the identity.
\begin{proof}
Let $L$ denote the kernel of $f$ and suppose that $L \neq E$. Note that $L \neq 0$ because $s \in H^0(L)$. Then $L$ has rank $r' < r=\rk E$ and is $\phi$-invariant containing the section $s$, therefore we have $\mu(L) \leq d/r+\sigma(1/r-1/r')$. In particular $f$ induces an injective map $E/L \rightarrow E$ so that $E/L$ can be thought as a sub sheaf of $E$. Since it is $\phi$ invariant, its saturation will be $\phi$ invariant as well and $\mu( E/L) \geq d/r + \sigma/r$. This means that the saturation of $E/L$ will be $\sigma$-destabilizing unless $f=0$.
The second part follows by noting that if $f$ is an endomorphism of $(E,\phi,s)$ and $f(s)=\alpha s$ then $f-\alpha \mathbb{I}_E$ satisfies the hypotheses of the first part of the proposition and therefore is $0$.
\end{proof}
\end{prop}
\subsection{BNR correspondence and moduli space of coherent systems on a surface}
\label{secbnr}
Recall from section \ref{higgs} that due to the BNR correspondence, pure one dimensional sheaves on $X$ that do not intersect the divisor at infinity correspond to Higgs bundles on $C$ through the pushforward along the projection map $\pi: T^*C \rightarrow C$. The pushforward also canonically identifies the sections of such sheaves with the sections of their pushforward, i.e. the underlying vector bundle of the Higgs bundle obtained by applying $\pi_*$. We can therefore alternatively work with pairs $(\mathcal{F},s)$ on $X$, where $\mathcal{F}$ is a rank one pure one dimensional sheaf of degree $d'$ whose support $\supp (\mathcal{F})$ does not intersect the divisor at infinity and $s$ is a non-zero section in $H^0(X,\mathcal{F})$.
These objects are known as a particular case of (one dimensional) \emph{coherent systems} on the surface $\P X$. The moduli spaces of these objects were defined and studied in \cite{le1995faisceaux} (and in a slightly different context in \cite{pandharipande2009curve}). It is worth recollecting some of the results we need.
First we recall the definition of the stability condition \cite[d\'{e}finition 4.4]{le1995faisceaux} and \cite[section 1.1]{pandharipande2009curve}.
\begin{defn}
Fix a polarization $L$ on $\P X$ and denote by $\mathcal{F}(k)=\mathcal{F} \otimes L^k$ for $\mathcal{F}$ a sheaf on $\P X$. Let $\sigma$ be a positive real number, we say $(\mathcal{F},s)$ is \emph{$\sigma$-(semi)stable} if:
\begin{align*}
\frac{\chi(\mathcal{G}(k))}{a(\mathcal{G})} \operatornamewithlimits{<}_{(=)} \frac{\chi(\mathcal{F}(k))+\sigma}{a(\mathcal{F})} \qquad & \text{if } s \notin H^0(\P X, \mathcal{G})\\
\frac{\chi(\mathcal{G}(k))+\sigma}{a(\mathcal{G})} \operatornamewithlimits{<}_{(=)} \frac{\chi(\mathcal{F}(k))+\sigma}{a(\mathcal{F})} \qquad & \text{if } s \in H^0(\P X, \mathcal{G})
\end{align*}
for all proper subsheaves $\mathcal{G} \subset \mathcal{F}$. Both conditions on the Hilbert polynomials are meant to hold for big values of $k$.
Note that, since $\mathcal{F}$ is pure one dimensional, both Hilbert polynomials of $\mathcal{F}$ and $\mathcal{G}$ have degree 1. Here we denote by $a(\cdot)$ the corresponding leading terms and by $b(\cdot)$ the constant terms.
\end{defn}
Observe that, as explained in \cite[section 1.1]{pandharipande2009curve}, $\chi(\mathcal{F}(k))=a(\mathcal{F})k+b(\mathcal{F})$ with:
\begin{align*}
&a(\mathcal{F})=\int_\beta c_1(L)\\
&b(\mathcal{F})=\chi(\mathcal{F})
\end{align*}
with $\beta \in H_2(\P X)$ the class of the support of $\mathcal{F}$.\\
After this we can define moduli spaces of $\sigma$-(semi)stable coherent systems:
\begin{defn}
Fix $m$ an integer and the homology class $\beta=n[C] \in H_2(\P X)$ of the support for our sheaves (here $C$ embeds in $\P X$ as the zero section of $T^*C$).\\
We denote by $\text{Syst}_{\P X,\sigma}(\beta,m,1)$ the moduli space of $\sigma$-(semi)stable pure one dimensional rank one degree $d'$ pairs $(\mathcal{F},s)$ with $\chi(\mathcal{F})=m$ and support in class $\beta$ (compare with \cite[th\'{e}or\`eme 4.11]{le1995faisceaux}).
\end{defn}
The $1$ in the definition of $\text{Syst}_{\P X,\sigma}(\beta,m,1)$ denotes that we only have one section $s$ of the underlying sheaf.\\
From \cite[th\'{e}or\`eme 4.11]{le1995faisceaux} we deduce that the $\text{Syst}_{\P X,\sigma}(\beta,m,1)$ are projective varieties, since $\P X$ is projective.
In order to obtain Bradlow-Higgs triples from pairs on $\P X$ we need the extra condition that the support of the underlying sheaf does not intersect the divisor at infinity. Once we identify the stability condition for coherent systems with the stability condition for Bradlow-Higgs triples we can identify the moduli spaces of $\sigma$-(semi)stable Bradlow-Higgs triples with the open subset of $\text{Syst}_{\P X,\sigma}(\beta,m,1)$ of coherent systems whose support does not intersect the divisor at infinity of $\P X$.
\begin{prop}
Let $(\mathcal{F},s)$ be a coherent system (with one section) such that $\supp(\mathcal{F})$ does not intersect the divisor at infinity. Assume $\chi(\mathcal{F})=m$ and the class of the support of $\mathcal{F}$ is $\beta=n[C]$, where $C$ is the zero section of $\P X$. Denote by $(E,\phi,s)$ the Bradlow-Higgs triple obtained by pushing forward $(\mathcal{F},s) $ along $\P X \rightarrow C$. Then $\deg E=m+n(g-1)$ and $\rk E=n$. Furthermore, the stability condition for $(\mathcal{F},s)$ is the same as the stability condition for $(E,\phi,s)$, in particular this means that the stability condition does not depend on the polarization on $\P X$ as long as the hypotheses on the support are satisfied.
\begin{proof}
Since $\supp(\mathcal{F})$ does not intersect the divisor at infinity it is a branched $n$ to $1$ cover of $C$ embedded as the zero section of $\P X$. More precisely there exists an affine open subset $U=\Spec A$ of $C$ over which $T^*C$ trivializes and $T^*C_{|U}=\Spec A[t]$. Then $\supp(\mathcal{F}) \cap T^*C_{|U}$ is defined by a degree $n$ polynomial in $A[t]$. Therefore if $\mathcal{F}$ restricts to the rank one $A[t]$-module $M$ on $\Spec A[t]$, its pushforward is $M$ seen as an $A$-module and therefore it will be isomorphic to $M \oplus M t \oplus \dots \oplus M t^{n-1}$ so that it will have rank $n$.\\
Under our hypotheses:
\begin{align*}
&a(\mathcal{F})=\int_\beta c_1(L)=n \int_\beta c_1(L) >0\\
&b(\mathcal{F})=\chi(\mathcal{F})=m
\end{align*}
therefore $\deg E+n(1-g)=\chi(E)=\chi(\mathcal{F})=m$. Let $A=\int_\beta c_1(L)>0$.\\
For the stability condition let $\mathcal{G}$ be a subsheaf of $\mathcal{F}$ and call $G$ its pushforward, which is a $\phi$-invariant subbundle of rank $0<\rk G < n$. Note that the class of the support of $\mathcal{G}$ is then $\rk G \cdot A$. Therefore we have:
\begin{align*}
&\frac{\chi(\mathcal{G}(k))}{a(\mathcal{G})}=k+\frac{\deg G}{\rk G \cdot A}+(1-g),\\
&\frac{\chi(\mathcal{G}(k))+\sigma}{a(\mathcal{G})}=k+\frac{\deg G+ \sigma}{\rk G \cdot A}+(1-g),\\
&\frac{\chi(\mathcal{F}(k))}{a(\mathcal{F})}=k+\frac{\deg E}{\rk E \cdot A}+(1-g) \text{ and}\\
&\frac{\chi(\mathcal{F}(k))+\sigma}{a(\mathcal{F})}=k+\frac{\deg E+ \sigma}{\rk E \cdot A}+(1-g).
\end{align*}
Thus we see that the inequalities for stability are equivalent.
\end{proof}
\end{prop}
Therefore we can give the following definition, using the identification above we know that such a moduli space actually exists and is also a quasi-projective variety.
\begin{defn}
Fix integers $r \geq 1$ and $d$. We denote by $\mathcal{M}_\sigma^{r,d}$ the moduli space of S-equivalence classes of $\sigma$-(semi)stable Bradlow Higgs triples of degree $d$ and rank $r$.
\end{defn}
As we saw $\mathcal{M}_\sigma^{r,d}$ corresponds to an open subset of $\text{Syst}_{\P X,\sigma}(r[C],d+r(1-g),1)$. We conclude with an
\begin{obs}
\label{degs}
The sheaves in $\text{Syst}_{\P X,\sigma}(r[C],d+r(1-g),1)$ can also be regarded as sheaves that are rank one and degree $d'$ on their support. So if their support is a smooth projective curve of genus $g'$ we have $d+r(1-g)=\chi(E)=\chi(\mathcal{F}_{\supp \mathcal{F}})=d'+1-g'$. If we further observe that for the curves above the genus is $g'=1+r^2(g-1)$ (section \ref{higgs} of the introductory chapter) we have that $d'=d+(r^2-r)(g-1)$.
\end{obs}
\subsection{Hitchin maps}
\label{hitchinmaps}
An analogy with the moduli space of stable Higgs bundles is the existence, for every $\sigma$, of a Hitchin map, defined in the analogous way as for Higgs bundles.
\begin{defn}
Let $\mathcal{A}^r$ be the Hitchin base. Then for every $\sigma>0$ we can define the \emph{Hitchin map}:
\begin{align*}
\chi^{r,d}_\sigma: \mathcal{M}_\sigma^{r,d} & \rightarrow \mathcal{A}^r\\
(E,\phi,s) & \mapsto \text{char poly}(\phi).
\end{align*}
\end{defn}
As we already observed for Higgs bundles, the Hitchin map is the analogue, through the BNR correspondence, of the map taking $(\mathcal{F},s)$ to the equation of $\supp(\mathcal{F})$. Let us first make the following observation.
\begin{prop}
\label{opendense}
Let $\mathcal{A}^r_{int} \subset \mathcal{A}^r$ be the locus of integral spectral curves. Then the open subset $(\chi^{r,d}_\sigma)^{-1} (\mathcal{A}^r_{int})$ does not depend on $\sigma$.
\begin{proof}
This follows from what we already observed for Higgs bundles. If $(E,\phi,s) \in \mathcal{M}_\sigma^{r,d}$ and $E$ has a $\phi$-invariant subbundle, then the characteristic polynomial of $\phi$ will factor, and hence $\chi^{r,d}_\sigma(E,\phi,s) \in \mathcal{A}^r \setminus \mathcal{A}^r_{int}$. It follows that $(\chi^{r,d}_\sigma)^{-1} (\mathcal{A}^r_{int})$ will consist of triples for which the Higgs bundle does not have any subobject and therefore are automatically $\sigma$-stable for all $\sigma$.
\end{proof}
\end{prop}
\subsection{Extremal stability conditions}
\label{extrstab}
We are interested in the stability for extremal values of $\sigma$ and for which combinations of $r ,d , \sigma$ the moduli spaces are non-empty.\\
Let us first define a filtration that will be useful in the following chapters.
\begin{theorem}
Let $(E,\phi,s)$ be a Bradlow-Higgs triple. Then there exists a filtration $0 \subset U_1 \subset \dots \subset U_l \subseteq E$ with $l \leq r$ and $U_r=E$ in case equality holds. The filtration satisfies the following conditions:
\begin{itemize}
\item[(i)] $\rk U_i=i$,
\item[(ii)] $\phi(U_i) \subseteq U_{i+1} \otimes K$ for $1 \leq i <l$,
\item[(iii)] for $1\leq i < l$, $\phi$ induces a nonzero map $U_i/U_{i-1} \rightarrow U_{i+1}/U_i \otimes K$,
\item[(iv)] $U_l$ is the smallest subbundle of $E$ that is $\phi$-invariant and contains the section $s$.
\end{itemize}
Furthermore, we have a bound $\deg U_i \geq i(i-1)(1-g)$.
\begin{proof}
We define the filtration with an inductive process. In the following we will denote by $\langle \cdot \rangle$ the saturation of a subsheaf of a vector bundle.
First define $U_1$ to be the line subbundle of $E$ generated by the section $s$, then clearly $\deg U_1 \geq 0$ so the bound is satisfied and also if $\phi(U_1) \subseteq U_1\otimes K$ then we can stop the construction and the filtration $0 \subset U_1 \subset E$ satisfies the properties.
If $\phi(U_1) \nsubseteq U_1 \otimes K$, $\phi$ induces a nonzero map $U_1 \rightarrow E/U_1 \otimes K$. Take:
\begin{equation*}
S_2=\left \langle \im( \phi: U_1 \rightarrow E/U_1 \otimes K) \right\rangle \otimes K^{-1}.
\end{equation*}
Then $S_2$ is a subbundle of $E/U_1$ and hence there exists a unique subbundle $U_2$ of $E$ containing $U_1$ and such that $S_2$ is the image of $U_2$ with respect to the projection $E \rightarrow E/U_1$. By construction we have a bound $\deg U_2 \geq \deg U_1 +2-2g \geq 2-2g$ and also if $\phi(U_2) \subseteq U_2 \otimes K$ then the filtration $0 \subset U_1 \subset U_2 \subset E$ satisfies the properties.
The construction can be carried through inductively. Assume that we have constructed $0 \subset U_1 \subset U_2 \subset \dots \subset U_p$ satisfying $\rk U_i=i$, $\deg U_i \geq i(i-1)(1-g)$ and that there are nonzero maps $U_i/U_{i-1} \rightarrow U_{i+1}/U_i \otimes K$ for $1 \leq i < p$. Then we claim that $\phi(U_p) \nsubseteq U_p \otimes K$ if and only if $\phi^p(s) \notin U_p \otimes K^p$. To prove this observe that $\phi(U_p) \subseteq U_p \otimes K$ if and only if $\phi$ induces the zero map $U_p \rightarrow E/U_p \otimes K$. Since by construction $\phi(U_{p-1}) \subseteq U_p \otimes K$ then $U_p \rightarrow E/U_p \otimes K$ is zero if and only if $U_p/U_{p-1} \rightarrow E/U_p \otimes K$. Consider the iterated applications of $\phi$:
\begin{equation*}
U_1 \rightarrow U_2/U_1 \otimes K \rightarrow \dots \rightarrow U_p/U_{p-1} \otimes K^{p-1} \rightarrow E/U_p \otimes K^p.
\end{equation*}
From this we see that since the composition of the first $p-1$ applications is nonzero and all the quotients are rank one except for the last, then $U_p/U_{p-1} \otimes K^{p-1} \rightarrow E/U_p \otimes K^p$ is zero if and only if $U_1 \rightarrow E/U_p \otimes K^p$ is zero and this is equivalent to $\phi^p(s) \in U_p \otimes K^p$.
Now that we proved the claim, assume $\phi(U_p) \nsubseteq U_p \otimes K$. Then this is equivalent to $\phi^p(s) \notin U_p \otimes K^p$ and hence consider:
\begin{equation*}
S_{p+1}=\left \langle \im( \phi^p: U_1 \rightarrow E/U_p \otimes K) \right\rangle \otimes K^{-p}
\end{equation*}
which is a subbundle of $E/U_p$. Then there exists a unique subbundle $U_{p+1}$ of $E$ containing $U_p$ and such that $S_{p+1}$ is the image of $U_{p+1}$ with respect to the projection $E \rightarrow E/U_p$. Then $U_{p+1}$ also satisfies the recursive hypotheses and the construction can proceed.
Note that this sequence of steps will terminate either when we get $U_r$ that has rank $r$ and hence $U_r=E$ or when we reach a $\phi$-invariant subbundle of $E$.
The statement about the minimality of $U_l$ is clear because it is generically generated by $s ,\phi(s),\dots, \phi^{p-1}(s)$.
\end{proof}
\end{theorem}
\begin{rmk}
Note that, by construction, the filtration above is unique for each triple $(E,\phi,s)$.
\end{rmk}
\begin{defn}[$U$-filtration]
Let $(E,\phi,s)$ be a triple. We will call \emph{$U$-filtration} the filtration whose existence was proved in the above theorem.\end{defn}
\begin{theorem}
\label{extremal}
\begin{itemize}
\item[(i)] Assume that $\sigma$ is very close to $0$ (i.e. smaller than the first critical value), then $\sigma$-stability for a triple $(E,\phi,s)$ implies the semistability of $(E,\phi)$ and so we have an Abel-Jacobi map:
\begin{equation*}
AJ: \mathcal{M}_\sigma^{r,d} \rightarrow \mathcal{M}^{r,d}.
\end{equation*}
For $d$ large enough (e.g. $d > r(2g-1)+(r-1)^2(2g-2) $) for any semistable Higgs bundle $(E,\phi)$ we have $H^1(E)=0$ and therefore $AJ$ is a projective bundle over the stable part of $\mathcal{M}^{r,d}$.\\
\item[(ii)] For $\sigma>(r-1)d+r(r-1)(r-2)(g-1)$ and a $\sigma$-stable triple $(E,\phi,s)$ corresponding to a pair $(\mathcal{F},s)$ the following three equivalent conditions are realized:
\begin{itemize}
\item there are no $\phi$-invariant subbundles of $E$ which contain the section
\item $s, \phi(s), \dots, \phi^{r-1}(s)$ generically generate $E$
\item $s$ as a map $\mathcal{O}_{\P X} \rightarrow \mathcal{F}$ has zero dimensional cokernel.
\end{itemize}
\end{itemize}
\begin{proof}
For (i) it is proved in \cite[corollary 3.4]{nitsure1991moduli} that if $d > r(2g-1)+(r-1)^2(2g-2) $ then any semistable Higgs bundle of degree $d$ and rank $r$ has vanishing $H^1$. Therefore on the stable locus of $\mathcal{M}^{r,d}$ the fibers of $AJ$ are projective spaces of constant dimension and so $AJ$ is a projective bundle.
For (ii) let us first show that the three conditions are equivalent. The first and second are equivalent because the minimal $\phi$-invariant subbundle of $E$ that contains $s$ is $U_l$ constructed above and it is generated by $s, \phi(s), \dots, \phi^{l-1}(s)$. The second and the third are equivalent since the subsheaf of $E$ generated by $s, \phi(s), \dots, \phi^{r-1}(s)$ is the pushforward of the image of $\mathcal{O}_X \rightarrow \mathcal{F}$.
To prove that the three conditions hold for $\sigma>(r-1)d+r(r-1)(r-2)(g-1)$ we use the $U$-filtration of the triple. Saying that the three conditions are satisfied is the same as saying that the $U$-filtration is of full length $r=\rk E$. Therefore assume that the filtration stops at the step $l < r$. We have the bound $\deg U_l \geq l(l-1)(1-g)$. Therefore:
\begin{equation*}
(l-1)(1-g)+\frac{\sigma}{l} \leq \frac{\deg U_l+\sigma}{l}\leq \frac{d+\sigma}{r}
\end{equation*}
where the first inequality follows from the bound on $\deg U_l$, while the second follows from assuming that the triple is $\sigma$-(semi)stable. Rearranging we get:
\begin{equation*}
\sigma \leq d \frac{l}{r-l}+r(g-1)\frac{l(l-1)}{r-l}\leq d(r-1)+r(r-1)(r-2)(g-1),
\end{equation*}
where the second inequality follows from $l < r$. This contradicts the bound on $\sigma$.
\end{proof}
\end{theorem}
\begin{defn}
We will refer to $\sigma$-stability as $\varepsilon$-stability, when $\sigma$ is very close to zero, and as $\infty$-stability, when $\sigma$ is very large (i.e. larger than the bound in the last proposition).
\end{defn}
From the previous proposition we see that $\infty$-stability implies that the section $s$ is a cyclic vector for $\phi$. Observe that, as opposed to Bradlow pairs, crossing the last critical value for triples will not cause the moduli space to be empty. Rather, the moduli space is constant (and non-empty) after $\sigma=(r-1)d+r(r-1)(r-2)(g-1)$. The cyclic property of $s$ and $\phi$ will have important consequences later, when we study the fixed points of a $\mathbb{C}^*$-action on the moduli space of triples.
For the moment we conclude by observing that the condition of $s: \mathcal{O}_X \rightarrow \mathcal{F}$ having zero dimensional cokernel is the one mentioned by R. Pandharipande and R. Thomas in \cite{pandharipande2009curve}. The only difference is that our ambient variety is a surface, namely $X$ and in \cite{pandharipande2009curve} the ambient variety is a Calabi-Yau threefold and the pairs are used for curve counting purposes.\\
In the case of surfaces however, the moduli space of $\infty$-stable triples satisfies a nice property, following from \cite[Proposition B.8]{pandharipande2010stable} using the BNR correspondence.
\begin{theorem}
\label{relhilb}
The Hitchin map:
\begin{equation*}
\chi^{r,d}_\infty: \mathcal{M}_\infty^{r,d} \rightarrow \mathcal{A}^r
\end{equation*}
is the relative Hilbert scheme of $d+r(r-1)(g-1)$ points over the family of spectral curves $\mathcal{A}^r$.
\begin{proof}
Follows from \cite[Proposition B.8]{pandharipande2010stable}, the BNR correspondence and the relation between $d'$ and $d$ we found in observation \ref{degs}.
\end{proof}
\end{theorem}
\section{Deformation theory of Bradlow-Higgs triples}
\label{deform}
Using \v{C}ech cohomology we can compute the tangent space to $\mathcal{M}^{r,d}_\sigma$ at $(E,\phi,s)$. Note that a similar analysis has been carried out in \cite[proposition 4.12]{le1995faisceaux}, under the point of view of coherent systems of $\P X$.
\begin{theorem}
\label{deformations}
Let $(E,\phi,s)$ be a $\sigma$-stable Bradlow-Higgs triple, then the tangent space at $(E,\phi,s)$ is given by the first cohomology $\H^1(E,\phi,s)$ of the complex:
\begin{equation*}
C^0(\End E)\rightarrow C^1(\End E) \oplus C^0(K\End E) \oplus C^0(E) \rightarrow C^1(E) \oplus C^1(K \End E)
\end{equation*}
where the first map is $$p(k)=(dk, [k,\phi],k \cdot s)$$ and the second one is $$q(\tau, \nu, \gamma)=(\tau \cdot s+d \gamma, [\tau,\phi]+d \nu).$$
Furthermore, $\H^0(E,\phi,s)=0$.
The same result can be obtained from the hypercohomology of the complex:
\begin{align*}
\End E &\rightarrow K \End E \oplus E\\
f & \mapsto ([f,\phi],f(s))
\end{align*}
from which we can also deduce the long exact sequence
\begin{align}
0 \rightarrow H^0(\End E) \rightarrow H^0(K \End E\oplus E) \rightarrow T_{(E,\phi,s)} \rightarrow \\ \nonumber
\rightarrow H^1(\End E) \rightarrow H^1(K \End E\oplus E) \rightarrow \H^2 \rightarrow 0
\end{align}
\begin{proof}
If we want to understand maps $\Spec \mathbb{C}[\varepsilon]/(\varepsilon^2) \rightarrow \mathcal{M}^{r,d}_\sigma$ such that the image of the closed point is $(E,\phi,s)$ then we can choose an open cover $\{U_\alpha\}$ over which $E$ trivializes. So the open sets $\Spec \mathbb{C}[\varepsilon]/(\varepsilon^2) \times U_\alpha$ constitute a trivializing open cover for any deformation $(\tilde{E}, \tilde{\phi}, \tilde{s})$ of our triple.
We can write the transition functions for $E$ as $1+\varepsilon \tau_{\alpha \beta}$ for some $\tau \in C^1(\End E)$, the section can be expressed as $s+\varepsilon \gamma_\alpha$ for $\gamma \in C^0(E)$ and finally the Higgs field as $\phi+\varepsilon \nu_\alpha$ for $\nu \in C^0(K \otimes \End E)$.
If we write the compatibility conditions necessary to obtain a deformation we get:
\begin{align*}
&(1+\varepsilon \tau_{\alpha \beta})(s+\varepsilon \gamma_\beta)=s+\varepsilon \gamma_\alpha\\
&(1+\varepsilon \tau_{\alpha \beta})(\phi+\varepsilon \nu_\beta)(1+\varepsilon \tau_{\alpha \beta})^{-1}=\phi+\varepsilon \nu_\alpha.
\end{align*}
The equations then yield $\tau \cdot s+d \gamma=0$ and $[\tau,\phi]+d\nu=0$ which is precisely the condition $q(\tau,\nu, \gamma)=0$.
Since $\sigma$-stable triples have no automorphisms we can assume that isomorphic deformations only come from the change of trivializations on the open cover $U_\alpha$. In other words we obtain isomorphic deformations (whose cocycles will be distinguished by an apex) if and only if there exists a $k \in C^0(\End E)$ such that
\begin{align*}
&(1+\varepsilon k_\alpha)(1+\varepsilon \tau'_{\alpha \beta})(1-\varepsilon k_\alpha)=1+\varepsilon \tau_{\alpha \beta}\\
&(1+\varepsilon k_\alpha)(s+\varepsilon \gamma'_\alpha)=s+\varepsilon \gamma_\alpha\\
&(1+\varepsilon k_\alpha)(\phi+\varepsilon \nu'_\alpha)(1-\varepsilon k_\alpha)=\phi+\varepsilon \nu_\alpha.
\end{align*}
These three equations are equivalent to requiring that $(\tau'-\tau, \nu'-\nu, \gamma'-\gamma)$ lies in the image of $p$.
Our complex fits into an exact sequence of complexes:
\footnotesize
\begin{center}
\begin{tikzcd}
& 0 \arrow{r} \arrow{d} & C^0(\End E) \arrow{r}{=} \arrow{d}{p} & C^0(\End E) \arrow{d} \\
& C^0(E) \arrow{r}{incl} \arrow{d}{d} & C^1(\End E) \oplus C^0(K\End E) \oplus C^0(E) \arrow{r}{proj} \arrow{d}{q} & C^1(\End E) \oplus C^0(K\End E) \arrow{d} \\
& C^1(E) \arrow{r}{incl} & C^1(E) \oplus C^1(K\End E) \arrow{r}{proj} & C^1(K\End E)
\end{tikzcd}
\end{center}
\normalsize
The right column is the complex associated with the deformations of the Higgs bundle $(E, \phi)$ and its cohomology will be denoted by $\H^i([\cdot, \phi])$.
The cohomology of the deformation complex for our moduli problem will be denoted by $\H^i(E,\phi,s)$ or simply $\H^i$ when no confusion can arise.
This yields an exact sequence of cohomology groups:
\begin{align}
\label{tan}
0 \rightarrow \H^0(E,\phi,s) \rightarrow \H^0([\cdot, \phi]) \rightarrow H^0(E) \rightarrow T_{(E,\phi,s)} \rightarrow \\ \nonumber
\rightarrow \H^1([\cdot,\phi]) \rightarrow H^1(E) \rightarrow \H^2(E,\phi,s) \rightarrow \H^2([\cdot, \phi]) \rightarrow 0.
\end{align}
Note that, of course, we obtain the same result by considering the long exact sequence of hypercohomologies associated to the exact sequence of complexes:
\begin{center}
\begin{tikzcd}
& 0\arrow{r} & 0 \arrow{r} \arrow{d} & \End E \arrow{r} \arrow{d} & \End E \arrow{r} \arrow{d} & 0\\
& 0\arrow{r} & E \arrow{r} & K\End E \oplus E \arrow{r} & K\End E \arrow{r} & 0\\
\end{tikzcd}
\end{center}
The right vertical arrow is the commutator with $\phi$, while the central vertical arrow is the map $f \mapsto ([f,\phi],f(s))$.
The map $\H^0([\cdot, \phi]) \rightarrow H^0(E)$ is evaluation of an endomorphism of $E$ on the section. $\H^0(E,\phi,s)$ is therefore the kernel of such map, i.e. $\H^0(E,\phi,s)$ consists of the endomorphisms of $E$ that commute with $\phi$ and that annihilate the section $s$ and we already know that, by $\sigma$-stability, these are zero. From this we deduce that $\H^0(E,\phi,s)=0$.
\end{proof}
\end{theorem}
Recall that the deformation complex for a Higgs bundle is self dual and in particular
$$\dim \H^2([\cdot, \phi])=\dim \H^0([\cdot, \phi]) \geq 1.$$
Since $\H^2$ surjects onto $\H^2([\cdot, \phi])$, we also have $\dim \H^2 \geq 1$.\\
This allows to get
$$\dim T_{(E,\phi,s)} \mathcal{M}^{r,d}_\sigma=d+(2r^2-r)(g-1)+\dim \H^0([\cdot,\phi]).$$
Note that stability of the underlying Higgs bundle is an open condition and therefore we know that $\dim \H^0([\cdot,\phi])$ is 1 on an open subset of $\mathcal{M}^{r,d}_\sigma$, as we already noted in proposition \ref{opendense}.
We can make a further step in the understanding of the deformations.
\begin{prop}
\label{h2}
Let $(E,\phi,s)$ be a $\sigma$-stable triple and let us denote by $\H^2$ the hypercohomology group in \ref{deformations}. Then $(\H^2)^*$ is the kernel of the following map:
\begin{align*}
H^0(\End E) \oplus H^0(K E^*) & \rightarrow H^0(K \End E)\\
(\alpha, \beta) &\mapsto [\alpha,\phi]+\beta \otimes s.
\end{align*}
In particular, $\dim \H^2 \geq \dim \H^2([\cdot, \phi])$.
\begin{proof}
It is enough to note that, by Serre's duality, $(\H^2)^*$ is the zeroth hypercohomology group of the dual of the deformation complex for $(E,\phi,s)$, i.e.:
\begin{align*}
\End E \oplus K E^* &\rightarrow K \End E\\
(\alpha, \beta) &\mapsto [\alpha,\phi]+\beta\otimes s
\end{align*}
and therefore $(\H^2)^*$ is the kernel of the previous map applied to the global sections.\\
It is worth noting that by $\beta \otimes s$ we mean the map $E \rightarrow E \otimes K$ obtained by tensoring $\beta: E \rightarrow K$ with $s: \mathcal{O} \rightarrow E$.\\
The last statement follows, for example, from the exact sequence \ref{tan}. It can also be deduced directly from the characterization of $(\H^2)^*$ which clearly contains all pairs $(\alpha, \beta)$ with $\alpha$ commuting with $\phi$ and $\beta=0$.
\end{proof}
\end{prop}
\begin{rmk}
\label{adhm}
From \cite[definition 2.1]{diaconescu2012moduli} we see that the data $(E,\phi,s,\alpha,\beta)$ with $(E,\phi,s)$ a Bradlow-Higgs triple, $\alpha \in H^0(\End E)$, $\beta \in H^0(K E^*)$ such that:
\begin{equation*}
[\alpha,\phi]+\beta \otimes s=0
\end{equation*}
define an \emph{ADHM sheaf} on the curve $C$. If $(E,\phi,s)\in \mathcal{M}_\sigma^{r,d}$ is a singular point, then there exist $\alpha$ and $\beta$ not both zero for which $(E,\phi,s,\alpha,\beta)$ is an ADHM sheaf.
For a triple $(E,\phi,s)$ we can deduce from \cite[lemma 2.5]{diaconescu2012moduli} that if $\dim \H^2 > \dim \H^2([\cdot,\phi])$ then there exists a subbundle of $E$ which is proper, $\phi$-invariant and contains $s$. Therefore the following holds:
\begin{itemize}
\item the triples $(E,\phi,s)$ for which the spectral curve is integral are always smooth points
\item if there are no $\phi$-invariant subbundles that contain the section then $\H^2(E,\phi,s)^*$ has the same dimension as the space of endomorphisms of $(E,\phi)$.
\end{itemize}
\end{rmk}
We can prove more precise results once we concentrate on specific combinations of $r,d$ and $\sigma$. Let us first state a technical lemma.
\begin{lemma}
\label{unstableext}
Let $E$ be a rank two vector bundle that is given by an extension:
\begin{equation*}
0 \rightarrow E_2 \rightarrow E \rightarrow E_1 \rightarrow 0
\end{equation*}
with $\deg E_2 > \deg E_1$. Denote by $i$ the inclusion $E_2 \rightarrow E$ and by $p$ the projection $E \rightarrow E_1$.\\
If the extension is nonsplit, we have an isomorphism:
\begin{align*}
\mathbb{C} \oplus H^0( E_1^*E_2) &\rightarrow H^0(\End E)\\
(\lambda, g) & \mapsto \lambda \cdot \mathbb{I}_E+igp.
\end{align*}
If, instead, the extension is split, then there is an isomorphism:
\begin{align*}
\mathbb{C} \oplus \mathbb{C} \oplus H^0( E_1^*E_2) &\rightarrow H^0(\End E)\\
(\lambda_1,\lambda_2, g) & \mapsto \begin{pmatrix} \lambda_1 & 0\\0& \lambda_2 \end{pmatrix}+igp.
\end{align*}
\begin{proof}
Let us first assume the extension is nonsplit. Then we have an exact sequence:
\begin{equation*}
0 \rightarrow \mathcal{O} \rightarrow E_2^* E \rightarrow E_2^*E_1 \rightarrow 0
\end{equation*}
and since $H^0(E_2^* E_1)=0$ we get $H^0(E_2^* E)=\mathbb{C}$. Consider then the exact sequence:
\begin{equation*}
0 \rightarrow E_1^*E_2 \rightarrow E_1^* E \rightarrow \mathcal{O} \rightarrow 0
\end{equation*}
where the second map is defined by taking $f: E_1 \rightarrow E$ and sending it to $pf : E_1 \rightarrow E_1$.
Note that the induced map on the global sections $H^0(E_1^* E) \rightarrow H^0(\mathcal{O})$ has to be zero because if not, we would have a map $f: E_1 \rightarrow E$ such that $p f = \mathbb{I}_{E_1}$ and this implies that the sequence is split. Therefore $H^0(E_1^* E_2) \cong H^0(E_1^* E)$.
Lastly, consider the exact sequence:
\begin{equation*}
0 \rightarrow E_1^*E \rightarrow \End E \rightarrow E_2^* E \rightarrow 0
\end{equation*}
where the second map is defined by sending $f : E \rightarrow E$ to $fi: E_2 \rightarrow E$. Note that the map $ H^0(\End E) \rightarrow H^0(E_2^* E) \cong \mathbb{C}$ is surjective because clearly it does not send $f=\mathbb{I}_E$ to zero. If we trace back the identifications we made we get the statement.\\
The proof for the split case is easier because $H^0(\End E)=H^0(\mathcal{O} \oplus \mathcal{O} \oplus E_1^*E_2 \oplus E_2^* E_1)=\mathbb{C} \oplus \mathbb{C} \oplus H^0(E_1^* E_2)$ and the statement follows immediately.
\end{proof}
\end{lemma}
\begin{cor}
\label{smoothr2}
Assume that $r=2$ and that $(E,\phi,s)$ is a triple whose underlying Higgs bundle is stable and of degree $d$. Then:
\begin{itemize}
\item[(i)] if $d<0$ then $(E,\phi,s)$ is a smooth point in $\mathcal{M}_\varepsilon^{2,d}$,
\item[(ii)] if $d>4g-4$ is odd then $(E,\phi,s)$ is a smooth point in $\mathcal{M}_\varepsilon^{2,d}$.
\end{itemize}
In particular $\mathcal{M}_\varepsilon^{2,d}$ is smooth for $d<0$ and for $d>4g-4$ odd.
\begin{proof}
Here we use the characterization in proposition \ref{h2}. Assume that $E$ is itself stable. Then $H^0(\End E)$ consists only of scalar multiples of the identity. Therefore $[\alpha,\phi]=0=\beta \otimes s$ and so $\beta =0$. So we deduce that $\dim \H^2=1$ which implies $(E,\phi,s)$ is a smooth point.
Assume that $E$ is unstable of rank 2. We have an extension:
\begin{equation*}
0 \rightarrow E_2 \rightarrow E \rightarrow E_1 \rightarrow 0
\end{equation*}
with $\deg E_2 > \deg E_1$ and denote by $i$ the inclusion and $p$ the projection. Here we are in the hypotheses of lemma \ref{unstableext}. Since $E_2$ is destabilizing but $(E,\phi)$ is stable, then $p \phi i$ cannot be zero. Suppose that the above extension is nonsplit. Pick $(\alpha, \beta) \in (\H^2)^*$. If $\alpha \in H^0(\End E)$ we know that $\alpha = \lambda \cdot \mathbb{I}_E+igp$ for $g \in H^0(E_1^*E_2^*)$. Therefore $$[\alpha, \phi]=igp \phi - \phi igp.$$
Since $$p [\alpha,\phi] i=pigp \phi i - p\phi igpi=0,$$ we know that
$$0=p(\beta \otimes s)=\beta i \otimes p(s)$$
and so either $\beta i=0$ or $p(s)=0$. In the first case:
$$igp\phi i=[\alpha, \phi] i = (\beta \otimes s) i =\beta i \otimes s=0$$
and since $p \phi i \neq 0$ we get $ig=0$ and hence $g=0$. In particular $[\alpha,\phi]=0=\beta \otimes s$ and hence $\alpha$ is scalar multiple of the identity while $\beta=0$. This implies $\dim \H^2=1$ and the triple is a smooth point.
If instead $p(s)=0$ then from
$$-p \phi igp=p[\alpha,\phi]=\beta \otimes p(s)=0$$
and since $p \phi i \neq 0$ we have $gp =0$ and hence $g=0$. So once again $\alpha$ is a multiple of the identity and $\beta=0$ so $\dim \H^2=1$ and we get again a smooth point.
We are therefore left with the case of the split extension. Write $s=s_1+s_2$ with $s_i \in H^0(E_i)$ and
$$\phi=\begin{pmatrix}
\phi_{11} & \phi_{21}\\
\phi_{12} & \phi_{22}
\end{pmatrix}.$$
In this case we must have $\phi_{21} \neq 0$ for the stability. Pick $(\alpha,\beta) \in (\H^2)^*$ and write
$$\alpha=\begin{pmatrix}
a & 0\\
c & d
\end{pmatrix},$$
$\beta= \beta_1+\beta_2$. By expanding $[\alpha,\phi]+\beta \otimes s$ we get the following relations:
\begin{align*}
&\phi_{21} c = \beta_1 \otimes s_1\\
&(d-a) \phi_{21} = \beta_2 \otimes s_1\\
&c \phi_{11}-\phi_{22} c+(d-a) \phi_{12} =- \beta_1 \otimes s_2\\
&-c \phi_{21} = \beta_2 \otimes s_2
\end{align*}
Note that if $d<0$ then $\deg E_1 <0$ and so $s_1=0$ so we deduce that $c=0$ and $a=d$ and hence $\alpha$ is a scalar multiple of the identity and $\beta=0$ so $\dim \H^2=1$.
If instead $d>4g-4$ then $d_2>2g-2$ and so $\beta_2: E_2 \rightarrow K$ has to be zero. From the above relations we deduce once again that $\alpha$ is a scalar multiple of the identity and $\beta=0$ so that $\dim \H^2=1$.
Assertions $(i)$ and $(ii)$ as well as the last statement follow from the previous remarks.
\end{proof}
\end{cor}
Let us conclude the section with some remarks. From theorem \ref{extremal} part (i) it follows that the Abel-Jacobi map $\mathcal{M}_\varepsilon^{r,d} \rightarrow \mathcal{M}^{r,d}$ is a projective bundle when $d$ and $r$ are coprime and $d >r(2g-1)+(r-1)^2(2g-2)$. In particular, under these hypotheses $\mathcal{M}_\varepsilon^{r,d}$ is smooth.
For rank 2 it easy to find a singular point in $\mathcal{M}_\varepsilon^{2,d}$ even when $d$ is odd and $0 \leq d \leq 4g-4$. We can build it as follows. Pick $d_1$ and $d_2$ integers satisfying $d_1 \geq 0$ and $d_2 \leq 2g-2$. Pick divisors $S$ and $B$ on $C$ of degrees $d_1$ and $2g-2-d_2$ respectively. Define $D=S+B$.
With these data we define a Bradlow-Higgs triple. Let $E_1$ be the line bundle associated to $S$ and $s_1$ a section of $E_1$ (unique up to scaling by a constant) whose divisor is $S$. Let $E_2$ be defined by $K(-B)$. Then $\deg E_1=d_1$ and $\deg E_2=d_2$. Define $E=E_1\oplus E_2$ and $\phi: E_2 \rightarrow E_1 \otimes K$ be the map associated to the divisor $D$. Lastly let $\beta_2: E_2 \rightarrow K$ be the map associated to the divisor $B$. Clearly we will have $\beta_2 \otimes s_1=\lambda\phi$ as maps $E_2 \rightarrow E_1 \otimes K$ for some constant $\lambda\neq 0$. Then we define $\alpha \in H^0(\End E)$ and $\beta: E \rightarrow K$ as:
\begin{align*}
\alpha &= \begin{pmatrix}
\lambda & 0\\
0 & 0
\end{pmatrix}\\
\beta &= \begin{pmatrix}
0 & \beta_2
\end{pmatrix}.
\end{align*}
It is immediate to check that $(E,\phi)$ is stable and $[\alpha,\phi]=\beta \otimes s$ and since $\alpha$ and $\beta$ are both nonzero we see that $\dim \H^2 > \dim \H^2([\cdot,\phi])$. It is also easy to verify that for this particular triple $\dim \H^2=2$.
We can also produce singular points for the other moduli spaces $\mathcal{M}_\sigma^{r,d}$.
The idea is to produce a split Higgs bundle such that the two summands are not too distant in degree and have both a nonzero section so that the resulting triple will be stable.
Let us build an example for $r=2$, $d$ odd and $\sigma>1$. Consider $E=L \oplus M$ with $\deg M=\deg L+1$, $\phi=diag(\alpha, \beta)$ with $\alpha, \beta \in H^0(K)$ distinct and nonzero. Finally take $s=u \oplus v$ with $u$ and $v$ nonzero sections of $L$ and $M$ respectively. Assume that $v \neq c(u)$ for all maps $c: L \rightarrow M$. Then $L$ and $M$ are clearly the only $\phi$-invariant sub bundles of $E$ and none of them contains the section $s$. If we have $\deg E=d=2k+1$ then $\deg L=k$ and $\deg M=k+1$ so we have:
\begin{align*}
\deg L < \frac{\deg E+\sigma}{2} \qquad \text{ and }\qquad \deg M < \frac{\deg E+\sigma}{2}
\end{align*}
provided that $\sigma>1$. This proves that under our assumption this triple is stable. Moreover any diagonal endomorphism of $E$ will commute with $\phi$ so that we have $\dim \H^2([\cdot, \phi])=\dim \H^0([\cdot, \phi]) \geq 2$ and so $\dim \H^2 \geq 2$ and the triple is a singular point for our moduli space.
Later, we will also construct explicitly fixed points for the $\mathbb{C}^*$-action that are singular in $\mathcal{M}_\sigma^{2,d}$ for $\sigma$ after the first wall.
\section{Properness of the Hitchin map}
In this brief section we will prove the following theorem.
\begin{theorem}
The Hitchin maps
\begin{equation*}
\chi^{r,d}_\sigma: \mathcal{M}_\sigma^{r,d} \rightarrow \mathcal{A}^r
\end{equation*}
are all proper.
\end{theorem}
In the subsequent sections we will see some consequences of this.
Recall from section \ref{secbnr} that $\mathcal{M}^{r,d}_\sigma$ is an open subset of $\text{Syst}_{\P X,\sigma}(r[C],d+r(1-g),1)$ where $\P X$ is the smooth projective surface obtained by adding a divisor at infinity to $T^* C$. We also have a map:
\begin{align*}
\mathcal{S}^{r,d}_\sigma: \text{Syst}_{\P X,\sigma}(r[C],d+r(1-g),1) &\rightarrow \mathcal{H}^r\\
(\mathcal{F},s) & \mapsto \supp{F}
\end{align*}
where $\mathcal{H}^r$ is a Hilbert scheme of curves in $\P X$ whose is Hilbert polynomial is determined by $r$. Observe that, even though we don't include it in the notation, both $\text{Syst}_{\P X,\sigma}(r[C],d+r(1-g),1)$ and $\mathcal{S}^{r,d}_\sigma$ depend on the choice of some polarization $L$ on $\P X$, as explained in section \ref{secbnr}.
It is clear that $\mathcal{S}^{r,d}_\sigma$ is a proper map. In fact, according to \cite[th\'{e}or\`eme 4.11]{le1995faisceaux}, $\text{Syst}_{\P X,\sigma}(r[C],d+r(1-g),1)$ is a projective variety and, according to \cite[theorem 1.1.2]{gottsche1994hilbert}, $\mathcal{H}^r$ is also projective. Using \cite[theorem 4.9]{hartshorne1977algebraic} we see that projective morphisms are proper therefore the composition
\begin{align*}
\text{Syst}_{\P X,\sigma}(r[C],d+r(1-g),1) \rightarrow \mathcal{H}^r \rightarrow \{pt\}
\end{align*}
is proper and $\mathcal{H}^r \rightarrow \{pt\}$ is proper and hence separated. According to \cite[corollary 4.8]{hartshorne1977algebraic}, $\mathcal{S}^{r,d}_\sigma$ is then proper.
Recall that $\mathcal{M}^{r,d}_\sigma$ is the locus of $\text{Syst}_{\P X,\sigma}(r[C],d+r(1-g),1)$ where the support of the sheaves does not intersect the divisor at infinity of $\P X$. As we observed in \ref{secbnr}, if we restrict to such a locus then the stability condition on $\text{Syst}_{\P X,\sigma}(r[C],d+r(1-g),1)$ does not actually depend on the choice of the polarization $L$ on $\P X$. Since the condition defining $\mathcal{M}^{r,d}_\sigma$ relies exclusively on the support of the underlying sheaves we see that the Hitchin map $\chi^{r,d}_\sigma$ is given by the following base change diagram:
\begin{center}
\begin{tikzcd}
& \mathcal{M}^{r,d}_\sigma \arrow{r}{\chi_\sigma^{r,d}} \arrow{d}& \mathcal{A}^r\arrow{d} \\
& \text{Syst} \arrow{r}{\mathcal{S}^{r,d}_\sigma} & \mathcal{H}^r
\end{tikzcd}
\end{center}
and, according to \cite[corollary 4.8]{hartshorne1977algebraic}, $\chi^{r,d}_\sigma$ is then a proper map.
\section{The $\mathbb{C}^*$-action}
As in the case of the moduli space of stable Higgs bundles, the moduli spaces $\mathcal{M}^{r,d}_\sigma$ admit a $\mathbb{C}^*$-action that scales the Higgs field.
\begin{defn}
We define an action:
\begin{align*}
\mathbb{C}^* \times \mathcal{M}_\sigma^{r,d} &\rightarrow \mathcal{M}_\sigma^{r,d}\\
(\lambda, (E,\phi,s)) &\mapsto \lambda \cdot (E,\phi,s)=(E,\lambda\phi,s).
\end{align*}
\end{defn}
As with the moduli space of Higgs bundles we can use the $\mathbb{C}^*$-action and the properness of the Hitchin map to decompose the $\mathcal{M}_\sigma^{r,d}$ into attracting sets. We will also examine the fixed point loci as $\sigma$ varies. Special attention is dedicated to the case of rank 2.
We also have the following theorem.
\begin{theorem}
For all $(E,\phi,s) \in \mathcal{M}_\sigma^{r,d}$ the limit
\begin{equation*}
\lim_{\lambda \rightarrow 0} \lambda \cdot (E,\phi,s)
\end{equation*}
exists in $\mathcal{M}_\sigma^{r,d}$.
\begin{proof}
We have proper maps
\begin{equation*}
\chi^{r,d}_\sigma: \mathcal{M}_\sigma^{r,d} \rightarrow \mathcal{A}^r
\end{equation*}
that are also equivariant with respect to the action of $\mathbb{C}^*$ on $\mathcal{A}^r$ we described in \ref{higgs}. The claimed existence of the limits follows from the valuative criterion for properness and the fact that $\mathbb{C}^*$ acts with positive weights on $\mathcal{A}^r$.
\end{proof}
\end{theorem}
\subsection{The fixed point locus}
Let us first understand which points in $\mathcal{M}_\sigma^{r,d}$ are fixed by the action. The proof of the following is completely analogous to the characterizations of the fixed points in $\mathcal{M}^{r,d}$
\begin{prop}
\label{fixedbht}
Suppose that $(E,\phi,s)$ is a $\sigma$-(semi)stable pair in $\mathcal{M}_\sigma^{r,d}$ such that $\lambda \cdot (E,\phi,s) \cong (E,\phi,s)$ for some $\lambda \in \mathbb{C}^*$ that is not a root of unity, then $E=E_1 \oplus \dots \oplus E_m$, $\phi(E_i) \subseteq E_{i-1} \otimes K$, $\phi(E_1)=0$ and $s \in E_i$ for some $1 \leq i \leq m$.
\begin{proof}
If we don't have the section the proof is the same as in \cite[Lemma 4.1]{simpson1992higgs}.\\
$\lambda \cdot (E,\phi,s) \cong (E,\phi,s)$ implies that there is an automorphism $f$ of $E$ such that $f(s)$ is a multiple of $s$ and $f \phi = \lambda \phi f$. Since the coefficients of the characteristic polynomial are sections of $\mathcal{O}_C$, we know that they are constants. Therefore also the eigenvalues of $f$ are constant and we have the decomposition of $E$ according to the generalized eigenspaces $E=\oplus_{\mu} E_\mu$ where $E_\mu=\ker (f-\mu)^n$. Now, $(f-\lambda \mu)^n \phi=\lambda^n \phi (f-\mu)^n$ so, in particular, $\phi(E_\mu) \subset E_{\lambda \mu}$. Ordering the eigenvalues of $f$ and assuming that $\lambda$ is not a root of unity, we get the statement about $\phi$. Furthermore, by definition of the $E_i$, we know that $f=\sum_i f_i$ with $f_i \in \Aut(E_i)$. This implies that, if $f(s)=k \cdot s$ for some $k \in \mathbb{C}^*$ then whenever we can write $s=\sum_i s_i$ with $s_i \in H^0(E_i)$ then each of the $f_i$ scales $s_i$ by a different constant. This clearly implies that only one of the $s_i$ can be nonzero.
\end{proof}
\end{prop}
Therefore we have
\begin{cor}
The fixed points of the $\mathbb{C}^*$-action on $\mathcal{M}_\sigma^{r,d}$ are exactly the points of the form above.
\begin{proof}
We simply need to check that every such point is fixed. But this is clear because, if $(E,\phi,s)$ is in the form above, then $f$ defined by $f_{|E_i}=\lambda^{i-1} \mathbb{I}_{E_i}$ is an isomorphism between $(E,\lambda\phi,s)$ and $(E,\phi,s)$.
\end{proof}
\end{cor}
\subsection{Structure of the attracting sets}
The results by Bia\l{}ynicki-Birula can only be applied for the few $\mathcal{M}_\sigma^{r,d}$ that are smooth. For example from proposition \ref{smoothr2} we know that when $r=2$ and $\sigma$ is small, we should have $d<0$ or $d$ odd and $>4g-4$ in order to have smooth moduli spaces. In the other cases in fact the moduli spaces are not smooth. Nevertheless, since we still have a proper Hitchin map, we have the decomposition of $\mathcal{M}_\sigma^{r,d}$ into attracting loci.
\begin{defn}
Given a partition $\underline{r}=(r_1, \dots ,r_m)$ of $r$, a partition $\underline{d}=(d_1, \dots, d_m)$ and $1 \leq k\leq m$ we denote by
\begin{equation*}
F_{\underline{r},\sigma}^{\underline{d},k}=\left \{
\begin{array}{c}
(E,\phi,s)\in \mathcal{M}_\sigma^{r,d} \text{ such that } E=E_1 \oplus \dots \oplus E_m\\
\text{with } \deg E_i=d_i, \rk{E_i}=r_i\\
\phi(E_i) \subseteq E_{i-1} \otimes K, \phi(E_1)=0\\
s \in H^0(E_k)
\end{array} \right \}
\end{equation*}
\end{defn}
\begin{defn}
Let us fix $F_{\underline{r},\sigma}^{\underline{d},k}$. We define two locally closed subsets by:
\begin{align*}
F_{\underline{r},\sigma}^{\underline{d},k+}&=\{(E,\phi,s) \in \mathcal{M}_\sigma^{r,d}\text{ such that } \lim_{\lambda \rightarrow 0}(E,\phi,s) \in F_{\underline{r},\sigma}^{\underline{d},k}\}\\
F_{\underline{r},\sigma}^{\underline{d},k-}&=\{(E,\phi,s) \in \mathcal{M}_\sigma^{r,d}\text{ such that } \lim_{\lambda \rightarrow \infty}(E,\phi,s) \in F_{\underline{r},\sigma}^{\underline{d},k}\}
\end{align*}
\end{defn}
Then, similar to the situation for $\mathcal{M}^{r,d}$ described in section \ref{semipr}, we have:
\begin{prop}
Let $(E,\phi,s)$ be a point in $F_{\underline{r},\sigma}^{\underline{d},k}$, for $\sigma$ not a critical value. With the previous notations, the weight $0$ part of the $\mathbb{C}^*$-action on $T_{(E,\phi,s)} \mathcal{M}_\sigma^{r,d}$ is given by the first hypercohomology of the complex:
\begin{equation*}
\bigoplus_{i=1}^{m} \Hom(E_i,E_i) \rightarrow \bigoplus_{i=2}^{m} \Hom(E_i,E_{i-1} \otimes K) \oplus E_k,
\end{equation*}
the positive weight part is the first hypercohomology of:
\begin{equation*}
\bigoplus_{i<j} \Hom(E_i,E_j) \rightarrow \bigoplus_{i\leq j} \Hom(E_i,E_{j} \otimes K) \oplus \bigoplus_{i>k} E_i
\end{equation*}
while the negative part is the first hypercohomology of:
\begin{equation*}
\bigoplus_{i>j}^{m} \Hom(E_i,E_j) \rightarrow \bigoplus_{i>j+1}^{m} \Hom(E_i,E_{j} \otimes K) \oplus \bigoplus_{i<k} E_i
\end{equation*}
\begin{proof}
Since $(E,\phi,s)$ is $\sigma$-stable, it only has scalar automorphisms. In particular the group of isomorphisms between $(E,\lambda \phi, s)$ and $(E,\phi,s)$ is also $\mathbb{C}^*$. Among these automorphisms, there is a unique one that preserves the section, which is exactly the one that acts with weight 0 on $E_k$.
In other words, suppose we have a fixed point $E=E_1 \oplus \dots \oplus E_m$, $\phi$ as usual and $s \in H^0(E_k)$. Then the canonical automorphism that we want is $f:E \rightarrow E$ such that $f_{|E_i}=\lambda^{i-k} \mathbb{I}_{E_i}$. Recall that the Zariski tangent space to $(E,\phi,s)$ is generated by the kernel of the map
\begin{align*}
C^1(\End E) \oplus C^0(K\End E) \oplus C^0(E) &\rightarrow C^1(E) \oplus C^1(K \End E)\\
(\tau, \nu, \gamma) & \mapsto (\tau \cdot s+d \gamma, [\tau,\phi]+d \nu).
\end{align*}
The $\mathbb{C}^*$-action induced on the Zariski tangent space is given by:
\begin{align*}
\lambda(\tau,0,0)=\lambda^{j-i}(\tau,0,0) &\text{ if } \tau \in C^1(E_i^* E_j)\\
\lambda(0,\nu,0)=\lambda^{j-i+1}(0,\nu,0) &\text{ if } \nu \in C^0(E_i^* E_j K)\\
\lambda(0,0,\gamma)=\lambda^{i-k}(0,0,\gamma) &\text{ if } \gamma \in C^0(E_i).\\
\end{align*}
In order to decompose the tangent complex for $(E,\phi,s)$ we can therefore consider $E_i^*E_j$ acted on with weight $i-j$ (observe that this does not depend on $k$), $E_i$ with weight $i-k$ and $K$ with weight 1 and we obtain the above complexes.
\end{proof}
\end{prop}
Later we will use the complexes to compute the dimension of the positive weight part of the $\mathbb{C}^*$-action on the tangent space at the fixed points.
\begin{rmk}
\label{smoothres}
An important remark is that, even though \cite[Theorem 4.1]{bialynicki1973some} does not apply to the singular $\mathcal{M}_\sigma^{r,d}$, it is still possible to apply them to the subvarieties that contain the smooth fixed points and all the points whose limit as $\lambda \rightarrow 0$ is one of the smooth fixed points. Note that these subvarieties are clearly smooth, since the dimension of the Zariski tangent space can only go up in dimension with specialization and it is already minimal at the smooth fixed points. As we will see later, some of the attracting sets lie entirely in these smooth subvarieties.
\end{rmk}
If the fixed point is smooth then we get the dimension of the affine fiber of $F_{\underline{r},\sigma}^{\underline{d},k+}\rightarrow F_{\underline{r},\sigma}^{\underline{d},k}$ at that point, by the previous remark.
\subsection{The rank 2 case}
\label{rank2disc}
We now specialize on the case $r=2$. First of all, we remark that the set of critical values for $\sigma$, when $r=2$ is, as for Bradlow pairs, the set of positive integers with the same parity as $d$.
First let us comment on the case of $d < 0$.
\begin{prop}
\label{negdegrk2}
Let $d < 0$. Then $\mathcal{M}_\infty^{2,d}=\mathcal{M}_\varepsilon^{2,d}$ and both are non-empty iff $d \geq 2-2g$.
\begin{proof}
Let $(E,\phi,s) \in \mathcal{M}_\varepsilon^{2,d}$ and consider the subbundle $S=\left \langle s \right \rangle$ generated by the section $s$. Then certainly $\deg S \geq 0 > d$ so that $S$ is stricly destabilizing for $E$ and cannot be fixed by $\phi$. This implies $s$ is cyclic for $\phi$ as we are in rank 2.\\
Viceversa, assume $(E,\phi,s) \in \mathcal{M}_\infty^{2,d}$. Suppose there exists a rank one Higgs subbundle $(L,\psi)$ of $(E, \phi)$ such that $\deg L \geq d/2$. Since $s$ is cyclic for $\phi$, we know that $\s \notin H^0(L)$. If we denote again by $S$ the subbundle generated by $s$, we would get a nonzero map $S \rightarrow E / L$. This would imply that $0 \leq \deg S \leq d - \deg L \leq d/2$ but this is impossible.
We now prove the statement about non-emptiness. If $d \geq 0$ then we know there exist $\sigma$-stable pairs and so both moduli spaces are non-empty (but they are not equal).\\
For $d<0$ the two moduli spaces coincide and if we have at least a stable triple $(E,\phi,s)$ there is a nonzero map $S \rightarrow E/S \otimes K$, since $s$ is cyclic for $\phi$. But then $0 \leq \deg S \leq d- \deg S +2g-2$ which implies $0 \leq 2 \deg S \leq d+2g-2$ or $d \geq 2-2g$.\\
Viceversa, note that for any point $P \in C$ there is an embedding
\begin{align*}
\mathcal{M}_\infty^{2,d} &\rightarrow \mathcal{M}_\infty^{2,d+2}\\
(E,\phi,s) &\mapsto (E(P),\phi,s(P))
\end{align*}
so that it is enough to prove that $\mathcal{M}_\infty^{2,d}$ is non-empty for $d=2-2g$ and $d=3-2g$.\\
For $d=2-2g$ take $E=\mathcal{O} \oplus K^*$, $s$ a constant section of $\mathcal{O}$ and $\phi$ annihilating $K^*$ and restricting to the constant map $\mathcal{O} \rightarrow K^* \otimes K=\mathcal{O}$. This is a triple in $\mathcal{M}_\infty^{2,2-2g}$.\\
For $d=3-2g$ pick $P \in C$ and define $E=\mathcal{O} \oplus K^*(P)$, $s$ a constant section of $\mathcal{O}$ and $\phi$ annihilating $K^*(P)$ and restricting to the only nonzero map $\mathcal{O} \rightarrow K^* \otimes K(P)=\mathcal{O}(P)$. This is instead a triple in $\mathcal{M}_\infty^{2,3-2g}$.
\end{proof}
\end{prop}
Note that this proposition also implies that if $(E,\phi,s) \in \mathcal{M}_\varepsilon^{2,d}$ for $d <0$ then $(E,\phi)$ is strictly stable.
We also prove a more precise version of \cite[corollary 3.4]{nitsure1991moduli} for rank $2$.
\begin{prop}
\label{sharpnits}
Let $(E,\phi)$ be a semistable Higgs bundle of rank 2 and degree $d$. If $d \geq 6g-5$ then $H^1(E)=0$.
\begin{proof}
If $E$ is semistable itself then there are no non-zero maps $E \rightarrow K$ since $\mu(E) > 3g-2 >2g-2 = \mu(K)$, therefore $H^1(E)=H^0(KE^*)^*=0$. Otherwise assume that $L \subset E$ is the maximal destabilizing line subbundle of $E$. Then $\deg L >d/2$ but since $(E,\phi)$ is stable there is also a non-zero map $L \rightarrow E/L \otimes K$ and so $\deg L \leq d/2 +g-1$. In particular we also get $1-g+d/2 \leq \deg E/L < d/2$. If $d \geq 6g-4$ then we have $\deg L > d/2 > 3g-2 >2g-2$ and $\deg E/L > 3g-3+1-g=2g-2$ and therefore $H^1(L)=H^1(E/L)=0$ which implies $H^1(E)=0$.
\end{proof}
\end{prop}
\begin{rmk}
\label{nits}
Note that the estimate in \cite[corollary 3.4]{nitsure1991moduli} would say that $H^1(E)=0$ for $d > 4g-2 + 2g-2=6g-4$ i.e. $d \geq 6g-3$ which leaves out a value of $d$ that we will need in the following chapters. It is also reasonable to think that the gap between the sharp estimate and the estimate in \cite[corollary 3.4]{nitsure1991moduli} will widen as the rank increases.
\end{rmk}
We can prove the following.
\begin{theorem}
\label{fpr2}
Let $d\geq 2-2g$ be an integer and $\sigma>0$ different from a critical value. Then we can classify the components of the fixed point locus of $\mathcal{M}_\sigma^{2,d}$ as follows:
\begin{itemize}
\item[(i)] if $d \geq 0$ and $\sigma < d$ then one of the components of the fixed points for the $\mathbb{C}^*$-action is $F_{(2),\sigma}^{(d),1}=M_\sigma^{2,d}$, i.e. the moduli space of $\sigma$-stable Bradlow pairs embedded as triples with zero Higgs field. If $d <0$ then there are no $\sigma$-stable Bradlow pairs and so this component is empty.
\item[(ii)] if there exists and integer $m$ such that $\max\{0,d/2 +1-g\} \leq m < \frac{d-\sigma}{2}$, then there exist components:
\begin{align*}
F_{(1,1),\sigma}^{(d_1,d_2),1}&=\left \{
\begin{array}{c}
(E,\phi,s)\in \mathcal{M}_\sigma^{2,d} \text{ such that } E=E_1 \oplus E_2\\
\text{with } \deg E_i=d_i, \rk{E_i}=1\\
\phi(E_2) \subseteq E_{1} \otimes K, \phi(E_1)=0\\
s \in H^0(E_1)
\end{array} \right \} \cong \\
&\cong S^{d_1}( C) \times S^{d_1-d_2+2g-2}( C).
\end{align*}
Here $d_1$ and $d_2$ are integers satisfying $d_1+d_2=d$ and one of the following equivalent inequalities:
\begin{align*}
&\max\{0,d/2 +1-g\} \leq d_1 < \frac{d-\sigma}{2} \\
&\frac{d+\sigma}{2} < d_2 \leq \min\{d,d/2+g-1\}\\
&\max\{0,2g-2-d\}\leq d_1-d_2+2g-2 < 2g-2-\sigma \text{ same parity as $d$}.
\end{align*}
\item[(iii)] There exist components:
\begin{align*}
F_{(1,1),\sigma}^{(d_1,d_2),2}&=\left \{
\begin{array}{c}
(E,\phi,s)\in \mathcal{M}_\sigma^{2,d} \text{ such that } E=E_1 \oplus E_2\\
\text{with } \deg E_i=d_i, \rk{E_i}=1\\
\phi(E_2) \subseteq E_{1} \otimes K, \phi(E_1)=0\\
s \in H^0(E_2)
\end{array} \right \} \cong\\
&\cong S^{d_2}( C) \times S^{d_1-d_2+2g-2}( C).
\end{align*}
Here $d_1$ and $d_2$ are integers satisfying $d_1+d_2=d$ and one of the following equivalent inequalities:
\begin{align*}
&d/2 +1-g \leq d_1 < \min \left \{\frac{d+\sigma}{2},d+1 \right \} \\
&\max \left \{-1,\frac{d-\sigma}{2} \right \}< d_2 \leq d/2+g-1\\
& 0\leq d_1-d_2+2g-2 < 2g-2+\min\{\sigma,d+1\} \text{ same parity as $d$}.
\end{align*}
\end{itemize}
\begin{proof}
Point (i) simply follows by noting that $M^{2,d}$ is nonempty if and only if $d \geq 0$ and $\sigma < d$ and $(E,s) \in M^{2,d}$ if and only if $(E,0,s) \in \mathcal{M}_\sigma^{2,d}$. Clearly, every $\sigma$-stable triple whose Higgs field vanishes is fixed by the $\mathbb{C}^*$-action.
For (ii), recall from proposition \ref{fixedbht} that for the partition $(1,1)$ we must have fixed points of the form above with $s \in H^0(E_1)$ if $k=1$. Furthermore, note that $E_1$ is the only $\phi$-invariant subbundle in this case. $s \in H^0(E_1)$ implies $d_1 \geq 0$, $\sigma$-stability implies $d_1<(d-\sigma)/2$ and the existence of a nonzero map $E_2 \rightarrow E_1 \otimes K$ implies $d_1-d_2+2g-2 \geq 0$. Putting all together and using $d_1+d_2=d$ we get the claimed set of inequalities.
For (iii) we know that $s \in E_2$ when $k=2$. Here $s \in H^0(E_2)$ forces $d_2 \geq 0$, $\sigma$-stability implies $d_1 < (d+\sigma)/2$ and as before $d_1-d_2+2g-2 \geq 0$. Using $d_1+d_2=d$ we get the claimed set of inequalities.
\end{proof}
\end{theorem}
From now on we will call \emph{non-split points} the points in $F_{(2),\sigma}^{(d),1}$, \emph{split type 1} the points in $F_{(1,1),\sigma}^{(d_1,d_2),1}$ and \emph{split type 2} the points in $F_{(1,1),\sigma}^{(d_1,d_2),2}$.
Note that split type 2 points always exist for all combinations of $d \geq 2-2g$ and $\sigma >0$ and nonsplit fixed points exist only for $0<\sigma<d$. Split type 1 points are more complicated, in fact they exist only if $\sigma < d$ and $\sigma < 2g-3$ for $d$ odd or $\sigma < 2g-2$ for $d$ even. In particular, for $d >2g-2$ this implies that there is a critical value for $\sigma$ after which the type 1 split points cease to exist.
We can briefly analyze the dimension of $\H^2$ for the deformations of the fixed points. Recall the characterization of $(\H^2)^*$ we gave in proposition \ref{h2}.
For a nonsplit point $(E,0,s)$, clearly $[\alpha,\phi]=0$ and so $\beta=0$. Therefore $\dim \H^2=\dim H^0( \End E)$ and so $(E,0,s)$ is smooth if and only if $E$ is simple, which happens for example if $E$ is stable.
For the split fixed points instead, we interpret $\phi$ as a matrix
\begin{equation*}
\begin{pmatrix}
0 & \phi\\
0 & 0
\end{pmatrix}
\end{equation*}
and an endomorphism of $E_1 \oplus E_2$ has the form
\begin{equation*}
\begin{pmatrix}
a & b\\
c & d
\end{pmatrix}
\end{equation*}
with $a \in \mathbb{C}$ and $c \in \Hom(E_1,E_2)$, $b \in \Hom(E_2,E_1)$. We have $\beta=\beta_1+\beta_2$ with $\beta_i : E_i \rightarrow K$ and $s=s_1+s_2$ with $s_i \in H^0(E_i)$.
Then:
\begin{equation*}
[\alpha,\phi]=\begin{pmatrix}
-\phi c & (a-d)\phi\\
0 & c \phi
\end{pmatrix}
\end{equation*}
and
\begin{equation*}
\beta \otimes s=\begin{pmatrix}
\beta_1 s_1 & \beta_2 s_1\\
\beta_1 s_2 & \beta_2 s_2
\end{pmatrix}.
\end{equation*}
If $s_2=0$, i.e. for a type 1 fixed point, we get $c=0$, $\beta_1=0$ and $\beta_2 s_1=(a-d) \phi$. Note that since $d_2-d_1>0$ for a type 1 fixed point, we will always have $b=0$. In particular
\begin{equation*}
\dim \H^2=1+
\begin{cases}
1 & \text{if } \text{div} s_1 + \text{div} \beta_2 = \text{div} \phi\\
0 & \text{otherwise}
\end{cases}
\end{equation*}
note that in the second case $\H^2 \cong \H^2([\cdot,\phi])$ and in the first $\dim \H^2 = \dim \H^2([\cdot,\phi])+1$.
If $s_1=0$, i.e. for a type 2 fixed point, we always get $c=0$ and $a=d$, therefore $[\alpha, \phi]=0$ and $\beta=0$ so $\dim \H^2 = \dim \H^2([\cdot,\phi])$.
We can summarize the discussion in the following proposition.
\begin{prop}
\label{deffixed}
A nonsplit fixed point is smooth if and only if the underlying vector bundle is simple.
A split type 1 fixed point $(E_1 \oplus E_2, \phi: E_2 \rightarrow E_1 \otimes K, s_1)$ is smooth if and only if $ \text{div}\phi-\text{div} s_1 $ is not an effective divisor. In other words the singular points in the component $S^{d_1}( C) \times S^{d_1-d_2+2g-2}( C)$ are those in the image of the map:
\begin{align*}
S^{d_1}( C) \times S^{-d_2+2g-2}( C) & \rightarrow S^{d_1}( C) \times S^{d_1-d_2+2g-2}( C)\\
(D_1, D_2) & \mapsto (D_1, D_1+D_2)
\end{align*}
which of course is empty when $d_2>2g-2$.
A split type 2 fixed point is smooth if and only if the underlying Higgs bundle is simple.
\end{prop}
We can also say something about the structure of the attracting sets, even though it is not always possible to conclude that they are affine fibrations over the fixed point loci since we are not inside smooth ambient varieties.
The fiber of the limit map $F_{(2),\sigma}^{(d),1+} \rightarrow F_{(2),\sigma}^{(d),1}$ over $(E,0,s)$ is clearly $H^0(K \otimes \End E)$. Observe that they are still affine spaces but not of constant dimension as
$$\dim H^0(K \otimes \End E)= \dim H^1(\End E)=4g-4+\dim H^0(\End E)$$
and this can vary if $E$ is not simple. Note that over the locus of $M_\sigma^{2,d} \cong F_{(2),\sigma}^{(d),1}$ where $E$ is stable, we have an affine fibration of dimension $4g-3$.
For fixed points of split type 1 we can compute the dimension of the fibers of $F_{(1,1),\sigma}^{(d_1,d_2),1+} \rightarrow F_{(1,1),\sigma}^{(d_1,d_2),1}$ over the smooth fixed points. We can decompose the deformation complex into weighted parts:
\begin{equation*}
\Hom(E_2,E_2)\oplus\Hom(E_1,E_1) \rightarrow \Hom(E_2,E_1 \otimes K) \oplus E_1
\end{equation*}
for the weight 0 part,
\begin{equation*}
\Hom(E_2,E_1)\rightarrow 0
\end{equation*}
for the negative weight part and
\begin{equation*}
\Hom(E_1,E_2) \rightarrow \Hom(E_1,E_1 \otimes K) \oplus \Hom(E_2,E_2 \otimes K) \oplus \Hom(E_1,E_2 \otimes K) \oplus E_2
\end{equation*}
for the positive weight part. The dimension of the weight 0 part, which clearly corresponds to the tangent space to the component of the fixed points, is $2d_1-d_2+2g-2$ equal to the dimension of $S^{d_1}( C) \times S^{d_1-d_2+2g-2} ( C)$. The dimension of the negative weight part is $d_2-d_1+g-1$ and finally the dimension of the positive weight part is $1+4(g-1)+d_2+1-g$.
For fixed points of split type 2 we can still decompose the deformation complex into weighted parts:
\begin{equation*}
\Hom(E_2,E_2)\oplus\Hom(E_1,E_1) \rightarrow \Hom(E_2,E_1 \otimes K) \oplus E_2
\end{equation*}
for the weight 0 part,
\begin{equation*}
\Hom(E_2,E_1)\rightarrow E_1
\end{equation*}
for the negative weight part and
\begin{equation*}
\Hom(E_1,E_2) \rightarrow \Hom(E_1,E_1 \otimes K) \oplus \Hom(E_2,E_2 \otimes K) \oplus \Hom(E_1,E_2 \otimes K)
\end{equation*}
for the positive weight part. Here type 2 split fixed points can be singular and it is actually easy to see that the positive part of the tangent space contains entirely the part that gives the extra dimension at the singular points. More specifically, the dimension of the weight 0 part is $d_1+2g-2$ equal to the dimension of $S^{d_2}( C) \times S^{d_1-d_2+2g-2} ( C)$. The dimension of the negative weight part is $d_2$ and finally the dimension of the positive weight part is $1+4(g-1)+\dim \Hom(E_2,E_1)$. As we already said, this gives the dimension of the fibers of $F_{(1,1),\sigma}^{(d_1,d_2),2+} \rightarrow F_{(1,1),\sigma}^{(d_1,d_2),2}$ over the smooth fixed points.
As for rank $2$ Higgs bundles we can prove that the decomposition of $\mathcal{M}_\sigma^{2,d}$ into attracting sets is the same as the decomposition according to the Harder-Narasimhan type of the Bradlow pair underlying the triple. We will call \emph{Hitchin stratification} the stratification of $\mathcal{M}_\sigma^{2,d}$ into attracting sets and \emph{Shatz stratification} the stratification of $\mathcal{M}_\sigma^{2,d}$ according to the Harder-Narasimhan type of the underlying pair.
\begin{theorem}
\label{shatz}
Let $\sigma$ be different from a critical value.\\
We have:
\begin{equation*}
F_{(2),\sigma}^{(d),1+}=
\left \{
\begin{array}{c}
(E,\phi,s) \text{ such that } (E,s) \text{ is $\sigma$-stable}
\end{array} \right \}
\end{equation*}
and the limit map $F_{(2),\sigma}^{(d),1+} \rightarrow F_{(2),\sigma}^{(d),1}$ takes $(E,\phi,s)$ to $(E,0,s)$.
Otherwise, the underlying Bradlow pair is $\sigma$-unstable and:
\begin{equation*}
F_{(1,1),\sigma}^{(d_1,d_2),1+}=
\left \{
\begin{array}{c}
(E,\phi,s) \text{ where } E \text{ is defined by}\\
0 \rightarrow E_2 \rightarrow E \rightarrow E_1 \rightarrow 0\\
s \notin H^0(E_2), \deg E_i =d_i, d_2 > (d+\sigma)/2\\
\text{and } \phi \text{ does not preserve } E_2
\end{array} \right \}
\end{equation*}
and the limit map $F_{(1,1),\sigma}^{(d_1,d_2),1+} \rightarrow F_{(1,1),\sigma}^{(d_1,d_2),1}$ sends $(E,\phi,s)$, described as above, to $(E_2\oplus E_1, p_{E_1} \phi i_{E_2}, p_{E_1}(s))$ where $i_{E_2}$ and $p_{E_1}$ are the inclusions and projections in the description of $E$ as an extension.
The other possibility is:
\begin{equation*}
F_{(1,1),\sigma}^{(d_1,d_2),2+}=
\left \{
\begin{array}{c}
(E,\phi,s) \text{ where } E \text{ is defined by}\\
0 \rightarrow E_2 \rightarrow E \rightarrow E_1 \rightarrow 0\\
s \in H^0(E_2), \deg E_i =d_i, d_2 > (d-\sigma)/2\\
\text{and } \phi \text{ does not preserve } E_2
\end{array} \right \}
\end{equation*}
and the limit map $F_{(1,1),\sigma}^{(d_1,d_2),2+} \rightarrow F_{(1,1),\sigma}^{(d_1,d_2),2}$ sends $(E,\phi,s)$, described as above, to $(E_2\oplus E_1, p_{E_1} \phi i_{E_2}, s)$ where $i_{E_2}$ and $p_{E_1}$ are the inclusions and projections in the description of $E$ as an extension.
\begin{proof}
The statement about $F_{(2),\sigma}^{(d),1+}$ is clear because, if $(E,s)$ is a $\sigma$-stable Bradlow pair, then $(E,\phi,s)$ will be a $\sigma$-stable triple for every $\phi \in H^0(K \End E)$. Furthermore, the entire family $\lambda \mapsto (E,\lambda \phi,s)$ consists of $\sigma$-stable triples, even for $\lambda=0$. Therefore the limit of $\lambda \cdot (E,\phi, s)$ as $\lambda \rightarrow 0$ is $(E,0,s)$ in this case.
For the other two types of cells we can use a \v{C}ech cohomology description of the triple to understand what the $\mathbb{C}^*$-action does. We spell all the details for $F_{(1,1),\sigma}^{(d_1,d_2),1+}$ and for the other cells the proof is analogous.
Let $(E,\phi,s)$ be defined by an extension:
\begin{equation*}
0 \rightarrow E_2 \rightarrow E \rightarrow E_1 \rightarrow 0,
\end{equation*}
call $i_{E_2}$ and $p_{E_1}$ the inclusions and projections and assume $d_2 > (d+\sigma)/2$, $s \notin H^0(E_2)$ and $\phi(E_2) \nsubseteq E_2 \otimes K$.
Then $(E,s)$ is not $\sigma$-stable and $E_2$ is the maximal destabilizing of $(E,s)$. If there were an $L \subset E$ that is $\phi$-invariant and contains the section, then clearly the map $L \rightarrow E \rightarrow E_1$ is nonzero and so $\deg L \leq \deg E_1 < (d-\sigma)/2$. This implies that every such triple is actually $\sigma$-stable.
Choose a fine enough affine cover $U_\alpha$ of $C$ such that $E$, $E_2$ and $E_1$ all trivialize. Suppose that $(E,\phi,s)$ is given by the \v{C}ech cohomology data $(U_\alpha, g_{\alpha \beta},\phi_\alpha,s_\alpha)$ relative to the open cover $U_\alpha$, where:
\begin{equation*}
g_{\alpha\beta}=
\begin{pmatrix}
e_{1\alpha\beta} & 0\\
\tau_{\alpha\beta} & e_{2\alpha\beta}
\end{pmatrix}
\end{equation*}
are the transition functions of $E$, where $e_i$ are the transition functions for $E_i$ and $\tau \in C^1( E_1^* E_2)$,
\begin{equation*}
\phi_\alpha=
\begin{pmatrix}
a_{\alpha} & b_{\alpha}\\
c_{\alpha} & d_{\alpha}
\end{pmatrix}
\end{equation*}
for the Higgs field,
\begin{equation*}
s_\alpha=
\begin{pmatrix}
s_{1\alpha}\\
s_{2\alpha}
\end{pmatrix}
\end{equation*}
for $s$ where $s_i$ is the part of the section lying in $E_i$.
Since these data define a triples, we have the compatibility relations:
\begin{align*}
& g_{\alpha \beta}=g_{\beta \alpha}^{-1}\\
& g_{\alpha \beta} g_{\beta \gamma}=g_{\alpha \gamma}\\
& s_\alpha = g_{\alpha \beta} s_{\beta}\\
& g_{\alpha \beta} \phi_\beta = \phi_\alpha g_{\alpha \beta}.
\end{align*}
Now consider the triple $(E,\lambda \phi, s)$ for $\lambda \in \mathbb{C}^*$, which is represented by the data $(U_\alpha, g_{\alpha \beta},\lambda \phi_\alpha,s_\alpha)$. We want to show that $(E,\lambda\phi,s)$ is also represented by the data $(U_\alpha, g^{\lambda}_{\alpha \beta},\phi^{\lambda}_\alpha,s^{\lambda}_\alpha)$ where:
\begin{equation*}
g^{\lambda}_{\alpha\beta}=
\begin{pmatrix}
e_{1\alpha\beta} & 0\\
\lambda \tau_{\alpha\beta} & e_{2\alpha\beta}
\end{pmatrix},
\end{equation*}
\begin{equation*}
\phi^{\lambda}_\alpha=
\begin{pmatrix}
\lambda a_{\alpha} & b_{\alpha}\\
\lambda^2 c_{\alpha} & \lambda d_{\alpha}
\end{pmatrix}
\end{equation*}
and
\begin{equation*}
s^{\lambda}_\alpha=
\begin{pmatrix}
s_{1\alpha}\\
\lambda s_{2\alpha}
\end{pmatrix}.
\end{equation*}
First of all the compatibility relations for $(U_\alpha, g_{\alpha \beta},\lambda \phi_\alpha,s_\alpha)$ imply the compatibility relations for $(U_\alpha, g^{\lambda}_{\alpha \beta},\phi^{\lambda}_\alpha,s^{\lambda}_\alpha)$ so that the second set of data actually defines a triple.
Using the change of trivialization (independent of $\alpha$):
\begin{equation*}
\gamma_\alpha=
\begin{pmatrix}
\lambda & 0\\
0 & 1
\end{pmatrix} \in C^0(\GL_2)
\end{equation*}
we can see that $(U_\alpha, g_{\alpha \beta},\lambda \phi_\alpha,s_\alpha)$ and $(U_\alpha, g^{\lambda}_{\alpha \beta},\phi^{\lambda}_\alpha,s^{\lambda}_\alpha)$ are equivalent set of data and so they define the same triple $(E,\lambda \phi, s)$. What we need to verify explicitly are the relations:
\begin{align*}
& \gamma_\alpha^{-1}g_{\alpha \beta} \gamma_\beta = g^{\lambda}_{\alpha \beta}\\
& \gamma_\alpha^{-1} \lambda \phi_\alpha \gamma_\alpha =\phi^\lambda_\alpha\\
& \gamma_\alpha^{-1} \lambda s_\alpha=s_\alpha^{\lambda}
\end{align*}
which are an immediate check. Note that in the last relation we are allowed to change $s_\alpha$ by $\lambda s_\alpha$ because two triples which are identical except for the sections which differ by multiplication by a constant are always isomorphic.
The data $(U_\alpha, g^{\lambda}_{\alpha \beta},\phi^{\lambda}_\alpha,s^{\lambda}_\alpha)$ produce a $\mathbb{C}^*$-family of $\sigma$-stable triples whose limit as $\lambda \rightarrow 0$ is given by the data:
\begin{equation*}
\begin{pmatrix}
e_{1\alpha\beta} & 0\\
0 & e_{2\alpha\beta}
\end{pmatrix},
\end{equation*}
\begin{equation*}
\begin{pmatrix}
0 & b_\alpha\\
0 & 0
\end{pmatrix},
\end{equation*}
and
\begin{equation*}
\begin{pmatrix}
s_{1 \alpha}\\
0
\end{pmatrix}
\end{equation*}
that correspond to the triple $(E_1\oplus E_2, p_{E_1} \phi i_{E_2}, p_{E_1}(s))$ which lies in $F_{(1,1),\sigma}^{(d_1,d_2),1}$.
Observe also that:
\begin{equation*}
F_{(1,1),\sigma}^{(d_1,d_2),1}\subset
\left \{
\begin{array}{c}
(E,\phi,s) \text{ where } E \text{ is defined by}\\
0 \rightarrow E_2 \rightarrow E \rightarrow E_1 \rightarrow 0\\
s \notin H^0(E_2), \deg E_i =d_i, d_2>(d+\sigma)/2\\
\text{and } \phi \text{ does not preserve } E_2
\end{array} \right \}
\end{equation*}
as is immediately checked by definition. This proves that
\begin{equation*}
F_{(1,1),\sigma}^{(d_1,d_2),1+}\supset
\left \{
\begin{array}{c}
(E,\phi,s) \text{ where } E \text{ is defined by}\\
0 \rightarrow E_2 \rightarrow E \rightarrow E_1 \rightarrow 0\\
s \notin H^0(E_2), \deg E_i =d_i, d_2>(d+\sigma)/2\\
\text{and } \phi \text{ does not preserve } E_2
\end{array} \right \}.
\end{equation*}
The same kind of argument allows to prove
\begin{equation*}
F_{(1,1),\sigma}^{(d_1,d_2),2+}\supset
\left \{
\begin{array}{c}
(E,\phi,s) \text{ where } E \text{ is defined by}\\
0 \rightarrow E_2 \rightarrow E \rightarrow E_1 \rightarrow 0\\
s \in H^0(E_2), \deg E_i =d_i, d_1 > (d-\sigma)/2\\
\text{and } \phi \text{ does not preserve } E_2
\end{array} \right \}
\end{equation*}
and that the limit map has the claimed form.
To conclude we observe that both the decomposition into attracting set and the suggested decomposition cover $\mathcal{M}^{2,d}_\sigma$. The first one because limits for the $\mathbb{C}^*$-action as $\lambda \rightarrow 0$ always exists, the second one because it exhausts all the possible Harder-Narasimhan types of the underlying Bradlow pair of a triple in $\mathcal{M}^{2,d}_\sigma$. Therefore we can conclude that all the inclusions we proved are actually equalities.
\end{proof}
\end{theorem}
\chapter{General facts and background results}
\section{Grothendieck ring of varieties over $\mathbb{C}$ and motivic invariants}
In the rest of the thesis we will use a ring containing universal additive invariants of varieties over $\mathbb{C}$. Good references for this are \cite[Section 2]{behrend2007motivic} and \cite[Section 1]{garcia2011motives}.
\begin{defn}
Define $K_0(\text{Var}_\mathbb{C})$ to be the ring with the following presentation:
\begin{itemize}
\item the generators are the isomorphism classes $[X]$ of varieties over $\mathbb{C}$
\item the relations are generated by $[X]-[Y]-[X \setminus Y]$ whenever $Y$ is a closed subvariety of $X$
\item the product is defined by $[X]\cdot [Y]=[X \times Y]$ where $\times$ is the product of $\mathbb{C}$-schemes over $\Spec(\mathbb{C})$ (or in this case the Cartesian product of varieties).
\end{itemize}
$K_0(\text{Var}_\mathbb{C})$ is called \emph{Grothendieck ring of varieties over $\mathbb{C}$}.
\end{defn}
A few remarks are necessary. The relations extend to $[X]=[Y]+[X \setminus Y]$ when $Y$ is a locally closed subvariety of $X$. The unit of this ring is the class of a point $[pt]$, while the $0$ is the class of the empty variety. The class of the affine line $\mathbb{A}^1$ is denoted by $\L$.
Here we mention some relevant examples of relations in $K_0(\text{Var}_\mathbb{C})$ some of which will be used in the rest of the work.
\begin{itemize}
\item If $X \rightarrow Y$ is a Zariski locally trivial fibration with fiber $F$, then $[X]=[Y][F]$.
\item If $X \rightarrow Y$ is bijective on closed points then $[X]=[Y]$ and this is also true under the more general condition that $X$ and $Y$ can be written as disjoint unions $X= \sqcup X_i$, $Y=\sqcup Y_j$ with a bijection between the index sets such that $[X_i]=[Y_i]$. An example of this occurs if there are isomorphisms $X_i \rightarrow Y_i$ that do not extend to maps with larger domain than the $X_i$.
\item We could have defined the Grothendieck ring using isomorphism classes of schemes of finite type. However this is not important for our purposes and also, using this last definition it is easy to see that $[X]=[X_{red}]$.
\item Pick a smooth projective curve $C$ of genus $g \geq 2$. We can consider the symmetric powers of $C$, $S^n( C)$ with the Abel-Jacobi map:
\begin{align*}
AJ_n: S^n( C) & \rightarrow J^n( C)\\
D & \mapsto \mathcal{O}(D).
\end{align*}
It is a well known fact that the fibers of $AJ_n$ are projective spaces, not necessarily of constant dimension. More precisely $AJ_n^{-1}(\{L\})=\P H^0(L)$. It is also known that $J^n ( C)$ can be stratified in such a way that $AJ_n$ is a projective bundle on each of the strata. We can be more precise. Define the \emph{Brill-Noether locus}:
\begin{equation*}
V^n_i=\{L \in J^n( C) : \dim H^0(L)=i\}
\end{equation*}
which are locally closed subvarieties of $J^n( C)$. Then we have the motivic relations:
\begin{equation*}
[S^n ( C)]=\sum_{i=1}^{n+1} [V_i^n] [\mathbb{C} \P^{i-1}]
\end{equation*}
and
\begin{equation*}
[J^n ( C)]=\sum_{i=0}^{n+1} [V_i^n].
\end{equation*}
Note that, in general, the strata $V_i^n$ are very complicated, maybe singular, and they depend on the complex structure of $C$, i.e. on the specific curve and not only on the genus. The sum of their motives, as well as the sum of the motives weighted with the appropriate projective space, can be expressed nicely as we have seen above.\\
For $n>2g-2$, $AJ_n$ is a $\mathbb{C}\P^{n-g}$ bundle since $H^1$ will vanish for line bundles of degree bigger than $2g-2$ and hence the dimension on $H^0$ is constant.
\end{itemize}
$K_0(\text{Var}_\mathbb{C})$ is often completed to the ring $\widehat{K_0(\text{Var}_\mathbb{C})}$ in the following way. First invert $\L$ in $K_0(\text{Var}_\mathbb{C})$. Then inside the ring $K_0(\text{Var}_\mathbb{C})_\L=K_0(\text{Var}_\mathbb{C})[\L^{-1}]$ consider the filtration $F^\bullet$ defined by imposing that $F^m$ is generated by the elements:
\begin{equation*}
\frac{[X]}{\L^n}
\end{equation*}
with $\dim X -n \leq -m$. Then $\widehat{K_0(\text{Var}_\mathbb{C})}$ is obtained by completing $K_0(\text{Var}_\mathbb{C})[\L^{-1}]$ with respect to this filtration. We omit most of the details, but we just remark that in $\widehat{K_0(\text{Var}_\mathbb{C})}$, $\L^n-1$ is invertible for all $n$ since:
\begin{equation*}
\frac{1}{\L^n-1}=\L^{-n}(1+\L^{-n}+\dots).
\end{equation*}
Also $[\GL_n]$ is invertible for all $n$ since:
\begin{equation*}
[\GL_n]=(\L^n-1)(\L^n-\L)\dots(\L^n-\L^{n-1}).
\end{equation*}
The idea behind this completion is that we would like to be able to define the motive of a certain class of stacks. Namely Artin stacks, locally of finite type, whose geometric stabilizers are linear algebraic groups. A result by Kresch \cite{kresch1999cycle} states that all of these stacks admit a stratification whose strata are quotients of varieties by $\GL_n$ for different $n$. Note that even though the stratification is not necessarily finite, the properties of these stacks imply that the motive is well defined in the completion.
As we said in the beginning, $K_0(\text{Var}_\mathbb{C})$ is the ring of universal additive invariants associated to varieties. Recall that an \emph{additive invariant} $\theta$ with values in a ring $R$ is a function $K_0(\text{Var}_\mathbb{C})$ that associates to the isomorphism class of a variety $X$ the invariant $\theta(X) \in R$. $\theta$ must also satisfy the scissor relations, i.e. $\theta(X)=\theta(X \setminus Y)+\theta(Y)$ whenever $Y$ is a closed subvariety of $X$, and the product rule, i.e. $\theta(X \times Y)=\theta(X)\cdot \theta(Y)$. This is equivalent to asking that $\theta$ is a ring homomorphism $K_0(\text{Var}_\mathbb{C}) \rightarrow R$.
Some examples are:
\begin{itemize}
\item the Euler characteristic $\chi: K_0(\text{Var}_\mathbb{C}) \rightarrow \mathbb{Z}$
\item the $E$-polynomial defined as:
\begin{equation*}
E(X, x,y)=\sum_{p,q,j} (-1)^j h_c^{p,q,j}(X) x^p y^q
\end{equation*}
where $h_c^{p,q,j}(X)=\dim \left ( Gr^F_p Gr^W_{p+q} H_c^j(X)\right )$ are the compactly supported Hodge numbers of $X$. The additive nature of the $E$-polynomial follows from the exact sequence in compactly supported cohomology arising from the decomposition $X= U \cup X \setminus U$ of a variety into an open subset and its complement. For more details see \cite[Section 2]{hausel2008mixed}. Here $E$ takes values in $R=\mathbb{Z}[x,y]$.
\end{itemize}
As a last remark, we recall that if the cohomology of a variety $X$ is pure, meaning that the weight filtration on $H^i(X)$ is limited to the $i$-th weight, then the $E$ polynomial (which is additive) will determine both the Hodge polynomial:
\begin{equation*}
H(X, x,y,t)=\sum_{p,q,j} h^{p,q,j}(X) x^p y^q t^j
\end{equation*}
by the substitution:
\begin{equation*}
H(X, x,y,t)=(xyt)^{2 \dim (X)} E\left ( X, -\frac{1}{xt}, -\frac{1}{yt}\right )
\end{equation*}
and the Poincar\'e polynomial by the substitution:
\begin{equation*}
P(X,t)=t^{2 \dim (X)} E\left ( X, -\frac{1}{t}, -\frac{1}{t}\right ).
\end{equation*}
\section{Moduli spaces of Higgs bundles}
\label{higgs}
Let us fix a smooth projective curve $C$ over $\mathbb{C}$ of genus $g \geq 2$. Denote by $K$ the canonical bundle of $C$.
\begin{defn}[Higgs bundles and stability]
A \emph{Higgs bundle} on $C$ is a pair $(E,\phi)$ consisting of a vector bundle $E$ and a twisted endomorphism $\phi: E \rightarrow E \otimes K$. $\phi$ is also referred to as the \emph{Higgs field}.
The slope of a Higgs bundle $(E,\phi)$ is defined to be the slope of $E$, i.e. $\mu(E)=\deg E / \rk E$.\\
We say that a Higgs bundle is (semi)stable if for every $F \subset E$ proper $\phi$-invariant subbundle of $E$ we have:
\begin{equation*}
\mu(F) \operatornamewithlimits{<}_{(=)} \mu(E).
\end{equation*}
\end{defn}
After fixing the rank $r$ and the degree $d$, we can define the moduli space $\mathcal{M}^{r,d}$ of S-equivalence classes of (semi)stable Higgs bundles. We summarize here the main properties of $\mathcal{M}^{r,d}$. Good references for this section are \cite{hitchin1987self}, \cite{nitsure1991moduli} among many others. For a nice summary of the case $r=2$ see \cite{hausel2001geometry}.
$\mathcal{M}^{r,d}$ is a complex variety of dimension $2+2 r^2(g-1)$ and it is smooth when $(r,d)=1$. The Zariski tangent space to a point $(E,\phi) \in \mathcal{M}^{r,d}$ is given by the hypercohomology of the following two step complex:
\begin{align*}
\End E &\rightarrow \End E \otimes K\\
f & \mapsto [f, \phi].
\end{align*}
Note that the complex is invariant under Serre's duality and this is responsible, in the smooth case, for $\mathcal{M}^{r,d}$ being a symplectic variety.
There are several useful maps between these moduli spaces, as $d$ varies. Here we outline just a few that will be used in the following chapters.\\
Tensoring with a line bundle $L$:
\begin{align*}
\mathcal{M}^{r,d} &\rightarrow \mathcal{M}^{r,d+r \deg L}\\
(E,\phi) & \mapsto (E \otimes L, \phi)
\end{align*}
via the identification $\End E \cong \End (E \otimes L)$. This is an isomorphism with inverse given by tensoring with $L^*$. In particular $\mathcal{M}^{r,d}$ is isomorphic to $\mathcal{M}^{r,d'}$ when $d-d'$ is divisible by $r$. This for example implies that in rank 2, up to isomorphism, there are only two moduli spaces of (semi)stable Higgs bundles, namely $\mathcal{M}^{2,0}$ which is singular and $\mathcal{M}^{2,1}$ which is smooth.
There is also a form of Serre's duality for Higgs bundles:
\begin{align*}
\mathcal{M}^{r,d} &\rightarrow \mathcal{M}^{r,2r(g-1)-d }\\
(E,\phi) & \mapsto (E^* \otimes K, \phi)
\end{align*}
using the identification $\End E \cong \End (E^* \otimes K)$. This is also an isomorphism since it is an involution. As an example, for rank $3$, this implies that $\mathcal{M}^{3,1}$ and $\mathcal{M}^{3,2}$ are smooth and isomorphic, while $\mathcal{M}^{3,0}$ is singular.
$\mathcal{M}^{r,d}$ always contains a copy of the moduli space of S-equivalence classes of rank $r$ degree $d$ (semi)stable vector bundles over $C$, denoted by $N^{r,d}$, embedded as Higgs bundles with zero Higgs field. More precisely, let $(r,d)=1$ so that both $N^{r,d}$ and $\mathcal{M}^{r,d}$ are smooth. Then the fiber of the cotangent bundle $T^*N^{r,d}$ over $E \in N^{r,d}$ is canonically identified with $H^1(\End E)^*$ which, in turn, is canonically identified with $H^0(\End E \otimes K)$.
Observe that if $E \in N^{r,d}$ is a stable vector bundle, then certainly $(E,\phi)$ is a stable Higgs bundle regardless of $\phi$. This means that the embedding:
\begin{align*}
N^{r,d} &\rightarrow \mathcal{M}^{r,d}\\
E &\mapsto (E,0)
\end{align*}
extends to an embedding:
\begin{align*}
T^*N^{r,d} \rightarrow \mathcal{M}^{r,d}
\end{align*}
whose image is an open dense symplectic subset of $\mathcal{M}^{r,d}$.
$\mathcal{M}^{r,d}$ is not proper but admits a proper map to an affine space. The map is known as \emph{Hitchin morphism} and is defined as follows. Let $\mathcal{A}^r = H^0(K) \oplus H^0(K^2) \oplus \dots \oplus H^0(K^r)$ be the so called \emph{Hitchin base}. Define:
\begin{align*}
h^{r,d}: \mathcal{M}^{r,d} &\rightarrow \mathcal{A}^r\\
(E,\phi) &\mapsto \text{char poly}(\phi).
\end{align*}
The dimension of $\mathcal{A}^r$ is $1+r^2(g-1)$ which is exactly half of $\dim \mathcal{M}^{r,d}$. The map $h^r$ is proper complete integrable system whose generic fibers, as we will see later, are (torsors over) Abelian varieties. The proof of the properness can be found in \cite{nitsure1991moduli}, while the proof of second assertion can be found in \cite{hitchin1987stable}.
Beauville, Narasimhan, Ramanan (and many others extended their initial results) proved in \cite{beauville1989spectral} that $\mathcal{M}^{r,d}$ can be identified with a moduli space of stable sheaves on the cotangent bundle $T^*C=\text{Tot}(K)$ of $C$. Indeed we have the following equivalence:
\begin{equation*}
\left \{
\begin{array}{c}
\mathcal{F} \text{ pure one dimensional}\\
\text{sheaves of rank one on } T^*C\\
\text{whose support does not}\\
\text{intersect the divisor at infinity}
\end{array} \right \}
\longleftrightarrow
\left \{
\begin{array}{c}
(E,\phi) \text{ Higgs bundles on } C\\
\end{array} \right \}
\end{equation*}
given by the pushforward $\pi_*$ with respect to the projection map $\pi: T^*C \rightarrow C$. More precisely, given a sheaf $\mathcal{F}$ as above we can see that, since $\pi$ is affine, $\supp \mathcal{F}$ does not intersect the divisor at infinity and $\mathcal{F}$ is pure dimensional, $E=\pi_* \mathcal{F}$ is a torsion free sheaf on $C$ and therefore a vector bundle. With the hypotheses above, $\supp\mathcal{F}$ must be an $r$ to $1$ cover of the zero section in $T^*C$ and the rank of $E$ is equal to $r$. Since $E=\pi_* \mathcal{F}$ it also carries an action of the sheaf of algebras $\pi_* \mathcal{O}_{T^*C}=\Sym^* K^*$. This is the same as a morphism $K^* \otimes E \rightarrow E$ which can be identified with a Higgs field $\phi: E \rightarrow E \otimes K$.
Observe that $\mathcal{A}^r$ is the parameter space of a flat family of curves in $T^*C$ given by their equations. Namely the point $(\psi_1, \dots, \psi_r) \in \mathcal{A}_r$ corresponds to the curve of equation $y^r+\psi_r(x) y^{r-1}+\dots+\psi_1(x)=0$ where $x$ is the coordinate on the curve $C$ and $y$ is the coordinate on the fibers of the cotangent bundle of $C$. These curves are all $r$ to $1$ covers of the zero section of $T^*C$ and are generically smooth projective curves of genus $1+r^2(g-1)$. The curves parametrized by $\mathcal{A}^r$ are called \emph{spectral curves}. The family however contains curves that are not irreducible and even non-reduced. Under the identification of Higgs bundles on $C$ with one dimensional sheaves on $T^*C$, the Hitchin morphism will send a sheaf $\mathcal{F}$ to $\supp(\mathcal{F})$ which is a curve in $T^*C$. Therefore the statement of the BNR correspondence can be slightly improved by adding that if $(E,\phi)=\pi_* \mathcal{F}$ then $\supp (\mathcal{F})$ is the curve whose equation is given by $h^{r,d}(E,\phi)$. It is also important to note that the generic curve in the Hitchin base is integral and that if $(E,\phi)$ has a $\phi$-invariant proper subbundle $F$, then the characteristic polynomial of $\phi$ will factor and will be divisible by the characteristic polynomial of $\phi_{|F}$. These two facts together imply that the generic Higgs bundle in $\mathcal{M}^{r,d}$ will have no $\phi$-invariant subbundles and therefore will be trivially stable.
Let $X$ be a spectral curve. The fiber of $h^{r,d}$ over $X$ is easily seen to be (a torsor over) the Jacobian of $X$ if $X$ is smooth. If $X$ is singular but integral, then the Jacobian is replaced by the compactified Jacobian of $X$ which parametrizes torsion free sheaves on $X$. When $X$ is only reduced, meaning it could be singular and have more than one irreducible component, then the fiber of $h^{r,d}$ is a fine compactified Jacobian of $X$ for a specific choice of the stability condition
There is an action of $\mathbb{C}^*$ on $\mathcal{M}^{r,d}$ defined by:
\begin{align*}
\mathbb{C}^* \times \mathcal{M}^{r,d} &\rightarrow \mathcal{M}^{r,d}\\
(\lambda, (E, \phi)) & \mapsto (E, \lambda \phi).
\end{align*}
After we endow the Hitchin base with the action:
\begin{align*}
\mathbb{C}^* \times \mathcal{A}^r &\rightarrow \mathcal{A}^r\\
(\lambda, (\psi_1, \dots ,\psi_r)) & \mapsto (\lambda\psi_1, \lambda^2 \psi_2, \dots ,\lambda^r\psi_r),
\end{align*}
we can see that the Hitchin map has the property that $h^{r,d}(\lambda \cdot (E,\phi)=\lambda \cdot h^{r,d}(E,\phi)$.
This, together with the fact that $h^{r,d}$ is proper, implies that for all $(E,\phi) \in \mathcal{M}^{r,d}$:
\begin{equation*}
\lim_{\lambda \rightarrow 0} \lambda \cdot (E,\phi)
\end{equation*}
exists and lies in $\mathcal{M}^{r,d}$. Clearly, the limit points for the $\mathbb{C}^*$-action are the fixed points for the action. In general the fixed points have the following form:
\begin{prop}
\label{split}
Suppose that $(E,\phi)$ is a semistable Higgs bundle of rank $r$ such that $\lambda \cdot (E,\phi) \cong (E,\phi)$ for some $\lambda \in \mathbb{C}^*$ that is not a root of unity, then $E=E_1 \oplus \dots \oplus E_m$, for some $m \leq r$, $\sum \rk E_i=r$ and $\sum \deg E_i =\deg E$. Furthermore, $\phi$ has the property that $\phi(E_i) \subseteq E_{i-1} \otimes K$, $\phi(E_1)=0$.
\begin{proof}
See for example \cite[Lemma 4.1]{simpson1992higgs}.
\end{proof}
\end{prop}
For example, if $m=1$ in the proposition above, then $E$ is (semi)stable as a vector bundle and $\phi=0$ so that we find the moduli space of semistable vector bundles as one of the fixed point components for the $\mathbb{C}^*$-action. For a discussion about the cohomology of $\mathcal{M}^{r,d}$ see section \ref{semipr}.\\
\section{Semiprojective varieties and their cohomology}
\label{semipr}
The moduli space of Higgs bundles, as well as some of the moduli spaces of Bradlow-Higgs triples satisfy some properties that have strong implications on their cohomology. In this section we give a short overview of semiprojective varieties. All the material can be found in greater detail in \cite{hausel2013cohomology}.
\begin{defn}
Let $X$ be a quasi-projective complex variety with an action of $\mathbb{C}^*$. We call $X$ \emph{semiprojective} if the following two properties are satisfied:
\begin{itemize}
\item[(i)] the fixed point set $X^{\mathbb{C}^*}$ is proper
\item[(ii)] for all $x \in X$, the limit $\lim_{\lambda \rightarrow 0} \lambda \cdot x$ for $\lambda \in \mathbb{C}^*$ exists in $X$.
\end{itemize}
Let $X^{\mathbb{C}^*}=\bigsqcup_i F_i$ be the decomposition of the fixed point locus of $X$ into connected components, where $i$ ranges over some index set $I$. Define $U_i$ to be the set of points $x \in X$ for which $\lim_{\lambda \rightarrow 0} \lambda \cdot x $ lies in $F_i$. Define also $D_i$ to be the set of points $x \in X$ for which $\lim_{\lambda \rightarrow \infty} \lambda \cdot x $ lies in $F_i$.
\end{defn}
The $U_i$ are often referred to as \emph{attracting sets (or cells, or loci)} and the $D_i$ are referred to as \emph{downward flows}. Both the $U_i$ and the $D_i$ are locally closed subsets of $X$.\\
Note that if $X$ is semiprojective then condition (ii) implies the decomposition:
\begin{equation*}
X=\bigsqcup_i U_i.
\end{equation*}
The decomposition of $X$ into attracting sets is referred to as \emph{Bia\l{}ynicki-Birula decomposition}, after \cite[Theorem 4.1]{bialynicki1973some}.\\
We also give an important definition.
\begin{defn}
Let $X$ be a semiprojective variety. The \emph{core} of $X$ is:
\begin{equation*}
\mathcal{C}:= \cup_i D_i.
\end{equation*}
\end{defn}
According to \cite[Corollary 1.2.2]{hausel2013cohomology}, if $X$ is semiprojective then the core $\mathcal{C}$ is a proper subvariety of $X$. Moreover, according to \cite[Theorem 1.3.1]{hausel2013cohomology} the embedding $\mathcal{C} \rightarrow X$ induces an isomorphism $H^*(\mathcal{C},\mathbb{Z})\cong H^*(X,\mathbb{Z})$ when $X$ is smooth.
If we assume that $X$ is smooth, in addition to being semiprojective, then \cite[Corollary 1.3.2]{hausel2013cohomology} states that $X$ has pure cohomology. In fact, on one hand, $H^*(X)=H^*(\mathcal{C})$ is the cohomology of a proper variety so the weight filtration $W_\bullet$ on $H^*(X)$ has the property that $W_r H^i(X)=W_i H^i(X)=H^i(X)$ for $r \geq i$. On the other hand, since $X$ is smooth we have $W_r H^i(X)=0$ for $r<i$ and therefore $H^i(X)$ is pure of weight $i$. If $X$ is smooth, according to \cite[Corollary 1.3.6]{hausel2013cohomology} the core $\mathcal{C}$ is a deformation retract of $X$. The smoothness of $X$ also implies, see \cite[Theorem 4.1]{bialynicki1973some}, that both the $U_i$ and the $D_i$ are Zariski locally trivial affine fibrations over the $F_i$.
Thanks to the purity of $H^*(X)$ we see that the $E$-polynomial of $X$ will determine the Poincar\'e polynomial and hence we can deduce the Betti numbers from the motive, if we are able to compute it.
Let us now apply the previous discussion to $\mathcal{M}^{r,d}$. As we already observed, the limits as $\lambda \rightarrow 0$ of the $\mathbb{C}^*$ action applied to any point of $\mathcal{M}^{r,d}$ exist. This is a consequence of the equivariance and properness of the Hitchin map. Also, since the fixed point components of the $\mathbb{C}^*$-action are closed subvarieties of the fiber $(h^{r,d})^{-1}(\{0\})$ which is proper, the fixed point locus is proper. Finally, if $r$ and $d$ are coprime, then $\mathcal{M}^{r,d}$ is smooth.\\
Given an ordered partition $\underline{r}$ of $r$ and an ordered partition $\underline{d}$ of $d$, we can denote by $F_{\underline{r}}^{\underline{d}}$ the component of the fixed point locus in $\mathcal{M}^{r,d}$ whose points split as in proposition \ref{split}. Furthermore we can denote by $F_{\underline{r}}^{\underline{d},+}$ and by $F_{\underline{r}}^{\underline{d},-}$ the loci of $\mathcal{M}^{r,d}$ of those $(E,\phi)$ for which $\lim_{\lambda \rightarrow 0}\lambda \cdot (E,\phi) \in F_{\underline{r}}^{\underline{d}}$ and $\lim_{\lambda \rightarrow \infty} \lambda \cdot (E,\phi) \in F_{\underline{r}}^{\underline{d}}$ respectively. The notations given here are consistent with \cite{heinloth2014intersection}.
The previous discussion on smooth semiprojective varieties then implies the following.
\begin{prop}
Let $(r,d)=1$. Then:
\begin{itemize}
\item The fixed point components of the $\mathbb{C}^*$-action on $\mathcal{M}^{r,d}$ are the $F_{\underline{r}}^{\underline{d}}$ (when they are non-empty), which are all contained in the fiber $(h^{r,d})^{-1}(\{0\})$ of the Hitchin map.
\item The limit maps $F_{\underline{r}}^{\underline{d},+} \rightarrow F_{\underline{r}}^{\underline{d}}$ and $F_{\underline{r}}^{\underline{d},+} \rightarrow F_{\underline{r}}^{\underline{d}}$ are Zariski locally trivial fibrations whose fibers are isomorphic to affine spaces.
\item $\mathcal{M}^{r,d}=\bigcup F_{\underline{r}}^{\underline{d},+}$ and $(h^{r,d})^{-1}(\{0\})=\bigcup F_{\underline{r}}^{\underline{d},-}$. In particular the closure of the $F_{\underline{r}}^{\underline{d},-}$ are the irreducible components of $(h^{r,d})^{-1}(\{0\})$ and $(h^{r,d})^{-1}(\{0\})$ is the core of $\mathcal{M}^{r,d}$.
\item The dimension of the fibers of $F_{\underline{r}}^{\underline{d},+} \rightarrow F_{\underline{r}}^{\underline{d}}$ is constant and equal to $1+r^2(g-1)$.
\item There is a motivic equality $\left [\mathcal{M}^{r,d} \right ]=\L^{1+r^2(g-1)} \cdot \sum \left [ F_{\underline{r}}^{\underline{d}}\right ]$.
\end{itemize}
\begin{proof}
Follows from the previous discussion. A proof of this is also found in \cite[Proposition 2.1]{garcia2011motives}.
\end{proof}
\end{prop}
Let us briefly recall how to derive the structure of the attracting sets $F_{\underline{r}}^{\underline{d},+}$ in $\mathcal{M}^{r,d}$. Fix $r,d$ coprime. Let us consider a fixed point $(E,\phi)\in F_{\underline{r}}^{\underline{d}}$. If $r=(r)$ then $\phi=0$ and $E$ is a stable vector bundle. Since $\dim H^0(K \End E)=1/2 \dim \mathcal{M}^{r,d}$ then we know that the tangent space at $(E,\phi)$ decomposes into two parts of the same dimension, one of which is acted on with weight 0 by $\mathbb{C}^*$ and the other one with positive weight.
In all other cases, when $E=E_1 \oplus \dots \oplus E_m$ is split and has the form above. By the definition of hypercohomology we see that $T_{(E,\phi)} \mathcal{M}^{r,d}$ is generated by the kernel of the map:
\begin{align*}
C^1(\End E) \oplus C^0(K\End E) &\rightarrow C^1(K \End E)\\
(\tau, \nu) & \mapsto [\tau,\phi]+d \nu.
\end{align*}
The action of $\mathbb{C}^*$ induced on $T_{(E,\phi)} \mathcal{M}^{r,d}$ is then given by the pullback with respect to $f$, where $f: E \rightarrow E$ is the (one of the) automorphism of $E$ giving the isomorphism $(E,\lambda \phi) \cong (E,\phi)$. More explicitly if $E=E_1 \oplus \dots \oplus E_m$ and $\phi(E_i) \subseteq E_{i-1} \otimes K$ then $f$ is diagonal and defined by $f_{E_i}=\lambda^i \mathbb{I}_{E_i}$. Therefore we see that:
\begin{align*}
\lambda(\tau,0)=\lambda^{j-i}(\tau,0) &\text{ if } \tau \in C^1(E_i^* E_j)\\
\lambda(0,\nu)=\lambda^{j-i+1}(0,\nu) &\text{ if } \nu \in C^0(E_i^* E_j K).
\end{align*}
An easy way to remember the exponents above is to imagine that $\mathbb{C}^*$ acts with weight $i$ on $E_i$ and with weight $1$ on $K$. We can then split the deformation complex:
\begin{equation*}
\End E \rightarrow K \End E
\end{equation*}
into parts of positive, zero and negative weights according to the exponents above.
The splitting is as follows:
\begin{equation*}
\bigoplus_{i=1}^{m} \Hom(E_i,E_i) \rightarrow \bigoplus_{i=2}^{m} \Hom(E_i,E_{i-1} \otimes K)
\end{equation*}
gives the 0 weight part,
\begin{equation*}
\bigoplus_{i<j} \Hom(E_i,E_j) \rightarrow \bigoplus_{i\leq j} \Hom(E_i,E_{j} \otimes K)
\end{equation*}
gives the positive weight part and
\begin{equation*}
\bigoplus_{i>j}^{m} \Hom(E_i,E_j) \rightarrow \bigoplus_{i>j+1}^{m} \Hom(E_i,E_{j} \otimes K)
\end{equation*}
gives the negative weight part.
We can, for example, compute the positive weight part of the tangent space at $(E,\phi)$ by writing the long exact sequence associated to the positive weight complex:
\begin{align*}
0 &\rightarrow \bigoplus_{i<j} \Hom(E_i,E_j) \rightarrow \bigoplus_{i\leq j} \Hom(E_i,E_{j} \otimes K) \rightarrow T^+_{(E,\phi)} \mathcal{M}^{r,d} \rightarrow\\
&\rightarrow \bigoplus_{i<j} \Ext^1(E_i,E_j) \rightarrow \bigoplus_{i\leq j} \Ext^1(E_i,E_{j} \otimes K) \rightarrow \mathbb{C} \rightarrow 0.
\end{align*}
From this we can compute the dimension of the positive weight part using the formula:
\begin{equation*}
\dim T^+_{(E,\phi)} \mathcal{M}^{r,d} =1+ \sum_{i \leq j} \chi( \Hom(E_i,E_{j} \otimes K) ) - \sum_{i<j} \chi( \Hom(E_i,E_{j}))=1+(g-1)r^2.
\end{equation*}
Here we see that, as it should be, the dimension of the positive weight part is constant, i.e. it does not depend on $\underline{r}$ and $\underline{d}$, and is always half of the dimension of the moduli space.
Since $\mathcal{M}^{r,d}$ is smooth, \cite[Theorem 4.1]{bialynicki1973some} implies that $F_{\underline{r}}^{\underline{d},+} \rightarrow F_{\underline{r}}^{\underline{d}}$ is a Zariski locally trivial affine fibration whose fibers have dimension equal to the dimension of the positive weight part of the $\mathbb{C}^*$ induced on the tangent space at one of the fixed points, which is what we computed with the above complexes.
\section{Character varieties and P = W}
\label{charvar}
For all details in the forthcoming discussion we refer to \cite{de2010topology} and \cite{hausel2008mixed}. Define the following variety.
\begin{defn}
Let $C$ be a smooth projective curve of genus $g \geq 2$. Let $\mathcal{M}^{2}_B$ be the variety defined by the following GIT quotient:
\begin{equation*}
\mathcal{M}^{2}_B = \left \{ A_1,\dots, A_g, B_1,\dots, B_g \in \GL_2(\mathbb{C}) : [A_1, B_1] \dots [A_g, B_g]=-\mathbb{I}\right \} \git \GL_2(\mathbb{C})
\end{equation*}
where $\GL_2(\mathbb{C})$ acts by conjugation. Such a variety is known as (rank 2) \emph{twisted character variety}.
\end{defn}
Note that $\mathcal{M}_B^2$ parametrizes isomorphism classes of representations of the fundamental group of $C$ into $\GL_2(\mathbb{C})$ and is a purely topological invariant of $C$, i.e. it does not depend on the complex structure on the curve but only the genus. Also, $\mathcal{M}^2_B$ is a smooth affine variety since it is obtained from a GIT quotient of an affine space.
From the non-Abelian Hodge theorem (see for example \cite[proposition 10]{simpson1990nonabelian}) follows that $\mathcal{M}^{2,1}$ is canonically diffeomorphic to $\mathcal{M}^2_B$.
The cohomology of $\mathcal{M}^2_B$ is not pure and carries a nontrivial weight filtration. Let $Y$ be a complex variety. Recall that, if we denote by $F^\bullet$ the decreasing Hodge filtration on $H^*(Y)$ and by $W_\bullet$ the increasing weight filtration on $H^*(Y)$, we can define the mixed Hodge numbers:
\begin{equation*}
h^{p,q,j}(Y)=\dim (Gr_p^F Gr_{p+q}^W H^j(Y))
\end{equation*}
and the Hodge polynomial:
\begin{equation*}
H(Y,x,y,t)=\sum_{p,q,j} h^{p,q,j}(Y)x^p y^q t^j.
\end{equation*}
From \cite[corollary 4.1.11]{hausel2008mixed} we see that $\mathcal{M}^2_B$ has the property that $h^{p,q,j}(\mathcal{M}^2_B)=0$ if $p \neq q$. In particular $Gr_{2i+1}^W H^*(\mathcal{M}^2_B)=0$. We can define a variant of the mixed Hodge polynomial:
\begin{equation*}
H(\mathcal{M}^2_B,q,t)=H(\mathcal{M}^2_B,\sqrt{q},\sqrt{q},t).
\end{equation*}
On the other hand $\mathcal{M}^{2,1}$ has pure cohomology, so carries trivial weight filtration, but the Hitchin map $h^{2,1}: \mathcal{M}^{2,1} \rightarrow \mathcal{A}^2$ allows to define a \emph{perverse filtration} (see \cite[section 1.4]{de2010topology}).
Consider the object $\mathbb{R} h^{2,1}_* \mathbb{Q} \in D_c^b(\mathcal{A}^2)$. Similarly to the case of the perverse filtration on the cohomology of the compactified Jacobian, $\mathbb{R} h^{2,1}_* \mathbb{Q}$ is filtered by its perverse truncations $\leftidx{^p}{\tau}{^{\leq i}}\mathbb{R} h^{2,1}_* \mathbb{Q}$ and the perverse filtration consists of the images:
\begin{equation*}
P_i H^*(\mathcal{M}^{2,1})= \im \left ( H^*\left (\leftidx{^p}{\tau}{^{\leq i}}\mathbb{R} h^{2,1}_* \mathbb{Q} \right) \rightarrow H^*\left (\mathbb{R} h^{2,1}_* \mathbb{Q} \right) \right ).
\end{equation*}
The following theorem is central:
\begin{theorem}[{\cite[theorem 1.1.1]{de2010topology}}]
The diffeomorphism given by non-abelian Hodge theory $\mathcal{M}^{2,1} \rightarrow\mathcal{M}^2_B$ induces an isomorphism in cohomology $H^*(\mathcal{M}^{2,1}) \rightarrow H^*(\mathcal{M}^2_B)$ with the property that:
\begin{equation*}
W_{2k}H^*(\mathcal{M}^2_B)=W_{2k+1}H^*(\mathcal{M}^2_B)=P_k H^*(\mathcal{M}^{2,1}).
\end{equation*}
\end{theorem}
The previous result is also known as "P = W", from the name of the filtrations involved. As of now, it is proved only for rank 2, while for higher rank it remains a conjecture. If we define the perverse Hodge polynomial by:
\begin{equation*}
PH(\mathcal{M}^{2,1},q,t)=\sum_{k,j} \dim P_k H^j(\mathcal{M}^{2,1}) q^k t^j
\end{equation*}
then the P = W result implies:
\begin{cor}
\begin{equation*}
H(\mathcal{M}^2_B,q,t) = PH(\mathcal{M}^{2,1},q,t).
\end{equation*}
\end{cor}
In \cite[Theorem 1.1.3]{hausel2008mixed} we find the computation of the mixed Hodge polynomial of $\mathcal{M}_B^2$:
\begin{align*}
H(\mathcal{M}_B^2,q,t)&=
( 1+qt )^{2g} \left ({\frac { ( 1+q^2t^3)^{2g}}{ ( 1-q^2t^2 )( 1-q^2t^4) }}+{\frac {q^{2g-2}t^{4g-4}( 1+q^2t)^{2g}}{ ( 1-q^2)( 1-q^2t^2) }} \right )+\\
&-\frac{1}{2} q^{2g-2} t^{4g-4}( 1+qt)^{2g} \left ( {\frac {( 1+qt )^{2g}}{( 1-qt^2 )( 1-q ) }}+ \frac{( 1-qt )^{2g}}{( 1+qt^2)( 1+q ) } \right ).
\end{align*}
which, in turn, is equal to $PH(\mathcal{M}^{2,1},q,t)$.
\section{Higgs bundles with poles}
\label{hbpoles}
There is a variant of Higgs bundles that is also of interest for us. Choose once and for all a point $P \in C$.
\begin{defn}
\label{gammahiggs}
A pair $(E,\phi)$ is a \emph{Higgs $\gamma$-bundle} for some integer $\gamma \geq 0$ if $E$ is a vector bundle on $C$ and $\phi: E \rightarrow E \otimes K(\gamma P)$ is a twisted endomorphism of $E$.
\end{defn}
The definition is very similar to that of a Higgs bundle, and actually is the same when $\gamma=0$. Stability is the defined in the same way as for Higgs bundles and we denote by $\mathcal{M}^{r,d}(\gamma)$ the moduli space of S-equivalence classes of semistable Higgs $\gamma$-bundles of rank $r$ and degree $d$. Higgs $\gamma$-bundles can be thought as Higgs bundles whose Higgs field is allowed to have a pole of order at most $\gamma$ at $P$. The choice of putting $\gamma$ in parentheses instead of at the subscript will be apparent when we introduce Bradlow-Higgs triples. The construction of Higgs $\gamma$-bundles (and more general twists) can be found in \cite{nitsure1991moduli}.
Most of the properties of the moduli space of Higgs bundles carry over to Higgs $\gamma$-bundles. In particular we have semiprojectivity, smoothness for $r$ and $d$ coprime, the fixed points of the $\mathbb{C}^*$-action have the same form. There is a slight difference in the BNR correspondence, since Higgs $\gamma$-bundles will now correspond to sheaves on the total space of $K(\gamma P)$. This difference also carries to the Hitchin base $\mathcal{A}^{r,d}(\gamma)= H^0(K(\gamma P)) \oplus H^0(K^2(2\gamma P)) \oplus \dots \oplus H^0(K^r(r\gamma P))$ and to the Hitchin map $h^{r,d}_\gamma$.
It is worth observing that since we have a canonical map $\mathcal{O} \rightarrow \mathcal{O}(P)$ we also have embeddings $\mathcal{M}^{r,d}(\gamma) \rightarrow \mathcal{M}^{r,d}(\gamma+1)$. This approach, for $r=2$, is carried over in \cite[Chapter 7]{hausel2001geometry}. There we find the definition of $ \mathcal{M}^{2,d}(\infty)$ as the direct limit of the increasing embeddings $\mathcal{M}^{2,d}(\gamma) \rightarrow \mathcal{M}^{2,d}(\gamma+1)$.
The Poincar\'{e} polynomial of $\mathcal{M}^{2,d}(\infty)$ is then computed in \cite[Section 7.2]{hausel2001geometry}:
\begin{equation*}
P(\mathcal{M}^{2,d}(\infty),t) =\lim_{\gamma \rightarrow \infty} P(\mathcal{M}^{2,d}(\gamma) ,t)=\frac{(1+t)^{2g}(1+t^3)^{2g}}{(1-t^2)(1-t^4)}=P(B\overline{\mathcal{G}},t)
\end{equation*}
where $B\overline{\mathcal{G}}$ is the classifying space of a gauge group $\mathcal{G}$ modulo scalars. In the same section it is also proved that $\mathcal{M}^{2,d}(\infty)$ and $B\overline{\mathcal{G}}$ are homotopically equivalent, but we will not attempt the same with Bradlow-Higgs triples.
\section{Moduli spaces of Bradlow pairs and wall crossing}
\label{Bradlow}
Let us first define Bradlow pairs on a smooth projective curve $C$ over $\mathbb{C}$. The main reference for rank 2 Bradlow pairs will be \cite{thaddeus1994stable}. In \cite{mozgovoy2013moduli} it is possible to find the computation of the motives of the moduli spaces of Bradlow pairs for arbitrary rank and degree. Further comments on this approach will be made in the last chapter.
\begin{defn}[Bradlow pairs and stability]
A \emph{Bradlow pair} $(E,s)$ is a pair consisting of a vector bundle $E$ on $C$ and a nonzero section $s \in H^0(C,E)$.
Let $\sigma >0$ be a real number. We say $(E,s)$ is \emph{$\sigma$-(semi)stable} if for every $F \subset E$ proper subbundle we have:
\begin{align*}
&\frac{\deg(F)+\sigma}{\rk(F)}\operatornamewithlimits{<}_{(=)} \frac{\deg(E)+\sigma}{\rk(E)} &\text{if } s \in H^0(F)\\
&\frac{\deg(F)}{\rk(F)} \operatornamewithlimits{<}_{(=)}\frac{\deg(E)+\sigma}{\rk(E)} &\text{if } s \notin H^0(F).\\
\end{align*}
\end{defn}
There are several other equivalent notions of Bradlow pairs and stability. Here we outline some:
\begin{itemize}
\item we say $(E,s)$ is $\tau$-(semi)stable if for all $F \subseteq E$ subbundles we have:
\begin{align*}
&\mu(F) \operatornamewithlimits{<}_{(=)} \tau & \\
&\mu(E/F) \operatornamewithlimits{>}_{(=)} \tau &\text{if } s \in H^0(F).\\
\end{align*}
This is an equivalent notion of stability ($\tau = (d+\sigma)/r$) mentioned in \cite{mozgovoy2013moduli}.
\item A Bradlow pair can be equivalently defined as the datum of a map $s: \mathcal{O} \rightarrow E$. This interpretation falls into a more general framework of triples $\mathcal{E}=(f:E_0 \rightarrow E_1)$. There is a notion of $\gamma$-slope for this kind of triples:
\begin{equation*}
\mu_\gamma(\mathcal{E})=\frac{\deg E_0 +\deg E_1+\gamma \rk E_0}{\rk E_0+\rk E_1}
\end{equation*}
and $\mathcal{E}$ is called $\gamma$-(semi)stable if for all subtriples $\mathcal{F} \subset \mathcal{E}$ we have
\begin{equation*}
\mu_\gamma(\mathcal{F}) \operatornamewithlimits{<}_{(=)} \mu_\gamma(\mathcal{E}).
\end{equation*}
This notion of stability is equivalent to the original one via $\gamma= \sigma+ (d+\sigma)/r$. This stability is mentioned in \cite{mozgovoy2013moduli}.
\end{itemize}
If we fix a $\sigma>0$, a rank $r$ and a degree $d$ we can define the moduli space of S-equivalence classes of $\sigma$-(semi)stable Bradlow pairs $M_\sigma^{r,d}$.
Note that for $d < 0$ or for $\sigma > d/(r-1)$ the moduli space is empty since the line bundle generated by the section $s$ is destabilizing. In the interval $[0,d/(r-1)]$ there is a finite set of values of $\sigma$ for which the stability inequalities can possibly become equalities. We call these values \emph{critical values}. For $\sigma$ different from the critical values, semistability and stability coincide and $M_\sigma^{r,d}$ is a smooth projective variety of dimension $d+(r^2-r)(g-1)$. See e.g. \cite{thaddeus1994stable} for the deformation theory of the rank 2 case.
It is also clear that, for $\sigma$ lying in the interval between two consecutive critical values, the inequalities for $\sigma$-stability will all be equivalent and therefore $M_\sigma^{r,d}$ only depends on the interval and not on the specific value of $\sigma$.
As we already noticed for $\sigma>d/(r-1)$ the moduli space of Bradlow pairs is empty. Near the other extremal value, meaning for $\sigma$ very close to $0$, $\sigma$-stability of a Bradlow pair $(E,s)$ implies semistability of the underlying vector bundle $E$. Therefore we have an \emph{Abel-Jacobi map}:
\begin{align*}
AJ: M_\sigma^{r,d} & \rightarrow N^{r,d}\\
(E,s) &\mapsto E
\end{align*}
In this notation we avoid as much as possible putting too many decorations on $AJ$ and it should be clear from the context what ranks and degrees we are referring to. Note that if $r$ and $d$ are coprime and $\sigma$ is small, then $\sigma$-stability implies stability of $E$ and $AJ^{-1}(E)=\P H^0(E)$. In particular, in this case, $AJ$ is a projective bundle for $d > r(2g-2)$.
For small $\sigma$ and $r$, $d$ not coprime the fibers of $AJ$ are more complicated. This is due to the fact that if $(E,s)$ is $\sigma$-stable, then it is possible that $E$ is strictly semistable and therefore $AJ$ will not only forget about the section, but will lose extra extension information about $E$ keeping only its S-equivalence class.
Let us discuss the role of critical values. First we introduce a notation that will be used throughout the whole thesis.
\begin{defn}
Fix a critical value $\bar \sigma$. Denote by $\bar\sigma_+$ any value of $\sigma$ bigger than $\bar\sigma$ but smaller than the consecutive critical value. Similarly $\bar\sigma_-$ will denote a value of $\sigma$ that is smaller than $\bar\sigma$ but bigger than the previous critical value.
\end{defn}
If $\sigma$ crosses $\bar\sigma$ and is increasing, then the first inequality for the stability becomes stronger, while the second one will become weaker. This means that as $\sigma$ crosses $\bar\sigma$, some pairs that were $\bar\sigma_-$-stable will become $\bar\sigma_+$-unstable and viceversa.
Let us see the rank 2 case, which is also the most relevant for the present thesis. The main reference here is of course \cite{thaddeus1994stable} even though some of the notations might vary. For rank 2 it is easy to see that the critical values are all the integers in $[0,d]$ with the same parity as $d$. Let us pick one of the critical values $\bar\sigma$.
There are two families $\P W^{d,+}_{\bar\sigma}$ and $\P W^{d,-}_{\bar\sigma}$ that parametrize the so called \emph{flip loci}, i.e. where $M^{2,d}_{\bar\sigma_-}$ and $M^{2,d}_{\bar\sigma_+}$ differ. More precisely, $\P W^{d,+}_{\bar\sigma}$ parametrizes pairs $(E,s)$ that are $\bar\sigma_-$-stable but not $\bar\sigma_+$-stable. This is equivalent to say that $E$ is a nonsplit extension
\begin{equation*}
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0
\end{equation*}
with $L$ and $M$ two line bundles satisfying $\deg L =(d-\bar\sigma)/2$ and $\deg M =(d+\bar\sigma)/2$. Furthermore, $s \in H^0(L)$. Here $L$ is the canonical $\bar\sigma_+$-destabilizing subbundle of $(E,s)$. The condition of the extension being non-split is equivalent to the pair being $\bar\sigma_-$-stable.\\
There is a map:
\begin{align*}
\P W^{d,+}_{\bar\sigma} & \rightarrow S^{(d-\bar\sigma)/2} ( C) \times J^{(d+\bar\sigma)/2} ( C) \\
(E,s) & \mapsto (L,s, M)
\end{align*}
sending $(E,s)$ to the divisor of $s \in H^0(L)$ (which also retrieves $L$ itself) and to $M$, the quotient of $E$ by $L$. The fibers of the previous map are the projectivized extension spaces $\Ext^1(M,L)$. It is easy to check that the previous map is actually a projective bundle of rank $\bar\sigma+g-2$. In particular if we want to compute the motive of $\P W^{d,+}_{\bar\sigma}$ we get:
\begin{equation*}
[\P W^{d,+}_{\bar\sigma}]=[S^{(d-\bar\sigma)/2} ( C)] [J^{(d+\bar\sigma)/2} ( C)][\mathbb{C}\P^{\bar\sigma+g-2}].
\end{equation*}
$\P W^{d,-}_{\bar\sigma}$ instead parametrizes pairs $(E,s)$ that are $\bar\sigma_+$-stable but not $\bar\sigma_-$-stable. This is equivalent to say that $E$ is an extension
\begin{equation*}
0 \rightarrow M \rightarrow E \rightarrow L \rightarrow 0
\end{equation*}
where $L$ and $M$ are two line bundles with $\deg L =(d+\bar\sigma)/2$, $\deg M =(d-\bar\sigma)/2$ and $s \notin H^0(M)$. Note that, in this case, $s \notin H^0(M)$ is equivalent to $p(s)=\bar{s} \in H^0(L)$ being nonzero, where $p: E \rightarrow L$ is the projection. Once again, there is a map:
\begin{align*}
\P W^{d,-}_{\bar\sigma} & \rightarrow S^{(d-\bar\sigma)/2} ( C) \times J^{(d+\bar\sigma)/2} ( C) \\
(E,s) & \mapsto (L,\bar{s}, M)
\end{align*}
but this time $M$ is the canonical $\bar\sigma_-$-destabilizing subobject of $(E,s)$ and the divisor in $S^{(d-\bar\sigma)/2} ( C)$ is the divisor $D$ of the projection of $s$ on the canonical quotient $E/M$. The fibers of this map are the projectivization of the vector spaces $H^0(M \mathcal{O}_D)$ that parametrize extensions as above plus a lift of the section $\bar{s} \in H^0(L)$ to $s \in H^0(E)$. Again, this map is a projective bundle of rank $(d-\bar\sigma)/2 -1$. From this follows the relation:
\begin{equation*}
[\P W^{d,-}_{\bar\sigma}]=[S^{(d-\bar\sigma)/2} ( C)] [J^{(d+\bar\sigma)/2} ( C)][\mathbb{C}\P^{(d-\bar\sigma)/2-1}].
\end{equation*}
Observe that we have the motivic relation
\begin{equation*}
[M^{2,d}_{\bar\sigma_-}] - [\P W^{d,+}_{\bar\sigma}]=[M^{2,d}_{\bar\sigma_+}] - [\P W^{d,-}_{\bar\sigma}]
\end{equation*}
since outside the two flip loci, the two moduli spaces are isomorphic.
The previous phenomenon, involving a varying stability condition together with sets of critical values whose crossing makes the moduli problem change geometry, is known as \emph{wall crossing}.
\section{Hilbert schemes of points}
Here we introduce a geometric object that will be of interest in the last chapter of the thesis, namely the Hilbert scheme of points. A general introduction can be found in \cite{gottsche1994hilbert} and \cite{grothendieck1960techniques}, for more specific results about Hilbert schemes of points on a surface we refer to \cite{nakajima1999lectures}, while for a recollection on the properties of Hilbert schemes of singular curves and their relative version we refer to \cite{migliorini2011support}, \cite{maulik2014macdonald} and \cite{migliorini2015support}.
The most general definition of the Hilbert scheme, due to Grothendieck, passes through the functor of points. Since we don't need this much generality we just quickly sketch it.
Let $X$ be a scheme and $P$ a polynomial with integer coefficients. $\hilb^P(X)$ is a scheme (note that the existence of such a scheme is already a theorem by Grothendieck) such that maps $U \rightarrow \hilb^P(X)$ from another scheme $U$ are canonically identified with closed subschemes $Z \subset X \times U$ that are flat over $U$ and such that for every $u \in U$ the fiber $Z_u$ has Hilbert polynomial equal to $P$. In other words $\hilb^P(X)$ is the moduli space of subschemes of $X$ whose Hilbert polynomial is $P$.
For the purpose of some of the work in this thesis we just need to outline some of the properties of $\hilb^P(X)$ when $P$ is a constant polynomial and $X$ is either a smooth surface or a projective curve over $\mathbb{C}$.
When $P=n$ is a constant, $\hilb^P(X)$ is also commonly denoted by $X^{[n]}$. In this case the Hilbert scheme parametrizes subschemes of $X$ which are $0$ dimensional and whose space of global sections has dimension $n$. We can think of these subschemes as $n$-tuples of points and the Hilbert scheme in this case is called \emph{Hilbert scheme of points}.
For the case of $X$ a smooth surface we refer to \cite{nakajima1999lectures} for the proof of the main results and to the many works of G\"{o}ttsche \cite{gottsche1990betti,gottsche2000motive,gottsche1994hilbert} and G\"{o}ttsche, Soergel \cite{gottsche1993perverse} for the results on the cohomology.
When $X$ is a smooth surface then $X^{[n]}$ is a smooth variety of dimension $2n$ and is projective if $X$ itself is projective. Let $X^{(n)}$ or $\Sym^n X$ denote the $n$-fold product $X^n$ divided by the action of the symmetric group of the set of cardinality $n$ permuting the factors. There exists a map, called \emph{Hilbert-Chow morphism}:
\begin{equation*}
\pi_X^{n}: X^{[n]} \rightarrow X^{(n)}
\end{equation*}
sending a subscheme of length $n$ in $X$ to its support. This map is an isomorphism over the locus of $X^{(n)}$ where the points are all distinct, but has more complicated fibers when some points coincide. The fiber $(\pi_X^{n})^{-1}\{(P, \dots , P)\}$ for $P \in X$ is called the \emph{punctual Hilbert scheme of $n$-points} and denoted by $X^{[n]}_P$ or $\hilb_p^n(X)$, among the most common. It parametrizes subschemes of length $n$ whose support lies entirely in $P$. In other words it parametrizes the possible structures of subscheme of length $n$ that can be given to the point $P \in X$. The fibers of the Hilbert-Chow map are in general products of punctual Hilbert schemes. The fundamental property of the Hilbert-Chow morphism is that it is actually a resolution of the singular space $X^{(n)}$.
For the case of a projective curve we refer to \cite{migliorini2011support,maulik2014macdonald} for the integral case and to \cite{migliorini2015support} for the reduced case.
If $X$ is a projective locally planar curve then $X^{[n]}$ is a projective scheme over $\mathbb{C}$. The locally planar property for $X$ insures that it can be locally embedded in a surface and this implies some regularity for $X^{[n]}$. In general Hilbert schemes of points of varieties of dimension at least $3$ can be extremely complicated.
For the case of a curve $X$ we still have the Hilbert-Chow morphism $\pi_X^{n}: X^{[n]}_{red} \rightarrow X^{(n)}$, that is only defined on the reduction of $X^{[n]}$ which might be non-reduced if $X$ is non-reduced. If $X$ is smooth, then the Hilbert scheme of $n$-points is isomorphic to the symmetric power $X^{(n)}$ and the Hilbert-Chow morphism is an isomorphism.
If $X$ is integral then $X^{[n]}$ is integral of dimension $d$ and locally complete intersection. If we have a flat family of projective integral curves $f: \mathcal{X} \rightarrow B$ then there exists a variety $\mathcal{X}^{[n]}$ together with a proper map $f^{[n]}: \mathcal{X}^{[n]} \rightarrow B$ such that $(f^{[n]})^{-1}(b)=(f^{-1}(b))^{[n]}$ for all $b \in B$. $\mathcal{X}^{[n]}$ is called the \emph{Hilbert scheme of $n$-points relative to $B$}. If $X$ is singular, in general $X^{[n]}$ can be singular as well and the Hilbert-Chow morphism could have very complicated fibers. In \cite{ran2004note}, for example, it is proved that for the simplest possible planar singularity, i.e. the simple node $xy=0$, the punctual Hilbert scheme is a chain of $\mathbb{C}\P^1$'s touching at a point.
For reduced curves $X$ the Hilbert scheme behaves similarly to the case of integral curves but certainly has more than one component if the curve $X$ does. Not much is known about the Hilbert scheme of points of a non-reduced curve.
\section{The CKS complex}
\label{cks}
A more precise reference for this section is \cite[section 3.4]{migliorini2015support}. Let us assume that $\pi: \mathcal{C} \rightarrow B$ is a locally versal family of curves whose central fiber is the curve $\mathcal{C}_b$ which is nodal and reduced. We are not interested in the definition of locally versal family, as will become clear soon, just assume that it is a nice enough family.
The goal of this section is to introduce a tool to compute the stalk of
\begin{equation*}
IC(\bigwedge^i \mathbb{R}^1 \pi_{sm*} \mathbb{Q})
\end{equation*}
where $\pi_{sm}$ is the restriction of $\pi$ to the locus of $B$ where the curves are smooth.
To the nodal curve $\mathcal{C}_b$ we can associate some combinatorial data, according to \cite[section 3]{migliorini2015support}. Denote by $\Gamma$ the graph of the curve, i.e. the set of vertices $V$ is determined by the components of $\mathcal{C}_b$ and the set of edges $E$ is formed by adding an edge between two vertices (not necessarily distinct) if the corresponding components intersect in a node. We will also denote by $\varnothing$ and $\exists$ the vector spaces spanned by the vertices and edges respectively.
Let us denote by $\mathcal{C}_\eta$ the versal deformation of our nodal curve. Then on $H^1(\mathcal{C}_\eta,\mathbb{Q})$ is defined a monodromy weight filtration $W$ with the following identifications:
\begin{align*}
&Gr_0^W H^1(\mathcal{C}_\eta,\mathbb{Q})=H^1(\Gamma)\\
&Gr_1^W H^1(\mathcal{C}_\eta,\mathbb{Q})=H^1(\mathcal{C}_b^{\nu})\\
&Gr_2^W H^1(\mathcal{C}_\eta,\mathbb{Q})=H_1(\Gamma) \otimes \L
\end{align*}
where $H^1(\Gamma)$ and $H_1(\Gamma)$ are the cohomology and homology of the graph $\Gamma$, $\mathcal{C}_b^{\nu}$ is the normalization of our nodal curve and $\L=\mathbb{Q}[-2](-1)$ as a Hodge structure. Note that the name $\L$ is no accident because it is in fact the image of $\L \in K_0(Var_\mathbb{C})$ under the additive invariant taking a variety to its class in the Grothendieck ring of Hodge structures.
For each edge $e \in E$ we define a linear map:
\begin{equation*}
N_e : H^1(\mathcal{C}_\eta,\mathbb{Q}) \rightarrow H^1(\mathcal{C}_\eta,\mathbb{Q}) \otimes \L.
\end{equation*}
It is defined as the composition:
\begin{equation*}
H^1(\mathcal{C}_\eta,\mathbb{Q}) \rightarrow Gr_2^W H^1(\mathcal{C}_\eta,\mathbb{Q}) \rightarrow Gr_0^W H^1(\mathcal{C}_\eta,\mathbb{Q})\otimes \L \rightarrow H^1(\mathcal{C}_\eta,\mathbb{Q}) \otimes \L
\end{equation*}
where the first map is the canonical projection, the last map is the canonical inclusion. The middle map instead is defined by knowing that $Gr_2^W H^1(\mathcal{C}_\eta,\mathbb{Q})=H_1(\Gamma) \otimes \L$ and it will embed into $\exists$, while $Gr_0^W H^1(\mathcal{C}_\eta,\mathbb{Q})=H^1(\Gamma)$ and it is a quotient of the dual vector space $\exists^*$. Therefore the middle map is defined by the composition:
\begin{equation*}
H_1(\Gamma) \rightarrow \exists \rightarrow \exists^* \rightarrow H^1(\Gamma)
\end{equation*}
where $\exists \rightarrow \exists^*$ is the natural duality map $t \mapsto \langle t, e^*\rangle e^*$. Here $e^*$ is the dual element of the edge $e$.
The operators $N_e$ extend to operators $N_e^{(i)}$ in a natural way:
\begin{align*}
N_e^{(i)}: \bigwedge^i H^1(\mathcal{C}_\eta,\mathbb{Q}) &\rightarrow \bigwedge^i H^1(\mathcal{C}_\eta,\mathbb{Q}) \otimes \L\\
c_1 \wedge \dots \wedge c_i & \mapsto \sum_{k=1}^i c_1 \wedge \dots \wedge N_e(c_k) \wedge \dots \wedge c_i.
\end{align*}
For a subset of edges $I \subseteq E$ we write $N_I^{(i)}$ to denote the iterated composition of the $N_e^{(i)}$ for $e \in I$. Note that this composition does depend on the order (in terms of a sign) but the image $\im N_I^{(i)} \subset H^1(\mathcal{C}_\eta,\mathbb{Q}) \otimes \L^{|I|}$ does not.
We are now finally ready to define the Cattani-Kaplan-Schmid complex associated to $\mathcal{C}_b$.
\begin{defn}
For a fixed $i$ define the following complex $CKS^i$:
\begin{equation*}
0 \rightarrow \bigwedge^i H^1(\mathcal{C}_\eta,\mathbb{Q}) \rightarrow \bigoplus_{\substack{I \subseteq E\\ |I|=1}} \im N_I^{(i)} \rightarrow \bigoplus_{\substack{I \subseteq E\\ |I|=2}} \im N_I^{(i)} \rightarrow \dots
\end{equation*}
the maps are defined by considering an ordering $e_1, e_2, \dots$ of the edges and then for $J \subseteq E$ with $|J|=j$:
\begin{align*}
\im N_J^{(i)} &\rightarrow \bigoplus_{\substack{I \subseteq E\\ |I|=j+1}} \im N_I^{(i)}\\
c &\mapsto (c_I)_{|I|=j+1}
\end{align*}
where $c_I=0$ if $J \nsubseteq I$ and if $I = J \cup \{e\}$ then $c_I= (-1)^a N_e^{(i)}(c)$. Here the sign is determined by ordering $J$ according to the order of $E$, then appending $e$ at the end of $J$ and counting the number of swaps necessary to order $J \cup \{e\}$.
\end{defn}
We have the following proposition.
\begin{prop}
The $i$-th CKS complex for the curve $\mathcal{C}_b$ is quasi isomorphic to the stalk of $IC(\bigwedge^i \mathbb{R}^1 \pi_{sm*} \mathbb{Q})$ at $\mathcal{C}_b$.
\begin{proof}
See \cite[section 3.4]{migliorini2015support}.
\end{proof}
\end{prop}
As is also explained in \cite[section 3.4]{migliorini2015support} one might want to compute the weight polynomial of the CKS complexes, instead of trying to compute their cohomology. In this case it is enough to know the dimension of the $\im N_I^{(i)}$. Here we prove a lemma that we will use in one of the next chapters.
\begin{lemma}
\label{ratcurve}
Denote by $\Sigma$ a curve with two rational components meeting in $2g-2$ simple nodes and by $\overline\Sigma$ an integral curve with $2g-2$ simple nodes whose normalization is isomorphic to $\mathbb{C}\P^1$. Define:
\begin{equation*}
U(\Sigma)=\sum_{n = 0}^{2g-3} q^n [CKS^n [-n]]
\end{equation*}
and
\begin{equation*}
U(\overline\Sigma)=\sum_{n= 0}^{2g-3} q^n [\overline{CKS}^n [-n]]
\end{equation*}
where $[CKS^n]$ is the weight polynomial of the $n$-th CKS complex associated to $\Sigma$ and $[\overline{CKS}^n]$ is the analogous object for $\overline\Sigma$.
Then:
\begin{equation*}
U(\overline\Sigma)=(1-q\mathbb{Q})(1-q\L)U(\Sigma) \qquad \text{mod } q^{2g-2}.
\end{equation*}
\begin{proof}
The dual graph $\Gamma$ of $\Sigma$ consists of two vertices and $2g-2$ edges connecting them, while the dual graph $\overline \Gamma$ of $\overline\Sigma$ consists of one vertex and $2g-2$ loops. For the rest of the proof we will crucially use the fact that $n$ is at most $2g-3$ and therefore there is no way we can disconnect the graphs of $\Sigma$ and $\overline\Sigma$ by removing $n$ edges. Choose $0 \leq n \leq 2g-3$, then:
\begin{equation*}
[CKS^n [-n]]=(-1)^n \sum_{i=0}^n (-1)^i \sum_{\substack{I \subseteq e(\Gamma)\\|I|=i}} \left [ \bigwedge^{n-i} \left( H^1(\Gamma \setminus I) \oplus H_1(\Gamma \setminus I) \otimes \L\right )\right ]
\end{equation*}
and
\begin{equation*}
[\overline{CKS}^n [-n]]=(-1)^n \sum_{i=0}^n (-1)^i \sum_{\substack{I \subseteq e(\overline\Gamma)\\|I|=i}} \left [ \bigwedge^{n-i} \left( H^1(\overline\Gamma \setminus I) \oplus H_1(\overline\Gamma \setminus I) \otimes \L\right )\right ].
\end{equation*}
Note that $H^1(\overline\Gamma \setminus I) = H^1(\Gamma \setminus I) \oplus \mathbb{Q}$ and $H_1(\overline\Gamma \setminus I) \otimes \L=H_1(\Gamma \setminus I) \otimes \L \oplus \L$. Therefore:
\begin{align*}
[\overline{CKS}^n [-n]]&=(-1)^n \sum_{i=0}^n (-1)^i \sum_{\substack{I \subseteq e(\overline\Gamma)\\|I|=i}} \left [ \bigwedge^{n-i} \left( H^1(\overline\Gamma \setminus I) \oplus H_1(\overline\Gamma \setminus I) \otimes \L\right )\right ]=\\
&=(-1)^n \sum_{i=0}^n (-1)^i \sum_{\substack{I \subseteq e(\Gamma)\\|I|=i}} \left [ \bigwedge^{n-i} \left( H^1(\Gamma \setminus I) \oplus H_1(\Gamma \setminus I) \otimes \L \oplus \mathbb{Q} \oplus \L \right )\right ]=\\
&=(-1)^n \sum_{i=0}^n (-1)^i \sum_{\substack{I \subseteq e(\Gamma)\\|I|=i}} \left [ \bigwedge^{n-i} \left( H^1(\Gamma \setminus I) \oplus H_1(\Gamma \setminus I) \otimes \L \right )\right ]+\\
&+(-1)^n \sum_{i=0}^n (-1)^i \sum_{\substack{I \subseteq e(\Gamma)\\|I|=i}} \left [ \bigwedge^{n-i-1} \left( H^1(\Gamma \setminus I) \oplus H_1(\Gamma \setminus I) \otimes \L \right )\otimes (\mathbb{Q} \oplus \L) \right ]+\\
&+(-1)^n \sum_{i=0}^n (-1)^i \sum_{\substack{I \subseteq e(\Gamma)\\|I|=i}} \left [ \bigwedge^{n-i-2} \left( H^1(\Gamma \setminus I) \oplus H_1(\Gamma \setminus I) \otimes \L \right )\otimes \L \right ]=\\
&=[CKS^n [-n]]-[CKS^{n-1} [-n+1]]\cdot (\mathbb{Q}+\L)+[CKS^{n-2} [-n+2]] \cdot \L.
\end{align*}
From this follows:
\begin{equation*}
U(\overline\Sigma)=(1-q\mathbb{Q})(1-q\L)U(\Sigma) \qquad \text{mod } q^{2g-2}.
\end{equation*}
\end{proof}
\end{lemma}
\chapter{Remarks on higher rank}
Some of the results of the previous section can be generalized for all ranks.
In this chapter we will comment about the behavior of Bradlow-Higgs pairs for higher rank. Namely we will point out several difficulties that arise when trying to compute the motive of the $\mathcal{M}^{r,d}_{\sigma}$, both with similar methods to those we used in this thesis or with methods available for other moduli problems with wall-crossing.
We will also comment on how it could be possible to generalize the results of chapter \ref{MSMY}.
\section{Higher rank motives}
In the case of Bradlow pairs it is still possible, for lower rank, to understand the flip loci explicitly. For rank 3 this has been done in \cite{munoz2008hodge}. It becomes clear soon, though, that a direct approach would not work for Bradlow pairs and so a different approach to the wall-crossing problem was used in \cite{mozgovoy2013moduli}. The case of Bradlow-Higgs triples is more complicated and arguably a direct approach will fail for this problem as well.
It is worth commenting about the approach in \cite{mozgovoy2013moduli} and explain why it will not apply to the case of Bradlow-Higgs triples. We will try to highlight the main points of the strategy, but a fully detailed survey would require a separate thesis, therefore we refer directly to \cite{mozgovoy2013moduli} for all the details.
First of all denote by $\mathcal{A}_0$ the category of coherent sheaves on the smooth projective curve $C$ and by $\mathcal{A}$ the category of triples $(E_0,E_1,s)$ where $E_1 \in \mathcal{A}_0$, $E_0$ is a direct sum of a certain number of copies of $\mathcal{O}_C$ and $s: E_0 \rightarrow E_1$. Both categories admit \emph{Chern characters}, namely group morphisms from the Grothendieck group $K(\mathcal{A}_0)$ to $\mathbb{Z}^2$ and from $K(\mathcal{A})$ to $\mathbb{Z}^3$ sending a coherent sheaf to its rank and degree and a triple $(E_0,E_1,s)$ to $(\rk E_1, \deg E_1, \rk E_0)$ respectively. There are also skew symmetric bilinear forms $\langle \cdot, \cdot \rangle$ on $\mathbb{Z}^2$ and $\mathbb{Z}^3$ defined in terms of simple formulas on the Chern characters (see \cite[section 4]{mozgovoy2013moduli}).
For each of $\mathcal{A}_0$ and $\mathcal{A}$ we define the respective \emph{motivic Hall algebras} which are $\mathbb{C}$-algebras generated by indicator functions of locally closed substacks of the stacks of objects of $\mathcal{A}_0$ and $\mathcal{A}$ respectively. We denote these two algebras as $H(\mathcal{A}_0)$ and $H(\mathcal{A})$ (see \cite[section 5]{mozgovoy2013moduli}). Note that we have an embedding $\mathcal{A}_0 \subset \mathcal{A}$ and accordingly an inclusion of $\mathbb{C}$-algebras. The product of two indicators functions $\mathbb{I}_X$ and $\mathbb{I}_Y$ is the indicator function $\mathbb{I}_Z$ of objects that have a subobject in $Y$ and the quotient by such a subobject lies in $X$.
There is an integration map $I: H(\mathcal{A}) \rightarrow \mathbb{A}$ where $\mathbb{A}$ is a certain completion of the ring $R[x_1,x_2^{\pm 1},x_3]$ and $R$ is the Grothendieck ring of stacks over $\mathbb{C}$. The product on $\mathbb{A}$ is defined as
\begin{equation*}
x^{\alpha} x^{\beta} = (-\L^{1/2})^{\langle \alpha, \beta \rangle} x^{\alpha+\beta}.
\end{equation*}
The map $I$ is defined on the set of indicator functions $\mathbb{I}_X$ where $X$ is a locally closed substack of the stack of objects of $\mathcal{A}$ for which the Chern character is constantly equal to $\alpha \in \mathbb{Z}^3$. We have
\begin{equation*}
I(\mathbb{I}_X)=(-\L^{1/2})^{\langle \alpha, \alpha \rangle} x^{\alpha} [X]
\end{equation*}
and then it is extended by linearity on the whole Hall algebra. It is clearly a linear map and under the hypothesis that $\Ext^2$ groups vanish between the objects in the indicator functions $\mathbb{I}_X$ and $\mathbb{I}_Y$ then $I(\mathbb{I}_X \mathbb{I}_Y)=I(\mathbb{I}_X)I(\mathbb{I}_Y)$ (see \cite[remark 5.1]{mozgovoy2013moduli}).
Define some distinguished elements in $H^0(\mathcal{A})$. Let $u^h(\alpha)$ be the indicator function of semistable vector bundles with character $(\alpha,0)$ and $f^h_\tau(\alpha)$ be the indicator function of $\tau$-semistable triples $(E_0,E_1,s)$ with character $(\alpha,1)$. Let also $u(\alpha)$ and $f_\tau(\alpha)$ be the integration of $u^h(\alpha)$ and $f^h_\tau(\alpha)$ respectively (both up to a factor). Observe that they are essentially the motives of the moduli spaces that we would like to compute. Define
\begin{equation*}
u^h_\tau=1+\sum_{\mu(\alpha)=\tau} u^h(\alpha) \quad f^h_\tau=\sum_\alpha f^h_\tau(\alpha)
\end{equation*}
in the Hall algebra and
\begin{equation*}
u_\tau=1+\sum_{\mu(\alpha)=\tau} u(\alpha) x^\alpha \quad f_\tau=\sum_\alpha f_\tau(\alpha) x^{(\alpha,1)}
\end{equation*}
to be the corresponding generating series in $\mathbb{A}$.
Using the geometry of Bradlow pairs it is proven (see \cite[lemmas 4.11 and 4.13]{mozgovoy2013moduli}) that every $\tau$-semistable triple $(E_0,E_1,s)$ has a canonical $\tau+$ filtration whose quotient is a semistable vector bundle of slope $\tau$ and the subobject is a $\tau+$-stable triple. Also, every $\tau$-semistable triple $(E_0,E_1,s)$ has a canonical $\tau-$ filtration whose subobject is a semistable vector bundle of slope $\tau$ and the subobject is a $\tau-$-stable triple. This implies the following relations in the motivic hall algebra:
\begin{equation*}
f^h_{\tau}=f^h_{\tau+} u_\tau^h \text{ and } f^h_{\tau| \mu<\tau}= u_\tau^h f^h_{\tau-}
\end{equation*}
where $|\mu < \tau$ denotes the truncation of the series to the term for which the slope is less than $\tau$ (see \cite[theorem 5.6]{mozgovoy2013moduli}).
A very important observation at this point is that for $\tau = \infty$ the only $\tau$-semistable triples are those for which $\rk E_1=1$. These correspond to rank 1 Bradlow pairs that in turn can be identified with divisors on the curve. Therefore
\begin{equation*}
f_\infty =x_1 x_3 \sum_{d \geq 0} [S^d( C) ] x_2^d.
\end{equation*}
Also, we have $f^h_\tau(\alpha)=0$ if $\mu(\alpha)>\tau$ which geometrically amounts to say that the moduli spaces of Bradlow pairs are empty if the stability parameter exceeds a prescribed threshold.
Combining all these results it is possible to find a formula (see \cite[theorem 5.6]{mozgovoy2013moduli}):
\begin{equation*}
f_\tau = (u^{-1}_{>\tau} f_\infty u_{\geq \tau})_{|\mu \leq \tau}
\end{equation*}
where
\begin{equation*}
u_{\geq \tau}= \prod_{\tau' \geq \tau} u_{\tau'}
\end{equation*}
and the product is taken in decreasing order of $\tau'$. The formula can then be inverted using the fact that we know the generating functions $u_\tau$. At the end the motive of the moduli spaces of Bradlow pairs is computed in \cite[theorem 6.2]{mozgovoy2013moduli}.
The same approach for Bradlow-Higgs triples will not work as some essential hypotheses are not satisfied. The main issue is that the use of the integration map is conditional to the fact that for the objects we need in the wall-crossing of Bradlow pairs, the $\Ext^2$ will all vanish. This is far from being true in the case of Bradlow-Higgs triples and in lemma \ref{exts} we saw an example of the presence of nonzero $\Ext^2$. In terms of the flip loci this will imply that they are not bundles in general. A positive fact is that for Higgs bundles, $\Ext^2$ is dual to $\Hom$ by Serre's duality.
The second issue is that the moduli spaces of $\sigma$-stable Bradlow-Higgs triples are not empty after a certain value of $\sigma$. This will imply that the analogue of the generating function $f_\infty$ will not be as simple as in the case of Bradlow pairs. Note also that we gave an approach to compute $\mathcal{M}_\infty^{2,d}$ directly but it is not at all trivial already in rank 2.
In conclusion, the analogues of the formulas
\begin{equation*}
f^h_{\tau}=f^h_{\tau+} u_\tau^h \text{ and } f^h_{\tau| \mu<\tau}= u_\tau^h f^h_{\tau-}
\end{equation*}
in the motivic Hall algebra are probably still valid but, even if they can be proved, it will then not be possible to apply an easy integration map to get information about the motives.
\section{Partial Hilbert scheme formula for higher rank}
In this section we will discuss how to possibly generalize the result in theorem \ref{BHMSMY} to higher rank. Consider the following higher rank generating functions.
\begin{defn}
Let $r \geq 2$ be an integer and $1 \leq m \leq r-1$ be an integer coprime with $r$.
\begin{equation*}
F_{r,m}^{sh}(q)= \sum_{\substack{d \geq r(r-1)(1-g)\\ d \text{ mod } r= m}} \mathbb{R} (\chi_\varepsilon^{r,d})_* (IC_{\mathcal{M}_\varepsilon^{r,d}}) q^{d+r(r-1)(g-1)},
\end{equation*}
\begin{equation*}
F_{r,m}^{mot}(q)= \sum_{\substack{d \geq r(r-1)(1-g)\\ d \text{ mod } r= m}} [\mathcal{M}_\varepsilon^{r,d}] q^{d+r(r-1)(g-1)},
\end{equation*}
\begin{equation*}
F_{r,m}^{vir}(q,t)=\sum_{\substack{d \geq r(r-1)(1-g)\\ d \text{ mod } r= m}} P^{vir}(\mathcal{M}_\varepsilon^{2,2n+1},t) q^{d+r(r-1)(g-1)}.
\end{equation*}
\end{defn}
These are clearly the generating functions in theorem \ref{BHMSMY} extended for higher rank. Let us introduce the following notation.
\begin{defn}
Let $r \geq 2$ be an integer and $0 \leq m \leq r-1$ be an integer. We define $\text{del}_{q,r,m}$ be the operator acting on $R[[q]]$ by deleting from power series all terms whose degree $d$ does not have remainder $m$ modulo $r$.
\end{defn}
Let $h^{r,d}:\mathcal{M}^{r,d} \rightarrow \mathcal{A}^r$ denote the usual Hitchin map and $PH(\mathcal{M}^{r,d},q,t)$ denote the perverse hodge polynomial of $\mathcal{M}^{r,d}$. Then we define the second class of generating functions we need.
\begin{defn}
Let $r \geq 2$ be an integer and $1 \leq m \leq r-1$ be an integer coprime with $r$. Define:
\begin{equation*}
G_{r,m}(q,t)=\text{del}_{q,r,m} \left (\frac{PH(\mathcal{M}^{r,m},q,t)}{(1-q)(1-qt^2)} \right )
\end{equation*}
and
\begin{equation*}
G_{r,m}^{sh}(q)=\text{del}_{q,r,m} \left (\frac{ \mathbb{R} h^{r,m}_* \mathbb{Q} }{(1-q\mathbb{Q})(1-q\mathbb{Q}[-2](-1))} \right )
\end{equation*}
\end{defn}
In this section we will prove a variant of theorem \ref{BHMSMY} for higher rank.
\begin{theorem}
Let $r \geq 2$ be an integer and $1 \leq m \leq r-1$ be an integer coprime with $r$. Then $F_{r,m}^{sh}$ and $G_{r,m}^{sh}$ coincide for $\deg q \leq (r-1)(2g-2)-1$. $P(F_{r,m}^{sh}(q),t)$ and $G_{r,m}$ coincide for $\deg q \leq (r-1)(2g-2)-1$ and for $\deg q > r(2g-1)+(3r^2-5r+2)(g-1)$.
If the following conditions are satisfied:
\begin{itemize}
\item for $d > r(r+1)(g-1)$ then a stable Higgs bundle $(E,\phi)$ of rank $r$ and degree $d$ satisfies $H^1(E)=0$
\item for $d>r^2(g-1)$ coprime with $r$, $\mathcal{M}_\varepsilon^{r,d}$ is smooth
\end{itemize}
then $P(F_{r,m}^{sh}(q),t)$ and $G_{r,m}$ coincide for $\deg q \leq (r-1)(2g-2)-1$ and for $\deg q > (2 r^2-r)(2g-2)$.
\end{theorem}
Recall that the strategy of the proof for rank 2 was to compare the cohomology of $\mathcal{M}_\varepsilon^{2,d}$ for $d < 0$ and for $d > 4g-4$ odd to the perverse filtration on the cohomology of $\mathcal{M}^{2,1}$. Let us summarize what was needed in order to prove the result.
First we saw in proposition \ref{negdegrk2} that $\mathcal{M}_\infty^{2,d}=\mathcal{M}_\varepsilon^{2,d}$ if $2-2g \leq d <0$ and that in this case they are both smooth. To prove theorem \ref{BHMSMY} in this range we then used proposition \ref{dimsupp} to get information about the possible supports in the decomposition theorem and concluded by studying the case of a generic nodal curve in $\mathcal{A}^2_{red}$.
For the range $d > 4g-4$ odd, we used again smoothness of $\mathcal{M}_\varepsilon^{2,d}$ and the fact that for $d \geq 6g-5$ odd the Abel-Jacobi map $\mathcal{M}_\varepsilon^{2,d}\rightarrow \mathcal{M}^{2,d}$ is a projective bundle (see proposition \ref{sharpnits}). In the range $d \geq 6g-5$ odd theorem \ref{BHMSMY} was then easy to prove. For the range $4g-3 \leq d \leq 6g-7$ odd, we then used smoothness of $\mathcal{M}_\varepsilon^{2,d}$ plus Serre's duality to conclude that if the cohomological formula holds for odd values in the range $3-2g, \dots , -1$ then it must also hold for odd values in the range $4g-3, \dots , 6g-7$.
We also used the fact that the extra summand $\mathscr{L}$ in the decomposition theorem for the Hitchin map $\mathcal{M}^{2,1} \rightarrow \mathcal{A}^2$ has no cohomology. A closer look to the argument though will show that we don't need this to show the formula in the ranges $d \geq 6g-5$ and $d <0$.
It is worth discussing which of the previous properties can be generalized to the case of higher rank. Let us first give a brief remark.
\begin{rmk}
\label{cflip}
Note that since $\sigma$-stability is an open condition, the flip loci are closed in each of the $\mathcal{M}_\sigma^{r,d}$. In particular, due to the fact that the Hitchin maps are proper, the restriction of the Hitchin maps to the flip loci is proper as well. It follows that if $(E,\phi,s)$ belongs to a flip locus, then
$$\lim_{\lambda \rightarrow 0} \lambda \cdot (E,\phi,s)$$
also belongs to the same flip locus.
\end{rmk}
We can generalize proposition \ref{negdegrk2} as follows.
\begin{prop}
Let $d < r(r-2)(1-g)$, then $\mathcal{M}_\varepsilon^{r,d} = \mathcal{M}_\infty^{r,d}$. Both are non-empty if and only if $d \geq r(r-1)(1-g)$.
\begin{proof}
The statement about non-emptiness follows from the $U$-filtration for $\mathcal{M}_\infty^{r,d}$ and for $\mathcal{M}_\varepsilon^{r,d}$ follows from the previous statement.
Let us first prove that if $d < r(r-2)(1-g)$ then $\mathcal{M}_\varepsilon^{r,d} \subseteq \mathcal{M}_\infty^{r,d}$. Suppose $(E,\phi,s)$ has stable underlying Higgs bundle and assume by contradiction that the $U$-filtration of the triple has length $l < r$. But then $\mu(U_l) \geq (l-1)(1-g )>(r-2)(1-g)> \mu(E)$ and therefore $U_l$ cannot be preserved by $\phi$ which, by construction, is impossible.
To conclude we use remark \ref{cflip}. From it we can deduce that it is enough to prove that if $(E,\phi,s) \in \mathcal{M}_\infty^{r,d}$ is a fixed point for the $\mathbb{C}^*$-action, then $(E,\phi,s) \in \mathcal{M}_\varepsilon^{r,d}$. Recall from proposition \ref{fixedbht} and from the fact that $s$ is cyclic for $\phi$ that $(E,\phi,s)$ has to be of the form $E= E_1 \oplus \dots \oplus E_r$ for line bundles $E_i$, $s \in E_r$, $\phi(E_i) \subseteq E_{i-1} \otimes K $ and $\phi(E_1)=0$. Note also that all the maps $E_i \rightarrow E_{i-1} \otimes K$ induced by $\phi$ have to be non-zero. This implies that $d_{r-i} = \deg E_{r-i} \geq 2i (1-g)$.
Note that the only proper $\phi$-invariant subbundles of $E$ are of the form $\oplus_{i=1}^k E_i$ for some $1 \leq k <r$. Assume by contradiction that any of those are destabilizing for $(E,\phi)$.
This means that for some $1 \leq k < r$ we have:
\begin{equation*}
\frac{1}{k} \sum_{i=1}^k d_i \geq \frac{1}{r} \sum_{i=1}^r d_i
\end{equation*}
which is equivalent to:
\begin{equation*}
(r-k) \sum_{i=1}^k d_i \geq k \sum_{i=k+1}^r d_i.
\end{equation*}
This implies
\begin{equation*}
(r-k) \sum_{i=1}^r d_i \geq r \sum_{i=k+1}^r d_i = r \sum_{i=0}^{r-k-1} d_{r-i} \geq r(r-k)(r-k-1)(1-g)
\end{equation*}
which contradicts $d < r(r-2)(1-g)$.
\end{proof}
\end{prop}
From remark \ref{adhm} we also deduce the following.
\begin{cor}
If $d < r(r-2)(1-g)$ then $\mathcal{M}_\varepsilon^{r,d} = \mathcal{M}_\infty^{r,d}$ is smooth when non-empty.
\begin{proof}
It follows from the fact that in this case there are no proper $\phi$-invariant subbundles that contain the section and so the Zariski tangent space has minimal dimension, since the only endomorphisms that commute with $\phi$ are scalars.
\end{proof}
\end{cor}
Assume $d < r(r-2)(1-g)$. In the case $r \geq 2$ proposition \ref{dimsupp} applied to the Hitchin map $\mathcal{M}_\infty^{r,d} \rightarrow \mathcal{A}^r$ allows to conclude that any properly supported summand of the pushforward of the constant sheaf $\mathbb{Q}$ should have support with dimension at least:
\begin{equation*}
2+2r^2(g-1)-1-r^2(g-1)-d-r(r-1)(g-1) \geq 2+r(r-1)(g-1).
\end{equation*}
Since
\begin{equation*}
\dim \mathcal{A}^r_{red} = \max_{1 \leq k \leq r-1} \{ \dim \mathcal{A}^k+\dim \mathcal{A}^{r-k}\}=2+(g-1)(r^2-2r+2)
\end{equation*}
and $2+r(r-1)(g-1) \geq 2+(g-1)(r^2-2r+2)$ we can already conclude that proper supported summands, if there are any, should have support equal to the closure of $\mathcal{A}^r_{red}$.
It is also immediate to check that the argument in theorem \ref{BHMSMY} about the comparison of weight polynomials for nodal curves can be applied again. Just note that in this case we should let $\mathcal{C}$ and $\mathcal{D}$ be generic nodal curves in $\mathcal{A}^r_{red}$ and $\mathcal{A}^r_{int}$. Here lemma \ref{ratcurve} should be modified as follows and the rest of the argument is the same.
\begin{lemma}
Denote by $\Sigma$ a curve with two rational components meeting in $(r-1)(2g-2)$ simple nodes and by $\overline\Sigma$ an integral curve with $(r-1)(2g-2)$ simple nodes whose normalization is isomorphic to $\mathbb{C}\P^1$. Define:
\begin{equation*}
U(\Sigma)=\sum_{n = 0}^{(r-1)(2g-2)-1} q^n [CKS^n [-n]]
\end{equation*}
and
\begin{equation*}
U(\overline\Sigma)=\sum_{n= 0}^{(r-1)(2g-2)-1} q^n [\overline{CKS}^n [-n]]
\end{equation*}
where $[CKS^n]$ is the weight polynomial of the $n$-th CKS complex associated to $\Sigma$ and $[\overline{CKS}^n]$ is the analogous object for $\overline\Sigma$.
Then:
\begin{equation*}
U(\overline\Sigma)=(1-q\mathbb{Q})(1-q\L)U(\Sigma) \qquad \text{mod } q^{(r-1)(2g-2)}.
\end{equation*}
\end{lemma}
This concludes the generalization of theorem \ref{BHMSMY} from the range $d<0$ for rank 2 to the range $d < r(r-2)(1-g)$ for rank $r$.
Now, using \cite[corollary 3.4]{nitsure1991moduli} we can immediately deduce the same result for $d> r(2g-1)+(r-1)^2(2g-2)$ coprime with $r$ since we know that in this case the Abel-Jacobi map is a projective bundle and the argument in theorem \ref{BHMSMY} carries through. As we already pointed out, the presence of extra summands (even with non-trivial global cohomology) in the decomposition theorem for the Hitchin map $\mathcal{M}^{r,d} \rightarrow \mathcal{A}^r$ would not cause problems at this point.
Serre's duality would allow to deduce the cohomological statement from theorem \ref{BHMSMY} for $d$ coprime with $r$ in the range $r^2(g-1) < d \leq r(r+1)(g-1)$ from the statement we already proved for $r(r-1)(1-g) \leq d < r(r-2)(1-g)$, provided $\mathcal{M}_\varepsilon^{r,d}$ is smooth for $r$, $d$ coprime and $d>r^2(g-1)$.
Ultimately, we would be left with $d$ coprime with $r$ in the range $r(r-2)(1-g) < d < r(2g-1)+(r-1)^2(2g-2)$. It would be reasonable to guess that the sharp estimate for $d$ for which $H^1$ of a semistable Higgs bundle of degree $d$ and rank $r$ vanishes is $d>r(r-2)(1-g)$ rather than the one in \cite[corollary 3.4]{nitsure1991moduli}, but we do not attempt to prove it here.
\chapter{Introduction}
\section*{Description of the project}
The main goal of the thesis is to understand the geometry of the moduli spaces of Bradlow-Higgs triples on a smooth projective curve $C$, with particular focus on the case of rank 2. These objects consist of a vector bundle $E$ on $C$, a morphism $\phi: E \rightarrow E \otimes K$ and a non-zero global section of $E$. Since there is a Higgs field $\phi$ and a section $s$, Bradlow-Higgs triples relate to both Higgs bundles and Bradlow pairs. See \cite{nitsure1991moduli} and \cite{thaddeus1994stable} for an introduction on the two moduli problems. Features from both original moduli problems are then inherited by the moduli problem of Bradlow-Higgs triples.
As for Bradlow pairs, there exists a family of stability conditions that depends on a positive real parameter $\sigma$. This will allow to vary the stability and hence produce several moduli spaces $\mathcal{M}_\sigma^{r,d}$ of $\sigma$-semistable triples of rank $r$ and degree $d$. For fixed $r$ and $d$, the set of positive real numbers is then partitioned into a finite number of intervals, the last one of which is unbounded. The defining property of these intervals is that varying $\sigma$ in the interior of each one of these intervals will yield the same moduli space $\mathcal{M}_\sigma^{r,d}$, but different intervals have non-isomorphic associated moduli spaces. The endpoints of these intervals are called \emph{critical values} and $\mathcal{M}_\sigma^{r,d}$ will change whenever $\sigma$ crosses one. This phenomenon is well known as \emph{wall-crossing}.
The moduli spaces $\mathcal{M}_\sigma^{r,d}$ will inherit properties from the moduli problem of Higgs bundles as well. On the other hand they are not all smooth, and this makes their geometry richer but more complicated. As we will see, each of them admits a proper \emph{Hitchin map}
$$\chi_\sigma^{r,d}: \mathcal{M}_\sigma^{r,d} \rightarrow \mathcal{A}^r$$
whose target, the \emph{Hitchin base}, is an affine space. Furthermore, $\mathbb{C}^*$ acts on each of the $\mathcal{M}_\sigma^{r,d}$ by scaling the Higgs field and there is an action of $\mathbb{C}^*$ on $\mathcal{A}^r$ for which the Hitchin map is equivariant. In particular we will prove that for every $(E,\phi,s) \in \mathcal{M}_\sigma^{r,d}$ there exists the limit
$$\lim_{\lambda \rightarrow 0} \lambda \cdot (E,\phi,s)$$
and this has important consequences on the geometry of our moduli spaces.
When $\sigma$ crosses one of the critical values, the geometry of the moduli space will change, in the sense that some triples will become unstable and they will be erased from $\mathcal{M}_\sigma^{r,d}$, while some others will become stable and hence will be added to $\mathcal{M}_\sigma^{r,d}$. The loci that are added and removed are often known as \emph{flip loci}. Furthermore there are two distinguished stability conditions. Namely, for $\sigma=\varepsilon$ smaller than the least critical value, the stability of the triple will imply the semistability of the underlying Higgs bundle and so we have a forgetful map, also known as \emph{Abel-Jacobi map},
$$\mathcal{M}_\varepsilon^{r,d} \rightarrow \mathcal{M}^{r,d}$$
whose target is the moduli space of semistable Higgs bundles of rank $r$ and degree $d$. If $r$, $d$ are coprime and $d$ is very large compared to $r$ then the Abel-Jacobi map is a projective bundle, providing a direct relation between $\mathcal{M}_\varepsilon^{r,d}$ and $\mathcal{M}^{r,d}$.
Instead, when $\sigma$ is large enough to be bigger than the last critical value then the stability condition is equivalent to requiring that the section $s$ is a cyclic vector for $\phi$. In this case, the map $\chi_\sigma^{r,d}$ can be proved to be a relative Hilbert scheme of points. It is also related to moduli spaces of stable pairs studied by Pandharipande and Thomas in \cite{pandharipande2009curve}. The connection between Bradlow-Higgs triples and Hilbert schemes of points is also an important part of the project and a consequence of the study of the geometry of Bradlow-Higgs triples.
\section*{Content of the thesis and main results}
In chapter 1 we collect some known results from the literature about the main topics that are needed to understand the thesis. First we introduce the invariants that we will compute throughout the work i.e. the Grothendieck motives. We recall some results about Higgs bundles and Bradlow pairs that will set the starting point for the definition of Bradlow-Higgs triples. We will also recollect some classical results about non-abelian Hodge theory and the relation between moduli spaces of Higgs bundles and character varieties. Some attention will be given to the wall-crossing for the moduli problem of Bradlow pairs, which is a good starting point for the understanding of the wall-crossing for Bradlow-Higgs triples. Finally we give some insight about the notion of Hilbert scheme, especially in the case of curves, since moduli spaces of Bradlow-Higgs triples are connected to relative Hilbert schemes.
Chapter 2 will contain the basic definition of Bradlow-Higgs triples and their $\sigma$-stabilty. We also outline the construction of the moduli spaces $\mathcal{M}_\sigma^{r,d}$ as open subsets of moduli spaces of coherent systems on the surface $\P(\mathcal{O}_C \oplus K)$ that were originally studied in \cite{le1995faisceaux}. This approach will also lead to the proof of the properness of the Hitchin maps $\chi_\sigma^{r,d}$. There will be a study of the deformation theory of Bradlow-Higgs triples, in order to understand if the moduli spaces are singular. We will prove that after the first critical value is crossed the moduli spaces are in fact singular. The $\mathbb{C}^*$-action on the moduli spaces of Bradlow-Higgs triples is understood and for rank 2 we provide a decomposition of the moduli spaces in terms of the connected components of the fixed point loci. For some combinations of $\sigma$, $r$ and $d$ we can prove that $\mathcal{M}_\sigma^{r,d}$ is smooth and this will allow to compute its cohomology from the motives. Here is the main outcome of the chapter:
\begin{theorem*}
Bradlow-Higgs triples $(E,\phi,s)$ on $C$ correspond to pairs $(\mathcal{F},s)$ where $\mathcal{F}$ is a rank one pure one dimensional sheaf on the surface $\P(\mathcal{O}_C \oplus K)$ whose support does not intersect the divisor at infinity and $s$ is a non-zero global section of $\mathcal{F}$. Under this identification the Hitchin base $\mathcal{A}^r$ is a family curves in $\P(\mathcal{O}_C \oplus K)$ whose support does not intersect the divisor at infinity. Furthermore the Hitchin maps $\chi_\sigma^{r,d}: \mathcal{M}_\sigma^{r,d} \rightarrow \mathcal{A}^r$ correspond to taking the scheme theoretic support of the sheaves $\mathcal{F}$.
For $\sigma \rightarrow 0$ there is an Abel-Jacobi map
$$AJ: \mathcal{M}_\sigma^{r,d} \rightarrow \mathcal{M}^{r,d}$$
forgetting the section and whose target is therefore the moduli space of semistable Higgs bundles.
For $\sigma \rightarrow \infty$, $\mathcal{M}_\sigma^{r,d} \rightarrow \mathcal{A}^r$ is a relative Hilbert scheme of points for the family of curves parametrized by $\mathcal{A}^r$.
For $r=2$, $\sigma \rightarrow 0$ and either $d<0$ or $d>4g-4$ odd, $\mathcal{M}_\sigma^{r,d}$ is smooth. In all other cases it is singular.
$\mathcal{M}_\sigma^{r,d}$ can be decomposed into attracting sets for a $\mathbb{C}^*$-action, explicitly for $r=2$:
$$\mathcal{M}_\sigma^{2,d}=F_{(2),\sigma}^{(d),1+} \sqcup \bigsqcup_{I_1} F_{(1,1),\sigma}^{(d_1,d_2),1+}\sqcup \bigsqcup_{I_2} F_{(1,1),\sigma}^{(d_1,d_2),2+}.$$
Each of the $F^+$ contains a connected component $F$ of the fixed point locus of $\mathcal{M}_\sigma^{2,d}$ and is characterized by the property that the limit as $\lambda \rightarrow 0$ of $\mathbb{C}^*$ acting on points in $F^+$ will belong to $F$.
\end{theorem*}
In chapter 3 we compute the motivic invariants of the $\mathcal{M}_\sigma^{r,d}$ for $r=2$. There are three main parts. First, we examine the flip loci and we understand what triples are added and erased from $\mathcal{M}_\sigma^{r,d}$ as a critical value $\bar\sigma$ is crossed. We can prove that the two flip loci $\mathcal{W}^{d,+}_{\bar\sigma}$ and $\mathcal{W}^{d,-}_{\bar\sigma}$ admit maps:
$$
\pi_{\bar\sigma}^{d,+}: \mathcal{W}_{\bar\sigma}^{d,+} \rightarrow X^d_{\bar\sigma}
$$
and
$$
\pi_{\bar\sigma}^{d,-}: \mathcal{W}_{\bar\sigma}^{d,-} \rightarrow X^d_{\bar\sigma}
$$
where $X^d_{\bar\sigma}$ is the cartesian product of a symmetric power of the curve $C$, of its Jacobian and of two copies of an affine space. The flip loci themselves contain triples that can be described as extensions of Higgs bundles for which the canonical subobject and quotient have prescribed degrees. The fibers of the maps $\pi_{\bar\sigma}^{d,+}$ and $\pi_{\bar\sigma}^{d,-}$ are projective spaces, although not of constant dimension, and this suffices to compute the motive of the flip loci. We will also point out how the flip loci interact with a decomposition of the $\mathcal{M}_\sigma^{2,d}$ that is obtained by exploiting the $\mathbb{C}^*$-action. Second we compute $[\mathcal{M}_\varepsilon^{2,d}]$ for small $\varepsilon$. The strategy relies once again on the $\mathbb{C}^*$-action. Last, we will point out a strategy to compute $[\mathcal{M}_\sigma^{2,d}]$ for $\sigma$ bigger than the last critical value. The main results of the chapter can be formulated as follows:
\begin{theorem*}
There are maps:
$$
\pi_{\bar\sigma}^{d,+}: \mathcal{W}_{\bar\sigma}^{d,+} \rightarrow X^d_{\bar\sigma}
$$
and
$$
\pi_{\bar\sigma}^{d,-}: \mathcal{W}_{\bar\sigma}^{d,-} \rightarrow X^d_{\bar\sigma}
$$
whose fibers are projective spaces. This allows to compute motives:
\begin{equation*}
[\mathcal{W}_{\bar\sigma}^{d,+}]=\L^{2g} \cdot [\mathbb{C}\P^{2g-3}]\cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot [J( C)]+\L^{3g-2} \cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot [S^{\bar\sigma}( C)],
\end{equation*}
\begin{equation*}
[\mathcal{W}_{\bar\sigma}^{d,-}]=\L^{2g}\cdot[S^{(d-{\bar\sigma})/2}( C)]\cdot[J^{(d+{\bar\sigma})/2}( C)]\cdot [\mathbb{C}\P^{(d+\bar\sigma)/2+g-2}].
\end{equation*}
The motive of $\mathcal{M}_\varepsilon^{2,d}$ for $0<\varepsilon<1$ and $d >0$ odd can be computed from the decomposition into attracting sets:
\begin{align*}
[\mathcal{M}_\varepsilon^{2,d}]=&\L^{1+4(g-1)}[M_\varepsilon^{2,d}]+\sum_{(d_1,d_2) \in I_1^o(d)} \L^{1+3(g-1)+d_2}[S^{d_1}( C)][S^{d_1-d_2+2g-2}]+\\
&+\sum_{(d_1,d_2) \in I_2^o(d)}\L^{1+4(g-1)} [S^{d_2}( C)][S^{d_1-d_2+2g-2}( C)]+\\
&+\sum_{(d_1,d_2) \in I_1^o(d)} (\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] \left ([S^{d_2}( C)]-[J( C)] \frac{\L^{d_2+1-g}-1}{\L-1}\right).
\end{align*}
A similar formula also holds for $d \geq 0$ even.
\begin{align*}
[\mathcal{M}_\varepsilon^{2,d}]&=\L^{1+4(g-1)} [M^{2,d}_\varepsilon]+\sum_{(d_1,d_2) \in I_1^e(d)} \L^{1+3(g-1)+d_2} [S^{d_1}( C)][S^{d_1-d_2+2g-2}( C)] +\\
&+\sum_{(d_1,d_2) \in I_2^e(d)}\L^{1+4(g-1)} [S^{d_2}( C)][S^{d_1-d_2+2g-2}( C)]+\\
&+(\L-1) \L^{1+4(g-1)}[\Sym^2( S^{d/2} ( C) )] +\L^{1+4(g-1)} [S^{d/2}( C)][J^{d/2}( C)][\mathbb{C}\P^{g-2}]+\\
&+[S^{d/2}( C)] \L^{3g-2} \left ( \L^{d/2+g-1}+\L^{2g-2}-1\right )+\\
&+\sum_{(d_1,d_2) \in I_1^e(d)} (\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] \left ([S^{d_2}( C)]-[J( C)] \frac{\L^{d_2+1-g}-1}{\L-1} \right).
\end{align*}
For $\sigma > d$, $[\mathcal{M}_\infty^{2,d}]$ can be computed either by combining the above formulas and the motive of the flip loci or directly.
\end{theorem*}
Chapter 4 contains some comments about the generalization of the results of the previous chapters to the case of Bradlow-Higgs triples with poles. Fixed an integer $\gamma \geq 1$ and a point $P \in C$, a Bradlow-Higgs $\gamma$-triple is the datum $(E,\phi,s)$ where $E$ is a vector bundle on $C$, $s $ is a non-zero global section of $E$ and $\phi: E \rightarrow E \otimes K(\gamma P)$ is a Higgs field that is allowed to have poles at $P$. Most of the properties of the $\mathcal{M}_\sigma^{2,d}$ are still valid for their analogues with poles $\mathcal{M}_\sigma^{2,d}(\gamma)$. We also prove that when $\gamma$ is chosen large enough, $\mathcal{M}_\sigma^{2,d}(\gamma)$ are smooth, the flip loci in the wall crossing are actually projective bundles and their motive can be computed in an easier way. The two main results are:
\begin{theorem*}
$\mathcal{M}_\varepsilon^{2,d}(\gamma)$ is smooth for all $\gamma \geq 1$. In particular $\mathcal{M}_\varepsilon^{2,d}(\gamma)$ is always semiprojective.
If $\gamma > d$ then $\mathcal{M}_\sigma^{2,d}(\gamma)$ is smooth regardless of $\sigma$ as long as it is different from a critical value. In this case then $\mathcal{M}_\sigma^{2,d}(\gamma)$ is semiprojective.
We have:
\begin{align*}
&\lim_{\gamma \rightarrow \infty} P(\mathcal{M}_{\varepsilon}^{2,d}(\gamma),t) = \lim_{\gamma \rightarrow \infty} P(\mathcal{M}_{\infty}^{2,d}(\gamma),t)=\\
&=\frac{(1+t^3)^{2g}(1+t)^{2g}}{(1-t^2)^2(1-t^4)}=P(\mathbb{C} \P^{\infty},t) P(B \overline{\mathcal{G}},t)
\end{align*}
where $B \overline{\mathcal{G}}$ is the classifying space of the group $\overline{\mathcal{G}}$ mentioned in \cite[section 7.2]{hausel2001geometry}.
\end{theorem*}
In chapter 5 we explore the relation between the moduli spaces of Bradlow-Higgs triples and relative Hilbert schemes of points on curves. We already mentioned that for very large $\sigma$, $\chi_{\sigma}^{2,d}$ is a Hilbert scheme relative to the family of locally planar curves parametrized by the Hitchin base $\mathcal{A}^2$. A lot of recent work by several authors, see \cite{migliorini2011support}, \cite{maulik2014macdonald} and \cite{migliorini2015support}, relates the cohomology of the Hilbert scheme of points on a curve to the cohomology of the compactified Jacobian of the same curve. The key assumption is that the curve has to be locally planar and at least reduced. Note that these formulas are valid for certain families of curves and their relative Hilbert scheme. In the chapter we exploit the connection between Bradlow-Higgs triples and relative Hilbert schemes of points on curves to prove a partial formula relating the cohomology of $\mathcal{M}_\varepsilon^{2,d}$ for $d$ odd to the perverse filtration on the cohomology of $\mathcal{M}^{2,1}$. The main formula is as follows:
\begin{theorem*}
Setting
\begin{equation*}
F^{sh}(q)= \sum_{n \geq 1-g} \mathbb{R} (\chi_\varepsilon^{2,2n+1})_* (IC_{\mathcal{M}_\varepsilon^{2,2n+1}}) q^{2n+2g-1},
\end{equation*}
\begin{equation*}
F^{vir}(q,t)= \sum_{n \geq 1-g} P^{vir}(\mathcal{M}_\varepsilon^{2,2n+1},t) q^{2n+2g-1}
\end{equation*}
and
\begin{equation*}
G(q,t)=\text{odd}_q \left (\frac{PH(\mathcal{M}^{2,1},q,t)}{(1-q)(1-qt^2)} \right ),
\end{equation*}
\begin{equation*}
G^{sh}(q)=\text{odd}_q \left (\frac{\bigoplus_{i=0}^{8g-6} IC\left (\bigwedge^i\mathbb{R}^1 \right) }{(1-q\mathbb{Q})(1-q\mathbb{Q}[-2](-1))} \right )
\end{equation*}
we can prove that $F^{sh}$ and $G^{sh}$ coincide for $\deg q \leq 2g-3$. $P(F^{sh}(q),t)$ and $G$ coincide for $\deg q \leq 2g-3$ and for $\deg q \geq 6g-5$. Furthermore $P(F^{sh}(q),t)-G(q,t)$ is a polynomial with non-negative coefficients.
\end{theorem*}
In the last chapter we will discuss the issues that arise when trying to generalize the content of the thesis to Bradlow-Higgs triples of rank bigger than 2. In the first part, we focus on the fact that while there exists an approach to understand the wall-crossing of Bradlow pairs for arbitrary rank and degree, see \cite{mozgovoy2013moduli}, this will fail for Bradlow-Higgs triples and we point out why. In the second part we will partly generalize the content of chapter 5 to higher rank.
\chapter{Hilbert schemes and compactified Jacobians}
\label{MSMY}
\section{Maulik-Yun and Migliorini-Shende-Viviani formulas}
Recently there has been a lot of steady progress in relating the cohomology of the Hilbert scheme of points of a locally planar curve and the cohomology of compactified Jacobians of the same curve.
In \cite{migliorini2011support}, \cite{maulik2014macdonald} and \cite{rennemo2013homology} we find three different approaches to generalize the well known formula by Macdonald relating the cohomology of the symmetric powers of a smooth projective curve and the cohomology of the Jacobian. The last progress made in this direction is \cite{migliorini2015support}.
Let $X$ be a complex projective and locally planar curve. If $X$ is smooth of genus $g$, denote by $S^n(X)$ its $n$-th symmetric power and by $J(X)$ its Jacobian. The classical Macdonald formula states that:
\begin{equation*}
\sum_{n \geq 0} P(S^n(X) ,t) q^n = \frac{(1+qt)^{2g}}{(1-q)(1-qt^2)}=\frac{P(J(X),qt)}{(1-q)(1-qt^2)}.
\end{equation*}
If $X$ is integral then the Macdonald formula has been generalized to this case independently in \cite{migliorini2011support} and \cite{maulik2014macdonald}, later reproved in \cite{rennemo2013homology}. Denote by $X^{[n]}$ the Hilbert scheme of $n$ points on $X$ and by $\overline{J}({X})$ the compactified Jacobian parametrizing degree zero rank one torsion free sheaves on ${X}$. A version of the generalized Macdonald formula can be stated as follows (see for example \cite[formula 1.4]{maulik2014macdonald}):
\begin{equation*}
\bigoplus_{n \geq 0} H^*({X}^{[n]}) q^n = \frac{\oplus_{i=0}^{2 g_a} \Gr_i^P H^*(\overline{J}({X})) q^i}{(1-\mathbb{Q} q)(1-\mathbb{Q}[-2](-1) q)}.
\end{equation*}
The equality is meant to be of (cohomologically) graded vector spaces. Also, $\Gr_i^P$ refers to the $i$-th graded piece with respect to the perverse filtration. This particular filtration on the cohomology of $\overline{J}({X})$ can be defined by deforming ${X}$ in a flat family of curves $\pi:\mathcal{X} \rightarrow \mathcal{B}$ with the following properties (see \cite[section 2.1]{maulik2014macdonald}):
\begin{itemize}
\item $\mathcal{B}$ is irreducible.
\item The fibers of $\pi$ are integral and locally planar.
\item Let $\pi^{[n]}: \mathcal{X}^{[n]} \rightarrow \mathcal{B}$ denote the Hilbert scheme of $n$ points relative to the family $\pi$. Then we assume the total space $\mathcal{X}^{[n]}$ is smooth for all $n \geq 0$.
\item If $b \in \mathcal{B}$ we denote by $\delta(b)$ the delta invariant of $\pi^{-1}(b)$. Then we assume that $codim_\mathcal{B}(\overline{\{b\}}) \geq \delta(b)$ where $\overline{\{b\}}$ is the Zariski closure of the point $b \in \mathcal{B}$.
\end{itemize}
After we have any of these families $\pi:\mathcal{X} \rightarrow \mathcal{B}$ we also get the relative compactified Jacobian $\pi^J:\mathcal{J} \rightarrow \mathcal{B}$ whose total space is smooth as a consequence of the previous axioms. The map $\pi^J$ is proper and the object $\mathbb{R} \pi^J_* \mathbb{Q}_\mathcal{J}$ is filtered by its perverse truncations $\leftidx{^p}{\tau}{^{\leq i}}\mathbb{R} \pi^J_* \mathbb{Q}_\mathcal{J}$. The perverse filtration is thus defined by restricting the images of the natural maps:
\begin{equation*}
H^*\left (\leftidx{^p}{\tau}{^{\leq i}}\mathbb{R} \pi^J_* \mathbb{Q}_\mathcal{J} \right) \rightarrow H^*\left (\mathbb{R} \pi^J_* \mathbb{Q}_\mathcal{J} \right).
\end{equation*}
to the fiber of $\pi$ corresponding to the starting curve ${X}$. In the papers is also explained why the perverse filtration does not depend on the choice of the particular family. Furthermore, a similar version of the formula holds for the cohomology of the total spaces $\mathcal{X}^{[n]}$ and $\mathcal{J} $ with the perverse filtration. In fact in both \cite{migliorini2011support} and \cite{maulik2014macdonald} the formula is obtained from a stronger sheaf theoretic statement, i.e. an equality in $D_c^b(\mathcal{B})[[q]]$ (see \cite[formula 4]{migliorini2011support}):
\begin{equation*}
\bigoplus_{n \geq 0} q^n \mathbb{R}\pi_*^{[n]} \mathbb{Q}= \frac{\bigoplus_i q^i \cdot \mathbb{R}^i \pi_*^J \mathbb{Q}[-i]}{(1-q)(1-q\mathbb{Q}[-2](-1))}.
\end{equation*}
To deduce the statements for a single curve or for the entire family of curves it is then enough to take the stalks or the global cohomology.
For us, the most relevant example of such a family comes from the Hitchin fibration. Regard the Hitchin base $\mathcal{A}^r$ as the base parametrizing the family of spectral curves $\mathcal{C} \rightarrow \mathcal{A}^r$. Then in \cite[proposition 3.3]{maulik2014macdonald} is proved that the restriction of this family to the locus where the spectral curves are integral $\mathcal{C}_{int} \rightarrow \mathcal{A}^r_{int}$ satisfies the properties that are necessary for the generalized Macdonald formula to hold.
Note that the fibers of the Hitchin map $h: \mathcal{M}^{n,d} \rightarrow \mathcal{A}^r$ are indeed (torsors over) compactified Jacobians over $\mathcal{A}^r_{int}$. Also, as we already noted in proposition \ref{relhilb}, the fibers of any of the $\chi_\sigma^{n,d}: \mathcal{M}_\sigma^{n,d} \rightarrow \mathcal{A}^r$ over integral spectral curves are Hilbert schemes and we also proved that the smooth locus of the $\mathcal{M}_\sigma^{n,d}$ is contained in the integral locus. Therefore in the case of $\mathcal{A}^r_{int}$ we know what the analogues of the relative compactified Jacobian and relative Hilbert schemes are. In the family of spectral curves, however, also appear curves that are reducible and even non-reduced.
There has been a recent and even broader generalization of the Macdonald formula to reduced locally planar curves in \cite{migliorini2015support}, even though the assumptions on the family of curves that is allowed are more restrictive. Let us give a brief overview for reduced curves, details can be found in \cite{migliorini2015support}.
Consider a flat family $\mathcal{C} \rightarrow \mathcal{B}$ of reduced locally planar curves containing at least a reducible curve. To generalize the Macdonald formula, several technical assumptions are needed and we highlight the most important ones here:
\begin{itemize}
\item The family has to be \emph{independently broken} \cite[definition 1.10]{migliorini2015support} meaning that we should have a finite set of flat families $\mathcal{C}_S \rightarrow \mathcal{B}_S$ indexed by subsets $S$ of a fixed finite set $V$ satisfying some compatibility properties. The idea behind this definition is that we should be able to distinguish the irreducible components of our curves in families.
\item The family has to be \emph{H-smooth} \cite[definition 1.12]{migliorini2015support}, i.e. the relative Hilbert schemes $\mathcal{C}^{[n]}$ should have smooth total space for all $n \geq 0$ and also the families of irreducible components $\mathcal{C}_S \rightarrow \mathcal{B}_S$ should have smooth relative Hilbert schemes.
\end{itemize}
In the case of reduced curves it is also necessary to modify the notion of Jacobian (see \cite[section 2.3]{migliorini2015support}). In fact if a curve has more than one irreducible component, the moduli problem of torsion free sheaves will not be of finite type and then one has to impose a further stability condition. We only mention the main existence result \cite[theorem 2.8]{migliorini2015support}, originally \cite[theorems A and B]{esteves2001compactifying}.
\begin{theorem}
Let $X$ be a geometrically connected projective locally planar reduced curve and $\underline{m}$ a general polarization on $X$. Then there exists a projective scheme $\overline{J}_X(\underline{m})$ which is a fine moduli space for rank one pure one dimensional sheaves that are semistable with respect to $\underline{m}$.
\end{theorem}
It is also crucial to note that given a flat family $\pi: \mathcal{C} \rightarrow \mathcal{B}$ of reduced geometrically connected and locally planar curves, we can always construct, up to an \'etale cover of $\mathcal{B}$, a relative fine compactified Jacobian $\pi^J: \overline{J}_\mathcal{C}\rightarrow \mathcal{B}$. A more detailed discussion on the properties of fine compactified Jacobians can be found in \cite{melo2014fine} and \cite{esteves2001compactifying}.
The formula proved in \cite[theorem 1.16]{migliorini2015support} compares once again $\bigoplus_{n=0}^\infty q^n \mathbb{R} \pi_*^{[n]} \mathbb{Q}$ and $\bigoplus_i q^i \cdot \leftidx{^p}{\mathbb{R}}{^i} \pi_*^J \mathbb{Q} [-i]$.
However, contrary to what happened for integral curves, $\bigoplus_{n=0}^\infty q^n \mathbb{R} \pi_*^{[n]} \mathbb{Q}$ has extra summands that are not supported on the full base of the family, but rather on proper closed subvarieties. These extra summands come from partial normalizations of the curves in the family that can be disconnected and therefore clearly cannot appear for integral curves.
Another fundamental difference is that, in the case of integral curves, the Hilbert scheme of points can be characterized as a moduli space of torsion free sheaves with the extra datum of a section. There is then a forgetful map that will send such a pair to the underlying sheaf which is then a torsion free sheaf and so an element of the compactified Jacobian. When the curve is smooth this is the classical Abel-Jacobi map, sending a divisor on the curve to the associated line bundle. For a sufficiently high number of points such a map is even a projective bundle. In particular this implies that for an integral locally planar curve (or a family of such curves) it is equivalent to require that:
\begin{itemize}
\item[(i)] the relative Hilbert schemes have smooth total space for all $n$
\item[(ii)] the relative Hilbert schemes have smooth total space up to $n=2g-1$
\item[(iii)] the relative compactified Jacobian has smooth total space.
\end{itemize}
When the curve instead is reduced, due to stability issues, we do not have a map from the Hilbert scheme of points to the fine compactified Jacobian. It is still true that if the family of curves is H-smooth then the relative fine compactified Jacobian has smooth total space but the converse implication is not true. For more details about the previous remarks see \cite[section 1]{migliorini2015support}.
Let us discuss briefly what happens for our current example: the family of spectral curves. The first difference we see is that in the family there are non reduced curves. This is possibly the biggest issue because, at the moment, it is not clear what the analogue of the Jacobian should be for such curves. The Hilbert scheme of points for a non reduced curve is defined but it is harder to understand.
For the relative Hilbert scheme of points instead the situation is more complicated. We already proved that the moduli spaces $\mathcal{M}_\sigma^{r,d}$, together with the Hitchin maps $\chi_\sigma^{r,d}: \mathcal{M}_\sigma^{r,d} \rightarrow \mathcal{A}^r$ provide different extensions of the relative Hilbert schemes of points for the family $\mathcal{C}_{int} \rightarrow \mathcal{A}^r_{int}$ of integral spectral curves to $\mathcal{A}^r \setminus \mathcal{A}^r_{int}$. Only one of these extensions, namely $\mathcal{M}_\infty^{r,d}$ will be a relative Hilbert scheme, but we proved that it is not smooth and the singular locus will intersect the locus where the spectral curves are reduced. If we want some kind of H-smoothness property to be satisfied, then the $\mathcal{M}_\varepsilon^{r,d}$ provide smooth extensions but unfortunately not for all combinations of $r$ and $d$.
Another interesting property that the $\mathcal{M}_\varepsilon^{r,d}$ satisfy is the existence of an Abel-Jacobi map $\mathcal{M}_\varepsilon^{r,d} \rightarrow \mathcal{M}^{r,d}$ that forgets the section of the triple. The fibers of this map are projective spaces for all $d$, not necessarily of the same dimension. For large $d$ however, the map is a projective bundle.
\section{A formula for rank 2 Bradlow-Higgs triples}
At this point the natural question is whether or not we can compare
\begin{equation*}
\bigoplus_{n \geq 0} \mathbb{R} (\chi_\varepsilon^{r,n+\theta(r)})_* IC_{\mathcal{M}_\varepsilon^{r,n+\theta(r)}} \cdot q^n
\end{equation*}
and
\begin{equation*}
\frac{\bigoplus_i \leftidx{^p}{\mathbb{R}}{^i} h^{r,d}_* \mathbb{Q} [-i] \cdot q^i}{(1-q)(1-q\L)}
\end{equation*}
or some variant of the two expressions.
Here $\theta$ is a shift function that only depends on the rank $r$. Recall from proposition \ref{degs} that for a rank on pure one dimensional sheaf $\mathcal{F}$ on $T^*C$ not intersecting the divisor at infinity and $E= \pi_*\mathcal{F}$ we have the relation $\deg E= \deg \mathcal{F} - r(r-1)(g-1)$. Therefore $\theta(r)=-r(r-1)(g-1)$ in the formula above will allow us to relate $\mathcal{M}_\varepsilon^{r,d}$ with the relative Hilbert scheme of $d+ r(r-1)(g-1)$ points.
From theorems \ref{motodd} and \ref{moteven} we have the motive of $\mathcal{M}^{2,d}_\varepsilon$. However, the motive will allow to compute the Poincar\'e polynomial only when $\mathcal{M}^{2,d}_\varepsilon$ is smooth, which happens only when either $d<0$ or for $d\geq 4g-5$ odd. Since in general computing intersection cohomology is a very hard task, we will concentrate mainly on the previous degrees.
For the following, denote by $P$ the Poincar\'e polynomial and by $E$ the E-polynomial. We will also use $P^{vir}$ to denote the following specialization of $E$. For a variety $X$:
\begin{equation*}
P^{vir}(X,t)=t^{2 \dim X} E(X, -1/t,-1/t).
\end{equation*}
Note that $P^{vir}(X,t)=P(X,t)$ when the cohomology of $X$ is pure.
Let us also define the following generating functions:
\begin{defn}
\begin{equation*}
F^{sh}(q)= \sum_{n \geq 1-g} \mathbb{R} (\chi_\varepsilon^{2,2n+1})_* (IC_{\mathcal{M}_\varepsilon^{2,2n+1}}) q^{2n+2g-1},
\end{equation*}
\begin{equation*}
F^{mot}(q)= \sum_{n \geq 1-g} [\mathcal{M}_\varepsilon^{2,2n+1}] q^{2n+2g-1},
\end{equation*}
\begin{equation*}
F^{vir}(q,t)= \sum_{n \geq 1-g} P^{vir}(\mathcal{M}_\varepsilon^{2,2n+1},t) q^{2n+2g-1}
\end{equation*}
\end{defn}
We note immediately that $F^{vir}$ is equal to the generating function of the Poincar\'e polynomials for $\deg q \leq 2g-3$ and $\deg q \geq 6g-5$. Also for $2n+1<0$ and $2n+1 \geq 4g-3$ the smoothness of $\mathcal{M}_\varepsilon^{2,2n+1}$ implies that
$$IC_{\mathcal{M}_\varepsilon^{2,2n+1}}=\mathbb{Q}_{\mathcal{M}_\varepsilon^{2,2n+1}}.$$
Using the motives it is possible to compute $F^{mot}$ and $F^{vir}$ explicitly.
\begin{theorem}
We have:
\begin{align*}
&F^{mot}(q)=\L^{4g-3}q^{2g-1}\left (\frac{\L^{g}[J ( C)] Z(C,q^2)}{(\L-1)(1-\L^2 q^2)}-\frac{[J ( C)] Z(C,\L q^2)}{(\L-1)(1- q^2)}\right )+\\
&+\L^{5g-4}q^{4g-4} Z(C,\L q^2) \Gamma(\L^{-1}q^{-1})+\L^{4g-3} Z(C,q^2) \Gamma(q)+\\
&+\Theta(q)
\end{align*}
where
\begin{equation*}
\Gamma(q)=\sum_{j=0}^{g-2} [S^{2j+1}( C)]q^{2j+1}
\end{equation*}
and
\begin{align*}
\Theta(q)=&\sum_{n=0}^{2g-3} \sum_{i=0}^n (\L^{4g-2}-\L^{4g-3}) [S^{i}( C)] \cdot\\
& \cdot\left ([S^{2n+1-i}( C)]-[J( C)]\frac{\L^{2n+2-i-g}-1}{\L-1} \right)q^{2n+2g-1}.
\end{align*}
Also:
\begin{align*}
F^{vir}(q,t)&= \frac{q^{2g-1}(1+t)^{2g} (1+q^2t^3)^{2g}}{(1-t^2)(1-q^2)(1-q^2t^2)(1-q^2t^4)} - \frac{q^{2g-1}t^{2g}(1+t)^{2g} (1+q^2t)^{2g}}{(1-t^2)(1-q^2)(1-q^2t^2)(1-q^2t^4)}+\\
&+\frac{1}{2} q^{2g-2}t^{4g-4} \left ( \frac{(1+q^2 t)^{2g}}{(1-q^2)(1-q^2t^2)}\right ) \left ( {\frac {( 1+qt )^{2g}}{( 1-qt^2 )( 1-q ) }}- \frac{( 1-qt )^{2g}}{( 1+qt^2)( 1+q ) } \right )+\\
&+ \frac{1}{2} q^{2g-2}t^{4g-4} \left ( \frac{(1+q^2 t)^{2g}}{(1-q^2)(1-q^2t^2)}\right ) \left ( \frac{(1+t)^{2g}}{1-t^2} \left ( \frac{qt^{4-2g}}{1-q^2t^4} - \frac{q}{1-q^2}\right ) \right )+\\
&+\frac{1}{2}\left ( \frac{(1+q^2 t^3)^{2g}}{(1-q^2t^2)(1-q^2t^4)}\right )\left ( {\frac {( 1+qt )^{2g}}{( 1-qt^2 )( 1-q ) }}- \frac{( 1-qt )^{2g}}{( 1+qt^2)( 1+q ) } \right )+\\
&- \frac{q^{2g-1}(1+t)^{2g}}{1-t^2}\left ( \frac{(1+q^2 t^3)^{2g}}{(1-q^2t^2)(1-q^2t^4)}\right ) \left ( \frac{1}{1-q^2} - \frac{t^{2g}}{1-q^2t^4} \right )+\\
&+t^{8g-6} E(\Theta(u),-1/t,-1/t)_{|u=qt^2}.
\end{align*}
\begin{proof}
Recall from theorem \ref{motodd} that if $d >0$ odd then:
\begin{align*}
[\mathcal{M}_\varepsilon^{2,d}]=&\L^{1+4(g-1)}[M_\varepsilon^{2,d}]+\sum_{(d_1,d_2) \in I_1^o(d)} \L^{1+3(g-1)+d_2}[S^{d_1}( C)][S^{d_1-d_2+2g-2}]+\\
&+\sum_{(d_1,d_2) \in I_2^o(d)}\L^{1+4(g-1)} [S^{d_2}( C)][S^{d_1-d_2+2g-2}( C)]+\\
&+\sum_{(d_1,d_2) \in I_1^o(d)} (\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] ([S^{d_2}( C)]-[J( C)][\mathbb{C}\P^{d_2-g}])
\end{align*}
and if $d <0 $ odd then:
\begin{align*}
[\mathcal{M}_\varepsilon^{2,d}]=&\sum_{(d_1,d_2) \in I_2^o(d)}\L^{1+4(g-1)} [S^{d_2}( C)][S^{d_1-d_2+2g-2}( C)].
\end{align*}
We can therefore split the computation into 4 main terms:
\begin{align*}
F^{mot}&(q)=\sum_{n \geq 0} \L^{4g-3}[M_\varepsilon^{2,2n+1}] q^{2n+2g-1} +\\
&+\sum_{n \geq 0}q^{2n+2g-1} \left (\sum_{(d_1,d_2) \in I_1^o(2n+1)} \L^{1+3(g-1)+d_2}[S^{d_1}( C)][S^{d_1-d_2+2g-2}] \right ) +\\
&+\sum_{n \geq 1-g}q^{2n+2g-1} \left (\sum_{(d_1,d_2) \in I_2^o(d)}\L^{1+4(g-1)} [S^{d_2}( C)][S^{d_1-d_2+2g-2}( C)]\right )+\\
&+\sum_{n \geq 0} \sum_{(d_1,d_2) \in I_1^o(2n+1)} (\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] \cdot \\
&\cdot \left ([S^{d_2}( C)]-[J( C)] \frac{\L^{d_2+1-g}-1}{\L-1} \right ) q^{2n+2g-1}.
\end{align*}
From \cite[Remark 6.3]{mozgovoy2013moduli}, we deduce:
\begin{align*}
\sum_{n \geq 0} [M_\varepsilon^{2,2n+1}] u^{n}=\frac{\L^{g}[J ( C)] Z(C,u)}{(\L-1)(1-\L^2 u)}-\frac{[J ( C)] Z(C,\L u)}{(\L-1)(1- u)}
\end{align*}
where
$$Z(C,u)=\sum_{n \geq 0} [S^n( C)]u^n$$
is the motivic zeta function of $C$. Therefore:
\begin{align*}
&\sum_{n \geq 0} \L^{4g-3}[M_\varepsilon^{2,2n+1}] q^{2n+2g-1}= \L^{4g-3}q^{2g-1}\sum_{n \geq 0} [M_\varepsilon^{2,2n+1}] q^{2n}=\\
&=\L^{4g-3}q^{2g-1}\left (\frac{\L^{g}[J ( C)] Z(C,q^2)}{(\L-1)(1-\L^2 q^2)}-\frac{[J ( C)] Z(C,\L q^2)}{(\L-1)(1- q^2)}\right ).
\end{align*}
We also have:
\begin{align*}
&\sum_{n \geq 0} \left (\sum_{(d_1,d_2) \in I_1^o(2n+1)} \L^{1+3(g-1)+d_2}[S^{d_1}( C)][S^{d_1-d_2+2g-2}] \right )q^{2n+2g-1}=\\
&=\L^{3g-2}q^{2g-1} \sum_{n=0}^{g-2} \sum_{j=g-2-n}^{g-2} [S^{j+n+2-g}( C)][S^{2j+1}] \L^{n-j+g-1} q^{2n}+\\
&+\L^{3g-2}q^{2g-1} \sum_{n\geq g-1} \sum_{j=0}^{g-2} [S^{j+n+2-g}( C)][S^{2j+1}] \L^{n-j+g-1} q^{2n}=\\
&=\L^{3g-2}q^{2g-1} \sum_{n=0}^{g-2} \sum_{j=g-2-n}^{g-2} [S^{j+n+2-g}( C)][S^{2j+1}] \L^{n-j+g-1} q^{2n}+\\
&+\L^{3g-2}q^{2g-1} \sum_{j=0}^{g-2} \sum_{i\geq j+1} [S^{i}( C)][S^{2j+1}] \L^{i-2j+2g-3} q^{2i-2j+2g-4}=\\
&=\L^{3g-2}q^{2g-1} \sum_{n=0}^{g-2} \sum_{j=g-2-n}^{g-2} [S^{j+n+2-g}( C)][S^{2j+1}] \L^{n-j+g-1} q^{2n}+\\
&+\L^{3g-2}q^{2g-1} \sum_{j=0}^{g-2} [S^{2j+1}] \L^{-2j+2g-3} q^{-2j+2g-4} \sum_{i\geq j+1} [S^{i}( C)]\L^{i} q^{2i}=\\
&=\L^{3g-2}q^{2g-1} \sum_{n=0}^{g-2} \sum_{j=g-2-n}^{g-2} [S^{j+n+2-g}( C)][S^{2j+1}] \L^{n-j+g-1} q^{2n}+\\
&+\L^{3g-2}q^{2g-1} \sum_{j=0}^{g-2} [S^{2j+1}] \L^{-2j+2g-3} q^{-2j+2g-4} \left(Z(C,\L q^2)- \sum_{i=0}^j [S^{i}( C)]\L^{i} q^{2i}\right)=\\
&=\L^{5g-4}q^{4g-4} Z(C,\L q^2) \Gamma(\L^{-1}q^{-1}).
\end{align*}
For the other type of split fixed points we get:
\begin{align*}
&\sum_{n \geq 1-g} \sum_{(d_1,d_2) \in I_2^o(d)}\L^{1+4(g-1)} [S^{d_2}( C)][S^{d_1-d_2+2g-2}( C)]q^{2n+2g-1}=\\
&=\L^{4g-3}q^{2g-1} \sum_{n=1-g}^{-1} \sum_{j=0}^{n-1+g}[S^{2j+1}( C)][S^{n-1+g-j}( C)] q^{2n}+\\
&+\L^{4g-3}q^{2g-1} \sum_{n\geq 0} \sum_{j=0}^{g-2}[S^{2j+1}( C)][S^{n-1+g-j}( C)] q^{2n}=\\
&=\L^{4g-3}q^{2g-1} \sum_{n=1-g}^{-1} \sum_{j=0}^{n-1+g}[S^{2j+1}( C)][S^{n-1+g-j}( C)] q^{2n}+\\
&+\L^{4g-3}q^{2g-1} \sum_{j=0}^{g-2} \sum_{i\geq g-1-j} [S^{2j+1}( C)][S^{i}( C)] q^{2i+2j+2-2g}=\\
&=\L^{4g-3}q^{2g-1} \sum_{n=1-g}^{-1} \sum_{j=0}^{n-1+g}[S^{2j+1}( C)][S^{n-1+g-j}( C)] q^{2n}+\\
&+\L^{4g-3}q^{2g-1} \sum_{j=0}^{g-2} [S^{2j+1}( C)] q^{2j+2-2g}\sum_{i\geq g-1-j} [S^{i}( C)] q^{2i}=\\
&=\L^{4g-3}q^{2g-1} \sum_{n=1-g}^{-1} \sum_{j=0}^{n-1+g}[S^{2j+1}( C)][S^{n-1+g-j}( C)] q^{2n}+\\
&+\L^{4g-3}q^{2g-1} \sum_{j=0}^{g-2} [S^{2j+1}( C)] q^{2j+2-2g}\left( Z(C,q^2) - \sum_{i=0}^{g-2-j} [S^{i}( C)] q^{2i}\right )=\\
&=\L^{4g-3} Z(C,q^2) \Gamma(q).
\end{align*}
For the last term we have:
\begin{align*}
&\sum_{n \geq 0}\sum_{(d_1,d_2) \in I_1^o(d)} (\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] \left([S^{d_2}( C)]-[J( C)]\frac{\L^{d_2+1-g}-1}{\L-1}\right)q^{2n+2g-1} =\\
&=\sum_{n=0}^{g-2} \sum_{i=0}^n (\L^{4g-2}-\L^{4g-3}) [S^{i}( C)] \left([S^{2n+1-i}( C)]-[J( C)]\frac{\L^{2n+2-i-g}-1}{\L-1}\right)q^{2n+2g-1}+\\
&+\sum_{n=g-1}^{2g-3} \sum_{i=2n-2g+3}^n (\L^{4g-2}-\L^{4g-3}) [S^{i}( C)]\cdot\\
&\cdot \left([S^{2n+1-i}( C)]-[J( C)]\frac{\L^{2n+2-i-g}-1}{\L-1}\right)q^{2n+2g-1}
\end{align*}
since for $d_2 \geq 2g-1$,
$$[S^{d_2}( C)]-[J( C)]\frac{\L^{d_2+1-g}-1}{\L-1}=0.$$
For the same reason,
\begin{align*}
&\sum_{n=g-1}^{2g-3} \sum_{i=2n-2g+3}^n (\L^{4g-2}-\L^{4g-3}) [S^{i}( C)] \cdot \\
&\cdot\left([S^{2n+1-i}( C)]-[J( C)]\frac{\L^{2n+2-i-g}-1}{\L-1} \right)q^{2n+2g-1}=\\
=&\sum_{n=g-1}^{2g-3} \sum_{i=0}^n (\L^{4g-2}-\L^{4g-3}) [S^{i}( C)] \left([S^{2n+1-i}( C)]-[J( C)]\frac{\L^{2n+2-i-g}-1}{\L-1}\right)q^{2n+2g-1}
\end{align*}
and therefore
\begin{align*}
&\sum_{n \geq 0}\sum_{(d_1,d_2) \in I_1^o(d)} (\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] \left ([S^{d_2}( C)]-[J( C)]\frac{\L^{d_2+1-g}-1}{\L-1}\right)q^{2n+2g-1} =\\
=&\sum_{n=0}^{2g-3} \sum_{i=0}^n (\L^{4g-2}-\L^{4g-3}) [S^{i}( C)] \left([S^{2n+1-i}( C)]-[J( C)]\frac{\L^{2n+2-i-g}-1}{\L-1}\right)q^{2n+2g-1}.
\end{align*}
In order to get the expression for $F^{vir}(q,t)$ we can argue as follows. Since
$$P^{vir}(\mathcal{M}_\varepsilon^{2,2n+1},t)= t^{2 \dim \mathcal{M}_\varepsilon^{2,2n+1}} E(\mathcal{M}_\varepsilon^{2,2n+1},-1/t,-1/t)$$
and $\dim \mathcal{M}_\varepsilon^{2,2n+1}=2n+1+1+6g-6=2n+2+6(g-1)$ we have:
\begin{equation*}
F^{vir}(q,t)=t^{8g-6} E(F^{mot}(u) , -1/t,-1/t)_{|u=qt^2}.
\end{equation*}
Recall that:
\begin{align*}
&E(\L)=xy\\
&E(Z(C,u))=\frac{(1-xu)^g(1-yu)^g}{(1-u)(1-uxy)}\\
&E(J( C))=(1-x)^g(1-y)^g.
\end{align*}
Also, since:
\begin{equation*}
\sum_{n=0}^{2g-2} [S^n( C)] u^n=Z(C,u)- \frac{\L^g[J( C)] u^{2g-1}}{(\L-1)(1-\L u)}+\frac{[J( C)] u^{2g-1}}{(\L-1)(1-u)},
\end{equation*}
we get:
\begin{align*}
&\Gamma(u)=\frac{1}{2}\left ( Z(C,u)-Z(C,-u)\right)-\frac{1}{2}\left (\frac{\L^g[J( C)] u^{2g-1}}{(\L-1)(1-\L u)}+\frac{\L^g[J( C)] u^{2g-1}}{(\L-1)(1+\L u)} \right)+\\
&+\frac{1}{2}\left ( \frac{[J( C)] u^{2g-1}}{(\L-1)(1-u)} + \frac{[J( C)] u^{2g-1}}{(\L-1)(1+u)} \right).
\end{align*}
The last ingredient we need to compute $F^{vir}$ is:
\begin{align*}
&E(\Gamma(u),-1/t,-1/t)=\frac{1}{2} \left ( \frac{(1+u/t)^{2g}}{(1-u)(1-u/t^2)}-\frac{(1-u/t)^{2g}}{(1+u)(1+u/t^2)}\right )+\\
&-\frac{1}{2} \frac{(1+t)^{2g}}{t^{4g-2}(1-t^2)}\left (\frac{u^{2g-1}}{(1-u/t^2)}+\frac{u^{2g-1}}{(1+u/t^2)} \right )+\\
&+\frac{1}{2} \frac{(1+t)^{2g}}{t^{2g-2}(1-t^2)}\left (\frac{u^{2g-1}}{(1-u)}+\frac{u^{2g-1}}{(1+u)} \right )
\end{align*}
Therefore:
\begin{align*}
F^{vir}(q,t)&= \frac{q^{2g-1}(1+t)^{2g} (1+q^2t^3)^{2g}}{(1-t^2)(1-q^2)(1-q^2t^2)(1-q^2t^4)} - \frac{q^{2g-1}t^{2g}(1+t)^{2g} (1+q^2t)^{2g}}{(1-t^2)(1-q^2)(1-q^2t^2)(1-q^2t^4)}+\\
&+\frac{1}{2} q^{2g-2}t^{4g-4} \left ( \frac{(1+q^2 t)^{2g}}{(1-q^2)(1-q^2t^2)}\right ) \left ( {\frac {( 1+qt )^{2g}}{( 1-qt^2 )( 1-q ) }}- \frac{( 1-qt )^{2g}}{( 1+qt^2)( 1+q ) } \right )+\\
&+ \frac{1}{2} q^{2g-2}t^{4g-4} \left ( \frac{(1+q^2 t)^{2g}}{(1-q^2)(1-q^2t^2)}\right ) \left ( \frac{(1+t)^{2g}}{1-t^2} \left ( \frac{qt^{4-2g}}{1-q^2t^4} - \frac{q}{1-q^2}\right ) \right )+\\
&+\frac{1}{2}\left ( \frac{(1+q^2 t^3)^{2g}}{(1-q^2t^2)(1-q^2t^4)}\right )\left ( {\frac {( 1+qt )^{2g}}{( 1-qt^2 )( 1-q ) }}- \frac{( 1-qt )^{2g}}{( 1+qt^2)( 1+q ) } \right )+\\
&- \frac{q^{2g-1}(1+t)^{2g}}{1-t^2}\left ( \frac{(1+q^2 t^3)^{2g}}{(1-q^2t^2)(1-q^2t^4)}\right ) \left ( \frac{1}{1-q^2} - \frac{t^{2g}}{1-q^2t^4} \right )+\\
&+t^{8g-6} E(\Theta(u),-1/t,-1/t)_{|u=qt^2}.
\end{align*}
\end{proof}
\end{theorem}
We can also prove some properties of $F^{mot}$ and $F^{vir}$. Note that, since we have an explicit formula, the proof of these properties is straightforward. However, they can be deduced \emph{a priori} from the geometry of the $\mathcal{M}_\varepsilon^{2,d}$
\begin{prop}
The following holds for $F^{mot}$ and $F^{vir}$:
\begin{itemize}
\item[(i)] There exists a polynomial $Q^{mot}(q)$ such that
\begin{equation*}
F^{mot}(q)=\frac{Q^{mot}(q)}{(1-q^2)(1-q^2\L^2)}
\end{equation*}
and a polynomial $Q^{vir}(q,t)$ such that
\begin{equation*}
F^{vir}(q,t)=\frac{Q^{vir}(q,t)}{(1-q^2)(1-q^2t^4)}
\end{equation*}
\item[(ii)] $Q^{mot}(1)=(1+\L)[\mathcal{M}^{2,1}]$ and $Q^{vir}(1,t)=(1+t^2)P(\mathcal{M}^{2,1},t)$
\item[(iii)] $Q^{vir}(q,t)$ satisfies $$Q^{vir}(q,t)=(qt)^{8g-4}Q^{vir}(q^{-1}t^{-2},t)$$ and $F^{vir}$ satisfies $$F^{vir}(q,t)=(qt)^{8g-8}F^{vir}(q^{-1}t^{-2},t).$$
\end{itemize}
\begin{proof}
\begin{itemize}
\item[(i)] First of all observe that for $n \geq 3g-3$ the Abel-Jacobi map $\mathcal{M}_\varepsilon^{2,2n+1} \rightarrow \mathcal{M}^{2,2n+1}$ is a projective bundle of rank $2n+2-2g$. Also, $\mathcal{M}^{2,2n+1}$ is always isomorphic to $\mathcal{M}^{2,1}$. In particular, for $n \geq 3g-1$ we have:
\begin{equation*}
[\mathcal{M}_\varepsilon^{2,2n+1}]-(1+\L^2)[\mathcal{M}_\varepsilon^{2,2n-1}] + \L^2[\mathcal{M}_\varepsilon^{2,2n-3})]=0
\end{equation*}
and
\begin{equation*}
P(\mathcal{M}_\varepsilon^{2,2n+1},t)-(1+t^4)P(\mathcal{M}_\varepsilon^{2,2n-1},t) + t^4P(\mathcal{M}_\varepsilon^{2,2n-3})=0.
\end{equation*}
Now the property for $F^{mot}$ can be deduced from the fact that:
\begin{align*}
&(1-q^2)(1-q^2\L^2)F^{mot}(q)= Q^{mot}(q)+\\
&+\sum_{n \geq 3g-1} q^{2n+1+2g-2} ( [\mathcal{M}_\varepsilon^{2,2n+1}]-(1+\L^2)[\mathcal{M}_\varepsilon^{2,2n-1}] + \L^2[\mathcal{M}_\varepsilon^{2,2n-3}] ) =\\
&= Q^{mot}(q)
\end{align*}
for some polynomial $Q^{mot}$. And for $F^{vir}$:
\begin{align*}
&(1-q^2)(1-q^2t^4)F^{vir}(q,t)= Q^{vir}(q,t)+\\
&+\sum_{n \geq 3g-1} q^{2n+1+2g-2} ( P(\mathcal{M}_\varepsilon^{2,2n+1},t)-(1+t^4)P(\mathcal{M}_\varepsilon^{2,2n-1},t)+ t^4P(\mathcal{M}_\varepsilon^{2,2n-3},t) ) =\\
&= Q^{vir}(q,t)
\end{align*}
for some polynomial $Q^{vir}$. We can retrieve an explicit expression for $Q^{vir}$ by observing that:
\begin{align*}
&(1-q^2)(1-q^2t^4)F(q,t) = \\
&=q P(\mathcal{M}_\varepsilon^{2,3-2g},t) + q^3 P(\mathcal{M}_\varepsilon^{2,5-2g},t)-q^3 (1+t^4) P(\mathcal{M}_\varepsilon^{2,3-2g},t) + \\
&+\sum_{n=3-g}^{3g-2}q^{2n+1+2g-2} (P(\mathcal{M}_\varepsilon^{2,2n+1},t)-(1+t^4)P(\mathcal{M}_\varepsilon^{2,2n-1},t) + t^4P(\mathcal{M}_\varepsilon^{2,2n-3},t)).
\end{align*}
\item[(ii)] For $Q^{mot}$ we have:
\begin{align*}
&F^{mot}(q)=\sum_{n=1-g}^{3g-4}q^{2n+1+2g-2} [\mathcal{M}_\varepsilon^{2,2n+1}] + \sum_{n\geq 3g-3}q^{2n+1+2g-2} [\mathcal{M}_\varepsilon^{2,2n+1}]=\\
&=\sum_{n=1-g}^{3g-4}q^{2n+1+2g-2} [\mathcal{M}_\varepsilon^{2,2n+1}] + \frac{[\mathcal{M}^{2,1}]}{\L-1}\left ( \frac{\L^{4g-3}q^{8g-7}}{1-\L^2q^2} - \frac{q^{8g-7} }{1-q^2}\right ).
\end{align*}
Therefore:
\begin{align*}
Q^{mot}(1)&=\left ((1-q^2)(1-\L^2 q^2)F^{mot}(q) \right )_{|q=1}=(1+\L)[\mathcal{M}^{2,1}].
\end{align*}
For $Q^{vir}$, similarly:
\begin{align*}
&F^{vir}(q,t)=\sum_{n=1-g}^{3g-4}q^{2n+1+2g-2} P^{vir}(\mathcal{M}_\varepsilon^{2,2n+1},t) + \sum_{n\geq 3g-3}q^{2n+1+2g-2} P(\mathcal{M}_\varepsilon^{2,2n+1},t)=\\
&=\sum_{n=1-g}^{3g-4}q^{2n+1+2g-2} P^{vir}(\mathcal{M}_\varepsilon^{2,2n+1},t) + \sum_{n\geq 3g-3}q^{2n+1+2g-2} P(\mathcal{M}^{2,1},t)P(\mathbb{C}\P^{2n+2-2g})=\\
&=\sum_{n=1-g}^{3g-4}q^{2n+1+2g-2} P^{vir}(\mathcal{M}_\varepsilon^{2,2n+1},t) + \frac{q^{2g-1}P(\mathcal{M}^{2,1},t)}{1-t^2}\sum_{n\geq 3g-3}q^{2n} (1-t^{4n+4-4g})=\\
&=\sum_{n=1-g}^{3g-4}q^{2n+1+2g-2} P^{vir}(\mathcal{M}_\varepsilon^{2,2n+1},t) + \frac{q^{2g-1}P(\mathcal{M}^{2,1},t)}{1-t^2}\left ( \frac{q^{6g-6}}{1-q^2} - \frac{q^{6g-6} t^{8g-8}}{1-q^2t^4}\right ).
\end{align*}
Therefore:
\begin{align*}
Q^{vir}(1,t)&=\left ((1-q^2)(1-q^2t^4)F^{vir}(q,t) \right )_{|q=1} = (1+t^2) P(\mathcal{M}^{2,1},t).
\end{align*}
\item[(iii)] The functional equation for $F^{vir}$ follows from the one for $Q^{vir}$. First of all we prove the following motivic identity, valid for all $n \geq g-1$:
\begin{equation*}
[\mathcal{M}_\varepsilon^{2,2n+1}]-\L^{2n-2g+3}[\mathcal{M}_\varepsilon^{2,4g-4-2n-1}] =[\mathcal{M}^{2,1}][\mathbb{C}\P^{2n-2g+2}].
\end{equation*}
We start by defining:
\begin{equation*}
Y_k^{2n+1}=\{(E,\phi) \in \mathcal{M}^{2,2n+1} : \dim H^0(E)=k\}.
\end{equation*}
Recall that we have a Serre duality isomorphism:
\begin{align*}
\mathcal{M}^{2,2n+1} &\rightarrow \mathcal{M}^{2,4g-4-2n-1}\\
(E,\phi) & \mapsto (K E^*,\phi).
\end{align*}
Since $\dim H^0(E)- \dim H^0(K E^*) = 2n+1+2-2g$ we see that:
\begin{equation*}
[Y_k^{2n+1}]=[Y_{k+2g-3-2n}^{4g-4-2n-1}].
\end{equation*}
This relation in turn implies that:
\begin{align*}
&[\mathcal{M}_\varepsilon^{2,2n+1}]-\L^{2n-2g+3}[\mathcal{M}_\varepsilon^{2,4g-4-2n-1}]=\\
&= \sum_i [Y_i^{2n+1}][\mathbb{C}\P^{i-1}] - \L^{2n-2g+3}\sum_j [Y_j^{4g-4-2n-1}][\mathbb{C}\P^{j-1}]=\\
&=\sum_i [Y_i^{2n+1}][\mathbb{C}\P^{i-1}- \L^{2n-2g+3} ] \sum_j [Y_{j+2n+3-2g}^{2n+1}][\mathbb{C}\P^{j-1}]=\\
&=\sum_i [Y_i^{2n+1}][\mathbb{C}\P^{i-1}]- \L^{2n-2g+3} \sum_i [Y_{i}^{2n+1}][\mathbb{C}\P^{i+2g-4-2n}] =\\
&=\sum_i [Y_i^{2n+1}] \frac{\L^{i}-1-\L^{i}+\L^{2n-2g+3}}{\L-1}=[\mathcal{M}^{2,1}][\mathbb{C}\P^{2n-2g+2}].
\end{align*}
The motivic relation, in turn, implies the following relation between the virtual Poincar\'e polynomials:
\begin{equation*}
P^{vir}(\mathcal{M}_\varepsilon^{2,2n+1},t)-t^{4n+6-4g}P^{vir}(\mathcal{M}_\varepsilon^{2,4g-5-2n},t)=P^{vir}(\mathcal{M}^{2,1},t)P(\mathbb{C}\P^{2n-2g+2},t).
\end{equation*}
To prove the identity for $Q^{vir}$, let us first write:
\begin{equation*}
Q^{vir}(q,t)=\sum_{l=0}^{4g-3} a_l(t) q^{2l+1}
\end{equation*}
for:
\begin{equation*}
a_l(t)=P^{vir}(\mathcal{M}_\varepsilon^{2,2l+3-2g},t)-(1+t^4)P^{vir}(\mathcal{M}_\varepsilon^{2,2l+1-2g},t)+t^4P^{vir}(\mathcal{M}_\varepsilon^{2,2l-1-2g},t).
\end{equation*}
Note that some of the coefficients of $Q^{vir}$ do not include three terms, but the previous formula holds for those coefficients as well since for the appropriate choice of $l$ some of the three moduli spaces appearing in the formula can be empty.\\
The identity $Q^{vir}(q,t)=(qt)^{8g-4}Q^{vir}(q^{-1}t^{-2},t)$ is equivalent to:
\begin{equation*}
a_l(t)=t^{4l+6-8g}a_{4g-3-l}(t).
\end{equation*}
Also note that it is enough to check the previous equation for $l \geq 2g-1$. In fact assume it holds for all $l \geq 2g-1$ and choose $k \leq 2g-2$. Then $4g-3-k \geq 2g-1$ and so:
\begin{equation*}
a_{4g-3-k}(t)=t^{8g-6-4k}a_{k}(t)
\end{equation*}
which is just the equation for $k$.\\
We can compute:
\begin{align*}
a_l(t)-&t^{4l+6-8g}a_{4g-3-l}(t)=\left(P^{vir}(\mathcal{M}_\varepsilon^{2,2l+3-2g},t)-t^{4l+10-8g}P^{vir}(\mathcal{M}_\varepsilon^{2,6g-7-2l},t)\right)+\\
&-(1+t^4)\left (P^{vir}(\mathcal{M}_\varepsilon^{2,2l+1-2g},t) -t^{4l+6-8g} P^{vir}(\mathcal{M}_\varepsilon^{2,6g-5-2l},t)\right )+\\&+t^4 \left ( P^{vir}(\mathcal{M}_\varepsilon^{2,2l-1-2g},t) -t^{4l+2-8g} P^{vir}(\mathcal{M}_\varepsilon^{2,6g-3-2l},t)\right).
\end{align*}
For $2l-1-2g \geq 2g-1$, i.e. for $l \geq 2g$ we can use the identity we proved above, obtaining:
\begin{align*}
a_l(t)-&t^{4l+6-8g}a_{4g-3-l}(t)= P^{vir}(\mathcal{M}^{2,1},t)P(\mathbb{C}\P^{2l-4g+4},t)+\\
&-(1+t^4)P^{vir}(\mathcal{M}^{2,1},t)P(\mathbb{C}\P^{2l-4g+2},t)+t^4 P^{vir}(\mathcal{M}^{2,1},t)P(\mathbb{C}\P^{2l-4g},t)=\\
&=\frac{P^{vir}(\mathcal{M}^{2,1},t)}{1-t^2} (1-t^{4l-8g+10}-(1+t^4)(1-t^{4l-8g+6})+t^4(1-t^{4l-8g+2}))=0.
\end{align*}
Using the same techniques, with a bit of care, we can check the last case $l=2g-1$:
\begin{align*}
a_{2g-1}(t)-&t^{2}a_{2g-2}(t)=\left(P^{vir}(\mathcal{M}_\varepsilon^{2,2g+1},t)-t^{6}P^{vir}(\mathcal{M}_\varepsilon^{2,2g-5},t)\right)+\\
&-(1+t^4)\left (P^{vir}(\mathcal{M}_\varepsilon^{2,2g-1},t) - t^2 P(\mathcal{M}_\varepsilon^{2,2g-3},t)\right )+\\
&+t^4 P^{vir}(\mathcal{M}_\varepsilon^{2,2g-3},t) -t^{2} P^{vir}(\mathcal{M}_\varepsilon^{2,2g-1},t)=\\
&=P^{vir}(\mathcal{M}^{2,1},t) \left( P(\mathbb{C}\P^2,t)-(1+t^4)-t^2\right)=0.
\end{align*}
This completes the proof.
\end{itemize}
\end{proof}
\end{prop}
We now define the second kind of generating functions involved in the formula.
\begin{defn}
Denote by $\text{odd}_q$ the operator that acts on $f \in R[[q]]$ (for some ring $R$) by deleting the even powers of $q$ in $f$. Define:
\begin{equation*}
G(q,t)=\text{odd}_q \left (\frac{PH(\mathcal{M}^{2,1},q,t)}{(1-q)(1-qt^2)} \right )
\end{equation*}
and
\begin{equation*}
G^{sh}(q)=\text{odd}_q \left (\frac{\bigoplus_{i=0}^{8g-6} IC\left (\bigwedge^i\mathbb{R}^1 \right) }{(1-q\mathbb{Q})(1-q\mathbb{Q}[-2](-1))} \right )
\end{equation*}
where $\mathbb{R}^1$ is the local system $\mathbb{R}^1(h^{2,1}_{sm})_* \mathbb{Q}$ obtained by first restricting $h^{2,1}: \mathcal{M}^{2,1}\rightarrow \mathcal{A}^2$ to the locus of smooth spectral curves and then pushing forward the constant sheaf.
\end{defn}
\begin{rmk}
\label{decthm}
Let us make a couple comments about the previous definition. First of all, the reason we take the odd powers in $G$ and $G^{sh}$ is that we want to compare them with $F^{vir}$ and $F^{sh}$ respectively and they only contains odd powers.
Second, it follows from \cite[theorem 1.1.2]{de2010topology} and \cite[lemma 1.3.5]{de2010topology} that:
\begin{equation*}
\mathbb{R} h^{2,1}_* \mathbb{Q} = \bigoplus_{i=0}^{8g-6} \leftidx{^p}{\mathbb{R}}{^i} h^{2,1}_* \mathbb{Q}[-i] \oplus \mathscr{L} = \bigoplus_{i=0}^{8g-6} IC\left (\bigwedge^i\mathbb{R}^1 \right) \oplus \mathscr{L}
\end{equation*}
where $\mathscr{L}$ is a complex supported on the locus of reduced curves in $\mathcal{A}^2$. In other words, the summands of $\mathbb{R} h^{2,1}_* \mathbb{Q}$ that are supported on the whole $\mathcal{A}^2$ are exactly the middle extensions of the external powers of $\mathbb{R}^1$, while $\mathscr{L}$ is an extra summand with proper support. Another important remark is that the perversity is actually determined by the index of the external power of $\mathbb{R}^1$. Furthermore, $\mathscr{L}$ will have no global cohomology, as I learned from private communication (\cite{demighein}). These facts together imply that taking the global cohomology of the object:
\begin{equation*}
\bigoplus_{i=0}^{8g-6} \leftidx{^p}{\mathbb{R}}{^i} h^{2,1}_* \mathbb{Q}[-i] q^i \in D_c^b(\mathcal{A}^2)[[q]]
\end{equation*}
will compute $PH(\mathcal{M}^{2,1},q,t)$.
Last, note that from the results summarized in section \ref{charvar}, we can actually explicitly compute $G$.
\end{rmk}
Observe that $G$ satisfies properties that are similar to the ones of $F^{vir}$.
\begin{prop}
The following holds for $G$:
\begin{itemize}
\item[(i)] There exists a polynomial $V(q,t)$ such that
\begin{equation*}
G(q,t)=\frac{V(q,t)}{(1-q^2)(1-q^2t^4)}
\end{equation*}
\item[(ii)] $V(1,t)=(1+t^2)P(\mathcal{M}^{2,1},t)$
\item[(iii)] $V(q,t)$ satisfies $$V(q,t)=(qt)^{8g-4}V(q^{-1}t^{-2},t)$$ and $G$ satisfies $$G(q,t)=(qt)^{8g-8}G(q^{-1}t^{-2},t)$$
\end{itemize}
\begin{proof}
\begin{itemize}
\item[(i)] Note that, if $U(q,t) \in \mathbb{Z}[t][[q]]$, then
\begin{equation*}
\text{odd}_q U(q,t)= \frac{1}{2} \left ( U(q,t)- U(-q,t)\right ).
\end{equation*}
Therefore
\begin{align*}
G(q,t)&=\frac{1}{2} \left (\frac{PH(\mathcal{M}^{2,1},q,t)}{(1-q)(1-qt^2)} - \frac{PH(\mathcal{M}^{2,1},-q,t)}{(1+q)(1+qt^2)}\right )=\\
&=\frac{1}{2} \left (\frac{(1+q)(1+qt^2) PH(\mathcal{M}^{2,1},q,t)- (1-q)(1-qt^2)PH(\mathcal{M}^{2,1},-q,t)}{(1-q^2)(1-q^2t^4)}\right )
\end{align*}
so
\begin{equation*}
V(q,t)=\frac{1}{2} \left ((1+q)(1+qt^2) PH(\mathcal{M}^{2,1},q,t)- (1-q)(1-qt^2)PH(\mathcal{M}^{2,1},-q,t) \right ).
\end{equation*}
\item[(ii)] We have:
\begin{equation*}
V(1,t)=\frac{1}{2} \left (2(1+t^2) PH(\mathcal{M}^{2,1},1,t) \right )=(1+t^2)P(\mathcal{M}^{2,1},t).
\end{equation*}
\item[(iii)] From \cite[Corollary 1.1.4]{hausel2008mixed} we see that
\begin{align*}
(qt)^{8g-6}PH(\mathcal{M}^{2,1},q^{-1}t^{-2},t)=PH(\mathcal{M}^{2,1},q,t).
\end{align*}
Therefore:
\begin{align*}
(qt)^{8g-4}& V(q^{-1}t^{-2},t)=\frac{1}{2} ((qt)^{8g-6}(1+q)(1+qt^2) PH(\mathcal{M}^{2,1},q^{-1}t^{-2},t)+\\
&- (qt)^{8g-6}(1-q)(1-qt^2)PH(\mathcal{M}^{2,1},-q^{-1}t^{-2},t) )=\\
&=\frac{1}{2} \left ((1+q)(1+qt^2) PH(\mathcal{M}^{2,1},q,t)- (1-q)(1-qt^2)PH(\mathcal{M}^{2,1},-q,t) \right )=\\
&= V(q,t).
\end{align*}
Clearly the formula for $G$ follows from the formula for $V$.
\end{itemize}
\end{proof}
\end{prop}
\begin{rmk}
Even though for rank 2 it is possible to compute the difference explicitly, it is quite interesting to understand an a priori reason why the two functions should have such a relation. The idea is to compare $F^{sh}$ with $G^{sh}$ and $P(F^{sh}(q),t)$ with $G(q,t)$. $P(F^{sh}(q),t)$ will be the generating function of the Poincar\'e polynomials of the sheaves contained in $F^{sh}$, which are the intersection cohomology sheaves of the $\mathcal{M}_\varepsilon^{2,2n+1}$. Note that, since for $2n+1<0$ and for $2n+1>4g-4$ $\mathcal{M}_\varepsilon^{2,2n+1}$ is smooth, then $F^{vir}$ and $P(F^{sh}(q),t)$ coincide for $\deg q \leq 2g-3$ and for $\deg q \geq 6g-5$.
\end{rmk}
In the next few pages we want to prove that $G$ and $P(F^{sh}(q),t)$ coincide for $\deg q \leq 2g-3$ and for $\deg q \geq 6g-5$. Let us first recall the following result. See \cite[Theorem 3]{chaureport}.
\begin{prop}
\label{dimsupp}
Let $f: X \rightarrow Y$ be a proper map of smooth algebraic varieties whose fibers have constant dimension. Denote by $m$ the dimension of $Y$ and by $n$ the dimension of $X$. Let $\mathscr{S}$ be summand of $f_* \mathbb{Q}_X$ whose support has dimension $s$. Then $s \geq 2m-n$.
\end{prop}
\begin{theorem}
\label{BHMSMY}
$F^{sh}$ and $G^{sh}$ coincide for $\deg q \leq 2g-3$. $P(F^{sh}(q),t)$ and $G$ coincide for $\deg q \leq 2g-3$ and for $\deg q \geq 6g-5$. Furthermore $P(F^{sh}(q),t)-G(q,t)$ is a polynomial with non-negative coefficients.
\begin{proof}
Let us start by considering the Hitchin map $h: \mathcal{M}^{2,1} \rightarrow \mathcal{A}^2$ and the object $\mathbb{R} h^{2,1}_* \mathbb{Q} \in D_c^b(\mathcal{A}^2)$. Consider also $\mathcal{A}^2_{sm} \subset \mathcal{A}^2$, the locus of smooth spectral curves and call $\mathbb{R}^1$ the local system $\mathbb{R}^1 (h_{sm})_* \mathbb{Q}$ where $h_{sm}$ is the restriction of $h$ to the inverse image of $\mathcal{A}^2_{sm}$.
As we noted in remark \ref{decthm}, we have:
\begin{equation*}
\mathbb{R} h^{2,1}_* \mathbb{Q} = \bigoplus_{i=0}^{8g-6} \leftidx{^p}{\mathbb{R}}{^i} h^{2,1}_* \mathbb{Q}[-i] \oplus \mathscr{L} = \bigoplus_{i=0}^{8g-6} IC\left (\bigwedge^i\mathbb{R}^1 \right) \oplus \mathscr{L}
\end{equation*}
First of all we prove that:
\begin{equation*}
\bigoplus_{2-2g \leq d <0} q^{d+2g-2} \mathbb{R} (\chi_\varepsilon^{2,d})_* \mathbb{Q} = \left (\frac{\bigoplus_{i=0}^{8g-6} \leftidx{^p}{\mathbb{R}}{^i} h^{2,1}_* \mathbb{Q}[-i] q^i}{(1-q \mathbb{Q})(1-q \L)} \right )_{|\deg(q) \leq 2g-3}.
\end{equation*}
which implies that $F^{sh}$ and $G^{sh}$ coincide for $\deg q \leq 2g-3$ and is a bit stronger because it also says something about the even degree case. Recall from proposition \ref{negdegrk2} that $\mathcal{M}_{\varepsilon}^{2,d}=\mathcal{M}_{\infty}^{2,d}$ for $d < 0$.
When we restrict to the locus $\mathcal{A}^2_{int}$ of integral spectral curves, the formula is true since the restriction of $\chi_\sigma^{2,d}$ to $\mathcal{A}^2_{int}$ is always the Hilbert scheme of $d+2g-2$ points relative to the family of integral spectral curves, independently of $\sigma$. The formula then descends from the results in \cite{migliorini2011support}. This means that if the left hand side and the right hand side differ then, since the $\leftidx{^p}{\mathbb{R}}{^i} h^{2,1}_* \mathbb{Q}[-i] $ are supported on the whole $\mathcal{A}^2$, there must be summands of the left hand side that are supported inside $\mathcal{A}^2\setminus \mathcal{A}^2_{int}$.
For the second step, let us restrict the formula to the locus $\mathcal{A}^2_{red} \subset \mathcal{A}^2$, i.e. the locus of reduced (but possibly reducible) spectral curves. Note that we cannot immediately conclude that the formula holds using the results from \cite{migliorini2015support} because the family of reduced spectral curves does not satisfy the conditions in the paper. However we can argue as follows. Assume that there is a summand with support contained in the closure of $\mathcal{A}^2_{red}$ and denote by $s$ its dimension. Using proposition \ref{dimsupp}, we get:
\begin{equation*}
s \geq 8g-6 - 4g +3 - d -2g +2=2g-1-d.
\end{equation*}
and for $d \leq -1$ we get $s \geq 2g$. Since $\mathcal{A}^2_{red}$ is irreducible, this implies that the support is actually the closure of $\mathcal{A}^2_{red}$.
Note that the generic reduced and reducible curve is nodal, since $\mathcal{A}^2_{red}$ is the image of the following map:
\begin{align*}
\Sym^2 \left ( H^0(K) \right )\setminus \Delta &\rightarrow \mathcal{A}^2\\
(\psi_1, \psi_2) & \mapsto (-\psi_1-\psi_2, \psi_1 \psi_2)
\end{align*}
and two generic sections of $H^0(K)$ intersect in $2g-2$ points.
Using \cite[Lemma 2.13]{migliorini2015support}, we can deduce that if the stalks of
\begin{equation*}
\bigoplus_{2-2g \leq d <0} q^{d+2g-2} \mathbb{R} (\chi_\varepsilon^{2,d})_* \mathbb{Q}
\end{equation*}
and
\begin{equation*}
\left (\frac{\bigoplus_{i=0}^{8g-6} \leftidx{^p}{\mathbb{R}}{^i} h^{2,1}_* \mathbb{Q}[-i] q^i}{(1-q \mathbb{Q})(1-q \L)} \right )_{|\deg(q) \leq 2g-3}
\end{equation*}
were to be different at a generic nodal curve of $\mathcal{A}^2_{red}$ then their weight polynomial would be different. Since we know that the potential support could only be the closure of $\mathcal{A}^2_{red}$, if we verify that the weight polynomials of the stalks at the generic nodal curves are the same, we can conclude that there is no such support. More details about the strategy outlined here can be found in \cite{migliorini2015support}.
Let $\mathcal{C}$ be a generic nodal curve in $\mathcal{A}^2_{red}$ and let $\mathcal{D}$ be an \emph{integral} nodal curve in $\mathcal{A}^2_{int}$. Define:
\begin{equation*}
Z(\mathcal{C})=\sum_{n\geq 0} q^n [\mathcal{C}^{[n]}]
\end{equation*}
the zeta function of the weight polynomials of the Hilbert schemes of points of $\mathcal{C}$, and $Z(\mathcal{D})$ analogously. Recall that if we have a disjoint union of varieties $X \sqcup Y$, then $Z(X \sqcup Y)=Z(X)\cdot Z(Y)$. Define:
\begin{equation*}
Z(N)=\sum_{n\geq 0} q^n [\text{Hilb}_{pt}^n (\text{node})]
\end{equation*}
to be the zeta function of the weight polynomials of the punctual Hilbert schemes of points of the curve $\{xy=0\} \subset \mathbb{C}^2$ supported at the origin. Since both $\mathcal{C}$ and $\mathcal{D}$ can be decomposed as the union of their smooth part (here denoted with a subscript \emph{sm}) and their nodes, and they both have $2g-2$ nodes, we get:
\begin{equation*}
\frac{Z(\mathcal{C})}{Z(\mathcal{D})}=\frac{Z(\mathcal{C}_{sm})\cdot Z(N)^{2g-2}}{Z(\mathcal{D}_{sm}) \cdot Z(N)^{2g-2}}=\frac{Z(\mathcal{C}_{sm})}{Z(\mathcal{D}_{sm})}.
\end{equation*}
If we denote by $\mathcal{C}^\nu$ and $\mathcal{D}^\nu$ the normalization of the two curves, then we get:
\begin{equation*}
Z(\mathcal{C}^\nu)=Z(\mathcal{C}_{sm})\cdot Z(pt)^{4g-4}
\end{equation*}
and analogously for $\mathcal{D}^\nu$. Putting all together and using the fact that $\mathcal{C}^\nu$ is the disjoint union of two copies of $C$ we get:
\begin{equation*}
\frac{Z(\mathcal{C})}{Z(\mathcal{D})}=\frac{Z(\mathcal{C}_{sm})}{Z(\mathcal{D}_{sm})}=\frac{Z(C)^2}{Z(\mathcal{D}^\nu)}.
\end{equation*}
In order to compute the (weight polynomial of the) stalk of $IC\left (\bigwedge^i\mathbb{R}^1 \right)$ at the generic nodal curve in $\mathcal{A}^2_{red}$ we can use the content of section \ref{cks}. We can also avoid the direct computation of the full weight polynomial by using the following idea.
Whenever we write the CKS complexes for curves whose irreducible components have positive genus, we will have $H^1$ of the normalization of the curve appearing only in degree $0$ and all the elements of the external powers of the degree $0$ term in which appear some classes of the $H^1$ of the normalization lie in the kernel of the first map. In our case this implies:
\begin{equation*}
\left( \sum_{i \geq 0} \left [IC(\bigwedge^i\mathbb{R}^1 )_{|\mathcal{C}}[-i] \right] q^i\right )= U(\Sigma) \left( \sum_{i \geq 0} \left [ \bigwedge^i H^1 (C)[-i] \right] q^i\right )^2\qquad \text{mod }q^{2g-2}
\end{equation*}
and
\begin{equation*}
\left( \sum_{i \geq 0} \left [IC(\bigwedge^i\mathbb{R}^1 )_{|\mathcal{D}}[-i] \right] q^i\right )= U(\overline\Sigma) \left( \sum_{i \geq 0} \left [ \bigwedge^i H^1 (\mathcal{D}^\nu)[-i] \right] q^i\right )\qquad \text{mod }q^{2g-2}
\end{equation*}
where $U(\Sigma)$ and $U(\overline\Sigma)$ are defined and examined in \ref{ratcurve}.
Putting all together we have, mod $q^{2g-2}$:
\begin{align*}
Z(\mathcal{D}^\nu)&\left( \sum_{i \geq 0} \left [IC(\bigwedge^i\mathbb{R}^1 )_{|\mathcal{C}}[-i] \right] q^i\right )=U(\Sigma)Z(C)^2Z(\mathcal{D}^\nu)(1-q\mathbb{Q})^2(1-q \L)^2=\\
&=U(\overline\Sigma)Z(C)^2Z(\mathcal{D}^\nu)(1-q\mathbb{Q})(1-q \L)=Z(C)^2 \left( \sum_{i \geq 0} \left [IC(\bigwedge^i\mathbb{R}^1 )_{|\mathcal{D}}[-i] \right] q^i\right )=\\
&=Z(C)^2 Z(\mathcal{D})(1-q\mathbb{Q})(1-q \L)=Z(\mathcal{C})Z(\mathcal{D}^\nu)(1-q\mathbb{Q})(1-q \L)
\end{align*}
and therefore:
\begin{equation*}
\left( \sum_{i \geq 0} \left [IC(\bigwedge^i\mathbb{R}^1 )_{|\mathcal{C}}[-i] \right] q^i\right )= Z(\mathcal{C})(1-q\mathbb{Q})(1-q \L) \qquad \text{mod } q^{2g-2}.
\end{equation*}
To summarize, we wanted to prove that weight polynomials of the stalks at $\mathcal{C}$ of
\begin{equation*}
\bigoplus_{2-2g \leq d <0} q^{d+2g-2} \mathbb{R} (\chi_\varepsilon^{2,d})_* \mathbb{Q}
\end{equation*}
and
\begin{equation*}
\left (\frac{\bigoplus_{i=0}^{8g-6} \leftidx{^p}{\mathbb{R}}{^i} h^{2,1}_* \mathbb{Q}[-i] q^i}{(1-q \mathbb{Q})(1-q \L)} \right )_{|\deg(q) \leq 2g-3}
\end{equation*}
are the same. In order to do that we related both polynomials to the weight polynomials of the stalks at $\mathcal{D}$ that we know coincide because $\mathcal{D}$ lies in $\mathcal{A}^2_{int}$ where we know the equality holds.
Let us now prove that $P(F^{sh}(q),t)$ and $G$ coincide for high enough degree. Recall that for odd $d \geq 6g-5$ the map $\mathcal{M}_\varepsilon^{2,d} \rightarrow \mathcal{M}^{2,d}$ is a $\mathbb{C} \P^{d+1-2g}$-bundle. This implies:
\begin{equation*}
\mathbb{R} (\chi_\varepsilon^{2,d})_* \mathbb{Q} = \mathbb{R} h^{2,1}_* \mathbb{Q} \otimes H^*(\mathbb{C} \P^{d+1-2g})
\end{equation*}
here $H^*(\mathbb{C} \P^{d+1-2g})$ denotes the constant local system $\mathbb{Q} \oplus \mathbb{Q}[-2] \oplus \dots \oplus \mathbb{Q}[-2(d+1-2g)]$.
Let us compare the coefficients of the left hand side
\begin{equation*}
\bigoplus_{n \geq 1-g} q^{2n+1+2g-2} \mathbb{R} (\chi_\varepsilon^{2,2n+1})_* IC_{\mathcal{M}_\varepsilon^{2,2n+1}}
\end{equation*}
and of the right hand side
\begin{equation*}
\text{odd}_q \left (\frac{\bigoplus_{i=0}^{8g-6} \leftidx{^p}{\mathbb{R}}{^i} h^{2,1}_* \mathbb{Q}[-i] q^i}{(1-q \mathbb{Q})(1-q \L)} \right ).
\end{equation*}
for the powers $q^{d+2g-2}$ with odd $d \geq 6g-5$.
The coefficient of $q^{d+2g-2}$ in the first expression is, as we said:
\begin{align*}
\mathbb{R} h^{2,1}_* \mathbb{Q} \otimes H^*(\mathbb{C} \P^{d+1-2g}) &= \left ( \bigoplus_{i=0}^{8g-6} IC\left (\bigwedge^i\mathbb{R}^1 \right)[-i] \oplus \mathscr{L} \right )\otimes \\
&\otimes \left ( \mathbb{Q} \oplus \mathbb{Q}[-2] \oplus \dots \oplus \mathbb{Q}[-2(d+1-2g)]\right ).
\end{align*}
For the coefficient of $q^{d+2g-2}$ in the second expression we have instead:
\begin{align*}
\sum_{i=0}^{8g-6}& IC\left (\bigwedge^i\mathbb{R}^1 \right)[-i] \otimes \left ( \mathbb{Q} \oplus \dots \mathbb{Q}[-2(d+2g-2-i)] \right)=\\
&=\sum_{i=0}^{4g-3} IC\left (\bigwedge^i\mathbb{R}^1 \right)[-i] \otimes \left ( \mathbb{Q} \oplus \dots \mathbb{Q}[-2(d+1-2g)] \right)+\\
&+\sum_{i=0}^{4g-4} IC\left (\bigwedge^i\mathbb{R}^1 \right)[-i] \otimes \left ( \mathbb{Q}[-2(d+2-2g)] \oplus \dots \mathbb{Q}[-2(d+2g-2-i)] \right)+\\
&+\sum_{j=4g-2}^{8g-6} IC\left (\bigwedge^j\mathbb{R}^1 \right)[-j] \otimes \left ( \mathbb{Q} \oplus \dots \mathbb{Q}[-2(d+2g-2-j)] \right).
\end{align*}
Using the fact that (see \cite[equation 1.4.7]{de2010topology}):
\begin{equation*}
IC\left (\bigwedge^j\mathbb{R}^1 \right) \cong IC\left (\bigwedge^{8g-6-j}\mathbb{R}^1 \right)
\end{equation*}
and the substitution $j=8g-6-i$, we can compute:
\begin{align*}
&\sum_{i=0}^{4g-4} IC\left (\bigwedge^i\mathbb{R}^1 \right)[-i] \otimes \left ( \mathbb{Q}[-2(d+2-2g)] \oplus \dots \mathbb{Q}[-2(d+2g-2-i)] \right)=\\
&=\sum_{j=4g-2}^{8g-6} IC\left (\bigwedge^j\mathbb{R}^1 \right)[6+j-8g] \otimes \left ( \mathbb{Q}[-2(d+2-2g)] \oplus \dots \mathbb{Q}[-2(d-6g+4+j)] \right)=\\
&=\sum_{j=4g-2}^{8g-6} IC\left (\bigwedge^j\mathbb{R}^1 \right)[-j] \otimes \left ( \mathbb{Q}[-2(d+2g-1-j)] \oplus \dots \mathbb{Q}[-2(d+1-2g)] \right).
\end{align*}
Putting all together we get:
\begin{align*}
\sum_{i=0}^{8g-6}& IC\left (\bigwedge^i\mathbb{R}^1 \right)[-i] \otimes \left ( \mathbb{Q} \oplus \dots \mathbb{Q}[-2(d+2g-2-i)] \right)=\\
&=\sum_{i=0}^{4g-3} IC\left (\bigwedge^i\mathbb{R}^1 \right)[-i] \otimes \left ( \mathbb{Q} \oplus \dots \mathbb{Q}[-2(d+1-2g)] \right)+\\
&+\sum_{j=4g-2}^{8g-6} IC\left (\bigwedge^j\mathbb{R}^1 \right)[-j] \otimes \left ( \mathbb{Q}[-2(d+2g-1-j)] \oplus \dots \mathbb{Q}[-2(d+1-2g)] \right)+\\
&+\sum_{j=4g-2}^{8g-6} IC\left (\bigwedge^j\mathbb{R}^1 \right)[-j] \otimes \left ( \mathbb{Q} \oplus \dots \mathbb{Q}[-2(d+2g-2-j)] \right)=\\
&=\sum_{i=0}^{4g-3} IC\left (\bigwedge^i\mathbb{R}^1 \right)[-i] \otimes \left ( \mathbb{Q} \oplus \dots \mathbb{Q}[-2(d+1-2g)] \right)+\\
&+\sum_{j=4g-2}^{8g-6} IC\left (\bigwedge^j\mathbb{R}^1 \right)[-j] \otimes \left ( \mathbb{Q} \oplus \dots \mathbb{Q}[-2(d+1-2g)] \right)=\\
&=\sum_{i=0}^{8g-6} IC\left (\bigwedge^i\mathbb{R}^1 \right)[-i] \otimes \left ( \mathbb{Q} \oplus \dots \mathbb{Q}[-2(d+1-2g)] \right).
\end{align*}
Note that this expression only differs from the first by the presence of the factor:
\begin{equation*}
\mathscr{L} \otimes H^*(\mathbb{C} \P^{d+1-2g})
\end{equation*}
which, however, does not have any global cohomology and therefore this proves the equality of the coefficients of $P(F^{sh}(q),t)$ and $G$ for the powers $q^{d+2g-2}$ with odd $d\geq 6g-5$.
Recall also that $F^{vir}-G$ satisfies:
\begin{equation*}
F^{vir}(q,t)-G(q,t)=(qt)^{8g-8}\left (F^{vir}(q^{-1}t^{-2},t)-G(q^{-1}t^{-2},t) \right).
\end{equation*}
In particular, if $F^{vir}$ and $G$ agree for $\deg q$ odd in the range $ 1, \dots, 2g-3$, then they will agree for $\deg q$ odd in the range $6g-5, \dots, 8g-9$. Since $F^{vir}$ and $P(F^{sh}(q),t)$ agree in the previous ranges we can conclude that $P(F^{sh}(q),t)$ and $G$ agree for $\deg q \leq 2g-3$ and $\deg q \geq 6g-5$ as we wanted.
For the last statement about non-negativity recall that $P(F^{sh}(q),t)$ is the global cohomology of:
\begin{equation*}
\bigoplus_{n \geq 1-g} q^{2n+1+2g-2} \mathbb{R} (\chi_\varepsilon^{2,2n+1})_* IC_{\mathcal{M}_\varepsilon^{2,2n+1}}
\end{equation*}
and $G$ is the global cohomology of $G^{sh}$.
Since these two expressions coincide when restricted to $\mathcal{A}^2_{int}$ and the second one only has summands that are supported on the full $\mathcal{A}^2$, we can conclude that $P(F^{sh}(q),t)-G$ is the global cohomology of the summands of $F^{sh}(q)$ that have proper support in $\mathcal{A}^2$ and do not come from $\mathscr{L}$ (if there are any).
\end{proof}
\end{theorem}
\chapter{Motivic invariants and wall-crossing}
\section{Wall crossing for rank $2$}
In this section we examine what happens when the stability parameter for rank 2 Bradlow-Higgs triples is modified. First we examine the geometry of the loci that change when crossing a critical value, the so called \emph{flip loci}. Second, we compute their class in the Grothendieck ring of varieties.
\subsection{The flip loci}
First of all note that, in the rank 2 case, the possible critical values for $\sigma$ are the integers greater than or equal to 1, with the same parity as the degree $d$.
As we discussed in section \ref{Bradlow}, the moduli spaces will be modified by adding and subtracting objects that are defined using extensions of Higgs bundles. Fix a critical value $\bar\sigma$. Let us start with a definition.
\begin{defn}[Flip loci]
Let $\bar\sigma$ be a critical value. We denote by $\mathcal{W}_{\bar\sigma}^{d,+}$ the locally closed subvariety of $\mathcal{M}_{\bar\sigma_-}^{2,d}$ consisting of those triples that are $\bar\sigma_-$-stable but not $\bar\sigma_+$-stable.
Analogously we denote by $\mathcal{W}_{\bar\sigma}^{d,-}$ the locally closed subvariety of $\mathcal{M}_{\bar\sigma_+}^{2,d}$ consisting of those triples that are $\bar\sigma_+$-stable but not $\bar\sigma_-$-stable.
\end{defn}
Before going into the details let us give some useful definitions.
\begin{defn}
Given a degree $d \geq 1$ and $\bar\sigma$ a critical value for $d$ we define:
\begin{align*}
p^d_{\bar\sigma}: S^{(d-\bar\sigma)/2}( C) &\rightarrow J^{(d-\bar\sigma)/2}( C)\\
D \mapsto &\mathcal{O}(D),
\end{align*}
and
\begin{equation*}
X^d_{\bar\sigma}:=p_{\bar\sigma}^{d*}T^*J^{(d-\bar\sigma)/2}( C)\times T^*J^{(d+\bar\sigma)/2}( C).
\end{equation*}
Finally, we denote by:
\begin{equation*}
q^d_{\bar\sigma}: X_{\bar\sigma} \times C \rightarrow X_{\bar\sigma}
\end{equation*}
the proper projection on the first factor.
\end{defn}
The variety $X^d_{\bar\sigma}$ has a cumbersome definition but is actually simple to understand. In fact it is a product of the total spaces of two vector bundles but they are both trivial, therefore $$X^d_{\bar\sigma}\cong S^{(d-\bar\sigma)/2}( C) \times J^{(d+\bar\sigma)/2}( C) \times H^0(K) \times H^0(K).$$
It is also clear from the definition of $X^d_{\bar\sigma}$ that it is the moduli space of pairs of rank one Higgs bundles, plus a section of the first one.
First we treat the case of wall crossing occurring when the parameter is increasing, i.e. for the flip locus $\mathcal{W}_{\bar\sigma}^{d,+}$. Looking at definition \ref{sigmast}, we see that the inequality involving the section becomes stronger while the other one becomes weaker.
We are looking for triples $(E,\phi,s)$ that are $\bar\sigma_-$-stable but not $\bar\sigma_+$-stable. These two conditions together impose the existence of a subobject that is preserved by $\phi$ and is just barely $\bar\sigma_+$-destabilizing.
More precisely, we want triples $(E,\phi,s)$ such that there is a $\phi$-invariant subbundle $L \subset E$, $s \in H^0(L)$ and $\deg L=(d-\bar\sigma)/2$. In other words, we need to have an extension of Higgs bundles:
\begin{equation*}
0 \rightarrow (L,\psi_1) \rightarrow (E,\phi) \rightarrow (M, \psi_2) \rightarrow 0
\end{equation*}
such that $s \in H^0(L)$. This, in particular, implies that $L=\mathcal{O}(D)$ where $D$ is the effective divisor associated to the section $s$.
Let us first establish a result about extensions of Higgs bundles of rank 1.
\begin{lemma}
\label{exts}
Let $(L,\psi_1)$ and $(M, \psi_2)$ be two rank one Higgs bundles. Then there exists a vector space, that we denote by $\H^1((M,\psi_2),(L,\psi_1))$ that parametrizes all extensions of Higgs bundles starting with $(L, \psi_1)$ and ending with $(M,\psi_2)$. Such a vector space is the first hypercohomology of the complex:
\begin{align*}
M^* L &\rightarrow M^* L K\\
f & \mapsto f \psi_2 - \psi_1 f
\end{align*}
and therefore fits into the long exact sequence:
\begin{align}
0 \rightarrow \H^0((M,\psi_2),(L,\psi_1)) \rightarrow H^0(M^* L) \rightarrow H^0(M^* L K) \rightarrow \H^1((M,\psi_2),(L,\psi_1)) \rightarrow\nonumber\\ \rightarrow H^1(M^* L) \rightarrow H^1(M^* L K) \rightarrow \H^2((M,\psi_2),(L,\psi_1)) \rightarrow 0. \label{eq:exext}
\end{align}
The zero class in $\H^1((M,\psi_2),(L,\psi_1))$ corresponds to a split extension of Higgs bundles and if two classes differ by the multiplication by a nonzero scalar, then they define isomorphic Higgs bundles.
\begin{proof}
Choose an open cover $\{U_\alpha\}$ of $C$ on which $L$, $M$ and $K$ are all trivial. Let us denote by $\tau \in C^1(M^* L)$ the off diagonal portion of the transition matrix for our vector bundle $E$. If $l_{\alpha\beta}$ and $m_{\alpha\beta}$ are the transition functions for $L$ and $M$ respectively then the transition functions for $E$ can be written as
\begin{equation*}
\begin{pmatrix}
l_{\alpha\beta} & \tau_{\alpha\beta}\\
0 & m_{\alpha\beta}
\end{pmatrix}.
\end{equation*}
Let then $\nu \in C^0(M^* L K)$ be the off diagonal part of the Higgs field. So that the restriction of the Higgs field to the open subset $U_\alpha$ will be:
\begin{equation*}
\begin{pmatrix}
\psi_{1\alpha} & \nu_{\alpha}\\
0 & \psi_{2\alpha}
\end{pmatrix}.
\end{equation*}
After the notation is fixed, we can write the compatibility relation between $\tau$ and $\nu$:
\begin{equation*}
\begin{pmatrix}
l_{\alpha\beta} & \tau_{\alpha\beta}\\
0 & m_{\alpha\beta}
\end{pmatrix}
\begin{pmatrix}
\psi_{1\beta} & \nu_{\beta}\\
0 & \psi_{2\beta}
\end{pmatrix}=
\begin{pmatrix}
\psi_{1\alpha} & \nu_{\alpha}\\
0 & \psi_{2\alpha}
\end{pmatrix}
\begin{pmatrix}
l_{\alpha\beta} & \tau_{\alpha\beta}\\
0 & m_{\alpha\beta}
\end{pmatrix}
\end{equation*}
which, after taking into account the trivializations of the bundles, becomes $d \nu=\psi_1 \tau-\tau \psi_2$, i.e. the kernel of the following map:
\begin{align*}
C^0(M^* L K) \oplus C^1(M^* L) &\rightarrow C^1(M^* L K)\\
(\nu, \tau) &\mapsto d\nu -\tau \psi_2+\psi_1 \tau.
\end{align*}
Note that with this procedure we produce two isomorphic Higgs bundles $(E,\psi)$ and $(E',\psi')$ (that of course still have $(L,\psi_1)$ as a subobject and $(M,\psi_2)$ as a quotient) if and only if there is $\gamma \in C^0(M^* L)$ that yields an equivalence between the \v{C}ech data representing $E$ and $E'$. The matrices of such an equivalence will look like:
\begin{equation*}
\begin{pmatrix}
1 & \gamma_{\alpha}\\
0 & 1
\end{pmatrix}.
\end{equation*}
The compatibility relation with the respective transition matrices of $E$ and $E'$ therefore becomes
\begin{equation*}
\begin{pmatrix}
1 & \gamma_{\alpha}\\
0 & 1
\end{pmatrix}
\begin{pmatrix}
l_{\alpha\beta} & \tau'_{\alpha\beta}\\
0 & m_{\alpha\beta}
\end{pmatrix}
=
\begin{pmatrix}
l_{\alpha\beta} & \tau_{\alpha\beta}\\
0 & m_{\alpha\beta}
\end{pmatrix}
\begin{pmatrix}
1 & \gamma_{\beta}\\
0 & 1
\end{pmatrix}
\end{equation*}
while for the Higgs fields we get:
\begin{equation*}
\begin{pmatrix}
1 & \gamma_{\alpha}\\
0 & 1
\end{pmatrix}
\begin{pmatrix}
\psi_{1\alpha} & \nu'_{\alpha}\\
0 & \psi_{2\alpha}
\end{pmatrix}=
\begin{pmatrix}
\psi_{1\alpha} & \nu_{\alpha}\\
0 & \psi_{2\alpha}
\end{pmatrix}
\begin{pmatrix}
1 & \gamma_{\alpha}\\
0 & 1
\end{pmatrix}.
\end{equation*}
The relations can be written in a compact form as $\tau'-\tau=d \gamma$ and $\nu'-\nu=\psi_1 \gamma-\gamma \psi_2$. These relations can also be expressed as the image of the map
\begin{align*}
C^0(M^* L) &\rightarrow C^0(M^* L K) \oplus C^1(M^* L)\\
\gamma &\mapsto (\psi_1 \gamma-\gamma \psi_2, d\gamma).
\end{align*}
This clearly completes the assertion about the hypercohomology, from which also follows the statement about the long exact sequence.
The zero class corresponds to the data
\begin{equation*}
\begin{pmatrix}
l_{\alpha\beta} & 0\\
0 & m_{\alpha\beta}
\end{pmatrix}, \quad
\begin{pmatrix}
\psi_{1\alpha} & 0\\
0 & \psi_{2\alpha}
\end{pmatrix}
\end{equation*}
which clearly represent the Higgs bundle $(L,\psi_1) \oplus (M,\psi_2)$. Changing $\{\tau_{\alpha\beta}\}$ and ${\nu_{\alpha}}$ with $\{\lambda\tau_{\alpha\beta}\}$ and ${\lambda\nu_{\alpha}}$, for some $\lambda \neq 0$, corresponds to acting with the change of trivialization:
\begin{equation*}
\theta_\alpha=
\begin{pmatrix}
\lambda & 0\\
0 & 1
\end{pmatrix}
\end{equation*}
and therefore does not change $(E,\phi)$.
\end{proof}
\end{lemma}
We have the following:
\begin{prop}
Fix a degree $d$ and a critical value $\bar\sigma$. The locus $\mathcal{W}_{\bar\sigma}^{d,+}$ consists of those triples $(E,\phi,s)$ for which $(E,\phi)$ fits into a non-split extension:
\begin{equation*}
0 \rightarrow (L,\psi_1) \rightarrow (E,\phi) \rightarrow (M, \psi_2) \rightarrow 0
\end{equation*}
where $\deg L =(d-\bar\sigma)/2$, $\deg M =(d+\bar\sigma)/2$ and $s \in H^0(L)$.
Furthermore, there is a projective map $\pi_{\bar\sigma}^{d,+}: \mathcal{W}_{\bar\sigma}^{d,+} \rightarrow X^d_{\bar\sigma}$ whose fibers are projective spaces (but not of constant dimension).
\begin{proof}
Consider a triple $(E,\phi,s)$ that fits in the non-split extension:
\begin{equation*}
0 \rightarrow (L,\psi_1) \rightarrow (E,\phi) \rightarrow (M, \psi_2) \rightarrow 0
\end{equation*}
with $s \in H^0(L)$, $\deg L=(d-\bar\sigma)/2$ and $\deg M=(d+\bar\sigma)/2$. Assume there is a $\phi$-invariant subbundle $L'$. Then if $s \in H^0(L')$ we get $L=L'$ and hence $\deg L' =(d-\bar\sigma)/2<(d-\bar\sigma_-)/2$. Otherwise assume $s \notin H^0(L')$. This implies there is a nonzero map $L' \rightarrow M$ and hence $\deg L' \leq (d+\bar\sigma)/2$. The equality can never occur because otherwise the map would be an isomorphism and the sequence would be split. This implies that all such triples are $\bar\sigma_-$-stable. It is clear by construction that they are not $\bar\sigma_+$-stable.
Viceversa, every triple in $\mathcal{W}_{\bar\sigma}^{d,+}$ must be $\bar\sigma_-$-stable and $\bar\sigma_+$-unstable so it must have a $\phi$-invariant subbundle containing the section of degree $(d-\bar\sigma)/2$.
The map
\begin{align*}
\pi_{\bar\sigma}^{d,+}: \mathcal{W}_{\bar\sigma}^{d,+} &\rightarrow X^d_{\bar\sigma}\\
(E,\phi,s) & \mapsto (L,s,M,\psi_1,\psi_2)
\end{align*}
is defined by sending a triple $(E,\phi,s)$ to the canonical subbundle $L$ containing the section (which has degree $(d-\bar\sigma)/2$ and is $\phi$-invariant) and to the canonical quotient $M$, both of which inherit a Higgs field from $(E,\phi)$. The fibers of such a map are, by lemma \ref{exts}, projectivized extension spaces and their dimension can vary, as we will see.
\end{proof}
\end{prop}
In lemma \ref{exts} we gave an interpretation only for the first hypercohomology of the complex as a space of extensions of Higgs bundles. We can give an interpretation for the 0th and the 2nd hypercohomology as well. In fact we can think of $\H^0((M,\psi_2),(L,\psi_1))$ the subspace of $H^0(M^* L)$ of morphisms $f: M \rightarrow L$ that satisfy $f \psi_2=\psi_1 f$, i.e. the morphisms of Higgs bundles $(M, \psi_2) \rightarrow (L, \psi_1)$. Furthermore, by using Serre duality on the starting complex we get the dual complex:
\begin{align*}
L^* M &\rightarrow L^* M K\\
f & \mapsto f \psi_2 - \psi_1 f
\end{align*}
which is the same complex but with $L$ and $M$ interchanged. So we see that
$$\H^2((M,\psi_2),(L,\psi_1))$$
is dual to $$\H^0((L,\psi_1),(M,\psi_2)).$$
Finally from the long exact sequence we get
$$\chi((M,\psi_2),(L,\psi_1))=\chi(M^* L)-\chi(M^* L K)=2-2g.$$
Since for $ \mathcal{W}_{\bar\sigma}^{d,+}$ we always have $\deg L < \deg M$, we also have $H^0(M^* L)=0$ and also $\H^0((M,\psi_2),(L,\psi_1))=0$ because of the injection $\H^0((M,\psi_2),(L,\psi_1)) \rightarrow H^0(M^* L)$. However $\H^2((M,\psi_2),(L,\psi_1))$ might very well jump. This, as we mentioned, means that the fibers of $\pi_{\bar\sigma}^{d,+}$ can vary in dimension.
Since the final goal will be to compute the motives of our moduli spaces, let us now examine the stratification of $X^d_{\bar\sigma}$ that is induced by the dimension of the fibers of $\pi_{\bar\sigma}^{d,+}$.
\begin{defn}
\label{defstr}
Let $\bar\sigma$ be a critical value, then we define locally closed subsets:
\begin{equation*}
S_{{\bar\sigma},i}^{d,+}:=\{(L,s,\psi_1, M, \psi_2) \in X^d_{\bar\sigma} | \dim \H^2((M,\psi_2),(L,\psi_1)) \geq i\}.
\end{equation*}
\end{defn}
Note that $S_{{\bar\sigma},i+1}^{d,+} \subset S_{{\bar\sigma},i}^{d,+}$ and $S_{{\bar\sigma},0}^{d,+}=X^d_{\bar\sigma}$. Let us fix a point $(L,s,\psi_1, M, \psi_2) \in X^d_{\bar\sigma}$. First of all, since $\dim \P\H^1((M,\psi_2),(L,\psi_1))=2g-2+\dim \H^2((M,\psi_2),(L,\psi_1))-1$, we see that $S_{{\bar\sigma},i}^{d,+}$ is exactly the stratification that agrees with the dimension of the fibers of $\pi_{\bar\sigma}^{d,+}$, after shifting the indices. In other words, the fiber of $\pi_{\bar\sigma}^{d,+}$ over each point of $S_{{\bar\sigma},i}^{d,+}\setminus S_{{\bar\sigma},i+1}^{d,+}$ is exactly $\mathbb{C} \P^{2g-3+i}$.
Furthermore, $\H^2((M,\psi_2),(L,\psi_1))$ is dual to $\H^0((L,\psi_1),(M,\psi_2))$ and we can easily compute the dimension of the latter. In fact $L$ and $M$ are both line bundles and so if $\psi_1 \neq \psi_2$ there are certainly no maps $L \rightarrow M$ that commute with the Higgs fields.
On the other hand, if $\psi_1=\psi_2$, then $\H^0((L,\psi_1),(M,\psi_2))=H^0(L^* M)$. From this we can deduce that $S_{\bar\sigma,1}^{d,+}$ is contained in the locus of $X^d_{\bar\sigma}$ where the Higgs fields agree, furthermore, when $\psi_1=\psi_2$, $\dim \H^0((L,\psi_1),(M,\psi_2))= \dim H^0(L^* M)$ and this is positive if and only if $L^* M=\mathcal{O}(D')$; therefore $M=L(D')$. We will soon use these considerations to compute the motive of $\mathcal{W}_{\bar\sigma}^{d,+}$.
We can now examine the second family involved in the wall crossing, which is the one appearing for decreasing parameter. Fix again a critical value $\bar\sigma$. From the definition of stabilty \ref{sigmast} we see that we should look for those triples $(E,\phi,s)$ such that there exists a $\phi$-invariant line subbundle $M \subset E$ with $s \notin H^0(M)$ and $\deg M=(d+\bar\sigma)/2$ so that the triple will be barely $\bar\sigma_-$-unstable. The fact that $M$ does not contain the section is equivalent to the section $s$ projecting on a nonzero section $\bar{s}$ of the quotient $L=E/M$.
Here we invert the notations for the subobject and the quotient, both for consistency with the degrees and for the fact that, for us, $L$ is the line bundle that comes with a section.\\
We will need a lemma about extensions as in the previous case.
\begin{lemma}
\label{extssec}
Let $(L,\psi_1)$ and $(M, \psi_2)$ be two rank one Higgs bundles such that $\bar{s} \in H^0(L)$ is a nonzero section. Then there exists a vector space, that we denote by
$$\widetilde\H^1((L, \bar{s},\psi_1),(M,\psi_2))$$
that parametrizes all extensions
\begin{equation*}
0 \rightarrow (M, \psi_2) \rightarrow (E,\phi) \rightarrow (L,\psi_1) \rightarrow 0
\end{equation*}
together with a section $s \in H^0(E)$ that projects onto $\bar{s}$. This vector space is the first hypercohomology of the complex:
\begin{align*}
L^* M &\rightarrow L^* M K \oplus M\\
f & \mapsto (f \psi_1 - \psi_2 f,f(\bar{s})).
\end{align*}
The zero class corresponds to the split extension of Higgs bundles together with a trivial lift of the section, meaning that the section is the given one in the quotient. Again, if two classes in $\widetilde\H^1((L, \bar{s},\psi_1),(M,\psi_2))$ differ by the multiplication by a nonzero scalar, then they define isomorphic triples.
\begin{proof}
The proof is similar to the \v{C}ech cohomology proof that we outlined in lemma \ref{exts}, however a bit of care is needed because now subobject and quotient are inverted.
Let us denote by $\tau \in C^1(L^* M)$ the off diagonal portion of the transition matrix for our vector bundle $E$. If $l_{\alpha\beta}$ and $m_{\alpha\beta}$ are the transition functions for $L$ and $M$ respectively then the transition functions for $E$ can be written as
\begin{equation*}
\begin{pmatrix}
m_{\alpha\beta} & \tau_{\alpha\beta}\\
0 & l_{\alpha\beta}
\end{pmatrix}.
\end{equation*}
Let then $\nu \in C^0(L^* M K)$ be the off diagonal part of the Higgs field. So that the restriction of the Higgs field to the open subset $U_\alpha$ will be:
\begin{equation*}
\begin{pmatrix}
\psi_{2\alpha} & \nu_{\alpha}\\
0 & \psi_{1\alpha}
\end{pmatrix}.
\end{equation*}
Let also $\theta \in C^0(M)$ be the lift of $\bar{s}$ restricted to $M$ so that on the open set $U_\alpha$ the section is written:
\begin{equation*}
\begin{pmatrix}
\theta_{\alpha}\\
\bar{s}_{\alpha}
\end{pmatrix}.
\end{equation*}
We can write the compatibility relation between $\tau$ and $\nu$ that becomes $d \nu=\psi_2 \tau-\tau \psi_1$. Also, the compatibility between $\theta$ and $\tau$ becomes $d \theta = \tau \bar{s}$. The two relations together are the kernel of the following map:
\begin{align*}
C^0(L^* M K) \oplus C^0(M) \oplus C^1(L^* M) & \rightarrow C^1(L^* M K) \oplus C^1(M)\\
(\nu, \theta, \tau) &\mapsto (d\nu -\tau \psi_1+\psi_2 \tau, d\theta -\tau \bar{s}).
\end{align*}
We produce two isomorphic extensions if and only if there is $\gamma \in C^0(L^* M)$ that produces an equivalence between the \v{C}ech data representing them. The matrices of such an equivalence will look like:
\begin{equation*}
\begin{pmatrix}
1 & \gamma_{\alpha}\\
0 & 1
\end{pmatrix}.
\end{equation*}
The compatibility relation with the respective transition matrices of $E$ and $E'$ therefore becomes $\tau'-\tau=d \gamma$ while for the Higgs fields is $\nu'-\nu=\psi_2 \gamma-\gamma \psi_1$ and for the sections is $\theta'-\theta=\gamma \bar{s}$. These relations can also be expressed as the image of the map
\begin{align*}
C^0(L^* M) &\rightarrow C^0(L^* M K) \oplus C^0(M) \oplus C^1(L^* M)\\
\gamma &\mapsto (\psi_2 \gamma-\gamma \psi_1, d\gamma, \gamma \bar{s}).
\end{align*}
This completes the assertion about the hypercohomology. The last statement is obtained in the same way as in the proof of lemma \ref{exts}.
\end{proof}
\end{lemma}
We have the following.
\begin{prop}
Fix a degree $d$ and a critical value $\bar\sigma$. The locus $\mathcal{W}_{\bar\sigma}^{d,-}$ consists of those triples $(E,\phi,s)$ for which $(E,\phi)$ fits into an extension:
\begin{equation*}
0 \rightarrow (M,\psi_2) \rightarrow (E,\phi) \rightarrow (L, \psi_1) \rightarrow 0
\end{equation*}
where $\deg L =(d-\bar\sigma)/2$, $\deg M =(d+\bar\sigma)/2$ and $s \in H^0(E)$ projects to a nonzero $\bar{s} \in H^0(L)$. The condition on $(E,\phi,s)$ is that either the extension is nonsplit or it is split but $s \neq \bar{s}$.
Furthermore, there is a projective map $\pi_{\bar\sigma}^{d,-}: \mathcal{W}_{\bar\sigma}^{d,-} \rightarrow X^d_{\bar\sigma}$ whose fibers are projective spaces.
\begin{proof}
Consider a triple $(E,\phi,s)$ that fits in the extension:
\begin{equation*}
0 \rightarrow (M,\psi_2) \rightarrow (E,\phi) \rightarrow (L, \psi_1) \rightarrow 0
\end{equation*}
with $s \notin H^0(M)$, $\deg L=(d-\bar\sigma)/2$ and $\deg M=(d+\bar\sigma)/2$. Call $\bar{s}$ the projection of $s$ on $L$ and assume that either the extension is nonsplit or that it is split and $\bar{s} \neq s$. Take a $\phi$-invariant line subbundle $L'$ of $E$. Then, if $s \in H^0(L')$ then certainly $L'$ is not contained in $M$ so there is a nonzero map $L'\rightarrow L$ and therefore $\deg L' \leq \deg L = (d-\bar\sigma)/2$.
Observe that if $\deg L' = \deg L$ then $L=L'$ and the extension has to be split. In this case however the section is contained entirely in $L$ which is a contradiction. Therefore it has to be $\deg L' < \deg L$ and hence $\deg L' < (d-\bar\sigma_+)/2$. In any case $\deg L'$ is either contained in $M$, and then $\deg L' \leq \deg M < (d+\bar\sigma)/2$ or it has a nonzero map $L' \rightarrow L$ so again $\deg L' < (d+\bar\sigma_+)/2$. This proves that $(E,\phi,s)$ is $\bar\sigma_+$-stable. Clearly $(E,\phi,s)$ is also $\bar\sigma_-$-unstable because of $M$.
Vice versa, every triple in $\mathcal{W}_{\bar\sigma}^{d,-}$ has to be $\bar\sigma_-$-unstable and $\bar\sigma_+$-stable so it has to be of the above form.
The map
\begin{align*}
\pi_{\bar\sigma}^{d,-}: \mathcal{W}_{\bar\sigma}^{d,-} &\rightarrow X^d_{\bar\sigma}\\
(E,\phi,s) & \mapsto (L,\bar{s},M,\psi_1,\psi_2)
\end{align*}
is defined by sending a triple $(E,\phi,s)$ to the canonical $\bar\sigma_-$-destabilizing subbundle $M$ and to the pair $(L,\bar{s})$ consisting of the quotient $E/M$ and the projection of $s$ onto the quotient. Both $L$ and $M$ are endowed with Higgs fields because $M$ is $\phi$-invariant.
The fibers of $\pi_{\bar\sigma}^{d,-}$ are the projectivized hypercohomology spaces of lemma \ref{extssec}.
\end{proof}
\end{prop}
In order to compute $\dim \widetilde\H^1((L, \bar{s},\psi_1),(M,\psi_2))$ we can use the following exact sequence of complexes:
\begin{center}
\begin{tikzcd}
& 0\arrow{r} & 0 \arrow{r} \arrow{d} & L^*M \arrow{r}\arrow{d} & L^*M \arrow{r} \arrow{d} & 0\\
& 0\arrow{r} & KL^*M \arrow{r} & KL^*M \oplus M \arrow{r} & M \arrow{r}& 0.
\end{tikzcd}
\end{center}
The 0th hypercohomology of the first complex of course vanishes, furthermore the same is true for the 0th hypercohomology of the last complex because the map $L^*M \rightarrow M$ is injective on the global sections. This in particular implies that the 0th hypercohomology of the middle complex also vanishes because it is between two zeroes in the long exact sequence of hypercohomologies associated to the exact sequence of complexes we wrote.
The rest of the sequence is:
\begin{align*}
0 \rightarrow& H^0(KL^*M) \rightarrow \widetilde{\H}^1(M,\bar{s},\psi_2, L, \psi_1) \rightarrow H^0(M\mathcal{O}_D) \rightarrow \\
\rightarrow& H^1(KL^*M) \rightarrow \widetilde{\H}^2(M,\bar{s},\psi_2, L, \psi_1) \rightarrow 0.
\end{align*}
where $D$ is the divisor associated to $\bar{s}$. Since $H^1(KL^*M) \cong H^0(M^*L)^*$ and, in our situation, $\deg M > \deg L$, we see that $H^1(KL^*M)=0$. Therefore $\widetilde{\H}^2(M,\bar{s},\psi_2, L, \psi_1)=0$ as well and we can compute
$$\dim \widetilde{\H}^1(M,\bar{s},\psi_2, L, \psi_1)=\dim H^0(KL^*M)+\dim H^0(M\mathcal{O}_D)=g-1+(d+\bar\sigma)/2.$$
For the map $\pi_{\bar\sigma}^{d,-}: \mathcal{W}_{\bar\sigma}^{d,-} \rightarrow X^d_{\bar\sigma}$ we find that the fibers are actually of constant dimension so that $\mathcal{W}_{\bar\sigma}^{d,-}$ is a projective bundle over $\rightarrow X^d_{\bar\sigma}$.
To sum up, both families have a map to $X^d_{\bar\sigma}$. However, in the context of the first family, the divisor yields a line bundle that is the subobject $L$ of the extension we produce, while the Jacobian part yields the quotient $M$. For the second family the roles are exchanged, meaning that the quotient comes with a section, while the subobject is parametrized by the Jacobian part.
Another difference is that, while the fibers of the map $\pi_{\bar\sigma}^{d,+} : \mathcal{W}_{\bar\sigma}^{d,+} \rightarrow X^d_{\bar\sigma}$ can vary in dimension, the fibers of $\pi_{\bar\sigma}^- : \mathcal{W}_{\bar\sigma}^{d,-} \rightarrow X^d_{\bar\sigma}$ are of constant dimension.
\subsection{Motives of the flip loci}
We can start by handling the computation of $[\mathcal{W}_{\bar\sigma}^{d,+}]$.
\begin{prop}
\label{wplus}
We have the following motivic equality:
\begin{equation*}
[\mathcal{W}_{\bar\sigma}^{d,+}]=\L^{2g} \cdot [\mathbb{C}\P^{2g-3}]\cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot [J( C)]+\L^{3g-2} \cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot [S^{\bar\sigma}( C)].
\end{equation*}
\begin{proof}
We have a map $\pi_{\bar\sigma}^{d,+}: \mathcal{W}^{d,+}_{\bar\sigma} \rightarrow X^d_{\bar\sigma}$ and we can stratify $X^d_{\bar\sigma}$ according to the dimension of the fibers of this map. The strata $S_{{\bar\sigma},i}^{d,+}$ appear in definition \ref{defstr}.
We get a preliminary formula
\begin{equation*}
[\mathcal{W}_{\bar\sigma}^{d,+}]=\sum_{i=0}^{\bar\sigma}[S_{{\bar\sigma},i}^{d,+}\setminus S_{{\bar\sigma},i+1}^{d,+}]\cdot [\mathbb{C}\P^{2g-3+i}].
\end{equation*}
For $i>0$, we can identify $S_{{\bar\sigma},i}^{d,+}$ with the locus of points $(L,s,\psi_1,M, \psi_2)$ of $X^d_{\bar\sigma}$ such that $\psi_1=\psi_2$ and $M=L(D')$ with $\dim H^0(\mathcal{O}(D')) \geq i$. For $i=0$ the same holds but there is an extra disjoint part coming from the open locus of $X^d_{\bar\sigma}$ where $\psi_1 \neq \psi_2$.
Let us consider the maps:
\begin{align*}
S^{(d-{\bar\sigma})/2}( C) \times S^{\bar\sigma}( C)\times H^0(K) &\rightarrow S^{(d-{\bar\sigma})/2}( C) \times J^{\bar\sigma}( C)\times H^0(K)\\
(L,s,D',\psi) &\mapsto (L,s, L(D'),\psi)\\
S^{(d-{\bar\sigma})/2}( C) \times J^{\bar\sigma}( C)\times H^0(K) & \rightarrow X^d_{\bar\sigma}\\
(L,s,A,\psi) & \mapsto (L,s, ,\psi, L \otimes A, \psi).
\end{align*}
Clearly the composition of the two is the map $(L,s,D',\psi) \mapsto (L,s,\psi, L(D'),\psi)$ and the second one is injective.
Since we have an alternative description of $S_{{\bar\sigma},i}^{d,+}$ for $i >0$ we can compute:
\begin{equation*}
[S_{{\bar\sigma},i}^{d,+}\setminus S_{{\bar\sigma},i+1}^{d,+}]=[S^{(d-{\bar\sigma})/2}( C)]\cdot[V^{\bar\sigma}_{i}]\cdot[H^0(K)].
\end{equation*}
where $V^{k}_{i}=\{A \in J^{k}( C) | \dim H^0(A) = i\}$.
For $i=0$, there is an extra contribution to $S_{{\bar\sigma},0}^{d,+}\setminus S_{{\bar\sigma},1}^{d,+}$ coming from the locus of $X_{\bar\sigma}^d$ where $\psi_1 \neq \psi_2$. Therefore:
\begin{align*}
[S_{{\bar\sigma},0}^{d,+}\setminus S_{{\bar\sigma},1}^{d,+}]=[S^{(d-{\bar\sigma})/2}( C)]\cdot[V^{\bar\sigma}_{0}]\cdot[H^0(K)]+\\
[S^{(d-{\bar\sigma})/2}( C)]\cdot[J^{(d+{\bar\sigma})/2}( C)]\cdot \left([H^0(K)]^2-[H^0(K)]\right).
\end{align*}
Putting all together we get:
\begin{align*}
[\mathcal{W}_{\bar\sigma}^{d,+}]&=[\mathbb{C} \P^{2g-3}]\cdot[S^{(d-{\bar\sigma})/2}( C)]\cdot[J^{(d+{\bar\sigma})/2}( C)]\cdot \left([H^0(K)]^2-[H^0(K)]\right) +\\
&+[S^{(d-{\bar\sigma})/2}( C)]\cdot[H^0(K)] \cdot \sum_{i=0}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot [\mathbb{C} \P^{2g-3+i}].
\end{align*}
First, we try to evaluate the second part of the sum. With a simple algebraic trick we get:
\begin{align*}
\sum_{i=0}^{\bar\sigma} &[V^{\bar\sigma}_{i}]\cdot [\mathbb{C} \P^{2g-3+i}]=[V^{\bar\sigma}_{0}]\cdot [\mathbb{C} \P^{2g-3}]+\\
&+\sum_{i=1}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot ([\mathbb{C} \P^{2g-3+i}]-[\mathbb{C}\P^{i-1}]) + \sum_{i=1}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot [\mathbb{C}\P^{i-1}].
\end{align*}
Now, clearly
$$[\mathbb{C} \P^{2g-3+i}]-[\mathbb{C}\P^{i-1}]=\L^i (1+ \dots+\L^{2g-3})=\L^i\cdot [\mathbb{C}\P^{2g-3}].$$
Recall that the canonical map:
\begin{align*}
S^{\bar\sigma}( C) &\rightarrow J^{\bar\sigma} ( C)\\
D &\mapsto \mathcal{O}(D)
\end{align*}
has $V^{\bar\sigma}_{1}$ as image and the fiber over $L \in J^{\bar\sigma} ( C)$ is $\P H^0(L)$. This implies that:
\begin{equation*}
\sum_{i=1}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot [\mathbb{C}\P^{i-1}] = [S^{\bar\sigma} ( C)].
\end{equation*}
Putting everything in the formula we get:
\begin{align*}
\sum_{i=0}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot [\mathbb{C} \P^{2g-3+i}]=[S^{\bar\sigma} ( C)]+\sum_{i=0}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot [\mathbb{C} \P^{2g-3}] \cdot \L^i.
\end{align*}
With tricks similar to the ones we already used we find:
\begin{align*}
\sum_{i=0}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot \L^i&=\sum_{i=1}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot (\L^i-1) + \sum_{i=0}^{\bar\sigma} [V^{\bar\sigma}_{i}]=\\
&=\sum_{i=1}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot (\L^i-1)+J^{\bar\sigma} ( C)
\end{align*}
and
\begin{align*}
\sum_{i=1}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot (\L^i-1)=(\L-1)\cdot \sum_{i=1}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot [\mathbb{C}\P^{i-1}] = (\L-1) \cdot [S^{\bar\sigma} ( C)].
\end{align*}
Recalling that $[H^0(K)]=\L^{g}$ and that the Jacobian varieties of any degree of the curve are all isomorphic we can find a final formula for:
\begin{align*}
&[\mathcal{W}_{\bar\sigma}^{d,+}]=[\mathbb{C} \P^{2g-3}]\cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot[J^{(d+{\bar\sigma})/2}( C)]\cdot \left([H^0(K)]^2-[H^0(K)]\right) +\\
&+[S^{(d-{\bar\sigma})/2}( C)]\cdot[H^0(K)] \cdot \sum_{i=0}^{\bar\sigma} [V^{\bar\sigma}_{i}]\cdot [\mathbb{C} \P^{2g-3+i}]=\\
&=\L^g\cdot(\L^g-1) \cdot [\mathbb{C} \P^{2g-3}]\cdot[S^{(d-{\bar\sigma})/2}( C)]\cdot[J^d( C)]+\\
&+\L^g\cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot \left( [S^{\bar\sigma} ( C)]+(\L-1) \cdot [\mathbb{C}\P^{2g-3}] \cdot [S^{\bar\sigma} ( C)] + [\mathbb{C}\P^{2g-3}] \cdot [J^{\bar\sigma} ( C)]\right)=\\
&=\L^{2g} \cdot [\mathbb{C}\P^{2g-3}]\cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot [J( C)]+\L^{3g-2} \cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot [S^{\bar\sigma}( C)].
\end{align*}
\end{proof}
\end{prop}
As for $\mathcal{W}_{\bar\sigma}^{d,-}$, we can easily compute the motive because the fibers of the map $\pi_\sigma^{d,-}$ are projective spaces of constant dimension. Therefore we get:
\begin{prop}
We have the motivic equality:
\begin{equation*}
[\mathcal{W}_{\bar\sigma}^{d,-}]=\L^{2g}\cdot[S^{(d-{\bar\sigma})/2}( C)]\cdot[J^{(d+{\bar\sigma})/2}( C)]\cdot [\mathbb{C}\P^{(d+\bar\sigma)/2+g-2}].
\end{equation*}
\end{prop}
\subsection{Interaction between the flip loci and the attracting sets}
One of the nice features of the moduli spaces of Bradlow-Higgs triples is that they combine the wall-crossing with the presence of the $\mathbb{C}^*$-action. In this section we study how the flip loci intersect the various attracting sets.
\begin{defn}
\label{weirdl}
We will denote by $S\mathcal{W}^{d,+}_{\bar\sigma}$ the subvariety of $\mathcal{W}^{d,+}_{\bar\sigma}$ containing those triples in $\mathcal{M}^{2,d}_{\bar\sigma_-}$ whose underlying Bradlow pair is $\bar\sigma_-$-stable.
$NS\mathcal{W}^{d,+}_{\bar\sigma}$ will instead denote the subvariety of $S\mathcal{W}^{d,+}_{\bar\sigma}$ containing the triples whose underlying Bradlow pair is not $\bar\sigma_-$-stable.
Analogously for $S\mathcal{W}^{d,-}_{\bar\sigma}$ and $NS\mathcal{W}^{d,-}_{\bar\sigma}$ but in relation to $\bar\sigma_+$-stability.
We define $B^{d,+}_{\bar\sigma}$ to be the locally closed subvariety containing those triples in $F_{(2),\bar\sigma_-}^{(d),1+}$ such that the underlying Bradlow pair is contained in $\P W^{d,+}_{\bar\sigma}$.
We also define $B^{d,-}_{\bar\sigma}$ as the locally closed subvariety containing those triples in $F_{(2),\bar\sigma_+}^{(d),1+}$ such that the underlying Bradlow pair is contained in $\P W^{d,-}_{\bar\sigma}$.
\end{defn}
Let us first remark that when $d <0 $ there are no $\sigma$-stable Bradlow pairs, for any $\sigma$. Therefore many of the subvarieties we examine in the following are empty for $d <0$.
We can prove two propositions summarizing the relations and properties of the loci defined above.
\begin{prop}
\label{weirdexpl1}
Let $\bar\sigma$ be a critical value. Then:
\begin{itemize}
\item[(i)] $\mathcal{W}^{d,+}_{\bar\sigma}=S\mathcal{W}^{d,+}_{\bar\sigma} \sqcup NS\mathcal{W}^{d,+}_{\bar\sigma}$.
\item[(ii)] $S\mathcal{W}^{d,+}_{\bar\sigma}=F_{(2),\bar\sigma_-}^{(d),1+} \cap \mathcal{W}^{d,+}_{\bar\sigma}$ and in particular $S\mathcal{W}^{d,+}_{\bar\sigma} \subset F_{(2),\bar\sigma_-}^{(d),1+}$.
\item[(iii)] The canonical map $\mathcal{W}^{d,+}_{\bar\sigma} \rightarrow X^d_{\bar\sigma}$ restricts to a map $NS\mathcal{W}^{d,+}_{\bar\sigma} \rightarrow X^d_{\bar\sigma}$ whose fiber over $(L,s,M,\psi_1,\psi_2)$ is $\P H^0(M^* L K)$.
\item[(iv)] $\mathcal{W}^{d,-}_{\bar\sigma}=S\mathcal{W}^{d,-}_{\bar\sigma} \sqcup NS\mathcal{W}^{d,-}_{\bar\sigma}$.
\item[(v)] $S\mathcal{W}^{d,-}_{\bar\sigma}=F_{(2),\bar\sigma_+}^{(d),1+} \cap \mathcal{W}^{d,-}_{\bar\sigma}$ and in particular $S\mathcal{W}^{d-}_{\bar\sigma} \subset F_{(2),\bar\sigma_+}^{(d),1+}$.
\item[(vi)] The canonical map $\mathcal{W}^{d,-}_{\bar\sigma} \rightarrow X^d_{\bar\sigma}$ restricts to a map $NS\mathcal{W}^{d,-}_{\bar\sigma} \rightarrow X^d_{\bar\sigma}$ whose fiber over $(L,\bar{s},M,\psi_1,\psi_2)$ is $\P H^0(L^* M K)$.
\item[(vii)] $NS\mathcal{W}^{d,+}_{\bar\sigma} \subset F_{(1,1),\bar\sigma_-}^{((d-\bar\sigma)/2,(d+\bar\sigma)/2),1+}$.
\item[(viii)] $NS\mathcal{W}^{d,-}_{\bar\sigma} \subset F_{(1,1),\bar\sigma_+}^{((d+\bar\sigma)/2,(d-\bar\sigma)/2),2+}$.
\end{itemize}
\begin{proof}
(i) follows from the fact that if $(E,\phi,s) \in \mathcal{W}^{d,+}_{\bar\sigma}$ then it is $\bar\sigma_-$-stable. Obviously, this can happen either because the underlying Bradlow pair is already $\bar\sigma_-$-stable or because the underlying Bradlow pair is unstable but the triple is $\bar\sigma_-$-stable anyway.
For (ii) it is enough to recall that $F_{(2),\bar\sigma_-}^{(d),1+}$ contains all the triples in $\mathcal{M}_{\bar\sigma_-}^{2,d}$ for which the underlying Bradlow pair is itself $\bar\sigma_-$-stable.
To prove (iii) consider a triple in $(E,\phi,s) \in \mathcal{W}^{d,+}_{\bar\sigma}$ whose subobject is $(L,\psi_1)$, with $s \in H^0(L)$ and whose quotient is $(M,\psi_2)$. From lemma \ref{exts} we know that there is an exact sequence:
\begin{align*}
& \H^0((M,\psi_2),(L,\psi_1)) \rightarrow H^0(M^* \otimes L) \rightarrow H^0(M^* \otimes L \otimes K) \rightarrow \H^1((M,\psi_2),(L,\psi_1)) \rightarrow\\
&\rightarrow H^1(M^* \otimes L) \rightarrow H^1(M^* \otimes L \otimes K) \rightarrow \H^2((M,\psi_2),(L,\psi_1)) \rightarrow 0
\end{align*}
and the triple is represented, up to isomorphism, by a class in $\P \H^1((M,\psi_2),(L,\psi_1))$.
Also, the underlying Bradlow pair is $\bar\sigma_-$-stable if and only if the class representing the triple does not lie in $$\ker( \H^1((M,\psi_2),(L,\psi_1)) \rightarrow H^1(M^* \otimes L)).$$
Since $\deg M > \deg L$ we know that $H^0(M^* L)=0$ so that $$H^0(M^* L K)=\ker( \H^1((M,\psi_2),(L,\psi_1)) \rightarrow H^1(M^* L)).$$
From this we deduce that the projectivization of the previous kernel is the fiber of $NS\mathcal{W}^{d,+}_{\bar\sigma} \rightarrow X^d_{\bar\sigma}$.
Statements (iv) and (v) follow as (i) and (ii). As for (vi), let $(E,\phi,s) \in \mathcal{W}^{d,-}_{\bar\sigma}$ and $(M,\psi_2)$ be the subobject, while $(L,\psi_1)$ be the quotient and $s \notin H^0(M)$ projects to $\bar{s} \in H^0(M)$. From lemma \ref{extssec} we know there is a different exact sequence:
\begin{align*}
0 \rightarrow H^0(KL^*M) \rightarrow \widetilde\H^1((L, \bar{s},\psi_1),(M,\psi_2)) \rightarrow H^0(M\mathcal{O}_D) \rightarrow 0
\end{align*}
where $D$ is the divisor associated to $\bar{s}$. The underlying Bradlow pair, $(E,s)$ is $\bar\sigma_+$-stable if and only if the class representing $(E,\phi,s)$ in $\P \widetilde\H^1((L, \bar{s},\psi_1),(M,\psi_2))$ is not in $$\ker( \widetilde\H^1((L, \bar{s},\psi_1),(M,\psi_2)) \rightarrow H^0(M\mathcal{O}_D))=H^0(KL^*M).$$
Therefore the canonical map $\mathcal{W}^{d,-}_{\bar\sigma} \rightarrow X^d_{\bar\sigma}$ also restricts to a canonical map
$$NS\mathcal{W}^{d,-}_{\bar\sigma} \rightarrow X^d_{\bar\sigma}$$
whose fiber over $(L,\bar{s},M,\psi_1, \psi_2)$ is $\P H^0(KL^*M)$.\\
For (vii) consider a triple in $NS\mathcal{W}^+_{\bar\sigma}$. The underlying Bradlow pair is split, the maximal $\bar\sigma_-$-destabilizing subbundle does not contain the section and is of degree $(d+\bar\sigma)/2$. So the Harder-Narasimhan type of the underlying pair matches with $F_{(1,1),\bar\sigma_-}^{((d-\bar\sigma)/2,(d+\bar\sigma)/2),1+}$.
Statement (viii) is proved as (vii).
\end{proof}
\end{prop}
\begin{prop}
\label{weirdexpl}
Let $\bar\sigma$ be a critical value. Then:
\begin{itemize}
\item[(i)] $B^{d,+}_{\bar\sigma}$ is the inverse image of $\P W^{d,+}_{\bar\sigma}$ with respect to the limit map $F_{(2),\bar\sigma_-}^{(d),1+} \rightarrow F_{(2),\bar\sigma_-}^{(d),1}\cong M^{2,d}_{\bar\sigma_-}$.
\item[(ii)] $S\mathcal{W}^{d,+}_{\bar\sigma} \subset B^{d,+}_{\bar\sigma}$ and $B^{d,+}_{\bar\sigma} \setminus S\mathcal{W}^{d,+}_{\bar\sigma} \subset F_{(1,1),\bar\sigma_+}^{((d+\bar\sigma)/2,(d-\bar\sigma)/2),2+}$. The complement can be described as:
\begin{equation*}
\left \{
\begin{array}{c}
(E,\phi,s) \text{ such that }\\
E=L \oplus M\\
s \in H^0(L), \deg L =(d-\bar\sigma)/2, \deg M=(d+\bar\sigma)/2\\
\text{and } \phi \text{ does not preserve } L
\end{array} \right \}.
\end{equation*}
and therefore contains $NS\mathcal{W}^{d,-}_{\bar\sigma}$. Furthermore, if $\bar\sigma > 2g-2$ then the locally closed subvariety we just described is equal to $NS\mathcal{W}^{d,-}_{\bar\sigma}$.
\item[(iii)] $B^{d,-}_{\bar\sigma}$ is the inverse image of $\P W^{d,-}_{\bar\sigma}$ with respect to the limit map $F_{(2),\bar\sigma_+}^{(d),1+} \rightarrow F_{(2),\bar\sigma_+}^{(d),1}\cong M^{2,d}_{\bar\sigma_+}$.
\item[(iv)] $S\mathcal{W}^{d,-}_{\bar\sigma} \subset B^{d,-}_{\bar\sigma} $ and $B^{d,-}_{\bar\sigma} \setminus S\mathcal{W}^{d,-}_{\bar\sigma} \subset F_{(1,1),\bar\sigma_-}^{((d-\bar\sigma)/2,(d+\bar\sigma)/2),1+}$. The complement can be described as:
\begin{equation*}
\left \{
\begin{array}{c}
(E,\phi,s) \text{ such that }\\
E=M \oplus L\\
s \in H^0(L), \deg L =(d-\bar\sigma)/2, \deg M=(d+\bar\sigma)/2\\
\text{and } \phi \text{ does not preserve } M
\end{array} \right \}.
\end{equation*}
and therefore contains $NS\mathcal{W}^{d,+}_{\bar\sigma}$. Furthermore, if $\bar\sigma > 2g-2$ then the locally closed subvariety we just described is empty and therefore $NS\mathcal{W}^{d,+}_{\bar\sigma}$ is empty as well.
\end{itemize}
\begin{proof}
(i) follows by the identification $F_{(2),\bar\sigma_-}^{(d),1}\cong M^{2,d}_{\bar\sigma_-}$.
For (ii), the first inclusion follows from the definitions. Note that $B^{d,+}_{\bar\sigma} \setminus S\mathcal{W}^{d,+}_{\bar\sigma}$ consists of triples that are $\bar\sigma_-$-stable and whose underlying pair is also $\bar\sigma_-$-stable. However those triples in $B^{d,+}_{\bar\sigma}$ that are $\bar\sigma_+$-unstable are contained in $S\mathcal{W}^{d,+}_{\bar\sigma}$ and by definition the triples in $B^{d,+}_{\bar\sigma}$ whose underlying Bradlow pair is $\bar\sigma_+$-stable are also contained in $S\mathcal{W}^{d,+}_{\bar\sigma}$. Therefore $B^{d,+}_{\bar\sigma} \setminus S\mathcal{W}^{d,+}_{\bar\sigma}$ consists of triples that are both $\bar\sigma_-$ and $\bar\sigma_+$-stable. The underlying Bradlow pairs are however $\bar\sigma_-$-stable but $\bar\sigma_+$-unstable. Since the maximal destabilizing has degree $(d-\bar\sigma)/2$ and contains the section, it will fit the Harder-Narasimhan type of $F_{(1,1),\bar\sigma_+}^{((d+\bar\sigma)/2,(d-\bar\sigma)/2),2+}$. For the second part we see immediately that the complement has to consist of triples whose underlying pair is both $\bar\sigma_+$-unstable and $\bar\sigma_-$-unstable and so has to be split and all triples in $NS\mathcal{W}^{d,-}_{\bar\sigma}$ are of this form. The statement about the case $\bar\sigma > 2g-2$ follows because, since $\deg M^*LK <0$ then any Higgs field assigned to such a split pair has to preserve the subobject not containing the section (denoted $M$) and therefore any triple in
$$F_{(1,1),\bar\sigma_+}^{((d+\bar\sigma)/2,(d-\bar\sigma)/2),2+} \setminus (B^{d,+}_{\bar\sigma} \setminus S\mathcal{W}^{d,+}_{\bar\sigma})$$
will also lie in $NS\mathcal{W}^{d,-}_{\bar\sigma}$ proving the converse inclusion.
Statement (iii) is proved by using the identification $F_{(2),\bar\sigma_+}^{(d),1}\cong M^{2,d}_{\bar\sigma_+}$ and statement (iv) has an analogous proof to (ii).
\end{proof}
\end{prop}
We conclude the section by giving a more intuitive idea of the previous proposition. Let us examine the flip locus $\mathcal{W}^{d,+}_{\bar\sigma}$ for instance. Then it can be decomposed as $\mathcal{W}^{d,+}_{\bar\sigma}=S\mathcal{W}^{d,+}_{\bar\sigma} \sqcup NS\mathcal{W}^{d,+}_{\bar\sigma}$. $S\mathcal{W}^{d,+}_{\bar\sigma}$ is the intersection of the flip locus with the attracting set $F_{(2),\bar\sigma_-}^{(d),1+}$ while $NS\mathcal{W}^{d,+}_{\bar\sigma}$ is the intersection of the flip locus with the attracting set $F_{(1,1),\bar\sigma_-}^{((d-\bar\sigma)/2,(d+\bar\sigma)/2),1+}$. After $\mathcal{W}^{d,+}_{\bar\sigma}$ has been removed from the moduli space some of the limit points are also removed, but then the limits for the triples in $B^{d,+}_{\bar\sigma} \setminus S\mathcal{W}^{d,+}_{\bar\sigma}$ are replaced by new limit points after the critical value is passed and in fact, for example, $B^{d,+}_{\bar\sigma} \setminus S\mathcal{W}^{d,+}_{\bar\sigma} \subset F_{(1,1),\bar\sigma_+}^{((d+\bar\sigma)/2,(d-\bar\sigma)/2),2+}$ indicating that the limit is in a different attracting set.
\section{The case of low stability parameter}
\subsection{Odd degree}
\label{odddeg}
As we saw in corollary \ref{smoothr2}, $\mathcal{M}^{2,d}_\varepsilon$ is smooth for $d<0$ and for $d$ odd bigger than $4g-4$. In particular in this case $\mathcal{M}^{2,d}_\varepsilon$ is semiprojective and we can use the Bia\l{}ynicki-Birula stratification to compute the motive. We start by introducing a notation:
\begin{defn}
Let $d$ be an odd integer. Denote by $I_1^o(d)$ the set of pairs of integers $(d_1,d_2)$ for which $F_{(1,1),\varepsilon}^{(d_1,d_2),1}$ is a nonempty fixed point component of $\mathcal{M}_\varepsilon^{2,d}$. Analogously, let $I_2^o(d)$ denote the set of pairs of integers $(d_1,d_2)$ for which $F_{(1,1),\varepsilon}^{(d_1,d_2),2}$ is a nonempty fixed point component of $\mathcal{M}_\varepsilon^{2,d}$.
\end{defn}
From theorem \ref{fpr2}, we see that $I_1^o(d)$ is empty if $d <1$ and if $d \geq 1$ consists of the pairs $(d_1,d_2)$ satisfying $d_1+d_2=d$ and the following three equivalent sets of inequalities:
\begin{align*}
&\max \left \{0,\frac{d+1}{2}+1-g\right \} \leq d_1\leq \frac{d-1}{2} \\
&\frac{d+1}{2} \leq d_2 \leq \min \left \{ \frac{d-1}{2}+g-1, d\right \} \\
&\max\{1,2g-2-d\} \leq d_1-d_2+2g-2 \leq 2g-3 \text{ only odd values}.
\end{align*}
$I_2^o(d)$ instead consists of the pairs of integers $(d_1,d_2)$ satisfying $d_1+d_2=d$ and the following three equivalent sets of inequalities:
\begin{align*}
&\frac{d+1}{2}+1-g \leq d_1\leq \min \left \{d,\frac{d-1}{2} \right\} \\
&\max \left \{0,\frac{d+1}{2} \right\} \leq d_2 \leq \frac{d-1}{2}+g-1 \\
&1 \leq d_1-d_2+2g-2 \leq \min\{ d+2g-2,2g-3\} \text{ only odd values}.
\end{align*}
From proposition \ref{deffixed} we know that every nonsplit fixed point is smooth. Since the dimension of the Zariski tangent space can only increase by specialization, we see that $F_{(2),\varepsilon}^{(d),1+}$ always lies in the smooth part of $\mathcal{M}^{2,d}_\varepsilon$. The same is true for split type 2 fixed points and their corresponding attracting sets. In order to compute the motive of those type 1 attracting sets containing singular points we need to work a bit more. Recall from theorem \ref{fpr2} that we have the following description of the type 1 attracting loci:
\begin{equation*}
F_{(1,1),\varepsilon}^{(d_1,d_2),1+}=
\left \{
\begin{array}{c}
(E,\phi,s) \text{ where } E \text{ is defined by}\\
0 \rightarrow E_2 \rightarrow E \rightarrow E_1 \rightarrow 0\\
s \notin H^0(E_2), \deg E_i =d_i\\
\text{and } \phi \text{ does not preserve } E_2
\end{array} \right \}
\end{equation*}
for $(d_1,d_2) \in I_1^o(d)$. We will make use of the following definition.
\begin{defn}
Let $(d_1,d_2)\in I_1^o(d)$. Let us denote by $NSPF_{(1,1),\varepsilon}^{(d_1,d_2),1+}$ the locus of $F_{(1,1),\varepsilon}^{(d_1,d_2),1+}$ where the underlying Bradlow pair lies in $\P W^{d,-}_{d_2-d_1}$ and by $SPF_{(1,1),\varepsilon}^{(d_1,d_2),1+}$ the complement, i.e. the locus where the pair is split.
\end{defn}
With the above definition we clearly have:
\begin{equation*}
NSPF_{(1,1),\varepsilon}^{(d_1,d_2),1+} = B_{d_2-d_1}^{d,-} \cap F_{(1,1),\varepsilon}^{(d_1,d_2),1+}
\end{equation*}
and also:
\begin{equation*}
F_{(1,1),\varepsilon}^{(d_1,d_2),1+} = B_{d_2-d_1}^{d,-} \cap F_{(1,1),\varepsilon}^{(d_1,d_2),1+} \sqcup SPF_{(1,1),\varepsilon}^{(d_1,d_2),1+}.
\end{equation*}
In order to compute the motives we need the following.
\begin{prop}
\begin{equation*}
B_{d_2-d_1}^{d,-} = NSPF_{(1,1),\varepsilon}^{(d_1,d_2),1+} \sqcup S \mathcal{W}^{d,-}_{d_2-d_1} = B_{d_2-d_1}^{d,-} \cap F_{(1,1),\varepsilon}^{(d_1,d_2),1+} \sqcup S \mathcal{W}^{d,-}_{d_2-d_1}.
\end{equation*}
\begin{proof}
The second equality is immediate from the previous observation. From part (iv) of proposition \ref{weirdexpl} we see that $S \mathcal{W}^{d,-}_{d_2-d_1} = B_{d_2-d_1}^{d,-}$, also by definition if $(E,\phi,s) \in NSPF_{(1,1),\varepsilon}^{(d_1,d_2),1+}$ then $(E,s) \in \P W^{d,-}_{d_2-d_1}$ and so by definition \ref{weirdl} $(E,\phi,s) \in B_{d_2-d_1}^{d,-}$. This proves one of the inclusions.
For the reverse inclusion, again from definition \ref{weirdl} we have $(E,\phi,s) \in B_{d_2-d_1}^{d,-}$ if and only if $(E,s) \in \P W^{d,-}_{d_2-d_1}$. If this is the case, then either $\phi$ preserves the subobject of degree $d_2$ that does not contain the section, and then $(E,\phi,s) \in S \mathcal{W}^{d,-}_{d_2-d_1}$ or $\phi$ does not preserve such subobject and then $(E,\phi,s) \in NSPF_{(1,1),\varepsilon}^{(d_1,d_2),1+}$. This proves the second inclusion and also the fact that the union is disjoint.
\end{proof}
\end{prop}
Now we can write the motivic identity:
\begin{equation*}
[F_{(1,1),\varepsilon}^{(d_1,d_2),1+}]=[B_{d_2-d_1}^{d,-}]-[S \mathcal{W}^{d,-}_{d_2-d_1}]+[SPF_{(1,1),\varepsilon}^{(d_1,d_2),1+}].
\end{equation*}
Therefore it suffices to compute the previous three motives. Since we will use $[B_{\bar\sigma}^{d,-}]$ in other sections, we can compute it now.
\begin{prop}
\label{bminus}
Let $\bar\sigma$ be a critical value. Then we have:
\begin{align*}
[B_{\bar\sigma}^{d,-}]&=\L^{1+4(g-1)} [S^{(d-\bar\sigma)/2}( C)] \left ( (\L-1)[S^{(d+\bar\sigma)/2}( C)]+[J^{(d+\bar\sigma)/2}( C)] \right ) + \\&+\L^{1+4(g-1)} \frac{\L^{(d-\bar\sigma)/2}-\L}{\L-1} [S^{(d-\bar\sigma)/2}( C)] \left ( (\L-1)[S^{\bar\sigma}( C)]+[J^{\bar\sigma}( C)] \right ).
\end{align*}
\begin{proof}
Recall that $(E,\phi,s) \in B_{\bar\sigma}^{d,-}$ if and only if $(E,s) \in \P W^{d,-}_{\bar\sigma}$. Therefore consider the composition of maps:
\begin{align*}
B_{\bar\sigma}^{d,-} &\rightarrow \P W^{d,-}_{\bar\sigma} \rightarrow S^{(d-\bar\sigma)/2}( C) \times J^{(d+\bar\sigma)/2}( C)\\
(E,\phi,s) & \mapsto (E,s) \mapsto (L,\bar{s}, M)
\end{align*}
where the second map is defined by using the fact that, if $(E,s) \in \P W^{d,-}_{\bar\sigma}$, then $E$ is an extension:
\begin{equation*}
0 \rightarrow M \rightarrow E \rightarrow L \rightarrow 0
\end{equation*}
with $s \notin H^0(M)$ projecting to $\bar{s} \in H^0(L)$ and $\deg M =(d+\bar\sigma)/2$, $\deg L =(d-\bar\sigma)/2$.
As explained in \cite[proposition 3.3]{thaddeus1994stable}, over $(L,\bar{s}, M)$ the fiber of
$$\P W^{d,-}_{\bar\sigma} \rightarrow S^{(d-\bar\sigma)/2}( C) \times J^{(d+\bar\sigma)/2}( C)$$
is $\P H^0(M \mathcal{O}_D)$ where $D$ is the divisor associated to $\bar{s}$ i.e. $\mathbb{C}\P^{(d-\bar\sigma)/2-1}$. Inside $\P H^0(M \mathcal{O}_D)$ there is the locus corresponding to those pairs whose underlying vector bundle is split. Such a locus is the projectivization of $$\ker \left ( H^0(M \mathcal{O}_D) \rightarrow H^1(L^*M) \right) \cong \quotient{H^0(M)}{H^0(L^*M)}.$$
Over the points of such a kernel the vector bundle will be a split sum and so the fiber of $B_{\bar\sigma}^{d,-} \rightarrow \P W^-_{\bar\sigma}$ is $H^0(K \End E)$ whose dimension, according to lemma \ref{unstableext}, is $2+4(g-1)+\dim H^0(L^*M)$, over the complement instead the fiber is still $H^0(K \End E)$ but the dimension is now $1+4(g-1)+\dim H^0(L^*M)$.
We introduce the stratification:
\begin{equation*}
Z^{d, \bar \sigma}_{i,j}:= \{ (L,\bar{s}, M) \in S^{(d-\bar\sigma)/2}( C) \times J^{(d+\bar\sigma)/2}( C): \dim H^0(L^*M)=i, \dim H^0(M)=j \}.
\end{equation*}
With this definition and the previous observations we can compute the motive of $B_{\bar\sigma}^{d,-}$:
\begin{align*}
[B_{\bar\sigma}^{d,-}]&=\sum_{i,j} [Z^{d, \bar \sigma}_{i,j}] \left ( [\mathbb{C}\P^{j-i-1}] \L^{2+4(g-1)+i} + ([\mathbb{C}\P^{(d-\bar\sigma)/2-1}] - [\mathbb{C}\P^{j-i-1}]) \L^{1+4(g-1)+i} \right )=\\
&=\frac{\L^{1+4(g-1)}}{\L-1} \sum_{i,j} [Z^{d, \bar \sigma}_{i,j}] \left ( \L^{j+1}-\L^{i+1} + \L^{(d-\bar\sigma)/2+i} - \L^{j} \right )
\end{align*}
Now observe that:
\begin{align*}
\sum_{i,j}& [Z^{d, \bar \sigma}_{i,j}] \L^j = [S^{(d-\bar\sigma)/2}( C)] \sum_j [V_j^{(d+\bar\sigma)/2}]\L^j =\\
&=[S^{(d-\bar\sigma)/2}( C)] \left ( (\L-1)[S^{(d+\bar\sigma)/2}( C)]+[J^{(d+\bar\sigma)/2}( C)] \right ).
\end{align*}
In order to compute
\begin{equation*}
\sum_{i,j} [Z^{d, \bar \sigma}_{i,j}] \L^i
\end{equation*}
we can argue as follows. Consider the isomorphism:
\begin{align*}
S^{(d-\bar\sigma)/2}( C) \times J^{(d+\bar\sigma)/2}( C) & \rightarrow S^{(d-\bar\sigma)/2}( C) \times J^{\bar\sigma}( C)\\
(L,\bar{s},M) & \mapsto (L,\bar{s},L^*M).
\end{align*}
The stratum $\cup_j Z^{d, \bar \sigma}_{i,j}$ in $S^{(d-\bar\sigma)/2}( C) \times J^{(d+\bar\sigma)/2}( C) $ is then isomorphic to $S^{(d-\bar\sigma)/2}( C) \times V_i^{\bar\sigma}$ in $S^{(d-\bar\sigma)/2}( C) \times J^{\bar\sigma}( C)$. Therefore:
\begin{align*}
\sum_{i,j}& [Z^{d, \bar \sigma}_{i,j}] \L^i = [S^{(d-\bar\sigma)/2}( C)] \sum_i [ V_i^{\bar\sigma}] \L^i = [S^{(d-\bar\sigma)/2}( C)] \left ( (\L-1)[S^{\bar\sigma}( C)]+[J^{\bar\sigma}( C)] \right ).
\end{align*}
Putting all together, we get
\begin{align*}
[B_{\bar\sigma}^{d,-}]&=\L^{1+4(g-1)} [S^{(d-\bar\sigma)/2}( C)] \left ( (\L-1)[S^{(d+\bar\sigma)/2}( C)]+[J^{(d+\bar\sigma)/2}( C)] \right ) + \\&+\L^{1+4(g-1)} \frac{\L^{(d-\bar\sigma)/2}-\L}{\L-1} [S^{(d-\bar\sigma)/2}( C)] \left ( (\L-1)[S^{\bar\sigma}( C)]+[J^{\bar\sigma}( C)] \right ).
\end{align*}
\end{proof}
\end{prop}
\begin{prop}
We have the identity:
\begin{equation*}
[SPF_{(1,1),\varepsilon}^{(d_1,d_2),1+}]=\L^{3g-1+d_2-d_1} [S^{d_1} ( C)] [S^{d_1-d_2+2g-2}( C)].
\end{equation*}
\begin{proof}
Recall that $SPF_{(1,1),\varepsilon}^{(d_1,d_2),1+}$ consists of those pairs in the form $(E_1\oplus E_2,s,\phi)$ where $\deg E_i=d_i$, $s \in H^0(E_1)$ and $\phi$ does not preserve $E_2$. First of all, note that the moduli space of such split pairs is
\begin{equation*}
S^{d_1} ( C) \times J^{d_2} (C ).
\end{equation*}
Second, we can always write the Higgs field into matrix form:
\begin{equation*}
\phi =
\begin{pmatrix}
\phi_{11} & \phi_{21}\\
\phi_{12} & \phi_{22}
\end{pmatrix}
\end{equation*}
where
\begin{equation*}
\phi_{21} : E_2 \rightarrow E_1 \otimes K
\end{equation*}
has to be nonzero, and
\begin{equation*}
\phi_{12} : E_1 \rightarrow E_2 \otimes K.
\end{equation*}
Since the endomorphisms of the underlying pair are just diagonal (possibly non scalar) then we can see that $(E_1 \oplus E_2, s , \phi)$ and $(E_1 \oplus E_2, s , \phi')$ are isomorphic if and only if there exists $\lambda \in \mathbb{C}^*$ such that:
\begin{equation*}
\phi' =
\begin{pmatrix}
\phi'_{11} & \phi'_{21}\\
\phi'_{12} & \phi'_{22}
\end{pmatrix}=
\begin{pmatrix}
\phi_{11} & \lambda\phi_{21}\\
\lambda^{-1}\phi_{12} & \phi_{22}
\end{pmatrix}.
\end{equation*}
Let's consider the map:
\begin{equation*}
SPF_{(1,1),\varepsilon}^{(d_1,d_2),1+} \rightarrow S^{d_1} ( C) \times J^{d_2} (C ) \times H^0(K)^2
\end{equation*}
forgetting the off-diagonal parts of the Higgs field. Then, by what we said, the fiber of such a map over $(E_1,s,E_2)$ will be:
\begin{equation*}
\quotient{\left (H^0(KE_1 E_2^*) \setminus \{0\} \right ) \times H^0(K E_2 E_1^*)}{ \mathbb{C}^*}.
\end{equation*}
This allows us to compute the motive of $SPF_{(1,1),\sigma}^{(d_1,d_2),1+}$ by stratifying $S^{d_1} ( C) \times J^{d_2} (C )$ according to the dimension of $H^0(KE_1 E_2^*)$. Note in fact that the dimension of $H^0(K E_2 E_1^*)$ is constantly equal to $g-1+d_2-d_1$. Call $V^{d_1-d_2+2g-2}_i$ the stratum of $J^{d_1-d_2+2g-2}$ where the dimension of the global sections of the line bundle is equal to $i$. Then we have:
\begin{align*}
[SPF_{(1,1),\varepsilon}^{(d_1,d_2),1+}]=&\sum_i \left (\L^{2g} [S^{d_1} ( C)] [V^{d_1-d_2+2g-2}_i] \L^{d_2-d_1+g-1} [\mathbb{C}\P^{i-1} ]\right )=\\
=&\L^{3g-1+d_2-d_1} [S^{d_1} ( C)] [S^{d_1-d_2+2g-2}( C)].
\end{align*}
\end{proof}
\end{prop}
Lastly, we have the following.
\begin{prop}
Let $\bar\sigma$ be a critical value. The following motivic equalities hold:
\begin{align*}
[NS\mathcal{W}_{\bar\sigma}^{d,-}]&=\L^{2g} [S^{(d-\bar\sigma)/2} ( C)] [J ( C)] [\mathbb{C}\P^{g-2+\bar\sigma}]\\
[S\mathcal{W}_{\bar\sigma}^{d,-}]&=\L^{2g} [S^{(d-\bar\sigma)/2} ( C)] [J ( C)] \cdot ([\mathbb{C}\P^{(d+\bar\sigma)/2+g-2}]-[\mathbb{C}\P^{g-2+\bar\sigma}]).
\end{align*}
\begin{proof}
From part (vi) of proposition \ref{weirdexpl1} we know that there exists a map $NS\mathcal{W}_{\bar\sigma}^{d,-} \rightarrow X^d_{\bar\sigma}$ whose fiber over $(L,\bar{s},M,\psi_1,\psi_2)$ is $\P H^0(KL^*M)$. Note that the dimension of the fibers is constant and is always $g-2+\bar\sigma$. Therefore:
\begin{equation*}
[NS\mathcal{W}_{\bar\sigma}^{d,-}]=\L^{2g} [S^{(d-\bar\sigma)/2} ( C)] [J ( C)] [\mathbb{C}\P^{g-2+\bar\sigma}]
\end{equation*}
and
\begin{equation*}
[S\mathcal{W}_{\bar\sigma}^{d,-}]=[\mathcal{W}_{\bar\sigma}^{d,-}]-[NS\mathcal{W}_{\bar\sigma}^{d,-}]=\L^{2g} [S^{(d-\bar\sigma)/2} ( C)] [J ( C)] \cdot ([\mathbb{C}\P^{(d+\bar\sigma)/2+g-2}]-[\mathbb{C}\P^{g-2+\bar\sigma}]).
\end{equation*}
\end{proof}
\end{prop}
We can now have a formula for the type 1 attracting sets.
\begin{prop}
\begin{align*}
&[F_{(1,1),\varepsilon}^{(d_1,d_2),1+}]=\L^{3g-2+d_2} [S^{d_1}( C)] [S^{d_1-d_2+2g-2}( C)]+\\
&+(\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] \left ( [S^{d_2}( C)]-[J( C)] \frac{\L^{d_2+1-g}-1}{\L-1} \right ).
\end{align*}
\begin{proof}
We can compute:
\begin{align*}
&[F_{(1,1),\varepsilon}^{(d_1,d_2),1+}]=[B_{d_2-d_1}^{d,-}]-[S \mathcal{W}^{d,-}_{d_2-d_1}]+[SPF_{(1,1),\varepsilon}^{(d_1,d_2),1+}]=\\
&=\L^{1+4(g-1)} [S^{d_1}( C)] \left ( (\L-1)[S^{d_2}( C)]+[J( C)] \right ) + \\
&+\L^{1+4(g-1)} \frac{\L^{d_1}-\L}{\L-1} [S^{d_1}( C)] \left ( (\L-1)[S^{d_2-d_1}( C)]+[J( C)] \right )+\\
&-\L^{2g} [S^{d_1} ( C)] [J ( C)] \cdot ([\mathbb{C}\P^{d_2+g-2}]-[\mathbb{C}\P^{g-2+d_2-d_1}])+\\
&+\L^{3g-1+d_2-d_1} [S^{d_1} ( C)] [S^{d_1-d_2+2g-2}( C)]=\\
&=(\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] [S^{d_2}( C)]+\L^{4g-3} [S^{d_1}( C)] [J( C)]+\\
&+(\L^{4g-3+d_1}-\L^{4g-2}) [S^{d_1}( C)] [S^{d_2-d_1}( C)]+\L^{4g-3} \frac{\L^{d_1}-\L}{\L-1}[S^{d_1}( C)] [J( C)]+\\
&-\L^{3g-1+d_2-d_1}\frac{\L^{d_1}-1}{\L-1}[S^{d_1}( C)] [J( C)]+\\
&+ \L^{3g-1+d_2-d_1} [S^{d_1} ( C)] [S^{d_1-d_2+2g-2}( C)]=\\
&=(\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] [S^{d_2}( C)]+\\
&+ (\L^{4g-3}-\L^{3g-1+d_2-d_1})\frac{\L^{d_1}-1}{\L-1}[S^{d_1}( C)] [J( C)]+\\
&+ [S^{d_1}( C)]\left ( (\L^{4g-3+d_1}-\L^{4g-2}) [S^{d_2-d_1}( C)] + \L^{3g-1+d_2-d_1} [S^{d_1-d_2+2g-2}( C)] \right).
\end{align*}
From Serre's duality we have the identity:
\begin{equation*}
[S^{d_2-d_1}( C)]=\L^{d_2-d_1-g+1}[S^{d_1-d_2+2g-2}( C)]+\frac{\L^{d_2-d_1-g}-1}{\L-1}[J( C)]
\end{equation*}
for $0 \leq d_2-d_1 \leq 2g-2$. Using this we can proceed in the computation:
\begin{align*}
&[F_{(1,1),\varepsilon}^{(d_1,d_2),1+}]=(\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] [S^{d_2}( C)]+ \\
&+(\L^{4g-3}-\L^{3g-1+d_2-d_1})\frac{\L^{d_1}-1}{\L-1}[S^{d_1}( C)] [J( C)]+\\
&+ (\L^{4g-3+d_1}-\L^{4g-2})\frac{\L^{d_2-d_1+1-g}-1}{\L-1}[S^{d_1}( C)] [J( C)]+\\
&+(\L^{4g-3+d_1}-\L^{4g-2})\L^{d_2-d_1+1-g}[S^{d_1}( C)][S^{d_1-d_2+2g-2}( C)]+\\
&+\L^{3g-1+d_2-d_1} [S^{d_1}( C)] [S^{d_1-d_2+2g-2}( C)]=\\
&=(\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] [S^{d_2}( C)]+ \\
&-(\L^{4g-2}-\L^{4g-3})\frac{\L^{d_2+1-g}-1}{\L-1} [S^{d_1}( C)] [J( C)]+\\
&+\L^{3g-2+d_2} [S^{d_1}( C)] [S^{d_1-d_2+2g-2}( C)]=\\
&=\L^{3g-2+d_2} [S^{d_1}( C)] [S^{d_1-d_2+2g-2}( C)]+\\
&+(\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] \left ([S^{d_2}( C)]-[J( C)] \frac{\L^{d_2+1-g}-1}{\L-1} \right).
\end{align*}
\end{proof}
\end{prop}
\begin{rmk}
Observe that, according to proposition \ref{deffixed}, the attracting set $F_{(1,1),\varepsilon}^{(d_1,d_2),1+}$ lies in the smooth part if and only if $d_2>2g-2$. In that case we can apply \cite[Theorem 4.1]{bialynicki1973some} and deduce that $F_{(1,1),\varepsilon}^{(d_1,d_2),1+} \rightarrow F_{(1,1),\varepsilon}^{(d_1,d_2),1}$ is a smooth affine fibration of rank $3g-2+d_2$ (see section \ref{rank2disc}). In this case we have $[S^{d_2}( C)]=[J( C)][\mathbb{C}\P^{d_2-g}]$ and hence:
\begin{align*}
&[F_{(1,1),\varepsilon}^{(d_1,d_2),1+}]= \L^{3g-2+d_2} [S^{d_1}( C)] [S^{d_1-d_2+2g-2}( C)]
\end{align*}
which is the same as computing $F_{(1,1),\varepsilon}^{(d_1,d_2),1+}$ as an affine fibration.
\end{rmk}
We can summarize the section in the following.
\begin{theorem}
\label{motodd}
For $d \geq 0 $ odd and $0<\varepsilon<1$ we have:
\begin{align*}
[\mathcal{M}_\varepsilon^{2,d}]=&\L^{1+4(g-1)}[M_\varepsilon^{2,d}]+\sum_{(d_1,d_2) \in I_1^o(d)} \L^{1+3(g-1)+d_2}[S^{d_1}( C)][S^{d_1-d_2+2g-2}]+\\
&+\sum_{(d_1,d_2) \in I_2^o(d)}\L^{1+4(g-1)} [S^{d_2}( C)][S^{d_1-d_2+2g-2}( C)]+\\
&+\sum_{(d_1,d_2) \in I_1^o(d)} (\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] \left ([S^{d_2}( C)]-[J( C)] \frac{\L^{d_2+1-g}-1}{\L-1}\right).
\end{align*}
\end{theorem}
\subsection{The moduli spaces for even degree}
For even degree $d$, the situation is quite different due to the presence of strictly semistable Higgs bundles. In fact, for $0<\varepsilon<2$ and even $d$, it is clear that for every $\varepsilon$-stable triple $(E,\phi,s)$ the underlying Higgs bundle $(E,\phi)$ is semistable. Therefore, there is still an Abel-Jacobi map $\mathcal{M}_\varepsilon^{2,d} \rightarrow \mathcal{M}^{2,d}$.
We can say something a bit more precise for $\varepsilon$-stability for even degree. It is easy to see that $(E,\phi,s)$ is $\varepsilon$-stable if and only if $(E,\phi)$ is semistable and if $L \subset E$ is a $\phi$-invariant line subbundle of $E$ with $\deg L = \deg E /2$, then $s \notin H^0(L)$.
Furthermore, the fibers of the Abel-Jacobi map are more complicated than just projectivized spaces of global sections because of the fact that the equivalence relation for the moduli spaces is isomorphism for triples and S-equivalence for Higgs bundles.\\
However, it is still true that above the stable locus of $\mathcal{M}^{2,d}$ the Abel-Jacobi map behaves exactly as in the case of odd degree.
Let us outline the strategy for the computation of $[\mathcal{M}_\varepsilon^{2,d}]$.
\begin{defn}
Let $0<\varepsilon<2$ and $d$ be an even integer. We define two locally closed subvarieties of $\mathcal{M}_\varepsilon^{2,d}$ by:
\begin{equation*}
\mathscr{X}^d_1=\left \{\begin{array}{c}
(E,\phi,s) \in \mathcal{M}_\varepsilon^{2,d} \text{ such that}\\
(E,s) \text{ is $\varepsilon$-stable and}\\
E \text{ is strictly semistable}
\end{array} \right \}
\end{equation*}
and
\begin{equation*}
\mathscr{X}^d_2=\left \{\begin{array}{c}
(E,\phi,s) \text{ such that $E$ fits into an exact sequence}\\
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0\\
\text{with $\deg L=\deg M$, $s \in H^0(L)$ and}\\
\phi \text{ does not preserve } L
\end{array} \right \}.
\end{equation*}
\end{defn}
First we examine the attracting sets of $\mathcal{M}_\varepsilon^{2,d}$ more closely. For the following, refer to theorem \ref{fpr2} and \ref{rank2disc}.
The attracting set corresponding to nonsplit fixed points is $F_{(2),\varepsilon}^{(d),1+}$ and is nonempty if and only if $d \geq 0$. There is a limit map $F_{(2),\varepsilon}^{(d),1+} \rightarrow M_\varepsilon^{2,d}$ whose fiber over a pair $(E,s)$ is $H^0(K \End E)$. We can write:
\begin{equation*}
M_\varepsilon^{2,d}=M_{\varepsilon,st}^{2,d} \sqcup M_{\varepsilon,ss}^{2,d}
\end{equation*}
where the first subvariety corresponds to the locus of pairs having underlying stable vector bundle, while the second one corresponds to the locus of pairs having strictly semistable vector bundle. Then $F_{(2),\varepsilon}^{(d),1+} \rightarrow M_\varepsilon^{2,d}$ is an affine fibration over $M_{\varepsilon,st}^{2,d}$ with fibers of dimension $4g-3$. The inverse image of $F_{(2),\varepsilon}^{(d),1+} \rightarrow M_\varepsilon^{2,d}$ over $M_{\varepsilon,ss}^{2,d}$ is instead precisely $\mathscr{X}^d_1$ defined above.
For the split cells we can introduce a notation as for the odd degree case.
\begin{defn}
Let $d \geq 0$ be an even integer. Denote by $I_1^e(d)$ the set of pairs of integers $(d_1,d_2)$ for which $F_{(1,1),\varepsilon}^{(d_1,d_2),1}$ is a nonempty fixed point component of $\mathcal{M}_\varepsilon^{2,d}$. Analogously, let $I_2^e(d)$ denote the set of pairs of integers $(d_1,d_2)$ for which $F_{(1,1),\varepsilon}^{(d_1,d_2),2}$ is a nonempty fixed point component of $\mathcal{M}_\varepsilon^{2,d}$
\end{defn}
From theorem \ref{fpr2}, we see that $I_1^e(d)$ is nonempty if and only if $d \geq 2$ and consists of pairs of integers $(d_1,d_2)$ satisfying $d_1+d_2=d$ and the following three equivalent sets of inequalities:
\begin{align*}
&\max \left \{0, d/2+1-g\right \} \leq d_1\leq d/2-1 \\
&d/2+1 \leq d_2 \leq \min \left \{ d, d/2+g-1\right \} \\
&\max\{0,2g-2-d\} \leq d_1-d_2+2g-2 \leq 2g-4 \text{ only even values}
\end{align*}
and $I_2^e(d)$ is the set of pairs of integers $(d_1,d_2)$ satisfying $d_1+d_2=d$ and the following three equivalent sets of inequalities:
\begin{align*}
&d/2+1-g \leq d_1\leq \min\{ d/2,d\} \\
&\max\{0,d/2\} \leq d_2 \leq d/2+g-1 \\
&0 \leq d_1-d_2+2g-2 \leq \min\{2g-2,2g-2+d\} \text{ only even values}.
\end{align*}
As we proved in theorem \ref{shatz}, the decomposition into attracting cells coincides with the decomposition according to the Harder-Narasimhan type of the underlying Bradlow pair. From this we see that $F_{(1,1),\varepsilon}^{(d/2,d/2),2+}$ corresponds to the Harder-Narasimhan type $(d/2,d/2)$ i.e. the destabilizing subobject of the pair $(E,s)$ is a line bundle of degree $d/2$ that contains the section $s$. Therefore $F_{(1,1),\varepsilon}^{(d/2,d/2),2+}=\mathscr{X}^d_2$.
If we denote by $\tilde{I}_2^e(d)=I_2^e(d) \setminus \{(d/2,d/2)\}$ then the previous discussion allows us to write the following preliminary formula for $d \geq 0$. Note that the computation of the type 1 attracting set is exactly the same as in section \ref{odddeg}.
\begin{align*}
[\mathcal{M}_\varepsilon^{2,d}]&=\L^{1+4(g-1)} [M^{2,d}_{\varepsilon,st}]+\sum_{(d_1,d_2) \in I_1^e(d)} \L^{1+3(g-1)+d_2} [S^{d_1}( C)][S^{d_1-d_2+2g-2}( C)] +\\
&+\sum_{(d_1,d_2) \in \tilde{I}_2^e(d)}\L^{1+4(g-1)} [S^{d_2}( C)][S^{d_1-d_2+2g-2}( C)]+\\
&+[\mathscr{X}^d_1]+[\mathscr{X}^d_2]+\\
&+\sum_{(d_1,d_2) \in I_1^e(d)} (\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] ([S^{d_2}( C)]-[J( C)][\mathbb{C}\P^{d_2-g}]).
\end{align*}
In order to complete the above formula, we are left with computing $[M^{2,d}_{\varepsilon,st}]$, $[M^{2,d}_{\varepsilon,ss}]$, $[\mathscr{X}^d_1]$ and $[\mathscr{X}^d_2]$. Let us first state some technical lemmas that we will need in the computation.
\begin{lemma}
\label{dim_end_even}
Let $E$ be a rank 2 vector bundle over $C$ fitting in the following exact sequence:
\begin{equation*}
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0
\end{equation*}
with $L$ and $M$ line bundles of the same degree.
Then:
\begin{equation*}
\dim H^0(\End E)= \begin{cases}
1 & \text{if $L \neq M$ and the extension is nonsplit}\\
2 & \text{if $L \neq M$ and the extension is split}\\
2 & \text{if $L = M$ and the extension is nonsplit}\\
4 & \text{if $L = M$ and the extension is split}
\end{cases}
\end{equation*}
\begin{proof}
The statement is clear in the last case, where $\End E = \mathcal{O}^4$ and in the second case where $\End E = \mathcal{O}^2\oplus L^*M \oplus M^*L$ since $M^*L$ and $L^*M$ have global sections if and only if $L=M$.
Suppose now that the extension is nonsplit and that $L \neq M$. Consider the exact sequence:
\begin{equation*}
0 \rightarrow \mathcal{O} \rightarrow E L^* \rightarrow ML^* \rightarrow 0.
\end{equation*}
Since $H^0(L^*M)=0$, the long exact sequence obtained by applying $H^0$ to the above short exact sequence will yield $H^0(EL^*)=H^0( \mathcal{O})=\mathbb{C}$. Consider now:
\begin{equation*}
0 \rightarrow LM^* \rightarrow E M^* \rightarrow \mathcal{O} \rightarrow 0.
\end{equation*}
Again, $H^0(LM^*)=0$ therefore we get an injection $H^0(EM^*) \rightarrow H^0( \mathcal{O})$. Such a map however sends $\gamma : M \rightarrow E$ to $p \gamma$ where $p: E \rightarrow M$ is the projection in the exact sequence for $E$.
If $H^0(EM^*) \neq 0$ then there would be a map $\gamma : M \rightarrow E$ such that $p \gamma \neq 0$ i.e. a splitting of the sequence which is a contradiction. Therefore $H^0(EM^*) = 0$.\\
We can conclude this case by observing that there is a further exact sequence:
\begin{equation*}
0 \rightarrow EM^* \rightarrow \End E \rightarrow EL^* \rightarrow 0
\end{equation*}
yielding the long exact sequence:
\begin{equation*}
0 \rightarrow H^0(EM^*) \rightarrow H^0(\End E) \rightarrow H^0(EL^*) \rightarrow \dots
\end{equation*}
Note that the map $H^0(\End E) \rightarrow H^0(EL^*)$ sends $f: E \rightarrow E$ to $fi : L \rightarrow E$ where $i : L \rightarrow E$ is the inclusion map in the exact sequence for $E$. Since $\mathbb{I}_E$ is sent to $i : L \rightarrow E$ we conclude that the map is surjective and therefore $H^0(\End E)=\mathbb{C}$.
When the sequence is nonsplit and $L=M$, from the sequence:
\begin{equation*}
0 \rightarrow LM^* \rightarrow E M^* \rightarrow \mathcal{O} \rightarrow 0,
\end{equation*}
arguing as before, we deduce that $H^0(EM^*)=\mathbb{C}$. However this also implies that $H^0(EL^*)= \mathbb{C}$ and therefore from the sequence
\begin{equation*}
0 \rightarrow EM^* \rightarrow \End E \rightarrow EL^* \rightarrow 0
\end{equation*}
we can deduce that $H^0(\End E)= \mathbb{C}^2$ because $H^0( \End E ) \rightarrow H^0(EL^*)$ is still surjective.
\end{proof}
\end{lemma}
\begin{lemma}
Let $n \geq 0$ be an integer and let $\Gr(2,n+2)$ denote the Grassmannian variety of vector spaces of dimension two inside $\mathbb{C}^{n+2}$. Then:
\begin{equation*}
[\Sym^2 ( \mathbb{C}\P^n)]=[\Gr (2,n+2)]=\frac{(\L^{n+2}-1)(\L^{n+1}-1)}{(\L^2-1)(\L-1)}.
\end{equation*}
\begin{proof}
From \cite[Lemma 4.4]{gottsche2000motive} we deduce, by putting $X=\{ pt \}$, that $\Sym^n(\L^m)=\L^{mn}$ in the ring of motives. Therefore:
\begin{align*}
[\Sym^2 ( \mathbb{C}\P^n)]&= \sum_{0 \leq i < j \leq n} \L^{i+j} + \sum_{j=0}^n \L^{2j}=\sum_{j=1}^n \L^j \left ( \frac{\L^j-1}{\L-1}\right ) + \frac{\L^{2n+2}-1}{\L^2-1}=\\
&= \sum_{j=0}^n \frac{\L^{2j}}{\L-1}- \sum_{j=0}^n \frac{\L^{j}}{\L-1}+\frac{\L^{2n+2}-1}{\L^2-1}=\\
&=\frac{\L^{2n+2}-1}{(\L^2-1)(\L-1)}-\frac{\L^{n+1}-1}{(\L-1)^2}+\frac{\L^{2n+2}-1}{\L^2-1}=\\
&=\frac{\L^{2n+2}-1-\L^{n+2}+\L-\L^{n+1}+1+\L^{2n+3}-\L^{2n+2}-\L+1}{(\L^2-1)(\L-1)}=\\
&=\frac{(\L^{n+2}-1)(\L^{n+1}-1)}{(\L^2-1)(\L-1)}.
\end{align*}
To compute the motive of $\Gr(2,n+2)$ we can view it as a global quotient of:
\begin{equation*}
\{(u,v) \in \mathbb{C}^{n+2} : u \neq v, u \neq 0 \neq v\}
\end{equation*}
by $GL_2$. Therefore:
\begin{align*}
[\Gr(2, n+2)]&=\frac{(\L^{n+2}-1)(\L^{n+2}-\L)}{(\L^2-1)(\L^2-\L)}=\frac{(\L^{n+2}-1)(\L^{n+1}-1)}{(\L^2-1)(\L-1)}
\end{align*}
and the result follows.
\end{proof}
\end{lemma}
\begin{lemma}
Let $E$ be a rank two vector bundle fitting in the nonsplit exact sequence:
\begin{equation*}
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0
\end{equation*}
with $L$ and $M$ line bundles and $\deg L = \deg M$.
Then, the locus of $H^0(K \End E)$ consisting of the Higgs fields preserving the subobject $L$ is a vector subspace of $H^0(K \End E)$ of dimension $3g-1$.
\begin{proof}
Let us call $i$ and $p$ the inclusion and projection in the exact sequence defining $E$. First observe that $\phi \in H^0(K \End E)$ preserves $L$ if and only if $p \phi i=0$. This means that the locus we are looking for is the kernel of the map:
\begin{align*}
H^0(K \End E) &\rightarrow H^0(KL^*M)\\
\phi & \mapsto p \phi i.
\end{align*}
Furthermore, such a map factors as:
\begin{align*}
H^0(K \End E) &\rightarrow H^0(KE^*M)\\
\phi & \mapsto p \phi
\end{align*}
followed by
\begin{align*}
H^0(K E^*M) &\rightarrow H^0(KL^*M)\\
\gamma & \mapsto \gamma i.
\end{align*}
Suppose that $L \neq M$. As we computed in \ref{dim_end_even}, $\dim H^0(K \End E)=1+4(g-1)$, $\dim H^0(K E^* L)=2g-1$, $\dim H^0(KE^* M)=2g-2$. From the exact sequence:
\begin{equation*}
0 \rightarrow H^0(K E^*L) \rightarrow H^0( K \End E) \rightarrow H^0(K E^* M) \rightarrow \dots
\end{equation*}
we see that $\dim \im (H^0( K \End E) \rightarrow H^0(K E^* M))=2g-2=\dim H^0(K E^* M)$. Therefore $H^0( K \End E) \rightarrow H^0(K E^* M)$ is surjective and this implies:
\begin{align*}
&\dim \im (H^0(K \End E) \rightarrow H^0(KL^*M))=\dim \im (H^0(K E^*M) \rightarrow H^0(KL^*M)).
\end{align*}
Furthermore, from the sequence:
\begin{equation*}
0 \rightarrow H^0(K) \rightarrow H^0( K E^*M) \rightarrow H^0(K L^* M) \rightarrow \dots
\end{equation*}
we deduce that $\dim \im (H^0(K E^*M) \rightarrow H^0(KL^*M))=g-2$.
To conclude, we have:
\begin{align*}
&\dim \ker (H^0(K \End E) \rightarrow H^0(KL^*M) )=\\
&=1+4(g-1) - \dim \im (H^0(K E^*M) \rightarrow H^0(KL^*M))=3g-1.
\end{align*}
In the case when $L=M$ the proof works in the same way but $\dim H^0(K \End E)=2+4(g-1)$, $\dim H^0(K E^* L)=2g-1$, $\dim H^0(KE^* M)=2g-1$. Here $H^0( K \End E) \rightarrow H^0(K E^* M)$ is still surjective and $\dim \im (H^0(K E^*M) \rightarrow H^0(KL^*M))=g-1$ from which we deduce that the dimension of our kernel is again $3g-1$.
\end{proof}
\end{lemma}
Let us start by defining a decomposition of $ \mathscr{X}^d_1$ and $M_{\varepsilon,ss}^{2,d}$ that we will use in the computation of the motive.
\begin{defn}
Let $(E,\phi,s)$ be a triple in $ \mathscr{X}^d_1$. Then there are three possibilities for $E$: it is a nonsplit extension of two line bundles of the same degree, it is a split extension of two different line bundles of the same degree or it is a split extension of two copies of the same line bundle. Accordingly we define:
\begin{equation*}
\mathscr{U}^d_1=\left \{\begin{array}{c}
(E,\phi,s) \text{ such that $E$ fits into a nonsplit exact sequence}\\
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0\\
\text{with $\deg L=\deg M$, $s \notin H^0(L)$ and}\\
\phi \in H^0(K\End E) \text{ arbitrary}
\end{array}\right \}
\end{equation*}
covering the first case,
\begin{equation*}
\mathscr{V}^d_1=\left \{\begin{array}{c}
(E,\phi,s) \text{ such that } E=L \oplus M \\
\text{with $\deg L=\deg M$, $s \notin H^0(L)$ and $s \notin H^0(M)$, $L \neq M$ and}\\
\phi \in H^0(K\End E) \text{ arbitrary}
\end{array}\right \}
\end{equation*}
covering the second and finally
\begin{equation*}
\mathscr{W}^d_1=\left \{\begin{array}{c}
(E,\phi,s) \text{ such that } E=L \oplus L \\
\text{$s= u\oplus v$ with $u$, $v$ linearly independent and}\\
\phi \in H^0(K\End E) \text{ arbitrary}
\end{array}\right \}
\end{equation*}
covering the third.
Similarly for $M_{\varepsilon,ss}^{2,d}$ we define:
\begin{equation*}
\widetilde{\mathscr{U}}^d_1=
\left \{\begin{array}{c}
(E,\phi,s) \text{ such that $E$ fits into a nonsplit exact sequence}\\
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0\\
\text{with $\deg L=\deg M$, $s \notin H^0(L)$}
\end{array}\right \}
\end{equation*}
covering the first case,
\begin{equation*}
\widetilde{\mathscr{V}}^d_1=\left \{\begin{array}{c}
(E,\phi,s) \text{ such that } E=L \oplus M \\
\text{with $\deg L=\deg M$, $L \neq M$, $s \notin H^0(L)$ and $s \notin H^0(M)$ }
\end{array}\right \}
\end{equation*}
covering the second and finally
\begin{equation*}
\widetilde{\mathscr{W}}^d_1=\left \{\begin{array}{c}
(E,\phi,s) \text{ such that } E=L \oplus L \\
\text{$s= u\oplus v$ with $u$, $v$ linearly independent}
\end{array}\right \}
\end{equation*}
covering the third.
\end{defn}
\begin{rmk}
A few comments are in order for the previous definitions. Suppose that $0 < \varepsilon <2$. For a nonsplit exact sequence:
\begin{equation*}
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0
\end{equation*}
with $\deg L=\deg M$ it is clear that $(E,s)$ is $\varepsilon$-stable if and only if $s \notin L$. For $E=L \oplus M$ with $L \neq M$ instead, since $\Aut E = (\mathbb{C}^*)^2$, we see that $(E,s)$ is $\varepsilon$-stable if and only if $s$ is not concentrated in $L$ nor $M$, which is the same as saying that $s=u \oplus v$ with $0 \neq u \in H^0(L)$ and $0 \neq v \in H^0(M)$. Finally, when $E=L \oplus L$ then $\Aut E = \GL_4(\mathbb{C})$ so $(E,s)$ is $\varepsilon$-stable if and only if $s= u \oplus v$ with $u, v \in H^0(L)$ generating a dimension 2 subspace.
\end{rmk}
\begin{prop}
\label{motiveU}
We have the following motivic equalities:
\begin{align*}
[\mathscr{U}^d_1]&=\L^{2+4(g-1)}\sum_{j \geq 1} [V_j^{d/2}] \left ( \frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{d/2}-1}{\L-1}-\frac{\L^{j-1}-1}{\L-1}\right )+\\
&+\L^{1+4(g-1)}\sum_{j \geq 0} [V_j^{d/2}] \left ( [S^{d/2}( C)]-\frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{d/2}-1}{\L-1}-\frac{\L^{j}-1}{\L-1}\right ).
\end{align*}
and
\begin{align*}
[\widetilde{\mathscr{U}}^d_1]&=\sum_{j \geq 1} [V_j^{d/2}] \left ( \frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{d/2}-1}{\L-1}-\frac{\L^{j-1}-1}{\L-1}\right )+\\
&+\sum_{j \geq 0} [V_j^{d/2}] \left ( [S^{d/2}( C)]-\frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{d/2}-1}{\L-1}-\frac{\L^{j}-1}{\L-1}\right ).
\end{align*}
\begin{proof}
Let $(E,s) \in \widetilde{\mathscr{U}}^d_1$ and $E$ be defined by a nonsplit exact sequence:
\begin{equation*}
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0.
\end{equation*}
Then, by Lemma \ref{dim_end_even}, we know that $\dim H^0( \End E)=1$ if $L \neq M$ and $2$ if $L=M$. Note that in both cases, if $p: E \rightarrow M$ is the projection map in the exact sequence defining $E$, $p(s) \in H^0(M)$ is preserved by the action of invertible global endomorphisms of $E$. In particular this implies that we have a well defined map:
\begin{align*}
\widetilde{\mathscr{U}}^d_1 &\rightarrow S^{d/2} ( C) \times J^{d/2} ( C)\\
(E,s) & \mapsto (M,p(s),L).
\end{align*}
As we can deduce easily from the exact sequence:
\begin{equation*}
0 \rightarrow H^0(M^*L) \rightarrow H^0(L) \rightarrow H^0(L \mathcal{O}_D ) \rightarrow H^1(M^*L) \rightarrow H^1(L) \rightarrow 0
\end{equation*}
where $D$ is the divisor of $p(s)$, the fiber of the previous map above $(M,p(s),L)$ is $$\mathbb{C}\P^{d/2-1} \setminus \mathbb{C}\P^{\dim H^0(L) - \dim H^0(M^*L) -1}.$$
Recall that in this case $\P H^0(L \mathcal{O}_D)$ parametrizes isomorphism classes of pairs $(E,s)$ that have $L$ as a subobject, $M$ as a quotient and the section $s$ projects to the one fixed for $M$.
We stratify $S^{d/2} ( C) \times J^{d/2} ( C)$ according to the two dimensions appearing in the fibers of the above map. Define:
\begin{equation*}
Z_{ij}^d:=\{ (M, p(s), L) \in S^{d/2} ( C) \times J^{d/2} ( C): \dim H^0(M^*L) =i,\ \dim H^0(L)=j \}.
\end{equation*}
Note that $\dim H^0(M^*L)$ can only be 1 if $L=M$ and 0 if $L \neq M$. In particular:
\begin{equation*}
Z_{0j}^d:=\{ (M, p(s), L) \in S^{d/2} ( C) \times J^{d/2} ( C): L \neq M,\ \dim H^0(L)=j \}
\end{equation*}
and
\begin{equation*}
Z_{1j}^d:=\{ (M, p(s), L) \in S^{d/2} ( C) \times J^{d/2} ( C): L=M,\ \dim H^0(L)=j \}.
\end{equation*}
Furthermore it is clear that:
\begin{equation*}
[Z_{0j}^d]+[Z_{1j}^d]=[S^{d/2}( C)][V_j^{d/2}]
\end{equation*}
and
\begin{equation*}
[Z_{1j}^d]=[V_j^{d/2}][\mathbb{C}\P^{j-1}].
\end{equation*}
From these considerations we can deduce that:
\begin{align*}
[\widetilde{\mathscr{U}}^d_1]&=\sum_{j \geq 2} [Z_{1j}^d] \left ( [\mathbb{C}\P^{d/2-1}]-[\mathbb{C}\P^{j-2}] \right ) + [Z_{11}^d][\mathbb{C}\P^{d/2-1}]+\\
&+\sum_{j \geq 1} [Z_{0j}^d] \left ( [\mathbb{C}\P^{d/2-1}]-[\mathbb{C}\P^{j-1}]\right ) + [Z_{00}^d][\mathbb{C}\P^{d/2-1}]=\\
&=\sum_{j \geq 1} [V_j^{d/2}] \left ( \frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{d/2}-1}{\L-1}-\frac{\L^{j-1}-1}{\L-1}\right )+\\
&+\sum_{j \geq 0} [V_j^{d/2}] \left ( [S^{d/2}( C)]-\frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{d/2}-1}{\L-1}-\frac{\L^{j}-1}{\L-1}\right ).
\end{align*}
Instead, since $\dim H^0( K\End E)=1+4(g-1)$ if $L \neq M$ and $2+4(g-1)$ if $L=M$ we can look at the following map as well:
\begin{align*}
\mathscr{U}^d_1&\rightarrow \widetilde{\mathscr{U}}^d_1\\
(E,\phi,s) & \mapsto (E,s)
\end{align*}
and deduce
\begin{align*}
[\mathscr{U}^d_1]&=\L^{2+4(g-1)}\sum_{j \geq 2} [Z_{1j}^d] \left ( [\mathbb{C}\P^{d/2-1}]-[\mathbb{C}\P^{j-2}] \right ) +\L^{2+4(g-1)} [Z_{11}^d][\mathbb{C}\P^{d/2-1}]+\\
&+\L^{1+4(g-1)}\sum_{j \geq 1} [Z_{0j}^d] \left ( [\mathbb{C}\P^{d/2-1}]-[\mathbb{C}\P^{j-1}]\right ) +\L^{1+4(g-1)} [Z_{00}^d][\mathbb{C}\P^{d/2-1}]=\\
&=\L^{2+4(g-1)}\sum_{j \geq 1} [V_j^{d/2}] \left ( \frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{d/2}-1}{\L-1}-\frac{\L^{j-1}-1}{\L-1}\right )+\\
&+\L^{1+4(g-1)}\sum_{j \geq 0} [V_j^{d/2}] \left ( [S^{d/2}( C)]-\frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{d/2}-1}{\L-1}-\frac{\L^{j}-1}{\L-1}\right ).
\end{align*}
\end{proof}
\end{prop}
\begin{rmk}
Note the similarities in the computation of $[\mathscr{U}^d_1]$ and $[B^{d,-}_{\bar\sigma}]$ in proposition \ref{bminus}.
\end{rmk}
\begin{prop}
\label{motiveV}
We have the following motivic equalities:
\begin{align*}
[\mathscr{V}^d_1]&=\L^{2+4(g-1)}\left ( [\Sym^2( S^{d/2} ( C) )]-\sum_{j \geq 1} [V_j^{d/2}] [\Sym^2 ( \mathbb{C}\P^{j-1})] \right ).
\end{align*}
and
\begin{align*}
[\widetilde{\mathscr{V}}^d_1]&=[\Sym^2( S^{d/2} ( C) )]-\sum_{j \geq 1} [V_j^{d/2}] [\Sym^2 ( \mathbb{C}\P^{j-1})].
\end{align*}
\begin{proof}
Note that for all $(E,s) \in \widetilde{\mathscr{V}}^d_1$ we have $\dim H^0(K \End E)=2+4(g-1)$ therefore $[\mathscr{V}^d_1]=\L^{2+4(g-1)}[\widetilde{\mathscr{V}}^d_1]$. Second, by the description of $\widetilde{\mathscr{V}}^d_1$ it is easy to see that it is isomorphic to $\Sym^2 ( S^{d/2} ( C) ) \setminus \Gamma$ where $\Gamma$ is the locus of divisors $(D_1, D_2) \in \Sym^2 ( S^{d/2} ( C) )$ such that $ \mathcal{ O }(D_1) \cong \mathcal{O} (D_2 )$. To understand $\Gamma$, we look at the map
\begin{align*}
\Sym^2 ( S^{d/2} ( C) ) & \rightarrow \Sym^2 ( J^{d/2} ( C) )\\
(D_1, D_2) &\mapsto (\mathcal{ O }(D_1), \mathcal{O} (D_2 ) ).
\end{align*}
$\Gamma$ is the inverse image of the diagonal in $\Sym^2 ( J^{d/2} ( C) )$ with respect to the previous map. In particular there is a map:
\begin{align*}
\Gamma & \rightarrow J^{d/2} ( C) \\
(D_1, D_2) &\mapsto \mathcal{ O }(D_1) = \mathcal{O} (D_2 ) .
\end{align*}
whose fiber over $L \in J^{d/2} (C )$ is $\Sym^2 ( \P H^0( L) )$. Therefore we get:
\begin{align*}
[\widetilde{\mathscr{V}}^d_1]&=[\Sym^2( S^{d/2} ( C) )]-\sum_{j \geq 1} [V_j^{d/2}] [\Sym^2 ( \mathbb{C}\P^{j-1})].
\end{align*}
\end{proof}
\end{prop}
\begin{prop}
\label{motiveW}
We have the following motivic equalities:
\begin{align*}
[\mathscr{W}^d_1]&=\L^{4g}\sum_{j\geq 2} [V_j^{d/2}] [\Gr(2,j)]
\end{align*}
and
\begin{align*}
[\widetilde{\mathscr{W}}^d_1]&=\sum_{j\geq 2} [V_j^{d/2}] [\Gr(2,j)].
\end{align*}
\begin{proof}
Since for all $(E,s) \in \widetilde{\mathscr{W}}^d_1$ we have $\dim H^0( K \End E)= 4g$ then we have $[\mathscr{W}^d_1]=\L^{4g}[\widetilde{\mathscr{W}}^d_1]$.
Furthermore, by the definition of $\widetilde{\mathscr{W}}^d_1$ we have a map:
\begin{align*}
\widetilde{\mathscr{W}}^d_1 & \rightarrow J^{d/2}\\
(L \oplus L,s) &\mapsto L
\end{align*}
and, since $s= u \oplus v$ with $u, v \in H^0(L)$ linearly independent, the fiber of such a map above $L$ is $\Gr (2, H^0(L))$, the Grassmannian of two dimensional subspaces of $H^0(L)$. The result follows.
\end{proof}
\end{prop}
\begin{prop}
We have the following motivic equalities:
\begin{align*}
[\mathscr{X}^d_1]&=\L^{2+4(g-1)} [S^{d/2}( C)] (\L^{d/2-1}-1)+\L^{1+4(g-1)} [J^{d/2} ( C)] [S^{d/2}( C)] [\mathbb{C}\P^{d/2-1}] +\\
&-\L^{1+4(g-1)} [S^{d/2}( C)] ^2 + \L^{2+4(g-1)} [\Sym^2( S^{d/2} ( C) )]
\end{align*}
and
\begin{align*}
[M_{\varepsilon,ss}^{2,d}]&=[J^{d/2} (C )][S^{d/2}( C)] [\mathbb{C}\P^{d/2-1}]-[S^{d/2}( C)]^2+[\Sym^2( S^{d/2} ( C) )].
\end{align*}
\begin{proof}
We use propositions \ref{motiveU}, \ref{motiveV} and \ref{motiveW} to compute directly:
\begin{align*}
&[\mathscr{X}^d_1]=[\mathscr{U}^d_1]+[\mathscr{V}^d_1]+[\mathscr{W}^d_1]=\\
&=\L^{2+4(g-1)}\sum_{j \geq 1} [V_j^{d/2}] \left ( \frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{d/2}-1}{\L-1}-\frac{\L^{j-1}-1}{\L-1}\right )+\\
&+\L^{1+4(g-1)}\sum_{j \geq 0} [V_j^{d/2}] \left ( [S^{d/2}( C)]-\frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{d/2}-1}{\L-1}-\frac{\L^{j}-1}{\L-1}\right )+\\
&+\L^{2+4(g-1)}\left ( [\Sym^2( S^{d/2} ( C) )]-\sum_{j \geq 1} [V_j^{d/2}] [\Sym^2 ( \mathbb{C}\P^{j-1}] \right )+\\
&+\L^{4g}\sum_{j\geq 2} [V_j^{d/2}] [\Gr(2,j)]=\\
&=\L^{2+4(g-1)}[\mathbb{C}\P^{d/2-1}]\sum_{j \geq 1} [V_j^{d/2}] [\mathbb{C}\P^{j-1}]-\L^{2+4(g-1)}\sum_{j \geq 1} [V_j^{d/2}] \left ( \frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{j-1}-1}{\L-1}\right )+\\
&+\L^{1+4(g-1)}\sum_{j \geq 0} [V_j^{d/2}] [S^{d/2}( C)] [\mathbb{C}\P^{d/2-1}]-\L^{1+4(g-1)}\sum_{j \geq 0} [V_j^{d/2}] [S^{d/2}( C)] [\mathbb{C}\P^{j-1}]+\\
&-\L^{1+4(g-1)}\sum_{j \geq 0} [V_j^{d/2}][\mathbb{C}\P^{d/2-1}][\mathbb{C}\P^{j-1}]+\L^{1+4(g-1)}\sum_{j \geq 0} [V_j^{d/2}] \left ( \frac{\L^{j}-1}{\L-1}\right )^2+\\
&+\L^{2+4(g-1)}[\Sym^2( S^{d/2} ( C) )]-\L^{2+4(g-1)}\sum_{j \geq 1} [V_j^{d/2}] [\Gr(2,j+1)]+\\
&+\L^{4g}\sum_{j\geq 2} [V_j^{d/2}] [\Gr(2,j)]=\\
&=\L^{2+4(g-1)}[\mathbb{C}\P^{d/2-1}][S^{d/2}( C)]+\L^{1+4(g-1)}[J^{d/2} ( C)] [S^{d/2}( C)] [\mathbb{C}\P^{d/2-1}]+\\
&-\L^{1+4(g-1)}[S^{d/2}( C)]^2-\L^{1+4(g-1)}[S^{d/2}( C)][\mathbb{C}\P^{d/2-1}]+\L^{2+4(g-1)}[\Sym^2( S^{d/2} ( C) )]+\\
&+\L^{1+4(g-1)} \sum_{j \geq 1} [V_j^{d/2}] \bigg ( -\L \left ( \frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{j-1}-1}{\L-1}\right ) +
\left ( \frac{\L^{j}-1}{\L-1}\right )^2 +\\
&- \L \frac{(\L^{j+1}-1)(\L^{j}-1)}{(\L^2-1)(\L-1)} + \L^3 \frac{(\L^{j}-1)(\L^{j-1}-1)}{(\L^2-1)(\L-1)} \bigg )=\\
&=\L^{1+4(g-1)}[\mathbb{C}\P^{d/2-1}][S^{d/2}( C)](\L-1)+\L^{1+4(g-1)} [J^{d/2} ( C)] [S^{d/2}( C)] [\mathbb{C}\P^{d/2-1}] +\\
&-\L^{1+4(g-1)} [S^{d/2}( C)] ^2 + \L^{2+4(g-1)} [\Sym^2( S^{d/2} ( C) )]+\\
&+\L^{1+4(g-1)} \sum_{j \geq 1} [V_j^{d/2}] [\mathbb{C}\P^{j-1}](1-\L)=\\
&=\L^{2+4(g-1)} [S^{d/2}( C)] (\L^{d/2-1}-1)+\L^{1+4(g-1)} [J^{d/2} ( C)] [S^{d/2}( C)] [\mathbb{C}\P^{d/2-1}] +\\
&-\L^{1+4(g-1)} [S^{d/2}( C)] ^2 + \L^{2+4(g-1)} [\Sym^2( S^{d/2} ( C) )].
\end{align*}
Furthermore:
\begin{align*}
[\widetilde{\mathscr{X}}^d_1]&=[\widetilde{\mathscr{U}}^d_1]+[\widetilde{\mathscr{V}}^d_1]+[\widetilde{\mathscr{W}}^d_1]=\\
&=[\mathbb{C}\P^{d/2-1}]\sum_{j \geq 1} [V_j^{d/2}] [\mathbb{C}\P^{j-1}]-\sum_{j \geq 1} [V_j^{d/2}] \left ( \frac{\L^j-1}{\L-1} \right )\left ( \frac{\L^{j-1}-1}{\L-1}\right )+\\
&+\sum_{j \geq 0} [V_j^{d/2}] [S^{d/2}( C)] [\mathbb{C}\P^{d/2-1}]-\sum_{j \geq 0} [V_j^{d/2}] [S^{d/2}( C)] [\mathbb{C}\P^{j-1}]+\\
&-\sum_{j \geq 0} [V_j^{d/2}][\mathbb{C}\P^{d/2-1}][\mathbb{C}\P^{j-1}]+\sum_{j \geq 0} [V_j^{d/2}] \left ( \frac{\L^{j}-1}{\L-1}\right )^2+\\
&+[\Sym^2( S^{d/2} ( C) )]-\sum_{j \geq 1} [V_j^{d/2}] [\Gr(2,j+1)]+\sum_{j\geq 2} [V_j^{d/2}] [\Gr(2,j)]=\\
&=[S^{d/2}( C)] [\mathbb{C}\P^{d/2-1}]+[J^{d/2} (C )][S^{d/2}( C)] [\mathbb{C}\P^{d/2-1}]+\\
&-[S^{d/2}( C)] [\mathbb{C}\P^{d/2-1}]-[S^{d/2}( C)]^2+[\Sym^2( S^{d/2} ( C) )]+\\
&+\sum_{j \geq 1} \frac{[V_j^{d/2}] (L^j-1)}{(\L^2-1)(\L-1)} \left ( (\L^j-\L^{j-1})(\L+1)-(\L^{j+1}-1)+(\L^{j-1}-1)\right )=\\
&=[J^{d/2} (C )][S^{d/2}( C)] [\mathbb{C}\P^{d/2-1}]-[S^{d/2}( C)]^2+[\Sym^2( S^{d/2} ( C) )].
\end{align*}
\end{proof}
\end{prop}
\begin{prop}
We have the following motivic equality:
\begin{align*}
[\mathscr{X}^d_2]&=[S^{d/2}( C)][J^{d/2}( C)][\mathbb{C}\P^{g-2}]\L^{1+4(g-1)}+[S^{d/2}( C)]\L^{3g-2}(\L^{2g-2}+\L^g-1).
\end{align*}
\begin{proof}
Recall the description:
\begin{equation*}
\mathscr{X}^d_2=\left \{\begin{array}{c}
(E,\phi,s) \text{ such that $E$ fits into an exact sequence}\\
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0\\
\text{with $\deg L=\deg M$, $s \in H^0(L)$ and}\\
\phi \text{ does not preserve } L
\end{array} \right \}.
\end{equation*}
We can divide the computation of the motive according to whether or not the vector bundle underlying the triple is split. In both cases it is relevant to distinguish when the quotient and the subobject are the same or different.
More precisely:
\begin{align*}
\mathscr{X}^d_2&=\left \{\begin{array}{c}
(E,\phi,s) \text{ such that $E=L\oplus L$}\\
s = u \oplus 0 \text{ with $u \in H^0(L)$}\\
\phi \text{ does not preserve}\\
\text{ the copy of $L$ containing $s$}
\end{array} \right \} \sqcup
\left \{\begin{array}{c}
(E,\phi,s) \text{ such that $E=L\oplus M$}\\
L \neq M, s \in H^0(L), \deg L = \deg M\\
\phi \text{ does not preserve $L$}
\end{array} \right \} \sqcup\\
& \sqcup
\left \{\begin{array}{c}
(E,\phi,s) \text{ with $E$ fitting}\\
\text{ in a nonsplit exact sequence}\\
0 \rightarrow L \rightarrow E \rightarrow L \rightarrow 0\\
\text{$s \in H^0(L)$ and}\\
\phi \text{ does not preserve the subobject}
\end{array} \right \} \sqcup
\left \{\begin{array}{c}
(E,\phi,s) \text{ with $E$ fitting}\\
\text{ in a nonsplit exact sequence}\\
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0\\
\text{$L \neq M$, $s \in H^0(L)$ and}\\
\phi \text{ does not preserve $L$}
\end{array} \right \}.
\end{align*}
Let us name the four strata $\mathscr{Y}^d_i$ for $i=1, \dots, 4$, in the order they appear. We have a map:
\begin{align*}
\mathscr{Y}^d_1&\rightarrow S^{d/2} ( C)\\
(L \oplus L, \phi, u \oplus 0) &\mapsto (L, u \oplus 0).
\end{align*}
The datum of $L$ and $u$ essentially recovers the pair but then $(E,s)$ will have automorphisms. More precisely $\Aut(E,s)=(\mathbb{C}^*)^2 \times \mathbb{C}$ given by the invertible upper triangular matrices. These automorphisms act by conjugation on the set of possible Higgs fields:
\begin{equation*}
\phi=\begin{pmatrix}
\psi_1 & \vartheta \\
\gamma & \psi_2
\end{pmatrix}
\end{equation*}
where the entries of the matrix are in $H^0(K)$ and $\gamma \neq 0$. A quick computation shows that the stabilizers of the action are only the scalar multiples of the identity. Therefore the motive of the fibers of:
\begin{align*}
\mathscr{Y}^d_1 &\rightarrow S^{d/2} ( C)
\end{align*}
is:
\begin{equation*}
\left [\frac{H^0(K)^3 \times (H^0(K) \setminus \{0\})}{\mathbb{C}^* \times \mathbb{C}} \right ]=\frac{\L^{3g}(\L^g-1)}{\L(\L-1)}
\end{equation*}
and
\begin{equation*}
[\mathscr{Y}^d_1]=[S^{d/2}( C)]\frac{\L^{3g}(\L^g-1)}{\L(\L-1)}.
\end{equation*}
For the computation of $[\mathscr{Y}_2^d]$ let us first note that the locus of:
\begin{equation*}
S^{d/2}( C) \times J^{d/2}( C)
\end{equation*}
defined by the triples $(L,s,M)$ where $L = M$ is the graph $\Gamma^d$ of the map:
\begin{align*}
S^{d/2}( C) &\rightarrow J^{d/2}( C)\\
(L,s) & \mapsto L
\end{align*}
and in particular it is isomorphic to $S^{d/2}( C)$.
We have a map
\begin{align*}
\mathscr{Y}^d_2&\rightarrow S^{d/2} ( C)\times J^{d/2}( C) \setminus \Gamma^d\\
(L \oplus M, \phi, s) &\mapsto (L, s, M).
\end{align*}
As before, the datum of $(L,s,M)$ determines the pair $(E,s)$ but $\Aut(E,s)=(\mathbb{C}^*)^2$. The set of Higgs fields not preserving $L$ is $H^0(K)^2 \times H^0(KM^*L) \times ( H^0(KL^*M) \setminus \{0\})$ and the action of $\Aut(E,s)$ again has only scalar multiples of the identity as stabilizers. Therefore the motive of the fibers of
\begin{align*}
\mathscr{Y}^d_2&\rightarrow S^{d/2} ( C)\times J^{d/2}( C) \setminus \Gamma^d
\end{align*}
is
\begin{equation*}
\left [\frac{H^0(K)^2 \times H^0(KM^*L) \times ( H^0(KL^*M) \setminus \{0\})}{\mathbb{C}^*} \right ]=\frac{\L^{3g-1}(\L^{g-1}-1)}{\L-1}
\end{equation*}
and so
\begin{equation*}
\left [ \mathscr{Y}_2^d \right ]=\left ([S^{d/2}( C)][J^{d/2}( C)]-[S^{d/2}( C)] \right )\frac{\L^{3g-1}(\L^{g-1}-1)}{\L-1}.
\end{equation*}
For $\mathscr{Y}_3^d$ we have a map:
\begin{align*}
\mathscr{Y}^d_3 &\rightarrow S^{d/2} ( C) \times \P\Ext^1( \mathcal{O},\mathcal{O})\\
(E,\phi,s) & \mapsto (L,s, [E] )
\end{align*}
remembering $(L,s)$ and the class of the extension. Note that the target is a trivial projective bundle of rank $g-1$ because it is the pullback to $S^{d/2} ( C)$ of the constant bundle $\P\Ext^1( \mathcal{O},\mathcal{O}) \rightarrow \{pt\}$. As we already observed earlier, $\Aut(E,s)=\mathbb{C}^* \times \mathbb{C}$ generated by nonzero multiples of the identity and multiples of $i \circ p$ where $i$ and $p$ are the inclusion and the projection in the exact sequence defining $E$. The locus of possible Higgs fields is isomorphic to $\mathbb{C}^{2+4(g-1)} \setminus \mathbb{C}^{3g-1}$ and the conjugation action of $\Aut(E,s)$ has scalar multiples of the identity as stabilizers. Therefore:
\begin{equation*}
[\mathscr{Y}^d_3]=[S^{d/2}( C)][\mathbb{C}\P^{g-1}]\left ( \L^{1+4(g-1)} - \L^{3g-2} \right ).
\end{equation*}
Finally, for $\mathscr{Y}^d_4$, there is a map:
\begin{align*}
\mathscr{Y}^d_4 &\rightarrow \mathcal{B}^d\\
(E,\phi,s) & \mapsto (L,s, M, [E] )
\end{align*}
where $\mathcal{B}^d$ is the projective bundle on $S^{d/2} ( C)\times J^{d/2}( C) \setminus \Gamma^d$ defined by the pulling back the projective bundle on $J^0( C)$ whose fiber over $A$ is $\P H^1(A)$, with respect to the map:
\begin{align*}
S^{d/2} ( C)\times J^{d/2}( C) \setminus \Gamma^d &\rightarrow J^0( C)\\
(L,s,M) & \mapsto M^*L.
\end{align*}
Once again, $\mathcal{B}^d$ will remember the pair $(E,s)$. In this case the automorphisms of the pairs are just scalar multiples of the identity and therefore we get:
\begin{align*}
[\mathscr{Y}^d_4]=\left ([S^{d/2}( C)][J^{d/2}( C)]-[S^{d/2}( C)] \right )[\mathbb{C}\P^{g-2}]\left ( \L^{1+4(g-1)} - \L^{3g-1} \right ).
\end{align*}
Putting everything together we get:
\begin{align*}
[\mathscr{X}^d_2]&=[S^{d/2}( C)]\frac{\L^{3g}(\L^g-1)}{\L(\L-1)}+\left ([S^{d/2}( C)][J^{d/2}( C)]-[S^{d/2}( C)] \right )\frac{\L^{3g-1}(\L^{g-1}-1)}{\L-1}\\
&+[S^{d/2}( C)][\mathbb{C}\P^{g-1}]\left ( \L^{1+4(g-1)} - \L^{3g-2} \right )+\\
&+\left ([S^{d/2}( C)][J^{d/2}( C)]-[S^{d/2}( C)] \right )[\mathbb{C}\P^{g-2}]\left ( \L^{1+4(g-1)} - \L^{3g-1} \right )=\\
&=[S^{d/2}( C)][\mathbb{C}\P^{g-1}]\left ( \L^{1+4(g-1)} + \L^{3g-1}- \L^{3g-2} \right )+\\
&+\left ([S^{d/2}( C)][J^{d/2}( C)]-[S^{d/2}( C)] \right )[\mathbb{C}\P^{g-2}]\L^{1+4(g-1)}=\\
&=[S^{d/2}( C)][J^{d/2}( C)][\mathbb{C}\P^{g-2}]\L^{1+4(g-1)}+[S^{d/2}( C)]\L^{3g-2}(\L^{2g-2}+\L^g-1).
\end{align*}
\end{proof}
\end{prop}
We can summarize the results of this section as follows.
\begin{theorem}
\label{moteven}
Let $d \geq 0$ be an even integer and $0 < \varepsilon < 2$. Then:
\begin{align*}
[\mathcal{M}_\varepsilon^{2,d}]&=\L^{1+4(g-1)} [M^{2,d}_\varepsilon]+\sum_{(d_1,d_2) \in I_1^e(d)} \L^{1+3(g-1)+d_2} [S^{d_1}( C)][S^{d_1-d_2+2g-2}( C)] +\\
&+\sum_{(d_1,d_2) \in I_2^e(d)}\L^{1+4(g-1)} [S^{d_2}( C)][S^{d_1-d_2+2g-2}( C)]+\\
&+(\L-1) \L^{1+4(g-1)}[\Sym^2( S^{d/2} ( C) )] +\L^{1+4(g-1)} [S^{d/2}( C)][J^{d/2}( C)][\mathbb{C}\P^{g-2}]+\\
&+[S^{d/2}( C)] \L^{3g-2} \left ( \L^{d/2+g-1}+\L^{2g-2}-1\right )+\\
&+\sum_{(d_1,d_2) \in I_1^e(d)} (\L^{4g-2}-\L^{4g-3}) [S^{d_1}( C)] \left ([S^{d_2}( C)]-[J( C)] \frac{\L^{d_2+1-g}-1}{\L-1} \right).
\end{align*}
\end{theorem}
\section{Direct computation of the motive of $\mathcal{M}^{2,d}_{\infty}$}
Using the content of the previous sections, we can compute the motive of $\mathcal{M}^{2,d}_{\infty}$ by simply starting from the motive of $\mathcal{M}^{2,d}_{\varepsilon}$ and then add and subtract the motive of the flip loci. In other words we have all the ingredients to compute $[\mathcal{M}^{2,d}_{\infty}]$.
It is interesting to note that it is possible to proceed in the reverse direction, i.e. first compute $[\mathcal{M}^{2,d}_{\infty}]$ directly, then add the motives of the flip loci and ultimately deduce $[\mathcal{M}^{2,d}_{\varepsilon}]$. Here we outline the strategy for the computation of $[\mathcal{M}^{2,d}_{\infty}]$.
Recall that $[\mathcal{M}^{2,d}_{\infty}]$ contains only split type 2 attracting sets. These have the form:
\begin{equation*}
F_{(1,1),\infty}^{(d_1,d_2),2+}=
\left \{
\begin{array}{c}
(E,\phi,s) \text{ $\sigma$-stable where } E \text{ is defined by}\\
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0\\
s \in H^0(L), \deg L =d_2, \deg M = d_1\\
\text{and } \phi \text{ does not preserve } L
\end{array} \right \}
\end{equation*}
with $d_1$ and $d_2$ satisfying:
\begin{align*}
&d/2 +1-g\leq d_1 \leq d \\
& 0 \leq d_2 \leq d/2+g-1\\
& 0 \leq d_1-d_2+2g-2 \leq 2g-2+d \text{ same parity as $d$}.
\end{align*}
Observe that for $d_2>d_1$, which is equivalent to $d_1<d/2$, $F_{(1,1),\infty}^{(d_1,d_2),2+}=F_{(1,1),\varepsilon}^{(d_1,d_2),2+}$ lies entirely in the smooth part of the moduli space (see proposition \ref{deffixed}) and therefore the motive can be computed using \cite[Theorem 4.1]{bialynicki1973some}. If $d$ is even and $d_1=d_2=d/2$ we know that $F_{(1,1),\infty}^{(d/2,d/2),2+}= \mathscr{X}^d_2$.
To compute the remaining motives we assume that $d_1>d_2$. Similar to the computation of the motive of the type 1 attracting sets, we can further decompose an attracting set according to the properties of the underlying pair. In fact, we know that for pairs of the form:
\begin{equation*}
0 \rightarrow L \rightarrow E \rightarrow M \rightarrow 0
\end{equation*}
with $\deg M-\deg L= k >0$, being in $\P W^{d,+}_k$ is equivalent to the extension being non-split. Therefore we can distinguish between the cases where the underlying pair is non-split and when $E=L\oplus M$.
\begin{defn}
Let $(d_1,d_2)$ satisfy $d_1+d_2=d$, the three inequalities above and $d_1> d_2$. Let us denote by $NSPF_{(1,1),\infty}^{(d_1,d_2),2+}$ the locus of $F_{(1,1),\infty}^{(d_1,d_2),2+}$ where the underlying pair lies in $\P W^{d,+}_{d_1-d_2}$ and by $SPF_{(1,1),\infty}^{(d_1,d_2),2+}$ the locus where the pair is split.
\end{defn}
It is clear from the definitions that we have:
\begin{equation*}
NSPF_{(1,1),\infty}^{(d_1,d_2),2+}=B_{d_1-d_2}^{d,+} \cap F_{(1,1),\infty}^{(d_1,d_2),2+}
\end{equation*}
and:
\begin{equation*}
F_{(1,1),\infty}^{(d_1,d_2),2+} =B_{d_1-d_2}^{d,+} \cap F_{(1,1),\infty}^{(d_1,d_2),2+} \sqcup SPF_{(1,1),\infty}^{(d_1,d_2),2+}.
\end{equation*}
We can also prove the following:
\begin{prop}
\begin{equation*}
B_{d_1-d_2}^{d,+} = NSPF_{(1,1),\infty}^{(d_1,d_2),2+} \sqcup S \mathcal{W}^{d,+}_{d_1-d_2}=B_{d_1-d_2}^{d,+} \cap F_{(1,1),\infty}^{(d_1,d_2),2+} \sqcup S \mathcal{W}^{d,+}_{d_1-d_2}.
\end{equation*}
\begin{proof}
The second equality follows from the previous remarks. Recall that $(E,\phi,s) \in B_{d_1-d_2}^{d,+}$ if and only if $(E,s) \in \P W^{d,+}_{d_1-d_2}$. Since by definition $(E,\phi,s) \in NSPF_{(1,1),\infty}^{(d_1,d_2),2+}$ if and only if $(E,s) \in F_{(1,1),\infty}^{(d_1,d_2),2+}$ and the extension for $(E,s)$ is nonsplit we immediately get that if $(E,\phi,s) \in NSPF_{(1,1),\infty}^{(d_1,d_2),2+}$ then $(E,s) \in \P W^{d,+}_{d_1-d_2}$ and so $(E,\phi,s) \in B_{d_1-d_2}^{d,+}$.
This proves $NSPF_{(1,1),\infty}^{(d_1,d_2),2+} \subset B_{d_1-d_2}^{d,+}$. We already observed in proposition \ref{weirdexpl1} that $S \mathcal{W}^{d,+}_{d_1-d_2} \subset B_{d_1-d_2}^{d,+}$. To conclude it is enough to observe that if $(E,\phi,s) \in NSPF_{(1,1),\infty}^{(d_1,d_2),2+}$ or $(E,\phi,s) \in S \mathcal{W}^{d,+}_{d_1-d_2}$ then $(E,s) \in \P W^{d,+}_{d_1-d_2}$ but in the first case the subobject containing the section is not preserved by $\phi$, since $NSPF_{(1,1),\infty}^{(d_1,d_2),2+} \subset F_{(1,1),\infty}^{(d_1,d_2),2+}$, while in the second case the subobject containing the section is preserved by $\phi$, since $S \mathcal{W}^{d,+}_{d_1-d_2} \subset \mathcal{W}^{d,+}_{d_1-d_2}$. This proves both the reverse inclusion and the fact that the union is disjoint.
\end{proof}
\end{prop}
Therefore we can write the relation:
\begin{equation*}
[F_{(1,1),\sigma}^{(d_1,d_2),2+}]=[B_{d_1-d_2}^{d,+}]-[S \mathcal{W}^{d,+}_{d_1-d_2}]+[SPF_{(1,1),\sigma}^{(d_1,d_2),2+}].
\end{equation*}
\begin{rmk}
Note that it is particularly hard to attempt the computation of $[B_{d_1-d_2}^{d,+}]$ because for $(E,\phi,s) \in B_{d_1-d_2}^{d,+}$ we get a presentation of the pair $(E,s)$ as an extension of line bundles for which the subobject is not destabilizing. This makes the problem particularly hard because we need to know the dimension of $H^0( K \End E)$ for these triples and for that we need to know whether or not $E$ is stable and if not who is the maximal destabilizing subbundle, all of which is not clear from the given presentation.
There is however a way around this. In fact the $B_{d_1-d_2}^{d,+}$ and $B_{d_1-d_2}^{d,-}$ are closely related to the wall-crossing of Bradlow pairs. In fact we have the following relations:
\begin{equation*}
0=[F_{(2),\varepsilon}^{(d),1+}]+\sum_{k=0}^{(d-1)/2} \left ( [B_{2k+1}^{d,-}] -[B_{2k+1}^{d,+}] \right )
\end{equation*}
for odd $d$, and
\begin{equation*}
0=[F_{(2),\varepsilon}^{(d),1+}]+\sum_{k=0}^{d/2} \left ( [B_{2k}^{d,-}] -[B_{2k}^{d,+}] \right )
\end{equation*}
for even $d$. Both these relations imply that if we only wish to compute the sum of all of the $[B_{d_1-d_2}^{d,+}]$ then it suffices to $[F_{(2),\varepsilon}^{(d),1+}]$, which we computed in the previous sections, and the sum of the $[B_{d_1-d_2}^{d,-}]$ which can be computed using proposition \ref{bminus}.
\end{rmk}
In the following propositions we compute the remaining motives.
\begin{prop}
Let $\bar\sigma$ be a critical value. The following motivic equalities hold:
\begin{align*}
[NS\mathcal{W}_{\bar\sigma}^{d,+}]&=\L^{2g} [S^{(d-\bar\sigma)/2} ( C)] [S^{2g-2-\bar\sigma} ( C)] \text{ if } \bar\sigma < 2g-2 \text{ or } 0 \text{ otherwise}\\
[S\mathcal{W}_{\bar\sigma}^{d,+}]&=\L^{2g} \cdot [\mathbb{C}\P^{2g-3}]\cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot [J( C)]+\L^{3g-2} \cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot [S^{\bar\sigma}( C)]+\\
&-\L^{2g} [S^{(d-\bar\sigma)/2} ( C)] [S^{2g-2-\bar\sigma} ( C)].
\end{align*}
\begin{proof}
We already proved that there is a map
$$NS\mathcal{W}_{\bar\sigma}^{d,+} \rightarrow X_{\bar\sigma}$$
whose fiber over $(L,s,M,\psi_1,\psi_2)$ is $\P H^0(KLM^*)$. Let us define the strata
\begin{equation*}
U_{\bar\sigma,i}^+:=\{(L,s,M) \in S^{(d-\bar\sigma)/2} ( C) \times J^{(d+\bar\sigma)/2}( C) | \dim H^0(KLM^*) = i\}.
\end{equation*}
Consider also the map:
\begin{align*}
S^{(d-\bar\sigma)/2} ( C) \times J^{(d+\bar\sigma)/2}( C) &\rightarrow S^{(d-\bar\sigma)/2}( C) \times J^{2g-2-\bar\sigma}( C)\\
(L,s,M) & \mapsto (L,s, KLM^*)
\end{align*}
which is clearly an isomorphism. Under this isomorphism $U_{\bar\sigma,i}^+$ corresponds to
$$V^{2g-2-\bar\sigma}_{i} \times S^{(d-\bar\sigma)/2} ( C).$$
Therefore we deduce that:
\begin{equation*}
[U_{\bar\sigma,i}^+]=[V^{2g-2-\bar\sigma}_{i} ]\cdot [S^{(d-\bar\sigma)/2} ( C)].
\end{equation*}
So we can compute:
\begin{align*}
[NS\mathcal{W}_{\bar\sigma}^{d,+}]&=\L^{2g} [S^{(d-\bar\sigma)/2} ( C)] \sum_{i=0}^{2g-2-\bar\sigma} [ V^{2g-2-\bar\sigma}_{i} ] \cdot [\mathbb{C} \P^{i-1}]=\\
&=\L^{2g} [S^{(d-\bar\sigma)/2} ( C)] [S^{2g-2-\bar\sigma} ( C)].
\end{align*}
Since $\mathcal{W}_{\bar\sigma}^{d,+}=S\mathcal{W}_{\bar\sigma}^{d,+} \sqcup NS\mathcal{W}_{\bar\sigma}^{d,+}$ we also deduce:
\begin{align*}
[S\mathcal{W}_{\bar\sigma}^{d,+}]=[\mathcal{W}_{\bar\sigma}^{d,+}]-&[NS\mathcal{W}_{\bar\sigma}^{d,+}]=\L^{2g} \cdot [\mathbb{C}\P^{2g-3}]\cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot [J( C)]+\\
&+\L^{3g-2} \cdot [S^{(d-{\bar\sigma})/2}( C)] \cdot [S^{\bar\sigma}( C)]-\L^{2g} [S^{(d-\bar\sigma)/2} ( C)] [S^{2g-2-\bar\sigma} ( C)].
\end{align*}
\end{proof}
\end{prop}
\begin{rmk}
Note that in the previous proposition we used the motive of $\mathcal{W}_{\bar\sigma}^{d,+}$ to compute $[S\mathcal{W}_{\bar\sigma}^{d,+}]$. Observe that it not strictly necessary and we can compute $[S\mathcal{W}_{\bar\sigma}^{d,+}]$ directly by observing that the restriction of $\mathcal{W}_{\bar\sigma}^{d,+} \rightarrow X_{\bar\sigma}^d$ to $S\mathcal{W}_{\bar\sigma}^{d,+}$ has fibers that can be well understood. Namely the fiber over $(L,s,M,\psi_1,\psi_2)$ is $\P\H^1((M,\psi_2),(L,\psi_1)) \setminus \P H^0(KLM^*)$ the dimension of which can be computed by considerations similar to those in proposition \ref{wplus}.
\end{rmk}
\begin{prop}
We have the identity:
\begin{equation*}
[SPF_{(1,1),\sigma}^{(d_1,d_2),2+}]=\L^{2g} [S^{d_2} ( C)] [\mathbb{C} \P^{d_1-d_2+g-2}] \left ( (\L-1) [S^{d_2-d_1+2g-2}( C)] + [J (C )]\right ).
\end{equation*}
\begin{proof}
Recall that $SPF_{(1,1),\sigma}^{(d_1,d_2),2+}$ consists of those pairs in the form $(L\oplus M,s,\phi)$ where $\deg L=d_2$ and $\deg M=d_1$ have the appropriate degrees, $s \in H^0(L)$ and $\phi$ does not preserve $L$. First of all, note that the moduli space of such split pairs is
\begin{equation*}
S^{d_2} ( C) \times J^{d_1} (C ).
\end{equation*}
Second, we can always write the Higgs field into matrix form:
\begin{equation*}
\phi =
\begin{pmatrix}
\psi_1 & \xi\\
\theta & \psi_2
\end{pmatrix}
\end{equation*}
where
\begin{equation*}
\theta : L \rightarrow M \otimes K
\end{equation*}
has to be nonzero, and
\begin{equation*}
\xi : M \rightarrow L \otimes K.
\end{equation*}
Since the endomorphisms of the underlying pair are just diagonal (possibly non scalar) then we can see that $(L \oplus M, s , \phi)$ and $(L \oplus M, s , \phi')$ are isomorphic if and only if there exists $\lambda \in \mathbb{C}^*$ such that:
\begin{equation*}
\phi' =
\begin{pmatrix}
\psi'_1 & \xi'\\
\theta' & \psi'_2
\end{pmatrix}=
\begin{pmatrix}
\psi_1 & \lambda \xi\\
\lambda^{-1} \theta & \psi_2
\end{pmatrix}.
\end{equation*}
Let's consider the map:
\begin{equation*}
SPF_{(1,1),\sigma}^{(d_1,d_2),2+} \rightarrow S^{d_2} ( C) \times J^{d_1} (C ) \times H^0(K)^2
\end{equation*}
forgetting the off-diagonal parts of the Higgs field. Then, by what we said, the fiber of such a map over $(L,s,M)$ will be:
\begin{equation*}
\quotient{\left (H^0(KML^*) \setminus \{0\} \right ) \times H^0(KLM^*)}{ \mathbb{C}^*}.
\end{equation*}
This allows us to compute the motive of $SPF_{(1,1),\sigma}^{(d_1,d_2),2+}$ in a familiar way by stratifying $S^{d_2} ( C) \times J^{d_1} (C )$ according to the dimension of $H^0(KLM^*)$. Note in fact that the dimension of $H^0(KML^*)$ is constantly equal to $g-1+d_1-d_2$. Call $V_i$ the stratum of $J^{d_1-d_2+2g-2}$ where the dimension of the global sections of the line bundle is equal to $i$.
Then we have:
\begin{align*}
[SPF_{(1,1),\sigma}^{(d_1,d_2),2+}]=&\sum_i \left (\L^{2g} [S^{d_2} ( C)] [V_i] \L^i \frac{\L^{d_1-d_2+g-1}-1}{\L-1} \right )=\\
=&\L^{2g} [S^{d_2} ( C)] [\mathbb{C} \P^{d_1-d_2+g-2}] \left ( (\L-1) \sum_i [V_i] [\mathbb{C}\P^{i-1}] + \sum_i [V_i]\right )=\\
=&\L^{2g} [S^{d_2} ( C)] [\mathbb{C} \P^{d_1-d_2+g-2}] \left ( (\L-1) [S^{d_2-d_1+2g-2}( C)] + [J (C )]\right ).
\end{align*}
\end{proof}
\end{prop}
\chapter{Bradlow-Higgs triples with poles}
In this chapter we will introduce a variation to Bradlow-Higgs triples that is the analogue to the one in section \ref{hbpoles} for Higgs bundles. We will see that the results of the thesis about Bradlow-Higgs triples carry over with almost no modifications.
Let's start with a definition.
\begin{defn}[Bradlow-Higgs triples with poles and $\sigma$-stability]
Let $\gamma > 0$ be an integer and $P$ a point on the smooth projective curve $C$. A triple $(E,\phi,s)$ is said to be a \emph{Bradlow-Higgs $\gamma$-triple} if $(E,\phi)$ is a Higgs $\gamma$-bundle (as in definition \ref{gammahiggs}) and $s \in H^0(C,E)$ is a nonzero section of the underlying vector bundle $E$.
Let $\sigma$ be a positive real number. We say $(E,\phi,s)$ is \emph{$\sigma$-(semi)stable} if, for all proper $\phi$-invariant subbundles $F \subset E$, we have:
\begin{align*}
\frac{\deg(F)}{\rk(F)} \operatornamewithlimits{<}_{(=)} \frac{\deg(E)+\sigma}{\rk(E)} \qquad & \text{if } s \notin H^0(C, F)\\
\frac{\deg(F)+\sigma}{\rk(F)} \operatornamewithlimits{<}_{(=)} \frac{\deg(E)+\sigma}{\rk(E)} \qquad & \text{if } s \in H^0(C, F)
\end{align*}
We denote by $\mathcal{M}_\sigma^{r,d}(\gamma)$ the corresponding moduli spaces.\
\end{defn}
Many of the results from the previous chapters are still valid with no or small changes. We will discuss the differences in the following.
The BNR correspondence explained in section \ref{secbnr} relates Bradlow-Higgs $\gamma$-triples to coherent systems on the surface $\P X(\gamma)=\P (\mathcal{O}_C \oplus K(\gamma P))$ which is the compactification of the total space of $K (\gamma P)$.
We define a $\gamma$-Hitchin base
$$\mathcal{A}^r(\gamma)=\oplus_{i=0}^{r}H^0(K(\gamma P)^i)$$
which parametrizes spectral curves in $\P X(\gamma)$ and has dimension $r^2(g-1)+ \gamma r(r+1)/2$. Accordingly, there are also modified Hitchin maps:
\begin{align*}
\chi^{r,d}_\sigma(\gamma): \mathcal{M}_\sigma^{r,d}(\gamma) & \rightarrow \mathcal{A}^r(\gamma)\\
(E,\phi,s) & \mapsto \text{char poly}(\phi).
\end{align*}
We can also define a $\mathbb{C}^*$-action on the $\mathcal{M}_\sigma^{r,d}(\gamma)$ by scaling the Higgs field in the usual way. Also, the $\gamma$-Hitchin maps satisfy the same properties as the original one, namely they are equivariant with respect to the $\mathbb{C}^*$-action and they are proper.
The fixed point loci and attracting sets have a similar structure and it is possible to compute the dimension of the positive, negative and zero weight part of the $\mathbb{C}^*$-action restricted to the Zariski tangent space of a fixed point.
\begin{defn}
Given a partition $\underline{r}=(r_1, \dots ,r_m)$ of $r$, a partition $\underline{d}=(d_1, \dots, d_m)$ and $1 \leq k\leq m$ we denote by
\begin{equation*}
F_{\underline{r},\sigma}^{\underline{d},k}(\gamma)=\left \{
\begin{array}{c}
(E,\phi,s)\in \mathcal{M}_\sigma^{r,d}(\gamma) \text{ such that } E=E_1 \oplus \dots \oplus E_m\\
\text{with } \deg E_i=d_i, \rk{E_i}=r_i\\
\phi(E_i) \subseteq E_{i-1} \otimes K(\gamma P), \phi(E_1)=0\\
s \in H^0(E_k)
\end{array} \right \}
\end{equation*}
\end{defn}
\begin{defn}
Let us fix $F_{\underline{r},\sigma}^{\underline{d},k}(\gamma)$. We define two locally closed subsets by:
\begin{align*}
F_{\underline{r},\sigma}^{\underline{d},k+}(\gamma)&=\{(E,\phi,s) \in \mathcal{M}_\sigma^{r,d}(\gamma)\text{ such that } \lim_{\lambda \rightarrow 0}(E,\phi,s) \in F_{\underline{r},\sigma}^{\underline{d},k}(\gamma)\}\\
F_{\underline{r},\sigma}^{\underline{d},k-}(\gamma)&=\{(E,\phi,s) \in \mathcal{M}_\sigma^{r,d}(\gamma)\text{ such that } \lim_{\lambda \rightarrow \infty}(E,\phi,s) \in F_{\underline{r},\sigma}^{\underline{d},k}(\gamma)\}
\end{align*}
\end{defn}
Then, similar to the situation for $\mathcal{M}^{r,d}$ described in section \ref{semipr}, we have:
\begin{prop}
Let $(E,\phi,s)$ be a point in $F_{\underline{r},\sigma}^{\underline{d},k}(\gamma)$, for $\sigma$ not a critical value. With the previous notations, the weight $0$ part of the $\mathbb{C}^*$-action on $T_{(E,\phi,s)} \mathcal{M}_\sigma^{r,d}(\gamma)$ is given by the first hypercohomology of the complex:
\begin{equation*}
\bigoplus_{i=1}^{m} \Hom(E_i,E_i) \rightarrow \bigoplus_{i=2}^{m} \Hom(E_i,E_{i-1} \otimes K(\gamma)) \oplus E_k,
\end{equation*}
the positive weight part is the first hypercohomology of:
\begin{equation*}
\bigoplus_{i<j} \Hom(E_i,E_j) \rightarrow \bigoplus_{i\leq j} \Hom(E_i,E_{j} \otimes K(\gamma)) \oplus \bigoplus_{i>k} E_i
\end{equation*}
while the negative part is the first hypercohomology of:
\begin{equation*}
\bigoplus_{i>j}^{m} \Hom(E_i,E_j) \rightarrow \bigoplus_{i>j+1}^{m} \Hom(E_i,E_{j} \otimes K(\gamma)) \oplus \bigoplus_{i<k} E_i
\end{equation*}
\end{prop}
The remarks about extremal values of $\sigma$ and the $U$-filtration are still valid once we change $K$ with $K(\gamma P)$. There are new estimates on the degrees of the $U$-filtration. Namely
$$\deg U_i \geq i(i-1)(1-g)-\gamma \frac{i(i-1)}{2}.$$
In particular theorem \ref{extremal} is modified as follows:
\begin{theorem}
\begin{itemize}
\item[(i)] Assume that $\sigma$ is very close to $0$, then $\sigma$-stability for a triple $(E,\phi,s)$ implies the semistability of $(E,\phi)$ and so we have an Abel-Jacobi map:
\begin{equation*}
AJ(\gamma): \mathcal{M}_\sigma^{r,d}(\gamma) \rightarrow \mathcal{M}^{r,d}(\gamma).
\end{equation*}
For $d$ large enough (e.g. $d > r(2g-1)+(r-1)^2(2g-2+\gamma) $) for any semistable Higgs $\gamma$-bundle $(E,\phi)$ we have $H^1(E)=0$ and therefore $AJ(\gamma)$ is a projective bundle over the stable part of $\mathcal{M}^{r,d}(\gamma)$.\\
\item[(ii)] For $\sigma>(r-1)d+r(r-1)(r-2)(g-1)+\gamma r(r-1)(r-2)/2$ and a $\sigma$-stable triple $(E,\phi,s)$ corresponding to a pair $(\mathcal{F},s)$ the following three equivalent conditions are realized:
\begin{itemize}
\item there are no $\phi$-invariant subbundles of $E$ which contain the section
\item $s, \phi(s), \dots, \phi^{r-1}(s)$ generically generate $E$
\item $s$ as a map $\mathcal{O}_{\P X(\gamma)} \rightarrow \mathcal{F}$ has zero dimensional cokernel.
\end{itemize}
\end{itemize}
\end{theorem}
Also, theorem \ref{relhilb} modifies as follows.
\begin{theorem}
The Hitchin map:
\begin{equation*}
\chi^{r,d}_\infty(\gamma): \mathcal{M}_\infty^{r,d}(\gamma) \rightarrow \mathcal{A}^r(\gamma)
\end{equation*}
is the relative Hilbert scheme of $d+r(r-1)(g-1)+\gamma r(r-1)/2$ points over the family of spectral curves $\mathcal{A}^r(\gamma)$.
\end{theorem}
Deformation theory deserves a special discussion, as increasing $\gamma$ will make it easier for a Bradlow-Higgs $\gamma$-triple to be a smooth point of the corresponding moduli space.
We still have a complex that computes the dimension of the Zariski tangent space.
\begin{theorem}
Let $(E,\phi,s)$ be a $\sigma$-stable Bradlow-Higgs $\gamma$-triple, then the tangent space at $(E,\phi,s)$ is given by the first cohomology $\H^1(E,\phi,s)$ of the complex:
\begin{equation*}
C^0(\End E)\rightarrow C^1(\End E) \oplus C^0(K(\gamma P)\End E) \oplus C^0(E) \rightarrow C^1(E) \oplus C^1(K(\gamma P) \End E)
\end{equation*}
where the first map is $$p(k)=(dk, [k,\phi],k \cdot s)$$ and the second one is $$q(\tau, \nu, \gamma)=(\tau \cdot s+d \gamma, [\tau,\phi]+d \nu).$$
Furthermore, $\H^0(E,\phi,s)=0$.
The same result can be obtained from the hypercohomology of the complex:
\begin{align*}
\End E &\rightarrow K(\gamma P) \End E \oplus E\\
f & \mapsto ([f,\phi],f(s))
\end{align*}
from which we can also deduce the long exact sequence
\begin{align*}
0 \rightarrow H^0(\End E) \rightarrow H^0(K(\gamma P) \End E\oplus E) \rightarrow T_{(E,\phi,s)} \rightarrow \\ \nonumber
\rightarrow H^1(\End E) \rightarrow H^1(K(\gamma P) \End E\oplus E) \rightarrow \H^2 \rightarrow 0
\end{align*}
\end{theorem}
The proof of the proposition is completely analogous to proposition \ref{deform}. This also allows us to compute the dimension by noting that
$$\dim T_{(E,\phi,s)} = d+ 2 r^2(g-1)+r(1-g) + r^2 \gamma + \dim \H^2 (E,\phi,s).$$
As in the case of Higgs bundles, we generically have $\H^2(E,\phi,s)=\H^2([\cdot,\phi])=0$. Note that specializing to $\gamma =0$ does not give the formula for the dimension of the moduli space of Bradlow-Higgs triples (compare with \cite[proposition 7.1]{nitsure1991moduli}).
Here $\H^2(E,\phi,s)$ is characterized in a slightly different way.
\begin{prop}
Let $(E,\phi,s)$ be a $\sigma$-stable $\gamma$-triple. Then $(\H^2(E,\phi,s))^*$ is the kernel of the following map:
\begin{align*}
H^0(\End E (-\gamma P) )\oplus H^0(K E^*) & \rightarrow H^0(K \End E)\\
(\alpha, \beta) &\mapsto [\alpha,\phi]+\beta \otimes s.
\end{align*}
In particular, $\dim \H^2(E,\phi,s) \geq \dim \H^2([\cdot, \phi])$.
\end{prop}
This is particularly important to observe because in general $H^0(\End E (-\gamma P) )$ is smaller than $H^0(\End E )$.
Let us discuss the case of rank 2 more in detail. We have the analogue of proposition \ref{negdegrk2}, with the same proof.
\begin{prop}
Let $d < 0$. Then $\mathcal{M}_\infty^{2,d}(\gamma)=\mathcal{M}_\varepsilon^{2,d}(\gamma)$ and both are non-empty iff $d \geq 2-2g-\gamma$.
\end{prop}
We can classify the fixed points as in theorem \ref{fpr2}.
\begin{theorem}
Let $d\geq 2-2g-\gamma$ be an integer and $\sigma>0$ different from a critical value. Then we can classify the components of the fixed point locus of $\mathcal{M}_\sigma^{2,d}(\gamma)$ as follows:
\begin{itemize}
\item[(i)] if $d \geq 0$ and $\sigma < d$ then one of the components of the fixed points for the $\mathbb{C}^*$-action is $F_{(2),\sigma}^{(d),1}(\gamma)=M_\sigma^{2,d}$, i.e. the moduli space of $\sigma$-stable Bradlow pairs embedded as triples with zero Higgs field. If $d <0$ then there are no $\sigma$-stable Bradlow pairs and so this component is empty.
\item[(ii)] if there exists and integer $m$ such that $\max\{0,(d-\gamma)/2 +1-g\} \leq m < \frac{d-\sigma}{2}$, then there exist components:
\begin{align*}
F_{(1,1),\sigma}^{(d_1,d_2),1}(\gamma)&=\left \{
\begin{array}{c}
(E,\phi,s)\in \mathcal{M}_\sigma^{2,d}(\gamma) \text{ such that } E=E_1 \oplus E_2\\
\text{with } \deg E_i=d_i, \rk{E_i}=1\\
\phi(E_2) \subseteq E_{1} \otimes K, \phi(E_1)=0\\
s \in H^0(E_1)
\end{array} \right \} \cong\\
&\cong S^{d_1}( C) \times S^{d_1-d_2+2g-2+\gamma}( C).
\end{align*}
Here $d_1$ and $d_2$ are integers satisfying $d_1+d_2=d$ and one of the following equivalent inequalities:
\begin{align*}
&\max\{0,(d-\gamma)/2+1-g\} \leq d_1 < \frac{d-\sigma}{2} \\
&\frac{d+\sigma}{2} < d_2 \leq \min\{d,(d+\gamma)/2+g-1\}\\
&\max\{0,2g-2+\gamma-d\}\leq d_1-d_2+2g-2+\gamma < 2g-2+\gamma-\sigma \text{ same parity as $d$}.
\end{align*}
\item[(iii)] There exist components:
\begin{align*}
F_{(1,1),\sigma}^{(d_1,d_2),2}(\gamma)&=\left \{
\begin{array}{c}
(E,\phi,s)\in \mathcal{M}_\sigma^{2,d}(\gamma) \text{ such that } E=E_1 \oplus E_2\\
\text{with } \deg E_i=d_i, \rk{E_i}=1\\
\phi(E_2) \subseteq E_{1} \otimes K, \phi(E_1)=0\\
s \in H^0(E_2)
\end{array} \right \} \cong\\
&\cong S^{d_2}( C) \times S^{d_1-d_2+2g-2+\gamma}( C).
\end{align*}
Here $d_1$ and $d_2$ are integers satisfying $d_1+d_2=d$ and one of the following equivalent inequalities:
\begin{align*}
&(d-\gamma)/2 +1-g \leq d_1 < \min \left \{\frac{d+\sigma}{2},d+1 \right \} \\
&\max\left \{-1,\frac{d-\sigma}{2} \right \}< d_2 \leq (d+\gamma)/2+g-1\\
& 0\leq d_1-d_2+2g-2+\gamma < 2g-2+\gamma+\min\{\sigma,d+1\} \text{ same parity as $d$}.
\end{align*}
\end{itemize}
\end{theorem}
We can now discuss the dimension of the Zariski tangent space at the fixed point. Recall that, since limits as $\lambda \rightarrow 0$ for $\mathbb{C}^*$ acting on any $\gamma$-triple always exist, if the dimension of the Zariski tangent space at a fixed point $(E,\phi,s)$ is already minimal, i.e. the fixed point is smooth, then all the triples whose limit as $\lambda \rightarrow 0$ is $(E,\phi,s)$ will also be smooth points. In the following always assume that $\sigma$ is different from a critical value.
Consider a non-split fixed point $(E,0,s) \in \mathcal{M}_\sigma^{2,d}(\gamma)$. Pick $\alpha \in H^0( \End E(-\gamma P))$ and $\beta \in H^0( KE^*)$. Certainly if $(\alpha, \beta) \in \H^2(E,\phi,s)^*$ then $\beta =0$. In particular $\H^2(E,\phi,s)^*=H^0( \End E(-\gamma P))$. Since $(E,s) \in M_\sigma^{2,d}$ is a $\sigma$-stable Bradlow pair, either $E$ is already semistable and then $H^0( \End E(-\gamma P))=0$ or $E$ fits into an extension:
\begin{equation*}
0 \rightarrow M \rightarrow E \rightarrow L \rightarrow 0
\end{equation*}
where $\deg M > \deg L$ and $\deg M < (d+\sigma)/2$. In this case
$$H^0( \End E(-\gamma P))=H^0(L^*M(-\gamma P))$$
regardless of whether or not the extension is split (see proof of lemma \ref{unstableext}). Since $\deg L^*M(-\gamma P) = \sigma - \gamma$, if $\gamma > \sigma$ then all these points are smooth.
A type 1 split fixed point is of the form $E=E_1 \oplus E_2$ with $\phi: E_2 \rightarrow E_1 K(\gamma P)$ and $s \in H^0(E_1)$. For these points we always have $\deg E_2 > \deg E_1$ and so
$$H^0( \End E(-\gamma P))=H^0(E_1^*E_2(-\gamma P)).$$
Pick $\alpha \in H^0( \End E(-\gamma P))$ and $\beta \in H^0( KE^*)$ then:
\begin{equation*}
[\alpha, \phi]=
\begin{pmatrix}
- \phi c & 0 \\
0 & c \phi
\end{pmatrix}
\end{equation*}
for $c \in H^0(E_1^*E_2(-\gamma P))$ and
\begin{equation*}
\beta \otimes s=
\begin{pmatrix}
\beta_1 s & \beta_2 s \\
0 & 0
\end{pmatrix}
\end{equation*}
for $\beta_i \in H^0(K E_i^*)$. If $(\alpha, \beta) \in \H^2(E,\phi,s)^*$ then $\beta=0$ and $\alpha=0$ and so every type 1 split fixed point is smooth, regardless of $\sigma$.
Finally, a type 2 split fixed point is of the form $E=E_1 \oplus E_2$ with $\phi: E_2 \rightarrow E_1 K(\gamma P)$ and $s \in H^0(E_2)$. Here $H^0( \End E(-\gamma P))=H^0(E_1^*E_2(-\gamma P)) \oplus H^0(E_2^*E_1(-\gamma P)) $. Pick $\alpha \in H^0( \End E(-\gamma P))$ and $\beta \in H^0( KE^*)$ then:
\begin{equation*}
[\alpha, \phi]=
\begin{pmatrix}
- \phi c & 0 \\
0 & c \phi
\end{pmatrix}
\end{equation*}
for $c \in H^0(E_1^*E_2(-\gamma P))$, $b \in H^0(E_2^*E_1(-\gamma P))$ and
\begin{equation*}
\beta \otimes s=
\begin{pmatrix}
0 & 0\\
\beta_1 s & \beta_2 s
\end{pmatrix}
\end{equation*}
for $\beta_i \in H^0(K E_i^*)$. If $(\alpha, \beta) \in \H^2(E,\phi,s)^*$ then $\beta=0$ and
$$\H^2(E,\phi,s)^*=H^0(E_2^*E_1(-\gamma P)).$$
Since $\deg E_2^*E_1(-\gamma P)=\deg E_1-\deg E_2 -\gamma < -\gamma + \min \{\sigma, d+1\}$, if $\gamma \geq \min\{\sigma ,d+1\}$ then the fixed point is certainly smooth.
We can summarize the main consequences of the above discussion in the following proposition.
\begin{cor}
$\mathcal{M}_\varepsilon^{2,d}(\gamma)$ is smooth for all $\gamma \geq 1$. In particular $\mathcal{M}_\varepsilon^{2,d}(\gamma)$ is always semiprojective.
If $\gamma > d$ then $\mathcal{M}_\sigma^{2,d}(\gamma)$ is smooth regardless of $\sigma$ as long as it is different from a critical value. In this case then $\mathcal{M}_\sigma^{2,d}(\gamma)$ is semiprojective.
\end{cor}
A further consequence of this is the following theorem.
\begin{theorem}
Let $\gamma \geq 1$ be an integer. Then if $d <0$ we have:
\begin{align*}
[\mathcal{M}_{\varepsilon}^{2,d}(\gamma)]=\sum_{(d_1,d_2) \in I_{2,\varepsilon}(d,\gamma)} \L^{4g-4+3\gamma} [S^{d_2}( C)][S^{d_1-d_2+2g-2+\gamma}( C)]
\end{align*}
and if $d \geq 0$:
\begin{align*}
&[\mathcal{M}_{\varepsilon}^{2,d}(\gamma)]=\L^{4g-4+4 \gamma} [M_{\varepsilon}^{2,d}] + \sum_{(d_1,d_2) \in I_{1,\varepsilon}(d,\gamma)} \L^{3g-3+3\gamma +d_2} [S^{d_1}( C)][S^{d_1-d_2+2g-2+\gamma}( C)]+\\
&+\sum_{(d_1,d_2) \in I_{2,\varepsilon}(d,\gamma)} \L^{4g-4+3\gamma} [S^{d_2}( C)][S^{d_1-d_2+2g-2+\gamma}( C)]
\end{align*}
where $I_{1,\varepsilon}(d,\gamma)$ is the set of pairs of integers $(d_1,d_2)$ satisfying $d_1+d_2=d$ and:
\begin{align*}
&\max\{0,(d-\gamma)/2+1-g\} \leq d_1 < \frac{d}{2} \\
&\frac{d}{2} < d_2 \leq \min\{d,(d+\gamma)/2+g-1\}\\
&\max\{0,2g-2+\gamma-d\}\leq d_1-d_2+2g-2+\gamma < 2g-2+\gamma \text{ same parity as $d$}
\end{align*}
while $I_{2,\varepsilon}(d,\gamma)$ is the set of pairs of integers $(d_1,d_2)$ satisfying $d_1+d_2=d$ and:
\begin{align*}
&(d-\gamma)/2 +1-g \leq d_1 < \min \left \{\frac{d+1}{2},d+1 \right\} \\
&\max\left \{-1,\frac{d-1}{2} \right\}< d_2 \leq (d+\gamma)/2+g-1\\
& 0\leq d_1-d_2+2g-2+\gamma < 2g-2+\gamma+\min\{1,d+1\} \text{ same parity as $d$}.
\end{align*}
Furthermore, if $\gamma > d$:
\begin{align*}
[\mathcal{M}_{\infty}^{2,d}(\gamma)]=\sum_{(d_1,d_2) \in I_{2,\infty}(d,\gamma)} \L^{4g-4+3\gamma} [S^{d_2}( C)][S^{d_1-d_2+2g-2+\gamma}( C)]
\end{align*}
where $I_{2,\infty}(d,\gamma)$ is the set of pairs of integers $(d_1,d_2)$ satisfying $d_1+d_2=d$ and:
\begin{align*}
&(d-\gamma)/2 +1-g \leq d_1 \leq d \\
&0 \leq d_2 \leq (d+\gamma)/2+g-1\\
& 0\leq d_1-d_2+2g-2+\gamma \leq 2g-2+\gamma+d \text{ same parity as $d$}.
\end{align*}
\end{theorem}
In particular, as done in \cite[section 7.2]{hausel2001geometry} we can consider the embeddings:
\begin{equation*}
\mathcal{M}_{\sigma}^{2,d}(\gamma) \rightarrow \mathcal{M}_{\sigma}^{2,d}(\gamma+1)
\end{equation*}
and compute, for $\sigma=\infty$ and $\sigma=\varepsilon$
\begin{equation*}
\lim_{\gamma \rightarrow \infty} P(\mathcal{M}_{\sigma}^{2,d}(\gamma),t).
\end{equation*}
\begin{theorem}
We have:
\begin{align*}
&\lim_{\gamma \rightarrow \infty} P(\mathcal{M}_{\varepsilon}^{2,d}(\gamma),t) = \lim_{\gamma \rightarrow \infty} P(\mathcal{M}_{\infty}^{2,d}(\gamma),t)=\\
&=\frac{(1+t^3)^{2g}(1+t)^{2g}}{(1-t^2)^2(1-t^4)}=P(\mathbb{C} \P^{\infty},t) P(B \overline{\mathcal{G}},t)
\end{align*}
where $B \overline{\mathcal{G}}$ is the classifying space of the group $\overline{\mathcal{G}}$ mentioned in \cite[section 7.2]{hausel2001geometry}.
\begin{proof}
Recall that $\dim \mathcal{M}_{\varepsilon}^{2,d}(\gamma) = \dim \mathcal{M}_{\infty}^{2,d}(\gamma) = d+ 6g-6+4 \gamma$. We can also directly compute the Poincar\'e polynomial of $\mathcal{M}_{\varepsilon}^{2,d}(\gamma)$ since it is semiprojective. Assuming $d \geq 0$, we have:
\begin{align*}
&P(\mathcal{M}_{\varepsilon}^{2,d}(\gamma),t)=P(M_{\varepsilon}^{2,d},t) +\\
&+ \sum_{(d_1,d_2) \in I_{1,\varepsilon}(d,\gamma)} t^{2g-2+2d-4 d_1} P(S^{d_1}( C),t)P(S^{d_1-d_2+2g-2+\gamma}( C),t)+\\
&+\sum_{(d_1,d_2) \in I_{2,\varepsilon}(d,\gamma)} t^{2 d_2} P(S^{d_2}( C),t)P(S^{d_1-d_2+2g-2+\gamma}( C),t).
\end{align*}
Therefore:
\begin{align*}
&\lim_{\gamma \rightarrow \infty} P(\mathcal{M}_{\varepsilon}^{2,d}(\gamma),t)=P(M_{\varepsilon}^{2,d},t) + t^{2g-2} P(S^{\infty}( C),t) \sum_{i=0}^{\lfloor{d/2}\rfloor} t^{2d-4 i} P(S^{i}( C),t)+\\
&+ P(S^{\infty}( C),t) \sum_{i>d/2} t^{2 i} P(S^{i}( C),t)=\\
&=P(M_{\varepsilon}^{2,d},t)+P(S^{\infty}( C),t) \sum_{i\geq 0} t^{2i} P(S^{i}( C),t)+\\
&+P(S^{\infty}( C),t) \left ( \sum_{i=0}^{\lfloor{d/2}\rfloor} (t^{2g-2+2d-4i}-t^{2i})P(S^{i}( C),t) \right ).
\end{align*}
From \cite[remark 6.3]{mozgovoy2013moduli} it is immediate to deduce that:
\begin{equation*}
P(M_{\varepsilon}^{2,d},t)=P(S^{\infty}( C),t) \left ( \sum_{i=0}^{\lfloor{d/2}\rfloor} (t^{2i}-t^{2g-2+2d-4i})P(S^{i}( C),t) \right )
\end{equation*}
and hence:
\begin{align*}
&\lim_{\gamma \rightarrow \infty} P(\mathcal{M}_{\varepsilon}^{2,d}(\gamma),t)=P(S^{\infty}( C),t) \sum_{i\geq 0} t^{2i} P(S^{i}( C),t)=\\
&=\frac{(1+t^3)^{2g}(1+t)^{2g}}{(1-t^2)^2(1-t^4)}
\end{align*}
as we wanted.
For $d<0$, $\mathcal{M}_{\varepsilon}^{2,d}(\gamma)=\mathcal{M}_{\infty}^{2,d}(\gamma)$. Otherwise:
\begin{align*}
&P(\mathcal{M}_{\infty}^{2,d}(\gamma),t)=\sum_{(d_1,d_2) \in I_{2,\infty}(d,\gamma)} t^{2 d_2} P(S^{d_2}( C),t)P(S^{d_1-d_2+2g-2+\gamma}( C),t)
\end{align*}
and so:
\begin{align*}
&\lim_{\gamma \rightarrow \infty} P(\mathcal{M}_{\infty}^{2,d}(\gamma),t)= P(S^{\infty}( C),t) \sum_{i\geq 0} t^{2i} P(S^{i}( C),t)=\\
&=\frac{(1+t^3)^{2g}(1+t)^{2g}}{(1-t^2)^2(1-t^4)}
\end{align*}
as claimed.
\end{proof}
\end{theorem}
We conclude the chapter by examining what happens during the wall-crossing, i.e. what is the structure of the flip loci.
As in the case of Bradlow-Higgs triples we have two flip loci coming with canonical maps:
\begin{equation*}
\pi_{\bar\sigma}^{d,+}(\gamma): \mathcal{W}_{\bar\sigma}^{d,+}(\gamma) \rightarrow X_{\bar\sigma}^{d}(\gamma),
\end{equation*}
\begin{equation*}
\pi_{\bar\sigma}^{d,-}(\gamma): \mathcal{W}_{\bar\sigma}^{d,-}(\gamma) \rightarrow X_{\bar\sigma}^{d}(\gamma)
\end{equation*}
where
\begin{equation*}
X_{\bar\sigma}^{d}(\gamma)=S^{(d-\bar\sigma)/2}( C) \times J^{(d-\bar\sigma)/2}( C) \times H^0(K(\gamma P))^2.
\end{equation*}
Triples $(E,\phi,s) \in \mathcal{W}_{\bar\sigma}^{d,+}(\gamma)$ are characterized as non-split extensions
\begin{equation*}
0 \rightarrow (L,\psi_1) \rightarrow (E,\phi) \rightarrow (M, \psi_2) \rightarrow 0
\end{equation*}
where $\deg L =(d-\bar\sigma)/2$, $\deg M =(d+\bar\sigma)/2$ and $s \in H^0(L)$. Here $\psi_i \in H^0(K(\gamma P))$. For such a triple $\pi_{\bar\sigma}^{d,+}(\gamma)(E,\phi,s)=(L,s,M,\psi_1,\psi_2)$.
If we fix a point $(L,s,M,\psi_1,\psi_2) \in X_{\bar\sigma}^{d}(\gamma)$ then the fiber of $\pi_{\bar\sigma}^{d,+}(\gamma)$ over such a point can be once again characterized using a hypercohomology group $\H^1((M,\psi_2),(L,\psi_1))$ whose dimension is:
\begin{equation*}
\dim \H^1((M,\psi_2),(L,\psi_1))=
\begin{cases}
2g-2+\gamma & \text{if } \psi_1 \neq \psi_2\\
2g-2+\gamma + \dim H^0(L^*M(-\gamma P)) & \text{if } \psi_1=\psi_2.
\end{cases}
\end{equation*}
In particular, if $\gamma > d$, $\mathcal{W}_{\bar\sigma}^{d,+}(\gamma)$ is a $\mathbb{C} \P^{2g-3+\gamma}$-bundle over $X_{\bar\sigma}^{d}(\gamma)$ regardless of $\bar\sigma$.
Triples $(E,\phi,s) \in \mathcal{W}_{\bar\sigma}^{d,-}(\gamma)$ are characterized as extensions
\begin{equation*}
0 \rightarrow (M,\psi_2) \rightarrow (E,\phi) \rightarrow (L, \psi_1) \rightarrow 0
\end{equation*}
where $\deg L =(d-\bar\sigma)/2$, $\deg M =(d+\bar\sigma)/2$ and $s \in H^0(E)$ projects to a nonzero $\bar{s} \in H^0(L)$. We must have that either the extension is non-split or it is split but $s \neq \bar{s}$.
If we fix a point $(L,s,M,\psi_1,\psi_2) \in X_{\bar\sigma}^{d}(\gamma)$ then the fiber of $\pi_{\bar\sigma}^{d,-}(\gamma)$ over such a point can be once again characterized using a hypercohomology group $ \widetilde\H^1((L, \bar{s},\psi_1),(M,\psi_2))$ whose dimension is:
\begin{equation*}
\dim \widetilde\H^1((L, \bar{s},\psi_1),(M,\psi_2))=(d+\bar\sigma)/2+g-1+\gamma
\end{equation*}
In particular, $\mathcal{W}_{\bar\sigma}^{d,-}(\gamma)$ is always a $\mathbb{C} \P^{(d+\bar\sigma)/2+g-2+\gamma}$-bundle over $X_{\bar\sigma}^{d}(\gamma)$.
\chapter{Abstract}
In the present thesis we study the geometry of the moduli spaces of Bradlow-Higgs triples on a smooth projective curve $C$. $(E,\phi, s)$ is a Bradlow-Higgs triple if $(E,\phi)$ is a Higgs bundle and $s$ is a non-zero global section of $E$. There is a family of stability conditions for triples that depends on a positive real parameter $\sigma$. The moduli spaces $\mathcal{M}_\sigma^{r,d}$ of $\sigma$-semistable triples of rank $r$ and degree $d$ vary with $\sigma$. The phenomenon arising from this is known as wall-crossing.
In the first half of the thesis we will examine how the moduli spaces $\mathcal{M}_\sigma^{r,d}$ and their universal additive invariants change as $\sigma$ varies, for the case $r=2$. In particular we will study the case of $\sigma$ very close to 0, for which $\mathcal{M}_\sigma^{r,d}$ relates to the moduli space of stable Higgs bundles, and $\sigma$ very large, for which $\mathcal{M}_\sigma^{r,d}$ is a relative Hilbert scheme of points for the family of spectral curves. Some of these results will be generalized to Bradlow-Higgs triples with poles.
In the second half we will prove a formula relating the cohomology of $\mathcal{M}_\sigma^{2,d}$ for small $\sigma$ and $d$ odd and the perverse filtration on the cohomology of the moduli space of stable Higgs bundles. The formula is not far from the generalized Macdonald formulas found in \cite{migliorini2011support}, \cite{maulik2014macdonald} and \cite{migliorini2015support}. We will also partially generalize this result to the case of rank greater than 2.\\[10pt]
\textbf{Keywords:} Moduli spaces, wall-crossing, Bradlow-Higgs triples, Macdonald formula, Hilbert scheme.
\chapter{R\'esum\'e}
Dans la pr\'esente th\`{e}se, nous \'etudions la g\'eom\'etrie des espaces de modules de triplets de Bradlow-Higgs sur une courbe lisse projective $C$. $(E, \phi, s)$ est un triplet de Bradlow-Higgs si $ (E, \phi) $ est un fibr\'e de Higgs et $s$ est une section globale de $ E $ non-nulle. Il y a une famille de conditions de stabilit\'e pour les triplets qui d\'epend d'un param\`{e}tre positif r\'eel $ \sigma $. Les espaces de modules $ \mathcal {M}_\sigma^{r, d} $ de triplets $\sigma$-semistable de rang $r$ et degr\'e $d$ varient avec $ \sigma $. Le ph\'enom\`{e}ne r\'esultant est connu comme wall-crossing.
Dans la premi\`{e}re moiti\'e de la th\`{e}se, nous examinerons comment les espaces de modules $\mathcal{M}_\sigma^{r, d} $ et leurs invariants additifs universels changent en fonction de $ \sigma $, pour le cas $ r = 2 $. En particulier, nous allons \'etudier le cas de $ \sigma $ tr\`{e}s proche de 0, pour lequel $ \mathcal {M}_\sigma^{r, d} $ est reli\'e \`{a} l'espace des modules de fibr\'es de Higgs stables, et $ \sigma $ tr\`{e}s grand, pour lequel $ \mathcal{M}_\sigma^{r, d} $ est un sch\'ema de Hilbert de points relatif pour la famille de courbes spectrales. Certains de ces r\'esultats seront g\'en\'eralis\'es aux triplets de Bradlow-Higgs avec des p\^oles.
Dans la seconde moiti\'e, nous allons prouver une formule concernant la cohomologie de $ \mathcal{M}_\sigma^{2, d} $ pour $ \sigma $ petit et $d$ impair et la filtration perverse sur la cohomologie de l'espace des modules de fibr\'es de Higgs stables de degr\'e impair. La formule est proche des formules de Macdonald g\'en\'eralis\'ees trouv\'ees dans \cite{migliorini2011support}, \cite{maulik2014macdonald} et \cite{migliorini2015support}. Nous allons aussi partiellement g\'en\'eraliser ce r\'esultat au cas du rang sup\'erieur \`{a} 2.\\[10pt]
\textbf{Mots-cl\'e:} Espaces de modules, wall-crossing, triplets de Bradlow-Higgs, formule de Macdonald, sch\'ema de Hilbert. |
\section*{INTRODUCTION}
\newHead{INTRO}
\vskip20pt
The object of this work is to give an adequate theory
of surgery for paracompact manifolds and proper maps.
By adequate we mean first that it should contain the theory of surgery
for compact manifolds. Secondly, the theory should be general enough to
permit extensions of the theoretical results of compact surgery.
These objectives are largely realized.
We obtain surgery groups which characterize the problem in dimensions
greater than or equal to five.
These groups depend only on the proper $2${--}type of the problem.
Using these groups one can classify all paracompact manifolds of a
given proper, simple homotopy type (see \cite{bthirtythree} or
\cite{bten} for a definition of simple homotopy type).
The first chapter constitutes the chief technical results of this work.
In \cite{bthirtythree}, Siebenmann gives a ``geometric'' characterization
of proper homotopy equivalence (Proposition IV).
This characterization was also discovered by Farrell{--}Wagoner
\cite{bnine} from whom I learned it.
In section 2 we develop an algebraic process to handle this
characterization.
In section 3 we apply this process to construct groups which are the
analogue of the homotopy and homology groups.
Thus we get actual groups measuring by how much a map fails to
be a proper homotopy equivalence.
These groups also satisfy a version of the Hurewicz and Namioka theorems,
so one can often use these homology groups, which satisfy a version of
excision, Mayer{--}Vietoris, etc.
In section 4 we construct a cohomology theory for our theory.
We get various products for this theory.
Section 5 is devoted to an analysis of simple homotopy type along the
lines set out by Milnor in \cite{btwentythree}.
Section 6 is devoted to constructing locally compact CW complexes
with a given chain complex (see Wall \cite{bthirtyeight} for a
treatment of the compact case of this problem).
Chapter 2 is devoted to an analysis of Poincar\'e duality for paracompact
manifolds and its generalization to arbitrary locally compact, finite
dimensional CW complexes.
In Chapter 3 the actual surgeries are performed.
It has been observed by several people (especially Quinn
\cite{btwentynine} and \cite{bthirty}) that all the surgery one needs to
be able to do is the surgery for a pair $(X,\partial X)$ for which
$\partial X\subseteq X$ is a proper $1${--}equivalence (in the compact
case this means the inclusion induces isomorphisms on components and
on $\pi_1$).
We do this in the first section.
In the second section, we sketch the general set up and applications
of the theory of paracompact surgery.
A word or two is in order here about internal referencing.
A reference reads from right to left, so that Corollary 3.4.1.5 is the fifth
corollary to the first theorem of section 4 in chapter 3.
If the reference is made from chapter 3 it would be Corollary 4.1.5,
and if from section 4, Corollary 1.5.
Theorem (Proposition, Lemma) 3.4.6 is the sixth theorem of section 4 of
chapter 3.
Perhaps we should also remark that our use of the term
$n${--}ad agrees with the use of the term in
Wall \cite{bfortyone} (see especially Chapter 0).
For an $n${--}ad, $K$, $\partial_iK$ denotes the $(n-1)${--}ad whose
total space is the $i^{\text{ th}}$ face of $K$ and with the $(n-1)${--}ad
structure induced by intersecting the other faces of $K$.
$\delta_iK$ is the $(n-1)${--}ad obtained by deleting the $i^{\text{ th}}$
face.
$s_nK$ denotes the $(n+1)${--}ad obtained by making $K$
the $(n+1)^{\text{ st}}$ face (it can also be regarded as the $(n+1)${--}ad
$K\times I$, where $I$ has the usual pair structure).
Lastly, several acknowledgements are in order.
This thesis was written under the direction of J.~Wagoner,
to whom I am indebted for many suggestions during the preparation of
this work.
I am greatly indebted to him and to T.~Farrell for sharing their results and
intuition on proper homotopy with me at the very beginning.
Thanks are also due to G.~Cooke for many helpful discussions.
Many other friends likewise deserve thanks for their help.
The National Science Foundation should also be thanked for its support
during my graduate career.
\newpage
\chapter{The Proper Homotopy Category and Its Functors}
\section{Introduction, elementary results, and homogamous spaces}
\newHead{I.1}
The purpose of this chapter is to recall for the reader some of the basic
results we will need and to describe a ``good'' category in which to do
proper homotopy theory.
The notion of a proper map is clearly essential.
We define a map to be \emph{proper} \iff\ the inverse image of every closed
compact set is contained in a closed compact set.
We note that this definition is also found in Bredon \cite{btwo}, page 56.
With this definition of a proper map we immediately have the notions of
proper homotopy, proper homotopy equivalence, etc.,
and we can define the category of all topological spaces and proper maps.
Classically there are several functors which apply to this situation.
As examples we have sheaf cohomology with compact supports
and Borel{--}Moore homology with closed supports
(see Bredon \cite{btwo}).
We prefer to use singular theory whenever possible.
Here too we have cohomology with compact supports and homology with
locally finite chains.
Most of the results concerning such groups are scattered
(or non{--}existent) in the literature.
As a partial remedy for this situation we will write out the definitions of
these groups and at least indicate the results we need.
\begin{xDefinition}
A collection of subsets of $X$ is said to be \emph{locally finite} if every
closed, compact subset of $X$ intersects only finitely many
elements of this collection.
\end{xDefinition}
\begin{xDefinition}
$S_q^\locf(X;\Gamma)$, where $\Gamma$ is a local system of
$R${--}modules on $X$
(see Spanier \cite{bthirtyfive} pages 58; 281{--}283),
is defined to be the $R${--}module which is the set of all formal sums
$\sum \alpha_\sigma \sigma$, where $\sigma$ is a singular $q${--}simplex
of $X$, and $\alpha_\sigma\in\Gamma\bigl(\sigma(V_0)\bigr)$ is
zero except for a set of $\sigma$ whose images in $X$ are locally finite.
\end{xDefinition}
$S^q(X;\Gamma)$ is the module of functions $\varphi$ assigning to every
singular $q${--}simplex $\sigma$ of $X$ an element
$\varphi(\sigma)\in \Gamma\bigl(\sigma(V_0)\bigr)$.
For a family of supports $\psi$ on $X$
(see Bredon \cite{btwo} page 15 for a definition) let
$S^\psi_q(X;\Gamma)$ denote the submodule of $S_q^\locf(X;\Gamma)$
such that the union of all the images of the $\sigma$
occurring with non{--}zero coefficient in a chain lies in some element of
$\psi$.
$S^q_\psi(X;\Gamma)$ consists of the submodule of all functions
$\varphi$ for which there exists an element $c\in\phi$ such that if
$\Image\, \sigma \cap c=\emptyset$, $\varphi(\sigma)=0$.
These modules become chain complexes in the usual fashion.
Note that for the family of compact supports, $\cmpsup$,
$S^\cmpsup_q(X;\Gamma)$ is just the ordinary singular chains with
local coefficients.
For a proper subspace $A\subseteq X$ (inclusion is a proper map)
we get relative chain groups $S^\psi_q(X, A;\Gamma)$ and
$S^q_\psi(X, A;\Gamma)$.
Actually proper subspace is sometimes stronger than we need;
i.e. $S^\cmpsup_q(X, A;\Gamma)$ and
$S^q(X, A;\Gamma)$ are defined for any $A\subseteq X$.
There is a similar definition for the chain groups of a (proper) $n${--}ad.
The homology of $S^\psi_\ast(X, A;\Gamma)$ will be denoted
$H^\psi_\ast(X, A;\Gamma)$ except when $\psi=\cmpsup$ when
we just write $H_\ast(X, A;\Gamma)$.
The homology of $S_\psi^\ast(X, A;\Gamma)$ will be written
$H_\psi^\ast(X, A;\Gamma)$.
Now $S_\cmpsup^q(X, A;\Gamma)\subseteq S^q(X, A;\Gamma)$.
The quotient complex will be denoted
$S_{\text{end}}^q(X, A;\Gamma)$ and its homology
$H_{\text{end}}^q(X, A;\Gamma)$.
We have similar definitions for proper $n${--}ads and also for homology.
We will next set out the properties of these groups we will use.
Some of the obvious properties such as naturality and long exact sequences
will be omitted.
\textbf{Cup products}: There is a natural cup product\hfill
\[H^q_\psi(X\Colon A_1,\cdots, A_n;\Gamma_1)\otimes
H^k(X\Colon A_{n+1},\cdots, A_m;\Gamma_2)\RA{\cup}
H^{q+k}(X\Colon A_{1}, \cdots, A_m;\Gamma_1\otimes\Gamma_2)
\]
for a proper $(m+1)${--}ad $(X\Colon A_1,\cdots A_m)$.
It is associative and commutative in the graded sense
(i.e. $a\cdot b=(-1)^{{\text{deg}}\>a\cdot{\text{deg}}\> b}b\cdot a$).
Since $S^q_\psi\subseteq S^q$, all this follows easily from the properties of the
ordinary cup product with local coefficients once one checks that
if a cochain was supported in $c\in\psi$, then its product with any other
cochain is supported in $c$ if one uses the Alexander{--}Whitney diagonal
approximation (Spanier \cite{bthirtyfive} page 250).
\footnote{I should have remarked here that the chain homotopies giving
the associativity and the graded commutativity are correctly supported.}
{\bf Cross products}:
There are natural products
\[\begin{aligned}%
H^q_\psi(A\Colon A_1&,\cdots,A_n;\Gamma_1)\otimes
H^k(Y\Colon B_1,\cdots, B_m;\Gamma_2)\quad\RA{\qquad\times\qquad}\\
&H^{q+k}_{\pi_1^{-1}\psi}(X\times Y\Colon X\times B_1,\cdots,X\times B_m,
A_1\times Y,\cdots, A_m\times Y;\Gamma_1\otimes\Gamma_2)\\\end{aligned}\]
and
\[\begin{aligned}%
H_q^\psi(A\Colon A_1&,\cdots,A_n;\Gamma_1)\otimes
H_k^\locf(Y\Colon B_1,\cdots, B_m;\Gamma_2)\quad\RA{\qquad\times\qquad}\\
&H_{q+k}^{\psi\times Y}(X\times Y\Colon X\times B_1,\cdots,X\times B_m,
A_1\times Y,\cdots, A_m\times Y;\Gamma_1\otimes\Gamma_2)\\\end{aligned}\]
where
$\pi_1^{-1}(\psi)=\{ K\subseteq X\times Y\ \vert\ \pi_1(K)\in\psi\}$
and $\psi\times Y=\{ K\times Y\subseteq X\times Y\ \vert\ K\in\psi\}$.
These satisfy the usual properties of the cross product.
We discuss this case in some detail.
Let us first define
\[\tau\colon S_n^\locf(X\times Y)\to \sum_{i+j=n}
S^{\cmpsup}_i(X)\ \widehat\otimes\ S^{\cmpsup}_j(Y)\]
where $\widehat\otimes$ is the completed tensor product,
i.e. infinite sums are allowed\footnote{As long as the resulting sum is locally finite.}.
If $\sigma\colon \Delta^n\to X\times Y$, and if $\pi_1$ and $\pi_2$
are the projections,
$\displaystyle\tau(\sigma)=
\sum_{i+j=n}\ _i(\pi_1\sigma)\ \otimes\ (\pi_2\sigma)_j$
where $\ _i(\ )$ is the front $i${--}face and $(\ )_j$ is the back $j${--}face
(see Spanier \cite{bthirtyfive} page 250).
This extends over all of $S^\locf_n$ and is a natural chain map.
The cohomology cross product is then defined on the chain level by
$(c\times d)(\sigma)=c\bigl(\>_i(\pi_1\sigma)\bigr)\ \otimes
d\bigl((\pi_2\sigma)_j\bigr)$,
where $c$ is an $i${--}cochain, $d$ a $j${--}cochain, and $\sigma$ an
$(i+j)${--}chain.
One checks it has the usual properties.
We next define
$\lambda\colon S^\locf_i(X)\ \otimes\ S^\locf_j(Y)\ \RA{\ }\
S^\locf_{i+j}(X\times Y)$ as follows.
Let $h_{i,j}\colon \Delta^{i+j}\ \to\ \Delta^i\times\Delta^j$
be a homeomorphism such that
$\ _i(h_{i,j}) \colon \Delta^i\to\Delta^i\times\Delta^j$ is given by
$x\mapsto(x,0)$ and such that
$(h_{i,j})_j\colon \Delta^j\to\Delta^i\times\Delta^j$ is given by
$y\mapsto(0,y)$.
Define $\lambda(\sigma_X\otimes\sigma_Y)=
h_{i,j}\circ(\sigma_X\times\sigma_Y)$ and extend ``linearly'';
i.e. $\displaystyle\lambda\left(\sum \alpha\>\sigma_\alpha\otimes
\sum\beta\>\sigma_\beta\right)=
\sum_{\alpha,\beta}\ \alpha\otimes\beta\cdot
\lambda(\sigma_\alpha\otimes\sigma_\beta)$.
$\lambda$ then becomes a chain map, and the homology cross product
is then defined on the chain level as above.
It has the usual properties.
{\bf Slant product}:
There are natural products
\[\begin{aligned}%
H^q_\cmpsup(Y\Colon B_1,\cdots, B_m;&\Gamma_1)\otimes
H^\locf_{q+k}(X\times Y\Colon A_1\times Y,\cdots, A_n\times Y,
X\times B_1,\cdots, X\times B_m;\Gamma_2)
\\&\RA{\ \vert\ }\
H_k^\locf(X\Colon A_1,\cdots, A_n;\Gamma_1\otimes\Gamma_2)\\
\end{aligned}\]
and
\[\begin{aligned}%
H^q(Y\Colon B_1,\cdots, B_m;&\Gamma_1)\otimes
H_{q+k}(X\times Y\Colon A_1\times Y,\cdots, A_n\times Y,
X\times B_1,\cdots, X\times B_m;\Gamma_2)
\\&\RA{\ \vert\ }\
H_k(X\Colon A_1,\cdots, A_n;\Gamma_1\otimes\Gamma_2)\quad.\\\end{aligned}\]
The product is defined on the chain level by
\[c\vert \sigma = c\Big\vert {\sum \alpha\>\sigma_\alpha}=
\sum_\alpha\biggl(\sum_{i+j=q+k}
\ _i(\pi_1\sigma_\alpha)\ \otimes\ \Bigl(
c\bigl((\pi_2\sigma)_j\bigr)\otimes\alpha\Bigr)\biggr)\]
where $c$ applied to a chain is zero if the dimensions do not agree.
The slant product is natural on the chain level and has all the usual
properties.
{\bf Cap product}:
There is a natural product
\[\begin{aligned}%
H^q_\psi(X\Colon A_1,\cdots,A_n;&\Gamma_1)\otimes
H^\varphi_{q+k}(X\Colon A_1,\cdots,A_n,B_1,\cdots, B_m;\Gamma_2)\\
&\RA{\ \cap\ }\
H^{\varphi\cap\psi}_{k}(X\Colon B_1,\cdots,B_m;
\Gamma_1\otimes\Gamma_2)\quad .\\\end{aligned}\]
It is given by $u\cap z =u\vert d_\ast v$, where
$d\colon X\to X\times X$ is the diagonal map.
The cap product has all the usual properties.
We get better support conditions for our cap product than we did
for an arbitrary slant product because $d_\ast$ of a chain in $X\times X$
is ``locally finite'' with respect to sets of the form $c\times X$ and
$X\times c$ for any closed, compact $c\subseteq X$.
One of the most useful of the usual properties of the cap product is the
\setcounter{footnote}{0}
{\bf Browder Lemma}:
(\cite{bthree}, \cite{bfour}).\footnote{Best reference is I.1.5 Theorem of \cite{Browder}.}
Let $(X,A)$ be a proper pair ($A$ is a proper subspace), and let
$Z\in H^\psi_n(X,A;\Gamma_2)$.
Then $\partial Z\in H^\psi_{n-1}(A;\Gamma_2\vert_A)$ is defined.
The following diagram commutes,
where $\Gamma_3=\Gamma_1\otimes\Gamma_2$
\[\begin{matrix}%
H^{\ast-1}_\varphi(A;\Gamma_1\vert_A)&\rsa&
H^\ast_\varphi(X,A;\Gamma_1)&\rsa&H^\ast_\varphi(X;\Gamma_1)
&\rsa&H^{\ast}_\varphi(A;\Gamma_1\vert_A)
\\
\downlabeledarrow[\bigg]{\cap(-1)^{n-1}\partial Z}{}&&\downlabeledarrow[\bigg]{\cap Z}{}
&&\downlabeledarrow[\bigg]{\cap Z}{}&&\downlabeledarrow[\bigg]{\cap \partial Z}{}\\
H^{\varphi\cap\psi}_{n-\ast}(A;\Gthree\vert_A)&\rsa&
H^{\varphi\cap\psi}_{n-\ast}(X;\Gthree)&\rsa&
H^{\varphi\cap\psi}_{n-\ast}(X,A;\Gthree)&\rsa&
H^{\varphi\cap\psi}_{n-\ast-1}(A;\Gthree\vert_A)\\
\end{matrix}\]
\LRTpageLabel{Universalcoefficientx}
In two cases, we also have a universal coefficient formula relating
cohomology and homology.
We first have the ordinary universal coefficient formula; namely
\[0\to{\text{Ext}}\bigl(H_{\ast-1}(\_\ ;\Gamma),\Z\bigr)\to
H^\ast\bigl(\_\ ;{\Homx}(\Gamma,\Z)\bigr)\to
{\Homx}\bigl(H_\ast(\_\ ;\Gamma),\Z\bigr)\to 0\]
is split exact (see Spanier \cite{bthirtyfive}, page 283).
We have a natural chain map
\[\alpha\colon S^\locf_\ast\bigl(\_\ ,{\Homx}(\Gamma,\Z)\bigr)\ \to\
{\Homx}\bigl(S^\ast_\cmpsup(\_\ ;\Gamma),\Z\bigr)\]
given by $\alpha(c)(\varphi)=\varphi(c)$.
If the space $X$ is \HCLx\ Bredon \cite{btwo}*{II.9.23}, shows that $\alpha$ induces
a homology isomorphism, so we get
\[0\to{\text{Ext}}(H^{\ast+1}_\cmpsup(X ;\Gamma),\Z)\to
H_\ast^\locf(X ;{\Homx}(\Gamma,\Z))\to
{\Homx}(H^\ast_\cmpsup(X ;\Gamma),\Z)\to0\]
is split exact.
Write $\Bar{\Gamma}$ for ${\Homx}(\Gamma,\Z)$.
Then if $c\in H^k(\_\ ;\Gamma)$, the following diagram commutes
\[\begin{matrix}%
0\to&{\text{Ext}}(H_{\ast-1}(\_\ ;\Z),\Z)&\to&
H^\ast(\_\ ;\Z)&\to&{\Homx}(H_\ast(\_\ ;\Z),\Z)&\to0\\
&\downlabeledarrow[\bigg]{{\text{Ext}}(\cap c)}{}&&\downlabeledarrow[\bigg]{c\cup}{}
&&\downlabeledarrow[\bigg]{{\Homx}(\cap c)}{}\\
0\to&{\text{Ext}}(H_{\ast+k-1}(\_\ ;\Gamma),\Z)&\to&
H^{\ast+k}(\_\ ;\Bar\Gamma)&\to&
{\Homx}(H_{\ast+k}(\_\ ;\Gamma),\Z)&\to0\\
\end{matrix}\]
If $c\in H^\locf_k(\_\ ;\Gamma)$, and if the spaces in question are \HCLx,
the following diagram commutes
\[\begin{matrix}%
0\to&\text{Ext}(H_{k-\ast-1}(\_\ ;\Z),\Z)&\to&
H^{k-\ast}(\_\ ;\Z)&\to&{\Homx}(H_{k-\ast}(\_\ ;\Z),\Z)&\to0\\
&\downlabeledarrow[\bigg]{{\text{Ext}}(\cap c)}{}&&\downlabeledarrow[\bigg]{\cap c}{}
&&\downlabeledarrow[\bigg]{{\Homx}(\cap c)}{}\\
0\to&{\text{Ext}}(H^{\ast+1}_\cmpsup(\_\ ;\Gamma),\Z)&\to&
H^\locf_\ast(\_\ ;\Bar\Gamma)&\to&
{\Homx}(H^{\ast}_\cmpsup(\_\ ;\Gamma),\Z)&\to0\\
\end{matrix}\]
These formulas can actually be seen on the chain level by picking
representatives and using the Alexander{--}Whitney
diagonal approximation.
These homology and cohomology groups enjoy other pleasant properties.
One which we shall exploit heavily throughout the remainder of this work
is the existence of a transfer map for any arbitrary cover.
For particulars, let $\pi\colon \widetilde X\to X$ be a covering map.
Then we have homomorphisms
\[{\text{tr}}\colon H^\locf_\ast(X;\Gamma)\ \to\
H^\locf_\ast(\widetilde X;\pi^\ast\Gamma)\]
and
\[{\text{tr}}\colon H_\cmpsup^\ast(\widetilde X;\pi^\ast\Gamma)
\ \to\ H_\cmpsup^\ast(X;\Gamma)
\ .\]
The first of these is given by defining ${\text{tr}}(\sigma)$ for a simplex
$\sigma$ and extending ``linearly.''
$\displaystyle{\text{tr}}(\sigma)=\sum_{p\in\pi^{-1}(v_0)}\sigma_p$,
where $p$ runs over all the points in $\pi^{-1}(v_0)$, where $v_0$
is a vertex of $\sigma$, and $\sigma_p$ is $\sigma$
lifted so that $v_0$ goes to $p$.
It is not hard to check ${\text{tr}}$ is a chain map.
For the cohomology trace define ${\text{tr}}(c)$ as the cochain whose value
on the simplex $\sigma$ in $X$ is $c\bigl({\text{tr}}(\sigma)\bigr)$; i.e.
$\bigl({\text{tr}}(c)\bigr)(\sigma)=c\bigl({\text{tr}}(\sigma)\bigr)$.
If $f\colon X\to Y$ is a proper map, and if $\pi\colon \widetilde Y\to Y$
is a cover, then, for the cover $\widetilde X\to X$ which is induced from
$\pi$ by $f$, $\tilde f_\ast({\text{tr}}Z)={\text{tr}}\tilde f_\ast Z$ and
${\text{tr}}(\tilde f^\ast c)=f^\ast({\text{tr}}\>c)$.
\insetitem{Warning}
The trace tends to be highly unnatural except in this one situation.
As an easy exercise, one may check that if
$c\in H^k_\cmpsup(\widetilde X; \pi^\ast\Gamma_1)$
and if $Z\in H^\locf_{q+k}(X;\Gamma_2)$, then, in
$H_q(X;\Gamma_1\otimes\Gamma_2)$,
$\pi_\ast(c\cap {\text{tr}} Z)= {\text{tr}}\>c\ \cap Z$.
In the coming pages, we will want to study spherical fibrations and
paracompact manifolds.
For the former objects we have
{\bf Thom Isomorphism Theorem}:
Let $\xi$ be a spherical fibration of dimension $(q-1)$ over $B$.
Let $S(\xi)$ be its total space, and let $D(\xi)$ be the total space of the
associated disc fibration.
Then there is a class
$U_\xi\in H^q\bigl(D(\xi),S(\xi); p^\ast_\xi(\Gamma_\xi)\bigr)$
[where $p\colon D(\xi)\to B$ is the projection, and $\Gamma_\xi$
is the local system on $B$ given at $b\in B$ by
$H^q\bigl(p^{-1}(b),p^{-1}(b)\cap S(\xi);\Z) \bigr)$] such that
\[\cup U_\xi \ \colon H_\varphi^\ast(B;\Gamma)\ \to\
H^{\ast+q}_{p^{-1}(\varphi)}\bigl(
D(\xi), S(\xi); p^\ast(\Gamma\otimes\Gamma_\xi)\bigr)\]
is an isomorphism.
One also has
\[U_\xi\cap\ \colon H_\ast\bigl(D(\xi), S(\xi);p^\ast(\Gamma)\bigr)\ \to\
H_{\ast-q}(B;\Gamma_\xi\otimes\Gamma)\]
is an isomorphism.
Note that we have been (and will continue to be) a little sloppy.
If $c\in H^\ast_\varphi(B;\Gamma)$, $c\cup U_\xi$ should actually be
$p^\ast(c)\ \cup\ U_\xi$.
A similar notational amalgamation has occurred when we write $U_\xi\cap$.
This theorem is proved by a spectral sequence argument
(see \cite{btwentysix}), so one need only check that we still have a Serre
spectral sequence with the appropriate supports.
For a paracompact manifold (i.e. a locally Euclidean, paracompact, Hausdorff
space), possibly with boundary, we have
{\bf Lefschetz Duality}: (\cite{btwenty}, \cite{bfortyfour}).
If $(M,\partial M)$ is a paracompact manifold pair of dimension $n$, there is
a class $[M]\in H^\locf_n(M,\partial M;\Gamma_M)$ (where $\Gamma_M$
is the local system for the bundle $\nu$, the normal bundle of $M$)
such that the maps
\[\cap[M]\colon H^\ast_\psi(M,\partial M;\Gamma)\ \to\
H^\psi_{n-\ast}(M;\Gamma\otimes\Gamma_M)\]
and
\[\cap[M]\colon H^\ast_\psi(M;\Gamma)\ \to\
H^\psi_{n-\ast}(M,\partial M;\Gamma\otimes\Gamma_M)\]
are isomorphisms.
This completes the first objective of this section, so we turn to the second.
The functors above already give us much non{--}trivial information on
the category of all spaces and proper maps, but they are insufficient
even to determine if a map is a proper homotopy equivalence on the
subcategory of locally compact, finite dimensional CW complexes, a
category in which we are surely going to be interested.
In fact, the next two sections will be concerned precisely with the
problem of constructing functors which will determine whether a map
is or is not a proper homotopy equivalence in this category.
If we restrict ourselves to finite complexes, the Whitehead Theorem
(\cite{bfortythree}) already provides the answer.
Notice that to solve the problem, even for finite complexes, we are
forced to consider homotopy, which means base points.
In order to solve the problem for locally finite complexes, we are going to
have to consider lots of base points simultaneously.
The category of spaces we are about to define is about the largest in
which we can place our points nicely.
It is also closed under proper homotopy equivalence.
\begin{xDefinition}
A set $B$ of points of $X$ is said to be a \emph{set of base points} for $X$
provided
\begin{enumerate}
\item[a)] every path component of $X$ contains a point of $B$
\item[b)] given any closed, compact set $c\subseteq X$,
there is a closed compact set $D$ such that there is a point of $B$
in every path component of $X-c$ which is not contained in $D$.
\end{enumerate}
\end{xDefinition}
\begin{xDefinition}
A set of base points, $B$, for a path connected space $X$
is said to be \emph{irreducible} if,
for any set of base points $C$ for $X$ with $C\subseteq B$, the cardinality
of $C$ is equal to the cardinality of $B$.
A set of base points for any space $X$ is said to be \emph{irreducible}
provided it is an irreducible set of base points for
each path component of $X$.
\end{xDefinition}
\begin{xDefinition}
Two locally finite sets of points are said to be equivalent $(\sim)$
provided there is a 1{--}1 correspondence between the two sets which
is given by a locally finite set of paths.
\end{xDefinition}
\begin{xDefinition}
Consider the following two properties of a space $X$:
\begin{enumerate}
\item[]
\begin{enumerate}
\item[1)] Every set of base points for $X$ has an irreducible,
locally finite subset.
\item[2)] Any two irreducible, locally finite sets of base points for $X$
are equivalent.
\end{enumerate}
\item[] A space $X$ is said to be \emph{\prex{--}homogamous}
\footnote[1]{\text{In the original manuscript we started by calling this property
homogamous and then}\\\text{redefined the term at the end of this section.}}
provided $X\times I$ satisfies 1) and 2).
\end{enumerate}
\end{xDefinition}
\bigskip
\BEGIN{P.1.1.1}
If $X$ has the proper homotopy type of a \hfill\\\prex{--}homogamous space,
then $X$ has properties 1) and 2).
\end{Proposition}
\medskip
\begin{proof}
We first prove two lemmas.
\BEGIN{L.1.1.1}
Let $f\colon X\to Y$ be a proper map which induces injections of
$H^0(Y)$ into $H^0(X)$ and
of $H^0_{{\text{end}}}(Y)$ into $H^0_{{\text{end}}}(X)$.
Then if $\{p\}$ is a set of base points for $X$, then $\{f(p)\}$ is a set
of base points for $Y$.
\end{Lemma}
\begin{proof}
Since $f$ induces an injection on $H^0$, there is an $f(p)$ in
every path component of $Y$.
Now look at the path components of $Y-c$, where $c$ is some closed,
compact subset of $Y$.
Let $\{W_\alpha\}$ be the set of path components of $Y-c$ such that
$f^{-1}(W_\alpha)$ contains no point of $\{ p\}$.
Since $\{p\}$ is a set of base points for $X$,
$\displaystyle\mathop{\cup}_\alpha f^{-1}(W_\alpha)\subseteq D$,
where $D$ is some closed, compact subset of $X$.
Then $f(X-D)\cap W_\alpha=\emptyset$ for all $\alpha$.
Define a cochain $\beta$ by
\[\beta(q)=\begin{cases}1& q\in W_\alpha\\ 0& q\notin W_\alpha\\\end{cases}\ .\]
Then $\delta\beta(\lambda)=
\beta\bigl(\lambda(1)\bigr)-\beta\bigl(\lambda(0)\bigr)=0$
if $\lambda\cap c = \emptyset$.
Hence $\delta\beta=0$ in $S^1_{{\text{end}}}(Y;\Z)$.
But since $f(X-D)\cap W_\alpha=\emptyset$, $f^\ast\beta=0$ in
$S^0_{{\text{end}}}(X;\Z)$.
Since $f^\ast$ is an injection on $H^0_{{\text{end}}}$, $\beta=0$ in
$H^0_{{\text{end}}}(Y;\Z)$.
But this implies $\displaystyle\mathop{\cup}_\alpha W_\alpha$
is contained in some compact set.
\end{proof}
\medskip
\BEGIN{L.1.1.2}
Let $f$ be a map properly homotopic to the identity.
Let $\{p\}$ be a locally finite set of base points.
Then $\{f(p)\}$ is equivalent to a subset of $\{p\}$.
\end{Lemma}
\begin{proof}
We have $F\colon X\times I\to X$ a proper map.
The set $\{p\times I\}$ is clearly locally finite.
Since $F$ is proper, $\{F(p\times I)\}$ is easily seen to be locally finite.
But $\{F(p\times I)\}$ provides an equivalence between $\{f(p)\}$
and some subset of $\{p\}$ [more than one $p$ may go to a given
$f(p)$].
\end{proof}
Now let $X$ have the proper homotopy type of $Y$,
a \prex{--}homogamous space.
Hence we have proper maps $f\colon X\to Y$ and $g\colon Y\to X$
with the usual properties.
Let $\{p\}$ be a set of base points for $X$.
Then by \fullRef{L.1.1.1},
$\{f(p)\}$ is a set for $Y$ and
$\{f(p)\times 0\}$ is a set for $Y\times I$.
Since $Y$ is \prex{--}homogamous, there is an irreducible, locally finite
subset $\{ f(p^\prime)\times 0\}$.
By
\fullRef{L.1.1.1},
$\{g\circ f(p^\prime)\}$ is a locally finite set
of base points for $X$.
But by
\fullRef{L.1.1.2},
there is a further refinement, $\{p^{\prime\prime}\}$,
of $\{p\}$ such that $\{p^{\prime\prime}\}\sim \{g\circ f(p^\prime)\}$.
But then $\{p^{\prime\prime}\}$ is easily seen to be a set of base points
also.
Now $\{p^{\prime\prime}\}$ is in 1{--}1 correspondence
with $\{f(p^{\prime\prime})\}$, and $\{f(p^{\prime\prime})\times 0\}$
is a set of base points for $Y\times I$ by \fullRef{L.1.1.1}.
$\{f(p^{\prime\prime})\times 0\}$ is
a subset of $\{f(p^{\prime})\times 0\}$ and is thus irreducible.
Hence $\{p^{\prime\prime}\}$ is easily seen to be irreducible, and
therefore $X$ satisfies 1).
Let $\{p\}$ be an irreducible, locally finite set of base points for $X$.
We claim that there is an irreducible, locally finite set of base points $\{q\}$
for $Y\times I$ such that $\{p\}\sim\{g\circ\pi(q)\}$, where
$\pi\colon Y\times I\to Y$ is projection.
By the argument in \fullRef{L.1.1.2},
we see that we have a locally finite set
of paths $\{\lambda_p\}$ from $\{p\}$ to
$\{g\circ f(p)\}$.
However, $(g\circ f)^{-1}(g\circ f)(p)$ may contain more points of
$\{p\}$ than just $p$.
But since $\{\lambda_p\}$ is locally finite, there are
only finitely many such points, say $p_1$, \dots, $p_n$.
Let $q=f(p)\times 0$ and define $q_i=f(p)\times 1/i$ for $1\leq i\leq n$.
The resulting set of points, $\{q\}$ is easily seen to be locally finite, and
by several applications of \fullRef{L.1.1.1},
$\{q\}$ is an irreducible set of
base points for $Y\times I$.
So suppose given $\{p\}$ and $\{p^\prime\}$, irreducible, locally
finite sets of base points for $X$.
Pick $\{q\}$ and $\{q^\prime\}$ as above to be irreducible, locally finite
sets of base points for $Y\times I$.
Since $Y$ is \prex{--}homogamous, $\{q\}\sim\{q^\prime\}$,
so $\{g\circ\pi(q)\}\sim\{g\circ\pi(q^\prime)\}$.
Thus $\{p\}\sim\{p^\prime\}$, so $X$ satisfies 2).
\end{proof}
\medskip
\BEGIN{C.1.1.1.1.1}
A space which is the proper homotopy type of a \prex{--}homogamous space
is \prex{--}homogamous.
\end{Corollary}
\medskip
\BEGIN{C.1.1.1.2}
The mapping cylinder of a proper map whose range is \prex{--}homogamous
is \prex{--}homogamous.
\end{Corollary}
\bigskip
\BEGIN{P.1.1.2}
Let $\{\mathcal O\}$ be a locally finite open cover of $X$.
Further assume that each $\mathcal O$ is path connected and that
each $\Bar{\mathcal O}$ is compact.
Then $X$ is \prex{--}homogamous.
\end{Proposition}
\medskip
\BEGIN{C.1.1.2.1}
A locally compact, locally path connected,
paracompact space is \prex{--}homogamous.
\end{Corollary}
\medskip
\BEGIN{C.1.1.2.2}
A locally compact CW complex is \prex{--}homogamous.
\end{Corollary}
\medskip
\BEGIN{C.1.1.2.3}
A paracompact topological manifold is \prex{--}homogamous.
\end{Corollary}
\bigskip
\begin{proof}
If $\{\mathcal O\}$ is the collection for $X$, $\{{\mathcal O}\times I\}$ is
a cover for $X\times I$ with the same properties, so, if we can show
1) and 2) hold for $X$, we are done.
Since each $\mathcal O$ is path connected, each path component of $X$ is open.
Also the complement of a path component is open, so each path component
is both open and closed.
Hence $X$ is \prex{--}homogamous \iff\ each path component is, so we assume
$X$ is path connected.
We claim $X$ is $\sigma${--}compact, i.e. the countable union
of compact sets.
In fact, we will show $\{{\mathcal O}\}$ is at most countable.
As a first step, define a metric $d$ on $X$ as follows.
If $p\neq q$, look at a path $\lambda$ from $p$ to $q$.
$\lambda$ is compact, so it is contained in a finite union of ${\mathcal O}$'s.
Hence $\lambda$ is contained in a closed, compact set
so $\lambda$ intersects only finitely many ${\mathcal O}$'s.
Let $r(\lambda;p,q) =$ the number of ${\mathcal O}$'s that $\lambda$
intersects (non{--}empty).
Define $\displaystyle d(p,q)=\mathop{{\text{min}}}_{\lambda}
r(\lambda;p,q)$.
This is a natural number, so there is actually some path, $\lambda$,
such that $d(p,q)=r(\lambda;p,q)$.
If $p=q$, set $d(p,q)=0$.
$d$ is easily seen to be a metric.
Let us fix $p\in X$.
Then to each ${\mathcal O}$ we can associate a number
$\displaystyle m({\mathcal O},p)=\mathop{{\text{min}}}_{q\in{\mathcal O}}d(p,q)$.
We claim that, for any $n$, $m({\mathcal O},p)\leq n$ for only finitely many
${\mathcal O}$.
For $n=0$ this is an easy consequence of the fact that
$\{{\mathcal O}\}$ is locally finite.
Now induct on $n$.
Let ${\mathcal O}_1$, \dots,${\mathcal O}_k$ be all the ${\mathcal O}$'s such that
$m({\mathcal O},p)\leq n-1$.
Let $\displaystyle c=\mathop{\cup}_{i=1}^k \Bar{\mathcal O}_i$.
$c$ is compact.
Suppose $\Bar{\mathcal O}\cap c=\emptyset$.
Then we claim $m({\mathcal O},p)\geq n+1$.
To see this, pick $q\in {\mathcal O}$, and any path $\lambda$ from $p$
to $q$.
If we can show $r(\lambda;p,q)\geq n+1$, we are done.
Let $[0,x]$ be the closed interval which is the first component of
$\lambda^{-1}(c)$, where $\lambda\colon I\to X$ is the path.
Since $c\cap \Bar{\mathcal O}=\emptyset$,
$\lambda^{-1}(\Bar{\mathcal O})\geq s$, where $s>x$.
Pick $x<t<s$.
Then $\lambda(t)\notin c$, so the path from $p$ to $\lambda(t)$
already intersects at least $n$ of the ${\mathcal O}$'s, so, from
$p$ to $q$ it must intersect at least $n+1$.
Therefore, if $m({\mathcal O},p)\leq n$, $\Bar{\mathcal O}\cap c\neq\emptyset$.
But since $\{{\mathcal O}\}$ is locally finite, there are only finitely many
${\mathcal O}$ for which this is true.
This completes the induction.
Hence the cover $\{{\mathcal O}\}$ is at most countable.
If $\{{\mathcal O}\}$ is finite, $X$ is compact and hence easily seen to satisfy
1) and 2).
Hence we assume $\{{\mathcal O}\}$ is infinite.
Enumerate $\{{\mathcal O}\}$, and set
$\displaystyle C_k=\mathop{\cup}_{i=0}^k \Bar{\mathcal O}_i$.
Since $C_k$ is compact, there are but finitely many ${\mathcal O}$'s such that
$\Bar{\mathcal O}\cap C_k\neq\emptyset$.
Let $E$ be the union of $c$ and these ${\mathcal O}$'s.
Then $E$ is compact, as is $\partial E$, the frontier of $E$ in $X$.
Let $\{W_\alpha\}$ be the path components of $X-C_k$ not
contained entirely in $E$.
Look at $W_\alpha\cap \partial E$.
It might be empty, in which case $W_\alpha$ is actually a component of $X$
since $\partial E$ separates the interior of $E$ and $X-E$.
But $X$ is connected, so $W_\alpha\cap\partial E\neq\emptyset$.
Now if $p\in\partial E$, $p\in{\mathcal O}$ with ${\mathcal O}\cap C_k=\emptyset$.
Now ${\mathcal O}$ is a path connected set missing $C_k$ with ${\mathcal O}$ not
contained entirely in $E$, so ${\mathcal O}\subseteq W_\alpha$ for
some $\alpha$.
Hence the $W_\alpha$ cover $\partial E$.
The $W_\alpha$ are disjoint, so, as $\partial E$ is compact,
there are only finitely many of them.
Some $\Bar{W}_\alpha$ may be compact.
Set $D_k=E\cup ({\text{compact\ }}\Bar{W}_\alpha)$.
Then $D_k$ is compact.
Since the $C_k$ are cofinal in the collection of all compact subsets of $X$,
we may assume, after refinement, that
$C_0\subseteq D_0\subseteq C_1\subseteq D_1\subseteq\cdots
C_k\subseteq D_k\subseteq C_{k+1}\subseteq\cdots$
Now let $\{ p\}$ be a set of base points for $X$.
Let $\{W_{\alpha,k}\}$ be the set of unbounded path components
of $X-C_k$, which we saw above was finite.
Since $\{p\}$ is a set of base points, in each $\{W_{\alpha,k}\}$
there are infinitely many $p\in \{p\}$ for which there exists an
${\mathcal O}\in \{{\mathcal O}\}$ such that $p\in{\mathcal O}\subseteq W_{\alpha,k}$.
We get a locally finite subset $\{p^\prime\}\subseteq\{p\}$ by
picking one element of $\{p\}\cap{\mathcal O}$ for each such non{--}empty
intersection as ${\mathcal O}$ runs over $\{{\mathcal O}\}$.
By the above remarks, this set is a set of base points.
It is clearly locally finite so $X$ satisfies 1).
Now let $\{p_k\}$ and $\{q_k\}$ be locally finite irreducible
sets of base points (they are of necessity both countable).
Look at all the $p_k$'s in $D_0$.
Join them by paths to some $q_\ell$ not in $D_0$.
Join the $q_k$'s in $D_0$ to some $p_\ell$'s not in $D_0$.
Note that the number of paths intersecting $C_0\leq
(\hbox{number of $p_k$ in $D_0$})+(\hbox{number of $q_k$ in $D_0$})$.
For the inductive step, assume we have joined all the $p_k$'s in $D_{n-1}$
to some $q_k$'s and vice versa.
Suppose moreover that the
\[\hbox{\small number of paths intersecting $C_{n-i}$}\leq
(\hbox{\small number of $p_k$ in $D_{n-i}$})+
(\hbox{\small number of $q_k$ in $D_{n-i}$})\]
for $1\leq i\leq n$.
Look at the $p_k$'s in $D_n-D_{n-1}$ which have not already been joined
to some $q_\ell$ in $D_{n-1}$.
Each of these lies in some $W_{\alpha,n-1}$; i.e. in an unbounded component
of $X-C_{n-1}$.
Join the $p_k$ in $W_{\alpha,n-1} \cap (D_n-D_{n-1})$ which
have not already been fixed up to some $q_\ell$ in $W_{\alpha,n-1}-D_{n}$
by a path in $W_{\alpha,n-1}$; i.e. outside of $C_{n-1}$.
(Recall there are an infinite number of $p_k$ [and $q_k$] in each
$W_{\alpha,\ell}$, so we can always do this.)
Do the same for the $q_k$ in $D_n-D_{n-1}$.
Now each of these new paths misses $C_{n-1}$, so the
\[\hbox{\small number of paths intersecting $C_{n-i}$}\leq
(\hbox{\small number of $p_k$ in $D_{n-i}$})+
(\hbox{\small number of $q_k$ in $D_{n-i}$})\]
for $1\leq i\leq n$.
For $i=0$, the
\[\hbox{\small number of paths intersecting $C_{n}$}\leq
(\hbox{\small number of $p_k$ in $D_{n}$})+
(\hbox{\small number of $q_k$ in $D_{n}$})\ .\]
This completes the induction and shows $X$ satisfies 2).
\end{proof}
Local compactness and $\sigma${--}compactness are easily seen
to be proper homotopy invariants so we define
\begin{xDefinition} A space is said to be \emph{homogamous} provided it is
locally compact, $\sigma${--}compact and \prex{--}homogamous.
Note now that any irreducible set of base points for an homogamous
space is countable.
\end{xDefinition}
\bigskip
\section{The \texorpdfstring{$\epsilon${--}$\Delta$}{e--D} construction}
\newHead{I.2}
In this section we describe our construction.
It will enable us to produce a proper homotopy functor on any
homogamous space from an ordinary homotopy functor (a homotopy
functor is a functor from the category of based topological spaces
and based homotopy classes of maps to some category).
Now our homotopy functor, say ${\bf H}$, takes values in some category
${\mathcal A}$.
Associated to any homogamous space, $X$, we have an irreducible set
of locally finite base points, ${\bf I}$.
We also have a diagram scheme, ${\mathcal D}$, consisting of the closed,
compact subsets of $X$ (see the definition below for the definition
of a diagram scheme).
Our basic procedure is to associate an element in ${\mathcal A}$ to the
collection ${\bf H}(X-C,p)$, where $C$ is a closed compact subset of
$X$, and $p\in{\bf I}$.
In order to be able to do this, we must impose fairly strenuous conditions
on our category ${\mathcal A}$, but we prefer to do this in two stages.
\begin{xDefinition}[see \cite{btwentyfive} page 42]
A \emph{diagram scheme} is a triple ${\mathcal D}=(J,M,d)$, where $J$ is a
set whose elements are called vertices, $M$ is a set whose elements
are called arrows, and $d\colon M\to J\times J$ is a map.
Given a diagram scheme ${\mathcal D}$ and a category ${\mathcal A}$, a diagram
over ${\mathcal D}$ is a map from $J$ to the objects of ${\mathcal A}$
$( j\mapsto A_j )$ and a map from $M$ to the morphisms of ${\mathcal A}$
such that, if $d(m)=(i,j)$ $m$ goes to an element of
${\Homx}(A_i, A_j)$.
\end{xDefinition}
\insetitem{Notation}
$\bigl[{\mathcal D}, {\mathcal A}\bigr]$ denotes the category
of all diagrams in ${\mathcal A}$ over ${\mathcal D}$.
(A map between diagrams over ${\mathcal D}$ is a collection of morphisms
$f_j\colon A_j\to B_j$ such that $f_j\circ m=\Bar m\circ f_i$, where
$m\in{\Homx}(A_i, A_j)$, and $\Bar m\in{\Homx}(B_i, B_j)$
correspond to the same element in $M$).
If ${\mathcal I}$ is an index set (i.e. a set) ${\mathcal A}^{\mathcal I}$ denotes the
category whose objects are sets of objects in ${\mathcal A}$ indexes by
${\mathcal I}$.
The morphisms are sets of morphisms in ${\mathcal A}$ indexed by ${\mathcal I}$.
Finally, if ${\mathcal A}$ and ${\mathcal B}$ are categories,
$\bigl\{{\mathcal A}, {\mathcal B}\bigr\}$ is the category of covariant functors
from ${\mathcal A}$ to ${\mathcal B}$ (see \cite{btwentyfive} page 63).
\begin{xDefinition}
A category ${\mathcal A}$ is \emph{weakly regular} with respect to an
index set ${\mathcal I}$ provided:
\begin{enumerate}
\item[1)] ${\mathcal A}$ has products and zero objects.
\item[2)] Let ${\mathcal F}({\mathcal I})=\bigl\{ T\ \vert\ T\subseteq {\mathcal I}
\hbox{ and $T$ is finite}\bigr\}$.
If $\{ G_i \}$ is an object in ${\mathcal A}^{\mathcal I}$,
each $T\in{\mathcal F}({\mathcal I})$ induces an endomorphism of $\{ G_i \}$ by
\[\begin{cases}G_i\ \to\ G_i& \text{if the identity if } i\notin T\\
G_i\ \to\ G_i& \text{if the zero map if }i\in T\ .\\\end{cases}
\hbox{This induces
a unique map}\]
$\displaystyle
X_T\colon \mathop{\bigtimes}_{i\in{\mathcal I}}G_i \longrightarrow
\mathop{\bigtimes}_{i\in{\mathcal I}}G_i$.
We require that there exist an object
$\displaystyle\mathop{\mu}_{i\in{\mathcal I}}(G_i)$ and a map
$\displaystyle
\mathop{\bigtimes}_{i\in{\mathcal I}}G_i\longrightarrow
\mathop{\mu}_{i\in{\mathcal I}}(G_i)$ which is the coequalizer of the
family of morphisms $X_T$ for all $T\in{\mathcal F}({\mathcal I})$.
\end{enumerate}
\end{xDefinition}
We easily check
\BEGIN{L.1.2.1}
$\mu\colon {\mathcal A}^{\mathcal I}\to {\mathcal A}$ is a functor when ${\mathcal A}$
is a weakly regular category with respect to ${\mathcal I}$.\qed
\end{Lemma}
\begin{xExamples}
The categories of groups, abelian groups, rings and pointed sets are all
weakly regular with respect to any index set ${\mathcal I}$.
$\mu$ is each case is given as follows.
We define an equivalence relation $R$ on
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}G_i$ by $x R y$ \iff\
(the $i^{\text{\thx}}$ component of $x$)=(the $i^{\text{\thx}}$ component of $y$)
for all but finitely many $i\in{\mathcal I}$.
Then $\displaystyle
\mathop{\mu}_{i\in{\mathcal I}}(G_i)=\left(\mathop{\bigtimes}_{i\in{\mathcal I}}G_i\right)/R$.
\end{xExamples}
\BEGIN{L.1.2.2}
If ${\mathcal D}$ is a diagram scheme, and if ${\mathcal A}$ is a weakly regular
category with respect to ${\mathcal I}$,
then $\bigl[{\mathcal D}, {\mathcal A}\bigr]$ is also weakly regular with respect
to ${\mathcal I}$.
\end{Lemma}
\begin{proof}
$\bigl[{\mathcal D}, {\mathcal A}\bigr]$ is easily seen to have a zero object.
$\bigl[{\mathcal D}, {\mathcal A}\bigr]$ has products, for to each object in
$\bigl[{\mathcal D}, {\mathcal A}\bigr]^{\mathcal I}$, $\bigl(\{G_{i j},\{m_i\}\bigr)$,
we associate the diagram
$\displaystyle\Bigl(\ \mathop{\bigtimes}_{i\in{\mathcal I}} G_{i j},
\mathop{\bigtimes}_{i\in{\mathcal I}} m_i\Bigr)$.
It is not hard to check that this diagram has
the requisite universal properties.
To see condition 2), to $\bigl\{G_{i j}\bigr\}$ associate
$\displaystyle\mathop{\mu}_{i\in{\mathcal I}}\bigl(G_{i j}\bigr)$.
Then $\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}} m_i$
induces $\displaystyle\mathop{\mu}_{i\in{\mathcal I}}(m_i)$, so we do
get a diagram.
To show it is a coequalizer, let $X_j$ be the objects of a diagram.
Set $H_j=\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}} G_{i j}$.
We are given $g_j\colon H_j\to X_j$ which commute with the diagram
maps.
If $T_1$, $T_2\in{\mathcal F}({\mathcal I})$ we also have
$g_j\circ X_{T_1}=g_j\circ X_{T_2}$.
Hence by the universality for $\mu$ for ${\mathcal A}$, we get unique maps
$f_j\colon \displaystyle\mathop{\mu}_{i\in{\mathcal I}}(G_{i j})\to X_j$
such that
\hskip1.3in$\xymatrix@C70pt@R50pt{
H_j\ar[r]^-{g_j}\ar[d]&X_j\\
\displaystyle\mathop{\mu}_{i}(G_{i j})\ar[ru]_-{f_j}
}$
\vskip4pt\noindent
commutes.
If we have a map in ${\mathcal D}$ from $j$ to $k$, we get
\hskip1.3in$\xymatrix@C20pt@R30pt{
H_j\ar[rr]\ar[rd]\ar[dd]_-{g_j}&&H_k\ar[rd]\ar[dd]_<<<<<<<<<<<<<<{g_k}\\
&\displaystyle\mathop{\mu}_{i}(G_{i j})\ar@{.>}[rr]\ar[ld]^-{f_j}&&
\displaystyle\mathop{\mu}_{i}(G_{i k})\ar[ld]^-{f_k}\\
X_j\ar[rr]&&X_k\\
}$
\noindent with the front and back squares and both end triangles
commutative.
By the uniqueness of the map
$\displaystyle\mathop{\mu}_{i\in{\mathcal I}}\bigl(G_{i j}\bigr)\to X_k$,
the bottom square also commutes and we are done.
\end{proof}
Suppose given a functor $F\in\bigl\{{\mathcal A}, {\mathcal B}\bigr\}$.
If $F$ does not preserve products, it seems unreasonable to expect
it to behave well with respect to $\mu$, so assume $F$ preserves
products.
Then we get a natural map
$\mu\circ F^{\mathcal I}\to F\circ \mu$
($F^{\mathcal I}$ is the obvious element in
$\bigl\{{\mathcal A}^{\mathcal I},{\mathcal B}^{\mathcal I}\bigr\}$).
$F$ preserves $\mu$ \iff\ this map is an isomorphism.
Now suppose ${\mathcal A}$ is complete with respect to a diagram scheme
${\mathcal D}$.
Then we have a functor
$\displaystyle\lim_{\mathcal D}\colon \bigl[{\mathcal D}, {\mathcal A}\bigr]\ \to\
{\mathcal A}$, the limit functor (see \cite{btwentyfive}, page 44).
\begin{xDefinition}
If ${\mathcal D}$ is a diagram scheme, and if ${\mathcal A}$ is a
${\mathcal D}${--}complete, weakly regular category with respect to ${\mathcal I}$,
then we define
$\epsilon\colon \bigl[{\mathcal D}, {\mathcal A}\bigr]^{\mathcal I}\to{\mathcal A}$
to be the composite $\displaystyle\lim_{\mathcal D}\circ \mu$.
\end{xDefinition}
\BEGIN{P.1.2.1}
Let ${\mathcal D}$ be a
diagram scheme, and let ${\mathcal A}$ and ${\mathcal B}$ be
two ${\mathcal D}${--}complete, weakly regular categories
with respect to ${\mathcal I}$.
Let $F\in\bigl\{{\mathcal A}, {\mathcal B}\bigr\}$.
Then $\epsilon\colon \bigl[{\mathcal D}, {\mathcal A}\bigr]^{\mathcal I}\to{\mathcal A}$
is a functor.
If $F$ preserves products and $\displaystyle\lim_{\mathcal D}$, there is
a natural map
$\epsilon\circ F^{\mathcal I}_{\#}\to F\circ\epsilon$
(where $F_\#\colon\bigl[{\mathcal D}, {\mathcal A}\bigr]\
\to\ \bigl[{\mathcal D}, {\mathcal B}\bigr]$ is the induced functor).
If $F$ preserves $\mu$, this map is an isomorphism.
\end{Proposition}
\begin{proof}
Trivial. \end{proof}
\medskip
Unfortunately, the limit we are taking is an inverse limit,
which is notorious for causing problems.
In some cases however (and in all cases in which we shall be interested)
it is possible to give a description of $\epsilon$ as a direct limit.
In fact, we will describe the $\Delta${--}construction as a direct limit
and then investigate the relationship between the two.
\begin{xDefinition}
A lattice scheme ${\mathcal D}$ is a diagram scheme $(J, M, d)$ such that
$J$ is a partially ordered set with least upper and greatest lower
bounds for any finite subset of $J$.
We also require that $d\colon M\to J\times J$ be a monomorphism
and that ${\Imx}\ d\ =\ \{ (j,k)\in J\times J\ \vert\ j>k\}$.
To the lattice scheme ${\mathcal D}$ and the index set ${\mathcal I}$, we
associate a diagram scheme ${\mathcal D}_{\mathcal I}$ as follows
( ${\mathcal D}_{\mathcal I}$ is the diagram of
``cofinal subsequences of ${\mathcal D}$'').
If $\displaystyle\alpha\in\mathop{\bigtimes}_{i\in{\mathcal I}} J$,
define $J_\alpha= \{j\in J\ \vert\ j=
p_i(\alpha)\hbox{ for some $i\in{\mathcal I}$}\}$.
$p_i$ is just the $i^{\text{\thx}}$ projection, so $J_\alpha$ is just the subset of
$J$ we used in making up $\alpha$.
Define $\rho_\alpha\colon {\mathcal I}\ \to\ J$ by
$\rho_\alpha(i)=p_i(\alpha)$.
Set
\[J_{\mathcal I}=\bigl\{ \displaystyle\alpha\in\mathop{\bigtimes}_{i\in{\mathcal I}} J\ \vert\
\hbox{$J_\alpha$ is cofinal in $J$ and
$\displaystyle\mathop{\cup}_{j\leq k} \rho^{-1}_\alpha(j)$ is finite for all $k\in J$}\}\ .\]
(A subset of $J$ is cofinal \iff\ given any $j\in J$, there is an element
$k$ of our subset so that $k\geq j$.
$J_{\mathcal I}$ may be thought of as the set of
``locally finite, cofinal subsets'' of $J$).
\end{xDefinition}
We say $\alpha\geq\beta$ \iff\ $p_i(\alpha)\geq p_i(\beta)$ in $J$ for all
$i\in{\mathcal I}$.
Set $M_{\mathcal I}=\{ (\alpha, \beta)\in J_{\mathcal I}\times J_{\mathcal I}\ \vert\
\alpha\geq\beta\}$ and let $d_{\mathcal I}$ be the inclusion.
Given $\alpha$, $\beta\in J_{\mathcal I}$, define $\gamma\in J_{\mathcal I}$
by $p_i(\gamma)=\hbox{least upper bound of $p_i(\alpha)$
and $p_i(\beta)$.}$
(It is not hard to see $\gamma\in J_{\mathcal I}$.)
Greatest lower bounds can be constructed similarly.
Hence ${\mathcal D}_{\mathcal I}$ is also a lattice scheme.
Now if $J$ does not have any cofinal subsets of cardinality
$\leq {\text{card}}({\mathcal I})$, $J_{\mathcal I}=\emptyset$.
Since $J$ has upper bounds for finite sets, if $J$ has finite cofinal subsets,
then $J$ has cofinal subsets of cardinality $\geq{\text{card}}({\mathcal I})-N$,
where $N$ is some natural number, then the condition that
$\displaystyle\mathop{\cup}_{j\leq k} \rho^{-1}_\alpha(j)$ be finite forces
$J_{\mathcal I}=\emptyset$.
Empty diagrams are a nuisance, so we define an ${\mathcal I}${--}lattice
scheme as a lattice scheme with cofinal subsets of cardinality
$= {\text{card}}({\mathcal I})$.
We can now define
$\delta\colon \bigl[{\mathcal D}, {\mathcal A}\bigr]^{\mathcal I}\ \to\
\bigl[{\mathcal D}_{\mathcal I}, {\mathcal A}\bigr]$ as follows.
If $\bigl\{ d_i\}\in\bigl[{\mathcal D}, {\mathcal A}\bigr]^{\mathcal I}$,
$\delta(d)$ has for objects
$\displaystyle\delta_\alpha=
\mathop{\bigtimes}_{i\in{\mathcal I}}G_{i p_i(\alpha)}$,
where $G_{i j}$ is the $j^{\text{\thx}}$ object in the diagram for $d_i$
($\alpha\in J_{\mathcal I}$, $j\in J$, $i\in{\mathcal I}$).
If $\alpha\geq\beta$, we define $\delta_\alpha\ \to\ \delta_\beta$
by the maps $G_{i p_i(\alpha)}\to G_{i p_i(\beta)}$
which come from the diagram $d_i$.
We can also define maps $\displaystyle\delta_\alpha\to \mathop{\mu}_{j}$
as follows.
Map $G_{i p_i(\alpha)} \to G_{i j}$ by the unique map in
$d_i$ if $p_i(\alpha)\geq j$, and by the zero map if $j>p_i(\alpha)$.
(Notice that there are at most finitely many $i$
such that $p_i(\alpha)<j$ by
the second defining condition on $J_{\mathcal I}$.)
These maps induce a unique map
$\delta_\alpha \to \mathop{\bigtimes}_{i\in{\mathcal I}}G_{i j}$.
Composing with the projection, we get a unique map
$\displaystyle\delta_\alpha\to \mathop{\mu}_{i\in{\mathcal I}}(G_{i j})=
\mathop{\mu}_{j}$.
If $k\geq j$,\hskip20pt \lower 12pt\vtop{\hsize=1in\noindent
$\displaystyle\begin{matrix}%
\delta_\alpha&\longrightarrow&\mathop{\mu}_{j}\\
&\searrow&\big\downarrow\\
&&\mathop{\mu}_{k}\\
\end{matrix}$}\hskip20pt commutes as one easily checks.
If $\alpha\geq\beta$,
\hskip20pt \lower 12pt\vtop{\hsize=1in\noindent$\displaystyle\begin{matrix}%
\delta_\alpha\\
&&\searrow\\
\big\downarrow&&&\mathop{\mu}_{j}\\
&&\nearrow\\
\delta_\beta\\
\end{matrix}$}\hskip20pt also commutes.
\bigskip
\BEGIN{L.1.2.3}
$\delta\colon\bigl[{\mathcal D}, {\mathcal A}\bigr]^{\mathcal I}\ \to\
\bigl[{\mathcal D}_{\mathcal I}, {\mathcal A}\bigr]$ is a functor.
\end{Lemma}
\smallskip\begin{proof}
The proof is easy and can be safely left to the reader. \end{proof}
Now suppose ${\mathcal A}$ is ${\mathcal D}_{\mathcal I}${--}cocomplete.
Then we have a colimit functor $\displaystyle\colim_{{\mathcal D}_{\mathcal I}}$.
\begin{xDefinition}
If ${\mathcal D}$ is an ${\mathcal I}${--}lattice scheme, and if ${\mathcal A}$ is a
${\mathcal D}_{\mathcal I}${--}cocomplete, weakly regular category with respect
to ${\mathcal I}$, then we define
$\Delta\colon \bigl[{\mathcal D}, {\mathcal A}\bigr]^{\mathcal I}\to{\mathcal A}$
to be the composition
$\displaystyle\colim_{{\mathcal D}_{\mathcal I}}\ \circ\ \delta$.
\end{xDefinition}
\bigskip
\BEGIN{P.1.2.2}
Let ${\mathcal D}$ be a diagram scheme and let ${\mathcal A}$ and ${\mathcal B}$ be
two ${\mathcal D}_{\mathcal I}${--}cocomplete, weakly regular categories with
respect to ${\mathcal I}$.
Let $F\in\bigl\{{\mathcal A}, {\mathcal B}\bigr\}$.
Then $\Delta\colon \bigl[{\mathcal D}, {\mathcal A}\bigr]^{\mathcal I}\to{\mathcal A}$
is a functor.
There is always a natural map
$\Delta\circ F_{\#}^{\mathcal I}\to F\circ\Delta$.
If $F$ preserves products and $\displaystyle\colim_{{\mathcal D}_{\mathcal I}}$,
this map is an isomorphism.
\end{Proposition}
\smallskip\begin{proof}
Trivial. \end{proof}
The maps we constructed from
$\displaystyle\delta_\alpha\to \mathop{\mu}_{j}$ combine to give us a
natural transformation from $\Delta$ to $\epsilon$ whenever both
are defined.
We would like to study this natural transformation in order to get information
about both $\Delta$ and $\epsilon$.
A $({\mathcal D}, {\mathcal I})${--}regular category is about the most general category
in which we can do this successfully, and it includes all the examples we
have in mind.
\begin{xDefinition}
A category ${\mathcal A}$ is said to be $({\mathcal D}, {\mathcal I})${--}regular
provided
\begin{enumerate}
\item[1)] ${\mathcal A}$ is weakly regular with respect to ${\mathcal I}$
\item[2)] ${\mathcal A}$ has images and inverse images
\item[3)] There is a covariant functor $F$ from ${\mathcal A}$ to
the category of pointed sets and maps such that
\begin{enumerate}
\item[a)] $F$ preserves kernels, images, products,
limits over ${\mathcal D}$, increasing unions, and $\mu$
\item[b)] $F$ reflects kernels, images, and isomorphisms
\end{enumerate}
\item[4)] ${\mathcal A}$ is ${\mathcal D}${--}complete and
${\mathcal D}_{\mathcal I}${--}cocomplete
\item[5)] ${\mathcal I}$ is countable
\end{enumerate}
\end{xDefinition}
\bigskip
\begin{xExamples}
The categories of groups, abelian groups, rings and pointed sets are
all $({\mathcal D}, {\mathcal I})${--}regular for any ${\mathcal I}${--}lattice scheme
\footnote[1]{For which ${\mathcal I}$ is countable.}.
The functor $F$ is just the forgetful functor.
\end{xExamples}
\bigskip
\BEGIN{L.1.2.4}
Let ${\mathcal A}$ be a $({\mathcal D}, {\mathcal I})${--}regular category.
Then $\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}$ and
$\mu$ preserve kernels and images.
\end{Lemma}
\begin{proof}
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}$ is known to preserve
kernels (Mitchell \cite{btwentyfive}, page 67, Corollary 12.3).
Since $F$ preserves images, if ${\Imx}(f)$ is the image of
$A\ \RA{f}\ B$, then $F(A)$ is onto $F\bigl({\Imx}(f)\bigr)$
and $F\bigl({\Imx}(f)\bigr)$ injects into $F(B)$.
Let $K_i$ be the image of $A_i\ \RA{\ f_i\ }\ B_i$.
Then, since $F$ preserves products,
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}} K_i\ \to\
\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}B_i$ is a monomorphism
since $\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}$ is a monofunctor.
Since $F$ preserves monomorphisms,
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}} F(K_i)\ \to\
\mathop{\bigtimes}_{i\in{\mathcal I}} F(B_i)$ is seen to be a monomorphism
as $F$ also preserves products.
Since $F$ reflects images,
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}} K_i$ is the image of
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}A_i\ \to\
\mathop{\bigtimes}_{i\in{\mathcal I}} B_i$.
Let $K_i\ \to\ A_i\ \to\ B_i$ be kernels.
Then
\[\begin{matrix}%
\bigtimes K_i&\RA{\ h\ }& \bigtimes A_i&\RA{\ g\ }&\bigtimes B_i\\
\big\downarrow&&\big\downarrow&&\big\downarrow\\
\mu(K_i)&\RA{\ \ \ }&\mu(A_i)&\RA{\ \ \ }&\mu(B_i)\\
\end{matrix}\leqno(\diagramOne)\]
commutes.
Since $F$ reflects kernels, we need only show that
$F\bigl(\mu(K_i)\bigr)$ injects into $F\bigl(\mu(A_i)\bigr)$ and is onto
$F(f)^{-1}(0)$.
Since $F$ preserves $\mu$, we may equally consider
$\mu\bigl(F(K_i)\bigr)$, etc.
Since $F$ preserves products, we may as well assume the diagram
$(\diagramOne)$ is in the category of pointed sets.
We show $\mu(K_i)$ is onto $f^{-1}(0)$.
Let $x\in f^{-1}(0)\subseteq \mu(A_i)$.
Lift $x$ to $y\in\bigtimes A_i$, which we may do since
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}$ is onto
$\displaystyle\mathop{\mu}_{i\in{\mathcal I}}$ in the category of
pointed sets.
Now $g(y)\in\bigtimes B_i$ can have only finitely many
non{--}zero components since it goes to $0$ in $\mu(B_i)$.
Define $\Bar y$ by
\[p_i(\Bar y)=\begin{cases}p_i(y)& \text{if } p_i\bigl(g(y)\bigr) = 0\\
0& \text{if }p_i\bigl(g(y)\bigr) \neq 0\\\end{cases}\ .\]
Then $\Bar y$ also lifts $x$ and $g(\Bar y)=0$.
There is a $z\in\bigtimes K_i$ such that $h(z)=\Bar y$, so
$\mu(K_i)$ maps onto $f^{-1}(0)$.
A similar argument shows $\mu(K_i)$ injects into $\mu(A_i)$.
Hence $\mu$ preserves kernels.
Now let $K_i$ be the image of $A_i\ \to\ B_i$.
Then
\[\begin{matrix}%
\bigtimes A_i&\RA{\ \ \ }& \bigtimes K_i&\RA{\ \ \ }&\bigtimes B_i\\
\big\downarrow&&\big\downarrow&&\big\downarrow\\
\mu(A_i)&\RA{\ \ \ }&\mu(K_i)&\RA{\ \ \ }&\mu(B_i)\\
\end{matrix}\]
commutes.
By general nonsense, it suffices to prove the result assuming
we are working in the category of pointed sets.
Since $\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}$ preserves
images, $\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}} A_i\ \to\
\mathop{\bigtimes}_{i\in{\mathcal I}} K_i$ is onto, so
$\mu(A_i)\ \to\ \mu(K_i)$ is easily seen to be onto.
Since $\mu$ preserves kernels, $\mu(K_i)$ injects into $\mu(B_i)$,
so $\mu(K_i)$ is the image of $\mu(A_i)\ \to\ \mu(B_i)$. \end{proof}
\bigskip
\BEGIN{T.1.2.1}
Let ${\mathcal A}$ be a $({\mathcal D}, {\mathcal I})${--}regular category.
Then $\epsilon$ preserves kernels and images.
\end{Theorem}
\begin{proof}
By Mitchell \cite{btwentyfive} (page 67, Corollary 12.2)
$\displaystyle\lim_{\mathcal D}$ preserves kernels so
$\epsilon = \displaystyle\lim_{\mathcal D}\ \circ\ \mu$ also does
using \fullRef{L.1.2.4} and general nonsense.
Now let $K_{i j}$ be the image of $A_{i j}\ \to\ B_{i j}$.
We claim that, if $x\in \epsilon(K_{i j})$,
then there exists $\alpha\in J_{\mathcal I}$ such that
$x$ is in the image of $\delta_\alpha(K_{i j})$.
Assuming this for now we proceed as follows.
Since $\delta_\alpha(K_{i j})=
\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}} K_{i p_i(\alpha)}$,
$\delta_\alpha$ preserves kernels and images by
\fullRef{L.1.2.4}.
Hence
\[\begin{matrix}%
\delta_\alpha(A_{i j})&\RA{\hskip 10pt}&\delta_\alpha(K_{i j})
&\RA{\hskip 10pt}&\delta_\alpha(B_{i j})\\
\big\downarrow&&\big\downarrow&&\big\downarrow\\
\epsilon(A_{i j})&\RA{\hskip 10pt}&\epsilon(K_{i j})
&\RA{\hskip 10pt}&\epsilon(B_{i j})\\
\end{matrix}\]
commutes and $\delta_\alpha(K_{i j})$ is the image of
$\delta_\alpha(A_{i j})\ \to\ \delta_\alpha(B_{i j})$.
By the usual abstract nonsense, we may as well assume we are in the
category of pointed sets (note $F$ preserves $\epsilon$ by
\fullRef{P.1.2.1}.
Now using our claim we can easily get
$\epsilon(A_{i j})\ \to\ \epsilon(K_{i j})$ is onto.
$\epsilon(K_{i j})\ \to\ \epsilon(B_{i j})$ injects since $\epsilon$
preserves kernels.
Hence $\epsilon(K_{i j})$ is the image of
$\epsilon(A_{i j})\ \to\ \epsilon(B_{i j})$. \end{proof}
We prove a stronger version of our claim than we have yet used.
\BEGIN{L.1.2.5}
Let ${\mathcal A}$ be the category of pointed sets.
Let $\{G_{i j}\}$ be an object in $\bigl[{\mathcal D}, {\mathcal A}\bigr]^{\mathcal I}$.
Then if $x\in\epsilon(G_{i j})$, there is an $\alpha\in J_{\mathcal I}$
such that $\delta_\alpha(G_{i j})$ contains $x$ in its image.
If $y$, $z\in\delta_\alpha(G_{i j})$ both hit $x$, then there is a
$\beta\leq\alpha$ such that, in $\delta_\beta(G_{i j})$, the images of
$y$ and $z$ differ in only finitely many coordinates.
In fact, if there is a $j\in J$ such that $j\leq p_i(\alpha)$ for all
$i\in{\mathcal I}$ and such that $y$ and $z$ agree in
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}G_{i j}$, then $\beta$
can be chosen so that $y=z$ in $\delta_\beta(G_{i j})$.
\end{Lemma}
\begin{proof}
If $x\in \epsilon(G_{i j})$, there exist unique
$a_j\in\displaystyle\mathop{\mu}_{i\in{\mathcal I}}(G_{i j})$ such that
$x$ hits $a_j$.
Since $\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}$ is onto
$\displaystyle\mathop{\mu}_{i\in{\mathcal I}}$, we may lift $a_j$ to
$b_j\in\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}} G_{i j}$.
Since $J$ has countable cofinal subsets, let the natural numbers
$j=1$, $2$, \dots\ be one such.
Since ${\mathcal I}$ is countable (and infinite or our result is easy) we also
assume it to be the natural numbers.
Now look at $b_2$ and $b_1$. Since they agree in $\mu(G_{i1})$,
$b_2$ projected into $\bigtimes G_{i j}$ differs from $b_1$ in only
finitely many coordinates.
Let ${\mathcal I}_1\subseteq{\mathcal I}$ be the finite subset which indexes these
unequal coordinates, together with the element $1\in{\mathcal I}$.
Next look at the pairs $(b_3,b_2)$ and $(b_3,b_1)$.
As before, projected into $\bigtimes G_{i2}$, $b_3$ and $b_2$ agree
in all but finitely many coordinates.
In $\bigtimes G_{i1}$, $b_3$ and $b_1$ differ in only
finitely many coordinates.
Set ${\mathcal I}_2\subseteq{\mathcal I}$ to be the finite subset of ${\mathcal I}$
which indexes the unequal coordinates of $(b_3, b_2)$ or $(b_3,b_1)$
which lie in ${\mathcal I}-{\mathcal I}_1$, together with the smallest integer in
${\mathcal I}-{\mathcal I}_1$.
Define ${\mathcal I}_k$ to be the finite subset of ${\mathcal I}$ which indexes
the unequal coordinates of
$(b_k, b_{k-1})$, \dots, $(b_k, b_2)$, $(b_k,b_1)$ which lie in
${\mathcal I}-\bigl({\mathcal I}_{k-1}\ \cup\ \cdots\ \cup\
{\mathcal I}_2\ \cup\ {\mathcal I}_1\bigr)$,
together with the smallest integer in
${\mathcal I}-\bigl({\mathcal I}_{k-1}\ \cup\ \cdots\ \cup\
{\mathcal I}_2\ \cup\ {\mathcal I}_1\bigr)$.
Then $\displaystyle{\mathcal I}=\mathop{\cup}_{k=1}^\infty {\mathcal I}_k$
as a disjoint union.
Define $\alpha$ by $p_i(\alpha)=k$, where $i\in{\mathcal I}_k$.
Since ${\mathcal I}$ is countable, but not finite, and since each ${\mathcal I}_k$
is finite, $\alpha\in J_{\mathcal I}$.
Define $y\in\delta(G_{i j})$ by $p_i(y)=p_i(b_{p_i(\alpha)})$.
A chase through the definitions shows $y$ hits each $a_j$ through the
map $\delta_\alpha(G_{i j})\ \to\
\displaystyle\mathop{\mu}_{i\in{\mathcal I}}(G_{i j})$.
Thus $y$ hits $x$ in $\epsilon(G_{i j})$.
Now suppose $y$, $z\in \delta_\alpha(G_{i j})$ both map to $x$.
Then they map to the same element in each
$\displaystyle\mathop{\mu}_{i\in{\mathcal I}}(G_{i j})$.
Let $a_j$ be the image of $y$ in
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}G_{i j}$ under the map
$\delta_\alpha(G_{i j})\ \to\
\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}G_{i j}$
which we defined just before \fullRef{L.1.2.3}.
Set $b_j$ to be the image of $z$ in
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}} G_{i j}$.
Then $a_j$ and $b_j$ differ in only finitely many coordinates.
Let ${\mathcal I}_1$ be the finite subset of ${\mathcal I}$ which indexes the
unequal coordinates of $a_1$ and $b_1$.
If there is a $j\leq p_i(\alpha)$ for all $i\in{\mathcal I}$ such that $y$
and $z$ agree in $\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}} G_{i j}$,
we may assume $j=1$, so $a_1=b_1$, and ${\mathcal I}_1=\emptyset$.
Define ${\mathcal I}_k$ as the finite subset of ${\mathcal I}$ which indexes the
unequal coordinates of $(a_k,b_k)$ which lie in
${\mathcal I}-\bigl({\mathcal I}_{k-1}\ \cup\ \cdots\ \cup\
{\mathcal I}_2\ \cup\ {\mathcal I}_1\bigr)$.
Define $\beta$ by
\[p_i(\beta)=\begin{cases}k-1& \text{if }i\in{\mathcal I}_k \text{ for some } k\geq2\\
p_i(\alpha)& \text{if }i\notin{\mathcal I}_k \text{ for any }k\geq2\\\end{cases}\ .\]
Note $p_i(\beta)\leq p_i(\alpha)$, since $i\in{\mathcal I}_k$, this says
$p_i(a_k)\neq p_i(b_k)$.
But if $k>p_i(\alpha)$, $p_i(a_k)=0=p_i(b_k)$
by the definition of our map from $\delta_\alpha$ to
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}$.
Hence $k\leq p_i(\alpha)$, so $p_i(\beta)\leq p_i(\alpha)$.
Let $\Bar y$ be the projection of $y$ into $\delta_\beta(G_{i j})$,
and let $\Bar z$ be the projection of $z$ into $\delta_\beta(G_{i j})$.
$p_i(\Bar y)= p_i(a_{p_i(\beta)})$ and $p_i(\Bar z)= p_i(b_{p_i(\beta)})$.
If $p_i(a_{p_i(\beta)})\neq p_i(b_{p_i(\beta)})$, then $i\notin{\mathcal I}_k$
for any $k\geq2$, since $i\in{\mathcal I}_k$ for $k\geq2$ says that
$p_i(a_k)\neq p_i(b_k)$ but $p_i(a_{k-1})=p_i(b_{k-1})$.
If $i\notin{\mathcal I}_k$ for any $k$, it says that $p_i(y)=p_i(z)$.
Thus $p_i(\Bar y)=p_i(\Bar z)$ if $i\notin {\mathcal I}_1$.
Hence they agree for all but finitely many $i\in{\mathcal I}$.
In fact, if ${\mathcal I}_1=\emptyset$, $\Bar y=\Bar z$.
\end{proof}
We can now describe $\epsilon(G_{i j})$ as a colimit (direct limit).
Let $\displaystyle\mathop{\mu}_{\alpha}(G_{i j})$ be the $\mu$
functor applied to $\bigl\{ G_{i\> p_i(\alpha)}\bigr\}$.
Then the map $\delta_\alpha(G_{i j})\ \to\ \epsilon(G_{i j})$ factors
through $\displaystyle\mathop{\mu}_{\alpha}(G_{i j})$.
\bigskip
\BEGIN{T.1.2.2}
Let ${\mathcal A}$ be a $({\mathcal D}, {\mathcal I})${--}regular category.
Then the natural map \\
$\displaystyle\colim_{{\mathcal D}_{\mathcal I}}\ \mu_{\alpha}\ \to\ \epsilon$
is an isomorphism.
Hence $\epsilon$ is a cokernel, kernel preserving functor.
\end{Theorem}
\begin{proof}
Let us first show $F$ preserves $\displaystyle\colim_{{\mathcal D}_{\mathcal I}}$;
i.e. we must show that the natural map
\[\displaystyle
\colim\ F(A_\alpha)\ \RA{f}\ F(\colim_{{\mathcal D}_{\mathcal I}} A_\alpha)\]
is an isomorphism.
To do this, we first compute ${\Imx}(f)$.
If ${\Imx}(f_\alpha)$ is the image of $F(A_\alpha)\ \to\
\colim\> F(A_\alpha)\ \to\ F(\colim\> A_\alpha)$, then by Mitchell
\cite{btwentyfive} (Proposition 2.8, page 46),
$\displaystyle{\Imx}(f)=\mathop{\cup}_\alpha\ {\Imx}(f_\alpha)$.
Let ${\Imx}(g_\alpha)$ be the image of
$A_\alpha\ \to\ \colim A_\alpha$.
Then, since $F$ preserves images, $F\bigl({\Imx}(g_\alpha)\bigr)=
{\Imx}(f_\alpha)$, so
$\displaystyle\mathop{\cup}_\alpha {\Imx}(f_\alpha) =
\mathop{\cup}_\alpha F\bigl({\Imx}(g_\alpha)\bigr)$.
Now $\{\alpha\}$ has a cofinal subsequence (which is countable and, if
${\mathcal I}$ is finite, it is also finite) $\{\alpha_i\}$ such that
$\alpha_o < \alpha_1< \cdots < \alpha_n < \cdots$\ .
Therefore
$\displaystyle\mathop{\cup}_\alpha {\Imx}(f_\alpha) =
\mathop{\cup}_{i=0}^\infty{\Imx}(f_{\alpha_i}) =
\mathop{\cup}_{i=0}^\infty F\bigl({\Imx}(g_{\alpha_i})\bigr)
$
since $\{\alpha_i\}$ is cofinal.
Again by Mitchell \cite{btwentyfive} (Proposition 2.8, page 46),
$\colim A_\alpha =
\mathop{\cup}_\alpha {\Imx}(g_\alpha)=
\mathop{\cup}_{i=0}^\infty{\Imx}(g_{\alpha_i})
$.
Thus $F(\colim\ A_\alpha)=
F\bigl(\mathop{\cup}_{i=0}^\infty{\Imx}(g_{\alpha_i})\bigr)$.
Since ${\mathcal A}$ is a $({\mathcal D}, {\mathcal I})${--}regular category,
the natural map
\[\displaystyle\mathop{\cup}_{i=0}^\infty F\bigl({\Imx}(g_{\alpha_i})
\bigr)\subseteq F\bigl(
\mathop{\cup}_{i=0}^\infty {\Imx}(g_{\alpha_i})\bigr)
\] is an isomorphism.
Thus the map
$\displaystyle\mathop{\cup}_\alpha{\Imx}(f_\alpha)\subseteq
F\bigl(\mathop{\cup}_\alpha{\Imx}(g_\alpha)\bigr)$
is an isomorphism.
But this map is just the natural map
$\displaystyle\colim_{{\mathcal D}_{\mathcal I}}\ F(A_\alpha)\ \to\
F(\colim_{{\mathcal D}_{\mathcal I}}\ A_\alpha)$.
The natural map $\colim \mu_\alpha\ \to\ \epsilon$
is the map which comes from the maps $\mu_\alpha\ \to\ \epsilon$.
To show it is an isomorphism, it is enough to show it is for pointed sets
by the result above and the fact that $F$ reflects isomorphisms.
But this is exactly what \fullRef{L.1.2.5} says.
Now $\epsilon$ preserves kernels by \fullRef{T.1.2.1},
and it preserves cokernels since colimits preserve cokernels by
Mitchell \cite{btwentyfive} (page 67, Corollary 12.2 dualized).
\end{proof}
\medskip
\BEGIN{C.1.2.2.1}
Let $\bigl\{ G_{i j}^n\bigr\}\in\bigl[{\mathcal D}, {\mathcal A}\bigr]^{\mathcal I}$
be a collection of exact sequences in a $({\mathcal D}, {\mathcal I})${--}regular
category ${\mathcal A}$ (i.e. there are maps
$f^n_{i j}\ \colon\ G^n_{i j}\ \to\ G^{n-1}_{i j}$
which are maps of diagrams
such that ${\Imx}(f^n_{i j})={\kerx}(f^{n-1}_{i j})$\ ).
Then the sequence
\[\cdots \to\ \epsilon(G^n_{i j})\ \RA{\ \epsilon(f^n_{i j})\ }\
\epsilon(G^{n-1}_{i j})\ \to\ \cdots\]
is also exact.
\end{Corollary}
\medskip
\BEGIN{C.1.2.2.2}
Let ${\mathcal A}$ be a $({\mathcal D}, {\mathcal I})${--}regular abelian category.
Let $\bigl\{ G^\ast_{i j}, f^\ast_{i j}\bigr\}$
be a collection of chain complexes in $\bigl[{\mathcal D}, {\mathcal I}\bigr]^{\mathcal I}$.
Then $\bigl\{\epsilon(G^\ast_{i j}), \epsilon(f^\ast_{i j})\bigr\}$ is
a chain complex, and
$H_\ast\bigl(\epsilon(G^\ast_{i j})\bigr) =
\epsilon\bigl(H_\ast(G^\ast_{i j})\bigr)$, where $H_\ast$ is the homology
functor (see Mitchell \cite{btwentyfive}, page 152).
\end{Corollary}
\renewcommand{\proofname}{Proofs}
\begin{proof}
The first corollary is easily seen to be true.
(It is, in fact, a corollary of \fullRef{T.1.2.1}.)
The second corollary is almost as easy.
If $\bigl\{ Z^n_{i j}\bigr\}$ are the $n${--}cycles, and if
$\bigl\{ B^{n+1}_{i j}\bigr\}$ are the $(n+1)${--}boundaries,
$0\to\ B^{n+1}_{i j}\ \to\ Z^n_{i j}\ \to\ H_n(G^\ast_{i j})\to0$
is exact.
Applying $\epsilon$, we get
$0\to\ \epsilon(B^{n+1}_{i j})\ \to\ \epsilon(Z^n_{i j})\ \to\
\epsilon\bigl(H_n(G^\ast_{i j})\bigr)\to0$
is exact.
But as $\epsilon$ preserves kernels and images,
$\epsilon\bigl(Z^n_{i j}\bigr)$ is the collection of $n${--}cycles for
$\epsilon(G^\ast_{i j})$ and
$\epsilon\bigl(B^{n+1}_{i j}\bigr)$ is
the collection of $(n+1)${--}boundaries.
Hence
$H_\ast\bigl(\epsilon(G^\ast_{i j})\bigr)\ \to\
\epsilon\bigl(H_n(G^\ast_{i j})\bigr)$ is an isomorphism.
\end{proof}
\renewcommand{\proofname}{Proof}
Now suppose $J$ has a unique minimal element $j_0$.
Then we get a square
\[\begin{matrix}%
\Delta(G_{i j})&\RA{\hskip1in}&\displaystyle \mathop{\bigtimes}_{i\in{\mathcal I}} G_{i j_0}\\
\Bigg\downarrow&&\Bigg\downarrow\\
\epsilon(G_{i j})&\RA{\hskip1in}&\displaystyle\mathop{\mu}_{i\in{\mathcal I}}(G_{i j_0})\\
\end{matrix}\]
\BEGIN{T.1.2.3}
In a $({\mathcal D}, {\mathcal I})${--}regular category, the above diagram is a
pullback in the category of pointed sets, so if $F$ reflects pullbacks
the above square is a pullback.
\end{Theorem}
\begin{proof}
As we showed in the proof of \fullRef{T.1.2.2}
that $F$ and $\displaystyle\colim_{{\mathcal D}_{\mathcal I}}$ commute, we
have $F\bigl(\Delta(G_{i j})\bigr) = \Delta\bigl(F(G_{i j})\bigr)$,
so we may work in the category of pointed sets.
The omnipresent \fullRef{L.1.2.5} can be used
to show the above square is a pullback.
The pullback is the subset of
$\epsilon(G_{i j})\times\bigl(\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}
G_{i j_0}\bigr)$
consisting of pairs which project to the same element in
$\displaystyle\mathop{\mu}_{i\in{\mathcal I}}(G_{i j_0})$.
Given any element, $x$, in $\epsilon(G_{i j})$
we can find $\alpha\in J_{\mathcal I}$ such that
the element is in the image of $\delta_\alpha(G_{i j})$.
Lift the image of $x$ in $\displaystyle\mathop{\mu}_{i\in{\mathcal I}}(G_{i j_0})$
to $y\in\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}}G_{i j_0}$.
Let $z\in\delta_\alpha(G_{i j_0})$ be an element which hits $x$.
Then $y$ pushed into $\bigtimes G_{i j_0}$ and $z$ agree, except in
finitely many places.
It is then easy to find $\beta\in J_{\mathcal I}$ with $\beta\leq\alpha$
and an element $q\in\delta_\beta(G_{i j})$such that $q$ hits $x$
and $y$.
This says precisely that our square is a pullback.
\end{proof}
\begin{xRemarks}
In all our examples, $F$ reflects pullbacks. The analogues of
Corollaries \shortFullRef{C.1.2.2.1}
and \shortFullRef{C.1.2.2.2}
may be stated and proved by the reader for the $\Delta$ functor.
\end{xRemarks}
\BEGIN{T.1.2.4}
In a $({\mathcal D}, {\mathcal I})${--}regular category, $\epsilon(G_{i j})=0$
\iff\ given any $j\in J$ there exists a $k\geq j$ such that
$G_{i k}\ \to\ G_{i j}$ is the zero map for all but finitely many $i$.
\end{Theorem}
\begin{proof}
Suppose given $j$ we can find such a $k$.
The we can produce a cofinal set of $j$'s, $j_0\leq j_1\leq \cdots$\ ,
such that the map $\mu(G_{i j_k})\ \to\ \mu(G_{i j_{k-1}})$ is
the zero map.
Hence $\epsilon=0$.
Conversely, suppose for some $j_0$ that no such $k$ exists.
This means that for every $k\geq j_0$ there are infinitely many
$i$ for which $G_{i k}\ \to\ G_{i j_0}$ is not the zero map.
As usual, it suffices to prove the result for pointed sets,
so assume we have $Z_{i k}\in G_{i k}$ which goes non{--}zero
into $G_{i j_0}$.
Pick $j_0\leq j_1\leq j_2\leq \cdots$ a countable cofinal subsequence
of $J$.
We define an element $\alpha$ of ${\mathcal D}_{\mathcal I}$ as follows.
Well order ${\mathcal I}$.
Then $\alpha(i)=j_0$ until we hit the first element of ${\mathcal I}$ for
which a $Z_{i j_k}$ is defined.
Set $\alpha(i)=j_k$ for this $i$ and continue defining $\alpha(i)=j_k$
until we hit the next element of ${\mathcal I}$ for which a $Z_{i j_{k_1}}$
is defined with $k_1\geq k$.
Set $\alpha(i)=j_{k_1}$ until we hit the next $Z_{i j_{k_2}}$
with $k_2\geq k_1$.
Continuing in this fashion is seen to give an element of ${\mathcal D}_{\mathcal I}$.
Define $Z_\alpha$ by $Z_{i\>\alpha(i)}=0$ unless $i$ is one of the
distinguished elements of ${\mathcal I}$, in which case set
$Z_{i\>\alpha(i)}=Z_{i j_k}$, where $j_k=\alpha(i)$.
Then $Z_\alpha\in\Delta_\alpha(G_{i j})$.
It is non{--}zero in $\displaystyle\mathop{\mu}_{i\in{\mathcal I}}(G_{i j_0})$
by construction, so $\epsilon(G_{i j})\neq0$.
\end{proof}
\bigskip
\section{Proper homotopy functors and their relations}
\newHead{I.3}
\medskip
We begin by clarifying the concept of an ordinary homotopy functor.
A homotopy functor is a functor, $h$, from
the category of pointed topological spaces to some category, ${\mathcal A}$.
Given a space $X$ and two base points $p_1$ and $p_2$, and a path
$\lambda$ from $p_1$ to $p_2$, there is a natural transformation
$\alpha_\lambda\colon h(X,p_1)\ \to\ h(X,p_2)$ which is an
isomorphism and depends only on the homotopy class of $\lambda$
rel end points.
Furthermore, $h(X,p)\ \to\ h(X\times I, p\times t)$ given by
$x\mapsto (x,t)$ is an isomorphism for $t=0$ and $1$.
For any homotopy functor we are going to associate a proper homotopy
functor defined on the category of homogamous spaces and countable
sets of locally finite irreducible base points.
To be able to do this in the generality we need, we shall have to digress
momentarily to discuss the concept of a covering functor.
Let $X$ be a homogamous space, and let $\{ x_i\}$ be a countable,
locally finite, irreducible set of base points for $X$. (From now on we
write just ``set of base points'' for ``countable, locally finite, irreducible
set of base points.'')
Let ${\mathcal D}_X$ be some naturally defined collection of subsets of $X$
(by naturally defined we mean that if $f\colon X \to Y$ is a proper map,
$f^{-1}{\mathcal D}_Y\subseteq{\mathcal D}_X$).
${\mathcal D}_X$ is a diagram with arrows being inclusion maps.
Assume ${\mathcal D}_X$ is an $\{x_i\}${--}lattice,
and assume $\emptyset\in{\mathcal D}_X$.
\medskip
\begin{xDefinition}
A \emph{covering functor} for ${\mathcal D}_X$ is a functor, $S$, which
assigns to each $\pi_1(X-C,x_i)$ a subgroup $S\pi_1(X-C,x_i)$
subject to
\[\begin{matrix}%
S\pi_1(X-C,x_i)&\subseteq& \pi_1(X-C,x_i)\\
\Big\downarrow&&\Big\downarrow\\
S\pi_1(X-D,x_i)&\subseteq& \pi_1(X-D,x_i)\\
\end{matrix}\]
commutes whenever $D\subseteq C$, $x_i\notin C$, and where the
vertical maps are induced by inclusion ($C$ and $D$ are any elements
of ${\mathcal D}_X$).
\end{xDefinition}
\medskip
\begin{xRemarks}
We have two examples for ${\mathcal D}_X$ in mind.
In this section we can use the set of all closed, compact subsets
of $X$ for ${\mathcal D}_X$.
For cohomology however, we will have to use
the set of open subsets of $X$ with compact closure for ${\mathcal D}_X$.
\end{xRemarks}
\begin{xExamples}
There are three useful examples we shall define.
\begin{enumerate}
\item[1)] no covering functor (the subgroup is the whole group)
\item[2)] the universal covering functor (the subgroup is the
zero group)
\item[3)] the universal cover of $X$ but no more covering functor
(the subgroup is the kernel of
$\pi_1(X-C,x_i)\ \to\ \pi_1(X,x_i)$\ ).
\end{enumerate}
\end{xExamples}
\medskip
\begin{xDefinition}
A compatible covering functor for ${\mathcal D}_X$ is a covering functor
$S$ such that, for any $C\in{\mathcal D}_X$, the cover of the component
of $X-C$ containing $x_i$ corresponding to $S\pi_1(X-C,x_i)$ exists.
We write $(X,\coverFA{})$ for
a compatible covering functor for ${\mathcal D}_X$
(which is inferred from context) to denote a collection of pointed spaces
$\bigl( (\coverFA{X-C})^i,\hat x_i\bigr)$, where $(\coverFA{X-C})^i$
is the covering space of the component of $X-C$ containing $x_i$, and
$\hat x_i$ is a lift of $x_i$ to this cover such that
$\pi_1\bigl((\coverFA{X-C})^i,\hat x_i\bigr)=S\pi_1(X-C,x_i)$.
Notice this notation is mildly ambiguous since if we change the $\hat x_i$
we get a different object.
As the two objects are homeomorphic this tends to cause no problems
so we use the more compact notation.
We say $(X,\coverFA{})\leq (X,\coverFB{})$
provided the subgroup of $\pi_1(X-C,x_i)$ corresponding to
$ \coverFB{}$
contains the one corresponding to $\coverFA{}$.
Hence any $(X,\coverFA{})\leq (X, {{\text{no\ cover}}})$,
and if the universal covering functor is compatible with ${\mathcal D}_X$,
$(X, {\text{universal\ cover}}) \leq (X,\coverFA{})$.
\end{xDefinition}
Now the no covering functor is compatible with any ${\mathcal D}_X$.
If $X$ is semi{--}locally 1{--}connected, the universal cover of $X$
but no more is compatible with any ${\mathcal D}_X$.
If ${\mathcal D}_X$ is the collection of closed, compact subsets of $X$,
and if $X$ is locally 1{--}connected, the universal covering functor is
compatible with ${\mathcal D}_X$, as is any other covering functor.
Hence a CW complex is compatible with any covering functor
(see Lundell and Weingram \cite{btwentyone} page 67, Theorem 6.6)
for ${\mathcal D}_X$.
We can now describe our construction.
Let $(X,\coverFA{})$ be a covering functor for $X$.
Assume from now on that our homotopy functor takes values in a
$({\mathcal D}_X, \{x_i\})${--}regular category for all homogamous $X$ with
base points $\{x_i\}$.
We apply the $\epsilon$ and $\Delta$ constructions to the collection
\[G_{i C}=\begin{cases}h\bigl((\coverFA{X-C})^i, \hat x_i\bigr)& \text{if } x_i\in X-C\\
0&\text{if }x_i\notin X-C\ .\end{cases}\]
If $D\subseteq C$ there is a unique map
$\bigl((\coverFA{X-D})^i, \hat x_i\bigr)\ \to\
\bigl((\coverFA{X-C})^i, \hat x_i\bigr)$
if $x_i\in X-D$ by taking the lift of the
inclusion which takes $\hat x_i$ in
$(\coverFA{X-D})^i$ to $\hat x_i$ in
$(\coverFA{X-C})^i$.
Hence we get a map $G_{i D}\ \to\ G_{i C}$.
We denote these groups by
$\epsilon(X\Colon h,\{\hat x_i\},\coverFA{})$ and
$\Delta(X\Colon h,\{\hat x_i\},\coverFA{}\ )$.
\BEGIN{T.1.3.1}
Let $\{ x_i\}$ and $\{ y_i\}$ be two sets of base points for $X$,
$X$ homogamous with countable base points.
Let $\Lambda$ be a locally finite collection of paths giving an equivalence
between $\{x_i\}$ and $\{y_i\}$.
Then there are natural transformations
$\alpha_\Lambda\colon
\epsilon(X\Colon h,\{\hat x_i\},\coverFA{})\ \to\
\epsilon(X\Colon h,\{\hat y_i\},\coverFB{})$ and
$\alpha_\Lambda\colon
\Delta(X\Colon h,\{\hat x_i\},\coverFA{})\ \to\
\Delta(X\Colon h,\{\hat y_i\},\coverFB{})$
which are isomorphisms and depend only on the proper homotopy
class of $\Lambda$ rel end points. ($\coverFB{}$ is the covering
functor induced by the set of paths $\Lambda$.)
\end{Theorem}
\begin{proof}
Define $\alpha_\Lambda$ as follows.
By relabeling if necessary we may assume $x_i$ goes to $y_i$
by a path in $\Lambda$.
Map $h\bigl((\coverFA{X-C})^i, \hat x_i\bigr)$ to
$h\bigl((\coverFB{X-C})^i, \hat y_i\bigr)$ by the zero map if the path from
$x_i$ to $y_i$ hits $C$.
If the path misses $C$, map $h\bigl((\coverFA{X-C})^i, \hat x_i\bigr)$ to
$h\bigl((\coverFB{X-C})^i, \hat y_i\bigr)$ by lifting the path from
$x_i$ to $y_i$ into $(\coverFA{X-C})^i$ beginning at $\hat x_i$, say
it now ends at $z$, and then map $\bigl((\coverFA{X-C})^i,z\bigr)$
to $\bigl((\coverFB{X-C})^i,\hat y_i\bigr)$ by the unique homeomorphism
covering the identity which takes $z$ to $\hat y_i$.
This defines a homomorphism $\alpha_\Lambda$ on $\epsilon$
and $\Delta$.
If by $\Lambda^{-1}$ we mean the collection of paths from $y_i$ to $x_i$
given by the inverse of the path from $x_i$ to $y_i$,
we can also define $\alpha_{\Lambda^{-1}}$.
$\alpha_{\Lambda}\ \circ\ \alpha_{\Lambda^{-1}}$ takes
$h\bigl((\coverFB{X-C})^i, \hat y_i\bigr)$ to itself by the zero
map if the path hits $C$ and by the identity otherwise.
Since all but finitely many paths miss $C$, this induces the identity on
$\epsilon$.
Since the empty set is is the minimal element of ${\mathcal D}_X$,
$\alpha_{\Lambda}\ \circ\ \alpha_{\Lambda^{-1}}$ is the identity on
$\displaystyle\mathop{\bigtimes}_{i\in{\mathcal I}} h(X,y_i)$ and
$\mu\bigl(h(X,\hat y_i)\bigr)$.
Hence it is also the identity on $\Delta$.
A similar argument shows
$\alpha_{\Lambda^{-1}}\ \circ\ \alpha_{\Lambda}$
is the identity, so they are both isomorphisms.
The same sort of argument shows $\alpha_\Lambda$ depends only on
the proper homotopy type of $\Lambda$.
It can be safely left to the reader.
\end{proof}
If $h$ is actually a homotopy functor on the category of pairs
(or $n${--}ads) we can define \\
$\gamma\bigl(X,A\Colon h,\{\hat x_i\},\coverFA{}\ \bigr)$
for the pair $(X,A)$ (where $\gamma$ denotes
$\epsilon$ or $\Delta$) using
$G_{i C}= h\bigl((\coverFA{X-C})^i, (\coverFA{A}\ \cap\ (\coverFA{X-C})^i)\
\cup\ (\hat x_i), \hat x_i\ \bigr)$ where
$\coverFA{A}\ \cap\ (\coverFA{X-C})^i= \pi^{-1}\bigl(A\ \cap\ (X-C)^i\bigr)$,
$\pi\colon (\coverFA{X-C})^i\ \to\ (X-C)^i$, if $x_i\notin C$ and is $0$
otherwise ( for $n${--}ads use
$h\bigl((\coverFA{X-C})^i, (\coverFA{A}_1\ \cap\ (\coverFA{X-C})^i)\
\cup\ \{\hat x_i\},\cdots,(\coverFA{A}_{n-1}\ \cap\ (\coverFA{X-C})^i)\
\cup\ \{\hat x_i\},
\hat x_i\ \bigr)$ or $0$).
Now suppose we have a connected sequence of homotopy functors
$h_\ast$; i.e. each $h_n$ is a homotopy functor on some category
of pairs and we get long exact sequences.
By applying our construction to $(X,A)$, one would hope to get
a similar long exact sequence for the $\epsilon$ or $\Delta$ theories.
Several problems arise with this naive expectation.
To begin, we can certainly define groups which fit into a long exact
sequence.
Define $\gamma\bigl(A;X\Colon h_\ast,\{\hat x_i\},\coverFA{}\ \bigr)$
where $\gamma = \epsilon$ or $\Delta$ from
$G_{i C} = h_\ast\Bigl(\bigl(\coverFA{A}\ \cap\ (\coverFA{X-C})^i\bigr)\
\cup\ \{\hat x_i\}, \hat x_i\Bigr)$ if $x_i\notin C$ and $0$ if $x_i\in C$.
Then \fullRef{C.1.2.2.1}, or its unstated analogue
for \fullRef{T.1.2.3}, shows we get a
long exact sequence
\[\begin{aligned}%
\cdots\ \to\ \gamma\bigl(A;X\Colon h_n,&\{ \hat x_i\},\coverFA{}\ \bigr)
\ \to\
\gamma\bigl(X\Colon h_n,\{ \hat x_i\},\coverFA{}\ \bigr)\ \to\
\gamma\bigl(X, A\Colon h_n,\{ \hat x_i\},\coverFA{}\ \bigr)\\
&\to
\gamma\bigl(A;X\Colon h_{n-1},\{ \hat x_i\},\coverFA{}\ \bigr)\ \to\
\cdots\\
\end{aligned}\]
The problem of course is to describe $\gamma(A;X\Colon {\text{etc.}})$
in terms of $A$.
We clearly have little hope unless $A$ is homogamous, and
for convenience we insist $A\subseteq X$ be a proper map.
Such a pair is said to be homogamous, and for such a pair we begin
to describe $\gamma(A;X\Colon {\text{etc.}})$.
Pick a set of base points for $A$, and then add enough new points to
get a set of base points for $X$.
Such a collection is a set of base points for $(X,A)$.
Two such are equivalent provided the points in $X-A$ can be made to
correspond via a locally finite collection of paths in $X$ all of which lie in
\setcounter{footnote}{0}
$X-A$\footnote{And two sets of base points for $A$ are equivalent.}.
A set of base points for $(X,A)$ is irreducible provided any subset
which is also a set of base points for $(X-A)$ has the same cardinality
\footnote{And the set of base points for $A$ is irreducible.}.
(Note an irreducible set of base points for $(X,A)$ is not always an
irreducible set of base points for $X$.
$(S^1, S^0)$ is an example.)
We can construct $\epsilon$ and $\Delta$ groups for $X$ based on
an irreducible set of base points for $(X,A)$, and whenever we have
a pair, we assume the base points are an irreducible set of base points
for the pair.
If $X$ has no compact component, then any irreducible set of base
points for $(X,A)$ is one for $X$.
Over the compact components of $X$, the $\Delta$ group is just
the direct product of $h(\tilde X,p)$ for one $p$ in each component of $A$.
As in the absolute case, we drop irreducible and write
``set of base points'' for ``irreducible set of base points''.
With a set of base points for $(X,A)$, there is a natural map
\[\gamma\bigl(A\Colon h, \{\hat x_i\}, \coverFA{}_F\ \bigr)\ \to\
\gamma\bigl(A; X\Colon h, \{\hat x_i\}, \coverFA{}\ \bigr)\ ,\]
where $\coverFA{}_F$ is the covering functor over
$A$ induced as follows.
Let ${\mathcal D}(X)$ denote the following category.
The objects are closed compact subsets $C\subseteq X$.
The morphisms are the inclusions.
Given $A\subseteq X$ a closed subset, there is a natural map
${\mathcal D}(X)\ \to\ {\mathcal D}(A)$ given by $C\mapsto C\cap A$.
A lift functor $F\colon {\mathcal D}(A)\ \to\ {\mathcal D}(X)$ is a functor
such that ${\mathcal D}(A)\ \RA{F}\ {\mathcal D}(X)\ \to\ {\mathcal D}(A)$
is the identity and such that the image of $F$ is cofinal in ${\mathcal D}(X)$.
$\coverFA{}_F$ is the covering functor whose subgroups are the
pullbacks of
\[\begin{matrix}%
&&S\pi_1\Bigl(\big(X-F(C)\bigr)^i, x_i\Bigr)\\
&&\downlabeledarrow[\Bigg]{\cap}{}\\
\pi_1\bigl((A-C)^i, x_i\bigr)&\RA{\hskip 40pt}&
\pi_1\Bigl(\big(X-F(C)\bigr)^i, x_i\Bigr)\\
\end{matrix}\]
for $x_i\in A-C$, $C\in{\mathcal D}(A)$.
The existence of our natural map
$\gamma(A\Colon \cdots \coverFA{}_F)\ \to\
\gamma(A\Colon \cdots \coverFA{}\ )$
presupposes $\coverFA{}_F$ is compatible of $A$, but this is always
the case since the appropriate cover of $(A-C)^i$ is sitting in
$\bigl((\coverFA{X-F(C)}\bigr)^i$.
We denote this natural map by $\tau(A,X)$.
Notice first that $\tau(A,X)$ is a monomorphism since each map is.
Moreover, $\tau(A,X)$ is naturally split.
The splitting map is induced as follows.
We need only define it on some cofinal subset of ${\mathcal D}(X)$, so we
define it on $\bigl\{ F(C)\ \vert\ C\in{\mathcal D}(A)\ \bigr\}$.
$h_\ast\Bigl(\bigl(\coverFA{A}\ \cap\ (\coverFA{X-F(C)})^i\bigr)\ \cup\
\{\hat x_i\},\ \hat x_i\ \Bigr)$ goes to $0$ if $x_i\notin A$, and it
goes to to
$h_\ast\bigl((\coverFA{A-C})^i,\hat x_i\bigr)$ if $x_i\in A$, where
$\coverFA{}$ in this last case is the cover given by the covering
functor $\coverFA{}_F$.
$\coverFA{A}\ \cap\ (\coverFA{X-F(C)})^i$ is just several disjoint copies
of $(\coverFA{A-C})^i$ union covers of other components of $A-C$.
The map collapses each of these covers of other components of $A-C$
to $\hat x_i$ and on the copies of $(\coverFA{A-C})^i$ it is just the
covering projection.
At this point, this is all we can say about $\tau(A,X)$.
This map however has many more properties and we shall return to it again.
Now let $f\colon X\ \to\ Y$ be a proper map between homogamous spaces.
We have the mapping cylinder $M_f$.
$(M_f, X)$ is an homogamous pair (\fullRef{C.1.1.1.2}).
Let $\{x_i\}$ be a set of base points for $(M_f, X)$.
We also have the homogamous pair $(M_f, Y)$.
By \fullRef{L.1.1.2} a set of base points for $Y$ is also a
set of base points for $(M_f, Y)$.
If $\{y_i\}$ is such a set, $\tau(Y, M_f)$ is an isomorphism.
This is seen by showing the splitting map is a monomorphism.
But if we use the lift functor
$F(C)=I\times f^{-1}(C)\ \cup\ C\subseteq M_f$ this is not hard to see.
Given a covering functor on $M_f$, it induces covering functors on $X$ and
$Y$, and these are the covering functors we shall use.
Given a covering functor on $Y$, we can get a covering functor on $M_f$
as follows.
The subgroups to assign to
$\pi_1( M_f-f^{-1}(C)\times I\ \cup\ C)$ are the subgroups for
$\pi_1(Y-C)$.
One can then assign subgroups to all other required sets in
such a way as to get a covering functor.
If we use the obvious lift functor for $Y$, the induced cover is the original.
By taking the cofinal collection $F(C)$, it is also not hard to see
$\gamma(M_f, Y\Colon h_n, \{ \hat y_i\}, \coverFA{}\ )=0$.
We define
$f\colon \gamma(X\Colon h_n, \{\hat x_i\},\coverFA{}\ )\ \to\
\gamma(Y\Colon h_n, \{\hat x_i\},\coverFA{}\ )$ if no component of
$Y$ is compact by
\[\begin{aligned}%
\gamma(X\Colon{\text{etc.}})\ \RA{\tau(X,M_f)}\ \gamma&(X;M_f\Colon{\text{etc.}})\
\RA{\hskip 10pt}\ \gamma(M_f\Colon h_n,\{\hat x_i\},\coverFA{}\ )\
\RA{\alpha_\Lambda}\\&\gamma(M_f\Colon h_n, \{\hat y_i\},\coverFA{}\ )
\ \LA{\ \cong}\ \gamma(Y\Colon {\text{etc.}})\\
\end{aligned}\]
Notice that this map may depend on the paths used to join
$\{\hat x_i\}$ to $\{\hat y_i\}$.
If $f$ is properly $1/2${--}connected,
(i.e. $f$ induces isomorphisms on $H^0$
and $H^0_{{\text{end}}}$: compare this definition and the one in
\cite{beleven}) there is a natural choice of paths.
This choice is obtained as follows.
Take a set of base points $\{x_i\}$ for $X$.
By \fullRef{L.1.1.1}, $\bigl\{f(x_i)\bigr\}$
is a set of base points for $Y$.
Let $\{ x^\prime_i\}\subseteq \{x_i\}$ be any subset obtained by
picking precisely one element of $\{ x_i\}$ in each
$f^{-1}f(x_i)$.
By \fullRef{L.1.3.1} below, $\{x_i^\prime\}$
is a set of base points for $X$.
Thus we can always find a set of base points for $X$ on which $f$ is
$1${--}$1$ and whose image under $f$ is a set of base points for $Y$.
Take such a set of points as a set of base points for $(M_f,X)$.
Take their image in $Y$ as a set of base points for $(M_f,Y)$.
The paths joining these two sets are just the paths
\[\lambda_{x_i}(t) = \begin{cases}x_i\times t& 0\leq t\leq 1\\
f(x_i)& t=1\ .\\\end{cases}
\]
Given a properly $1/2${--}connected map $f$,
we can get another definition of the induced map.
Pick a set of base points $\{x_i\}$ as in the last paragraph.
Then we have
\[f_\ast\colon \gamma(X\Colon h,\{\hat x_i\}, \coverFA{}\ )
\ \RA{\hskip 10pt}\
\gamma(Y\Colon h,\{\longHatFA{f(x_i)} \}, \coverFA{}\ )\]
defined by taking $h\bigl((\coverFA{X-C})^i, \hat x_i\bigr)\ \to\
h\bigl((\coverFA{Y-F(C)})^i, \longHatFA{f(x_i)}\bigr)$
by $f$, where $F$ is
a lift functor which splits ${\mathcal D}(Y)\ \to\ {\mathcal D}(X)$
and $F$ is the lift functor used to get the covering functor for $X$
from the one for $Y$.
One sees easily the two definitions of $f_\ast$ agree.
Now suppose we consider $i\colon A\subseteq X$ for an homogamous pair.
Then we can define $i_\ast$ as above.
It is not hard to see
\[\begin{matrix}%
\gamma(A;X\Colon {\text{etc.}})&\longrightarrow&\gamma(X\Colon{\text{etc.}})\\
\uplabeledarrow[\Bigg]{}{\tau(A,X)}&\hskip10pt\nearrow
\lower 8pt\hbox{$i_\ast$}\\
\gamma(A\Colon {\text{etc.}})\\
\end{matrix}\]
commutes, where the paths we use in defining $i_\ast$ are $\lambda_{x_i}(t)= x_i\times t$ in
$A\times I\ \cup\ X\times 1 = M_i$.
\medskip
\BEGIN{L.1.3.1}
If $f\colon X\ \to\ Y$ is a proper map which induces epimorphisms on
$H^0$ and $H^0_{{\text{end}}}$, then, if $\bigl\{ f(p)\bigr\}$ is a set
of base points for $Y$, $\{ p\}$ is a set of base points for $X$.
\end{Lemma}
\begin{proof}
Since $f$ is an epimorphism on $H^0$, each path component of $X$
has a point of $\{p\}$ in it.
Now define a cochain in $S^0(X)$ for some closed compact set
$D\subseteq X$, $\varphi_D$ as follows.
$\varphi_D(q)=1$ if $q$ is in a path component of $X-D$
with no point of $\{ p\}$ in it and is $0$ otherwise.
$\delta\varphi_D = 0$ in $S^1_{{\text{end}}}$.
Since $f$ is an epimorphism on $H^0_{{\text{end}}}$, there must be a chain in
$S^0(Y)$, $\psi$, such that $f^\ast\psi=\varphi$ in $S^0_{{\text{end}}}(X)$.
But this means there is some closed compact set $C\subseteq X$ such
that $f^\ast\psi$ and $\varphi$ agree for any point in $X-C$.
Hence there is a closed, compact set $E\subseteq Y$ such that
$f^{-1}(E)\supseteq C\ \cup\ D$.
There is also a closed, compact $F\subseteq Y$
such that there is an $f(p)$
in each path component of $Y-E$ which is not contained in $F$.
$\psi$ restricted to $Y-E$ must be $0$ since some component of
$X-D$ which is not contained in $f^{-1}(E)$ has a point of $\{p\}$ in it.
Hence $\varphi$ restricted to $X-f^{-1}(E)$ is $0$, so we are done.
\end{proof}
\begin{xDefinition}
An homogamous pair $(X,A)$ is properly $0${--}connected if the inclusion
induces monomorphisms on $H^0$ and $H^0_{{\text{end}}}$.
We have already defined properly $1/2${--}connected.
If $(X,A)$ is properly $0${--}connected we can choose a set of base
points for the pair to be a set of base points for $A$.
We say $(X,A)$ is \emph{properly $n${--}connected}, $n\geq 1$
provided it is properly $1/2${--}connected, and, with base points chosen
as above,
$\Delta(X,A\Colon \pi_k, \{ x_i\}, {\text{no\ cover}}\ )=0$, $1\leq k\leq n$.
It is said to be \emph{properly $n${--}connected at $\infty$} provided it
is properly $1/2${--}connected and
$\epsilon(X,A\Colon \pi_k, \{ x_i\}, {\text{no\ cover}}\ )=0$, $1\leq k\leq n$.
\end{xDefinition}
\BEGIN{P.1.3.1}
If $(X,A)$ is properly $1/2${--}connected, and if
\[i_\ast\colon \Delta(A\Colon \pi_1,\{ x_i\}, {\text{no\ cover}}\ )\ \to\
\Delta(X\Colon \pi_1,\{ x_i\}, {\text{no\ cover}}\ )\] is onto,
$(X,A)$ is properly $1${--}connected and conversely.
\end{Proposition}
\begin{proof}
If $(X,A)$ is properly $1/2${--}connected,
\[\Delta(A\Colon \pi_0,\{ x_i\}, {\text{no\ cover}}\ )\ \to\
\Delta(X\Colon \pi_0,\{ x_i\}, {\text{no\ cover}}\ )\]
is seen to be an isomorphism by applying
\fullRef{T.1.2.4} to the kernel and cokernel
of this map, together with the definition of a set of base points.
Hence $\Delta(A;X \Colon \pi_1)\ \to\
\Delta(X\Colon \pi_1 )\ \to\ \Delta(X, A\Colon \pi_1)\to 0$ is exact.
\[\begin{matrix}%
\Delta(A;X\Colon \pi_1)&\longrightarrow&\Delta(X\Colon \pi_1)\\
\uplabeledarrow[\Bigg]{}{\tau(X,A)}&
\hskip14pt\nearrow\lower 4pt\hbox{$i_\ast$}\\
\Delta(A\Colon \pi_1)\\
\end{matrix}\]
commutes, and $i_\ast$ is an epimorphism.
Hence $\Delta(X,A\Colon \pi_1)=0$, so $(X,A)$ is properly $1${--}connected.
The converse follows trivially from
\fullRef{P.1.3.2} below and the definitions.
\end{proof}
\BEGIN{P.1.3.2}
Let $(X,A)$ be a properly $1${--}connected pair.
Then $\tau(A,X)$ is an isomorphism if the base points for the pair
are a set of base points for $A$.
We may use any lift functor to induce the covering functor.
\end{Proposition}
\begin{proof}
If $\tau$ is an isomorphism on the $\epsilon$ objects, we need only show
$h(\coverFA{A},\hat x_i)= h(\coverFA{A}\ \cap\ \coverFA{X}, \hat x_i)$.
But $\coverFA{A}=\coverFA{A}\ \cap \coverFA{X}$ if $\pi_1(A)\to\pi_1(X)$
is onto, so if we can show the result for
the $\epsilon$ objects we are done.
We need only show $\tau$ is onto.
By \fullRef{T.1.2.4} applied to the cokernels of
the maps inducing $\tau$, we need only show
that for each $C\in{\mathcal D}(X)$, there is a $D\supseteq C$ in ${\mathcal D}(X)$
such that
\[\begin{matrix}%
h\bigl((\coverFA{A-D})^i, \hat x_i)&\RA{\hskip10pt}&
h\biggl(\Bigl(\coverFA{A}\ \cap\ \bigl(\coverFA{X-F(D)}\bigr)^i\Bigr)\
\cup \{\hat x_i\},\hat x_i\biggr)\\
\downarrow[\Big]&&\downlabeledarrow[\Big]{i_\ast}{}\\
h\bigl((\coverFA{A-C})^i, \hat x_i)&\RA{\ \tau_\ast\ }&
h\biggl(\Bigl(\coverFA{A}\ \cap\ \bigl(\coverFA{X-F(C)}\bigr)^i\Bigr)\
\cup \{\hat x_i\},\hat x_i\biggr)\\
\end{matrix}\]
satisfies ${\Imx}\ i_\ast\ \subseteq {\Imx}\ \tau_\ast$
for all $x_i\notin D$.
We saw $\coverFA{A}\ \cap\ \bigl(\coverFA{X-F(C)}\bigr)^i$
was just some copies of $(\coverFA{A-C})^i$, together with covers
of components of $A-C\subseteq \bigl(X-F(C)\bigr)^i$.
Since $(X,A)$ is properly $1/2${--}connected, we can find $D$ so that
${\Imx}\ i_\ast\subseteq h\bigl({\text{copies\ of\ }}
(\coverFA{A-C})^i\bigr)$; i.e. we can find $D$ so that
$\bigl(X-F(D)\bigr)^i\ \cap\ (A-C) =
\bigl(X-F(D)\bigr)^i\ \cap\ (A-C)^i$.
Since $(X,A)$ is properly $1${--}connected, we can find
$D_1\supseteq D$ so that
\[\begin{matrix}%
\pi_1\bigl( X-F(D_1), A-D_1, x_i\bigr)\\
\Big\downarrow\\
\pi_1\bigl( X-F(C), A-C, x_i\bigr)\\
\end{matrix}\]
is zero for all $x_i\notin D_1$.
But this says all the copies of
$(\coverFA{A-D_1})^i$ in
$\coverFA{A}\ \cap\ \bigl(\coverFA{X-F(D_1)}\bigr)^i$ go to the same copy
of $(\coverFA{A-C})^i$ in
$\coverFA{A}\ \cap\ \bigl(\coverFA{X-F(C)}\bigr)^i$, namely
the one containing $\hat x_i$.
\end{proof}
\fullRef{T.1.2.4} can also be used to get
\begin{Named Theorem}[The subspace principle]
Let $(X,A)$ be an arbitrary homogamous pair.
Then \\
$\gamma(A;X \Colon h, \{\hat x_i\},\coverFA{}\ )=0$ \iff\
$\gamma(A\Colon h, \{\hat x_i\},\coverFA{}\ )=0$ provided, for the if part,
\begin{enumerate}
\item[1)]
if $A_\alpha$ is a collection of disjoint subsets of $A$,
$\displaystyle h(\mathop{\cup}_{\alpha} \coverFA{A}_\alpha\ \cup\ p,p)
\cong \mathop{\oplus}_\alpha h(\coverFA{A}_\alpha\ \cup\ p,p)$.
\item[2)] if $E\subseteq B$ are subsets of $A$, and if there is a
$q\in \coverFA{E}$ such that $h(\coverFA{E}, q)\ \to\ h(\coverFA{B},q)$
is the zero map, then
$h(\coverFA{E}\ \cup\ p, p)\ \to\ h(\coverFA{B}\ \cup\ p, p)$ is the zero
map for any $p$.
$h$ need only be natural on subsets of $A$.
\end{enumerate}
\end{Named Theorem}
\begin{proof}
Only if is clear as $\tau(A,X)$ is naturally split, so we concentrate
on the if part.
$\gamma(A\Colon h, \{\hat x_i\},\coverFA{}\ )=0$ implies by
\fullRef{T.1.2.4} that we can find a cofinal sequence
$C_0\subseteq C_1\subseteq\cdots$ of closed, compact subsets
of $A$ such that
$h\bigl((\coverFA{A-C_j})^i, \hat x_i\bigr)\ \to\
h\bigl((\coverFA{A-C_{j-1}})^i, \hat x_i\bigr)$
is the zero map for all $x_i\notin C_j$.
If $\gamma=\Delta$, $h(A)$ is also zero.
We then claim
$h\Bigl(\coverFA{A}\ \cap\ \bigl((\coverFA{X-F(C_j)}\bigr)^i
\ \cup\ \{\hat x_i\}, \hat x_i\Bigr)\ \to\
h\Bigl(\coverFA{A}\ \cap\ \bigl((\coverFA{X-F(C_{j-1})}\bigr)^i
\ \cup\ \{\hat x_i\}, \hat x_i\Bigr)$
is the zero map, and, if $\gamma=\Delta$,
$h(\coverFA{A}\ \cap\ \coverFA{X}, \hat x_i)=0$.
This last is easy since $\coverFA{A}\ \cap\ \coverFA{X}$ is the disjoint union
of copies of $\coverFA{A}$.
Now $\displaystyle\coverFA{A}\ \cap\ \bigl(\coverFA{X-F(C_j)}\bigr)^i=
\mathop{\cup}_\beta \mathop{\cup}_{\alpha_\beta} Z_{\alpha_\beta}$
where $\beta$ runs over the path components of $A-C_j$ in
$\bigl(X-F(C_j)\bigr)^i$, and $\alpha_\beta$ runs over
the path components of $\pi^{-1}\bigl((A-C_j)^\beta\bigr)$ where
\[\pi\colon \bigl(\coverFA{X-F(C_j)}\bigr)^i \to\
\bigl(X-F(C_j)\bigr)^i\]
is the covering projection and $(A-C_j)^\beta$
is the component of $A-C_j$ corresponding to $\beta$.
$Z_{\alpha_\beta}$ is the $\alpha_\beta^{\text{\thx}}$ component
of $\pi^{-1}\bigl((A-C_j)^i\bigr)$.
Similarly $\coverFA{A}\ \cap\ \bigl(\coverFA{X-F(C_{j-1})}\bigr)^i=
\displaystyle\mathop{\cup}_{\alpha}\mathop{\cup}_{\alpha_b}
Z_{\alpha_b}$.
The map we are looking at is just the map induced on the direct
sum by the maps
$h( Z_{\alpha_\beta}\ \cup\ \{\hat x_i\}, \hat x_i)\ \to\
h( Z_{a_b}\ \cup\ \{\hat x_i\}, \hat x_i)$
for the unique $a_b$ such that $Z_{a_b}$
is mapped into by $Z_{\alpha_\beta}$.
$Z_{\alpha_\beta}=(A-C_j)^\beta$, so if $\hat x_i\in Z_{\alpha_\beta}$,
the map is the zero map since it is then a map of the form
$h\bigl((\coverFA{A-C_j})^i, \hat x_i\bigr)\ \to\
h\bigl((\coverFA{A-C_{j-1}})^i, \hat x_i\bigr)$ which we know to be zero.
If $\hat x_i\notin Z_{\alpha_\beta}$, the map is now a map of the form
\par\noindent
$h\bigl((\coverFA{A-C_j})^\beta\ \cup \ \{\hat x_i\}, \hat x_i)
\ \to\
h\bigl((\coverFA{A-C_{j-1}})^\beta\ \cup \ \{\hat x_i\}, \hat x_i)$,
which is still zero by the properties of $h$.
\end{proof}
We now investigate the invariance of our construction.
\medskip
\BEGIN{T.1.3.2}
Let $f$, $g\colon X\ \to\ Y$ be properly homotopic maps
between homogamous spaces.
Then there is a set of paths $\Lambda$ such that
\[\begin{matrix}%
\gamma(X\Colon h, \{\hat x_i\},\coverFA{}\ )&\RA{\ f_\ast\ }&
\gamma(Y\Colon h, \{\hat y_i\},\coverFA{}\ )\\
&\hbox to 0pt{\hss${g_\ast}$}\searrow&\downlabeledarrow[\Big]{\alpha_\Lambda}{}\\
&&\gamma(Y\Colon h, \{\hat y_i\},\coverFA{}\ )\\
\end{matrix}\]
commutes.
\end{Theorem}
\begin{proof}
Let $F \colon X\times I\to Y$ be the homotopy, and let $M_F$ be its
mapping cylinder.
Then it is possible to pick paths so that
\[\begin{matrix}%
\gamma(M_F\Colon h, \{\hat x_i\}\times 0,\coverFA{}\ )&\RA{\ \ \ }&
\gamma(M_F\Colon h, \{\hat y_i\},\coverFA{}\ )\\\noalign{\vskip4pt}
\Big\downarrow&&\downlabeledarrow[\Big]{\alpha_\Lambda}{}\\
\gamma(M_F\Colon h, \{\hat x_i\}\times 1,\coverFA{}\ )&\RA{\ \ \ }&
\gamma(M_F\Colon h, \{\hat y_i\},\coverFA{}\ )\\
\end{matrix}\]
commutes, where the horizontal maps are the maps induced by the
paths joining $\{\hat x_i\}\times t$ to $\{\hat y_i\}$ ($t=0$, $1$) and
the left hand vertical map is induced by the canonical path
$\hat x_i\times 0$ to $\hat x_i\times 1$ in $X\times I\ \to\ M_F$.
It is now a chase of the definitions to show the desired diagram
commutes.
\end{proof}
\BEGIN{C.1.3.2.1}
Let $f\colon X\to Y$ be a proper homotopy equivalence between
two homogamous spaces.
Then $f_\ast$ is an isomorphism.
\end{Corollary}
\begin{proof}
There is a standard derivation of the corollary from the theorem.
\end{proof}
\BEGIN{C.1.3.2.2}
A proper homotopy equivalence between homogamous spaces
is properly $n${--}connected for
all $n$ (i.e. its mapping cylinder modulo its domain is
a properly $n${--}connected pair).
\end{Corollary}
\begin{proof}
$(M_f, X)$ is clearly properly $1/2${--}connected.
$i_\ast\colon \Delta(X\Colon \pi_1)\ \to\ \Delta(Y\Colon \pi_1)$ is onto, so it is
easy to show $(M_f,X)$ is properly $1${--}connected.
Then
$\gamma(X; M_f\Colon \pi_k)\ \cong\ \gamma(X\Colon \pi_k)\ \cong\
\gamma(M_f\Colon \pi_k)$, so
$\gamma(M_f, X\Colon \pi_k)=0$.
\end{proof}
\BEGIN{C.1.3.2.3}
If $f\colon X\to Y$ is a proper homotopy equivalence,
\[\gamma( M_f, X\Colon h, \{\hat x_i\}, \coverFA{}\ )=0\ .\]
\end{Corollary}
\begin{proof}
Since $f$ is properly $1${--}connected, $\gamma(X;M_f\Colon h, {\text{etc.}})\ \cong\
\gamma(X\Colon h, {\text{etc.}})$ by \fullRef{P.1.3.2}
$\gamma(X\Colon h, {\text{etc.}})\ \cong\ \gamma(M_f\Colon h, {\text{etc.}})$
by \fullRef{C.1.3.2.1}.
Hence
$\gamma(M_f, X\Colon h, {\text{etc.}})=0$.
\end{proof}
In the other direction we have
\BEGIN{T.1.3.3}(Proper Whitehead)
Let $f\colon X\to Y$ be properly $n${--}connected.
Then for a locally finite CW complex, $K$, of dimension $\leq n$,
$f_\#\colon [K, X]\ \to\ [K,Y]$ is an epimorphism.
If $f$ is properly $(n+1)${--}connected, $f_\#$ is a bijection.
\end{Theorem}
\begin{xRemarks}
$[K,X]$ denotes the proper homotopy classes of proper maps of $K$
to $X$.
For a proof of this result, see \cite{beleven} Theorem 3.4 and note
the proof is valid for $X$ and $Y$ homogamous.
\end{xRemarks}
\begin{xDefinition}
An homogamous space $Z$ is said to satisfy $D n$ provided the statement
of \fullRef{T.1.3.3} holds for $Z$ in place of $K$ and for
each properly $n${--}connected map $f$ between homogamous spaces.
\end{xDefinition}
\BEGIN{P.1.3.3}
Let $Z$ be properly dominated by a space satisfying $D n$.
Then $Z$ satisfies $D n$.
\end{Proposition}
\begin{proof}
We leave it to the reader to modify the proof of
\fullRef{P.1.1.1} to show $Z$ is homogamous
\iff\ it is properly dominated by an homogamous space.
Let $K$ be a space satisfying $D n$ and properly dominating $Z$.
Then $[Z,X]$ us a natural summand of $[K,X]$
for any homogamous $X$ and the result follows.
\end{proof}
We finish this section by proving a proper Hurewicz and a proper
Namioka theorem.
\begin{xDefinition}
A (as opposed to the) universal covering functor for $X$ is a
covering functor $\coverFA{}$ such that
$\epsilon(X\Colon \pi_1,\coverFA{}\ )=\Delta(X\Colon \pi_1,\coverFA{}\ )=0$.
Note that if the universal covering functor is
compatible with $X$, then it is a universal covering functor for $X$.
There are other examples however.
\end{xDefinition}
We start towards a proof of the Hurewicz theorem.
The proof mimics Spanier \cite{bthirtyfive} pages 391{--}393.
We first prove
\medskip
\BEGIN{L.1.3.2}
Suppose ${\mathcal G}=\{G_{i j}\}$ is a system of singular chain complexes on
spaces $X_{i j}$.
Suppose the projection maps $G_{i j}\ \to\ G_{i\>j-1}$ are induced
by continuous maps of the spaces $X_{i j}\ \to\ X_{i\>j-1}$.
Assume $i\geq 0$, $j\geq 0$.
Assume we are given a system $C=\{C_{i j}\}$, where each $C_{i j}$
is a subcomplex of $G_{i j}$ which is generated by the singular simplices
of $G_{i j}$ which occur in $C_{i j}$.
Also assume that the projection $G_{i j}\ \to\ G_{i\>j-1}$
takes $C_{i j}\ \to\ C_{i\>j-1}$.
Lastly assume that to every singular simplex
$\sigma\colon\Delta^q\to X_{i j}$ for $j\geq n$ ($n$ is given at the
start and held fixed throughout) there is assigned a map
$P_{i j}(\sigma)\colon
\Delta^q\times I\ \to\ X_{i\>j-n}$ which satisfies
\begin{enumerate}
\item[a)]
$P_{i j}(\sigma)(z,0)=\Bar\sigma(x)$, where
$\Bar\sigma\colon \Delta^q\ \RA{\sigma}\ X_{i j}\ \RA{\text{projection}}\ X_{i\>j-n}$.
\item[b)]
Define $\sigma_1\colon \Delta^q\ \to\ X_{i\>j-n}$ by
$\sigma_1(z)=P_{i j}(\sigma)(z,1)$.
Then we require that $\sigma_1\in C_{i\>j-n}$,
and, if $\sigma\in C_{i j}$, then $\sigma_1=\Bar\sigma$.
\item[c)] If $e^k_q\colon \Delta^{q-1}\ \to\ \Delta^q$
omits the $k^{\text{\thx}}$ vertex, then
$P_{i j}(\sigma)\ \circ\ (e^k_q\times 1) = P_{i j}(\sigma^{(k)})$.
\end{enumerate}
Then $\epsilon(C )\subseteq \epsilon({\mathcal G})$ is an homology equivalence.
(Compare Spanier \cite{bthirtyfive}, page 392, Lemma 7).
\end{Lemma}
\begin{proof}
Let $\alpha(i,k)\colon C_{i k}\subseteq G_{i k}$ be the inclusion, and
let $\tau(i,k)\colon G_{i k}\ \to\ C_{i\>k-n}$ be defined by
$\tau(i,k)(\sigma)=\Bar\sigma_1$ and extend linearly.
(Here we must assume $k\geq n$).
Define $\rho_r\colon G_{i k}\ \to\ G_{i\>k-r}$ to be the
projection.
One easily checks that condition c) makes $\tau(i,k)$ into a chain map.
$\tau(i,k)\ \circ\ \alpha(i,k)\colon C_{i k}\penalty-2000 \to\ C_{i\>k-n}$ is
just the map induced on the $C_{i k}$ by $\rho_n$ on the $G_{i k}$.
This follows from condition b).
We claim $\alpha(i, k-n)\ \circ\ \tau(i,k)\colon G_{i k}\ \to\ G_{i\>k-n}$
is chain homotopic to $\rho_n$.
To show this, let $D_g\colon S(\Delta^q)\ \to\ S(\Delta^q\times I)$
be a natural chain homotopy between $\Delta(h_1)$ and
$\Delta(h_0)$, where $h_0$, $h_1\colon \Delta^q\ \to\ \Delta^q\times I$
are the obvious maps ($S$ is the singular chain functor).
Define a chain homotopy $D_{i k}\colon S(X_{i k})\ \to\ S(X_{i\>k-n})$
by $D_{i k}(\sigma)= S\bigl(P_{i k}(\sigma)\bigr)\bigl( D_q(\xi_q)\bigr)$
(where $\xi_q \colon \Delta^q\subset\Delta^q$ is the identity) where
$\sigma$ is a $q${--}simplex.
One checks, using c) and the naturality of $D_q$ that
$\partial D_{i k}+ D_{i k}\partial = \rho_n-
\alpha(i, k-n)\ \circ\ \tau(i,k)$.
By definition, $\displaystyle\epsilon({\mathcal G})=
\mathop{\lim}_{{\Atop{\leftarrow}{k}}}\ \mu(G_{i k})$ and
$\displaystyle\epsilon(C)=
\mathop{\lim}_{{\Atop{\leftarrow}{k}}}\ \mu(C_{i k})$.
Since
\[\begin{matrix}%
C_{i k}&\RA{\alpha(i,k)}&G_{i k}\\
\Big\downarrow&&\Big\downarrow\\
C_{i\>k-1}&\RA{\alpha(i,k-1)}&G_{i\>k-1}\\
\end{matrix}\]
commutes, we get a chain map
$\alpha\colon\epsilon(C)\ \to\epsilon({\mathcal G})$,
which is just the inclusion.
Since $\tau(i,k)\ \circ\ \alpha(i,k)=\rho_n$,
\[\begin{matrix}%
C_{i k}&\RA{\alpha(i,k)} & G_{i k} &\RA{\tau(i,k)}&C_{i\>k-n}\\
\noalign{\vskip4pt}
\downlabeledarrow[\Big]{\rho_1}{}&&\downlabeledarrow[\Big]{\rho_1}{}
&&\downlabeledarrow[\Big]{\rho_1}{}\\
C_{i\>k-1}&\RA{\alpha(i,k-1)} & G_{i\>k-1} &\RA{\tau(i,k-1)}&C_{i\>k-1-n}\\
\end{matrix}\]
commutes along the outside square.
Unfortunately the right{--}hand square may not commute
as we have made no stipulation as to the behavior of $P_{i j}$
with respect to $\rho_1$.
Similarly
\[\begin{matrix}%
G_{i k}&\RA{\tau(i,k)}&C_{i\>k-n}&\RA{\alpha(i, k-n)}&G_{i\>k-n}\\
\downlabeledarrow[\Big]{\rho_1}{}&&&&\downlabeledarrow[\Big]{\rho_1}{}\\
G_{i\>k-1}&\RA{\tau(i,k-1)}&C_{i\>k-1-n}&\RA{\alpha(i, k-1-n)}&G_{i\>k-1-n}\\
\end{matrix}\]
may not commute.
However, since $\alpha(i, k-n)\ \circ\ \tau(i,k)$
is chain homotopic to $\rho_n$,
\[\begin{matrix}%
H_\ast(G_{i k})&\RA{H_\ast\bigl(\tau(i,k)\bigr)}&H_\ast(C_{i\>k-n})&
\RA{H_\ast\bigl(\alpha(i, k-n)\bigr)}&H_\ast(G_{i\>k-n})\\
\downlabeledarrow[\Big]{H_\ast(\rho_1)}{}&&&&\downlabeledarrow[\Big]{H_\ast(\rho_1)}{}\\
H_\ast(G_{i\>k-1})&\RA{H_\ast\bigl(\tau(i,k-1)\bigr)}&
H_\ast(C_{i\>k-1-n})&\RA{H_\ast\bigl(\alpha(i, k-1-n)\bigr)}&
H_\ast(G_{i\>k-1-n})\\
\end{matrix}\]
does commute.
Define $\beta(i,k)\colon G_{i k}\ \to\ C_{i\> k-2n}$ for $k\geq 2n$
by
$\beta(i,k)=\tau(i,k-n)\ \circ\ \alpha(i, k-n)\ \circ\ \tau(i,k)$.
We claim
\[\begin{matrix}%
H_\ast(G_{i k})& \RA{H_\ast\bigl(\beta(i,k)\bigr)}&H_\ast(C_{i\> k-2n})\\
\downlabeledarrow[\Big]{H_\ast(\rho_1)}{}&&
\downlabeledarrow[\Big]{H_\ast(\rho_1)}{}\\
H_\ast(G_{i\>k-1})& \RA{H_\ast\bigl(\beta(i,k-1)\bigr)}
&H_\ast(C_{i\> k-1-2n})\\
\end{matrix}\]
commutes.
To see this, look at
\[\begin{matrix}%
H_\ast(G_{i k})&\RA{H_\ast(\tau)}&
H_\ast(C_{i\>k-n})&\RA{H_\ast(\alpha)}&
H_\ast(G_{i\> k-n})&\RA{H_\ast(\tau)}&
H_\ast(G_{i\>k-2n})\\
\downlabeledarrow[\Big]{H_\ast(\rho_1)}{}&{\text{I}}&
\downlabeledarrow[\Big]{H_\ast(\rho_1)}{}&{\text{II}}&
\downlabeledarrow[\Big]{H_\ast(\rho_1)}{}&{\text{III}}&
\downlabeledarrow[\Big]{H_\ast(\rho_1)}{}\\
H_\ast(G_{i\>k-1})&\RA{H_\ast(\tau)}&
H_\ast(C_{i\>k-1-n})&\RA{H_\ast(\alpha)}&
H_\ast(G_{i\> k-1-n})&\RA{H_\ast(\tau)}&
H_\ast(G_{i\>k-1-2n})\\
\end{matrix}\]
The square II commutes since it already does on the chain level.
Similarly, the square II+III commutes.
The square I+II commutes on the homology level.
The desired commutativity is now a diagram chase.
Now define $\tau\colon \epsilon\bigl(H_\ast({\mathcal G})\bigr)\ \to\
\epsilon\bigl(H_\ast(C)\bigr)$ using the $H_\ast(\beta)$'s.
We also have $H_\ast(\alpha)\colon \hfill\penalty-10000
H_\ast\bigl(\epsilon(C)\bigr)\to\
H_\ast\bigl(\epsilon({\mathcal G})\bigr)$.
By \fullRef{C.1.1.2.2} we have
$H_\ast(\alpha)\colon \epsilon\bigl(H_\ast(C)\bigr)\ \to\
\epsilon\bigl(H_\ast({\mathcal G})\bigr)$.
$\tau\ \circ\ H_\ast(\alpha)$ and $H_\ast(\alpha)\ \circ\ \tau$
are both induced from the maps $H_\ast(\rho_{2n})$, and
hence are the identities on the inverse limits.
\end{proof}
\medskip
\BEGIN{L.1.3.3}
Let $X$ be an homogamous space.
Then we can find a countable, cofinal collection of closed,
compact sets $C_j\subseteq X$ with $C_j\subseteq C_{j+1}$.
Let $G_{i j}=S\bigl((\coverFA{X-C_j})^i, \hat x_i\bigr)$, the singular
chain groups on $(\coverFA{X-C_j})^i$.
Let $C_{i j}=S\bigl((\coverFA{X-C_j})^i,
\coverFA{A}\ \cap\ (\coverFA{X-C_j})^i, \hat x_i\bigr)^n$.
(see Spanier \cite{bthirtyfive}, page 391 for a definition).
Suppose $(X,A)$ is properly $1${--}connected and
properly $n${--}connected at $\infty$ for $n\geq 0$.
Then the inclusion map
$\epsilon(C)\subseteq\epsilon({\mathcal G})$ is an homology equivalence.
(Notice that if we pick a set of base points $x_i$ for $A$, they are
a set for the pair, and $\epsilon({\mathcal G})=
\epsilon(X\Colon H_\ast,\{\hat x_i\},\coverFA{}\ )$.)
\end{Lemma}
\begin{proof}
Let $r=\min(q,n)$.
Then we produce for every $\sigma\in G_{i j}$ a map
\[
P_{i j}(\sigma) \colon \Delta^q\times I\ \to\
\bigl((\coverFA{X-C_{j-r}})^i,\hat x_i\bigr)
\]
which satisfies
\begin{enumerate}
\item[a)]
$P_{i j}(\sigma)(z,0)=\Bar\sigma\colon\Delta^q\ \RA{\sigma}\
\bigl((\coverFA{X-C_{j}})^i\bigr)\ \RA{\text{projection}}\
\bigl((\coverFA{X-C_{j-r}})^i\bigr)$.
\item[b)] If $\sigma_1(z)=P_{i j}(\sigma)(z,1)$, $\sigma_1\in C_{i\>j-r}$,
and if $\sigma\in C_{i j}$,
\[P_{i j}(\sigma)\colon
\Delta^q\times I\ \RA{{\text{proj}}}\ \Delta^q\ \RA{\sigma}\
\bigl((\coverFA{X-C_j})^i\bigr)\ \RA{{\text{projection}}}\
\bigl((\coverFA{X-C_{j-r}})^i\bigr)\]
\item[c)] $P_{i j}(\sigma)\ \circ\ (C^k_q\times 1)=
\begin{cases} ({\text{projection\ 1\ step}}\ )\ \circ\ P_{i j}(\sigma^{(k)})
& q\leq n\\
P_{i j}(\sigma^{(k)})&q>n\\\end{cases}$
\end{enumerate}
From such a $P$ it is easy to see how to get a $P$ as required by our
first lemma.
We remark that $C_{i j}$ and $G_{i j}$ satisfy all the other requirements
to apply the lemma.
Hence \fullRef{L.1.3.2} will then give us the
desired conclusion.
We define $P_{i j}$ by induction on $q$.
Let $q=0$.
Then $\sigma\in G_{i j}$ is a map
$\sigma\colon \Delta^0\to \bigl((\coverFA{X-C_j})^i\bigr)$.
Since the point $\sigma(\Delta^0)$ lies in the same path component of
$(\coverFA{X-C_j})^i$ as $\hat x_i$, there is a path joining them.
Let $P_{i j}(\sigma)$ be such a path.
If $\sigma(\Delta^0)=\hat x_i$, $P_{i j}(\sigma)$
should be the constant path.
This defines $P_{i j}$ for $q=0$, and $P$ is easily seen to satisfy a){--}c).
Now suppose $P_{i j}$ is defined for all $\sigma$ of degree $<q$,
$0<q\leq n$ so that it has properties a){--}c).
If $\sigma\in C_{i j}$, b) defines $P(\sigma)$, and $P$ then satisfies
a) and c).
So suppose $\sigma\notin C_{i j}$.
a) and c) define $P_{i j}$ on $\Delta^q\times 0\ \cup\
\dot\Delta^q\times I$; i.e. we get a map
$f\colon \Delta^q\times 0\ \cup\ \dot\Delta^q\times I\ \to\
(\coverFA{X-C_{j-q+1}})^i$.
There is a homeomorphism $h\colon E^q\times I\approx \Delta^q\times I$
such that $h(E^q\times 0)=\Delta^q\times 0\ \cup\ \dot\Delta^q\times I$;
$h(S^{q-1}\times 0)=\dot\Delta^q\times 1$; and
$h(S^{q-1}\times I\ \cup\ E^q\times 1)=\Delta^q\times 1$.
Let $g\colon (E^q,S^{q-1})\ \to\
\bigl((\coverFA{X-C_{j-q+1}})^i,
\coverFA{A}\ \cup\ (\coverFA{X-C_{j-q+1}})^i\bigr)$
be defined by $g=f\ \circ\ h$.
Because $q\leq n$ and $(X,A)$ is properly $n${--}connected at $\infty$,
we could have chosen (and did) the $C_j$ so that
\[\pi_q(X-C_k,A\cap(X-C_k),\ast)\ \to
\pi_q(X-C_{k-1},A\cap(X-C_{k-1}),\ast)\]
is the zero map for $q\leq n$.
Thus we get a homotopy
\[H\colon (E^q,S^{q-1})\times I\ \to\
\bigl((\coverFA{X-C_{j-q}})^i, \coverFA{A}\ \cap\ (\coverFA{X-C_{j-q}})^i\bigr)\]
between $\rho_1\ \circ\ g$ and an element of $C_{i\>j-q}$.
Define $P_{i j}(\sigma)$ to be the composite
$\Delta^q\times I\ \RA{h^{-1}\times{\text{id}}}\ E^q\times I\
\RA{H}\
(\coverFA{X-C_{j-q}})^i$.
$P_{i j}$ clearly satisfies a) and b). Since $h$ was chosen carefully
c) is also satisfied.
In this way $P$ is defined for all simplices of degree $\leq n$.
Note that a singular simplex of degree $>n$ is in $C_{i j}$ \iff\ every
proper face is in $C_{i j}$.
Suppose that $P$ has been defined for all degrees $<q$, where $q>n$.
If $\sigma\in C_{i j}$, we define $P_{i j}(\sigma)$
by b) as usual.
It satisfies a) and c).
So suppose $\sigma\notin C_{i j}$.
Then a) and c) define a map
$f\colon \Delta^q\times 0\ \cup\ \dot\Delta^q\times I\ \to\
(\coverFA{X - C_{j-n}})^i$.
By the homotopy extension property we can extend $f$ to some map
$P(\sigma)\colon \Delta^q\times I\ \to\ (\coverFA{X - C_{j-n}})^i$.
It clearly satisfies a) and c).
It also satisfies b) since every proper face of $\sigma_1$ is in $C_{i\> j-n}$.
Hence we have defined our $P$.
\end{proof}
\bigskip
Now define $\epsilon^{(n)}(X,A\Colon H_q,\coverFA{}\ )$ to be
\[\begin{aligned}%
\epsilon\Biggl(H_q\biggl( &S\Bigl( (\coverFA{X-C_j})^i,
\coverFA{A}\ \cap\ (\coverFA{X-C_j})^i,\hat x_i\Bigr)^n /\\
&S\Bigl( (\coverFA{X-C_j})^i,
\coverFA{A}\ \cap\ (\coverFA{X-C_j})^i,\hat x_i\Bigr)^n\ \cap\
S\Bigl(\coverFA{A}\ \cap\ (\coverFA{X-C_j})^i, \hat x_i\Bigr)
\biggr)\Biggr)\ .\\\end{aligned}\]
Then there are natural maps
\[\epsilon^{(n)}(X,A\Colon H_q,\coverFA{}\ )\ \to\
\epsilon^{(n-1)}(X,A\Colon H_q,\coverFA{}\ )\ \to\
\cdots \ \to\ \epsilon(X,A\Colon H_q,\{\hat x_i\},\coverFA{}\ )\ .\]
\bigskip
\BEGIN{L.1.3.4}
Assume $(X,A)$ is a properly $1${--}connected pair which is
properly $n${--}connected at $\infty$ for some $n\geq0$.
Then the natural map
$\epsilon^{(n)}(X,A\Colon H_q,\coverFA{}\ )\ \to\
\epsilon(X,A\Colon H_q,\{ \hat x_i\},\coverFA{}\ )$ is an isomorphism for all $q$.
\end{Lemma}
\begin{proof}
We have the following commutative diagram
\[\begin{aligned}%
&\begin{matrix}0\to&S\bigl((\coverFA{X-C_j})^i, {\text{etc.}}\bigr)^n
\ \cap\ S\bigl(\coverFA{A}\ \cap\ (\coverFA{X-C_j})^i\bigr)
&\to&S\bigl((\coverFA{X-C_j})^i, {\text{etc.}}\bigr)\\
&\downlabeledarrow[\Big]{\alpha}{}&&\downlabeledarrow[\Big]{\beta}{}\\
0\ \hbox to 0pt{$\RA{\hskip 2cm}$\hss}\hfill&
S\bigl( \coverFA{A}\ \cap\ (\coverFA{X-C_j})^i\bigr)&
\hbox to 1cm{\hss$\RA{\hskip 2.5cm}$}&
S\bigl((\coverFA{X-C_j})^i\bigr)\\
\end{matrix}\\\noalign{\vskip 14pt}
&\begin{matrix}\to&({\text{the\ quotient\ complex}})&\to 0\\%
&\downlabeledarrow[\Big]{\gamma}{}\\
\to&S\bigl((\coverFA{X-C_j})^i, \coverFA{A}\ \cap\ (\coverFA{X-C_j})^i\bigr)&
\to0\\
\end{matrix}\\
\end{aligned}\]
where (the quotient complex) was used in defining
$\epsilon^{(n)}(X,A\Colon H_q,\coverFA{}\ )$.
Now, since $(X,A)$ is properly $1${--}connected, the subspace groups
$\epsilon^{(n)}(A;X\Colon H_q,\coverFA{}\ )$ and \\
$\epsilon(A; X\Colon H_q,\coverFA{}\ )$ are the absolute groups.
Since $(X,A)$ is properly $1${--}connected and properly $n${--}connected
at infinity for all $n$, \fullRef{L.1.3.3} says
$\epsilon(\alpha)$ is an isomorphism on homology.
Similarly, \fullRef{L.1.3.3} says
$\epsilon(\beta)$ is an isomorphism on homology.
Thus $\epsilon(\gamma)$ is an isomorphism on homology as asserted.
\end{proof}
\BEGIN{T.1.3.4}
Suppose $(X,A)$ is properly $1${--}connected and
properly $(n-1)${--}connected at $\infty$ for some $n\geq 2$.
Then the Hurewicz map
\[\epsilon(X,A\Colon \pi_n^\prime,\{\hat x_i\},\coverFA{}\ )\ \to\
\epsilon(X,A\Colon H_n,\{\hat x_i\},\coverFA{}\ )\]
is an isomorphism, where $\pi^\prime_n\bigl(
(\coverFA{X-C_j})^i, \coverFA{A}\ \cap\ (\coverFA{X-C_j})^i \ \cup\
\hat x_i,\hat x_i\bigr)$ is $\pi_n$ quotiented out by the action of
$\pi_1\bigl( \coverFA{A}\ \cap\ (\coverFA{X-C_j})^i \ \cup\
\hat x_i,\hat x_i\bigr)$.
\end{Theorem}
\begin{proof}
The usual Hurewicz theorem contains the fact that
\[\pi^\prime_n\bigl(
(\coverFA{X-C_j})^i, \coverFA{A}\ \cap\ (\coverFA{X-C_j})^i ,\hat x_i\bigr)\to
H_n^{(n-1)}\bigl((\coverFA{X-C_j})^i, \coverFA{A}\ \cap\
(\coverFA{X-C_j})^i ,\hat x_i\bigr)\] is an isomorphism.
Thus $\epsilon(X,A\Colon \pi^\prime_n, \{\hat x_i\},\coverFA{}\ )\ \to\
\epsilon^{(n-1)}(X,A\Colon H_n, \coverFA{}\ )$ is an isomorphism.
But \fullRef{L.1.3.4} says
$\epsilon^{(n-1)}(X,A\Colon H_n,\coverFA{}\ )\ \to\
\epsilon(X,A\Colon H_n, \{\hat x_i\},\coverFA{}\ )$ is an isomorphism.
\end{proof}
\BEGIN{T.1.3.5}
Suppose that $\epsilon(A\Colon \pi_1,\{ \hat x_i\},\coverFA{}_F\ )=0$
where $\coverFA{}_F$ is the cover over $A$ induced by the lift functor
from a cover $\coverFA{}$ over $X$.
Then the natural surjection\\
$\epsilon(X,A\Colon \pi_n, \{\hat x_i\},\coverFA{}\ )\ \to\
\epsilon(X,A\Colon \pi^\prime_n; \{\hat x_i\},\coverFA{}\ )$
is an isomorphism.
\end{Theorem}
\begin{proof}
Set $G_{i j}=\pi_n\bigl( (\coverFA{X-C_j})^i,
\coverFA{A}\ \cap\ (\coverFA{X-C_j})^i\ \cup\ \{\hat x_i\},\hat x_i\bigr)$
and $H_{i j}=\pi^\prime_n\bigl( (\coverFA{X-C_j})^i,
\coverFA{A}\ \cap\ (\coverFA{X-C_j})^i\ \cup\ \{\hat x_i\},\hat x_i\bigr)$.
Define $K_{i j}$ to be the kernel of $G_{i j}\ \to\ H_{i j}\ \to\ 0$.
$K_{i j}$ is generated by elements of the form $x-\alpha x$ where
$x\in \pi_n\bigl( (\coverFA{X-C_j})^i,
\coverFA{A}\ \cap\ (\coverFA{X-C_j})^i\ \cup\ \{\hat x_i\},\hat x_i\bigr)$
and $\alpha\in \pi_1\bigl(
\coverFA{A}\ \cap\ (\coverFA{X-C_j})^i\ \cup\ \{\hat x_i\},\hat x_i\bigr)$.
Since $\epsilon(A\Colon \pi_1,\{\hat x_i\},\coverFA{}_F\ )=0$,
the subspace principle says that we can assume the map
$\pi_1\bigl(
\coverFA{A}\ \cap\ (\coverFA{X-C_j})^i\ \cup\ \{\hat x_i\},\hat x_i\bigr)\ \to\
\pi_1\bigl(
\coverFA{A}\ \cap\ (\coverFA{X-C_{j-1}})^i\ \cup\ \{\hat x_i\},\hat x_i\bigr)$
is the zero map.
Then $K_{i j}\ \to\ K_{i\>j-1}$ takes $x - \alpha x$ to
$i_\ast(x)-i_\ast(\alpha x)=i_\ast(x)-i_\#(\alpha)\cdot i_\ast( x)=
i_\ast(x)-i_\ast(x)=0$, so this map is the zero map.
\end{proof}
\setcounter{footnote}{0}
\BEGIN{T.1.3.6}
Let $(X,A)$ be a properly $1${--}connected pair.
Then, for any covering functor $\coverFA{}$ on $X$, the natural map
$\epsilon(X,A\Colon \pi_n,\{\hat x_i\},\coverFA{}\ )\ \to\
\epsilon(X,A\Colon \pi_n,\{\hat x_i\},{\text{no\ cover}}\ )$
is an isomorphism\footnote{For $n\geq 2$.}.
\end{Theorem}
\begin{proof}
We have
\[\begin{matrix}%
\cdots\to\epsilon(A\Colon \pi_k,\{\hat x_i\},\coverFA{}_F\ )&\too&
\epsilon(X\Colon \pi_k,\{\hat x_i\},\coverFA{}\ )&\too&
\epsilon(X,A\Colon \pi_k,\{\hat x_i\},\coverFA{}\ )\to\cdots\\
\Big\downarrow&&\Big\downarrow&&\Big\downarrow\\
\cdots\to
\epsilon(A\Colon {\text{ditto}},{\text{no\ cover}})&\too&
\epsilon(X\Colon {\text{ditto}},{\text{no\ cover}})&\too&
\epsilon(X,A\Colon {\text{ditto}},{\text{no\ cover}})\to\cdots\\
\end{matrix}\]
commutes.
The first two maps are clearly isomorphisms for $k\geq2$, so the third
is for $k\geq3$.
Moreover
\[\begin{matrix}%
\epsilon(A\Colon \pi_1,\{\hat x_i\},\coverFA{}_F\ )&\to&
\epsilon(X\Colon \pi_1,\{\hat x_i\},\coverFA{}\ )\\
\Big\downarrow&&\Big\downarrow\\
\epsilon(A\Colon \pi_1,\{\hat x_i\},{\text{no\ cover}}\ )&\to&
\epsilon(X\Colon \pi_1,\{\hat x_i\},{\text{no\ cover}}\ )\\
\end{matrix}\]
is a pullback since it is obtained as the $\epsilon$
construction applied to pullbacks.
Hence the theorem remains true for $k=2$.
\end{proof}
\BEGIN{C.1.3.6.1}
Suppose $(X,A)$ is a properly $1${--}connected pair which is
$(n-1)${--}connected at $\infty$ for some $n\geq2$.
If $n=2$ assume $\epsilon(A\Colon \pi_1\{\hat x_i\},{\text{no\ cover}})\ \to\ \\
\epsilon(X\Colon \pi_1\{\hat x_i\},{\text{no\ cover}})$ is an isomorphism.
Then the Hurewicz map
\[\epsilon(X,A\Colon \pi_n,\{\hat x_i\},{\text{no\ cover}})\ \to\
\epsilon(X,A\Colon H_n,\{\hat x_i\},\coverFA{}\ )\] is an isomorphism
where\ \ $\coverFA{}$\quad is any universal covering functor for $X$.
\end{Corollary}
\BEGIN{T.1.3.7}
Theorems \shortFullRef{T.1.3.4}, \shortFullRef{T.1.3.5}
and \shortFullRef{T.1.3.6} are true (after appropriate
changes) with $\Delta$ instead of $\epsilon$.
They are also true for the absolute groups.
\end{Theorem}
\begin{proof} Easy.\end{proof}
\bigskip
Now suppose $(X,A)$ is a locally compact CW pair.
Then we might hope to improve our Hurewicz theorems by
getting information about the second non{--}zero map
(see \cite{bfortytwo}).
We do this following Hilton \cite{bthirteen}.
\begin{xDefinition}
Two proper maps $f$, $g\colon X\ \to\ Y$ are said to be \emph{properly
$n${--}homotopic} if for every proper map $\phi\colon K\ \to\ X$,
where $K$ is a locally compact CW complex of dimension $\leq n$,
$f \circ\ \phi$ is properly homotopic to $g\ \circ\ \phi$.
$X$ and $Y$ are of the \emph{same proper $n${--}homotopy type}
provided there exist proper maps $f\colon X\ \to\ Y$ and
$g\colon Y\ \to\ X$ such that $f\ \circ\ g$ and $g\ \circ\ f$ are properly
$n${--}homotopic to the identity.
Two locally compact CW complexes, $K$ and $L$, are said to be of
the same \emph{proper $n${--}type} \iff\ $K^n$ and $L^n$ have the same
proper $(n-1)${--}type.
A proper cellular map $f\colon K\ \to\ L$ is said to be a {\sl
proper $n${--}equivalence} provided there is a proper map
$g\colon L^{n+1}\ \to\ K^{n+1}$ with $f\vert_{K^{n+1}}\ \circ\ g$ and
$g\ \circ\ f\vert_{K^{n+1}}$ properly $n${--}homotopic to the identity.
\end{xDefinition}
A \emph{proper $J_m${--}pair}, $(X,A)$, is a properly $1${--}connected,
locally compact CW pair such that the maps
$\Delta(X^{n-1}\ \cup\ A, A\Colon \pi_n,\{\hat x_i\},{\text{no\ cover}})\to\
\Delta(X^{n}\ \cup\ A, A\Colon \pi_n,\{\hat x_i\},{\text{no\ cover}})$ are zero
for $2\leq n\leq m$.
A \emph{proper $J_m${--}pair at $\infty$} is the obvious thing.
\BEGIN{L.1.3.5}
The property of being a proper $J_m${--}pair is an invariant of
proper $m${--}type.
\end{Lemma}
\begin{proof} See Hilton \cite{bthirteen}.
\end{proof}
\BEGIN{T.1.3.8}
Let $(X,A)$ be a proper $J_m${--}pair at $\infty$.
Then the Hurewicz map
$h_n\colon \epsilon(X,A\Colon \pi_n,\{\hat x_i\},{\text{no\ cover}})\to
\epsilon(X,A\Colon H_n,\{\hat x_i\},\coverFA{}\ )$,
where\ $\coverFA{}$\quad
is a universal covering functor for $X$, satisfies
$h_n$ is an isomorphism for $n\leq m$ and $h_{m+1}$ is an epimorphism.
\end{Theorem}
\begin{proof}
See Hilton \cite{bthirteen}.
\end{proof}
\BEGIN{C.1.3.8.1}
The same conclusions hold for a proper $J_m${--}pair with the $\Delta$
groups.
\end{Corollary}
\BEGIN{C.1.3.8.2}
Let $(X,A)$ be a properly $(n-1)${--}connected, locally compact CW pair,
for $n\geq 2$.
If $n=2$ let $\Delta(A\Colon \pi_1,\{\hat x_i\},{\text{no\ cover}})\ \to\
\Delta(X\Colon \pi_1,\{\hat x_i\},{\text{no\ cover}})$ be an isomorphism.
Then the Hurewicz map
$h_n\colon \Delta(X,A\Colon \pi_n,\{\hat x_i\},{\text{no\ cover}})\to
\Delta(X,A\Colon H_n,\{\hat x_i\},\coverFA{}\ )$,
is an isomorphism, where\ $\coverFA{}$\quad
is a universal covering functor for $X$.
$h_{n+1}$ is an epimorphism.
\end{Corollary}
\begin{proof}
In section 5 we will see there is a locally finite $1${--}complex
$T\subseteq A$ such that $(A,T)$ is a proper $1/2${--}equivalence
and $\Delta(T\Colon \pi_k.\{\hat x_i\}),{\text{no\ cover}})=0$ for $k\geq1$.
Then $(T,T)$ is certainly a proper $J_n${--}complex.
$(T,T)\subseteq (X,A)$ is a proper $(n-1)${--}equivalence, so
$(X,A)$ is a $J_n${--}complex by \fullRef{L.1.3.5}.
\end{proof}
\BEGIN{T.1.3.9}
(Namioka \cite{btwentyeight})
Let $\phi\colon (X,A)\ \to\ (Y,B)$ be a map of pairs of
locally compact CW complexes.
Let $\phi\vert_X$ and $\phi\vert_A$ be properly $n${--}connected,
$n\geq 1$. ($\phi\vert_X$ and $\phi\vert_A$ should induce
isomorphisms on $\Delta(\quad\Colon \pi_1,\{\hat x_i\},{\text{no\ cover}})$
if $n=1$).
Then the Hurewicz map
$h_{n+1}\colon
\Delta\bigl((M_\phi\Colon M_{\phi\vert_A}, X)\Colon
\pi_{n+1},\{\hat x_i\}, {\text{no\ cover}}\bigr)\ \to\
\Delta\bigl((M_\phi\Colon M_{\phi\vert_A}, X)\Colon
H_{n+1},\{\hat x_i\}, \coverFA{}\ \bigr)$, where \ $\coverFA{}$\quad
is a universal covering functor of $M_\phi$, is an epimorphism.
\end{Theorem}
\begin{proof}
$(M_\phi\Colon M_{\phi\vert_A},X)$ is a triad and the groups in question
are the proper triad groups.
The reader should have no trouble defining these groups.
We can pick a set of base points for $(X,A)$ and it will also be a set for
our triad.
The triad groups fit into a long exact sequence
\[\cdots\to\
\Delta\bigl((M_{\phi\vert_A}); M_\phi)\ \to\
\Delta\bigl((M_{\phi}), X)\ \to\
\Delta\bigl((M_\phi\Colon M_{\phi\vert_A},X)\bigr)\ \to\cdots \quad,\]
where again we get the subspace groups.
Since $\phi\vert_{A}$ and $\phi\vert_{X}$ are properly $n${--}connected,
$h_m$ for $(M_\phi,A)$ is an isomorphism for $m\leq n$ and
an epimorphism for $m=n+1$.
By the subspace principle, $h_m$ for
$\bigl((M_{\phi\vert_A}, A ); M_\phi\bigr)$ is an isomorphism for
$m\leq n$ and an epimorphism for $m=n+1$.
The strong version of the $5${--}lemma now shows the triad $h_n$ an
isomorphism and the triad $h_{n+1}$ an epimorphism.
\end{proof}
\insetitem{Notation}
$\Delta_\ast(X,A\Colon \coverFA{}\ )$ will hereafter denote
$\Delta(X,A\Colon H_\ast,\{\hat x_i\},\coverFA{}\ )$
for some set of base points for the pair $(X,A)$.
Similar notation will be employed for homology $n${--}ad groups,
subspace groups, etc.
\medskip
We conclude this section with some definitions and computations.
\begin{xDefinition}
An homogamous space $X$ is said to have \emph{monomorphic ends},
provided \\
$\displaystyle\Delta(X\Colon \pi_1,\{ x_i\},{\text{no\ cover}})\to
\mathop{\bigtimes}_{i\in{\mathcal I}}\ \pi_1(X,x_i)$ is a monomorphism
(equivalently $\displaystyle\epsilon\ \to\ \mathop{\mu}_{i}$
is a monomorphism).
A space has \emph{epimorphic ends} provided the above map is onto,
and \emph{isomorphic ends} if the map is an isomorphism.
\end{xDefinition}
As examples, if $X$ is an homogamous space which is not compact,
$X\times\R$ has one, isolated end (see \cite{bthirtytwo}) which is
epimorphic.
$X\times\R^2$ has isomorphic ends.
These results use Mayer{--}Vietoris to compute the number of ends of
$X\times\R$ and van{--}Kampen to yield the $\pi_1$ information,
using the following pushout
\[\begin{matrix}%
(X-C)\times Y-D)&\to&X\times (Y-D)\\
\Big\downarrow&&\Big\downarrow\\
(X-C)\times Y&\to&X\times Y\ -\ C\times D\\
\end{matrix}\]
In fact, this diagram shows that if $X$ and $Y$ are not compact
(but are path connected),
$X\times Y$ has one end, which is seen to be epimorphic since
$\pi_1(X\times Y\ -\ C\times D)\ \to\ \pi_1(X\times Y)$ is
easily seen to be onto.
If $X$ has epimorphic ends, $\pi_1(X-C,p)\to\pi_1(X,p)$
must always be onto, so if $X$ and $Y$ have epimorphic ends,
$X\times Y$ has one isomorphic end.
Monomorphic ends are nice for then the third example of covering functor
that we gave (the universal cover of $X$ bit no more) becomes a
universal covering functor.
Farrell and Wagoner (\cite{bnine} or \cite{beleven}) then showed \
that a proper map $f\colon X\ \to\ Y$, $X$, $Y$ locally compact CW,
with $X$ having monomorphic ends is a proper homotopy equivalence
provided it is a properly $1${--}connected map; a homotopy equivalence;
and $f^\ast\colon H^\ast_\cmpsup(\tilde Y)\ \to\
H^\ast_\cmpsup(\tilde X)$ is an isomorphism
where $\tilde{\ }$ denotes the universal cover (coefficients
are the integers).
\section{Proper cohomology, coefficients and products}
\newHead{I.4}
In attempting to understand ordinary homotopy theory, cohomology theory
is an indispensable tool.
In ordinary compact surgery, the relationship between homology and
cohomology in Poincar\'e duality spaces forms the basis of many of the results.
To extend surgery to paracompact objects, we are going to need
a cohomology theory.
If one grants that the homology theory that we
constructed in section 3 is the right one,
then the correct cohomology theory is not hard to intuit.
To be loose momentarily, in homology we associate to each compact set
$C$ the group $H_\ast(\coverFA{X-C})$.
If $M-C$ is a manifold with boundary, Lefschetz duality tells us this is
dual to
$H^\ast_\cmpsup(\coverFA{\Bar{X-C}},\coverFA{\partial C})$,
where $\Bar{M-C}$ is the closure of $M-C$.
If $C\subseteq D$, we have a map
$H_\ast(\coverFA{M-D})\ \to\ H_\ast(\coverFA{M-C})$,
so we need a map
$H^\ast_c(\coverFA{\Bar{M-D}},\coverFA{\partial D})\ \to\
H^\ast_c(\coverFA{\Bar{M-C}},\coverFA{\partial C})$.
A candidate for this map is
\[
H^\ast_\cmpsup(\coverFA{\Bar{M-D}},\coverFA{\partial D})\ \RA{\text{tr}}\
H^\ast_\cmpsup(\dottedBar{\Bar{M-C}},\dottedBar{\partial D})\ \LA{ex}\
H^\ast_\cmpsup(\coverFA{\Bar{M-C}},\coverFA{\Bar{D - C}})\ \RA{\text{inc}}\
H^\ast_\cmpsup(\coverFA{\Bar{M-C}},\coverFA{\partial C})
\]
where $\dottedBar{\Bar{M - D}}=\pi^{-1}(M - D)$
($\pi\colon \coverFA{M-C}\ \to\ M-C$),
${\text{inc}}$ is the map induced by inclusion, ${\text{tr}}$ is the trace and
${\text{ex}}$ is an excision map.
The first problem that arises is that ${\text{ex}}$ need not be an isomorphism.
This problem is easily overcome.
We define ${\mathcal O}(X)$ to be the category whose objects are
open subsets of $X$ whose closure (in $X$) is compact.
If $U$, $V\in{\mathcal O}(X)$, there is a morphism $U\ \to\ V$ \iff\
$\Bar U\subseteq V$ or $U=V$.
${\mathcal O}(X)$ will be our diagram scheme.
Note we have a functor ${\mathcal O}(X)\ \to {\mathcal D}(X)$ which sends
$U\mapsto \Bar U$.
Since $X$ is locally compact, this functor has a cofinal image ($X$ is
homogamous, hence locally compact).
The second problem which arises concerns covering functors.
Since $X - U$, $U\in{\mathcal O}(X)$ is closed, it is hard to get conditions
on $X$ so that $X-U$ has arbitrary covers.
There are two solutions to this problem.
We can restrict ${\mathcal O}(X)$ (e.g. if $X$ is an homogamous CW complex,
and if we pick sets $U$ so that $X-U$ is a subcomplex, then we
always have covers), or we can ignore the problem.
We choose the latter alternative, and when we write $\coverFA{}$ is
a covering functor for $X$, we mean \ $\coverFA{}$\quad is
compatible with $X-U$ for each $U\in{\mathcal O}(X)$.
It is not hard to see that if $X$ is locally $1${--}connected, then
universal covering functors exist despite the fact that the universal
covering functor need not.
Now we could have defined homology and homotopy groups
using ${\mathcal O}(X)$ instead of ${\mathcal D}(X)$.
Given a covering functor for ${\mathcal O}(X)$ there is an obvious one for
${\mathcal D}(X)$.
It is not hard to show that the homology and homotopy groups for $X$
are the same whether one uses ${\mathcal O}(X)$ or ${\mathcal D}(X)$.
\begin{xDefinition}
$\Delta_\ast(X;A_1,\cdots, A_n\Colon \coverFA{}\ , \Gamma)$,
where $\Gamma$ is a local coefficient system on $X$, denotes the
$\Delta${--}construction applied to $G_{i U}=
H_\ast\bigl((\coverFA{X-U})^i;
\coverFA{A}_1\ \cap\ (\coverFA{X-U})^i, \cdots,
\coverFA{A}_n\ \cap\ (\coverFA{X-U})^i\Colon i^\ast\Gamma\bigr)$,
where the homology group is the ordinary (singular) $n${--}ad
homology group with coefficients $i^\ast\Gamma$, where
$i^\ast\Gamma$ is the local system induced from $\Gamma$ by
the composite $(\coverFA{X-U})^i\ \RA{\pi}\ X-U\subseteq X$.
\end{xDefinition}
\begin{xDefinition}
$\Delta^\ast(X\Colon \coverFA{}\ ,\Gamma)$ is the $\Delta${--}construction
applied to \[G_{i U}= H_\cmpsup^\ast\bigl((\coverFA{X-U})^i,
\coverFA{\partial U}\ \cap\ (\coverFA{X-U})^i; i^\ast\Gamma\bigr)\]
($\partial U = {\text{frontier\ of\ }} U {\text{\ in\ }}X$).
$\Delta^\ast(X,A\Colon \coverFA{}\ ,\Gamma)$ is the $\Delta${--}construction
applied to
\[G_{i U} = H^\ast_\cmpsup\bigl( (\coverFA{X-U})^i;
\coverFA{\partial U}\ \cap\ (\coverFA{X-U})^i,
\coverFA{A}\ \cap\ (\coverFA{X-U})^i; i^\ast\Gamma\bigr)\ .\]
\end{xDefinition}
\insetitem{{\bf Caution}} $(X,A)$ must be a proper pair (i.e. $A\subseteq X$ is proper)
before $H^\ast_\cmpsup(X,A)$ makes sense.
A similar remark applies to $n${--}ads.
$\Delta^\ast(X; A_1, \cdots, A_n\Colon \coverFA{}\ ,\Gamma)$ is defined
similarly.
In our definition we have not defined our maps
$G_{i V}\ \to\ G_{i U}$ if $\Bar U\subseteq V$.
If $\Bar{X - V} = \pi^{-1}(X-V)$, where $\pi\colon (\coverFA{X-U})^i\ \to\
X-U$, then the map is the composite
\[\begin{aligned}%
H^\ast_\cmpsup&\bigr((\coverFA{X-V})^i, \coverFA{\partial V}\ \cap\
(\coverFA{X-V})^i; i^\ast\Gamma\bigr)\ \RA{{\text{tr}}}\
H^\ast_\cmpsup(\Bar{X-V},\Bar{\partial V}; \Gamma_1)
\
\rightlabeledarrow{\ ex \ }{\cong}
\ \\
&H^\ast_c\bigr((\coverFA{X-U})^i, (\coverFA{V-U})\ \cap\ (\coverFA{X-U})^i;
\Gamma_2\bigr)\ \RA{\ {\text{inc}}\ }\
H^\ast_c\bigr((\coverFA{X-U})^i, \coverFA{\partial U}\
\cap\ (\coverFA{X-U})^i; i^\ast\Gamma\bigr)\\\end{aligned}\]
where $\Gamma_1$ and $\Gamma_2$ are the obvious local systems.
A similar definition gives the map in the pair and $n${--}ad cases.
Once again we get long exact sequences modulo
the usual subspace difficulties.\\
We let $\Delta^\ast(A; X\Colon \coverFA{}\ ,\Gamma)$ denote the subspace
group with a similar notation for the sub{--}$n${--}ad groups.
Again we get a subspace principle.
Lastly, the cohomology groups are ``independent'' of base points
(compare \fullRef{T.1.3.2}) and are invariant under
proper homotopy equivalence.
The proofs of these results should be easy after section 3, and hence
they are omitted.
One reason for the great power of cohomology is that we have various
products.
The first product we investigate is the cup product.
\BEGIN{T.1.4.1}
There is a natural bilinear pairing, the cup product
\[H^m(X,A;\Gamma_1)\times
\Delta^n(X,B\Colon \coverFA{}\ ,\Gamma_2)\ \to\
\Delta^{m+n}(X;A,B\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)\ .\]
If $\{ A, B\}$ us a properly excisive pair, the natural map
\[\Delta^n(X,A\ \cup\ B\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)\ \to\
\Delta^n(X;A,B\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)\]
is an isomorphism, so we get the ``usual'' cup product.
\end{Theorem}
\begin{proof}
Given $\varphi\in H^m(X,A;\Gamma_1)$, define, for any $U\in{\mathcal O}(X)$,
$\varphi_U\in H^m\bigl((\coverFA{X-U})^i,
\coverFA{A}\ \cap\ (\coverFA{X-U})^i; i^\ast\Gamma_1\bigr)$ as the image of
$\varphi$ under the composite
\[H^m(X,A;\Gamma_1)\ \to\ H^m\bigl((X-U), (A-U);\Gamma_1)
\ \RA{\ \pi^\ast\ }\
H^m\bigl((\coverFA{X-U})^i,
\coverFA{A}\ \cap\ (\coverFA{X-U})^i,i^\ast\Gamma_1\bigr)\ .\]
One then checks that if $G_{i U}=
H^n_\cmpsup\bigl((\coverFA{X-U})^i,
\coverFA{B}\ \cap\ (\coverFA{X-U})^i, \coverFA{\partial U}\ \cap\
(\coverFA{X-U})^i, i^\ast\Gamma_2\bigr)$ and if
$H_{i U} = $ the corresponding group for
$\Delta^{m+n}(X;A,B\Colon \coverFA{}\ ,
\Gamma_1\otimes\Gamma_2)$, then
\[\begin{matrix}%
G_{i U}&\RA{\ \cup\>\varphi_U\ }&H_{i U}\\
\Big\downarrow&&\Big\downarrow\\
G_{i V}&\RA{\ \cup\>\varphi_V\ }&H_{i V}\\
\end{matrix}\]
commutes.
Hence the maps $\cup\>\varphi_U$ give us a map
$\Delta^n(X,B\Colon \coverFA{}\ ,\Gamma_2)\ \to\
\Delta^{m+n}(X; A, B\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)$.
One easily checks this map gives us a natural bilinear pairing.
Now we have a natural map
$\Delta^\ast(X, A\ \cup\ B)\ \to\ \Delta^\ast(X; A,B )$.
We get a commutative diagram
\[\begin{matrix}%
\cdots\to&\Delta^\ast(X, A\ \cup\ B)&\to& \Delta^\ast(X,A)&\to&
\Delta^\ast(A\ \cup\ B, A;X)&\to\cdots\\
&\Big\downarrow&&\Big\downarrow&&\Big\downarrow\\
\cdots\to&\Delta^\ast(X; A, B)&\to& \Delta^\ast(X,A)&\to&
\Delta^\ast(B, A\ \cap\ B;X)&\to\cdots\\
\end{matrix}\]
where the rows are exact.
$\{A, B\}$ a properly excisive pair implies
$\Delta^\ast(A\ \cup\ B, A)\ \to\ \Delta^\ast(B, A\ \cap\ B)$
is an isomorphism for a set of base points in $A\ \cap\ B$ which is
a set for $A$, $B$, and $A\ \cup\ B$.
The subspace principle now shows the right hand map is an isomorphism.
The middle map is the identity, so the left hand map is an isomorphism.
This establishes the last part of our claim.
\end{proof}
For completeness we give the definition of a properly excisive pair.
\begin{xDefinition}
A pair $\{A, B\}$ of homogamous spaces is said to be properly excisive
with respect to some covering functor \ $\coverFA{}$\ ,
provided
\[\Delta^\ast(A\ \cup\ B\Colon A, B\Colon \coverFA{}\ ,\Gamma)=
\Delta_\ast(A\ \cup\ B\Colon A, B\Colon \coverFA{}\ ,\Gamma)=
H^0_{{\text{end}}}(\coverFA{A\ \cup\ B}; \coverFA{A}, \coverFA{B}; \Gamma)=0\]
for any local system $\Gamma$.
\medskip
The pair is properly excisive if it is properly excisive with respect to all
covering functors compatible with $A\ \cup\ B$.
\end{xDefinition}
\medskip
The other product of great importance is the cap product.
We get two versions of this ( Theorems \shortFullRef{T.1.4.2}
and \shortFullRef{T.1.4.3}).
\BEGIN{T.1.4.2}
There is a natural bilinear pairing, the cap product
\[\Delta^m(X,A\Colon \coverFA{}\ ,\Gamma_1)\times
H^\locf_{n+m}(X; A, B;\Gamma_2)\ \to\
\Delta_{n}(X,B\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)\]
If $\{A,B\}$ is a properly excisive pair, we can define the ``usual''
cap product.
\end{Theorem}
\begin{proof}
Let $C\in H^\locf_{n+m}(X;A,B;\Gamma_2)$.
Define \[C_U\in H^\locf_{n+m}\bigl((\coverFA{X-U})^i;
\coverFA{A}\ \cap\ (\coverFA{X-U})^i, \coverFA{B}\ \cap\ (\coverFA{X-U})^i,
\coverFA{\partial U}\ \cap\ (\coverFA{X-U})^i; i^\ast\Gamma_2\bigr)\]
as the image of $C$ under the composite
\[\begin{matrix}%
H^\locf_{n+m}(X; A, B;\Gamma_2)\\
\downarrow\\
H_{n+m}^\locf(X;A,B,U;\Gamma_2)&
\leftlabeledarrow{\text{\ ex\ }}{\cong}&
H_{n+m}^\locf(X-U;A-U,B-U, \partial U;\Gamma_2)\\
&&\downlabeledarrow{{\text{tr}}}{}\\
\hbox to 0pt{$H^\locf_{n+m}\bigl((\coverFA{X-U})^i;
\coverFA{A}\ \cap\ (\coverFA{X-U})^i, \coverFA{B}\ \cap\ (\coverFA{X-U})^i,
\coverFA{\partial U}\ \cap\ (\coverFA{X-U})^i; i^\ast\Gamma_2\bigr)$ .\hss}
\\\end{matrix}\]
One can check that $\cap\ C_U$ satisfies the necessary commutativity
relations to define a map
\[\Delta^m(X,A\Colon \coverFA{}\ ,\Gamma_1)\ \to\
\Delta_n(X,B\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)\ .\]
If $\{A, B\}$ is properly excisive, $H^\ast_\cmpsup(A\ \cup\ B;A, B)=0$
follows from $\Delta^\ast=0$.
Universal coefficients shows $H^\locf_\ast(A\ \cup\ B;A,B)=0$, so the
standard exact sequence argument shows
$H^\locf_\ast(X;A,B;\Gamma_2)\cong
H^\locf_\ast(X,A\ \cup\ B;\Gamma_2)$.
\end{proof}
\BEGIN{T.1.4.3}
There is a natural bilinear pairing, the cap product
\[H^m(X,A;\Gamma_1)\times
\Delta_{n+m}(X;A,B\Colon \coverFA{}\ ,\Gamma_2)\ \to\
\Delta_n(X,B\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)\ .\]
If $\{A,B\}$ is a properly excisive pair, we can define the
``usual'' cap product.
\end{Theorem}
\begin{proof}
Given $\varphi\in H^m(X,A;\Gamma_1)$, define, for any $U\in{\mathcal O}(X)$,
$\varphi_U\in H^m\bigl((\coverFA{X-U})^i,
\coverFA{A}\ \cap\ (\coverFA{X-U})^i; i^\ast\Gamma_1\bigr)$ as the image of
$\varphi$ under the composite
\[H^m(X,A;\Gamma_1)\ \to\ H^m\bigl((X-U), (A-U);\Gamma_1)
\ \RA{\ \pi^\ast\ }\
H^m\bigl((\coverFA{X-U})^i,
\coverFA{A}\ \cap\ (\coverFA{X-U})^i,i^\ast\Gamma_1\bigr)\ .\]
One checks again that the necessary diagrams commute.
The statement about $\{A,B\}$ follows from the $5${--}lemma and
the subspace principle.
\end{proof}
We will also need a version of the slant product for our theory.
To get this we need to define a group for the product of two ``ads''.
As usual we apply the $\Delta${--}construction to a particular situation.
Pick a set of base points for $X$ and a set for $Y$.
Our indexing set is the Cartesian product of these two sets.
Our diagram is ${\mathcal O}(X)\times{\mathcal O}(Y) =
\{ U\times V\subseteq X\times Y\ \vert\ U\in{\mathcal O}(X),
\ V\in{\mathcal O}(Y)\}$.
\[\begin{aligned}%
G^{i\times j}_{U\times V}=
H_\ast\bigl(&(\coverFA{X-U})^i\times(\coverFC{Y-V\ })^j;
\bigl(\coverFA{A}_1\ \cap\ (\coverFA{X-U})^i\bigr)\times (\coverFC{Y-V\ })^j,
\cdots, \\&
\bigl(\coverFA{A}_n\ \cap\ (\coverFA{X-U})^i\bigr)\times (\coverFC{Y-V\ })^j,
(\coverFA{X-U})^i\times\bigl(\coverFA{B_1}\ \cap\
\coverFC{Y-V})^j,\cdots,\\&
(\coverFA{X-U})^i\times\bigl(\coverFA{B_m}\ \cap\ \coverFC{Y-V})^j;
i^\ast\Gamma\bigr)\quad .\\
\\\end{aligned}\]
The resulting groups will be denoted
$\Delta_\ast\bigl( (X;A_1,\cdots, A_n)\times (Y;B_1,\cdots, B_m)\Colon
\coverFA{}\ , \coverFC{\hskip10pt }\ ,\Gamma)$ ($\Gamma$ is
some local system on $X\times Y$).
\BEGIN{T.1.4.4}
There is a natural bilinear pairing, the slant product
\[\begin{matrix}%
\slantp\colon H^m(\coverFC{Y\ }; \coverFC{B_1},\cdots, \coverFC{B_m};
\Gamma_1)\times
\Delta_{n+m}\bigl( (X;A_1,\cdots, A_n)\times (Y;B_1,\cdots, B_m)\Colon
\coverFA{}\ , \coverFC{\hskip10pt }\ ,\Gamma_2)\\
\downarrow\\
\Delta_n(X;A_1,\cdots, A_n,\coverFA{}\ ,\Gamma_1\otimes\Gamma_2)
\\
\end{matrix}\]
\end{Theorem}
\begin{proof}
For $\varphi\in H^m(\coverFC{Y\ }; \coverFC{B_1},\cdots, \coverFC{B_m};
\Gamma_1)$,
define $\varphi_V$ by analogy with the definition in
\fullRef{T.1.4.1}.
These give us the necessary maps.
\end{proof}
\BEGIN{C.1.4.4.1}
If $d\colon X\ \to\ X\times X$ is the diagonal, and \\if
$C\in \Delta_{n+m}(X; A, B\Colon \coverFA{}\ ,\Gamma_1)$, and if
$\varphi\in H^m(X,A;\Gamma_2)$, then
\[\varphi\ \cap\ C\ =\ \varphi\slantp d_\ast C\ .\quad\mathqed\]
\end{Corollary}
Using our slant product, we can define the cap product of
\fullRef{T.1.4.3} on the chain level.
There are two basic chain groups we would like to use,
For an homogamous CW complex we would like to use the cellular chains,
and when $X$ is a paracompact manifold with a locally finite handlebody
decomposition, we want to use the chains based on the handles.
We do the former case and leave the reader to check the theory still
holds in the latter.
If $X$ is an homogamous CW complex,
we define
\[P_\ast(X;A,B\Colon \coverFA{}\ ,\Gamma)=
\Delta_\ast(X^\ast; X^{\ast-1}, A^\ast, B^\ast\Colon \coverFA{}\ , \Gamma)\]
(where $A^\ast=A\ \cap X^\ast$, etc.) for $\ast\geq 2$.
If $\ast=0$ or $1$, we must use subspace groups
\[\Delta_\ast\bigl((X^\ast;X^{\ast-1},A^\ast,B^\ast);X\Colon
\coverFA{}\ ,\Gamma)\ .\]
$A$ and $B$ are subcomplexes.
Similarly define
\[\begin{aligned}%
P^\ast(A;A,B\Colon \coverFA{}\ ,\Gamma)&=
\Delta^\ast(X^\ast; X^{\ast-1}, A^\ast, B^\ast\Colon \coverFA{},\Gamma)\\
C^\locf_\ast(X\Colon A,B)&=
H^\locf_\ast(X^\ast; X^{\ast-1}, A^\ast, B^\ast)\\
C^\ast(X\Colon A,B)&= H^\ast(X^\ast; X^{\ast-1}, A^\ast, B^\ast)\quad .\\
\end{aligned}\]
The triple $(X^\ast, X^{\ast-1}, X^{\ast-2})$ gives us a boundary map
$P_\ast\ \to\ P_{\ast-1}$, $P^\ast\ \to\ P^{\ast+1}$, etc.
This boundary map makes the above objects into chain complexes
($\partial\partial=0$), and by \fullRef{C.1.2.2.2},
the homology of these complexes is just what one expects.
A diagonal approximation
\[h_\ast\colon P_\ast(X; A, B\Colon \coverFA{}\ ,\Gamma)\ \to\
\Delta_\ast\Bigl(\bigl( (X,A)\times (X,B)\bigr)^\ast\Colon
\coverFA{}\ , \Gamma\Bigr)\]
is a cellular approximation to $d\colon X\ \to\ X\times X$ with a homotopy
between $d$ and the cellular map,
$H\colon X\times I\ \to\ X\times X$, such that $\pi_1\ \circ H$ and
$\pi_2\ \circ H$ are proper.
$\bigl( (X,A)\times (X,B)\bigr)^\ast$ is just
$\displaystyle\mathop{\cup}_k (X,A)^k\times (X,B)^{\ast-k}$.
Any two such diagonal approximations are cellularly homotopic so that
the homotopy composed with the projections is proper.
\BEGIN{T.1.4.5}
Given any diagonal approximation $h$, there is a bilinear pairing
\[B_h\colon
C^m(X,A;\Gamma_1)\times
P_{n+m}(X;A,B\Colon \coverFA{}\ ,\Gamma_2)\ \to\
P_n(X,B\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)\ .\]
If $f\in C^m(X,A;\Gamma_1)$ and
$c\in P_{n+m}(X;A,B\Colon \coverFA{}\ ,\Gamma_2)$, then
\[\partial B_h(f,c)= (-1)^{n} B_h(\delta f, c) + B_h(f, \partial c)\ .\]
Hence we get an induced pairing on the homology level.
Any two $B_h(f,\quad)$ are chain homotopic, so the pairing on homology
does not depend on the diagonal approximation.
This pairing on homology is the cap product of
\fullRef{T.1.4.3}.
\end{Theorem}
\begin{proof}
Consider the element $h_\ast(c)\in
\Delta_{n+m}\Bigl(\bigl((X,A)\times(X,B)\bigr)^{n+m},{\text{etc.}}\Bigr)$.
The group
\[\Delta_{n+m}\bigl((X^n;X^{n-1},A^n)\times (X^m; X^{m-1}, B^m)\Colon
\coverFA{}\ ,\Gamma_2)\]
lies as a natural summand of this first group.
Let $p^n_m$ be this projection.
Then $B_h(f,c) = f\Big\vert_{p^n_m\bigl(h_\ast(c)\bigr)}$.
The rest of the proof involves checking this definition has all the asserted
properties.
\end{proof}
We also want to define the cap product of
\fullRef{T.1.4.2} on the chain level.
Unfortunately, there is no slant product of the needed type,
so we must use brute force.
\BEGIN{T.1.4.6}
Given any diagonal approximation $h$, there is a bilinear pairing
\[B_h\colon
P^m(X,A\Colon \coverFC{}\ ,\Gamma_1)\times
C^\locf_{n+m}(X;A,B;\Gamma_2)\ \to\
P_n(X,B\Colon \coverFC{}\ ,\Gamma_1\otimes\Gamma_2)\ .\]
If $f\in P^m(X,A\Colon \coverFC{}\ ,\Gamma_1)$ and
$c\in C^\locf_{n+m}(X;A,B;\Gamma_2)$, then
\[\partial B_h(f,c)=(-1)^{n}B_h(\delta f,c) + B_h(f, \partial c)\quad .\]
Hence we get an induced pairing (independent of $h$) on the homology
level.
This pairing is the cap product of \fullRef{T.1.4.2}.
\end{Theorem}
\begin{proof}
Let $c\in C^\locf_{n+m}(X;A,B;\Gamma_2)$.
Define \[c_U\in C^\locf_{n+m}\bigl( (\coverFC{X-U})^i;
\coverFC{A\ }\ \cap\ (\coverFC{X-U})^i, \coverFC{B\ }\ \cap\ (\coverFC{X-U})^i,
\coverFC{\partial U\ }\ \cap\ (\coverFC{X-U})^i; i^\ast\Gamma_2\bigr)\]
by excision and trace as in the proof of \fullRef{T.1.4.2}.
We define $B_h(\quad, c)$ from the maps\hfill\\\noindent
$H^m_\cmpsup\bigl( \coverFC{(X-U)^i\ \cap\ X^m};
\coverFC{(X-U)^i\ \cap\ X^{m-1}}, {\text{etc.}}\bigr)\ \RA{\ \slantp b_U\ }\
$\hfil\par\noindent$H_n\bigl( \coverFC{(X-U)^i\ \cap\ X^n};
\coverFC{(X-U)^i\ \cap\ X^{n-1}}, {\text{etc.}} \bigr)$
where $\slantp$ is the slant product and $b_U$ is the homology class
which is the image of $c_U$ under the following composite.
\[\begin{aligned}%
H^\locf_{n+m}\bigl(& \coverFC{(X-U)^i\ \cap\ X^m\ }; {\text{etc.}} \bigr)
\RA{\ h_\ast\ }
H^\locf_{n+m}\Bigl( \bigl( (\coverFC{X-U})^i\times(\coverFC{X-U})^i\bigr)
\cap
\coverFC{(X\times X)^{n+m}}; {\text{etc.}} \Bigr)\\
&\RA{\ p^n_m\ }\
H^\locf_{n+m}\Bigl( \bigl( \coverFC{(X-U)^i\ \cap\ X^m\ }\bigr) \times
\bigl( \coverFC{(X-U)^i\ \cap\ X^n\ }\bigr); {\text{etc.}}\Bigr)
\\\end{aligned}\]
(superscript $i$ denotes a component containing $\hat x_i$,
and superscripts $n$, $m$, and $n+m$ denote skeletons.)
Note in passing that $h_\ast({\text{tr}}\ b_U) \neq {\text{tr}}(h_\ast b_U)$,
which is why we are unable to define a general slant product like
\fullRef{T.1.4.4} to cover this case.
The rest of the proof involves verifying diagrams commute and
verifying our equation.
\end{proof}
\medskip
Lastly we prove the Browder lemma, which will be essential in our
study of Poincar\'e duality.
\BEGIN{T.1.4.7}
Let $(X,A)$ be a proper pair, and let $c\in H^\locf_{n}(X,A;\Gamma_2)$.
Then
\[\begin{aligned}\begin{matrix}%
\Delta^{\ast-1}(A;X\Colon \coverFA{}\ ,\Gamma_1)&\to&
\Delta^\ast(X,A\Colon \coverFA{}\ ,\Gamma_1)&\to\\
\downlabeledarrow[\Big]{\ \cap\ (-1)^{n-1}\partial c}{}
&&\downlabeledarrow[\Big]{\cap\ c}{}\\\noalign{\vskip4pt
\Delta_{n-\ast}(A;X\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)&\to&
\Delta_{n-\ast}(X\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)&\to&
\\\end{matrix}\\\noalign{\vskip20pt}
\begin{matrix}%
\Delta^\ast(X\Colon \coverFA{}\ ,\Gamma_1)&\to&
\Delta^\ast(A;X\Colon \coverFA{}\ ,\Gamma_1)&\\
\downlabeledarrow[\Big]{\cap\ c}{}&&
\downlabeledarrow[\Big]{\cap\ \partial c}{}\\\noalign{\vskip4pt
\Delta_{n-\ast}(X,A\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)&\to&
\Delta_{n-1-\ast}(A;X\Colon \coverFA{}\ ,\Gamma_1\otimes\Gamma_2)
&
\\\end{matrix}\\\end{aligned}
\]
commutes.
\end{Theorem}
\begin{proof}
The usual Browder lemma (see section 1) says that the corresponding
diagram commutes for ordinary homology and cohomology with compact
supports.
Commutativity is then trivial for the above diagram.
(While we have not defined a cap product for subspace groups,
the reader should have no difficulty writing down the necessary maps.)
\end{proof}
\bigskip
\section{Chain complexes and simple homotopy type}
\newHead{I.5}
In our $\Delta${--}construction as applied to the homology or homotopy
functors, we still have some structure that we have not utilized.
As an example of this extra structure, let us consider $\epsilon(X\Colon \pi_1)$.
This is an inverse limit\\
$\displaystyle\lim_{\to} \mu\bigl( \pi_1(X-C, x_i)\Bigr)$.
Now many of the $\pi_1(X-C, x_i)$ are isomorphic.
(Unfortunately this isomorphism are not unique but depends on paths
joining $x_i$ to $x_j$.)
Our $\epsilon${--}construction makes no use of this fact.
In order to be able to make effective use of this extra structure, we
need a way to choose the above isomorphisms.
We will do this through the concept of a tree.
A \emph{tree for an homogamous space} $X$ will be a $1${--}dimensional,
locally finite, simplicial complex, $T$, such that
\begin{enumerate}
\item[1)] $\Delta(T\Colon \pi_k)=0$ for $k>0$
\item[2)] If $T^\prime\subseteq T$ is a subcomplex of $T$, $T^\prime$
has the proper homotopy type of $T$ \iff\ $T=T^\prime$.
\end{enumerate}
\noindent
(This last condition is to insure that
$\displaystyle\begin{matrix}&&\circ&&\circ\\
&&\vert&&\vert\\
\circ&\vrule width 10pt height 4.4pt depth -3.9pt&
\circ&\vrule width 10pt height 4.4pt depth -3.9pt&
\circ&\vrule width 10pt height 4.4pt depth -3.9pt&
\circ&\cdots
\\
\end{matrix}$
is not a tree for $\R^2$ but rather
$\circ\ \vrule width 10pt height 4.4pt depth -3.9pt\
\circ\ \vrule width 10pt height 4.4pt depth -3.9pt\
\circ\ \cdots$
is.)
We also require a map $f\colon T\ \to\ X$
which is properly $1/2${--}connected.
Two trees $(T,f)$ and $(S,g)$ are equivalent provided there is a
proper homotopy equivalence $h\colon T\ \to\ S$
with $g\ \circ h$ properly homotopic to $f$.
A space $X$ is said to have a tree provided $X$ is homogamous and
there is a tree for $X$.
Any locally path connected homogamous space has a tree.
To see this, let $\{ p\}$ be a set of base points for our space $X$.
Let $\{C_i\}$ be a cofinal collection of compact subsets of $X$.
We can assume $X$ is path connected since we can do each path
component separately.
We may assume $\{ p\} \ \cap\ C_0\neq\emptyset$.
Pick a point $p_0\in\{ p\}\ \cap\ C_0$.
Look at the components of $X-C_0$ with a point of $\{p\}$ in them.
As we showed in the proof of \fullRef{P.1.1.2},
there are only finitely many such components of $X- C_0$.
The components whose closure is not compact are called essential
components.
We may assume $\{p\}\ \cap\ ({\text{each\ essential\ component\ of\ }}
X-C)\ \cap\ C_1 \notin C_1\neq \emptyset$
since this is true for some compact set.
Let $p_1^{\alpha_1}$, $p_1^{\alpha_2}$, \dots , $p_1^{\alpha_n}$
be a subset of $\{p\}\ \cap\ C_1$, one for each essential component
of $X-C_0$.
Join $p_1^{\alpha_i}$ to $p_0$ by a path $\lambda_{1,i}$.
Now look at the essential components of $X-C_1$.
Pick $p_2^{\alpha_1}$, \dots , $p_2^{\alpha_m}$
(which we may assume are in $C_2$), one for each essential component
of $X-C_1$.
Each $p_2^{\alpha_i}$ lies in an essential component of $X-C_0$, so
pick paths $\lambda_{2,i}$ which join $p_2^{\alpha_i}$ to the
appropriate element in $\{p_1^\alpha\}$.
These paths should lie in $X-C_0$.
(Since $X$ is locally path connected, the components of $X-C_0$ are
path connected.)
Continue in this fashion to get $\{ p^\alpha_j\}$, one for each
essential component of $X-C_{j-1}$.
$\{p^\alpha_j\}$ may be assumed to lie in $X-C_j$.
We can also get paths $\lambda_{j,\alpha_i}$ which join
$p^{\alpha_i}_j$ to the appropriate $p^\alpha_{j-1}$ and which
lie in $X-C_{j-1}$.
Now $T$ has $\{ p^\alpha_j \}$ for vertices and
$( p^{\alpha_i}_j, p^{\alpha_\ell}_k)$ is a $1${--}simplex \iff\ $k=j-1$
and $\lambda_{j,\alpha_i}$ joins $p^{\alpha_i}_j$ to $p^{\alpha_\ell}_k$.
The map $f\colon T\ \to\ X$ is the obvious one.
We claim $H_1(T;\Z)=0$, and in fact, if $H_1$ is computed from the
simplicial chains then there are no $1${--}cycles.
This is fairly clear , so it will be left to the reader.
Now any locally finite $1${--}complex with $H_1(T;\Z)=0$ satisfies
$\Delta(T\Colon \pi_k)=0$ for $k>0$.
One shows $f$ is properly $1/2${--}connected by showing that
$Z^0_{\text{end}}(X)\ \to\ Z^0_{\text{end}}(T)$ is an isomorphism
($Z^0_{\text{end}}$ are the $0${--}cycles in $S^0_{\text{end}}$).
But this follows from our construction.
Lastly suppose $T^\prime\subseteq T$ is a connected subcomplex,
and suppose $p\in T- T^\prime$.
Now by definition $p$ is in an essential component of $X-C_i$ for
all $i\leq n$ for some $n$.
Since each essential component of $X-C_i$ has infinitely many base points
in it, let $\{q\}$ be the set of base points in the component of $X-C_n$
containing $p$.
Then $\{q\}\subseteq T-T^\prime$, as is easily seen.
Hence $H^0_{\text{end}}(T)\ \to\ H^0_{\text{end}}(T^\prime)$ has a kernel,
and so $T^\prime\subseteq T$ is not a proper homotopy equivalence.
Hence $X$ has a tree.
From now on in this section we restrict ourselves to the category of
homogamous CW complexes.
We will use \hCWx\ complex to denote objects in this category.
Given $X$, an \hCWx\ complex, we have the category ${\mathcal C}(X)$ whose
objects are all sets $A\subseteq X$ such that
\begin{enumerate}
\item[1)] $A$ is a subcomplex
\item[2)] $A$ is connected
\item[3)] There exists an element of ${\mathcal O}(X)$, $U$, such that
$A$ is an essential component of $X-U$.
\end{enumerate}
\noindent The morphisms are inclusions.
Now given a tree $(T,f)$ for $X$, we get a functor
${\mathcal C}(X)\ \RA{{\mathcal C}(f)}\ {\mathcal C}(T)$ ($f$ is always assumed to be
cellular).
\begin{xDefinition}
A \emph{lift of} ${\mathcal C}(f)$ is a covariant functor
$F\colon {\mathcal C}(T)\ \to\ {\mathcal C}(X)$ such that
${\mathcal C}(f)\ \circ F$ is the identity and such that the image of $F$
is cofinal.
The set of all such lifts is a diagram scheme by defining $F\leq G$
\iff\ $F(A)\subseteq G(A)$ for all $A\in{\mathcal C}(T)$.
We denote this diagram scheme by ${\mathcal L}(f)$.
\end{xDefinition}
\begin{xDefinition}
A \emph{tree of rings} is a covariant functor
$R\colon{\mathcal C}(T)\ \to\ {\mathcal R}$, where ${\mathcal R}$ is the category of
all rings (rings have units and all ring homomorphisms preserve units).
A \emph{tree of modules} over $R$ is a collection of modules $M_A$,
$A\in{\mathcal C}(T)$, where $M_A$ is a unitary $R_A${--}module.
A tree of right (left) $R${--}modules requires each $M_A$ to be a right
(left) $R_A${--}module.
If $A\subseteq B$ in ${\mathcal C}(T)$, there is a unique map
$p_{AB}\colon M_A\ \to\ M_B$, which is an
$R(A\subseteq B)${--}linear map; i.e. if $f\colon R_A\ \to\ R_B$ is the
ring homomorphism associated to $A\subseteq B$ by $R$,
\[p_{AB}(a\cdot\alpha + b\cdot\beta)=
p_{AB}(a)\cdot f(\alpha) + p_{AB}(b)\cdot f(\beta)\]
for $\alpha$, $\beta\in R_A$; $a$, $b\in M_A$.
\end{xDefinition}
An $R${--}module homomorphism $f\colon M\ \to\ M^\prime$
is a set of maps $f_A\colon M_A\ \to\ M^\prime_A$
for each $A\in{\mathcal C}(T)$ such that
\begin{enumerate}
\item[1)] $f_A$ is an $R_A${--}module homomorphism
\vskip 4pt
\item[2)] For $A\subseteq B$,
\lower 12pt\vtop {\hsize=1.75in$\displaystyle\begin{matrix}%
M_A&\RA{\ f_A\ }&M^\prime_A\\
\downlabeledarrow[\big]{p_{AB}}{}&&\downlabeledarrow[\big]{p^\prime_{AB}}{}\\
M_B&\RA{\ f_B\ }&M^\prime_B\\\noalign{\vskip6pt}\end{matrix}$\vss}
commutes, where the vertical maps come from the tree structures
on $M$ and $M^\prime$.
\end{enumerate}
\medskip
\begin{xExample}
Given an \hCWx\ complex $X$ with a tree $(T,f)$ and given $F\in{\mathcal L}(f)$,
we get a tree of rings from $R_A=\Z\bigl[\pi_1\bigl( F(A),f(p)\bigr)\bigr]$
where if $A\neq T$, $p$ is the vertex of $\partial A$, the set
theoretic frontier of $A$.
If $A=T$ pick a vertex for a base point and use it.
This will be the tree of rings we will consider for our geometry,
and we will denote it by $\Z\pi_1$.
The tree of $\Z\pi_1${--}modules we will consider will be
various chain modules.
The basic idea is given by
$M_A= H_i\bigl( \coverFA{F(A)^i\ }, \coverFA{F(A)^{i-1}\ }, f(p)\bigr)$,
where \ $\coverFA{}$\quad denotes the universal cover of $F(A)$,
and $\coverFA{F(A)^i}$ is $\pi^{-1}$ of the $i${--}skeleton of $F(A)$
in $\coverFA{F(A)}$. ($\pi\colon \coverFA{F(A)}\ \to\ F(A)$).
Now given an $R${--}module $M$, we can form $\Delta(M)$ by applying
the $\Delta${--}construction with index set the vertices of $T$, and
with diagram scheme ${\mathcal O}(X)$.
Given $U\in{\mathcal O}(X)$, there are finitely many $A\in{\mathcal C}(T)$ for
which $A\ \cap\ \Bar U = {\text{a\ vertex}}$.
Set \[G_{p U}=\begin{cases}M_A& \text{if }p\in A\\0&\text{otherwise}\end{cases}\]
for some $A$ such that $A\ \cap\ \Bar U = {\text{a\ vertex}}$.
An $R${--}module homomorphism $f\colon M\ \to\ M^\prime$
clearly induces a map $\Delta(f)\colon \Delta(M)\ \to\ \Delta(M^\prime)$.
An $R${--}module homomorphism, $f$, which induces an isomorphism
$\Delta(f)$ is said to be a \emph{strong equivalence} and
the two modules
are said to be \emph{strongly equivalent}.
Note that this relation on $R${--}modules seems not to be symmetric.
Nevertheless we can define two $R${--}modules to be \emph{equivalent}
\iff\ there is a (finite) sequence of $R${--}modules $M=M_0$, $M_1$, \dots,
$M_n=M^\prime$ such that either $M_{i}$ is strongly equivalent to
$M_{i+1}$ or $M_{i+1}$ is strongly equivalent to $M_{i}$.
We tend only to be really interested in the equivalence class of $M$
(indeed, we are often interested merely in $\Delta(M)$ ).
The relation of equivalence is not however very nice.
We would like $M$ to be equivalent to $M^\prime$ \iff\ there were
``maps'' $f\colon M\ \to\ M^\prime$ and $g\colon M^\prime\ \to\ M$
whose composites were the identity.
To do this properly we need a short digression.
\end{xExample}
\begin{xDefinition}
A functor $F$ which assigns to each $A\in{\mathcal C}(T)$
a cofinal subcomplex of $A$, $F(A)$, such that $F(A)\subseteq F(B)$
whenever $A\subseteq B$ and such that $F(T)=T$ will be called a
\emph{shift functor}.
${\mathcal S}(T)$ will denote the set of all shift functors on $T$.
${\mathcal S}(T)$ is partially ordered via $F\geq G$ \iff\ $F(A)\subseteq G(A)$
for all $A\in{\mathcal C}(T)$.
Define $(F\cap G)(A)=F(A)\ \cap\ G(A)$, and one checks it is
a shift functor.
$F\cap G\geq F$ and $F\cap G\geq G$.
\end{xDefinition}
Given a tree of $R${--}modules and a shift functor $F$,
we get a tree of $R${--}modules, $M_F$, in a natural way; i.e.
$F$ is going to induce a functor from the category of $R${--}modules
to itself.
$M_F$ is defined as follows.
Let $A\in{\mathcal C}(T)$.
Then $F(A)=\displaystyle\mathop{\cup}_{i=1}^n A_i$,
with $A_i\in{\mathcal C}(T)$.
$\displaystyle(M_F)_A=\mathop{\oplus}_{i=1}^n\ M_{A_i}\otimes R_A$,
where the tensor product is formed using the homomorphisms
$R_{A_i}\ \to\ R_A$.
Note that there is an $R_A${--}module map
$(M_F)_A\ \to\ M_A$.
$(p_F)_{A B}\colon \displaystyle\mathop{\oplus}_{i=1}^n\ M_{A_i}\otimes R_A
\ \to\ \mathop{\oplus}_{i=1}^n\ M_{B_i}\otimes R_B$
is defined as follows.
Since $A\subseteq B$, $F(A)\subseteq F(B)$, so each $A_i$ is
contained in a unique $B_j$.
Let $p_{i j}$ be $p_{A_i B_j}$ if $A_i\subseteq B_j$ and $0$ otherwise.
$f_{i j}$ is the map $R_{A_i}\ \to\ R_{B_j}$ if $A_i\subseteq B_j$
and $0$ otherwise.
$g$ is the map $R_A\ \to\ R_B$.
Then $(p_F)_{A B}=\displaystyle\mathop{\oplus}_{i=1}^n\ \mathop{\oplus}_{j=1}^m\
p_{i j}\otimes f_{i j}\otimes g$.
Notice that
\lower 16pt\hbox{$\displaystyle\begin{matrix}%
(M_F)_A&\RA{\hskip 40pt}& M_A\\
\downlabeledarrow[\Big]{(p_F)_{A B}}{}&&\downlabeledarrow[\Big]{p_{A B}}{}\\
(M_F)_B&\RA{\hskip 40pt}& M_B\\\end{matrix}$\hskip 20pt}
commutes.
If $f\colon M\ \to\ M^\prime$ is a map,
$\displaystyle(f_F)_A=\mathop{\oplus}_{i=1}^n f_{A_i}\otimes g_{A_i A}$,
where $g_{A_i A}\colon R_{A_i}\ \to\ R_A$, defines a map
$f_F\colon M_F\ \to\ M^\prime_F$ so that
\lower 12pt\hbox{%
$\displaystyle\begin{matrix}%
M_F&\RA{f_F}&M^\prime_F\\
\big\downarrow&&\big\downarrow\\
M&\RA{\ f\ }&M^\prime\\\end{matrix}$\hskip10pt} commutes.
For the natural map of $M_F$ into $M$ we write $M_F\subseteq M$.
If $G\geq F$ there is a natural map $M_G\ \to\ M_F$ induced by
the inclusion of each component of $G(A)$ in $F(A)$.
\medskip
\BEGIN{L.1.5.1}
$M_F\subseteq M$ is a strong equivalence.
\end{Lemma}
\begin{proof}
We must show $\Delta(M_F)\ \to\ \Delta(M)$ is an isomorphism.
Suppose $B\in{\mathcal C}(T)$ and $B\subseteq F(A)$.
Then
\lower 16pt \hbox{$\begin{matrix}%
(M_F)_B&\RA{\hskip10pt}& M_B \\
\big\downarrow&&\big\downarrow\\
(M_F)_A&\RA{\hskip10pt}& M_A\\\end{matrix}$}
commutes and there is a map
$h\colon M_B\ \to\ (M_F)_A$
so that the resulting triangles commute.
But then clearly $\Delta(M_F)\cong\Delta(M)$.
\end{proof}
\medskip
As motivation for our next definition we prove
\BEGIN{L.1.5.2}
Let $f\colon M\ \to\ N$ be a strong equivalence.
Then there is a shift functor $F$ and a map $N_F\ \to\ M$ such that
\lower 12pt\hbox{$\begin{matrix}&&M\\
&\nearrow&\downarrow\\
N_F&\subseteq&N\\\end{matrix}$}
commutes.
\end{Lemma}
\begin{proof}
By \fullRef{T.1.2.4} applied to kernel and cokernel,
$f$ is a strong equivalence \iff\ for any $A\in{\mathcal C}(T)$ there is a
$U\in{\mathcal O}(T)$ such that for any $B\in{\mathcal C}(T)$ with
$B\subseteq A-U$
\[\begin{matrix}%
M_B&\RA{\ f_B\ }&N_B\\
\downlabeledarrow[\Big]{p^M_{A B}}{}&&\downlabeledarrow[\Big]{p^N_{A B}}{}\\
M_A&\RA{\ f_A\ }&N_A\\\end{matrix}\]
satisfies
\begin{enumerate}
\item[1)] ${\kerx}\ f_B\ \subseteq\ {\kerx}\ p^M_{A B}$ and
\item[2)] ${\Imx}\ p^N_{A B}\ \subseteq\ {\Imx}\ f_A$.
\end{enumerate}
\medskip
For each $A\in{\mathcal C}(T)$, pick such an element in ${\mathcal O}(T)$, $U_A$.
Now let $F(A)= A-\displaystyle\mathop{\cup}_{A\subseteq D} U_D$.
$F$ is easily seen to be a shift functor, and for any $B\in{\mathcal C}(T)$
with $B\subseteq F(A)$, 1) and 2) hold.
Now look at
\lower 25pt\hbox{$\begin{matrix}%
M_{A_2}&\RA{\hskip10pt}& N_{A_2}\\
\big\downarrow&&\downlabeledarrow[\big]{p}{}\\
M_{A_1}&\RA{f_{A_1}}& N_{A_1}\\
\downlabeledarrow[\big]{q}{}&&\big\downarrow\\
M_{A}&\RA{\hskip10pt}& N_{A}\\\end{matrix}$}
where $A_1\subset F(A)$, $A_2\subset F(A_1)$.
Then there exists a map $h\colon N_{A_2}\ \to\ M_A$
defined by $h(x)= q(f_{A_1})^{-1}p(x)$ for all $x\in N_{A_2}$.
By properties 1) and 2), $h$ is well{--}defined, and if
$g\colon R_{A_2}\ \to\ R_A$ is the homomorphism given by the tree,
$h$ is easily seen to be $g${--}linear.
Define a shift functor $F\ \circ\ G$ by $F\ \circ\ G(A)=
\displaystyle\mathop{\cup}_{i=1}^n F(A_i)$,
where $G(A)=\displaystyle\mathop{\cup}_{i=1}^n A_i$.
Then one checks that the $h$ defined above yields a map
$N_{F\circ G}\ \to\ M$.
\end{proof}
\bigskip
\begin{xDefinition}
A $T${--}map $f\colon M\ \to\ N$ is a map $M_F\ \to\ N$,
where $F\in{\mathcal S}(T)$.
$M_F\ \to\ N$ induces a natural map $M_G\ \to\ N$ for all $G\geq F$.
We say $f$ is defined on $M_G$ for all $G\geq F$.
Two $T${--}maps $f$, $g\colon M\ \to\ N$ are equal provided that,
for some $F\in{\mathcal S}(T)$ such that $f$ and $g$ are defined on
$M_F$, the two maps $M_F\ \to\ N$ are equal.
\end{xDefinition}
\begin{xRemarks}
If $f$ is defined on $M_F$, and if $g$ is defined on $M_G$, $f$ and
$g$ are both defined on $M_{F\cap G}$.
With this remark it is easy to see equality of $T${--}maps is an
equivalence relation.
It is also easy to see how to add or subtract two $T${--}maps,
and it is easy to check that if $f_1=f_2$ and $g_1=g_2$, then
$f_1\pm g_1 = f_2\pm g_2$.
Hence, if ${\Homx}_T(M,N)$ is the set of equivalence classes of
$T${--}maps from $M$ to $N$, ${\Homx}_T(M,N)$ has the
structure of an abelian group.
An equivalence class of $T${--}maps is called a \emph{map{--}germ}.
We can compose two $T${--}maps $f\colon M\ \to\ N$ and
$g\colon N\ \to\ P$ as follows.
$g$ is defined on $N_G$ and $f$ is defined on $M_F$.
Hence $f\colon N_F\ \to\ P$ is an actual map, and we define the
$T${--}map $g\ \circ\ f$ to be the map
$g\ \circ\ f_G\colon (M_F)_G\ \to\ N_G\ \to\ P$.
Note $(M_F)_G=M_{F\circ G}$.
One can check that the map{--}germ $g\ \circ\ f$ is well{--}defined.
\end{xRemarks}
\medskip
Hence \fullRef{L.1.5.2} becomes
\BEGIN{L.1.5.3}
$M$ and $N$ are equivalent \iff\ they are $T${--}equivalent.
\end{Lemma}
\begin{proof}
If $M$ and $N$ are equivalent, \fullRef{L.1.5.2}
shows how to get $T${--}maps $M\ \to\ N$ and $N\ \to\ M$ using
the sequence of strong equivalences.
If $M$ and $N$ are $T${--}equivalent, we have map $T${--}maps
$f\colon M\ \to\ N$ and $g\colon N\ \to\ M$ such that
$f\ \circ\ g ={\text{id}}_N$ and $g\ \circ\ f = {\text{id}}_M$.
Now a $T${--}map $f\colon M\ \to\ N$ induces a unique map
$\Delta(f)\colon \Delta(M)\ \to\ \Delta(N)$ via
$\Delta(f)=\Delta(f)\ \circ\ \Delta({\text{inc}})^{-1}$
where $f$ is defined on $M_F$ and ${\text{inc}}\colon M_F\subseteq M$.
It is clear that $\Delta(f)$ depends only on the map{--}germ of $f$.
Hence in our case, $g$ induces an equivalence of $M$ and $N$ by
$N\supseteq N_G\ \RA{\ g\ }\ M$.
\end{proof}
\medskip
Also useful is
\medskip
\BEGIN{L.1.5.4}
Let $f$ and $g$ be $T${--}maps.
Then $f=g$ \iff\ $\Delta(f)=\Delta(g)$.
\end{Lemma}
\begin{proof}
$f=g$ \iff\ $f-g=0$.
$\Delta(f-g)=\Delta(f)-\Delta(g)$
Thus we need only show $h=0$ \iff\ $\Delta(h)=0$.
Since $\Delta(h)$ depends only on the map{--}germ, and since
$\Delta(0)=0$, one way is easy.
So assume we are given a $T${--}map $h\colon M\ \to\ N$ with
$\Delta(h)=0$.
We may as well assume that $h$ is an actual map, since otherwise
set $M=M_H$ and proceed.
We have a submodule ${\kerx}\ h\subseteq M$ defined in
the obvious way.
Since ${\kerx}\ h\subseteq M$ is a strong
equivalence, \fullRef{L.1.5.2} says we
can find $F$ such that $M_F\ \to\ {\kerx}\ h\subseteq M$.
But then $M_F\ \to\ N$ is the zero map.
\end{proof}
\bigskip
\begin{xDefinition}
If $R$ is a tree of rings, let ${\mathcal M}_R$ be the category of trees of
$R${--}modules and germs of maps.
Let ${\mathcal M}_{\Delta(R)}$ be the category of $\Delta(R)${--}modules.
\end{xDefinition}
\bigskip
\BEGIN{P.1.5.1}
${\mathcal M}_R$ is an abelian category.
The natural functor \[\Delta\colon {\mathcal M}_R\ \to\ {\mathcal M}_{\Delta(R)}\]
is an exact, additive, faithful functor.
\end{Proposition}
\begin{proof}
The functor just takes $M$ to $\Delta(M)$ and $[f]$ to $\Delta(f)$
($[f]$ denotes the map{--}germ of $f$).
$\Delta$ is additive more or less by definition, and faithful by
\fullRef{L.1.5.4}.
$\Delta$ preserves kernels: Let $M\ \RA{\ [g]\ }\ N$ be a map{--}germ
in ${\mathcal M}_R$.
We can find $G$ such that $M_g\ \RA{\ g\ }\ N$ is a representative.
Clearly any kernel for $[g]$ is equivalent to
${\kerx}\ g\subseteq M_G$, where ${\kerx}\ g$
is the obvious submodule.
But $\Delta({\kerx}\ g)$ is clearly a kernel for $\Delta(g)$.
An entirely similar argument shows $\Delta$ preserves cokernels,
so $\Delta$ is exact.
To see ${\mathcal M}_R$ is normal and conormal, take representatives for
the germs and construct the quotient or the kernel module.
${\mathcal M}_R$ has pullback and pushouts, again by finding representatives
for the germs and constructing the desired modules.
Now by \cite{btwentyfive}, Theorem 20.1 (c), page 33, ${\mathcal M}_R$
is abelian.
\end{proof}
We want to do stable algebra, and for this we need an analogue of
finitely{--}generated projective.
Projective is easy, we just insist that a projective $R${--}module
is projective in the category ${\mathcal M}_R$ (see \cite{btwentyfive},
pages 69{--}71 for definitions and elementary properties).
For the analogue of finitely{--}generated, we first produce the analogue
of a finitely{--}generated, free module.
\begin{xDefinition}
Let $T$ be a tree and let $S$ be a set.
A \emph{partition of} $S$ is a functor $\pi\colon {\mathcal C}(T)\ \to\ 2^S$
(where $2^S$ is the category of subsets of $S$ and inclusion maps)
satisfying
\begin{enumerate}
\item[1)] $\pi(T)=S$.
\item[2)] If $A\ \cap\ B=\emptyset$, $\pi(A)\ \cap\ \pi(B)=\emptyset$
( $A$, $B\in{\mathcal C}(T)$).
\item3)] Let $A_i\in{\mathcal C}(T)$, $i=1$, \dots, $n$.
If $\displaystyle T-\mathop{\cup}_{i=1}^n A_i$ is compact,
$\displaystyle \pi(T)-\mathop{\cup}_{i=1}^n \pi(A_i)$ is finite.
\item[4)] Let $s\in S$.
Then there exist $A_i\in{\mathcal C}(T)$, $i=1$, \dots, $n$ such that
$\displaystyle T - \mathop{\cup}_{i=1}^n A_i$ is compact and
$s\notin \pi(A_i)$ for any $i=1$, \dots, $n$.
\end{enumerate}
\end{xDefinition}
\bigskip
\begin{xDefinition}
Let $R$ be a tree of rings over $T$.
Let $\pi$ be a partition of $S$.
The \emph{free $R${--}module based on $\pi$}, $F_\pi$, is the tree
of $R${--}modules defined by $(F_\pi)_A$ is the free $R_A${--}module
based on $\pi(A)$, and if $A\subseteq B$,
$p_{A B}\colon (F_\pi)_A\ \to\ (F_\pi)_B$ is induced by the inclusion
$\pi(A)\subseteq\pi(B)$.
\end{xDefinition}
\begin{xDefinition}
A tree of $R${--}modules, $M$, is said to be locally{--}finitely generated
\iff\ there is a set of generators, $S$, and a partition $\pi$, of $S$,
such that there is an epimorphism $F_\pi\ \to\ M$.
\end{xDefinition}
Let us briefly discuss partitions.
If $\pi$ and $\rho$ are two partitions of a set $S$, we say
$\pi\subseteq\rho$ \iff\ $\pi(A)\subseteq\rho(A)$ for all $A\in{\mathcal C}(T)$.
(Hence we could talk about the category of partitions, but we shall
largely refrain.)
Two partitions are \emph{equivalent} \iff\ there exist a finite sequence
$\pi=\pi_0$, $\pi_1$, \dots, $\pi_n=\rho$ of partitions with
$\pi_{i}\subseteq\pi_{i+1}$, or $\pi_{i+1}\subseteq\pi_{i}$.
(This is clearly an equivalence relation.)
Given two sets $X$ and $Y$, and partitions $\pi$ and $\rho$,
$\pi\cup\rho$ is the partition $X\cup Y$ given by $(\pi\cup\rho)(A)=
\pi(A)\cup\rho(A)$.
\bigskip
\BEGIN{L.1.5.5}
Let $R$ be a tree of rings over $T$, and let $X$ and $Y$ be sets.
Then if $\pi$ and $\pi^\prime$ are equivalent partitions of $X$,
$F_\pi$ is isomorphic to $F_{\pi^\prime}$ in ${\mathcal M}_R$.
If $\rho$ is a partition of $Y$,
$F_{\pi\cup\rho}=F_\pi\oplus F_\rho$ ($X$ and $Y$ disjoint).
\end{Lemma}
\begin{proof}
To show the first statement we need only show it for
$\pi\subseteq \pi^\prime$.
In this case there is a natural map $f\colon F_\pi\ \to\ F_{\pi^\prime}$.
For each $A\in{\mathcal C}(T)$, $(F_\pi)_A\ \to\ (F_{\pi^\prime})_A$ is
injective, so $f$ is a monomorphism.
If $\pi\subseteq \pi^\prime$, then $\pi^\prime(A)-\pi(A)$ has only
finitely many elements.
To see this, observe we can find $A_i\in{\mathcal C}(T)$, $i=1$, \dots, $n$
such that $A\ \cap\ A_i=\emptyset$, and
$\displaystyle T - \mathop{\cup}_{i=1}^n A_i - A$ is compact.
Then by 2) $\displaystyle\pi^\prime(A)\subseteq \pi^\prime(T)-
\mathop{\cup}_{i=1}^n \pi^\prime(A_i)$, so
$\displaystyle\pi^\prime(A)-\pi(A)\subseteq \pi^\prime(T)-
\mathop{\cup}_{i=1}^n \pi^\prime(A_i)-\pi(A)\subseteq
\pi(T)- \mathop{\cup}_{i=1}^n \pi(A_i)-\pi(A)$, which is finite.
Since $\pi^\prime(A)-\pi(A)$ is finite,
$f_A\colon (F_\pi)_A\ \to (F_{\pi^\prime})_A$ has finitely generated
cokernel, so when the $\Delta${--}construction is applied to it,
4) guarantees that $\Delta(f)$ is onto, so $f$ is an equivalence.
The second statement is the definition of $\pi\cup\rho$ and
$F_\pi\oplus F_\rho$.
\end{proof}
It is not hard to see that if we have a partition of $S$ for the tree $T$,
then $S$ has at most countably many elements if $T$ is infinite, and
at most finitely many if $T$ is a point.
In the case $S$ is infinite, we have a very handy countable infinite
set lying around, namely the vertices of $T$.
There is an obvious partition, $\pi$, where $\pi(A)=\{ p\ \vert\
p{\text{\ is\ a\ vertex\ of\ }}A\ \}$.
Denote $F_\pi$ by $F^{(1)}$.
If $T$ is a point, let $F^{(1)}$ denote the free module on one
generator; i.e. still $F_\pi$ for the above partition $\pi$.
$F^{(n)}= F^{(n-1)}\oplus F^{(1)}$ for $n\geq 2$.
\bigskip
\BEGIN{L.1.5.6}
Let $\pi$ be any partition of a set $S$ for the tree $T$, and let $R$
be a tree of rings.
Then $F_\pi\oplus F^{(1)}$ is equivalent to $F^{(n)}$ for some
$n\geq 1$.
If $T$ is infinite, $n$ can be chosen to be $1$.
\end{Lemma}
\begin{proof}
If $T$ is a point, this is obvious, so assume $T$ is infinite.
$F_\pi\oplus F^{(1)}$ is just $F_{\pi\cup\rho}$, where $\rho$
is the standard partition on $V$, the vertices of $T$.
Since $V\ \cup\ S$ is infinite (and countable), there is a $1${--}$1$
correspondence $\alpha\colon V\ \cup\ S\ \to\ V$.
Any such $\alpha$ induces an equivalence of categories
$\alpha\colon 2^{V\cup S}\ \to\ 2^V$.
We show that we can pick $\alpha$ so that
$\alpha\ \circ\ (\pi\ \cup\ \rho)$ is equivalent to $\rho$.
(We will show in \fullRef{L.1.5.7} that
$\alpha\ \circ\ (\pi\ \cup\ \rho)$ is a partition for any $\alpha$.)
Our $\alpha$ is defined by picking a strictly increasing sequence
of finite subcomplexes, $C_0\subseteq C_1\subseteq\cdots$, so
that $\displaystyle\mathop{\cup}_{i=0}^\infty C_i=T$.
Let $A_k(i)$ be the essential components of $T-C_i$.
Set $A_1(-1)=T$, and let $\displaystyle K_{k i}=
(\pi\cup\rho)\bigl(A_k(i)\bigr) -\mathop{\cup}_\ell\ (\pi\cup\rho)\bigl(
A_\ell(i+1)\bigr)$.
Note $K_{k i}\ \cap\ K_{k^\prime i}=\emptyset$ and
$K_{k i}\ \cap\ K_{k^\prime i+1}=\emptyset$ by 2), so
$K_{k i}\ \cap\ K_{\ell j}\neq \emptyset$ \iff\ $k=\ell$ and $i=j$.
Now $K_{k i}$ is finite.
We define $\alpha$ on $K_{k i}$ by induction on $i$.
Let $\displaystyle L_{k i}=\rho\bigl(A_k(i)\bigr) -\mathop{\cup}_\ell
\rho\bigl(A_\ell(i+1)\bigr)$, and note that the cardinality of $K_{k i}$
is greater than or equal to the cardinality of $L_{k i}$.
Define $\alpha$ on $K_{1\ -1}$ by mapping some subset of it to
$L_{1\ -1}$ and mapping any left over elements to any elements of $V$
($\alpha$ should be injective).
Suppose $\alpha$ defined on $k_{k\>i-1}$ so that $\alpha(K_{k j})
\subseteq\rho\bigl(A_k(j)\bigr)$ for $j\leq i-1$.
We need only define $\alpha$ on $K_{k i}$ so that $\alpha(K_{k i})
\subseteq\rho\bigl(A_k(i)\bigr)$ to be done.
Look at
$\displaystyle M = L_{k i} -
\mathop{\cup}_{\Atop{\scriptscriptstyle{\text{all\ }}\ell}{j\leq i-1}}
\ {\Imx}\ \alpha(K_{\ell j})$.
Map some subset of $K_{k i}$ to $M$.
Map the rest of $K_{k i}$ to any elements of $\rho\bigl(A_k(i)\bigr)$ at all.
By 4), $\displaystyle V\cup S=
\mathop{\cup}_{\Atop{\scriptscriptstyle{\text{all\ }}k}{{\text{all\ }}i}} K_{k i}$
and
$\displaystyle S=
\mathop{\cup}_{\Atop{\scriptscriptstyle{\text{all\ }}k}{{\text{all\ }}i}} L_{k i}$
(as disjoint unions as we saw).
Since $\alpha$ is onto each $L_{k i}$, and since it injects when restricted to
each $K_{k i}$, $\alpha$ is $1${--}$1$.
Furthermore, $\tau=\alpha\ \circ\ (\pi\cup\rho)$ satisfies
$\tau\bigl(A_k(i)\bigr)\subseteq \rho\bigl(A_k(i)\bigr)$ by construction.
Set $\lambda(A)=\tau(A)\cap\rho(A)$.
We claim $\lambda$ is a partition.
Clearly $\lambda$ is a functor ${\mathcal C}(T)\ \to\ 2^V$.
1) and 2) are trivial and 4) is not much harder (\ 1), 2) and 4) hold for
the intersection of any two partitions, it is only 3) which might fail).
To show 3), note $\lambda\bigl(A_k(i)\bigr)=\tau\bigl(A_k(i)\bigr)$.
If $\displaystyle T-\mathop{\cup}_{j=1}^n B_j$ is compact, there is
a minimal $i$ such that $B_j$ contains $A_k(i)$ for some $k$
(perhaps several, say $k=1$, \dots, $m$).
Then
$\displaystyle \mathop{\cup}_{k=1}^m \lambda\bigl(A_k(i)\bigr)\subseteq
\lambda(B_j)$.
$\displaystyle\lambda(T)-\mathop{\cup}_{j=1}^n\lambda(B_j)\subseteq
\lambda(T)-\cup\ \lambda\bigl(A_k(i)\bigr)=
\tau(T)-\cup\ \tau\bigl(A_k(i)\bigr)$.
The last two unions are over all the $A_k(i)\subseteq B_j$ for
$j=1$, \dots, $n$.
This last set is finite, so 3) holds.
Hence $\lambda$ is a partition and thus $\tau$ is equivalent to $\rho$.
The map from $F_{\pi\cup\rho}\ \to F_\tau$ induced by $\alpha$
is the obvious map: $(F_{\pi\cup\rho})_A\ \to\ (F_\tau)_A$ is the
isomorphism induced by the equivalence of bases $\alpha\colon
(\pi\cup\rho)(A)\leftrightarrow\tau(A)$.
\fullRef{L.1.5.5} completes the proof modulo
the proof of \fullRef{L.1.5.7}.
\end{proof}
\bigski
\BEGIN{L.1.5.7}
Let $X$ and $Y$ be two (disjoint) sets, and let $\pi$ be a partition
of $X$ for the tree $T$.
Any $1${--}$1$ correspondence $\alpha\colon X\ \to\ Y$ induces a
partition $\alpha\ \circ\ \pi$ of $Y$ for the tree $T$.
\end{Lemma}
\begin{proof} The easy proof is omitted.\end{proof}
\bigski
\BEGIN{L.1.5.8}
$F_\pi$ is projective.
\end{Lemma}
\begin{proof}
By \fullRef{L.1.5.6} and standard nonsense, it is
enough to prove the result for $F^{(1)}$.
By Mitchell \cite{btwentyfive} Proposition 14.2, page 70, we need only show
$M\ \RA{\ [f]\ }\ F^{(1)}$ splits whenever $[f]$ is an epimorphism
(note ${\mathcal M}_R$ is abelian by \fullRef{P.1.5.1}
so we may apply Mitchell).
By taking a representative for $[f]$, we may as well assume that we
have a map $f\colon M\ \to\ F=F^{(1)}$ which is an epimorphism.
Now there is a partition $\pi$ with $\pi\subseteq \rho$
($\rho$ the standard partition for $F^{(1)}$), such that the
inclusion of $(F_\pi)_A$ in $F_A$ lies in the image of $M_A$
under $f_A$; i.e. define $\pi(A)=\{ x\in\rho(A)\ \vert\
x\in{\Imx}\ f_A\ \}$.
Since $f$ is an epimorphism, one can easily check $\rho(A)-\pi(A)$
is finite, and from this result one easily deduces $\pi$ is a partition.
Now pick a base point $\ast\in T$.
This choice immediately partially orders all the vertices of $T$ by saying
$p\geq q$ provided the minimal path from $p$ to $\ast$ hits $q$.
$A_p\in{\mathcal C}(T)$ for each $p$ a vertex of $T$, $p\neq\ast$, is defined
as the unique $A\in{\mathcal C}(T)$ such that $q\in A_p$ implies $q\geq p$.
\IpdfLRT{diagram151}{2.5in}{2in}
\vskip -60pt\noindent Given a partition $\pi$, define a new partition $\tau$ by
$\displaystyle\tau(A) = \mathop{\cup}_{A_p\subseteq A}\ \pi(A_p)$
(again $\pi(A)-\tau(A)$ is finite, $\tau(A)\subseteq \pi(A)$, so one
can check $\tau$ is a partition).
Since $\tau\subseteq\pi$, $(F_\tau)_A\subseteq F_A$ lies in
${\Imx\ }(f_A)$.
Now given any vertex $v$ of $T$, there is a unique $p$ such that
$v\in\tau(A_p)$ and $v\in\tau(A)$ \iff\ $A_p\subseteq A$, unless
$v\notin \tau(A_p)$ for any $A_p$ (there are only finitely many of
the latter).
To see this, set $A=\displaystyle\mathop{\cap}_{v\in\tau(A_p)} A_p$.
Now $A_p\ \cap\ A_q\neq\emptyset$ implies $A_p\subseteq A_q$
(or $A_q\subseteq A_p$).
By 4) the intersection runs over finitely many objects, so $A=A_p$ for
some $p$.
This $A_p$ has the properties we claimed.
Define $x_v\in M_{A_p}$ to be any element such that $f_{A_p}(x_v)$
hits the image of the generator in $(F_\pi)_{A_p}$ corresponding to $v$.
Define $h\colon F_\tau\ \to\ M$ by $h_A\colon (F_\tau)_A\ \to\ M_A$
takes the generator corresponding to $v$ to $p_{A_p\>A}(x_v)$.
We extend linearly.
Notice that if the generator corresponding to $v$ lies in $(F_\tau)_A$,
$A_p\subseteq A$, so $p_{A_p\>A}$ makes sense.
It is not hard to check that the $h_A$ induce a map
$h\colon F_\tau\ \to\ M$, and $f\ \circ\ h\colon F_\tau\ \to\ F$
is just the inclusion.
\end{proof}
\medskip
If ${\mathcal P}_R$ is the category of locally{--}finitely generated trees of
projective $R${--}modules, we have
\medskip
\BEGIN{L.1.5.9}
Let $0\to\ M\ \to\ N\ \to\ Q\ \to0$ be a short exact sequence
of $R${--}modules.
Then, if $N$, $Q\in{\mathcal P}_R$, $M\in{\mathcal P}_R$.
If $M$, $Q\in{\mathcal P}_R$, $N\in{\mathcal P}_R$.
Lastly, any $P\in{\mathcal P}_R$ is a summand of a locally{--}finitely generated
free module.
\end{Lemma}
\begin{proof} The proof is easy.\end{proof}
\medskip\begin{xRemarks}
${\mathcal P}_R$ is a suitable category in which to do stable algebra (see
Bass \cite{bone}).
${\mathcal P}_R$ has a product, the direct sum. ${\mathcal P}_R$ is also a full
subcategory of ${\mathcal M}_R$, which is abelian by
\fullRef{P.1.5.1}.
Hence we may use either of Bass's definitions of the $K${--}groups.
Note ${\mathcal P}_R$ is semi{--}simple (Bass \cite{bone}) so
the two definitions agree.
\end{xRemarks}
\insetitem{Notation}
$K_0(R)=K_0({\mathcal P}_R)$ and $K_1(R)=K_1({\mathcal P}_R)$ for
$R$ a tree of rings.
\bigskip
Given a map of trees of rings $R\ \to\ S$ ($R_A\ \to\ S_A$
takes units to units) we can define $M\otimes_R S$ for
$M$ a right $R${--}module by taking $(M\otimes_R S)_A=
M_A\otimes_{R_A} S_A$.
$\otimes$ induces a functor ${\mathcal M}_R\ \to\ {\mathcal M}_S$.
The only non{--}trivial part of this is to show $\otimes$ is well{--}defined
on map{--}germs.
But since
\[\begin{matrix}%
{\mathcal M}_R&\rightlabeledarrow{\hskip40pt\otimes_R S \hskip40pt}{}&{\mathcal M}_S\\
\Big\downarrow&&\Big\downarrow\\
{\mathcal M}_{\Delta(R)}&
\rightlabeledarrow{\hskip 40pt\otimes_{\ \Delta(R)} \Delta(S)\hskip40pt}{}&{\mathcal M}_{\Delta(S)}\\
\end{matrix}\]
commutes, this is easy.
$\otimes$ is, as usual, an additive, right exact functor.
Now given a partition $\pi$, $F^R_\pi\otimes_R S= F^S_\pi$, where
$F^R_\pi$ is the free $R${--}module based on $\pi$ ( $F^S_\pi$similarly).
Hence it is easy to see that $\otimes$ takes ${\mathcal P}_R$ to ${\mathcal P}_S$.
$\otimes$ is cofinal in the sense of Bass \cite{bone}, so we get a
relative group $K_0(f)$, where $f\colon R\ \to\ S$ is the map of trees
of rings.
There is an exact sequence
\[K_1(R)\ \to\ K_1(S)\ \to\ K_0(f)\ \to\ K_0(R)\ \to\ K_0(S)\ .\]
We denote by $K_i(T)$, $i=0$, $1$, the result of applying the
$K${--}groups to ${\mathcal P}_T$, where ${\mathcal P}_T$ is the category of
locally{--}finitely generated projective modules
over the tree of rings $``T''$,
where $(``T'')_A=\Z$ for all $A$ and $p_{A B}={\text{id}}$.
There is always a functor ${\mathcal P}_T\ \to\ {\mathcal P}_R$
induced by the unit map $``T''\ \to\ R$.
The relative $K_0$ of this map will be called the reduced $K_1$
of $R$, written $\Bar{K}(R)$.
\medskip\begin{xRemarks}
If the tree of rings is a point the functor
${\mathcal M}_R\ \to\ \Delta({\mathcal M}_R)$ induces a functor
${\mathcal P}_R\ \to\ {\mathcal P}_{\Delta(R)}$, where ${\mathcal P}_{\Delta(R)}$
is the category of finitely{--}generated projective $\Delta(R)${--}modules.
This functor induces an isomorphism on $K_0$ and $K_1$.
For the compact case ($T={\text{pt}}$.),
torsions lie in quotients of $K_1({\mathcal P}_{\Delta(R)})$.
This, together with \fullRef{P.1.5.2} below is
supposed to motivate our choice of ${\mathcal P}_R$ as the category
in which to do stable algebra.
\end{xRemarks}
\medskip\begin{xDefinition}
Let $W$ be an \hCWx\ complex of finite dimension.
Let $X$ and $Y$ be subcomplexes.
Let $(T,f)$ be a tree for $W$.
Lastly let $F\in {\mathcal L}(f)$.
Then $\Z \pi_1(W,F,f)$ is the tree of rings we had earlier as an example.
Pick a locally{--}finite set of paths, $\Lambda$, from the cells of $W$
to the vertices of $f(T)$ (the paths all begin at the barycenter
of each cell).
\end{xDefinition}
\medskip
$C_\ast(W;X,Y\Colon \Lambda,F)$ is the tree of $\Z\pi_1(W,F,f)${--}modules
given at $A$ by
\[H_\ast\left(
\coverFC{F(A)\ }^{\lower5pt\hbox{$\scriptstyle\ast$}}; \coverFC{F(A)\ }^{\lower5pt\hbox{$\scriptstyle\ast-1$}},
\coverFC{F(A)^\ast\cap X\hskip 6pt },
\coverFC{F(A)^\ast\cap Y\hskip 6pt }
\right)
\]
where \ $\coverFC{\hskip10pt}$\quad is the universal cover of $F(A)$, so,
for example, $\coverFC{F(A)^\ast\cap Y\hskip 6pt}$
is the part of the universal
cover of $F(A)$ lying over $Y\ \cap\ ({\text{the\ }}\ast{\text{{-}skeleton\ of\ }}F(A)$.
In each $\coverFC{F(A)}$ pick a base point covering the vertex
$\partial A$.
These choices give us maps
$\coverFC{F(A)}\ \to\ \coverFC{F(B)}$ whenever $A\subseteq B$.
\medskip
$C^\ast(W;X,Y\Colon \Lambda,F)$ is defined from the cohomology groups
\[H^\ast_\cmpsup\left(
\coverFC{F(A)\ }^{\lower5pt\hbox{$\scriptstyle\ast$}}; \coverFC{F(A)\ }^{\lower5pt\hbox{$\scriptstyle\ast-1$}},
\coverFC{\partial F(A)\ }^{\lower5pt\hbox{$\scriptstyle\ast$}},
\coverFC{F(A)^\ast\cap X\hskip 6pt},
\coverFC{F(A)^\ast\cap Y\hskip 6pt}
\right) .
\]
The maps are the ones we defined in section 4.
\bigskip
\BEGIN{P.1.5.2}
$C_\ast(W;X,Y\Colon\,\Lambda,F)$ ($C^\ast(W;X,Y\Colon\,\Lambda,F)$) is a
locally{--}finitely generated, free, right (left) $\Z\pi_1(W,F,f)${--}module.
If $G\in{\mathcal L}(f)$ satisfies $G\geq F$, there is an induced map
$\Z\pi_1(W,F,f)\ \to\ \Z\pi_1(W,G,f)$.
$C_\ast(W;X,Y\Colon\,\Lambda,F)\otimes_{\Z\pi_1(W,F,f)}\Z\pi_1(W,G,f)$
is equivalent to $C_\ast(W;X,Y\Colon\,\Lambda,G)$.
$\Z\pi_1(W,G,f)\otimes_{\Z\pi_1(W,F,f)}C^\ast(W;X,Y\Colon\,\Lambda,F)$
is equivalent to \\$C^\ast(W;X,Y\Colon\,\Lambda,G)$.
The $\Delta${-}functor applied to $C_\ast(W;X,Y\Colon\,\Lambda,F)$
is \\$P_\ast(W;X,Y\Colon\,\coverFC{\hskip10pt})$;
$\Delta\bigl(C^\ast(W;X,Y\Colon\,\Lambda,F)\bigr) =
P^\ast(W;X,Y\Colon\,\coverFC{\hskip10pt})$ (the $P$ were defined in section 4) ).
\end{Proposition}
\begin{proof}
The assertions are all fairly obvious.
Note in passing that the set $S$ for $C_\ast$ ($C^\ast$) is the set of
all $\ast${--}cells in $W-(X\ \cup\ Y)$.
\end{proof}
\medskip
\BEGIN{P.1.5.3}
The choice of paths $\Lambda$ determines a basis for $C_\ast$ ($C^\ast$).
\end{Proposition}
\begin{proof}
Let $S$ be the set of all $\ast${--}cells in $W-(X\ \cup\ Y)$.
Partition $S$ by $\pi(A)=$ the set of all $\ast${--}cells in $W-(X\ \cup\ Y)$
such that the cell and its associated path both lie in $F(A)$.
$\pi$ is seen to be a partition, and $F_\pi$ is equivalent to $C_\ast$.
The path also determines a lift of the cell into $\coverFC{F(A)}$,
so each $(F\pi)_A$ is based.
\end{proof}
\medskip
Apparently our tree of rings and modules is going to depend on the
lift functor we choose.
This is not the case and we proceed to prove this.
Given a shift functor $F$ and a tree of rings $R$, $R_F$ is the tree
of rings given by
$\displaystyle (R_F)_A=\mathop{\oplus}_{i=1}^n\ R_{A_i}$
where the $A_i$ are the essential components of $F(A)$.
$p_{A B}$ is just $\oplus p_{i j}$, where $p_{i j}$ is the projection
$p_{A_i B_j}$ where $A_i\subseteq B_j$.
We now redefine $M_F$.
$M_F$ is going to be an $R_F${--}module.
$\displaystyle (M_F)_A=\mathop{\oplus}_{i=1}^n\ M_{A_i}$
with the obvious $R_F${--}module structure.
Note $M_F\otimes_{R_F} R$ is just our old $M_F$.
Now a $T${--}map of rings is just a map $R_F\ \to\ S$.
As in the case of modules, we can define a map{--}germ between
two rings.
\medski
\BEGIN{L.1.5.10}
The maps $K_i(R_F)\ \to\ K_i(R)$, $i=0$, $1$, are isomorphisms.
\end{Lemma}
\begin{proof}
$M\mapsto M_F$, $f\mapsto f_F$ defines a functor
${\mathcal P}_R\ \to\ {\mathcal P}_{R_F}$.
Using this functor, one checks ${\mathcal P}_{R_F}\ \to\ {\mathcal P}_R$
is an equivalence of categories.
The result is now easy.
\end{proof}
\medskip
Hence given a map{--}germ $f\colon R\to S$, we get well{--}defined
induced maps $K_i(R)\ \to\ K_i(S)$, $i=0$, $1$, and
$\Bar K_1(R)\ \to\ \Bar K_1(S)$.
\medski
\BEGIN{L.1.5.11}
Let $f\colon R\ \to\ S$ be a map such that $\Delta(f)$ is an
isomorphism.
Then there is a shift functor $F$ and a map $g\colon S_F\ \to\ R$
such that
\[\begin{matrix}%
&&R\\
&\hbox to 0pt{\hss$\scriptstyle g$}\nearrow&\downlabeledarrow{}{f}\\
S_F&\to&S
\end{matrix}\]
commutes.
\end{Lemma}
\begin{proof}
The proof is just like that of \fullRef{L.1.5.2}.
\end{proof}
\medski
\BEGIN{L.1.5.12}
Let $[f]\colon R\ \to\ S$ be a map{--}germ such that $\Delta(f)$
is an isomorphism.
Then the maps
$K_0(R)\ \to\ K_0(S)$; $K_1(R)\ \to\ K_1(S)$; and
$\Bar K_1(R)\ \to\ \Bar K_1(S)$ are isomorphisms.
\end{Lemma}
\begin{proof}
This proof is easy and will be left to the reader.
\end{proof}
\medskip\begin{xRemarks}
By \fullRef{L.1.5.12}, the $K${--}groups we get
will not depend on which lift functor we use.
Let $\displaystyle K_i(X\Colon f)=
\lim_{\Atop{\longrightarrow}{F\in{\mathcal L}(f)}}\
K_i\bigl(\Z\pi_1(X,F,f)\bigr)$.
Since all the maps in our direct limit are isomorphisms, $K_i(X\Colon f)$
is computable in terms of $K_i\bigl(\Z\pi_1(X,F,f)\bigr)$ for any $F$.
$\Bar K_1(X\Colon f)$ is defined similarly.
\end{xRemarks}
\bigskip
\begin{xDefinition}
A stably free (s{--}free) tree of $R${--}modules is an element, $P$,
of ${\mathcal P}_R$ such that $[P]$ is in the image of $K_0(T)$.
Let $P$ be an s{--}free\ $R${--}module.
An s{--}basis for $P$ is an element $F\in{\mathcal R}_T$ and an isomorphism
$b\colon F\otimes_T R\ \to\ P\oplus F_1\otimes_T R$, where
$F_1\in{\mathcal P}_T$.
\end{xDefinition}
Two s{--}bases $b\colon F\otimes_T R\ \to\ P\oplus F_1\otimes_T R$
and $c\colon F_2\otimes_T R\ \to\ P\oplus F_3\otimes_T R$ are
\emph{equivalent} ($B\sim c$) \iff\
$0=(F\oplus F_3, (b\oplus {\text{id}}_{F_3})\ \circ\ {\tw\ }\circ\
(c\oplus {\text{id}}_{F_1})^{-1}, F_2\oplus F_1)$ in $\Bar K_1(R)$,
where
${\tw\ }\colon (P\oplus F_1\otimes_T R)\oplus F_3\otimes_T R\ \to\
(P\oplus F_3\otimes_T R)\oplus F_1\otimes_T R$ is the obvious map.
\medskip
We can now give an exposition of torsion following
Milnor \cite{btwentythree}.
Given a short exact sequence
$0\to\ E\ \RA{\ i\ }\ F\ \RA{\ p\ }\ G\ \to0$ and s{--}bases $b$ for $E$
and $c$ for $G$, define an s{--}basis $b c$ for $F$ by picking a
splitting $r\colon G\to F$ for $p$ and then taking the composition
$F_1\oplus F_2\ \RA{\ (b,c)\ }\ (E\oplus F_3)\oplus (G\oplus F_4)\
\RA{\ h\ }\ F\oplus(F_3\oplus F_4)$, where
$h(e,x,g,z)$ goes to $\bigl( i(e)+r(g), x, z\bigr)$.
It is not hard to check that this s{--}basis does not depend on the choice
of splitting map.
We use Milnor's formulation.
Let $F_0\subseteq F_1\subseteq$ \dots $\subseteq F_k$ and suppose
each $F_i/F_{i-1}$ has an s{--}basis $b_i$.
Then $b_1b_2\cdots b_k$ is seen to be well{--}defined;
i.e. our construction is associative.
Let $E$ and $F$ be submodules of $G$.
Then $E+F$ is the submodule of $G$ generated by $E$ and $F$.
$E\ \cap\ F$ is the pullback of
\lower \baselineskip\hbox{$\begin{matrix}%
&&E\\
&&\downarrow\\
F&\to&G
\end{matrix}$}
\smallski
\BEGIN{L.1.5.13}
(Noether) The natural map
$E/(E\cap F)\ \to\ (E+F)/F$ is an isomorphism.
\end{Lemma}
\begin{proof}
Apply the ordinary Noether isomorphism to each term.
\end{proof}
\medskip
Now let $E/(E\cap F)$ have an s{--}basis $b$ and let $F/(E\cap F)$
have an s{--}basis $c$.
Base $(E+F)/F$ by $b$ composed with the Noether map (we will
continue to denote it by $b$).
Similarly base $(E+F)/E$ by $c$.
Then $b c\sim c b$ as s{--}bases for $(E+F)/(E\cap F)$.
\medskip\begin{xDefinition}
Let $b$ and $c$ be two s{--}bases for $P$.
Then $[b/c]\in\Bar K_1(R)$ is defined as follows:
if $F\ \RA{\ b\ }\ P\oplus F_1$; $G\ \RA{\ c\ }\ P\oplus F_2$, then
$[b/c]= (F\oplus F_2, h, G\oplus F_1)$ where
$\vrule height 12pt depth 0pt width 0pt
h\colon F\oplus F_2\ \RA{\ b\oplus{\text{id}}\ }\
(P\oplus F_1)\oplus F_2\ \to\ (P\oplus F_2)\oplus F_1\
\RA{\ c^{-1}\oplus {\text{id}}\ }\ G\oplus F_1
\vrule height 14pt depth 0pt width 0pt$.
Two s{--}bases are equivalent \iff\ $[b/c]=0$.
The formulas
$[b/c]+[c/d]=[b/d]$ and $[b/c]+[d/e]=[b d/c e]$ are easy to derive
from the relations in the relevant $K_1$.
\end{xDefinition}
\setcounter{footnote}{0}
We next define a torsion for chain complexes.
A \emph{free} chain complex is a set of s{--}free modules, $P_n$,
together with map{--}germs $\partial_n\colon P_n\ \to\ P_{n-1}$
such that $\partial_n\ \circ\ \partial_{n-1}=0$\footnote{Probably should have been called s{--}free.}.
A \emph{finite} free chain complex is one with only finitely many non{--}zero
$P_n$.
A \emph{positive} free chain complex has $P_n=0$ for $n<0$.
\medskip\begin{xDefinition}
Let $\{P_n, \partial_n\}$ be a finite free chain complex.
Let $P_n$ be s{--}based by $c_n$, and suppose each homology group
$H_i$ is s{--}free and s{--}based by $h_i$.
\end{xDefinition}
The sequences $0\to\ B_{n+1}\ \to\ Z_n\ \to\ H_n\ \to0$ and
$0\to\ Z_n\ \to\ P_n\ \to\ B_n\ \to0$, where
$B_n={\Imx}\ (P_n\ \to\ P_{n-1})$ and
$Z_n={\kerx} (\partial_n)$, are short exact.
Let $b_n$ be an s{--}basis for $B_n$, which exists by an inductive
argument.
\[\tau(P_\ast)=\sum_{n}(-1)^n\bigl[ b_n h_n b_{n-1}/c_n\bigr]
\in\Bar K_1(R)\ .\]
It is easy to show $\tau(P_\ast)$ does not depend on the choice of $b_n$.
Let $0\to\ P^\prime_\ast\ \to\ P_\ast\ \to\
P^{\prime\prime}_\ast\ \to0$ be a short exact sequence of finite free
chain complexes.
There is a long exact sequence
\[\begin{matrix}H_\ast(P^\prime)&\RA{\hskip20pt}&H_\ast(P)\\
\hskip10pt\hbox to 0pt{$\hbox to 0pt{\hss$\scriptstyle\partial$}\nwarrow$\hss}
&&\hskip-20pt\hbox to0pt{\hss$\swarrow$}\\
&H_\ast(P^{\prime\prime})\\\end{matrix}\]
Suppose each homology group is s{--}based.
Then we have a torsion associated to ${\mathcal H}$, where
\[{\mathcal H}_{3n} = H_n(P^{\prime})\ ,\quad
{\mathcal H}_{3n-1} = H_n(P)\ ,\quad
{\mathcal H}_{3n-2} = H_n(P^{\prime\prime})\ ,\]
since ${\mathcal H}$ is acyclic.
\bigskip
\BEGIN{T.1.5.1}
$\tau(P_\ast)=\tau(P^\prime_\ast)+\tau(P^{\prime\prime}_\ast)+
\tau({\mathcal H})$.
\end{Theorem}
\begin{proof} See Milnor \cite{btwentythree}, Theorems 3.1 and 3.2.
\end{proof}
\medskip
We next describe the algebraic Subdivision Theorem of
Milnor \cite{btwentythree} (Theorem 5.2).
Given a chain complex $C_\ast$, suppose it is filtered by
$C_\ast^{(0)}\subseteq C_\ast^{(1)}\subseteq \cdots
\subseteq C_\ast^{(n)}= C_\ast$ such that the homology groups
$H_i\bigl( C^{(\lambda)}/C^{(\lambda-1)}\bigr)=0$ for
$i\neq \lambda$. ($C_\ast^{(-1)}=0$).
Then we have a chain complex $(\Bar C_\ast, \Bar\partial_\ast)$
given by $\Bar C_\lambda= H_\lambda
\bigl( C^{(\lambda)}/C^{(\lambda-1)}\bigr)$ and $\Bar\partial$ is given
by the boundary in the homology exact sequence of the triple
$\bigl( C^{(\lambda)}, C^{(\lambda-1)}, C^{(\lambda-2)}\bigr)$.
There is a well{--}known canonical isomorphism
$H_i(\Bar C)\ \RA{\ \cong\ }\ H_i(C)$ (see Milnor, Lemma 5.1).
Now suppose each $C^{(\lambda)}_i/C^{(\lambda-1)}_i$ has an s{--}basis
$c^\lambda_i$: each $\Bar C_\lambda$ has an s{--}basis $\Bar c_\lambda$:
each $H_i(\Bar C)$ has an s{--}basis $h_i$.
Assume $C_\ast$ is a finite complex.
Then so is $\Bar C_\ast$.
Each $C^{(\lambda)}/C^{(\lambda-1)}$ has a torsion.
If $C_i$ is s{--}based by $c^0_i c^1_i$ \dots $c^n_i$, and $H_i(C)$ is
based by $h_i$ composed with the canonical isomorphism, then the
torsion of $C$ is defined.
Lastly, the torsion of $\Bar C$ is also defined.
\bigskip
\BEGIN{T.1.5.2}
(Algebraic Subdivision Theorem)
\[\tau(C)= \tau(\Bar C\> )+\sum_{\lambda=0}^n
\tau\bigl(C^{(\lambda)}/C^{(\lambda-1)}\bigr)\quad .\]
\end{Theorem}
\begin{proof}
The proof is the same as Milnor's \cite{btwentythree}, Theorem 5.2.
One does the same induction, but one just shows
$\displaystyle \tau(C^{(k)})= \tau(\Bar C^{(k)})+\sum_{\lambda=0}^k
\tau\bigl(C^{(\lambda)}/C^{(\lambda-1)}\bigr)$
(notation is the same as Milnor's).
\end{proof}
\medskip
Now let $(K,L)$ be a pair of finite dimensional \hCWx\ complexes with $L$
a proper deformation retract of $K$.
We have the modules $C_\ast(K, L\Colon \Lambda, F)$.
The exact sequence of a triple makes $C_\ast$ into a chain complex,
whose homology is zero since $L$ is a proper deformation retract of
$K$.
The paths $\Lambda$ give us a basis for $C_\ast$ up to sign; i.e.
we must orient each cell, which we can do arbitrarily.
$\tau(K,L\Colon \Lambda,f)\in \Bar K_1\bigl(\Z\pi_1(K,F,f)\bigr)$
is the torsion of this complex with the basis given by $\Lambda$.
We proceed to show that it does not depend on the choice of signs.
Let $\tau^\prime$ be the torsion with a different choice of signs.
Then, by \fullRef{L.1.5.14} below,
$\tau^\prime - \tau=
\displaystyle\sum_\ast (-1)^\ast[c_\ast/c^\prime_\ast]$
where $c_\ast$ and $c^\prime_\ast$ are maps $F_\pi\ \to\ C_\ast$,
one with the signs for $\tau$ and the other
with the signs for $\tau^\prime$.
But $c^{-1}_\ast\ \circ\ c^\prime_\ast\colon F_\pi\ \to\ F_\pi$
lies in the image of ${\mathcal P}_T\ \to\ {\mathcal P}_R$, and so
$[c_\ast/c^\prime_\ast]=0$ in $\Bar K_1(R)$.
\medskip
\BEGIN{L.1.5.14}
Let $C_\ast$ be a finite chain complex.
Let $c_\ast$ and $c^\prime_\ast\colon F_\pi\ \to\ C_\ast$ be two free
bases for $C_\ast$.
Suppose $H_\ast(C)$ is s{--}based.
Let $\tau$ and $\tau^\prime$ be the torsions from the bases
$c_\ast$ and $c^\prime_\ast$ respectively.
Then $\tau-\tau^\prime =
\displaystyle\sum_{\ast}(-1)^\ast \bigl[c_\ast/c^\prime_\ast\bigr]$.
\end{Lemma}
\begin{proof}
This is a fairly dull computation.
\end{proof}
\medskip
Now suppose $G$ is a different lift functor with $F\leq G$.
Then by \fullRef{P.1.5.2}, the basis
$c_\ast\colon F_\pi\ \to\ C_\ast(F)$ goes to $c_\ast\colon
F_\pi\ \to\ C_\ast(G)$ under $\otimes_{\Z\pi_1(F)}\Z\pi_1(G)$.
Let $c^\prime_\ast\colon F_\rho\ \to\ C_\ast(G)$ be the usual
basis.
Then $\pi\subseteq\rho$, and $F_\pi\ \to\ F_\rho
\ \RA{\ c^\prime_\ast\ }\
C_\ast(G)$ is just $c_\ast$.
The inclusion $F_\pi\ \to\ F_\rho$ lies in the image of ${\mathcal P}_T$
in ${\mathcal P}_R$, so $[c_\ast/c^\prime_\ast]=0\in \Bar K_1\bigl(
\Z\pi_1(K,G,f)\bigr)$.
Hence
$i_\ast \tau(K,L\Colon F,\Lambda) - \tau(K,L\Colon G,\Lambda)=0$ where
$i_\ast\colon \Bar K_1\bigl(\Z\pi_1(K,F,f)\bigr)\ \to\
\Bar K_1\bigl(\Z\pi_1(K,G,f)\bigr)$.
Therefore we can define $\tau(K,L\Colon \Lambda)\in \Bar K_1(K\Colon f)$.
$\tau(K,L\Colon \Lambda)$ depends strongly on $\Lambda$.
We would like this not to be the case, so we pass to a quotient of
$\Bar K_1$.
\medskip\begin{xDefinition}
Let $G$ be a tree of groups with associated tree of rings $\Z G$.
The \emph{Whitehead group} of $G$,
$\wh(G)=\Bar K_1(\Z G)/\bigl(\Delta(G)\bigr)$,
where $\bigl(\Delta(G)\bigr)$ is the subgroup generated by all
objects of the form $( F^{(1)},[g], F^{(1)})$ where $[g]$ is the
map{--}germ of $F^{(1)}$ to itself induced by any element
$g\in\Delta(G)$ as follows:
$g$ can be represented by a collection $\{g_p\}$, where
$g_p\in G_{A(p)}$, with $p\in A(p)$ and $\{ A(p) \}$ cofinal and
locally finite.
Define a partition, $\pi$, of the vertices of $T$ by
$\pi(A)=\{\ p\in T\ \vert\ A(p)\subseteq A\ \}$.
$\pi$ is seen to be a partition and $\pi\subseteq \rho$,
the standard partition.
Define a map $g \colon F_\pi\ \to\ F_\pi$ by
$g_A\colon (F_\pi)_A\ \to\ (F_\pi)_A$ takes $e_p$ to
$e_p\cdot f_{A_p\>A}(g_p)$ where
$f_{A B}\colon (\Z G)_A\ \to\ (\Z G)_B$.
It is not hard to show this is a well{--}defined map{--}germ.
What we have actually done is to construct a homomorphism
$\Delta(G)\ \to\ \Bar K_1(\Z G)$ defined by
$g\mapsto \bigl(F^{(1)},[g], F^{(1)}\bigr)$.
Be definition
$\Delta(G)\ \to\ \Bar K_1(\Z G)\ \to\ \wh(G)\ \to0$ is exact.
\end{xDefinition}
Given a homomorphism $f\colon G\ \to\ H$ between two trees of groups,
we clearly get a commutative square
\[\begin{matrix}%
\Delta(G)&\to&\Bar K_1(\Z G)\\
\downarrow&&\downarrow\\
\Delta(H)&\to&\Bar K_1(\Z H)\hbox to 0pt{$\quad ,$\hss}\\
\end{matrix}\]
so we get a homomorphism $\wh(G)\ \to\ \wh(H)$.
\medskip
\BEGIN{L.1.5.15}
Let $f\colon G\ \to\ H$ be a map between two trees of groups for
which $\Delta(f)$ is an isomorphism.
Then $\wh(G)\ \to\ \wh(H)$ is an isomorphism.
\end{Lemma}
\begin{proof}
$\Delta(f)\colon \Delta(G)\ \to\ \Delta(H)$ is also an isomorphism,
so apply \fullRef{L.1.5.12}
and the $5${--}lemma.
\end{proof}
\medskip
We can now define $\wh(X\Colon f)$ as
$\displaystyle\lim_{
\Atop{\longrightarrow}{F\in{\mathcal L}(f)}
}\
\wh\bigl(\Z\pi_1(X,F,f)\bigr)$.
\bigskip
\BEGIN{P.1.5.4}
Let $(K,L)$ be a pair of finite dimensional \hCWx\ complexes with $L$
a proper deformation retract of $K$.
Then, if $\Lambda$ and $\Lambda^\prime$ are two choices of
paths, $\tau(K,L\Colon \Lambda)=\tau(K,L\Colon \Lambda^\prime)$ in $\wh(X\Colon f)$,
Hence we can then define $\tau(K,L)\in\wh(X\Colon f)$.
\end{Proposition}
\begin{proof}
We can pick any lift functor we like, say $F$.
$C_\ast(K,L\Colon \Lambda, F) = C_\ast(K,L\Colon \Lambda^\prime, F)$,
and each is naturally based.
Let $\pi_\ast$ be the partition associated to $\Lambda$ (see
\fullRef{P.1.5.3} ) and let $\pi^\prime_\ast$ be the
partition associated to $\Lambda^\prime$.
Let $\rho_\ast$ be the partition $\rho_\ast(A)=\{ e\ \vert\ e$
is a $\ast${--}cell in $F(A)$ and the path for $e$ in $\Lambda$ lies in
$F(A)$ and the path for $e$ in $\Lambda^\prime$ also lies in $F(A)\ \}$.
$\rho_\ast=\pi_\ast\ \cap\ \pi_\ast^\prime$.
The basis $F_{\rho_\ast}\ \to\ C_\ast$ is equivalent to the basis
$F_{\pi_\ast}\ \to\ C_\ast$.
Similarly $F_{\rho_\ast}\ \to\ C^\prime_\ast$ is equivalent to the basis
$F_{\pi^\prime_\ast}\ \to\ C^\prime_\ast$.
($C_\ast=C_\ast(\cdots, \Lambda)$;
$C^\prime_\ast=C_\ast(\cdots, \Lambda^\prime)$.)
$\tau^\prime- \tau = \tau(K,L\Colon \Lambda^\prime) -\tau(K,L\Colon \Lambda)=
\displaystyle\sum_\ast (-1)^\ast [\pi_\ast/\pi_\ast^\prime]$, by
\fullRef{L.1.5.14}.
If we can show $[\pi_\ast/\pi_\ast^\prime]$ is in the image of
$\Delta(\pi_1)$ we are done.
But this is not hard to see ( $\wh(\ )$ was defined by factoring out
such things).
\end{proof}
\medskip
Having defined a torsion, we prove it invariant under subdivision.
We follow Milnor \cite{btwentythree}.
\bigski
\BEGIN{T.1.5.3}
The torsion $(K,L)$ is invariant under subdivision of the pair $(K,L)$;
$(K,L)$ a finite dimensional \hCWx\ pair.
\end{Theorem}
\begin{proof}
Following Milnor \cite{btwentythree} we prove two lemmas.
\medskip
\BEGIN{L.1.5.16}
Suppose that each component of $K-L$ has compact closure and is simply
connected.
If $L$ is a proper deformation retract of $K$, then $\tau(K,L)=0$.
\end{Lemma}
\begin{proof}
(Compare Milnor \cite{btwentythree} Lemma 7.2).
Let $f\colon T\ \to\ K$ be a tree.
We wish to find a set of paths $\Lambda$ so that the boundary maps
in $C_\ast(K,L\Colon F,\Lambda)$ come from ${\mathcal P}_T$.
Let $\{ M_i\}$ be the components of $K-L$.
Pick a point $q\in M_i$ and join $\{ q_i\}$ to $T$ by a locally
finite set of paths $\lambda_i$.
Now join each cell in $M_i$ to $q_i$ by a path lying in $M_i$.
Let $\Lambda$ be the set of paths gotten by following the path from the
cell to a $q_i$ and then following the path $\lambda_i$.
Clearly $\Lambda$ is a locally finite set of paths joining the cells of $K-L$
to $T$.
Let $e$ be a cell of $K-L$.
Then if $f$ is a cell of $\partial e$, to compute the coefficient of $f$
in $\partial e$ we join the barycenter of $f$ to the barycenter of $e$
by a path in $e$ and look at the resulting loop.
The path from $e$ and the path from $f$ hit the same $q_i$, and since
$\pi_1(M_i,q_i)=0$, the coefficient is $\pm1$, so the boundary maps come
from ${\mathcal P}_T$.
\end{proof}
\medskip
\BEGIN{L.1.5.17}
Suppose that $H_\ast\bigl(C_\ast(K,L\Colon \Lambda)\bigr)$ is not $0$,
but is a free $\Z\pi_1(K)${--}module with a preferred basis.
Suppose each basis element can be represented by a cycle lying over
a single component of $K-L$.
Assume as before that each component of $K-L$ is compact and
simply connected.
Then $\tau(K,L)=0$.
\end{Lemma}
\begin{proof}
Pick a set of paths as in \fullRef{L.1.5.16}
so that the boundary maps come from ${\mathcal P}_T$.
Look at a cycle $z$, representing a basis element of $H_\ast$.
What this means is the following.
Let $C_0\subseteq C_1\subseteq$ \dots be an increasing sequence
of compact subcomplexes with $\cup\ C_i=K$ and $M_i\subseteq C_i$.
Then $z\in H_\ast(\coverFA{K-C_i}, \coverFA{L-C_i})$ for a maximal
$C_i$.
Then $z$ is represented by a cycle lying in some component of
$\pi^{-1}(M_{i+1})$, where $\pi\colon \coverFA{K-C_i}\ \to\ K-C_i$.
All the lifted cells of $M_{i+1}$ lie in a single component of
$\pi^{-1}(M_{i+1})$, so let $g\in \pi_1(K-C_i)$ be such that $g z$ also lies
in this distinguished component.
Then the torsion computed with this altered basis is zero since it again
comes from $\wh(T)=0$.
But the new basis for $H_\ast$ is clearly equivalent to the old one
in $\wh(K)$.
\end{proof}
\medskip
The proof of \fullRef{T.1.5.3} now follows Milnor's proof of
Theorem 7.1 word for word except for a renumbering of
the requisite lemmas.
\end{proof}
\medskip
\BEGIN{L.1.5.18}
If $M\subseteq L\subseteq K$, where both $L$ and $M$ are
proper deformation retracts of $K$, then
$\tau(K,L)=\tau(K,M)+ i_\ast\tau(L,M)$, where
$i_\ast\colon\wh(L\Colon f)\ \to\ \wh(K\Colon i \circ f)$ is the map induced by
$i\colon L\subseteq K$. (Note the tree must be in $L$.)
\end{Lemma}
\begin{proof}
This is a simple application of \fullRef{T.1.5.1}.
\end{proof}
\medskip
Let $f\colon X\ \to\ Y$ be a proper, cellular map between two
finite dimensional \hCWx\ complexes.
Let $M_f$ be the mapping cylinder.
$Y$ is a proper deformation retract of $M_f$ and we have
\medski
\BEGIN{L.1.5.19}
$\tau(M_f,Y)=0$ in $\wh(M_f, t)$ where $t\colon T\ \to\ Y$
is a tree for $Y\subseteq M_f$.
\end{Lemma}
\begin{proof}
Word for word Milnor \cite{btwentythree} Lemma 7.5.
\end{proof}
\bigskip\begin{xDefinition}
For any cellular proper homotopy equivalence $f\colon X\ \to\ Y$,
$X$ and $Y$ as above, there is a torsion, $\tau(f)$, defined as follows.
Let $t\colon T\ \to\ Y$ be a tree for $Y$.
Then, as in \fullRef{L.1.5.19}, $t$ is also
a tree for $M_f$ under $T\ \to\ Y\subseteq M_f$.
$\tau(f)=r_\ast\tau(M_f,X)\in\wh(Y\Colon t)$ where
$r_\ast\colon \wh(M_f\Colon t)\ \to\ \wh(Y\Colon t)$, where $r$ is the retraction.
\end{xDefinition}
Just as in Milnor we have
\BEGIN{L.1.5.20}
If $i\colon L\ \to\ K$ is an inclusion map
$\tau(i)=\tau(K,L)$ if either is defined.
\end{Lemma}
\medski
\BEGIN{L.1.5.21}
If $f_0$ and $f_1$ are properly homotopic, $\tau(f_0)=\tau(f_1)$.
\end{Lemma}
\medski
\BEGIN{L.1.5.22}
If $f\colon X\ \to\ Y$ and $g\colon Y\ \to\ Z$ are cellular proper
homotopy equivalences, then
\[\tau(g\ \circ\ f)=\tau(g)+g_\ast\tau(f)\ ,\] where $t\colon T\ \to\ Y$
is a tree for $Y$ and $g_\ast\colon \wh(Y\Colon t)\ \RA{\ \cong\ }\
\wh(Z\Colon g \circ t)$.
\end{Lemma}
\medskip\begin{xRemarks}
It follows from \fullRef{L.1.5.21} that we may define
the torsion of any proper homotopy equivalence between finite dimensional
\hCWx\ complexes, since we have a proper cellular approximation
theorem \cite{beleven}.
\end{xRemarks}
\medskip
Now in \cite{bthirtythree}, Siebenmann defined the notion of simple
homotopy type geometrically.
In particular, he got groups $\sieb(X)$ associated to any locally compact
CW complex.
If $X$ is finite dimensional, we can define a map
$\tau\colon\sieb(X)\ \to\ \wh(X\Colon f)$ by choosing a tree
$f\colon T\ \to\ X$.
If $g\colon X\ \to\ Y$ is an element of $\sieb(X)$, $g$
goes to $\tau(M_{g^{-1}},Y)$ where $g^{-1}\colon Y\ \to\ X$
is a proper homotopy inverse for $g$.
$\tau$ is additive by \fullRef{L.1.5.22}
and depends only on the proper homotopy class of $g$ by
\fullRef{L.1.5.21}.
That $\tau$ is well{--}defined reduces therefore to showing that
$g$ a simple homotopy equivalence implies $\tau(g)=0$.
We defer for the proof to Farrell{--}Wagoner \cite{bten},
where it is also proved $\tau$ is an isomorphism.
The inverse for $\tau$ is easy to describe.
Let $\alpha\in\wh(X\Colon f)$ be an automorphism of $F^{(n)}$ for some $n$.
Wedge $n$ $2${--}spheres to each vertex of the tree.
Attach $3${--}cells by $\alpha$ to get an \hCWx\ complex Y with $Y-X$
$3${--}dimensional.
Then $i\colon X\subseteq Y$ is an element of $\sieb(X)$ and
$\tau(i)=\alpha$.
Again we defer to \cite{bten} for the proof that this map is well{--}defined.
In \cite{bthirtythree} Siebenmann also constructs an exact sequence
\[0\to\ \wh^\prime \pi_1(X)\ \to\ \sieb(X)\ \to\
K_0\pi_1E(X)\ \to\ K_0 \pi_1(X)\ .\]
We have
\[\begin{matrix}%
&&\sieb(X)\\
&\hskip 60pt\hbox to 0pt{$\nearrow$\hss}&&\searrow\\
0\to&\wh^\prime \pi_1(X)&\uplabeledarrow[\Big]{\tau^{-1}}{}&&
K_0\pi_1E(X)&\to&K_0 \pi_1(X)\\
&\hskip 60pt\hbox to 0pt{$\hbox to 0pt{\hss$\scriptstyle\alpha$}\searrow$\hss}&&
\nearrow\hbox to 0pt{$\scriptstyle\beta$\hss}\\
&&\wh(X\Colon f)\\
\end{matrix}\]
commutes.
Farrell and Wagoner describe $\alpha$ and $\beta$ and prove this diagram
commutes.
They show that the bottom row is exact, so $\tau^{-1}$ is an
isomorphism.
Note now that if $g\colon T\ \to\ X$ is another tree for $X$, we
have natural maps \\$\wh(X\Colon f)\
{\crB}\ \wh(X\Colon g)$
which take $\tau(X,Y)$ computed with $f$ to $\tau(X,Y)$ computed
with $g$ and vice{--}versa.
This shows $\wh(X\Colon f)$ does not really depend on the choice of tree.
We content ourselves with remarking that the map
$\wh(X\Colon f)\ \to\ \wh(X\Colon g)$ is not easy to describe algebraically.
In \cite{bthirtythree} Siebenmann derives some useful formulas which
we name
\begin{enumerate}
\item[1)] Sum formula
\item[2)] Product formula
\item[3)] Transfer formula
\end{enumerate}
Note if $\pi\colon \coverFA{Y}\ \to\ Y$ is a cover, $\pi$ induces
$\pi^\ast\colon \sieb(Y)\ \to\ \sieb(\coverFA{Y})$.
We are unable to say much about this map algebraically.
The product formula is algebraically describable however.
\medski
\BEGIN{L.1.5.23}
Let $C_\ast$ be an s{--}based, finite chain complex over the tree of
rings $R$.
Let $D_\ast$ be an s{--}based, finite chain complex on the ring $S$
(the tree of rings over a point).
Then $(C\otimes D)_\ast$ is defined.
If $C_\ast$ is acyclic with torsion $\tau$, $(C\otimes D)_\ast$
is acyclic with torsion $\chi(D)\cdot i_\ast \tau(C)\in \wh(R\times S)$
where $(R\times S)_A= R_A\times S$, and
$i_\ast\colon \wh(R)\ \to\ \wh(R\times S)$ is the obvious split
monomorphism.
If $D_\ast$ is acyclic, then so is $(C\otimes D)_\ast$, and if
$\tau(D)=0$, then $\tau(C\otimes D)=0$.
\end{Lemma}
\begin{proof}
The first formula is Siebenmann's product formula and is proved
by induction on the number of cells in $D_\ast$.
The second formula is new, but it is fairly easy.
It basically requires the analysis of maps
$\wh(S)\ \to\ \wh(R\times S)$ of the form
$D_\ast\ \to\ P\otimes D_\ast$ for $P$ and s{--}based $R${--}module.
These maps are homomorphisms, and so, if $\tau(D_\ast)=0$,
$\tau(P\otimes D_\ast)=0$.
But $\tau\bigl((C\otimes D)_\ast\bigr)=\displaystyle
\sum_k(-1)^k\tau(C_k\otimes D_\ast)$.
(There is evidence for conjecturing that the map
$\wh(S)\ \to\ \wh(R\times S)$ is always $0$\setcounter{footnote}{0}\footnote{If the tree is infinite.}.)
\end{proof}
\medskip
We conclude this section by discussing the notion of duality.
In particularly, we would like a functor
$\du\colon {\mathcal M}_R\ \to\ {\mathcal M}^\ell_{\mathcal R}$
which generalizes the usual duality $P\ \to\ {\Homx}(P, R)$ in
the compact case.
Up until now, ${\mathcal M}_R$ has denoted without prejudice either the
category of right or left $R${--}modules.
We now fix it to be the category of right $R${--}modules.
${\mathcal M}^\ell_R$ then denotes the category of left $R${--}modules.
Actually, we are really only interested in
$\du\colon {\mathcal P}_R\ \to\ {\mathcal P}^\ell_R$.
Hence we begin by discussing a functor
$\du\colon {\mathcal F}_R\ \to\ {\mathcal F}^\ell_R$, where ${\mathcal F}_R$
is the category of locally{--}finitely generated free right $R${--}modules.
$\du$ will satisfy
\begin{enumerate}
\item[1)] $\du$ is a contravariant, additive, full faithful functor
\item[2)] $\du\du$ is naturally equivalent to the identity.
\end{enumerate}
By this last statement we mean the following.
Given $\du\colon{\mathcal F}_R\ \to\ {\mathcal F}^\ell_R$ there will be another
obvious duality $\du\colon {\mathcal F}^\ell_R\ \to\ {\mathcal F}_R$.
The composition of these two is naturally equivalent to the identity.
We proceed to define $\du$.
If $F_A$ is a free right $R_A${--}module based on the set $A$, there is
also a free left $R_A${--}module based on the same set,
$F_A^\ast$.
$F^\ast_A$ can be described as ${\Homx}^\cmpsup_{R_A}(F_A,R_A)$,
where ${\Homx}^\cmpsup_{R_A}$ is the set of all $R_A${--}linear
homomorphisms which vanish on all but finitely many generators.
${\Homx}^\cmpsup_{R_A}(F_A,R_A)$, is easily seen to have
the structure of a left $R_A${--}module.
Let $A\subseteq B$, and let $f\colon R\ \to\ S$ be a ring homomorphism.
Then we have
\[\begin{aligned}%
{\Homx}^\cmpsup_{R_A}(F_A,R_A)\ \to\
{\Homx}^\cmpsup_{R_B}(F_A\otimes R_B,R_B)\ \LA{\ {\text{ex}}\ }\\
{\Homx}^\cmpsup_{R_B}(F_B/F_{B-A},R_B)\ \to\
{\Homx}^\cmpsup_{R_B}(F_B,R_B)\\
\end{aligned}\]
The map ${\text{ex}}$ is an isomorphism since $0\to\ F_A\otimes R_B\ \to\
F_B\ \to\ F_{B-A}\ \to0$ is split exact.
Thus we get a well{--}defined homomorphism
\[{\Homx}^\cmpsup_{R_A}(F_A,R_A)\ \to\
{\Homx}^\cmpsup_{R_B}(F_B,R_B)\ .\]
Now given $F_\pi$, let $F^\ast_\pi$ be the tree of left modules over
the tree of rings $R$ defined by
$(F^\ast_\pi)_A={\Homx}^\cmpsup_{R_A}(F_{\pi(A)},R_A)$, and use
the map discussed above to define $p_{A B}$.
Given a map $f\colon F_\pi\ \to\ F_\rho$, define
$f^\ast\colon F_\rho^\ast\ \to\ F_\pi^\ast$ by
\[(f^\ast)_A={\Homx}(f_A) \colon
{\Homx}^\cmpsup_{R_A}(F_{\rho(A)},R_A)\ \to\
{\Homx}^\cmpsup_{R_A}(F_{\pi(A)}, R_A)\ .\]
We must check that $(f^\ast)_A$ is defined and that the requisite
diagrams commute.
This last is trivial, so we concentrate on the first objective.
To this end, let $\alpha\in {\Homx}^\cmpsup_{R_A}(F_{\rho(A)},R_A)$.
We must show ${\Homx}(f_A)(\alpha)$ lies in
${\Homx}^\cmpsup_{R_A}(F_{\pi(A)},R_A)\subseteq
{\Homx}_{R_A}(F_{\pi(A)},R_A)$.
Since $\alpha$ has compact support, $\alpha$ vanishes on the
generators corresponding to a subset $S\subseteq \rho(A)$ with
$\rho(A)-S$ finite.
Hence there is a $B\in{\mathcal C}(T)$ so that $\rho(B)\subseteq S$;
i.e. $\alpha$ vanishes on generators corresponding to $\rho(B)$.
Let $\Bar F_{\pi(B)} = F_{\pi(B)}\otimes_{R_B} R_A$; let
$\Bar F_{\rho(B)} = F_{\rho(B)}\otimes_{R_B} R_A$; and let
$\Bar f_B=f_b\otimes {\text{id}}$.
Then
\[\begin{matrix}%
{\Homx}(F_{\rho(A)}, R_A)&\RA{\ {\Homx}(f_A)\ }&
{\Homx}(F_{\pi(A)}, R_A)\\
\downlabeledarrow[\Big]{i}{}&&\downlabeledarrow[\Big]{j}{}\\
{\Homx}(\Bar F_{\rho(B)}, R_A)&\RA{\ {\Homx}(\Bar f_B)\ }&
{\Homx}(\Bar F_{\pi(B)}, R_A)\\
\end{matrix}\]
commutes.
$\alpha$ is in the kernel of $i$, so ${\Homx}(f_A)(\alpha)\in{\kerx}\ j$.
But this means ${\Homx}(f_A)(\alpha)$ has compact support.
There is a natural map $F\ \to\ F^{\ast\ast}$ induced by the natural
inclusion of a module into its double dual.
This map is an isomorphism\setcounter{footnote}{0}\footnote{Since $F$ is finitely-generated and free.} and
\[\begin{matrix}%
F&\RA{\hskip 10pt}&F^{\ast\ast}\\
\downlabeledarrow{f}{}&&\downlabeledarrow{}{f^{\ast\ast}}{}\\
G&\RA{\hskip 10pt}&G^{\ast\ast}\\
\end{matrix}\]
commutes.
$\du$ is clearly contravariant and a functor.
If $\pi\subseteq\rho$, one sees $F_\rho^\ast\ \to\ F^\ast_{\pi}$
is an equivalence.
Hence we can define $\du$ for map{--}germs.
$(f+g)^\ast= f^\ast+g^\ast$ is easy to see, so $\du$ is additive.
Since $\du\du$ is naturally isomorphic to the identity, $\du$ must be both
faithful and full, so 1) is satisfied.
We next define the subcategory on which we wish to define $\du$.
Let $\Bar{\mathcal M}_R$ be the full subcategory of ${\mathcal M}_R$
such that $M\in \Bar{\mathcal M}_R$ \iff\ there exists
$f\colon F_{\rho}\ \to\ F_{\pi}$ with ${\cokerx\ }f\cong M$.
Note ${\mathcal P}_R\subseteq \Bar{\mathcal M}_R$.
We define $\du \colon \Bar{\mathcal M}_R\to {\mathcal M}^\ell_R$ by
$M^\ast={\kerx}(f^\ast)$.
Given $M$, $N\in \Bar{\mathcal M}_R$, a map $g\colon M\ \to\ N$, and
resolutions $F_\rho\ \to\ F_\pi\ \to\ M\ \to0$ and
$F_\alpha\ \to\ F_\beta\ \to\ M\ \to0$, note that we can compare
resolutions.
That is, we can find $h$ and $f$ so that
\[
\begin{matrix}%
&F_\rho&\RA{\ f\ }&F_{\alpha}\\
&\downarrow&&\downarrow\\
\hbox to 0pt{\hss A)\hskip 20pt}&F_\pi&\RA{\ h\ }&F_{\beta}\\
&\downarrow&&\downarrow\\
&M&\RA{\ g\ }&N\\
&\downarrow&&\downarrow\\
&0&&0\\
\end{matrix}\] commutes.
\par\noindent Define $g^\ast\colon N^\ast\ \to\ M^\ast$ by
\[
\begin{matrix}%
&0&&0\\
&\downarrow&&\downarrow\\
&N^\ast&\RA{\ g^\ast\ }&M^\ast\\
&\downarrow&&\downarrow\\
\hbox to 0pt{\hss B)\hskip 20pt}&
F^\ast_\beta&\RA{\ h^\ast\ }&F^\ast_{\pi}\\
&\downarrow&&\downarrow\\
&F^\ast_\alpha&\RA{\ f^\ast\ }&F^\ast_{\rho}\ .\\
\end{matrix}\]
We first note that the definition of $g^\ast$ does not depend on $h$
and $f$, for if we pick $h_1$ and $f_1$ such that A) commutes, there
is a commutative triangle
\[\begin{matrix}%
&&F_\alpha\\
&\hbox to 0pt{\hss $\scriptstyle p$}\nearrow&\downarrow\\\noalign{\vskip 6pt}
F_\pi&\hbox to 0pt{\hss$\RA{h-h_1}$\hss}& F_\beta
\hbox to 0pt{\quad .\hss}\\
\end{matrix}\]
Dualizing, we get
\[\begin{matrix}%
F^\ast_\beta&\hbox to 0pt{\hss$\RA{h^\ast-h^\ast_1}$\hss}
& F^\ast_\pi\\\noalign{\vskip6pt}
\downarrow&\nearrow\hbox to 0pt{$\scriptstyle p^\ast$\hss}\\
F^\ast_\alpha&&\hbox to 0pt{\quad .\hss}\\
\end{matrix}\]
Now this triangle shows that the map we get from $f_1$ and $h_1$
is the same as we got from $f$, $h$.
To show $M^\ast$ does not depend on the resolution is now done
by comparing two resolutions and noting $({\text{id}})^\ast={\text{id}}$.
Unfortunately, $(M^\ast)^\ast$ may not even be defined, so we have
little hope of proving a result like 2).
One useful result that we can get however is
\medski
\BEGIN{L.1.5.24}
Let $f\colon P\ \to\ M$ be an epimorphism with $M\in\Bar{\mathcal M}_R$
and $P\in{\mathcal P}_R$.
Then $f^\ast\colon M^\ast\ \to\ P^\ast$ is a monomorphism.
\end{Lemma}
\begin{proof}
The proof is easy.
\end{proof}
\medskip
If we restrict ourselves to ${\mathcal P}_R$, we can get 1) and 2) to hold.
It is easy to see $P^\ast\in{\mathcal P}^\ell_R$ for $P\in{\mathcal P}_R$.
Now the equation $(P\oplus Q)^\ast=P^\ast\oplus Q^\ast$ is easily
seen since direct sum preserves kernels.
Thus $(P\oplus Q)^{\ast\ast}=P^{\ast\ast}\oplus Q^{\ast\ast}$, so it is
not hard to see $P\ \to\ P^{\ast\ast}$ must be an isomorphism since
if $P$ is free the result is known.
Lastly, $\du$ is natural, i.e. if $f\colon R\ \to\ S$ is a map,
\lower \baselineskip\hbox{$\begin{matrix}%
\Bar{\mathcal M}_R&\RA{\hskip10pt}&\Bar{\mathcal M}_S\\
\downlabeledarrow{\du}{}&&\downlabeledarrow{\du}{}\\
{\mathcal M}^\ell_R&\RA{\hskip10pt}&{\mathcal M}^\ell_S\\
\end{matrix}$} commutes.
That $\Bar{\mathcal M}_R$ hits $\Bar{\mathcal M}_S$ follows since $\otimes$
is right exact.
\medskip\begin{xDefinition}
Let $\{ M_i, \partial_i\}$ be a chain complex with
$M_i\in\Bar{\mathcal M}_R$.
Then $\{ M^\ast_i,\partial^\ast_i\}$ is also a chain complex.
The \emph{cohomology} of $\{ M_i, \partial_i\}$ is defined
as the homology of $\{M^\ast_i, \partial^\ast_i\}$.
\end{xDefinition}
\medski
\BEGIN{P.1.5.5}
Let $(X,Y)$ be an \hCWx\ pair; let $F$ be a lift functor; and let $\Lambda$
be a set of paths.
Then $\{C_\ast(X,Y\Colon F,\Lambda), \partial_\ast\}$ is a chain complex
as we saw.
Its dual is
$\{C^\ast(X,Y\Colon F,\Lambda), \delta^\ast\}$.
Hence the cohomology of a pair is the same as the cohomology of
its chain complex.
\end{Proposition}
\begin{proof} Easy.
\end{proof}
\medskip
Notice that our geometric chain complexes lie in ${\mathcal P}_R$.
For such complexes we can prove
\bigski
\BEGIN{T.1.5.4}
Let $\{ P_r, \partial_r\}$ be a finite chain complex in ${\mathcal P}_R$.
$H_k(P)=0$ for $k\leq n$ \iff\ there exist maps
$D_r\colon P_r\ \to\ P_{r+1}$ for $r\leq n$ with
$D_{r-1}\partial_r + \partial_{r+1}D_r={\text{id}}_{P_r}$.
\end{Theorem}
\begin{proof} Standard.
\end{proof}
\medski
\BEGIN{C.1.5.4.1}
(Universal Coefficients).
With $\{ P_r,\partial_r\}$ as above, $H_k(P)=0$ for $k\leq n$
implies $H^k(P)=0$ for $k\leq n$.
$H^k(P)=0$ for $k\geq n$
implies $H_k(P)=0$ for $k\geq n$.
\end{Corollary}
\begin{proof} Standard.
\end{proof}
\bigskip
Now suppose $\{P_r,\partial_r\}$ is a chain complex in ${\mathcal P}_R$.
Then ${\cokerx}\ \partial_{r+1}\in\Bar{\mathcal M}_R$.
By \fullRef{L.1.5.24},
${\kerx\ }\delta^{r}=({\cokerx\ }\partial_{r+1})^\ast$.
Now
\topD{12}{$\begin{matrix}%
&&P_{r-1}\\
&\hbox to 0pt{\hss$\scriptstyle\partial _r$}\nearrow&\uparrow\\\noalign{\vskip 6pt}
P_r&\to&{\cokerx\ }\partial_{r+1}\\
&&\uparrow\\
&&H_r(P)\\
&&\uparrow\\
&&0\\\end{matrix}
$}{5}
commutes and is exact.
If $H_r(P)\in\Bar{\mathcal M}_R$, applying duality to this
diagram yields
\\
\topD{12}{$\begin{matrix}%
P^\ast_{r-1}&\RA{\ \alpha\ }&{\kerx}(\delta^r)&\RA{\ \beta\ }&
\bigl(H_r(P)\bigr)^\ast\\
&\hbox to 0pt{\hss$\scriptstyle \delta^{r-1}$}\searrow&\downarrow\\\noalign{\vskip 3pt}
&&P^\ast_r\\
\end{matrix}$\ .}{5}
By definition,
${\cokerx\ }\alpha = H^r(P)$. $\beta\ \circ\ \alpha=0$, so there
is a unique, natural map
$H^r(P)\ \to\ \bigl(H_r(P)\bigr)^\ast$.
\medski
\BEGIN{C.1.5.4.2}
With $\{P_r,\partial_r\}$ as above, if $H_k(P)=0$ for $k<n$,
$H_n(P)\in\Bar{\mathcal M}_R$.
If $H_n(P)\in{\mathcal P}_R$, the natural map
$H^n(P)\ \to\ \bigl(H_n(P)\bigr)^\ast$ is an isomorphism.
\end{Corollary}
\begin{proof}
By induction, one shows $Z_n\in{\mathcal P}_R$, and since
$P_{n+1}\ \RA{\ \partial_{n+1}\ }\ Z_n\ \to\ H_n(P)\ \to0$ is exact,
it is not hard to see $H_n(P)\in\Bar{\mathcal M}_R$.
If $H_n(P)\in{\mathcal P}_R$,
$0\to\ \bigl(H_n(P)\bigr)^\ast\ \to\ Z^\ast_n\ \to\ P^\ast_{n+1}$
is exact, so $H^n(P)\cong \bigl(H_n(P)\bigr)^\ast$.
\end{proof}
\bigski
\BEGIN{T.1.5.5}
With $\{ P_r,\partial_r\}$ as above, suppose $H_k(P)=0$ for $k<n$
and $H^k(P)=0$ for $k>n$.
Then $H_n(P)\in{\mathcal P}_R$ and the natural map
$H^n(P)\ \to\ \bigl(H_n(P)\bigr)^\ast$ is an isomorphism.
In $K_0(R)$,
$\bigl[H_n(P)\bigr]=(-1)^n\chi(P)$, where
$\chi(P)\in K_0(R)$ is $\displaystyle\sum_r (-1)^r[P_r]$.
\end{Theorem}
\begin{proof}
Since $H_k(P)=0$ for $k<n$, the sequence
$\cdots\to\ P_{n+1}\ \to\ P_n\ \to\ P_{n-1}\ \to\cdots$
splits up as
\[\begin{aligned}
\cdots\to\ P_{n+1}\ \to\ Z_n&\ \to 0\\
0\to\ Z_n&\ \to\ P_n\ \to\ P_{n-1}\ \to\cdots\\\end{aligned}\]
The second sequence is exact, and
$\cdots\to\ P_{n+1}\ \to\ Z_n\ \to H_n(P)\ \to 0$
is exact since $H_k(P)=0$ for $k>n$ by
\fullRef{C.1.5.4.1}.
By \fullRef{C.1.5.4.2}, $H_n\in\Bar{\mathcal M}_R$.
Dualizing, we get
$\cdots\leftarrow\ P^\ast_{n+1}\ \leftarrow\ Z^\ast_n\ \leftarrow\
(H_n)^\ast\ \leftarrow 0$ is exact by
\fullRef{C.1.5.4.1} and
\fullRef{L.1.5.24}.
As in the proof of
\fullRef{T.1.5.4}, we get a chain retraction up to
$D\colon P^\ast_{n+1}\ \to\ Z^\ast_n$.
This shows $(H_n)^\ast\in{\mathcal P}_R$.
But
\[\begin{aligned}%
\cdots\leftarrow\ P^\ast_{n+1}\ \leftarrow\ &Z^\ast_n\ \leftarrow\
(H_n)^\ast\ \leftarrow 0\\
0\leftarrow\ &Z^\ast_{n}\ \leftarrow\ P^\ast_n\ \leftarrow\
P^\ast_{n-1}\ \leftarrow\cdots\\
\end{aligned}\]
splice together to give the cochain complex.
$H^n\ \to (H_n)^\ast$ is now easily seen to be an isomorphism.
Now $\displaystyle
\sum_{r\geq n+1} (-1)^r[P_r] +(-1)^n[Z_n] + (-1)^{n-1}[H_n]=0$
and $\displaystyle
\sum_{r\leq n} (-1)^r[P_r] +(-1)^{n+1}[Z_n] =0$
in $K_0(R)$ by Bass \cite{bone}, Proposition 4.1, Chapter VIII.
Summing these two equations shows
$\chi(P)+(-1)^{n-1}[H_n]=0$.
\end{proof}
Now let us return and discuss the products we defined in section 4.
We defined two versions of the cap product on the chain level
(see Theorems \shortFullRef{T.1.4.5} and \shortFullRef{T.1.4.6}).
Notice that the maps we defined on $P_\ast(X;A,B)$ and $P^\ast(X;A,B)$
actually come from maps on the tree modules
$C_\ast(X;A,B\Colon \Lambda,F)$ and $C^\ast(X;A,B\Colon \Lambda,F)$.
Thus if $f$ is a cocycle in $C^m(X;A;\Gamma)$, and if $h$ is a
diagonal approximation, \fullRef{T.1.4.5} yields a
chain map
$C_{\ast+m}(X;A,B\Colon \Lambda,F)\ \RA{\ \capf f h\ }\
C_{\ast}(X,B\Colon \Lambda,F)$.
Note that in order for this to land in the asserted place, $\Gamma$
pulled up to the universal cover of $X$ must just be ordinary
integer coefficients.
$\capf f h$ dualizes to
$\cupf f h\colon C^\ast(X,B\Colon \Lambda,F)\ \to\
C^{m+\ast}(X\Colon A,B\Colon \Lambda,F)$.
Since we did not define cup products on the chain level, we may
take this as a definition.
Nevertheless we assert that on homology $f\cup_h$ induces
the cup product of \fullRef{T.1.4.1}.
This follows from the duality relations we wrote down between
ordinary cohomology and homology (see the discussion around
the universal coefficient theorems in section 1)
\setcounter{footnote}{0}\footnote{See page \pageref{Universalcoefficientx}.}.
Now one easily sees $\du$ induces a map $\wh(\du)\colon
\wh(R)\ \to\ \wh^\ell(R)$, where $\wh^\ell(R)$ is the group formed
from left modules.
If $\capf f{}$ (or $\cupf f{}$) is a chain equivalence, we can compare
$\tau(\cupf f{})$ and $\tau(\capf f{})$.
We get $\wh(\du)\bigl(\tau(\capf f{})=(-1)^m\tau(\cupf f{})$
by definition.
Next we study the cap product of \fullRef{T.1.4.6}.\footnote
{We did not get this quite right in the original so there are some changes
here.}\\
A cycle $c\in C^\locf_m(X;A,B;\Gamma)$ yields maps
$C^\ast(X,A\Colon \Lambda,F)\ \to\ C_{m-\ast}(X,B\Colon \Lambda,F)$.
$C^\ast$ is a left module while $C_\ast$ is a right module, so
$\capf c h$ is not a map of tree modules.
If $\Gamma$ has all its groups isomorphic to $\Z$, which it must to
yield the asserted product, we get a homomorphism
$\wo\colon \Gamma(X)\ \to\ \cy2={\Autx}(\Z)$
given by the local system.
We can make $C^\ast$ into a right module (or $C_\ast$ into a left module)
by defining $m_A\cdot a=\Bar{a}\cdot m_A$, where $m_A\in(C^\ast)_A$,
$a\in(\Z\pi_1)_A$ and $\Bar{\barnone}$ \
is the involution on $(\Z\pi_1)_A$ induced by
$g\in (\pi_1)_A$ goes to $\wo(g)\cdot g^{-1}$, where $\wo(g)\in
\cy2=\{1,-1\}$ is the image of $g$ under the composition
$(\pi_1)_A\ \to\ \pi_1(X)\ \RA{\ \wo\ }\ \cy2$.
With this right module structure, $\capf{c}{h}$ is a right module map.
It is not however the case that $\capf{c}{h}$ is a chain map.
The requisite diagrams commute up to sign, but they only commute in
half the dimensions.
To overcome this annoyance, alter the boundary maps in
$C^\ast(X,A\Colon \Lambda,F)$ to be
$\delta^\ast_{(m)}=(-1)^{\ast+m}\delta^\ast$ where $\delta^\ast$ are
the duals of the $\partial$ boundary maps.
Let $\twistedC{\ast}{\wo}{m}$ be
$C^\ast$ with our right module structure
and boundary maps $\delta^\ast_{(m)}$.
Given any finite, projective chain or cochain complex,
$\{ P_\ast, \partial_\ast\}$ ( or $\{ P^\ast, \delta^\ast\}$ ),
we can get a new complex $\{ P_\ast, (-1)^{\ast+m}\partial_\ast\}$.
There are evident chain isomorphisms among the three complexes
and these isomorphisms are simple (even measured in $\Bar{K}_1$).
Given a complex, its $(\wo,m)${--}dual is formed by taking the
dual modules, converting them to modules of the same sideness as
the original using $\wo$
and altering the dual boundary maps by $(-1)^{\ast+m}$.
The double $(\wo,m)${--}dual of a complex is chain isomorphic to the
original complex.
The map on a particular module is just the isomorphism
$P\ \to\ P^{\ast\ast}$.
Now $\capf ch\colon \twistedC{\ast}{\wo}{m}(X,A)\ \to\
C_{m-\ast}(X,B)$
is a chain map.
If we $(\wo,m)${--}dualize, we get a map
$(\capf c h)^\ast\colon \twistedC{m-\ast}{\wo}{m}(X,B)\ \to\
\bigl(\twistedC{\ast}{\wo}{m}(X,A)\bigr)^{(\wo,m){-}{\text{dualized}}}$.
$\twistedC{\ast}{\wo}{m}$ $(\wo,m)${--}dualized is just
$C_{\ast}$, and $(\capf c h)^\ast=\capf c h$.
The involution is seen to induce an isomorphism $\wh^\ell(G)\ \to\ \wh(G)$,
and the composition
$\wh(G)\ \RA{\ \wh(\du)\ }\ \wh^\ell(G)\ \to\ \wh(G)$ is the map
induced by $\Z G\ \to\ \Z G$
via $\Bar{\barnone}$ (it is not hard to see that
this map induces a map on the Whitehead group level).
We will denote the map on $\wh(G)$ also by
$\Bar{\barnone}$.
If $\capf h c$ is a chain isomorphism, either from
$\twistedC{\ast}{\wo}{m}(X,A)\ \to\ C_{m-\ast}(X,B)$ or from
$\twistedC{m-\ast}{\wo}{m}(X,B)\ \to\ C_{\ast}(X,A)$, we can
compare the two torsions.
We get the confusing equation $\tau(\capf c h)=
(-1)^m\Bar{\tau(\capf c h) }$ where despite their similar appearance,
the two $\capf c h$'s are not the same (which is which is irrelevant).
We conclude by recording a notational convention.
We will sometimes have a map on homology such as $\cap c{}\colon
\Delta^\ast(M)\ \to\ \Delta_\ast(M)$.
If this map is an homology isomorphism we will often speak of the torsion
of $\capf c{}$ (or $\cupf f{}$, etc.).
By this we mean that there is a chain map, possibly after
twisting the cochain complex, (the chain map and twists
will be clear from context)
and these maps on the chain level are equivalences.
Note that by the usual nonsense, the torsions of these product maps
do not depend on a choice of cycle (or cocycle) within
the homology (cohomology) class.
Nor do they depend on lift functor or choice of paths.
They are dependent on the tree at this stage of our discussion, but
this too is largely fictitious.
A better proof of independence is given at the end of section 6.
Especially relevant for this last discussion are
\fullRef{T.2.1.2} and
the discussion of the Thom isomorphism theorem
in the appendix to Chapter 2.
\section{The realization of chain complexes}
\newHead{I.6}
In \cite{bthirtyseven} and \cite{bthirtyeight}, Wall discussed
the problem of constructing a CW complex whose chain complex
corresponds to a given chain complex.
We discuss this same problem for locally compact CW complexes.
Throughout this section, complex will mean a finite dimensional,
locally compact CW complex.
If we have a chain complex $A_\ast$, there are many conditions it
must satisfy if it is to be the chain complex of a complex.
Like Wall \cite{bthirtyeight} we are unable to find an algebraic
description of these conditions in low dimensions.
We escape the dilemma in much the same way.
\medskip
\begin{xDefinition}
A \emph{geometric chain complex} is a positive, finite,
chain complex $A_\ast$ together with a $2${--}complex $K$,
a tree $f\colon T\ \to\ K$, and a lift functor $F\in{\mathcal L}(f)$
such that
\begin{enumerate}
\item[1)] each $A_k$ is a locally{--}finitely generated free
$\Z\pi_1(K,F,f)${--}module
\item[2)] each $\partial_k\colon A_k\ \to\ A_{k-1}$ is a map
(not a map{--}germ)
\item[3)] in dimensions $\leq 2$, $C_\ast(K\Colon F)=A_\ast$.
\end{enumerate}
\end{xDefinition}
For 3) to make sense, we must define equality for two free tree modules.
If $A$ is free and based on $(S,\pi)$ and if $B$ is free and based on
$(R,\rho)$, $A=B$ \iff\ there exists a $1${--}$1$ map $\alpha\colon
S\leftrightarrow R$ such that $\alpha\ \circ\ \pi$ is equivalent to $\rho$.
One easily checks that this is an equivalence relation.
Notice that if $A_\ast$ is going to be the chain complex of some complex,
then all the above conditions are necessary.
Given two geometric chain complexes $A_\ast$ and $B_\ast$,
a map $f_\ast\colon A_\ast\ \to\ B_\ast$ is a map (not a germ) on
each $A_k$ and $\partial_k f_k = f_{k-1}\partial_k$ as maps.
\begin{xDefinition}
A map $f_\ast\colon A_\ast\ \to\ B_\ast$ between two geometric
chain complexes is \emph{admissible} provided
\begin{enumerate}
\item[1)] if $L$ is the $2${--}complex for $B_\ast$, $L=K$ wedged
with some $2${--}spheres in a locally finite fashion
\item[2)] $f_0$ and $f_1$ are the identity
\item[3)] $f_2$ is the identity on the $2${--}cells of $K$ and takes
any $2${--}sphere to its wedge point.
(The tree for $L$ is just the tree for $K$.
The lift functor for $L$ is just $g^{-1}\bigl($ lift functor for $K \bigr)$,
where $g\colon K\ \to\ L$ is the collapse map.)
\end{enumerate}
\end{xDefinition}
\begin{xRemarks}
It seems unlikely that we really need such strong conditions on a map
before we could handle it, but in our own constructions we usually get this,
and these assumptions save us much trouble.
\end{xRemarks}
\medskip
The chief geometric construction is the following.
\medskip
\BEGIN{T.1.6.1}
Let $X$ be a connected complex.
Let $A_\ast$ be a geometric chain complex with an admissible map
$f_\ast\colon A_\ast\to C_\ast(X)$ which is an equivalence.
Then we can construct a complex $Z$ and a proper, cellular map
$g\colon Z\ \to\ X$ so that $C_\ast(Z)=A_\ast$ and
\hskip2.5in$\xymatrix@C4pt{
A_\ast\ar[rr]^-{f_\ast}\ar@2@{-}[rd]&&C_\ast(X)\\
&C_\ast(Z)\ar[ru]_-{g_\ast}\\
}$
\noindent
commutes.
$g$ is a proper homotopy equivalence.
\end{Theorem}
\begin{proof}
We construct $Z$ skeleton by skeleton.
Since $f_\ast$ is admissible, $Z^2=X^2$ wedge $2${--}spheres.
$g_2\colon Z^2\ \to\ X$ is just the collapse map onto $X^2$.
To induct, assume we have an $r${--}dimensional complex
$Z^r$ and $g_r\colon Z^r\ \to\ X$ so that
$C_\ast(Z_r)=A_\ast$ in dimensions $\leq r$ and
$(g_r)_\ast=f_\ast$ in these dimensions.
If we can show how to get $Z^{r+1}$ and $g_{r+1}$ we are
done since $A_\ast$ is finite.
Now $A_{r+1}$ is free, so pick generators $\{e_i\}$.
We have a map $\partial\colon A_{r+1}\ \to\ A_r$
and $C_r(Z^r)=A_r$.
Hence each $\partial e_i$ is an $r${--}chain in $Z^r$.
We will show that these $r${--}chains are locally finite and spherical
(i.e. there is a locally finite collection of $r${--}spheres, and, after
subdivision, cellular maps $\cup S^r_i\ \to\ Z^r$ such that
$\partial e_i$ is homologous to $S^r_i$, and, if $h_i$ is an
$(r+1)$-chain giving the homology, the $\{h_i\}$
may be picked to be locally finite.)
We will then attach cells by these spheres and extend the map.
Let us now proceed more carefully.
For each $i$, $\partial e_i\in A_r$ and $\partial e_i\in (A_r)_{W_i}$
for some $W_i\in{\mathcal C}(T)$ with $\{ W_i\}$ cofinal in the subcategory
of ${\mathcal C}(T)$ consisting of all $A$ such that $\partial e_i\in (A_r)_A$.
Since $C_r(Z^r)=A_r$, $\partial e_i=c_i\in \bigl(C_r(Z^r)\bigr)_{B_i}$
for some $B_i\in {\mathcal C}(T)$ with $B_i\leq W_i$ ( we write
$B_i\leq W_i$ provided $B_i\subseteq W_i$ and $\{ B_i\}$ is cofinal in the
subcategory of all $A\in{\mathcal C}(T)$ for which $e_i\in (A_r)_A$.
$c_i$ is now a real geometric chain.
$\partial c_i=0$ since $\partial$ is actually a map.
Let $[c_i]$ be the homology class of $c_i$ in
$H_r\bigl(\coverFA{F_r(B_i)}\bigr)$, where $F_r$
is the lift functor for $Z^r$.
Now $g_\ast[c_i]=0$ in $H_r\bigl(\coverFA{F(U_i)}\bigr)$, where $F$ is
the lift functor for $X$ and $U_i\leq B_i$.
Hence there is an $f_i\in H_{r+1}\bigl(\coverFA{g_r}\colon
\coverFA{F_r(U_i)}\ \to\
\coverFA{F(U_i)}\bigr)$ with $f_i\mapsto [c_i]$.
But $g_r\colon Z^r\ \to\ X$ is properly $r${--}connected (it induces an
isomorphism of $\Delta(\quad\Colon \pi_1)$'s and $H^0_{{\text{end}}}$'s
by assumption, so it is always 1-1/2{--}connected.
Hence the universal covering functor for $X$
is a universal covering functor for $Z^r$, so $\Delta(\quad\Colon \pi_k)=0$ \iff\
$\Delta(M_{g_r},X\Colon H_k,\coverFA{}\ )=0$ for $k\leq r$ by the
Hurewicz theorem.
But $\Delta(M_{g_r},X\Colon H_k,\coverFA{}\ )=0$ for $k\leq r$
\iff\ $\Delta\bigl(H_k(g_r)\bigr)$ is an isomorphism for $k<r$ and
an epimorphism for $k=r$, which it is.)
Hence the Hurewicz theorem gives us elements
$s_i\in \pi_{r+1}\bigl( g_r\colon F_r(V_i)\ \to\ F(V_i)\bigr)$
where $V_i\leq U_i$ and $s_i$ hits the image of $f_i$ in
$H_{r+1}\bigl( \coverFA{g}_r\colon \coverFA{F_r(V_i)}\ \to\
\coverFA{F(V_i)}\bigr)$ under the Hurewicz map.
Let $Z^{r+1} = Z^r\ \cup$ a collection of $(r+1)${--}cells, $\{e_i\}$
attached by $s_i$.
$g_{r+1}\colon Z^{r+1}\ \to\ X$ is $g_r$ on $Z^r$.
Since $g_r\ \circ\ s_i\colon S^r\ \to\ Z^r\ \to\ X$ are properly
null homotopic, choose a locally finite collection $\{Q_i\}$
of null homotopies of $g_r\ \circ\ s_i$ to zero in $F(V_i)$.
$g_{r+1}\colon Z^{r+1}\ \to\ X$ is defined by $Q_i$ on each $e_i$.
$g_{r+1}$ is obviously still proper.
$C_\ast(Z^r)\ \to\ C_\ast(Z^{r+1})$ induces
an isomorphism for $\ast\leq r$.
$C_{r+1}(Z^{r+1})= A_{r+1}$ by taking the cell $e_i$ to the generator
$e_i$.
$F_{r+1}(B)= F_r(B)\ \cup\ ($ all cells $e_i$ for which the generator
$e_i$ lies in $B$ less those for which $g_{r+1}(e_i)\not\subseteq
F(B)$ ).
Then $g^{-1}_{r+1}\bigl(F(B)\bigr)\supseteq F_{r+1}(B)$.
Notice that if a cell $e$ does not attach totally in $F_r(B)$,
$g_{r+1}(e)\not\subseteq F(B)$, so $F_{r+1}(B)$ is a subcomplex.
$F_{r+1}(B)$ is cofinal in $B$, so $F_{r+1}$ is a lift functor.
Look at the chain map
$(g_{r+1})_\ast\colon C_{r+1}(Z^{r+1})\ \to\ C_{r+1}(X)$.
$e_i$ as a cell goes under $(g_{r+1})_\ast$ to the same element in
$\bigl(C_{r+1}(X)\bigr)_B$ as the generator $e_i$ does under $f_\ast$
for all $B\in{\mathcal C}(T)$ such that $e_i$ is a cell in $F_{r+1}(B)$.
Hence
\topD{12}{$\begin{matrix}%
A_{r+1}&\hbox to 0pt{\hss$\RA{\ f_{r+1}\ }$\hss}&\ C_{r+1}(X)\\
\parallel&\hskip 10pt\hbox to 0pt{$\nearrow\hbox to 0pt{$\scriptstyle g_{r+1}$\hss}$\hss}\\
\noalign{\vskip 4pt}
\hbox to 30pt{$C_{r+1}(Z^{r+1})$\hss}\\
\end{matrix}$}{4}
commutes.
\end{proof}
\begin{xDefinition}
A \emph{relative geometric chain complex} is a triple
$(A_\ast, K, L)$ consisting
of a finite, positive chain complex $A_\ast$
and a pair of complexes $(K,L)$.
Understood is a tree and a lift functor.
Then each $A_k$ is a locally{--}finitely generated free
$\Z\pi_1(K)$ tree module; each $\partial_k$ is a map; and in
dimensions $\leq2$, $A_\ast = C_\ast(K,L)$.
An \emph{admissible map} $f_\ast$ from
$(A_\ast, K, L)$ to $(B_\ast, K^\prime, L)$
is a chain map, not a germ,
$f_\ast\colon A_\ast\ \to\ B_\ast$ and $K=K^\prime$ wedge a locally
finite collection of $2${--}spheres.
$f_0$ and $f_1$ are the identity, and $f_2$ is the map induced by the
collapse $K\ \to\ K^\prime$.
\end{xDefinition}
\medskip
\BEGIN{C.1.6.1.1}
Let $(X,Y)$ be a pair of complexes, $X$ connected.
Let $A_\ast$ be a relative geometric chain complex with an admissible
map $f_\ast\colon A_\ast\ \to\ C_\ast(X,Y)$
which is an equivalence.
Then we can construct a complex $Z$ with $Y$ as a subcomplex and a
proper cellular map $g\colon Z\ \to\ X$ which is the identity on $Y$
such that $C_\ast(Z,Y)=A_\ast$ and
\topD{12}{$\begin{matrix}%
A_{\ast}&\hbox to 0pt{\hss$\RA{\ f_{\ast}\ }$\hss}&\ C_{\ast}(X,Y)\\
\parallel&\hskip 10pt\hbox to 0pt{$\nearrow\hbox{$\scriptstyle g_{\ast}$\hss}$\hss}\\
\noalign{\vskip 4pt}
\hbox to 30pt{$C_{\ast}(Z,Y)$\hss}\\
\end{matrix}$}{4}
commutes.
$g$ is a proper homotopy equivalence of pairs.
\end{Corollary}
\begin{proof}
The proof parallels the proof of \fullRef{T.1.6.1},
except that we must now use Namioka to show our elements
are spherical.
\end{proof}
Now let $f_\ast\colon A_\ast\ \to\ C_\ast(X)$ be an arbitrary
chain equivalence.
As in Wall \cite{bthirtyeight}, we would like to replace $A_\ast$
by an admissible complex with $f_\ast$ admissible while changing
$A_\ast$ as little as possible.
Look at
\[\begin{matrix}%
\cdots\to&A_3&\RA{\partial}&A_2&\to&A_1&\to& A_0&\to&A_{-1}
\to\cdots\to0\\
&\downlabeledarrow{f_3}{}&&\downlabeledarrow{f_2}{}&&\downarrow
&&\downarrow\\
\cdots\to&C_3&\to&C_2&\to&C_1&\to& A_0&\to0\\
\end{matrix}\]
One might like to try the complex
\[\begin{matrix}%
\cdots\to&A_3&\RA{f_2\ \circ\ \partial}&C_2&\to&C_1&\to& C_0&\to0\\
&\downlabeledarrow{f_3}{}&&\downlabeledarrow{\text{id}}{}&&
\downlabeledarrow{\text{id}}{}&&\downlabeledarrow{\text{id}}{}\\
\cdots\to&C_3&\to&C_2&\to&C_1&\to& C_0&\to0\quad .\\
\end{matrix}\]
The top complex is clearly admissible, but unfortunately the map
is no longer an equivalence.
The cycles in $A_3$ are now bigger with no new boundaries, and
the boundaries in $C_2$ are smaller with no fewer cycles.
Note first that $X$ is not of great importance.
If we replace $X$ by something in its proper homotopy class,
we will not be greatly concerned.
Let $X^\prime$ be $X$ with $2${--}spheres wedged on
to give a basis for $A_2$ and $3${--}cells attached to kill them.
Then $X^\prime$ has the same simple homotopy type as $X$,
$C_k(X^\prime)=C_k(X)$ except for $k=2$, $3$, and
$C_k(X^\prime)=C_k(X)\oplus A_2$ for $k=2$, $3$.
Let $f^\prime_k=f_k$, $k\neq 2$, $3$, and let $f^\prime_3=(f_3,\partial)$
and $f^\prime_2=(f_2, {\text{id}})$.
Then $A_\ast\ \RA{\ f^{\prime}_\ast\ }\ C_\ast(X^\prime)$
is still an equivalence and now $f^\prime_2$ is a monomorphism.
Let $A^\prime_\ast$ be the complex
\[\cdots\to\ A_3\ \RA{\ f^\prime_2\ \circ\ \partial\ }\
C_2^\prime=C_2(X^\prime)\ \to\ C^\prime_1\ \to\
C^\prime_0\ \to0\quad.\]
Then $h_\ast\colon A_\ast\ \to\ C^\prime_\ast$ has homology in only one
dimension:
$0\to\ H_2(h)\ \to\ H_2(A^\prime)\ \to\ H_2(C^\prime)\ \to0$.
Since $A_\ast\ \to\ A^\prime_\ast$ and since the composition
$A_\ast\ \to\ A^\prime_\ast\ \to\ C^\prime_\ast$ is an
equivalence, $H_2(A^\prime)=H_2(h)\oplus H_2(C^\prime)$.
Now by \fullRef{T.1.5.5}, $h_2(h)$ is s{--}free, provided
we can show $H^k(h)=0$ for $k\geq 3$.
But since we have a chain equivalence $A_\ast\ \to\ C^\prime_\ast$
we have a chain homotopy inverse in each dimension.
We then clearly get a chain homotopy inverse for
$A^\prime_k\ \to\ C^\prime_k$, $k\geq 4$, and
$h_3\ \circ\ g^\prime_3$ chain homotopic to ${\text{id}}_{{C_3}^\prime}$.
But this implies $H^k(h)=0$, $k\geq 3$.
Since $H_2(h)$ is projective, we get a map
$\rho=\partial\ \circ\ \rho^\prime$, where
$\rho^\prime\colon H_2(h)\ \to\ C^\prime_3$ is given as follows.
Both $A_3$ and $C^\prime_3$ map into $C_2=C^\prime_2$, and
$0\to\ {\Imx} A_3\ \to\ {\Imx} C_3\ \to\ H_2(h)\ \to0$
is exact.
Split this map by $\sigma\colon H_2(h)\ \to\ {\Imx} C_3^\prime$
and note $ {\Imx}\ A_3\ \cap\ {\Imx}\ \sigma = \{1\}$.
Now $C^\prime_3\ \to\ {\Imx}\ C^\prime_3\ \to0$ is exact,
so we can lift $\sigma$ to $\rho^\prime\colon H_2(h)\ \to\ C^\prime_3$.
Since $\sigma$ is a monomorphism, note
$ {\Imx}\ \rho^\prime\ \cap\ {\Imx}\ f_3 = \{1\}$.
Form $A^{\prime\prime}_\ast$ and $h^\prime_\ast$ by
\[\begin{matrix}%
\cdots\to&A_4&\to&A_3\oplus H_2(h)&
\rightlabeledarrow{\hskip 15pt \partial + \rho \hskip 15pt}{}&C^\prime_2&\to&C^\prime_1&\to
&C^\prime_0\to0\\
&\downlabeledarrow[\bigg]{f_4}{}&
&\downlabeledarrow[\bigg]{f_3 + \rho^\prime}{}
&&\downlabeledarrow[\bigg]{\text{id}}{}&&\downlabeledarrow[\bigg]{\text{id}}{}
&&\downlabeledarrow[\bigg]{\text{id}}{}\hfill
\\\noalign{\vskip4pt}
\cdots\to&C^\prime_4&\to&C^\prime_3&
\RA{\hskip 30pt}&
C^\prime_2&\to&C^\prime_1&\to
&C^\prime_0\to0\\
\end{matrix}\]
Note ${\kerx}(\partial+\rho)= ({\kerx}\ \partial, 0)$ since $\rho$
is a monomorphism and if $\rho(x)\in{\Imx}\ \partial$,
$\rho(x)=\{1\}$ as ${\Imx}\ A_3\ \cap\ {\Imx}\ \sigma=
\{1\}$.
Likewise note ${\Imx}(\partial+\rho)={\Imx}\ C^\prime_3$
since ${\Imx}\ A_3\ \oplus\ {\Imx}\ \rho=
{\Imx}\ C^\prime_3$.
Hence $h^\prime_\ast$ is an equivalence.
Note $\rho\colon H_2(h)\ \to\ C_2^\prime$ is a direct summand.
We split $\rho$ as follows.
\[\begin{matrix}%
\cdots\to&A_4&\to& A_3\oplus H_2(h)&
\rightlabeledarrow{\hskip 15pt \partial + \rho \hskip 15pt}{}& C^\prime_2&\to\cdots\\
&\uplabeledarrow[\big]{}{{\text{id}}}&&\uplabeledarrow[\big]{}{\alpha}
&&\uplabeledarrow[\big]{}{f_2}
\\\noalign{\vskip 4pt}
\cdots\to&A_4&\to& A_3&
\hbox to 32pt{\hss$\RA{\hskip 48pt}$}& A_2&\to\cdots\quad,\\
\end{matrix}\]
where $\alpha$ is inclusion on the first factor, commutes.
These maps must define a chain equivalence, so the dual situation
is also an equivalence.
\[\begin{matrix}%
\cdots\leftarrow&A^\ast_4&\leftlabeledarrow{\hskip 10pt \delta^3 \hskip 10pt}{}&
A^\ast_3\oplus \bigl(H_2(h)\bigr)^\ast&
\leftlabeledarrow{\hskip 20pt \partial^\ast+\rho^\ast\hskip 20pt}{}&(C^\prime_2)^\ast&
\leftarrow\cdots\\
&\bigg\downarrow&&\downlabeledarrow[\bigg]{\alpha^\ast}{}&&
\downlabeledarrow[\bigg]{f^\ast_2}{}
\\\noalign{\vskip 4pt}
\cdots\leftarrow&A^\ast_4&
\hbox to 22pt{$\leftlabeledarrow{\hskip 20pt \delta^3_A\hskip 20pt}{}$\hss}&
A^\ast_3&
\hbox to 42pt{\hss$\leftlabeledarrow{\hskip 30pt \delta^2_A\hskip 30pt}{}$}&A_2^\ast&
\leftarrow\cdots\\
\end{matrix}\]
${\kerx}\ \delta^3 = {\kerx}\ (\delta^3)_A \oplus
\bigl(H_2(h)\bigr)^\ast$, and ${\Imx}\ (\partial^\ast+\rho^\ast)=
{\Imx}\ \partial^\ast \oplus {\Imx}\ \rho^\ast$.
\[\begin{aligned}H_3({\text{Top\ complex}}) =&
{\kerx}\ \delta^3 / {\Imx}(\partial^\ast+\rho^\ast)\\
=&\Bigl({\kerx}\ (\delta_3)_A/{\Imx}\ \partial^\ast\Bigr)
\oplus
\Bigl(\bigl(H_2(h)\bigr)^\ast/{\Imx}\ \rho^\ast\Bigr)
\\\noalign{\medskip}
H_3({\text{Bottom\ complex}})=&
{\kerx}\ (\delta^3_A)/{\Imx}\ (\delta^2_A)\\
\end{aligned}\]
$H_3(\alpha)\colon H_3({\text{Top\ complex}})\ \to\
H_3({\text{Bottom\ complex}})$ is
\[\begin{matrix}%
{\kerx}\ (\delta^3_A)/{\Imx}\ \partial^\ast&&
{\kerx}\ (\delta^3_A)/{\Imx}\ (\delta^2_A)\\
\noalign{\vskip 6pt}
\oplus&\hbox to 20pt{\hss$\RA{\hskip 60pt}$\hss}&\oplus\\
\noalign{\vskip 6pt}
\bigl(H_2(h)\bigr)^\ast/{\Imx}\ \rho^\ast&&0\\
\end{matrix}\]
Hence, since $H_3(\alpha)$ is an isomorphism,
${\Imx}\ \partial^\ast = {\Imx}\ (\delta^2_A)$,
and $\rho^\ast\colon (C^\prime_2)^\ast\ \to\ \bigl(H_2(h)\bigr)^\ast$
is onto.
$\bigl(H_2(h)\bigr)^\ast$ is projective so split $\rho^\ast$.
Dualizing splits $\rho\colon H_2(h)\ \to\ C^\prime_2$.
$H_2(h)$ may not be free (it is only s{--}free).
$A_3\oplus H_2(h)$ is often free, but we prefer to keep $A_3$.
Hence form $A^s_\ast$ and $f^s_\ast$ by
\[\begin{matrix}%
\cdots\to&A_4&\to&
A_3\oplus \bigl(H_2(h)_S\oplus F^{(n)}\bigr)&\to&
C^\prime_2\oplus F^{(n)}&\to\cdots\\
&\downlabeledarrow[\Big]{f_4}{}&&
\downlabeledarrow[\Big]{f_3+(\rho^\prime_S+0)}{}&&
\downlabeledarrow[\Big]{{\text{id}}+0}{}\\
\cdots\to&C^\prime_4&\hbox to 0pt{$\RA{\hskip 40pt}$\hss}&
C^\prime_3&
\hbox to 10pt{\hss\hbox to 50pt{\hss$\RA{\hskip 70pt}$}\hss}&
C^\prime_2&\to\cdots\\
\end{matrix}\]
where $S$ is a shift functor so that the map germs $\rho$
and $\rho^\prime$ are actual maps.
By wedging on $n$ $2${--}spheres at each vertex of the tree, we
see $A^s_\ast$ and $f^s_\ast$ are admissible.
Notice that exactly the same procedure makes a map
$f\ast\colon A_\ast\ \to\ C_\ast(X,Y)$ admissible.
In section 3, \fullRef{P.1.3.3} we defined what it
meant by $X$ satisfies $D n$.
We briefly digress to prove
\medski
\BEGIN{T.1.6.2}
The following are equivalent for $n\geq 2$, $X$ a complex
\begin{enumerate}
\item[1)] $X$ satisfies $D n$
\item[2)] $X$ is properly dominated by an $n${--}complex
\item[3)] $\Delta^k(X\Colon {\text{universal\ covering\ functor}})=0$
for $k>n$.
\end{enumerate}
\end{Theorem}
\begin{proof}
1) implies 2) as $X^n\subseteq X$ is properly $n${--}connected and
hence dominates $X$ if $X$ satisfies $D n$.
2) implies 3) by computing $\Delta^k$ from the cellular chain complex
of the dominating complex.
3) implies 2): Since $\Delta^k(X\Colon \quad)=0$ for $k>n$, by
\fullRef{T.1.5.4}\ dualized to cohomology, we
get chain retracts
\[0\to\ C_r\
\crA\
C_{r-1}\
\crA\
\cdots\cdots\cdots\ \crA\
C_{n+1}\
\crA\
C_n\
\crA\
\cdots
\]
where $r={\text{dim}}\ X<\infty$.
By an induction argument, ${\Imx}\ \partial_{n+1}$ is s{--}free,
and $C_n={\Imx}\ \partial_{n+1}\oplus A_n$
(dualize everything to get these results in the cochain complex and then
dualize back).
$A_n$ is s{--}free and
\[\begin{matrix}%
&0&\to&A_n&\to&C_{n-1}&\to\ \cdots\ \to&C_0&\to 0\\
&&&\hbox to 0pt{\hss$\scriptstyle r$}\downarrow&&\downlabeledarrow{{\text{id}}}{}&&
\downlabeledarrow{{\text{id}}}{}\\
\cdots \to&C_{n+1}&\to&C_n&\to&C_{n-1}&\to\ \cdots\ \to&
C_0&\to 0\\
\end{matrix}\]
gives us an $n${--}dimensional chain complex and a chain equivalence.
$A_n$ is only s{--}free, so form
$0\to\ A_n\oplus F^{(m)}\ \to\ C_{n-1}\oplus F^{(m)}\ \to\ \cdots$
which is now a free complex.
If $n\geq 3$, the complex and the map are clearly admissible, so by
\fullRef{T.1.6.1} we get an $n${--}complex $Y$
and a proper homotopy equivalence $g\colon Y\ \to\ X$, so $X$
satisfies 2).
If $n=2$, $X$ has the proper homotopy type of a $3${--}complex by
the above, so we assume $X$ is a $3${--}complex.
Its chain complex is then
$0\to\ C_3\ \to\ C_2\ \to\ C_1\ \to\ C_0\ \to0$
with $H^3(C)=0$.
Wedge $2${--}spheres to $X$ at the vertices of the tree to get a
chain complex
$0\to\ C_3\ \to\ C_2\oplus C_3\ \to\ C_1\ \to\ C_0\ \to0$.
Since $H^3(C)=0$, $C_2=C_3\oplus M$.
Let $j\colon C_3 \to\ C_2$ be the inclusion.
Then we have
\[\begin{matrix}%
\hbox to 0pt{\hss{\text{A:}}\hskip 40pt}
&&0\to&C_3&\RA{j}&C_2\oplus C_3&\to&C_1&\to&C_0\to0\\
\noalign{\vskip 4pt}
&&&\Big\uparrow&&\uplabeledarrow[\Big]{}{r}&&\uplabeledarrow[\Big]{}{{\text{id}}}
&&\uplabeledarrow[\Big]{}{{\text{id}}}\\
\noalign{\vskip 6pt}
\hbox to 0pt{\hss{\text{B:}}\hskip 40pt}
&&&0&\to&C_2&\to&C_1&\to&C_0\to0\\
\end{matrix}\]
where $r\colon ( C_3\oplus M)\ \to\
( C_3\oplus M)\oplus C_3$
is given by $r(x,y)=(0,y,x)$.
This is a chain equivalence between B and A.
Both A and B are the chain complexes for a space (A for $X\vee S^2$'s
and B for the $2${--}skeleton of $X$).
The chain map is easily realized on the $1${--}skeleton as a map,
and we show we can find a map
$g\colon X^2\ \to\ X\vee_j S^2_j$ realizing the whole chain map.
Let $\{ e_i\}$ be the two cells of $X^2$.
Their attaching maps determine an element in
$\Delta( X^1\Colon \pi_1, \coverFA{}\ )$, where this group denotes the
$\Delta${--}construction applied to the groups
$\pi_1\bigl(p^{-1}(X^1-C),\hat x_i\bigr)$, where
$p\colon \coverFA{X}\ \to X$ is the projection for the universal cover
of $X$. (i.e. $\coverFA{}$\quad denotes the covering functor
over $X^1$ induced in the above manner form the universal covering
functor on $X$.)
Let $g_1\colon X^1\ \to\ X\vee_j S^2_j$ be the natural inclusion.
As in the proof of \fullRef{T.1.6.1}, the $\{e_i\}$
determine an element of $\Delta(g_1\Colon H_2, \coverFA{}\ )$.
Our two elements agree in $\Delta( X^1\Colon H_1, \coverFA{}\ )$.
The following diagram commutes and the rows are exact
\[\begin{matrix}%
&&\Delta(X^2\Colon \pi_2,\coverFA{}\ )&\zto&\Delta(g_1\Colon \pi_2,\coverFA{}\ )
&\zto&\Delta(X^1\Colon \pi_1,\coverFA{}\ )&\zto&
{\scriptstyle\Delta(X^2\Colon \pi_1,\coverFA{}\ )=}0\\
&&\downlabeledarrow[\Big]{h}{}&&\Big\downarrow&&\Big\downarrow\\
0={\scriptstyle\Delta(X^1\Colon H_2,\coverFA{}\ )}&\zto&
\Delta(X^2\Colon H_2,\coverFA{}\ )&\zto&\Delta(g_1\Colon H_2,\coverFA{}\ )
&\zto&\Delta(X^1\Colon H_1,\coverFA{}\ )\\
\end{matrix}\]
where $\Delta(X^1\Colon H_1,\coverFA{}\ )$ and
$\Delta(X^1\Colon H_2,\coverFA{}\ )$ are defined similarly to
$\Delta(X^1\Colon \pi_1,\coverFA{}\ )$.
$X^1\subseteq X^2$ is properly $1${--}connected, so the subspace
groups are the groups asserted.
$h$ is an isomorphism by the Hurewicz theorem, so a diagram chase
yields a unique element in $\Delta(g_1\Colon \pi_2,\coverFA{}\ )$
which hits our elements in both $\Delta(X^1\Colon \pi_1,\coverFA{}\ )$
and $\Delta(g_1\Colon H_2,\coverFA{}\ )$.
Use this element to extend the map to
$g_2\colon X^2\ \to\ X\vee_j S^2_j$.
By our choices, $g_2$ induces an isomorphism of $\Delta(\pi_1)$'s.
Hence $g_2$ is a proper homotopy equivalence.
This 3) implies 2) for $n\geq 2$.
2) implies 1) is trivial.
\end{proof}
\medskip
\BEGIN{C.1.6.2.1}
If $X$ satisfies $D n$ for $n\geq 3$, $X$ has the proper homotopy
type of an $n${-}complex. \mathqed\end{Corollary}
\medskip
Combining our admissibility construction with
\fullRef{T.1.6.1} gives
\bigskip
\BEGIN{T.1.6.3}
Let $f_\ast\colon A_\ast\ \to\ C_\ast(X)$ be a chain equivalence
for a complex $X$ ($A_\ast$ free, finite and positive).
Then there exists a complex $Y_0$ satisfying $D2$; a complex
$Y\supseteq Y_0$ such that $C_\ast(Y,Y_0)=A_\ast$
in dimensions greater than or equal to $3$; and a proper,
cellular homotopy equivalence $g\colon Y\ \to\ X$ such that
$g_\ast=f_\ast$ in dimensions greater than or equal to $4$.
The torsion of $g$ may have any preassigned value.
\end{Theorem}
\begin{proof}
Make $A_\ast$, $f_\ast$ admissible.
The new complex is
$\cdots\to\ A_4\ \to\ A_3\oplus (?)\ \to\
C_2\oplus(?)\ \to\ X_1\ \to\cdots$.
Construct a $Y$ from this complex as in \fullRef{T.1.6.1}.
When we pick a basis for $A_3\oplus (?)$, pick a basis for $A_3$
and one for $(?)$ and use their union.
Then there is a subcomplex $Y_0\subseteq Y$ whose chain complex is
$0\to\ (?)\ \to\ C_2\oplus (?)\ \to\ C_1\ \to\ C_0\ \to0$
The first $(?)$ is $H_2(h)\oplus F^{(m)}$.
It is not hard to show $\Delta^3(Y_0\Colon \coverFA{}\ )=0$,
so $Y_0$ satisfies $D2$.
The remainder of the theorem is trivial except for the remark about torsion.
But for some $m\geq 0$, we can realize a given torsion by an
automorphism $\alpha\colon F^{(m)}\to\ F^{(m)}$.
Hence by altering the basis in $F^{(m)}$ we can cause our map to
have any desired torsion (we may have to take $m$ bigger,
although in the infinite case $m=1$ will realize all torsions).
\end{proof}
\bigski
\BEGIN{T.1.6.4}
Let $f_\ast\colon A_\ast\ \to\ C_\ast(X,Z)$ be a chain map for a pair
$(X,Z)$ ($A_\ast$ free, finite and positive).
Then there exist complexes $Y_0$ and $Y$ such that
$Y\supseteq Y_0\supseteq Z$; $C_\ast(Y,Y_0)=A_\ast$
in dimensions greater than or equal to 3; a proper cellular homotopy
equivalence $g\colon Y\ \to\ X$ which is the identity on $Z$ such
that $g_\ast=f_\ast$ in dimensions greater than or equal to $4$.
The torsion of $g$ may have any preassigned value.
\end{Theorem}
\begin{proof} Use \fullRef{C.1.6.1.1}.
\end{proof}
\bigskip
We conclude this chapter by returning briefly to the question of
the invariance of torsion for chain maps under change of trees.
The natural map $\wh(X\Colon f)\ \to\ \wh(X\Colon g)$ is a homomorphism,
so the property of being a simple chain equivalence is
independent of the tree.
But now use \fullRef{T.1.6.3} to get a proper
homotopy equivalence $X\ \to\ X_\tau$ with torsion $\tau$.
Suppose given a chain map, say for example, $\capf c{}\colon
\Delta^\ast(X)\ \to\ \Delta_{m-\ast}(X)$ with torsion $-\tau$.
Then the composition $\Delta^\ast(X)\ \RA{\ \capf c{}\ }\
\Delta_{m-\ast}(X)\ \to\ \Delta_{m-\ast}(X_\tau)$ is simple, so a change
of trees leaves it simple.
But the second map is a proper homotopy equivalence of spaces,
and so is preserved by our change of tree map.
Hence so must be the torsion of $\capf c{}$.
(Note we are using our convention of writing chain equivalences on
the homology level.)
\chapter{Poincar\'e Duality Spaces}
\section{Introduction, definitions and elementary properties}
\newHead{II.1}
\bigskip
In this chapter we discuss the analogue of manifold in the proper homotopy
category.
We seek objects, to be called Poincar\'e duality spaces, which have the
proper homotopy attributes of paracompact manifolds.
To this end, we begin by discussing these attributes.
There is a well known Lefschetz duality between $H_\ast$ and
$H^\ast_\cmpsup$ or between $H^\ast$ and $H_\ast^\locf$
which is valid for any paracompact manifold with boundary (see for
instance Wilder \cite{bfortyfour}).
This duality is given via the cap product with a generator of
$H^\locf_N$,
perhaps with twisted coefficients.
This generator is called the {\sl fundamental class}.
Given any paracompact handlebody $M$, $M$ can be covered by an
increasing sequence of compact submanifolds with boundary.
Let $\{C_i\}$ be such a collection.
If $[M]\in H^\locf_{\dim M}(M;\Z^t)$ is the fundamental class, its
image in $H^\locf_{\dim M}(\Bar{M-C_i},\partial C_i; \Z^t)$ via the
inclusion and excision is the fundamental class for the pair
$(\Bar{M-C_i},\partial C_i)$.
A word about notation: $\Z^t$ occurring as a coefficient group will
always denote coefficients twisted by the first Stiefel{--}Whitney
class of the manifold.
\bigskip
\BEGIN{T.2.1.1}
The fundamental class $[M]$ in $H^\locf_N(M;\Z^t)$ induces via
the cap product an isomorphism
\[\capf {[M]}{}\colon \Delta^{N-\ast}(M\Colon \coverFA{})\ \to\
\Delta_\ast(M\Colon \coverFA{})\]
where \ $\coverFA{}$\quad is any covering functor.
If $M$ has a boundary, we get a fundamental class
$[M]\in H^\locf_N(M, \partial M;\Z^t)$ and isomorphisms
\[\begin{matrix}%
\capf {[M]}{}\colon& \Delta^{N-\ast}(M,\partial M\Colon \coverFA{})&\to&
\Delta_\ast(M\Colon \coverFA{})\hfill\cr
\capf {[M]}{}\colon& \Delta^{N-\ast}(M\Colon \coverFA{})\hfill&
\hbox to10pt{\hss$\RA{\hskip 25pt}$}&
\Delta_\ast(M,\partial M\Colon \coverFA{})\cr
\end{matrix}\]
A similar result holds for a manifold $n${--}ad.
\end{Theorem}
\begin{proof}
The proof is easy.
On the cofinal subset of compact submanifolds with boundary of $M$,
$[M]$ induces, via inclusion and excision, the fundamental class
for the pair ( $(n+1)${--}ad in general) $(\Bar{M-C_i}, \partial C_i)$,
where $C_i$ is a compact submanifold with boundary of $M$,
and $(\Bar{M-C_i}$ is the closure in $M-C_i$ in $M$.
$\partial C_i$ is equally the boundary of $C_i$ as a manifold or the
frontier of $C_i$ as a set.
By the definition of $\capf{[M]}{}$, it induces an isomorphism for each
base point and set $C_i$.
Hence it must on the inverse limit.
\end{proof}
\medskip
If one computes the homology and cohomology from chain complexes
based on a PL triangulation, on a handlebody decomposition, or on
a triangulation of the normal disc bundle, $\capf{[M]}{}$ induces a
chain isomorphism.
We can ask for the torsion of this map.
We have
\BEGIN{T.2.1.2}
If $(M,\partial M)$ is a manifold with (possibly empty) boundary, and if
$[M]\in H^\locf_N(M,\partial M;\Z^t)$ is the fundamental class,
$\capf{[M]}{}\colon \Delta^{N-\ast}(M,\partial M\Colon \coverFA{} )\ \to\
\Delta_\ast(M\Colon \coverFA{})$ and
$\capf{[M]}{}\colon \Delta^{N-\ast}(M\Colon \coverFA{})\ \to\
\Delta_\ast(M,\partial M\Colon \coverFA{})$
are simple equivalences, where \ $\coverFA{}$\quad
is the universal covering functor.
\end{Theorem}
\begin{proof}
Given a handlebody decomposition, the proof is easy.
The cap product with the fundamental class takes the cochain which
is $1$ on a given handle and zero on all the other handles to the
dual of the given handle.
Hence $\capf{[M]}{}$ takes generators in cohomology to generators
in homology (up to translation by the fundamental group).
The fact that the simple homotopy type as defined by a PL triangulation
or by a triangulation of the normal disc bundle is the same as that defined
by a handlebody has been shown by Siebenmann \cite{bthirtyfour}.
\end{proof}
\medskip
We are still left with manifolds which have no handlebody decomposition.
Let $N=\CP^4\ \#\ S^3\times S^5\ \#\ S^3\times S^5$.
Then $\chi(N)=1$. $N\times M$ has $[N]\times [M]$
as a fundamental class.
For $M$ we use the simple homotopy type defined by a triangulation
of the normal disc bundle.
Then $\capf{[N]\times[M]}{}$ is a simple homotopy equivalence \iff\
$\capf{[M]}{}$ is by \fullRef{L.1.5.23},
since $\capf{[N]}{}$ is known to induce a simple equivalence.
But $\capf{[N]\times[M]}{}$ is a simple equivalence since $N\times M$
has a handlebody structure (Kirby{--}Siebenmann \cite{beighteen}).
Note Theorems \shortFullRef{T.2.1.1}
and \shortFullRef{T.2.1.2} now hold for arbitrary
paracompact manifolds.
With these two theorems in mind, we make the following definition.
\medskip\begin{xDefinition}
A locally finite, finite dimensional CW pair $(X,\partial X)$ with
orientation class $\wone\in H^1(X;\cy2)$
is said to satisfy
Poincar\'e duality with respect to $[X]$ and the covering functor
\ $\coverFA{}$\quad provided there is a class
$[X]\in H^\locf_N(X,\partial X; \Z^\wone)$ such that the maps
\[\begin{matrix}%
\capf{[X]}{}\colon &\Delta^{N-\ast}(X\Colon \coverFA{})\hfill&\to&
\Delta_\ast(X,\partial X\Colon \coverFA{})\cr\noalign{\vskip6pt}
\capf{[X]}{}\colon &
\Delta^{N-\ast}(X,\partial X\Colon \coverFA{} )&\to&
\Delta_\ast(X\Colon \coverFA{})\hfill\cr
\end{matrix}\]
are isomorphisms.
$\Z^\wone$ denotes integer coefficients twisted by the class $\wone$.
\end{xDefinition}
If $X$ is an $n${--}ad we require that all the duality maps be isomorphisms.
\bigskip\begin{xRemarks}
The two maps above are dual to one another, so if one is an isomorphism
the other is also.
\end{xRemarks}
Suppose \ $\coverFA{}$\quad is a regular covering functor for $X$,
and suppose \ $\coverFC{}$\quad is another regular
covering functor with $\coverFC{} > \coverFA{}$.
Then the chain and cochain groups have the structure of
$\Z\pi^\prime_1(X\Colon F,f\Colon \coverFA{})${--}modules, when
$f\colon T\ \to\ X$ is a tree and $F\in{\mathcal L}(f)$.
The tree of groups $\pi^\prime_1(X\Colon F,f\Colon \coverFA{})$
is the tree given by
$(\pi^\prime_1)_A=\pi_1\bigl(F(A),p\bigr)/\pi_1\Bigl(\coverFA{F(A)},p\Bigr)$
where $p$ is the minimal vertex for $A$.
There is a map of rings $\Z\pi^\prime_1(X\Colon F,f\Colon \coverFA{} )\ \to\
\Z\pi^\prime_1(X\Colon F,f\Colon \coverFC{} )$, and the tensor product takes
$\Delta(X\Colon \coverFA{})$ to $\Delta(X\Colon \coverFC{})$.
Since $\capf{[X]}{}$ is an isomorphism for $\coverFA{}$,
we can get chain homotopy inverses, so under tensor product,
$\capf{[X]}{}$ still induces isomorphisms for $\coverFC{}$.
\medskip
As we have the Browder Lemma (\fullRef{T.1.4.7}),
with patience we can prove a variety of cutting and gluing theorems.
The following are typical.
\bigskip
\BEGIN{T.2.1.3
Let $(X\Colon \partial_0X,\partial_1X)$ be a triad.
Then the following are equivalent.
\begin{enumerate}
\item[1)] $(X\Colon \partial_0X,\partial_1X)$ satisfies Poincar\'e duality with respect
to \hfil\penalty-10000
\hbox{$V\in H^\locf_N(X\Colon \partial_0X,\partial_1X\Colon \Z^\wone)$ and
$\coverFA{}$}.
\item[2)] $(\partial_0X,\partial_{\{0,1\}}X)$ satisfies Poincar\'e duality
with respect to\hfil\penalty-10000
\hbox{$\partial V\in H^\locf_{N-1}(\partial_0X\Colon
\partial_{\{0,1\}}X\Colon \Z^\wone)$ and
$\coverFA{}$ where $\coverFA{}$ is induced from}
$\coverFA{}$ over $X$\hfil\penalty-1000 and $\wone$ is the orientation
class induced from $\wone$ over $X$.
Moreover, one of the maps
\[\begin{matrix}%
\capf{V}{}\colon&\Delta^\ast(X,\partial_1X\Colon \coverFA{})&
\to&\Delta_{N-\ast}(X,\partial X_0\Colon \coverFA{})\cr
\noalign{\vskip6pt}
\capf{V}{}\colon&\Delta^\ast(X,\partial_0X\Colon \coverFA{})&
\to&\Delta_{N-\ast}(X,\partial X_1\Colon \coverFA{})\cr
\end{matrix}\]
is an isomorphism. (Hence they are both isomorphisms.)
\item[3)] The same conditions as 2) but considering
$(\partial_1X,\partial_{\{0,1\}}X)$.
\end{enumerate}
\end{Theorem}
\begin{proof}
The proof is fairly standard.
We look at one of the sequences associated to a triple, say
\[\begin{matrix}%
\Delta^\ast(X\Colon \partial_0X,\partial_1X\Colon \coverFA{})&\to&
\Delta^\ast(X,\partial_1X\Colon \coverFA{})&\to&
\Delta^\ast(\partial_0X\Colon \partial_{\{0,1\}}X\Colon X\Colon \coverFA{})\cr
\downlabeledarrow[\Big]{}{\capf{V}{}}&&\downlabeledarrow[\Big]{}{\capf{V}{}}&&
\downlabeledarrow[\Big]{}{\capf{\partial V}{}}\cr\noalign{\vskip4pt}
\Delta_{N-\ast}(X\Colon \coverFA{})&
\hbox to 10pt{\hss$\RA{\hskip 26pt}$}&
\Delta_{N-\ast}(X,\partial_0X\Colon \coverFA{})&\to&
\Delta_{N-1-\ast}(\partial_0X\Colon X\Colon \coverFA{})\cr
\end{matrix}\]
1) implies both $\capf{V}{}$'s are isomorphisms, so the $5${--}lemma
shows $\capf{\partial V}{}$ is an isomorphism.
2) implies one of the $\capf{V}{}$'s is an isomorphism and that
\[\capf{\partial V}{}\colon \Delta^\ast(\partial_0X,\partial_{\{0,1\}}X\Colon
\coverFA{})\ \to\ \Delta_{N-1-\ast}(\partial_0X\Colon \coverFA{})\]
is an isomorphism.
Hence we must investigate how the subspace groups depend
on the absolute groups.
Make sure that the set of base points for $X$ contains
a set for $\partial_0X$.
Then we have a diagram
\[\begin{matrix}%
\Delta^\ast(\partial_0X,\partial_{\{0,1\}}X\Colon \coverFA{})&
\to&\Delta^\ast(\partial_0X,\partial_{\{0,1\}}X\Colon X\Colon
\coverFA{})\cr\noalign{\vskip6pt}
\downlabeledarrow[\big]{}{\cap\partial V}&&\downlabeledarrow[\big]{}{\cap\partial V}\cr
\noalign{\vskip6pt}
\Delta_{N-1-\ast}(\partial_0X \Colon \coverFA{})&
\to&\Delta_{N-1-\ast}(\partial_0X\Colon X \Colon \coverFA{})\cr
\end{matrix}\]
which commutes.
The horizontal maps are naturally split, so if
$\capf{\partial V}{}$ on the subspace groups is an isomorphism,
then it is also an isomorphism on the absolute groups.
Hence 1) implies 2) and 3).
Now if $\capf{\partial V}{}$ on the absolute groups is an isomorphism,
then it is also an isomorphism on the subspace groups by
\fullRef{T.2.1.3}.
Hence 2) or 3) implies 1).
\end{proof}
\bigskip
\BEGIN{T.2.1.4
Let $Z= Y\ \cup\ Y^\prime$ and set $X = Y\ \cap\ Y^\prime$.
Then any two of the following imply the third.
\begin{enumerate}
\item[1)] $Z$ satisfies Poincar\'e duality with respect to $[Z]$ and
$\coverFA{}$.
\item[2)] $(Y,X)$ satisfies Poincar\'e duality with respect to $\partial[Z]$
and $\coverFA{}$.
\item[3)] $(Y^\prime,X)$ satisfies Poincar\'e duality with respect to
$\partial[Z]$ and $\coverFA{}$.
\end{enumerate}
where $\coverFA{}$ is a covering functor over $Z$, which
then induces $\coverFA{}$ over $Y$ and $Y^\prime$.
An orientation class over $Z$ which induces one over $Y$ and
$Y^\prime$ has been assumed in our statements.
\end{Theorem}
\begin{proof}
The reader should have no trouble proving this.
\end{proof}
\medskip
A map $\varphi\colon M\ \to\ X$, where $M$ and $X$ are locally compact
CW $n${--}ads which satisfy Poincar\'e duality with respect to $[M]$ and
$\coverFA{}$, and $[X]$ and $\coverFC{}$
respectively, is said to be {\sl degree $1$\/} provided it is a map
of $n${--}ads and
\begin{enumerate}
\item[1)] $\varphi^\ast(\coverFC{})=\coverFA{}$,
where $\varphi^\ast(\coverFC{})$ is the covering functor
over $M$ induced by $\varphi$ from $\coverFC{}$
over $X$.
\item[2)] If $\wone\in H^1(X;\cy2)$ is the orientation class for $X$,
$\varphi^\ast\wone$ is the orientation class for $M$.
\item[3)] $\varphi_\ast[M]=[X]$.
\end{enumerate}
\bigski
\BEGIN{T.2.1.5}
Let $\varphi\colon M\ \to\ X$ be a map of degree $1$
of Poincar\'e duality spaces.
Then the diagram
\[\begin{matrix}%
\Delta^r(M\Colon \coverFA{})&\LA{\ \varphi^\ast\ }&
\Delta^r(X\Colon \coverFA{})\cr
\downlabeledarrow[\Big]{}{\capf{[M]}{}}&&\downlabeledarrow[\Big]{}{\capf{[X]}{}}\cr
\noalign{\vskip6pt}
\Delta_{n-r}(M\Colon \coverFA{})&\RA{\ \varphi_\ast\ }&
\Delta_{n-r}(X\Colon \coverFA{})\cr
\end{matrix}\]
commutes. ($\coverFA{}$ over $M$ is the covering functor
induced from $\coverFA{}$ over $X$.)
Hence $\capf{[M]}{}$ induces an isomorphism on the cokernel of
$\varphi^\ast$, $K^r(M\Colon \coverFA{})$ onto the kernel of
$\varphi_\ast$, $K_{n-r}(M\Colon \coverFA{})$.
Thus if $\varphi$ is $k${--}connected, $\varphi_\ast$ and $\varphi^\ast$
are isomorphism for $r<k$ and $r>n-k$.
Similarly let $\varphi\colon (N,M)\ \to\ (Y,X)$
be a degree $1$ map of pairs.
Then $\varphi_\ast$ gives split surjections of homology groups with
kernels $K_\ast$, and split injections of cohomology with cokernels
$K^\ast$.
The duality map $\capf{[N]}{}$ induces isomorphisms
$K^\ast(N\Colon \coverFA{})\ \to\
K_{n-\ast}(N,M\Colon\coverFA{})$ and
$K^\ast(N, M\Colon \coverFA{})\ \to\
K_{n-\ast}(N\Colon \coverFA{})$.
Analogous results hold for $n${--}ads.
\end{Theorem}
\begin{proof}
The results follow easily from definitions and the naturality
of the cap product.
\end{proof}
\bigskip
\section{The Spivak normal fibration}
\newHead{II.2}
One important attribute of paracompact manifolds is
the existence of normal bundles.
In \cite{bthirtysix} Spivak constructed an analogue for these bundles
in the homotopy category.
Although he was interested in compact spaces, he was often forced
to consider paracompact ones.
It is then not too surprising that his definition is perfectly adequate for
our problem.
This is an example of a general principle in the theory of paracompact
surgery, namely that all bundle problems encountered are exactly the
same as in the compact case.
One does not need a ``proper'' normal bundle or a ``proper''
Spivak fibration.
\bigskip\begin{xDefinition}
Let $(X,\partial X)$ be a locally compact, finite dimensional CW pair.
Embed $(X,\partial X)$ in $(\HS^n,\R^{n-1})$, where $\HS^n$
is the upper half plane and $\R^{n-1}=\partial\HS^n$.
Let $(N;N_1,N_2)$ be a regular neighborhood of $X$ as a subcomplex
of $\HS^n$; i.e. $X\subseteq N$, $\partial X\subseteq N_2$ and
$N$ (resp. $N_2$) collapses to $X$ (resp. $\partial X$).
Let ${\mathcal P}(N_1,N,X)$ be the space of paths starting in $N_1$,
lying in $N$, and ending in $X$ endowed with the compact-open topology.
(If $A$, $B$, $C$ are spaces with $A$, $C\subseteq B$, a
similar definition holds for ${\mathcal P}(A,B,C)$.)
There is the endpoint map $w\colon {\mathcal P}(N_1,N,X)\ \to\ X$.
$w$ is a fibration and is called the {\sl Spivak normal fibration}.
Its fibre is called the {\sl Spivak normal fibre}.
\end{xDefinition}
Spivak showed that a necessary and sufficient condition for a finite complex
to satisfy Poincar\'e duality with respect to the universal covering functor
was that the Spivak normal fibre of the complex should have the
homotopy type of a sphere.
He also showed that if one started with a compact manifold, then the
normal sphere bundle had the same homotopy type as the Spivak
normal fibration, at least stably.
Before we can do this for paracompact manifolds, we will need to do some
work.
In practice, the fact that the Spivak normal fibration is constructed from
a regular neighborhood is inconvenient.
More convenient for our purposes is a {\sl semi{--}regular neighborhood}
\begin{xDefinition}
Let $(X,\partial X)$ be a pair of finite dimensional, locally compact
CW complexes.
A {\sl semi{--}regular neighborhood\/} ( s-r neighborhood) is a manifold
triad $(M\Colon M_1, M_2)$ and proper maps
$i\colon X\ \to\ M$ and $j\colon \partial X\ \to\ M_2$
such that
\topD{10}{$\begin{matrix}%
X&\RA{\ i\ }&M\cr
\cup\vrule width .1pt depth 0pt height 6pt&&
\cup\vrule width .1pt depth 0pt height 6pt\cr
\partial X&\RA{\ j\ }&M_2\cr
\end{matrix}$}{4}
commutes, and such that
$i$ and $j$ are simple homotopy equivalences.
Lastly we require that $M$ be parallelizable (equivalent to being
stably parallelizable).
The definition for an $n${--}ad is similar: we have a manifold $(n+1)${--}ad
$(M\Colon M_1,\cdots M_{n})$ with a simple homotopy equivalence of $n${--}ads
$i\colon X\to \delta_1 M$ .
\end{xDefinition}
\bigski
\BEGIN{T.2.2.1}
The fibration $w\colon{\mathcal P}(M_1,M,X)\ \to\ X$ is stably fibre
homotopy equivalent to the Spivak normal fibration.
\end{Theorem}
\begin{proof}
The proof needs
\medski
\BEGIN{L.2.2.1}
If $\bigl((M\Colon M_1,M_2), i, j \bigr)$ is an s-r neighborhood of $X$, then so is
\[\bigl((M\Colon M_1,M_2)\times(D^m,S^{m-1}), i\times c, j_1\bigr)\]
where $c$ denotes the constant map.
The $(n+1)$ structure on the product is $(M\times D^m\Colon
M_1\times D^m\ \cup\ M\times S^{m-1},
M_2\times D^m,\cdots)$.
Let $\xi$ be ${\mathcal P}(M_1, M,X)\ \to\ X$ and let $\eta$ be
${\mathcal P}(M_1\times D^m\ \cup\ M\times S^{m-1},M\times D^m)\ \to\ X$.
Then $\xi\ast(m)$ is fibre homotopy equivalent to $\eta$, where $(m)$ is
the trivial spherical fibration of dimension $m-1$ and $\ast$ denotes the
fibre join.
\end{Lemma}
\begin{proof}
The first statement is trivial and the second is Spivak \cite{bthirtysix},
Lemma 4.3.
\end{proof}
\smallskip
Now if $(M\Colon M_1,M_2)$ is an s-r neighborhood of $X$, then for some $n$,
$(M\Colon M_1,M_2)\times (D^n,S^{n-1})$ is homeomorphic to a regular
neighborhood of $X$ in $\R^{n+m}$, where $m=\dim M$.
If we can show this then the lemma easily implies that $\xi$ is stably
equivalent to the Spivak normal fibration formed from this regular
neighborhood.
By crossing with $D^n$ if necessary, we may assume
$\dim M\geq 2\dim X+1$, so we may assume $i$ and $j$ are embeddings.
Since $M$ is parallelizable, $(M,\partial M)$ immerses in $(\HS^m,\R^{m-1})$.
Since $m\geq 2\dim X+1$ we can subject $i$ and $j$ to a proper homotopy
so that $i\colon X\ \to\ M\subset \HS^m$ and
$j\colon \partial X\ \to\ M_2\subset \R^{m-1}$ become embeddings
on open neighborhoods $U$ and $U_2$, where $U$ is a neighborhood
of $X$ in $M$ and $U_2=M_2\ \cap\ U$ is a neighborhood of
$\partial X$.
In $U$ sits a regular neighborhood of $X$, $(N\Colon N_1, N_2)$.
Hence $(N\Colon N_1, N_2)\subseteq (M\Colon M_1, M_2)$ and excision gives a
simple homotopy equivalence $\partial N\subseteq \Bar{M-N}$ and
$\partial N_2\subseteq \Bar{M_2-N_2}$ (this uses the fact that $i$ and
$j$ are simple homotopy equivalences).
By the $s${--}cobordism theorem (see \cite{bthirtythree} or
\cite{bten}) these are products (assume $m\geq 6$) so
$(M\Colon M_1,M_2)$ is homeomorphic to a
regular neighborhood of $X$ in $\R^m$.
\end{proof}
\medski
\BEGIN{C.2.2.1.1}
The Spivak normal fibration is stably well defined.
\end{Corollary}
\medskip\begin{xRemarks}
By definition we have a Spivak normal fibration for any regular
neighborhood, so we can not properly speak of ``the'' Spivak normal
fibration.
By the corollary however they are all stably equivalent, so we will
continue to speak of the Spivak normal fibration when we really
mean any fibration in this stable class.
This includes fibrations formed from s-r neighborhoods.
\end{xRemarks}
Now, for finite complexes we know the complex satisfies Poincar\'e duality
\iff\ the Spivak normal fibre has the homology of a sphere.
Unfortunately, this is not true for our case.
In fact, Spivak has already shown what is need to get the normal fibre a
sphere.
This information is contained in Theorems
\shortFullRef{T.2.2.2} and \shortFullRef{T.2.2.3}
\begin{xDefinition}\footnote[1]{Expanded.}
A locally compact, finite dimensional CW complex is a {\sl Spivak space\/}
provided the fibre of any Spivak normal fibration has the homology of
a sphere.
A {\sl Spivak pair\/} is a pair, $(X,\partial X)$, of locally compact,
finite dimensional CW complexes such that the fibre of any Spivak
normal fibration has the homology of a sphere, and such that the
Spivak normal fibration for $X$ restricted to $\partial X$ is the Spivak
normal fibration for $\partial X$.
To be slightly more precise, given an s-r neighborhood for
the pair $(X,\partial X)$, there is a natural fibre map from the
Spivak normal fibration for $\partial X$ to the Spivak normal fibration
for $X$ restricted to $\partial X$: it is this map we are requiring to
be an equivalence.
A {\sl Spivak $n${--}ad\/} is defined analogously.
Any Spivak space (pair, $n${--}ad) has a first Stiefel{--}Whitney class
and a fundamental class.
The first Stiefel{--}Whitney class of the Spivak space, $\wone$, is the
first Stiefel{--}Whitney class of the Spivak normal spherical fibration.%
There is a Thom isomorphism $H^\locf_{m+k}(M,\partial M;\Z)\ \to\
H^\locf_{m}(M,M_2;\Z^\wone)$.
Since a parallelizable manifold is oriented, we get a fundamental class
$[M]\in H^\locf_{m+k}(M,\partial M;\Z)$ and the fundamental class of
the Spivak space is the image of this class in
$H^\locf_{m}(X,\partial X;\Z^\wone)\cong H^\locf_{m}(M,M_2;\Z^\wone)$.
We denote it by $[X]$ and note that it is defined up to sign.
A choice of sign will be called an {\sl orientation}.
\end{xDefinition}
\bigski
\BEGIN{T.2.2.2}
The following are equivalent
\begin{enumerate}
\item[1)] $X$ is a Spivak space
\item[2)]
$\capf{[X]}{}\colon
H^\ast_\cmpsup(\tilde X)\ \to\ H_{N-\ast}(\tilde X)$ is an isomorphism
\item[3)] $\capf{[X]}{}\colon
H^\ast(\tilde X)\ \to\ H^\locf_{N-\ast}(\tilde X)$ is an isomorphism
\end{enumerate}
\end{Theorem}
\begin{proof}
2) implies 3) thanks to the following commutative diagram.
\[\begin{matrix}%
0\to&{\text{ Ext}}\bigl(H^{\ast+1}_\cmpsup(\tilde X;\Z),\Z\bigr)&\to&
H^\locf_\ast(\tilde X;\Z) &\to&
{\Homx}\bigl(H^\ast_\cmpsup(\tilde X;\Z),\Z)&\to0\cr
&\uplabeledarrow[\bigg]{}{{\text{ Ext}}\bigl(\capf{[X]}{}\bigr)}&&
\uplabeledarrow[\bigg]{}{\capf{[X]}{}}&&
\uplabeledarrow[\bigg]{}{{\Homx}\bigl(\capf{[X]}{}\bigr)}\cr
0\to&{\text{ Ext}}\bigl(H_{N-1-\ast}(\tilde X;\Z),\Z\bigr)&\to&
H^{N-\ast}(\tilde X;\Z) &\to&
{\Homx}\bigl(H_{N-\ast}(\tilde X;\Z),\Z)&\to0\cr
\end{matrix}\]
3) implies 1) thanks to Spivak, Proposition 4.4, and the observation that
the Spivak normal fibration for $X$ pulled back over $\tilde X$ is
the Spivak normal fibration for $\tilde X$.
This observation is an easy consequence of
\fullRef{T.2.2.1}, the definition of an s-r neighborhood,
and the fact that the transfer map $\sieb(X)\ \to\ \sieb(\tilde X)$
is a homomorphism.
1) implies 2) as follows.
Look at
\[\begin{matrix}%
H_{N-\ast}\bigl(D(\tilde X);\Z\bigr)&
\ZRA{0}{\hskip70pt}{10}&H_{N-\ast}(\tilde X;\Z)\cr
\uplabeledarrow[\Big]{}{\capf{U}{}}\cr
H_{N+k-\ast}\bigl(D(\tilde X), S(\tilde X);\Z)&\LA{\hskip30pt}&
H_{N+k-\ast}(\tilde N,\partial N;\Z)\cr
&&\uplabeledarrow[\Big]{}{\capf{[N]}{}}\cr
H^\ast_\cmpsup(\tilde X;\Z)&\ZLA{-4}{\hskip108pt}{0}
&H^\ast_\cmpsup(N)\cr
\end{matrix}\]
where $U$ is the Thom class for the normal disc fibration $D(\tilde X)$
with spherical fibration $S(\tilde X)$.
$(N,\partial N)$ is an s-r neighborhood for $\tilde X$.
The horizontal maps are induced by the inclusion $\tilde X\subseteq N$
and the proper homotopy equivalence $(N,\partial N)\ \to\
\bigl(D(\tilde X), S(\tilde X)\bigr)$.
All horizontal maps are isomorphisms.
The composite map $H^\ast_\cmpsup(\tilde X)$ to $H_{N-\ast}(\tilde X)$
is essentially the cap product with $U\cap [N]$, where $U\cap[N]$
should be actually be written $i_\ast\bigl(U\cap j_\ast[N]\bigr)$,
where $i_\ast\colon H^\locf_\ast\bigl(D(\tilde X)\bigr)\ \to\
H^\locf_\ast(\tilde X)$ and $j_\ast\colon
H^\locf(N, \partial N)\ \to\ H^\locf_\ast\bigl(D(\tilde X), S(\tilde X)\bigr)$.
$H^\locf_\ast\bigl(D(\tilde X)\bigr)$ is the homology group of the infinite
singular chains on $D(\tilde X)$ which project to give locally finite chains
on $\tilde X$.
( $H^\locf_\ast\bigl(D(\tilde X), S(\tilde X)\bigr)$ is similar.)
Now $[X]=U\cap[N]$ shows 2) is satisfied.
[ 1) was used to get the Thom class $U$.]
\end{proof}
\medski
\BEGIN{C.2.2.2.1}
\footnote[1]{New Corollary.}
For a Spivak space $X$, \\
$\capf{[X]}{}\colon
H^\ast_\cmpsup(\tilde X;\Gamma)\ \to\
H_{N-\ast}(\tilde X;\Gamma^\wone)$ and
$\capf{[X]}{}\colon
H^\ast(\tilde X;\Gamma)\ \to\
H^\locf_{N-\ast}(\tilde X;\Gamma^\wone)$ are isomorphisms for
any local coefficients $\Gamma$.
\end{Corollary}
\smallskip\begin{proof} The results follows from the universal coefficient
formulas.
\end{proof}
\bigski
\BEGIN{T.2.2.3}
\footnote[2]{Restated Theorem and reworked proof.}
Fix an $[X]\in H^\locf_N(X,\partial X;\Z^\wone)$ and
consider the following four families of maps
\begin{enumerate}\item[]
\begin{enumerate}
\item[A)] $\capf{[X]}{}\colon
H^\ast_\cmpsup(\tilde X,\widetilde{\partial X};\Z)\ \to\
H_{N-\ast}(\tilde X;\Z)$
\item[B)] $\capf{[X]}{}\colon
H^\ast(\tilde X,\widetilde{\partial X};\Z)\ \to\
H^\locf_{N-\ast}(\tilde X;\Z)$
\item[C)] $\capf{[X]}{}\colon
H^\ast_\cmpsup(\tilde X;\Z)\ \RA{\hskip16pt}\
H_{N-\ast}(\tilde X,\widetilde{\partial X};\Z)$
\item[D)] $\capf{[X]}{}\colon
H^\ast(\tilde X;\Z)\ \RA{\hskip16pt}\
H^\locf_{N-\ast}(\tilde X,\widetilde{\partial X};\Z)$
\end{enumerate}
\end{enumerate}
Suppose $\partial X$ is a Spivak space with fundamental
class $\partial [X]$.
Then the following are equivalent.
\begin{enumerate}
\item[1)] $(X,\partial X)$ is a Spivak pair with fundamental class $[X]$
\item[2)] any one of the above four maps is an isomorphism
\end{enumerate}
\end{Theorem}
\begin{proof}
1) implies by definition that all four maps are isomorphisms.
By a diagram similar to the one in the proof of
\fullRef{T.2.2.2}, A) implies D) and C) implies B).
Similar diagrams show D) implies A) and B) implies C).
In fact, 2) implies all four maps are isomorphisms.
If A) is an isomorphism, then the Browder lemma shows C) is too.
Conversely, if C) is an isomorphism the Browder lemma shows A) is.
This shows the claim.
D) implies that the Spivak normal fibre is
a sphere by Spivak, Proposition 4.4.
Either 1) or 2) implies that the Spivak normal fibration
has a spherical fibre, $S^{k-1}$, so we have a Thom isomorphism
\[H^\locf_{N+k}(M,\partial M;\Z)\ \RA{\ \capf{U}{}\ }\
H^\locf_N(X,\partial X;\Z^\wone)\]
and that we can pick a fundamental class
$[M]\in H^\locf_{N+k}(M,\partial M;\Z)$
with $[X]=[M]\cap U$.
This shows 1) implies 2) and the only remaining point in 2) implies 1)
is to check that the Spivak normal fibration for $X$
restricted to $\partial X$, ${\mathcal P}(M_1, M, \partial X)$,
is the Spivak normal fibration for $\partial X$,
${\mathcal P}(M_1\cap M_2, M_2, \partial X)$.
There is an evident inclusion \[{\mathcal P}(M_1\cap M_2, M_2, \partial X)
\subseteq {\mathcal P}(M_1, M, \partial X)\ ,\leqno(\ast)\]
where both fibrations are spherical with fibre $S^{k-1}$.
Let $\wone$ and $U$ be the orientation class and the Thom class
for ${\mathcal P}(M_1, M, \partial X)$.
The following
diagram of restrictions and Thom isomorphisms commutes
\[\begin{matrix}%
H^\locf_{N+k}(M,\partial M;\Z)&\RA{\ \capf{U}{}\ }&
H^\locf_N(X,\partial X;\Z^\wone)\cr
\big\downarrow&&\big\downarrow\cr
H^\locf_{N-1+k}(M_2,\partial M_2;\Z)&\RA{\ \capf{U}{}\ }&
H^\locf_{N-1}(\partial X;\Z^\wone)
\end{matrix}\]
where the vertical maps are Poincar\'e isomorphisms and the top map is
a Thom isomorphism.
It follows that the bottom map is an isomorphism, so the Thom class of
one of our fibrations restricts to be the Thom class of the other,
which shows $(\ast)$ is a fibre homotopy equivalence.
\end{proof}
\medski
\BEGIN{C.2.2.3.1}
\footnote[1]{New Corollary.}
For a Spivak pair $(X,\partial X)$, the analogues of all four maps
are isomorphisms with
any local coefficients $\Gamma$.
\end{Corollary}
\smallskip\begin{proof} The results follows from the universal coefficient
formulas.
\end{proof}
\bigskip
Now suppose $\xi$ is an arbitrary spherical fibration over a locally compact,
finite dimensional CW complex $X$.
The total space in general is not such a complex.
Our techniques apply best to such spaces however, and we want to
study these total spaces.
Hence we wish to replace any such space by a space with the proper
homotopy type of a locally compact, finite dimensional CW complex.
\medskip\begin{xDefinition}
Let $S(\xi)$ be the total space of a spherical fibration $\xi$ over
a locally compact, finite dimensional $n${--}ad $X$.
A \emph{\cwation}\ of $\xi$ is an $n${--}ad $Y$ and a proper map
$Y\ \to\ X$ such that the following conditions are satisfied.
$Y$ has the proper homotopy type of a locally finite, finite dimensional
$n${--}ad.
There are maps
$\vrule width 0pt depth 10pt height 12pt
\displaystyle S(\xi)\ {\scriptscriptstyle\Atop%
{\Limitsarrow{10}{10}{\scriptscriptstyle g}{}{\rightarrowfill}}
{\Limitsarrow{10}{10}{}{\scriptscriptstyle h}{\leftarrowfill}}
}\ Y$
such that $h\ \circ\ g$ is a fibre map, fibre homotopic to the
identity and such that $g\ \circ\ h$ is properly homotopic to
the identity.
Lastly
\topD{12}{$\begin{matrix}%
S(\xi)&{\scriptscriptstyle\Atop
{\Limitsarrow{10}{10}{\scriptscriptstyle g}{}{\rightarrowfill}}
{\Limitsarrow{10}{10}{}{\scriptscriptstyle h}{\leftarrowfill}}
}& Y\cr
&\hbox to 10pt{\hss$\searrow\hskip 20pt\swarrow\hbox to 0pt{$\scriptstyle{f}$\hss}$\hss}\cr
&X\cr
\end{matrix}$}{4}
should commute.
The pair $(M_f,Y)$ is seen to satisfy the Thom isomorphism for
$\Delta^\ast$ and $\Delta_\ast$ theories. (See the appendix page \pageref{PoincareAppendix} for
a discussion of the Thom isomorphism in these theories.)
The simple homotopy type of $Y$ is defined by any locally compact,
finite dimensional CW complex having the same proper homotopy
type as $Y$ and for which the Thom isomorphisms are simple
homotopy equivalences.
For a fibration $\xi$, $\bigl(D(\xi),C(\xi)\bigr)$ will denote the pair
$(M_f,Y)$ with this simple homotopy type.
Such a pair is said to be a {\sl simple \cwation}.
It has a {\sl Thom class\/} $U_\xi\in H^k\bigl(D(\xi),C(\xi);
\Z^{w_1(\xi)}\bigr)$, where $w_1(\xi)$ is the first Stiefel{--}Whitney
class of the spherical fibration $\xi$.
There is a Thom isomorphism with twisted coefficients.
\end{xDefinition}
\medskip\begin{xRemarks}
Any spherical fibration of dimension two or more has a simple \cwation.
The proof of this fact is long and is the appendix page \pageref{PoincareAppendix} to this chapter.
\end{xRemarks}
\bigski
\BEGIN{T.2.2.4}
Let $\xi$ be any spherical fibration of dimension $>1$ over a Spivak
space $X$.
Then $\bigl(D(\xi), C(\xi)\bigr)$ is a Spivak pair
with fundamental class $[\xi\>]\in H^\locf_{N+k}\bigl(
D(\xi), C(\xi);\Z^\omega\bigr)$, where
$\omega(g)=\wone(g)\cdot \bigl(w_1(\xi)\bigr)(g)$, where
$\wone$ is the first Stiefel{--}Whitney class for $X$.
If $U_{\xi}$ is a Thom class for $\xi$, $U_\xi \cap\ [\xi] = [X]$,
a fundamental class for $X$.
If $(X,\partial X)$ is a Spivak pair, then
$\bigl(D(\xi)\Colon D(\xi\vert_{\partial X}),C(\xi)\bigr)$ is a Spivak triad with
fundamental class $[\xi\>]$.
In general, a \cwation\ of a Spivak $n${--}ad has
an $(n+1)${--}ad structure with fundamental class $[\xi\>]$.
Fundamental classes and Thom classes are related as in the absolute case.
\end{Theorem}
\medskip\begin{proof}\footnote[1]{New proof.}
To show $\bigl(D(\xi), C(\xi)\cup D(\xi\>\vert_{\partial X})\bigr)$
is a Spivak pair, look at
\[\begin{matrix}%
H^\ast\big(\coverFC{D(\xi)}\bigr)
\RA{\hskip 10pt\psi\hskip10pt}&
H^\locf_{N+k-\ast}\bigl(\coverFC{D(\xi)},
\coverFC{C(\xi)\cup D(\xi\>\vert_{\partial X})\quad}\hskip-5pt\bigr)\cr
\downlabeledarrow[\Big]{\cong}{\cap[X]}&&
\downlabeledarrow[\Big]{\cong}{U_\xi\cap}\cr
H^\locf_{N-\ast}(\coverFC{X\ },\coverFC{\partial X})
&\hbox to 40pt{$\RA{\hskip 60pt}$\hss}&
H^\locf_{N-\ast}\bigl(\coverFC{D(\xi)},
\coverFC{D(\xi\>\vert_{\partial X})}\bigr)\cr
\end{matrix}\]
The bottom horizontal map is an isomorphism, hence so is $\psi$.
We claim $\psi(x)=x\ \cap\ \psi(1)$, where
$1\in H^0\bigl(\coverFC{D(\xi)}\bigr)$ is a choice of generator
which we pick so $U_\xi\cap \psi(1)=[X]$.
But from the diagram,
$U_\xi\ \cup\ \psi(x)=x\cap [X]=x\cap U_\xi\cap \psi(1)=
U_\xi\cap\bigl(x\cap \psi(1)\bigr)$ which proves the result.
Suppose we can show there is a class
$[\xi\>]\in H^\locf_{N+k}\bigl(D(\xi),
C(\xi)\cup D(\xi\vert_{\partial X});\Z^\omega\bigr)$
with $tr\bigl([\xi\>]\bigr)=\psi(1)$.
Note
\[\begin{matrix}%
H^\ast_\cmpsup\big(\coverFC{D(\xi)}\bigr)&
\RA{\hskip10pt \cap [\xi\>]\hskip 10pt}&
H_{N+k-\ast}\bigl(\coverFC{D(\xi)},
\coverFC{C(\xi)\cup D(\xi\>\vert_{\partial X}})\bigr)\cr
\downlabeledarrow[\Big]{\cong}{\cap[X]}&&
\downlabeledarrow[\Big]{\cong}{U_\xi\cap}\cr
H_{N-\ast}(X,\partial X)
&\hbox to 40pt{$\RA{\hskip 60pt}$\hss}&
H_{N-\ast}\bigl(\coverFC{D(\xi)},
\coverFC{D(\xi\>\vert_{\partial X})}\bigr)\cr
\end{matrix}\]
commutes and since three of the maps are isomorphisms,
so is $\cap\ [\xi\>]$.
Since $\xi$ has dimension $>1$,
$\coverFC{D(\xi\>\vert_{\partial X}) \cup C(\xi)\quad}$
is simply connected.
By the Browder lemma,
\[\cap \partial [\xi\>]\colon
H^\ast\bigl(\coverFC{D(\xi\>\vert_{\partial X})
\cup C(\xi)\quad};\Z\bigr)\ \RA{\hskip10pt}\
H_{N+k-1-\ast}\bigl(\coverFC{D(\xi\>\vert_{\partial X})
\cup C(\xi)\quad};\Z\bigr)\]
is an isomorphism, so by \fullRef{T.2.2.2},
$D(\xi\>\vert_{\partial X}) \cup C(\xi)$ is a Spivak space with
fundamental class $\partial [\xi\>]$.
It follows from \fullRef{T.2.2.3} that
$\bigl(D(\xi), D(\xi\>\vert_{\partial X}) \cup C(\xi)\bigr)$ is a
Spivak pair with fundamental class $[\xi\>]$.
Hence we are done if we can construct the class $[\xi\>]$
with the right trace.
To do this, look at the Thom isomorphism
\[H^\locf_{N+k-\ast}\bigl(D(\xi),
C(\xi)\cup D(\xi\>\vert_{\partial X});\Z^\omega\bigr)\
\RA{\ U_\xi \cap\ }\
H^\locf_{N-\ast}\bigl(D(\xi), D(\xi\>\vert_{\partial X});\Z^\wone\bigr)
\]
There is a unique element $[\xi\>]\in H^\locf_{N+k}\bigl(D(\xi),
C(\xi)\cup D(\xi\>\vert_{\partial X});\Z^\omega\bigr)$
such that $U_\xi\cap [\xi\>]=[X]$.
Since cap product commutes with trace,
\[\begin{matrix}%
H^\locf_{N+k-\ast}\bigl(D(\xi),
C(\xi)\cup D(\xi\>\vert_{\partial X});\Z^\omega\bigr)&
\RA{\ U_\xi \cap\ }&
H^\locf_{N-\ast}\bigl(D(\xi), D(\xi\>\vert_{\partial X});\Z^\wone\bigr)\cr
\downlabeledarrow[\big]{}{{\text{ tr}}}&&\downlabeledarrow[\big]{}{{\text{ tr}}}\cr
H^\locf_{N+k-\ast}\bigl(\coverFC{D(\xi)},
\coverFC{C(\xi)\cup D(\xi\>\vert_{\partial X})};\Z\bigr)&
\RA{\ U_\xi \cap\ }&
H^\locf_{N-\ast}\bigl(\coverFC{D(\xi)},
\coverFC{D(\xi\>\vert_{\partial X})};\Z\bigr)\cr
\end{matrix}\]
commutes and it is easy to see ${\text{ tr}}\bigl([\xi\>]\bigr)=\psi(1)$
since
\[\begin{aligned}%
H^\locf_{N}\bigl(D(\xi), D(\xi\>\vert_{\partial X});\Z^\wone\bigr)\cong
H^\locf_{N-\ast}\bigl(X, \partial X);\Z^\wone\bigr)\cong\Z\cr
H^\locf_{N}\bigl(\coverFC{D(\xi)}, \coverFC{D(\xi\>\vert_{\partial X})};
\Z^\wone\bigr)\cong
H^\locf_{N-\ast}\bigl(\coverFC{X}, \coverFC{\partial X});
\Z^\wone\bigr)\cong\Z\cr
\end{aligned}\]
and ${\text{ tr}}$ is an isomorphism by \fullRef{T.2.2.5} below.
\end{proof}
\medski
\BEGIN{C.2.2.4.1}
Let $X$ be a locally compact, finite dimensional CW $n${--}ad.
Let $\xi$ be a spherical fibration of dimension $\geq 2$ over $X$.
Then if $D(\xi)$ is a Spivak $(n+1)${--}ad, $X$ is a Spivak $n${--}ad.
\end{Corollary}
\smallskip\begin{proof}
Let $[X]=U_\xi\ \cap\ [\xi\>]$.
Then $\cap [X]$ induces isomorphisms
\[H^\ast_\cmpsup\bigl(\coverFA{D(\xi)};\Z\bigr)\ \to\
H_{N-\ast}\bigl(\coverFA{D(\xi)},\coverFA{D(\xi\>\vert_{\partial X})};\Z)\ ,\]
or equivalently
$H^\ast_\cmpsup\bigl(\coverFA{X};\Z\bigr)\ \to\
H_{N-\ast}\bigl(\coverFA{X},\coverFA{\partial X};\Z)$.
Inducting over the $n${--}ad structure of $X$ and applying Theorems
\shortFullRef{T.2.2.2} and \shortFullRef{T.2.2.3},
we get $X$ is a Spivak $n${--}ad.
\end{proof}
\bigski
\BEGIN{T.2.2.5}
\hskip-5pt\setcounter{footnote}{0}\footnote{Expanded statement of theorem and added to proof.}
$X$ is a Spivak $n${--}ad \iff\ $\coverFA{X}$ is for any cover of $X$:
moreover ${\text{ tr}}\bigl([X]\bigr)=[\coverFA{X}]$.
If $X$ is an $n${--}ad and $Y$ is an $m${--}ad,
$X\times Y$ is a Spivak $(n+m-1)${--}ad \iff\
$X$ is a Spivak $n${--}ad and $Y$ is a Spivak $m${--}ad:
$[X\times Y]=[X]\times[Y]$.
\end{Theorem}
\medskip\begin{proof}
Our first statement if immediate from \fullRef{T.2.2.1},
since if $N$ is an s-r neighborhood for $X$, $\coverFA{N}$ is one
for $\coverFA{X}$.
Since ${\text{ tr}}$ preserves the fundamental class of manifolds and
since $[X]=U_\xi\ \cap\ [N]$,
$[\coverFC{X}]=U_\xi\ \cap\ [\coverFC{N}]$, we see
${\text{ tr}}\bigl([X]\bigr)=[\coverFA{X}]$.
The ``adic'' part of the product result is easy once we see
$(X\times Y, \partial X\times Y\ \cup X\times\partial Y)=
(X,\partial X)\times (Y,\partial Y)$ is a Spivak pair.
To see this we first need
\medski
\BEGIN{L.2.2.2}
If $\nu_Z$ is the Spivak normal fibration
for an finite dimensional $n${--}ad $Z$, and
if $X$ and $Y$ are such complexes,
$\nu_X\ \ast\ \nu_Y=\nu_{X\times Y}$.
\end{Lemma}
\smallskip\begin{proof}
Let $D_Z$ be the ``disc'' fibration of $\nu_Z$.
Then
$\nu_X\ \ast\ \nu_Y=\nu_X\times D_Y\ \cup\ D_X\times\nu_Y
\subseteq D_X\times D_Y$ denotes the fibrewise join.
Let $(N\Colon N_1,N_2)$ be the s-r neighborhood from which $\nu_X$
was formed.
$(M\Colon M_1, M_2)$ is the corresponding object for $\nu_Y$.
Then $\nu_{X\times Y}$ has
${\mathcal P}(N\times M_1\ \cup\ N_1\times N, N\times M,X\times Y)$ for
total space.
$\nu_X\times D_Y$ consists of triples $e\in\nu_X$, $f\in\nu_Y$ and
$t\in[0,1]$ with $(f,0)=(g,0)$ if $f(1)=g(1)\in Y$ ($D_Y$ is the fibrewise
cone on $\nu_Y$).
There is a similar map for $D_X\times\nu_Y$ which agrees with the first
on $\nu_X\times\nu_Y$.
Hence we get a fibre map
$\nu_X\ \ast\ \nu_Y\ \to\ \nu_{X\times Y}$.
Now $\nu_X$ restricts from a fibration $\nu^\prime_X$ over all of $N$.
$\nu^\prime_Y$ and $\nu^\prime_{X\times Y}$ are defined similarly,
we have a fibre map
$\nu^\prime_X\ \ast\ \nu^\prime_Y\ \to\ \nu^\prime_{X\times Y}$, and
a homotopy equivalence $\nu_Z\subseteq \nu^\prime_Z$.
There is an initial point map $\nu^\prime_{X\times Y}\ \to\
N\times M_1\ \cup\ N_1\times M$, which is a homotopy equivalence.
$\nu^\prime\ \ast\ \nu^\prime_Y\ \to\
N\times M_1\ \cup\ N_1\times M$ via the composition is likewise a
homotopy equivalence.
Hence by Dold \cite{bseven},
$\nu^\prime\ \ast\ \nu^\prime_Y\ \to\ \nu^\prime_{X\times Y}$
is a fibre homotopy equivalence.
Hence so is $\nu\ \ast\ \nu_Y\ \to\ \nu_{X\times Y}$.
\end{proof}
Now $X\times Y$ is a Spivak ad \iff\ the fibre of $\nu_{X\times Y}$
has the homology of a sphere,
and, if $Z\subseteq X\times Y$ is a piece of the ``adic'' structure,
$\nu_{X\times Y}\vert_Z\cong \nu_Z$.
Since the fibre of the fibrewise join is the join of the fibres, the
fibre of $\nu_{X\times Y}$ has the homology of a sphere \iff\
the fibres of $\nu_X$ and $\nu_Y$ do.
The equation $[X\times Y]=[X]\times[Y]$ also follows.
If $\partial X=\partial Y=\emptyset$, we are done.
For any ``adic'' piece $Z\subseteq X\times Y$, there is a fibre map
$\nu_Z\subseteq \nu_{X\times Y}\vert_Z$.
Next suppose $\partial Y=\emptyset$ and $X$ and $Y$ are Spivak.
It follows easily from the lemma that $\partial X\times Y$ is a Spivak space,
and, since $\nu_X\vert_{\partial X}\cong \nu_{\partial X}$ is
an equivalence, that $X\times Y$ is a Spivak pair.
Conversely, if $X\times Y$ is a Spivak pair, $\partial X\times Y$
is a Spivak space so $\partial X$ and $Y$ are.
$\nu_X$ has fibres that have the homology of spheres and since
$\nu_X\ \ast\ \nu_Y\vert_{\partial X\times Y}\cong
\nu_{\partial X}\ \ast\ \nu_Y$ it follows that $\nu_X\vert_{\partial X}\cong
\nu_{\partial X}$ so $(X,\partial X)$ is a Spivak pair.
\medski
\begin{Lemma}
\footnote[2]{This was a remark tossed off in the original which
now seems a bit harder.}
If $(Y,X)$ and $(Y^\prime, X)$ are Spivak pairs, $Y\ \cup_X\ Y^\prime$
is a Spivak space.
\end{Lemma}
\smallskip\begin{proof}
We choose our neighborhoods with care.
Embed $X$ in $\R^{K-1}$ for $K$ much bigger than the dimension of $X$.
Then we can extend this embedding to embeddings $Y$ in $\HS^K$ and
$Y^\prime$ in $\HS^K$.
Let $(M\Colon M_1,M_2)$ be a regular neighborhood for $(Y,X)$ and
let $(M^\prime\Colon M^\prime_1,M^\prime_2)$ be a regular neighborhood for $(Y^\prime,X)$.
Write $Z=Y\ \cup\ Y^\prime$ and note $M\ \cup\ M^\prime \subset \R^K$
is a regular neighborhood for $Z$.
Let $\nu_X={\mathcal P}(M_1\cap M_2,M_2,X)$;
$\nu^\prime_X={\mathcal P}(M^\prime_1\cap M^\prime_2,M^\prime_2,X)$;
$\nu_Y={\mathcal P}(M_1,M,Y)$;
$\nu_{Y^\prime}={\mathcal P}(M^\prime_1,M^\prime,Y^\prime)$;
and $\nu_Z={\mathcal P}(M_1\ \cup_{h\vert_{M_1\cap M_2}}\ M^\prime_1,
M\ \cup_h\ M^\prime,Z)$.
$h$ gives a fibre equivalence between $\nu_X$ and $\nu^\prime_X$
and the natural maps $\nu_X\subseteq\nu_Y\vert_X$,
$\nu^\prime_X\subseteq\nu_{Y^\prime}\vert_X$ are fibre
homotopy equivalences.
Let $\nu^\prime_Z$ be the pushout
\topD{10}{$\begin{matrix}%
\nu_x&\to&\nu_Y\cr
\downarrow&&\downarrow\cr
\nu_{Y^\prime}&\to&\nu^\prime_Z\cr
\end{matrix}$}{4}
and note we have a projection $\nu^\prime_Z\to Z$
and a fibre map $\nu^\prime_Z\subseteq \nu_Z$.
$\nu^\prime_Z\ \to\ Z$ is a Dold fibration \cite{bseven}
and the map $\nu^\prime_Z\subseteq \nu_Z$ is a homotopy
equivalence, both being homotopy equivalent to $Z$.
Hence the map is a fibre homotopy equivalence and the
fibres of $\nu_Z$ have the homology of spheres.
\end{proof}
Now assume $X$ and $Y$ are pairs.
Then $\partial X\times (Y,\partial Y)$ and $(X,\partial X)\times \partial Y$
are Spivak pairs.
Then $\partial (X\times Y)$ is a Spivak space by the last lemma.
Since the fibre of $\nu_{X\times Y}$ has the homology of a sphere,
$\bigl(X\times Y, \partial (X\times Y)\bigr)$ is a Spivak pair,
and so $X\times Y$ is a Spivak triad.
Conversely, if $X\times Y$ is a Spivak triad, $\partial X\times\partial Y$
is a Spivak space, so $\partial X$ and $\partial Y$ are.
$X\times\partial Y$ and $\partial X\times Y$ are Spivak pairs,
so $X$ and $Y$ must be Spivak pairs as well.
The general case follows by induction.
\end{proof}
\bigski
\BEGIN{T.2.2.6}
Let $X$ be a Spivak $n${--}ad, and let $N$ be a regular neighborhood
for $X$.
If $\bigl(D(X),S(X)\bigr)$ is a simple \cwation\ for this normal fibration,
there is a proper map of $(n+1)${--}ads $g\colon N\ \to\ D(X)$
such that the composition $N\ \to\ D(X)\ \to\ N$ is a proper
homotopy inverse for $X\ \to\ N$.
If $[N]$ and $\bigl[D(X)\bigr]$ are the fundamental classes for $N$ and
$D(X)$ respectively, $g_\ast[N]=\bigl[D(X)\bigr]$.
$g$ is a homotopy equivalence of $(n+1)${--}ads (not necessarily a proper
homotopy equivalence)
$g$ is however properly $(\dim N - \dim X -1)$ connected.
\end{Theorem}
\medskip\begin{xRemarks}
If $[X]$ lives in $k${--}dimensional homology, then the normal fibration
has a simple \cwation\ if $\dim N -k \geq 3$.
\end{xRemarks}
\medskip\begin{proof}
To be momentarily sloppy, let $D(X)$ denote the total space of the
normal disc fibration for $X$.
Since $X\ \to\ N$ is a proper homotopy equivalence, pick
an inverse $N\ \to\ X$.
Pull $D(X)$ back over $N$.
It is also a disc fibration and so has a section (see Dold \cite{bseven}
Corollary 6.2).
Map $N\ \to\ D(X)$ by the section followed by the map into $D(X)$.
Under the composition $N\ \to\ D(X)\ \to\ X$, we just get our original
map.
But now we can take the map from the total space of the fibration to
the \cwation.
Letting $D(X)$ be the disc \cwation\ again, we get a map $N\ \to\ D(X)$
so that the map $N\ \to\ D(X)\ \to\ X$ is a proper homotopy inverse
to our original map $X\ \to\ N$.
The map $N\ \to\ D(X)$ is easily seen to be proper and is $g$.
$g$ is a homotopy equivalence of $(n+1)${--}ads by construction.
$g_0\colon N\ \to\ D(X)$ is also a proper homotopy equivalence
($g\colon N\ \to\ D(X)$ is not necessarily a proper homotopy
equivalence of $n$-ads).
The following diagram commutes
\[\begin{matrix}%
H^\ast(N,\partial N)&\LA{\ g^\ast\ }&H^\ast\bigl(D(X),C(X)\bigr)\cr
\downlabeledarrow[\big]{}{\capf{[N]}{}}&&
\downlabeledarrow[\big]{}{g_\ast\capf{[N]}{}}\cr\noalign{\vskip4pt}
H^\locf_{N-\ast}(N)&\RA{(g_0)_\ast}&
H^\locf_{N-\ast}\bigl(D(X)\bigr)\cr
\end{matrix}\]
$\capf{[N]}{}$, $g^\ast$ and $(g_0)_\ast$ are all isomorphisms, so
$g_\ast[N]$ is also an isomorphism.
Therefore $g_\ast[N]=\pm\bigl[D(X)\bigr]$ and we may orient $N$ so that
$g_\ast[N]=\bigl[D(X)\bigr]$.
The map $C(X)\ \to\ X$is properly $q${--}connected, where the normal
spherical fibration has fibre $S^{q}$.
This is seen from the fibration sequence $S^q\ \to\ S(\xi)\ \to\ X$,
where $S(\xi)$ is the total space of the normal spherical fibration,
by noticing that
\[\begin{matrix}%
\Delta(S(\xi)\Colon \pi_k)&\to&\Delta(C(X)\Colon \pi_x)\cr
&\hbox to 0pt{\hss$\searrow\hskip 60pt\swarrow$\hss}\cr
&\Delta(X\Colon \pi_k)\cr
\end{matrix}\]
commutes, where $\Delta(S(\xi)\Colon \pi_k)$ is formed from the groups
$\pi_k(S(\xi)\vert_{X-C},\hat p)$, where $\hat p\in S(\xi)$ covers
$p\in X$ (i.e. just pick one $\hat p$ for each base point in $X$).
The horizontal map is an isomorphism since $C(X)$ is a \cwation.
The first vertical map is an isomorphism for $k<q$ and an epimorphism
for $k=q$, so we are done.
The map $N_1\subseteq N\ \to\ X$ is properly $r${--}connected, where
$r=(\dim N - \dim X - 1)$.
This is seen by showing that the map $N_1\subseteq N$ is properly
$r${--}connected.
But this is easy.
If $K$ is a locally compact complex with $\dim K\leq r$, any map of
$K\ \to\ N$ deforms properly by general position to a map whose
image lies in $N-X$, and so can be properly deformed into $N_1$.
Hence $\Delta(N_1\Colon \pi_k)\ \to\ \Delta(N\Colon \pi_k)$ is onto for $k\leq r$
and $1${--}$1$ for $k\leq r-1$.
Now $g_0\colon N\ \to\ D(X)$ is a proper equivalence so the map
is properly $r${--}connected.
Since $r\leq q$, $g\colon N_1\to\ C(X)$ is properly $r${--}connected.
If $X$ is a space, we are done.
If $(X,\partial X)$ is a pair, the regular neighborhood is $(N\Colon N_1,N_2)$
and the \cwation\ is $\bigl(D(X)\Colon C(X), D(\partial X)\bigr)$.
$C(X)\ \cap\ D(\partial X)= C(\partial X)$.
$g\colon N_1\ \cap\ N_2\ \to\ C(\partial X)$
is properly $r${--}connected as it is an example of the absolute case.
$N_2\ \to D(\partial X)$ is a proper homotopy equivalence, hence
properly $r${--}connected.
$g\colon N_1\ \to\ C(X)$ and $g\colon N\ \to\ D(X)$ we saw were
properly $r${--}connected, so the case for pairs is done.
For the $n${--}ad case, just induct.
\end{proof}
\medskip
We are now ready to define Poincar\'e duality spaces.
\medskip\begin{xDefinition}
A Spivak $n${--}ad is a {\sl Poincar\'e duality $n${--}ad\/} \iff\ the $g$ of
\fullRef{T.2.2.6} is a proper homotopy equivalence
of $(n+1)${--}ads for some regular neighborhood.
\end{xDefinition}
\medskip\begin{xRemarks}
A priori our definition depends on which regular neighborhood we have
used in \fullRef{T.2.2.6}.
In fact this is not the case as our next theorem demonstrates.
\end{xRemarks}
\bigski
\BEGIN{T.2.2.7}
Let $X$ be a locally finite, finite dimensional CW $n${--}ad.
Then $X$ is a Poincar\'e space \iff\ $X$ satisfies Poincar\'e duality with respect
to $[\tilde X]\in H^\locf_{N}(\coverFA{X},\coverFA{\partial X};\Z)$
and with respect to a universal covering functor.
A pair $(X,\partial X)$ is a Poincar\'e pair \iff\ $\partial X$ is a Poincar\'e space
and $X$ satisfies Poincar\'e duality with respect to a universal covering
functor and a class $[\tilde X]\in H^\locf_N(\coverFA{X},
\coverFA{\partial X};\Z)$ such that
$\partial[\tilde X]=[\coverFA{\partial X}]$.
A similar result holds for $n${--}ads.
\end{Theorem}
\medskip\begin{proof}
Since $\capf{[\tilde X]}{}\colon
\Delta^\ast(X\Colon \coverFA{})\ \to\
\Delta_{N-\ast}(X\Colon \coverFA{})$ an isomorphism implies
$\capf{[\tilde X]}{}\colon
H^\ast_\cmpsup(\coverFA{X})\penalty-1000 \to\
H^\locf_{N-\ast}(\coverFA{X})$ is an isomorphism, if $X$ satisfies Poincar\'e
duality then, by \fullRef{T.2.2.2}, $X$ is a Spivak space.
Similarly, by \fullRef{T.2.2.3}, we may show
$(X,\partial X)$ is a Spivak space if $\partial X$ is a Poincar\'e
and if $(X,\partial X)$ satisfies Poincar\'e duality.
In both cases, the fundamental class ,$[X]$,
transfers up to give $\pm[\tilde X]$.
Now look at
\[\begin{matrix}%
\Delta^\ast(X,\partial X\Colon \coverFA{})&\RA{\ r^\ast\ }&
\Delta^\ast(N,N_2\Colon \coverFA{})\cr
&&\downlabeledarrow[\big]{}{\capf{[N]}{}}\cr\noalign{\vskip4pt}
&&\Delta_{n+k-\ast}(N,N_1\Colon \coverFA{})&
\RA{\ g_\ast\ }&
\Delta_{n+k-\ast}(D(X),C(X)\Colon \coverFA{})\cr
&&&&\downlabeledarrow[\big]{}{U_{\nu}\cap}\cr\noalign{\vskip4pt}
&&&&\Delta_{n-\ast}(X\Colon \coverFA{})\cr
\end{matrix}\]
where $r\colon (N,N_2)\ \to\ (X,\partial X)$ is a proper homotopy inverse
for $(X,\partial X)\subseteq (N,N_2)$, and $U_\nu$ is the Thom class
for the normal fibration $\nu$.
By \fullRef{T.2.2.4}, the composition is just
$\capf{[X]}{}$, and $r^\ast$, $\capf{[N]}{}$, and $U_\nu\cap$ are
all isomorphisms.
Hence $(X,\partial X)$ satisfies Poincar\'e duality \iff\ $g_\ast$ is an isomorphism.
If $g$ is a proper homotopy equivalence, $g_\ast$ is clearly an isomorphism.
If $(X,\partial X)$ satisfies Poincar\'e duality, and if $\dim N - \dim X\geq 3$,
$g$ is a proper homotopy equivalence by the Whitehead theorem.
To see this, first note \ $\coverFA{}$ is a universal covering
functor for both $N$ and $D(X)$.
Since $\dim N - \dim X \geq 3$, $N_1\subseteq N$ and
$C(X)\subseteq D(X)$ are at least properly $2${--}connected.
Since $\partial X$ is by hypothesis a Poincar\'e duality space,
$g\ast\colon\Delta_\ast(N_1\Colon \coverFA{})\ \to\
\Delta_\ast(C(X)\Colon \coverFA{})$ is an isomorphism.
By the connectivity of $N_1\subseteq N$ and $C(X)\subseteq D(X)$.
these groups are already the subspace groups for a wise choice
of base points.
By the Browder lemma
$g_\ast\Delta(N\Colon \coverFA{})\ \to\
\Delta_\ast(D(X)\Colon \coverFA{})$ is an isomorphism, and $g$ is
at least properly $2${--}connected, so the Whitehead theorem applies
to show that $g$ is a proper homotopy equivalence.
\end{proof}
\medskip\begin{xRemarks}Note that the proof shows that if $X$
is Poincar\'e, $g$ must be a proper homotopy equivalence whenever
$\dim N - \dim X\geq 3$.
\end{xRemarks}
\medskip
We have seen that manifolds satisfy Poincar\'e duality
with respect to any covering functor.
The Thom isomorphism theorem also holds for any covering functor.
Hence it is easy to see
\medski
\BEGIN{C.2.2.7.1}
A Poincar\'e duality $n${--}ad satisfies Poincar\'e duality with respect to
any covering functor.
\end{Corollary}
\medskip\begin{xDefinition}
The torsion of the equivalence $\capf{[X]}{}\colon
\Delta^\ast(X,\partial X\Colon \coverFA{})\ \to\
\Delta_{N-\ast}(X\Colon \coverFA{})$ is defined to be the
torsion of the Poincar\'e duality space $X$ ($\coverFA{}$
is the universal covering functor).
Since $\bigl(D(X), C(X)\bigr)$ is a simple \cwation,
and since $\capf{[N]}{}$ is a simple equivalence
(\fullRef{T.2.1.2}),
$\tau(X)=(-1)^{N+k}\tau(g)$, where $\tau(X)$ is the torsion of $X$
and everything else comes from the diagram in the proof of
\fullRef{T.2.2.7}.
A simple Poincar\'e $n${--}ad is one for which all the duality maps are simple.
\end{xDefinition}
\bigskip\begin{xExamples}
By Theorems \shortFullRef{T.2.1.1} and \shortFullRef{T.2.1.2}, any paracompact manifold
$n${--}ad is a simple Poincar\'e $n${--}ad.
There are also examples of Spivak spaces which are not Poincar\'e duality
spaces.
One such is the following.
Let $X$ be a finite complex whose reduced homology
with integer coefficients is zero, but which is not contractible.
(The dodecahedral manifold minus an open disc is such an example.)
Look at $\open{C}(X\vee S^2)$, the open cone on $X\vee S^2$.
The obvious map
$\R^3=\open{C}(S^2)\ \to \ \open{C}(X\vee S^2)$ is seen to induce
isomorphisms on $\pi_1$, $H_\ast$ and $H^\ast_\cmpsup$.
Since $\R^3$ is a Spivak space, so is $\open{C}(X\vee S^2)$.
$\open{C}(X\vee S^2)$ is not a Poincar\'e duality space as $X\vee S^2$
is not a Poincar\'e duality space.
\end{xExamples}
\medskip
In the other direction, we have as an application of
a theorem of Farrell{--}Wagoner \cite{bnine}
\medski
\BEGIN{T.2.2.8}
Let $X$ be a locally compact complex with monomorphic ends.
Then $X$ is a Poincar\'e duality space \iff\ $X$ is a Spivak space.
\end{Theorem}
An analogous result is true for $n${--}ads.
\medski
\BEGIN{C.2.2.8.1}
Let $X$ be a Spivak $n${--}ad.
Then $X\times \R^2$ is a Poincar\'e duality $n${--}ad.
\end{Corollary}
\medskip\begin{proof}
We only prove $X$ Spivak implies $X$ Poincar\'e.
If $X$ has monomorphic ends, and if $N$ is an s-r neighborhood with
$\dim N - \dim X\geq 3$, $\partial N$ has monomorphic ends.
$C(X)$ also has monomorphic ends.
The $g$ of \fullRef{T.2.2.6} is at least properly
$2${--}connected.
Hence by \cite{bnine} we need only prove $g$ induces isomorphisms
on $H_\ast$ and $H^\ast_\cmpsup$.
But $g\ast[N]=\bigl[D(X)\bigr]$, and $g$ on homology is an
isomorphism since it is a homotopy equivalence.
Since $N$ and $C(X)$ are both Spivak spaces,
\fullRef{T.2.2.1} shows $g$ induces isomorphisms
on $H^\ast_\cmpsup$.
To show the corollary, observe that if $X$ is not compact, $\times\R^2$
has monomorphic ends.
It is a Spivak space by \fullRef{T.2.2.5}, so, in
this case, we are done.
If $X$ is compact, $X$ is already a Poincar\'e duality space, so the result
will follow from the next theorem.
\end{proof}
\bigski
\BEGIN{T.2.2.9}
Let $X$ be a Poincar\'e duality $n${--}ad, and
let $Y$ be a Poincar\'e duality $m${--}ad.
Then $X\times Y$ is a Poincar\'e duality $(n+m-1)${--}ad.
If $X$ or $Y$ is compact, the converse is true.
\end{Theorem}
\medskip\begin{proof}
From \fullRef{L.2.2.2} we have
$C(X\times Y)=D(X)\times C(Y)\ \cup C(X)\times D(Y)\subseteq
D(X)\times D(Y)=D(X\times Y)$.
If $N$ is an s-r neighborhood for $X$ and if $M$ is one for $Y$,
$N\times Y$ is one for $X\times Y$.
Hence we have $g\times f\colon N\times M\ \to\ D(X)\times D(Y)$
is a map of $(n+m+1)${--}ads.
It is a proper homotopy equivalence if $f$ and $g$ are.
Now suppose $X$ is compact.
By \fullRef{T.2.2.5}, $X$ is a Spivak $n${--}ad,
and hence a Poincar\'e $n${--}ad.
Since $g\times f$ is a proper homotopy equivalence, it induces
isomorphisms on the proper homotopy groups.
We claim $\Delta(N\times M\Colon \pi_k)=\pi_k(N)\oplus\Delta(M\Colon \pi_k)$
for $N$ compact.
This is easily seen by using the cofinal collection of compact subsets
of $N\times M$ of the form $N\times C$, $C\subseteq M$ compact.
A similar result computes $\Delta\bigl(D(X\times Y)\Colon \pi_k\bigr)$.
Since $g\times f$ and $g$ induce isomorphisms,
$f_\ast\Delta(M\Colon \pi_k)\ \to\ \Delta(D(Y)\Colon \pi_k)$ is an isomorphism.
By inducting this argument over the various subspaces of $D(Y)$,
$f$ is seen to be a proper homotopy equivalence of $(m+1)${--}ads.
Hence $Y$ is a Poincar\'e duality $m${--}ad.
\end{proof}
\bigski
\BEGIN{T.2.2.10}
$X$ a Poincar\'e duality $n${--}ad implies $\coverFA{X}$ is a Poincar\'e duality
$n${--}ad for any cover of $X$.
If $X$ is compact or if $\coverFA{X}$ is a finite sheeted cover,
then the converse is true.
\end{Theorem}
\medskip\begin{proof}
Let $N$ be an s-r neighborhood for $X$.
Then $\coverFA{N}$ is an s-r neighborhood for $\coverFA{X}$, so
$D(\coverFA{X})=\coverFA{D(X)}$.
$X$ a Poincar\'e duality $n${--}ad implies $N\ \to\ D(X)$ is a proper
homotopy equivalence of $n${--}ads.
But then so is
$\coverFA{N}\ \to\ \coverFA{D(X)}$, so $\coverFA{X}$ is a Poincar\'e
duality $n${--}ad.
If $\coverFA{X}$ is a Poincar\'e duality $n${--}ad, $X$ is a Spivak $n${--}ad
by \fullRef{T.2.2.5}.
Hence if $X$ is compact, it is a Poincar\'e duality $n${--}ad.
Now if $\coverFA{X}\ \to\ X$ is finite sheeted and we know
$\coverFA{N}\ \to\ \coverFA{D(X)}$ is a proper homotopy
equivalence of $(n+1)${--}ads, we must show $N\ \to\ D(X)$ is a proper
homotopy equivalence of $(n+1)${--}ads.
But if $\dim N - \dim X\geq 3$ (which we may freely assume), this map
is properly $2${--}connected.
Since $\Delta(\coverFA{N}\Colon \pi_k)\ \to\ \Delta(N\Colon \pi_k)$
is an isomorphism for $k\geq2$ when $\coverFA{N}$ is a finite sheeted
cover, $N\to D(X)$ is seen to be a proper homotopy equivalence.%
\footnote[1]{It is easy to be too naive about covers
verses covering functors: indeed
this last sentence is not right.}
To proceed correctly, we repair the error in this last sentence.
The problem begins with the trees: the tree for $\coverFA{N}$
is a cover of the tree for $N$ and so may have a different end
structure.
The groups at corresponding vertices are isomorphic but there tend
to be several vertices in $\coverFA{N}$ for each one in $N$.
What we do see is that the proper map $\coverFA{N}\ \to\ N$
induces epimorphisms $\Delta(\coverFA{N}\Colon \pi_k)\ \to\ \Delta(N\Colon \pi_k)$,
$k\geq 2$.
Now apply this remark to the relative homotopy groups.
The map is still epic and the domain is $0$, hence so is the range.
This argument can be applied to any piece of the $(n+1)${--}ad structure,
so $X$ is a Poincar\'e duality $n${--}ad.
\end{proof}
\bigskip\begin{xRemarks}
The full converse to Theorems \shortFullRef{T.2.2.9}
and \shortFullRef{T.2.2.10} are false.
Let $X$ be any Spivak space which is not a Poincar\'e duality space.
Then $X\times\R^2$ is a counterexample to the converse of
\Ref{T.2.2.9} as it is a Poincar\'e duality space
by \Ref{C.2.2.8.1}.
$X\times T^2$ is a counterexample to
\Ref{T.2.2.10}, since $X\times T^2$ is not a
Poincar\'e duality space by \Ref{T.2.2.9},
but its cover $X\times\R^2$ is.
\end{xRemarks}
\bigski
\BEGIN{T.2.2.11}
Let $\xi$ be any spherical fibration of dimension $\geq 2$ over
a locally compact, finite dimensional CW $n${--}ad $X$.
Then $X$ is a Poincar\'e duality $n${--}ad \iff\ $D(\xi)$ is a Poincar\'e duality
$(n+1)${--}ad.
\end{Theorem}
\medskip\begin{proof}
By \fullRef{T.2.2.4} or \fullRef{C.2.2.4.1}, we may assume $X$ and
$D(\xi)$ are Spivak ads, and we have the formula $U_\xi\ \cap\
[\xi\>] = [X]$.
Since the Thom isomorphism is valid for the $\Delta$ theory
(see the appendix page \pageref{PoincareAppendix}), $\capf{[X]}{}$ is an isomorphism \iff\
$\capf{[\xi\>]}{}$ is an isomorphism.
Since $\dim \xi\geq 2$, a universal covering functor for $X$
induces one for $D(\xi)$.
\fullRef{T.2.2.7} now gives the desired
conclusions.
\end{proof}
\bigskip\begin{xRemarks}
The torsions of the Poincar\'e spaces occurring in Theorems
\shortFullRef{T.2.2.9}, \shortFullRef{T.2.2.10} and \shortFullRef{T.2.2.11} can be ``computed''.
In particular, $\tau(X\times Y)=A\bigl(\tau(X),\tau(Y)\bigr)$
where $A$ is the pairing $\sieb(X)\times\sieb(Y)\ \to\ \sieb(X\times Y)$
(see \fullRef{L.1.5.23} and the preceding discussion).
$\tau(\coverFA{X})={\text{ tr}}\ \tau(X)$,
where ${\text{ tr}}\colon \sieb(X)\ \to\ \sieb(\coverFA{X})$.
$\tau\bigl(D(\xi)\bigr) = (-1)^n\tau\bigl(D(\xi)\bigr)^t$, where
$n$ is the dimension of the fundamental class of $X$, and $t$ is
the transpose operation on $\sieb(D(\xi))$.
These formulas are not very hard to deduce and will be left to the
reader.
\end{xRemarks}
\medskip
We conclude this section by investigating the ``uniqueness'' of the Spivak
normal fibration.
We first prove
\bigski
\BEGIN{L.2.2.3}
Let $D(\xi)$ be a \cwation\ for some spherical fibration $\xi$ over
a Poincar\'e duality $n${--}ad.
If there is a stably parallelizable manifold $(n+1)${--}ad $N$ and a
proper, degree one, homotopy equivalence $N\ \to\ D(\xi)$,
then $\xi$ is stably equivalent to the Spivak normal fibration.
\end{Lemma}
\medskip\begin{xRemarks}
Given all spherical fibrations over a Poincar\'e duality $n${--}ad $X$,
we wish to determine which of these could be the normal fibration
of some complex having the same proper homotopy type as $X$.
In the compact case, Spivak showed that there was only one,
the one with the reducible Thom space.
\fullRef{L.2.2.3} shows that if $D(\xi)$
has the degree one proper homotopy type of a stably
parallelizable manifold, then $\xi$ is the Spivak normal fibration
for $X$.
If $\xi$ is the normal fibration for some complex $Y$, $D(\xi)$
has the degree one proper homotopy type of a parallelizable
manifold, so in the non{--}compact case there is one and only
one Spivak normal fibration.
\end{xRemarks}
\medskip\begin{proof}
If the equivalence were simple, $N$ would be an s-r neighborhood
and this would follow from \fullRef{T.2.2.1}.
Now by Siebenmann \cite{bthirtythree},
$N\times S^1\ \to D(\xi)\times S^1$ is a simple equivalence.
$D(\xi)\times S^1$ is a simple \cwation\ for $\xi\times S^1$ over
$X\times S^1$.
$N\times S^1$ is an s-r neighborhood for $X\times S^1$.
${\mathcal P}(N_1,N,N)\times S^1\to\ N\times S^1$ makes the map
$N_1\times S^1\subseteq N\times S^1$ into a fibration,
so $\nu_X\times S^1$ is fibre homotopy equivalent to $\nu_{X\times S^1}$.
But $\xi\times S^1$ is stably fibre homotopy equivalent
to $\nu\vert_{X\times S^1}$ by \fullRef{T.2.2.1}.
Hence $\nu_X$ is stably $\xi$.
\end{proof}
\bigski
\BEGIN{T.2.2.12}
If $f\colon X\ \to\ Y$ is a proper homotopy equivalence
between Poincar\'e duality $n${--}ads, then
$f^\ast\nu_Y\cong \nu_X$.
\end{Theorem}
\medskip\begin{proof}
Let $\xi=f^\ast(\nu_Y)$.
Then
\[\begin{matrix}%
D(\xi)&\to&\xi&\to&\nu_Y&\to&D(\nu_Y)\cr
&\searrow&\downarrow&&\downarrow&\swarrow\cr
&&X&\to&Y\cr
\end{matrix}\]
commutes.
The top horizontal row is a proper homotopy equivalence, as one easily
checks by applying $\Delta(\quad\Colon \pi_k)$ to everything.
Since $D(\nu_Y)$ has the degree one proper homotopy type of
a parallelizable manifold, so does $D(\xi)$.
Hence by \fullRef{L.2.2.3}, $\xi\cong\nu_x$.
\end{proof}
\medskip
Spivak's identification of the normal fibration actually proves a stronger
theorem.
We can prove this result as
\bigski
\BEGIN{T.2.2.13}
Let $f\colon X\ \to\ Y$ be a degree one map of Poincar\'e duality
$n${--}ads.
If there is a spherical fibration $\xi$ over $Y$ such that
$f^\ast(\xi)\cong \nu_X$, then $\xi\cong\nu_Y$.
\end{Theorem}
\medskip\begin{proof}
\[\begin{matrix}%
D(\nu_X)&\to&\nu_X&\to&\xi&\to&D(\xi)\cr
&\searrow&\downarrow&&\downarrow&\swarrow\cr
&&X&\to&Y\cr
\end{matrix}\]
commutes, so it is not hard to show that the top row is a degree one map.
$U_{\nu_X}\ \cap\ \big[D(\nu_X)\bigr]=[X]$;
$U_{\xi}\ \cap\ \big[D(\xi)\bigr]=[Y]$; and $f_\ast[X]=[Y]$.
Hence the top row must take $\bigl[D(\nu_X)\bigr]$ to $\bigl[D(\xi)\bigr]$.
$D(\nu_X)$ has the proper homotopy type of a parallelizable $(n+1)${--}ad,
$N$ so there is a degree $1$ map $g\colon N\ \to\ D(\xi)$.
Since $N$ is parallelizable, there is a topological microbundle over $D(\xi)$
which pulls back to the normal bundle of $N$ (namely the trivial bundle).
If $\dim\xi\geq 2$ (which we may always assume) then the pair
$\Big(D\bigl(\xi(Z)\bigr),C\bigl(\xi(Z)\bigr)\Bigr)$, for $Z\subseteq Y$
as part of the $n${--}ad structure on $Y$, is properly $2${--}connected.
Hence by the remarks following \fullRef{T.3.1.2},
we can find a parallelizable manifold $M$ and a degree one proper homotopy
equivalence $M\ \to\ D(\xi)$.
By \fullRef{L.2.2.3}, $\xi\cong\nu_Y$.
\end{proof}
\medskip\begin{xRemarks}
Logically \fullRef{T.2.2.13} should follow \fullRef{T.3.1.2} in chapter 3.
We do not use the result until we are past that point so it does no harm
to include it here.
\end{xRemarks}
The chief purpose of \fullRef{T.2.2.13} is to
severely limit the bundles which can occur in a surgery problem.
\newpage
\section{The normal form for Poincar\'e duality spaces}
\newHead{II.3}
In order to get a good theory of surgery, one needs to be able to do
surgery on Poincar\'e duality spaces; at least one must be able to
modify fundamental groups.
The results of this section show that Poincar\'e duality spaces look like
manifolds through codimension $1$.
These results are a direct generalization of Wall \cite{bthirtynine}
Section 2, especially pages 220{--}221.
\medskip\begin{xDefinition}
Let $X$ be a Poincar\'e duality $n${--}ad.
Then if $[X]\in H^\locf_n$, $X$ is said to have \emph{formal dimension $n$}.
($X$ is often said to have dimension $n$.)
\end{xDefinition}
\bigski
\BEGIN{T.2.3.1}
Let $X$ be a Poincar\'e space of dimension $n\geq 2$.
The $X$ satisfies $D n$.
If $X$ is a connected Poincar\'e duality $m${--}ad $m\geq2$, of
dimension $n\geq 3$, then $X$ satisfies $D(n-1)$.
\end{Theorem}
\medskip\begin{proof}
This follows from definitions and \fullRef{T.1.6.2}.
\end{proof}
\bigski
\BEGIN{T.2.3.2}
Let $X$ be a Poincar\'e duality space of dimension $n$, $\geq 4$.
Then $X$ has the proper homotopy type of $Y$, where $Y$ is a
Poincar\'e duality space which is the union of two Poincar\'e duality pairs
$(Z,\partial H)$ and $(H,\partial H)$ where $H$ is a smooth manifold
of dimension $n$ formed from a regular manifold in $\R^n$ of
a given tree for $Y$ by adding $1$ handles along the boundary,
and where $Z$ is a subcomplex satisfying $D(n-2)$.
The torsion of this equivalence may have any preassigned value.
The map induced by inclusion $\Delta(H\Colon \pi_1)\ \to\ \Delta(Y\Colon \pi_1)$
is surjective.
\end{Theorem}
\medskip\begin{proof}
Let $\hat C_\ast$ be the dual chain complex for $X$, reindexed so that
there is a chain map $\capf{[X]}{}\colon \hat C_\ast\ \to\
C_{\ast}(X)$.
By \fullRef{T.1.6.3}, we can find a chain complex $Y$ with
$C_\ast(Y)=\hat C_\ast$ in dimensions greater than $3$.
$C_3(Y)=\hat C_3\oplus junk$, and the complex $Y^2\ \cup\ junk$
satisfies $D2$.
Now we could have arranged things so that the only vertices of $X$
were the vertices of the tree.
This is seen as follows.
First we claim we can find a subcomplex $V\subseteq X$ which contains
all the vertices and such that $T\subseteq V$ is a proper deformation
retract.
We do this as follows.\footnote[1]{The original argument here was wrong.}
\hide
Let ${\mathcal S}=\{ U\ \vert\ U$ is a $1${--}dimensional subcomplex of $X$,
$T\subset U$ with the inclusion a proper $0${--}equivalence, and
$U$ contains all the vertices of $X\ \}$.
${\mathcal S}\neq\emptyset$ as $X^1\in{\mathcal S}$.
${\mathcal S}$ is ordered by inclusion.
Let $U_1\supseteq U_2\supseteq\cdots$ be a totally ordered sequence
in ${\mathcal S}$.
Then $\cap\ U_i$ is also in ${\mathcal S}$.
Let $V$ be a minimal element of ${\mathcal S}$, which exists by Zorn.
We claim $H_1(V)=0$, so if not, look at a cycle in $V$.
At least one of the $1${--}simplexes of the cycle is not in $T$ for $T$ has
no $1${--}cycles.
Let $V_1\subseteq V$ be all of $V$ less one of the one simplexes in
the cycle which is not in $T$.
Then $V_1$ is a subcomplex, $T\subseteq V_1$, and $V_1$ contains all
the vertices.
This contradiction shows $H_1(V)=0$.
Put a metric on the vertices of $X^1$ as follows.
If $p$ and $q$ are vertices, the length of a cellular path joining $p$
to $q$ is obtained by counting the number of $1${--}simplices
in the domain (or equivalently the number of vertices minus $2$).
Define $d(p,q)$ to be the minimum of the length amongst all cellular
paths joining $p$ to $q$.
It is not hard to see $d$ is a metric.
\endhide
Choose an increasing sequence of compact subcomplexes of the $1$-skeleton,
$X^1$, $K_0\subseteq K_1\subseteq \cdots$, whose union is $X^1$.
Let $V$ be a subcomplex of $X^1$ with $T\subseteq V$.
Let $\{v_0, \cdots \}$ denote the vertices of $X^1-V$.
By the definition of a tree, there is a locally finite set of paths
$\{\lambda_i\}$, with $\lambda_i$ joining $v_i$ to some vertex of $V$.
It is no problem to assume the $\lambda_i$ are cellular.
\hide
We can refine the collection of $K_j$ if necessary to insure that
any $\lambda_i$ beginning at a vertex in $K_j$ lies entirely in $K_{j+1}$.
\endhide
Construct a sequence of increasing subcomplexes, $V_r$, of $X^1$,
two sets of points ( $\{ x^{(r)}_i\}$ the vertices in $X^1-V_{r}$
and $\{y^{(r)}_i\}$ the vertices of $V_r$ )
and two sets of locally finite cellular paths $\lambda^{(r)}_i\subset X^1$
(beginning at $x^{(r)}_i$ and ending at a vertex of $V_{r}$)
and $\Lambda^{(r)}\subset V_r$ (beginning at $y^{(r)}_i$ and
ending at a vertex of $T$ )
inductively as follows.
$V_0=T$, $\lambda^{(0)}=\lambda_i$, $\Lambda^{(0)}_i$ the constant
path at the relevant vertex.
$V_r$ is obtained from $V_{r-1}$ in two steps.
First adjoin all vertices $x^{(r-1)}_i$ of $X^1$
for which there exists a $j$ such that the path
$\lambda^{(r-1)}_j$ has its next to the last vertex $x^{(r-1)}_i$.
Next add some $1${--}cells.
First, for a fixed $x^{(r-1)}_i$, there may be several $j$ satisfying
our condition: pick one, say $i_j$.
Then add the $1${--}cell from $x^{(r-1)}_i$ to $V_{r-1}$
given by the last $1${--}cell in $\lambda^{(r-1)}_{i_j}$.
We need only define $\lambda^{(r)}_i$ for the $ x^{(r)}_i$:
just define it to be the sub{--}path of $\lambda^{(r-1)}_i$ which
starts at $x^{(r)}_i$ and continues until it encounters a vertex in $V_r$.
This path is definitely properly contained in $\lambda^{(r-1)}_i$ since
the next to the last vertex in $\lambda^{(r-1)}_i$ is certainly in $V_r$.
(Of course $\lambda^{(r)}_i$ may be shorter than this.)
The $\lambda^{(r)}_i$ are a locally finite collection
since they form a subcomplex of such a collection.
Define $\Lambda^{(r)}_i$ as follows.
If the vertex in question lies in $V_{r-1}$, use $\Lambda^{(r-1)}_i$.
Otherwise, look at the $1${--}cell out of $y^{(r)}_i$. Its other end
lies in $V_{r-1}$ by definition so is $y^{(r-1)}_j$ for some $j$.
$\Lambda^{(r)}_i$ is the path which
follows the $1${--}cell and then $\Lambda^{(r-1)}_j$.
Check that the collection $\Lambda^{(r)}_i$ is a locally finite collection.
Let $V_\infty=\cup\ V_r$.
Consider any vertex $x^{(0)}_i$ and its path $\lambda^{(0)}_i$.
The distance from this vertex to the tree along this path is finite,
say $R$.
Then $x^{(0)}_i\in V_R$ (it may of course land in a smaller $V_R$).
This is easy to check by induction on $R$.
It follows that $V_\infty$ contains all the vertices of $X^1$ and we
let $y^{(\infty)}_i$ be an enumeration of them.
Define $\Lambda^{(\infty)}_i\subset V_\infty$ as follows.
$y^{(\infty)}_i$ lies in some $V_r$ so define
$\Lambda^{(\infty)}_i=\Lambda^{(r)}_i$ and note that
$\Lambda^{(r+1)}_i$ is the same path so this is well{--}defined.
There is a deformation retraction
$d\colon V_r \times[0,1]\to\ V_{r}$ of $V_r$ to $V_{r-1}$
obtained by collapsing the new $1${--}cells to the end attached
to $V_{r-1}$, so $V_\infty$ is a $1${--}complex with $H_1(V_\infty)=0$.
The paths $\Lambda^{(\infty)}_i$ show the inclusion $T\subseteq X_\infty$
is a proper $0${--}equivalence.
It follows that $T\subseteq V_\infty$ is a proper homotopy equivalence
as desired.
Set $K=\Bar{V-T}$ and look at $X/K$.
The collapse map $X\to X/K$ is a proper homotopy equivalence.
For a proof, see \cite{bsix} Proposition 2.11, page 220.
Note that all the maps there may be taken to be proper.
$X/K$ has only the vertices of the tree for $0${--}cells.
Now, to return to our proof, we may assume $\hat C_n=C_0(X)$
has a generator for each vertex of our tree.
$\hat C_{n-1}$ has a generator for each $1${--}cell of $X$.
As in Wall \cite{bthirtynine} Corollary 2.3.2, each $(n-1)${--}cell
is incident to either two $n${--}cells, or to the same $n${--}cell twice.
Look at an attaching map $S^{n-1}\ \to\ X^{n-1}$ for an $n${--}cell.
This can be normalized to take a finite, disjoint, collection of discs
onto the $(n-1)${--}cells homeomorphically and to take the rest of
$S^{n-1}$ into the $(n-2)${--}skeleton.
Each $(n-1)${--}cell eventually gets just two such discs mapped into it.
The $n${--}discs together with the $(n-1)${--}cells corresponding to the
$1${--}cells of the tree are seen to form a regular neighborhood in
$\R^n$ of the tree, and $H$ is obtained from this
by attaching $1${--}handles.
If $Z$ is the part of $Y$ in dimensions $\leq n-2$ (or is $Y^2\ \cup\ junk$
if $n=4$), $Y=Z\ \cup_{\partial H}\ H$ where $H$ is formed
from $n${--}discs corresponding to the $n${--}cells by attaching
$1${--}cells as indicated by the $(n-1)${--}cells.
Actually, we want to form the mapping cylinder of $\partial H\ \to\ Z$
and then tale the union along $\partial H$.
Since $H$ is a manifold, the result is clearly homeomorphic to $Y$.
We denote the mapping cylinder by $Z$ so $Y=Z\ \cup_{\partial H}\ H$,
and $\partial H$ is a subcomplex of $Z$ and hence $Y$.
Note that $Z$ still satisfies $D(n-2)$.
\vfill
Now $Z\subseteq Y$ is at least properly $2${--}connected, for
$Z$ always contains the $2${--}skeleton of $Y$.
Since $(H,\partial H)$ is a Poincar\'e duality space,
\fullRef{T.2.1.4} says $(Z,\partial H)$ satisfies Poincar\'e
duality with respect to the covering functor induced from the universal
covering functor for $Y$.
But this is just the universal covering functor for $Z$ as $Z\subseteq Y$
is properly $2${--}connected.
$\partial H$ is a Poincar\'e duality space, so
\fullRef{T.2.2.7} says $(Z,\partial H)$ is a Poincar\'e
duality pair.
The statement about the torsion is contained in \fullRef{T.1.6.3},
so we finish by showing $\Delta(H\Colon \pi_1)\ \to\ \Delta(Y\Colon \pi_1)$
is onto.
Our proof is basically Wall \cite{bthirtynine} Addendum 2.3.3, but
is more complicated.
We too will use the construction of $Z$ and $H$ via the dual cell
decomposition.
In our original complex, there were $0${--}cells, $e^0_p$,
one for each $p$ a vertex of $T$.
There were $1${--}cells $e^1_i$ satisfying
$\partial e^1_i=g_i e^0_p - e^0_q$
where $g_i$ is a loop at $p$.
The $g_i$ which occur generate $\Delta(Y\Colon \pi_1)$.
In the dual complex we have $n${--}cells, $e^n_p$ and $(n-1)${--}cells
$e^{n-1}_i$ with
$\displaystyle e^n_p=\sum_i (\pm g_i e^{n-1}_i)-\sum_j e^{n-1}_j$,
where the sign is given by the local coefficients on $Y$, and where the
sum runs over all $(n-1)${--}cells incident to $e^n_p$.
The core $1${--}disc of the handle corresponding to $e^{n-1}_i$
followed by the unique minimal path in $T$ from
the endpoint of the $1${--}disc to its initial point point has homotopy
class $g_i$.
Hence $\Delta(H\Colon \pi_1)$ is onto $\Delta(Y\Colon \pi_1)$.
\end{proof}
\medski
\BEGIN{C.2.3.2.1}
Let $X$ be a Poincar\'e duality space of dimension $3$.
Then $X$ has the proper homotopy type of $Y$, where $Y$
is the union of two Poincar\'e duality pairs $(Z,\partial H)$ and $(H,\partial H)$,
where $H$ is a regular neighborhood in $\R^3$ of a given tree for $X$,
and $Z$ is a subcomplex of $Y$ satisfying $D2$.
The torsion of this equivalence can be arbitrary.
\end{Corollary}
\medskip\begin{proof}
Using the dual cell decomposition as before, let $Z$ be the subcomplex
of $Y$ such that $\hat C_3=C_3(Y,Z)$ and such that $Z$ satisfies $D2$.
$\hat C_3$ has one $3${--}cell for each vertex of the tree.
Now there is a locally finite collection of paths from each $n${--}cell
to the vertex of the tree it represents.
Given $H$, a regular neighborhood of the tree in $\R^3$, we
describe a map $\partial H\ \to\ Z$ which extends to a map
$H\ \to\ Y$ such that the induced map $C_3(H,\partial H)\ \to\
C_3(Y,Z)$ is an isomorphism.
Hence $Z\ \cup_{\partial H}\ H$ has the proper homotopy type of $Y$
and we will be done.
The map is the following.
$H$ can be viewed as the connected sum of a collection of $n${--}discs,
one for each vertex of the tree, by tubes corresponding to the
$1${--}cells of the tree.
$H$ can then be properly deformed to the subcomplex consisting of
$n${--}discs joined by the cores of the connecting tubes.
$\partial H$ under this deformation goes to a collection
of $(n-1)${--}spheres joined by arcs.
Map the $(n-1)${--}sphere to $Z$ by the attaching map of the
corresponding $n${--}cell in $Y$.
Map an arc between two such spheres to the paths to the tree,
and then along the unique path in the tree between the two vertices.
This map clearly has the necessary properties.
\end{proof}
\bigski
\BEGIN{T.2.3.3}
Let $(X,\partial X)$ be a Poincar\'e duality pair of dimension $n$, $n\geq 4$.
Then $(X,\partial X)$ has the proper homotopy type of a Poincar\'e duality
pair $(Y,\partial Y)$ which is the union of a Poincar\'e duality pair
$(Z, \partial H\ \cup\ \partial Y)$ and
a Poincar\'e duality pair $(H,\partial H)$, where $H$ is a regular neighborhood
in $\R^n$ of any given tree for $Y$ with $1${--}handles added
along the boundary, and $Z$ is a subcomplex of $Y$ satisfying $D(n-1)$.
The torsion of this equivalence may be given any preassigned value.
$\Delta(H\Colon \pi_1)\ \to\ \Delta(Y\Colon \pi_1)$ is onto.
\end{Theorem}
\medskip\begin{proof}
By \fullRef{T.2.3.2} or \fullRef{C.2.3.2.1}, we may assume $\partial X$
already looks like $K\ \cup\ M$, where $M$ is a regular neighborhood
for a tree of $\partial X$ in $\R^{n-1}$, and $K$ satisfies $D(n-2)$.
Let $\hat C_\ast$ be the dual complex for $C_{n-\ast}(X)$.
Then there is a chain map $\capf{[X]}{}\colon
\hat C_\ast\ \to\ C_\ast(X,\partial X)$.
We apply \fullRef{T.1.6.4} to find a complex $Y$ with
$C_\ast(Y)=C_\ast$ in dimensions greater than $3$ and with
$\partial X \subseteq Y$.
$C_3(Y)=C_3\ \cup\ junk$.
Set $L$ to be $Y^{(n-1)}$.
Then $M\subseteq L$.
Normalize the attaching maps for the $n${--}cells as before.
If $Z=Y^{(n-2)}\ \cup\ M$ ($Y^2\ \cup\ junk\ \cup\ M$ if $n=4$),
then $Y=Z\ \cup\ H$ where $H$ has the advertised description.
Notice $\partial H\ \cap\ \partial X$ can be $M$ if one likes.
As before, $(H, \partial H)$ is a Poincar\'e duality pair.
$\partial H\ \cap\ \partial X =M$, so
$\partial H=(\Bar{\partial H - M\ },\partial M)\ \cup\ (M,\partial M)$
and $\partial X=(K,\partial M)\ \cup\ (M,\partial M)$.
The rest of the proof proceeds as in the proof of
\fullRef{T.2.3.2}.
\end{proof}
\newpage
\section*{Appendix. The \cwation\ of a spherical fibration}
\label{II.A}
\renewcommand\rightmark{\Roman{chapter}.\ \nameref*{II.A}}
\LRTpageLabel{PoincareAppendix}
\bigskip
We recall the definition.
Let $\xi$ be a spherical fibration over a finite dimensional,
locally finite CW $n${--}ad.
Assume $\xi\geq 2$.
Let $S(\xi)$ be the total space.
We seek an $n${--}ad $Y$, a proper map $f\colon Y\ \to\ X$,
and maps
$\vrule width 0pt depth 10pt height 12pt
\displaystyle S(\xi)\ {\scriptscriptstyle
\Atop{\Limitsarrow{10}{10}{\scriptscriptstyle g}{}{\rightarrowfill}}
{\Limitsarrow{10}{10}{}{\scriptscriptstyle h}{\leftarrowfill}}
}\ Y$
which commute with the two projections.
We also require that $Y$ have the proper homotopy type of a
locally compact, finite dimensional CW $n${--}ad. $g\ \circ\ h$ must
be properly homotopic to the identity, and $h\ \circ\ g$ must
be fibre homotopic to the identity.
We give $Y$ a simple homotopy type by finding an equivalent CW
complex for which the Thom isomorphism is simple.
We digress briefly to include a discussion of the Thom isomorphism.
If $D(\xi)$ is the total space of the disc fibration associated to $\xi$,
we define $\Delta(D(\xi)\Colon h,\coverFA{})$ and
$\Delta(D(\xi), S(\xi)\Colon h,\coverFA{})$ to be the groups one
gets by applying the $\Delta$ construction to the groups
$h\bigl(\pi^{-1}(\coverFA{X-C},\hat p)\bigr)$ for $D(\xi)$ and
$h\bigl(\pi^{-1}(\coverFA{X-C},
\rho^{-1}(\coverFA{X-C},\hat p)\bigr)$ for $\bigl(D(\xi), S(\xi)\bigr)$,
where $p$ is a vertex of $X$, $\hat p$ is a lift of $p$ into
$\coverFA{D(\xi)}$, and
$\pi\colon \coverFA{D(\xi)}\ \to\ \coverFA{X-C}$ and
$\rho\colon \coverFA{S(\xi)}\ \to\ \coverFA{X-C}$ are the
projections for the fibrations over $\coverFA{X-C}$ by restriction
and pullback from $D(\xi)$ and $S(\xi)$ respectively.
Now the Thom class for $\xi$, $U_\xi$ goes under
$\coverFA{X-C}\ \to\ X-C\ \to\ X$
to the Thom class for $\coverFA{S(\xi)}$.
If $h$ is cohomology we modify the $\Delta$ groups above
in the obvious manner.
We will denote by $\Delta_\ast(D(\xi)\Colon \coverFA{})$
the $\ast${-}\thx\ homology group with covering functor $\coverFA{}$\ .
$\Delta^\ast$ is the cohomology theory.
Then we have maps
$\cupf{U_\xi}{}\colon \Delta^{\ast}(D(\xi)\Colon \coverFA{})\ \to\
\Delta^{\ast+k}(D(\xi),S(\xi)\Colon \coverFA{})$ and
$\capf{U_\xi}{}\colon \Delta_{\ast+k}(D(\xi),S(\xi)\Colon \coverFA{})\ \to\
\Delta_\ast(D(\xi)\Colon \coverFA{})$.
They are easily seen to be isomorphisms.
The maps $h$ and $g$ induce isomorphisms of
$\Delta_{\ast}(S(\xi)\Colon \coverFA{})$ and $\Delta_\ast(Y\Colon \coverFA{})$
with a similar result for cohomology (the reader should have no
trouble defining $\Delta_\ast(S(\xi)\Colon \coverFA{})$ or its cohomological
analogue).
We also get isomorphisms of $\Delta_\ast(D(\xi),S(\xi)\Colon \coverFA{})$
and $\Delta_\ast(M_f,Y\Colon \coverFA)$, again with a similar result in
cohomology.
Hence we can speak of a Thom isomorphism for the \cwation.
We first prove that if we can find a \cwation, we can give it a unique simple
homotopy type.
Let $C$ be a CW $n${--}ad the proper homotopy type of the \cwation\ $Y$
( $C$ locally compact, finite dimensional).
\topD{0}{$\begin{matrix}%
C&\RA{\ \rho\ }&Y\cr
&\hbox to 0pt{\hss$\hbox to 0pt{\hss$\scriptstyle f\circ\rho$}\searrow\hskip 10pt%
\swarrow\hbox to 0pt{$\scriptstyle{f}$\hss}$\hss}\cr
&X&
\end{matrix}$}{4}
commutes, where $\rho$ is a proper homotopy equivalence with
$f\ \circ\ \rho$ cellular.
(It is easy to find such a $\rho$.)
Let $\tau_\rho$ denote the torsion of the corresponding Thom homology
isomorphism.
If $\lambda\colon K\ \to\ C$ is a proper homotopy equivalence, the
Thom isomorphism associated to $\rho\ \circ \lambda$ has torsion
$\tau_\rho+\rho_\ast \tau(\lambda)$ by
\fullRef{L.1.5.22}.
Since we may pick $\tau(\lambda)$ arbitrarily, we can find a $\rho$
with $\tau_\rho=0$.
Suppose now we have $\lambda\colon K\ \to\ Y$ with $\tau_\lambda=0$,
$f\ \circ\ \lambda$ cellular).
Let $a\colon Y\ \to\ K$ be a proper homotopy inverse to $\lambda$.
Then
\[\begin{matrix}%
C&\RA{\ a\ \circ\ \rho\ }&K\cr
&\hbox to 0pt{\hss\hbox to 0pt{\hss$\scriptstyle f\circ\rho$}$\searrow\hskip 24pt%
\swarrow$\hbox to 0pt{$\scriptstyle{f\circ\lambda}$\hss}\hss}\cr
&X&
\end{matrix}\]
properly homotopy commutes.
We get a proper homotopy equivalence of pairs\hfill\penalty-10000
$F\colon (M_{f\circ\rho},C)\ \to\ (M_{f\circ\lambda},K)$
such that
$F\vert_C= a\ \circ\ \rho$, and
\topD{12}{$\begin{matrix}%
M_{f\circ\rho}&\RA{\ F\ }&M_{f\circ\lambda}\cr
&\hbox to 0pt{\hss$\searrow$}\hskip 10pt%
\hbox to 0pt{$\swarrow$\hss}\cr
&X&
\end{matrix}$}{4}
commutes.
By \fullRef{L.1.5.19}, $M_{f\circ\rho}\ \to\ X$
and $M_{f\circ\lambda}\ \to\ X$ are simple, so $F$ is a simple
equivalence.
The torsion of $F$ from $(M_{f\circ\rho},C)$ to $(M_{f\circ\lambda},K)$
is $\tau_\rho-\tau_\lambda=0$, so by \fullRef{T.1.5.1},
the torsion of $a\ \circ\ \rho$ on the subspace groups is zero.
But as $\dim\xi\geq 2$, $f\ \circ\ \rho$ and $f\ \circ\ \lambda$
are at least properly $2${--}connected.
Hence the subspace groups with the induced covering functor are the
absolute groups with the universal covering functor.
Hence $a\ \circ\ \rho$ is a simple homotopy equivalence, so the simple
homotopy type of a \cwation\ is unique.
\medskip
We now construct the promised $Y$.
Notice first that we can replace $X$ by any locally compact,
finite dimensional CW $n${--}ad of the same proper homotopy type.
Hence we may as well assume $X$ is a locally finite simplicial $n${--}ad
of finite dimension.
This is seen as follows.
By \cite{beleven} Theorem 4.1 and Lemma 5.1, $X$ is the union of $A$
and $B$ where $A$ and $B$ are the disjoint union of finite complexes.
Each finite complex has the homotopy type of a finite simplicial complex,
and if a subcomplex is already simplicial, we need not disturb it.
Hence we get a locally finite simplicial complex $Y$ and a map
$f\colon X\ \to\ Y$ by making subcomplexes of the form $C\ \cap\ D$,
$C\subset A$ and $D\subset B$ simplicial and then making $C$ and $D$
simplicial.
Then $Y=A^\prime\ \cup\ B^\prime$ where $f\colon A\ \to\ A^\prime$
and $f\colon B\ \to\ B^\prime$ are proper homotopy equivalences.
Also $f\colon E\ \to\ E^\prime$ is a proper homotopy equivalence
where $E=\{ C\ \cap\ D\ \vert\ C\subset A,\ D\subset B \}$.
The proper Whitehead Theorem shows $f$
is a proper homotopy equivalence.
$X$ being what it is, we can find open sets $C_i$ such that $X-C_i$
and $\Bar{C}_i$ are subcomplexes, each $\Bar{C}_i$ is compact,
and $\xi\vert_{\Bar{C}_i}$ is trivial.
Furthermore, $\cup\ C_i =X$, the $C_i$ are locally finite, and the $C_i$
are indexed by the positive integers.
We set $\displaystyle V_i=\mathop{\cup}_{j\leq i} C_j$
We can also find an increasing collection of open sets $U_i$ such that
$U_i\subseteq V_i-C_i$, $\Bar{U}_i$ is compact, and
$\displaystyle \mathop{\cup}_{i} U_i=X$.
We first construct spaces $Y_i$ and maps $g_i$ and $f_i$ inductively
so that
\[\begin{matrix}%
\xi\vert_{\bar{V}_i}&\RA{\ g_i\ }& Y_i\cr\noalign{\vskip6pt}
&\hbox to 0pt{\hss$%
\hbox to 0pt{\hss$\scriptstyle\pi\vert_{\Bar{V}_i}$}\searrow\hskip 20pt
\swarrow\hbox to 0pt{$\scriptstyle f_i$\hss}
$\hss}\cr
&\Bar{V}_i\cr
\end{matrix}\leqno{A)}\]
commutes.
Let $Y_1=\Bar{V}_1\times S^k$, $k=\dim\xi\geq 2$.
$g_1$ and $f_1$ exist since $\xi\vert_{V_1}$ is trivial.
$f_1$ is just projection.
We now induct; i.e. we have
\begin{enumerate}
\item[1)] A space $Y_i$ and maps $g_i$ and $f_i$ such that A) commutes.
\item[2)] $g_i$ is a homotopy equivalence.
\item[3)] $Y_i=Y_{i-1}\ \cup_\rho\ \Bar{C}_i\times S^k$ via some
homotopy equivalence
\item[]$\rho\colon Y_{i-1}\ \cap\ f^{-1}_{i-1}\bigl(
\Bar{V}_{i-1}\ \cap \Bar{C}_i\bigr)\ \to\
\bigl(\Bar{V}_{i-1}\ \cap\ \Bar{C}_i\bigr)\times S^k$.
\item[4)] $g_{i-1}\vert_{f^{-1}_{i-1}(U_{i-1})}=
g_{i}\vert_{f^{-1}_{i-1}(U_{i-1})}$ and $f_{i-1}\vert_{Y_{i-1}}=
f_{i}\vert_{Y_{i-1}}$.
\item[5)] Let ${\mathcal S}_r=\{ \Bar{C}_{i_1}\ \cap\ \Bar{C}_{i_2}\ \cap\
\cdots \Bar{C}_{i_r}\ \vert\ i_1 < i_2 < \cdots < i_r\ \}$.
\item[]If $C\in{\mathcal S}_r$, $g_i$ restricted to $f^{-1}_i(C\ \cap\ \Bar{V}_i)$
is a homotopy equivalence.
\end{enumerate}
Notice that $Y_1$, $g_1$ and $f_1$ satisfy 1){--}5).
(Let $Y_0=\emptyset$.)
If we can verify 1){--}5), we can construct $Y$ as the increasing union of $Y_i$
with identifications.
$g$ and $f$ can be defined from the $g_i$ and $f_i$ respectively by 4).
Inductively, $Y$ has the proper homotopy type of a locally compact,
finite dimensional complex, since it is covered by finite complexes,
$\Bar{C}_i\times S^k$, of bounded dimension in a locally finite fashion.
For a better proof, see \fullRef{P.2.4.1} below.
Now given $Y_{i-1}$, $f_{i-1}$, and $g_{i-1}$, we construct
$Y_i$, $f_i$ and $g_i$.
By Dold \cite{beight}, $\xi\vert_{\Bar{V}_i}$ can be gotten from
$\xi\vert_{\Bar{V}_{i-1}}$ and $\xi\vert_{\Bar{C}_i}$ as follows.
Over $\Bar{C_i\ \cap\ V_{i-1}\ }$, we have an equivalence
$\varphi\colon (\xi\vert_{\Bar{V}_{i-1}})\vert_{\Bar{C_i\ \cap\ V_{i-1}\ }}
\ \to\ (\Bar{C_i\ \cap\ V_{i-1}\ })\times S^k$.
Let $H_1=\xi\vert_{\Bar{V}_i-1}$, $H_2=\Bar{C}_i\times S^k$, and let
$H_3=\bigl\{ (x,w)\ \vert\ x\in H_1\vert_{\Bar{C_i\cap V_{i-1}\ }},
\ w\in\bigl((\Bar{C_i\cap V_{i-1}\ })\bigr)^I,\
\pi(x)=\pi\bigl(w(t)\bigr)$ for all $t\in I,\ \varphi(x)=w(1)\ \bigr\}$.
Then $\xi\vert_{\Bar{V}_i} \cong H_1\cup H_3 \cup H_2$, where
$H_1\vert_{\Bar{C_i\cap V_{i-1}}}$ is embedded in $H_3$ via
$x\mapsto \bigl(x, $ constant path at $\varphi(x)\bigr)$.
The embedding of $H_2\vert_{\Bar{C_i\cap V_{i-1}}}$ is harder
to describe.
Let $\varphi^\prime$ be the inverse to $\varphi$.
The $\varphi\ \circ\ \varphi^\prime$ is fibre homotopic to the identity.
Let $\psi$ be a fibre homotopy between these two maps, with
$\psi(\quad,0)={\text{ id}}$.
Then $H_2\vert_{\Bar{C_i\cap V_{i-1}}}$ is embedded in $H_3$ via
$x\mapsto \bigl(\varphi^\prime(x), \psi(x,t)\bigr)$.
We must now define the $\rho$ in 3).
We are given
\[\begin{matrix}%
H_1\vert_{\Bar{C_i\cap V_{i-1}}}&\RA{\hskip 20pt \varphi\hskip20pt }&
\Bar{C_i\cap V_{i-1}}\times S^k\cr
\noalign{\vskip2pt}
\downlabeledarrow[\Big]{}{g_{i-1}}&&\downlabeledarrow[\Big]{}{{\text{ id}}}\cr
\noalign{\vskip6pt}
Y_{i-1}\big\vert_{f^{-1}_{i-1}(\Bar{C_i\cap V_{i-1}})}&
\dottedBar{\hskip50pt}\hskip-10pt\to&
\Bar{C_i\cap V_{i-1}}\times S^k\cr
\noalign{\vskip2pt}
\downlabeledarrow[\Big]{}{f_{i-1}}&&\downlabeledarrow[\Big]{}{{\text{ proj}}}\cr
\noalign{\vskip6pt}
\Bar{C_i\cap V_{i-1}}&\RA{\hskip 20pt{\text{ id}}\hskip20pt }
&\Bar{C_i\cap V_{i-1}}\cr
\end{matrix}\leqno{B)}\]
We would like to fill in the dotted arrow with $\rho$
so that the diagram actually commutes.
To do this, we may have to alter $\varphi$ within its fibre homotopy
class, but this will not change our bundle.
Since $g_{i-1}$ is a homotopy equivalence, it has an inverse, $h$.
$h$ may be assumed to g=be a fibre map, so $h\ \circ\ g_{i-1}$ is
a fibre homotopy equivalence.
Let $G$ be its fibre homotopy inverse.
The $G\ \circ\ h\ \circ\ g_{i-1}$ is fibre homotopic to the identity.
$g_{i-1}\ \circ\ (G\ \circ\ h)$ is homotopic to the identity.
Set $\rho=({\text{ id}})\ \circ\ \varphi\ \circ\ (G\ \circ\ h)$.
Then $\rho$ is a fibre map so the bottom square commutes.
Set $\varphi_1=({\text{ id}})^{-1}\ \circ\ \rho\ \circ\ g_{i-1}$.
Then $\varphi_1$ is fibre homotopic to $\varphi$, and B) commutes
with $\varphi_1$ in place of $\varphi$.
$\rho$ is a homotopy equivalence, so 3) is satisfied.
From now on, we assume $\varphi$ chosen so that B) commutes with
the $\rho$ along the dotted arrow.
Set $Y_i= Y_{i-1}\ \cup_\rho\ \Bar{C}_i\times S^k$.
$f_i$ is defined by $f_i\vert_{Y_{i-1}}= f_{i-1}$ and
$f_i\vert_{\Bar{\scriptstyle C}_i\times S^k}={\text{ proj}}$.
B) insures that this is well{--}defined on the intersection.
$g_i$ is unfortunately harder to define.
$\xi\vert_{\Bar{V}_i}\cong H_1\ \cup\ H_2\ \cup\ H_3$, so let
$\alpha\colon \xi\vert_{\Bar{V}_i}\ \to\ H_1\ \cup\ H_2\ \cup\ H_3$
be an equivalence.
$\alpha$ may be chosen to be the identity on $\xi\vert_{U_{i-1}}$.
We define a map $h\colon H_1\ \cup\ H_2\ \cup\ H_3\ \to\ Y_i$
as follows.
$g\vert_{H_1}=g_{i-1}$.
To define $g_i$ on the other two pieces, look at $\psi$, the fibre
homotopy between $\varphi\ \circ\ \varphi^\prime$ and ${\text{ id}}$.
This can be extended to a fibre map of
$(\Bar{C}_i\times S^k)\times I\ \to\ (\Bar{C}_i\times S^k)$ since
${\text{ id}}\colon (\Bar{C_i\cap V_{i-1}})\times S^k\ \to\
(\Bar{C_i\cap V_{i-1}})\times S^k$ can clearly be extended.
Now define $(g\vert_{H_3})(x,w)=g_{i-1}(x)$.
Note our two definitions agree on $H_1\cap H_3$.
We could have defined $(g\vert_{H_3})(x,w)=w(1)$ equally well.
We define $(g\vert_{H_2})(x)=F(x)$.
If $x\in H_2\cap H_3$, then $(g\vert_{H_3})(x)=(g\vert_{H_3})\bigl(
\varphi^\prime(x),\psi(x,t)\bigr)=\psi(x,1)=
\varphi\ \circ\ \varphi^\prime(x)$.
$(g\circ_{H_2})(x)=F(x)=\varphi\ \circ\ \varphi^\prime(x)$ by
the definition of $F$.
Hence $g$ is well{--}defined and we set $g_i= g\ \circ\ \alpha$.
Now 4) clearly holds since $\alpha\vert_{f^{-1}_{i-1}(U_{i-1})}$
is the identity.
1) holds as $g\colon H_1\cup H_2\ \cup H_3\ \to\ Y_i$
preserves fibres by construction.
Hence we are left with showing 2) and 5).
For $r$ sufficiently large, $C\in{\mathcal S}_r$ implies
$C\cap\Bar{V}_i=\emptyset$, since the collection $\{C_i\}$is locally finite
\setcounter{footnote}{0}\footnote{${\mathcal S}_r$ is defined in 5), page \pageref{ItemFive}}.
We show 5) by downward induction on $r$, since if
$C\cap\Bar{V}_i=\emptyset$, 5) is obvious.
Assume we have established the result for $r=k+1$.
Let $C\in{\mathcal S}_k$.
If $C\cap\bar{C_i}=\emptyset$, then $C\cap\Bar{V}_{i-1}=C\cap\Bar{V}_i$
and we are done since 5) holds for $g_{i-1}$ and $\alpha$ is a fibre
homotopy equivalence.
If $C\cap\Bar{V}_i=\Bar{C}_i\cap\Bar{V}_i$ we are done since $F$
is a fibre map.
So let $L=C\cap\Bar{V}_{i-1}$, and let $K=C\cap\Bar{C}_i$ with both $K$
and $L$ non{--}empty.
$g_i\vert_{f^{-1}_i(L)}$ is a homotopy equivalence, again since $\alpha$
is a fibre homotopy equivalence and $g_{i-1}\vert_{f^{-1}_i(K)}$ is also
a homotopy equivalence.
$K\cap L\subseteq \Bar{V}_{i-1}$, and $K\cap L\in{\mathcal S}_{k+1}$.
Hence $g_i\vert_{f^{-1}_i(K\cap L)}$ is a homotopy equivalence.
Therefore $g_i\vert_{f^{-1}_i(C)}$ is a homotopy equivalence and we
are done with 5).
Therefore we have a space $Y$ and maps
\topD{12}{$\begin{matrix}%
S(\xi)&\RA{\ g\ }&Y\cr
&\hbox to 0pt{\hss$\hbox to 0pt{\hss$\scriptstyle{\pi}$\hss}
\searrow\hskip10pt\swarrow\hbox to 0pt{$\scriptstyle{f}$\hss}$\hss}\cr
&X\cr
\end{matrix}\ .$}{4}
We claim $g$ is a homotopy equivalence.
Since by Milnor \cite{btwentytwo}, $S(\xi)$ has the homotopy type of
a CW complex, this is equivalent to showing $g$ induces isomorphisms
in homotopy.
But $\displaystyle\pi_k(g)=\mathop{\lim}_{\Atop{\longrightarrow}{i}}
\pi_k(g_i)$, and since $\pi_k(g_i)=0$, $\pi_k(g)=0$.
Let $h\colon Y\ \to\ S(\xi)$ be a homotopy inverse for $g$.
By an easy argument like the one after diagram B), we may assume $h$
preserves fibres and the $h\ \circ\ g$ is fibre homotopic to the identity.
Notice that by construction $f^{-1}(x)$ is homeomorphic to a sphere
of dimension $\dim\xi$.
$\pi^{-1}(x)$ has the homotopy type of such a sphere.
Since $h\ \circ\ g$ is fibre homotopic to the identity,
$g_x\colon \pi^{-1}(x)\ \to\ f^{-1}(x)$ has a left inverse.
As both spaces are spheres of dimension $2$ or more, $g_x$ is a
homotopy equivalence.
Now in the terminology of Bredon \cite{btwo}, $f$ is $\psi${--}closed, and
$f^{-1}(x)$ is $\psi${--}taut, where $\psi$ is the family of compact supports.
(Note $Y$ is locally compact, so $\psi$ is paracompactifying, and then
apply (d) on page 52 to show $f^{-1}(x)$ is $\psi${--}taut.
$f$ is $\psi${--}closed easily from the definition, which is on page 53,
since $X$ is Hausdorff.)
Hence we have a Leray spectral sequence for the map
$f\colon X\ \to\ Y$.
We have the Serre spectral sequence for $\pi\colon S(\xi)\ \to\ X$,
and $g$ induces a map between these two.
$g$ induces an isomorphism on the $E_2$ terms since it is a homotopy
equivalence on each fibre.
Hence $g\colon H^\ast_\cmpsup(Y)\ \to\
H^\ast_{\varphi}\bigl(S(\xi)\bigr)$ is an isomorphism, where $\varphi$ is
the set of supports whose image in $X$ is compact.
As $\dim\xi\geq2$, $\pi^\ast\colon H^\ast_\cmpsup(X)\ \to\
H^\ast_\varphi\bigl(S(\xi)\bigr)$ is an isomorphism for $\ast<2$.
Hence $f^\ast\colon H^\ast_\cmpsup(X)\ \to\ H^\ast_\cmpsup(Y)$ is
an isomorphism for $\ast<2$, so
$f^\ast\colon H^0_{{\text{ end}}}(X)\ \to\ H^0_{{\text{ end}}}(Y)$ is an
isomorphism, so $f$ is a proper $0${--}equivalence.
We claim $f$ is a proper $1${--}equivalence.
To see this, note $f\vert_C$ is a $1${--}equivalence for $C\in{\mathcal C}_r$
all $r\geq1$.
Now an easy van{--}Kampen induction shows $f$ is a $1${--}equivalence
when restricted to any union of $\Bar{C}_i$'s.
Hence $f$ is a proper $1${--}equivalence.
Thus $g_\#\colon \Delta(S(\xi)\Colon \pi_1)\ \to\ \Delta(Y\Colon \pi_1)$ is an
isomorphism as both groups are isomorphic, via $\pi_\#$ and $f_\#$,
to $\Delta(X\Colon \pi_1)$.
Now we still have maps
\[\begin{matrix}%
Y- f^{-1}(K_i)&{\scriptscriptstyle
\Atop{\Limitsarrow{10}{10}{\scriptscriptstyle h}{}{\rightarrowfill}}
{\Limitsarrow{10}{10}{}{\scriptscriptstyle g}{\leftarrowfill}}
}&
S(\xi\vert_{X-K_i})\cr
&\hbox to0pt{\hss$%
\hbox to 0pt{\hss$\scriptstyle{f}$}\searrow\hskip40pt\swarrow\hbox to 0pt{$\scriptstyle{\pi}$\hss}
$\hss}\cr
&X-K_i\cr
\end{matrix}\]
\newpage
\noindent
where $\displaystyle K_i=X-\ \mathop{\cup}_{j\geq i} \Bar{C}_i$.
$g$ restricted to each fibre is still a homotopy equivalence with
inverse induced from $h$.
For any cover $\coverFC{}$, of $X-K_i$, we get
\[\begin{matrix}%
\coverFC{Y- f^{-1}(K_i)\ }&{\scriptscriptstyle
\Atop{\Limitsarrow{10}{10}{\coverFC{\scriptscriptstyle \ h\ }}{}{\rightarrowfill}}
{\Limitsarrow{10}{10}{}{\coverFC{\ \scriptscriptstyle g\ }}{\leftarrowfill}}
}&
\coverFC{S(\xi\vert_{X-K_i})\ }\cr
&\hbox to0pt{\hss$%
\hbox to 0pt{\hss$\scriptstyle{\coverFC{\ f\ }}$}
\searrow\hskip40pt\swarrow\hbox to 0pt{$\scriptstyle{\coverFC{ \pi\ \ }}$\hss}
$\hss}\cr
&\coverFC{X-K_i\ }\cr
\end{matrix}\]
where the covers on the top row are induced covers from $\coverFC{}$ over
$X-K_i$.
$\coverFC{S(\xi\vert_{X-K_i})}$ is the same as
$S(\coverFC{\ \xi\ }\vert_{\coverFC{X-K_i}})$,
the spherical fibration induced from $\xi\vert_{X-K_i}$
over $\coverFC{X-K_i}$.
$\coverFC{\ g\ }$ likewise induces a homotopy equivalence of fibres,
so as before we get
\[h^\ast\colon H^\ast_\varphi
\Bigl(\coverFC{S\bigl(\xi\vert_{X-K_i}\bigr)\ },
\coverFC{S\bigl(\xi\vert_{\partial(X-K_i)}\bigr)\ }\Bigr)\ \to\
H^\ast_\cmpsup\Bigl(\coverFC{Y-f^{-1}(K_i)\ },
\partial\bigl(\coverFC{Y-f^{-1}(K_i)\ }\bigr)\Bigr)
\]
is an isomorphism.
A word about the existence of these covers is in order.
Since $X-K_i$ is a CW complex, its cover exists.
The cover for $S(\xi\vert_{X-K_i})$ then also clearly exists.
We claim $Y-f^{-1}(K_i)$ is semi{--}locally $1${--}connected, from
which it follows that its cover also exists.
To see our claim, observe $f\colon Y-f^{-1}(K_i)\ \to\ X-K_i$ is
a $1${--}equivalence.
Given any point $y\in Y-f^{-1}(K_i)$, let $N\subseteq X-K_i$ be a
neighborhood of $f(y)$ such that $\pi_1(N)\ \to\ \pi_1(X-K_i)$
is the zero map.
Since $X-K_i$ is semi{--}locally $1${--}connected, such an $N$ exists.
Now $f^{-1}(N)$ is a neighborhood for $y$, and
$\pi_1\bigl(f^{-1}(N)\bigr)\ \to\ \pi_1\bigl(Y-f^{-1}(K_i)\bigr)$
is also zero.
Hence $Y-f^{-1}(K_i)$ is semi{--}locally $1${--}connected.
A similar argument shows $Y-f^{-1}(K_i)$ is locally path connected.
Therefore, $h^\ast\colon\Delta(S(\xi)\Colon \coverFC{})\ \to\
\Delta(Y\Colon \coverFC{})$ is an isomorphism for any covering functor
induced from one over $X$.
Since $f$ is a proper $1${--}equivalence, if we take a universal
covering functor for $X$, we get one for $Y$.
(The actual covering functor on $Y$ is the following. Any $A\in{\mathcal C}(Y)$
is contained in a unique minimal $f^{-1}(X-K_i)$ so let the cover over
$A$ be induced from the cover over this space.)
$g^\ast\colon \Delta(Y\Colon \coverFA{})\ \to\ \Delta(S(\xi)\Colon \coverFC{})$
is defined where $\coverFC{}$ is the covering functor induced by $g$ from
$\coverFA{}$ over $Y$.
$g^\ast\ \circ\ h^\ast=(h\ \circ\ g)^\ast\colon
\Delta^\ast(S(\xi)\Colon \coverFA{})\ \to\ \Delta^\ast(S(\xi)\Colon \coverFC{})$
is an isomorphism as $\coverFA{}$ and $\coverFC{}$ are equivalent
covering functors.
Hence $h^\ast\ \circ\ g^\ast=(g\ \circ\ h)^\ast\colon
\Delta^\ast(Y\Colon \coverFA)\ \to\ \Delta^\ast(Y\Colon \coverFA)$
is an isomorphism, so $h\ \circ\ g$ is a proper homotopy equivalence.
$g\ \circ\ h$ is already a fibre homotopy equivalence, and it
is not hard to change $h$ until $h\ \circ\ g$ is properly homotopic
to the identity and $g\ \circ\ h$ is fibre homotopic to the identity.
To finish, we need only show \fullRef{P.2.4.1} below.
We first need
\bigski
\BEGIN{T.2.4.1}
Let $Y$ be a locally compact, separable ANR.
Then $Y$ is properly dominated by a locally{--}finite simplicial complex.
\end{Theorem}
\medskip\begin{proof}
Let $\alpha$ be an open covering of $Y$
by sets whose closure is compact.
Since $Y$ is metrizable , $Y$ is paracompact, so we can assume $\alpha$
is locally finite.
We now apply Hu \cite{bfifteen}, Theorem 6.1, page 138, to get a
locally finite simplicial complex $X$ and maps $\varphi\colon X\ \to\ Y$
and $\psi\colon Y\ \to\ X$ with $\varphi\ \circ\ \psi$
$\alpha${--}homotopic to the identity, i.e. if $H$ is the homotopy,
for each $y\in Y$, there exists $U\in\alpha$ such that $H(y,t)\in U$
for all $t\in[0,1]$.
By our choice of $\alpha$, $\varphi\ \circ\ \psi$ is properly
homotopic to the identity.
Now $X$ is actually the nerve of some cover $\delta$ in the proof of
Hu, Theorem 6.1.
In the proof, we may take $\delta$ to be star{--}finite and locally finite.
Then the nerve $X$ is a locally finite simplicial complex, and the map
$\varphi\colon X\ \to\ Y$ is proper.
To see this last statement, it is enough to show $\varphi^{-1}(U)$ is
contained in a compact subset of $X$ for any $U\in\alpha$.
Recall $\varphi$ is defined by picking a point in each $V\in\delta$
and sending vertex of the nerve which corresponds to $V$ to our
chosen point and then extending.
Our extension satisfies the property that any simplex lies entirely in
some element of $\alpha$.
So let $U_1$ be the union of all elements of $\alpha$ intersecting $U$.
$\Bar{U}_1$ is compact as $\alpha$ is locally finite, so let $U_2$
be the union of all elements of $\alpha$ intersecting $\Bar{U}_1$.
$\Bar{U_2}$ is again compact, so there are only finitely many elements
of $\delta$ which intersect $U_2$.
Let $K\subseteq X$ be the subcomplex generated by these elements
of $\delta$.
$K$ is finite, hence compact, and $\varphi^{-1}(U)\subseteq K$.
\end{proof}
\medski
\BEGIN{C.2.4.1.1}
Let $Y$ be a locally compact, separable ANR, and suppose the covering
dimension of $Y$, $\dim Y$, is finite
(see Hurewicz and Wallman \cite{bsixteen} for a definition).
Then $Y$ is properly dominated by a locally finite simplicial complex of
dimension $\dim Y$.
\end{Corollary}
\medskip\begin{proof}
Make the same changes in Hu \cite{bfifteen} Theorem 6.1, page 164
that we made to the proof of Theorem 6.1, page 138.
We get a simplicial complex $P$ and a proper map $\varphi\colon P\ \to\ Y$
such that for any map $f\colon X\ \to\ Y$ with $X$ a metric space
of dimension $\leq\dim Y$, there exists a map $\psi\colon X\ \to\ P$
with $\varphi\ \circ\ \psi$ $\alpha${--}homotopic to $f$.
Moreover, $P$ has no simplices of dimension $>\dim Y$.
Apply this for $X=Y$, $f={\text{ id}}$.
\end{proof}
\medski
\BEGIN{C.2.4.1.2}
A locally compact, separable ANR of dimension $\leq n$ satisfies $D n$.
\end{Corollary}
\medskip\begin{proof}
By \fullRef{C.2.4.1.1} and nonsense, it remains to show $Y$
is homogamous.
But an ANR is locally contractible (Hu \cite{bfifteen}, Theorem 7.1,
page 96), and any metric space is paracompact so
\fullRef{C.1.1.2.1} applies.
\end{proof}
\bigski
\BEGIN{P.2.4.1}
The space $Y$ which we constructed has the proper homotopy type of
a locally compact, finite dimensional CW complex.
\end{Proposition}
\medskip\begin{proof}
We first show $Y$ is a finite dimensional, locally compact, separable ANR.
We then find a finite dimensional simplicial complex $Z$ and a proper
map $h\colon Z\ \to\ Y$ which is properly $n${--}connected for
any finite $n$.
Since both $Y$ and $Z$ satisfy $D n$ for some finite $n$, $h$ is a proper
homotopy equivalence.
\medskip
\insetitem{Step 1}
$Y$ is a finite dimensional, locally compact, separable ANR.
\smallskip
By Hu \cite{bfifteen} Lemma 1.1, page 177, Theorem 1.2, page 178, and
induction, each $Y_i$ is an ANR.
The induction is complicated by the necessity of showing
$Y_{i-1}\ \cap\ f^{-1}_{i-1}(\Bar{V}_{i-1}\cap\Bar{C}_i)$ is an ANR.
Hence our induction hypothesis must be
\begin{enumerate}
\item[a$)_{k\hphantom{,r}}$] $Y_k$ is an ANR
\item[b$)_{k, r}$] $Y_k \cap f^{-1}_k(\Bar{V}_k\cap C)$ is an ANR
for all $C\in{\mathcal S}_r$.
\end{enumerate}
One then shows that for some finite $r$, b$)_{k, r}$ holds vacuously.
b$)_{k, s}$, $s>r$, and b$)_{k-1 , r}$ imply b$)_{k, r}$, so we get
b$)_{k, r}$ for all $r$.
b$)_{k, 1}$ and a$)_{k-1}$ imply a$)_k$, so we are done.
Since each $Y_i$ is an ANR, each $Y_i$ is a local ANR (Hu, Proposition 7.9,
page 97).
If $Y$ is metrizable, $Y$ is an ANR by Hu, Theorem 8.1, page 98.
Now $Y$ is $T_1$ and regular.
To see this observe each $Y_i$ is $T_1$ and regular since it is metrizable.
Now if $U\subseteq Y$ is any compact set, there is a $Y_i$ with
$V\subseteq Y_i$ and $V$ homeomorphic to $U$.
With this result and the observation that $Y$ is locally compact, it is easy to
show $Y$ is $T_1$ and regular. $Y$ is locally compact because
it has a proper map to a locally compact space $X$.
$Y$ is $\sigma${--}compact since $X$ is, so $Y$ is second countable.
Hence $Y$ is metrizable (see Kelly \cite{bseventeen} page 125)
and separable.
We are left with showing $Y$ has finite covering dimension.
By Nagami \cite{btwentyseven} (36-15 Corollary, page 206),
we need only show the small cohomological dimension with respect to
the integers (Nagami, page 199) is finite ($Y$ is paracompact since
it is $\sigma${--}compact and regular (see Kelly \cite{bseventeen},
page 172, exercise Y a) and b))).
To compute $d(Y\Colon \Z)$, look at the map $f\colon Y\ \to\ X$.
$f$ is a closed, onto map. $f$ is onto by construction, and $f$ is
closed since $Y$ is the increasing union of compact sets $\{D_i\}$,
so $F\subseteq Y$ is closed \iff\ $F\cap D_i$ is closed for all $i$, and
$f(F\cap D_i)$ is closed since $F\cap D_i$ is compact and $X$ is Hausdorff.
We can find an increasing sequence of compact sets $\Bar{V}_i$ such
that $E\subseteq X$ is closed \iff\ $E \cap \Bar{V}_i$ is closed.
Since $f$ is proper, $D_i=f^{-1}(\Bar{V}_i$ has the expected properties.
But $f(F\cap D_i)=f(F)\cap\Bar{V}_i$ if $D_i=f^{-1}(\Bar{V}_i)$, so
$f$ is closed.
Hence by Nagami \cite{btwentyseven} (38-4 Theorem, page 216),
$d(Y\Colon \Z)\leq \dim X+k$ where $k$ is the dimension of the fibration $\xi$.
To see this, note $f^{-1}(x)$ is homeomorphic to $S^k$ for all$x\in X$,
so $d(F^{-1}(x)\Colon \Z)=k$.
Since $X$ is paracompact and metrizable, ${\text{ Ind}} X=\dim X = d(X\Colon \Z)=$
dimension of $X$ as a CW complex (see Nagami 8-2 Theorem for the
first equality; Nagami 36-15 Corollary shows the second; Nagami 37-12
Theorem and subdivision show the third
[this uses the fact that $X$ is a regular complex]).
\medskip\insetitem{Step 2}
\LRTpageLabel{ItemFive}
There is a locally compact, finite dimensional CW complex $Z$ and a proper
map $h\colon Z\ \to\ Y$ which is properly $n${--}connected for all $n$.
\smallskip
We define $Z$ and $h$ by induction; i.e. we have
\begin{enumerate}
\item[1)] a finite CW complex $Z_i$ and a map $h_i\colon Z_i\ \to\ Y_i$
\item[2)] $h_i$ is a homotopy equivalence
\item[3)] $h_i$ restricted to $(f_i\ \circ\ h_i)^{-1}(C\cap\Bar{V}_i)$
is a homotopy equivalence for all $C\in {\mathcal S}_r$, $r\geq 1$
\item[4)] $h_i\vert_{(h_{i-1}\circ f_{i-1})^{-1}(U_{i-1})}=
h_{i-1}\vert_{(h_{i-1}\circ f_{i-1})^{-1}(U_{i-1})}$
\item[5)] $Z_i=Z_{i-1}\ \cup_\lambda \Bar{C}_i\times S^k$ where
\item[] $\lambda\colon Z_{i-1}\ \cap\ (h_{i-1}\circ f_{i-1})^{-1}(
\Bar{V}_{i-1}\cap \Bar{C}_i)\ \to\
(\Bar{V}_{i-1}\cap\Bar{C}_i)\times S^k$
\item[] is a cellular homotopy equivalence.
\end{enumerate}
\medskip
If we can find such $Z_i$ and $h_i$, we can find $Z$ and $h\colon Z\ \to\ Y$.
$h$ is clearly proper.
$h\vert_{(f\circ h)^{-1}(C)}\colon (f\circ h)^{-1}(C)\ \to \ f^{-1}(C)$
is a homotopy equivalence by 3) for all $C\in{\mathcal S}_r$, $r\geq 1$,
so $h\vert_{(f\circ h)^{-1}(D_i)}$ is a homotopy equivalence where
$\displaystyle D_i=\mathop{\cup}_{j\geq i} \Bar{C}_i$.
Thus $h$ induces isomorphisms on $H^0$ and $H^0_{{\text{ end}}}$,
and $\Delta(h\Colon \pi_s)=0$ for $s\geq 1$.
Hence we are done if we can produce $Z_i$ and $h_i$.
We proceed by induction on $i$. $Z_1=\Bar{V}_1\times S^k$ and
$h_1={\text{ id}}$.
1){--}5) are trivial, so suppose we have $Z_{i-1}$ and $h_{i-1}$.
We have
\[
Z_{i-1}\cap (f_{i-1}\circ h_{i-1})^{-1}(\Bar{V}_{i-1}\cap\Bar{C}_i)
\ \RA{\hskip10pt}\
Y_{i-1}\cap f^{-1}_{i-1}(\Bar{V}_{i-1}\cap\Bar{C}_i)
\ \RA{\ \rho\ }\
(\Bar{V}_{i-1}\cap\Bar{C}_i)\times S^k
\]
Let $\rho^\prime$ be this composition.
Deform $\rho^\prime$ to a cellular map as follows.
For some $r\geq1$, $C\in{\mathcal S}_r$ implies $C\cap\Bar{C}_i=\emptyset$.
Now deform $\rho^\prime$ to a cellular map over each
$C \cap \Bar{C}_i \cap \Bar{V}_{i-1}$ for $C\in{\mathcal S}_r$, all $r\geq1$
and finally to a cellular map over $\Bar{C}_i\cap \Bar{V}_{i-1}$.
Denote this map by $\lambda$.
Let $Z_i=Z_{i-1}\ \cup_{\lambda}\ (\Bar{C}_i\times S^k)$.
We extend $h_{i-1}$ to a homotopy equivalence $h_i\colon Z_i\ \to\ Y_i$
which leaves $h_{i-1}$ fixed on
$(f_{i-1}\circ h_{i-1})^{-1}(U_{i-1})$.
$h_i$ in fact can be chosen to be a homotopy equivalence on each
$(f_i\circ h_i)^{-1}(C\cap\Bar{V}_i)$ be extending inductively over the
various $C\in{\mathcal S}_r$.
1){--}5) hold and we are done.
\end{proof}
\chapter{The Geometric Surgery Groups}
\section{The fundamental theorems of surgery}
\newHead{III.1}
In this section we will prove three results which may be called the
fundamental theorems of surgery.
They constitute all the geometry needed to define surgery groups
and to prove these groups depend only on the proper $1${--}type
of the spaces in question.
These results together with the $s${--}cobordism theorem constitute
the geometry necessary to give a classification of paracompact
manifolds in a given proper homotopy class \`a la Wall \cite{bfortyone},
Chapter 10.
Let \CAT\ denote either TOP, PL or DIFF.
If $X$ is a locally finite, finite dimensional CW $n${--}ad, and if $\nu$ is a
\CAT\ bundle over $X$, then
$\bor_M(X,\nu)$ is the space of cobordism classes of the following triples:
a \CAT\ manifold $n${--}ad $M$, $\dim M=m$; a proper map of $n${--}ads
$f\colon M\ \to\ X$; a stable bundle map $F\colon \nu_M\ \to\ \nu$,
where $\nu_M$ is the normal bundle of $M$ and $F$ covers $f$.
Such a triple is called a {\sl normal map}, and the cobordisms
are called {\sl normal cobordisms}.
\bigski
\BEGIN{T.3.1.1}
Given $\alpha\in\bor_m(X,\nu)$, there is a representative $(M,f,F)$ of
$\alpha$ with $f$ properly $\bigl[\frac{m}{2}\bigr]${--}connected
if $X$ is a space. (\, $[\quad ]=$ greatest integer.)
For a pair $(X,\partial X)$, we have a representative
$\bigl((M,\partial M), f,F)$ with
$f\colon M\ \to\ X$ properly $\bigl[\frac m2\bigr]${--}connected;
$f\colon \partial M\ \to\ \partial X$ properly
$\bigl[\frac {m-1}2\bigr]${--}connected; and the pair map
$f\colon (M,\partial M)\ \to\ (X,\partial X)$ properly
$\bigl[\frac m2\bigr]${--}connected.
If $\partial X\subseteq X$ is properly $0${--}connected, then the map
of pairs may be made properly homologically
$\bigl[\frac{m+1}{2}\bigr]${--}connected provided $m\geq 3$.
\end{Theorem}
\medskip\begin{proof}
The proof follows Wall \cite{bforty}, Theorem 1.4.
(See the remark following his proof.)
We first remark that his Lemma 1.1 is equally valid in our case.
\medski
\BEGIN{L.3.1.1}
Suppose $M$ and $X$ locally compact, finite dimensional CW complexes,
$\psi\colon M\ \to\ X$ a map.
Then we can attach cells of dimension $\leq k$ to $M$ so that the
resulting complex is locally finite and so that the map is
properly $k${--}connected.
\end{Lemma}
\smallskip\begin{proof}
We may assume $\psi$ cellular by the cellular approximation theorem.
Then the mapping cylinder of $\psi$ is a locally compact, finite dimensional
complex, and $(M_\psi,M)$ is a CW pair.
Set $M^\prime =M^k_\psi\ \cup M$.
Note then that $M^\prime$is obtained from $M$ by adding cells of dimension
$\leq k$ and that $M^\prime\ \to\ M_\psi$
is properly $k${--}connected.
\end{proof}
Now given a representative $(N,g,G)$ for $\alpha$, attach handles of
dimension $\leq k$ to $N$ to get $\psi\colon W\ \to\ X$ with
$\partial W=N\ \cup\ M$, $\psi\vert_N=g$, and with $\psi$ covered
by a bundle map which is $G$ over $N$, and $\psi$
is properly $k${--}connected.
The argument that we can do this is the same as for the compact case.
Wall \cite{bfortyone} Theorem 1.1 generalizes immediately to
\medski
\BEGIN{L.3.1.2}
Given $\alpha\in\bor_m(X,\nu)$ with any representative $(M,f,F)$,
any element of $\Delta(f:\pi_k)$ determines proper regular homotopy
class of immersions of a disjoint collection of $S^k\times D^{m-k}$'s into
$M$ for $k\leq m-2=\dim M-2$.
\end{Lemma}
\smallskip\begin{proof}
Precisely as in Wall, \cite{bfortyone} Theorem 1.1, we get a stable
trivialization of the tangent bundle of $M$ over each sphere $S^k$
in our collection.
Given any sphere $S^k$, we see in fact that there is an open submanifold
$U\subseteq M$ such that we get a trivialization of the tangent bundle
of $U$ restricted to $S^k$ which agrees with the one for $\tau_M$.
In fact $U=f^{-1}\bigr($ a small neighborhood of the disc
bounding $f(S^k)\bigl)$ will do (we have momentarily confused $S^k$
with its image in $M$).
Notice that we can pick such a collection of $U$'s to be locally finite.
Now apply Hirsch \cite{bfourteen}, Haefliger \cite{btwelve}, or Lees
\cite{bnineteen} to immerse each $S^k$ in its $U$
with trivial normal bundle.
This is a proper homotopy, so each $\alpha$ determines a proper map
which immerses each sphere.
It is not hard to show any two such immersions which are
properly homotopic are regularly properly homotopic.
\end{proof}
\medskip
If there is an embedding in the proper regular homotopy class of
$\alpha$, we can attach a collection of handles by $\alpha$
and extend our map and bundle map over the resulting trace of
the surgeries.
Notice that in an embedding, all the spheres have disjoint images, so
we can certainly do the surgery.
The map can be extended properly by construction, and one shows the
bundle map extends precisely as in the compact case
(Wall \cite{bfortyone} Theorem 1.1).
\medski
\BEGIN{L.3.1.3}
Given $\alpha\in\bor_m(X,\nu)$ with any representative $(M,f,F)$,
we can do surgery on any element $\beta\in\Delta(f:\pi_k)$
for $m>2k$.
\end{Lemma}
\smallskip\begin{proof}
General position supplies us with an embedding.
\end{proof}
\medskip
We now return to the proof of \fullRef{T.3.1.1}.
By our lemmas, we see that if $m>2k$, we can get $W$ as advertised.
Now $W$ is obtained from $M$ by adding handles of dimension
$\geq (m+1)-k>k+1$, so $M\subseteq W$ is properly $k${--}connected.
In the pairs case, given a representative, we first fix up the boundary
as above.
Then we can attach handles away from the boundary to get the
absolute map fixed up.
The long exact homotopy sequence shows that the pair map is properly
$\bigl[\frac m2\bigr]${--}connected.
If $m$ is even, we are done.
The case for $m=2k+1$ follows Wall \cite{bfortyone} Theorem 1.4.
We may assume that we have $f\colon (M,\partial M)\ \to\ (X,\partial X)$
properly connected up to the middle dimension on each piece.
Let $E$ be the disjoint union of the $(k+1)${--}cells of $M_f - M$.
Then we have a proper map $\partial E\ \to\ M_f$.
Since $\partial E$ is $k${--}dimensional, and since $(M_f,M)$ is properly
$k${--}connected, there is a proper homotopy of the attaching maps into
$M$.
$\partial E=\displaystyle\mathop{\disjointunion}_p S^k_p$, so embed these
spheres in $M$ with trivial normal bundle by Lemmas
\shortFullRef{L.3.1.1} and \shortFullRef{L.3.1.2}.
Join each sphere to $\partial M$ by a locally finite collection of tubes,
one for each sphere.
(Since $H^0_{{\text{ end}}}(X,\partial X)=0$ by hypothesis, and since
$M\ \to\ X$ is properly $1${--}connected (at least), and since
$\partial M\ \to\ \partial X$ is properly $0${--}connected,
$H^0_{{\text{ end}}}(M,\partial M)=0$ so we can do this.)
Note in fact that we need only disturb $\partial M$ in a (\prex{--}assigned)
neighborhood of a set of base points.)
By general position we may assume all these tubes disjoint ($m\geq3$).
Hence we get framed embeddings of a collection of disjoint $D^k$'s.
We may assume (by adding trivial discs if necessary) that the centers
of our discs form a set of base points for $M$.
We claim that if we do these relative surgeries we will have killed
$K_k(M,\partial M)$ without affecting our other conditions.
Our proof of this claim is the same as Wall's.
Let $H$ denote the union of the handles, $N_0$ the constructed manifold,
$f_0\colon (N_0,\partial N_0)\ \to\ (X,\partial X)$ the resulting map.
Note that $(N_0,\partial N_0)\ \to\ (M, H\ \cup\ \partial N_0)$
is a proper excision map.
We can pick a set of base points for $\partial M$ away from
$\partial M\ \cap\ H$.
As usual we can pick them so that they are a set of base points for
$f\colon \partial M\ \to\ \partial X$.
They are then also seen to be a set of base points for
$M$, $N_0$, $\partial N_0$, and $H\ \cup \partial N_0$.
With these base points and the above excision map we get an exact
sequence\hskip 10pt
$\Delta_k(H\cup\partial N_0,\partial M:M:\coverFA{})\ \to
\Delta_{k+1}(f:\coverFA{})\ \to\ \Delta_{k+1}(f_0:\coverFA{})\ \to\
\Delta_{k-1}(H\cup\partial N_0,\partial M:M:\coverFA{})$.
Clearly the lower relative proper homotopy groups of $f_0$ vanish.
Notice $(H, H\cap\partial M)\ \to\ (H\cup\partial N_0,\partial M)$
is also a proper excision map.
Since $(H, H\cap\partial M)$ is a collection of copies of
$(D^k\times D^{k+1}, S^{k-1}\times D^{k+1})$,
$\Delta_\ast(H, H\cap\partial M)$ is $0$ except in dimension $k$.
If we pick base points in $H$,
$\Delta_\ast(H \cup\partial N_0,\partial M:\coverFB{})=0$
also except in dimension $k$ ($\coverFB{}$ is any covering functor).
Hence $\Delta_{k-1}(H\cup\partial N_0,\partial M:M:\coverFA{})=0$.
Let $g\colon M\ \to\ X$ denote $f$ on $M$ to distinguish it from
$f$ on $(M,\partial M)$.
The collection of elements above generates $\Delta(g:\pi_{k+1})$.
Clearly the composite map \qquad
$\Delta(g:\pi_{k+1}) \ \to\
\Delta_{k+1}(g:\coverFA{})\ \to\
\Delta_{k+1}(f:\coverFA{})\ \to\
\Delta_{k+1}(f_0:\coverFA{})$ is the zero map.
But by Hurewicz, the first map is an isomorphism, and the second map is
onto since $\partial M\ \to\ \partial X$ is properly $k${--}connected.
Hence $\Delta_{k+1}(f:\coverFA{})\ \to\
\Delta_{k+1}(f_0:\coverFA{})$ is the zero map.
Now the last two paragraphs and our exact sequence show
$\Delta_{k+1}(f_0:\coverFA{})=0$ as claimed.
\end{proof}
\bigskip\begin{xRemarks}
Note throughout the proof that should
$\partial X=\partial_1X\ \cup\ \partial_2X$ and
$\partial M=\partial_1M\ \cup\ \partial_2M$, and if
$\partial_2M\ \to\ \partial_2X$ is already properly $r${--}connected,
then we need attach no cells of dimension less than $r$ to $\partial_2M$
in our construction
(provided $H^0_{{\text{ end}}}(X,\partial_1X)=0$, otherwise to get this
last part of the result we must attach some $k${--}cells in $\partial_2M$).
In particular, if $\partial_2M\ \to\ \partial_2X$ is a proper homotopy
equivalence, we can do our construction away from $\partial_2M$
(except possibly for the last step).
\end{xRemarks}
\bigski
\BEGIN{T.3.1.2}
Let $f\colon (M,\partial M)\ \to\ (X,\partial X)$ be a degree one
normal map; i.e. a bundle over $X$ and
a bundle map over $f$ are understood.
Let $(X,\partial X)$ be a Poincar\'e duality pair of formal
dimension at least $6$.
Suppose $\partial X\subseteq X$ is a proper $1${--}equivalence.
Then $f$ is normally cobordant to
$g\colon (N,\partial N)\ \to\ (X,\partial X)$ with $g$ a proper homotopy
equivalence of pairs.
The torsion of $g$ may have any \prex{--}assigned value.
The torsions of $g\colon \partial N\ \to\ \partial X$ and of $g$ as
a map of pairs is then determined.
\end{Theorem}
\medskip\begin{proof}
The proof of the theorem divides into two cases.
\medskip\insetitem{\hskip\parindent Case 1} $\dim(X)=2k$.
\smallskip
By \fullRef{T.3.1.1}, we can do surgery on $f$ to make
the map $f\colon M\ \to\ X$ $k${--}connected, and to make the
map $\partial f\colon \partial M\ \to\ \partial X$
$(k-1)${--}connected (properly connected actually, but we shall be sloppy).
Since $k\geq 3$, $f$, $\partial f$ and $\partial M\subseteq M$ are
all (proper) $1${--}equivalences.
Now subdivide $(M,\partial M)$ until the chain map
$C_\ast(M,\partial M)\ \to\ C_\ast(X,\partial X)$ is onto.
$C_\ast(X,\partial X)$ is $C_\ast(X,\partial X:\Lambda,F)$ for a
collection of paths $\Lambda$ and a lift functor $F$.
The tree for $X$ should come from a tree for $\partial M$,
which we can clearly assume.
$C_\ast(M,\partial M)$ is defined in the same way only with lift
functor $f^{-1}F$.
Let $D_\ast(f)$ be the kernel complex.
Then $H_r\bigl(D_\ast(f)\bigr)=0$ for $r<k$ and
$H^r\bigl(D_\ast(f)\bigr)=0$ for $r>k$.
Now \fullRef{T.1.5.5} shows $H_k\bigl(D_\ast(f)\bigr)$
is an s{--}free\ tree module.
Doing surgery on trivial $(k-1)${--}spheres in $\partial M$ replaces $M$
by its boundary connected sum with a collection of $(S^k\times D^k)$'s.
Hence we may as well assume $H_k\bigl(D_\ast(f)\bigr)$ is free and based.
Let $\{ e_i\}$ be a preferred basis for this module.
By the Namioka Theorem,
$\Delta(f:\pi_{k+1})\ \to\ H_k\bigl(D_\ast(f)\bigr)$ is an
isomorphism.
Thus the $e_i$ determine classes in $\Delta(f:\pi_{k+1})$.
These in turn determine a proper regular homotopy class of immersions
$e_i\colon (D^k\times D^k,\partial D^k\times D^k)\ \to\ (M,\partial M)$.
We claim the $e_i$ are properly regularly homotopic to disjoint embeddings.
It is clearly enough to show this for the restricted immersions
$\bar e_i\colon (D^k,\partial D^k)\ \to\ (M,\partial M)$, for then we
just use small neighborhoods of the $\bar e_i$ to get the $e_i$.
The proof for the $\bar e_i$ proceeds as follows.
Let $C_j$ be an increasing sequence of compact subsets of $M$
with $\displaystyle\mathop{\cap}_j C_j=M$.
Let $C_j$ be such that any element of $\pi_1(M-C_j)$, when pushed into
$\pi_1(M-C_{j-1})$, lies in the image of $\pi_1\bigl(\partial M\ \cap\
(M- C_{j-1})\bigr)$ (compatible base points are understood).
We can do this as $\partial M\subseteq M$
is a proper $1${--}equivalence.
We now proceed.
Only a finite number of the $\bar e_i$ do not lie in $M-C_2$.
Embed these disjointly by the standard piping argument.
Again only finitely many $\bar e_i$ which do lie in $M-C_2$ do not
lie in $M-C_3$.
Put these in general position.
The intersections and self{--}intersections can be piped into
$\partial M \cap(M-C_1)$ without disturbing the $\bar e_i$ we
embedded in the previous step.
This follows from Milnor \cite{btwentyfour}, Theorem 6.6, where we
see that, to do the Whitney trick,
we need only move one of the protagonists.
Hence we can always leave the $\bar e_i$ from the previous steps fixed.
Continuing in this fashion, we can always embed an $\bar e_i$ which
lies in $M-C_{j}$ but not in $M-C_{j+1}$, in $M-C_{j-1}$.
This gives us a proper regular homotopy and establishes our claim.
We next perform handle subtraction.
Let $N$ be obtained from $M$ by deleting the interiors
of the images of the $e_i$.
Let $U$ be the union of the images of the $e_i$.
Let $\partial N=N\cap\partial M$.
By our construction, there is a chain map
$C_\ast\colon U\cup\partial M,\partial M)\ \to\ D_\ast(f)$ such that
\[\begin{matrix}%
0\to&C_\ast(U\cup\partial M,\partial M)&\to&
C_\ast(M,\partial M)&\to& C_\ast(M, U\cup\partial M)&\to0\cr
&\downarrow&&\parallel&&\downarrow\cr
0\to&D_\ast(f)&\to&C_\ast(M,\partial M)&\to&
C_\ast(X,\partial X)&\to0\cr
\end{matrix}\]
chain homotopy commutes.
$C_\ast(U\cup\partial M,\partial M)$ has homology only in
dimension $k$ where it is $H_k\bigl(D_\ast(f)\bigr)$.
The map $C_\ast(U\cup\partial M,\partial M)\ \to\ D_\ast(f)$
gives this isomorphism in homology by construction.
Hence $C_\ast(M, U\cup\partial M)\ \to\ C_\ast(X,\partial X)$
is a chain equivalence.
Now $(N,\partial N)\subseteq (M, U\cup\partial M)$ is a proper
excision map, so $g\colon (N,\partial N)\ \to\ (X,\partial X)$ is
a proper homotopy equivalence from $n$ to $X$.
It induces proper homology isomorphisms on $\partial N\ \to\ \partial X$
and is thus a proper homotopy equivalence there since
$\partial X\subseteq X$ is $1${--}connected.
Hence $g$ is a proper homotopy equivalence of pairs.
By adding an $h${--}cobordism to $\partial N$, we can achieve any
torsion we like for the map $g\colon N\ \to\ X$.
Notice we have not assumed $X$ is a simple Poincar\'e duality $n${--}ad,
but even so, the torsion of $g$ determines the torsions of the
remaining maps.
We leave it to the reader to derive the standard formulas and
remark that if $g$ is simple and if $X$ is simple,
then $g$ is a simple proper homotopy
equivalence of $n${--}ads.\footnote[1]{This bit replaces the original
discussion which had assumed $X$ was simple.}
\medskip\insetitem{\hskip\parindent Case2} $\dim X=2k+1$.
\smallskip
This time, \fullRef{T.3.1.1} permits us to suppose that $f$
induces $k${--}connected maps $M\ \to\ X$ and
$\partial M\ \to\ \partial X$, and moreover we may assume
$K_k(M,\partial M)=0$.
Hence we get a short exact sequence of modules
$0\to\ K_{k+1}(M,\partial M)\ \to\ K_k(\partial M)\ \to\
K_k(M)\ \to0$.
( $K_\ast(M)$ is the tree of modules which is the kernel of the map
$H_\ast\bigl(C(M:\Lambda, f^{-1}F)\bigr)\ \to\
H_\ast\bigl(C(X:\Lambda^\prime, F)\bigr)$.
The other $K${--}groups are defined similarly.)
\fullRef{T.1.5.5} now tells us that each of these modules is s{--}free.
As before we can perform surgery on trivial $(k+1)${--}spheres in
$\partial M$ to convert all of the above modules to free modules.
Again we get a locally finite collection of immersions
$\bar e_i\colon (D^{k+1},\partial D^{k+1})\ \to\ (M,\partial M)$
representing a basis of $K_{k+1}(M,\partial M)$.
We can no longer modify the $\bar e_i$ by a proper regular homotopy
to get disjoint embeddings (we could do this if $\partial M
\subseteq M$ were properly $2${--}connected) but by the same
sort of argument as in the first part, we can modify the $\bar e_i$
until $\bar e_i\vert_{\partial D^{k+1}}$ is a collection of disjoint
embeddings.
The rest of the proof is the same as Wall's.
We have represented a basis of $K_{k}(\partial M)$ by framed,
disjoint embeddings $S^k\ \to\ \partial M$.
Attach corresponding $k+1${--}handles to $M$, thus performing surgery.
Let $U$ be the union of the added handles, and let $(N,\partial N)$
be the new pair.
Since our spheres are null homotopic in $M$, $M$ is just replaced
(up to proper homotopy type) by $M$ with $(k+1)${--}spheres
wedged on in a locally finite fashion.
Hence $K_k(N)$ is free, with a basis given by these spheres.
Dually, the exact sequence of the triple
$\partial N\subseteq \partial N\cup U\subseteq N$, reduces,
using excision, to
\[0\to\ K_{k+1}(N,\partial N)\ \to\
K_{k+1}(M,\partial M)\ \to\
K_{k}(U, U\cap\partial N:M)\ \to\
K_{k}(N,\partial N)\ \to0\ .\]
The map $K_{k+1}(M,\partial M)\ \to\
K_{k}(U, U\cap \partial N:M)$ is seen to be zero since it factors as
$K_{k+1}(M,\partial M)\ \to\ K_k(\partial M)\ \to\
K_k(U:M)\ \to\ K_k(U, U\cap\partial N:M)$ and $K_k(U:M)$ is zero.
(Note that in this composition, $K_k(\partial M)$ should be a subspace
group, but such a group is isomorphic to the absolute group in our case.)
Since $K_k(U, U\cap\partial N:M)$ is free, so is $K_k(N,\partial N)$
and $K_{k+1}(N,\partial N)\cong K_{k+1}(M,\partial M)$.
The attached handles correspond to a basis of $K_{k+1}(M, \partial M)$,
so the map \\$K_{k+1}(N)\ \to\ K_{k+1}(M,\partial M)$ is an
epimorphism, since $K_{k+1}(N)$ is free and based on a set of generators
for $K_{k+1}(M,\partial M)$ and the map takes each basis element
to the corresponding generator.
But $K_{k+1}(M, \partial M)$ is free on these generators, so this map
is an isomorphism.
Hence $K_{k+1}(N)\ \to\ K_{k+1}(N, \partial N)$ is an isomorphism.
Now, by Poincar\'e duality, $K^k(N, \partial N)\ \to\ K^k(N)$ is an
isomorphism.
The natural maps
$K^k(N,\partial N)\ \to\ \bigl(K_k(N,\partial N)\bigr)^\ast$ and
$K^k(N)\ \to\ \bigl(K_k(N)\bigr)^\ast$
are isomorphisms by \fullRef{C.1.5.4.2} since all the modules
are free.
Hence the map $K_k(N)\ \to\ K_k(N,\partial N)$ is an isomorphism.
Thus $K_k(\partial N)=0$, so $f$ restricted to $\partial N$ is a
proper homotopy equivalence.
Next choose a basis for $K_k(N)$ and perform surgery on it.
Write $P$ for the cobordism so obtained of $N$ to $N^\prime$ say.
Consider the induced map degree $1$ and Poincar\'e triads
$(P:N\cup(\partial N\times I), N^\prime)\ \to\
(X\times I:X\times 0\cup\partial X\times I,X\times 1)$.
We will identify $N\cup(\partial N\times I)$ with $N$.
In the exact sequence
\[0\to\
K_{k+1}(N)\ \to\
K_{k+1}(P)\ \to\
K_{k+1}(P,N))\ \RA{\ d\ }\
K_{k}(N)\ \to\
K_{k}(P)\ \to0\]
the map $d$ is by construction an isomorphism.
Hence $K_k(P)=0$ and $K_{k+1}(N)\ \to\ K_{k+1}(P)$
is an isomorphism.
The dual of $d$ is $K_{k+1}(N,\partial N)\ \to\ K_{k+1}(P,N^\prime)$,
so it is an isomorphism (the map is the map induced by the inclusion).
Now, since $f$ on $\partial N$ is a proper homotopy equivalence,
$K_{k+1}(N)\ \to\ K_{k+1}(N,\partial N)$ is an isomorphism.
$K_{k+1}(N)\ \to\ K_{k+1}(P)$ is an isomorphism, so
$K_{k+1}(P)\ \to\ K_{k+1}(P,N^\prime)$ is an isomorphism.
Thus in the sequence
\[0\to\
K_{k+1}(N^\prime)\ \to\
K_{k+1}(P)\ \to\
K_{k+1}(P,N^\prime)\ \to\
K_{k}(N^\prime)\ \to0\]
we have $K_{k+1}(N^\prime)=K_k(N^\prime)=0$, so
$N^\prime\ \to\ X$ is a proper homotopy equivalence.
$\partial N^\prime\ \to\ \partial X$ is the same as
$\partial N\ \to\ \partial X$ (i.e. we did nothing to $\partial N$ as all
our additions were in the interior of $N$) and therefore is a proper
homotopy equivalence.
Hence we have an equivalence of pairs.
The statement about torsions is proved the same way as for Case 1.
\end{proof}
\bigskip\begin{xRemarks}
Note that our proof is still valid in the case $\partial X=\partial_1X\ \cup\
\partial_2X$ provided $\partial_1M\ \to\ \partial_1X$ is a
proper homotopy equivalence (of pairs if $\partial_1X\ \cap\ \partial_2X
\neq\emptyset$) and $\partial_2X\subseteq X$ is a proper
$1${--}equivalence ( $(X:\partial_1X,\partial_2X)$ should be a Poincar\'e
triad).
The proof is word for word the same after we note that $K_i(\partial_2N)
\ \to\ K_i(\partial M)$ is always an isomorphism and that
we may attach all our handles away from $\partial_1M$.
By induction we can prove a similar theorem for $n${--}ads, which is
the result we needed to prove \fullRef{T.2.2.13}.
\end{xRemarks}
\medskip
Our approach to surgery is to consider the surgery groups as bordism
groups of surgery maps.
To make this approach work well, one needs a theorem like
\fullRef{T.3.1.3} below.
\bigskip\begin{xDefinition}
Given a Poincar\'e duality $n${--}ad,
a {\sl surgery map\/} is a map $f\colon M\ \to\ X$
where $M$ is a \CAT{--}manifold $n${--}ad,
$f$ is a degree $1$ map of $n${--}ads,
and there is a \CAT{--}bundle $\nu$ over $X$ and a bundle map
$F\colon \nu_M\ \to\ \nu$ which covers $f$.
\end{xDefinition}
Given a locally finite CW $n${--}ad $K$ with a class $\wone\in H^1(K;\cy2)$,
we say \hfill\penalty-10000
$M\ \RA{\ f\ }\ X\ \RA{\ g\ }\ K$ is a {\sl surgery map over
$(K,\wone)$\/} provided $g$ is a map of $n${--}ads with $g^\ast\wone$
equal to the first Stiefel{--}Whitney class of $X$, and provided
$f$ is a surgery map.
Two surgery maps over $(K,\wone)$ are said to be {\sl bordant\/}
(over $(K,\wone)$) if there is a surgery $(n+1)${--}ad
$W\ \to\ \RA{\ F\ }\ Y\ \RA{\ G\ }\ (K\times I,\wone)$
which is one of the surgery maps on $K\times 0$ and the other
on $K\times 1$.
\bigski
\BEGIN{T.3.1.3}
Let $M\ \RA{\ f\ }\ X\ \RA{\ g\ }\ K$ be a surgery map over $(K,\wone)$,\
a $3${--}ad.
Suppose the formal dimension of $X$ is at least $6$.
Then, if $f\vert_{\partial_1M}$ is a proper homotopy equivalence,
and if $\partial_2K\subseteq K$ is a proper $1${--}equivalence,
we can find another surgery map
$N\ \RA{\ h\ }\ Z\ \RA{\ i\ }\ K$ over $(K,\wone)$ with $h$
a proper homotopy equivalence of $3${--}ads, and with $i$
bordant over $(K,\wone)$ to $g$ so that over $\partial_1K\times I$
the bordism map is $\partial_1M\ \to\ \partial_1X$ crossed with $I$.
\end{Theorem}
\medskip\begin{proof}
If $\partial_2X\subseteq X$ were a proper $1${--}equivalence we could
finish easily using \fullRef{T.3.1.2}.
The proof then consists of modifying $X$ and $\partial_2X$
to get this condition.
The idea is to do surgery, first on $\partial_2X$ (and then on $X$)
to get $\partial_2X\ \to\ \partial_2K$ a proper $1${--}equivalence
(similarly for $X\ \to\ K$) and then show that we can cover these
surgeries on $\partial_2M$ and $M$.
Look at the map $g\colon \partial_2X\ \to\ \partial_2K$.
By \fullRef{T.2.3.2}, $\partial_2X$ can be replaced by
$L\ \cup\ H$, where $H$ is a manifold and $L$ satisfies $D(n-3)$,
where $n$ is the formal dimension of $X$.
This replacement does not alter the bordism class in which we are working.
Let $\wone$ also denote the restriction of $\wone\in H^1(K;\cy2)$
to $\partial_2K$.
Let $\nu$ be the line bundle over $\partial_2K$ classified by $\wone$.
Let $g\colon H\ \to\ \partial_2K$ denote the induced map.
Then $\tau_H\oplus g^\ast\nu$ is trivial, for $H$ has the homotopy type
of a $1${--}complex so the bundle is trivial \iff\ its first Stiefel{--}Whitney
class vanishes (and $\wone(\tau_H\oplus g^\ast\nu)=0$ by
construction).
Hence we can find a bundle map $F\colon \nu_H\ \to\ \nu$.
By \fullRef{T.3.1.1} we can add $1$ and $2$ handles to
to $H$ to get $W$ with
$\partial W=H\ \cup\ H^\prime\ \cup\ \partial H\times I$ and a map
$G\colon W\ \to\ \partial_2K$ with $G\vert_H=h$ and
$G\vert_{H^\prime}$ a proper $1${--}equivalence.
Let $Y= L\times I\ \cup\ W$ by gluing $\partial H\times I$ to $L\times I$
via the map $\partial H\ \to\ L$ crossed with $I$.
$(Y: L\ \cup_{\partial H}\ H,
L\ \cup_{\partial H}\ H^\prime\ \cup\ L\times I)$ is a Poincar\'e duality triad.
This follows since $(L,\partial H)$ is a Poincar\'e duality pair and $Y$ is
$(L,\partial H)\times I$ glued to the manifold triad
$(W: H, \partial H\times I, H^\prime)$ along $\partial H\times I$.
$(L,\partial H)\times I$ is a Poincar\'e triad by \fullRef{T.2.2.9},
and we can glue by \fullRef{T.2.1.3} and \fullRef{T.2.2.7}.
Let $Z=L\ \cup_{\partial H}\ H^\prime$.
We have a map of $Y\ \to\ K\times I$ given by $L\ \to\ K$
crossed with $I$ on $L\times I$ and by $W\ \to\ K\times I$ on $W$.
We claim the restriction $Z\ \to\ \partial_2K\times 1$ is a proper
$1${--}equivalence.
To see this, note first that
\topD{0}{$\begin{matrix}%
\partial H&\subseteq&H^\prime\cr
\cap\hskip .2pt\vrule width .3pt height 6pt depth .1pt&&
\cap\hskip .2pt\vrule width .3pt height 6pt depth .1pt\cr
L&\subseteq&Z\cr
\end{matrix}$}{5}
is a pushout.
$\partial H\subseteq L$ is properly $1${--}connected by construction
(see \fullRef{T.2.3.2}).
It follows from a Mayer{--}Vietoris argument that $H^\prime \subseteq Z$
induces isomorphisms on $H^0_{{\text{ end}}}$ and $H^0$.
Since $\Delta(\partial H:\pi_1)\ \to\ \Delta(L:\pi_1)$ is onto, it
follows from a van{--}Kampen argument that
$\Delta(H^\prime:\pi_1)\ \to\ \Delta(Z:\pi_1)$ is onto.
Now consider $H^\prime\subseteq Z\ \to\ \partial_2K$.
The composite is a proper $1${--}equivalence by construction.
The first map is properly $1${--}connected, as we saw in the last
paragraph.
It then follows that $Z\ \to\ \partial_2K$ is a proper $1${--}equivalence.
It is easy to extend our bundle $\nu$ over all of $Y$.
Wall \cite{bfortyone} pages 89{--}90 shows how to cover our
surgeries back in $\partial_2M$.
One changes $f\colon\partial_2M\ \to\ \partial_2X$ through a proper
homotopy until it is transverse regular to all our core spheres in
$H\subseteq\partial_2X$.
The inverse image of a core sphere back in $\partial_2M$ will be a collection
of disjoint spheres, and Wall shows that, if we do our surgery
correctly on these spheres, then we can extend all our maps and
bundles.
Hence we get $F\colon P\ \to\ Y$ and a bundle map $\nu_P\ \to\ \nu$,
where $\nu$ is the extended $\nu$ over $Y$.
Thus our original problem $M\ \to\ X\ \to\ K$ is normally cobordant
over $(K,\wone)$ to a problem for which $\partial_2X\ \to\ \partial_2K$
is a proper $1${--}equivalence.
We have not touched $\partial_1M\ \to\ \partial_1X$ so we still have
that this map is a proper homotopy equivalence.
In fact, the part of $\partial P$ over $\partial_1M$ is just a product.
Now use \fullRef{T.2.3.3} on $X$ and proceed as above to
get a problem for which $X\ \to\ K$ is a proper $1${--}equivalence.
Note that we need never touch $\partial X$ so
$\partial_1M\ \to\ \partial_1X$ is still a proper homotopy equivalence
and $\partial_2X\ \to\ \partial_2K$
is still a proper $1${--}equivalence.
\end{proof}
\bigskip
\section{Paracompact surgery{--}patterns of application}
\newHead{III.2}
It has been noted by several people (see especially Quinn
\cite{btwentynine} or \cite{bthirty}) that the theorems in section 1,
the $s${--}cobordism theorem, and transverse regularity are all the
geometry one needs to develop a great deal of the theory of surgery.
We define surgery groups as in Wall \cite{bfortyone} Chapter 9.%
\footnote[1]{%
As pointed out
by Farrell and Hsiang, \cite{boneone} (pages 102-103) there is
a problem with Wall's definition of geometric surgery groups
which will affect us as well.
Many people have used variants of this definition.
Ranicki's annotation of Wall's book
\cite{bonefour}, page 92, mentions the problem.
In \cite{bonethreeA} I show that Wall's original definition is almost right.
Wall's problem is a cavalier treatment of local coefficients and the
treatment presented in \cite{bonethreeA} works here as well.}
Let $K$ be a locally compact CW $n${--}ad, and let $w\in H^1(K;\cy2)$
be an orientation.
An {\sl object of type $n$ over $(K,\wone)$\/} is a surgery map $f$
(see section 1) over $s_n K$ for which, if
$M\ \to\ X\ \to\ s_n K$ is the surgery map,
$\varphi\colon \partial_n M\ \to\ \partial_n X$ is a proper homotopy
equivalence of $n${--}ads.%
We write $(\varphi,f)\sim 0$ to denote the existence of a surgery map
over $(s_n K,\wone)$ such that $\partial_{n+1}$ is $(\varphi,f)$;
i.e. if $W\ \to\ Z\ \to\ s_{n+1}s_n K$ is the surgery map,
$\partial_{n+1}W\ \to\ \partial_{n+1}Z\ \to\ s_n K$ is our
original surgery problem; and such that
$\partial_n$ is a proper homotopy equivalence of $(n+1)${--}ads.
$(\varphi,f)\sim(\varphi_1,f_1)$ provided
$(\varphi,f) + -(\varphi_1,f_1)\sim 0$, where $+$ denotes disjoint union
and $-(\varphi,f)$ denotes the same object but with the reverse
orientation.
Write $L^h_m(K,\wone)$ for the group of objects of type $n$ and
dimension $m$ (i.e. $m$ is the dimension of $M$) modulo the
relation $\sim$.
One checks that $\sim$ is an equivalence relation and that disjoint
union makes these sets into abelian groups.
If we require all the torsions of all the proper homotopy equivalences
in the above definitions to be $0$ (including the torsions for the
Poincar\'e duality $n${--}ads), we get groups $L^s_m(K,\wone)$.
If $c\subseteq\sieb(K)$ is a subgroup closed under the involution
induced by the orientation $\wone$, then we get groups
$L^c_m(K,\wone)$ by requiring all the torsions (including those
for the Poincar\'e duality $n${--}ads) to lie in $c$.
( $\sieb(K)$ is Siebenmann's group of proper simple homotopy types;
see Chapter 1, section 5, or \cite{bthirtythree}).
\bigski
\BEGIN{T.3.2.1}
Let $\alpha\in L^c_m(K,\wone)$, $n+m\geq 6$.
Then if $M\ \RA{\ \varphi\ }\ X\ \RA{ f\ }\ K$ is a representative
of $\alpha$ with $f$ a proper $1${--}equivalence, $\alpha=0$ \iff\
there is a normal cobordism $W\ \to\ X\times I$ with
$\partial_-W\ \to\ X\times0$ our original map $\varphi$, and
$\partial_+W\ \to\ X\times 1$ a proper homotopy equivalence of
$n${--}ads with torsions lying in $c$.
\end{Theorem}
\medskip\begin{proof}
Standard from \fullRef{T.3.1.2}, by doing surgery on the
boundary object.
\end{proof}
\bigski
\BEGIN{T.3.2.2}
\[\cdots \to\ L^c_m(\partial_n K,\wone)\ \to\
L^c_m(\delta_n K,\wone)\ \to\
L^c_m(K,\wone)\ \to\
L^c_{m-1}(\partial_n K,\wone)\ \to\ \cdots\]
is exact.
\end{Theorem}
\medskip\begin{proof}
A standard argument.
\end{proof}
\bigski
\BEGIN{T.3.2.3}
If $f\colon K_1\ \to\ K_2$ is a proper map of $n${--}ads, we get an
induced map
$L^c_m(K_1,f^\ast\wone)\ \to\ L^{c^\prime}_m(K_2,\wone)$ where
$f_\#(c)\subseteq c^\prime$, $f_\#\colon\sieb(K_1)\ \to\ \sieb(K_2)$
If $f$ is a proper $1${--}equivalence, the induced map is an isomorphism
for $c=f^{-1}_\#(c^\prime)$.
\end{Theorem}
\medskip\begin{proof}
The induced map is easily defined by
$M\ \to\ X\ \to\ K_1$ goes to
$M\ \to\ X\ \to\ K_1\ \RA{\ f\ }\ K_2$.
For the last statement, if $m\geq 5$ this is just \fullRef{T.3.1.3},
if $K_1$ and $K_2$ are $1${--}ads.
If $K_1$ and $K_2$ are $n${--}ads, an induction argument shows the
result for $n+m\geq6$.
The result is actually true in all dimensions and a proof can be given
following Quinn's proof in the compact case ( see \cite{btwentynine} or
\cite{bthirty}).
We will not carry it out here.
\end{proof}
\bigski
\BEGIN{T.3.2.4}
Let $K$ be a $1${--}ad, and let $M^m\ \RA{\ \varphi\ }\
X\ \RA{\ f\ }\ K$ be a surgery map over $(K,\wone)$ with
$\varphi$ a proper homotopy equivalence and with $f$ a proper
$1${--}equivalence.
Suppose given $\alpha\in L^c_{m+1}(K,\wone)$, $m\geq5$,
and suppose all the torsions for $\varphi$ lie in $c$.
Then there is an object of type $1$, $W\ \to\ X\times I\ \to\ K$,
over $(K,\wone)$ with $\partial W= M\ \cup\ N$, $N\to\ X\times 1$
a proper homotopy equivalence whose torsion also lies in $c$,
and such that the surgery obstruction for this problem is $\alpha$.
\end{Theorem}
\medskip\begin{proof}
The proof is basically Quinn's (see \cite{btwentynine}).
Given $\alpha$, there is always an object of type $1$,
$P\ \to\ Z\ \to\ K$, whose obstruction is $-\alpha$.
(We may always assume $\partial P$ and $\partial Z$ are non{--}empty
by removing a disc from $Z$ and its inverse image in $P$, which we
can modify to be a disc.)
$M\times I\ \to\ X\times I\ \to\ K$ is also an object of type $1$
over $K$.
Take the boundary connected sum of $Z$ and $X\times I$ by extending
$\partial Z\ \#\ X\times 0$ (we may always assume $X$ and $Z$ are
in normal form so we may take this sum in their discs).
Similarly we may extend $\partial P\ \#\ M\times 0$.
We get a new object of type $1$,
$P\ \#_{M\times 0}\ M\times I\ \to\
Z\ \#_{X\times 0}\ X\times I\ \to\ K$.
By the proof of \fullRef{T.3.1.3}, we may do surgery on
$Z\ \#_{X\times0}\ X\times I$ until the map of it to $K$ is a proper
$1${--}equivalence, and we may cover this by a normal cobordism
of $P\ \#_{M\times0}\ M\times I$.
In doing this, we need never touch $M\times 1$ or $X\times 1$.
Let $P^\prime\ \to\ Z^\prime\ \to\ K$ denote this new
object of type $1$.
Note that it still has surgery obstruction $-\alpha$.
Now using \fullRef{T.3.1.2}, we can do surgery on
$\varphi\colon P^\prime\ \to\ Z^\prime$ where $Z^\prime$ is
considered to be the triad $(Z^\prime: X\times1,$ any other
boundary components$)$.
$\varphi$ restricted to the other boundary components is a
proper homotopy equivalence, so we may do surgery leaving them fixed
($X\times 1\subseteq Z^\prime$ is a proper $1${--}equivalence).
Let $W$ be the normal cobordism obtained over $M\times 1$.
Then $W\ \to\ X\times 1\times I$ is a surgery map,
$\partial_{-}W\ \to\ X\times 1\times 0$ is our old map, and
$\partial_+W\ \to\ X\times 1\times 1$ is a proper homotopy equivalence.
We can make all our torsions lie in $c$, and then the surgery obstruction
for $W\ \to\ X\times I\ \to\ K$ must be $\alpha$.
\end{proof}
\bigskip\begin{xDefinition}
Let $\sts_{\CAT}(X)$, for $X$ a Poincar\'e duality space of dimension $n$, be
the set of all simple, degree $1$, homotopy equivalences
$\varphi\colon N^n\ \to\ X$ ($N$ a \CAT{--}manifold) modulo the
relation $\varphi\sim\psi$ \iff\ there is a \CAT{--}homeomorphism $h$
such that
\topD{0}{$\begin{matrix}%
N&\lower 10pt\hbox{$\searrow$}\hbox to 0pt{$\scriptstyle \varphi$\hss}\cr
\downlabeledarrow{}{h}&&X\cr
M&\raise 10pt\hbox{$\nearrow$}\hbox to 0pt{$\scriptstyle {\psi}$\hss}\cr
\end{matrix}$}{5}
properly homotopy commutes.
\end{xDefinition}
A similar definition holds for $X$ a Poincar\'e $n${--}ad.
\bigski
\BEGIN{T.3.2.5}
There is an exact structure sequence
\[\cdots\ \to\ \big[\Sigma X, F/\CAT\bigr]\ \to\
L^s_{m+1}(X,\wone)\ \to\ \sts_{\CAT}(X)\ \to\
\bigl[X, F/\CAT\bigr]\ \RA{\ \theta\ }\ L^s_{m}(X,\wone)\]
where $\wone$ is the first Stiefel{--}Whitney class of the Poincar\'e duality
space $X$ with dimension of $X$ being $m\geq5$.
We also insist that the Spivak normal fibration of $X$ lift to a
\CAT{--}bundle.
By exactness we mean the following.
First of all $\sts_{\CAT}(X)$ may be empty, but in any case,
$\theta^{-1}(0)$ is the image of $\sts_{\CAT}(X)$.
If $\sts_{\CAT}(X)$ is not empty, then $L^s_{m+1}(X,\wone)$
acts on it, and two elements of $\sts_{\CAT}(X)$ which agree in
$\bigl[ X, F/\CAT\bigr]$ differ by an element of this action.
The sequence continues infinitely to the left.
($\Sigma X$ is the ordinary suspension of $X$.)
\end{Theorem}
\medskip\begin{proof}
See Wall \cite{bfortyone}, Chapter 10.
\end{proof}
\bigski
\BEGIN{T.3.2.6}
Let $\bar{\hskip10pt}$ be the involution defined on $\sieb(K)$
in Chapter 1, section 5. Define
$A_m(K,\wone) = H^m\bigl(\cy2, \sieb(K)\bigr)$
where $\sieb(K)$ is made into a $\cy2${--}module by the involution
$\bar{\hskip10pt}$ (which depends on $\wone$).
If $K$ is an $n${--}ad, then
\[\cdots \to\
A_{m+1}(K,\wone)\ \to\
L^s_m(K,\wone)\ \to\
L^h_m(K,\wone)\ \to\
A_m(K,\wone)\ \to\ \cdots\]
is exact for $m+n\geq 6$.
\end{Theorem}
\medskip\begin{proof}
The map $L^s\ \to\ L^h$ is just the forgetful map.
The map $L^h\ \to\ A$ just takes the torsion of the part of the boundary
that was a proper homotopy equivalence and maps it into $A$
(if the proper homotopy equivalence is over more than one
component, sum the torsions).
The map $A\ \to\ L^s$ takes a proper homotopy equivalence
$M^m\ \to\ X$ whose torsion hits an element in $A_{m+1}$, and
maps it to the obstruction to surgering the map to a simple
homotopy equivalence.
See Shaneson \cite{bthirtyone} for the details of proving these
maps well{--}defined and the sequence exact.
\end{proof}
\medski
\BEGIN{C.3.2.6.1}
If $A^c_m(K,w)= H^m(\cy2, c)$,
\[\cdots \to\
A^c_{m+1}(K,\wone)\ \to\
L^s_m(K,\wone)\ \to\
L^c_m(K,\wone)\ \to\
A^c_m(K,\wone)\ \to\ \cdots\]
is exact for $m+n\geq 6$.
\end{Corollary}
\medskip
We now produce our major computation.
\bigski
\BEGIN{T.3.2.7}
\footnote[1]{\text{
Note added in proof\ $^2$:
Compare Maumary, \emph{The open surgery
obstruction in odd dimensions}}. Notices Amer. Math. Soc.
17 (number 5) p.848.\hfill\par $^2$Footnote $1$ is an original footnote. A better
reference is Maumary \cite{bonetwo}.}
Let $(L,\partial L)$ be a finite CW pair.
Form a new CW $n${--}ad $K$ by
$K=L\ \cup\ \partial L\times[0,\infty)$.
Suppose $\partial L$ is the disjoint union of subcomplexes
$\partial_i L$, $i=1$, \dots, $n$.
Let $L^c_m(L,\wone)$ denote the Wall group for $L$
for homotopy equivalences which are simple over $L$ and
which, over $\partial_i L$, have torsions in $c_i$, where
$c_i=\ker\Bigl(\wh\bigl(\pi_1(\partial_i L)\bigr)\ \to\
\wh\bigl(\pi_1(L)\bigr)\Bigr)$.
Then there is an isomorphism
$L^c_m(L,\wone)\ \to\ L^s_m(K,\wone)$.
Combining this with Wall's long exact sequence we see
\[\cdots\zt
\mathop{\oplus}_{i=1}^n
L^{c_i}_m\bigl(\pi_1(\partial_i L),\wone\bigr)
\to L^s_m\bigl(\pi_1(L),\wone\bigr)\to
L^s_m\bigl(K,\wone\bigr)\to
\mathop{\oplus}_{i=1}^n
L^{c_i}_{m-1}\bigl(\pi_1(\partial_i L),\wone\bigr)\zt\cdots
\]
for $m\geq 7$.
\end{Theorem}
\medskip\begin{proof}
The map $L^s_m\bigl(\pi_1(L),\wone\bigr)\ \to\
L^s_m\bigl(K,\wone\bigr)$ is given by
$M\ \to\ X\ \to\ L$ goes to
$M\ \cup\ \partial M\times[0,\infty)
\ \to\
X\ \cup\ \partial X\times[0,\infty)\ \to\
L\ \cup\ \partial L\times[0,\infty)$.
Siebenmann's thesis \cite{bthirtytwo} shows this map is a monomorphism.
To show that the map is onto we can assume
$W\ \RA{\ \varphi\ }\ Z\ \to\ K$ is a surgery map and that
$Z$ is a manifold using \fullRef{T.3.2.4} (this
representation theorem is also needed to show injectivity).
By Siebenmann \cite{bthirtytwo}, we can assume $Z$ is collared; i.e.
$\displaystyle Z=N\ \cup\ \bigl(\mathop{\cup}_{i=1}^n
\partial_i N\times[0,\infty)\bigr)$.
By making $\varphi$ transverse regular to the $\partial_i N$, we
get a problem over $L$, say $V\ \to\ N\ \to\ L$.
We claim
$\displaystyle
V\ \cup\ \bigl(\mathop{\cup}_{i=1}^n \partial_i V\times[0,\infty)\bigr)
\ \to\
N\ \cup\ \bigl(\mathop{\cup}_{i=1}^n \partial_i N\times[0,\infty)\bigr)
$
has the same surgery obstruction as $W\ \to\ Z$.
But this is seen by actually constructing the normal cobordism
using Siebenmann's concept of a $1${--}neighborhood and
some compact surgery.
\end{proof}
\medski
\BEGIN{C.3.2.7.1}
We can improve $m\geq7$ to $m\geq6$.
\end{Corollary}
\smallskip\begin{proof}
Using recent work of Cappell{--}Shaneson \cite{bfive},
one can get a modified version of Siebenmann's main theorem.
One can not collar a $5${--}manifold, but one can at least get an
increasing sequence of cobordisms whose ends are
$\partial_i N\ \#\ S^2\times S^2\ \#\ \cdots\ \#\ S^2\times S^2$.
This is sufficient.
\end{proof}
\bigskip
Actually, one would hope that these surgery groups would be
periodic, just as the compact ones are.
This is actually the case, but the only proof I know involves describing
surgery in terms of algebra.
This can be done, but the result is long and will be omitted.
We briefly consider splitting theorems.
The two{--}sided codimension $1$ splitting theorem holds; i.e. if
$W$ has the simple homotopy type of
$Z=(X,\partial X)\ \cup\ (Y,\partial X)$ with
$\partial X\subseteq X$ a proper $1${--}equivalence, then the map
$W\ \to\ Z$ can be split.
The proof is the same as for the compact case.
Hence we also get codimension greater than or equal to $3$
splitting theorems for proper submanifolds.
In fact, most of Wall \cite{bfortyone} Chapter 11
goes over with minor modifications.
We are unable to obtain a one{--}sided splitting theorem in general,
due to the lack of a Farrell fibering theorem in the non{--}compact case.
We also note in passing that one could define surgery spaces as in
\cite{btwentynine} and \cite{bthirty}.
We than get the same basic geometric constructions; e.g. assembly maps
and pullback maps.
We have nothing new to add to the theory, so we leave the reader the
exercise of restating \cite{btwentynine} so that it is valid
for paracompact surgery spaces\setcounter{footnote}{0}\footnote{We presumably
could define proper algebraic bordism groups following Ranicki.
See for example \cite{Ranicki}.}.
\vfill
\newpage
\renewcommand{\rightmark}{Bibliography}
\begin{bibdiv}
\begin{biblist}
\bib{bone}{book}{
author={Bass, Hyman},
title={Algebraic $K$-theory},
publisher={W. A. Benjamin, Inc., New York-Amsterdam},
date={1968},
pages={xx+762},
review={\MR{0249491}},
}
\bib{btwo}{book}{
author={Bredon, Glen E.},
title={Sheaf theory},
publisher={McGraw-Hill Book Co., New York-Toronto, Ont.-London},
date={1967},
pages={xi+272},
review={\MR{0221500}},
}
\bib{bthree}{misc}{
author={Browder, W.},
title={Cap products and Poincar\'e duality},
date={1964},
note={Cambridge University notes \BR{Browder}},
}
\bib{bfour}{misc}{
author={Browder, W.},
title={Surgery on simply{--}connected manifolds},
date={1969},
note={Princeton University notes \BR{Browder}},
}
\bib{bfive}{unpublished}{
author={Cappell, S.},
author={Shaneson, J.},
title={On four dimensional surgery and applications},
note={preprint \BR{CS}},
}
\bib{bsix}{book}{
author={Cooke, George E.},
author={Finney, Ross L.},
title={Homology of cell complexes},
series={Based on lectures by Norman E. Steenrod},
publisher={Princeton University Press, Princeton, N.J.; University of
Tokyo Press, Tokyo},
date={1967},
pages={xv+256},
review={\MR{0219059}},
}
\bib{bseven}{article}{
author={Dold, Albrecht},
title={Partitions of unity in the theory of fibrations},
journal={Ann. of Math. (2)},
volume={78},
date={1963},
pages={223--255},
issn={0003-486X},
review={\MR{0155330}},
}
\bib{beight}{book}{
author={Dold, Albrecht},
title={Halbexakte Homotopiefunktoren},
language={German},
series={Lecture Notes in Mathematics},
volume={12},
publisher={Springer-Verlag, Berlin-New York},
date={1966},
pages={156},
review={\MR{0198464}},
}
\bib{bnine}{article}{
author={Farrell, T.},
author={Wagoner, J.},
title={A torsion invariant for proper $h${--}cobordisms/ an algebraic criterion for a map to be a proper homotopy equivalence},
date={1969},
journal={Notices Amer. Math. Soc.},
volume={16},
pages={988 and 1090},
note={\BR{FWB}},
}
\bib{bten}{unpublished}{
author={Farrell, T.},
author={Wagoner, J.},
title={Infinite matrices in algebraic $K${--}theory and topology},
note={To appear: \BR{FWB}},
}
\bib{beleven}{unpublished}{
author={Farrell, T.},
author={Taylor, L.},
author={Wagoner, J.},
title={The {W}hitehead theorem in the proper category},
note={To appear: \BR{FWA}},
}
\bib{btwelve}{article}{
author={Haefliger, Andr{\'e}},
title={Lissage des immersions. I},
language={French},
journal={Topology},
volume={6},
date={1967},
pages={221--239},
issn={0040-9383},
review={\MR{0208607}},
}
\bib{bthirteen}{book}{
author={Hilton, P. J.},
title={An introduction to homotopy theory},
series={Cambridge Tracts in Mathematics and Mathematical Physics, no. 43},
publisher={Cambridge, at the University Press},
date={1953},
pages={viii+142},
review={\MR{0056289}},
}
\bib{bfourteen}{article}{
author={Hirsch, Morris W.},
title={Immersions of manifolds},
journal={Trans. Amer. Math. Soc.},
volume={93},
date={1959},
pages={242--276},
issn={0002-9947},
review={\MR{0119214}},
}
\bib{bfifteen}{book}{
author={Hu, Sze-tsen},
title={Theory of retracts},
publisher={Wayne State University Press, Detroit},
date={1965},
pages={234},
review={\MR{0181977}},
}
\bib{bsixteen}{book}{
author={Hurewicz, Witold},
author={Wallman, Henry},
title={Dimension Theory},
series={Princeton Mathematical Series, v. 4},
publisher={Princeton University Press, Princeton, N. J.},
date={1941},
pages={vii+165},
review={\MR{0006493}},
}
\bib{bseventeen}{book}{
author={Kelley, John L.},
title={General topology},
publisher={D. Van Nostrand Company, Inc., Toronto-New York-London},
date={1955},
pages={xiv+298},
review={\MR{0070144}},
}
\bib{beighteen}{article}{
author={Kirby, R.},
author={Siebenmann, L.},
title={Foundations of topology},
date={1969},
journal={Notices Amer. Math. Soc.},
volume={16},
note={\BR{KS}},
}
\bib{bnineteen}{article}{
author={Lees, J. Alexander},
title={Immersions and surgeries of topological manifolds},
journal={Bull. Amer. Math. Soc.},
volume={75},
date={1969},
pages={529--534},
issn={0002-9904},
review={\MR{0239602}},
}
\bib{btwenty}{book}{
author={Lefschetz, Solomon},
title={Introduction to Topology},
series={Princeton Mathematical Series, vol. 11},
publisher={Princeton University Press, Princeton, N. J.},
date={1949},
pages={viii+218},
review={\MR{0031708}},
}
\bib{btwentyone}{book}{
author={Lundell, A.},
author={Weingram, S.},
title={The topology of CW complexes},
publisher={Van Nostrand Reinhold Co.},
address={New York},
date={1969},
}
\bib{btwentytwo}{article}{
author={Milnor, John},
title={On spaces having the homotopy type of a ${\rm CW}$-complex},
journal={Trans. Amer. Math. Soc.},
volume={90},
date={1959},
pages={272--280},
issn={0002-9947},
review={\MR{0100267}},
}
\bib{btwentythree}{article}{
author={Milnor, J.},
title={Whitehead torsion},
journal={Bull. Amer. Math. Soc.},
volume={72},
date={1966},
pages={358--426},
issn={0002-9904},
review={\MR{0196736}},
}
\bib{btwentyfour}{book}{
author={Milnor, John},
title={Lectures on the $h$-cobordism theorem},
series={Notes by L. Siebenmann and J. Sondow},
publisher={Princeton University Press, Princeton, N.J.},
date={1965},
pages={v+116},
review={\MR{0190942}},
}
\bib{btwentyfive}{book}{
author={Mitchell, Barry},
title={Theory of categories},
series={Pure and Applied Mathematics, Vol. XVII},
publisher={Academic Press, New York-London},
date={1965},
pages={xi+273},
review={\MR{0202787}},
}
\bib{btwentysix}{article}{
author={Moore, John C.},
title={Some applications of homology theory to homotopy problems},
journal={Ann. of Math. (2)},
volume={58},
date={1953},
pages={325--350},
issn={0003-486X},
review={\MR{0059549}},
}
\vfill
\bib{btwentyseven}{book}{
author={Nagami, Kei{\^o}},
title={Dimension theory},
series={With an appendix by Yukihiro Kodama. Pure and Applied
Mathematics, Vol. 37},
publisher={Academic Press, New York-London},
date={1970},
pages={xi+256},
review={\MR{0271918}},
}
\bib{btwentyeight}{article}{
author={Namioka, I.},
title={Maps of pairs in homotopy theory},
journal={Proc. London Math. Soc. (3)},
volume={12},
date={1962},
pages={725--738},
issn={0024-6115},
review={\MR{0144345}},
}
\bib{btwentynine}{thesis}{
author={Quinn, F.},
title={A geometric formulation of surgery},
type={Ph.D. Thesis},
date={1969},
}
\bib{bthirty}{inproceedings}{
author={Quinn, Frank},
title={A geometric formulation of surgery},
conference={
title={Topology of Manifolds (Proc. Inst., Univ. of Georgia, Athens,
Ga., 1969)},
},
book={
publisher={Markham, Chicago, Ill.},
},
date={1970},
pages={500--511},
review={\MR{0282375}},
}
\bib{bthirtyone}{article}{
author={Shaneson, Julius L.},
title={Wall's surgery obstruction groups for $G\times Z$},
journal={Ann. of Math. (2)},
volume={90},
date={1969},
pages={296--334},
issn={0003-486X},
review={\MR{0246310}},
}
\bib{bthirtytwo}{thesis}{
author={Siebenmann, L.},
title={The obstruction to finding a boundary for an open manifold of dimension greater than five},
type={Ph.D. Thesis},
date={1965},
}
\bib{bthirtythree}{article}{
author={Siebenmann, L. C.},
title={Infinite simple homotopy types},
journal={Nederl. Akad. Wetensch. Proc. Ser. A 73 = Indag. Math.},
volume={32},
date={1970},
pages={479--495},
review={\MR{0287542}},
}
\bib{bthirtyfour}{unpublished}{
author={Siebenmann, L.},
title={A report on topological manifolds},
note={To appear in {\sl Proc. ICM} (Nice 1970), \BR{Sieb}},
}
\bib{bthirtyfive}{book}{
author={Spanier, Edwin H.},
title={Algebraic topology},
publisher={McGraw-Hill Book Co., New York-Toronto, Ont.-London},
date={1966},
pages={xiv+528},
review={\MR{0210112}},
}
\bib{bthirtysix}{article}{
author={Spivak, Michael},
title={Spaces satisfying Poincar\'e duality},
journal={Topology},
volume={6},
date={1967},
pages={77--101},
issn={0040-9383},
review={\MR{0214071}},
}
\bib{bthirtyseven}{article}{
author={Wall, C. T. C.},
title={Finiteness conditions for ${\rm CW}$-complexes},
journal={Ann. of Math. (2)},
volume={81},
date={1965},
pages={56--69},
issn={0003-486X},
review={\MR{0171284}},
}
\bib{bthirtyeight}{article}{
author={Wall, C. T. C.},
title={Finiteness conditions for ${\rm CW}$ complexes. II},
journal={Proc. Roy. Soc. Ser. A},
volume={295},
date={1966},
pages={129--139},
review={\MR{0211402}},
}
\bib{bforty}{article}{
author={Wall, C. T. C.},
title={Surgery of non-simply-connected manifolds},
journal={Ann. of Math. (2)},
volume={84},
date={1966},
pages={217--276},
issn={0003-486X},
review={\MR{0212827}},
}
\bib{bthirtynine}{article}{
author={Wall, C. T. C.},
title={Poincar\'e complexes. I},
journal={Ann. of Math. (2)},
volume={86},
date={1967},
pages={213--245},
issn={0003-486X},
review={\MR{0217791}},
}
\bib{bfortyone}{book}{
author={Wall, C. T. C.},
title={Surgery on compact manifolds},
note={London Mathematical Society Monographs, No. 1},
publisher={Academic Press, London-New York},
date={1970},
pages={x+280},
review={\MR{0431216}},
}
\bib{bfortytwo}{article}{
author={Whitehead, George W.},
title={On spaces with vanishing low-dimensional homotopy groups},
journal={Proc. Nat. Acad. Sci. U. S. A.},
volume={34},
date={1948},
pages={207--211},
issn={0027-8424},
review={\MR{0028027}},
}
\bib{bfortythree}{article}{
author={Whitehead, J. H. C.},
title={Combinatorial homotopy. I},
journal={Bull. Amer. Math. Soc.},
volume={55},
date={1949},
pages={213--245},
issn={0002-9904},
review={\MR{0030759}},
}
\bib{bfortyfour}{book}{
author={Wilder, Raymond Louis},
title={Topology of Manifolds},
series={American Mathematical Society Colloquium Publications, vol. 32},
publisher={American Mathematical Society, New York, N. Y.},
date={1949},
pages={ix+402},
review={\MR{0029491}},
}
\newpage
\centerline{\bf Additional References}
\bib{Browder}{book}{
author={Browder, William},
title={Surgery on simply-connected manifolds},
note={Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 65},
publisher={Springer-Verlag},
place={New York},
date={1972},
pages={ix+132},
review={\MR{0358813} (50 \#11272)},
}
\bib{CS}{article}{
author={Cappell, Sylvain E.},
author={Shaneson, Julius L.},
title={On four dimensional surgery and applications},
journal={Comment. Math. Helv.},
volume={46},
date={1971},
pages={500--528},
issn={0010-2571},
review={\MR{0301750} (46 \#905)},
}
\bib{boneone}{article}{
author={Farrell, F. T.},
author={Hsiang, W. C.},
title={Rational $L$-groups of Bieberbach groups},
journal={Comment. Math. Helv.},
volume={52},
date={1977},
number={1},
pages={89--109},
issn={0010-2571},
review={\MR{0448372}},
}
\bib{FTW}{article}{
author={Farrell, F. T.},
author={Taylor, L. R.},
author={Wagoner, J. B.},
title={The Whitehead theorem in the proper category},
journal={Compositio Math.},
volume={27},
date={1973},
pages={1--23},
issn={0010-437X},
review={\MR{0334226} (48 \#12545)},
}
\bib{FWA}{article}{
author={Farrell, F. T.},
author={Wagoner, J. B.},
title={Infinite matrices in algebraic $K$-theory and topology},
journal={Comment. Math. Helv.},
volume={47},
date={1972},
pages={474--501},
issn={0010-2571},
review={\MR{0319185} (47 \#7731a)},
}
\bib{FWB}{article}{
author={Farrell, F. T.},
author={Wagoner, J. B.},
title={Algebraic torsion for infinite simple homotopy types},
journal={Comment. Math. Helv.},
volume={47},
date={1972},
pages={502--513},
review={\MR{0319185} (47 \#7731b)},
}
\bib{KS}{book}{
author={Kirby, Robion C.},
author={Siebenmann, Laurence C.},
title={Foundational essays on topological manifolds, smoothings, and triangulations},
note={With notes by John Milnor and Michael Atiyah;
Annals of Mathematics Studies, No. 88},
publisher={Princeton University Press},
place={Princeton, N.J.},
date={1977},
pages={vii+355},
review={\MR{0645390} (58 \#31082)},
}
\bib{bonetwo}{inproceedings}{
author={Maumary, Serge},
title={Proper surgery groups and Wall-Novikov groups},
conference={
title={Algebraic $K$-theory, III: Hermitian $K$-theory and geometric
applications (Proc. Conf., Battelle Memorial Inst., Seatlle, Wash.,
1972)},
},
book={
publisher={Springer, Berlin},
},
date={1973},
pages={526--539. Lecture Notes in Math, Vol. 343},
review={\MR{0377938}},
}
\bib{Ranicki}{article}{
author={Ranicki, Andrew},
title={Algebraic Poincar\'e cobordism},
conference={
title={Topology, geometry, and algebra: interactions and new
directions },
address={Stanford, CA},
date={1999},
},
book={
series={Contemp. Math.},
volume={279},
publisher={Amer. Math. Soc., Providence, RI},
},
date={2001},
pages={213--255},
review={\MR{1850750}},
doi={10.1090/conm/279/04563},
}
\bib{Sieb}{article}{
author={Siebenmann, L. C.},
title={Topological manifolds},
conference={
title={Actes du Congr\`es International des Math\'ematiciens},
address={Nice},
date={1970},
},
book={
publisher={Gauthier-Villars},
place={Paris},
},
date={1971},
pages={133--163},
review={\MR{0423356} (54 \#11335)},
}
\bib{bonethree}{inproceedings}{
author={Taylor, L.~R.},
title={Surgery groups and inner automorphisms},
date={1973},
booktitle={{A}lgebraic $K$-theory {III}: Hermitian $K$-theory and
Geometric Applications (proc. conf., Battelle Memorial Inst., Seattle, Wash.,
1972)},
series={{L}ecture {N}otes in {M}ath.},
volume={343},
publisher={{S}pringer{--}{V}erlag},
address={{B}erlin},
pages={471\ndash 477},
}
\bib{bonethreeA}{article}{
author={Taylor, Laurence R.},
title={Unoriented geometric functors},
journal={Forum Math.},
volume={20},
date={2008},
number={3},
pages={457--467},
issn={0933-7741},
review={\MR{2418201}},
doi={10.1515/FORUM.2008.023},
eprint={{https://arxiv.org/abs/math/0606651}{arXiv:math/0606651v1}},
}
\bib{bonefour}{book}{
author={Wall, C. T. C.},
title={Surgery on compact manifolds},
series={Mathematical Surveys and Monographs},
volume={69},
edition={2},
note={Edited and with a foreword by A. A. Ranicki},
publisher={American Mathematical Society, Providence, RI},
date={1999},
pages={xvi+302},
isbn={0-8218-0942-3},
review={\MR{1687388}},
doi={10.1090/surv/069},
}
\newpage
\centerline{\bf Some Papers Citing This Work}
\bib{MR2014912}{article}{
author={Ayala, R.},
author={C{\'a}rdenas, M.},
author={Muro, F.},
author={Quintero, A.},
title={An elementary approach to the projective dimension in proper
homotopy theory},
journal={Comm. Algebra},
volume={31},
date={2003},
number={12},
pages={5995--6017},
issn={0092-7872},
review={\MR{2014912}},
doi={10.1081/AGB-120024863},
}
\bib{MR1848146}{book}{
author={Baues, Hans-Joachim},
author={Quintero, Antonio},
title={Infinite homotopy theory},
series={$K$-Monographs in Mathematics},
volume={6},
publisher={Kluwer Academic Publishers, Dordrecht},
date={2001},
pages={viii+296},
isbn={0-7923-6982-3},
review={\MR{1848146}},
doi={10.1007/978-94-009-0007-3},
}
\bib{MR1758300}{article}{
author={Block, Jonathan},
author={Weinberger, Shmuel},
title={Arithmetic manifolds of positive scalar curvature},
journal={J. Differential Geom.},
volume={52},
date={1999},
number={2},
pages={375--406},
issn={0022-040X},
review={\MR{1758300}},
}
\bib{MR671656}{article}{
author={Farrell, F. T.},
author={Hsiang, W. C.},
title={The stable topological-hyperbolic space form problem for complete
manifolds of finite volume},
journal={Invent. Math.},
volume={69},
date={1982},
number={1},
pages={155--170},
issn={0020-9910},
review={\MR{671656}},
doi={10.1007/BF01389189},
}
\bib{MR973309}{article}{
author={Farrell, F. T.},
author={Jones, L. E.},
title={A topological analogue of Mostow's rigidity theorem},
journal={J. Amer. Math. Soc.},
volume={2},
date={1989},
number={2},
pages={257--370},
issn={0894-0347},
review={\MR{973309}},
doi={10.2307/1990978},
}
\bib{MR1388311}{article}{
author={Ferry, Steven C.},
author={Pedersen, Erik K.},
title={Epsilon surgery theory},
conference={
title={Novikov conjectures, index theorems and rigidity, Vol.\ 2 },
address={Oberwolfach},
date={1993},
},
book={
series={London Math. Soc. Lecture Note Ser.},
volume={227},
publisher={Cambridge Univ. Press, Cambridge},
},
date={1995},
pages={167--226},
review={\MR{1388311}},
doi={10.1017/CBO9780511629365.007},
}
\bib{MR666159}{article}{
author={Freedman, Michael H.},
title={A surgery sequence in dimension four;\ the relations with knot
concordance},
journal={Invent. Math.},
volume={68},
date={1982},
number={2},
pages={195--226},
issn={0020-9910},
review={\MR{666159}},
doi={10.1007/BF01394055},
}
\bib{MR872483}{article}{
author={Freedman, Michael H.},
title={A geometric reformulation of $4$-dimensional surgery},
note={Special volume in honor of R. H. Bing (1914--1986)},
journal={Topology Appl.},
volume={24},
date={1986},
number={1-3},
pages={133--141},
issn={0166-8641},
review={\MR{872483}},
doi={10.1016/0166-8641(86)90054-4},
}
\bib{MR689390}{article}{
author={Hambleton, Ian},
title={Projective surgery obstructions on closed manifolds},
conference={
title={Algebraic $K$-theory, Part II},
address={Oberwolfach},
date={1980},
},
book={
series={Lecture Notes in Math.},
volume={967},
publisher={Springer, Berlin-New York},
},
date={1982},
pages={101--131},
review={\MR{689390}},
}
\bib{MR1268592}{article}{
author={Hambleton, Ian},
author={Madsen, Ib},
title={On the computation of the projective surgery obstruction groups},
journal={$K$-Theory},
volume={7},
date={1993},
number={6},
pages={537--574},
issn={0920-3036},
review={\MR{1268592}},
doi={10.1007/BF00961217},
}
\bib{MR2212277}{article}{
author={Jahren, Bj{\o}rn},
author={Kwasik, S{\l}awomir},
title={Manifolds homotopy equivalent to $\bold R{\rm P}^4\#\bold R{\rm
P}^4$},
journal={Math. Proc. Cambridge Philos. Soc.},
volume={140},
date={2006},
number={2},
pages={245--252},
issn={0305-0041},
review={\MR{2212277}},
doi={10.1017/S0305004105008893},
}
\bib{MR520500}{article}{
author={Jones, Lowell},
title={The nonsimply connected characteristic variety theorem},
conference={
title={Algebraic and geometric topology},
address={Proc. Sympos. Pure Math., Stanford Univ., Stanford, Calif.},
date={1976},
},
book={
series={Proc. Sympos. Pure Math., XXXII},
publisher={Amer. Math. Soc., Providence, R.I.},
},
date={1978},
pages={131--140},
review={\MR{520500}},
}
\bib{MR869712}{article}{
author={Kwasik, S{\l}awomir},
title={On periodicity in topological surgery},
journal={Canad. J. Math.},
volume={38},
date={1986},
number={5},
pages={1053--1064},
issn={0008-414X},
review={\MR{869712}},
doi={10.4153/CJM-1986-051-6},
}
\bib{MR1191376}{article}{
author={Kwasik, S{\l}awomir},
author={Schultz, Reinhard},
title={Vanishing of Whitehead torsion in dimension four},
journal={Topology},
volume={31},
date={1992},
number={4},
pages={735--756},
issn={0040-9383},
review={\MR{1191376}},
doi={10.1016/0040-9383(92)90005-3},
}
\bib{MR1451758}{article}{
author={Kwasik, S{\l}awomir},
author={Schultz, Reinhard},
title={Inductive detection for homotopy equivalences of manifolds},
journal={$K$-Theory},
volume={11},
date={1997},
number={3},
pages={287--306},
issn={0920-3036},
review={\MR{1451758}},
doi={10.1023/A:1007702426750},
}
\bib{MR3349793}{article}{
author={Kwasik, S{\l}awomir},
author={Schultz, Reinhard},
title={Tangential thickness of manifolds},
journal={Proc. Lond. Math. Soc. (3)},
volume={110},
date={2015},
number={5},
pages={1281--1313},
issn={0024-6115},
review={\MR{3349793}},
doi={10.1112/plms/pdv009},
}
\bib{MR579574}{article}{
author={Pedersen, Erik Kjaer},
author={Ranicki, Andrew},
title={Projective surgery theory},
journal={Topology},
volume={19},
date={1980},
number={3},
pages={239--254},
issn={0040-9383},
review={\MR{579574}},
doi={10.1016/0040-9383(80)90010-5},
}
\bib{MR1747539}{article}{
author={Pedersen, Erik Kj{\ae}r},
title={Continuously controlled surgery theory},
conference={
title={Surveys on surgery theory, Vol. 1},
},
book={
series={Ann. of Math. Stud.},
volume={145},
publisher={Princeton Univ. Press, Princeton, NJ},
},
date={2000},
pages={307--321},
review={\MR{1747539}},
}
\bib{MR1944688}{article}{
author={Pedersen, Erik Kj{\ae}r},
title={Controlled methods in equivariant topology, a survey},
conference={
title={Current trends in transformation groups},
},
book={
series={$K$-Monogr. Math.},
volume={7},
publisher={Kluwer Acad. Publ., Dordrecht},
},
date={2002},
pages={231--245},
review={\MR{1944688}},
doi={10.1007/978-94-009-0003-5-15},
}
\bib{MR1818774}{article}{
author={Weiss, Michael},
author={Williams, Bruce},
title={Automorphisms of manifolds},
conference={
title={Surveys on surgery theory, Vol. 2},
},
book={
series={Ann. of Math. Stud.},
volume={149},
publisher={Princeton Univ. Press, Princeton, NJ},
},
date={2001},
pages={165--220},
review={\MR{1818774}},
}
\end{biblist}
\end{bibdiv}
\vfill
\end{document}
|
\section{Introduction}
Each layer (or module) in a directed neural network can be considered a computation step, that transforms its incoming data. These modules are connected via directed edges, creating a forward processing graph which defines the flow of data from the network inputs, through each module, producing network outputs. Defining a loss on outputs allows errors to be generated, and propagated back through the network graph to provide a signal to update each module.
This process results in several forms of \emph{locking}, namely: (i) \emph{Forward Locking} -- no module can process its incoming data before the previous nodes in the directed forward graph have executed; (ii) \emph{Update Locking} -- no module can be updated before all dependent modules have executed in forwards mode; also, in many credit-assignment algorithms (including backpropagation~\citep{Rumelhart86}) we have (iii)~\emph{Backwards Locking} -- no module can be updated before all dependent modules have executed in both forwards mode and backwards mode.
\begin{figure}[t!]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.4\textwidth]{figures/comms_small.pdf}
\end{tabular}
\caption{\small General communication protocol between $A$ and $B$. After receiving the message $h_{A}$ from $A$, $B$ can use its model of $A$, $M_{B}$, to send back \emph{synthetic gradients} $\hat{\delta}_{A}$ which are trained to approximate real error gradients $\delta_{A}$. Note that $A$ does not need to wait for any extra computation after itself to get the correct error gradients, hence decoupling the backward computation. The feedback model $M_{B}$ can also be conditioned on any privileged information or context, $c$, available during training such as a label.}
\label{fig:comms}
\end{figure}
Forwards, update, and backwards locking constrain us to running and updating neural networks in a sequential, synchronous manner. Though seemingly benign when training simple feed-forward nets, this poses problems when thinking about creating systems of networks acting in multiple environments at different and possibly irregular or asynchronous timescales. For example, in complex systems comprised of multiple asynchronous cooperative modules (or agents), it is undesirable and potentially unfeasible that all networks are update locked. Another example is a distributed model, where part of the model is shared and used by many downstream clients -- all clients must be fully executed and pass error gradients back to the shared model before the model can update, meaning the system trains as fast as the slowest client. The possibility to parallelise training of currently sequential systems could hugely speed up computation time.
The goal of this work is to remove update locking for neural networks. This is achieved by removing backpropagation. To update weights $\theta_i$ of module $i$ we drastically approximate the function implied by backpropagation:
\begin{equation*}
\begin{split}
\frac{\partial L}{\partial \theta_i} & = f_\text{Bprop}((h_i, x_i, y_i, \theta_i),\ldots)\frac{\partial h_i}{\partial \theta_i}
\\ &\simeq~ \hat{f}_\text{Bprop}(h_i)\frac{\partial h_i}{\partial \theta_i}
\end{split}
\end{equation*}
where $h$ are activations, $x$ are inputs, $y$ is supervision, and $L$ is the overall loss to minimise. This leaves dependency only on $h_i$ -- the information local to module $i$.
The premise of this method is based on a simple protocol for learnt communication, allowing neural network modules to interact and be trained without update locking. While the communication protocol is general with respect to the means of generating a training signal, here we focus on a specific implementation for networks trained with gradient descent -- we replace a standard \emph{neural interface} (a connection between two modules in a neural network) with a Decoupled Neural Interface (DNI). Most simply, when a module (\emph{e.g.}~a layer) sends a message (activations) to another module, there is an associated model which produces a predicted error gradient with respect to the message immediately. The predicted gradient is a function of the message alone; there is no dependence on downstream events, states or losses. The sender can then immediately use these \emph{synthetic gradients} to get an update, without incurring any delay. And by removing update- and backwards locking in this way, we can train networks without a synchronous backward pass. We also show preliminary results that extend this idea to also remove forward locking -- resulting in networks whose modules can also be trained without a synchronous forward pass. When applied to RNNs we show that using synthetic gradients allows RNNs to model much greater time horizons than the limit imposed by truncating backpropagation through time (BPTT). We also show that using synthetic gradients to decouple a system of two RNNs running at different timescales can greatly increase training speed of the faster RNN.
Our synthetic gradient model is most analogous to a value function which is used for gradient ascent~\citep{Baxter00} or critics for training neural networks~\citep{schmidhuber1990networks}. Most other works that aim to remove backpropagation do so with the goal of performing biologically plausible credit assignment, but this doesn't eliminate update locking between layers. \emph{E.g.} target propagation~\citep{Lee15,Bengio14} removes the reliance on passing gradients between layers, by instead generating target activations which should be fitted to. However these targets must still be generated sequentially, propagating backwards through the network and layers are therefore still update- and backwards-locked. Other algorithms remove the backwards locking by allowing loss or rewards to be broadcast directly to each layer -- \emph{e.g.}~REINFORCE~\citep{Williams92} (considering all activations are actions), Kickback~\citep{Balduzzi14}, and Policy Gradient Coagent Networks~\citep{Thomas11}~--~but still remain update locked since they require rewards to be generated by an output (or a global critic). While Real-Time Recurrent Learning~\citep{Williams89} or approximations such as \citep{Ollivier15,Tallec17} may seem a promising way to remove update locking, these methods require maintaining the full (or approximate) gradient of the current state with respect to the parameters. This is inherently not scalable and also requires the optimiser to have global knowledge of the network state. In contrast, by framing the interaction between layers as a local communication problem with DNI, we remove the need for global knowledge of the learning system. Other works such as \citep{Taylor16,Carreira14} allow training of layers in parallel without backpropagation, but in practice are not scalable to more complex and generic network architectures.
\section{Decoupled Neural Interfaces}\label{sec:communication}
\begin{figure}[t]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.5\textwidth]{figures/rnns.png}
\end{tabular}
\caption{\small (a) An RNN trained with truncated BPTT using DNI to communicate over time: Every timestep a recurrent core takes input and produces a hidden state $h_t$ and output $y_t$ which affects a loss $L_t$. The core is unrolled for $T$ steps (in this figure $T=3$). Gradients cannot propagate across the boundaries of BPTT, which limits the time dependency the RNN can learn to model. However, the recurrent core includes a synthetic gradient model which produces synthetic gradients $\hat{\delta}_t$ which can be used at the boundaries of BPTT to enable the last set of unrolled cores to communicate with the future ones. (b) In addition, as an auxiliary task, the network can also be asked to do future synthetic gradient prediction: an extra output $\hat{\hat{\delta}}_{t+T}$ is computed every timestep, and is trained to minimise $\|\hat{\hat{\delta}}_{t+T}-\hat{\delta}_{t+T}\|$.}
\label{fig:rnn}
\end{figure}
We begin by describing the high-level communication protocol that is used to allow asynchronously learning agents to communicate.
As shown in \figref{fig:comms}, Sender $A$ sends a message $h_{A}$ to Receiver $B$. $B$ has a model $M_{B}$ of the utility of the message $h_{A}$. $B$'s model of utility $M_{B}$ is used to predict the feedback: an error signal $\hat{\delta}_{A} = M_{B}(h_{A}, s_B, c)$ based on the message $h_{A}$, the current state of $B$, $s_B$, and potentially any other information, $c$, that this module is privy to during training such as the label or context. The feedback $\hat{\delta}_{A}$ is sent back to $A$ which allows $A$ to be updated immediately. In time, $B$ can fully evaluate the true utility $\delta_{A}$ of the message received from $A$, and so $B$'s utility model can be updated to fit the true utility, reducing the disparity between $\hat{\delta}_{A}$ and $\delta_{A}$.
This protocol allows $A$ to send messages to $B$ in a way that $A$ and $B$ are \emph{update decoupled} -- $A$ does not have to wait for $B$ to evaluate the true utility before it can be updated -- and $A$ can still learn to send messages of high utility to $B$.
We can apply this protocol to neural networks communicating, resulting in what we call Decoupled Neural Interfaces (DNI). For neural networks, the feedback error signal $\hat{\delta}_{A}$ can take different forms, \emph{e.g.}~gradients can be used as the error signal to work with backpropagation, target messages as the error signal to work with target propagation, or even a value (cumulative discounted future reward) to incorporate into a reinforcement learning framework. However, as a clear and easily analysable set of first steps into this important and mostly unexplored domain, we concentrate our empirical study on differentiable networks trained with backpropagation and gradient-based updates. Therefore, we focus on producing error gradients as the feedback $\hat{\delta}_{A}$ which we dub \emph{synthetic gradients}.
\paragraph{Notation}
To facilitate our exposition, it's useful to introduce some notation. Without loss of generality, consider neural networks as a graph of function operations (a finite chain graph in the case of a feed-forward models, an infinite chain in the case of recurrent ones, and more generally a directed acyclic graph). The forward execution of the network graph has a natural ordering due to the input dependencies of each functional node. We denote the function corresponding to step $i$ in a graph execution as $f_i$ and the composition of functions (\emph{i.e.}~the forward graph) from step $i$ to step $j$ inclusive as $\mathcal{F}^j_i.$ We denote the loss associated with layer, $i$, of the chain as $L_i$.
\subsection{Synthetic Gradient for Recurrent Networks}\label{sec:rnn}
We begin by describing how our method of using synthetic gradients applies in the case of recurrent networks; in some ways this is simpler to reason about than feed-forward networks or more general graphs.
An RNN applied to infinite stream prediction can be viewed as an infinitely unrolled recurrent core module $f$ with parameters $\theta$, such that the forward graph is $\mathcal{F}^\infty_1 = (f_i)_{i=1}^\infty$ where $f_i = f~\forall i$ and the core module propagates an output $y_i$ and state $h_i$ based on some input $x_i$: $y_i,h_i=f_i(x_i, h_{i-1})$.
At a particular point in time $t$ we wish to minimise $\sum_{\tau=t}^\infty L_\tau$. Of course, one cannot compute an update of the form
$\theta \leftarrow \theta - \alpha \sum_{\tau=t}^\infty \frac{\partial L_\tau}{\partial \theta}$
due to the infinite future time dependency. Instead, generally one considers a tractable time horizon $T$
\begin{equation*}
\begin{split}
\theta - \alpha \sum_{\tau=t}^\infty \frac{\partial L_\tau}{\partial \theta}
&= \theta - \alpha (\sum_{\tau=t}^{t+T} \frac{\partial L_\tau}{\partial \theta} + (\sum_{\tau=T+1}^\infty \frac{\partial L_\tau}{\partial h_T})\frac{\partial h_T}{\partial \theta})\\
&= \theta - \alpha (\sum_{\tau=t}^{t+T} \frac{\partial L_\tau}{\partial \theta} + \delta_T\frac{\partial h_T}{\partial \theta})
\end{split}
\end{equation*}
and as in truncated BPTT, calculates $\sum_{\tau=t}^{t+T} \frac{\partial L_\tau}{\partial \theta}$ with backpropagation and approximates the remaining terms, beyond $t+T$, by using $\delta_T = 0$. This limits the time horizon over which updates to $\theta$ can be learnt, effectively limiting the amount of temporal dependency an RNN can learn. The approximation that $\delta_T=0$ is clearly naive, and by using an appropriately \emph{learned} approximation we can hope to do better. Treating the connection between recurrent cores at time $t+T$ as a Decoupled Neural Interface we can approximate $\delta_T$, with $\hat{\delta}_T = M_T(h_T)$ -- a learned approximation of the future loss gradients -- as shown and described in \figref{fig:rnn}~(a).
This amounts to taking the infinitely unrolled RNN as the full neural network $\mathcal{F}^\infty_1$, and chunking it into an infinite number of sub-networks where the recurrent core is unrolled for $T$ steps, giving $\mathcal{F}^{t+T-1}_t$. Inserting DNI between two adjacent sub-networks $\mathcal{F}^{t+T-1}_t$ and $\mathcal{F}^{t+2T-1}_{t+T}$ allows the recurrent network to learn to communicate to its future self, without being update locked to its future self. From the view of the synthetic gradient model, the RNN is predicting its own error gradients.
The synthetic gradient model $\hat{\delta}_T = M_T(h_T)$ is trained to predict the true gradients by minimising a distance $d(\hat{\delta}_T, \delta_T)$ to the target gradient $\delta_T$ -- in practice we find L$_2$ distance to work well. The target gradient is ideally the true gradient of future loss, $\sum_{\tau=T+1}^\infty \frac{\partial L_\tau}{\partial h_T}$, but as this is not a tractable target to obtain, we can use a target gradient that is itself bootstrapped from a synthetic gradient and then backpropagated and mixed with a number of steps of true gradient, \emph{e.g.}~$\delta_T = \sum_{\tau=T+1}^{2T} \frac{\partial L_\tau}{\partial h_T} + \hat{\delta}_{2T+1}\frac{\partial h_{2T}}{\partial h_T}$. This bootstrapping is exactly analogous to bootstrapping value functions in reinforcement learning and allows temporal credit assignment to propagate beyond the boundary of truncated BPTT.
This training scheme can be implemented very efficiently by exploiting the recurrent nature of the network, as shown in \figref{fig:rnnprocess} in the Supplementary Material. In \sref{sec:exprnn} we show results on sequence-to-sequence tasks and language modelling, where using synthetic gradients extends the time dependency the RNN can learn.
\paragraph{Auxiliary Tasks}
We also propose an extension to aid learning of synthetic gradient models for RNNs, which is to introduce another auxiliary task from the RNN, described in \figref{fig:rnn}~(b). This extra prediction problem is designed to promote coupling over the maximum time span possible, requiring the recurrent core to explicitly model short term and long term synthetic gradients, helping propagate gradient information backwards in time. This is also shown to further increase performance in \sref{sec:exprnn}.
\subsection{Synthetic Gradient for Feed-Forward Networks}\label{sec:ff}
\begin{figure}[t]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.45\textwidth]{figures/ffscenarios.pdf}
\end{tabular}
\caption{\small (a) A section of a vanilla feed-forward neural network $\mathcal{F}^N_1$. (b) Incorporating one synthetic gradient model for the output of layer $i$. This results in two sub-networks $\mathcal{F}^i_1$ and $\mathcal{F}^N_{i+1}$ which can be updated independently. (c) Incorporating multiple synthetic gradient models after every layer results in $N$ independently updated layers.}
\label{fig:fftypes}
\end{figure}
As another illustration of DNIs, we now consider feed-forward networks consisting of $N$ layers $f_i, i \in \{1,\ldots,N\}$, each taking an input $h_{i-1}$ and producing an output $h_{i} = f_i(h_{i-1})$, where $h_0=x$ is the input data. The forward execution graph of the full network can be denoted as as $\mathcal{F}^N_1$, a section of which is illustrated in \figref{fig:fftypes}~(a).
Define the loss imposed on the output of the network as $L=L_N$. Each layer $f_i$ has parameters $\theta_i$ that can be trained jointly to minimise $L(h_N)$ with a gradient-based update rule
\begin{equation*}
\theta_i \leftarrow \theta_i - \alpha~\delta_i \frac{\partial h_i}{\partial \theta_i}~;~~\delta_i=\frac{\partial L}{\partial h_i}
\label{eqn:sgd}
\end{equation*}
where $\alpha$ is the learning rate and $\frac{\partial L}{\partial h_i}$ is computed with backpropagation. The reliance on $\delta_i$ means that the update to layer $i$ can only occur after the remainder of the network, \emph{i.e.}~$\mathcal{F}^N_{i+1}$ (the sub-network of layers between layer $i+1$ and layer $N$ inclusive) has executed a full forward pass, generated the loss $L(h_N)$, then backpropagated the gradient through every successor layer in reverse order. Layer $i$ is therefore update locked to $\mathcal{F}^N_{i+1}$.
To remove the update locking of layer $i$ to $\mathcal{F}^N_{i+1}$ we can use the communication protocol described previously. Layer $i$ sends $h_i$ to layer $i+1$, which has a communication model $M_{i+1}$ that produces a synthetic error gradient $\hat{\delta}_i = M_{i+1}(h_i)$, as shown in \figref{fig:fftypes}~(b), which can be used immediately to update layer $i$ and all the other layers in $\mathcal{F}^i_1$
\begin{equation*}
\theta_n \leftarrow \theta_n - \alpha~\hat{\delta}_i \frac{\partial h_i}{\partial \theta_n},~~n \in \{1,\ldots,i\}.
\label{eqn:sgdbprop}
\end{equation*}
To train the parameters of the synthetic gradient model $M_{i+1}$, we simply wait for the true error gradient $\delta_i$ to be computed (after a full forwards and backwards execution of $\mathcal{F}^N_{i+1}$), and fit the synthetic gradient to the true gradients by minimising $\|\hat{\delta}_i - \delta_i\|^2_2$.
Furthermore, for a feed-forward network, we can use synthetic gradients as communication feedback to decouple every layer in the network, as shown in \figref{fig:fftypes}~(c). The execution of this process is illustrated in \figref{fig:ffprocess} in the Supplementary Material. In this case, since the target error gradient $\delta_i$ is produced by backpropagating $\hat{\delta}_{i+1}$ through layer $i+1$, $\delta_i$ is not the true error gradient, but an estimate bootstrapped from synthetic gradient models later in the network. Surprisingly, this does not cause errors to compound and learning remains stable even with many layers, as shown in \sref{sec:expff}.
Additionally, if any supervision or context $c$ is available at the time of synthetic gradient computation, the synthetic gradient model can take this as an extra input, $\hat{\delta}_i = M_{i+1}(h_i, c)$.
This process allows a layer to be updated as soon as a forward pass of that layer has been executed. This paves the way for sub-parts or layers of networks to be trained in an asynchronous manner, something we show in \sref{sec:expff}.
\subsection{Arbitrary Network Graphs}
Although we have explicitly described the application of DNIs for communication between layers in feed-forward networks, and between recurrent cores in recurrent networks, there is nothing to restrict the use of DNIs for arbitrary network graphs. The same procedure can be applied to any network or collection of networks, any number of times. An example is in \sref{sec:expmulti} where we show communication between two RNNs, which tick at different rates, where the communication can be learnt by using synthetic gradients.
\subsection{Mixing Real \& Synthetic Gradients}
In this paper we focus on the use of synthetic gradients to replace real backpropagated gradients in order to achieve update unlocking. However, synthetic gradients could also be used to augment real gradients. Mixing real and synthetic gradients results in $BP(\lambda)$, an algorithm anolgous to $TD(\lambda)$ for reinforcement learning~\citep{Sutton98}. This can be seen as a generalized view of synthetic gradients, with the algorithms given in this section for update unlocked RNNs and feed-forward networks being specific instantiations of $BP(\lambda)$. This generalised view is discussed further in \sref{sec:unified} in the Supplementary Material.
\section{Experiments}\label{sec:exp}
In this section we perform empirical expositions of the use of DNIs and synthetic gradients, first by applying them to RNNs in \sref{sec:exprnn} showing that synthetic gradients extend the temporal correlations an RNN can learn. Secondly, in \sref{sec:expmulti} we show how a hierarchical, two-timescale system of networks can be jointly trained using synthetic gradients to propagate error signals between networks. Finally, we demonstrate the ability of DNIs to allow asynchronous updating of layers a feed-forward network in \sref{sec:expff}. More experiments can be found in \sref{sec:appexp} in the Supplementary Material.
\subsection{Recurrent Neural Networks}\label{sec:exprnn}
\begin{table*}[t]
\begin{center}\small
\setlength{\tabcolsep}{3pt}
\begin{tabular}{r|ccccccc|ccccc|ccccc}
\toprule
& \multicolumn{7}{c|}{BPTT} & \multicolumn{5}{c|}{DNI} & \multicolumn{5}{c}{DNI + Aux} \\
$T=$ & 2 & 3 & 4 & 5 & 8 & 20 & 40 & 2 & 3 & 4 & 5 & 8 & 2 & 3 & 4 & 5 & 8 \\
\midrule
Copy & 7 & 8 & 10 & 8 & - & - & - & 16 & 14 & 18 & 18 & - & 16 & 17 & 19 & 18 & - \\
Repeat Copy & 7 & 5 & 19 & 23 & - & - & - & 39 & 33 & 39 & 59 & - &39 & 59 & 67 & 59 & - \\
Penn Treebank & 1.39 & 1.38 & 1.37 & 1.37 & 1.35 & 1.35 & 1.34 & 1.37 & 1.36 & 1.35 & 1.35 & 1.34 & 1.37 & 1.36 & 1.35 & 1.35 & 1.33 \\
\bottomrule
\end{tabular}
\end{center}
\caption{\small Results for applying DNI to RNNs. Copy and Repeat Copy task performance is reported as the maximum sequence length that was successfully modelled (higher is better), and Penn Treebank results are reported in terms of test set bits per character (lower is better) at the point of lowest validation error. No learning rate decreases were performed during training.}
\label{table:rnnresults}
\end{table*}
\begin{figure*}[t]
\centering
\begin{tabular}{cccc}
\includegraphics[width=1.0\textwidth]{figures/rnngraphs.png}
\end{tabular}
\caption{\small \emph{Left:} The task progression during training for the Repeat Copy task. All models were trained for 2.5M iterations, but the varying unroll length $T$ results in different quantities of data consumed. The x-axis shows the number of samples consumed by the model, and the y-axis the time dependency level solved by the model -- step changes in the time dependency indicate that a particular time dependency is deemed solved. DNI+Aux refers to DNI with the additional future synthetic gradient prediction auxiliary task. \emph{Right:} Test error in bits per character (BPC) for Penn Treebank character modelling. We train the RNNs with different BPTT unroll lengths with DNI (solid lines) and without DNI (dashed lines). Early stopping is performed based on the validation set. Bracketed numbers give final test set BPC.}
\label{fig:rnngraphs}
\end{figure*}
Here we show the application of DNIs to recurrent neural networks as discussed in \sref{sec:rnn}. We test our models on the Copy task, Repeat Copy task, as well as character-level language modelling.
For all experiments we use an LSTM~\citep{Hochreiter97} of the form in~\citep{Graves13}, whose output is used for the task at hand, and additionally as input to the synthetic gradient model (which is shared over all timesteps). The LSTM is unrolled for $T$ timesteps after which backpropagation through time (BPTT) is performed. We also look at incorporating an auxiliary task which predicts the output of the synthetic gradient model $T$ steps in the future as explained in \sref{sec:rnn}. The implementation details of the RNN models are given in \sref{sec:apprnndetails} in the Supplementary Material.
\paragraph{Copy and Repeat Copy}
We first look at two synthetic tasks -- Copy and Repeat Copy tasks from~\citep{Graves14}. Copy involves reading in a sequence of $N$ characters and after a stop character is encountered, must repeat the sequence of $N$ characters in order and produce a final stop character. Repeat Copy must also read a sequence of $N$ characters, but after the stop character, reads the number, $R$, which indicates the number of times it is required to copy the sequence, before outputting a final stop character. Each sequence of reading and copying is an episode, of length $T_\text{task}=N+3$ for Copy and $T_\text{task}=NR + 3$ for Repeat Copy.
While normally the RNN would be unrolled for the length of the episode before BPTT is performed, $T=T_\text{task}$, we wish to test the length of time the RNN is able to model with and without DNI bridging the BPTT limit. We therefore train the RNN with truncated BPTT: $T\in \{2,3,4,5 \}$ with and without DNI, where the RNN is applied continuously and across episode boundaries. For each problem, once the RNN has solved a task with a particular episode length (averaging below 0.15 bits error), the task is made harder by extending $N$ for Copy and Repeat Copy, and also $R$ for Repeat Copy.
\tblref{table:rnnresults} gives the results by reporting the largest $T_\text{task}$ that is successfully solved by the model. The RNNs without DNI generally perform as expected, with longer BPTT resulting in being able to model longer time dependencies. However, by introducing DNI we can extend the time dependency that is able to be modelled by an RNN. The additional computational complexity is negligible but we require an additional recurrent core to be stored in memory (this is illustrated in \figref{fig:rnnprocess} in the Supplementary Material). Because we can model larger time dependencies with a smaller $T$, our models become more data-efficient, learning faster and having to see less data samples to solve a task. Furthermore, when we include the extra task of predicting the synthetic gradient that will be produced $T$ steps in the future (DNI + Aux), the RNNs with DNI are able to model even larger time dependencies. For example with $T=3$ (\emph{i.e.}~performing BPTT across only three timesteps) on the Repeat Copy task, the DNI enabled RNN goes from being able to model 33 timesteps to 59 timesteps when using future synthetic gradient prediction as well. This is in contrast to without using DNI at all, where the RNN can only model 5 timesteps.
\paragraph{Language Modelling}
We also applied our DNI-enabled RNNs to the task of character-level language modelling, using the Penn Treebank dataset~\citep{Marcus93}. We use an LSTM with 1024 units, which at every timestep reads a character and must predict the next character in the sequence. We train with BPTT with and without DNI, as well as when using future synthetic gradient prediction (DNI + Aux), with $T\in \{2,3,4,5,8 \}$ as well as strong baselines with $T=20,40$. We measure error in bits per character (BPC) as in~\citep{Graves13}, perform early stopping based on validation set error, and for simplicity do not perform any learning rate decay. For full experimental details please refer to \sref{sec:apprnndetails} in the Supplementary Material.
The results are given in \tblref{table:rnnresults}. Interestingly, with BPTT over only two timesteps ($T=2$) an LSTM can get surprisingly good accuracy at next character prediction. As expected, increasing $T$ results in increased accuracy of prediction. When adding DNI, we see an increase in speed of learning (learning curves can be found in \figref{fig:rnngraphs}~(Right) and \figref{fig:ptb} in the Supplementary Material), and models reaching greater accuracy (lower BPC) than their counterparts without DNI. As seen with the Copy and Repeat Copy task, future synthetic gradient prediction further increases the ability of the LSTM to model long range temporal dependencies -- an LSTM unrolled 5 timesteps with DNI and future synthetic gradient prediction gives the same BPC as a vanilla LSTM unrolled 20 steps, only needs 58\% of the data and is $2\times$ faster in wall clock time to reach 1.35BPC.
Although we report results only with LSTMs, we have found DNI to work similarly for vanilla RNNs and Leaky RNNs~\citep{Ollivier15}.
\subsection{Multi-Network System}\label{sec:expmulti}
\begin{figure*}[t]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.45\textwidth]{figures/multinet.pdf}&
\hspace{-0.5em}\includegraphics[width=0.45\textwidth]{figures/multinet.png}\\
\small(a)&\small(b)\\
\end{tabular}
\caption{\small (a) System of two RNNs communicating with DNI. Network A sees a datastream of MNIST digits and every $T$ steps must output the number of odd digits seen. Network B runs every $T$ steps, takes a message from Network A as input and must output the number of 3s seen over the last $T^2$ timesteps. Here is a depiction where $T=2$. (b) The test error over the course of training Network A and Network B with $T=4$. Grey shows when the two-network system is treated as a single graph and trained with backpropagation end-to-end, with an update every $T^2$ timesteps. The blue curves are trained where Network A and Network B are decoupled, with DNI (blue) and without DNI (red). When not decoupled (grey), Network A can only be updated every $T^2$ steps as it is update locked to Network B, so trains slower than if the networks are decoupled (blue and red). Without using DNI (red), Network A receives no feedback from Network B as to how to process the data stream and send a message, so Network B performs poorly. Using synthetic gradient feedback allows Network A to learn to communicate with Network B, resulting in similar final performance to the end-to-end learnt system (results remain stable after 100k steps).}
\label{fig:multinet}
\end{figure*}
In this section, we explore the use of DNI for communication between arbitrary graphs of networks. As a simple proof-of-concept, we look at a system of two RNNs, Network A and Network B, where Network B is executed at a slower rate than Network A, and must use communication from Network A to complete its task. The experimental setup is illustrated and described in \figref{fig:multinet}~(a). Full experimental details can be found in \sref{sec:appmultidetails} in the Supplementary Material.
First, we test this system trained end-to-end, with full backpropagation through all connections, which requires the joint Network A-Network B system to be unrolled for $T^2$ timesteps before a single weight update to both Network A and Network B, as the communication between Network A to Network B causes Network A to be update locked to Network B. We the train the same system but using synthetic gradients to create a learnable bridge between Network A and Network B, thus decoupling Network A from Network B. This allows Network A to be updated $T$ times more frequently, by using synthetic gradients in place of true gradients from Network B.
\figref{fig:multinet}~(b) shows the results for $T=4$. Looking at the test error during learning of Network A (\figref{fig:multinet}~(b) Top), it is clear that being decoupled and therefore updated more frequently allows Network A to learn much quicker than when being locked to Network B, reaching final performance in under half the number of steps. Network B also trains faster with DNI (most likely due to the increased speed in learning of Network A), and reaches a similar final accuracy as with full backpropagation (\figref{fig:multinet}~(b) Bottom). When the networks are decoupled but DNI is not used (\emph{i.e.}~no gradient is received by Network A from Network B), Network A receives no feedback from Network B, so cannot shape its representations and send a suitable message, meaning Network B cannot solve the problem.
\subsection{Feed-Forward Networks}\label{sec:expff}
In this section we apply DNIs to feed-forward networks in order to allow asynchronous or sporadic training of layers, as might be required in a distributed training setup. As explained in \sref{sec:ff}, making layers decoupled by introducing synthetic gradients allows the layers to communicate with each other without being update locked.
\begin{figure*}[t]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.7\textwidth]{figures/drasticnet.pdf}
\end{tabular}
\caption{\small Completely unlocked feed-forward network training allowing forward and update decoupling of layers.}
\label{fig:drasticnet}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{tabular}{c}
\includegraphics[width=1.0\textwidth]{figures/drasticdrastic.pdf}
\end{tabular}
\caption{\small \emph{ Left:} Four layer FCNs trained on MNIST using DNI between every layer, however each layer is trained stochastically -- after every forward pass, a layer only does a backwards pass with probability $p_{\text{update}}$. Population test errors are shown after different numbers of iterations (turquoise is at the end of training after 500k iterations). The purple diamond shows the result when performing regular backpropagation, requiring a synchronous backwards pass and therefore $p_{\text{update}}=1$. When using cDNIs however, with only 5\% probability of a layer being updated the network can train effectively. \emph{Right:} The same setup as previously described however we also use a synthetic \emph{input} model before every layer, which allows the network to also be \emph{forwards decoupled}. Now every layer is trained completely asynchronously, where with probability $1-p_{\text{update}}$ a layer does not do a forward pass or backwards pass -- effectively the layer is ``busy'' and cannot be touched at all.}
\label{fig:drastic}
\end{figure*}
\paragraph{Asynchronous Updates}
To demonstrate the gains by decoupling layers given by DNI, we perform an experiment on a four layer FCN model on MNIST, where the backwards pass and update for every layer occurs in random order and only with some probability $p_{\text{update}}$ (\emph{i.e.}~a layer is only updated after its forward pass $p_{\text{update}}$ of the time). This completely breaks backpropagation, as for example the first layer would only receive error gradients with probability $p_{\text{update}}^{3}$ and even then, the system would be constrained to be synchronous. However, with DNI bridging the communication gap between each layer, the stochasticity of a layer's update does not mean the layer below cannot update, as it uses synthetic gradients rather than backpropagated gradients. We ran 100 experiments with different values of $p_{\text{update}}$ uniformly sampled between 0 and 1. The results are shown in \figref{fig:drastic}~(Left) for DNI with and without conditioning on the labels. With $p_{\text{update}}=0.2$ the network can still train to 2\% accuracy. Incredibly, when the DNI is conditioned on the labels of the data (a reasonable assumption if training in a distributed fashion), the network trains perfectly with only 5\% chance of an update, albeit just slower.
\paragraph{Complete Unlock}
As a drastic extension, we look at making feed-forward networks completely asynchronous, by removing forward locking as well. In this scenario, every layer has a synthetic gradient model, but also a synthetic \emph{input} model -- given the data, the synthetic input model produces an approximation of what the input to the layer will be. This is illustrated in \figref{fig:drasticnet}. Every layer can now be trained independently, with the synthetic gradient and input models trained to regress targets produced by neighbouring layers. The results on MNIST are shown in \figref{fig:drastic}~(Right), and at least in this simple scenario, the completely asynchronous collection of layers train independently, but co-learn to reach 2\% accuracy, only slightly slower. More details are given in the Supplementary Material
\section{Discussion \& Conclusion}\label{sec:conclusion}
In this work we introduced a method, \emph{DNI using synthetic gradients}, which allows decoupled communication between components, such that they can be independently updated.
We demonstrated significant gains from the increased time horizon that DNI-enabled RNNs are able to model, as well as faster convergence.
We also demonstrated the application to a multi-network system: a communicating pair of fast- and slow-ticking RNNs can be decoupled, greatly accelarating learning.
Finally, we showed that the method can be used facilitate distributed training by enabling us to completely decouple all the layers of a feed-forward net -- thus allowing them to be trained asynchronously, non-sequentially, and sporadically.
It should be noted that while this paper introduces and shows empirical justification for the efficacy of DNIs and synthetic gradients, the work of \cite{wojtek17} delves deeper into the analysis and theoretical understanding of DNIs and synthetic gradients, confirming the convergence properties of these methods and modelling impacts of using synthetic gradients.
To our knowledge this is the first time that neural net modules have been decoupled, and the update locking has been broken. This important result opens up exciting avenues of exploration -- including improving the foundations laid out here, and application to modular, decoupled, and asynchronous model architectures.
\section{Unified View of Synthetic Gradients}\label{sec:unified}
The main idea of this paper is to learn a \emph{synthetic gradient}, \emph{i.e.}~a separate prediction of the loss gradient for every layer of the network. The synthetic gradient can be used as a drop-in replacement for the backpropagated gradient. This provides a choice of two gradients at each layer: the gradient of the true loss, backpropagated from subsequent layers; or the synthetic gradient, estimated from the activations of that layer.
In this section we present a unified algorithm, $BP(\lambda)$, that mixes these two gradient estimates as desired using a parameter $\lambda$. This allows the backpropagated gradient to be used insofar as it is available and trusted, but provides a meaningful alternative when it is not. This mixture of gradients is then backpropagated to the previous layer.
\subsection{BP(0)}
We begin by defining our general setup and consider the simplest instance of synthetic gradients, $BP(0)$. We consider a feed-forward network with activations $h_k$ for $k\in \{1,\ldots,K\}$, and parameters $\theta_k$ corresponding to layers $k\in \{1,\ldots,K\}$. The goal is to optimize a loss function $L$ that depends on the final activations $h_K$. The key idea is to approximate the gradient of the loss, $g_k \approx \dd{L}{h_k}$, using a \emph{synthetic gradient}, $g_k$. The synthetic gradient is estimated from the activations at layer $k$, using a function $g_k = g(h_k, \phi_k)$ with parameters $\phi_k$. The overall loss can then be minimized by stochastic gradient descent on the synthetic gradient,
\begin{align*}
\dd{L}{\theta_k} = \dd{L}{h_k} \dd{h_k}{\theta_k} \approx g_k \dd{h_k}{\theta_k}.
\end{align*}
In order for this approach to work, the synthetic gradient must also be trained to approximate the true loss gradient. Of course, it could be trained by regressing $g_k$ towards $\dd{L}{h_k}$, but our underlying assumption is that the backpropagated gradients are not available. Instead, we ``unroll'' our synthetic gradient just one step,
\begin{align*}
g_k &\approx \dd{L}{h_k} = \dd{L}{h_{k+1}} \dd{h_{k+1}}{h_k} \approx g_{k+1} \dd{h_{k+1}}{h_k},
\end{align*}
and treat the unrolled synthetic gradient $z_k = g_{k+1} \dd{h_{k+1}}{h_k}$ as a constant training target for the synthetic gradient $g_k$. Specifically we update the synthetic gradient parameters $\phi_k$ so as to minimise the mean-squared error of these one-step unrolled training targets, by stochastic gradient descent on $\dd{(z_k - g_k)^2}{\phi_k}$. This idea is analogous to bootstrapping in the TD(0) algorithm for reinforcement learning \citep{sutton:td}.
\subsection{BP($\lambda$)}
In the previous section we removed backpropagation altogether. We now consider how to combine synthetic gradients with a controlled amount of backpropagation. The idea of BP($\lambda$) is to mix together many different estimates of the loss gradient, each of which unrolls the chain rule for $n$ steps and then applies the synthetic gradient,
\begin{align*}
g^n_k &= g_{k+n} \dd{h_{k+n}}{h_{k+n-1}} ... \dd{h_{k+1}}{h_k} \nonumber \\
&\approx \dd{L}{h_{k+n}} \dd{h_{k+n}}{h_{k+n-1}} ... \dd{h_{k+1}}{h_k} \nonumber \\
&= \dd{L}{h_k}.
\end{align*}
We mix these estimators together recursively using a weighting parameter $\lambda_k$ (see Figure 1),
\begin{align*}
\bar{g}_k &= \lambda_k \bar{g}_{k+1} \dd{h_{k+1}}{h_k} + (1 - \lambda_k) g_k.
\end{align*}
The resulting $\lambda$-weighted synthetic gradient $\bar{g}_k$ is a geometric mixture of the gradient estimates $g^1_k, ..., g^2_K$,
\begin{align*}
\bar{g}_k &= \sum_{n=k}^K c^n_k g^n_k.
\end{align*}
where $c^n_k = (1-\lambda_n) \prod_{j=k}^{n-1} \lambda_j$ is the weight of the $n$th gradient estimator $g^n_k$, and $c^K_k = 1 - \sum_{n=1}^{K-1} c^n_k$ is the weight for the final layer. This geometric mixture is analogous to the $\lambda$-return in $TD(\lambda)$ \citep{sutton:td}.
To update the network parameters $\theta$, we use the $\lambda$-weighted synthetic gradient estimate in place of the loss gradient,
\begin{align*}
\dd{L}{\theta_k} &= \dd{L}{h_k} \dd{h_k}{\theta_k} \approx \bar{g}_k \dd{h_k}{\theta_k}
\end{align*}
To update the synthetic gradient parameters $\phi_k$, we unroll the $\lambda$-weighted synthetic gradient by one step, $\bar{z}_k = \bar{g}_{k+1} \dd{h_{k+1}}{h_k}$, and treat this as a constant training target for the synthetic gradient $g_k$. Parameters are adjusted by stochastic gradient descent to minimise the mean-squared error between the synthetic gradient and its unrolled target, $\dd{(\bar{z}_k - g_k)^2}{\phi_k}$.
The two extreme cases of $BP(\lambda)$ result in simpler algorithms. If $\lambda_k=0~\forall k$ we recover the $BP(0)$ algorithm from the previous section, which performs no backpropagation whatsoever. If $\lambda_k=1~\forall k$ then the synthetic gradients are ignored altogether and we recover error backpropagation. For the experiments in this paper we have used binary values $\lambda_k \in \{0,1\}$.
\subsection{Recurrent $BP(\lambda)$}
We now discuss how $BP(\lambda)$ may be applied to RNNs. We apply the same basic idea as before, using a synthetic gradient as a proxy for the gradient of the loss. However, network parameters $\theta$ and synthetic gradient parameters $\phi$ are now shared across all steps. There may also be a separate loss $l_k$ at every step $k$. The overall loss function is the sum of the step losses, $L = \sum_{k=1}^\infty l_k$.
The synthetic gradient $g_k$ now estimates the cumulative loss from step $k+1$ onwards, $g_k \approx \dd{\sum_{j=k+1}^\infty l_j}{h_k}$. The $\lambda$-weighted synthetic gradient recursively combines these future estimates, and adds the immediate loss to provide an overall estimate of cumulative loss from step $k$ onwards,
\begin{align*}
\bar{g}_k &= \dd{l_k}{h_k} + \lambda_k \bar{g}_{k+1} \dd{h_{k+1}}{h_k} + (1 - \lambda_k) g_k.
\end{align*}
Network parameters are adjusted by gradient descent on the cumulative loss,
\begin{align*}
\dd{L}{\theta} &= \sum_{k=1}^\infty \dd{L}{h_k} \dd{h_k}{\theta} = \sum_{k=1}^\infty \dd{\sum_{j=k}^\infty l_j}{h_k} \dd{h_k}{\theta} \approx \sum_{k=1}^\infty \bar{g}_k \dd{h_k}{\theta}.
\end{align*}
To update the synthetic gradient parameters $\phi$, we again unroll the $\lambda$-weighted synthetic gradient by one step, $\bar{z}_k = \dd{l_k}{h_k} + \bar{g}_{k+1} \dd{h_{k+1}}{h_k}$, and minimise the MSE with respect to this target, over all time-steps, $\sum_{k=1}^\infty \dd{(\bar{z}_k - g_k)^2}{\phi}$.
We note that for the special case $BP(0)$, there is no backpropagation at all and therefore weights may be updated in a fully online manner. This is possible because the synthetic gradient estimates the gradient of cumulative future loss, rather than explicitly backpropagating the loss from the end of the sequence.
Backpropagation-through-time requires computation from all time-steps to be retained in memory. As a result, RNNs are typically optimised in N-step chunks $[mN,(m+1)N]$. For each chunk $m$, the cumulative loss is initialised to zero at the final step $k=(m+1)N$, and then errors are backpropagated-through-time back to the initial step $k=mN$. However, this prevents the RNN from modelling longer term interactions. Instead, we can initialise the backpropagation at final step $k=(m+1)N$ with a synthetic gradient $g_k$ that estimates long-term future loss, and then backpropagate the synthetic gradient through the chunk. This algorithm is a special case of $BP(\lambda)$ where $\lambda_k=0$ if $k \mod N = 0$ and $\lambda_k=1$ otherwise. The experiments in \sref{sec:exprnn} illustrate this case.
\subsection{Scalar and Vector Critics}
One way to estimate the synthetic gradient is to first estimate the loss using a \emph{critic}, $v(h_k, \phi) \approx {\expect{L | h_k}}$, and then use the gradient of the critic as the synthetic gradient, $g_k = \dd{v(h_k, \phi)}{h_k} \approx \dd{L}{h_k}$. This provides a choice between a scalar approximation of the loss, or a vector approximation of the loss gradient, similar to the scalar and vector critics suggested by Fairbank \citep{fairbank:thesis}.
These approaches have previously been used in control \citep{werbos:adp, fairbank:thesis} and model-based reinforcement learning \citep{heess:svg}. In these cases the dependence of total cost or reward on the policy parameters is computed by backpropagating through the trajectory. This may be viewed as a special case of the $BP(\lambda)$ algorithm; intermediate values of $\lambda<1$ were most successful in noisy environments \citep{heess:svg}.
It is also possible to use other styles of critics or error approximation techniques such as Feedback Alignment~\citep{lillicrap2016random}, Direct Feedback Alignment~\citep{NIPS2016_6441}, and Kickback~\citep{balduzzi2014kickback}) -- interestingly \cite{wojtek17} shows that they can all be framed in the synthetic gradients framework presented in this paper.
\section{Synthetic Gradients are Sufficient}\label{sec:sufficient}
In this section, we show that a function $f(h_t, \theta_{t+1:T})$, which depends only on the hidden activations $h_t$ and downstream parameters $\theta_{t+1:T}$, is sufficient to represent the gradient of a feedforward or recurrent network, without any other dependence on past or future inputs $x_{1:T}$ or targets $y_{1:T}$.
In (stochastic) gradient descent, parameters are updated according to (samples of) the expected loss gradient,
\begin{tiny}
\begin{align*}
\expectx{x_{1:T},y_{1:T}}{\dd{L}{\theta_t}} &= \expectx{x_{1:T},y_{1:T}}{\dd{L}{h_t} \dd{h_t}{\theta_t}} \\
&= \expectx{x_{1:T},y_{1:T}}{\expectx{x_{t+1:T},y_{t:T} | x_{1:t},y_{1:t-1}}{\dd{L}{h_t} \dd{h_t}{\theta_t}}} \\
&= \expectx{x_{1:T},y_{1:T}}{\expectx{x_{t+1:T},y_{t:T} | h_t }{\dd{L}{h_t}} \dd{h_t}{\theta_t}} \\
&= \expectx{x_{1:T},y_{1:T}}{g(h_t, \theta_{t+1:T}) \dd{h_t}{\theta_t}} \\
\end{align*}
\end{tiny}
where $g(h_t, \theta_{t+1:T}) = \expectx{x_{t+1:T},y_{t:T} | h_t }{\dd{L}{h_t}}$ is the expected loss gradient given hidden activations $h_t$. Parameters may be updated using samples of this gradient, $g(h_t, \theta_{t+1:T}) \dd{h_t}{\theta_t}$.
The synthetic gradient $g(h_t, v_t) \approx g(h_t, \theta_{t+1:T})$ approximates this expected loss gradient at the current parameters $\theta_{t+1:T}$. If these parameters are frozen, then a sufficiently powerful synthetic gradient approximator can learn to perfectly represent the expected loss gradient. This is similar to an actor-critic architecture, where the neural network is the actor and the synthetic gradient is the critic.
In practice, we allow the parameters to change over the course of training, and therefore the synthetic gradient must learn online to track the gradient $g(h_t, \theta_{t+1:T})$
\begin{figure*}[t!]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.8\textwidth]{figures/ffprocess.pdf}
\end{tabular}
\caption{\small The execution during training of a feed-forward network. Coloured modules are those that have been updated for this batch of inputs. First, layer $i$ executes it's forward phase, producing $h_i$, which can be used by $M_{i+1}$ to produce the synthetic gradient $\hat{\delta}_i$. The synthetic gradient is pushed backwards into layer $i$ so the parameters $\theta_i$ can be updated immediately. The same applies to layer $i+1$ where $h_{i+1}=f_{i+1}(h_i)$, and then $\hat{\delta}_{i+1}=M_{i+2}(h_{i+1})$ so layer $i+1$ can be updated. Next, $\hat{\delta}_{i+1}$ is backpropagated through layer $i+1$ to generate a target error gradient $\delta_i = f'_{i+1}(h_i)\hat{\delta}_{i+1}$ which is used as a target to regress $\hat{\delta}_i$ to, thus updating $M_{i+1}$. This process is repeated for every subsequent layer.}
\label{fig:ffprocess}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.7\textwidth]{figures/rnnprocess.pdf}
\end{tabular}
\caption{\small The execution during training of an RNN, with a core function $f$, shown for $T=3$. Changes in colour indicate a weight update has occurred. The final core of the last unroll is kept in memory. Fresh cores are unrolled for $T$ steps, and the synthetic gradient from step $T$ (here $\hat{\delta}_{t+3}$ for example) is used to approximate the error gradient from the future. The error gradient is backpropagated through the earliest $T$ cores in memory, which gives a target error gradient for the last time a synthetic gradient was used. This is used to generate a loss for the synthetic gradient output of the RNN, and all the $T$ cores' gradients with respect to parameters can be accumulated and updated. The first $T$ cores in memory are deleted, and this process is repeated. This training requires an extra core to be stored in memory ($T+1$ rather than $T$ as in normal BPTT). Note that the target gradient of the hidden state that is regressed to by the synthetic gradient model is slightly stale, a similar consequence of online training as seen in RTRL~\citep{Williams89}.}
\label{fig:rnnprocess}
\end{figure*}
\section{Additional Experiments}\label{sec:appexp}
\begin{table}[t]
\begin{center}\small
\setlength{\tabcolsep}{4pt}
\begin{tabular}{lr|rrrr|rrrrr}
\toprule
& & \multicolumn{4}{c|}{MNIST (\% Error)} & \multicolumn{4}{c}{CIFAR-10 (\% Error)} \\
\multicolumn{2}{r|}{Layers} & \rotatebox[origin=c]{90}{~No Bprop} & \rotatebox[origin=c]{90}{Bprop} & \rotatebox[origin=c]{90}{DNI} & \rotatebox[origin=c]{90}{cDNI} & \rotatebox[origin=c]{90}{~No Bprop} & \rotatebox[origin=c]{90}{Bprop} & \rotatebox[origin=c]{90}{DNI} & \rotatebox[origin=c]{90}{cDNI} \\
\midrule
\hline
\multirow{4}{*}{\rotatebox[origin=c]{90}{FCN}} & 3 & 9.3 & 2.0 & 1.9 & 2.2 & 54.9 & 43.5 & 42.5 & 48.5 \\
& 4 & 12.6 & 1.8 & 2.2 & 1.9 & 57.2 & 43.0 & 45.0 & 45.1 \\
& 5 & 16.2 & 1.8 & 3.4 & 1.7 & 59.6 & 41.7 & 46.9 & 43.5 \\
& 6 & 21.4 & 1.8 & 4.3 & 1.6 & 61.9 & 42.0 & 49.7 & 46.8 \\
\midrule
\multirow{2}{*}{\rotatebox[origin=c]{90}{CNN}} & 3 & 0.9 & 0.8 & 0.9 & 1.0 & 28.7 & 17.9 & 19.5 & 19.0 \\
& 4 & 2.8 & 0.6 & 0.7 & 0.8 & 38.1 & 15.7 & 19.5 & 16.4 \\
\bottomrule
\end{tabular}
\qquad\quad
\hspace{-3em}
\vtop{\vspace{0.5em}\hbox{\includegraphics[width=0.4\textwidth]{figures/mnistffsummary.pdf}}}
\end{center}
\caption{\small Using DNI between every layer for FCNs and CNNs on MNIST and CIFAR-10. \emph{Left:} Summary of results, where values are final test error (\%) after 500k iterations. \emph{Right:} Test error during training of MNIST FCN models for regular backpropagation, DNI, and cDNI (DNI where the synthetic gradient model is also conditioned on the labels of the data).}
\label{table:ff}
\end{table}
\paragraph{Every layer DNI}
We first look at training an FCN for MNIST digit classification~\citep{lecun1998mnist}. For an FCN, ``layer'' refers to a linear transformation followed by batch-normalisation~\citep{Ioffe15} and a rectified linear non-linearity (ReLU)~\citep{Glorot11}. All hidden layers have the same number of units, 256. We use DNI as in the scenario illustrated in \figref{fig:fftypes}~(d), where DNIs are used between \emph{every} layer in the network. \emph{E.g.}~for a four layer network (three hidden, one final classification) there will be three DNIs. In this scenario, every layer can be updated as soon as its activations have been computed and passed through the synthetic gradient model of the layer above, without waiting for any other layer to compute or loss to be generated. We perform experiments where we vary the depth of the model (between 3 and 6 layers), on MNIST digit classification and CIFAR-10 object recognition~\citep{Krizhevsky09}. Full implementation details can be found in \sref{sec:appffdetails}.
Looking at the results in \tblref{table:ff} we can see that DNI does indeed work, successfully update-decoupling all layers at a small cost in accuracy, demonstrating that it is possible to produce effective gradients \emph{without either label or true gradient information}. Further, once we condition the synthetic gradients on the labels, we can successfully train deep models with very little degradation in accuracy. For example, on CIFAR-10 we can train a 5 layer model, with backpropagation achieving 42\% error, with DNI achieving 47\% error, and when conditioning the synthetic gradient on the label (cDNI) get 44\%. In fact, on MNIST we successfully trained up to 21 layer FCNs with cDNI to 2\% error (the same as with using backpropagation). Interestingly, the best results obtained with cDNI were with \emph{linear} synthetic gradient models.
As another baseline, we tried using historical, stale gradients with respect to activations, rather than synthetic gradients. We took an exponential average historical gradient, searching over the entire spectrum of decay rates and the best results attained on MNIST classification were 9.1\%, 11.8\%, 15.4\%, 19.0\% for 3 to 6 layer FCNs respectively -- marginally better than using zero gradients (no backpropagation) and far worse than the associated cDNI results of 2.2\%, 1.9\%, 1.7\%, 1.6\%. Note that the experiment described above used stale gradients with respect to the activations which do not correspond to the same input example used to compute the activation. In the case of a fixed training dataset, one could use the stale gradient from the same input, but it would be stale by an entire epoch and contains no new information so would fail to improve the model. Thus, we believe that DNI, which uses a parametric approximation to the gradient with respect to activations, is the most desirable approach.
This framework can be easily applied to CNNs~\citep{Lecun98}. The spatial resolution of activations from layers in a CNN results in high dimensional activations, so we use synthetic gradient models which themselves are CNNs without pooling and with resolution-preserving zero-padding. For the full details of the CNN models please refer to \sref{sec:appffdetails}. The results of CNN models for MNIST and CIFAR-10 are also found in \tblref{table:ff}, where DNI and cDNI CNNs perform exceptionally well compared to true backpropagated gradient trained models -- a three layer CNN on CIFAR-10 results in 17.9\% error with backpropagation, 19.5\% (DNI), and 19.0\% (cDNI).
\begin{figure*}[t]
\centering
\begin{tabular}{cccc}
\hspace{-4em}
\includegraphics[width=\textwidth]{figures/mnistffsingle.png}\\
\small (a)\\
\includegraphics[width=0.8\textwidth]{figures/mnistffupto.png}\\
\small (b)
\end{tabular}
\caption{\small Test error during training of a 6 layer fully-connected network on MNIST digit classification. Bprop (grey) indicates traditional, synchronous training with backpropagation, while DNI (blue) shows the use of a (a) single DNI used after a particular layer indicated above, and (b) every layer using DNI up to a particular depth. Without backpropagating any gradients through the connection approximated by DNI results in poor performance (red).}
\label{fig:fcnsingle}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{tabular}{cccc}
\hspace{-4em}
\includegraphics[width=\textwidth]{figures/dnigrad.png}\\
\end{tabular}
\caption{\small Error between the synthetic gradient and the true backpropagated gradient for MNIST FCN where DNI is inserted at a single position. Sign error refers to the average number of dimensions of the synthetic gradient vector that do not have the same sign as the true gradient.}
\label{fig:graderr}
\end{figure*}
\paragraph{Single DNI}
We look at training an FCN for MNIST digit classification using a network with 6 layers (5 hidden layers, one classification layer), but splitting the network into two unlocked sub-networks by inserting a single DNI at a variable position, as illustrated in \figref{fig:fftypes}~(c).
\figref{fig:fcnsingle}~(a) shows the results of varying the depth at which the DNI is inserted. When training this 6 layer FCN with vanilla backpropagation we attain 1.6\% test error. Incorporating a single DNI between two layers results in between 1.8\% and 3.4\% error depending on whether the DNI is after the first layer or the penultimate layer respectively. If we decouple the layers without DNI, by just not backpropagating any gradient between them, this results in bad performance -- between 2.9\% and 23.7\% error for after layer 1 and layer 5 respectively.
One can also see from \figref{fig:fcnsingle}~(a) that as the DNI module is positioned closer to the classification layer (going up in layer hierarchy), the effectiveness of it degrades. This is expected since now a larger portion of the whole system never observes true gradient. However, as we show in \sref{sec:expff}, using extra label information in the DNI module almost completely alleviates this problem.
We also plot the synthetic gradient regression error (L$_2$ distance), cosine distance, and the sign error (the number of times the sign of a gradient dimension is predicted incorrectly) compared to the true error gradient in \figref{fig:graderr}. Looking at the L$_2$ error, one can see that the error jumps initially as the layers start to train, and then the synthetic gradient model starts to fit the target gradients. The cosine similarity is on average very slightly positive, indicating that the direction of synthetic gradient is somewhat aligned with that of the target gradient, allowing the model to train. However, it is clear that the synthetic gradient is not tracking the true gradient very accurately, but this does not seem to impact the ability to train the classifiers.
\begin{figure*}[t]
\centering
\begin{tabular}{cccc}
\multicolumn{2}{c}{\small MNIST FCN}\\
\multicolumn{2}{c}{\includegraphics[width=\textwidth]{figures/mnistffcani.png}}\\
\multicolumn{2}{c}{\small CIFAR-10 FCN}\\
\multicolumn{2}{c}{\includegraphics[width=\textwidth]{figures/cifarffcani.png}}\\
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\small MNIST CNN&
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\small CIFAR-10 CNN\\
\multicolumn{2}{c}{\includegraphics[width=\textwidth]{figures/ffconv.pdf}}\\
\end{tabular}
\caption{\small Corresponding test error curves during training for the results in \tblref{table:ff}. (a) MNIST digit classification with FCNs, (b) CIFAR-10 image classification with FCNs. DNI can be easily used with CNNs as shown in (c) for CNNs on MNIST and (d) for CNNs on CIFAR-10.}
\label{fig:fcnall}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{tabular}{cccc}
\hspace{-4em}
\includegraphics[width=\textwidth]{figures/mnistfflinear.png}\\
\end{tabular}
\caption{\small Linear DNI models for FCNs on MNIST.}
\label{fig:mnistlinear}
\end{figure*}
\subsection{Underfitting of Synthetic Gradient Models}\label{sec:appunderfitting}
If one takes a closer look at learning curves for DNI model (see \figref{fig:cifartrain} for training error plot on CIFAR-10 with CNN model) it is easy to notice that the large test error (and its degradation with depth) is actually an effect of underfitting and not lack of ability to generalise or lack of convergence of learning process. One of the possible explanations is the fact that due to lack of label signal in the DNI module, the network is over-regularised as in each iteration DNI tries to model an expected gradient over the label distribution. This is obviously a harder problem than modelling actual gradient, and due to underfitting to this subproblem, the whole network also underfits to the problem at hand. Once label information is introduced in the cDNI model, the network fits the training data much better, however using synthetic gradients still acts like a regulariser, which also translates to a reduced test error. This might also suggest, that the proposed method of conditioning on labels can be further modified to reduce the underfitting effect.
\begin{figure*}[t]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.6\textwidth]{figures/cifarconvtrain.png}\\
(a)
\end{tabular}
\caption{\small (a) Training error for CIFAR-10 CNNs.}
\label{fig:cifartrain}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.8\textwidth]{figures/ptbnofuture.png}\\
\includegraphics[width=0.8\textwidth]{figures/ptbfuture.png}\\
\end{tabular}
\caption{\small Test error in bits per character (BPC) for Penn Treebank character modelling. We train the RNNs with different BPTT unroll lengths with DNI (solid lines) and without DNI (dashed lines). Early stopping is performed based on the validation set. Top shows results with DNI, and bottom shows results with DNI and future synthetic gradient prediction (DNI+Aux). Bracketed numbers give final test set BPC.}
\label{fig:ptb}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{tabular}{cccc}
\includegraphics[width=1.0\textwidth]{figures/copyrepeatcopy.pdf}
\end{tabular}
\caption{\small The task progression for Copy (top row) and Repeat Copy (bottom row) without future synthetic gradient prediction (left) and with future synthetic gradient prediction (right). For all experiments the tasks' time dependency is advanced after the RNN reaches 0.15 bits error. We run all models for 2.5M optimisation steps. The x-axis shows the number of samples consumed by the model, and the y-axis the time dependency level solved by the model -- step changes in the time dependency indicate that a particular time dependency is deemed solved. DNI+Aux refers to DNI with the additional future synthetic gradient prediction auxiliary task.}
\label{fig:copyloop}
\end{figure*}
\section{Implementation Details}
\subsection{Feed-Forward Implementation Details}\label{sec:appffdetails}
In this section we give the implementation details of the experimental setup used in the experiments from \sref{sec:expff}.
\paragraph{Conditional DNI (cDNI)}
In order to provide DNI module with the label information in FCN, we simply concatenate the one-hot representation of a sample's label to the input of the synthetic gradient model. Consequently for both MNIST and CIFAR-10 experiments, each cDNI module takes ten additional, binary inputs. For convolutional networks we add label information in the form of one-hot encoded channel masks, thus we simply concatenate ten additional channels to the activations, nine out of which are filled with zeros, and one (corresponding to sample's label) is filled with ones.
\paragraph{Common Details}
All experiments are run for 500k iterations and optimised with Adam~\citep{Kingma14} with batch size of 256. The learning rate was initialised at $3\times 10^{-5}$ and decreased by a factor of 10 at 300k and 400k steps. Note the number of iterations, learning rate, and learning rate schedule was not optimised. We perform a hyperparameter search over the number of hidden layers in the synthetic gradient model (from 0 to 2, where 0 means we use a linear model such that $\hat{\delta} = M(h) = \phi_w h + \phi_b$) and select the best number of layers for each experiment type (given below) based on the final test performance. We used cross entropy loss for classification and L$_2$ loss for synthetic gradient regression which was weighted by a factor of 1 with respect to the classification loss. All input data was scaled to $[0, 1]$ interval. The final regression layer of all synthetic gradient models are initialised with zero weights and biases, so initially, zero synthetic gradient is produced.
\paragraph{MNIST FCN}
Every hidden layer consists of fully-connected layers with 256 units, followed by batch-normalisation and ReLU non-linearity. The synthetic gradient models consists of two (DNI) or zero (cDNI) hidden layers and with 1024 units (linear, batch-normalisation, ReLU) followed by a final linear layer with 256 units.
\paragraph{MNIST CNN}
The hidden layers are all convolutional layers with 64 $5\times 5$ filters with resolution preserving padding, followed by batch-normalisation, ReLU and $3 \times 3$ spatial max-pooling in the first layer and average-pooling in the remaining ones. The synthetic gradient model has two hidden layers with 64 $5\times 5$ filters with resolution preserving padding, batch-normalisation and ReLU, followed by a final 64 $5 \times 5$ filter convolutional layer with resolution preserving padding.
\paragraph{CIFAR-10 FCN}
Every hidden layer consists of fully-connected layers with 1000 units, followed by batch-normalisation and ReLU non-linearity. The synthetic gradient models consisted of one hidden layer with 4000 units (linear, batch-normalisation, ReLU) followed by a final linear layer with 1000 units.
\paragraph{CIFAR-10 CNN}
The hidden layers are all convolutional layers with 128 $5\times 5$ filters with resolution preserving padding, followed by batch-normalisation, ReLU and $3 \times 3$ spatial max-pooling in the first layer and avg-pooling in the remaining ones. The synthetic gradient model has two hidden layers with 128 $5\times 5$ filters with resolution preserving padding, batch-normalisation and ReLU, followed by a final 128 $5 \times 5$ filter convolutional layer with resolution preserving padding.
\paragraph{Complete Unlock.}
In the completely unlocked model, we use the identical architecture used for the synthetic gradient model, but for simplicity both synthetic gradient and synthetic input models use a single hidden layer (for both DNI and cDNI), and train it to produce synthetic inputs $\hat{h}_i$ such that $\hat{h}_i \simeq h_i$. The overall training setup is depicted in \figref{fig:drasticnet}. During testing all layers are connected to each other for a forward pass, \emph{i.e.}~the synthetic inputs are not used.
\subsection{RNN Implementation Details}\label{sec:apprnndetails}
\paragraph{Common Details}
All RNN experiments are performed with an LSTM recurrent core, where the output is used for a final linear layer to model the task. In the case of DNI and DNI+Aux, the output of the LSTM is also used as input to a single hidden layer synthetic gradient model with the same number of units as the LSTM, with a final linear projection to two times the number of units of the LSTM (to produce the synthetic gradient of the output and the cell state). The synthetic gradient is scaled by a factor of 0.1 when consumed by the model (we found that this reliably leads to stable training). We perform a hyperparameter search of whether or not to backpropagate synthetic gradient model error into the LSTM core (the model was not particularly sensitive to this, but occasionally backpropagating synthetic gradient model error resulted in more unstable training). The cost on the synthetic gradient regression loss and future synthetic gradient regression loss is simply weighted by a factor of 1.
\paragraph{Copy and Repeat Copy Task}
In these tasks we use 256 LSTM units and the model was optimised with Adam with a learning rate of $7\times 10^{-5}$ and a batch size of 256. The tasks were progressed to a longer episode length after a model gets below 0.15 bits error. The Copy task was progressed by incrementing $N$, the length of the sequence to copy, by one. The Repeat Copy task was progressed by alternating incrementing $N$ by one and $R$, the number of times to repeat, by one.
\paragraph{Penn Treebank}
The architecture used for Penn Treebank experiments consists of an LSTM with 1024 units trained on a character-level language modelling task.
Learning is performed with the use of Adam with learning rate of $7\times 10^-5$ (which we select to maximise the score of the baseline model through testing also $1\times 10^{-4}$ and $1\times 10^{-6}$) without any learning rate decay
or additional regularisation. Each 5k iterations we record validation error (in terms of average bytes per character) and store the network which achieved the smallest one. Once validation error starts to increase we stop training and report test error using previously saved network. In other words, test error is reported for the model yielding minimum validation error measured with 5k iterations
resolution. A single iteration consists of performing full BPTT over $T$ steps with a batch of 256 samples.
\subsection{Multi-Network Implementation Details}\label{sec:appmultidetails}
The two RNNs in this experiment, Network A and Network B, are both LSTMs with 256 units which use batch-normalisation as described in \citep{Cooijmans16}. Network A takes a $28\times 28$ MNIST digit as input and has a two layer FCN (each layer having 256 units and consisting of linear, batch-normalisation, and ReLU), the output of which is passed as input to its LSTM. The output of Network A's LSTM is used by a linear classification layer to classify the number of odd numbers, as well as input to another linear layer with batch-normalisation which produces the message to send to Network B. Network B takes the message from Network A as input to its LSTM, and uses the output of its LSTM for a linear classifier to classify the number of 3's seen in Network A's datastream. The synthetic gradient model has a single hidden layer of size 256 followed by a linear layer which produces the 256-dimensional synthetic gradient as feedback to Network A's message.
All networks are trained with Adam with a learning rate of $1\times 10^{-5}$. We performed a hyperparameter search over the factor by which the synthetic gradient should by multiplied by before being backpropagated through Network A, which we selected as 10 by choosing the system with the lowest training error.
\section{Electronic Submission}
\label{submission}
Submission to ICML 2017 will be entirely electronic, via a web site
(not email). Information about the submission process and \LaTeX\ templates
are available on the conference web site at:
\begin{center}
\textbf{\texttt{http://icml.cc/2017/}}
\end{center}
Send questions about submission and electronic templates to
\texttt{<EMAIL>}.
The guidelines below will be enforced for initial submissions and
camera-ready copies. Here is a brief summary:
\begin{itemize}
\item Submissions must be in PDF.
\item The maximum paper length is \textbf{8 pages excluding references and acknowledgements, and 10 pages
including references and acknowledgements} (pages 9 and 10 must contain only references and acknowledgements).
\item Do \textbf{not include author information or acknowledgements} in your initial
submission.
\item Your paper should be in \textbf{10 point Times font}.
\item Make sure your PDF file only uses Type-1 fonts.
\item Place figure captions {\em under} the figure (and omit titles from inside
the graphic file itself). Place table captions {\em over} the table.
\item References must include page numbers whenever possible and be as complete
as possible. Place multiple citations in chronological order.
\item Do not alter the style template; in particular, do not compress the paper
format by reducing the vertical spaces.
\item Keep your abstract brief and self-contained, one
paragraph and roughly 4--6 sentences. Gross violations will require correction at the camera-ready phase.
Title should have content words capitalized.
\end{itemize}
\subsection{Submitting Papers}
{\bf Paper Deadline:} The deadline for paper submission to ICML 2017
is at \textbf{23:59 Universal Time (3:59 p.m.\ Pacific Standard Time) on February 24, 2017}.
If your full submission does not reach us by this time, it will
not be considered for publication. There is no separate abstract submission.
{\bf Anonymous Submission:} To facilitate blind review, no identifying
author information should appear on the title page or in the paper
itself. Section~\ref{author info} will explain the details of how to
format this.
{\bf Simultaneous Submission:} ICML will not accept any paper which,
at the time of submission, is under review for another conference or
has already been published. This policy also applies to papers that
overlap substantially in technical content with conference papers
under review or previously published. ICML submissions must not be
submitted to other conferences during ICML's review period. Authors
may submit to ICML substantially different versions of journal papers
that are currently under review by the journal, but not yet accepted
at the time of submission. Informal publications, such as technical
reports or papers in workshop proceedings which do not appear in
print, do not fall under these restrictions.
\medskip
To ensure our ability to print submissions, authors must provide their
manuscripts in \textbf{PDF} format. Furthermore, please make sure
that files contain only Type-1 fonts (e.g.,~using the program {\tt
pdffonts} in linux or using File/DocumentProperties/Fonts in
Acrobat). Other fonts (like Type-3) might come from graphics files
imported into the document.
Authors using \textbf{Word} must convert their document to PDF. Most
of the latest versions of Word have the facility to do this
automatically. Submissions will not be accepted in Word format or any
format other than PDF. Really. We're not joking. Don't send Word.
Those who use \textbf{\LaTeX} to format their accepted papers need to pay close
attention to the typefaces used. Specifically, when producing the PDF by first
converting the dvi output of \LaTeX\ to Postscript the default behavior is to
use non-scalable Type-3 PostScript bitmap fonts to represent the standard
\LaTeX\ fonts. The resulting document is difficult to read in electronic form;
the type appears fuzzy. To avoid this problem, dvips must be instructed to use
an alternative font map. This can be achieved with the following two commands:
{\footnotesize
\begin{verbatim}
dvips -Ppdf -tletter -G0 -o paper.ps paper.dvi
ps2pdf paper.ps
\end{verbatim}}
Note that it is a zero following the ``-G''. This tells dvips to use
the config.pdf file (and this file refers to a better font mapping).
A better alternative is to use the \textbf{pdflatex} program instead of
straight \LaTeX. This program avoids the Type-3 font problem, however you must
ensure that all of the fonts are embedded (use {\tt pdffonts}). If they are
not, you need to configure pdflatex to use a font map file that specifies that
the fonts be embedded. Also you should ensure that images are not downsampled
or otherwise compressed in a lossy way.
Note that the 2017 style files use the {\tt hyperref} package to
make clickable links in documents. If this causes problems for you,
add {\tt nohyperref} as one of the options to the {\tt icml2017}
usepackage statement.
\subsection{Reacting to Reviews}
We will continue the ICML tradition in which the authors are given the
option of providing a short reaction to the initial reviews. These
reactions will be taken into account in the discussion among the
reviewers and area chairs.
\subsection{Submitting Final Camera-Ready Copy}
The final versions of papers accepted for publication should follow the
same format and naming convention as initial submissions, except of
course that the normal author information (names and affiliations)
should be given. See Section~\ref{final author} for details of how to
format this.
The footnote, ``Preliminary work. Under review by the International
Conference on Machine Learning (ICML). Do not distribute.'' must be
modified to ``\textit{Proceedings of the
$\mathit{34}^{th}$ International Conference on Machine Learning},
Sydney, Australia, 2017. JMLR: W\&CP.
Copyright 2017 by the author(s).''
For those using the \textbf{\LaTeX} style file, simply change
$\mathtt{\backslash usepackage\{icml2017\}}$ to
$$\mathtt{\backslash usepackage[accepted]\{icml2017\}}$$
\noindent
Authors using \textbf{Word} must edit the
footnote on the first page of the document themselves.
Camera-ready copies should have the title of the paper as running head
on each page except the first one. The running title consists of a
single line centered above a horizontal rule which is $1$ point thick.
The running head should be centered, bold and in $9$ point type. The
rule should be $10$ points above the main text. For those using the
\textbf{\LaTeX} style file, the original title is automatically set as running
head using the {\tt fancyhdr} package which is included in the ICML
2017 style file package. In case that the original title exceeds the
size restrictions, a shorter form can be supplied by using
\verb|\icmltitlerunning{...}|
just before $\mathtt{\backslash begin\{document\}}$.
Authors using \textbf{Word} must edit the header of the document themselves.
\section{Format of the Paper}
All submissions must follow the same format to ensure the printer can
reproduce them without problems and to let readers more easily find
the information that they desire.
\subsection{Length and Dimensions}
Papers must not exceed eight (8) pages, including all figures, tables,
and appendices, but excluding references and acknowledgements. When references and acknowledgements are included,
the paper must not exceed ten (10) pages.
Acknowledgements should be limited to grants and people who contributed to the paper.
Any submission that exceeds
this page limit or that diverges significantly from the format specified
herein will be rejected without review.
The text of the paper should be formatted in two columns, with an
overall width of 6.75 inches, height of 9.0 inches, and 0.25 inches
between the columns. The left margin should be 0.75 inches and the top
margin 1.0 inch (2.54~cm). The right and bottom margins will depend on
whether you print on US letter or A4 paper, but all final versions
must be produced for US letter size.
The paper body should be set in 10~point type with a vertical spacing
of 11~points. Please use Times typeface throughout the text.
\subsection{Title}
The paper title should be set in 14~point bold type and centered
between two horizontal rules that are 1~point thick, with 1.0~inch
between the top rule and the top edge of the page. Capitalize the
first letter of content words and put the rest of the title in lower
case.
\subsection{Author Information for Submission}
\label{author info}
To facilitate blind review, author information must not appear. If
you are using \LaTeX\/ and the \texttt{icml2017.sty} file, you may use
\verb+\icmlauthor{...}+ to specify authors. The author information
will simply not be printed until {\tt accepted} is an argument to the
style file. Submissions that include the author information will not
be reviewed.
\subsubsection{Self-Citations}
If your are citing published papers for which you are an author, refer
to yourself in the third person. In particular, do not use phrases
that reveal your identity (e.g., ``in previous work \cite{langley00}, we
have shown \ldots'').
Do not anonymize citations in the reference section by removing or
blacking out author names. The only exception are manuscripts that are
not yet published (e.g. under submission). If you choose to refer to
such unpublished manuscripts \cite{anonymous}, anonymized copies have
to be submitted
as Supplementary Material via CMT. However, keep in mind that an ICML
paper should be self contained and should contain sufficient detail
for the reviewers to evaluate the work. In particular, reviewers are
not required to look a the Supplementary Material when writing their
review.
\subsubsection{Camera-Ready Author Information}
\label{final author}
If a paper is accepted, a final camera-ready copy must be prepared.
For camera-ready papers, author information should start 0.3~inches
below the bottom rule surrounding the title. The authors' names should
appear in 10~point bold type, electronic mail addresses in 10~point
small capitals, and physical addresses in ordinary 10~point type.
Each author's name should be flush left, whereas the email address
should be flush right on the same line. The author's physical address
should appear flush left on the ensuing line, on a single line if
possible. If successive authors have the same affiliation, then give
their physical address only once.
A sample file (in PDF) with author names is included in the ICML2017
style file package.
\subsection{Abstract}
The paper abstract should begin in the left column, 0.4~inches below
the final address. The heading `Abstract' should be centered, bold,
and in 11~point type. The abstract body should use 10~point type, with
a vertical spacing of 11~points, and should be indented 0.25~inches
more than normal on left-hand and right-hand margins. Insert
0.4~inches of blank space after the body. Keep your abstract brief and
self-contained,
limiting it to one paragraph and roughly 4--6 sentences. Gross violations will require correction at the camera-ready phase.
\subsection{Partitioning the Text}
You should organize your paper into sections and paragraphs to help
readers place a structure on the material and understand its
contributions.
\subsubsection{Sections and Subsections}
Section headings should be numbered, flush left, and set in 11~pt bold
type with the content words capitalized. Leave 0.25~inches of space
before the heading and 0.15~inches after the heading.
Similarly, subsection headings should be numbered, flush left, and set
in 10~pt bold type with the content words capitalized. Leave
0.2~inches of space before the heading and 0.13~inches afterward.
Finally, subsubsection headings should be numbered, flush left, and
set in 10~pt small caps with the content words capitalized. Leave
0.18~inches of space before the heading and 0.1~inches after the
heading.
Please use no more than three levels of headings.
\subsubsection{Paragraphs and Footnotes}
Within each section or subsection, you should further partition the
paper into paragraphs. Do not indent the first line of a given
paragraph, but insert a blank line between succeeding ones.
You can use footnotes\footnote{For the sake of readability, footnotes
should be complete sentences.} to provide readers with additional
information about a topic without interrupting the flow of the paper.
Indicate footnotes with a number in the text where the point is most
relevant. Place the footnote in 9~point type at the bottom of the
column in which it appears. Precede the first footnote in a column
with a horizontal rule of 0.8~inches.\footnote{Multiple footnotes can
appear in each column, in the same order as they appear in the text,
but spread them across columns and pages if possible.}
\begin{figure}[ht]
\vskip 0.2in
\begin{center}
\centerline{\includegraphics[width=\columnwidth]{icml_numpapers}}
\caption{Historical locations and number of accepted papers for International
Machine Learning Conferences (ICML 1993 -- ICML 2008) and
International Workshops on Machine Learning (ML 1988 -- ML
1992). At the time this figure was produced, the number of
accepted papers for ICML 2008 was unknown and instead estimated.}
\label{icml-historical}
\end{center}
\vskip -0.2in
\end{figure}
\subsection{Figures}
You may want to include figures in the paper to help readers visualize
your approach and your results. Such artwork should be centered,
legible, and separated from the text. Lines should be dark and at
least 0.5~points thick for purposes of reproduction, and text should
not appear on a gray background.
Label all distinct components of each figure. If the figure takes the
form of a graph, then give a name for each axis and include a legend
that briefly describes each curve. Do not include a title inside the
figure; instead, the caption should serve this function.
Number figures sequentially, placing the figure number and caption
{\it after\/} the graphics, with at least 0.1~inches of space before
the caption and 0.1~inches after it, as in
Figure~\ref{icml-historical}. The figure caption should be set in
9~point type and centered unless it runs two or more lines, in which
case it should be flush left. You may float figures to the top or
bottom of a column, and you may set wide figures across both columns
(use the environment {\tt figure*} in \LaTeX), but always place
two-column figures at the top or bottom of the page.
\subsection{Algorithms}
If you are using \LaTeX, please use the ``algorithm'' and ``algorithmic''
environments to format pseudocode. These require
the corresponding stylefiles, algorithm.sty and
algorithmic.sty, which are supplied with this package.
Algorithm~\ref{alg:example} shows an example.
\begin{algorithm}[tb]
\caption{Bubble Sort}
\label{alg:example}
\begin{algorithmic}
\STATE {\bfseries Input:} data $x_i$, size $m$
\REPEAT
\STATE Initialize $noChange = true$.
\FOR{$i=1$ {\bfseries to} $m-1$}
\IF{$x_i > x_{i+1}$}
\STATE Swap $x_i$ and $x_{i+1}$
\STATE $noChange = false$
\ENDIF
\ENDFOR
\UNTIL{$noChange$ is $true$}
\end{algorithmic}
\end{algorithm}
\subsection{Tables}
You may also want to include tables that summarize material. Like
figures, these should be centered, legible, and numbered consecutively.
However, place the title {\it above\/} the table with at least
0.1~inches of space before the title and the same after it, as in
Table~\ref{sample-table}. The table title should be set in 9~point
type and centered unless it runs two or more lines, in which case it
should be flush left.
\begin{table}[t]
\caption{Classification accuracies for naive Bayes and flexible
Bayes on various data sets.}
\label{sample-table}
\vskip 0.15in
\begin{center}
\begin{small}
\begin{sc}
\begin{tabular}{lcccr}
\hline
\abovespace\belowspace
Data set & Naive & Flexible & Better? \\
\hline
\abovespace
Breast & 95.9$\pm$ 0.2& 96.7$\pm$ 0.2& $\surd$ \\
Cleveland & 83.3$\pm$ 0.6& 80.0$\pm$ 0.6& $\times$\\
Glass2 & 61.9$\pm$ 1.4& 83.8$\pm$ 0.7& $\surd$ \\
Credit & 74.8$\pm$ 0.5& 78.3$\pm$ 0.6& \\
Horse & 73.3$\pm$ 0.9& 69.7$\pm$ 1.0& $\times$\\
Meta & 67.1$\pm$ 0.6& 76.5$\pm$ 0.5& $\surd$ \\
Pima & 75.1$\pm$ 0.6& 73.9$\pm$ 0.5& \\
\belowspace
Vehicle & 44.9$\pm$ 0.6& 61.5$\pm$ 0.4& $\surd$ \\
\hline
\end{tabular}
\end{sc}
\end{small}
\end{center}
\vskip -0.1in
\end{table}
Tables contain textual material that can be typeset, as contrasted
with figures, which contain graphical material that must be drawn.
Specify the contents of each row and column in the table's topmost
row. Again, you may float tables to a column's top or bottom, and set
wide tables across both columns, but place two-column tables at the
top or bottom of the page.
\subsection{Citations and References}
Please use APA reference format regardless of your formatter
or word processor. If you rely on the \LaTeX\/ bibliographic
facility, use {\tt natbib.sty} and {\tt icml2017.bst}
included in the style-file package to obtain this format.
Citations within the text should include the authors' last names and
year. If the authors' names are included in the sentence, place only
the year in parentheses, for example when referencing Arthur Samuel's
pioneering work \yrcite{Samuel59}. Otherwise place the entire
reference in parentheses with the authors and year separated by a
comma \cite{Samuel59}. List multiple references separated by
semicolons \cite{kearns89,Samuel59,mitchell80}. Use the `et~al.'
construct only for citations with three or more authors or after
listing all authors to a publication in an earlier reference \cite{MachineLearningI}.
Authors should cite their own work in the third person
in the initial version of their paper submitted for blind review.
Please refer to Section~\ref{author info} for detailed instructions on how to
cite your own papers.
Use an unnumbered first-level section heading for the references, and
use a hanging indent style, with the first line of the reference flush
against the left margin and subsequent lines indented by 10 points.
The references at the end of this document give examples for journal
articles \cite{Samuel59}, conference publications \cite{langley00}, book chapters \cite{Newell81}, books \cite{DudaHart2nd}, edited volumes \cite{MachineLearningI},
technical reports \cite{mitchell80}, and dissertations \cite{kearns89}.
Alphabetize references by the surnames of the first authors, with
single author entries preceding multiple author entries. Order
references for the same authors by year of publication, with the
earliest first. Make sure that each reference includes all relevant
information (e.g., page numbers).
\subsection{Software and Data}
We strongly encourage the publication of software and data with the
camera-ready version of the paper whenever appropriate. This can be
done by including a URL in the camera-ready copy. However, do not
include URLs that reveal your institution or identity in your
submission for review. Instead, provide an anonymous URL or upload
the material as ``Supplementary Material'' into the CMT reviewing
system. Note that reviewers are not required to look a this material
when writing their review.
\section*{Acknowledgements}
\textbf{Do not} include acknowledgements in the initial version of
the paper submitted for blind review.
If a paper is accepted, the final camera-ready version can (and
probably should) include acknowledgements. In this case, please
place such acknowledgements in an unnumbered section at the
end of the paper. Typically, this will include thanks to reviewers
who gave useful comments, to colleagues who contributed to the ideas,
and to funding agencies and corporate sponsors that provided financial
support.
\nocite{langley00}
|
\section{Introduction}
In a recent paper \cite{Chapling:2016LORE} we proved three conjectures of Wu \cite{ISI:000187232400003} required in the method of Loop Regularisation. In this note we give a direct analysis of the irreducible loop integrals (ILIs) involved in the formulation of the theory; this enables us to produce closed-form expressions for the limits required in loop regularisation.
The second section contains an outline of the theory of Loop Regularisation, focussing on the introduction and properties of the ILIs. In the third, we summarise the original method of evaluating the integrals, indicating the source of the conjectures proven in our previous paper \cite{Chapling:2016LORE}, but also deriving the form of the integral that we consider in the final section, namely
\begin{equation}
\lim_{\substack{ m \to \infty \\ M_{R}^{2} \to \infty }} \int_{0}^{\infty} \tau^{\alpha-1} e^{-\tau \mu^{2}} \left( 1-e^{-\tau M_{R}^{2}} \right)^{m} d\tau.
\end{equation}
Wu discusses this as an outcome of the \emph{proper-time formalism} in \cite{ISI:000224011700003}. We first consider some properties of the bracketed factor, which then allow us to use ideas from probability theory to understand its behaviour in the large-\(m\) limit: armed with this, we may deduce natural conditions on the way we can take \(M_{R} \to \infty\) while ensuring the finiteness of this integral.
Calculations shall be carried out in \(d\) dimensions; it will become apparent that with judicious choice of the quantities involved, the actual number of dimensions is not an essential feature of the theory as we investigate it here, although of course it is significant in applications. The metric \(g_{\mu\nu}\) is Minkowski with signature \( (+,-,-,-) \). Our notation is consistent with \cite{Chapling:2016LORE}, but not necessarily with \cite{ISI:000187232400003}.
\section{Loop Regularisation}
We outline briefly the details of a specific regularisation scheme known as Loop Regularisation, which was first discussed by Wu in a paper of 2003, \cite{ISI:000187232400003}; we follow loosely parts of that paper in our presentation in this section.
Initially, the second section of Wu's paper outlines a new set of conditions for a regularisation scheme to preserve gauge invariance. We consider the following set of integrals, known as \emph{irreducible loop integrals} or ILIs:
\begin{align*}
I_{-2\alpha} &= \frac{1}{(2\pi)^{d}}\int_{\ensuremath{\mathbb{R}}^{d}} \frac{dp}{(p^{2}-\Delta)^{d/2+\alpha}} \\
I_{-2\alpha \, \mu\nu} &= \frac{1}{(2\pi)^{d}}\int_{\ensuremath{\mathbb{R}}^{d}} \frac{p_{\mu}p_{\nu}}{(p^{2}-\Delta)^{d/2+1+\alpha}} \, dp \\
I_{-2\alpha \, \mu\nu\rho\sigma} &= \frac{1}{(2\pi)^{d}}\int_{\ensuremath{\mathbb{R}}^{d}} \frac{p_{\mu}p_{\nu}p_{\rho}p_{\sigma}}{(p^{2}-\Delta)^{d/2+2+\alpha}} \, dp.
\end{align*}
Wu shows by application of Ward identities and calculation of the Yang--Mills vacuum diagrams that the regularised versions of these, \( I^{R}_{2} \) \&c., should satisfy the following identities:
\begin{enumerate}
\item
Quadratically divergent:
\begin{align*}
I^{R}_{2 \, \mu\nu} &= \frac{1}{2} g_{\mu\nu} I^{R}_{2}, \\
I^{R}_{2 \, \mu\nu\rho\sigma} &= \frac{1}{8} (g_{\mu\nu}g_{\rho\sigma} + g_{\mu\rho}g_{\sigma\nu} + g_{\mu\sigma}g_{\nu\rho}) I^{R}_{2},
\end{align*}
\item
Logarithmically divergent:
\begin{align*}
I^{R}_{0 \,\mu\nu} &= \frac{1}{4} g_{\mu\nu} I^{R}_{2}, \\
I^{R}_{0 \, \mu\nu\rho\sigma} &= \frac{1}{24} (g_{\mu\nu}g_{\rho\sigma} + g_{\mu\rho}g_{\sigma\nu} + g_{\mu\sigma}g_{\nu\rho}) I^{R}_{0}
\end{align*}
\end{enumerate}
(We observe later that dimensional regularisation satisfies this, for example.)
The remainder of Wu's paper is dedicated to producing a regularisation scheme (\emph{Loop Regularisation}, or LORE) consistent with these, that operates strictly in four dimensions, and can also maintain non-Abelian gauge invariance.\footnote{Unlike Pauli--Villars regularisation, for example: \cite{Dai:2005ve,ISI:000258778900002,ISI:000293182400002}. For further details of the advantages and consequences of LORE, see Wu's review article \cite{Wu:2014uq}.}
The method is as follows:
\begin{enumerate}
\item
Introduce a set of \(m\) regulating masses \(M_{k}\), and replace the integrated loop momenta \( p^{2} \) by \( [p^{2}]_{k} := p^{2}+M_{k}^{2} \)
\item
Replace the loop integral by a regularised measure,
\begin{equation*}
\int dp \mapsto \int [dp]_{k} := \lim_{m \to \infty} \lim_{M_{k}^{2} \to \infty} \sum_{k=0}^{m} c^{m}_{k} \int dp,
\end{equation*}
where the \(c_{k}^{m}\) are constants to be determined.
\item
Calculate the integrals with a finite number of finite regulating masses
\item
Take the number of regulating masses, and their magnitudes, to \( \infty \), in a way specified below.
\end{enumerate}
One recovers the original integral from taking \( M_{0}=0 \), \( c^{m}_{0}=1 \), and the rest of the \(M_{l}\) to \( \infty \). If the original integrals are IR divergent, one may prefer to take \( M_{0} = \mu_{s} \), some small cutoff, to avoid this. However, the key idea of this regularisation is that we can get rid of the slowly-decaying terms at \( p^{2} \to \infty \): notice that for large \(p^{2}\),
\begin{align*}
\sum_{k=0}^{m} c^{m}_{k} \frac{1}{(p^{2}+M_{k}^{2})^{\beta}} &= \sum_{k=0}^{m} c^{m}_{k} \sum_{n=0}^{\infty} \frac{1}{(p^{2})^{\beta}} \binom{\beta}{n} \left( \frac{M_{k}^{2}}{p^{2}} \right)^{n} \\
&= \sum_{n=0}^{\infty} \binom{\beta}{n} \frac{1}{(p^{2})^{\beta+n}} \sum_{k=0}^{m} c^{m}_{k} ( M_{k}^{2})^{n}.
\end{align*}
Therefore if we take \(m\) sufficiently large, we are able to eliminate the ultraviolet divergence of an integrand by removing as many of the low-\(n\) powers of \(1/p^{2}\) as necessary, by choosing \(M_{k}\) and \(c_{k}^{m}\) to satisfy the equations\footnote{In fact, this only needs to be satisfied in the limit as the \(M_{k}\) become large, but we can demand it identically to save ourselves from having to change the order of limits more often than necessary.}
\begin{equation*}
\sum_{k=0}^{m} c_{k}^{m} (M_{k}^{2})^{n} = 0, \qquad c^{m}_{0}=1.
\end{equation*}
Probably the simplest choice for \( M_{k}^{2} \) is to allow it to depend linearly on \(k\), so that \( M_{k}^{2} = \mu_{s}^{2}+k M_{R}^{2} \), where \(M_{R}\) is an overall mass that we can take to \( \infty\) in some manner later. Indeed, it will be necessary to take both \(m \to \infty\) and \( M_{R} \to \infty \) to eliminate any explicit dependence on the specific nature of the regularisation. With this choice, the above constraints become
\begin{equation}
\label{eq:ckmk^nrelation}
\sum_{k=0}^{m} c_{k}^{m} k^{n} = 0, \qquad c^{m}_{0}=1
\end{equation}
by expanding out, and we recognise that a solution is provided by \( c_{k}^{m} = (-1)^{k} \binom{m}{k} \) (indeed, it is the unique solution).
The crux of the matter now becomes to evaluate the ILIs and demonstrate that the consistency conditions are satisfied; the evaluation also gives useful expressions for calculating amplitudes, of course.
\section{Evaluation of the Integrals Using Sums}
\subsection{Evaluation of the scalar integrals}
\label{sec:scalarinteval}
Scalar integrals are in general simpler to evaluate, so we we give some details of the calculation. The tensor ones are done by using tensor relationships, which may be applied to convergent integrals, as is shown in the next section.
The main calculation proceeds as follows: after Wick rotating and using Schwinger's parametrisation, we have
\begin{equation}
\frac{1}{(p^{2}+M'^{2}_{k})^{d/2+\alpha}} = \frac{1}{\Gamma(d/2+\alpha)}\int_{0}^{\infty} \tau^{d/2+\alpha-1} e^{-\tau(p^{2}+M'^{2}_{k})} d\tau,
\end{equation}
where we write \( M'^{2}_{k} := \Delta+M^{2}_{k} =: \mu^{2}+k M^{2}_{R} \). Interchanging the order of integration and using the Gaussian integral \( \int_{\ensuremath{\mathbb{R}}^{d}} e^{-\lambda p^{2}} \, dp = (\pi/\lambda)^{d/2} \), we have
\begin{align*}
I^{R}_{-2\alpha}(m) &= \frac{i(-1)^{-\alpha-d/2}}{(4\pi)^{d/2}\Gamma(d/2+\alpha)}\int_{0}^{\infty} \tau^{\alpha-1} \left( \sum_{k=0}^{m} (-1)^{k} \binom{m}{k} e^{-\tau M'^{2}_{k}} \right) d\tau \\
&= \frac{i(-1)^{-\alpha-d/2}}{(4\pi)^{d/2}\Gamma(d/2+\alpha)}\int_{0}^{\infty} \tau^{\alpha-1} e^{-\tau \mu^{2}} \left( \sum_{k=0}^{m} (-1)^{k} \binom{m}{k} e^{-\tau k M^{2}_{R}} \right) d\tau
\end{align*}
The constraints being satisfied, we can see using series expansions and \eqref{eq:ckmk^nrelation} that the bracket is of the order of \( \tau^{m} \) as \(\tau \downarrow 0\), so we can treat the integral as convergent, even if \( \alpha<0\). Hence it is continuous in \(\alpha\), and we can take limits. Evaluating the integral gives
\begin{equation*}
I^{R}_{-2\alpha}(m) = \frac{i(-1)^{-\alpha-d/2}}{(4\pi)^{d/2}}\frac{ \Gamma(\alpha)}{\Gamma(d/2+\alpha)} \sum_{k=0}^{m} (-1)^{k} \binom{m}{k} (M'^{2}_{k})^{-\alpha}
\end{equation*}
when \( \alpha>-m \) and is not a negative integer. If \(\alpha\) is a nonpositive integer, we take the limit of \( \Gamma(\alpha) \) times the sum, which gives a derivative with respect to \(\alpha\) due to the pole, but we shall deal with this shortly; the significant point is that this identity, correctly interpreted, holds for \(m\) finite and sufficiently large.
\subsection{Evaluation of the tensor integrals}
When the integral is convergent, we can apply the principle
\begin{equation}
\int_{\ensuremath{\mathbb{R}}^{d}} p_{\mu}p_{\nu} F(p^{2}) \, dp = \frac{1}{d} g_{\mu\nu} \int_{\ensuremath{\mathbb{R}}^{d}} p^{2} F(p^{2}) \, dp
\end{equation}
for isotropic integrals to compute the integral directly. Then we apply \( \int_{\ensuremath{\mathbb{R}}^{d}} p^{2} e^{-\lambda p^{2}} \, dp = \pi^{d/2}\lambda^{-d/2-1}d/2 \), so
\begin{equation*}
I^{R}_{-2\alpha \, \mu\nu}(m) = g_{\mu\nu} \frac{i(-1)^{-\alpha-d/2}}{(4\pi)^{d/2}} \frac{\Gamma(\alpha)}{2\Gamma(d/2+\alpha+1)} \sum_{k=0}^{m} (-1)^{k} \binom{m}{k} (M'^{2}_{k})^{-\alpha},
\end{equation*}
with the same provisos regarding \( \alpha>-m \) and the nonpositive integers.
We now see that
\begin{equation}
I^{R}_{-2\alpha \, \mu\nu}(m) = \frac{1}{d+2\alpha} g_{\mu\nu} I^{R}_{-2\alpha}(m),
\end{equation}
and putting \( d=4 \) and \(\alpha=-1,0\) gives the identities required. The rank-four integrals are similar.
\subsection{Evaluation of the sums and the interesting functional limits}
We now give a brief outline of the large-\(m\) expansions which require the results proven in \cite{Chapling:2016LORE}; these originate from Wu's requirements for the convergence of certain sums obtained in evaluating the limit of the regularised integrals. This is mainly by way of contrast with the method we consider in the next section, where the integral is treated without using series expansions.
We wish to evaluate the function
\begin{equation*}
\Gamma(\alpha) \sum_{k=0}^{m} (-1)^{k} \binom{m}{k} (M'^{2}_{k})^{-\alpha},
\end{equation*}
to find the explicit values of the regularised integrals. The main question is how to take the limit \( M_{R}^{2} \to \infty \) to obtain finite, sensible quantities; a simple way to consider this is to examine this sum for the quadratically divergent \( I^{R}_{2} \), i.e. with \(\alpha=-1\), when it takes on the form
\begin{align*}
\sum_{k=0}^{m} & (-1)^{k} \binom{m}{k} (\mu^{2}+k M_{R}^{2}) \log{(\mu^{2}+k M_{R}^{2})} \\
&= M_{R}^{2} \sum_{k=1}^{m} (-1)^{k} \binom{m}{k} k \log{k} +
M_{R}^{2} \sum_{k=1}^{m} (-1)^{k} \binom{m}{k} k \log{\left(1+\frac{\mu^{2}}{k M_{R}^{2}}\right)} \\
&\quad + \mu^{2} \log{\left(\frac{\mu^{2}}{M_{R}^{2}}\right)} + \mu^{2} \sum_{k=1}^{m} (-1)^{k} \binom{m}{k} \left(\log{k} + \log{\left(1+\frac{\mu^{2}}{k M_{R}^{2}}\right)} \right),
\end{align*}
after using \( \sum_{k=0}^{m} (-1)^{k} \binom{m}{k} (a+bk) = 0 \). We shall certainly require the first term to be finite in the limit, i.e. there is a constant \(M_{c}^{2}\) so that
\begin{equation}
M_{R}^{2} \sum_{k=0}^{m} (-1)^{k} \binom{m}{k} k \log{k} \to M_{c}^{2} \quad \text{as } M_{R}^{2}, m \to \infty.
\end{equation}
This is the source of the first conjecture: on numerical evidence, Wu hypothesises that we should take
\begin{equation*}
M_{R}^{2} = M_{c}^{2} \log{m};
\end{equation*}
this choice is shown to be correct by our \cite{Chapling:2016LORE}
\begin{theorem}
Let $m$ be a positive integer. Then
\begin{align}
\label{eq:conj1}
\sum_{k \geqslant 1} (-1)^{k}\binom{m}{k} k \log{k} \sim \frac{1}{\log{m}} \quad \text{as } m \to \infty.
\end{align}
\end{theorem}
Of course, we then have to take this into account when taking the limit of the other expressions found above. The next term to worry about is also included in the \( \alpha=0 \) sum, which is
\begin{align*}
\sum_{k=0}^{m} & (-1)^{k} \binom{m}{k} \log{(\mu^{2}+k M_{R}^{2})} \\
&= \log{\left(\frac{\mu^{2}}{M_{c}^{2}}\right)} + \left( \sum_{k=1}^{m} (-1)^{k} \binom{m}{k} \log{k} -\log{\log{m}} \right) + \sum_{k=1}^{m} (-1)^{k} \binom{m}{k}\log{\left(1+\frac{\mu^{2}}{k M_{R}^{2}}\right)}.
\end{align*}
For consistency, the middle bracket has to converge to a constant; this is resolved by
\begin{theorem}
Let $m$ be a positive integer. Then
\begin{align}
\label{eq:conj2}
\sum_{k \geqslant 1} (-1)^{k}\binom{m}{k} \log{k} \sim \log{\log{m}} + \gamma \quad \text{as } m \to \infty,
\end{align}
where $\gamma$ is the Euler--Mascheroni constant.
\end{theorem}
Lastly, in expanding the logarithm and binomial terms, one has to find limits of sums of the form
\begin{equation*}
n!\sum_{k=1}^{m} (-1)^{k-1} \binom{m}{k} \frac{1}{(k\log{m})^{n}}
\end{equation*}
as \(m \to \infty\), which is a consequence of
\begin{theorem}
Let $m,n$ be positive integers. Then
\begin{align}
\label{eq:conj3}
\sum_{k \geqslant 1} (-1)^{k-1}\binom{m}{k} \frac{1}{k^{n}} \sim \frac{\log^{n}{m}}{n!} \quad \text{as } m \to \infty.
\end{align}
\end{theorem}
It is apparent that this method of treating the integral, while a source of some interesting mathematical results, is a rather dissatisfactory method of producing the evaluation of the limit, and gives inelegant expressions in terms of series, where much care has to be taken with the interchange of the various limiting operations. In the next section, we avoid this by returning to the integral and evaluating the limit without expanding in sums.
\section{A New Approach to Evaluating the ILIs in Terms of Special Functions}
In this section we shall discuss an alternative approach, avoiding introducing any more summations than strictly necessary. We consider the integral
\begin{equation}
\int_{0}^{\infty} \tau^{\alpha-1} e^{-\tau \mu^{2}} \left( \sum_{k=0}^{m} (-1)^{k} \binom{m}{k} e^{-\tau k M^{2}_{R}} \right) d\tau \\
= \int_{0}^{\infty} \tau^{\alpha-1} e^{-\tau \mu^{2}} \left( 1-e^{-\tau M_{R}^{2}} \right)^{m} d\tau,
\end{equation}
from the expression for \(I^{R}_{-2\alpha}\) that we found in \S~\ref{sec:scalarinteval}. Changing to the non-dimensional variable \( u=\tau \mu^{2} \), we have to understand the limit of
\begin{equation}
\label{eq:nondimlintegral}
(\mu^{2})^{-\alpha} \int_{0}^{\infty} u^{\alpha-1} e^{-u} \left( 1-e^{- f(m) u} \right)^{m} du,
\end{equation}
as \(m \to \infty\), where \( f(m):= M_{R}^{2}/\mu^{2} \). Notably, if \( \alpha \leq 0 \), the singularity in \(u^{\alpha-1}\) is cancelled by the zero in the factor \( \left( 1-e^{- f(m) u} \right)^{m} \) providing that \(m\) is large enough; this is corresponds to the regularisation of the ultraviolet divergence that the regularisation provides in the original form of the integral.
Therefore the key lies in analysing the factor \( F_{m}(u) = ( 1-e^{- f(m) u } )^{m} \), and its behaviour as \(m \to \infty\). We first note the following properties:
\begin{enumerate}
\item
\(F_{m}(u)\) is positive on \( (0,\infty) \).
\item
\( F_{m}(u) = O(u^{m}) \) as \( u \downarrow 0 \).
\item
\( F_{m}(u) \to 1 \) as \( u \to \infty \).
\item
\( F_{m}'(u) = m f(m) e^{-f(m)u} (1-e^{-f(m)u})^{m-1} > 0 \) on \( (0,\infty) \).
\end{enumerate}
Therefore \(F_{m}\) does behave as a smooth cutoff at \(0\). Given that \(0\) is precisely the point we wish to avoid, we need to make sure that there is an interval \( (0,\varepsilon) \) in which \( F_{m}(u) \to 0 \) as \( m \to \infty \). Fundamentally this comes down to the position and width of the transition of \( F_{m} \) from \(0\) to \(1\), and we need to choose \(f(m)\) to affect this correctly as \( m \to \infty \). It is not obvious from the above how to determine the correct behaviour.
However, there is another way to approach this problem, from which emerges a clear solution. Namely, it is also possible to consider \(F_{m}\) as the cumulative distribution function of a random variable \(X_{m}\) on \( (0,\infty) \). With this perspective, it is natural to ask about the expectation and variance of \(X_{m}\), which will tell us about the location and shape of \(F_{m}\)'s transition from \(0\) to \(1\). We find that
\begin{align*}
\E[X_{m}^{n}] &= \int_{0}^{\infty} nu^{n-1} F_{m}(u) \, du \\
&= \int_{0}^{\infty} nu^{n-1} (1-e^{-f(m)u})^{m} du \\
&= \frac{n!}{f(m)^{n}} \sum_{k=1}^{m} (-1)^{k-1} \binom{m}{k} \frac{1}{k^{n}}.
\end{align*}
Hence all moments exist, and by the results in \S~4 of \cite{Chapling:2016LORE}, we can calculate the first two explicitly to obtain
\begin{align*}
\E[X_{m}] &= \frac{H_{m}^{(1)}}{f(m)}, \\
\E[X_{m}^{2}] &= \frac{H_{m}^{(2)}}{f(m)^{2}} + \frac{(H_{m}^{(1)})^{2}}{f(m)^{2}},
\end{align*}
\( H_{m}^{(n)} := \sum_{k=1}^{m} k^{-n} \) being the harmonic number of order \(n\). Thus the variance is
\begin{equation*}
\var(X_{m}) = \E[X_{m}^{2}]-(\E[X_{m}])^{2} = \frac{H_{m}^{(2)}}{f(m)^{2}};
\end{equation*}
since \( H_{m}^{(2)} < \pi^{2}/6 \) is bounded and \( f(m) \to \infty \), we conclude that as \(m \to \infty\), the variance tends to zero.
We now apply some results of analysis, phrased in the language of probability: suppose that \begin{equation}
z = \lim_{m \to \infty} \E[X_{m}] = \lim_{m \to \infty} \frac{H_{m}^{(1)}}{f(m)}
\end{equation}
exists. If we let \(X\) be a random variable that takes on the value \(z\) with probability \(1\), then \( \E[(X_{m}-X)^{2}] \to 0 \), so \(X_{m} \to X\) in mean square, and it follows that \(X_{m} \to X \) in distribution.\footnote{\cite{GrimmettStirzaker:2001pr}, \S~7.2, p.~310, Theorem (4) \emph{et seq}.} This means that, \( F_{m} \to F \), where \( F=\chi_{[z,\infty)} \) is the distribution function of \(X\).
We can now return to the integral we actually wish to evaluate. We have shown that \( F_{m} \to F \) pointwise. We also know that if (and only if) we also have \(z>0\), for sufficiently large \(m\), \(u^{\alpha-1} e^{-u} F_{m}(u)\) is summable and bounded by a constant multiple of \( e^{-u} \). Then we may use the Dominated Convergence Theorem to pass the limit inside the integral sign of the expression in \eqref{eq:nondimlintegral}, obtaining
\begin{equation*}
\lim_{m \to \infty} \int_{0}^{\infty} u^{\alpha-1} e^{-u} F_{m}(u) du = \int_{z}^{\infty} u^{\alpha-1} e^{-u} \, du = \Gamma(\alpha,z),
\end{equation*}
the standard special function known as the \emph{upper incomplete Gamma-function}.
To summarise, to make this integral well-defined and finite, we require that \(z\) exists and is greater than zero; this occurs, essentially by definition, if \( f(m) \) is asymptotic to \( \log{m} \) as \(m \to \infty \). We can most easily achieve this by putting \( f(m) = z^{-1} \log{m} \). Translating back to physical variables, we see that this has naturally lead to us taking Wu's choice (\cite{ISI:000187232400003}, equation (4.1))
\begin{equation}
M_{R}^{2} = M_{c}^{2} \log{m}.
\end{equation}
With this, the loop integral becomes
\begin{equation}
I^{R}_{-2\alpha} = \frac{i(-1)^{-\alpha-d/2}}{(4\pi)^{d/2}\Gamma(d/2+\alpha)} (\mu^{2})^{-\alpha} \Gamma\left(\alpha, \frac{\mu^{2}}{M_{c}^{2}} \right).
\end{equation}
One can check that this agrees with the explicit expressions given by Wu for \(I^{R}_{2}\), \(I^{R}_{0}\) and \(I^{R}_{-2}\) ((3.22--24) in \cite{ISI:000187232400003}). They also clearly exhibit the singularities typical of loop regularisation: the incomplete Gamma-function is given by
\begin{equation}
\Gamma(\alpha,z) = \int_{z}^{1} u^{\alpha-1} e^{-u} \, du + \int_{1}^{\infty} u^{\alpha-1} e^{-u} \, du;
\end{equation}
the latter is a constant, and the former can be expanded in what may be called a Frobenius series, with most singular term \( -z^{\alpha}/\alpha \), which is, for example, quadratic if \(\alpha=-1\). If \(\alpha=0\), the most singular term is instead the logarithmically divergent \( -\log{z} \); this term will also be included if \( \alpha\) is a negative integer, from integrating the \(1/u\) term.
\section{Conclusion}
We have arrived in quite a natural way at Wu's postulated form for the regulator mass \(M_{R}^{2}\), resolving the irreducible loop integrals into a simple closed-form expression involving the incomplete Gamma-function. The approach is reasonably general, and the dimension of the spacetime does not enter in a significant way. We hope that our closed-form expressions will both ease computation in the theory, and enable physicists to draw on a wider mathematical theory in using LORE.
|
\section{Introduction}
\label{Introduction}
In the last years several pieces of evidence have shown the presence of multiphase gas in the halo regions of low-redshift galaxies. This multiphase halo gas, called the `circumgalactic medium' (CGM), is explained by the existence of flows of gas towards and away from galaxies, in agreement with the idea of a strong interplay between galaxies and their intergalactic environment. Thus, studying the nature of the CGM is essential to understand how galaxies evolve in their environments.
The presence of a diffuse and hot circumgalactic phase is predicted by current cosmological theories. Low-redshift galaxies are expected to be surrounded by gas atmospheres at the virial temperature ($T \gtrsim 10^6$ K for Milky Way galaxies), the so-called `cosmological coronae' \citep{Fukugita&Peebles06}, extending out to hundreds of kpc from the galaxy center. Unfortunately the X-ray surface brightness of these coronae is rather faint, limiting their possibility of detection \citep{Bregman07}. In the Milky Way, the existence of a hot corona was originally postulated by \citet{Spitzer56} as the medium that provides the pressure required to confine the High-Velocity Clouds \citep[HVCs, e.g.][]{Wakker&vanWoerden97}. Most of the evidence of a hot corona collected over the last decades has been indirect and comes, for instance, from the rotation measure of pulsars in the Large Magellanic Cloud \citep{Gaensler+08,Anderson&Bregman10}, \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ absorption lines in spectra of halo stars or extragalactic sources \citep{Sembach+03, Fox+10}. Moreover, the head-tail structure of almost all the HVCs \citep[][]{Bruens+00, Putman+11}, the asymmetry of the Magellanic Stream \citep[MS,][]{Mastropietro+05} and the loss of external gas from the dwarf spheroidal galaxies of the Local Group \citep{Grcevich&Putman09, Gatto+13, Salem+15}, may all be explained by ram-pressure exerted by the coronal medium.
Recently, the presence of a hot corona around our Galaxy was revealed by combining detections of \ifmmode{\rm OVII}\else{O\/{\sc vii}}\fi\ and \ifmmode{\rm OVIII}\else{O\/{\sc viii}}\fi\ absorption lines in quasar spectra and emission lines in the soft X-ray background \citep{Miller&Bregman15}.
Haloes of hot gas have also been observed around some massive spiral galaxies. In these cases X-ray emission has been detected at more than 50 kpc from the center, indicating the presence of extended structures. Once extrapolated to the virial radius, the mass of these coronae is comparable with the mass of the baryonic discs of these galaxies ($\sim 10^{11}$ M$_\mathrm{\odot}$), accounting for $\sim 10-50\%$ of their associated `missing baryons' \citep[e.g.][]{Dai+12, Bogdan+13, Anderson+16}. By combining observations of \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ absorbers around star-forming galaxies (from the COS-Halos survey, see below) together with the \ifmmode{\rm OVII}\else{O\/{\sc vii}}\fi\ and \ifmmode{\rm OVIII}\else{O\/{\sc viii}}\fi\ absorption associated with our Galaxy in a single model of corona, \citet{Faerman+16} found that the typical coronal gas mass of a Milky-Way-like galaxy is $\sim 1.35\times 10^{11} \mo$.
Regarding the cooler phase of the CGM, \hi\ observations in the Milky Way revealed a population of HVCs, characterized by velocities inconsistent with the rotation of the Galactic disc \citep[e.g.][]{Wakker&vanWoerden97} and generally located at distances of $\sim 5-10$ kpc from the Galactic disc \citep[e.g][] {Wakker01,Wakker07}. The MS and its leading arm (LA) are an exception. They are situated in the interval at $\sim 50-200$ kpc from the disc \citep{Putman+12} and produced by interaction between the the Large Magellanic Cloud, the Small Magellanic Cloud and the Milky Way. Notably, the LA is composed by a number of head-tail compact HVCs larger than the MS, suggesting that they are caused by ram pressure stripping exerted by the coronal medium of the Milky Way \citep{Venzmer+12,For+13}.
The presence of the HVCs in the CGM around our Galaxy provides therefore direct evidence of a cold medium ($T \lesssim 10^4$ K). Furthermore, a significant fraction of these HVCs exhibits absorptions from elements at high ionization state \citep[e.g. \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi, \ifmmode{\rm CVI}\else{C\/{\sc vi}}\fi, \ifmmode{\rm SiVI}\else{Si\/{\sc vi}}\fi,][]{Sembach+03,Fox+04} indicating the presence of a more highly ionized and hotter medium at $10^{5-6}$ K, likely due to turbulent mixing at the interface between the HVCs and the corona \citep[e.g.][]{Kwak+11}. In recent years, lines of different ions (e.g. \cii, \ciii, \siii, \siiii), probing low ionization state material at $T \sim 10^{4-5}$ K, have been observed in absorption towards extragalactic sources in the Milky Way's halo \citep[e.g.][]{Shull+09,Lehner&Howk+11}. They fill about $70-90 \%$ of the sky and are often associated with the neutral gas emission from HVCs \citep[e.g.][]{Lehner+12}.
Beyond the Milky Way, \hi\ emission from the CGM is harder to detect due of its low column densities. However, nearby galaxies observed with enough sensitivity have systematically revealed the presence of gas complexes and layers of so-called extraplanar gas that typically extend up to a few~$\sim10$ kpc above and below the plane of the galaxy discs \citep[e.g.][]{Oosterloo+07, Gentile+13}.
Extraplanar gas layers likely consist of individual clouds whose origin may be internal or external to the galaxy \citep{Fraternali02, Boomsma+08}.
Their masses, sizes and location closely resemble those of the largest HVCs of the Milky Way \citep[see discussion in][]{Sancisi+08}. In the Local Group, M\,31 has a population of \hi\ clouds with masses down to $\sim 10^5$ $\mo$ and located at distance of tens of kpc from the galaxy \citep{Thilker+04, Lewis+13}.
Recently, the COS-Halos survey has constrained the physical properties of the CGM in nearby galaxies \citep[e.g.][]{Werk+12, Tumlinson+13}, detecting gas through absorption lines against background QSO spectra for a sample of galaxies in the low-redshift ($0.1<z<0.35$) Universe. The picture that has emerged is that a large fraction of the sample galaxies, regardless of their type, are surrounded by a cool and ionized circumgalactic phase (T $ \lesssim 10^5$ K): both $\Lya$ absorbers and low/intermediate ionization elements (\cii, \ciii, \mgii, \siii, \siiii) have shown strong column densities out to projected distances of 150 kpc from the galaxy centre \citep{Tumlinson+13, Werk+13}. Detections of highly ionized material (e.g. \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi) have also been reported \citep{Tumlinson+11, Werk+14}. The amount of \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ around star-forming galaxies is larger than around galaxies with little or no star formation \citep{Tumlinson+11}, indicating that the warmer CGM reflects the bimodality of the two types of galaxies.
The aim of this paper is to shed light on the coexistence and the ubiquity of different gas phases in the CGM, and, in particular, to understand under what conditions the cool gas can survive in the hot corona. Our working hypothesis is that the observed cool/warm and ionized phase is associated with the interaction and mixing between the hot coronal medium and a cooler neutral phase. The mechanism originating this cool phase can be various: for instance the Milky-Way's HVCs, gas streams due to tidal/ram pressure stripping (e.g. the MS) or, potentially, clouds ejected by powerful outflows \citep[e.g. the HVC toward the Large Magellanic Cloud,][]{Barger+16}. In this paper, we do not investigate the origin of the cool clouds but focus on the problem of their survival. In Sec. \ref{Hydrodynamical simulations} we introduce the set of hydrodynamical simulations performed justifying the choices of the parameters and we briefly describe the main features of the code that we used. In Sec. \ref{Results} we present the results of our simulations and we compare them with the COS-Halos observations. In Sec. \ref{Discussion} we discuss our simulations by referring to other numerical works investigating the survival of cool clouds, while in Sec. \ref{Conclusions} we summarize our main results.
\section{Hydrodynamical simulations}
\label{Hydrodynamical simulations}
We performed a suite of two-dimensional hydrodynamical simulations of cool neutral clouds (T $=10^4$ K) travelling through a hot and low-density coronal medium at given initial velocities (see Sec.~\ref{Discussion1} for a discussion about the possible limitations of the 2D hydrodynamical simulations). Hydrodynamical instabilities (both Kelvin-Helmholtz and Rayleigh-Taylor instabilities) tear up the cool cloud during its motion, producing a long turbulent wake where cloud and corona materials mix efficiently producing gas at intermediate temperature \citep[see][]{Murray+93,Heitsch&Putman09,Kwak+11}. This gas mixture may evaporate in the surrounding hot medium or, if radiative cooling is effective, condensate in cooler and smaller structures \citep{Marinacci+10,Armillotta+16}.
Hereafter we distinguish three ranges of temperature:
\begin{itemize}
\item \textit{cool gas}: gas at $T<10^5$ K, both neutral and ionized, responsible for the absorptions in \hi\ and low-intermediate ionization-state elements;
\item \textit{warm gas}: gas at $10^5<T<10^6$ K, responsible for the absorptions in \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ and other high ionization-state elements;
\item \textit{hot gas}: gas at $T>10^6$ K, typical of a galactic corona at the virial temperature.
\end{itemize}
The main goal of our work is to understand, by sampling different cloud sizes and velocity, whether the cool gas may survive its journey through the coronal medium. When we refer to the cool gas, we take into account both the cool gas belonging to the remnant of the initial cloud, and the cool gas present in the turbulent wake.
\begin{table}
\centering
\begin{tabular}{cccccc}
\hline
\hline
$T_\mathrm{cor}$& $n_\mathrm{cor}$& $Z_\mathrm{cor}$& $T_\mathrm{cl}$& $n_\mathrm{cl}$& $Z_\mathrm{cl}$\\
(K) & (cm$^{-3}$) & (Z$_\mathrm{\odot}$) &(K) & (cm$^{-3}$) & (Z$_\mathrm{\odot}$)\\
\hline
$2\times10^6$& $10^{-4}$&$0.1$ &$10^4$&$2\times10^{-2}$&$0.3$\\
\hline
\hline
\end{tabular}
\caption{Initial parameters of all our simulations: coronal temperature $T_\mathrm{cor}$, coronal density $n_\mathrm{cor}$, coronal metallicity $Z_\mathrm{cor}$, cloud temperature $T_\mathrm{cl}$, cloud density $n_\mathrm{cl}$, cloud metallicity $Z_\mathrm{cl}$.}
\label{FixedParameters}
\end{table}
The parameters fixed in each simulation are listed in Table \ref{FixedParameters}. In all simulations the parameters of the coronal medium were chosen assuming the typical properties of the Milky Way at large distances from the galactic disc (50-150 kpc, following the observed impact parameters of the COS-Halos survey). The coronal temperature was set to $2\times10^6$ K, roughly the coronal temperature of the Milky Way \citep{Fukugita&Peebles06, Miller&Bregman15}, while the coronal metallicity was set to $0.1$ Z$_\mathrm{\odot}$, according to the value estimated for those galaxies in which the hot halo was actually observed in X-rays \citep{Bogdan+13, Hodges-Kluck&Bregman13, Anderson+16}. For the Milky Way the value is not well constrained, but studies through both Far Ultraviolet absorption spectra and emission lines of \ifmmode{\rm OVII}\else{O\/{\sc vii}}\fi\ and \ifmmode{\rm OVIII}\else{O\/{\sc viii}}\fi\ returned values between $0.1$ and $0.3$ Z$_\mathrm{\odot}$ \citep{Sembach+03, Miller&Bregman15}. We assumed a coronal number density of $10^{-4}$ cm$^{-3}$, this value is compatible with the value of $\sim 10^{-4}$ cm$^{-3}$ necessary to explain the multiphase medium at the MS location \citep[e.g.][]{Fox+05, Kalberla+06}, the average value of $\sim 2\times10^{-4}$ cm$^{-3}$ within $100$ kpc found by \citet{Grcevich&Putman09}, the coronal number density of $1.3-3.6\times10^{-4}$ cm$^{-3}$ in the range $50-90$ kpc from the Galactic disc found by \citet{Gatto+13} and the upper limit of $\sim 10^{-4}$ cm$^{-3}$ found by \citet{Besla+12} by reconciling simulations of MS formation and observations.
We assumed pressure equilibrium between the cloud and the external medium. Our numerical experiments have indeed shown that if the cloud is initially out of pressure equilibrium with the ambient medium it readjusts itself and it reaches pressure equilibrium in a few Myr. With this prescription the cloud number density is fixed by environmental parameters to $2\times10^{-2}$ cm$^{-3}$. The cloud metallicity was fixed to 0.3 Z$_\mathrm{\odot}$ in all simulations. This value is in agreement with the metallicities measured for most high-velocity complexes in the Milky Way, typically between 0.1 and 0.5 Z$_\mathrm{\odot}$ \citep[e.g.][]{Wakker01, Collins+07, Shull+11}. We point out that a value of 0.3 Z$_\mathrm{\odot}$ may indicate that the metallicity is not primordial and it could have been enhanced from the star-forming disc \citep{Fraternali+15, Marasco+17}. A metallicity of 0.3 is also representative of the Magellanic Clouds \citep{Hunter+09}.
\begin{table}
\centering
\begin{tabular}{cccc}
\hline
\hline
Sim.&$v_\mathrm{cl}$& $r_\mathrm{cl}$& $M_\mathrm{cl}$\\
&($\kms$) & (pc)& ($\mo$)\\
\hline
1&$100$&$50$&$1.6\times10^2$\\
2&$200$&$50$&$1.6\times10^2$\\
3&$300$&$50$&$1.6\times10^2$\\
4&$100$&$100$&$1.3\times10^3$\\
5&$200$&$100$&$1.3\times10^3$\\
6&$300$&$100$&$1.3\times10^3$\\
7&$100$&$250$&$2\times10^4$\\
8&$200$&$250$&$2\times10^4$\\
9&$300$&$250$&$2\times10^4$\\
10&$100$&$500$&$1.6\times10^5$\\
11&$200$&$500$&$1.6\times10^5$\\
12&$300$&$500$&$1.6\times10^5$\\
\hline
\hline
\end{tabular}
\caption{List of the performed simulations. We varied both the initial cloud velocity, $v_\mathrm{cl}$, and the initial cloud radius, $r_\mathrm{cl}$. Since the cloud number density is fixed in all the performed simulations, different initial cloud radii correspond to different initial cloud masses, $M_\mathrm{cl}$.}
\label{ChangedParameters}
\end{table}
The values of the parameters that characterize the different simulations are listed in table \ref{ChangedParameters}. The two parameters really changing in our set of simulations are the initial cloud velocity with respect to the hot gas, $v_\mathrm{cl}$, and the initial cloud radius, $r_\mathrm{cl}$. Our goal is to investigate how these two physical quantities influence the cloud survival. We assumed three different values for the initial velocity, 100, 200 and 300 $\kms$. These values are compatible with the results found by \citet{Tumlinson+13} and \citet{Werk+13}: the material detected by the COS-Halos observations is within approximately $\pm 200 \kms$ of the galaxy systemic velocity. This range also encompasses the typical rotational velocities for not-extremely elongated orbits in Milky Way-like halos \citep[e.g.,][]{Lux+13}. The values chosen for the initial radius of the cloud are 50, 100, 250 and 500 pc, corresponding to cloud masses of $1.6\times10^2$, $1.3\times10^3$, $2\times10^4$ and $1.6\times10^5$ M$_\mathrm{\odot}$, respectively. The estimated masses of most of the Milky Way HVCs, for which good distance constraints exist, lie in a range of higher masses, between $10^5$ and $5\times10^6 \mo$ \citep{Putman+12}. However, these complexes are always composed by smaller clouds \citep{Thom+08,Hsu+11}, then the masses which we chose for our simulations are fully justified. Also the MS shows clear evidence for a hierarchy of structures, in the form of cloudlets around the main filaments of the Stream \citep{Putman+03,Fox+14}. The same cloud sizes are observed around external galaxies, in particular M\,31 \citep{Thilker+04}.
In our simulations the clouds are initially spherical. This assumption is unrealistic because the geometry of the \hi\ clouds around our Galaxy is strongly irregular \citep[e.g.][]{Putman+12}. In order to make our simulations more realistic, we allowed for a quick breakdown of the sphericity and homogeneity of the cloud by introducing internal random motions inside it. Both in most Milky Way HVCs and in the MS, the typical velocity dispersion, $\sigma$, observed for the gas phase at $T \lesssim 10^4$ K is $\sim 12 \kms$ \citep[e.g.][]{Kalberla+06}. This dispersion is mainly due to turbulent motions, rather than to thermal broadening. We introduced turbulence inside the cloud both along the x-axis and the y-axis: the initial velocity inside the cloud follows a gaussian distribution with dispersion $10 \kms$ for both the axes and mean value $0 \kms$ for the y-axis and the initial cloud velocity, $v_\mathrm{cl}$, for the x-axis, which is the cloud direction of motion \citep[see also][]{Armillotta+16}.
\subsection{The numerical scheme}
\label{The numerical scheme}
All simulations were performed with the ATHENA code \citep{Stone+08}, using a two-dimensional Cartesian geometry with fixed grid. The 2D geometry suppresses one of the dimensions perpendicular to the cloud velocity and it results in simulating an infinite cylinder that is moving perpendicular to its long axis. From the simulations we obtained quantities per unit length of the cylinder, then, to relate these to the corresponding quantities for an initially spherical cloud of radius $r_\mathrm{cl}$, we multiplied the cylindrical results by the length $4r_\mathrm{cl}/3$ within which the mass of the cylinder equals the mass of the spherical cloud. We used this correction to calculate the gas mass below a given temperature for all the simulations presented in this work \citep[see also][]{Marinacci+10}.
We implemented an algorithm to make the grid adaptive and moving, in order to follow the cloud along its motion. This is necessary in order to reduce the size of the computational domain and the time needed to run the simulations. After every Myr the algorithm repositions the cloud's head at the grid centre. Open boundary conditions were imposed at the four sides of the simulation domain. When the cloud head is repositioned at the grid centre, new gas is added in the grid portion towards which the cloud is moving. We force this gas to assume the same physical conditions of the unperturbed hot corona.
The spatial resolution is 2pc x 2pc. Our results on the study of the code convergence showed that this resolution appear to be optimal for limiting divergence problems related to numerical diffusion \citep[see][for details]{Armillotta+16}.
We did not take into account gravitational acceleration and coronal density variation along the cloud motion. Including a gravitational field is indeed beyond our current purposes. We were interested to study the average environmental effects on the cloud survival so we did not specify its trajectory (outflow or inflow) or its exact location in the circumgalactic halo. Furthermore, we neglected the presence of self-gravity inside the cloud. The effect of self-gravity is to stabilize the cloud against the formation of hydrodynamical instabilities, slowing down its ablation \citep{Murray+93}. However, in our simulations, the virial ratio of the clouds ranges from a few $\times~10^4$ (clouds with $r_\mathrm{cl}= 50$ pc) to a few $\times~10^2$ (clouds with $r_\mathrm{cl}= 500$ pc). In all cases, it is orders of magnitude larger than the unity, then we expect that self-gravity plays a minor role in our simulations.
We included thermal physics in our simulations by adding terms to take into account radiative cooling/heating and thermal conduction to the energy equation:
\begin{equation}
\dfrac{\partial e}{\partial t} + \mathbf {\nabla} \cdot [ (e+P) \mathbf{v}]= \,-\, \rho^2 \Lambda_\mathrm{net} - \mathbf {\nabla} \cdot \textbf{\textit{q}}
\label{Energy}
\end{equation}
where \textit{e} is the energy density, \textbf{v} the velocity, \textit{$\rho$} the density and $P=(\gamma-1) U$ the pressure with $U$ the internal energy density and $\gamma=5/3$, \textbf{\textit{q}} is the so-called `heat conduction flux', $\Lambda_\mathrm{net}$ is the net cooling/heating rate ($\Lambda_\mathrm{net}= \Lambda - \Gamma$, where $\Lambda$ and $\Gamma$ are, respectively, the radiative cooling and the photo-heating rate).
Despite modules for these processes were present in ATHENA, we modified them to make the code more suitable for our purposes, as explained below.
\subsubsection{Radiative processes}
The effects of radiative cooling and heating were included in the code by using the cooling and heating rates calculated through the CLOUDY spectral synthesis code \citep[version c13; ][]{Ferland+13} and tabulated across a range of physical conditions. We evaluated the effect of photoionization from a uniform UV background at $z=0.2$ (the average redshift of the COS-Halos sample of galaxies), accounting for the UV radiation emitted by all stars and AGN throughout the evolution of the Universe and attenuated by the $\Lya$ forest \citep{Haardt&Madau12}. Collisional ionization of all atoms and ions is also included in the model. All chemical elements were assumed to be in ionization equilibrium, which means that atomic processes (recombination, photoionization, collisional ionization) become steady in a timescale lower than the hydrodynamical timescale.
Cooling and heating rates were tabulated as a function of total gas temperature T, metallicity Z, and hydrogen numerical density n$_\mathrm{H}$. T ranges from $10^{4}$ to $10^{7}$ K with a resolution of 0.05 dex in the log space. n$_\mathrm{H}$ ranges from $10^{-6}$ to $1$ cm$^{-3}$ with a resolution of 0.5 dex. The gas metallicity, Z, assumes only two values, $0.3$ Z$_\mathrm{\odot}$ and $0.1$ Z$_\mathrm{\odot}$. The abundances at ${Z =}$ Z$_\mathrm{\odot}$ and ${Z = 0.1}$ Z$_\mathrm{\odot}$ were taken from \citet{Sutherland&Dopita93}. For the $Z=0.3$ Z$_\mathrm{\odot}$ model, we linearly interpolated the abundances at ${Z =}$ Z$_\mathrm{\odot}$ and ${Z = 0.1}$ Z$_\mathrm{\odot}$ on logarithmic scale.
\begin{figure}
\includegraphics[width=0.48\textwidth]{./CLOUDY.pdf}
\caption{Absolute value of net cooling/heating rate as a function of gas temperature for $n=2\times10^{-2}$ cm$^{-3}$ and $Z=0.3$ $Z_\mathrm{\odot}$ (orange line) and $n=10^{-4}$ cm$^{-3}$ and $Z=0.1$ Z$_\mathrm{\odot}$ (brown line). We plot the absolute value of the radiative cooling minus the photoheating rate per unit volume, divided by $n_\mathrm{H}^2$. We use a dashed line to indicate that at lower gas density and metallicity heating dominates over cooling for $T< 10^{4.3}$ K and the net effect is heating.}
\label{CLOUDY}
\end{figure}
Since the tabulated cooling/heating rates depend on n$_\mathrm{H}$, we need to know the value of n$_\mathrm{H}$ in each cell of the computational domain and at each time step. For this reason, we calculated the temporal and spatial evolution of n$_\mathrm{H}$ by modelling it as a passive scalar field. Numerically, we treated advection of hydrogen mass in the same manner as advection of mass. The initial conditions were calculated in each cell by multiplying the total gas density by the hydrogen abundance at given metallicity.
The results of the photoionization model are presented in Fig. \ref{CLOUDY}. The plot shows the net cooling/heating rates as a function of temperature in the log scale for two extreme couple of values present in our simulations, one with density and metallicity of the cloud ($n=2\times10^{-2}$ cm$^{-3}$ and $Z=0.3$ $Z_\mathrm{\odot}$) and the other with density and metallicity of the coronal medium ($n=10^{-4}$ cm$^{-3}$ and $Z=0.1$ Z$_\mathrm{\odot}$). Most of our cells falls typically within these values. At $T\sim 10^{4.3}$ K the curve at low density and metallicity show a discontinuity, below which heating dominates over cooling. We plotted this region using a dotted line, in order to distinguish it from the region where cooling dominates.
\subsubsection{Thermal conduction}
According to the classical theory, the heat conduction flux is given by the Spitzer formula \citep{Spitzer62}:
\begin{equation}
\textbf{\textit{q}} = - \kappa_\mathrm{Sp} \cdot \mathbf {\nabla} T \:,
\label{Spitzer}
\end{equation}
where ${\nabla} T$ is the temperature gradient, and the heat conduction coefficient is
\begin{equation}
\kappa_\mathrm{Sp} = \dfrac{1.84 \times 10^{-5} }{\mathrm{ln\Psi}} \cdot T^{5/2}\: \: \: \mathrm{erg \,s^{-1} \,K^{-1}\, cm^{-1}} \:,
\label{Spitzer2}
\end{equation}
where $\ln\Psi$ is the Coulomb logarithm and it can be expressed as
\begin{equation}
\mathrm{ln\Psi} = 29.7 + \mathrm{ln} \left[ \dfrac{T_{e} / 10^6 K}{\sqrt{n_{e} / cm ^{-3}}}\right]
\end{equation}
with $n_\mathrm{e}$ being the electron density and $T_\mathrm{e}$ the electron temperature.
The classical prescription breaks down in the presence of a magnetic field and the efficiency of thermal conduction may be strongly reduced \citep[e.g.][]{Chandran&Cowley98}. To take into account this effect we assumed an efficiency of 10\% for the Spitzer thermal conduction (Eq. \ref{Spitzer}). This value is in agreement with the result found by \citet{Narayan&Medvedev01}, which studied the thermal conduction suppression when the magnetic turbulence extends on a wide range of length scales, as it might happen with strong-intermediate MHD turbulence. Furthermore, the efficiency of the classical thermal conduction can be reduced when the local temperature scale-length falls below the mean free path of the conducting electrons. In this case the heat flux is replaced by a flux-limited form the so-called `saturated heat flux' \citep{Cowie&McKee77}:
\begin{equation}
\vert \,\textbf{\textit{q}}_\mathrm{sat}\vert = 5 \Phi_{\mathrm{s}}\rho c^3\;,
\label{Saturated1}
\end{equation}
with the sound speed \textit{c} and the density $\rho$. $ \Phi_{\mathrm{s}}$ is an efficiency factor less than or of the order of unity, related to the flux-limited treatment and flux suppression due to magnetic fields.
To take into account all these mentioned effects, we modified the Spitzer formula (Eq. \ref{Spitzer}):
\begin{equation}
\textbf{\textit{q}} = - 0.1\,\dfrac{\kappa_\mathrm{Sp} }{1+\sigma} \cdot \mathbf {\nabla} T \;,
\label{Eqcode}
\end{equation}
where $\sigma$ is the absolute ratio between the classical heat flux and the saturated heat flux, according to \citet{Balbus86}. We solved this equation by including a module for isotropic thermal conduction in ATHENA. This module was implemented through an implicit scheme in order to avoid very restrictive CFL constraints on the time step. Further details on the hydrodynamical treatment of thermal conduction and the explanation of the algorithm have been described in \citet{Armillotta+16}.
\section{Results}
\label{Results}
\begin{figure}
\includegraphics[width=0.48\textwidth]{./cloudsurv.pdf}
\caption{Parameter space (radius-velocity-mass) explored in our simulations to test cool gas survival for 250 Myr. The number written in each square and the color bar indicate the percentage of cool gas (T $\leq 10^5$ K) that survives. The grey band represents the range of parameters in which the cloud is not able to survive.}
\label{Fig1}
\end{figure}
We run our simulations of cool clouds interacting with the surrounding coronal gas for 250 Myr.
We found that the mass of cool gas decreases with time in all the performed simulations. As mentioned in Sec.\ref{Hydrodynamical simulations}, the cool gas, stripped from the cloud by hydrodynamical instabilities, mixes with the coronal gas in a turbulent wake. In all the analyzed cases, the wake density is not high enough to allow for gas condensation \citep[like in][]{Armillotta+16}, and most of cool gas lost from the cloud evaporates in the hot coronal medium. However, in several cases, a consistent fraction of cool gas survives until the end of the simulation: most of this gas resides inside the cloud head and only a very small fraction is located in the wake, as we show in Sec \ref{The role of thermal conduction}.
Fig.~\ref{Fig1} shows the values of initial radius and velocity of the cloud that allow the cool gas survival for at least 250 Myr, and the fraction of gas able to survive after this time. Clouds with radius equal to or less than 100 pc are entirely destroyed after 250 Myr. Clouds with radius equal to 250 pc are able to survive. However, the fraction of surviving mass strongly depends on the initial cloud velocity, decreasing with increasing velocity. The cloud with $v_\mathrm{cl}=100 \kms$ is able to keep $\sim50\%$ of its own initial mass after 250 Myr, while the cloud with $v_\mathrm{cl}=300 \kms$ retains only $\sim 7\%$. The situation changes completely for larger clouds ($r_\mathrm{cl}=500$ pc): a significant fraction of cool gas survives ($\gtrsim 70 \%$) and its dependence on the initial velocity becomes rather weak.
\begin{figure}
\includegraphics[width=0.48\textwidth]{./Bernoulli.pdf}
\caption{Temperature snapshots of the simulation with initial cloud velocity $100\kms$ and initial cloud radius $250$ pc (Sim.~7 in Tab.~\ref{ChangedParameters}) with thermal conduction. The time at which the snapshots have been taken is indicated in each panel.}
\label{Fig1.2}
\end{figure}
\begin{figure*}
\includegraphics[width=0.9\textwidth]{./snapshots_T_Dens.pdf}
\caption{Temperature (top panels) and number density (bottom panels) snapshots of the simulation with initial cloud velocity $100\kms$ and initial cloud radius $250$ pc (Sim.~7 in Tab.~\ref{ChangedParameters}) without (left panels) and with (right panels) thermal conduction. The time at which the snapshots have been taken is 200 Myr.}
\label{Fig2}
\end{figure*}
\begin{figure}
\includegraphics[width=0.48\textwidth]{./massevo.pdf}
\caption{Evolution of the mass of cool gas ($T<10^5$ K) with time for two simulations with initial cloud velocity $100\kms$ and initial cloud radius $250$ pc (Sim.~7 in Tab.~\ref{ChangedParameters}): one without (green line) and one with thermal conduction (red line).}
\label{Fig3}
\end{figure}
A key issue in this analysis is the efficiency of coronal ram pressure to warp the cloud, triggering its ablation. As an example, in Fig. \ref{Fig1.2} we show the temperature distribution on the grid at the initial time (top panel) and after 40 and 80 Myr (middle and bottom panel respectively) for the simulations with initial cloud velocity $100\kms$ and initial cloud radius $250$ pc (Sim.~7 in Tab.~\ref{ChangedParameters}). The coronal flow flattens the cloud perpendicularly to the motion direction, increasing its cross section as a consequence of the Bernoulli effect. The cloud distortion triggers the formation of Kelvin-Helmholtz instabilities resulting from the shear flow over its leading face and Rayleigh-Taylor instabilities along its symmetry axis due to deceleration exerted from the coronal pressure. The development of these two instabilities lead to the cloud ablation. Therefore, the time scale associated to the cloud survival is strongly related to time scale associated to ram pressure exerted on the cloud by the coronal medium, the so-called drag time \citep[e.g.][]{Fraternali&Binney06}:
\begin{equation}
t_\mathrm{drag}=\dfrac{M_\mathrm{cl}}{v_\mathrm{cl}\sigma\rho_\mathrm{cor}}\; ,
\end{equation}
where $\sigma \simeq \pi {r_\mathrm{cl}}^2$ is the cross section of the cloud. The drag time increases with increasing cloud size ($t_\mathrm{drag}\propto{M_\mathrm{cl}}/{\sigma}\propto r_\mathrm{cl}$) and decreases with increasing cloud velocity. The effect of coronal ram pressure on the largest clouds is very weak and they are able to survive for very long times.
The velocity can also play an important role. Fig. \ref{Fig1} shows that the fraction of surviving mass decreases with increasing velocity for the clouds with $r_\mathrm{cl}=250$ pc and for the clouds with $r_\mathrm{cl}=500$ pc and $v_\mathrm{cl}=100 \kms$ and $200 \kms$.
A dependence on the velocity is instead roughly absent for the cloud with $r_\mathrm{cl}=500$ pc and $v_\mathrm{cl}\geqslant200 \kms$. The cloud with $v_\mathrm{cl}=300 \kms$ keeps $\sim70\%$ of its own initial mass, while the cloud with $v_\mathrm{cl}=200 \kms$ keeps $\sim 75\%$, a small difference. In this last case the velocity is supersonic (the sound speed is $\sim200\kms$ at $T=2\times10^6$ K) and the radius is large enough to keep the velocities supersonic for almost the entire simulation time. In fact, in our simulations the cloud velocity evolves with time approximately as \citep[][]{Fraternali&Binney06}:
\begin{equation}
v_\mathrm{cl}(t)\simeq \dfrac{v_\mathrm{cl}(0)}{1+t/ t_\mathrm{drag}} \; .
\end{equation}
For $t\ll t_\mathrm{drag}$, $dv_\mathrm{cl}(t)/ dt \propto 1/t_\mathrm{drag} \propto 1/R_\mathrm{cl}$, then, the larger the cloud radius the slower the cloud deceleration. In the supersonic case the formation of a bow shock in front of the cloud suppresses the formation of Kelvin-Helmholtz instabilities and hampers the cloud destruction \citep[see also e.g.][]{Scannapieco&Bruggen15}.
\subsection{The role of thermal conduction}
\label{The role of thermal conduction}
Thermal conduction is a key mechanism to consider in the interaction between different gas phases because it allows for transfer of heat from a hot to a cool medium, accelerating their mixing.
In order to better understand how thermal conduction can influence the cloud survival, we chose a fiducial simulation, $r_\mathrm{cl}=250$ pc and $v_\mathrm{cl}=100 \kms$ (Sim.~7 in Tab. \ref{ChangedParameters}), and we analyzed it both in the absence and the presence of thermal conduction. Fig.\ref{Fig2} shows the temperature (top panels) and density (bottom panels) distributions on the grid after 200 Myr for the simulations without (left panels) and with (right panels) thermal conduction. The general evolution of the cloud is very different in the two cases. In the absence of thermal conduction the initial cloud is totally destroyed while the turbulent wake is characterized by the presence of a gradient in temperature and in density extended to several kpc downstream in the coronal medium. Cooler and denser regions are situated in the head of the wake, where the gas lost from the cloud is not yet well mixed with the coronal gas. Most of this cool gas is approximately an order of magnitude less dense than the initial cloud ($n\sim10^{-3}$ cm$^{-3}$), then, despite it encompasses a large volume, its contribution to the total mass will be low. Instead, in the presence of thermal conduction the bulk of the initial cloud is partially intact after 200 Myr and the cool gas mainly resides inside the dense cloud, while the wake is nearly lacking cool gas and it is composed by an extremely homogeneous mixture at $T \sim 10^{6}$ K and $n \sim 10^{-4}$ cm$^{-3}$. The effect of thermal conduction is indeed to create a more widespread and warmer wake in which temperature and density gradients are smoothed.
Fig. \ref{Fig3} shows the quantitative results for these two simulations. We compared the evolution of the mass of cool gas with time and we found that it decreases in both cases. The cool gas lost from the cloud evaporates in the coronal medium because the warm wake, produced by the mixing of cool and hot gas, is unable to cool down. As we saw in Fig. \ref{Fig2}, regardless of the presence of thermal conduction, regions in the wake at temperatures close to the peak of the cooling function ($T\sim10^{5.5}$ K) exhibit very low densities ($n\sim10^{-3}$ cm$^{-3}$) and metallicities ($Z<0.3$ Z$_\mathrm{\odot}$). Under these conditions, the condensation of gas \citep[see e.g.][]{Armillotta+16} is hard to occur. The efficiency of condensation strongly depends on gas density and metallicity (see Fig. \ref{CLOUDY}): decreasing density and metallicity, the cooling rates decrease and the effect of heating photoionization becomes more and more important. Therefore, the net result is the loss of cool gas. However, Fig.\ref{Fig3} shows that the reduction of cool gas mass occurs faster without thermal conduction. After 250 Myr the mass of cool gas is $\sim50\%$ of the initial cloud mass in the presence of thermal conduction compared to less than $40\%$ in the absence of it.
Thermal conduction smooths the velocity and density gradients at the interface between the two fluids, hindering the formation of hydrodynamical instabilities and subsequent mixing and making the cloud core more compact \citep[see also][]{Marcolini+05, Vieser&Hensler07b}. This phenomenon explains the slow destruction of the cloud in the presence of thermal conduction. Once the gas is stripped from the cloud, thermal conduction changes its role, accelerating the heating of cool gas and its evaporation in the coronal medium. As shown in Fig.\ref{Fig2}, cool gas is nearly absent in the warm and rarefied cloud wake in the presence of thermal conduction.
\subsection{Column densities: comparison with COS-Halos data}
The key observational properties of the COS-Halos detections is given by their column densities. Here we compare the column densities of our fiducial simulation ($R_\mathrm{cl}=250$ pc, $v_\mathrm{cl}=100 \kms$) at 200 Myr with the COS-Halos observations. We estimated the column density of a given ion, $N_\mathrm{X}$, along vertical lines across the simulation box by summing, for each pixel $i$ along the line of sight, the product between the hydrogen column density, $N_{\mathrm{H},i}$, the abundance of the ion with respect to its own species at the temperature and hydrogen number density of the pixel, $X/A \,(T_i,n_{\mathrm{H},i})$, and the abundance of the species with respect to hydrogen at the pixel metallicity, $A/H \,(Z_i)$, as in the following formula:
\begin{equation}
N_\mathrm{X} \,= \, \sum_{i} N_{\mathrm{H},i} \dfrac{X}{A}(T_i,n_{\mathrm{H},i}) \dfrac{A}{H}(Z_i)
\end{equation}
For this calculation, we excluded the gas at $T>10^6$ K, which is the temperature upper limit for the detected \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\
\citep{Werk+14}.
The abundance of the ion with respect to its own species, $X/A$, was obtained through the CLOUDY package, by using the same model including photoionization and collisional ionization that we described in Sec.~\ref{The numerical scheme} for the cooling/heating rates. The abundance of the species with respect to hydrogen, $A/H$, at a given metallicity was obtained by linearly interpolating the abundances at ${Z =}$ Z$_\mathrm{\odot}$ and ${Z = 0.1}$ Z$_\mathrm{\odot}$, taken from \citet{Sutherland&Dopita93}, on logarithmic scale.
The left panel of Fig.~\ref{Fig4} shows the distribution of \siii, \siiii\ and \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ in bins of column density by directly comparing our simulation predictions to the observations. The blue bars represent the observed data, where we excluded all the upper and lower limits \citep[from][]{Tumlinson+11,Werk+13}. The red bars represent the simulated column densities. The dotted line indicates that the sensitivity limit of the observations, while the simulations allow calculations of much lower values. We found that the ranges of observed and simulated column densities overlay for the low/intermediate ionization elements, while they do not overlap for the \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi. In the first case the mean values of two distributions differ by a factor of a few, while in the second case they differ by almost two orders of magnitude. Furthermore, the observed column densities extend to a higher range of values.
\begin{figure*}
\includegraphics[width=\textwidth]{./COS_SIMcfr_Shift.pdf}
\caption{Histograms of distribution of the observed (blue bars) and simulated (red bars) data in bins of column density. The top panels show the distribution of \siii, the middle panels the distribution of \siiii\ and the bottom panels the distribution of \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi. The simulated data were obtained from the fiducial simulation ($R_\mathrm{cl}=250$ pc, $v_\mathrm{cl}=100 \kms$) after 200 Myr. The left panel was obtained by directly comparing the simulation results with the observations. The central panel was obtained by accounting for the contribution of 3 (average detections in COS-Halos) intevening clouds along the line of sight. The right panel was obtained both by accounting for the contribution of 3, 9 and 60 simulated clouds along the line of sight containing respectively \siii, \siiii\ and \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi. The dotted lines indicate the observational detection limit.}
\label{Fig4}
\end{figure*}
\begin{figure}
\includegraphics[width=0.49\textwidth]{./plot4.pdf}
\caption{Snapshots of \siii\ (top panel) and \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ (bottom panel) number densities for the simulation with initial cloud velocity $100\kms$ and initial cloud radius $250$ pc with thermal conduction. The time at which the snapshots have been taken is 200 Myr. The white color indicates that the ion density is null or negligible (i.e. the cool gas in the cloud head or hot material with $T>10^6$ K, the upper limit of temperature for the detected \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi).}
\label{Fig5}
\end{figure}
The different column densities may be explained by the fact that the observations can probe multiple clouds along the line of sight, while in our simulations we have effectively followed the evolution of a single cloud. The central panel of Fig. \ref{Fig4} shows the distribution of \siii, \siiii\ and \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ in bins of column density, accounting for presence of multiple clouds along the line of sight: the new simulated column densities were obtained by summing the column densities along 3 vertical lines taken randomly from the simulation box.
We made this choice because it corresponds to the typical number of kinematic components present in individual absorption lines \citep{Werk+13}. We found that the column densities obtained from the simulation are in very good agreement with the observations for the \siii. However, the agreement worsens with increasing ionization state. Indeed, the ranges of observed and simulated column densities barely overlap for the \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi, they differ by about one order of magnitude.
\begin{figure}
\includegraphics[angle = 270, width=0.52\textwidth]{./cartoon.pdf}
\caption{Schematic picture of \siii\ and \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ distribution in the CGM. Within each cloud, \siii\ occupies a volume much smaller than \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi. As a consequence, a generic line of sight (red line) crossing multiple clouds intersects a larger number of \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ absorbers rather than \siii\ absorbers.}
\label{Cartoon}
\end{figure}
\begin{figure}
\includegraphics[width=0.45\textwidth]{./COS_SIMcfr_largecloud.pdf}
\caption{Histograms of distribution of the observed (blue bars) and simulated (red bars) data in bins of column density. The top panels show the distribution of \siii, the middle panels the distribution of \siiii\ and the bottom panels the distribution of \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi. The simulated data were obtained from the simulation with $R_\mathrm{cl}=500$ pc, $v_\mathrm{cl}=200 \kms$ (Sim.~11 in Tab.~\ref{ChangedParameters}) after 200 Myr. The histogram of the \siii\ was obtained by directly comparing the simulation results with the observations. The histograms of the \siiii\ and \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ were obtained by considering the presence respectively of 3 and 20 clouds along the line of sight.
The dotted lines indicate the observational detection limit.}
\label{Fig4.2}
\end{figure}
In order to examine the different behaviour of a weakly ionized element, as the \siii, and a highly ionized element, as the \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi, we analyzed their spatial distribution in our simulations. Fig. \ref{Fig5} shows the \siii\ (top panel) and \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ (bottom panel) number density distributions on the grid after 200 Myr in our fiducial simulation. The comparison between these snapshots and the right panels of Fig. \ref{Fig2} shows that the \siii\ traces the coolest and densest gas phase, it is present in the head of the cloud and in a few cool filaments in the wake, while it is totally absent everywhere else. Instead, the \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ is nearly absent in the coolest regions, while it is widely distributed in the wake, with peaks of density at the boundaries of the cool structures, where the mixing with hotter gas occurs and thermal conduction is stronger. Therefore, the \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ traces a warmer (T $\sim 10^{5-6}$ K) and more widespread gas phase, which occupies regions more spatially extended with respect to cooler gas (see Fig. \ref{Fig5}).
Since our simulations are two-dimensional, we can assume that the 2D cloud in Fig.~\ref{Fig5} is the projection on the sky of a 3D cloud. It is therefore natural to expect that, generic lines of sight should cross an higher number of warm wakes containing \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ rather than cool cloud cores containing low ionization elements. In our simulation we estimated that the ratio between the area occupied by \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ and the area occupied by a generic low ionization element, as the \siii, is $\sim 20$. Therefore, detection of $\sim 3$ clouds containing \siii\ could correspond to detection of $\sim 60$ clouds containing \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi. Fig.~\ref{Cartoon} shows a schematic picture of the expected 3D distribution of \siii\ and \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ in the CGM.
The right panel of Fig.\ref{Fig4} shows the numerical distribution of \siii, \siiii\ and \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ in bins of column density, taking into account a larger probability to detect \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ and \siiii\ along the line of sight with respect to \siii. In this case the \siii\ column densities are the same as those in the central panel. The \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ column densities were obtained by summing the column densities along 60 vertical lines taken randomly from the simulation box, where only 3 of these 60 lines contain \siii. The \siiii\ column densities were obtained in the same way as \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ accounting for the ratio between the area occupied by \siiii\ and the area occupied by \siii\ is $\sim 3$. We point out that for these calculations we made the double assumption that our 2D cloud is both the projection on the sky and a slice along the line of sight of the 3D cloud. The second assumption is necessary to extract a more realistic density profile along the line of sight, since, strictly speaking, this direction is suppressed in the 2D cartesian geometry of our simulations.
The ranges of column densities are in very good agreement with the observations both for the low-intermediate and for the high ionization elements.
We also calculated the column densities for the simulation with $r_\mathrm{cl}=500$ pc and $v_\mathrm{cl}=200 \kms$ (Sim.~11 in Tab.~\ref{ChangedParameters}), accounting for the presence of a single cloud containing \siii\ along the line of sight and multiple clouds containing \siiii\ and \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi, as shown in Fig.~\ref{Fig4.2}. The agreement between simulations and observations is excellent for \siii: the larger radius of the cloud implies a greater amount of cool gas and, then, higher column densities. We found a good agreement between the simulations and the observations also for the \siiii\ and \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi.
We conclude by noting that due to the way thermal conduction operates, in particular suppressing cool cloudlets in the wake, warm ($10^{5-6}$ K) gas can be found at large distances (several kpc) from the location of the cool material (see Fig.~\ref{Fig2} and Fig.~\ref{Fig5}). This may be relevant in explaining the detections of \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ \citep[see][for other possible explanations]{Werk+16} not directly associated with the classical Galactic HVCs, observed in \hi, or with the Magellanic Stream \citep{Sembach+03, Savage+03}.
Indeed, our simulations exhibit the presence of material at intermediate temperature in the wake for times of the order of hundreds of Myr.
\section{Discussion}
\label{Discussion}
\subsection{Comparison of models}
\begin{figure}
\includegraphics[width=0.49\textwidth]{./Metallicity.pdf}
\caption{Evolution of the mass of cool gas ($T<10^5$ K) with time for two simulations with different cloud metallicity, $Z_\mathrm{cl} = 0.1$ Z$_\odot$ (yellow line) and $Z_\mathrm{cl} = 0.3$ Z$_\odot$ (red line, Sim.7 in Tab.~\ref{ChangedParameters}). Both the simulations starts with initial cloud velocity $100\kms$ and initial cloud radius $250$ pc.}
\label{Fig6}
\end{figure}
All our simulations were performed with cloud metallicity equal to $0.3$ Z$_\odot$. We adopted this value because it is in agreement with the average metallicity measured for most of the high-velocity complexes in the Milky Way (see Sec.~\ref{Hydrodynamical simulations}). So far, the COS-Halos data have not allowed to obtain accurate information about the metallicity of the cool CGM \citep{Werk+14}. In a very recent work, \citet{Prochaska+17} analyzed new far-ultraviolet quasar spectra from the COS-Halos survey and found that the median metallicity of the 32 selected CGM systems is $~ 0.31$ Z$_\odot$, in agreement with the value adopted in this paper. However, the metallicity distribution spans over a wide range of values, with a $95 \%$ interval in the range from $~1/50$ Z$_\odot$ to 3 Z$_\odot$.
Different values of gas metallicity could affect the results of our simulations. For instance, lower values of the gas metallicity would imply a decrease of gas cooling rates. In our simulations, reducing the cloud metallicity entails that the metallicity of the gas mixture behind the cloud decreases. Lower cooling rates of the wake could lead to a faster evaporation of the cool cloudlets within it and, therefore, shorter survival times for the cool gas. On the other hand, if the cloud metallicities is higher than $0.3$ Z$_\odot$, the cool gas could survive for a longer time.
In order to evaluate how a lower cloud metallicity can influence the survival of cool gas, we repeated our fiducial simulation (Sim.7 in Tab.~\ref{ChangedParameters}) assuming the same metallicity for cloud and corona, $0.1$ Z$_\odot$. In Fig.~\ref{Fig6} we compare the evolution of the mass of cool gas with time obtained using the two different cloud metallicities.
The amount of cool gas decreases slightly faster in the case at lower metallicity but the difference between the two trends after 250 Myr is lower than $10 \%$. We conclude that our results do not change significantly if we lower the metallicity of the cool gas.
Our only result incompatible with the COS-Halos findings concerns the number gas density. \citet{Werk+14} found that it is very low ($n \sim 10^{-4}$ cm$^{-3}$) and inconsistent with its being in pressure equilibrium with the hot gas phase. This result is very puzzling because it is not clear how the cool material could survive in these conditions. Indeed, we found that clouds initially out of equilibrium reach pressure equilibrium in a few Myrs. Moreover, despite the cool gas density can decrease once the gas is stripped from the cloud, it remains at least an order of magnitude larger than cool gas density found by \citet{Werk+14}. However, this latter remains quite uncertain due to uncertainty in EUVB radiation field used for the modelling of the \hi\ column densities.
\subsection{Comparison with previous works}
\label{Discussion1}
Simulations of the survival of cool HVCs moving in a hot environment were already performed in the past, showing that the survival timescales are relatively short. \citet{Heitsch&Putman09} found that, for cloud velocities and halo densities typical of the Milky Way, \hi\ clouds are destroyed on timescales of $\sim$ 100 Myr. However, they investigated the survival of \hi\ gas ($T \leq 10^4$ K), while we took into account all gas at $T \leq 10^5$ K, i.e. both neutral and ionized cool gas. Furthermore, they analysed clouds with mass $\lesssim 3\times 10^{4}\, \mo$, while our mass range extends up to $ 1.6\times10^{5} \, \mo$. In order to better compare their result with ours, we analysed the evolution of gas at $T \leq 10^4$ K in our simulations. We obtained that the mass surviving after 250 Myr is $\sim 35 \%$ of the initial mass for the cloud with $M_\mathrm{cl}=2\times 10^4 \, \mo$ and $v_\mathrm{cl} = 100 \kms$, while it is null for higher velocities. This result seems to agree with those of \citet{Heitsch&Putman09}, accounting that they neglected thermal conduction, which hinders the cloud destruction.
In our simulations with $M_\mathrm{cl}=1.6\times 10^5 \, \mo$ the mass of gas at $T \leq 10^4$ K is less than $50\%$ after 250 Myr.
Short survival timescales (100-200 Myr) were also found by \citet{Bland-Hawthorn+07}. They focused on the origin of $\ha$ emission along the MS and explained it as due to the interaction between the \hi\ Stream and the hot corona: the cool ionized gas is shocked material as it is ablated from the Stream clouds by Kelvin-Helmholtz instabilities at the cloud-corona interface. This result agrees with ours: the large amount of ionized gas at $T\sim10^{4-5}$ K found by \citet{Bland-Hawthorn+07} is comparable to the cool gas mass at $T\leq10^5$ K that survives in our simulations.
Both the simulations above mentioned were performed in a three-dimensional cartesian geometry, while all our simulations are two-dimensional. Indeed, 3D simulations would have imply prohibitive computational times for a code with static grid,
as ATHENA. \citet{Armillotta+16} compared low-resolution simulations of cloud-corona interaction obtained by using the two different geometries. We found that, since in 3D the contact surface between the two fluids is larger, hydrodynamical instabilities and subsequent loss of gas from the cool cloud evolve more quickly. In this work, the effect of 2D geometry could therefore be to slow down the cloud destruction. However, in \citet{Armillotta+16} the difference in terms of mass of cool gas between the two geometries is lower than $10\%$ during the whole computational time, so it does not affect significantly the final result. Moreover, we point out that in the present work thermal conduction is much more important to slow down the cloud destruction than for the simulations of \citet{Armillotta+16}. Thus, it is likely that the larger contact surface in 3D simulations would increase the development of velocity gradients at cloud-corona interface and the efficiency of thermal conduction.
Therefore, it is unclear whether the combination of thermal conduction and 3D geometry would result in faster or slower cloud destruction, but we do not expect a significant difference with respect to our 2D simulations.
The simulations in the literature most similar to ours are those of \citet{Kwak+11}, who simulated in a 2D cartesian geometry the motion of HVCs through the Galactic corona. They found that clouds with masses larger than $4\times10^5\, \mo$ remain largely intact after 240 Myr, while we found that also clouds with smaller mass ($M_\mathrm{cl} = 1.6\times10^5\, \mo$) are able to keep a large fraction of their own initial mass ($\gtrsim 70\%$ after 250 Myr). However, as \citet{Heitsch&Putman09}, \citet{Kwak+11} analysed the survival of \hi\ gas, at $T \leq 10^4$ K and neglected thermal conduction.
It is interesting to point out that, similarly to our analysis, they found that high ionization elements, as \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi, are produced by cloud-corona mixing and lines of sight that cross the turbulent wake detect a large number of high ionization elements but a small numbers of low ionization elements, which are nearly absent in the wake.
\citet{Marinacci+10}, \citet{Fraternali+15} and \citet{Armillotta+16} analysed the cloud motion at the disc-corona interface (few kpc above disc), showing that the cloud is able to trigger the condensation of a large portion of coronal material. This result does not contradict those presented here. Indeed, the coronal density in the region close to the galactic disc is around one order of magnitude larger than the coronal density in the outer halo that we considered here. As a consequence, the density of the gas mixture behind the cloud becomes larger when the cloud is close to the galactic disc. The balance between evaporation and condensation strongly depends on this density, since the cooling rate is proportional to its square. Decreasing the wake density, the gas condensation becomes ineffective and the cool gas stripped from the cloud evaporates.
\section{Conclusions}
\label{Conclusions}
In the last years, sensitive observations have revealed that low-redshift galaxies are embedded in extended haloes of multiphase gas, the circumgalactic medium. An important improvement to our knowledge of the CGM was obtained by the COS-Halos survey, which has detected gas through absorption lines against background QSO spectra for a sample of galaxies at low-redshift, finding that most of the sample galaxies, regardless of their type, are surrounded by large amount of cool and ionized gas ($T \leq 10^5$ K), extending out to impact parameters of 150 kpc from the galaxy centre.
In this paper we have explored the physical conditions that allow the survival of cool gas in hot galactic coronae. We performed high-resolution hydrodynamical simulations of cool neutral clouds ($T=10^4$ K) travelling through a hot ($T=2\times10^6$ K) and low-density ($n=10^{-4}$ cm$^{-3}$) ambient medium including cooling, heating and thermal conduction. From our simulations, we conclude that the interaction and mixing between cool clouds and a hot corona lead to a gradual loss of gas from the clouds. The survival time of cool gas strongly depends on initial size (mass) of the cloud: clouds with radius $\gtrsim 250$ pc (mass $\gtrsim 2\times10^4 \mo$) are able to keep a large fraction of their own initial mass for hundreds of Myr.
Thermal conduction appears quite important for the cloud survival since it slows down the cloud destruction.
We compared the column densities of our simulations with the column densities of the COS-Halos observations. The agreement is very good for low-intermediate ionization elements, as \siii\ and \siiii, but not for high ionization elements, as \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi. However, while \siii\ traces a cool gas phase, mainly located inside the cloud, \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ traces a warmer and widespread gas phase, situated in the turbulent wake behind the cloud, produced by the mixing between the cool gas ablated from the cloud and the hot coronal medium. Accounting that observations can detect multiple clouds along the line of sight, the probability of detecting diffuse and extended wakes is much higher than the probability to intercept compact cloud heads. This fact could explain why the range of \ifmmode{\rm OVI}\else{O\/{\sc vi}}\fi\ column densities predicted by our single-cloud simulations does not overlap with the observations.
We can conclude that our results suggest that the existence and the ubiquity of large amount of cool/warm gas in the CGM is possible because large enough clouds, made compact by the effect of thermal conduction, are able to survive their interaction with the hot corona for several hundreds of Myr. This may have important implications for the gas accretion necessary to fuel star-formation in Milky Way galaxies \citep[see e.g.][]{Fraternali&Tomassetti12}. A recurrent question is, for instance, whether a massive structure like the MS can reach the disc of the Milky Way, feeding the star formation \citep[e.g.,][]{Fox+14}. We can speculate that the cloudlets composing the MS with mass $\sim 10^{5}\, \mo$ can survive the journey through the Galactic corona and provide a significant amount of gas accretion.
\section*{Acknowledgements}
LA acknowledges financial support from MARCO POLO 2015-2016. LA is pleased to thank the University of California Santa Cruz for the hospitality during the first phase of this work. JXP acknowledges partial funding by NASA grants HST-GO-13033.06-A and HST-GO-13846.005-A. LA is grateful to Annalisa Citro for helping her to use CLOUDY and to Yusuke Fujimoto for useful discussions. We acknowledge the CINECA award under the ISCRA initiative, for the availability of high performance computing resources and support. We acknowledge the University of California Santa Cruz to allow us to use the Hyades supercluster to perform some test simulations.
|
\section{Introduction}
Two-view triangulation is the problem of estimating a point $X\in \R^3$ from two noisy image projections;
see \cite[Chapter 12]{hartley-zisserman-2003} for its significance in structure from motion in computer vision.
Assuming a Gaussian error distribution, one way to solve the problem is to compute the maximum likelihood estimates (MLE)
for the true image point correspondences. After that the point $X\in \R^3$ can be recovered via linear algebra
\cite{hartley-zisserman-2003}. In this paper we study the above problem of finding the MLEs.
According to the discussion in \cite{aholt-agarwal-thomas} or \cite[Chapter 12]{hartley-zisserman-2003},
the problem is formulated as follows.
Consider a rank two matrix $F\in \R^{3\times 3}$ which is called a {\em fundamental matrix} in multi-view geometry.
This matrix $F$ encodes a pair of projective cameras \cite[Chapter 9]{hartley-zisserman-2003}. Given two points $u_1,u_2\in \R^2$ which denote the noisy image projections, we solve the problem
\begin{align} \label{triangulation}
\begin{split}
\min_{x_1,x_2\in \RR^2} \ & \|x_1 - u_1\|^2_2 + \|x_2- u_2\|^2_2\\
\text{subject to} \ \ & \wh{x}_2^\top F \wh{x}_ 1= 0
\end{split}
\end{align}
where $\wh{x}_k := (x_k^\top \ 1)^\top\in \R^3$ for $k=1,2$.
The equation $ \wh{x}_2^\top F \wh{x}_ 1= 0$ is called the {\em epipolar constraint}, which indicates that $x_1$ and $x_2$ are the true image projections under the projective cameras associated with $F$. The minimizers of \eqref{triangulation} are the MLEs for the true image correspondences, assuming the error is Gaussian.
In \cite[Chapter 12]{hartley-zisserman-2003} (or \cite{hartley1997triangulation}) there is a technique for finding the global minimizers of
\eqref{triangulation} using a non-iterative approach. They use multi-view geometry to reformulate the problem \eqref{triangulation} as minimizing a fraction in a single real variable say $t$. Using the Fermat rule in elementary calculus, it turns out that the minimizers can be computed via finding the real roots of a polynomial in $t$ of degree 6.
In this note, we view the problem \eqref{triangulation} as minimizing a multivariate quadratic polynomial over one single equality constraint, and then employ the classical method of Lagrange multipliers to locate the potential local minimizers. These candidates are called {\em critical points}.
For general rank two matrices $F$ and general points $u_1,u_2$, there are six critical points. They can be computed via finding the roots of a polynomial of degree 6 in the Lagrange multiplier.
Assuming that a global minimizer exists, the minimizer of \eqref{triangulation} can be obtained from the critical points.
\section{Six critical points for two-view triangulation}
\subsection{Reformulation of the problem \eqref{triangulation}}
Given a fundamental matrix $F\in \R^{3\times 3}$ and $u_1=\begin{pmatrix} u_{11} & u_{12}\end{pmatrix}^\top$, $u_2
= \begin{pmatrix} u_{21} & u_{22}\end{pmatrix}^\top\in \R^2$,
consider the invertible matrices
$W_1 :=\lt( \begin{smallmatrix} 1 & 0 & -u_{11} \\ 0 & 1 & -u_{12} \\ 0 & 0 & 1\end{smallmatrix} \rt)$ and
$W_2 :=\lt( \begin{smallmatrix} 1 & 0 & -u_{21} \\ 0 & 1 & -u_{22} \\ 0 & 0 & 1\end{smallmatrix} \rt)$.
Note that $\|x_k-u_k\|^2 = \|\wh{x}_k-\wh{u}_k\|^2$.
and that problem \eqref{triangulation} is equivalent to the problem
\begin{align*}
\begin{split}
\min_{x_1,x_2\in \RR^2} \ & \|W_1 \wh{x}_1\|^2_2 + \|W_2 \wh{x}_2\|^2_2\\
\text{subject to} \ \ & \wh{x}_2^\top F\wh{x}_ 1= 0
\end{split}
\end{align*}
For all $k=1,2$, the last coordinate of $W_k\wh{x}_i$ equals one. As a result, we let $y_k\in \R^2$ be such that
$\wh{y}_k = W_k \wh{x}_k$.
Then \eqref{triangulation}
is further equivalent to the problem
\begin{align} \label{t2}
\begin{split}
\min_{y_1,y_2\in \RR^2} \ & \f{1}{2}\lt( \|\wh{y}_1\|^2_2 + \|\wh{y}_2\|^2_2\rt) \\
\text{subject to} \ \ & \wh{y}_2^\top F'\wh{y}_ 1= 0
\end{split}
\end{align}
where $F':= W_2^{-\top} F W_1^{-1}= \lt( \begin{smallmatrix} a & b & c \\ d & e & f\\ g & h & i\end{smallmatrix} \rt)$ is another fundamental matrix.
\subsection{Derivation of a six degree polynomial}
Let $G(y_1,y_2) := \f{1}{2}\lt( \|\wh{y}_1\|^2_2 + \|\wh{y}_2\|^2_2\rt)$ and
$H(y_1,y_2):=\wh{y}_2^\top F'\wh{y}_ 1$.
The Karush-Kuhn-Tucker (KKT) equation for \eqref{t2}
is $\nabla G + \lambda \nabla H = 0$ for some $\lambda\in \CC$ called the Lagrange multiplier; see
any nonlinear programming text e.g. \cite{bertsekas1999nonlinear}.
Unwinding this equation we obtain
a linear system in four variables, namely,
\begin{align} \label{lsls}
\begin{pmatrix}
1 & 0 & \lambda a & \lambda b \\
0 & 1 & \lambda d & \lambda e \\
\lambda a & \lambda d & 1 & 0 \\
\lambda b & \lambda e & 0 & 1
\end{pmatrix}
\begin{pmatrix}
y_{21}\\
y_{22}\\
y_{11}\\
y_{12}
\end{pmatrix}
= -\lambda
\begin{pmatrix}
c\\
f\\
g\\
h
\end{pmatrix}
\end{align}
where $y_k = \begin{pmatrix} y_{k1} & y_{k2}\end{pmatrix}^\top$ for $k=1,2$,
and $\lambda$ is the Lagrange multiplier.
To acquire the critical points we derive a polynomial equation in $\lambda$. It comes from first expressing
$y_k$, $k=1,2$, in terms of $u_1,u_2,F$ and then substituting these expressions into the epipolar constraint
$ \wh{y}_2^\top F'\wh{y}_ 1= 0$.
Let $A_\lambda$ be the $4\times 4$ coefficient matrix of the above system. One has
$$
\det(A_\lambda) = (bd-ae)^2\lambda^4 - (a^2+b^2 + d^2 +e^2) \lambda^2 + 1.
$$
Define $p_{kl}:= \det(A_\lambda) y_{kl}$ for $k,l=1,2$. By Cramer's rule one has
\begin{align*}
& p_{21} = \lambda[ (bd-ae)(eg-dh)\lambda^3 + (d^2 c+ e^2 c - adf - bef) \lambda^2 + (ag+bh)\lambda-c]\\
& p_{22} = \lambda[ (bd-ae)(ah-bg)\lambda^3 + (a^2 f + b^2 f - acd - bce) \lambda^2 + (dg+eh)\lambda-f]\\
& p_{11}= \lambda[ (bd-ae)(ce-bf)\lambda^3 + (b^2 g + e^2 g - abh - deh) \lambda^2 +(ac+df)\lambda-g]\\
& p_{12}= \lambda[ (bd-ae)(af-cd)\lambda^3 + (a^2 h + d^2 h - abg - deg ) \lambda^2 + (bc+ef)\lambda-h ].
\end{align*}
Consider the polynomial
\begin{align*}
T & := -\det(A_\lambda)^2 \wh{y}_2^\top F' \wh{y}_1 = - p_2^\top F' p_1
\end{align*}
where $p_k := \begin{pmatrix} p_{k1} & p_{k2} & \det(A_\lambda)\end{pmatrix}^\top$ for $k=1,2$.
Since $\det(A_\lambda)$ is a quartic in $\lambda$, and $p_{kl}$ is also a quartic in $\lambda$ for $k,l = 1,2$,
we know $T$ is a polynomial in $\lambda$ of degree at most 8.
By a careful and slightly tedious computation without using any machines, or by using the following {\tt Macaulay2} \cite{M2} code:
\medskip
\noindent
{\small
{\tt
R = QQ[a,b,c,d,e,f,g,h,i,L];\\
A = matrix\{\{1,0,L*a,L*b\},\{0,1,L*d,L*e\},\{L*a,L*d,1,0\},\{L*b,L*e,0,1\}\};\\
detA = det A;\\
p21 = det matrix\{\{-L*c,0,L*a,L*b\},\{-L*f,1,L*d,L*e\},\{-L*g,L*d,1,0\},\{-L*h,L*e,0,1\}\};\\
p22 = det matrix\{\{1,-L*c,L*a,L*b\},\{0,-L*f,L*d,L*e\},\{L*a,-L*g,1,0\},\{L*b,-L*h,0,1\}\};\\
p11 = det matrix\{\{1,0,-L*c,L*b\},\{0,1,-L*f,L*e\},\{L*a,L*d,-L*g,0\},\{L*b,L*e,-L*h,1\}\};\\
p12 = det matrix\{\{1,0,L*a,-L*c\},\{0,1,L*d,-L*f\},\{L*a,L*d,1,-L*g\},\{L*b,L*e,0,-L*h\}\};\\
T = -(a*p11*p21+b*p12*p21+c*p21*detA+d*p11*p22+\\
\hspace*{2.5em} e*p12*p22+f*p22*detA+g*p11*detA+h*p12*detA+i*detA*detA);
}
}
\medskip
\noindent
we know the coefficient of
$\lambda^7$ is zero. The coefficient of $\lambda^8$ is
\begin{align*}
& -(bd-ae)^2 (eg-dh) (ace-abf+baf - bcd + cbd - cae) +\\
& -(bd-ae)^2 (ah-bg)(dce - dbf + eaf - ecd + fbd - fae) +\\
& -(bd-ae)^3 (gce - gbf +haf - hcd + ibd - iae) = (bd-ae)^3 \det(F)=0
\end{align*}
since $F$ has rank two.
This implies $T$ is a polynomial in $\lambda$ of degree at most six. Here we record the explicit expression of $T$:
\begin{align*}
T = \ & (bd-ae)^2 (acg + dfg + bch + efh - a^2 i - b^2 i - d^2 i - e^2 i ) \lambda^6 + \\
& a^{2} c^{2} d^{2} \lambda^{5} + c^{2} d^{4}
\lambda^{5} + 2 a b c^{2} d e \lambda^{5} + b^{2} c^{2} e^{2} \lambda^{5} + 2 c^{2} d^{2}
e^{2} \lambda^{5} + c^{2} e^{4} \lambda^{5} - \\
& 2 a^{3} c d f \lambda^{5} - 2 a b^{2} c d f
\lambda^{5} - 2 a c d^{3} f \lambda^{5} - 2 a^{2} b c e f \lambda^{5} - 2 b^{3} c e f
\lambda^{5} - 2 b c d^{2} e f \lambda^{5} - \\
& 2 a c d e^{2} f \lambda^{5} - 2 b c e^{3} f
\lambda^{5} + a^{4} f^{2} \lambda^{5} + 2 a^{2} b^{2} f^{2} \lambda^{5} + b^{4} f^{2}
\lambda^{5} + a^{2} d^{2} f^{2} \lambda^{5} +\\
& 2 a b d e f^{2} \lambda^{5} + b^{2} e^{2}
f^{2} \lambda^{5} + a^{2} b^{2} g^{2} \lambda^{5} + b^{4} g^{2} \lambda^{5} + 2 a b d e
g^{2} \lambda^{5} + 2 b^{2} e^{2} g^{2} \lambda^{5} + \\
& d^{2} e^{2} g^{2} \lambda^{5} +
e^{4} g^{2} \lambda^{5} - 2 a^{3} b g h \lambda^{5} - 2 a b^{3} g h \lambda^{5} - 2 a b
d^{2} g h \lambda^{5} - 2 a^{2} d e g h \lambda^{5} -\\
& 2 b^{2} d e g h \lambda^{5} - 2
d^{3} e g h \lambda^{5} - 2 a b e^{2} g h \lambda^{5} - 2 d e^{3} g h \lambda^{5} + a^{4}
h^{2} \lambda^{5} + a^{2} b^{2} h^{2} \lambda^{5} + \\
& 2 a^{2} d^{2} h^{2} \lambda^{5} +
d^{4} h^{2} \lambda^{5} + 2 a b d e h^{2} \lambda^{5} + d^{2} e^{2} h^{2} \lambda^{5} +
a^{3} c g \lambda^{4} + a b^{2} c g \lambda^{4} +\\
& a c d^{2} g \lambda^{4} -5 b c d e g
\lambda^{4} + 6 a c e^{2} g \lambda^{4} + a^{2} d f g \lambda^{4} + 6 b^{2} d f g \lambda^{4} +
d^{3} f g \lambda^{4} - \\
& 5 a b e f g \lambda^{4} + d e^{2} f g \lambda^{4} + a^{2} b c h
\lambda^{4} + b^{3} c h \lambda^{4} + 6 b c d^{2} h \lambda^{4} - 5 a c d e h \lambda^{4} +\\
& b c
e^{2} h \lambda^{4} - 5 a b d f h \lambda^{4} + 6 a^{2} e f h \lambda^{4} + b^{2} e f h
\lambda^{4} + d^{2} e f h \lambda^{4} + e^{3} f h \lambda^{4} - a^{4} i \lambda^{4} - \\
& 2 a^{2}
b^{2} i \lambda^{4} - b^{4} i \lambda^{4} - 2 a^{2} d^{2} i \lambda^{4} - 4 b^{2} d^{2} i
\lambda^{4} - d^{4} i \lambda^{4} + 4 a b d e i \lambda^{4} - 4 a^{2} e^{2} i \lambda^{4} - \\
& 2
b^{2} e^{2} i \lambda^{4} - 2 d^{2} e^{2} i \lambda^{4} - e^{4} i \lambda^{4} - 2 c^{2}
d^{2} \lambda^{3} - 2 c^{2} e^{2} \lambda^{3} + 4 a c d f \lambda^{3} + 4 b c e f \lambda^{3} -\\
&
2 a^{2} f^{2} \lambda^{3} - 2 b^{2} f^{2} \lambda^{3} - 2 b^{2} g^{2} \lambda^{3} - 2
e^{2} g^{2} \lambda^{3} + 4 a b g h \lambda^{3} + 4 d e g h \lambda^{3} - 2 a^{2} h^{2}
\lambda^{3} - \\
& 2 d^{2} h^{2} \lambda^{3} - 3 a c g \lambda^{2} - 3 d f g \lambda^{2} - 3 b c h
\lambda^{2} - 3 e f h \lambda^{2} + 2 a^{2} i \lambda^{2} + 2 b^{2} i \lambda^{2} + \\
& 2 d^{2} i
\lambda^{2} + 2 e^{2} i \lambda^{2} + c^{2} \lambda + f^{2} \lambda + g^{2} \lambda + h^{2} \lambda - i.
\end{align*}
\subsection{The six critical points}
By solving $T = 0$ for $\lambda$, we get six (complex) solutions (counting multiplicities) for $\lambda$, say $\lambda_1, \ldots, \lambda_6$.
Plugging in these six values of $\lambda$ into the linear system \eqref{lsls}, solving the linear system for $y_1$ and $y_2$, and computing $x_1$ and $x_2$, one obtains the critical points for two-view triangulation.
If $\det(A_{\lambda_k})\neq 0$ for every $k=1,\ldots,6$ then there are precisely six critical points counting multiplicities.
Now we claim that for general fundamental matrices $F$ and points $u_1,u_2\in \R^2$, there are six distinct critical points for two-view triangulation.
The claim is false if and only if
the discriminant of $T$ or the resultant
of $T$ and $\det(A_\lambda)$ are zero polynomials.
Instead of computing the desired discriminant and resultant which depend on $u_1,u_2$ and $F$, one can find an example of $(u_1,u_2,F)$ such that
the discriminant of $T$ and the resultant
of $T$ and $\det(A_\lambda)$ take a nonzero value, that is,
$\det(A_\lambda)\neq 0$ for every solution $\lambda$ of $T$, and the six critical points obtained are distinct.
If we consider the data $u_1 = \begin{pmatrix} 0 & 0 \end{pmatrix}^\top$,
$u_2 = u_1$ and $F= \lt( \begin{smallmatrix} 1 & 1 & 1 \\ 0 & 1 & 1 \\ 1 & 3 & 3\end{smallmatrix}\rt)$, then the polynomial
$T$ becomes $-2\lambda^6+6\lambda^5+3\lambda^4-12\lambda^3-3\lambda^2+12\lambda-3$, and
there are six distinct complex critical points for the problem \eqref{triangulation}; see Table \ref{t1}.
\begin{table}
\begin{tabular}{|c|c|c|c|}
\hline
$x_{21}$ & $x_{22}$ & $x_{11}$ & $x_{12}$\\
\hline
$0.0596$ & $-0.0321$ & $-0.312$ & $-0.891$\\
\hline
$-0.0843$ & $-2.06$ & $-0.438$ & $-0.0259$ \\
\hline
$-2.42 + 0.0137i$ & $-1.02 -1.56i$ & $-1.57 + 0.714i$ & $-1.246 - 1.51i$ \\
\hline
$-2.42 - 0.0137i$ & $-1.02 +1.56i$ & $-1.57 - 0.714i$ & $-1.246 + 1.51i$ \\
\hline
$-1.69+0.0226i$ & $-0.935 + 0.414i$ & $0.748 + 0.169i$ & $-0.279 - 0.574i$\\
\hline
$-1.69-0.0226i$ & $-0.935 - 0.414i$ & $0.748 - 0.169i$ & $-0.279 + 0.574i$\\
\hline
\end{tabular}
\caption{Six critical points for \eqref{triangulation} when $u_1=(0 \ 0)^\top$, $u_2=u_1$ and
$F=$ \usebox{\smlmat}.} \label{t1}
\end{table}
We summarize the discussion in the following theorem.
\begin{theorem}
For general points $u_1,u_2\in \R^2$ and fundamental matrices $F$, there are six complex critical points
for the problem \eqref{triangulation}.
\end{theorem}
\section{Discussion}
One can make sense of the critical points for $n$-view triangulation where $n$ is greater than two. The authors in
\cite{stewenius2005hard} (cf. \cite{hartley-kahl}) computed the number of critical points for 2 to 7 view triangulation
are 6, 47, 148, 336, 638, 1081. Draisma et al. \cite{DHOST} call this list of numbers the {\em Euclidean distance degrees} of the
multi-view variety associated to 2 to 7 cameras. They conjecture that the general term of this sequence is
$$
C(n):=\f{9}{2}n^3 - \f{21}{2} n^2 + 8n-4.
$$
One can apply the B\'{e}zout's theorem to conclude that $C(n)$ has order $n^3$, and our paper verified $C(2)=6$.
However a proof of the above general formula is still unknown.
\bibliographystyle{plain}
|
\section{B{\normalfont \textbf{I}}DEL$\;$\xspace - Bidirectional SMO{\normalsize \textbf{s}}}\label{sec:singlesmo}
\textsc{BiDEL}\xspace{}'s unique feature is the bidirectional semantics of its SMOs, which is the basis for \textsc{InVerDa}\xspace's co-existing schema versions.
We highlight the design principles behind \textsc{BiDEL}\xspace SMOs and formally validate their bidirectionality.
All \textsc{BiDEL}\xspace SMOs follow the same design principles.
Without loss of generality, the \sql{SPLIT\xspace} SMO is used as a representative example in this section to explain the concepts.
The remaining SMOs are introduced in Appendix~\ref{apx:SMOs}.
Figure~\ref{fig:smomapping} illustrates the principle structure of a single SMO instance resulting from the sample statement
\sqll{\sqlk{SPLIT TABLE} $T$ \sqlk{INTO} $R$ \sqlk{WITH} \ensuremath{c_R}\xspace, $S$ \sqlk{WITH} \ensuremath{c_S}\xspace}%
which horizontally splits a source table $T$ into two target tables $R$ and $S$ based on conditions \ensuremath{c_R}\xspace and \ensuremath{c_S}\xspace.
Assuming both schemas are materialized, reads and writes on both schema versions can simply be delegated to the corresponding data tables $T_D$, $R_D$, and $S_D$, respectively.
However, \textsc{InVerDa}\xspace materializes data non-redundantly on one side of the SMO instance, only.
If the data is physically stored on the source side of an SMO instance, the SMO instance is called \emph{virtualized}; with data stored on the target side it is called \emph{materialized}.
In any case, reads and writes on the unmaterialized side are mapped to the materialized side.
The semantics of each SMO is defined by two functions \getX{tgt} and \getX{src} which describe precisely the mapping from the source side to the target side and vice versa, respectively.
Assuming the target side of \sql{SPLIT\xspace} is materialized, all reads on $T$ are mapped by \getX{src} to reads on $R_D$ and $S_D$; and writes on $T$ are mapped by \getX{tgt} to writes on $R_D$ and $S_D$.
While the payload data of $R$, $S$, and $T$ is stored in the physical tables $R_D$, $S_D$, and $T_D$, the tables $R^-$, $S^+$, $S^-$, $R^*$, $S^*$, and $T'$ are auxiliary tables for the \sql{SPLIT\xspace} SMO to prevent information loss.
Note that the semantics of SMOs is complete, if reads and writes on both source and target schema work correctly regardless on which of both sides the data is physically stored.
This basically means that each schema version acts like a full-fledged database schema; however it does not enforce that data written in any version is also fully readable in other versions.
In fact, \textsc{BiDEL}\xspace ensures this for all SMOs except of those that create redundancy---in these cases the developer specifies a preferred replica beforehand.
Obviously, there are different ways of defining \getX{tgt} and \getX{src}; in this paper, we propose one way that systematically covers all potential inconsistencies and is bidirectional.
We aim at a non-redundant materialization, which also includes that the auxiliary tables merely store the minimal set of required auxiliary information.
Starting from the basic semantics of the SMO---e.g. the splitting of a table---we incrementally detect inconsistencies that contradict the bidirectional semantics and introduce respective auxiliary tables.
The proposed rule sets can serve as a blueprint, since they clearly outline which information needs to be stored to achieve bidirectionality.
To define \getX{tgt} and \getX{src}, we use Datalog---a compact and solid formalism that facilitates both a formal evaluation of bidirectionality and easy delta code generation.
Precisely, we use Datalog rule templates instantiated with the parameters of an SMO instance.
For brevity of presentation, we use some extensions to the standard Datalog syntax:
For variables, small letters represent single attributes and capital letters lists of attributes.
For equality predicates on attribute lists, both lists need to have the same length and same content, i.e.~for $A=(a_1,\ldots,a_n)$ and $B=(b_1,\ldots,b_m)$, $A=B$ holds if $n=m\wedge a_1=b_1 \wedge\ldots\wedge a_n=b_n$.
All tables have an attribute \ensuremath{p}\xspace, an \textsc{InVerDa}\xspace-managed identifier to uniquely identify tuples across versions.
Additionally, \ensuremath{p}\xspace ensures that the multiset semantics of a relational database fits with the set semantics of Datalog, as the unique key \ensuremath{p}\xspace prevents equal tuples in one relation.
For a table $T$ we assume $T(\ensuremath{p}\xspace,\_)$ and $\neg T(\ensuremath{p}\xspace,\_)$ to be safe predicates since any table can be projected to its key.
For the exemplary \sql{SPLIT\xspace}, let's assume this SMO instance is materialized, i.e.\ data is stored on the target side, and let's consider the \getX{tgt} mapping function first.
\sql{SPLIT\xspace} horizontally splits a table $T$ from the source schema into two tables $R$ and $S$ in the target schema based on conditions \ensuremath{c_R}\xspace and \ensuremath{c_S}\xspace:
\begin{align}
R(\ensuremath{p}\xspace,A)&\leftarrow T(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A)\label{rule:rSimple}\\
S(\ensuremath{p}\xspace,A)&\leftarrow T(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A)\label{rule:sSimple}
\end{align}
The conditions \ensuremath{c_R}\xspace and \ensuremath{c_S}\xspace can be arbitrarily set by the user so that Rule~\ref{rule:rSimple} and Rule~\ref{rule:sSimple} are insufficient wrt.\ the desired bidirectional semantics, since the source table $T$ may contain tuples neither captured by \ensuremath{c_R}\xspace nor by \ensuremath{c_S}\xspace.
In order to avoid inconsistencies and make the SMO bidirectional, such tuples are stored on the target side in the auxiliary table $T'$:
\begin{align}
T'(\ensuremath{p}\xspace,A)&\leftarrow T(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A)\label{rule:T'Simple}
\end{align}
\begin{figure}
\centering
\begin{tikzpicture}[node font=\scriptsize,font=\scriptsize,
every node/.style={font=\scriptsize},
opedge/.style={draw, thick, -latex},
catedge/.style={draw, darkgray, thin, -latex},
tab/.style={draw,minimum height=4mm,inner sep=0mm},
contab/.style={tab,minimum width=6mm,fill=white},
dattab/.style={tab,minimum width=5mm,fill=white},
auxtab/.style={tab,minimum width=5mm,fill=white},
cattab/.style={tab,minimum width=3mm,fill=white,font=\tiny},
tabbox/.style={draw,dashed,fill=gray!20},
getop/.style={fill=white},
getopedge/.style={opedge,latex-latex
]
\newcommand{\viewsandtriggers}[1]{
\node[scale=0.7] at ($(#1.north west)+(0.5mm,-0.4mm)$) {\tikz\view;};
\node[scale=0.9] at ($(#1.south east)+(0, 0.5mm)$) {\tikz\gear;};
}
\node(contabT) [contab] {$\,T$};
\viewsandtriggers{contabT}
\node(datatabT) [dattab,anchor=north] at ($(contabT.south)+(0,-8mm)$) {$T_D$};
\node(auxtabsSp) [auxtab,anchor=north] at ($(datatabT.south)+(0,-1mm)$) {$S^+$};
\node(auxtabsRp) [auxtab,anchor=east] at ($(auxtabsSp.west)+(-1mm,0)$) {$R^-$};
\node(auxtabsSm) [auxtab,anchor=west] at ($(auxtabsSp.east)+( 1mm,0)$) {$S^-$};
\node(auxtabsRs) [auxtab,anchor=north east] at ($(auxtabsSp.south)+(-.5mm,-1mm)$) {$R^*$};
\node(auxtabsSs) [auxtab,anchor=north west] at ($(auxtabsSp.south)+(.5mm,-1mm)$) {$S^*$};
\node(contabR) [contab,anchor=south west] at ($(contabT.east)+(38mm,1mm)$) {$R$};
\viewsandtriggers{contabR}
\node(contabS) [contab,anchor=north] at ($(contabR.south)+(0,-2mm)$) {$S$};
\viewsandtriggers{contabS}
\node(datatabR) [dattab,anchor=east,xshift=-0.5mm] at (datatabT.west-|contabS.south) {$R_D$};
\node(datatabS) [dattab,anchor=west,xshift= 0.5mm] at (datatabT.west-|contabS.south) {$S_D$};
\node(auxtabsTp) [auxtab,anchor=center] at ($(auxtabsSp.south-|contabS.south)+(0mm,-.5mm)$) {$T'$};
\node(schemaSrc) [anchor=south] at ($(contabT.north)+(0,5mm)$) {source schema $v_i$};
\node(schemaTrg) [anchor=base] at (schemaSrc.base-|contabR.north) {target schema $v_{i+1}$};
\begin{pgfonlayer}{background}
\node (tbox) [tabbox,fit=(contabT),minimum width=20mm,minimum height=13mm] {};
\node (phytbox) [tabbox,fit=(datatabT)(auxtabsRp)(auxtabsSm)(auxtabsRs)(auxtabsSs),minimum width=20mm,minimum height=13mm] {};
\node (rsbox) [tabbox,fit=(contabR)(contabS),minimum width=14mm,minimum height=13mm] {};
\node (phyrsbox) [tabbox,fit=(datatabR)(datatabS)(auxtabsTp)(auxtabsSs.south-|auxtabsTp),minimum width=14mm,minimum height=11mm] {};
\end{pgfonlayer}
\draw[opedge,-latex] ($(phytbox.north)+(-6mm,-3mm)$)--($(tbox.north)+(-6mm,-3mm)$)
node[anchor=south,inner sep=0.5mm] {read};
\draw[opedge,latex-] ($(phytbox.north)+( 6mm,-3mm)$)--($(tbox.north)+( 6mm,-3mm)$)
node[anchor=south,inner sep=0.5mm] {write};
\coordinate (x_smo) at ($(tbox.east)!0.5!(rsbox.west)$);
\node(partsmo) [draw,darkgray,thin,rounded corners=4pt,inner sep=1mm,font=\tiny\ttfamily] at (contabT.east-|x_smo) {SPLIT\xspace};
\begin{pgfonlayer}{background}
\draw[catedge] (contabT.east) -- (partsmo.west);
\draw[catedge] (partsmo.north) |- (contabR.west);
\draw[catedge] (partsmo.south) |- (contabS.west);
\end{pgfonlayer}
\node(catlbl) [anchor=north,darkgray,inner sep=1mm,font=\tiny] at (tbox.north-|x_smo) {schema catalog};
\coordinate (x_cat) at ($(tbox.west)+(-.5mm,0)$);
\node(contabslbl) [align=center,anchor=east,font=\tiny] at (x_cat|-contabT.center) {views\\\&\\triggers};
\node(dattabsblb) [align=center,anchor=center,font=\tiny] at (contabslbl|-datatabT) {data\\tables};
\node(auxtabsblb) [align=center,anchor=center,font=\tiny] at (contabslbl|-auxtabsTp) {auxiliary\\tables};
\coordinate (x_level) at ($(contabslbl.west)+(-0mm,0)$);
\node(catlbl) [rotate=90,anchor=south,align=center,font=\tiny] at (x_level|-contabslbl.west) {schemas\\versions};
\coordinate (y_storagelbl) at ($(dattabsblb)!0.5!(auxtabsblb)$);
\node(storlbl)[rotate=90,anchor=south,align=center,font=\tiny] at (x_level|-y_storagelbl) {physical\\storage};
\coordinate (y_phyBorder) at ($(tbox.south)!0.5!(phytbox.north)$);
\coordinate (x1_phyBorder) at ($(catlbl.north)+(1mm,0)$);
\coordinate (x2_phyBorder) at ($(schemaTrg.east)+(-1mm,0)$);
\begin{pgfonlayer}{background}
\draw[dotted] (x1_phyBorder|-y_phyBorder) -- (x2_phyBorder|-y_phyBorder);
\end{pgfonlayer}
\draw[getopedge] ($(phytbox.north east)+(0,-2mm)$)
.. controls ($(x_smo|-phytbox.north east)+(0,-2mm)$) and ($(x_smo|-tbox.south east)+(0,2mm)$) ..
($(rsbox.south west)+(0,2mm)$);
\begin{pgfonlayer}{background}
\node(gettrg)[getop,anchor=base west,yshift=-0.5mm] at (tbox.east|-y_phyBorder) {\getX{src}};
\end{pgfonlayer}
\draw[getopedge] ($(phyrsbox.north west)+(0,-2mm)$)
.. controls ($(x_smo|-phyrsbox.north west)+(0,-2mm)$) and ($(x_smo|-rsbox.south west)+(0,2mm)$) ..
($(tbox.south east)+(0,2mm)$);
\begin{pgfonlayer}{background}
\node(getsrc)[getop,anchor=base east,yshift=-0.5mm] at (rsbox.west|-y_phyBorder) {\getX{tgt}};
\end{pgfonlayer}
\node(remark)[align=center,darkgray,inner sep=1mm,font=\tiny] at (partsmo.south|-auxtabsSs.west) {alternative\\materializations};
\draw[thin,gray,shorten >=1mm,shorten <=-3mm] (remark)--(phytbox);
\draw[thin,gray,shorten >=1mm,shorten <=-3mm] (remark)--(phyrsbox);
\end{tikzpicture}\\[-2mm]
\caption{Mapping functions of single \sql{SPLIT\xspace} SMO.}
\label{fig:smomapping}
\vspace{-1mm}
\end{figure}
Let's now consider the \getX{src} mapping function for reconstructing $T$ while the target side is still considered to be materialized.
Reconstructing $T$ from the target side is essentially a union of $R$, $S$, and $T'$.
Nevertheless, $c_R$ and $c_S$ are not necessarily disjoint.
One source tuple may occur as two equal but independent instances in $R$ and $S$.
We call such two instances \emph{twins}.
Twins can be updated independently resulting in \emph{separated twins}, i.e.\ two tuples---one in $R$ and one in $S$---with equal key \ensuremath{p}\xspace but different value for the other attributes.
To resolve this ambiguity and make the SMO bidirectional, we consider the first twin in $R$ to be the primus inter pares and define \getX{src} of \sql{SPLIT\xspace} to propagate back all tuples in $R$ as well as those tuples in $S$ not contained in $R$:
\begin{align}
T(\ensuremath{p}\xspace,A)&\leftarrow R(\ensuremath{p}\xspace,A)\label{rule:tFromR}\\
T(\ensuremath{p}\xspace,A)&\leftarrow S(\ensuremath{p}\xspace,A), \neg R(\ensuremath{p}\xspace,\_)\label{rule:tFromS}\\
T(\ensuremath{p}\xspace,A)&\leftarrow T'(\ensuremath{p}\xspace,A)\label{rule:tFromTprime}
\end{align}
The Rules~\ref{rule:rSimple}--\ref{rule:tFromTprime} define sufficient semantics for \sql{SPLIT\xspace} as long as the target side is materialized.
Let's now assume the SMO instance is virtualized, i.e.\ data is stored on the source side, and let's keep considering the \getX{src} mapping function.
Again, $R$ and $S$ can contain separated twins---unequal tuples with equal key \ensuremath{p}\xspace.
According to Rule~\ref{rule:tFromS}, $T$ stores only the separated twin from $R$.
To avoid losing the other twin in $S$, it is stored in the auxiliary table $S^+$:
\begin{align}
S^+(\ensuremath{p}\xspace,A)&\leftarrow S(\ensuremath{p}\xspace,A), R(\ensuremath{p}\xspace,A'),A\neq A'\label{rule:S+FromS}
\end{align}
Accordingly, \getX{tgt} has to reconstruct the separated twin in $S$ from $S^+$ instead of $T$ (concerns Rule~\ref{rule:sSimple}).
Twins can also be deleted independently resulting in a \emph{lost twin}.
Given the data is materialized on the source side, a lost twin would be directly recreated from its other twin via $T$.
To avoid this information gain and keep lost twins lost, \getX{src} keeps the keys of lost twins from $R$ and $S$ in auxiliary tables $R^-$ and $S^-$:
\begin{align}
R^-(\ensuremath{p}\xspace) &\leftarrow S(\ensuremath{p}\xspace,A),\neg R(\ensuremath{p}\xspace,\_),\ensuremath{c_R}\xspace (A)\\
S^-(\ensuremath{p}\xspace) &\leftarrow R(\ensuremath{p}\xspace,A),\neg S(\ensuremath{p}\xspace,\_),\ensuremath{c_S}\xspace (A)
\end{align}
Accordingly, \getX{tgt} has to exclude lost twins stored in $R^-$ from $R$ (concerns Rule~\ref{rule:rSimple}) and those in $S^-$ from $S$ (concerns Rule~\ref{rule:sSimple}).
Twins result from data changes issued to the target schema containing $R$ and $S$ which can also lead to tuples that do not meet the conditions \ensuremath{c_R}\xspace resp. \ensuremath{c_S}\xspace.
In order to ensure that the reconstruction of such tuples is possible from a materialized table $T$, auxiliary tables $R^*$ and $S^*$ are employed for identifying those tuples using their identifiers (concerns Rules~\ref{rule:rSimple} and~\ref{rule:sSimple}).
\begin{align}
S^*(\ensuremath{p}\xspace) &\leftarrow S(\ensuremath{p}\xspace,A),\neg \ensuremath{c_S}\xspace(A)\\
R^*(\ensuremath{p}\xspace) &\leftarrow R(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace(A)
\end{align}
The full rule sets of \getX{tgt} respectively \getX{src} are now bidirectional and
defined as follows:
\begin{align}
\mathbf{\getX{tgt}:}&\nonumber\\
R(\ensuremath{p}\xspace,A)&\leftarrow T(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A), \neg R^-(\ensuremath{p}\xspace)\label{rule:getrg:begin}\\
R(\ensuremath{p}\xspace,A)&\leftarrow T(\ensuremath{p}\xspace,A), R^*(\ensuremath{p}\xspace)\label{rule:getrg:rs}\\
S(\ensuremath{p}\xspace,A)&\leftarrow T(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A), \neg S^-(\ensuremath{p}\xspace), \neg S^+(\ensuremath{p}\xspace,\_)\\
S(\ensuremath{p}\xspace,A)&\leftarrow S^+(\ensuremath{p}\xspace,A)\\
S(\ensuremath{p}\xspace,A)&\leftarrow T(\ensuremath{p}\xspace,A), S^*(\ensuremath{p}\xspace), \neg S^+(\ensuremath{p}\xspace,\_)\label{rule:getrg:ss}\\
T'(\ensuremath{p}\xspace,A)&\leftarrow T(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A), \neg R^*(\ensuremath{p}\xspace), \neg S^*(\ensuremath{p}\xspace)\label{rule:getrg:end}\\
\mathbf{\getX{src}:}&\nonumber\\
T(\ensuremath{p}\xspace,A)&\leftarrow R(\ensuremath{p}\xspace,A)\label{rule:tFromRFinal}\\
T(\ensuremath{p}\xspace,A)&\leftarrow S(\ensuremath{p}\xspace,A), \neg R(\ensuremath{p}\xspace,\_)\label{rule:tFromSFinal}\\
T(\ensuremath{p}\xspace,A)&\leftarrow T'(\ensuremath{p}\xspace,A)\label{rule:tFromTprimeFinal}\\
R^-(\ensuremath{p}\xspace) &\leftarrow S(\ensuremath{p}\xspace,A),\neg R(\ensuremath{p}\xspace,\_),\ensuremath{c_R}\xspace (A)\\
R^*(\ensuremath{p}\xspace) &\leftarrow R(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace(A)\\
S^+(\ensuremath{p}\xspace,A)&\leftarrow S(\ensuremath{p}\xspace,A), R(\ensuremath{p}\xspace,A'),A\neq A'\\
S^-(\ensuremath{p}\xspace) &\leftarrow R(\ensuremath{p}\xspace,A),\neg S(\ensuremath{p}\xspace,\_),\ensuremath{c_S}\xspace (A)\\
S^*(\ensuremath{p}\xspace) &\leftarrow S(\ensuremath{p}\xspace,A),\neg \ensuremath{c_S}\xspace(A)\label{rule:getsrc:end}
\end{align}
The semantics of all other \textsc{BiDEL}\xspace SMOs are defined in a similar way, see Appendix~\ref{apx:SMOs}.
This precise definition of \textsc{BiDEL}\xspace's SMOs, is the basis for the formal validation of their bidirectionality.
\newpage
\section{Formal evaluation of B{\normalfont \textbf{I}}DEL's$\;$\xspace\newline bidirectionality}\label{sec:bidirectionality}
\textsc{BiDEL}\xspace's SMOs are bidirectional, because, no matter whether the data is (1) materialized on the source side (SMO is virtualized) or (2) materialized on the target side (SMO is materialized), both sides behave like a full-fledged single-schema database.
To formally evaluate this claim, we consider the two cases (1) and (2) independently.
Let's start with case (1); the data is materialized on the source side.
For a correct target-side propagation, the data \ensuremath{D_{tgt}}\xspace at the target side has to be mapped by \getX{src} to the data tables and auxiliary tables at the source side (write) and mapped back by \getX{tgt} to the data tables on the target side (read) without any loss or gain visible in the data tables at the target side.
Similar conditions have already been defined for symmetric relational lenses~\cite{Hofmann2011}---given data at the target side, storing it at the source side, and mapping it back to target should return the identical data at the target side.
For the second case (2) it is vice versa.
Formally, an SMO has bidirectional semantics if the following holds:
\begin{align}
\ensuremath{D_{tgt}}\xspace&=\getX{tgt}^{data}(\getX{src}(\ensuremath{D_{tgt}}\xspace))\label{eq:dt}\\
\ensuremath{D_{src}}\xspace&=\getX{src}^{data}(\getX{tgt}(\ensuremath{D_{src}}\xspace))\label{eq:ds}
\end{align}
Data tables that are visible to the user need to match these bidirectionality conditions.
As indicated by the index $\gamma^{data}$, we project away potentially created auxiliary tables; however, they are always empty except for SMOs that calculate new values: e.g. adding a column requires to store the calculated values when data is stored at the source side to ensure repeatable reads.
The bidirectionality conditions are shown by applying and simplifying the Datalog rule sets that define the mappings \getX{src} and \getX{tgt}.
We label the original relations to distinguish them from the resulting relation, apply \getX{src} and \getX{tgt} in the order according to Condition~\ref{eq:dt} or~\ref{eq:ds}, and compare the outcome to the original relation.
It has to be identical.
As neither the rules for a single SMO nor the version genealogy have cycles, there is no recursion at all, which simplifies evaluating the combined Datalog rules.
In the following, we introduce some basic notion about Datalog rules as basis for the formal evaluation.
A Datalog rule is a clause of the form $H\leftarrow L_1,\ldots,L_n$ with $n \geq 1$ where $H$ is an atom denoting the rule's head, and $L_1,\ldots,L_n$ are literals, i.e. positive or negative atoms, representing its body.
For a given rule $r$, we use \rhead{r} to denote its head $H$ and \rbody{r} to denote its set of body literals $L_1,\ldots,L_n$.
In the mapping rules defining \getX{src} and \getX{tgt}, every \rhead{r} is of the form $q^r(\ensuremath{p}\xspace,Y)$ where $q^r$ is the derived predicate, \ensuremath{p}\xspace{} is the \textsc{InVerDa}\xspace-managed identifier, and $Y$ is a potentially empty list of variables.
Further, we use \rpred{r} to refer to the predicate symbol of \rhead{r}.
For a set of rules \ruleset{R}, $\ruleset{R}^q$ is defined as $\set{r \mid r\in\ruleset{R} \wedge \rpred{r}=q}$.
For a body literal $L$, we use \rpred{L} to refer to the predicate symbol of $L$ and \rvars{L} to denote the set of variables occurring in $L$.
In the mapping rules, every literal $L\in\rbody{r}$ is of the form either $q^r_i(\ensuremath{p}\xspace,Y^r_i,X^r_i)$ or $c^r(Y^r_i,X^r_i)$, where $Y^r_i\subset Y$ are the variables occurring in $L$ and \rhead{r} and $X^r_i$ are the variables occurring in $L$ but not in \rhead{r}.
Generally, we use capital letters to denote multiple variables.
For a set of literals $\mathcal{K}$, \rvars{\mathcal{K}} denotes $\bigcup_{L \in \mathcal{K}}\rvars{L}$.
The following lemmas are used for simplifying a given rule set \ruleset{R} into a rule set \ruleset{R'} such that \ruleset{R'} derives the same facts as \ruleset{R}.
\begin{lemma}[Deduction]\label{lemma:deducation}
Let $L\equiv q^r(\ensuremath{p}\xspace,Y)$ be a literal in the body of a rule $r$.
For a rule $s\in\ruleset{R}^{\rpred{L}}$ let $\renm{s,L}$ be rule $s$ with all variables occurring in the head of $s$ at positions of $Y$ variables in $L$ be renamed to match the corresponding $Y$ variable and all other variables be renamed to anything not in \rvars{\rbody{r}}.
If $L$ is
\begin{enumerate}[leftmargin=*]
\item a positive literal, $s$ can be applied to $r$ to get rule set $r(s)$ of the form $\set{\rhead{r} \leftarrow \rbody{r}\setminus\set{L} \cup \rbody{\renm{s,L}}}$.
\item a negative literal, $s$ can be applied to $r$ to get rule set $r(s)=$ $\set{\rhead{r} \leftarrow \rbody{r}\setminus\set{L} \cup t(K) \mid K \in \rbody{\renm{s,L}})}$\\ with either $t(K)=\set{\neg q^s_i(\ensuremath{p}\xspace,Y^s_i,\_)}$ if $K\equiv q^s_i(\ensuremath{p}\xspace,Y^s_i,X^s_i)$ or
$t(K)=\{q^s_j(\ensuremath{p}\xspace,Y^s_j,X^s_j) \mid q^s_j(\ensuremath{p}\xspace,Y^s_i,X^s_j)\in\rbody{\renm{s,L}} \wedge X^s_j\cap X^s_i\neq\emptyset\}\cup\set{c^r(Y^s_i,X^s_i)}$ if $K\equiv c^r(Y^s_i,X^s_i)$.\footnote{Correctness can be shown with help of first order logic.}
\end{enumerate}
For a given $p$, let $r$ be every rule in \ruleset{R} having a literal $L\equiv p(X,Y)$ in its body.
Accordingly, \ruleset{R} can be simplified by replacing all rules $r$ and all $s\in\ruleset{R}^p$ with all $r(s)$ applications to $\ruleset{R} \setminus (\set{r}\cup\ruleset{R}^p) \cup (\bigcup_{s\in\ruleset{R}^{\rpred{L}}}r(s))$.
\end{lemma}
\begin{lemma}[Empty Predicate]\label{lemma:empty}
Let $r\in \ruleset{R}$ be a rule, $L$ be a literal in the body $L\in\rbody{r}$ and the relation \rpred{L} is known to be empty.
If $L$ is a positive literal, $r$ can be removed from \ruleset{R}.
If $L$ is a negative literal, $r$ can be simplified to $\rhead{r}\leftarrow\rbody{r}\setminus\set{L}$.
\end{lemma}
\begin{lemma}[Tautology]\label{lemma:tautology}
Let $r,s\in\ruleset{R}$ be rules and $L$ and $K$ be literals in the bodies of $r$ and $s$, respectively, where $r$ and $s$ are identical except for $L$ and $K$, i.e. $\rhead{r}=\rhead{s}$ and $\rbody{r}\setminus\set{L}=\rbody{s}\setminus\set{K}$, or can be renamed to be so.
If $K\equiv\neg L$, $r$ can be simplified to $\rhead{r}\leftarrow\rbody{r}\setminus\set{L}$ and $s$ can be removed from \ruleset{R}.
\end{lemma}
\begin{lemma}[Contradiction]\label{lemma:unsatisfiability}
Let $r\in \ruleset{R}$ be a rule and $L$ and $K$ be literals in its body $L,K\in\rbody{r}$.
If $K\equiv\neg L$, $r$ can be removed from \ruleset{R}.
\end{lemma}
\begin{lemma}[Unique Key]\label{lemma:p}
Let $r\in \ruleset{R}$ be a rule and $q(\ensuremath{p}\xspace,X)$ and $q(\ensuremath{p}\xspace,Y)$ be literals in its body.
Since, by definition, \ensuremath{p}\xspace is a unique identifier, $r$ can be modified to $\rhead{r}\leftarrow\rbody{r}\cup\set{X=Y}$.
\end{lemma}
In this paper, we use these lemmas to show bidirectionality for the materialized \sql{SPLIT\xspace} SMO in detail.
Hence, Equation~\ref{eq:ds} needs to be satisfied.
Writing data $T_D$ from source to target-side results in the mapping $\getX{tgt}(\ensuremath{T_{D}}\xspace)$.
With target-side materialization all source-side auxiliary tables are empty.
Thus, $\getX{tgt}(\ensuremath{T_{D}}\xspace)$ can be simplified with Lemma~\ref{lemma:empty}:
\begin{align}
R(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A)\\
S(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A)\\
T'(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A)
\end{align}
Reading the source-side data back from $R$, $S$, and $T'$ to $T$ adds the rule set \getX{src} (Rule~\ref{rule:tFromRFinal}--\ref{rule:getsrc:end}) to the mapping.
Using Lemma~\ref{lemma:deducation}, the mapping $\getX{src}(\getX{tgt}(\ensuremath{T_{D}}\xspace))$ simplifies to:
\begin{align}
T(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace(A) \label{align:pr1}\\
T(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace(A),\neg\ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A) \label{align:partition_Source_T_empty}\\
T(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace(A),\neg\ensuremath{c_R}\xspace(A) \label{align:ps}\\
T(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\neg\ensuremath{c_S}\xspace(A),\neg\ensuremath{c_R}\xspace(A) \label{align:ns}\\
R^-(\ensuremath{p}\xspace) \leftarrow& \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace(A),\neg\ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace(A)\\
R^-(\ensuremath{p}\xspace) \leftarrow& \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace(A),\neg\ensuremath{c_R}\xspace (A),\ensuremath{c_R}\xspace (A) \label{align:rminus}\\
R^*(\ensuremath{p}\xspace) \leftarrow& \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace(A),\neg\ensuremath{c_R}\xspace(A) \label{align:rstar}
\end{align}
\begin{align}
S^+(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace(A),\ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A'),\ensuremath{c_R}\xspace(A'), A\!\neq\!A'\label{align:splus}\\
S^-(\ensuremath{p}\xspace) \leftarrow& \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace(A),\neg\ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace(A) \label{align:sminus}\\
S^-(\ensuremath{p}\xspace) \leftarrow& \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace(A),\neg\ensuremath{c_S}\xspace (A),\ensuremath{c_S}\xspace (A)\\
S^*(\ensuremath{p}\xspace) \leftarrow& \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace(A),\neg\ensuremath{c_S}\xspace(A)\label{align:sstar}
\end{align}
With Lemma~\ref{lemma:unsatisfiability}, we omit Rule~\ref{align:partition_Source_T_empty} as it contains a contradiction.
With Lemma~\ref{lemma:tautology}, we reduce Rules~\ref{align:ps} and~\ref{align:ns} to Rule~\ref{align:nr} by removing the literal $c_S(A)$.
The resulting rules for $T$
\begin{align}
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A)\label{align:pr2}\\
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace (A) \label{align:nr}
\end{align}
can be simplified again with Lemma~\ref{lemma:tautology} to
\begin{align}
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A)\quad. \label{partition_Source_T_logic_final}
\end{align}
For Rule~\ref{align:splus}, Lemma~\ref{lemma:p} implies $A=A'$, so this rule can be removed based on Lemma~\ref{lemma:unsatisfiability}.
Likewise, the Rules \ref{align:rminus}--\ref{align:sstar} have contradicting literals on \ensuremath{T_{D}}\xspace, \ensuremath{c_R}\xspace, and \ensuremath{c_S}\xspace respectively so that Lemma~\ref{lemma:unsatisfiability} applies here as well.
The result clearly shows that data $\ensuremath{T_{D}}\xspace$ in \ensuremath{D_{src}}\xspace is mapped by $\getX{src}(\getX{tgt}(\ensuremath{D_{src}}\xspace))$ to the target side and back to \ensuremath{D_{src}}\xspace without any information loss or gain:
\begin{align}
\mathbf{\getX{src}(\getX{tgt}(\ensuremath{D_{src}}\xspace)):}\;\
&T(\ensuremath{p}\xspace,A)\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A) \qed
\end{align}
So, $\ensuremath{D_{src}}\xspace=\getX{src}(\getX{tgt}(\ensuremath{D_{src}}\xspace))$ holds.
Remember that the auxiliary tables only exist on the materialized side of the SMO (target in this case).
Hence, it is correct that there are no rules left producing data for the source-side auxiliary.
The same can be done for Equation~\ref{eq:dt} as well (Appendix~A).
As expected, the simplification of $\getX{tgt}(\getX{src}(\ensuremath{D_{tgt}}\xspace))$ results in
\begin{align}
\mathbf{\getX{tgt}(\getX{src}(\ensuremath{D_{tgt}}\xspace)):}\;\
&R(\ensuremath{p}\xspace,A)\leftarrow \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)\\
&S(\ensuremath{p}\xspace,A)\leftarrow \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)
\end{align}
This formal evaluation works for the remaining \textsc{BiDEL}\xspace SMOs, as well (Appendix~\ref{apx:SMOs}).
\textsc{BiDEL}\xspace's SMOs ensure that given data at any schema version $V_{n}$ that is propagated and stored at a direct predecessor $V_{n-1}$ or direct successor schema version $V_{n+1}$ can always be read completely and correctly in $V_n$.
To our best knowledge, we are the first to design a set of powerful SMOs and validate their bidirectionality according to the criteria of symmetric relational lenses.
\textbf{Write operations: }
Bidirectionality also holds after write operations: When updating a not-materialized schema version, this update is propagated to the materialized schema in a way that it is correctly reflected when reading the updated data again.
Given a materialized SMO, we apply a write operation $\Delta_{src}(\ensuremath{D_{src}}\xspace)$ to given data on the source side.
$\Delta_{src}(\ensuremath{D_{src}}\xspace)$ can both insert and update and delete data.
Initially, we store \ensuremath{D_{src}}\xspace at the target side using $\ensuremath{D_{tgt}}\xspace=\getX{tgt}(\ensuremath{D_{src}}\xspace)$.
To write at the source side, we have to temporarily map back the data to the source with $\getX{src}^{data}(\ensuremath{D_{tgt}}\xspace)$, apply the write $\Delta_{src}$, and map the updated data back to target with $\ensuremath{D_{tgt}}\xspace'=\getX{tgt}(\Delta_{src}(\getX{src}^{data}(\ensuremath{D_{tgt}}\xspace)))$.
Reading the data from the updated target $\getX{src}^{data}(\ensuremath{D_{tgt}}\xspace')$ has to be equal to applying the write operation $\Delta_{src}(\ensuremath{D_{src}}\xspace)$ directly on the source side.
\begin{align}
\Delta_{src}(\ensuremath{D_{src}}\xspace)&=\getX{src}^{data}(\getX{tgt}(\Delta_{src}(\getX{src}^{data}(\getX{tgt}(\ensuremath{D_{src}}\xspace)))))\label{cond:1}
\end{align}
We have already shown that $D=\getX{src}^{data}(\getX{tgt}(D))$ holds for any data $D$ at the target side, so that Equation~\ref{cond:1} reduces to $\Delta_{src}(\ensuremath{D_{src}}\xspace) = \Delta_{src}(\ensuremath{D_{src}}\xspace)$.
Hence writes are correctly propagated through the SMOs.
The same holds vice versa for writing at the target-side of virtualized SMOs:
\begin{align}
\Delta_{tgt}(\ensuremath{D_{tgt}}\xspace)&=\getX{tgt}^{data}(\getX{src}(\Delta_{tgt}(\getX{tgt}^{data}(\getX{src}(\ensuremath{D_{tgt}}\xspace)))))
\end{align}
\textbf{Chains of SMOs: }
Further, the bidirectionality of \textsc{BiDEL}\xspace SMOs also holds for chains of SMOs: $smo_1,\ldots smo_n$, where $\getX{i, src/tgt}$ is the respective mapping of $smo_i$.
Analogous to symmetric relational lenses~\cite{Hofmann2011}, there are no side-effects between multiple \textsc{BiDEL}\xspace SMOs.
So, \textsc{BiDEL}\xspace's bidirectionality is also guaranteed along chains of SMOs:
\begin{align}
\ensuremath{D_{tgt}}\xspace&=\gamma_{n,tgt}^{data}(\ldots\gamma_{1,tgt}(\gamma_{1,src}(\ldots\gamma_{n,src}(\ensuremath{D_{tgt}}\xspace))))\label{eq:dtChain}\\
\ensuremath{D_{src}}\xspace&=\gamma_{1,src}^{data}(\ldots\gamma_{n,src}(\gamma_{n,tgt}(\ldots\gamma_{1,tgt}(\ensuremath{D_{src}}\xspace))))\label{eq:dsChain}
\end{align}
This bidirectionality ensures logical data independence, since any schema version can now be read and written without information loss or gain, no matter where the data is actually stored.
The auxiliary tables keep the otherwise lost information and we have formally validated their feasibility.
With the formal guarantee of bidirectionality---also along chains of SMOs and for write operations---we have laid a solid formal foundation for \textsc{InVerDa}\xspace's delta code generation.
\subsection{Split and Merge}
We already introduced the Datalog rules describing the split SMO in detail.
\begin{align*}
&\sql{\sqlk{SPLIT\xspace TABLE} $T$ \sqlk{INTO} $R$ \sqlk{WITH} \ensuremath{c_R}\xspace, $S$ \sqlk{WITH} \ensuremath{c_S}\xspace}
\end{align*}
We will now show bidirectionality for this SMO using the rule set discussed in Section~\ref{sec:architecture}:
\begin{align}
\mathbf{\getX{tgt}:}\phantom{T(\ensuremath{p}\xspace,A)}&\nonumber\\
R(\ensuremath{p}\xspace,A)&\leftarrow T(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A), \neg R^-(\ensuremath{p}\xspace)\\
S(\ensuremath{p}\xspace,A)&\leftarrow T(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A), \neg S^-(\ensuremath{p}\xspace), \neg S^+_\ensuremath{p}\xspace(\ensuremath{p}\xspace)\\
S(\ensuremath{p}\xspace,A)&\leftarrow S^+(\ensuremath{p}\xspace,A)\\
T'(\ensuremath{p}\xspace,A)&\leftarrow T(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A)\\
\mathbf{\getX{src}:}\phantom{T(\ensuremath{p}\xspace,A)}&\nonumber\\
T(\ensuremath{p}\xspace,A)&\leftarrow R(\ensuremath{p}\xspace,A)\\
T(\ensuremath{p}\xspace,A)&\leftarrow S(\ensuremath{p}\xspace,A), \neg R_\ensuremath{p}\xspace(\ensuremath{p}\xspace)\\
T(\ensuremath{p}\xspace,A)&\leftarrow T'(\ensuremath{p}\xspace,A)\\
S^+(\ensuremath{p}\xspace,A)&\leftarrow S(\ensuremath{p}\xspace,A), R(\ensuremath{p}\xspace,A'),A\neq A'\\
S^-(\ensuremath{p}\xspace) &\leftarrow R(\ensuremath{p}\xspace,A),\neg S_\ensuremath{p}\xspace(\ensuremath{p}\xspace),\ensuremath{c_S}\xspace (A)\\
R^-(\ensuremath{p}\xspace) &\leftarrow S(\ensuremath{p}\xspace,A),\neg R_\ensuremath{p}\xspace(\ensuremath{p}\xspace),\ensuremath{c_R}\xspace (A)
\end{align}
The first step to formally show bidirectionality is to assume data in the source schema table $\ensuremath{T_{D}}\xspace$ and map it to the target schema.
For data stored at the source schema, the \auxtabs $S^+$, $S^-$, and $R^-$ are not necessary, hence we assume them to be empty.
Using Lemma~\ref{lemma:empty}, we immediately eliminate the rules having one of these tables as positive literal and reduce the rules having a negative literal on one of them.
The reduced rule set for reading data $T_D$ from \ensuremath{D_{src}}\xspace to \ensuremath{D_{tgt}}\xspace is
\begin{align}
\mathbf{\getX{tgt}(\ensuremath{T_{D}}\xspace):}&\nonumber\\
R(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A)\\
S(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A)\\
T'(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A)
\end{align}
The above rule set describes how to transform the source data according to the target schema version.
To show bidirectionality we transform the data back and show that it survives the round trip completely and correctly without any information loss or gain.
For that purpose we insert the rule bodies of $\getX{tgt}(\ensuremath{T_{D}}\xspace)$ into $\getX{src}$ following Lemma~\ref{lemma:projection}, which result in the following rule set:
\begin{align}
\noalign{$\mathbf{\getX{src}(}\mathbf{\getX{tgt}(\ensuremath{T_{D}}\xspace)):}$}
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A)\\
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A), \neg (\ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A))\label{align:t}\\
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A)\\
S^+(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A), \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A'),\nonumber\\
\phantom{S^+(\ensuremath{p}\xspace,A)}&\phantom{\leftarrow }\ensuremath{c_R}\xspace (A'), A\neq A'\label{align:splus}\\
S^-(\ensuremath{p}\xspace) &\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A),\nonumber\\
\phantom{S^-(\ensuremath{p}\xspace) }&\phantom{\leftarrow }\neg (\ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A)),\ensuremath{c_S}\xspace (A)\label{align:smius}\\
R^-(\ensuremath{p}\xspace) &\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A),\nonumber\\
\phantom{R^-(\ensuremath{p}\xspace) }&\phantom{\leftarrow }\neg (\ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A)),\ensuremath{c_R}\xspace (A)\label{align:rminus}
\end{align}
To finally show the bidirectionality, we reduce these rules.
Let us start with the rules, producing $T(\ensuremath{p}\xspace,A)$.
By applying De Morgan's laws (Lemma~\ref{lemma:morgan}), we resolve the negation in Rule~\ref{align:t} and split it into two new rules, which leads to the following rules $T$:
\begin{align}
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A)\label{align:pr1}\\
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A), \neg \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A) \label{align:partition_Source_T_empty}\\
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A), \neg \ensuremath{c_R}\xspace (A)\label{align:ps}\\
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_S}\xspace (A), \neg \ensuremath{c_R}\xspace (A) \label{align:ns}
\end{align}
According to Lemma~\ref{lemma:unsatisfiability}, Rule~\ref{align:partition_Source_T_empty} will never add data to $T$, since there are two conflicting conditions $\ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A)$ and $\neg \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A)$.
The remaining rules can be reduced by applying Lemma~\ref{lemma:tautology}.
We combine Rule~\ref{align:ps} and Rule~\ref{align:ns} to Rule~\ref{align:nr}, since $c_S(A)$ occurs both as positive and negative literal respectively.
So, we reduced Rules~\ref{align:pr1}-\ref{align:ns} to:
\begin{align}
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A)\label{align:pr2}\\
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace (A) \label{align:nr}
\end{align}
Rules~\ref{align:nr} and~\ref{align:pr2} can be reduced following to same principle.
So, finally we obtain the following for $T$:
\begin{align}
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A) \label{partition_Source_T_logic_final}
\end{align}
Rule~\ref{align:splus} for $S^+$ reduces to $S^+(\ensuremath{p}\xspace,A)\leftarrow \emptyset$ since \ensuremath{p}\xspace uniquely identifies $A$ (Lemma~\ref{lemma:p}) and it is impossible that there are different values for $A$ for the same \ensuremath{p}\xspace:
\begin{align}
S^+(\ensuremath{p}\xspace,A)&\leftarrow \emptyset
\end{align}
The Rules~\ref{align:smius} and~\ref{align:rminus} for $S^-$ and $R^-$ can be replaced using De Morgan's laws (Lemma~\ref{lemma:morgan}) to obtain the following:
\begin{align}
S^-(\ensuremath{p}\xspace) &\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A),\neg \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A)\\
S^-(\ensuremath{p}\xspace) &\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A),\ensuremath{c_S}\xspace (A)\\
\nonumber\\
R^-(\ensuremath{p}\xspace) &\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A),\neg \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A)\\
R^-(\ensuremath{p}\xspace) &\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A),\neg \ensuremath{c_R}\xspace (A),\ensuremath{c_R}\xspace (A)
\end{align}
In each rule, there are contradicting literals on \ensuremath{T_{D}}\xspace, \ensuremath{c_R}\xspace, and \ensuremath{c_S}\xspace respectively.
Lemma~\ref{lemma:unsatisfiability} reduces these rules to the final rule set for $\getX{src}(\getX{tgt}(\ensuremath{T_{D}}\xspace))$:
\begin{align}
T(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,A)\\
S^+(\ensuremath{p}\xspace,A)&\leftarrow \emptyset\\
S^-(\ensuremath{p}\xspace) &\leftarrow \emptyset\\
R^-(\ensuremath{p}\xspace) &\leftarrow \emptyset \qed
\end{align}
This formally shows that data from the source schema can be stored at the target schema and translated back without any information loss or gain.
The same condition holds for the other direction, which we do not deduce in detail here, but directly present the result:
\begin{align}
R(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)\\
S(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)\\
T'(\ensuremath{p}\xspace) &\leftarrow \emptyset
\end{align}
So, the data is never lost independently in which version it is written, stored, or read.
The merge operation is just the opposite of split, hence we would exchange \getX{tgt} and \getX{src} to obtain its mapping rules resulting in essentially the same argument.
In our technical report~\cite{Herrmann2016}, we used the same method to show bidirectionality for all presented \textsc{BiDEL}\xspace operations.
This is a solid formal foundation for the forward and backward propagation of read and write accesses among multiple co-existing schema versions independently of the actual materialization.
Starting with established DELs, we defined \textsc{BiDEL}\xspace and formally showed that all \textsc{BiDEL}\xspace operations are bidirectional.
So, given data at any schema version $V_n$ which is propagated and stored at another schema version $V_{n+1}$, we can always read the initial data completely and correctly at schema version $V_n$.
\section{Migration Procedure}\label{sec:migration}
The materialization states of all SMO instances in a schema genealogy form the materialization schema.
The materialization schema determines the physical table schema, i.e. which table versions are directly stored in physical storage.
For the \elt{\tasky}\xspace example---Figure~\ref{fig:example}---this entails five different possible materialization schemas $M$, each implying a different physical table schema $P$ as shown in Table~\ref{tab:materializationSchemas}.
\begin{table}
\centerin
\scalebox{0.9}{
\begin{tabular}{ll}
\toprule
$M$ & $P$ \\
\midrule
$\emptyset$ & \set{\taskR{0}} \\
\set{\sql{SPLIT\xspace}} & \set{\taskR{0}} \\
\set{\sql{SPLIT\xspace},\sql{DROP COLUMN}} & \set{\todoR{1}} \\
\set{\sql{DECOMPOSE}} & \set{\taskR{1}, \authorR{0}} \\
\set{\sql{DECOMPOSE},\sql{RENAME}} & \set{\taskR{1}, \authorR{1}} \\
\bottomrule
\end{tabular}
}
\caption{Possible materialization schemas and the corresponding physical table schema in the \elt{\tasky}\xspace example.}
\label{tab:materializationSchemas}
\end{table}
The materialization schema has huge impact on the performance of a given workload.
Workload changes such as increased usage of newer schema versions demand adaptations of the materialization schema.
\textsc{InVerDa}\xspace facilitates such an adaptation with a foolproof migration command.
The migration command allows moving data non-redundantly along the schema genealogy to those table versions where the given workload causes the least overhead for propagating reads and writes.
Initially, all SMOs except of the create table SMOs are virtualized, i.e. only initially created table versions are in the physical table schema.
A new materialization schema is derived from a given valid materialization schema by changing the materialization state of selected SMO instances.
Formally, two conditions must hold for a materialization schema to be valid.
For an SMO $s$, we denote the source table versions as \srctabs{s}.
For each table version $t$ we denote the incoming SMO with \insmo{t} and the set of outgoing SMOs with \outsmo{t}.
A materialization is valid iff:
\begin{align}
&\forall s\!\in\!M\ \forall t\!\in\!\srctabs{s}\ \brk{\insmo{t} \in M}\\
&\forall s\!\in\!M\ \forall t\!\in\!\srctabs{s}\ \nexists o\!\in\!\brk{\outsmo{t}\!\setminus\!\set{s}}\ \brk{o \in M}
\end{align}
}%
The first condition ensures that all source table versions are in the materialization schema.
The second condition ensures that no source table version is already taken by another materialized SMO.
In the migration command, the DBA lists the table versions that should be materialized. For instance:
\sqll{\sqlk{MATERIALIZE} \sqlc{TasKy2.task}, \sqlc{TasKy2.author}; }%
\textsc{InVerDa}\xspace determines with the schema version catalog the corresponding materialization schema and checks, whether it is valid according to the conditions above.
If valid, \textsc{InVerDa}\xspace automatically creates the new physical tables including auxiliary tables in the physical storage, migrates the data, regenerates all necessary delta code, and deletes all old physical tables.
The actual data migration relies on the same SQL generation routines as used for view generation.
From a user perspective, all schema versions still behave the same after a migration.
However, any data access is now propagated to the new physical table schema resulting in a better performance for schema versions that are evolution-wise close to this new physical schema.
This migration is triggered by one single line of code, so adaptation to the current workload becomes a comfortable thing to do for database administrators.
\subsection{Simplicity and Robustness}\label{sec:simplicity}
Most importantly, we show that \textsc{InVerDa}\xspace unburdens developers by rendering the expensive and error-prone task of manually writing delta code unnecessary.
We show this using both the \elt{\tasky}\xspace example and Wikimedia.
\textbf{\elt{\tasky}\xspace:} We implement the evolution from \elt{\tasky}\xspace to \elt{\taskyy}\xspace with handwritten and hand-optimized SQL and compare this code to the equivalent \textsc{BiDEL}\xspace statements.
We manually implemented (1) creating the initial \elt{\tasky}\xspace schema, (2) creating the additional schema version \elt{\taskyy}\xspace with the respective views and triggers, and (3) migrating the physical table schema to \elt{\taskyy}\xspace and adapting all existing delta code.
This handwritten SQL code is much longer and much more complex than achieving the same goal with \textsc{BiDEL}\xspace.
Table~\ref{tab:simplicity} shows the lines of code (LOC) required with SQL and \textsc{BiDEL}\xspace, respectively, as well as the ratio between these values.
As there is no general coding style for SQL, LOC is a rather vague measure.
We also include the number of statements and the number of characters (consecutive white-space characters counted as one) as more objective measures to get a clear picture.
Obviously, creating the initial schema is equally complex for both approaches.
However, evolving to the new schema version \elt{\taskyy}\xspace and migrating the data accordingly requires \num{359} and \num{182} lines of SQL code respectively, while we can express the same with \num{3} and \num{1} lines with \textsc{BiDEL}\xspace.
Moreover, the SQL code is also more complex, as indicated by the average number of character per statement.
While \textsc{BiDEL}\xspace is working exclusively on the visible schema versions, with handwritten SQL developers also have to manage auxiliary tables, triggers, etc.
The automated delta code generation does not only eliminate the error-prone and expensive manual implementation, but it is also reasonably fast.
Creating the initial \elt{\tasky}\xspace took \num{154} ms on our test system.
The evolution to \elt{\taskyy}\xspace, which includes two SMOs, requires \num{230} ms for both the generation and execution of the evolution script.
The same took \SI{177}{\milli\second} for \elt{\dodo}\xspace.
Please note that the complexity of generating and executing evolution scripts depends linearly on the number of SMOs $N$ and the number of untouched table versions $M$.
The complexity is $O(N+M)$, since we generate the delta code for each SMO locally and exclusively work on the neighboring table versions.
This principle protects from additional complexity in longer chains of SMOs.
The same holds for the complexity of executing migration scripts.
It is $O(N)$ since \textsc{InVerDa}\xspace merely moves the data and updates the delta code for the materialized SMOs stepwise.
\begin{table}
\centering
\scalebox{0.7}{
\begin{tabular}{l l}
\toprule
\textbf{SMO}&occurrences \\
\midrule
\sql{CREATE TABLE}&42\\
\sql{DROP TABLE}&10\\
\sql{RENAME TABLE}&1\\
\sql{ADD COLUMN}&95\\
\sql{DROP COLUMN}&21\\
\bottomrule
\end{tabular}
}
\hspace{1cm}
\scalebox{0.7}{
\begin{tabular}{l l}
\toprule
\textbf{SMO}&occurrences \\
\midrule
\sql{RENAME COLUMN}&36\\
\sql{JOIN}&0\\
\sql{DECOMPOSE}&4\\
\sql{MERGE}&2\\
\sql{SPLIT\xspace}&0\\
\bottomrule
\end{tabular}
}\vspace{-1mm}
\caption{Used SMOs in Wikimedia database evolution.}
\label{tab:wikimediaSMOs}
\end{table}
\begin{figure}[b]
\includegraphics[width=\linewidth, page=1, trim=0mm 10mm 0mm 20mm, clip=true]{figures/taskyDetails.pdf}\vspace{-5mm}
\caption{Overhead of generated code.}
\label{fig:taskyOverhead}
\end{figure}
\textbf{Wikimedia:} Even long evolutions can be easily modeled with \textsc{BiDEL}\xspace.
To show this, we implement \num{171} schema versions of the Wikimedia~\cite{CarloA.Curino}, so data that is written in any of these schema versions, is also visible in all \num{170} other schema versions.
\textsc{BiDEL}\xspace proved to be capable of providing the database schema in each version exactly according to the benchmark and migrating the data accordingly.
In Table~\ref{tab:wikimediaSMOs}, we summarize how often each SMO has been used in the 211 SMOs long evolution.
Even though simple SMOs, like adding and removing tables/columns, are clearly dominating---probably due to the restricted database evolution support of current DBMSes---there are more complex evolutions including the other SMOs as well.
Hence, there is a need for more sophisticated database evolution support and \textsc{BiDEL}\xspace shows to be feasible.
\begin{figure}
\includegraphics[width=\linewidth, page=1, trim=0mm 0mm 0mm 20mm, clip=true]{figures/scenario.pdf}\vspace{-5mm}
\caption{Flexible materialization.}
\label{fig:scenario2}
\end{figure}
\begin{figure}
\includegraphics[width=\linewidth, page=1, trim=0mm 0mm 0mm 20mm, clip=true]{figures/scenarioThree.pdf}\vspace{-5mm}
\caption{Flexible materialization.}
\label{fig:scenario3}
\end{figure}
\subsection{Overhead of Generated Delta Code}\label{sec:overhead}
\textsc{InVerDa}\xspace's delta code is generated from Datalog rules and aims at a general and solid solution.
So far, our focus is on the correct propagation of data access on multiple co-existing schema versions.
We expect the database optimizer to find a fast execution plan, however, there will be an overhead of \textsc{InVerDa}\xspace compared to hand-optimized SQL.
\textbf{\elt{\tasky}\xspace:} In Figure~\ref{fig:taskyOverhead}, we use the previously presented \elt{\tasky}\xspace example with \num{100000} tasks and compare the performance of \textsc{InVerDa}\xspace generated delta code to the handwritten one.
There are two aspects to observe.
First, the hand-optimized delta code causes slightly less (up to \SI{4}{\%}) overhead than the generated one.
Considering the difference in length and complexity of the code (\SI{359}{x} LOC for the evolution), a performance overhead of \SI{4}{\%} in average is more than reasonable for most users.
Second, the materialization significantly influences the actual performance.
Reading the data in the materialized version is up to twice as fast as accessing it from the respective other version in this scenario.
For the write workload (insert new tasks), we observe again a reasonably small overhead compared to handwritten SQL.
Interestingly, the evolved materialization is always faster because the initial materialization requires to manage an additional auxiliary table for the foreign key relationship.
A DBA can optimize the overall performance for a given workload by adapting the materialization, which is a very simple task with \textsc{InVerDa}\xspace.
An advisor tool supporting the optimization task is very well imaginable, but out of scope for this paper.
\begin{figure*}
\begin{subfigure}{.33\textwidth}
\includegraphics[width=\columnwidth, page=1, trim=0mm 11mm 0mm 20mm, clip=true]{figures/tasky.pdf}\vspace{-1mm}
\caption{Materializations for \elt{\tasky}\xspace Mix.}
\label{fig:taskyMix}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\includegraphics[width=\columnwidth, page=1, trim=0mm 11mm 0mm 20mm, clip=true]{figures/taskyRead.pdf}\vspace{-1mm}
\caption{Materializations for \elt{\tasky}\xspace Read.}
\label{fig:taskyRead}
\end{subfigure}
\begin{subfigure}{.32\textwidth}
\includegraphics[width=\columnwidth, page=1, trim=0mm 11mm 0mm 20mm, clip=true]{figures/taskyWrite.pdf}\vspace{-1mm}
\caption{Materializations for \elt{\tasky}\xspace Write.}
\label{fig:taskyWrite}
\end{subfigure}
\vspace{-3mm}
\caption{Different workloads on all possible materialization of \elt{\tasky}\xspace.}
\label{fig:tasky}
\end{figure*}
\subsection{Benefit of Flexible Materialization}\label{sec:independence}
Adapting the physical table schema to the current workload is hard with handwritten SQL, but almost for free with \textsc{InVerDa}\xspace (\num{1} LOC instead of \num{182} in our \elt{\tasky}\xspace example).
Let's assume a development team spares the effort for rewriting delta code and works with a fixed materialization.
\textbf{\elt{\tasky}\xspace:} Again, we use the \elt{\tasky}\xspace example with \num{100000} tasks.
Figure~\ref{fig:scenario2} shows the accumulated propagation overhead for handwritten SQL with the two fixed materializations and for \textsc{InVerDa}\xspace with an adaptive materialization.
Assume, over time the workload changes from \SI{0}{\%} access to \elt{\taskyy}\xspace and \SI{100}{\%} to \elt{\tasky}\xspace to the opposite \SI{100}{\%} and \SI{0}{\%} according to the Technology Adoption Life Cycle.
The adoption is divided into \num{1000} time slices where \num{1000} queries are executed respectively.
The workload mixes \SI{50}{\%} reads, \SI{20}{\%} inserts, \SI{20}{\%} updates, and \SI{10}{\%} deletes.
As soon as the evolved materialization is faster for the current workload mix, we instruct \textsc{InVerDa}\xspace to change the materialization.
As can be seen, \textsc{InVerDa}\xspace facilitates significantly better performance---including migration cost---than a fixed materialization.
This effect increases with the length of the evolution, since \textsc{InVerDa}\xspace can also materialize intermediate stages of the evolution history.
Assume, all users use exclusively the mobile phone app \elt{\dodo}\xspace; but as \elt{\taskyy}\xspace gets released users switch to \elt{\taskyy}\xspace which comes with its own mobile app.
In Figure~\ref{fig:scenario3}, we simulate the accumulated overhead for either materializing one of the three schema versions or for a flexible materialization.
The latter starts at \elt{\dodo}\xspace, moves to \elt{\tasky}\xspace after several users started using \elt{\taskyy}\xspace, and finally moves to \elt{\taskyy}\xspace when the majority of users did so.
Again, \textsc{InVerDa}\xspace's flexible materialization significantly reduces the overhead for data propagation without any interaction of a developer.
The DBA can choose between multiple materialization schemas.
The number of valid materialization schemas greatly depends on the actual structure of the evolution.
The lower bound is a linear sequence of depending SMOs, e.g. one table with $N$ \sql{ADD COLUMN} SMOs has $N$ valid materializations.
The upper bound are $N$ independent SMOs, each evolving another table, with $2^N$ valid materializations.
Specifically, the \elt{\tasky}\xspace example has five valid materializations.
Figure~\ref{fig:tasky} shows the data access performance on the three schema versions for each of the five materialization schema.
The materialization schemas are represented as the lists of SMOs that are materialized.
We use abbreviations for SMOs: e.g. $[D,RC]$ on the very right corresponds to schema version \elt{\taskyy}\xspace since both the decompose SMO (D) and the rename column SMO (RC) are materialized.
The initial materialization is in the middle, while e.g. the materialization according to \elt{\dodo}\xspace is on the very left.
The workload mixes \SI{50}{\%} reads, \SI{20}{\%} inserts, \SI{20}{\%} updates, \SI{10}{\%} deletes in Figure~\ref{fig:taskyMix}, \SI{100}{\%} reads in Figure~\ref{fig:taskyRead}, and \SI{100}{\%} inserts in Figure~\ref{fig:taskyWrite} on the depicted schema versions.
Again, the measurements show that accesses to each schema version are fastest when its respective table versions are materialized, i.e. when the physical table schema fits the accessed schema version.
However, there are differences in the actual overhead, so the globally optimal materialization depends on the workload distribution among the schema version.
E.g. writing to \elt{\taskyy}\xspace is \num{49} times faster when the physical table schema matches \elt{\taskyy}\xspace instead of \elt{\dodo}\xspace.
This gain increases with every SMO, so for longer evolutions with more SMOs it will be even higher.
\begin{figure}[b]
\centering
\includegraphics[width=\columnwidth, page=1, trim=0mm 6mm 0mm 6mm, clip=true]{figures/wikipedia.pdf}
\caption{Optimization potential for Wikimedia.}
\label{fig:wikiPerformance}
\end{figure}
\textbf{Wikmedia:}
The benefits of the flexible materialization originate from the increased performance when accessing data locally without the propagation through SMOs.
We load our Wikimedia with the data of Akan Wiki in schema version \emph{v16524} (109th version) with \num{14359} pages and \num{536283} links.
We measure the read performance for the template queries from~\cite{CarloA.Curino} both in schema version \emph{v04619} (28th version) and \emph{v25635} (171th version).
The chosen materializations match version \emph{v01284} (1st), \emph{v16524} (109th), and \emph{v25635} (171th) respectively.
In Figure~\ref{fig:wikiPerformance}, a great performance difference of up to two orders of magnitude is visible, so there is a huge optimization potential.
We attribute this asymmetry to the dominance of add column SMOs, which need an expensive join with an auxiliary table to propagate data forwards, but only in a cheap projection to propagate backwards.
\begin{figure*}
\includegraphics[width=\textwidth, page=1, trim=0mm 22mm 5mm 25mm, clip=true]{figures/ADD_COLUMN.png}
\caption{Scaling behavior of the ADD COLUMN SMO.}
\label{fig:add_column}
\end{figure*}
\textbf{All possible evolutions with two SMOs:}
To show that it is always possible to gain a better performance by optimizing the materialization, we conduct a micro benchmark on all possible evolutions with two SMOs---except of creating and dropping tables as well as renaming columns and tables, since they have no relevant performance overhead in the first place.
We show that there is always a performance benefit when accessing data locally compared to propagating it through SMOs and we disprove that \textsc{InVerDa}\xspace might add complexity to the data access, so two SMOs do not impact each other negatively.
We generate evolutions with two SMOs and three schema versions: 1st version -- 1st SMO -- 2nd version -- 2nd SMO -- 3rd version.
The second version always contains a table $R(a,b,c)$; the number of generated tuples in this table is the x-axis of the charts.
In Figure~\ref{fig:add_column}, we exemplarily consider all the combinations with add column as 2nd SMO, since this is the most common one.
Again, accessing data locally is up to twice as fast as propagating it through an SMO, so the optimization potential exists in all scenarios.
The average speedup over all SMOs is \SI{2.1}{}.
We calculate the expected performance for the combination of both SMOs as the sum of both query execution times minus reading data locally at the 2nd schema version.
This is reasonable since the data for the 2nd SMO is already in memory after executing the 1st one.
Figure~\ref{fig:add_column} shows that the measured time for propagating the data through two SMOs is always in the same range as the calculated combination of the overhead of the two SMOs individually, so we showed that there is great optimization potential for all combinations of those SMOs and we can safely use it without fearing additional overhead when combining SMOs.
This holds for all pairs of SMOs: on average the measured time differs only \SI{6.3}{\%} from the calculated one.
In sum, \textsc{InVerDa}\xspace enables the DBA to easily adapt the materialization schema to a changing workload and to significantly speed up query processing without hitting other stakeholders' interests.
\section{Bidirectionality of Split}
In this paper (Section~\ref{sec:bidirectionality}), we merely showed one of the two bidirectionality conditions for the \sql{SPLIT} SMO to explain the concept.
As a reminder, the two conditions are:
\begin{align}
\ensuremath{D_{tgt}}\xspace&=\getX{tgt}^{data}(\getX{src}(\ensuremath{D_{tgt}}\xspace))\label{ap:dt}\\
\ensuremath{D_{src}}\xspace&=\getX{src}^{data}(\getX{tgt}(\ensuremath{D_{src}}\xspace))\label{ap:ds}
\end{align}
We have already shown Condition~\ref{ap:ds}, so we do the same for Condition~\ref{ap:dt}, now.
Writing data \ensuremath{R_{D}}\xspace and \ensuremath{S_{D}}\xspace from the target-side to the source-side is done with the mapping $\getX{src}(\ensuremath{R_{D}}\xspace, \ensuremath{S_{D}}\xspace)$.
With source-side materialization all target-side auxiliary tables are not required, so we apply Lemma~\ref{lemma:empty} to obtain:
\begin{align}
\mathbf{\getX{src}(\ensuremath{R_{D}}\xspace, \ensuremath{S_{D}}\xspace):}&\nonumber\\
T(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)\\
T(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
R^-(\ensuremath{p}\xspace) \leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\ensuremath{c_R}\xspace (A)\\
R^*(\ensuremath{p}\xspace) \leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace(A)\\
S^+(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A'),A\neq A'\\
S^-(\ensuremath{p}\xspace) \leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_),\ensuremath{c_S}\xspace (A)\\
S^*(\ensuremath{p}\xspace) \leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_S}\xspace(A)
\end{align}
Reading the target-side data back from the source-side adds the rule set \getX{tgt}(Rule~\ref{rule:getrg:begin}--\ref{rule:getrg:end}) to the mapping.
Using Lemma~\ref{lemma:deducation}, the mapping $\getX{tgt}(\getX{src}(\ensuremath{T_{D}}\xspace))$ extends to:
\begin{align}
\mathbf{\getX{tgt}(\getX{src}}&\mathbf{(\ensuremath{R_{D}}\xspace, \ensuremath{S_{D}}\xspace)):}\nonumber\\
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A), \neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A'), \neg \ensuremath{c_R}\xspace (A')\\
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace(A)\\
\nonumber\\[-3mm]
R(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)}, \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\ensuremath{c_R}\xspace (A), \mathbf{\neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)}\\
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)},\ensuremath{c_R}\xspace (A), \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)}\\
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\mathbf{\ensuremath{c_R}\xspace (A)}, \mathbf{\neg \ensuremath{c_R}\xspace (A)}\\
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)}, \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)},\neg \ensuremath{c_R}\xspace(A)\\
\nonumber\\
S(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)},\ensuremath{c_S}\xspace (A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)}, \neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
S(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)},\ensuremath{c_S}\xspace (A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)}, \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
S(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)},\ensuremath{c_S}\xspace (A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)}, \nonumber\\& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A'), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),A'=A\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A), \mathbf{\ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)}, \mathbf{\neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)},\ensuremath{c_S}\xspace (A), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_), \nonumber\\& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A'), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),A'=A\label{ap:so}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\mathbf{\ensuremath{c_S}\xspace (A)}, \mathbf{\neg \ensuremath{c_S}\xspace(A)}, \neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\mathbf{\ensuremath{c_S}\xspace (A)}, \mathbf{\neg \ensuremath{c_S}\xspace(A)}, \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\mathbf{\ensuremath{c_S}\xspace (A)}, \mathbf{\neg \ensuremath{c_S}\xspace(A)}, \nonumber\\& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A'), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),A'=A
\end{align}
\begin{align}
S(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)}, \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\ensuremath{c_S}\xspace (A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_), \nonumber\\& \mathbf{\neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\ensuremath{c_S}\xspace (A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_), \nonumber\\& \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)},\ensuremath{c_S}\xspace (A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_), \nonumber\\& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A'')},A=A''\\
S(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)}, \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\ensuremath{c_S}\xspace (A), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_), \nonumber\\& \mathbf{\neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\ensuremath{c_S}\xspace (A), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_), \nonumber\\& \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)},\ensuremath{c_S}\xspace (A), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_), \nonumber\\& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A'')},A=A''\\
S(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)}, \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\ensuremath{c_S}\xspace (A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A'),\nonumber\\& \ensuremath{c_S}\xspace (A'), \mathbf{\neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)},\ensuremath{c_S}\xspace (A), \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A')},\nonumber\\& \ensuremath{c_S}\xspace (A'), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)},\ensuremath{c_S}\xspace (A), \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A')},\nonumber\\& \ensuremath{c_S}\xspace (A'), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A''),A=A''\\
\nonumber\\[-3mm]
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A'),A\neq A'\\
\nonumber\\[-3mm]
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)},\neg \ensuremath{c_S}\xspace(A), \mathbf{\neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)}, \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_S}\xspace(A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_S}\xspace(A), \nonumber\\& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),A=A\\
S(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)}, \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_S}\xspace(A), \nonumber\\& \mathbf{\neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_S}\xspace(A), \nonumber\\& \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)}, \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_S}\xspace(A), \nonumber\\& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A')},A=A')
\\
T'(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)},\neg \ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)}, \nonumber\\& \neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
T'(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)},\neg \ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)}, \nonumber\\& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A'), \ensuremath{c_S}\xspace(A')\\
T'(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\mathbf{\neg \ensuremath{c_R}\xspace (A)},\neg \ensuremath{c_S}\xspace (A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A), \nonumber\\& \mathbf{\ensuremath{c_R}\xspace(A)}, \neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
T'(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\mathbf{\neg \ensuremath{c_R}\xspace (A)},\neg \ensuremath{c_S}\xspace (A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A), \nonumber\\& \mathbf{\ensuremath{c_R}\xspace(A)}, \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A''), \ensuremath{c_S}\xspace(A'')\\
\nonumber\\[-3mm]
T'(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)}, \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\neg \ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A), \nonumber\\& \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_), \mathbf{\neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)}\\
T'(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\neg \ensuremath{c_R}\xspace (A),\mathbf{\neg \ensuremath{c_S}\xspace (A)}, \nonumber\\& \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\ensuremath{c_S}\xspace(A)})\\
T'(\ensuremath{p}\xspace,A)\leftarrow& \mathbf{\ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)}, \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\neg \ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A), \nonumber\\& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A'), \ensuremath{c_R}\xspace(A'), \mathbf{\neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)}\\
T'(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \mathbf{\neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)},\neg \ensuremath{c_R}\xspace (A),\neg \ensuremath{c_S}\xspace (A), \nonumber\\& \mathbf{\ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A')}, \ensuremath{c_R}\xspace(A'), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A''), \ensuremath{c_S}\xspace(A'')
\end{align}
Using Lemma~\ref{lemma:unsatisfiability} we remove all rules that have contradicting literals (marked bold).
Particularly, there remains no rule for $T'$ as expected.
Further, we remove duplicate literals within the rules, so we obtain the simplified rule set:
\begin{align}
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A), \neg \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,\_)\label{ap:r1}\\
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A)\label{ap:r2}\\
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_R}\xspace (A), \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A'), \neg \ensuremath{c_R}\xspace (A')\label{ap:r3}\\
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\neg \ensuremath{c_R}\xspace(A)\label{ap:r4}\\
\nonumber\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),\ensuremath{c_S}\xspace (A)\label{ap:s0}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\ensuremath{c_S}\xspace (A)\label{ap:s1}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_),\ensuremath{c_S}\xspace (A)\label{ap:s2}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A'),A\neq A'\label{ap:s3}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{c_S}\xspace(A)\label{ap:s4}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_), \neg \ensuremath{c_S}\xspace(A)\label{ap:s5}
\end{align}
Rule~\ref{ap:s0} is derived from Rule~\ref{ap:so} by applying the equivalence of $A$ and $A'$ to the remaining literals.
Let's now focus on the rules for $R$.
Rules~\ref{ap:r1} and~\ref{ap:r3} are subsumed by Rule~\ref{ap:r2}, since they contain the identical literals as Rule~\ref{ap:r2} plus additional conditions.
Lemma~\ref{lemma:tautology} allows us to further reduce Rules~\ref{ap:r2} and~\ref{ap:r4}, so we achieve that all tuples in $R$ survive one round trip without any information loss or gain:
\begin{align}
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)
\end{align}
We also reduce the rules for $S$.
Rule~\ref{ap:s2} can be removed, since it is equal to Rule~\ref{ap:s1}.
With Lemma~\ref{lemma:tautology}, Rules~\ref{ap:s1} and~\ref{ap:s5} as well as Rules~\ref{ap:s0} and~\ref{ap:s4} can be combined respectively.
This results in the following rules for $S$:
\begin{align}
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)\label{ap:s10}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)\label{ap:s11}\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A'),A\neq A'\label{ap:s12}
\end{align}
Rules~\ref{ap:s10} and~\ref{ap:s12} basically state that the payload data in $R$ ($A$ and $A'$ respectively) is either equal to or different from the payload data in $S$ for the same key $\ensuremath{p}\xspace$.
When we rewrite Rule~\ref{ap:s10} to:
\begin{align}
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A'), A=A'
\end{align}
we can apply Lemma~\ref{lemma:tautology} to obtain the two rules:
\begin{align}
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A), \neg \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,\_)
\end{align}
With the help of Lemma~\ref{lemma:tautology}, we reduce $\getX{tgt}(\getX{src}(\ensuremath{R_{D}}\xspace, \ensuremath{S_{D}}\xspace))$ to
\begin{align}
R(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)\\
S(\ensuremath{p}\xspace,A)\leftarrow& \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A) \qed
\end{align}
So, both Condition~\ref{ap:dt} and Condition~\ref{ap:ds} for the bidirectionallity of the split SMO are formally validated by now.
Since the merge SMO is the inverse of the split SMO and uses the exact same mapping rules vice versa, we also implicitly validated the bidirectionality of the merge SMO.
We can now safely say: wherever we materialize the data, the mapping rules always guarantee that each table version can be accessed just like a regular table---no data will be lost or gained.
This is a strong guarantee and the basis for \textsc{InVerDa}\xspace to provide co-existing schema versions within a single database.
\section{Remaining SMO{\normalsize \textbf{s}}}\label{apx:SMOs}
We introduce the syntax and semantics of the remaining SMOs from Figure~\ref{fig:smos}.
Further, we include the results of the formal evaluation of their bidirectionality.
Please note that creating, dropping, and renaming tables as well as renaming columns exclusively affects the schema version catalog and does not include any kind of data evolution, hence there is no need to define mapping rules for these SMOs.
In Section~\ref{ap:ac}, we introduce the \sql{ADD COLUMN} SMO.
Since SMOs are bidirectional, exchanging the rule sets \getX{src} and \getX{tgt} yields the inverse SMO: \sql{DROP COLUMN}.
There are different extends for the \sql{JOIN} and its inverse \sql{DECOMPOSE} SMO: a join can have inner or outer semantics and it can be done based on the primary key, a foreign key, or on an arbitrary condition.
As summarized in Table~\ref{tab:decomposeAndJoin}, each configuration requires different mapping functions, however some are merely the inverse or variants of others.
The inverse of \sql{DECOMPOSE} is \sql{OUTER JOIN} and joining at a foreign key is merely a specific condition.
\subsection{Add Column / Drop Column}\label{ap:ac}
\noindent\textbf{SMO:} \sql{\sqlk{ADD COLUMN} $b$ \sqlk{AS} \tbl{$f$}{$r_1$,\ldots,$r_n$} \sqlk{INTO} $R$} \\
\textbf{Inverse:} \sql{\sqlk{DROP COLUMN} $b$ \sqlk{FROM} $R$ \sqlk{DEFAULT} \tbl{$f$}{$r_1$,\ldots,$r_n$}}\\
The \sql{ADD COLUMN} SMO adds a new column $b$ to a table $R$ and calculates the new values for $b$ according to the given function $f$.
The inverse \sql{DROP COLUMN} SMO uses the same parameters to ensure bidirectionality.
\begin{align}
\mathbf{\getX{tgt}:}\hspace{3mm} R'(\ensuremath{p}\xspace,A,b)&\leftarrow R(\ensuremath{p}\xspace,A),b=f_B(\ensuremath{p}\xspace,A),\neg B(\ensuremath{p}\xspace,\_)\\
R'(\ensuremath{p}\xspace,A,b)&\leftarrow R(\ensuremath{p}\xspace,A), B(\ensuremath{p}\xspace,b)\\
\mathbf{\getX{src}:}\hspace{6.7mm}R(\ensuremath{p}\xspace,A)&\leftarrow R'(\ensuremath{p}\xspace,A,\_)\\
B(\ensuremath{p}\xspace,b)&\leftarrow R'(\ensuremath{p}\xspace,\_,b)\\[3mm]
\noalign{$\mathbf{\getX{src}(\getX{tgt}(\ensuremath{R_{D}}\xspace)):}$}
R(\ensuremath{p}\xspace,A)&\leftarrow \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A)\\
B(\ensuremath{p}\xspace,b)&\leftarrow \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A),b=f_B(\ensuremath{p}\xspace,A) \qed\label{align:ac}\\
\noalign{$\mathbf{\getX{tgt}(\getX{src}(\ensuremath{R_{D}}\xspace')):}$}
R'(\ensuremath{p}\xspace,A,b)&\leftarrow \ensuremath{R_{D}}\xspace'(\ensuremath{p}\xspace,A,b) \qed
\end{align}
The auxiliary table $B$ stores the values of the new column when the SMO is virtualized to ensure bidirectionality.
With the projection to data tables, the SMO satisfies Conditions~\ref{ap:dt} and~\ref{ap:ds}.
For repeatable reads, the table $B$ is also needed when data is given in the source schema version (Rule~\ref{align:ac}).
\subsection{Decompose on the Primary Key}\label{sec:decomposePK}
\noindent\textbf{SMO:} \sql{\sqlk{DECOMPOSE TABLE} $R$ \sqlk{INTO} \tbl{$S$}{$A$}, \tbl{$T$}{$B$} \sqlk{ON} \sqlk{PK}}\\
\textbf{Inverse:} \sql{\sqlk{OUTER} \sqlk{JOIN TABLE} $S$, $T$ \sqlk{INTO} $R$ \sqlk{ON} \sqlk{PK}}\\
To fill the gaps potentially resulting from the inverse outer join, we use the null value $\omega_R$.
The bidirectionality conditions are satisfied: after one round trip, no data is lost or gained.
\begin{align}
\mathbf{\getX{tgt}:}\hspace{22mm} S(\ensuremath{p}\xspace,A) &\leftarrow R(\ensuremath{p}\xspace,A,\_), A\neq \omega_R\\
T(\ensuremath{p}\xspace,B) &\leftarrow R(\ensuremath{p}\xspace,\_,B), B\neq\omega_R\\
\mathbf{\getX{src}:}\hspace{17.5mm} R(\ensuremath{p}\xspace,A,B)&\leftarrow S(\ensuremath{p}\xspace,A), T(\ensuremath{p}\xspace,B)\\
R(\ensuremath{p}\xspace,A,\omega_R)&\leftarrow S(\ensuremath{p}\xspace,A), \neg T(\ensuremath{p}\xspace,\_)\\
R(\ensuremath{p}\xspace,\omega_R,B)&\leftarrow \neg S(\ensuremath{p}\xspace,\_), T(\ensuremath{p}\xspace,B)\\[3mm]
\mathbf{\getX{src}(\getX{tgt}(\ensuremath{R_{D}}\xspace)):}\hspace{2mm} R(\ensuremath{p}\xspace,A,B)&\leftarrow \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A,B) \qed\\
\mathbf{\getX{tgt}(\getX{src}(\ensuremath{S_{D}}\xspace, \ensuremath{T_{D}}\xspace)):} S(\ensuremath{p}\xspace,A) &\leftarrow \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)\\
T(\ensuremath{p}\xspace,B) &\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,B) \qed
\end{align}
\begin{table}
\centering\scriptsize
\begin{tabular}{lccc}
\toprule
&\textbf{Decompose}&\textbf{Outer Join}&\textbf{Inner Join}\\
\midrule
\textbf{ON PK}&\emph{\ref{sec:decomposePK}}&Inverse of \ref{sec:decomposePK}&\emph{\ref{sec:joinPK}}\\
\textbf{ON FK}&\emph{\ref{sec:decomposeFK}}&Inverse of \ref{sec:decomposeFK}&Variant of \ref{sec:joinCond}\\
\textbf{ON Cond.}&\emph{\ref{sec:decomposeCond}}&Inverse of \ref{sec:decomposeCond}&\emph{\ref{sec:joinCond}}\\
\bottomrule
\end{tabular}
\caption{Overview of different Decompose and Join SMOs.}
\label{tab:decomposeAndJoin}
\end{table}
\subsection{Decompose on a Foreign Key}\label{sec:decomposeFK}
\noindent\textbf{SMO:} \sql{\sqlk{DECOMPOSE TABLE} $R$ \sqlk{INTO} \tbl{$S$}{$A$}, \tbl{$T$}{$B$} \sqlk{ON} \sqlk{FK} $t$}\\
\textbf{Inverse:} \sql{\sqlk{OUTER} \sqlk{JOIN TABLE} $S$, $T$ \sqlk{INTO} $R$ \sqlk{ON} \sqlk{FK} $t$}\\
A \sql{DECOMPOSE}, which creates a new foreign key, needs to generate new identifiers.
Assume we cut away the addresses from persons stored in one table, we eliminate all duplicates in the new address table, assign a new identifier to each address, and finally add a foreign key column to the new persons table.
On every call, the function $id_T(B)$ returns a new unique identifier for the payload data $B$ in table $T$.
In our implementation, this is merely a regular SQL sequence and the mapping rules ensure that an already generated identifier is reused for the same data.
In order to guarantee proper evaluation of these functions, we enforce a sequential evaluation of the rules by distinguishing between existing and new data.
For a literal $L$, we use the indexes $L_{o}$ (old) and $L_{n}$ (new) to note the difference, however they have now special semantics and are evaluated like any other literal in Datalog.
For a \sql{DECOMPOSE ON FK}, we propose the rule set:
\begin{align}
\noalign{$\mathbf{\getX{tgt}:}$}
T_{n}(t,B)&\leftarrow R(\ensuremath{p}\xspace,\_,B),ID_R(\ensuremath{p}\xspace,t)\\
T_{n}(t,B)&\leftarrow R(\ensuremath{p}\xspace,\_,B),\neg ID_R(\ensuremath{p}\xspace,t),\nonumber\\&\phantom{\leftarrow} \neg T_{o}(\_,B),t=id_T(B)\label{align:dc4}\\
T_{n}(t,B)&\leftarrow R(\_,\_,B),T_{o}(t,B)\label{align:dc3}\\
S(\ensuremath{p}\xspace,A,t) &\leftarrow R(\ensuremath{p}\xspace,A,\_), ID_R(\ensuremath{p}\xspace,t)\\
S(\ensuremath{p}\xspace,A,\omega) &\leftarrow R(\ensuremath{p}\xspace,A,\_), ID_R(\ensuremath{p}\xspace,\omega)\\
S(\ensuremath{p}\xspace,A,t) &\leftarrow R(\ensuremath{p}\xspace,A,B), \neg ID_R(\ensuremath{p}\xspace,\_), T_{n}(t,B)\label{align:dc5}\\
\noalign{$\mathbf{\getX{src}:}$}
R(\ensuremath{p}\xspace,A,B)&\leftarrow S(\ensuremath{p}\xspace,A,t), T(t,B)\\
R(\ensuremath{p}\xspace,A,\omega)&\leftarrow S(\ensuremath{p}\xspace,A,\omega)\\
R(t,\omega,B)&\leftarrow \neg S(\_,\_,t), T(t,B)\\
ID_R(\ensuremath{p}\xspace,t)&\leftarrow S(\ensuremath{p}\xspace,\_,t), T(t,\_)\\
ID_R(\ensuremath{p}\xspace,\omega)&\leftarrow S(\ensuremath{p}\xspace,\_,\omega)\\
ID_R(t,t)&\leftarrow \neg S(\_,\_,t), T(t,\_)\\[3mm]
\noalign{$\mathbf{\getX{src}(\getX{tgt}(\ensuremath{R_{D}}\xspace)):}$}
R(\ensuremath{p}\xspace,A,B)&\leftarrow \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A,B)\\
ID_R(\ensuremath{p}\xspace,t)&\leftarrow \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A,B),t=id_T(B) \qed\\
\noalign{$\mathbf{\getX{tgt}(\getX{src}(\ensuremath{S_{D}}\xspace, \ensuremath{T_{D}}\xspace)):}$}
T(t,B)&\leftarrow \ensuremath{T_{D}}\xspace(t,B)\\
S(\ensuremath{p}\xspace,A,t) &\leftarrow \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A,t) \qed
\end{align}
Projecting the outcomes to the data tables, again satisfies our bidirectionality Conditions~\ref{ap:dt} and~\ref{ap:ds}.
Hence, no matter whether the SMO is virtualized or materialized, both the source and the target side behave like common single-schema databases.
Storing data in $R$ implicitly generates new values to the auxiliary table $ID_R$, which is intuitive: we need to store the assigned identifiers for the target version to ensure repeatable reads on those generated identifiers.
\subsection{Decompose on Condition}\label{sec:decomposeCond}
\noindent\textbf{SMO:} \sql{\sqlk{DECOMPOSE TABLE} $R$ \sqlk{INTO} \tbl{$S$}{$A$}, \tbl{$T$}{$B$} \sqlk{ON} $\ensuremath{c}\xspace(A,B)$}\\
\textbf{Inverse:} \sql{\sqlk{OUTER} \sqlk{JOIN TABLE} $S$, $T$ \sqlk{INTO} $R$ \sqlk{ON} $\ensuremath{c}\xspace(A,B)$}\\
To e.g. normalize a table that holds books and authors ($N:M$), we can either use two subsequent \sql{DECOMPOSE ON FK} to maintain the relationship between books and authors, or---if the new evolved version just needs the list of authors and the list of books---we simply split them giving up the relationship.
In the following, we provide rules for the latter case.
Either way we have to generate new identifiers for both the books and the authors.
We use the same identity generating function as in Section~\ref{sec:decomposeFK}.
\begin{align}
\noalign{$\mathbf{\getX{tgt}:}$}
S_{n}(s,A) \leftarrow& R(r,A,\_), ID_{o}(r,s,\_)\\
S_{n}(s,A) \leftarrow& R(r,A,\_), \neg ID_{o}(r,\_,\_),\nonumber\\& A\neq\omega_R, s=id_S(A)\\
S_{n}(r,A) \leftarrow& R(r,A,\_), \neg ID_{o}(r,\_,\_), A=\omega_R\\
T_{n}(t,B)\leftarrow& R(r,\_,B),ID_{o}(r,\_,t)\\
T_{n}(t,B)\leftarrow& R(r,\_,B),\neg ID_{o}(r,\_,\_),\nonumber\\& B\neq\omega_R,t=id_T(B)\\
T_{n}(r,B) \leftarrow& R(r,\_,B),\neg ID_{o}(r,\_,\_), B=\omega_R\\
ID_{n}(r,s,t) \leftarrow& R(r,A,B),S_{n}(s,A),T_{n}(t,B)\\
R^-(s,t)\leftarrow& \neg R(\_,A,B),S_{n}(s,A),\nonumber\\&T_{n}(t,B), \ensuremath{c}\xspace(A,B)\\
\noalign{$\mathbf{\getX{src}:}$}
R_{o}(r,A,B)\leftarrow& S(s,A), T(t,B), ID_{o}(r,s,t)\\
R_{o}(r,A,B)\leftarrow& S(s,A), T(t,B), \ensuremath{c}\xspace(A,B), \neg R^-(s,t), \nonumber\\& \neg ID_{o}(\_,s,t),r=id_R(A,B)\\
ID_{n}(r,s,t)\leftarrow& S(s,A), T(t,B), \ensuremath{c}\xspace(A,B),R_{o}(r,A,B)\\
ID_{n}(r,s,t)\leftarrow& ID_{o}(r,s,t)\\
R_{n}(r,A,B)\leftarrow& R_{o}(r,A,B)\\
R_{n}(s,A,\omega_R)\leftarrow&S(s,A),\neg ID_{n}(\_,s,\_)\\
R_{n}(t,\omega_R,B)\leftarrow&T(t,B),\neg ID_{n}(\_,\_,t)\\[3mm]
\noalign{$\mathbf{\getX{src}(\getX{tgt}(\ensuremath{S_{D}}\xspace, \ensuremath{T_{D}}\xspace)):}$}
R_{n}(r,A,B) \leftarrow& \ensuremath{R_{D}}\xspace(r,A,B)\\
ID_{n}(r,s,t)\leftarrow& \ensuremath{R_{D}}\xspace(r,A,B), \nonumber\\&s=id_S(A),t=id_T(B) \qed\\
\noalign{$\mathbf{\getX{tgt}(\getX{src}(\ensuremath{R_{D}}\xspace)):}$}
S(s,A)\leftarrow& \ensuremath{S_{D}}\xspace(s,A)\\
T(t,B)\leftarrow& \ensuremath{T_{D}}\xspace(t,B)\\
ID(r,s,t) \leftarrow& \ensuremath{S_{D}}\xspace(s,A),\ensuremath{T_{D}}\xspace(t,B), \nonumber\\&\ensuremath{c}\xspace(A,B),r=id_R(A,B)\qed
\end{align}
The bidirectionality Conditions~\ref{ap:dt} and~\ref{ap:ds} are satisfied.
For repeatable reads, the auxiliary table $ID$ stores the generated identifiers independently of the chosen materialization.
\subsection{Inner Join on Primary Key}\label{sec:joinPK}
\noindent\textbf{SMO:} \sql{\sqlk{JOIN TABLE} $R$, $S$ \sqlk{INTO} $T$ \sqlk{ON} \sqlk{PK}}\\
For this join, we merely need one auxiliary table to store those tuples that do not match with a join partner.
Since both bidirectionality conditions hold in the end, we have formally shown the bidirectionality of \sql{JOIN ON PK}.
\begin{align}
\mathbf{\getX{tgt}:}\hspace{18mm} R(\ensuremath{p}\xspace,A,B)&\leftarrow S(\ensuremath{p}\xspace,A), T(\ensuremath{p}\xspace,B)\\
S^+(\ensuremath{p}\xspace,A)&\leftarrow S(\ensuremath{p}\xspace,A), \neg T(\ensuremath{p}\xspace,\_)\\
T^+(\ensuremath{p}\xspace,B)&\leftarrow \neg S(\ensuremath{p}\xspace,\_), T(\ensuremath{p}\xspace,B)\\
\mathbf{\getX{src}:}\hspace{22mm} S(\ensuremath{p}\xspace,A) &\leftarrow R(\ensuremath{p}\xspace,A,\_)\\
S(\ensuremath{p}\xspace,A) &\leftarrow S^+(\ensuremath{p}\xspace,A)\\
T(\ensuremath{p}\xspace,B) &\leftarrow R(\ensuremath{p}\xspace,\_,B)\\
T(\ensuremath{p}\xspace,B) &\leftarrow T^+(\ensuremath{p}\xspace,B)\\[3mm]
\mathbf{\getX{src}(\getX{tgt}(\ensuremath{S_{D}}\xspace, \ensuremath{T_{D}}\xspace)):}S(\ensuremath{p}\xspace,A) &\leftarrow \ensuremath{S_{D}}\xspace(\ensuremath{p}\xspace,A)\\
T(\ensuremath{p}\xspace,B) &\leftarrow \ensuremath{T_{D}}\xspace(\ensuremath{p}\xspace,B) \qed \\
\mathbf{\getX{tgt}(\getX{src}(\ensuremath{R_{D}}\xspace)):}\hspace{1.5mm} R(\ensuremath{p}\xspace,A,B)&\leftarrow \ensuremath{R_{D}}\xspace(\ensuremath{p}\xspace,A,B) \qed
\end{align}
\subsection{Inner Join on Condition}\label{sec:joinCond}
\noindent\textbf{SMO:} \sql{\sqlk{JOIN TABLE} $R$, $S$ \sqlk{INTO} $T$ \sqlk{ON} $\ensuremath{c}\xspace(A,B)$}\\
A join on a condition creates new tuples, so we have to generate new identifiers as well.
We use the notion introduced in Section~\ref{sec:decomposeFK} and satisfy the bidirectionality conditions.
\begin{align}
\noalign{$\mathbf{\getX{tgt}:}$}
R_{n}(r,A,B)\leftarrow& S(s,A), T(t,B), ID_{o}(r,s,t)\\
R_{n}(r,A,B)\leftarrow& S(s,A), T(t,B), \ensuremath{c}\xspace(A,B),\neg R^-(s,t), \nonumber\\& \neg ID_{o}(\_,s,t),r=id_R(A,B)\\
ID_{n}(r,s,t)\leftarrow& S(s,A), T(t,B), \ensuremath{c}\xspace(A,B),R_{n}(r,A,B)\\
ID_{n}(r,s,t)\leftarrow& ID_{o}(r,s,t)\\
S^+(s,A)\leftarrow& S(s,A), \neg ID_{n}(\_,s,\_)\\
T^+(t,B)\leftarrow& T(t,B), \neg ID_{n}(\_,\_,t)\\
\noalign{$\mathbf{\getX{src}:}$}
S_{n}(s,A) \leftarrow& R(r,A,\_), ID(r,s,\_)\\
S_{n}(s,A) \leftarrow& R(r,A,\_), \neg ID(r,s,\_), s=id_S(A)\\
S_{n}(s,A) \leftarrow& S^+(s,A)\\
T_{n}(t,B)\leftarrow& R(r,\_,B),ID(r,\_,t)\\
T_{n}(t,B)\leftarrow& R(r,\_,B),\neg ID(r,\_,t),t=id_T(B)\\
T_{n}(t,B) \leftarrow& T^+(t,B)\\
ID(r,s,t) \leftarrow& R(r,A,B),S_{n}(s,A),T_{n}(t,B)\\
R^-(s,t)\leftarrow& \neg R(\_,A,B),S_{n}(s,A),T_{n}(t,B), \ensuremath{c}\xspace(A,B)\\[3mm]
\noalign{$\mathbf{\getX{src}(\getX{tgt}(\ensuremath{S_{D}}\xspace, \ensuremath{T_{D}}\xspace)):}$}
S(s,A)\leftarrow& \ensuremath{S_{D}}\xspace(s,A)\\
T(t,B)\leftarrow& \ensuremath{T_{D}}\xspace(t,B)\\
ID(r,s,t) \leftarrow& \ensuremath{S_{D}}\xspace(s,A),\ensuremath{T_{D}}\xspace(t,B), \nonumber\\&\ensuremath{c}\xspace(A,B),r=id_R(A,B)\qed\\
\noalign{$\mathbf{\getX{tgt}(\getX{src}(\ensuremath{R_{D}}\xspace)):}$}
R_{n}(r,A,B) \leftarrow& \ensuremath{R_{D}}\xspace(r,A,B)\\
ID_{n}(r,s,t)\leftarrow& \ensuremath{R_{D}}\xspace(r,A,B), \nonumber\\&s=id_S(A),t=id_T(B) \qed
\end{align}
In sum, \textsc{BiDEL}\xspace's SMOs are formally guaranteed to be bidirectional: a solid ground for co-existing schema versions.
\section{Introduction}
\input{chapters/1_introduction}
\section{User Perspective on I{\normalfont \textbf{N}}V{\normalfont \textbf{ER}}D{\normalfont \textbf{A}}\xspace}\label{sec:user}
\input{chapters/2_userperspective}
\section{I{\normalfont \textbf{N}}V{\normalfont \textbf{ER}}D{\normalfont \textbf{A}}\xspace Architecture}\label{sec:architecture}
\input{chapters/3_architecture}
\input{chapters/4_bidel}
\section{Delta Code Generation}\label{sec:generation}
\input{chapters/6_generation}
\section{Evaluation}\label{sec:evaluation}
\input{chapters/7_evaluation}
\section{Related Work}\label{sec:rw}
\input{chapters/8_relatedwork}
\section{Conclusions}\label{sec:conclusion}
\input{chapters/9_conclusion}
\bibliographystyle{abbrv}
\newpage
\begin{minipage}{\columnwidth}
|
\section{Abstract}
\begin{abstract}
The Cooperative Orienteering Problem with Time Windows (COPTW)is a class of problems with some important applications and yet has received relatively little attention. In the COPTW a certain number of team members are required to collect the associated reward from each customer simultaneously and cooperatively. This requirement to have one or more team members simultaneously available at a vertex to collect the reward, poses a challenging OR task. Exact methods are not able to handle large scale instances of the COPTW and no heuristic schemes have been developed for this problem so far. In this paper, a new modification to the classical Clarke and Wright saving heuristic is proposed to handle this problem. A new benchmark set generated by adding the resource requirement attribute to the existing benchmarks. The heuristic algorithm followed by boosting operators achieves optimal solutions for 64.5\% of instances for which the optimal results are known. The proposed solution approach attains an optimality gap of 2.61\% for the same instances and solves benchmarks with realistic size within short computational times.
\end{abstract}
\begin{keyword}
\small{Heuristics, Team Orienteering Problem with Time Windows, Saving algorithms, Vehicle Routing Problem, Travelling salesman}
\end{keyword}
\end{frontmatter}
\section{Introduction}
The \textit{Orienteering Problem} (OP) is a well-known integer programming problem in combinatorial optimisation, introduced by \cite{golden1987orienteering}. OP emerges from a combination between the travelling salesman problem and the knapsack problem. Therefore, it is a routing problem where travelling to all vertices is often not feasible due to limited resources and the available time. The objective of the OP is to find the combination of nodes that maximises the total rewards collected. Tourist trip design (\cite{verbeeck2014extension,vansteenwegen2007mobile,de2015multiobjective}), home fuel delivery (\cite{golden1987orienteering}) and asset protection during wildfire (\cite{van2014mixed})are some examples of application found in the literature.
Various extensions of the OP have been inspired by real-world problems, such as the \textit{ Team Orienteering Problem with Time Window} (TOPTW) (\cite{labadie2012team,souffriau2013multiconstraint,duque2015solving,gunawan2015iterated}), time-dependent orienteering problem (\cite{abbaspour2011time,mei2016efficient,gavalas2014efficient}) and other variants (\cite{salazar2014multi,van2014mixed,varakantham2015direct}). Readers are referred to the survey articles by \cite{vansteenwegen2011orienteering,Gunawan2016315}, for a complete list of applications and extensions.
The Orienteering problem is NP-hard (\cite{golden1987orienteering}). An extensive number of exact and heuristic approaches have been proposed to solve this problem. In the literature, a few papers have focused on exact methods(\cite{keshtkaran2016enhanced,poggi2010team,dang2013branch}) but most papers deal with heuristic approaches(\cite{bouly2010memetic,liang2013multiple,marinakis2015memetic,dang2013effective,gunawan2015iterated,labadie2012team}).
Given a set of nodes in a TOPTW problem a subset of vertices should be visited in order to collect corresponding rewards (\cite{vansteenwegen2011orienteering}). An extension to TOPTW is the \textit {\textbf{ Cooperative}} {\it Orienteering Problem with Time Windows} (COPTW), proposed by \cite{van2015optimisation}. This problem arises when some tasks that need to be undertaken at certain locations can only be accomplished by two or more individuals acting cooperatively. In COPTW each node has a unique resource requirement. Collecting the corresponding reward at any individual vertex is upon the condition of meeting the requirements within its time window. Depending on the requirement of each node, one or more team members must be available there to start the service simultaneously. This study seeks to propose a heuristic approach for solving the COPTW, for the first time.
In all orienteering problems with time windows the routes are very sensitive, either a minor change in service start time at any node or even a single swap in orientation of the nodes requires additional computational effort to double-check all constraints for every single route. This very attribute motivates developing a fast and efficient heuristic to solve the problem. We handled the existing complexity of the COPTW by proposing a new robust enhancement to the \textit{Clarke and Wright} (CW) algorithm followed by boosted operators. The main feature of the heuristic algorithm is its simplicity which facilitates the handling of the problem constraints.
The remainder of the paper is organised as follows: section 2 describes the mathematical formulation of the COPTW problem. In section 3 we introduce the general scheme of the proposed heuristic approach. Our findings from an extensive experimental study are discussed in section 4. Finally, the conclusion and potential future research directions are stated in section 5.
\section{Cooperative Orienteering Problem with Time Windows}
COPTW generalises the TOPTW formulation (see \cite{van2014mixed} and \cite{ van2015optimisation} for a detailed discussion), where a certain number of team members (resources) $R=\{r_1,\dots,r_n\} $ are required at vertices $V=\{v_1,\dots,v_n\} $. The location $v_i$ is considered served if $r_i$ team members arrive at the vertex within the time window $[o_i,c_i]$ and start the service simultaneously at time $s_i$ for a duration of $a_i$ units of time. In COPTW, a homogeneous fleet of $P$ team members start their route from $v_1$ and must return to depot $v_N$ by time $T_{max}$, where both $v_1$ and $v_N$ represent a same location. For any two vertices, $t_{ij}$ indicates the required time for each team member to travel from $i$ to $j$. The binary decision variable $y_i$, takes the value $1$ if $r_i$ team members visit $v_i$ within the appropriate time window for the required duration, $0$ otherwise. The other decision variable is $z_{ij}$ which takes $1$ if arc $ij$ is traversed, otherwise $0$. Furthermore, $x_{ij}$ represents the number of team members travelling from $i$ to $j$. Lastly, $\xi_{ij}$ defines a set of arcs that can be traversed if a team member departs vertex $i$ at $o_i+a_i$ and arrives vertex $j$ by $c_j$. Following the definition of $\xi_{ij}$, feasible arcs that can be traversed to and from node $i$ are shown by $\Omega^{-}_{i}$ and $\Omega^{+}_{i}$, respectively.
The mathematical formulation of the COPTW as a mixed-integer program is as follows:
\begin{equation}
Maximise \sum_{i=2}^{N-1} \sigma_iy_i
\end{equation}
\begin{equation}
s.t.:\sum_{j \in \Omega^{+}_{1}} x_{1j} = \sum_{i \in \Omega^{-}_{N}} x_{iN}=P,
\end{equation}
\begin{equation}
\sum_{i \in \Omega^{-}_{k}} x_{ik} = \sum_{j \in \Omega^{+}_{k}} x_{kj}, \; k=2,\dots,N-1,
\end{equation}
\begin{equation}
r_{k}y_{k} = \sum_{j \in \Omega^{+}_{k}} x_{kj}, \; k=2,\dots,N-1,
\end{equation}
\begin{equation}
x_{ij}\leq P z_{ij}, \;(i,j)\in \xi,
\end{equation}
\begin{equation}
s_i+t_{ij}+a_i-s_j\leq M(1-z_{ij}), \;(i,j)\in \xi,
\end{equation}
\begin{equation}
o_i \leq s_i, \; i=1,\dots,N,
\end{equation}
\begin{equation}
s_i \leq c_i, \; i=1,\dots,N,
\end{equation}
\begin{equation}
x_{ij} \in \{0,1,\dots,P\}, \; (i,j)\in \xi,
\end{equation}
\begin{equation}
y_i,z_{ij}\in \{0,1\}, \; (i,j)\in \xi.
\end{equation}
The objective function (1) maximises the sum of the rewards $\sigma_i$ collected at each vertex $i$. Constraint (2) ensures all members depart from and return to a designated depot. Constraint (3) guarantees flow conservation by enforcing the equality of incoming and outgoing arcs to each node. Constraints (4) and (5) ensure that the collection of a reward (ie a score) at each location is dependent upon the condition of fulfilling its resource requirement, and travelling members of a fleet through an arc never exceeds the total $P$. Constraint (6) ensures that at each vertex the service can only be started when the previously visited location has been served completely and there is sufficient time to travel to the vertex. Constraints (7) and (8) enforce every vertex is visited within its time window. Integer and binary conditions are defined in constraints (9) and (10).
A graphical representation of a sample solution for COPTW is sketched in Fig. 1 as below.
\begin{figure}[h!]
\centering
\includegraphics[width=0.46\textwidth]{FIG01
\caption{A sample solution of the COPTW}
\end{figure}
Two of the main attributes of each node, the resource requirement ($r_i$) and corresponding reward ($\sigma_i$), are specified. Three team members leave the depot and ($r_i$) members must be at each vertex $i$ before starting the service to collect associated the reward. The numbers over each arc represent the travelling members of the team through an arc. It is infeasible to visit all vertices within their time frames. Considering the set of all feasible solutions a schedule is produced that maximises the score (ie the sum of rewards collected).
\section{A heuristic approach for the COPTW}
The requirement in COPTW problems that two or more members of a team are required to be at some of the locations and within specified time windows pose complications observed in neither VRP's nor OP's.
Robust enhancements are applied to the classical Clarke and Wright Saving (CWS) heuristic (\cite{clarke1964scheduling}) followed by modified operators to tackle the challenges introduced by COPTW. The resulting algorithm is named as the \textit{Modified Clarke and Wright} (MCW) heuristic.
While COPTW aims to maximise the collected score by meeting the resource requirements of vertices within their associated time windows, this can be achieved along with construction of the shortest possible routes, at least in most cases. In the proposed saving function the first two terms are inspired by extensive studies in this area. particularly, \cite{doyuran2011robust} introduced a new saving function which not only covers the accomplished improvements by \cite{altinel2005new} but also integrates a sweep algorithm for further enhancements. However, the last term is developed to explain the orienteering problem attribute to maximise the total collected score. The first term of the saving function improves the reshaping ability of the classical CW heuristic and its circumference characteristic. Motivation of the second term is to give early placement to pairs in vicinity of the depot by including $\cos\theta_{ij}$, which is the value of constructed angles between pairs. Therefore, pairs with acute angle climb in the saving pair list while the ones with absolute angles get penalise. The last term is award-based, as pairs with higher score should be at high priorities in the saving pair list, with less impression from other factors. The proposed saving function for the MCW heuristic is therefore as follows:
\begin{equation}
\begin{gathered}
S_{i,j}=\dfrac{d_{i0}+d_{0j}-\lambda d_{ij}}{d^{max}}+\\
\mu\ast\dfrac{\cos \theta_{ij} | d^{max}-(d_{i0}-d_{0j})/2|}{d^{max}}+
\vartheta\ast \dfrac{\Gamma_{i}+\Gamma_{j}}{\overline{\Gamma}}
\hspace{-25pt}
\rule[0em]{0pt}{20pt}
\end{gathered}
\end{equation}
Where $\theta_{ij}$ represents the angle between customer $i$ and $j$ resulted by two rays originating from the depot and crossing them. In the above formulation,$S_{i,j}$, $d$ and $\Gamma$ demonstrate saving value, distance and score, respectively. In equation (11), $d^{max}$ is the maximum distance and $\overline{\Gamma}$ is the average score for each node. Moreover, three parameters namely $\lambda$, $\mu$ and $\vartheta$ are finely tuned to equip the MCW heuristic to perturbation move. To ensure the assignment of nodes with highest saving values prior to others, we implemented parallel route construction in the MCW algorithm which fit the best with underlying nature of the problem.
In the well-known classical CW saving heuristic \cite{clarke1964scheduling}, all vertices are initially connected to the depot. then, saving values of pairs are obtained by joining each of the two nodes in succession. The procedure adds a new vertex at a time to the ends of tours. But, the MCW heuristic is equipped with more than one insertion operator. Covering vertices in the MCW begins by checking the ends of the constructed tours for insertion. In case of failure, the new vertices will be added within tours, as inspired by the cheapest insertion algorithm (see \cite{rosenkrantz1974approximate}) in accordance with the associated saving values. It is worthwhile to mention that none of the constraints should be violated and no removal is allowed before starting the $improve ()$ function. In the other words, the algorithm is equipped with the same logic as exists in the ejection chains (\cite{glover1996ejection}).
\begin{figure}[h!]
\centering
\includegraphics[width=0.46\textwidth]{all
\caption{Comparison of the routes formed by three approaches for COPTW}
\end{figure}
Figure 2 illustrates the effect of taking advantage from the sweep algorithm and the saving algorithm in the MCW heuristic. The depot is denoted as "D" and associated rewards are defined in figure 2(c). The CW heuristic constructs circumferenced routes (figure 2(b)) while the sweep algorithm forms routes based on polar angles with the depot regardless of distance between any pairs(figure 2(c)). The proposed saving function reshapes the constructed routes in figure 2(a) by integrating the cosine value. As can be observed nodes with higher rewards are covered in figure 2(c), which proves the importance of the last term in equation 11. The existing routes in figure 2(c) can be improved after implementation of the $improve ()$ function operators.
As far as the parameters in the saving function are concerned, the computational effort is positively correlated to the intervals for $(\lambda, \mu, \vartheta)$ triplets. Based on extensive tests and considering the compromise between the search effort and solution quality, promising incremental size and intervals for the coefficients are defined. It was decided to change the parameters within $[0,1.4]$ for $(\lambda, \mu)$ pairs and $[0,3.5]$ for $\vartheta$, all with an increment of $0.7$. The large interval and step-size assist the heuristic to explore a broader area of the solution space. The pseudo-code of the presented solution is given in Algorithm 1.
\makeatletter
\def\BState{\State\hskip-\ALG@thistlm}
\makeatother
\begin{algorithm}
\caption{Pseudocode for the MCW heuristic}\label{euclida}
\begin{algorithmic}[1]
\Function{MCW \textit{heuristic} } {temporary routes $\tau$, best routes $\beta$, collected award $\alpha$, set of all nodes $\nu$ , set of assigned nodes $\gamma$, Saving Pair List \textit{SPL}, unvisited vertices $P$, service requirement $r$}
\For {\textbf{all} ($\mu,\, \vartheta,\, \lambda$)}
\BState Call the {$CalcSavingPairs (i,j)$ }function
\For{\textbf{all} ($i\in \textit{SPL}$) }
\BState \small{assignment of feasible $(j \in P)$ to ($\textit{subroutes} \in\tau$)}
\BState Call the { $FeasMatrix (i,j)$} function
\BState \textit{$VisitCount_{j}$} $\gets $ \textit{$ VisitCount_{j}+1$}
\If {(\textit{$ VisitCount_{j}\geq$ $r_{j}$)} }
\BState $\alpha_{current} \gets \alpha_{current} + \Gamma_{j}$, $\gamma \gets j $
\EndIf
\EndFor \State \textbf{end for}
\For{\textbf{all} $(j\in \nu)$ }
\If{$(VisitCout_{j}\leq$ $r_{j})$}
\BState {remove $j$ from \textit{tours} }, update $\tau$
\EndIf
\EndFor \State \textbf{end for}
\If {($\alpha_{current}>\,\alpha_{best}$)}
\BState $\alpha_{best} \gets \alpha_{current}$, $\beta \gets \tau$
\EndIf
\BState Call the {$Improve$} function
\EndFor \State \textbf{end for}
\BState \textbf{Return $\alpha_{best} \,and\, \beta$};
\EndFunction
\end{algorithmic}
\end{algorithm}
At first, a saving pair list is initialised by calling the function $CalcSavingPairs$. after that, unvisited nodes based on the pairs that they belong to and their saving values should be assigned to sufficient number of routes. Vertices with satisfied resource requirements are marked as serviced and relevant information are updated. However, nodes with less number of visits than their resource requirements must be removed, before applying the local search to the solution.
The $ CalcSavingPairs()$ function is represented in Algorithm 2. The saving values of feasible pairs are defined in tuples. The saving function is designed based on the sweep algorithm and CW heuristic (\cite{doyuran2011robust}) where a new additional term is added to give early placement to vertices with higher scores. The coefficient interval for this term is broader compare to the other two as it plays a crucial rule in maximisation of the objective function. Since frequent update in the $feasMatrix$ is required after each insertion or removal, this procedure is described in Algorithm 3.
\makeatletter
\def\BState{\State\hskip-\ALG@thistlm}
\makeatother
\begin{algorithm}
\caption{Pseudocode for the CalcSavingPairs}\label{euclidb}
\begin{algorithmic}[1]
\Function{CalcSavingPairs} {saving function parameters ($\vartheta,\mu,\lambda$), Saving Pair List \textit{SPL}, travel velocity $V$, distance matrix $d_{\nu\times\nu}$}
\BState generate $d_{\nu\times\nu}$
\BState \textbf{Initialise: } $\vartheta,\mu,\lambda$
\For{ ($(i,j)\in \nu$) }
\If {(($ O_{j}+a_{j}+d_{j0}/ V \leq T_{max}$) and ($ O_{i}+a_{i}+d_{ij} / V \leq C_{j}$))}
\BState $S_{i,j} \gets\dfrac{d_{i0}+d_{0j}-\lambda d_{ij}}{d^{max}}+\vartheta\ast \dfrac{\Gamma_{i}+\Gamma_{j}}{\overline{\Gamma}} +\mu\ast\dfrac{\cos \theta_{ij} | d^{max}-(d_{i0}-d_{0j})/2|}{d^{max}}$
\BState insert $S_{i,j}$ to a vector of tuples $(i,j,S_{i,j})$
\EndIf
\EndFor \State \textbf{end for}
\BState \textbf{Return $SPL$};
\EndFunction
\end{algorithmic}
\end{algorithm}
Insertion and removal of each vertex can trigger significant changes in the problem environment as revision of the feasibility matrix may be required. In algorithm 3, the start time of the service at each node is the maximum value of the latest arrival time of team members to a node and its opening time. The generated solution by the MCW heuristic is already promising due to the applied logic in the saving function and the insertion operators; however algorithm 4 is designed to further improve the solution quality.
\makeatletter
\def\BState{\State\hskip-\ALG@thistlm}
\begin{algorithm}
\caption{Pseudocode for the FeasMatrix }\label{euclidc}
\begin{algorithmic}[1]
\Function{FeasMatrix } {service start time $\varphi$, existing nodes in sub-tours $\kappa$, latest arrival time of members to a location $T$, feasibility to travel $ feas$}
\BState \textbf{Require: }{(temporary routes $\tau$ )}
\For{ \textbf{all} ($i\in \nu$) }
\BState $\varphi_{i} \gets O_{i} $
\EndFor \State \textbf{end for}
\For{ \textbf{all} ($\kappa\in \tau$) }
\BState compute $T$
\BState $\varphi_{i} \gets max{ (\varphi_{i}, T)} $
\EndFor \State \textbf{end for}
\For{ \textbf{all} ($\kappa\in \tau$) }
\If{$(\varphi_{\kappa}\ge C_{\kappa}$)}
\BState $feas_{\kappa-1 , \kappa} \gets false $
\Else
\BState $feas_{\kappa-1 , \kappa} \gets true $
\EndIf
\EndFor \State \textbf{end for}
\BState \textbf{Return $FeasMatrix$};
\EndFunction
\end{algorithmic}
\end{algorithm}
\makeatletter
\def\BState{\State\hskip-\ALG@thistlm}
\makeatother
\begin{algorithm}
\caption{Pseudocode for the improve function}\label{euclid}
\begin{algorithmic}[1]
\Function{\textit{improve} } { unvisited vertices $P$ }
\BState \textbf{Require: }{( best collected award $\alpha$, best routes $\beta$)}
\For{\textbf{all} ($i\in P$) }
\BState $ R \, \prime \gets \beta$
\BState Seek to insert i
\BState Call the {$FeasMatrix\, (i,j)$} function
\If {(insertion cause infeasibility of node j)}
\If { ($\Gamma_{j}\le \Gamma_{i}$)}
\BState insert i and remove j
\BState update $\beta$
\BState update $\alpha$
\Else
\BState $ \beta \gets R\prime$
\EndIf
\Else
\BState update $\beta$
\BState update $\alpha$
\EndIf
\BState Call the {$FeasMatrix (i,j)$} function
\EndFor \textbf{end for}
\State \textbf{Return $\beta$ and $\alpha$};
\EndFunction
\end{algorithmic}
\end{algorithm}
The solution then undergoes a local search improvement for further enhancements which is described in Algorithm \ref{euclid}. Saving values of pairs bring on assignment of vertices in an efficient manner which cause the least additional travel distance and thereby saving more time for score collection at as many locations as possible. The $improve ()$ function checks whether any of the inserted vertices can be substituted with the unvisited ones with the possibility of improving the objective function while satisfying all constraints. Finally, it is noteworthy that, while we switch the search space by changing the parameters (perturbation move), the $improve ()$ function acts as a local search algorithm.
\section{Computational results}
Extensive numerical studies were conducted to evaluate the efficacy of the proposed solution approach. A set of benchmark instances was generated by adding the problem-specific attribute to the well-known existing benchmark sets (see \cite{vansteenwegen2009iterated}). The resource requirement attribute was added to each vertex by picking a random number from a set of $\{1,2,3\}$, which indicates how many members of the team are required to collect the associated reward at each node.
In the first study, truncated benchmark sets are designed to solve sufficiently small-size instances by means of both the CPLEX commercial solver and the MCW algorithm. Furthermore, we explored the trade-off between an increased number of available members for service on the one hand and the computation time and objective value on the other hand. We furthermore showed the efficient performance of the proposed heuristic in term of time and accuracy on the large-size designed benchmark instances. All the above computational work was performed on a single node of the National Computational Infrastructure. Each node is equipped with dual 8-core Intel Xeon (Sandy Bridge 2.6 GHz) processors and 32GB of RAM. The algorithm was programmed in C++, using a GCC 5.2.0 compiler. Where applicable, MILP models were solved by the CPLEX 12.6.3 commercial solver in deterministic parallel optimisation mode. All tables show the execution times as elapsed times in seconds.
For the parameter studies, after running 584 instances with various parameter settings, we tuned them in a way to define the best possible trade-off between runtime and solution quality. It was decided to change the parameters within $[0,1.4]$ for $(\lambda, \mu)$ pairs and $[0,3.5]$ for $\vartheta$. Based on the authors' observations the value of $\vartheta$ plays a significant role in the solution quality, thus a broader interval is considered for the newly introduced term in the saving function. Additionally, an incremental size of $0.7$ is large enough to search the feasible region sufficiently and to avoid redundant iterations.
For validation and performance evaluation a collection of 456 small-size benchmark instances were generated and solved by means of both CPLEX and the MCW heuristic . A summary of the tests for $10-12$ nodes with $3$ and $4$ team members on instance sets c100, r100 and rc100 is provided in Table 1. The size of truncated instances are chosen in a way to investigate the correlation between the increase in problem size and exponential growth in computational effort. It is worthwhile to mention that, infeasible edges are excluded in MILP formulations to simplify models for CPLEX.
In Table 1 computation times are reported for all problems in seconds and the optimality gap is defined by "OPT Gap \%". The average gap from optimal solutions in Table 1 is $2.38\%$ which shows the promising performance of the proposed heuristic. Moreover, it can be seen that the computation time increases drastically with minor changes in the problem size for CPLEX, while that of MCW heuristic experiences negligible change.
Table 2 gives the summary of results for $24-26$ vertices with the same number of available team members. One can see that CPLEX solves larger problems from the sets c200, r200 and rc200 compared to those in Table 1. This is due the nature of the studied benchmark groups as the time window intervals are different in length and a larger portion of customers can be covered by the same number of team members. In Table 2, the heuristic average computational time remains below one second for all instances, while it takes hours to solve some sets by CPLEX. In Table 2, the average deviation of MCW heuristic from optimal solutions is just $0.45\%$ which is reasonable for a heuristic solution.
To further verify the reliability of MCW, more truncated instances from (pr01-pr10) and (pr11-pr20) sets were tested and results are demonstrated in Table 3 and 4. The proposed algorithm performs similarly in all examined cases which assures its reliability for further runs on larger problems.
\begin{table*}[]
\centering
\caption{A summary of MCW performance for small-size instances on c100, r100 and rc100 datasets. All computational times are in seconds.}
\begin{tabular}{@{}ccccccccccc@{}}
\toprule
\multirow{2}{*}{Set} & \multirow{2}{*}{\# Vehicles} & \multicolumn{2}{c}{{ 10}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} & \multicolumn{2}{c}{{11}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} & \multicolumn{2}{c}{{ 12}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} \\ \cmidrule(lr){3-4} \cmidrule(lr){6-7} \cmidrule(lr){9-10}
& & CPLEX & MCW & & CPLEX & MCW & & CPLEX & MCW & \\ \midrule
\multirow{2}{*}{c100} & p=3 & 1.54 & 0.01 & 0.00 & 0.90 & 0.01 & 0.00 & 2.82 & 0.10 & 0.00 \\
& p=4 & 1.70 & 0.01 & 0.00 & 0.75 & 0.01 & 0.00 & 1.18 & 0.11 & 0.00 \\
\multirow{2}{*}{r100} & p=3 & 9.23 & 0.17 & 2.67 & 40.07 & 0.24 & 3.89 & 1740.62 & 0.36 & 1.19 \\
& p=4 & 12.88 & 0.13 & 3.22 & 59.92 & 0.19 & 6.74 & 1455.69 & 0.30 & 5.43 \\
\multirow{2}{*}{rc100} & p=3 & 15.97 & 0.12 & 2.94 & 43.38 & 0.17 & 2.44 & 304.96 & 0.24 & 2.20 \\
& p=4 & 8.12 & 0.09 & 1.97 & 503.42 & 0.13 & 3.52 & 1908.95 & 0.19 & 6.77
\\ \midrule
\end{tabular}
\end{table*}
\begin{table*}[]
\centering
\caption{A summary of MCW performance for small-size instances on c200, r200 and rc200 dataset. All computational times are in seconds.}
\label{my-label}
\begin{tabular}{@{}ccccccccccc@{}}
\toprule
\multirow{2}{*}{Set} & \multirow{2}{*}{\# Vehicles} & \multicolumn{2}{c}{{ 24}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} & \multicolumn{2}{c}{{25}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} & \multicolumn{2}{c}{{ 26}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} \\ \cmidrule(lr){3-4} \cmidrule(lr){6-7} \cmidrule(lr){9-10}
& & CPLEX & MCW & & CPLEX & MCW & & CPLEX & MCW & \\ \midrule
\multirow{2}{*}{c200} & p=3 & 3.11 & 0.35 & 0.00 & 4.00 & 0.45 & 0.00 & 5.44 & 0.59 & 0.00 \\
& p=4 & 2.89 & 0.19 & 0.00 & 3.93 & 0.34 & 0.00 & 3.15 & 0.36 & 0.00 \\
\multirow{2}{*}{r200} & p=3 & 11.84 & 0.48 & 0.74 & 12.17 & 0.61 & 0.54 & 19.13 & 0.75 & 0.78 \\
& p=4 & 5.24 & 0.24 & 0.00 & 6.88 & 0.30 & 0.00 & 7.71 & 0.37 & 0.00 \\
\multirow{2}{*}{rc200} & p=3 & 22.75 & 0.38 & 0.77 & 478.37 & 0.52 & 1.20 & 13585.62 & 0.88 & 3.77 \\
& p=4 & 7.78 & 0.15 & 0.00 & 8.31 & 0.25 & 0.00 & 9.83 & 0.37 & 0.45 \\ \midrule
\end{tabular}
\end{table*}
\begin{table*}[]
\centering
\caption{A summary of MCW performance for small-size instances on pr01-10 dataset. All computational times are in seconds.}
\label{my-label}
\begin{tabular}{@{}ccccccccccc@{}}
\toprule
\multirow{2}{*}{Set} & \multirow{2}{*}{\# Vehicles} & \multicolumn{2}{c}{{ 10}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} & \multicolumn{2}{c}{{11}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} & \multicolumn{2}{c}{{ 12}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} \\ \cmidrule(lr){3-4} \cmidrule(lr){6-7} \cmidrule(lr){9-10}
& & CPLEX & MCW & & CPLEX & MCW & & CPLEX & MCW & \\ \midrule
\multirow{2}{*}{pr01-10} & p=3 & 55.36 & 0.79 & 3.74 & 84.27 & 0.86 & 2.83 & 259.63 & 0.97 & 3.43 \\
& p=4 & 64.98 & 0.73 & 3.75 & 254.00 & 0.81 & 3.16 & 854.03 & 0.90 & 4.48 \\ \midrule
\end{tabular}
\end{table*}
\begin{table*}[]
\centering
\caption{A summary of MCW performance for small-size instances on pr11-20 dataset. All computational times are in seconds.}
\label{my-label}
\begin{tabular}{@{}ccccccccccc@{}}
\toprule
\multirow{2}{*}{Set} & \multirow{2}{*}{\# Vehicles} & \multicolumn{2}{c}{{ 19}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} & \multicolumn{2}{c}{{20}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} & \multicolumn{2}{c}{{ 21}} & \multirow{2}{*}{\begin{tabular}[c]{@{}c@{}}Opt\\ Gap \% \end{tabular}} \\ \cmidrule(lr){3-4} \cmidrule(lr){6-7} \cmidrule(lr){9-10}
& & CPLEX & MCW & & CPLEX & MCW & & CPLEX & MCW & \\ \midrule
\multirow{2}{*}{pr11-20} & p=3 & 10.92 & 0.20 & 3.65 & 113.72 & 0.31 & 4.93 & 1497 & 0.43 & 4.97 \\
& p=4 & 19.50 & 0.14 & 1.58 & 131.95 & 0.25 & 4.43 & 6917.79 & 0.33 & 4.60 \\ \midrule
\end{tabular}
\end{table*}
\begin{table}[]
\centering
\caption{Computational results for the large-size sets}
\label{my-label}
\begin{tabular}{@{}cccc@{}}
\toprule
Set & \# Vehicles & 50 & 100 \\ \midrule
\multirow{2}{*}{c100} & p=6 & 21.42 & 138.15 \\
& p=12 & 10.92 & 153.60 \\
\multirow{2}{*}{c200} & p=6 & 3.97 & 131.59 \\
& p=12 & 3.32 & 44.06 \\
\multirow{2}{*}{r100} & p=6 & 26.77 & 148.34 \\
& p=12 & 22.04 & 184.40 \\
\multirow{2}{*}{r200} & p=6 & 10.01 & 190.71 \\
& p=12 & 7.76 & 86.68 \\
\multirow{2}{*}{rc100} & p=6 & 16.25 & 122.84 \\
& p=12 & 13.13 & 154.61 \\
\multirow{2}{*}{rc200} & p=6 & 8.57 & 185.64 \\
& p=12 & 3.80 & 56.74 \\
\multirow{2}{*}{pr01-10} & p=6 & 14.85 & 104.89 \\
& p=12 & 5.13 & 96.72 \\
\multirow{2}{*}{pr11-20} & p=6 & 38.11 & 274.12 \\
& p=12 & 14.21 & 246.51 \\ \midrule
\end{tabular}
\end{table}
An instance where the MCW achieved an optimal solution on a small set is demonstrated in Fig. 3. For the sake of better presentation, nodes are assigned into the cells of arrays. The times for starting the service and the corrsponding time window for each succeeding vertex are provided.
\begin{figure}[h!]
\centering
\includegraphics[width=0.46\textwidth,height=0.15\textheight]{FIG0002
\caption{A sample scheduled tour by MCW}
\end{figure}
Consider for example, three members leave the depot and arrive at customer number 5 within its time window [15,67], where they start the service simultaneously to collect the associated score at 15.1. After that, second and third members leave customer 5 toward vertex number 3. Finally, all the team members finish their tour by returning to depot before $T_{max}=1236$. One can see that the illustrated routes are highly dependent in a sense that any minor change in orientation of nodes at any tour requires reconstruction of other routes.
The proposed solution method was examined on truncated sets and results were discussed. After the superior performance of the heuristic approach analysed, we carried out further tests on larger instances by using the same approach. Finally, in Table 5 we report a summary of tests for 50 and 100 vertices with 6 and 12 team members. More team members are considered in order to cover a substantial portion of available nodes. Among the (pr01-pr20) test instances those with an insufficient number of nodes are excluded in the reported results. As can be observed, our implementation attains an optimality gap of 2.61\% on instances for which the optimal results are known and solves benchmarks with realistic size within short computational times.
\section{Conclusion}
In this paper we investigated the cooperative orienteering problem with time windows, and a new efficient heuristic algorithm is presented to tackle the problem-specific complexities. Although the problem is not newly introduced, we have developed the first heuristic solution to this problem. The proposed algorithm generates optimal solutions for 64.5\% of instances for which the best results are known. In this study, we introduced a new modification to the CW saving heuristic to improve solution quality without losing its simplicity.
To evaluate the solution approach, a new benchmark set was generated for the COPTW problem. This was achieved by adding the resource requirement attribute to the existing benchmarks (\cite{vansteenwegen2009iterated,montemanni2009ant}) for the TOPTW poblem. Then, the small-sized instances were solved by both the CPLEX solver and the MCW heuristic and results analysed. The optimality gaps and computational times demonstrated the efficiency of the algorithm.
The COPTW is an important class of problem that arises naturally in many applications. For example, in disaster relief aid workers transported by bus must be present at the same locations and time windows as a truck carrying relief supplies so that these supplies can be distributed. As the applications grow, further developments will most likely be neded to represent factors such as soft time windows, and real-time changes to route and capacity constraints. Adding new constraints will require more efficient algorithms where our MCW heuristic at the very least will be able to generate high quality and reliable results that can be used as initial solutions.
\section*{Acknowledgement}
{
The second author is supported by the Australian Research Council under the Discovery Projects funding scheme (project DP140104246).
}
\input{38.bbl}
\end{document}
|
\section{Introduction}
Cloud computing is an emerging technology in distributed environments where hardware and software computing resources are virtualized and delivered in the form of services to clients. Cloud computing provides a dynamic computing infrastructure, optimal resources allocation, rapid elastic access to computing architectures, reusable resources, flexibility, on-demand services, pay-per-use model, lower costs, and etc. \cite{pap1,pap2,pap3}. Therefore, due to many advantages of cloud computing, numerous users have been attracted to the use of cloud-based resources and services. The increasing number of users have made the management and allocation of cloud-based resources a challenging task. Resource allocation plays an important role in efficiency of the overall system and the level of consumer satisfaction of the deployed system. Therefore, it must be in such a way that in addition to delivering profit to the providers, consumer satisfaction will also be considered \cite{pap4}. Different approaches and models try to solve the resource allocation problem efficiently. Economics-based approaches such as market-based and auction-based resource allocation mechanisms have been two recent popular techniques to solve the aforementioned problem. In market-based approach, providers determine the resources price based on the number of resources that consumers are asking. While in the auction-based approach both providers and consumers have an impact on the agreed final prices \cite{pap5}.
Auction is an efficient economic system based on bidding where buyer/vendor compete with each other to purchase/sell commodities \cite{pap3,pap6}. In an auction the vendors goal is to gain the most profit in the long run while the buyers expect their tasks to be run at the lowest cost with respect to the expected level of service. The auction system in the cloud environment is comprised of three major components, namely, the cloud provider (CP), the cloud consumer (CC) and the cloud auctioneer (CA). In Fig. \ref{fig1} the responsibility and the relationship between these three components is briefly specified. The auctioneer is responsible for allocating resources and determining the final price that is payable for each participant.
The bidder drop issue is one of the important concerns in auctions that should be dealt with. This problem occurs when participants with powerful and high quality bids have a better chance of winning the auction. As a result, weaker participants will be gradually pushed aside and the strong ones dominate the auction \cite{pap6}. Hence, adding a fairness term to the problem formulation when choosing the winners can help alleviate this issue by enforcing the equality conditions among participants. Even though the impact of such a policy may not be tangible in the short run, it tends to increase the willingness of participants to take part in the auctions in the long run \cite{pap9}.
Auctions in general can be placed in one of the major categories of one-sided or double-sided. In a one-sided auction, multiple buyers compete to purchase commodities of from vendor or multiple vendors compete with each other to sell their commodities to a single buyer. In contrast, in a double-sided auction, multiple buyers compete with multiple vendors to purchase the commodities. A double-sided auction has been shown to be more efficient than holding multiple one-sided auctions \cite{pap7}. Moreover, an auction system can be either single-item or combinatorial. In a single item auction each bid is attributed to only one resource whereas in the combinatorial one, participants can request combinations of resources in their bids to run their tasks on. This type of auction is more efficient than old style single item auction \cite{pap7,pap8}.
In the rest of the paper, first a brief survey of the related works in multi-attribute auction domain is given focusing on approaches that integrate the fairness attribute into their model. In section 2 we introduce our proposed model in a combinatorial double auction which includes a fairness factor. In section 3 we evaluate our model and compare its performance with a similar auction model that disregards the individual fairness. Finally, in section 4 we conclude the paper.
\begin{figure}[t!]
\centerline{
\includegraphics[scale=0.6,trim=8cm 18.5cm 9cm 3cm]{Diagram.pdf}
}
\caption{A schematic view of an auction components and their interactions. In step 1 each consumer and provider send their bid to the auctioneer. After determination of the winners, in step 2 the auctioneer takes the resources from providers and gives them to the related consumers. In step 3 the transaction prices are specified and the consumers pay the payment to the providers.}
\label{fig1}
\end{figure}
\section{Related Works}
There has been many studies that have tried to maximize the overall profits in some formulated auction resource allocation problem in a cloud environment as their main objective. Even though in some of these works, in addition to the price, factors such as delivery time, quality of service and even penalty policies are taken into account, but few of them have considered the fairness attribute.
In \cite{pap9} a fair multi-attribute combinatorial double auction model is introduced in order to increase the consumer satisfaction. Fairness and providers' reputation, are two attributes that are used by the authors of the article. In this model due to the policies that are considered, providers who give incorrect information will be punished and their reputation will be decreased in the next round of auction. Hence, this model seeks increasing user satisfaction and truthfulness of providers. In \cite{pap6} a multi-dimensional fairness mechanism is employed in an auction-based environment. In order to overcome the bidder drop problem, the author introduced three approaches of calculating of the fairness, namely, the quantitative, the qualitative, and the stochastic approaches. In their work a consumer which is in a miserable situation and as a result may leave the auction, is given a higher priority to reduce the possibility of him leaving the auction.
On the other hand, the limited available resources as well as the high demand from users result in loss of fairness in the allocation of network resources. To address this issue, Wu et al. \cite{pap12} proposed an efficient relay resource allocation that considered fairness competition between users. Their method tried to add fairness by using the incomplete private information obtained from other nodes. In another study \cite{pap13}, authors designed an auction mechanism to allocate the spectrum resources with a trade-off between maximization of revenue and fairness. In their algorithm, bidders ranks will be changed based on the critical condition in order to decrease starvation. Finally, Lee et al. \cite{pap14} used a service oriented auction mechanism to deal with the bidder drop problem. In their proposed work, each resource has a reservation price and will be sold to the customer who has suggested a higher price. The resources that remain in the end will be shared among the weakest customers.
\section{Auction Model and Problem Formulation}
In this section, we first describe our model and its fairness mechanism. Then the winner determination problem is formulated as an integer programming (IP) problem.
\subsection{Multi-Dimensional Fairness Combinatorial Double Auction Model (MDFCDA)}
In this article we propose a new combinatorial double auction in cloud environment that considers multi-dimensional fairness, henceforth, a multi-dimensional fairness combinatorial double auction model (MDFCDA). In our model we used multi-dimensional fairness inspired by work report in \cite{pap10} and formulate it in the form of an integer programming framework to solve the WDP and resource allocation problem.
In the proposed model, multiple providers, consumers and different resources (in terms of capacity and the number of CPU, memory size and bandwidth) are considered. Auction takes place in several rounds; in each round the providers and consumers individually deliver their bids to the auctioneer in order to sell and buy the resources. The providers offer the available resources in their bids and the consumers request for the bundle of resources in their bids to run their applications (tasks). The main task of the auctioneer after receiving and collecting the bids from different participants is first, to extend them and then to select the winners and allocate resources to them and determine the transaction price.
There is a repository for storing historical information of participants. In this repository, information such as the number of wins, the number of losses, the overall quality of the participant bids and other factors related to the previous auction round is stored. Hence, the auctioneer extends the bids and makes decision on determining the winners based on the information available in this repository. Moreover, by our assumption, if a resource is assigned to a consumer it will not be released until the end of the auction round and the execution time of the user tasks on the assigned resources ends by then.
The fairness attribute helps reducing the number of bidders that drop and as a result the number of participants who could not compete with more powerful participants and leave the auction will be reduced. A. Pla \cite{pap6}, has introduced a quantitative, a qualitative and a stochastic approach for enforcing the fairness. In a quantitative approach, fairness is calculated based on the number of victories and failures of participants in the previous rounds of auction. This approach encourages the weak participants to participate in the upcoming auction rounds by increasing the chances of their winning. But the main issue with this approach is that each participant can easily increases his chance of winning with a dummy bid. A dummy bid is a bid which has the lowest possible price for the resources to increase the odds of winning the next rounds after losing several rounds. To solve this problem and also to affect the bids attributes to increase or decrease the chance of winning, the qualitative approach is proposed. In contrast, in a quantitative approach the chance of winning for each participant depends on the bid attributes (e.g. the suggested price, QoS, etc). Finally, in a stochastic approach, measurements that are performed based on the previous two approaches are randomly applied.
In our proposed approach, fairness attribute is guaranteed by a fairness factor that takes on a real value. The higher the value, the more the chance of winning will be. This defined fairness factor fuses all the previously mentioned perspectives (i.e. qualitative, quantitative and stochastic approaches) into one single measure. In cases where two participants have similar conditions in terms of the number of wins and losses, the qualitative approach favors the one with a higher bid value. Therefor the fairness factor is calculated by the qualitative and quantitative approaches and then it is tuned according to the stochastic approach.
The following policies are considered in computing the fairness factor:
\begin{itemize}
\item On the one hand, the participants who have more losses are given more chance of winning and on the other hand, participants who have more wins are given less chance of winning.
\item Among the losers of previous auction round, the participant who has the bid with higher quality (such as having higher bid price, lesser service time, etc.) are given more chance of winning. Also among the winners of previous auction round, the participant who has the bid with lower quality are given less chance of winning.
\item The more the participant be in on the verge of a drop (i.e. has lost a higher number of consecutive rounds) the higher the value that the fairness factor will take on.
\end{itemize}
Algorithm 1 shows the procedure to calculate the fairness factor for consumers. In this algorithm $CL_n$ is the last number of consecutive losses for the consumer $n$ (assuming that after each victory, $CL_n$ takes on value 0). $ml$ is the maximum consecutive auction rounds that consumers can loose and if $CL_n>ml$ the bidder drop problem occurs. Based on this algorithm if the consumer has participated in the current round two possible scenarios can happen:
\begin{itemize}
\item If the consumer $n$ lost in the previous auction round, fairness factor will be applied with probability of $Prob\_W$. The higher $CL_n$ becomes the more $Prob\_W(CL_n)$ will be. Hence a consumer who is in a more critical condition to drop the auction has a higher chance of winning. Fairness factor is hence calculated according to the following equation:
\begin{equation}
\label{eq3}
\begin{split}
Fun\_W (\#losses_n,eval\_fun_n,CL_n)=\\
(CL_n+1)\times ((\alpha_1\times \#looses_n)+(\alpha_2\times eval\_fun_n))
\end{split}
\end{equation}
$\#looses_n$ is the total number of looses in the previous auction rounds and also $eval\_fun_n$ is evaluation function to determine the quality of previous auction rounds' bids of consumer $n$ based on the historical information (e.g. the number of requested resources, suggested prices, etc.) and will be calculated through the equations that is introduced in \cite{pap10}. Where $\alpha_1$ and $\alpha_2$ coefficients are factors to strike a balance between the aforementioned two terms.
\item If the consumer $n$ won in the previous auction round, fairness factor will be applied with the probability of $Prob\_L$. The smaller $CL_n$ becomes the higher $Prob\_L(CL_n)$ and hence, the smaller chance of winning for a consumer. For this case, the fairness factor is calculated according to the following equation:
\begin{equation}
\label{eq4}
\begin{split}
Fun\_L(\#wins_n,eval\_fun_n,CL_n)=\\
-1\times\frac{1}{CL_n+1}\times ((\beta_1\times \#wins)+\frac{\beta_2}{eval\_fun_n})
\end{split}
\end{equation}
\end{itemize}
where $\#wins$ is the total number of wins in the previous auction rounds and $\beta_1$ and $\beta_2$ play similar roles to $\alpha_1$ and $\alpha_2$.
\begin{algorithm}
\KwResult{Fairness factor}
\For {consumer$_n \in $ participants\_in\_current\_auction\_round}{
\eIf {consumer$_n$\_failed\_in\_previous\_auction\_round }{
\If {$Random(0,1) < Prob\_W(CL_n)$}{
FairnessFactor$_n$= $Fun\_W(\#looses_n,eval\_func_n,CL_n)$
}
}{
\If {$Random(0,1) < Prob\_L(CL_n)$}{
FairnessFactor$_n$= $Fun\_L(\#wins_n,eval\_func_n,CL_n)$
}
}
}
\caption{Calculation of the fairness factor algorithm}
\label{alg1}
\end{algorithm}
\section{Problem Formulation}
The proposed model consists of three main phases of 1) bidding policies, 2)winning determination problem (WDP) and 3) pricing mechanism. In the bidding policies phase, there is a way each participant should bid according to the predefined rules. In phase of WDP phase, winners are determined with the aim of maximizing the total utility considering fairness factor using integer programming. After determining the winners, in the pricing phase trade prices that consumers should pay to providers are calculated.
\subsection{Bidding Policies}
In the proposed model $N$ consumers, $M$ providers and $L$ resource types are considered. Each of the consumers and the providers request and offer different types of resources for buying and selling, independently from each other. The consumer $n (1\le n \le N)$ requests the following two-tuple bid:
\begin{itemize}
\item $Bid_n^{'} = (Price_n^{'},Quantity_n^{'})$, where $Price_n^{'}=(P_{n_1}^{'},P_{n_2}^{'},\dots,P_{n_L}^{'})$ is the suggested prices for the requested resources by the consumer $n$, $P_{n_l}^{'}$ is the suggested price per unit of resource type $l$ $(1\le l \le L)$ by the consumer $n$, $Quantity_n^{'}=(Q_{n_1}^{'},Q_{n_2}^{'},\dots,Q_{n_L}^{'})$ is the required resources bundle by the consumer $n$ and $Q_{n_l}^{'}$ indicates the required resource type $l$ by the consumer $n$.
\end{itemize}
Also The provider $m$, $1\le m \le M$ offers the following two-tuple bid:
\begin{itemize}
\item $Bid_m = (Price_m, Quantity_m )$, where $Price_m=(P_{m_1},P_{m_2}\dots,P_{m_L} )$ is the suggested prices for the offered resources by the provider $m$, $P_{m_l}$ indicates the suggested price per unit of resource type $l$ by the provider $m$, $Quantity_n=(Q_{m_1},Q_{m_2},\dots,Q_{m_L} )$ is the offered resources bundle by the provider $m$ and $Q_{m_l}$ indicates the offered resource type $l$ by the provider $m$.
\end{itemize}
Next, the auctioneer collects the initial bids of the consumers and then according to the stored historical information in the repository extends the bids by fairness attribute. We defined the consumer extended bid as $BidExt_n^{'}=(Bid_n^{'},FairnessFactor_n )$ where $FairnessFactor_n$ is the fairness factor for the consumer $n$ calculated by Algorithm 1.
\subsection{Winning Determination Problem (WDP)}
Winner determination problem (WDP) plays an important role in an auction and its performance will change according to which algorithm is used \cite{pap3}. The purpose of WDP is to determine the participants who have the best bids as the winners. The best bids are selected based on criteria such as the suggested price for each unit of the resources. In our proposed method, we formulate WDP in integer programming (IP) framework to maximize the total utility of the participants. Also in order to guarantee fairness in the auction we apply the fairness factor based on the criteria mentioned previously in a random fashion. Applying the fairness factor besides maximization of the utility, decreases the chance of winning of a participant who has dominated the auction and by contrast, increases the winning odds for the weaker participants. Accordingly, we define the objective function as follows: $Obj = \max$ \{Total Utility + Total Satisfaction\}.
Each provider and consumer has a utility function which represents the difference between the suggested price and the trade price and is calculated according to \eqref{eq5} for the provider $m$ and the consumer $n$.
\begin{equation}
\begin{split}
Utility_m = trade\;price - suggested\;price\\
Utility_n = suggested\; price - trade\; price
\end{split}
\label{eq5}
\end{equation}
The Total Utility of participants is the summation of all consumers and providers utilities and is obtained through the following formula:
\begin{equation}
\begin{split}
Total\; Utility=Consumers\; Utilities+Providers\;Utilities=&\\
(\sum_{n,l} x_n P_{n_l}^{'} Q_{n_l}^{'} - \sum_{n,l,m}y_{nlm} TradePrice_{nlm} )+& \\
(\sum_{n,l,m} y_{nlm} (TradePrice_{nlm}- P_{ml}))&
\end{split}
\label{eq6}
\end{equation}
\begin{equation}
v_n = \sum_l P_{nl}'Q'_{nl}
\label{eq7}
\end{equation}
In the above equations, $TradePrice_{nlm}$ is the transaction price between the consumer $n$ and provider $m$ per unit of resource type $l$. Also $v_n$ is the total budget that the consumer $n$ can pay for all the required resources. Accordingly, we have,
\begin{equation*}
Total\; Utility=\sum_n x_n.v_n -\sum_{m,n,l} y_{nlm}.P_{ml} 〗
\end{equation*}
and the Total Satisfaction of consumers is computed as:
\begin{equation*}
Total\; Satisfaction = \left(\sum_n x_n FairnessFactor_n \right)
\end{equation*}
where the constraints are tied together according to the following integer programming formulation:
\begin{eqnarray*}
x_n\in\{0,1\} & \forall n \in 1\dots N\\
y_{nlm} \in \{0,1,\dots,Q_{ml} \} & \forall m \in 1\dots M ,\\
&\forall n \in 1\dots N, \forall l \in 1\dots L\\
x_n \ge \left( \sum_{l,m} y_{nlm} \right)/\left(\sum_l Q_{nl}^{'} \right) & \forall n \in 1\dots N\\
x_n \le \sum _ {l,m} y_{nlm} & \forall n \in 1\dots N\\
\sum_n y_{nlm} \le Q_{ml} & \forall l \in 1\dots L,\\
& \forall m \in 1\dots M\\
\sum_m y_{nlm} = x_n.Q_{nl}^{'} & \forall n \in 1\dots N, \\
&\forall l \in 1\dots L\\
y_{nlm}\left(P_{nL}^{'} - P_{mL}\right)\ge 0 & \forall n \in 1\dots N ,\forall l \in 1\dots L,\\
& \forall m \in 1\dots M\\
\end{eqnarray*}
In the above relations, $x_n=1$ indicates that consumer $n$ has won the auction round and gets all the resources he requested in his bid and lost otherwise ($x_n=0$), $y_{nlm}$ is the number of resource type $l$ sold by the provider $m$ to consumer $n$ and $v_n$ is the suggested price for all requested resources by the consumer $n$.
\subsection {Pricing Mechanism}
After determining the winners of the auction, each consumer must pay an amount to use resources of the related provider. This amount should be fair for both the customer and the provider to bring about some level of satisfaction among both parties. In the calculation of trade price, the average suggested prices approach is considered for each type of resources. It is assumed that consumer $n$ received the resources of type $l$ from provider $m$ and therefore the final payable amount for this resource is given by:
\begin{equation}
TradePrice_{nlm}=\frac{(P_{nl}^{'}+P_{ml})}{2}
\label{eq17}
\end{equation}
Moreover, the total payment that consumer $n$ (provider $m$) must pay (receives) is computed as bellow:
\begin{equation}
TradePrice_{n}=\sum_{m=1}^{M}\sum_{l=1}^LTotal\_q_{nlm}TradePrice_{nlm}
\label{eq18}
\end{equation}
\begin{equation}
TradePrice_{m}=\sum_{n=1}^{N}\sum_{l=1}^LTotal\_q_{nlm}TradePrice_{nlm}
\label{eq19}
\end{equation}
where in the above equation $Total\_q_{nlm}$ is the number of resources of type $l$ that consumer $n$ takes from provider $m$.
\section{Performance Evaluation }
In this section, we simulate our proposed model to evaluate and compare it with a model which do not incorporate the fairness factor into its model, in terms of utility, the average percentage of resource utilization, the average percentage of winnings for consumers, the number of bidder drops, and the average round that bidder drop occurs. We use the Java language to implement the model and ILOG CPLEX v12.6 library to solve the corresponding WDP.
\subsection{Experimental setup}
We simulate an auction for 100 rounds and we repeated it for 10 times with different random seeds. We considered 4 different types of resources, 300 consumers and 5 providers. Each provider offers a random number (according to a uniform distribution, $U(30,100)$) of resource for each resource category, and each consumer requests from each type of resource 1-3 times (according to a uniform distribution, $U(1,3)$) in range of [1, 3]. Also the suggested price of each provider is distributed randomly within the range $[50, 200]$ and the suggested price for each consumer is selected, randomly ($U[100,250]$), according to the suggested prices from the previous round. For each consumer, the bidder drop problem occurs when he loosed 7 consecutive auction rounds in a row. Table \ref{tab1} demonstrates the selected parameters for the described scenario.
\begin{table}
\centering
\caption{Algorithm parametrization}
\label{my-label}
\begin{tabular}{||c|c||}
\hline
Parameter name & Parameter value \\\hline\hline
$\alpha_1$ & 9 \\\hline
$\alpha_2$ & 7 \\\hline
$\beta_1$ & 4 \\\hline
$\beta_2$ & 28 \\\hline
$ml$ & 6 \\\hline
$N$ & 300 \\\hline
$M$ & 5 \\\hline
$L$ & 4 \\\hline
\# rounds & 100 \\\hline
\# simulation runs & 10 \\\hline
\end{tabular}
\label{tab1}
\end{table}
\subsection{Experimental Result}
Figure \ref{fig2} and \ref{fig3} make comparisons between the MDFCDA and the no fairness models in terms of the number of bidder drops and the average round number that bidder drop occurs. As depicted in these figures, the number of drops in MDFCDA is consistently smaller than the baseline model and the bidder drop always occurs in the later rounds. Figure \ref{fig4} depicts the total utility (i.e. summation of providers and consumers utilities) in different runs of auction. According to the figure, in terms of the utilities the baseline model slightly outperforms the MDFCDA which is a trade-off due to the higher fairness in the MDFCDA model. Figure \ref{fig5} and \ref{fig6} compare the average percentage of resource utilization and the average percentage of consumer wins between the two models. As clearly implied by the figures, the proposed model outperforms the baseline model. Figure \ref{fig7} and \ref{fig8} illustrate cumulative sum of the number of bidder drops and the average percent of resource utilization, for a single run respectively. Overall, from the simulation results we can conclude that our model improves the average percentage of resource utilization and the average percentage of consumer wins and also decreases the number of bidder drop in different auction rounds.
\begin{figure}[t!]
\centering
\includegraphics[scale=0.2,trim=0.5cm 3.5cm 0.5cm 4cm]{fig2}
\caption{The number of bidder drops across different runs }
\label{fig2}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.2,trim=0.5cm 3.5cm 0.5cm 4cm]{fig3}
\caption{The average round number that bidder drop occurs across different runs.}
\label{fig3}
\end{figure}%
\begin{figure}[t]
\centering
\includegraphics[scale=0.2,trim=0.5cm 3.5cm 0.5cm 4cm]{fig4}
\caption{The total utility across different runs.}
\label{fig4}
\end{figure}%
\begin{figure}[t]
\centering
\includegraphics[scale=0.2,trim=0.5cm 3.5cm 0.5cm 4cm]{fig5}
\caption{The average percentage of resource utilization across different runs.}
\label{fig5}
\end{figure}%
\begin{figure}
\centering
\includegraphics[scale=0.23,trim=0.5cm 4.5cm 0.5cm 3.5cm]{fig6}
\caption{The average percentage of winning across different runs.}
\label{fig6}
\end{figure}
\begin{figure}[t!]
\centering
\includegraphics[scale=0.23,trim=0.5cm 4.5cm 0.5cm 3.5cm]{fig7}
\caption{The number of bidder drops across different rounds of the first run.}
\label{fig7}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.23,trim=0.5cm 4.5cm 0.5cm 3.5cm]{fig8}
\caption{The average percentage of resource utilization across different rounds of the first run.}
\label{fig8}
\end{figure}
\section{Conclusion and Future work}
Fair resource allocation has been always a challenging task in auction systems, yet few papers have addressed this important criterion in auction systems. Here we made an effort to integrate the fairness as a factor by leveraging the power of integer programming and efficient tools to solve the optimum solutions for this NP-complete problem. In so doing, we developed a combinatorial double-sided auction model with fairness factor to maximize the overall profit while at the same time maintaining some level of fairness. Our results showed that our proposed model promisingly outperforms a similar baseline model where the fairness factor is not considered. The MDFCDA model can efficiently deal with the bidder drop problem and yield better resource utilization. Moreover, It results in a higher average number of winners for most cases.
\bibliographystyle{abbrv}
|
\section{Introduction}
In the 1950's, W. Heisenberg has offered the procedure of nonperturbative (NP) quantization for a nonlinear spinor field \cite{heis}. His purpose was to obtain all physical properties of electron from the first principles, i.e. from a fundamental equation, which he suggested to be the equation for a nonlinear spinor field. Following this approach, he was able to obtain, to some accuracy, the main properties of electron.
Heisenberg's main idea was to write the operator nonlinear Dirac equation. Then, on multiplying by field operators and performing subsequent quantum averaging,
one can obtain an infinite set of equations for all Green functions.
In order to make practical use of the infinite set of equations, he proposed to cut off this system of equations to obtain a finite set of equations.
In principle, this idea can be used for any strongly interacting fields. Here we employ this approach for quantum chromodynamics (QCD).
Our main goals are: (a) starting from the Yang-Mills operator equation, we write an infinite set of equations for all Green functions;
(b) using some assumptions, we truncate this set of equations up to two equations describing gauge fields from the subgroup $\mathcal G \subset SU(N)$
and gluon condensate for gauge fields belonging to the coset $SU(N)/\mathcal G$; (c) we apply these two equation to obtain a flux tube solution describing the
field distribution between quark and antiquark.
Similar approach for the perturbative calculations of field correlators can be found in Refs. \cite{Simonov:2009nf}-\cite{DelDebbio:1994zn}
where the authors consider the formalism of gauge-invariant nonlocal correlators in non-Abelian gauge theories and derive nonlinear equations for field correlators in the large $N_c$ gluodynamics.
One of unsolved problems in quantum chromodynamics (QCD) is the problem of field distribution between quark and antiquark. Similar problem is easily solved in electrodynamics: the distribution of electric field between positive and negative charges can be easily found since Maxwell's electrodynamics is a linear theory. In QCD the problem is that the calculations should be done for non-perturbatively quantized fields because Yang-Mills theories are strongly nonlinear ones. The standard point of view
is that in QCD there is the dual Meissner effect: longitudinal electric field lines get compressed to a flux tube.
The flux tube field distribution is investigated within the framework of lattice QCD. In Ref. \cite{Ichie:2002dy}, the Abelian color flux of two- and
three-quark systems in the maximally Abelian gauge in lattice QCD with dynamical fermions is investigated. In Refs.~\cite{Shibata:2014tpa} and \cite{Shibata:2012ae},
the non-Abelian dual Meissner effect in the SU(3) Yang-Mills theory is investigated by measuring the chromoelectric flux created by an quark-antiquark source.
Lattice calculations strongly support the idea of the dual Meissner effect in QCD. However, for a more complete understanding of the nature of confinement,
it is necessary to have at least approximate analytical calculations confirming this point of view.
Here we will show that applying the non-perturbative quantization \`{a} la Heisenberg for QCD and using the two-equation approximation, a solution describing the flux tube between quark and antiquark located at $\pm \infty$ can be obtained. The solution is characterized by a longitudinal color electric field directed from quark to antiquark. All fields in this solution are expelled by a condensate of coset gauge fields into the flux tube. This is a non-Abelian version of the field distribution between positive and negative charges in Maxwell's electrodynamics. Then we want to consider a non-Abelian version of the field distribution between charges with the same sign in Maxwell's electrodynamics. We expect that in this case we will have two longitudinal electric fields directed oppositely. We will consider some special case when these fields are the same that leads to zero longitudinal color electric field in the flux tube.
\section{Nonperturbative quantization \`{a} la Heisenberg for non-Abelian gauge theories}
Following Heisenberg, we write the SU(N) Yang-Mills equations as operator equations
\begin{equation}\label{1-10}
D_\nu \widehat {F}^{a \mu\nu} = 0,
\end{equation}
where
$\hat F^B_{\mu \nu} = \partial_\mu \hat A^B_\nu - \partial_\nu \hat A^B_\mu +
g f^{BCD} \hat A^C_\mu \hat A^D_\nu$ is the field strength operator;
$\hat A^B_\mu$ is the gauge potential operator; $B, C, D = 1, \ldots , N$ are the SU(N) color indices; $g$ is the coupling constant; $f^{BCD}$ are the structure constants for the SU(N) gauge group.
How we can solve this equation? Following Heisenberg \cite{heis}, we have to write an infinite set of equations for all Green functions
\begin{eqnarray}
\left\langle
D_\nu \widehat {F}^{A \mu\nu} (x)
\right\rangle &=& 0 ,
\label{1-20}\\
\left\langle
\hat A^{B_1}_{\alpha_1} (x_1)
D_\nu \widehat {F}^{A \mu\nu} (x)
\right\rangle &=& 0 ,
\label{1-30}\\
\left\langle
\hat A^{B_1}_{\alpha_1} (x_1) \hat A^{B_2}_{\alpha_2} (x_2)
D_\nu \widehat {F}^{A \mu\nu} (x)
\right\rangle &=& 0 ,
\label{1-40}\\
\ldots &=& 0 ,
\label{1-50}\\
\left\langle
\hat A^{B_1}_{\alpha_1} (x_1) \ldots \hat A^{B_n}_{\alpha_n} (x_n)
D_\nu \widehat {F}^{A \mu\nu} (x)
\right\rangle &=& 0
\label{1-60}\\
\ldots &=& 0 .
\label{1-70}
\end{eqnarray}
Here
$
\left\langle (\ldots) \right\rangle =
\left\langle Q \left| (\ldots) \right| Q \right\rangle
$ and $\left. \left. \right| Q \right\rangle$ is a quantum state of the given physical system. For brevity, we will write $\left. \right\rangle$ instead of
$\left. \left. \right| Q \right\rangle$.
The first equation \eqref{1-20} is the equation for the 1-st order Green function $G^B_{\mu} (x) = \left\langle A^B_{\mu}(x) \right\rangle$.
But it contains Green functions of the 2-nd order $\left\langle A \partial A \right\rangle$ and 3-rd order $\left\langle A^3 \right\rangle$ (for brevity, we omit all indices).
The second equation \eqref{1-30} is the equation for the 2-nd order Green function
$G^{BC}_{\mu \nu} (x,x) = \left\langle A^B_{\mu}(x) A^c_{\nu}(x) \right\rangle$.
But it contains Green functions of the 3-rd order $\left\langle A^2 \partial A \right\rangle$ and 4-th order $\left\langle A^4 \right\rangle$.
We see that such situation holds true for all other equations \eqref{1-40}-\eqref{1-70}. Thus, to close this set of equations, we have to write an infinite system of equations.
The solution of the full set of equations \eqref{1-20}-\eqref{1-70} gives us full information on a quantum state $\left. \left. \right| Q \right\rangle$ and field operators $\hat A^B_\mu$.
In this sense, we can say that the solution of the full set of equations \eqref{1-20}-\eqref{1-70} is the solution of the operator field equations \eqref{1-10}.
We would like to note that in Ref. \cite{Frasca:2015yva} the author considers Dyson - Schwinger equations set \eqref{1-20}-\eqref{1-70} for the Yang-Mills theory stopping to the two-point function. We have to note that in our approach presented here we use some approximation in order to work with 4 - point Green function.
\section{Two-equation approximation}
In practice, we cannot solve the infinite set of equations \eqref{1-20}-\eqref{1-70}. We need to cut it off to obtain a \emph{finite} set of equations.
Then the solution of the truncated set of equations \eqref{1-20}-\eqref{1-70} will approximately describe the solution of the full system.
In order to do so, we have to use some physical intuition. For example, we can assume that $n$-th Green function is a polylinear combination of $m < n$
Green functions and use it either to cut off the set of equations \eqref{1-20}-\eqref{1-70} or to write an effective Lagrangian which will be an averaged $SU(N)$ Lagrangian.
Probably there exist other variants to cut off the infinite set of equations \eqref{1-20}-\eqref{1-70}.
In this section we use the following strategy: we assume that in some physical situations all $SU(N)$ degrees of freedom can be decomposed into two groups.
In the first group, the gauge fields
$\hat A^a_\mu = \langle \hat A^a_\mu \rangle + i \delta \hat A^a_\mu \in \mathcal G \subset SU(N)$, where $\langle \hat A^a_\mu \rangle = A^a_\mu$
will be treated as classical fields, and
$\delta \hat A^a_\mu$ is the quantum fluctuation around the classical field $A^a_\mu$. In the second group, the gauge fields $A^m_\mu \in SU(N) / \mathcal G$
are pure quantum ones in the sense that
$\langle \hat A^m_\mu \rangle = 0$.
We will consider physical systems where the quantum average of odd degrees of the gauge field are zero,
\begin{equation}\label{1-80}
\left\langle
\left( \hat A^{m_1}_{\mu_1}(x_1) \ldots \hat A^{m_{2k+1}}_{\mu_{2k+1}}(x_{2k+1}) \right)
\right\rangle =
\left\langle
\left( \delta \hat A^{a_1}_{\mu_1}(x_1) \ldots
\delta \hat A^{a_{2k+1}}_{\mu_{2k+1}}(x_{2k+1}) \right)
\right\rangle =
0 .
\end{equation}
The decomposition of field strengths $\hat F^a_{\mu \nu}$ and $\hat F^m_{\mu \nu}$ into $(\cdots)^a$ and $(\cdots)^m$ parts can be written as follows:
\begin{eqnarray}
\hat F^a_{\mu \nu} &=& \hat{\mathcal F}^a_{\mu \nu} +
g f^{amn} \hat A^m_\mu \hat A^n_\nu ,
\label{1-82}\\
\hat F^m_{\mu \nu} &=& \partial_{[ \mu} \hat A^m_{\nu ]} -
g f^{amn} \hat A^a_{[ \mu} \hat A^n_{\nu ]} +
g f^{mpq} \hat A^p_\mu \hat A^q_\nu.
\label{1-84}
\end{eqnarray}
Here $
\hat{\mathcal F}^a_{\mu \nu} = \partial_\mu \hat A^a_\nu -
\partial_\nu \hat A^a_\mu + g f^{abc} \hat A^b_\mu \hat A^c_\nu
$ is the field strength tensor of the subgroup $\mathcal G$;
$
\hat A^B_{[\mu} \hat A^C_{\nu ]} ] = \hat A^B_\mu \hat A^C_\nu - \hat A^B_\nu \hat A^C_\mu
$ is the antisymmetrization procedure; $a, b, c, \ldots$ are the subgroup indices $\mathcal G$, and $m, n, p, \dots$ are the coset indices.
We will consider such subgroup $\mathcal G$ in order to have structure constants $f^{mpq} = 0$, where the indices $m, p, q$ are in the coset $SU(N) / \mathcal G$.
For example, it can be pairs $U(1) \subset SU(2)$ or $SU(2) \times U(1) \subset SU(3)$.
Let us consider the equation \eqref{1-20} for $A = a$. After algebraic manipulations, we have
\begin{equation}\label{1-90}
\tilde D_\nu \mathcal F^{a \mu \nu} - \left( m^2 \right)^{a b \mu \nu} A^b_\nu
+ \left( \mu^2 \right)^{ab \mu\nu} A^b_\nu = j^{a \mu},
\end{equation}
where $
\tilde D_\mu = \partial_\mu + g f^{abc} A^b_\mu$ is the covariant derivative in the subgroup
$\mathcal G$ (for details see Appendix \ref{averaged}).
2-point Green functions for the gauge fields $\delta \hat A^a_\mu \in \mathcal G$ and for the coset $\hat A^m_\mu \in SU(N) /\mathcal G$ are defined as
\begin{eqnarray}
G^{mn \mu \nu}(y,x) &=& \left\langle
\hat A^{m \mu}(y) \hat A^{n \nu}(x)
\right\rangle,
\label{1-120}\\
G^{ab \mu \nu}(y,x) &=& \left\langle
\delta \hat A^{a \mu}(y) \delta \hat A^{b \nu}(x)
\right\rangle .
\label{1-125}
\end{eqnarray}
Equation \eqref{1-20} for $A = m$ gives us
\begin{equation}\label{1-130}
\left\langle
D_\nu \widehat {F}^{m \mu\nu}
\right\rangle = 0.
\end{equation}
Here we took into account the condition \eqref{1-80} and the simplification that quantum fields from the subgroup
$\mathcal G$ and the coset $SU(n) / \mathcal G$ do not correlate,
\begin{equation}\label{1-135}
\left\langle
\hat A^m_\nu \delta \hat A^a_\mu
\right\rangle = 0 .
\end{equation}
Next step is a consideration of equation \eqref{1-30} for $A = m, B_1 = r, \alpha_1 = \alpha$. After tedious calculations, we have (for details see Appendix \ref{averaged})
\begin{equation}
\begin{split}
&
\left[
\partial_{x^\nu} \partial^{x^\mu} G^{rm \alpha \nu} (y, x) -
\Box_x G^{rm \alpha \mu} (y, x)
\right]_{y = x} +
\\
&
g f^{amn} \biggl \{
- \partial_{x^\nu} \left[
A^{a \mu} (x) G^{rn \alpha \nu} (y, x) -
A^{a \nu} (x) G^{rn \alpha \mu} (y, x)
\right]_{y = x} -
A^a_\nu (x) \left[
\partial^{x^\mu} G^{rn \alpha \nu} (y, x) -
\partial^{x^\nu} G^{rn \alpha \mu} (y, x)
\right]_{y = x} +
\\
&
G^{rn \alpha}_{\phantom{rn \alpha} \nu} (x, x)
\mathcal F^{a \mu \nu}(x)
\biggl \} +
g^2 f^{amn} f^{bnp} A^a_\nu (x)
\biggl[
A^{b \mu} (x) G^{rp \alpha \nu} (x, x) - A^{b \nu} (x) G^{rp \alpha \mu} (x, x)
\biggl] +
\\
&
g^2 f^{amn} f^{apq} G^{rnpq \alpha \phantom{\nu} \mu \nu}_{\phantom{rnpq \alpha} \nu} (x,x,x,x)
= 0.
\label{1-140}
\end{split}
\end{equation}
Here we have introduced a 4-point Green function
\begin{equation}\label{1-150}
G^{mnpq}_{\phantom{mnpq}\mu \nu \rho \sigma}(x, y, z, u) = \left\langle
\hat A^m_\mu(x) \hat A^n_\nu(y) \hat A^p_\rho(z) \hat A^q_\sigma(u)
\right\rangle .
\end{equation}
In order to close the equations \eqref{1-90} and \eqref{1-140}, we have to define the last term in \eqref{1-140}: 4-point Green function \eqref{1-150}.
Finally, two-equation approximation of the NP quantization \`{a} la Heisenberg is (see the equations \eqref{a-120} and \eqref{a-130} from Appendix \ref{averaged})
\begin{eqnarray}
&&
\tilde D_\nu \mathcal F^{a \mu \nu} - \left[
\left( m^2 \right)^{ab \mu \nu} -
\left( \mu^2 \right)^{ab \mu \nu}
\right] A^b_\nu = j^{a \mu} ,
\label{1-160} \\
&&
\left[
\partial_{x^\nu} \partial^{x^\mu} G^{rm \alpha \nu} (y, x) -
\Box_x G^{rm \alpha \mu} (y, x)
\right]_{y = x} +
\nonumber \\
&&
g f^{amn} \biggl \{
- \partial_{x^\nu} \left[
A^{a \mu} (x) G^{rn \alpha \nu} (y, x) -
A^{a \nu} (x) G^{rn \alpha \mu} (y, x)
\right]_{y = x} -
A^a_\nu (x) \left[
\partial^{x^\mu} G^{rn \alpha \nu} (y, x) -
\partial^{x^\nu} G^{rn \alpha \mu} (y, x)
\right]_{y = x} +
\nonumber \\
&&
G^{rn \alpha}_{\phantom{rn \alpha} \nu} (x, x)
\mathcal F^{a \mu \nu}(x)
\biggl \} +
g^2 f^{amn} f^{bnp} A^a_\nu (x)
\biggl[
A^{b \mu} (x) G^{rp \alpha \nu} (x, x) - A^{b \nu} (x) G^{rp \alpha \mu} (x, x)
\biggl] +
\nonumber \\
&&
g^2 f^{amn} f^{apq} G^{rnpq \alpha \phantom{\nu} \mu \nu}_{\phantom{rnpq \alpha} \nu} (x,x,x,x)
= 0 .
\label{1-170}
\end{eqnarray}
The first equation \eqref{1-160} describes the dynamics of $\mathcal G$ gauge fields interacting with a coset condensate of quantum degrees of freedom belonging to the coset $SU(N) / \mathcal G$.
The condensate is described by 2- and 4-point Green functions and it is defined by the second equation \eqref{1-170}.
In order to have a closed set of equations, we have to make some assumptions about a 4-point Green function. For example, it can be expressed as a bilinear combination of 2-point Green functions $G^{(2)}$:
\begin{equation}\label{1-180}
G^{(4)} \approx
C_1 \left(G^{(2)} \right)^2 + C_2 G^{(2)} + C_3,
\end{equation}
where $C_{1,2,3}$ are constants (here we omit all indices). Probably the simplest example is a Mexican hat approximation that schematically looks like
\begin{eqnarray}
G^{(4)} \approx \frac{\lambda}{4} \left(
G^{(2)} - m^2 \
\right)^2 + \text{const}.
\label{1-185}
\end{eqnarray}
Similar decomposition is used in Ref. \cite{Calzetta:1999xh} to derive the Boltzmann - Langevin equation as the correct description of the kinetic limit of quantum field theory.
\subsection{Qualitative consideration of possible regular solutions}
Here we want to consider possible solutions of the equations \eqref{1-160} and \eqref{1-170}.
Taking into account the expressions for the mass matrix $\left( \mu^2 \right)^{ab\mu\nu}$ and the current $j^{a\mu}$, we can assume that regular solutions with
\begin{equation}\label{2a-10}
G^{mp \mu \nu}(x,x) \stackrel{r, \rho \rightarrow \infty}{\longrightarrow} m^2
\end{equation}
may exist, where $r$ is for the spherical case, $\rho$ is for the cylindrical case, and $m^2 = \text{const}$. Then the classical field $A^a_\mu$ will decay as
\begin{eqnarray}\label{2a-20}
A^a_\mu &\propto& \frac{e^{- m r}}{r}, \text{ for the spherical case} ,
\\
A^a_\mu &\propto& \frac{e^{- m \rho}}{\sqrt \rho}, \text{ for the cylindrical case},
\label{2a-25}
\end{eqnarray}
and the equation \eqref{1-170} is also satisfied since
$\partial_\mu G^{(2)} \rightarrow 0$ and $G^{(4)} \rightarrow 0$ by virtue of \eqref{1-180}.
Physically such solutions will present either a hedgehog filled with the classical non-Abelian
$\mathcal G$ gauge field or a flux tube. The gluon condensate created by the coset gauge fields
$A^m_\mu \in SU(N) / \mathcal G$ confines classical $\mathcal G$ gauge fields either in the ball or in the tube.
\subsection{Energy density}
In this section we want to calculate the energy density for the two-equation approximation:
\begin{equation}
\varepsilon = \left \langle
\frac{\left( E^A_{i} \right )^2 + \left( H^A_{i} \right )^2}{8 \pi}
\right \rangle,
\label{2-10}
\end{equation}
where $E^A_i = F^A_{0i}$ is the color electric field; $H^A_i = \frac{1}{2} \epsilon_{ijk} F^A_{jk}$ is the color magnetic field,
and $\epsilon_{ijk}$ is the Levi-Civita symbol. Taking into account \eqref{1-82}-\eqref{1-84} and the following algebraic properties of the subgroup $\mathcal G \subset SU(N)$
\begin{equation}
f^{abc} , f^{amn } \neq 0 , f^{mnp} = 0, \quad a,b,c \in \mathcal G ;
m, n, p \in SU(N) / \mathcal G,
\label{2-20}
\end{equation}
we obtain
\begin{equation}
\begin{split}
8 \pi \varepsilon = & \left\langle
\left (\mathcal E^a_i \right)^2
\right\rangle +
\left\langle
\left (\mathcal H^a_i \right)^2
\right\rangle - 2 g f^{amn}
\left\langle \mathcal E^a_{i} \right\rangle G^{mn 0i}(x,x) +
2 g f^{amn} \epsilon_{ijk} \left\langle \mathcal H^a_{k} \right\rangle G^{mn ij}(x,x) -
\\
&
\left\langle
\partial_{[0} A^m_{i]} \partial^{[0} A^{(m)i}
\right\rangle +
\left\langle
\partial_{[i} A^m_{j]} \partial^{[i} A^{(m)j}
\right\rangle +
\\
&
g^2 \left\{
- f^{amn} f^{apq} G^{mnpq \phantom{0i} 0i}_{\phantom{mnpq} 0i}(x,x,x,x) +
f^{amn} f^{bmq} \left[
- \left\langle
A^a_{[ 0} A^n_{i ]} A^{b [ 0} A^{(q) i]}
\right\rangle +
\left\langle
A^a_{[ i} A^n_{j ]} A^{b [ i} A^{(q) j]}
\right\rangle
\right]
\right\}
\end{split}
\label{2-130}
\end{equation}
here we take into account $f^{amn} G^{mn}_{\phantom{mn}\mu \nu} = 0$. The interesting thing here is that, in principle, one can obtain regular solutions with finite energy even if the 2-point Green function $G^{mn}_{\phantom{mn} \mu \nu}$ is a nonzero constant at infinity. In such case it is necessary that the 4-point Green function $G^{mnpq}_{\phantom{mnpq} \alpha \beta \gamma \delta}$ and color electric $\mathcal E^a_i$ and magnetic $\mathcal H^a_i$ fields tend to zero at infinity. It can happen if we choose the bilinear decomposition \eqref{1-180}. In this case, even if $G^{(2)} \rightarrow m^2$, the 4-point Green function $G^{(4)} \rightarrow 0$, and if in addition the color electric and magnetic fields tend to zero, we will have the solution with finite energy. The equations \eqref{a-40} and \eqref{a-132} show that at infinity the mass matrix goes to a constant, and the current $j^{a \mu} \rightarrow 0$. Then asymptotically the equation \eqref{1-160} gives us exponentially decreasing color gauge fields $\mathcal E^a_i$ and $\mathcal H^a_i$.
\section{Infinite SU(3) flux tube between quark and antiquark}
In this section we would like to show that the set of equations \eqref{c-10} and \eqref{3c-70}
\begin{eqnarray}
\tilde D_\nu \mathcal F^{a \mu \nu} - \left[
\left( m^2 \right)^{ab \mu \nu} -
\left( \mu^2 \right)^{ab \mu \nu}
\right] A^b_\nu &=& 0 ,
\label{4-4}\\
\Box \phi - \left( m^2_\phi \right)^{ab \mu \nu} A^a_\nu A^b_\mu \phi -
\lambda \phi \left( M^2 - \phi^2 \right) &=& 0
\label{4-8}
\end{eqnarray}
has a flux tube solution where a longitudinal chromoelectric field is pushed out from a gluon condensate in the flux tube;
$
\left( m^2 \right)^{ab \mu \nu} A^b_\nu
$,
$
\left( \mu^2 \right)^{ab \mu \nu} A^b_\nu
$ and
$
\left( m^2_\phi \right)^{ab \mu \nu} A^a_\nu A^b_\mu \phi
$ are calculated in Appendix \ref{adaptation}. We seek a solution in the form
\begin{equation}
A^1_t(\rho) = \frac{f(\rho)}{g} ; \quad A^2_z(\rho) = \frac{v(\rho)}{g} ;
\quad \phi(\rho) = \phi(\rho).
\label{4-10}
\end{equation}
Here we use the cylindrical coordinate system $t, z, \rho , \varphi$. The corresponding color electric and magnetic fields are then
\begin{eqnarray}
E^3_z(\rho) &=& F^3_{tz} = \frac{f(\rho) v(\rho)}{g} ,
\label{4-70}\\
E^1_\rho(\rho) &=& F^1_{t \rho} = -\frac{f'(\rho)}{g} ,
\label{4-80}\\
H^2_\varphi (\rho) &=& \rho \epsilon_{\varphi \rho z} F^{2\rho z} =
- \rho \frac{v'(\rho)}{g}.
\label{4-90}
\end{eqnarray}
Substituting \eqref{4-10} into the equations \eqref{4-4} and \eqref{4-8}, we have
\begin{eqnarray}
f'' + \frac{f'}{\rho} &=& f \left( v^2 + m^2 \phi^2 - \mu^2_1 \right),
\label{sec1-30}\\
v'' + \frac{v'}{\rho} &=& v \left( - f^2 + m^2 \phi^2 - \mu^2_2 \right),
\label{4-20}\\
\phi'' + \frac{\phi'}{\rho} &=& \phi \left[
\frac{m^2_\phi}{g^2} \left( - f^2 + v^2 \right)
+ \lambda \left( \phi^2 - M^2 \right)\right].
\label{4-30}
\end{eqnarray}
Here the prime denotes differentiation with respect to $\rho$.
By redefining
$\phi = \tilde \phi / m$, $M = \tilde M / m$, and $\lambda = m^2 m^2_\phi \tilde \lambda / g$, we have
\begin{eqnarray}
f'' + \frac{f'}{\rho} &=& f \left( v^2 + \tilde \phi^2 - \mu^2_1 \right),
\label{4-40}\\
v'' + \frac{v'}{\rho} &=& v \left( - f^2 + \tilde \phi^2 - \mu^2_2 \right),
\label{4-50}\\
\tilde \phi'' + \frac{\tilde \phi'}{\rho} &=&
\frac{m^2_\phi}{g^2} \tilde \phi \left[ - f^2 + v^2
+ \tilde \lambda \left( \tilde \phi^2 - \tilde M^2 \right)\right]
\label{4-60}
\end{eqnarray}
here according to \eqref{c-110} $m^2_\phi/g^2 = 1/4$. Equations \eqref{4-40}-\eqref{4-60} are solved as a nonlinear eigenvalue problem with the eigenvalues $\mu_{1,2}, \tilde M$ and the eigenfunctions
$f, v, \tilde \phi$. The boundary conditions are
\begin{equation}
\begin{split}
f(0) &= 0.2, f'(0) = 0;
\\
v(0) &= 0.5, v'(0) = 0;
\\
\phi(0) &= 1.0, \phi'(0) = 0 .
\label{4-100}
\end{split}
\end{equation}
The results of calculations are presented in Figs. \ref{potentials} and \ref{fields}.
\begin{figure}[h]
\begin{minipage}[ht]{.45\linewidth}
\begin{center}
\fbox{
\includegraphics[width=.9\linewidth]{potentials}
}
\end{center}
\caption{The functions $f(\rho), v(\rho), \tilde \phi(\rho)$.
The solid curve is $f(\rho)$,
the dashed curve is $v(\rho)$,
the dotted curve is $\tilde \phi(\rho)$.
$\mu_1 = 1.2325683, \mu_2 = 1.180660003, M = 1.3137067$, $\tilde \lambda = 0.1$.
}
\label{potentials}
\end{minipage}\hfill
\begin{minipage}[ht]{.45\linewidth}
\begin{center}
\fbox{
\includegraphics[width=.9\linewidth]{fields}
}
\end{center}
\caption{The chromoelectric and chromomagnetic fields
$E^3_z(\rho), E^1_\rho(\rho), H^2_\varphi (\rho)$.
The solid curve is $E^3_z(\rho)$,
the dashed curve is $E^1_\rho(\rho)$,
the dotted curve is $H^2_\varphi (\rho)$
}
\label{fields}
\end{minipage}
\end{figure}
Thus, we have obtained the flux tube stretched between two infinitely separated quark and antiquark.
This is the non-Abelian analogue of electric field distribution between positive and negative electric charges in the Maxwell electrodynamics. In contrast with Abelian electrodynamics, the electric field is confined within the tube due to the self-interaction of the non-Abelian fields.
From Eq's \eqref{sec1-30}-\eqref{4-60} we can obtain the asymptotic behavior of functions
$f(\rho), v(x)$ and $\tilde \phi(\rho)$ is
\begin{eqnarray}
f(\rho) &\approx& f_0 \frac{e^{- \rho \sqrt{M^2 - \mu_1^2}}}{\sqrt{\rho}} ,
\label{4-110}\\
v(\rho) &\approx& v_0 \frac{e^{- \rho \sqrt{M^2 - \mu_2^2}}}{\sqrt{\rho}} ,
\label{4-120}\\
\tilde \phi(\rho) &\approx& M -
\phi_0 \frac{e^{- \rho \sqrt{2\lambda M^2}}}{\sqrt{\rho}}
\label{4-130}
\end{eqnarray}
where $f_0, v_0$ and $\phi_0$ are some constants.
\par
The linear energy density that follows from \eqref{2-130} and \eqref{4-10} - \eqref{4-90} is
\begin{equation}
8 \pi \epsilon(\rho) = \frac{{f^\prime}^2}{2} + \frac{{v^\prime}^2}{2} +
\frac{{\phi^\prime}^2}{2} + \frac{f^2 v^2}{8} + m^2 \frac{f^2 \phi^2}{8} +
m^2 \frac{v^2 \phi^2}{8} + \frac{\mu_1^2}{2} f^2 - \frac{\mu_2^2}{2} v^2 +
\frac{\lambda}{4} \left(
\phi^2 - M^2
\right)
\label{4-140}
\end{equation}
and is plotted in Fig. \ref{energy}. According to \eqref{4-110} - \eqref{4-130} the linear energy density exponentially decreases at the infinity.
\begin{figure}[h]
\begin{center}
\fbox{
\includegraphics[width=10cm,height=7cm]{energy.eps}}
\caption{The energy density $\epsilon(x)$.}
\label{energy}
\end{center}
\end{figure}
\par
The flux of the longitudinal electric field is
\begin{equation}
\Phi = \int E^3_z ds = \frac{2 \pi}{g}
\int^{\infty}_{0} \rho f(\rho)v(\rho) d\rho < \infty.
\label{4-150}
\end{equation}
It is useful to list the properties of the physical system under consideration for which the flux solution is obtained:
\begin{itemize}
\item the gauge fields belonging to the subgroup $\hat A^a_\mu \in SU(2) \times U(1) \subset SU(3)$ and the coset $\hat A^m_\mu \in SU(3) / (SU(2) \times U(1))$ have different dynamics;
\item quantum expectation values $\left\langle \hat A^a_\mu \right\rangle \neq 0$, but
$\left\langle \hat A^m_\mu \right\rangle = 0$;
\item 2-point Green functions of the coset fields $G^{mn \mu \nu}$ are approximately proportional to the total dispersion $D$ ($D$ is defined in Appendix \ref{approximation}):
$G^{mn \mu \nu}(x, x) \approx \delta^{mn} \mathcal A^\mu \mathcal A^\nu D(x, x)$;
\item 2-point Green functions of the subgroup fields $G^{ab \mu \nu}$ are approximately constant: $G^{ab \mu \nu}(x, x) \approx \Delta^{ab} \mathcal B^\mu \mathcal B^\nu$;
\item $\hat A^a_\mu = A^a_\mu + i \delta \hat A^a_\mu$, and the dispersion defect
$
\left\langle \delta \hat A^a_\mu \delta \hat A^b_\nu \right\rangle
$ gives rise to:
\begin{itemize}
\item the appearing of masses for the non-Abelian fields $A^a_\mu$;
\item that the masses
$
\left( \mu^2 \right)^{ab\mu\nu}
$ have to have a necessary sign for getting the flux tube solution;
\end{itemize}
\item the constants $m^2$ and $m^2_\phi$ are defined by the dispersions $G^{mn\mu\nu}$ of quantum fluctuations;
\item the closure procedure for the 4-point Green function
$G^{mnpq}_{\phantom{mnpq} \alpha \beta \mu \nu}$ which represents a 4-point Green function as a bilinear combination of the total dispersion:
$
G^{mnpq \alpha \beta \mu \nu}(x,x,x,x) \approx C^{mnpq \alpha \beta \mu \nu}
D(x,x) \left[ M^2 - D(x,x) \right]
$.
\item in order to have different masses $\mu_1 \neq \mu_2$, the color anisotropy for the quantum gauge fields $\delta \hat A^a_\mu \in SU(2)$ is necessary.
\end{itemize}
The flux tube solution obtained in this section shows us that in this situation we deal with the dual Meissner effect: the gluon condensate pushes out the color electric field. In this connection we want to note that in Ref. \cite{Baker:1998jw} the interaction between gluon condensate and dual Meissner effect is investigated.
\section{Infinite flux tube solution between quark and quark (antiquark and antiquark)}
Here we want to consider a non-Abelian version of the field distribution between charges with the same sign in Maxwell's electrodynamics. We expect that in this case we will have two longitudinal electric fields directed oppositely. We will consider some special case when these fields are the same that leads to zero longitudinal color electric field in the flux tube.
The set of equations describing such a situation is
\begin{eqnarray}
\tilde D_\nu F^{a \mu \nu} - \left[
\left( m^2 \right)^{ab \mu \nu} -
\left( \mu^2 \right)^{ab \mu \nu}
\right] A^b_\nu &=& 0 ,
\label{5-10}\\
\Box \phi - \left( m^2_\phi \right)^{ab \mu \nu} A^a_\nu A^b_\mu \phi -
\lambda \phi \left( M^2 - \phi^2 \right) &=& 0,
\label{5-20}
\end{eqnarray}
where
\begin{eqnarray}
\left( m^2 \right)^{ab \mu\nu} &=& - g^2 \left[
f^{abc} f^{cpq} G^{pq \mu\nu} -
f^{amn} f^{bnp} \left(
\eta^{\mu \nu} G^{mp \phantom{\alpha} \alpha}_{\phantom{mn} \alpha} -
G^{mp \nu \mu}
\right)
\right] ,
\label{5-30} \\
\left( \mu^2 \right)^{ab \mu \nu} &=& - g^2 \left(
f^{abc} f^{cde} G^{de \mu \nu} +
\eta^{\mu \nu} f^{adc} f^{cbe} G^{de \phantom{\alpha} \alpha}_{\phantom{de} \alpha} +
f^{aec} f^{cdb} G^{ed \nu \mu}
\right) ,
\label{5-40} \\
\left( m^2_\phi \right)^{ab \mu \nu} &=&
g^2 f^{amn} f^{bnp} \frac{
G^{mp \mu \nu} - \eta^{\mu \nu} G^{mp \alpha}_{\phantom{mp \alpha} \alpha}
}{G^{mm \alpha}_{\phantom{mm\alpha} \alpha}} .
\label{5-60}
\end{eqnarray}
2-point Green functions for the gauge fields $\delta \hat A^a_\mu \in SU(2) \times U(1)$ and for the coset $\hat A^m_\mu \in SU(3) / (SU(2) \times U(1))$ are defined as
\begin{eqnarray}
G^{mn \mu \nu}(y,x) &=& \left\langle
\hat A^{m \mu}(y) \hat A^{n \nu}(x)
\right\rangle ,
\label{5-70}\\
G^{ab \mu \nu}(y,x) &=& \left\langle
\delta \hat A^{a \mu}(y) \delta \hat A^{b \nu}(x)
\right\rangle ,
\label{5-80}
\end{eqnarray}
where
$F^a_{\mu \nu} = \partial_\mu A^a_\nu - \partial_\nu A^a_\mu + g f^{abc} A^b_\mu A^c_\nu$
is the field strength; $a, b,c,d =$ either $1, 2, 3$ or $2,5,7$ are the $SU(2)$ colour indices; $m,n = $ either $4,5, \cdots , 8$ or $1,3,4,6,8$; $g$
is the coupling constant; $f^{ABC}$ are the structure constants for the $SU(3)$ gauge group; $A, B, C = 1,2, \cdots , 8$. The equation \eqref{5-10}
describes $SU(2) \in SU(3)$ degrees of freedom that have non-zero expectation values, and equation \eqref{5-20} describes coset $SU(3) / SU(2)$
degrees of freedom with zero expectation values:
\begin{eqnarray}
\hat A^{a \mu} &=& \left\langle \hat A^{a \mu} \right\rangle +
i \delta \hat A^{a \mu} ,
\label{5-90}\\
\left\langle \hat A^{m \mu} \right\rangle &=& 0 .
\label{5-100}
\end{eqnarray}
We seek a cylindrically symmetric solution of equations \eqref{5-10} and \eqref{5-20} in the subgroup $SU(2) \in SU(3)$ spanned on either $\lambda^{1,2,3}$ or $\lambda^{2,5,7}$ in the form
\begin{eqnarray}
A^{1,2}_t(\rho) &=& \frac{f(\rho)}{g} , \quad
A^{1,2}_z(\rho) = \frac{u(\rho)}{g} ,
\label{1-110}\\
A^{2,5}_t(\rho) &=& \frac{w(\rho)}{g} , \quad
A^{2,5}_z(\rho) = \frac{v(\rho)}{g} ,
\label{5-120}\\
\phi(\rho) &=& \frac{\phi(\rho)}{g}.
\label{5-130}
\end{eqnarray}
Here the first superscript indices are for $\lambda^{1,2,3}$ and the second ones -- for $\lambda^{2,5,7}$. We work in a cylindrical coordinate
system $z, \rho, \varphi$, and the corresponding colour electric and magnetic fields are then
\begin{eqnarray}
E^{3,7}_z(\rho) &=& F^{3,7}_{tz} = \left( E^{3,7}_z \right)_1 - \left( E^{3,7}_z \right)_2 =
\frac{fv - w u}{g} ,
\label{5-140}\\
E^{1,2}_\rho(\rho) &=& F^{1,2}_{t \rho} =
- \frac{f'(\rho)}{g} , \quad
E^{2,5}_\rho(\rho) = F^{2,5}_{t \rho} = - \frac{w'(\rho)}{g} ,
\label{5-150}\\
H^{2,5}_\varphi (\rho) &=& \epsilon_{\varphi \rho z} F^{2,5 \; \rho z} =
- \frac{v'(\rho)}{g} , \quad
H^{1,2}_\varphi (\rho) = \epsilon_{\varphi \rho z} F^{1,2 \; \rho z} =
- \frac{u'(\rho)}{g},
\label{5-160}
\end{eqnarray}
where
$\left( E^{3,7}_z \right)_1 = A^{1,2}_t A^{2,5}_z = \frac{fv}{g},
\left( E^{3,7}_z \right)_2 = A^{2,5}_tA^{1,2}_z = - \frac{w u }{g}$.
For simplicity, we consider the case with
\begin{equation}
w = f, u = v.
\label{5-170}
\end{equation}
In both cases we have the following set of equations (for details see Appendix \ref{a1})
\begin{eqnarray}
- f'' - \frac{f'}{\rho} + m^2 \phi^2 f &=& \mu^2 f ,
\label{5-180}\\
-v'' - \frac{v'}{\rho} + m^2 \phi^2 v &=& \mu^2 v ,
\label{5-190}\\
\phi'' + \frac{\phi'}{\rho} &=& \phi \left[
\tilde \alpha \left(
- f^2 + v^2
\right) + \lambda \left(
\phi^2 - M^2
\right)
\right] .
\label{5-200}
\end{eqnarray}
We see that equations \eqref{5-180} and \eqref{5-190} are Schr\" odinger-type equations with a solution
$v(\rho) = k f(\rho)$, where $k$ is a constant, $\phi$ is the potential and $\mu$ is an eigenvalue. In this case we can rewrite the set of equations \eqref{5-180}-\eqref{5-200} as follows
\begin{eqnarray}
- f'' - \frac{f'}{x} + \phi^2 f &=& \mu^2 f ,
\label{5-210}\\
\phi'' + \frac{\phi'}{x} &=& \phi \left[
\alpha f^2 + \lambda \left( \phi^2 - M^2 \right)
\right].
\label{5-220}
\end{eqnarray}
Here $\alpha = \tilde \alpha (k^2 - 1)$ and it can be an arbitrary real number; we redefined
$m \phi / f(0) \rightarrow \phi$, $\lambda /m^2 \rightarrow \lambda$, $m M \rightarrow M$,
$f / f(0) \rightarrow f$, $x = \rho f(0)$. Numerical investigation shows that regular solution to \eqref{5-210} and \eqref{5-220} does exist only for some positive $\alpha > 0$.
It is necessary to note that because of \eqref{5-170} the total longitudinal electric field
$E^{3,7}_z = 0$, and this leads to the fact that equations \eqref{5-180}, \eqref{5-190},
and \eqref{5-210} do not have non-linear terms like $f v^2$.
The results of numerical calculations are presented in Figs.~\ref{potentials_zeroFT} and \ref{fields_zeroFT}. We see that we have the dual Meissner effect: the longitudinal electric fields $\left( E^{3,7}_z \right)_1$ and
$\left( E^{3,7}_z \right)_2$ are confined into a tube by the scalar field $\phi$
(which is a condensate of the coset fields).
\begin{figure}[h!]
\begin{minipage}[ht]{.45\linewidth}
\begin{center}
\fbox{
\includegraphics[width=.9\linewidth]{f_phi_zeroFT}
}
\end{center}
\caption{The profile of the functions $f(\rho), \phi(\rho)$.
}
\label{potentials_zeroFT}
\end{minipage}
\begin{minipage}[ht]{.45\linewidth}
\begin{center}
\fbox{
\includegraphics[width=.9\linewidth]{fields_zeroFT}
}
\end{center}
\caption{The profiles of electric and magnetic fields
$g E^{1,2}_\rho(\rho) = g E^{2,5}_\rho(\rho) = - f'(\rho)$,
$g H^{2,5}_\varphi (\rho) = g H^{1,2}_\varphi (\rho) = - f'(\rho)$,
$g \left( E^{3,7}_z \right)_1 = - g \left( E^{3,7}_z \right)_2 = f^2(\rho)$.
}
\label{fields_zeroFT}
\end{minipage}\hfill
\end{figure}
\section{NP quantization \`{a} la Heisenberg vs turbulence modelling}
In this section we would like to show that the two-equation approximation for the NP quantization \`{a} la Heisenberg and the Reynolds
approximation in stochastic turbulence modeling have the same mathematical basis:
the truncation of the corresponding infinite set of equations for all either Green functions or cumulants.
The Navier-Stokes equation describing a turbulent flow is (in this section we follow the textbook~\cite{Wilcox})
\begin{equation}
\rho \left(
\frac{\partial v_i}{\partial t}
+ v_j \cdot \frac{\partial v_i}{\partial x_j}
\right) = - \frac{\partial p}{\partial x_i} +
\frac{\partial t_{ij}}{\partial x_j},
\label{3-10}
\end{equation}
where $v_i$ is the flow velocity, $\rho$ is the fluid density, $p$ is the pressure, $t_{ij} = 2 \mu s_{ij}$ is the viscous stress tensor,
$s_{ij} = \frac{1}{2} \left(
\frac{\partial v_i}{\partial x_j} + \frac{\partial v_j}{\partial x_i}
\right)$, and $\mu$ is the molecular viscosity. All physical quantities are random ones and statistically fluctuate.
In order to show the similarity between the NP quantization and turbulence modelling, we will obtain two equations:
first of them will be the averaged Navier-Stokes equation, and the second one is obtained by multiplying the Navier-Stokes equation
by the velocity with subsequent averaging. The details can be found in the textbook \cite{Wilcox}, and the result is as follows:
\begin{eqnarray}
\rho \frac{\partial V_i}{\partial t} +
\rho V_j \frac{\partial V_i}{\partial x_j}
&=& - \frac{\partial p}{\partial x_i} +
\frac{\partial }{\partial x_j} \left(
2 \mu S_{ji} - \overline{\rho v_j^\prime v_i^\prime}
\right) ,
\label{3-20}\\
\frac{\partial \tau_{ij}}{\partial t} +
V_k \frac{\partial \tau_{ij}}{\partial x_k} &=&
- \tau_{ik} \frac{\partial V_j}{\partial x_k} -
\tau_{jk} \frac{\partial V_i}{\partial x_k} +
2 \mu \overline{\frac{\partial v^\prime_i}{\partial x_k}
\frac{\partial v^\prime_j}{\partial x_k}} -
\overline{
p^\prime \left(
\frac{\partial v_i^\prime}{\partial x_j} +
\frac{\partial v_j^\prime}{\partial x_i}
\right)
} +
\frac{\partial}{\partial x_k} \left(
\nu \frac{\partial \tau_{ij}}{\partial x_k} +
C_{ijk}
\right),
\label{3-30}
\end{eqnarray}
where $v_i(\vec x, t)$ is the instantaneous velocity; $V_i(\vec x, t)$ is a mean value of velocity; $v^\prime_i(\vec x, t)$ is a fluctuating part of the velocity;
$\tau_{ij} = - \overline{\rho v_i^\prime v_j^\prime}$ is the Reynolds stress tensor;
$\overline{( \cdots )}$ is the statistical averaging;
$
C_{ijk} =
\overline{\rho v^\prime_i \rho v^\prime_j \rho v^\prime_k} +
\delta_{jk} \overline{p^\prime v^\prime_i} +
\delta_{ik} \overline{p^\prime v^\prime_j}
$. Two equations \eqref{3-20} and \eqref{3-30} are not closed, since we have new unknown functions
$\overline{\rho v^\prime_i \rho v^\prime_j \rho v^\prime_k}$, $\overline{\mu \frac{\partial v^\prime_i}{\partial x_k}
\frac{\partial v^\prime_j}{\partial x_k}}$, and
$\overline{v^\prime_i \frac{\partial p^\prime}{\partial x_j}}$ . The equation \eqref{3-20} is the averaged Navier-Stokes equation, and \eqref{3-30} is the Reynolds equation. In order to close \eqref{3-20} and \eqref{3-30}, we have to make some physically reasonable assumptions about the unknowns, namely $C_{ijk}$,
$\overline{\rho v^\prime_i \rho v^\prime_j \rho v^\prime_k}$, $\overline{\mu \frac{\partial v^\prime_i}{\partial x_k}
\frac{\partial v^\prime_j}{\partial x_k}}$, and
$\overline{v^\prime_i \frac{\partial p^\prime}{\partial x_j}}$.
We can compare the NP quantization and turbulence modelling in the following table:
\fbox{
\addtolength{\linewidth}{-1\fboxsep}%
\addtolength{\linewidth}{-1\fboxrule}%
\begin{minipage}{\linewidth}
\begin{eqnarray*}
\text{quantum averaged Yang - Mills} &\leftrightarrow &
\text{stochastically averaged Navier-Stokes eqn.},
\\
\text{quantum averaged $A^m_\mu \cdot $ Yang - Mills} &\leftrightarrow &
\text{stochastically averaged Reynolds eqn., }
\\
\text{closure problem with a 4-point Green function } G^{(4)} &\leftrightarrow &
\text{closure problem with } C_{ijk},
\overline{\rho v^\prime_i \rho v^\prime_j \rho v^\prime_k},
\overline{\mu \frac{\partial v^\prime_i}{\partial x_k}
\frac{\partial v^\prime_j}{\partial x_k}},
\overline{v^\prime_i \frac{\partial p^\prime}{\partial x_j}} .
\end{eqnarray*}
\end{minipage}
}
\vspace{4mm}
The table gives us a full understanding of close interrelation between the NP quantization and turbulence modeling.
In addition, we want to note that the diagram technique from perturbative quantum field theory can be applied in turbulence modeling, for details see the textbook \cite{mccomb}.
\section{Discussion and conclusions}
The physical meaning of the two-equation approximation is to describe physical systems in which one
group of degrees of freedom is practically in a classical phase, and the remaining group of degrees of freedom is in a pure quantum phase.
In addition, in the first group, we have quantum fluctuations around the mean values.
The dispersion of these fluctuations gives rise to the appearance of masses of the corresponding gauge fields.
The dispersion of quantum fluctuations in the second group gives rise to the gluon condensate.
In fact, this system is a system where classical non-Abelian gauge fields belonging to a subgroup interact
with the quantum condensate of gauge fields belonging to the coset.
The most interesting case here can be the case when classical non-Abelian gauge fields are confined by a condensate of quantum gauge fields.
For example, it can be: (a) a flux tube with a longitudinal color electric field stretched between infinitely separated quark and antiquark;
(b) a hedgehog with an exponentially decreasing non-Abelian gauge field which is pushed out by the condensate, or something like this.
We have shown that applying the two-equation approximation in the non-perturbative quantization \`{a} la Heisenberg for QCD one can obtain the flux tube stretched between: (a) quark and antiquark, (b) quark and quark, and (c) antiquark and antiquark located at $\pm \infty$ with non-zero and zero longitudinal color electric field. It is shown that all color electric and magnetic fields are expelled by the scalar field that describes a condensate of coset non-Abelian fields. This effect is the analog of the Meissner effect in superconductivity for non-Abelian color fields.
One of the problems in QCD is to show that a flux tube filled with a longitudinal electric field does appear between quark and antiquark.
The conventional opinion in this case is that the appearance of the flux tube is the manifestation of the dual Meissner effect --
the pushing out of color electric field from the gluon condensate. We have shown that in our two-equation approximation such a solution does really exist.
That means that our approach can actually describe the dual Meissner effect.
As we mentioned above, probably the simplest way to close two equations \eqref{1-160} and \eqref{1-170}
is the Mexican hat approximation \eqref{1-180}. But one can assume that there are many other possibilities.
For example, Green functions of the operator Yang-Mills equation \eqref{1-10} can be connected with solutions in some classical field theory.
For instance, one can investigate a holographic idea that QCD Green functions are connected with a classical action of some classical theory calculated on a hypersurface of a bulk.
It would be interesting to compare Green functions calculated in this approach with results obtained in lattice calculations, see Ref's \cite{D'Elia:2015dxa} - \cite{D'Elia:2002ck}.
\section*{Acknowledgements}
This work was supported by Grant 3101/GF4 IPC-11 in fundamental research in natural sciences by the Ministry Education ans Science of Republic of Kazakhstan. I am very grateful to N. Kochelev and V. Folomeev for fruitful discussions and comments.
|
\section{Introduction}
Active Galactic Nucleus (AGN) is one of the most luminous objects in the universe. Observed as bright stellar-like point source, it is located in the core of massive galaxy. The extremely high luminosity and rapid variability require that gigantic energy is generated and released in a quite small volume with linear size not larger than a few parsecs. The accretion of interstellar medium onto super-massive black holes (SMBHs) is widely accepted as the process driving these central engines (Lynden-Bell 1969). The gravitational potential of the infalling matter is transfered into radiation by the viscous stress on the accretion disks surrounding the SMBHs (Rees 1984). Moreover, massive outflows from the inner part of the accretion disks are expected, not only to carry away the angular momentum of infalling matter to maintain the accretion, but also to explain the observed correlation between the properties of central SMBHs and their host galaxies. This correlation requires a mechanism of feedback from the central engines to regulate the star formation in the host galaxies (Granato \textit{et al.} 2004; Scannapieco \& Oh 2004; Hopkins \textit{et al.} 2008).
The feedback is believed to have solid direct observational basis, since both the asymmetric emission line profiles (Gaskell 1982; Richards \textit{et al.} 2002; Wang \textit{et al.} 2011) and the intrinsic absorption lines blueshifted to the quasars' rest frames (Weymann \textit{et al.} 1981) are detected and explained as the gaseous outflows. The latter, the blueshifted absorption troughs, are found with large outward velocity varying from a few thousand km s$^{-1}$ to about 0.2 $c$. Usually the profiles of these absorption troughs also display great widths, larger than $2000\ \mathrm{km\ s}^{-1}$. Thus, they are classified as broad absorption lines (BALs). A great deal of research works, for samples (Weymann \textit{et al.} 1991; Hewett \& Foltz 2003; Reichard \textit{et al.} 2003; Zhang \textit{et al.} 2010) or for individual objects (Wampler \textit{et al.} 1995; de Kool \textit{et al.} 2001, 2002; Leighly \textit{et al.} 2011; Zhang \textit{et al.} 2015a), have been published to investigate the geometry and evolution of the global structure of the outflows, or the physical conditions of individual outflowing clouds.
But the observational evidences of the more basic process, the accretion, remain questionable. Till now, few have been confirmed to directly demonstrate the existence of the accretion inflow in the vicinity of AGNs. Since the inflows are suggested originate from the torus (Vollmer \textit{et al.} 2004; Beckert \& Duschl 2004), the most possible place to discover it is the gap between the inner surface of the torus and the outer region of the accretion disk. But the ionizing flux here from the central engine is relatively weak so that the emission of the illuminated inflow would be overwhelmed by the broad emission lines. The absorption lines provide a more accessible way for the detection if the medium intercepts our line of sight (LOS) toward the central radiation source. The problem is that these inflows are expected to lie not far from the equatorial plane so that the LOS would be obscured by the torus.
Recent models tend to prefer a clumpy torus to explain the lack of $10\ \micron$ Si emission features in the spectral energy distribution (SED) of type 1 AGN (Nenkova \textit{et al.} 2002). The impact of such models to our concern is that we could thus have chances to look through the low-density part among the dense clumps especially when the angle of the LOS from the equator is relative large where the torus medium would be a bit more diffuse. The probability must be quite low. But once this occurs, the LOS could also intercepts the assumed inflow to unveil its existence by the redshifed absorption profile as a result of the inward motion.
The Sloan Digital Sky Survey (SDSS) provides the largest quasar sample, containing 105783 objects in the 7th data release for SDSS-I/II (Schneider \textit{et al.} 2010), and 297301 objects in the 12th data release for SDSS-III Baryon Oscillation Spectroscopic Survey (BOSS) (Alam \textit{et al.} 2015). The scale of the sample makes the detection of absorptions for the inflows much more available. In fact, Hall \textit{et al.} (2002) reported a few quasars with BAL troughs extending redshifted to the objects' rest frames. The most interesting one is SDSS J112526.12+002901.3 (hereafter J1125+0029), which shows two redshifted absorption troughs of hydrogen Balmer lines at about 70 and $650\ \mathrm{km\ s}^{-1}$, respectively. Both components are also present in the absorptions of metastable \ion{He}{1} $\lambda 3889$, while very strong \ion{Mg}{2} and UV \ion{Fe}{2} absorptions are found at around the same velocity although their exact profiles are unclear due to the overlapping. The rarely detected Balmer and \ion{He}{1}* absorptions, combined with low-ionization metal lines like \ion{Fe}{2}, are powerful diagnostics to the physical conditions of absorbing medium (Leighly \textit{et al.} 2011; Ji \textit{et al.} 2012; Liu \textit{et al.} 2015). Therefore J1125+0029 gives a suitable example to investigate the nature of the redshifted absorption line system.
This paper is organized as follows. In \S\ref{Observation} we describe the data of repeated spectroscopic observations including the SDSS-I/II, BOSS, and MMT. In \S\ref{Measurement} we measure the Balmer and \ion{He}{1}* absorption lines in the spectra using the curve of growth (COG) analysis.
The physical conditions of the absorbing medium are estimated using the photoionization simulations in \S\ref{Photoionization}, which provide the basis for the discussion about the origin of the absorption line system in \S\ref{Discussion}. And finally we give a brief summary in \S\ref{Summary}. The flux calibration and variability of the spectroscopic data is discussed in Appendix. Throughout this paper we assume a cosmology with $H_{0}=70\ \mathrm{km\ s}^{-1}\mathrm{Mpc}^{-1}$, $\Omega_{\mathrm{M}}=0.3$ and $\Omega_{\mathrm{\Lambda}}=0.7$.
\section{The Observations}\label{Observation}
J1125+0029 was observed as a quasar candidate for spectroscopy in SDSS-I/II (York \textit{et al.} 2000) on UT 2000 March 11 and SDSS-III BOSS (Dawson \textit{et al.} 2013) on UT 2011 January 14. The SDSS-I/II spectrum has a resolution of $R \approx 1800$, covering a wavelength range from 3800 to 9200 \AA. To reduce the contamination from the sky line subtraction residual on H$\gamma$ and H$\beta$ peaks, we employ the sky-residual subtracted SDSS DR7 spectrum{\footnote{The spectrum can be accessed through the Johns Hopkins University SDSS server. http://www.sdss.jhu.edu/skypca/spSpec}} post-processed by Wild \& Hewett (2005). The BOSS spectrum covers a wider wavelength range from 3600 to 10000 \AA. It seems the residual of sky line subtraction at the red end of the BOSS spectrum has little effect on our measurement. The spectroscopic data are then corrected for Galactic extinction using the mean extinction curve in Fitzpatrick \& Massa (2007), with selective extinction $E(B-V)=0.031$ in the Galactic dust map of Schlegel \textit{et al.} (1998). The narrow [\ion{O}{2}] $\lambda 3728$ emission presents a good measure of the systemic redshift (Hewett \& Wild 2010, Shen \textit{et al.} 2016). By fitting a Gaussian profile to the [\ion{O}{2}] $\lambda 3728$ line, we suggest a systemic redshift of $0.8632\pm 0.0002$. In Figure \ref{SED}, the SDSS-I/II and BOSS spectra are presented together.
The overall fluxes of the two spectra are significantly different. The fluxes of SDSS-I/II spectrum are about 60\% higher than those of BOSS spectrum. The cause could be the variability and the flux calibration problems in the reduction of the spectra. A detailed discussion about the variability is presented in Appendix.
Despite the great difference in the levels of fluxes, numerous similar absorption features can be identified in both spectra. In the part of spectra shortward of rest-frame 3000 \AA, a vast number of overlapping absorption troughs are the most prominent spectral features. Few windows are left absorption free in this part. The majority can be ascribed to the absorptions from \ion{Fe}{2} multiplets UV1 at $\sim 2600\ \mathrm{\AA}$ up to UV168 at $\sim 2200\ \mathrm{\AA}$. Others are identified as absorptions of \ion{Mg}{2} ($\sim 2800\ \mathrm{\AA}$), \ion{Cr}{2} ($\sim 2060\ \mathrm{\AA}$, $\sim 2665\ \mathrm{\AA}$ and $\sim 2850\ \mathrm{\AA}$), and \ion{Ni}{2} (2160 to $2325\ \mathrm{\AA}$).
The rarely detected Balmer series and \ion{He}{1}* $\lambda 3889$ absorptions are also present in both spectra. In Figure \ref{abs_var} we show the spectra around these absorptions in the velocity with respect to the quasar's rest frame, with the BOSS spectrum being scaled so that the fluxes in the absorption free sections $-4000$ to $-200\ \mathrm{km\ s}^{-1}$ and 1400 to $3400\ \mathrm{km\ s}^{-1}$ neighboring the troughs match those in the SDSS-I/II spectrum. It could thus be clearly seen that in both observations, these absorptions show two redshifted velocity components centered at $\sim 70$ and $650\ \mathrm{km\ s}^{-1}$, respectively, while no detectable changes in these central velocities are found. And the changes in the relative depth of these absorption profiles are much smaller compared with the difference in the overall fluxes.
To further monitor the absorption variability, a near-infrared (NIR) spectrum was obtained on the MMT red channel echellette using an $180\arcsec \times 1\arcsec$ slit on UT 2012 March 1. The exposure time is 1200 s, and the typical seeing is $0\arcsec .8$. The resolution is $R\approx 3270$, slightly higher than the SDSS data. We use the standard IRAF package to extract the 1-D spectrum and the fluxes are carefully calibrated. The spectrum covers H$\beta$ and narrow [\ion{O}{3}] $\lambda\lambda 4960,5008$ emission. The MMT spectrum around the H$\beta$ absorption is also plotted in the top panel of Figure \ref{abs_var}, scaled in the same way as the BOSS spectrum. The profile seems in good agreement with the BOSS observation, indicating that no variation of absorption can be detected from 2011 to 2012, about 7 months in the quasar's rest frame.
\section{Absorption Measurement}\label{Measurement}
In Shi \textit{et al.} (2016), we study the Balmer BAL quasar SDSS J125942.80+121312.6 (hereafter J1259+1213) in detail with the help of photoionization models. Many similarities can be found in the absorption features between J1125+0029 and J1259+1213. First, in both spectra we observe the overlapping troughs of low-ionization metal lines like UV \ion{Fe}{2} and \ion{Mg}{2} as well as the isolated Balmer lines and \ion{He}{1}*. Second, the UV \ion{Fe}{2} troughs between rest-frame 2400 and 2550 \AA, of which the majority are dominated by transitions from the terms with exciting energy $E_{\mathrm{ex}}>2.5\ \mathrm{eV}$, show relative depths of $\sim 0.5$. This implies a highly excited high column density \ion{Fe}{2} absorber, in which the resonant absorptions should be saturated, while large residual fluxes can be measured under the UV1 and UV2,3 multiplets from the ground term. Using the SDSS quasar composite (Vanden Berk \textit{et al.} 2001) intrinsically reddened by the SMC-type extinction curve (Gordon \textit{et al.} 2003) to match the fluxes in the absorption free windows around rest-frame 2100 \AA and longward of 3700 \AA, we find that the residuals under the UV1 and UV2,3 troughs are considerably larger than the UV \ion{Fe}{2} emission bump of the composite. Third, the H$\beta$ troughs are only slightly stronger than the H$\gamma$ troughs, given that the oscillator strength of H$\beta$ is more than twice as large as the oscillator strength of H$\gamma$.
In Shi \textit{et al.} (2016), we find that a high density and high column density gaseous medium can account for the BAL troughs of all observed ions in J1259+1213, including Balmer lines, \ion{He}{1}*, \ion{Mg}{2} and \ion{Fe}{2}. The absorbing medium covers part of the continuum source and little of the broad emission-line region (BELR), as the latter is two orders of magnitudes larger in size. Accordingly we also suppose that the low-ionization metal absorption lines in J1125+0029 are originated from the same medium as the redshifted Balmer lines and \ion{He}{1}* $\lambda 3889$, and this medium only obscures a fraction of the continuum source.
According to such assumption, to extract the normalized profile for the measurement of absorptions, the emission lines and unabsorbed continuum should be properly modeled. We would first remove the contribution of emission lines from the spectrum and then divide the residual by the model continuum. We fit the spectrum longward of rest-frame 3500 \AA following the steps described by Dong \textit{et al.} (2008) with small modification. The continuum is modeled using a single power-law continuum multiplied by the SMC-type extinction law (Gordon \textit{et al.} 2003). The continuum windows assumed nearly emission-free are rest-frame 3540 to 3560, 3810 to 3830, 4005 to 4035, 4150 to 4170 and 4550 to $4570\ \mathrm{\AA}$. The narrow emission lines such as [\ion{O}{2}] $\lambda 3728$ and [\ion{Ne}{3}] $\lambda 3868$ are modeled with a single Gaussian profile. The broad Balmer emissions H$\beta$ and H$\gamma$ are assumed to have the same redshift and profile, and modeled using three Gaussian profiles. Other weak broad emissions are modeled using one Gaussian profile. The fitting results are shown in Figure \ref{emis_model}. The measured value of full width at half maximum (FWHM) for the broad H$\beta$ emission in the SDSS-I/III spectrum is $7230\ \mathrm{km\ s}^{-1}$.
The emission model subtracted spectrum around H$\beta$, H$\gamma$, H$\delta$, H$\epsilon$, \ion{He}{1}* $\lambda 3889$ plus H$\zeta$, and H$\eta$ is normalized using model power-law continuum and the result for SDSS-I/II observation is plotted in Figure \ref{abs_prof}. Since the \ion{He}{1}* and Balmer absorption lines show similar profiles, we assume the \ion{He}{1}* and Balmer absorbers share the same kinematic structure and can fit two Gaussians to the normalized flux of all these lines simultaneously. The velocity shifts with respect to the QSO's rest frame are $72\pm 39$ and $651\pm 41\ \mathrm{km\ s}^{-1}$ for the two components respectively, including the uncertainty of systemic redshift. The FWHMs are $199.4\pm 16.4\ \mathrm{km\ s}^{-1}$ for the blue component and $398.6\pm 32.6\ \mathrm{km\ s}^{-1}$ for the red component. Spanning $\sim 1200\ \mathrm{km\ s}^{-1}$, the whole absorption line system can be classified as mini-BAL. Since the resolution of SDSS-I/II is $\sim 1800$, the $\mathrm{FWHM}_{\mathrm{inst}}$ for the instrumental profile is $\sim 167\ \mathrm{km\ s}^{-1}$. Employing the simple relation $\mathrm{FWHM}_{\mathrm{obs}}^{2}=\mathrm{FWHM}_{\mathrm{true}}^{2}+\mathrm{FWHM}_{\mathrm{inst}}^{2}$, the intrinsic $b$-values for two components are 64 and $215\ \mathrm{km\ s}^{-1}$, respectively.
The equivalent widths (EWs) and the $1\sigma$ uncertainties can be measured directly using the normalized fluxes and the fluctuations in the wavelength range defined by the Gaussian profiles. Since the absorbing medium only covers part of the continuum source, the apparent EWs are the reduced values of the true EWs by a factor of $C_f$, $\mathrm{EW_{app}}=\mathrm{EW_{true}}\times C_f$, where $C_f\leq 1$ is the covering factor. The values of covering factors and ionic column densities $N_{\mathrm{col}}(\mathrm{ion})$ can be derived using the COG analysis. In Figure \ref{COG}, for each component with the known $b$-value, the solid line shows the COG, while the dotted line shows the apparent EWs predicted by COG given the covering factors, $\log \mathrm{EW_{app}}/\lambda=\log \mathrm{EW_{true}}/\lambda+\log C_f$. For those measured absorption lines (represented using filled circles), the ordinated values show the measurements for apparent EWs, while the abscissa values, $N_{\mathrm{col}}(\mathrm{ion}) f \lambda_{\mathrm{rest}}$, would be determined by the ionic column densities $N_{\mathrm{col}}(\mathrm{ion})$. Appointing the unknown $C_f$, $N_{\mathrm{col}}(\mathrm{H}^{0}_{n=2})$, and $N_{\mathrm{col}}(\mathrm{He}^{0}_{2^3 \mathrm{S}})$ as adjustable parameters, we can get the optimal values and $1\sigma$ uncertainties for them by fitting the COG to the measured data points.
Fitting the measured EW values with the COG for the two spectra respectively, for the blue component we have $C_f=0.53\pm 0.18$, $\log N_{\mathrm{col}}(\mathrm{H}^{0}_{n=2})(\mathrm{cm}^{-2})=14.74\pm 0.24$, and $\log N_{\mathrm{col}}(\mathrm{He}^{0}_{2^3 \mathrm{S}})(\mathrm{cm}^{-2})=14.69\pm 0.34$ in the SDSS-I/II observation, and $C_f=0.73\pm 0.20$, $\log N_{\mathrm{col}}(\mathrm{H}^{0}_{n=2})(\mathrm{cm}^{-2})=14.39\pm 0.20$, and $\log N_{\mathrm{col}}(\mathrm{He}^{0}_{2^3 \mathrm{S}})(\mathrm{cm}^{-2})=14.59\pm 0.24$ in the BOSS observation. For the red component, $C_f=0.37\pm 0.23$, $\log N_{\mathrm{col}}(\mathrm{H}^{0}_{n=2})(\mathrm{cm}^{-2})=14.72\pm 0.31$, and $\log N_{\mathrm{col}}(\mathrm{He}^{0}_{2^3 \mathrm{S}})(\mathrm{cm}^{-2})=14.77\pm 0.30$ in the SDSS-I/II observation, and $C_f=0.22\pm 0.06$, $\log N_{\mathrm{col}}(\mathrm{H}^{0}_{n=2})(\mathrm{cm}^{-2})=14.92\pm 0.20$, and $\log N_{\mathrm{col}}(\mathrm{He}^{0}_{2^3 \mathrm{S}})(\mathrm{cm}^{-2})=15.11\pm 0.24$ in the BOSS observation (see Figure \ref{COG} panels (a) and (c)).
It can be found that for all parameters, the changes from the SDSS-I/II to the BOSS observations are not larger than about $1\sigma$ error. That means the absorption variability between the SDSS-I/II and BOSS observations is not significant. If we suppose the changes of the covering factors as the only cause of absorption variability (with the ionic column densities unchanged), a picture frequently prompted to explain the absorption variability (Hall \textit{et al.} 2002; Zhang \textit{et al.} 2015b), the covering factors are 0.55 and 0.30 for the blue and red components in the BOSS observation, respectively, with the reduced $\chi^2$ still being around 1 (see Figure \ref{COG} panels (b) and (d)). The uncertainties for $C_f$ are reduced to 0.14 and 0.07, respectively. It seems that the assumption that the physical conditions of the absorbing medium remain unchanged is acceptable.
\section{Photoionization Models for the Absorbing Medium}\label{Photoionization}
We use the photoionization code CLOUDY (version 10.00, last described by Ferland \textit{et al.} 1998) to simulate the ionization process, assuming a simple model of slab-shaped geometry, unique density and homogeneous chemical composition of solar values for the absorbing medium. The incident SED applied is the combination of a UV bump described as $\nu^{\alpha_{\mathrm{UV}}}\mathrm{exp}(-h\nu /kT_{\mathrm{BB}})\mathrm{exp}(-kT_{\mathrm{IR}}/h\nu)$ and power-law $a\nu^{\alpha_{\mathrm{X}}}$, incorporated in CLOUDY. This is considered typical for observed AGN continuum. The UV bump is parameterized by UV power-law index $\alpha_{\mathrm{UV}}=-0.5$, and exponentially cut off with temperature $T_{\mathrm{BB}}=1.5\times 10^5\ \mathrm{K}$ at high energy and $kT_{\mathrm{IR}}=0.01\ \mathrm{Ryd}$ at infrared. The power-law component has an index $\alpha_{\mathrm{X}}=-2$ beyond 100 keV, and $-1$ between 1.36 eV and 100 keV. The overall flux ratio of X-ray to optical is $\alpha_{\mathrm{OX}}=-1.4$. The physical conditions of the absorbing medium are characterized by the ionization parameter $U$ at the irradiated surface, the total hydrogen density $n(\mathrm{H})$, and the total hydrogen column density $N_{\mathrm{col}}(\mathrm{H})$ which indicates the thickness of the medium.
The \ion{He}{1}* $\lambda 3889$ absorption is originated from the metastable He$^0$ 2$^3$S level, which in the photoionization dominated medium is populated through the recombination of He$^+$ in the ionized zone. Ji \textit{et al.} (2015) presented a detailed investigation on the \ion{He}{1}* ionization structure using the photoionization simulations. They found that if the medium is thick enough that the ionizing front is well developed, the value of $U$ can solely determine the column density of \ion{He}{1}*. Given $\log N_{\mathrm{col}}(\mathrm{He}^{0}_{2^3 \mathrm{S}})(\mathrm{cm}^{-2})=14.69\pm 0.34$ and $14.77\pm 0.30$ for the blue and red components, we obtain $\log U=-1.9\pm 0.3$ and $-1.8\pm 0.3$ according to the Figure 10 in Ji \textit{et al.} (2015), respectively.
The hydrogen $n=2$ shell could be populated through a couple of mechanisms, including recombination, collisional excitation, and Ly$\alpha$ resonant scattering. Therefore, Balmer absorption lines can be originated in both ionized and neutral zones, showing more complicated dependence on the density and total column density of the medium as well as the strength of ionizing flux, while the density of $n(\mathrm{H})>10^6\ \mathrm{cm}^{-3}$ is generally required. In Figure \ref{model_distri} panels (a) and (b), we plot the column density of H$^0_{n=2}$ as functions of $n(\mathrm{H})$ and $N_{\mathrm{col}}(\mathrm{H})$ predicted by the photoionization simulations given the values of $U$, and the measured values are highlighted by the colored areas.
To further constrain the physical conditions of the absorbing medium, the measurements for other ions in the same medium are required, such as Fe$^+$. But unlike the case of J1259+1213 where the optical absorption troughs of Fe$^+$ $\lambda 4233$, $\lambda 4924$, $\lambda 5018$, and $\lambda 5169$ from excited Fe$^+$ are presented isolated which can be directly demonstrated associated with Balmer lines and reliably measured, in the spectra of J1125+0029 these lines can hardly be detected due to the poorer S/N. The UV \ion{Fe}{2} absorption troughs are so heavily saturated and blended that we cannot find isolated absorption lines to estimate the column density on any individual level of Fe$^+$. But being aware of the obvious similarity in appearance between these two objects, we guess in J1125+0029 the UV \ion{Fe}{2} troughs also come from the same absorbing gas as Balmer lines. That means each individual \ion{Fe}{2} trough would also has two components with the same central velocities, widths, and covering factors as Balmer lines since Fe$^+$ and H$^0_{n=2}$ are originated from the similar region in the photoionized medium.
Following the method described in Shi \textit{et al.} (2016), we can construct the synthetic model UV \ion{Fe}{2} absorption spectra based on the absorption profile extracted from Balmer lines and the column densities of Fe$^+$ on various levels predicted by the simulations. Since considerable contribution from Fe$^+$ excited levels with $E_{\mathrm{ex}}>2.5\ \mathrm{eV}$ can be identified in the overlapping UV \ion{Fe}{2} troughs, the full 371 levels Fe$^+$ model incorporated in CLOUDY is used in our simulations, including all levels up to 11.6 eV. Comparing the model spectra with the observation, we can use the UV \ion{Fe}{2} absorption features to further constrain the physical conditions of the absorbing medium.
In Figure \ref{synth_spec} top panel we plot an example of the synthetic model UV \ion{Fe}{2} absorption spectra. The best-fitting reddened SDSS quasar composite shown in Figure \ref{SED} is employed as the unabsorbed template. The reduced $\chi^2_{\nu}$ calculated in the range from rest-frame 2320 to 2780 \AA, where \ion{Fe}{2} contributes to nearly all the absorption features, is used to assess the agreement between the synthetic model spectra and the observation. With no further constraint, we find that the physical parameters leading to the minimum $\chi^2_{\nu}$ are $\log n(\mathrm{H})(\mathrm{cm}^{-3})\sim 9$ and $\log N_{\mathrm{col}}(\mathrm{H})(\mathrm{cm}^{-2})\sim 22$ for both components. Thus, it would be convenient and reasonable to suppose the physical states of the two components are identical. In Figure \ref{model_distri} panel (c), we plot the distribution of $\chi^{2}_{\nu}$ for the models of which the parameters for the blue and red components are the same.
The optimal values for the parameters $n(\mathrm{H})$ and $N_{\mathrm{col}}(\mathrm{H})$ are the values which present the measured value of $N_{\mathrm{col}}(\mathrm{H}^{0}_{n=2})$ and the minimum of $\chi^{2}_{\nu}$, while the $1\sigma$ uncertainty for these parameters are given by the area defined by the $1\sigma$ uncertainty of $N_{\mathrm{col}}(\mathrm{H}^{0}_{n=2})$ and the contour of $\chi^{2}_{\nu,\mathrm{min}}+1$. Thus we have $\log n(\mathrm{H})(\mathrm{cm}^{-3})=9\pm 0.3$ and $\log N_{\mathrm{col}}(\mathrm{H})(\mathrm{cm}^{-2})=21.9\pm 0.2$.
Other low-ionization ions, such as Ti$^+$, Cr$^+$, and Ni$^+$, can also be included in the synthetic model spectra following the same method as Fe$^+$. In Figure \ref{synth_spec} we also plot this complete model for the \ion{Fe}{2} selected optimal models. The inclusion of the absorptions from these ions makes the model better match the observation, with the only exception of \ion{Mg}{2}. The model \ion{Mg}{2} absorption is much shallower than observation. Since we suppose that the absorbing medium covers part of the accretion disk, the fluxes under \ion{Mg}{2} troughs should consist of the fluxes of \ion{Mg}{2} emission and unobscured part of continuum. But the observed fluxes under \ion{Mg}{2} troughs are smaller than \ion{Mg}{2} emission peak in the quasar composite. Such deviation could be ascribed to the difference between the \ion{Mg}{2} emission in the composite and the true \ion{Mg}{2} emission in our source, because there is considerable object to object variation in quasar emission lines. If we suppose the \ion{Mg}{2} emission in J1125+0029 is much weaker than that in the composite, the disagreement can be reduced. Leighly \textit{et al.} (2011) used a dozen of real non-absorption quasars to match the NIR spectrum of FBQS J1151+3822, and chose the best matched one as template to measure the \ion{He}{1}* $\lambda 10830$. Zhang \textit{et al.} (2014) and Liu \textit{et al.} (2015) also developed a similar pair-matching method to improve the estimate of unabsorbed level. We search all SDSS DR7 non-BAL quasar spectra with mean S/N per pixel greater than 15 for the suitable template. The spectrum of SDSS J142923.92+024023.1 showing very weak \ion{Mg}{2} emission best meets our request if it is reddened with $E(B-V)=0.045$. The synthetic spectrum is plotted in the bottom panel of Figure \ref{synth_spec}. Compared with the synthetic spectrum based on the reddened composite, the result is improved not only for the \ion{Mg}{2} doublets but also for the UV \ion{Fe}{2} from rest-frame 2320 to 2640 \AA.
\section{Discussion}\label{Discussion}
The mass of the central SMBH can be estimated according to the relation in Wang \textit{et al.} (2009), assuming the broad emission line region (BELR) is virialized, $\log (M_{\mathrm{BH}}/10^{6}M_{\sun})=(1.39\pm 0.14)+0.5\log (\lambda L_{5100}/10^{44}\ \mathrm{erg\ s}^{-1})+(1.09\pm 0.23)\log (\mathrm{FWHM}(\mathrm{H}\beta)/1000\ \mathrm{km\ s}^{-1})$, where $L_{5100}$ is the $\lambda L_{\lambda}$ at rest-frame 5100 \AA. Thus, we have $M_{\mathrm{BH}}=1.5\times 10^{9}\ M_{\sun}$ with an uncertainty of a factor of $\sim 2.5$ (from the intrinsic scatter $\sim 0.4$ dex for this single-epoch method compared with the results of reverberation mapping, Ho \& Kim 2015). And adopting the correction factor by Runnoe \textit{et al.} (2012), the bolometric luminosity is $L_{\mathrm{bol}}=(8.1\pm 0.4)\times L_{5100}=3.9\pm 0.2\times 10^{46}\ \mathrm{erg\ s}^{-1}$. Assuming an accretion efficiency of 0.1, the mass accretion rate is then $\dot{M}_{\mathrm{BH}}\approx 6.8\ M_{\sun}\ \mathrm{yr}^{-1}$. Extrapolating the best-fitting SDSS composite presented in Figure \ref{SED}, we can derive the continuum flux at 1215.67 \AA. Then given the model incident SED used in simulations and the systemic redshift, the pre-extinction monochromatic luminosity of ionizing continuum at Lyman limit can be roughly estimated. For the SDSS-I/II spectrum, this value is $L_{\nu}(912)=5.0\times 10^{30}\ \mathrm{erg\ s}^{-1}\mathrm{Hz}^{-1}$.
Given the luminosity and SED of the ionizing continuum, we can derive the distance of the absorbing medium to the central engine according to the physical conditions constrained by the photoionization models, as $\frac{L(<912)}{4\pi r_{\mathrm{abs}}^2}=Un(\mathrm{H})c\overline{E_{\mathrm{ph}}(<912)}$. $L(<912)$ is the ionizing luminosity of the continuum source, determined by $L_{\nu}(912)$ and the incident SED used in simulation models. And $\overline{E_{\mathrm{ph}}(<912)}$ is the average energy for all ionizing photons, which also can be evaluated according to the model SED. With $\log U=-1.8\pm 0.3$ and $\log n(\mathrm{H})(\mathrm{cm}^{-3})=9\pm 0.3$ from the optimal photoionization models, the distance of the inner surface of the absorbing medium is $r_{\mathrm{abs}}=4^{+6.6}_{-2.5} \mathrm{pc}$. The listed uncertainty only includes the uncertainties of $U$ and $n(\mathrm{H})$. The uncertain of the AGN ionizing luminosity introduced due to the extrapolation of power-law continuum is more difficult to assess. A change of 100\% for the luminosity can lead to a change of 41\% for the distance, making it a relative minor factor.
In their first paper reported the redshifted absorption line systems in quasars' spectra including J1125+0029, Hall \textit{et al.} (2002) suggested a rotation-dominated disk wind at the phase when the outflow just rises from the accretion disk to explain the redshifted troughs. With the release of BOSS spectra, Hall \textit{et al.} (2013) returned to the issue. Comparing the BOSS and SDSS-I/II spectra of J1125+0029, especially the Mg II troughs, they suggested that the blueshifted part of the Mg II absorption weakened more than the redshifted part, to account for the variability (for more details, see Appendix). The different behavior in the blueshifted and redshifted absorption is believed consistent with the picture of a rotational wind when the cloud moves from the approaching side to the receding side. Such wind is suggested located $1255\ R_{\mathrm{Sch}}$ ($5.5\times 10^{17}\ \mathrm{km\ s}^{-1}$ with an uncertainty of a factor of 2.5) from the SMBH (Murray \& Chiang 1998; Elvis 2000). Another explanation for the redshifted troughs mentioned in Hall \textit{et al.} (2013), the gravitational redshift, requires the absorbing medium to be located at even smaller radii, $\sim 100\ R_{\mathrm{Sch}}$.
The absorbing medium described by our photoionization models seems much more distant than that implied in these explanations. Furthermore, according to the COG analysis in \S\ref{Measurement}, we can estimate the transverse velocities of the absorbing medium by the variations of covering factors for Balmer lines between the SDSS-I/II and BOSS observations. In the simplest picture, these changes ($\Delta C_f=0.02$ for the blue component and 0.07 for the red component) stand for a continuous movement of the medium across our LOS. The radius of the accretion disk where the radiation peaks at 4863 \AA can be evaluated from $\sigma T_{\mathrm{eff}}^{4}=\frac{3GM\dot{M}}{8\pi R^{3}}f(R,a)$ (Eq. 2 in Collin \textit{et al.} 2002). With $k_{\mathrm{B}}T_{\mathrm{eff}}=h\nu$ and the boundary condition $f(R,a)\approx 1$, we obtain $R(4863)\approx 6.1\pm 2.1\times 10^{15}\ \mathrm{cm}$. Considering the uncertainty of $C_f$, although for the blue component the probability that the transverse velocity is smaller than the radial velocity is only 0.56, for the red component the upper limit of transverse velocity at a confidence level of 0.98 is $145\ \mathrm{km\ s}^{-1}$ which is much smaller than the corresponding redshifted velocity. The infall seems a more reasonable explanation.
The distance of the infalling medium is also larger than the radius of H$\beta$ BELR by a factor of $\sim 10$, but how about the torus? The torus has long been suggested as the reservoir of interstellar medium (ISM) feeding the central engine and the direct source of the accretion inflow (Krolik \& Begelman 1988). In recent works, a clumpy model for the torus is required to reproduce the IR SED especially the lack of $10\ \micron$ Si emission feature (Nenkova \textit{et al.} 2002). Such model is supported by Vollmer \textit{et al.} (2004) and Beckert \textit{et al.} (2004) who linked the clumpy torus in AGN with the circumnuclear disk (CND) surrounding the central black hole of our Galaxy which consist of several hundred clouds of gas and dust. The transfer from the thin CND to the thick obscuring torus depends on the accretion rate. The radial accretion flow is now naturally regarded as the result of cloud-cloud collision through which these clouds lost energy and then fall inward. If the falling clouds are located a bit faraway from the equatorial plane, they could be observed in the foreground of the accretion disk since at such view angles the disk would not be severely obscured due to the the relative low local filling factor.
The inner radius of the torus used to be approximated as the evaporation radius, $R_{\mathrm{evap}}=1.3L_{\mathrm{UV,46}}^{1/2}T_{1500}^{-2.8}\ \mathrm{pc}\approx 1.3\ \mathrm{pc}$ following Barvainis 1987 for our object, where $L_{\mathrm{UV,46}}$ is the UV luminosity in unit of $10^{46}\ \mathrm{erg\ s}^{-1}$ estimated using $\lambda L_{\lambda}(1450)$, and $T_{1500}$ is the grain evaporation temperature in unit of 1500 K which is $\sim 1$. By comparing $R_{\mathrm{evap}}$ with the results of reverberation mapping, Kishimoto \textit{et al.} (2007) suggested the inner radius was overestimated in this way by a factor of $\sim 3$. Kawaguchi \& Mori (2011) argued the inner radius should increase with the view angle of the torus due to the anisotropic illumination of the disk, to explain the intrinsic scatter in reverberation mapping (Reminding in J1125+0029, the SED is intermediately reddened by the dust associated with the object, implying that the LOS is not very close to the equatorial plane of the torus). Infrared interferometry, another direct radius measurement, presents the results a factor of 2 larger than those from reverberation mapping (Koshida \textit{et al.} 2014). Thus to date the uncertainty of the inner radius estimate is also quite large, comparable to the uncertainty of our photoionization models. However we can still conclude that the absorbing medium is not far away from the inner surface of the torus.
The infalling absorbing medium is estimated close to the inner surface of the torus, and the density of the medium is found similar to the gas density of the torus. Therefore we strongly suppose the redshifted absorption line system in J1125+0029 representing the accretion flow originated from the torus as our LOS toward the continuum source lying through the low-density part of the clumpy torus. The relation between the two components of the absorbing medium remains unclear. Since the physical conditions and the distance of both components are almost the same, they might reflect the approaching and receding parts of a spinning cloud in the accretion flow with the centroidal infalling velocity $v_{\mathrm{infall}}\approx 350\ \mathrm{km\ s}^{-1}$.
The mass of the absorbing medium is $M_{\mathrm{abs}}=\mu m_{\mathrm{p}}N_{\mathrm{col}}(\mathrm{H})S$, where $\mu$ is the mean atomic mass per proton, $m_{\mathrm{p}}$ is the mass of proton, and $S$ is the projection area. Assuming the observed absorbing medium is typical of all accretion clouds and these clouds are uniformly distributed at the inner surface of torus, we can estimate the mass inflow rate as $\dot{M}_{\mathrm{inflow}}\approx \pi R_{\mathrm{evap}}^2 C_{f,\mathrm{torus}} \mu m_{\mathrm{p}}N_{\mathrm{col}}(\mathrm{H})/t_{\mathrm{infall}}$, where $C_{f,\mathrm{torus}}\approx 0.6$ is the global covering factor of the torus (Lawrence \& Elvis 2010), $\pi R_{\mathrm{evap}}^2 C_{f,\mathrm{torus}}\approx \sum_i S_i$, and $t_{\mathrm{infall}}$ is the infalling timescale of the accretion clouds. If we approximate that $t_{\mathrm{infall}}=R_{\mathrm{evap}}/v_{\mathrm{infall}}$, $\dot{M}_{\mathrm{inflow}}\approx 0.43\ M_{\odot}\ \mathrm{yr}^{-1}$, too small compared with $\dot{M}_{BH}$. Reminding that $v_{\mathrm{infall}}$ is the radial velocity at the initial stage of infall, we think $t_{\mathrm{infall}}$ is overestimated in this way. Since the accretion inflow lies close to the equatorial plane, the radiation from the central engine would be obscured which reduces the radiation pressure, we can use the free fall timescale $t_{\mathrm{ff}}=(R_{\mathrm{evap}}^3/GM_{\mathrm{BH}})^{1/2}$ as $t_{\mathrm{infall}}$. Thus, $\dot{M}_{\mathrm{inflow}}\approx 2.8\ M_{\odot}$ is of the same magnitude as $\dot{M}_{BH}$. And in reality, the properties of clouds in the inflow would be dependent on the height. In J1125+0029, our LOS maybe only passes trough the outskirt of the torus. The closer to the equatorial plane the clouds are, the denser they would be. Thus, the mass inflow rate could be even higher.
\section{Summary}\label{Summary}
The redshifted absorption line systems are rarely detected features in the quasar spectra. A couple of theoretical pictures were prompted to explain the phenomena, but no decisive conclusion has been achieved, unless the physical conditions and the spacial structure of the absorbing medium are available. In the SDSS-I/II and BOSS spectra of quasar J1125+0029, a redshifted absorption line system is identified including lines of hydrogen Balmer series, metastable \ion{He}{1}, \ion{Mg}{2}, \ion{Fe}{2}, \textit{et al.}. These lines are powerful diagnostics to the physical properties of the absorbing medium. Performing a careful measurement of the ionic column densities and covering factors for Balmer and \ion{He}{1}* lines and using the photoionization simulations, we find that the medium is located $\sim 4\ \mathrm{pc}$ away from the central engine and its motion is dominated by infall. Since the distance is consistent with the radius of the inner surface of torus and the physical conditions of the medium are also similar to the torus, we suggest this absorption line system as a candidate for the accretion inflow originated from the torus which fuel the SMBH.
\acknowledgments
This work is supported by the National Basic Research Program of China (the 973 Program 2013CB834905) and the National Natural Science Foundation of China (NSFC-11421303 and 11473025). This research uses data obtained through the Telescope Access Program (TAP), which has been funded by the Strategic Priority Research Program The Emergence of Cosmological Structures (Grant No. XDB09000000), the National Astronomical Observatories, the Chinese Academy of Sciences, and the Special Fund for Astronomy from the Ministry of Finance.
|
\section{Introduction}
\label{intro}
The `box' $A \Box B$ of events $A$ and $B$, consisting of the set of outcomes where $A$ and $B$ `occur for disjoint reasons,' was first considered in \cite{vdBK} on the space $S = \{0,1\}^n$ with the uniform probability measure. More generally, on the product
$S=\prod_{i=1}^n S_i$ of general spaces $S_i,i=1,\ldots,n$,
with $[n]=\{1,\ldots,n\}$ and for $K \subseteq [n]$, define the $K$ cylinder of ${\bf x} \in S$ as
\begin{eqnarray*}
[{\bf x}]_K=\{{\bf y} \in S: y_i=x_i\,\, \forall i \in K\}.
\end{eqnarray*}
The box operation is given by
\begin{eqnarray} \label{def:classical_box}
A \Box B =\{{\bf x} \in S: \mbox{there exist disjoint $K,L \subseteq [n]$ such that $[{\bf x}]_K \subseteq A, [{\bf x}]_L \subseteq B$} \}.
\end{eqnarray}
An equivalent definition of the box operation in \eqref{def:classical_box} in terms of
cylinders $[A]_K$ of subsets $A$ of $S$, appearing as (2) in \cite{AGH} is
\begin{eqnarray} \label{def:AGH.cyl.box1}
A \Box B = \bigcup_{0 \le |K| \le n} [A]_K \cap [B]_{K^c}\quad \mbox{where} \quad[A]_K=\{{\bf x }: [{\bf x}]_K \subseteq A\},
\end{eqnarray}
and $K^c$ denotes the complement of $K$.
The set $A \Box B$ represents the event that $A$ and $B$ occur for disjoint reasons, as for any ${\bf x} \in A \Box B$, the event $A$ can be verified to occur based only on the coordinates of ${\bf x}$ in $K$, regardless of the values of the remaining coordinates, and similarly so for coordinates in $L$ for $B$. Clearly $A \Box B \subseteq A \cap B$.
Note that if $K \subseteq M \subseteq [n]$ then $[{\bf x}]_M \subseteq [{\bf x}]_K$. This implies readily that in \eqref{def:classical_box} we can assume without loss of generality that $L=K^c$.
Note also that the choice of $K$ and $L$ in \eqref{def:classical_box} may depend on ${\bf x}$, and we may write $K({\bf x})$ and $L({\bf x})$ for specific choices.
The famous {BKR} inequality, conjectured in \cite{vdBK} and proved in \cite{Rei} for the uniform probability measure on $\{0,1\}^n$, states that
\begin{eqnarray} \label{eq:box.inequality}
P(A \Box B) \le P(A)P(B).
\end{eqnarray}
The validity of \eqref{eq:box.inequality} for the uniform measure on $\{0,1\}^n$ was shown to imply the inequality on a finite product of discrete probability spaces in \cite{vdBF}.
{Henceforth we denote the end of an example by $\blacktriangle$.}
\begin{example} [A simple example: bond percolation.]\label{ex:simpleperc} Consider the following graph:
\centerline{\begin{tikzpicture}
\draw [thick](0,0) --(1,0) -- (2,0);
\draw [thick](0,1) --(1,1) -- (2,1);
\draw [thick](0,0) --(0,1);
\draw [thick](1,0) --(1,1);
\draw [thick](2,0) --(2,1);
\node at (-.08,1.08) {a};
\node at (2.08,1.08) {c};
\node at (-.08,-.08) {b};
\node at (2.08,-.08) {d};
\end{tikzpicture}}
\noindent Edges are open or closed independently with probability $1/2$.
Since there are 7 edges we may let $S=\{0,1\}^7$, and numbering the edges arbitrarily, the coordinates $x_i$ of ${\bf x}=(x_1,\ldots,x_7) \in S$ indicate whether edge $i$ is open or closed. Let $A=\{$there exists a path of open edges from $a$ to $c$$\}$.
A direct count of the events that corresponds to $A$ shows that under the uniform distribution, $P(A)=44/2^7$. For $B=\{$there exists a path of open edges from $b$ to $d$$\}$ it is easy to see that $P(A \Box B)=2^3/2^7=1/16$ as $A \Box B$ is the set of all paths for which all bonds on the top and bottom rows of the graph are open.
As $P(A)P(B)=({44/2^7})^2$,
in this case `the excess multiple' in the bound \eqref{eq:box.inequality} is given by
\begin{eqnarray} \label{eq:excess.multiple}
\frac{P(A)P(B)}{P(A \Box B)}= \frac{({44}/{2^7})^2}{1/16} \approx 1.9.
\end{eqnarray}
\hfill \ensuremath{\blacktriangle}
\end{example}
A version of the BKR inequality for a finite product of arbitrary probability spaces, including finite spaces, and discrete or non-discrete infinite spaces, was considered in \cite{GR}. The
case where $A$ and $B$ are subsets of finite or countable products of ${\mathbb R}$, and more generally of Polish spaces, was considered in \cite{AGH}. This later work raises important issues
regarding the measurability of $A \Box B$. In particular it is shown that the BKR combination $A \Box B$ of Lebesgue measurable sets need not be Lebesgue measurable.
For an interesting practical application in a non discrete case see \cite{AGMS}.
We now turn to the consideration of $A \Box B$ in the framework of \cite{GR} and its slightly corrected version \cite{GRa}, on a finite product
$S=\prod_{i=1}^n S_i$ of general spaces $S_i,i=1,\ldots,n$ endowed with the product measure $P=\prod_{i=1}^{n}P_i$ on the product sigma algebra of $S$ generated by given sigma algebras on $S_i$.
The question of the measurability of $A \Box B$ did not arise in \cite{GR} since, as discussed in \cite{GRa}, the set $A \small \framebox[1.45\width]{$11$} B$, which expresses the event that $A$ and $B$ occur for disjoint reasons almost surely, defined formally in \eqref{eq:specilize.function.version}, was considered instead. The results of \cite{GR} show that the combination $A \small \framebox[1.45\width]{$11$} B$ is measurable whenever $A$ and $B$ are, and that
\eqref{was.5} and \eqref{eq:oovsprod} below hold, that is
\begin{eqnarray}
\label{eq:allwekno}
A \Box B \subseteq A \small \framebox[1.45\width]{$11$} B \quad \mbox{and} \quad P(A \small \framebox[1.45\width]{$11$} B) \le P(A)P(B).
\end{eqnarray}
Further, in \cite{GR} it is shown that the latter inequality also holds when $P$ is the completion of any product measure. The proof in \cite{GR} of the inequality in \eqref{eq:allwekno} relies on Reimer's proof in \cite{Rei}.
The implications of \eqref{eq:allwekno} are seen to be stronger than the BKR inequality \eqref{eq:box.inequality}. In particular, if $A \Box B$ is measurable, then it follows from \eqref{eq:allwekno} that $P(A \Box B) \le P(A)P(B)$.
Lemma 1 in \cite{AGH} shows that if $A$ and $B$ are Borel subsets of $[0, 1]^n$ then $A \Box B$ is Lebesgue measurable. As stated in Section 8 of \cite{AGH}, the same proof holds on the product of Polish subspaces of $\mathbb{R}$ equipped with the completion of any Borel product probability. Proposition \ref{prop:partof4} shows that the qualifier `of $\mathbb{R}$' can be dropped.
\begin{proposition}\label{prop:partof4}
Let $S$ be a finite product of Polish spaces endowed with the product Borel sigma algebra and a product Borel probability measure $P$, and let $A$ and $B$ be Borel subsets of S.
Then $A \Box B$ is measurable with respect to the completion $\overline{P}$ of $P$, and
$\overline{P}(A \Box B) \le P(A)P(B)$.
\end{proposition}
The measurability claim follows from definition \eqref{def:AGH.cyl.box1} and the result, shown in Section \ref{sec:meas}, the Appendix on measurability, that $[A]_K$ is measurable with respect to the completed measure under the conditions of the Proposition; the inequality then follows by \eqref{eq:allwekno}. Here we have used common terminology where measurability of a set with respect to the completion of $P$ means measurability with respect to the completion of the relevant sigma algebra with respect to $P$.
In Section \ref{sec:00box} we review the definition and meaning of the $\small \framebox[1.45\width]{$11$}$ operation and provide some examples.
In Section \ref{sec:cyl} we look for an improvement of \eqref{eq:allwekno}, where for sets $A$ and $B$ we define in \eqref{def:overset2T} minimal sets ${A}^{(\tiny{0})}$ and ${B}^{(\tiny{0})}$, `cylindrical cores',
satisfying ${A}^{(\tiny{0})} \subseteq A, \, {B}^{(\tiny{0})} \subseteq B$, and $A \Box B = {A}^{(\tiny{0})} \Box {B}^{(\tiny{0})}$.
Measurability of ${A}^{(\tiny{0})}$ under the conditions of Proposition \ref{prop:partof4} is shown in Proposition \ref{prop:LebegT}. The main purpose of this paper is achieved in Section \ref{sec:lenient}, which extends the set operation in \cite{GR} that captures the event that $A$ and $B$ occur for disjoint reasons almost surely, to one where $A$ and $B$ occur for disjoint reasons with at least some prescribed probabilities, and to provide a bound of BKR type. Examples are given in Section \ref{sec:lenexam}, some brief history on the BKR inequality in continuous spaces in Section \ref{sec:GuRa}, and various technical
measurability issues are relegated to Section \ref{sec:meas}.
\section{The 11 Box Operation}\label{sec:00box}
With the notation given further meaning later, we {define} the box type operation $A \small \framebox[1.45\width]{$11$} B$ on events $A$ and $B$. Apart from the notation $A \small \framebox[1.45\width]{$11$} B$, these definitions appear in \cite{GR} and \cite{GRa}. Unlike $A \Box B$, the operation $A \small \framebox[1.45\width]{$11$} B$ depends on the measure $P$, taking into account probability zero events in a natural probabilistic way. The dependence on $P$ will be suppressed in the notation, here and in related definitions such as \eqref{def:stbox}.
Throughout this work, for each $i=1,\ldots,n$, let $(S_i,\mathbb{S}_i,P_i)$ be a probability
space, and set $S=\prod_{i=1}^n S_i$,
$\mathbb{S}=\bigotimes_{i=1}^n \mathbb{S}_i$, $P=\prod_{i=1}^n P_i$ to be the product space, product sigma algebra, and product probability, respectively. For $K \subseteq [n]$ we set $S_K=\prod_{i \in K}^n S_i$, \, $\mathbb{S}_K=\bigotimes_{i \in K} \mathbb{S}_i$ and $P_K =\prod_{i \in K} P_i$. For $A$ and $B$ in $\mathbb{S}$ define $A \Box B$ as in \eqref{def:classical_box} and note that it is shown in \cite{AGH} that the latter set may not be measurable. However, as a set it is well defined under the present general setting.
As in \cite{GR}, we consider real valued measurable functions on $S$, that is, functions
that are $(\mathbb{S},\mathbb{B})$ measurable, where $\mathbb{B}$
denotes the Borel sigma algebra of the real line ${\mathbb R}$. Let
$f$ and $g$ be two such functions, and for $K$ and
$L$ subsets of $[n]$, let
\begin{eqnarray} \label{eq:function.version}
\underline{f}_K({\bf x}) = {\rm ess} \inf_{{\bf y}
\in [{\bf x}]_K}f({\bf y}), \quad \mbox{and} \quad
\underline{g}_L({\bf x}) = {\rm ess} \inf_{{\bf y} \in [{\bf
x}]_L}g({\bf y}),
\end{eqnarray}
where the essential infimums defining $\underline{f}_K({\bf x})$ and $\underline{g}_L({\bf x})$ are taken with respect to the product probability measure on the coordinates in the complements $K^c$ and $L^c$ respectively.
{See the Appendix of \cite{GR} where it is shown that $\underline{f}$ and $\underline{g}$ of \eqref{eq:function.version} are measurable, and also extensions for handling completions.}
Below we use $\max_{K \cap L = \emptyset}$ to indicate maximum over all pairs of disjoint sets $K,L \subseteq [n]$.
In Theorem \ref{thm1} below we consider random vectors ${\bf X}=(X_1,\ldots,X_n)$ where each $X_i$ is a measurable function from some probability space to the measurable
spaces $(S_i,\mathbb{S}_i)$, so that $X_i \in S_i$, and for sets $B_i \in \mathbb{S}_i$ we have $P_i(B_i)=P(X_i \in B_i)$. We assume that ${\bf X} = (X_1,\ldots,X_n)$ is distributed according to the product probability $P$, so that the components $X_i$ are independent.
Clearly, in moving to this framework no generality over that obtained by working on the basic probability space is lost. Moreover, function formulations such as in the theorem allow for statements that are more general than relations between the probabilities of sets; see \cite{GR} for some examples.
\begin{theorem}[{Goldstein and Rinott, 2007}]\label{thm1}
Let ${\bf X} = (X_1,\ldots,X_n) \in S$ be a random
vector with a product distribution $P$. Then
\begin{eqnarray} \label{eq:finess} E\left\{\max_{K \cap L = \emptyset}\underline{f}_K({\bf X})\underline{g}_L({\bf
X})\right\} \leq E\left\{f({\bf X})\right\}\,E\left\{g({\bf
X})\right\}. \end{eqnarray}
\end{theorem}
We emphasize that Theorem \ref{thm1} applies to functions of a vector ${\bf X}$ having independent coordinates taking values in any measure
spaces.
Specializing {\eqref{eq:function.version} and} \eqref{eq:finess} to the case where $f$ and $g$ are the indicator functions ${\bf 1}_A({\bf x})$ and ${\bf 1}_B({\bf x})$ of $A$ and $B$ respectively,
we define $A \small \framebox[1.45\width]{$11$} B$ by its indicator
\begin{eqnarray} \label{eq:specilize.function.version}
{\bf 1}_{A \small \framebox[1.45\width]{$11$} B}({\bf x}) = \max_{K \cap L = \emptyset}
\underline{{\bf 1}_A}_K({\bf x}) \underline{{\bf 1}_B}_L({\bf x}),\end{eqnarray}
Theorem \ref{thm1} now yields
\begin{eqnarray} \label{eq:oovsprod} P(A \small \framebox[1.45\width]{$11$} B) \le P(A)P(B).
\end{eqnarray}
We mentioned earlier that in \eqref{def:classical_box} we can assume without loss of generality that $L=K^c$. We claim that this is true also for \eqref{eq:specilize.function.version}, where again we can replace $K \cap L = \emptyset$ by $L=K^c$. Clearly, such a replacement restricts the collection of sets over which the maximum is taken, so the quantity so obtained lower bounds the right hand side of \eqref{eq:specilize.function.version}. On the other hand, for any disjoint $K$ and $L$ we have $L \subseteq K^c$, which in turn implies $\underline{{\bf 1}_B}_L({\bf x}) \le \underline{{\bf 1}_B}_{K^c}({\bf x})$, as the infimum on the right is over a smaller set, see \eqref{eq:function.version}, implying
\begin{eqnarray*}
\underline{{\bf 1}_A}_K({\bf x}) \underline{{\bf 1}_B}_L({\bf x}) \le \underline{{\bf 1}_A}_K({\bf x}) \underline{{\bf 1}_B}_{K^c}({\bf x}).
\end{eqnarray*}
In particular if the right-hand side of \eqref{eq:specilize.function.version} takes the value 1, then for some disjoint $K$ and $L$ we have $\underline{{\bf 1}_A}_K({\bf x}) \underline{{\bf 1}_B}_L({\bf x})=1$ which implies $\underline{{\bf 1}_A}_K({\bf x}) \underline{{\bf 1}_B}_{K^c}({\bf x})=1$, so the replacement also upper bounds the right hand side of \eqref{eq:specilize.function.version}, and our claim follows.
We now consider definition \eqref{def:classical_box} of $A \Box B$ in general spaces, which, as noted above, is a set operation that does not depend on the underlying measure, in contrast to $A \small \framebox[1.45\width]{$11$} B$.
As shown in (5) of \cite{GRa}, we have the inclusion
\begin{eqnarray} \label{was.5}
{A \Box B \subseteq A \small \framebox[1.45\width]{$11$} B}.
\end{eqnarray}
Now \eqref{eq:oovsprod} and \eqref{was.5} imply \eqref{eq:box.inequality} in the general setting above provided $A \Box B$ is measurable. To see \eqref{was.5}, note that replacing essential infimum by infimum in \eqref{eq:function.version}, inequality \eqref{was.5} becomes equality. Hence \eqref{was.5} holds as stated because the essential infimum is at least as large as the infimum. In words, elements of $A \Box B$
demand that $A$ and $B$ hold for all outcomes satisfying disjoint reasons for their occurrence, while $A \small \framebox[1.45\width]{$11$} B$ requires only that the `reasons' imply that $A$ and $B$ hold almost surely.
To develop some intuition on the $\small \framebox[1.45\width]{$11$}$ box type operation, consider two individuals, one of whom desires that event $A$ happens, while the other desires $B$; we label the individuals according to their desired event with little danger of confusion. If ${\bf x} \in A \Box B$ then disjoint sets of coordinates of ${\bf x}$ may be revealed to $A$ and $B$ so that each will know with certainty that their event has occurred. Definition \eqref{eq:specilize.function.version} extends the box operation to accommodate situations where the disjoint sets of coordinates of $\bf x$ revealed to $A$ and $B$ indicate only that their event has occurred almost surely, that is, with probability one, but which may not certify it absolutely. It is easy to see that unlike $A \Box B$, the set $A \small \framebox[1.45\width]{$11$} B$ is not necessarily contained in $A \cap B$; an example that follows gives one such instance.
It is natural to ask what might be the differences, and possible advantages, of defining the box operation $\small \framebox[1.45\width]{$11$}$ as in \eqref{eq:specilize.function.version}, and more generally for functions as in \eqref{eq:function.version} as {considered} in Theorem \ref{thm1}, where the underlying measure is involved, in contrast to the more `straightforward' extension that preserves \eqref{def:classical_box} as a pure set operation.
For one, the approach taken in Theorem \ref{thm1} avoids certain subtle measurability difficulties that arise in the `straightforward' approach. In particular, Example 2 in \cite{AGH} presents a situation where the classical box of two Lebesgue measurable sets fails to be Lebesgue measurable.
In contrast, {as shown in \cite{GR}}, the function operation as defined in \eqref{eq:function.version} preserves measurability.
Technical matters aside, the approach of Theorem \ref{thm1} also has advantages from a probabilistic perspective, as illustrated by the following simple example, somewhat related to Example 2 in \cite{AGH}.
Let
\begin{eqnarray} \label{eq:soda1}
S = [0,1]^2 \qmq{and}
A=B = \{(x_1,x_2) \in [0,1]^2: x_1 \not = x_2\}.
\end{eqnarray}
It is easy to see that $A \Box B$ as defined in \eqref{def:classical_box} is empty. For instance, for any ${{\bf x}} \in [0,1]^2$, the cylinder $[{{\bf x}}]_{\{1\}}$ is not contained in $A$ (only) because it contains the point $(x_1,x_1)$, which is not in $A$. However, for any continuous measure on $S$, this single point is a set of measure zero, with the result that $A \small \framebox[1.45\width]{$11$} B = [0,1]^2$.
Here is an intuitive way to think about this example: suppose Alice and Bob each need to choose a single instant of time in the interval $[0,1]$ required for completion of a certain task. Assume there is only a single resource, and both will fail if and only if they require the resource at the same precise instant. To succeed for disjoint reasons as in \eqref{def:classical_box} means that each one can be sure of their own success independently of the other's choice of time. That here $A \Box B = \emptyset$ is a reflection of the fact that this is impossible. However, if they select times independently
by any continuous distribution
in the unit interval, then each one of them will succeed with probability one for any choice of the other. And indeed, for this case $A \small \framebox[1.45\width]{$11$} B = [0,1]^2$. This example, though simple, illustrates how sets of measure zero, which typically are a technical nuisance that can effectively be ignored, can greatly affect box type operations, here in particular causing the maximum possible discrepancy in the inequality
\begin{eqnarray*}
0=P(\emptyset)=P(A \Box B) < P(A \small \framebox[1.45\width]{$11$} B)=1.
\end{eqnarray*}
\begin{example}[Continuous percolation]\label{ex:perc}
Consider the random geometric graph ${\cal G}$ whose vertex set is a
collection of $n$ independent points ${{\bf X}}={\{}X_1, \ldots, X_n{\}}$, each sampled from a distribution having a strictly positive density on $[0,1]^2$. An edge is drawn between distinct points $X_i$ and $X_j$ if $||X_i-X_j||\le 2r$, where $r>0$ and $|| \cdot ||$ is any norm in $\mathbb{R}^2$. An equivalent way to view the graph is to connect two distinct points when the two circles of radius $r$ having these points as their centers intersect.
For percolation models based on randomly chosen balls see, for example, \cite{MP} and \cite{MR} and \cite{LPZ}.
Let $A$ be the event that there exists a `path along edges from left to right of $[0,1]^2$, that is, a path starting from a point that is within distance of at most $r$ from the $y$-axis, and ending at a point within distance at most $r$ from the line $x=1$. Similarly, let $B$ denote the event that there exists a path from bottom to top.
It is easy to see that for $n$ large, with positive probability, there exist realizations that allow crossing from left to right and from bottom to top on disjoint paths, that is, such that the two paths do not share a vertex.
In this case the event that $A$ and $B$ occur for disjoint reasons, which coincides with the event $A \Box B$, is non-empty. These events can be seen as continuous versions of those studied in \cite{BK}.\hfill \ensuremath{\blacktriangle}
\end{example}
\begin{example}[Continuous percolation with annihilation]\label{ex:percann}
We next introduce annihilation to the percolation model on the graph ${\cal G}$ of Example \ref{ex:perc}, and return to this setting again in Section \ref{sec:lenexam}.
If for some $i \not =j$ the points $X_i$ and $X_j$ coincide, that is, are equal, then any edge attached to these points is annihilated. The coinciding points themselves are not annihilated, so the number of points remains $n$. Processes with annihilation are ubiquitous, appearing in connections to random walks, Brownian motion, branching processes and more, based on motivations from physics and chemistry. As $X_i,i=1,\ldots,n$ are independent, each with a continuous distribution, the probability of annihilation is zero. Let $A$ and $B$ be as defined in Example \ref{ex:perc}. The fact that $A \Box B = \emptyset$ in the presence of annihilation, hence making inequality \eqref{eq:box.inequality} trivial, is covered by the following more general result in the context of this model.$\hfill \ensuremath{\blacktriangle}$
\end{example}
\begin{proposition}\label{prop:1}
For the percolation model with annihilation $A \Box B=\emptyset$, where $A$ and $B$ are any events that require that the edges of at least one point are not annihilated.
\end{proposition}
\proof For any nontrivial event $A$ that requires that edges of at least one point not be annihilated, any ${{\bf x}}$ and any index set $K$ such that $|{K^c}| \ge |K|$, it is impossible that $[{{\bf x}}]_K \subseteq A$ as the {$|{K^c}|$} unspecified coordinates in $[{{\bf x}}]_K$, those indexed by the larger set ${K^c}$, can annihilate all the edges attached to the points indexed by $K$. Hence $A \Box B$ must be empty, as given complementary sets $K$ and $L$, at least one set, say $K$, satisfies $|K^c| \ge |K|$. \qed
We remark that Proposition \ref{prop:1} has a flavor similar to the example based on \eqref{eq:soda1}, in that the outcome of the classical box operation may be determined by events that occur with probability zero. Returning to the sets $A$ and $B$ in Example \ref{ex:percann}, Proposition \ref{prop:1} implies $A \Box B=\emptyset$. However, for $n$ sufficiently large, $P(A \small \framebox[1.45\width]{$11$} B)>0$, and moreover, this probability converges to one as $n \rightarrow \infty$. To see this, consider a finite set of circles $C_i$, $i=1,\ldots,m$ of radius $r/2$ and centers in $[0,1]^2$ such that $C_1$ intersects the $y$-axis, and $C_m$ intersects the line $x=1$, and $C_{i} \cap C_{i+1} \ne \emptyset$ for $i=1,\ldots,m-1$. As $n \rightarrow \infty$, for all $i=1,2,\ldots,m$ there will be a point $X_i \in C_i$ with probability approaching one, and the edges between these points will not be annihilated with probability one. Hence, these points will form a path from left to right corresponding to the event $A$, and a similar argument for $B$ implies that $P(A \small \framebox[1.45\width]{$11$} B) \rightarrow 1$.
Finally we comment that ${\bf x} \in A \Box B$ under the model without annihilation if and only if ${\bf x} \in A \small \framebox[1.45\width]{$11$} B$ under the model with annihilation, that is, when edges attached to coincident points are annihilated. In other words, $A \Box B=A \small \framebox[1.45\width]{$11$} B$, where the set on the left-hand side is defined for the model without annihilation, and the right-hand set is defined for the model with annihilation.
\section{Cylindrification }\label{sec:cyl}
Before proceeding to our extension of the box operation, we pause to make a simple observation about potential improvements to the original box inequality, and our subsequent generalizations. Here we show that `cylindrical cores' of the sets that comprise the box operation may be separately extracted, yielding no loss to the left-hand side of \eqref{eq:box.inequality}, but a possibly smaller quantity on the right. For this purpose, the cylindrical core $E^{(0)}$ of a given $E \in \mathbb{S}$, that extracts the non-trivial cylinders that $E$ contains, is defined for $n \ge 1$ by
\begin{eqnarray} \label{def:overset2T}
{E}^{(\tiny{0})}=\bigcup_{{({\bf x},K)}\,:\,{K \subseteq [n]},0 \le |K| \le n-1, [{\bf x}]_K \subseteq E} [{\bf x}]_K.
\end{eqnarray}
Note that for any ${\bf x} \in S$ we have ${\bf x} \in [{\bf x}]_K \subseteq S$ for any $K$, and therefore ${\bf x} \in {S}^{(\tiny{0})}$,
implying that ${S}^{(\tiny{0})}=S$.
Recalling the definition of $[A]_K$ in \eqref{def:AGH.cyl.box1}, it is easy to verify that \eqref{def:overset2T} is equivalent to
\begin{eqnarray} \label{def:overset3T}
{E}^{(\tiny{0})}=\bigcup_{{K \subseteq [n]},0 \le |K| \le n-1} [E]_K.
\end{eqnarray}
In Proposition \ref{prop:coreT} we show that ${E}^{(\tiny{0})}$ is the smallest set that preserves the result of the box operation applied to $E$ and any proper subset of $S$.
Since for all $E$ it holds that $E \Box S = E$, the equality $A \Box S = A^{(\tiny{0})} \Box S$ would imply $A=A^{(\tiny{0})}$. Thus the claim \eqref{eq:AboxB.overset.AT} has been restricted to proper subsets $B$ of $S$.
Recalling $S=\prod_{i=1}^n S_i$, we assume for the last part of the next proposition that for all $i=1,\ldots,n$ the set $S_i$ is non-trivial in that it contains more than a single point.
\begin{proposition}\label{prop:coreT}
{For all $A \subseteq S$} the set ${A}^{(\tiny{0})}$ defined in \eqref{def:overset2T} satisfies ${A}^{(\tiny{0})} \subseteq A$, and
\begin{eqnarray} \label{eq:AboxB.overset.AT}
A \Box B = A^{(\tiny{0})} \Box B \quad \forall B \not = S.
\end{eqnarray}
If $A$ and $B$ are proper subsets of $S$, then
\begin{eqnarray} \label{eq:AboxBsuoo}
A \Box B = A^{({\tiny{0}})} \Box B=A \Box B^{({\tiny{0}})} = A^{({\tiny{0}})} \Box B^{({\tiny{0}})}.
\end{eqnarray}
If the space $S=\prod_{i=1}^n S_i$ satisfies $|S_i| >1$ for all $i=1,\ldots,n$ then ${A}^{(\tiny{0})}$ is the smallest set that satisfies
\eqref{eq:AboxB.overset.AT} in the sense that if \eqref{eq:AboxB.overset.AT} also holds for a set ${A}^{(\tiny{1})}$ in place of $A^{(\tiny{0})}$ then $A^{(\tiny{1})} \supseteq A^{(\tiny{0})}$.
\end{proposition}
\noindent {\em Proof: } {First it is clear that $A^{(\tiny{0})}$, being a union of subsets of $A$, is a subset of $A$.}
Hence, by the evident monotonicity of the box operation, in order to prove \eqref{eq:AboxB.overset.AT} we are required only to show that
$A \Box B \subseteq {A}^{(\tiny{0})} \Box B$ for all $B \not = S$.
If ${\bf x} \in A \Box B$ then there exist disjoint $K$ and $L$ such that $[{\bf x}]_K \subseteq A$ and $[{\bf x}]_L \subseteq B$. Note that $B \not = S$ implies $|L| \ge 1$ and hence $|K| \le n-1$. Now by definition $[{\bf x}]_K \subseteq A^{(\tiny{0})}$, and ${\bf x} \in A^{(\tiny{0})} \Box B$.
The first and third equalities in \eqref{eq:AboxBsuoo} follow by using \eqref{eq:AboxB.overset.AT} and that $B$ is a proper subset of $S$, that $A^{({\tiny{0}})} \subseteq A$ and that $A$ is a proper subset of $S$. The second equality then follows by reversing the roles of $A$ and $B$.
We finally show the minimality of $A^{(\tiny{0})}$ when $|S_i|>1$ for all $i=1,\ldots,n$. If ${\bf x} \in A^{(\tiny{0})}$ then for some $K$ satisfying $|K|\le n-1$ we have $[{\bf x}]_K \subseteq A^{(\tiny{0})}$. Now, with $L={K^c}$ let $B = [{\bf x}]_L$; then ${\bf x} \in A^{(\tiny{0})} \Box B$. Since $|L|=n-|K| \ge 1$ and $|S_i|>1$ {for all $i=1,\ldots,n$} we have that $B \not = S$ and therefore ${\bf x} \in A^{(\tiny{0})} \Box B = A^{(\tiny{1})} \Box B$, so in particular ${\bf x} \in A^{(\tiny{1})}$. The inclusion $ A^{(\tiny{0})} \subseteq A^{(\tiny{1})}$ is hence demonstrated.
\qed
Applying this result to any proper subsets $A$ and $B$ of $S$, we obtain, for possible improvements in \eqref{eq:box.inequality},
\begin{eqnarray} \label{eq:AboxB.oversetT}
{A}^{(\tiny{0})} \subseteq A, \quad {B}^{(\tiny{0})} \subseteq B, \lcolor{\quad} A \Box B = {A}^{(\tiny{0})} \Box {B}^{(\tiny{0})} \qmq{and} P(A\Box B)\le P(A^{({\tiny{0}})}) P(B^{({\tiny{0}})}),
\end{eqnarray}
with the last inequality holding provided all sets involved are measurable; see Proposition \ref{prop:LebegT}.
For the rest of this section we focus on the case where $S=\prod_{i=1}^n S_i$, a product of Polish spaces, $\mathbb{S}$ is the Borel product sigma algebra on $S$, $P$ is a product Borel measures on $S$ and the sets $A$ and $B$ considered below are Borel.
Recall that Proposition \ref{prop:partof4} states that $A \Box B$ is measurable with respect to the completion $\overline{P}$ of $P$ in this framework. Using similar arguments, the next proposition gives the measurability of $A^{({\tiny{0}})}$ under the same conditions.
\begin{proposition}\label{prop:LebegT}
Let $S$ be a finite product of Polish spaces endowed with the product Borel sigma algebra and a product Borel probability measure $P$, and let $A$ and $B$ be Borel subsets of S. Then the sets
$A^{({\tiny{0}})}$ and
$B^{({\tiny{0}})}$
are measurable with respect to the completion $\overline{P}$ of $P$, and
for all $A$ and $B$ proper Borel subsets of $S$, we have
$$\overline{P}(A \Box B) \le {\overline P}(A^{({\tiny{0}})}) {\overline P}(B^{({\tiny{0}})}).$$
\end{proposition}
\noindent {\em Proof: }
We prove that the set $[A]_K$, defined in \eqref{def:AGH.cyl.box1}, is $\overline{P}$ measurable in Section \ref{sec:meas}, the Appendix on measurability. The measurability claim now follows from \eqref{def:overset3T}.
To prove the inequality, by Proposition \ref{prop:coreT} and \eqref{was.5} we have
$A \Box B {=} A^{({\tiny{0}})} \Box B^{({\tiny{0}})}\subseteq A^{({\tiny{0}})} \small \framebox[1.45\width]{$11$} B^{({\tiny{0}})}$, and the result now follows by \eqref{eq:oovsprod}.
\qed
We now illustrate the $E \mapsto E^{(\tiny{0})}$ operation with the following simple example.
\begin{example}[Bond Percolation]
Consider bond percolation on the graph below. There are 13 edges, each of which is open or closed independently, yielding $|S|=2^{13}$. A path from left to right in this example means a path from either $a$ or $b$ to either $c$ or $d$ along open edges.
\centerline{\begin{tikzpicture}
\draw [thick](0,0) --(1,0) -- (2,0) -- (3,0) -- (4,0);
\draw [thick](0,1) --(1,1) -- (2,1) -- (3,1) -- (4,1);
\draw [thick](0,0) --(0,1);
\draw [thick](1,0) --(1,1);
\draw [thick](2,0) --(2,1);
\draw [thick](3,0) --(3,1);
\draw [thick](4,0) --(4,1);
\node at (-.08,1.08) {a};
\node at (-.07,-.07) {b};
\node at (4.07,-.07) {d};
\node at (4.07,1.07) {c};
\end{tikzpicture}}
Let $A=$$\{$there exists a path from left to right$\}\cup$$\{$all edges are closed$\}\cup$$\{${exactly three edges are open}$\}$.
In this case $A^{(\tiny{0})}=\{$there exists a path from left to right$\}$ which {is a strict subset} of $A$. More specifically, $A \setminus A^{(\tiny{0})}$ contains $1+ {13 \choose 3}$ points. Assuming, for example, a uniform probability, we obtain the gap $P(A)-P(A^{(\tiny{0})})=(1+ {13 \choose 3})/2^{13}=0.035$. It is easy to increase the gap by taking the union of $A$ with the set $\{$exactly two edges are open$\}$, or that exactly a particular set of edges is open, for any set that does not guarantee a path from left to right. \hfill \ensuremath{\blacktriangle}
\end{example}
\section{The Lenient $s,t$ Box} \label{sec:lenient}
In Section \ref{intro} we considered the box operation {$\small \framebox[1.45\width]{$11$}$} that is not influenced by events of zero probability. Here, for $(s,t) \in [0,1]^2$, we define the more general $\small \framebox[1.45\width]{$st$}$ operation from which {$\small \framebox[1.45\width]{$11$}$} can be recovered as the
special case where $(s,t)={(1,1)}$. The parameters $s$ and $t$ give thresholds for the `leniency' on certain conditional probabilities that $A$ and $B$ occur.
Our notations and definition of conditional probability follow \cite{Brei} Chapter 4, and will be explained in detail below.
We now describe the `lenient' set operation $A \small \framebox[1.45\width]{$st$} B$. Consider a random vector ${\bf X} = (X_1,\ldots,X_n)$ $\in S$ under the conditions of Theorem \ref{thm1}, and the probability that ${\bf X}$ is in $A$ conditioned on the values of the coordinates of ${\bf X}$ in some $K \subseteq [n]$, expressed as $P({\bf X} \in {A} \mid {\bf X}_{K}={\bf x}_{K})$, where for $K \subseteq [n]$ and for ${\bf x} \in S$ the vector ${\bf x}_K$ is the projection of ${\bf x}$ on $K$, that is, the $|K|$-vector consisting of the coordinates of $\bf x$ in $K$. If this conditional probability is at least $s$ then revealing that the coordinates in $K$ of ${\bf X}$ and ${\bf x}$ agree implies that the conditional probability that ${\bf X} \in A$ is at least $s$.
We define
\begin{multline}\label{def:firstst}
$A\small \framebox[1.45\width]{$st$} B$=\{{\bf x} \in S: \exists K,L \subseteq [n],\,\,
K \cap L = \emptyset \,\, \mbox{such that} \,\,\, P({\bf X} \in A \mid {\bf X}_{K}={\bf x}_{K})\ge s\,\,\, \mbox{and} \,\, P({\bf X} \in B \mid {\bf X}_{L}={\bf x}_{L})\ge t \}.
\end{multline}
In particular, when both $s$ and $t$ are close to one, events $A$ and $B$ are both conditionally very likely to occur for disjoint reasons.
{The set $A \small \framebox[1.45\width]{$st$} B$ as described is defined using conditioning on ${\bf X}={\bf x}$, rather than on a sigma algebra; we refer the reader to Definition 4.18. of \cite{Brei}. We also note, by Proposition 4.9 of \cite{Brei}, that when the conditioning sigma algebra is generated by ${\bf X}$, conditional expectation may be written as a function of ${\bf X}$. See also Corollary 4.38 of \cite{Brei}, which is also relevant for our case. Measurability of the sets $E_{r,K},E_r$ and $A \small \framebox[1.45\width]{$st$} B$ introduced in this section are handled in Proposition \ref{prop:measure.st.sets}.
Our formal definition of the $\small \framebox[1.45\width]{$st$}$ operation requires specifying a particular version of the conditional probabilities in \eqref{def:firstst}. For $K \subseteq [n]$, ${\bf y} \in S_K$ and ${\bf v} \in S_{K^c}$ let
$\langle {\bf y},{\bf v}\rangle_K$ be the point in $S$ with $i^{th}$ coordinate given by
$$\left({\langle {\bf {y}},{\bf v} \rangle_K} \right)_i = \left\{
\begin{array}{cc}
{y}_i & i \in K\\
v_i & i \in {K^c},
\end{array}
\right.$$
and for $E \in \mathbb S, 0 \le r \le 1$ and
$K {\subseteq} [n]$ let
\begin{eqnarray} \label{def:uv.cond}
E_{r,K}=\left\{
\begin{array}{cl}
\left\{{\bf x} \in S: P_{K^c}\left(\{{\bf v} \in S_{K^c}: {\langle {\bf x}_{{K}}, {\bf v} \rangle_K} \in E\}\right) \ge r \right\} & \qm{for $|K| \le n-1, 0 \le r \le 1$} \\
E & \qm{for $K=[n], 0<r \le 1$, and}\\
S & \qm{for $K=[n], r = 0$.}
\end{array}
\right.
\end{eqnarray}
We note that the relation ${\bf x} \in E_{r,K}$ is equivalent to $[{\bf x}]_K \in E_{r,K}$ since it depends on ${\bf x} \in S$ only through ${\bf x}_K$, as the probability appearing in its definition only so depends. To describe the set $E_{r,K}$ in words, given ${\bf x} \in S$, consider the set of vectors ${\bf v} \in S_{K^c}$ such that when substituting the coordinates of ${\bf x}$ with indices in ${K^c}$ by those of ${\bf v}$ yields a vector that is in $E$. If the ${K^c}$ marginal probability of the set of such ${\bf v}$ is at least $r$, then ${\bf x}$ is in $E_{r,K}$.
Illustrative special cases of the application of definition \eqref{def:uv.cond} are $E_{0,K}=S$ for all $K$ as the substitution of any coordinates in $K^c$ will yield a non-negative conditional probability. Also $E_{r,\emptyset} = S$ if $P(E) \ge r$, as even having no coordinates revealed one is still assured that $E$ will occur with probability at least $r$,
and otherwise $E_{r,\emptyset} = \emptyset$, as $\langle {\bf x}_\emptyset, {\bf v}\rangle_\emptyset ={\bf v}$, which is in $E$ if and only if ${\bf v} \in E$, whose probability does not reach the threshold $r$.
With $A_{s,K}$ and $B_{t,L}$ defined as in \eqref{def:uv.cond} for measurable set $A$ and $B$ of $S$, and $(s,t) {\in} [0,1]^2$, we have
\begin{eqnarray} \label{def:stbox}
A \small \framebox[1.45\width]{$st$} B = \bigcup_{K \cap L = \emptyset} {A_{s,K} \cap B_{t,L}}.
\end{eqnarray}
Definitions \eqref{def:firstst} and \eqref{def:stbox} are equivalent.
This equivalence becomes clear when noting, as stated in Proposition \ref{prop:measure.st.sets},
that $P_{K^c}\left(\{{\bf v} \in S_{K^c}: \langle {\bf {y}}, {\bf v} \rangle_K {\in} E\}\right)$, defined for all ${\bf y}\in S_K$, is a version of the conditional probability $P({\bf X} \in E|{\bf X}_K={\bf {y}})$.
The event $A \small \framebox[1.45\width]{$st$} B$ consists of all outcomes ${\bf x}$ for
which there are disjoint sets of coordinate indices $K$ and $L$ such that the conditional probabilities of $A$ given the values of $x_i$ for $i \in K$, and $B$ given the values of $x_j$ for $j \in L$, are at least $s$ and $t$, respectively.
It is easy to see that for any events $A$ and $B$ we have $A \small \framebox[1.45\width]{$oo$} B=S$. Moreover,
if $P(A)\ge s$ and $P(B)\ge t$ then $A \small \framebox[1.45\width]{$st$} B = S$, as can be seen by choosing $K=L=\emptyset$. However, this choice is impossible if we impose $L=K^c$, and in fact, it is easy to see that the relations $P(A)\ge s$ and $P(B)\ge t$ do not imply $A \small \framebox[1.45\width]{$st$} B = S$ in general if we impose $L=K^c$ in \eqref{def:stbox}. This shows that
in contrast to definition \eqref{def:classical_box} of $A \Box B$ and definition \eqref{eq:specilize.function.version} of $A \small \framebox[1.45\width]{$11$} B$, in general $A \small \framebox[1.45\width]{$st$} B$ of definition \eqref{def:stbox} is not equivalent to the one obtained by replacing the condition that $K \cap L = \emptyset$ by $L={K^c}$.
See Example \ref{ex:threeside} below for another illustration. Nevertheless, in Example \ref{ex:coin} below, assuming that $K$ and $L$ are complementary in the definition of the set $A \small \framebox[1.45\width]{$st$} B$ will not affect the set defined.
The next proposition shows that definitions \eqref{def:stbox} and \eqref{eq:specilize.function.version} are consistent, that is, that they coincide when $(s,t)=(1,1)$, thus explaining our earlier choice of notation for $A\small \framebox[1.45\width]{$11$} B$.
\begin{proposition}\label{prop:easyoo} The set $A \small \framebox[1.45\width]{$st$} B$ obtained by specializing definition \eqref{def:stbox} for $(s,t)=(1,1)$ agrees with the set $A \small \framebox[1.45\width]{$11$} B$ whose indicator is given in \eqref{eq:specilize.function.version}.
\end{proposition}
\proof By \eqref{eq:specilize.function.version},
${\bf x} \in A\small \framebox[1.45\width]{$11$} B$ if and only if for some disjoint $K$ and $L$,
${\rm ess} \inf_{{\bf y} \in [{\bf x}]_K}{\bf 1}_A({\bf y})={\rm ess} \inf_{{\bf y} \in [{\bf x}]_L}{\bf 1}_B({\bf y})=1$,
if and only if for some disjoint $K$ and $L$, $P_{K^c}\left(\{{\bf v} \in S_{K^c}: \langle {\bf x}_{{K}}, {\bf v} \rangle_K \in A\}\right)=1$ and \\ $P_{L^c}\left(\{{\bf v} \in S_{{L}^c}: \langle {\bf x}_{{L}}, {\bf v} \rangle_{{L}} \in B\}\right)=1$. These latter two equalities hold if and only if ${\bf x} \in A\small \framebox[1.45\width]{$st$} B$ for ${(s,t)=(1,1)}$. \hfill $\Box$
Theorem \ref{thm:bound.by.enlarged} below gives a version of the BKR inequality for the lenient box operation. Given $E \in \mathbb S$ and $r {\ge} 0$, define the $r$-inflated set
\begin{equation}\label{eq:nowinflate}
E_r= \bigcup_{K \subseteq [n]} E_{r,K}.
\end{equation}
Since $E_{r,[n]}=E$ for $r>0$ and equals $S$ otherwise, it follows that $E \subseteq E_r$. Below we denote the set $(E_r)^{(\tiny{0})}$ by $E_{{r}}^{(\tiny{0})}$.
As $E_r$ is measurable for a measurable $E$ by Proposition \ref{prop:measure.st.sets} below, so are the sets $A_s$ and $B_t$ appearing in Theorem \ref{thm:bound.by.enlarged}, and by Proposition \ref {prop:LebegT}, the sets $A_s^{(\tiny{0})}$ and $B_t^{(\tiny{0})}$ are measurable with respect to the completion of $P$ when $S$ is a finite product of Polish spaces endowed with the product Borel sigma algebra, on which $P$ is a product probability measure.
\begin{theorem} \label{thm:bound.by.enlarged}
Let $A$ and $B$ be in $\mathbb S$ and $(s,t) \in [0,1]^2$. Then with
$A_s$ and $B_t$
as in \eqref{eq:nowinflate},
\begin{eqnarray} \label{eq:BKRst}
P(A \small \framebox[1.45\width]{$st$} B) \le P(A_s)P(B_t).
\end{eqnarray}
If the sets $A_s^{(\tiny{0})}$ and $B_t^{(\tiny{0})}$ are measurable then
\begin{eqnarray} \label{eq:nowinflate0}
P(A \small \framebox[1.45\width]{$st$} B) \le P(A_s^{(\tiny{0})})P(B_s^{(\tiny{0})}) \le P(A_s)P(B_t).
\end{eqnarray}
\end{theorem}
\noindent {\em Proof:} We first show that
$A \small \framebox[1.45\width]{$st$} B \subseteq A_s \Box B_t $. Indeed, if ${\bf x} \in A \small \framebox[1.45\width]{$st$} B$ then there exist disjoint $K$ and $L$ such that
\begin{eqnarray*}
P_{K^c}\left(\{{\bf v} \in S_{K^c}: \langle {\bf x}_{{K}}, {\bf v} \rangle_K \in A\}\right) \ge s \qmq{and} P_{L^c}\left(\{{\bf v} \in S_{L^c}: \langle {\bf x}_{{L}}, {\bf v} \rangle_L \in B\}\right) \ge t.
\end{eqnarray*}
In particular, $[{\bf x}]_K \subseteq A_{s,K} \subseteq A_s$ and $[{\bf x}]_L \subseteq B_{t,L} \subseteq B_t$, and as $K$ and $L$ are disjoint we have ${\bf x} \in A_s \Box B_t$ as desired. Together with \eqref{was.5} we now have $A \small \framebox[1.45\width]{$st$} B \subseteq A_s \Box B_t \subseteq A_s \small \framebox[1.45\width]{$11$} B_t $, and now \eqref{eq:oovsprod} implies \eqref{eq:BKRst}.
To show \eqref{eq:nowinflate0}, we have
$$A \small \framebox[1.45\width]{$st$} B \subseteq A_s \Box B_t= A_s^{(\tiny{0})} \Box B_t^{(\tiny{0})} \subseteq A_s^{(\tiny{0})} \small \framebox[1.45\width]{$11$} B_t^{(\tiny{0})}$$
where we first apply the inclusion just shown, followed by
\eqref{eq:AboxB.oversetT} and \eqref{was.5}. We now obtain the first inequality in \eqref{eq:nowinflate0} applying \eqref{eq:oovsprod}; the second one follows from the first two relations in \eqref{eq:AboxB.oversetT}. \qed
Measurability issues of the sets appearing in this section are summarized in the following proposition, which is proved in Section \ref{sec:meas}, the Appendix on measurability. Below we set ${\bf y}={\bf x}_{K} \in S_K$, and the set $\{{\bf v} \in S_{K^c}: {\langle {\bf x}_{{K}}, {\bf v} \rangle_K} \in E\}$ becomes $\{{\bf v} \in S_{K^c}: {\langle {\bf y}, {\bf v} \rangle_K} \in E\}$.
\begin{proposition} \label{prop:measure.st.sets}
The set $\{{\bf v} \in S_{K^c}: {\langle {\bf y}, {\bf v} \rangle_K} \in E\}$, whose $P_{K^c}$ probability appears in \eqref{def:uv.cond}, is measurable, and $P_{K^c}\left(\{{\bf v} \in S_{K^c}: {\langle {\bf y}, {\bf v} \rangle_K} \in E\}\right)$ defined for all ${\bf y}\in S_K$, is a version of the conditional probability $P({\bf X} \in E|{\bf X}_K={\bf {y}})$. For
a measurable set $E$, the sets $E_{r,K}$ of \eqref{def:uv.cond} and $E_r$ of \eqref{eq:nowinflate} are measurable. Also, for measurable sets $A,B$ the set $A \small \framebox[1.45\width]{$st$} B$ of \eqref{def:stbox} is measurable. Furthermore, the replacement of our version of the conditional probability by another in \eqref{def:uv.cond} changes $E_{r,K}$ by a $P$-null set, and hence the same applies to $A \small \framebox[1.45\width]{$st$} B$, and our results hold independently of the version chosen.
\end{proposition}
\section{Lenient Box Examples} \label{sec:lenexam}
\begin{example}\label{ex:coin}[An odd coin tossing problem]\label{example:odd.coin}
A fair coin is to be tossed independently $n=2m+1$ times for some integer $m \ge 1$. Let $A$ be the event that the first $m+1$ tosses are all heads, and $B$ the event that the last $m+1$ tosses are all tails. With $s=t=1/2$ it is not hard to see that $A \small \framebox[1.45\width]{$st$} B$ consists of the following sequences: (a) the two sequences in which the first $m$ tosses are heads, the last $m$ tosses are tails and the middle toss could be either; (b) the $m$ sequences in which the $m+1^{st}$ toss is a head, the first $m$ tosses are all heads except for exactly one, and the last $m$ tosses are all tails, and (c) the $m$ sequences in which the $m+1^{st}$ toss is a tail, the first $m$ tosses are all heads, and the last $m$ tosses are all tails except for exactly one. Altogether there are $2(m+1)$ such sequences, hence
\begin{eqnarray*}
P(A \small \framebox[1.45\width]{$st$} B) = 2 (m+1) (1/2)^n=(m+1)(1/2)^{2m}.
\end{eqnarray*}
To illustrate the notions defined in previous sections, we determine the sets $A_{1/2,K}$ and $B_{1/2,L}$ as defined in \eqref{def:uv.cond}, and compute $P(A_{1/2})$. For $A_{1/2,K}$ to contain some $\bf x$, its projection ${\bf x}_K$ must either specify all the first $m+1$ coordinates as heads, so that $A$ occurs with probability one, or heads in all but one of these coordinates, so that the probability that the unspecified coordinate is heads, and therefore that $A$ occurs, is $1/2$. In particular, $A_{1/2,K} = \emptyset$ if $|K| < m$, so contributions to $A_{1/2}$ can only arise from sets $K$ satisfying $|K| \ge m$.
Hence, $A_{1/2}$, following definition \eqref{eq:nowinflate},
consists of $A$ and all sequences that have at most one tail in the first $m+1$ positions, and thus we obtain
\begin{eqnarray*}
P(A_{1/2})=(1/2)^{m+1}+ (m+1)(1/2)^{m+1} = (m+2)(1/2)^{m+1}.
\end{eqnarray*}
Arguing similarly for $B$, the inequality
\begin{eqnarray*}
(m+1)(1/2)^{2m} = P(A \small \framebox[1.45\width]{$st$} B) \le P(A_{1/2})P(B_{1/2})= \{(m+2)(1/2)^{m+1}\}^2=(m+2)^2 (1/2)^{2m+2}
\end{eqnarray*}
is consistent with the conclusion of Theorem \ref{thm:bound.by.enlarged}. These calculations show that in this case the excess multiple in the bound, such as the one computed in \eqref{eq:excess.multiple} for the standard box, is given by
\begin{eqnarray*}
\frac{P(A_{1/2})P(B_{1/2})}{P(A \small \framebox[1.45\width]{$st$} B)} = \frac{(m+2)^2}{4(m+1)}.
\end{eqnarray*}
In this case the lenient box is not empty, yet not only is the box empty, but also the intersection, that is
$A \cap B = \emptyset$. Intuitively this means that
$A$ and $B$ can never occur together and certainly not for disjoint reasons. Thus if Alice is betting on $A$ and Bob on $B$, then it is impossible for both of them to win their bets. However, certain revealed coordinates can make each of them hopeful. The set $A \small \framebox[1.45\width]{$st$} B$ with {$(s,t)=(1/2,1/2)$} contains those ${\bf x}$'s for which there exists a set of coordinates such that if Alice conditions on them her probability of winning the bet is at least 1/2, and there is a disjoint set of coordinates which does the same for Bob. It is easy to modify the example in a way that that it is not impossible for both of them to win.
To see that in this case we can restrict $L$ to be $K^c$ in \eqref{def:stbox}, note that for any sequence of tosses ${\bf x} \in A \small \framebox[1.45\width]{$st$} B$ described above, we can take $K$ and $L$ to be the sets of coordinates where heads and tails appear respectively, and then
${{\bf x}} \in A_{1/2,K} \cap B_{1/2,L}$.
$\hfill \ensuremath{\blacktriangle}$
\end{example}
\begin{example} [A three sided coin] \label{ex:threeside}
The goal of this example is to show that changing the definition of $A \small \framebox[1.45\width]{$st$} B$ from allowing any disjoint $K$ and $L$, to always setting $L=K^c$, may result in a different set, unlike the case of $A \Box B$.
Consider a three sided fair coin that shows Heads, Tails, or Sides. The coin is tossed 5 times, and similar to Example \ref{ex:coin}, $A$ is the event that the first three tosses are heads, and $B$ that the last three are tails. Let {$(s,t)=(1/3,1/3)$}. Then if we take $K=\{1,2\}$ and $L=\{4,5\}$ we see that
${\bf x}= (HHSTT) \in A \small \framebox[1.45\width]{$st$} B$. However, if we only allowed $L=K^c$
in the definition of $A \small \framebox[1.45\width]{$st$} B$, then ${\bf x}$ would not be in this set.$\hfill \ensuremath{\blacktriangle}$
\end{example}
\begin{example}[Continuous percolation with positive-probability annihilation] \label{ex:contq} In Example \ref{ex:percann} we considered a graph model having randomly placed vertices where certain edges were annihilated with probability zero. In particular, if the points $X_i$ and $X_j$, sampled independently from a continuous distribution, coincide, that is, if they are within distance $q=0$, then all their incident edges are annihilated.
Generalizing this model, we annihilate all edges incident on any two distinct points that lie within distance $q \ge 0$ of each other. Assume that $X_1,\ldots,X_n$ are i.i.d. with a strictly positive density, bounded above by $c$ on $[0,1]^2$. Edges are formed as in Example \ref{ex:perc}, with annihilation as just described. When $q>0$, annihilation occurs with positive probability, and hence it is easy to see that when $A$ and $B$ are the crossing events given in Example \ref{ex:perc}, we have $A \small \framebox[1.45\width]{$11$} B = \emptyset$. Indeed, similar to the reasoning in the proof of Proposition \ref{prop:1}, for each ${\bf x} \in A \small \framebox[1.45\width]{$11$} B$ at least one of $|K|$ or $|L|$ cannot be greater than $n/2$. Taking this set to be $K$ without loss of generality, its complementary set contains enough coordinates so that the points it indexes can annihilate, with positive probability, all edges incident on points indexed by $K$.
$\hfill \ensuremath{\blacktriangle}$
\end{example}
The next proposition shows that in Example \ref{ex:contq}, if the thresholds $s$ and $t$ are sufficiently small relative to $q$, then $A \small \framebox[1.45\width]{$st$} B$ is non-empty. To simplify matters, we now restrict to metrics $d$ in $\mathbb{R}^2$ that are induced by norms. Given a norm, let $\tau$ denote the area of the unit ball $\{x:d(0,x) \le 1\}$ in the plane. Then the area of balls of radius $q \ge 0$ in the metric space $([0,1]^2,d)$ is bounded by $\tau q^2$, with equality if the entire disk is in $[0,1]^2$.
We use the following notation: for $K \subset [n]$ let $\mathscr{A}_K$ denote the event that there is crossing from left to right corresponding to the event $A$, however now only the points $\{X_i: i \in K\}$ are taken into consideration, that is, edges are only formed between pairs of points indexed by $K$ having distance $\le 2r$, and only points indexed by $K$ may annihilate any of these edges.
The event $A$ that there is crossing from left to right when all $n$ points are taken into consideration, as in Example \ref{ex:perc}, through edges that are not annihilated, coincides with $\mathscr{A}_{[n]}$. It is easy to see that for $n$ and $|K|$ large enough, $\mathscr{A}_K$ has positive probability. Similarly, for $L \subset [n]$ let $\mathscr{B}_L$ denote the corresponding event for crossing from bottom to top.
\begin{proposition}\label{prop:emptybst} Let $A$ be the event that there is crossing from left to right as in Example \ref{ex:perc} through edges that are not annihilated, and let $B$ denote such a crossing from bottom to top.
Let
${\bf x} = (x_1,\ldots,x_n) \in ([0,1]^2)^n$ be such that ${\bf x} \in \mathscr{A}_K$ for some $K \subset [n]$, and ${\bf x} \in \mathscr{B}_L$
for a disjoint set $L \subset [n]$.
Then ${\bf x} \in A \small \framebox[1.45\width]{$st$} B$ provided
$s \,\,\mbox{and}\,\, t$
lie in the interval
$[0, 1-\frac{c}{4}n^2 \tau q^2]$, and $q$ is small enough so that $\frac{c}{4}n^2 \tau q^2 <1$, where $c$ is an upper bound on the common density function of $X_1,\ldots,X_n$.
\end{proposition}
\proof Consider $A$, reasoning for $B$ being similar, and let ${\bf x}$ and $K$ satisfy the conditions of the proposition. The number of points that make up the path corresponding to $A$ is at most $|K|$, and hence there is at most area $|K| \tau q^2$ into which a point having index in ${K^c}$ can land to cause annihilation, which happens with probability at most $c|K|\tau q^2$. Hence, letting $E_j$ be the event that the point indexed by $j \in {K^c}$ causes the annihilation of the edges incident to some point on the path, the conditional probability of the event $F$ that some point on the path is annihilated, given $\{{X}_i,\, i \in K\}$ satisfies
\begin{multline} \label{eq:to.not.hang}
P({\bf X \in}F\mid {\bf X}_K={\bf x}_K)=P\Big(\bigcup_{j \in {K^c}}E_j\mid {\bf X}_K={\bf x}_K\Big) \\
\le \sum_{j \in {K^c}} P(E_j\mid {\bf x}_K) \le (n-|K|)|K| c \tau q^2 \le
\frac{c}{4}n^2 \tau q^2 \le 1-s.
\end{multline}
Given ${\bf x} \in \mathscr{A}_K$, that is, a realization in which the points indexed by $K$ `already' form a path, $A$ can only fail to occur if some edge incident to a point on this path is annihilated. Hence
$P({\bf X} \in {A}\mid {\bf {X}}_K={\bf x}_K) {\ge} P({\bf X \in}F^c\mid{\bf X}_K={\bf x}_K) \ge s$. Repeating the argument for $B$, the result follows in view of \eqref{def:firstst}.
\hfill $\Box$
Another way to think about the situation considered in Proposition \ref{prop:emptybst} is to consider the case where points indexed by $K$ have been revealed, and, perhaps despite annihilation caused by the revealed points, form a path corresponding to $A$. We then ask about the conditional probability that the remaining, unobserved points will destroy this path by annihilation. It is easy to see that for
$q$ satisfying the condition of the propostion the probability that some unobserved point
annihilates the edges of an observed one is at most $1-(1-|K|c \tau q^2)^{n-|K|} \le (n-|K|)|K| c \tau q^2$, agreeing with the bound in \eqref{eq:to.not.hang}.
Note that $r$ of Example \ref{ex:perc} can be chosen so small that $\lceil{n/2}\rceil$ points may not suffice to form a path as required by the events $A$ and $B$. In this case disjoint $K$ and $L$, both of which must demonstrate paths, cannot exist. Nevertheless, in such a case $A \small \framebox[1.45\width]{$st$} B$ may still be non-empty, as the coordinates of ${\bf x}$ revealed by $K$ and $L$ when considered only marginally may be sufficiently promising for the chances that the
yet unrevealed coordinates will yield $A$, and the same for $B$.
\section{Earlier Results on Continuous Spaces} \label{sec:GuRa}
We conclude with a brief mention of an earlier BKR result for the
special case of continuous spaces given in \cite{GuRa}; see also references therein. The Poisson Boolean percolation model consists of an inhomogeneous Poisson point process in some bounded region of ${\mathbb{R}}^d$, each point having a ball of a random radius around it, independent of the point process. For events $A$ and $B$ defined by this process, $A \Box B$ comprises those configurations of process points and their associated radii for which the occurrence of $A$ and $B$ can respectively be claimed upon revealing all process points, and their associated balls, lying in two disjoint spatial regions. In this setting, inequality \eqref{eq:box.inequality} is proved in a way that relies heavily on properties of the Poisson process. Consequently, it appears that this Poisson process case does not imply our results for a fixed number of points.
Conversely, even leaving the spatial structure in \cite{GuRa} aside, it seems that our results do not lead to results that involve a random number of points.
We can condition on the number of Poisson points and obtain $P(A \Box B\mid N) \le P(A\mid N)P(B\mid N)$ and take expectation with respect to $N$. However, the inequality in $E\{P(A\mid N)P(B\mid N)\} \le E\{P(A\mid N)\}E\{P(B\mid N)\} =P(A)P(B)$ holds only under very special conditions.
\section{ Appendix on Measurability}\label{sec:meas}
In this section we prove Propositions \ref{prop:partof4}, \ref{prop:LebegT} and \ref{prop:measure.st.sets}, the first two of which pertain to Polish spaces. \smallskip
\noindent {\em Proof of Propositions \ref{prop:partof4} and \ref{prop:LebegT}:}
Both proposition follow once we show that if $A \in \mathbb{S}$ then $[A]_K$ is measurable with respect to the Borel sigma algebra $\mathbb{S}$ when completed relative to the product probability measure $P$ of Borel measures on a product of Polish spaces. For Polish spaces restricted to be subspaces of $\mathbb{R}$ this was shown
in Section 8 and Lemma 1 of \cite{AGH}.
First, as in \cite{AGH}, with ${\rm Proj}_K$ denoting the projection of $x \in S$ on the coordinates in $K$, we may express the cylinders $[A]_K$ defined in \eqref{def:AGH.cyl.box1} as
\begin{eqnarray} \label{def:cylinder}
[A]_K=\{ y \in S : y_K \in S_K \setminus {\rm Proj}_K(A^c) \}.
\end{eqnarray}
Letting $A$ be a set in the product Borel sigma algebra, the complement $A^c$ is Borel. By \cite{Cohn}, Proposition 8.4.4 (see also the text preceding) the projection $C={\rm Proj}_K(A^c)$ is analytic.
In particular, $C$ is universally measurable, and hence measurable in the completion of $P_K$.
The set $E=S_K \setminus {\rm Proj}_K(A^c)$, the complement of $C$ in $S_K$, is therefore also measurable in the completion of $P_K$.
Letting $K=\{1,\ldots,k\}$ without loss of generality, note that we may explicitly write the ``pullback" $[A]_K$ of $E$ as
\begin{eqnarray*}
[A]_K=E \times \prod_{j=k+1}^n S_j.
\end{eqnarray*}
As $E$ is measurable with respect to the completion of $P_K$, there exists $F$ and $G$ in ${\cal S}_K$ such that $F \subset E \subset G$ and $P_K(F)=P_K(G)$, yielding the inclusion
\begin{eqnarray*}
F \times \prod_{j=k+1}^n S_j \subset [A]_K \subset G\times \prod_{j=k+1}^n S_j.
\end{eqnarray*}
As the sets on the left and right are elements of the product Borel sigma algebra $\mathbb{S}$, and have the same $P$ measure, $[A]_K$ is measurable with respect to the completion of $P$. This argument, as in \cite{AGH}, holds also for countable products.
\noindent {\em Proof of Proposition \ref{prop:measure.st.sets}.}
A proof that for ${\bf y} \in S_K$ and $E \in \mathbb S$ the set $\{{\bf v} \in S_{K^c}: {\langle {\bf y}, {\bf v} \rangle_K} \in E\}$ is measurable is given in \cite{Fo} Proposition 2.34a.
Next we show that $P_{K^c}\left(\{{\bf v} \in S_{K^c}: {\langle {\bf y}, {\bf v} \rangle_K} \in E\}\right)$ is measurable in ${\bf y} \in S_K$, and that it is a version of the conditional probability $P({\bf X} \in E|{\bf X}_K={\bf {y}})$.
Fix some $K \subseteq [n]$ and $E \in \mathbb S$. For ${\bf y} \in S_K$ let $Q_{K^c}({\bf y};E):=P_{K^c}\{{\bf v} \in S_{K^c}: \langle {\bf y}, {\bf v} \rangle_K \in E\}$. Dropping the subscript ${K^c}$ when not needed for clarity, that $Q({\bf y};E)$ is a measurable function of ${\bf y}$ follows directly from Tonelli's Theorem, see \cite{Fo}, 2.37b, as
\begin{eqnarray} \label{eq:Q.is.int}
Q({\bf y};E) = \int_{S_{K^c}} \phi_E({\bf y}, {\bf v})P_{K^c}(d{\bf v}) \qmq{{upon setting} ${\phi}_E({\bf y},{\bf v})={\bf 1}(\langle {\bf y}, {\bf v} \rangle_K \in E)$.}
\end{eqnarray}
It is immediate that $Q({\bf y};\cdot)$ is a probability measure on $(S,{\mathbb S})$ for each ${\bf y} \in S_K$.
We next show that $Q({\bf y};E)$ is a version of $P({\bf X} \in C \mid {\bf X}_K={\bf {y}})$, that is, following 4.18 in \cite{Brei}, that
\begin{eqnarray} \label{eq:DC}
\int_{{D}}
{Q({\bf y};E) P_K({d\bf y})}
= P(\{{\bf X} \in E\} \cap \{{\bf X}_K \in D\}) {\qmq{for all $D \in \mathbb{S}_K$.}}
\end{eqnarray}
Using \eqref{eq:Q.is.int}, starting from the left hand side of \eqref{eq:DC}, we have, using $P=P_{K^c}P_K$ for the second equality, followed by Tonelli's Theorem,
\begin{multline*}
\int_D Q({\bf y};E)P_K(d{\bf y}) = \int_D \int_{S_{K^c}} \phi_E({\bf y}, {\bf v})P_{K^c}(d{\bf v})P_K(d{\bf y})
\\
={\int_D \int_{S_{K^c}} {\bf 1}(\langle {\bf y}, {\bf v} \rangle_K \in E)
P_{K^c}(d{\bf v})P_K(d{\bf y})
= \int {\bf 1}\left({\bf x} \in E, [{\bf x}]_K \in D \right)
P(d{\bf x})}
=P(\{{\bf X} \in E\} \cap \{{\bf X}_K \in D\}),
\end{multline*}
which is \eqref{eq:DC}.
Next we show measurability of $E_{r,{K}}$ in \eqref{def:uv.cond}, $E_r$ of \eqref{eq:nowinflate} and $A \small \framebox[1.45\width]{$st$} B$ of \eqref{def:stbox}; as by \eqref{def:uv.cond} the sets $E_{r,K}$ for $K=[n]$ are clearly measurable, we restrict to the case $|K| \le n-1$.
Letting again ${\rm Proj}_K:S \rightarrow S_K$ be the projection that sends ${\bf x}$ to ${\bf x}_K$, returning to \eqref{def:uv.cond} and considering ${\bf y} \mapsto Q_{K^c}({\bf y};E)$ as a function only of ${\bf y} \in S_K$ for fixed $K$ and $E$, we see that $E_{r,K}={\rm Proj}_K^{-1}(Q_{K^c}^{-1}([r,1];{E}))$, the inverse image of a composition of measurable mappings of a measurable set, and is therefore measurable.
By \eqref{eq:nowinflate} and \eqref{def:stbox} it readily follows that $E_r$ and $A \small \framebox[1.45\width]{$st$} B$ are measurable.
Lastly, we show that the replacement in \eqref{def:uv.cond} of our
version of the conditional probability by another changes $E_{r,K}$ by a $P$-null set. Again, as $E_{r,[n]}$ given in \eqref{def:uv.cond} does not depend on the version, we restrict to the case $|K| \le n-1$. Letting $G$ be the subset of $S_K$ where the two versions differ we have $P_K(G)=0$, and without loss of generality, let $K=\{1,\ldots,k\}$. With $E^i_{r,K}, i =1,2$ denoting the corresponding sets under these versions, the symmetric difference $E^1_{r,K} \triangle E^2_{r,K}$ is contained in $G \times S_{K^c}$, a $P$-null set. Hence the symmetric difference has $P$ measure zero. \qed
\noindent {\bf Acknowledgements:} We are deeply indebted to an anonymous referee for a very careful reading of two versions of this paper. The penetrating comments provided enlightened us on various measurability issues and other important, subtle points.
We thank Mathew Penrose for a useful discussion and for providing some relevant references.
The work of the first author was partially supported by NSA grant H98230-15-1-0250. The second author would like to thank the Isaac Newton Institute for Mathematical Sciences, Cambridge, for support and hospitality during the programme Data Linkage and Anonymisation where part of the work on this paper was undertaken,
supported by EPSRC grant no EP/K032208/1.
|
\section{Introduction}
Recently, the relativistic Beth-Uhlenbeck (BU) approach for two-particle
correlations in two-flavor quark matter within the Nambu--Jona-Lasinio (NJL) model
\cite{Hufner:1994ma,Zhuang:1994dw,Blaschke:2013zaa,Dubinin:2013yga}
and the Polyakov-loop improved NJL (PNJL) model
\cite{Wergieluk:2012gd,Yamazaki:2012ux,Yamazaki:2013yua,Blaschke:2014zsa,Dubinin:2015glr,Torres-Rincon:2016ahl}
has appeared as a promising candidate for a unified description of hadron and quark-gluon matter.
In particular, the description of correlations in deconfined quark matter, sometimes denoted as
"bound states above $T_c$" \cite{Ratti:2011au,Nahrgang:2013xaa,Nahrgang:2016lst},
with $T_c$ being the pseudocritical temperature of the chiral and
deconfinement crossover transitions, is naturally achieved within this formulation.
Attempts to formulate a unified thermodynamics of the hadron-to-quark-gluon matter transition have been made in postulating a spectral function for hadrons across the chiral/deconfinement transition
\cite{Blaschke:2003ut,Turko:2011gw,Blaschke:2015nma}.
In order to reach this description, however, there are a few more important steps to be made.
One of them is the inclusion of the strangeness degree of freedom.
We devote the present work to this aim, restricting ourselves to the case of the lightest hadron
channels only, the pseudoscalar $\pi$ and $K$ mesons and their chiral partner states,
the corresponding scalar mesons.
As is widely known, the NJL model is capable of describing the chiral restoration transition
in a hot and dense medium, where the dynamically generated quark masses drop as a function of temperature and chemical potentials, thus restoring the mass degeneracy of the chiral partner states.
At the same time, the continuum thresholds for quark-antiquark scattering channels drop, which results
in a lowering of the binding energy for the pseudoscalar meson bound states and finally in their dissociation when they enter the continuum and change their character to resonances with a finite
lifetime (Mott effect).
This aspect of the Mott dissociation is usually identified by mass poles for mesons becoming complex,
with the real part being the "mass" $M_i$ of the resonance $i$ and the imaginary part being related to a finite width $\Gamma_i$ due to its decay into the quark constituents.
The phase shift of these states is well described by a Breit-Wigner function as long as $\Gamma_i\ll M_i$.
Following previous works on the BU approach to pion dissociation in quark matter in the NJL
\cite{Hufner:1994ma,Zhuang:1994dw,Blaschke:2013zaa,Dubinin:2013yga} and PNJL \cite{Wergieluk:2012gd,Yamazaki:2012ux,Yamazaki:2013yua,Blaschke:2014zsa,Dubinin:2015glr,Torres-Rincon:2016ahl} models,
we point out that the Breit-Wigner description is not complete as it neglects the fact that analytic properties of continuum states (resonances or just scattering states) are properly described by Jost functions, defined along the cut for continuum states in the complex energy plane.
This fact is most appropriately taken into account by the introduction of a scattering phase shift function for each mesonic channel $\delta_i(s)$, which depends on the squared center-of-mass energy $s$ in the quark-antiquark system and thermodynamic parameters of the medium, here the temperature $T$ and chemical potentials $\mu_u=\mu_d=\mu$ and $\mu_s$ for light and strange quarks, respectively.
The behaviour of the phase shift at the threshold can be used as an indicator for the Mott transition of a bound state to the scattering state continuum.
The phase shift vanishes at infinity, while it has a value of $\pi$ at the continuum threshold as long as there is a bound state below the continuum.
When the bound state merges with the scattering states continuum, the phase shift jumps to zero, in accordance with the Levinson theorem.
A striking advantage of the BU approach over phenomenological models like \cite{Blaschke:2003ut,Turko:2011gw,Blaschke:2015nma} is the formulation of the thermodynamics of correlations in terms of phase shift functions that are in accordance with the Levinson theorem \cite{Hufner:1994ma,Zhuang:1994dw,Wergieluk:2012gd,Blaschke:2013zaa,Dubinin:2013yga,Blaschke:2014zsa,Dubinin:2015glr},
which guarantees that the partial pressure of correlations vanishes at asymptotic temperatures (and chemical potentials). It may be understood as an anticipation of asymptotic freedom for strong, finite range interactions \cite{Dashen:1969ep}.
As a result, within the present approach the EoS and thermodynamic properties (like the composition) of the system of hadrons, quarks and gluons can be described with the correct asymptotics of the hadron gas (of pions and kaons) at low temperatures and the quark-gluon plasma at high temperatures, with the transition in-between.
A quantitative comparison with lattice QCD thermodynamics data is premature at this stage of the development of the BU approach, mainly due to the lack of hadronic states and missing self-consistency.
It is therefore deferred to a future stage of work.
In this work we investigate thermodynamic properties of quark-gluon-meson matter for the PNJL model with $N_{f}=2+1$ quark flavors, where the strange chemical potential is fixed to $\mu_s=0.2~\mu$ as motivated by statistical model analyses of chemical
freeze-out parameters for describing hadron production in heavy-ion collision experiments.
We study the temperature dependence of quark and meson masses along lines of constant $\mu/T$ in the $T$-$\mu$ plane which are seen as approximations to lines of constant entropy per baryon along which the hydrodynamic evolution of the heavy-ion collision shall proceed.
The EoS for pressure versus temperature in the Beth-Uhlenbeck approach is evaluated for
$\mu/T=0,~0.5,~1.0,~2.0$, whereby the quark, gluon and meson contribution to the total pressure is given.
Special emphasis is on the Mott dissociation of the pseudoscalar meson states which is illustrated by the behaviour of the corresponding phase shifts as functions of energy for different temperatures and chemical potentials.
The corresponding scalar meson states are unbound already in the vacuum, as is characteristic for the local PNJL model.
Nevertheless one can observe how at high temperatures and chemical potentials the chiral partner states become degenerate in their analytic properties as encoded by the phase shifts.
The paper is organized as follows.
In Section \ref{sec:partition} we outline the path integral approach to the partition function of the 2+1 flavor PNJL model, including mesonic fluctuations beyond the meanfield in the scalar and pseudoscalar channels in Gaussian approximation.
In Section \ref{sec:results} we present the results for the phase diagram and for the temperature dependence of the order parameters $\Phi$, $\bar{\Phi}$, $m_u=m_d$ and $m_s$ as well as the meson mass spectra, the energy dependent phase shifts and the pressure with its contributions from quarks, gluons and mesons along a set of trajectories of constant $\mu/T$ in the phase diagram.
In Section \ref{sec:conclusion} we present the conclusions of this study.
\section{Partition function of the $2+1$ flavor PNJL model}
\label{sec:partition}
We employ a $2+1$ flavor chiral quark model with NJL - type interaction kernel
based on the one used in works on the SU$_f$(3) scalar and
pseudoscalar meson spectrum \cite{Costa:2002gk,Costa:2003uu,Costa:2005cz}
developed on the basis of Ref.~\cite{Rehberg:1995kh} and its
generalization by coupling to the Polyakov loop
\cite{Hansen:2006ee,Costa:2008dp},
\begin{eqnarray}
\label{lagr}
{\cal L} &=& \bar{q} \left( i \gamma^\mu D_\mu + \hat{m}_0\right) q +
G_S \sum_{a=0}^{8} \left[ \left( \bar{q} \lambda^a q\right)^2+
\left( \bar{q} i \gamma_5 \lambda^a q\right)^2
\right]
\nonumber\\
&-& \mathcal{U}\left(\Phi[A],\bar\Phi[A];T\right).
\end{eqnarray}
Here $q$ denotes the quark field with three flavors,
$f=u,d,s$, and three colors, $N_c=3$; $\lambda^a$ are the flavor SU$_f$(3)
Gell-Mann matrices ($a=0,1,2,\ldots,8$), $G_S$ is a coupling constant.
The scalar-pseudoscalar meson interaction channels in (1) are color
singlet and fulfill the requirement of being chirally symmetric.
Thus, only the diagonal matrix of current quark masses
$\hat{m}_0 = \mbox{diag}(m_{0,u}, m_{0,d}, m_{0,s})$ induces an explicit breaking
of the otherwise global symmetry of the Lagrangian (\ref{lagr}).
This is a property shared with the QCD Lagrangian.
The covariant derivative is defined as
$D^{\mu}=\partial^\mu-i A^\mu$, with $A^\mu=\delta^{\mu}_{0}A^0$
(Polyakov gauge); in Euclidean notation $A^0 = -iA_4$.
The strong coupling constant $g_s$ is absorbed in the definition of
$A^\mu(x) = g_s {\cal A}^\mu_a(x)\frac{\lambda_a}{2}$, where
${\cal A}^\mu_a$ is the (SU$_c$(3)) gauge field and $\lambda_a$ are the
Gell-Mann matrices in SU$_c$(3) color space.
The Polyakov loop field $\Phi$ appearing in the potential term of
(\ref{lagr}) is related to the gauge field through the gauge covariant
average of the Polyakov line~\cite{Ratti:2005jh}
\begin{equation}
\Phi(\vec x)=\left\langle \left\langle l(\vec x)\right\rangle\right\rangle
=\frac{1}{N_c}{\rm Tr}_c\left\langle \left\langle L(\vec x)
\right\rangle\right\rangle,
\label{eq:phi}
\end{equation}
where
\begin{equation}
L(\vec x) ={\cal P}\exp\left[i\int_0^\beta d\tau A_4(\vec x, \tau)\right]\,.
\label{eq:loop}
\end{equation}
Concerning the effective potential for the (complex) $\Phi$ field, we adopt
the polynomial form and the parametrization proposed in Ref.~\cite{Ratti:2005jh}.
Alternatively, the logarithmic form of this potential \cite{Roessner:2006xn}
could also be used.
The (P)NJL model is a primer for describing the dynamical breakdown of the
(approximate) chiral symmetry in the vacuum and its partial restoration at high
temperatures and chemical potentials.
This feature is governed by the occurrence of a nonvanishing expectation value
for the mean field in the scalar meson channel.
At the same time this model provides a field-theoretic description of
pseudoscalar meson properties which is in accordance with the low energy
theorems of QCD, such as the Goldstone theorem.
This is achieved by considering the (Gaussian) fluctuations above the mean field
in the scalar and pseudoscalar meson sector and analysing their properties as
encoded in the matrix elements of the polarization operator which after analytic
continuation from the bosonic Matsubara frequencies to the real $q_0$ axis reads
\cite{Klevansky:1992qe}
\begin{eqnarray}
\Pi^{M^a}_{ff'} (q_0 ,{\bf q}) &=& 2N_cT \sum_{n} \int \frac{d^3p}{(2\pi)^3}\mbox{tr}_{D}
\left[ S_{f}(p_n ,{\bf p}) \Gamma^{M^a}_{ff'} \right.\nonumber\\
&&\times \left.S_{f'}(p_n + q_0,{\bf p+q}) \Gamma^{M^a}_{ff'} \right],
\end{eqnarray}
where
\begin{eqnarray}
\Gamma^{P^a}_{ff'}&=& i\gamma_5\ T^a_{ff'} ~,~~\Gamma^{S^a}_{ff'} = T^a_{ff'}~,\\
T^a_{ff'} &=&
\left\{
\begin{array}{l}
(\lambda_3)_{ff'}, \\
(\lambda_1\pm i \lambda_2)_{ff'}/\sqrt{2}, \\
(\lambda_4\pm i \lambda_5)_{ff'}/\sqrt{2}, \\
(\lambda_6\pm i \lambda_7)_{ff'}/\sqrt{2},
\end{array}
\right.
\end{eqnarray}
for $P^a=\pi^0, \pi^\pm, K^\pm, K^0, \bar{K}^0$ denoting the pseudoscalar meson states and
$S^a=a_0^0, a_0^\pm, \kappa^\pm, \kappa^0, \bar{\kappa}^0$ the scalar meson states.
Here $\mbox{tr}_{D}$ is the trace over Dirac matrices,
the sum over $n$ denotes the sum over the fermionic Matsubara frequencies $\omega_n=(2n+1)\pi T$,
and
$S_f (p_n,{\bf p})=[\gamma_0(i\omega_n+\mu_f + A^0)-{\bf \gamma}{\bf p} - m_f]^{-1}$ is the quark
Green function with the dynamical quark mass $m_f$ and chemical potential $\mu_f$ for the flavor $f$.
The matrix elements of the polarization operator can be represented in terms of two integrals which after summation over the Matsubara frequencies for mesons at rest in the medium (${\bf q}={\bf 0}$) are given by
\begin{eqnarray}
&&\Pi^{P^a,S^a}_{ff'} (q_0+i\eta, {\bf 0}) = 4 \bigl\{ I_{1}^f(T,\mu_{f})+I_{1}^{f'}(T,\mu_{f'}) \nonumber \\
&&\, \mp \left[ (q_0+\mu_{ff'})^2 -(m_f \mp m_{f'})^2 \right] I_2^{ff'}(z,T,\mu_{ff'}) \bigr\}\, , \nonumber
\end{eqnarray}
where $\mu_{ff'}=\mu_{f}-\mu_{f'}$.
To parametrize the model with known pseudoscalar meson
masses the vacuum expressions of the integrals are used
\begin{eqnarray}
\label{i1}
I_1^f(0,0)
&=&\frac{N_c}{4 \pi^2} \int^{\Lambda}_0 \frac{dp \, p^2}{E_f} ,
\end{eqnarray}
{\color{black}
\begin{eqnarray}\label{i2}
I_2^{ff'}(z,0,0)
&=& \frac{N_c}{4 \pi^2} \int^{\Lambda}_0 \frac{dp \, p^2}{E_f E_{f'}}
\frac{E_f+E_{f'}}{z^2-(E_f+ E_{f'})^2} ,
\end{eqnarray}}
with $E_{f}=\sqrt{{p}^2+m_{f}^2}$ being the quark energy.
The integrals for finite temperature and chemical potential have the following form
\begin{eqnarray}\label{firstt}
I_1^f(T,\mu_f) &=& \frac{N_c}{4\pi^2} \int_0^\Lambda \frac{dp \, p^2}{E_f} \left(n^-_f - n^+_f \right),
\\
I_2^{ff'} (z,T,\mu_{ff'}) &=& \frac{N_c}{4\pi^2} \int_0^\Lambda \frac{dp \, p^2}{E_fE_{f'}}
\nonumber\\
&&\Biggl[ \frac{E_{f'}}{(z-E_f-\mu_{ff'})^2-E_{f'}^2} \,\, n^-_f \nonumber \\
&& - \frac{E_{f'}}{(z+E_f-\mu_{ff'})^2-E_{f'}^2} \,\, n^+_f
\nonumber \\
&& +\frac{E_f}{(z+E_{f'}-\mu_{ff'})^2-E_f^2} \,\, n^-_{f'}
\nonumber \\
&&
- \frac{E_f}{(z-E_{f'}-\mu_{ff'})^2-E_f^2} \,\, n^+_{f'}
\Biggr],
\end{eqnarray}
where $n_f^{\pm}=f^+_\Phi(\pm E_f)$ are the generalized fermion distribution
functions \cite{Costa:2008dp, Blaschke:2014zsa} for quarks of flavor $f$ with positive (negative)
energies in the presence of the Polyakov loop values $\Phi$ and $\bar{\Phi}$
\begin{eqnarray}
\label{f-Phi}
f^+_\Phi(E_f)&=&
\frac{(\bar{\Phi}+2{\Phi}Y)Y+Y^3}{1+3(\bar{\Phi}+{\Phi}Y)Y+Y^3}
~,\\
f^-_\Phi(E_f)&=&
\frac{({\Phi}+2\bar{\Phi}\bar{Y})\bar{Y}+\bar{Y}^3}{1+3({\Phi}+\bar{\Phi}\bar{Y})\bar{Y}+\bar{Y}^3}
~,
\label{f-Phi-bar}
\end{eqnarray}
where the abbreviations $Y={\rm e}^{-(E_f-\mu_f)/T}$ and $\bar{Y}={\rm e}^{-(E_f+\mu_f)/T}$ are used.
The functions (\ref{f-Phi}) and (\ref{f-Phi-bar}) fulfill the relationship
$f^+_\Phi(-E_f)=1-f^-_\Phi(E_f)$,
and they go over to the ordinary Fermi functions $f_1^\pm(E_f)$ for $\Phi=\bar \Phi=1$, where
\begin{eqnarray}\label{fermi}
f_1^\pm(E_f) = \frac{1}{1+ {\rm e}^{\beta (E_f \mp \mu_f)}}~.
\end{eqnarray}
The quark masses $m_f$ are found by solving the gap equation
\begin{eqnarray}
m_f = m_{0,f} + 16\, m_f G_S I_1^f(T,\mu),
\end{eqnarray}
while the meson mass spectrum is obtained from the pole condition
\begin{eqnarray}\label{mass}
\mathcal{P}^{M^a}_{ff'}(M_{M^a}+i\eta, {\bf 0})=
1- 2G_S \Pi^{M^a}_{ff'}(M_{M^a}+i\eta, {\bf 0}) =0.
\nonumber\\
\end{eqnarray}
Note that when there is no bound state solution below the continuum threshold for
$q_0=M_{M^a}<m_{{\rm thr}, ff'}$, $m_{{\rm thr}, ff'} = m_f+m_{f'}$,
then in the vicinity of the threshold, for $M_{M^a}>m_{{\rm thr},ff'}$, Eq.~(\ref{mass}) can still be solved in its
complex form in order to determine the mass of the resonance $M_{M^a}$ and the
respective decay width $\Gamma_{M^a}$.
Thus, we assume that this equation can be written as a system of two coupled equations
(upper (lower) sign for $M^a=P^a\, (S^a)$)
\begin{eqnarray}
\label{BSE1}
&&M_{M^a}^2-\frac{1}{4}\Gamma_{M^a}^2-(m_f\mp m_{f'})^2 = -\mbox{Re} I_2(M_{M^a},T,\mu_{ff'})\nonumber\\
&&\hspace{1cm}\times \frac{(8G_S)^{-1}\mp(I_1^f(T,\mu_f)+I_1^{f'}(T,\mu_{f'}))}{|I_2(M_{M^a}+i\eta,T,\mu_{ff'})|^2} ~,
\label{BSE2}
\end{eqnarray}
\begin{eqnarray}
\label{BSE2}
M_{M^a} \Gamma_{M^a} &=& - \mbox{Im} I_2(M_{M^a}+i\eta,T,\mu_{ff'})
\nonumber \\
&& \times \frac{(8G_S)^{-1}\mp (I_1^f(T,\mu_f)+I_1^{f'}(T,\mu_{f'}))}{|I_2(M_{M^a}+i\eta,T,\mu_{ff'})|^2}~,
\nonumber\\
\end{eqnarray}
which have solutions of the form
\begin{eqnarray}
q_0=M_{M^a}-i\frac{1}{2}\Gamma_{M^a}~.
\end{eqnarray}
To study the thermodynamics we use the following expression for the quark pressure of flavor $f$
\begin{eqnarray}
\label{pressure-quark}
P_f &=&
-\frac{(m_f-m_{0,f})^2}{8G}
+\frac{N_c}{\pi^2}\int_0^\Lambda {dp\, p^2}\, E_f
\nonumber\\
&+&\frac{N_c}{3\pi^2}\int_0^\infty\frac{dp\, p^4}{E_f}
\left[f_\Phi^+(E_f) + f_\Phi^-(E_f)\right]~,
\end{eqnarray}
and for the mesonic pressure we use the BU form in no-sea approximation with the phase shifts
\begin{eqnarray}
\label{GBU_M}
P_{M}
&=&
d_{\rm M} \int\frac{{\rm d}^3q}{(2\pi)^3}~
\int_0^\infty\frac{{\rm d}\omega}{2 \pi}~
\bigg\{\nonumber\\
&& g(\omega-\mu_{\rm M})+g(\omega+\mu_{\rm M})
\bigg\}
\delta_{\rm M}(\omega, {\bf q})
~,
\label{BUU}
\end{eqnarray}
where $\delta_{\rm M}(\omega, {\bf q})$ is the mesonic phase shift which has the form
\begin{eqnarray}
\delta_{\rm M}(\omega,{\bf q}) = -\arctan \left\{
\frac
{\mathrm{Im}\left(\mathcal{P}^M_{ff'}(\omega-i\eta,{\bf q})\right)}
{\mathrm{Re}\left(\mathcal{P}^M_{ff'}(\omega+i\eta,{\bf q})\right)}
\right\}~,
\end{eqnarray}
and $g(E)$ is the Bose function
\begin{eqnarray}
\label{Bose}
g(E)=\frac{1}{{\rm e}^{\beta E}-1}~
\end{eqnarray}
Note that the physical mesons are color singlet states and therefore their distribution functions do not
depend on the Polyakov loop.
This is in striking difference to the case of diquarks which are colored objects and their distribution is
strongly suppressed by the coupling to the Polyakov loop in the confined domain.
In the deconfined domain, diquarks are suppressed too by the Mott dissociation to their quark constituents
\cite{Blaschke:2014zsa}.
Therefore we do not include them in our considerations in this work.
\section{Results and Discussion}
\label{sec:results}
The parameters used for the numerical studies in this work are the bare quark masses
$m_{0(u,d)}= 5.5~$MeV and $m_{0s}= 138.6~$MeV, the three-momentum cut-off
$\Lambda= 602~$MeV and the scalar coupling constant $G_{S}\Lambda^{2}= 2.317$.
With these parameters one finds in vacuum a constituent quark mass for light quarks of 367 MeV, a pion mass of 135 MeV and pion decay constant $f_{\pi}=92.4~$MeV.
We present the results along trajectories in the PNJL phase diagram shown in Fig.~\ref{Fig:Phasdig}. Each trajectory corresponds to a constant ratio $\mu/T = 0, 0.5, 1.0, 2.0$, where $\mu=\mu_u=\mu_d$ is the light quark chemical potential.
For the strange quark chemical potential the relation $\mu_s=0.2~\mu$ is used \cite{Naskret:2015pna},
see also \cite{Karsch:2010ck}.
\begin{figure}[!t]
\centerline{\includegraphics[width=0.5\textwidth]{PhaseDiagV3}}%
\caption{Phase diagram of the PNJL model: the black solid line corresponds to the first order phase transition which turns into a crossover transition shown by the black dotted line.
Three thin lines correspond to fixed ratios $\mu/T$= 0.5 (red dotted line), $\mu/T=1.0$ (green dashed line) and $\mu/T=2.0$ (blue dash-dotted line).
}%
\label{Fig:Phasdig}%
\end{figure}
\begin{figure}[!bh]
\centerline{\includegraphics[width=0.5\textwidth]{MassesPhi}}%
\caption{Temperature dependence of the dynamical masses for light quarks $m_u=m_d$ and for the strange quark $m_s$ (lower panel) together with the sum and the difference of the traced Polyakov loop
$\Phi$ and its conjugate $\bar{\Phi}$ (upper panel) along different trajectories in phase diagram:
$\mu/T=0$ (solid black line), $\mu/T=0.5$ (blue dash-dotted line),
$\mu/T=1.0$ (green dashed line) and $\mu/T=2.0$ (red dotted line).
The value of the strange quark chemical potential is set to $\mu_s=0.2\mu$.
}
\label{Fig:QuarkMasses2}
\end{figure}
\subsection{Phase diagram}
In Fig.~\ref{Fig:Phasdig} we show the phase diagram of the present model.
To this end we find the positions of the minima of the temperature derivative (the steepest decrease) of the quark mass as the chiral order parameter $dM/dT$ in the $T-\mu$ plane and display them by the dashed line. These pseudocritical temperatures go over to the critical temperatures of the first order phase transition characterized by a jump of the quark mass at the corresponding position in the $T-\mu$ plane.
A characteristic feature of the phase diagram is that lowering the ratio $T/\mu \to 0$, the phase transition turns from crossover to first order. The chiral restoration is a result of the phase space occupation due to Pauli blocking which effectively reduces the interaction strength in the gap equation.
\subsection{Mass spectrum for quarks and mesons at finite temperatures}
In Fig.~\ref{Fig:QuarkMasses2} we show the masses of quarks together with the sum and the difference of the traced Polyakov loop $\Phi$ and its conjugate $\bar{\Phi}$ along different trajectories
$\mu/T={\rm const}$ in the $T-\mu$ plane.
We note that the chiral symmetry restoration in the light quark sector is correlated with a rise in the traced
Polyakov loop which, due to the polynomial form of the Polyakov-loop potential, also exceeds the value 1.
For finite chemical potentials $\Phi \neq \bar{\Phi}$ holds in the domains where $\Phi\neq 0$.
In Figs.~\ref{Fig:MesonMassPiMu0} and \ref{Fig:MesonMassKaMu01} we present masses of the pseudoscalar $\pi$ and $K$ mesons as well as of their scalar partners, the $a_0$ and $\kappa$ (kappa) mesons.
These masses may be found in different ways.
The first one is by solving the Bethe-Salpeter equation, where the mass is obtained as the pole of the meson propagator.
This method works well as long as the particle is a true bound state, that is below the Mott temperature
($T_{\rm Mott}^{M}$) for the meson $M$.
Above the Mott temperature the meson becomes an unbound state and the definition of mass is complicated by the fact that the pole becomes complex and the solution is not unique.
Therefore, to find approximate solutions in this case one is generally using the Breit-Wigner form of the propagator with the width defined by $\Gamma={\rm Im}(q_0)$.
However, this approximation is valid only close to $T_{\rm Mott}^{M}$ as it was discussed in our previous work \cite{Blaschke:2013zaa}.
\begin{figure}[!th]
\centerline{\includegraphics[height=0.192\textheight, width=0.4\textwidth]{MesonMassPiMu0}}\centerline{\includegraphics[height=0.192\textheight, width=0.4\textwidth]{MesonMassPiMu05}}
\centerline{\includegraphics[height=0.192\textheight, width=0.4\textwidth]{MesonMassPiMu1}}
\centerline{\includegraphics[height=0.192\textheight, width=0.4\textwidth]{MesonMassPiMu2}}
\caption{
Temperature dependence of the masses for pion (black dash double-dotted line) and $a_0$ meson (red dash double-dotted line) together with continuum thresholds $m_{\rm thr,\pi}=2 m_u$ (green dotted line)
for $\mu/T=0$, $0.5$, $1.0$ and $2.0$ (from top to bottom panel).
Different line styles correspond to different mass definitions: as a maximum of the derivative of the phase shift (lower line to the right) or when the phase shift value hits $\pi/2$ (upper line).
The point on the line denotes the temperature when the phase shift is lower than $\pi/2$ and in this case
the line after that point corresponds to the position of the maximum of the phase shift derivative.
\label{Fig:MesonMassPiMu0}}%
\end{figure}
\begin{figure}[!h]
\centerline{\includegraphics[height=0.222\textheight, width=0.4\textwidth]{MesonMassKaMu0}}
\centerline{\includegraphics[height=0.222\textheight, width=0.4\textwidth]{MesonMassKaMu05}}
\centerline{\includegraphics[height=0.222\textheight, width=0.4\textwidth]{MesonMassKaMu1}}
\centerline{\includegraphics[height=0.222\textheight, width=0.4\textwidth]{MesonMassKaMu2}}
\caption{
Same as Fig.~\ref{Fig:MesonMassPiMu0}, but for kaons and $\kappa$ mesons.
The continuum threshold is given by $m_{\rm thr,\pi}=m_u + m_s$ (green dotted line).
\label{Fig:MesonMassKaMu01}}
\end{figure}
The second method of finding the meson masses involves the phase shift of the quark-antiquark correlation in the considered mesonic interaction channel.
To define the mass we determine the energy $\omega$ where the phase shift assumes the value $\pi/2$.
In the rest frame of the meson this energy corresponds to the mass.
Below the Mott temperature, the phase shift jumps from zero to $\pi$ at this position so that its derivative
is a delta function, characteristic for a true bound state.
We do not consider the positive mode at low energy which arises in the strange sector for
${K}$, ${\kappa}$.
When the phase shift stays below $\pi/2$, the "mass" can be defined by the position of the maximum
of the phase shift.
For the behaviour of the phase shifts see Figs.~\ref{Fig:ShiftsPiSiV} - \ref{Fig:ShiftsKapKV1}.
The third method to determine the meson mass also involves the phase shift.
One finds the value of the mass from the position of the maximum of the derivative of the phase shift.
For a true bound state, the derivative of the phase shift corresponds to a delta function
while for a resonance in the continuum it is a smooth function \cite{Blaschke:2013zaa,Wergieluk:2012gd}. For a meson at rest in the medium, the position of the peak of the energy derivative of the phase shift corresponds to the value of the meson mass.
The finite chemical potential in the case nonzero $\mu/T$ removes the degeneracy of the meson masses
in the strange channels.
A mass difference arises between $K^{+}$ and $K^{-}$ as well as between their chiral partners
$\kappa^{+}$ and $\kappa^{+}$ see Fig.~\ref{Fig:MesonMassKaMu01}.
The chemical potential shifts the pole in the propagator, which results in a reduction of the pseudocritical temperature $T_{c}$ and therefore also in a reduction of the meson Mott temperatures $T_{\rm Mott}^{M}$.
\subsection{Phase shifts for mesons and the Beth-Uhlenbeck equation of state}
In this subsection we discuss the results for the phase shifts for $\pi$ and $K$ mesons in comparison with those of their chiral partners, the $a_0$ and the $\kappa$ mesons, resp. (see also Ref.~\cite{Dubinin:2015glr}) and their consequences for the thermodynamics of quark-meson matter at finite temperature,
with the coupling to the Polyakov-loop.
The solutions for the meson phase shifts at finite temperature are shown in Figs.~\ref{Fig:ShiftsPiSiV} - \ref{Fig:ShiftsKapKV1} as functions of the energy $\omega$ for different temperatures and values of the ratio $\mu/T$.
As in Ref.~\cite{Blaschke:2013zaa} we have made the simplifying assumption that, even in the medium,
the phase shifts are Lorentz invariant and depending on $\omega$ and $\bf q$ only via the Mandelstam variable $s=\omega^{2}-{\bf q}^2$ in the form $\delta_{\rm M}(\omega,{\bf q})=\delta_{\rm M}(\sqrt{s},{\bf q}=0)\equiv\delta_{\rm M}(\sqrt{s};T,\mu_{\rm M})$ for given temperature and chemical potential of the medium.
Then the BU formula for the mesonic pressure can be given the following form
\begin{eqnarray}
\label{GBU_M}
P_{\rm M}
&=&
d_{\rm M} \int\frac{{\rm d}^3q}{(2\pi)^3}~
\int_0^\infty\frac{{\rm d}s}{4 \pi}~
\frac{1}{\sqrt{s+q^2}}\bigg\{
g(\sqrt{s+q^2}-\mu_{\rm M})
\nonumber\\
&&+g(\sqrt{s+q^2}+\mu_{\rm M})
\bigg\}
\delta_{\rm M}(\sqrt{s};T,\mu)
~.
\label{BUU}
\end{eqnarray}
The bound state mass is located at the jump of the phase shift from $0$ to $\pi$ and this jump corresponds to a delta-function in the BU formulas (\ref{BUU}) for the correlations.
In the case when the continuum of the scattering states can be neglected, that is when it is separated by a sufficiently large energy gap from the bound state, we obtain as a limiting case the thermodynamics of a statistical ensemble of on-shell correlations (resonance gas).
\begin{figure}[!th]
\includegraphics[height=0.212\textheight, width=0.38\textwidth]{ShiftsPiSiV3010}
\includegraphics[height=0.212\textheight, width=0.38\textwidth]{ShiftsPiSiV3150}
\includegraphics[height=0.212\textheight, width=0.38\textwidth]{ShiftsPiSiV3250}
\includegraphics[height=0.212\textheight, width=0.38\textwidth]{ShiftsPiSiV3300}
\caption{Dependence of the phase shift in the pion and $a_0$ meson channel on the center of mass energy for different temperatures $T= 10$, $150$, $250$, $300$ MeV (from top to bottom) along different trajectories in the phase diagram: $\mu/T=0$ (black solid line), $\mu/T=0.5$ (blue dash-dotted line), $\mu/T=1.0$ (green dashed line) and $\mu/T=2.0$ (red dotted line).
}
\label{Fig:ShiftsPiSiV}%
\end{figure}
In Fig.~\ref{Fig:ShiftsPiSiV} we show the phase shifts for the nonstrange mesons $\pi$ and $a_0$.
As expected for pseudoscalar and scalar isovector states, their phase shifts reflect the chiral symmetry
breaking at low temperatures (and low chemical potentials), with a deeply bound pion and an $a_0$ state just above the continuum threshold, separated from the pion by a large gap.
At high temperature $T=300$ MeV these phase shifts become indistinguishable, reflecting the chiral symmetry restoration.
In Fig.~\ref{Fig:ShiftsKaKaV} we show the phase shifts for the charged kaons. For $\mu=0$ their modification with increasing temperature parallels that of the pions, with the gap between the bound state and the continuum diminishing with temperature and becoming zero for $T=300$ MeV, above the kaon Mott temperature, where the kaon becomes a resonance in the continuum.
At finite chemical potentials the opposite charge kaons develop a mass splitting and a new low-energy mode appears in the spectrum due to the finite mass difference, see also \cite{Yamazaki:2013yua}.
In Fig.~\ref{Fig:ShiftsKapKV1} the phase shifts for the chiral partner states of the kaons, the charged
$\kappa$ mesons, are shown.
Comparing these results with those for the kaons parallels the comparison of the scalar $a_0$ meson
with the pion.
The mass splitting of the opposite charge states with increasing chemical potential mirrors the behavior of the kaons. At low temperatures the chiral symmetry breaking is manifest in the mass splitting and energy gap between the kaons and the $\kappa$ mesons as chiral partner states.
At high temperatures, when chiral symmetry is restored, these differences vanish and kaon and
$\kappa$ phase shifts become indistinguishable.
\begin{figure}[!th]
\includegraphics[height=0.23\textheight, width=0.38\textwidth]{ShiftsKaKaV3010}
\includegraphics[height=0.23\textheight, width=0.38\textwidth]{ShiftsKaKaV3150}
\includegraphics[height=0.23\textheight, width=0.38\textwidth]{ShiftsKaKaV3250}
\includegraphics[height=0.23\textheight, width=0.38\textwidth]{ShiftsKaKaV3300}\\%}%
\caption{
Same as Fig.~\ref{Fig:ShiftsPiSiV} for the $K^+$ and $K^-$ states.
}
\label{Fig:ShiftsKaKaV}
\end{figure}
\begin{figure}[!th]
\includegraphics[height=0.23\textheight, width=0.38\textwidth]{ShiftsKapKV3010}
\includegraphics[height=0.23\textheight, width=0.38\textwidth]{ShiftsKapKV3150}
\includegraphics[height=0.23\textheight, width=0.38\textwidth]{ShiftsKapKV3250}
\includegraphics[height=0.23\textheight, width=0.38\textwidth]{ShiftsKapKV3300}
\caption{
Same as Fig.~\ref{Fig:ShiftsKaKaV} for the $\kappa^+$ and $\kappa^-$ states.
}
\label{Fig:ShiftsKapKV1}
\end{figure}
Now we want to study the thermodynamics of mesons in a hot and dense medium encoded in the thermodynamic potential or equivalently the pressure (\ref{BUU}).
Eq.~(\ref{BUU}) has the BU
form, where the phase shift obeys the Levinson theorem in a medium \cite{Wergieluk:2012gd,Dubinin:2013yga}
\begin{eqnarray}
\label{Levinson}
\int^{\infty}_{0}d\omega\frac{d\delta_{\rm M}(\omega)}{d\omega}~=~0.
\end{eqnarray}
Here we may introduce the energy level of the continuum threshold $\omega_{\rm thr}$ and split
(\ref{Levinson}) in the sum of two integrals with the intervals of integration $[0,\omega_{\rm thr}]$ and $[\omega_{\rm thr},\infty]$, respectively.
After integrating out we obtain the Levinson theorem in the form \cite{Blaschke:2014zsa}
\begin{eqnarray}
\label{Levinson2}
\pi ~n_{\rm B,M} = \delta_{\rm M}(\omega_{\rm thr}) - \delta_{\rm M}(\infty)~,
\end{eqnarray}
which applies also in the case of a hot and dense medium.
The continuum threshold is $\omega_{\rm thr} =\sqrt{q^2+m_{\rm thr}^2}$, where $m_{\rm thr}=2m_u$ for the light-light quark mesons ($\pi,a_0$) and $m_{\rm thr}=m_u+m_{s}$ for the mixed light-strange two-particle states (${K},{\kappa}$).
For energies below the threshold there can be only a discrete number $n_{\rm B,M}$ of bound states in the channel ${\rm M}$, each contributing an amount of $\pi$ to the change in the phase shift at the bound state energies $\omega_{\rm M, i}=\sqrt{q^2+M_{\rm i}^2}$ with ${\rm i}=1,\dots, n_{\rm B,M}$.
In particular, when due to the chiral symmetry restoration the dropping quark masses entail a lowering of the continuum thresholds $\omega_{\rm thr}$ this triggers the dissolution of the bound states into the continuum (the Mott effect) so that $n_{\rm B,M}=0$ results.
\begin{figure}[!th]
\includegraphics[height=0.2\textheight, width=0.4\textwidth]{Pr00}
\includegraphics[height=0.2\textheight, width=0.4\textwidth]{Pr05}
\includegraphics[height=0.2\textheight, width=0.4\textwidth]{Pr10}
\includegraphics[height=0.2\textheight, width=0.4\textwidth]{Pr20}
\caption{Pressure of the $2+1$ flavor PNJL model with scalar and pseudoscalar meson correlations as a function of temperature (black solid line) normalized to the Stefan-Boltzmann pressure (\ref{P_SB}) for the cases $\mu/T=0, 0.5,1.0, 2.0$ (from top to bottom). Also shown are the partial pressures of the components: light and strange quarks (red dotted); 'gluon' contribution from Polyakov loop potential $\mathcal{U}$ (green dashed); light and strange quarks plus gluons (blue dashed-dotted); mesonic contributions (magenta dash-double-dotted).
}%
\label{Fig:Press0}%
\end{figure}
In Fig.~\ref{Fig:Press0} we show the contributions from the mesons ($\pi,{K},a_0,\kappa$) and partons ($u,d,s$ quarks and gluons) to the total pressure in hot quark matter as functions temperature for different values of the ratio $\mu/T$.
We show the pressure in units of the Stefan-Boltzmann limit for the quark-gluon plasma ($N_c=N_f=3$)
\begin{eqnarray}
P_{\rm SB} &=& \frac{\pi^2}{90}T^4\left\{2(N_c^2-1)\right.\nonumber\\
&&\left.+N_c\sum_{f=u,d,s}\left[ \frac{7}{2} +\frac{15}{\pi^2}\left(\frac{\mu_f}{T}\right)^2 + \frac{15}{2\pi^4}\left(\frac{\mu_f}{T}\right)^4\right]\right\}\ .\nonumber\\
\label{P_SB}
\end{eqnarray}
At finite temperature and vanishing $\mu/T$ the partial pressure of the mesons shows a typical behavior,
first increasing with temperature, then, when the chiral phase transition occurs,
decreasing even before the Mott temperature is reached.
Above the Mott temperature, the growing meson width leads to a stronger reduction of the pressure with a rather sharp onset of this effect.
In case of finite chemical potentials, along trajectories with fixed ratio $\mu/T$, the total pressure grows faster with increasing temperature because the chiral transition temperature drops with increasing $\mu/T$.
\subsection{A possible explanation of the "horn effect" for the $K^+/\pi^+$ ratio?}
We want to come back to the observation made in the discussion of the phase shifts for the kaons
shown in Fig.~\ref{Fig:ShiftsKaKaV} and their scalar partner states shown in Fig.~\ref{Fig:ShiftsKapKV1}
that at finite baryon density and sufficiently high temperature a low-energy resonance or even bound state occurs in the positively charged channels and that is practically absent in the negatively charged channels.
The appearance of this anomalous mode is a consequence of the unequal masses of the constituents of these composite states and it is absent for the pions and their scalar partner states.
We are curious to see whether these anomalous states have a potential to contribute to the resolution of the puzzling observation of an enhancement of the ratio $K^+/\pi^+$ over the ratio $K^-/\pi^-$ of particle yields in heavy-ion collisions at $\sqrt{s_{NN}}\sim 8$ GeV (equivalent to $E_{\rm lab}\sim 30$ AGeV in fixed target experiments (the "horn" effect \cite{Gazdzicki:1998vd}), see \cite{Cleymans:2004hj} for the references to the experimental data and an early attempt to explain the location of the "horn" within a statistical model.
This work suggests that the transition from baryon to meson dominated entropy density may explain the position of the peak for the $K^+/\pi^+$ ratio. The sharpness of the peak, however, is not well reproduced by the statistical model.
While standard kinetic approaches to particle production in heavy-ion collisions have failed to explain the horn effect, the inclusion of chiral symmetry restoration effects for the string decay in the PHSD approach recently resulted in a striking improvement \cite{Palmese:2016rtq}.
Chiral symmetry restoration as the precondition for the Mott dissociation of hadrons has also been the key element in an alternative attempt to provide a mechanism for chemical freeze-out
\cite{Blaschke:2011ry,Blaschke:2011hm} and for the "horn" effect by Mott-Anderson localization \cite{Naskret:2015pna}.
On the basis of the present approach to in-medium phase shifts for quark-antiquark scattering in hot, dense quark matter and the resulting mesonic contributions to the thermodynamics described within the Beth-Uhlenbeck approach, we propose to consider the ratios of partial pressures for the meson states according to Eq.~(\ref{BUU}) for describing the ratio of meson yields at freeze-out
\begin{eqnarray}
\frac{n_{K^\pm}}{n_{\pi^\pm}} =
\frac{\int dM \int d^3p\ (M/E)[{\rm e}^{(E\mp \mu_K)/T}-1]^{-1}\delta_{K^\pm}(M)}
{\int dM \int d^3p\ (M/E)[{\rm e}^{(E - \mu_\pi)/T}-1]^{-1}\delta_{\pi^\pm}(M)}\ ,
\nonumber
\end{eqnarray}
where temperature and baryochemical potential are related to the collision energy $\sqrt{s_{NN}}$ by the
fit formula for the freeze-out parameters in the statistical model given, e.g., by \cite{Cleymans:2005xv}.
The nonequilibrium pion chemical potential is chosen as $\mu_pi=120$ MeV \cite{Kataja:1990tp}.
Since in the present model the absolute value of the pseudocritical temperature is larger than in lattice QCD, we propose to rescale the freeze-out temperature by the corresponding factor and to keep the ratio
$\mu/T$. With this prescription we obtain the energy scan for the ratios $K^+/\pi^+$ and $K^-/\pi^-$
shown in Fig.~\ref{Fig:K+pi+}.
In order to highlight the possible role of the anomalous low-energy states for explaining the "horn" effect
we show by thin lines the same ratios without these states.
\begin{figure}[!th]
\includegraphics[width=0.5\textwidth]{Kpi_MuPi012}
\caption{
Ratio of yields $K^+/\pi^+$ and $K^-/\pi^-$ with and without the anomalous low-energy states as a function of the nucleon-nucleon center of mass energy $\sqrt{s_{NN}}$ in heavy-ion collisions.
}%
\label{Fig:K+pi+}%
\end{figure}
\section{Conclusions}
\label{sec:conclusion}
In this work we have described the thermodynamics of $N_f=2+1$ meson-quark-gluon matter at finite
temperature and chemical potential in the framework of the Beth-Uhlenbeck approach.
In this formulation the quark-antiquark correlations in scalar and pseudoscalar channels are accounted for by the corresponding phase shifts as solutions of the Bethe-Salpeter equations for the meson propagators. The notion of the phase shifts is generalized such as to display both the bound and scattering states spectrum with their characteristic medium effects.
Most prominently this concerns the Mott transition where the bound state transforms to a resonance in the continuum.
This transition manifests itself in the vanishing of the binding energy as a consequence of the lowering of the continuum edge of scattering states due to chiral symmetry restoration.
It appears as a jump of the phase shift at threshold by $\pi$ in accordance with the Levinson theorem.
Thus the mesonic pressure expressed in the BU form reflects this Mott dissociation effect.
While the quark and gluon contribution to the total pressure is increased towards the Stefan-Boltzmann limit above the chiral restoration, the mesonic contribution to the pressure dies out.
In the behaviour of the phase shifts for $K^\pm$ and $\kappa^\pm$ mesons we obtain an anomalous low-energy mode that is particularly pronounced for the positive charge states at finite densities and temperatures.
We have discussed this phenomenon as a possible explanation for the "horn" effect, a pronounced peak in the energy scan of the $K^+/\pi^+$ ratio in heavy-ion collisions around $\sqrt{s_{NN}}\sim 8$ GeV.
This behavior of the pressure in the meson-quark-gluon system at finite temperature and chemical potential is in qualitative agreement with the results from lattice QCD simulations, for the most recent ones see
\cite{Borsanyi:2013bia,Bazavov:2014pvz}.
Quantitatively, there are differences the origin of which is well understood.
The absolute value of the temperature for which the pressure starts rising towards the Stefan-Boltzmann limit is well above the one obtained in lattice QCD.
This discrepancy can be traced to the difference in the pseudocritical temperatures for chiral symmetry restoration.
Local NJL and PNJL models predict chiral restoration temperatures above 200 MeV, while lattice QCD simulations give $T_c=154\pm 9 $ MeV \cite{Bazavov:2011nk}.
The so-called "entanglement" PNJL model has been suggested to cure this problem \cite{Sakai:2010rp}
by modifying the scalar coupling by a function of the traced Polyakov loop $G_S \to G_S(\Phi,\bar{\Phi})$.
Another improvement of the PNJL model concerns its nonlocal generalization which not only provides a solution to the problem of the pseudocritical temperature absolute value
\cite{Radzhabov:2010dd,Horvatic:2010md},
but also describes the running of the quark mass and the wave function renormalization of the quark propagator, in very close agreement with lattice QCD simulations at $T=0$ \cite{Bowman:2005vx}.
At high temperatures, lattice QCD thermodynamics does not approach the massless Stefan-Boltzmann limit but rather a modified one which is well decribed by thermal masses in the two-loop approximation to the thermodynamic potential, see \cite{Blaizot:1999ap} and references therein.
In the transition region, it is a matter of a detailed quantitative comparison, how many more hadronic resonances might be of relevance for a satisfactory description of the lattice QCD data.
It is clear that for very low temperatures and finite baryon densities the inclusion of baryonic states is customary.
Then the EoS constraints from compact star phenomenology and heavy-ion collisions \cite{Klahn:2006ir}
shall be the guideline for the model development where lattice QCD data are absent.
These issues will be subject of subsequent work along the lines of the Beth-Uhlenbeck approach to the thermodynamics of the hadron-to-quark-gluon matter transition.
\section*{Acknowledgements}
The work of A.D. was supported by the Polish National Science Centre (NCN) under grant number UMO - 2014/13/B/ST9/02621 and the Institute for Theoretical Physics of the University of Wroclaw under contract No. 1439/M/IFT/15.
D.B. and A.R. are grateful for support from the NCN under grant number UMO-2011/02/A/ST2/00306.
|
\section{Introduction}
Let $X$ denote the blow up $\mathbb{P}^2$ at $r$ points $p_1,\ldots,p_r \in
\mathbb{P}^2$. It is interesting to ask when a given line bundle $L$ on $X$
has positivity properties such as
ampleness, very ampleness, global generation, and more generally, {\it
$k$-very ampleness} (see the definition below). If the points
$p_1,\ldots,p_r$ are
general in $\mathbb{P}^2$, this question has been extensively studied
and is related to important conjectures in algebraic
geometry. See \cite{Han,ST} for a detailed introduction and some results
in this case.
There has also been some work on these questions when the points are
special in some way.
In \cite{H1,H2}, Harbourne considered blow ups of $\mathbb{P}^2$ at $r$
points (not necessarily distinct) on an irreducible and reduced plane
cubic and a characterization of
line bundles with various positivity properties (ample, global
generated, effective, very ample) was given. De Volder \cite{DV}
partially generalized results of \cite{H1,H2} by
considering blow ups of points (not necessarily distinct) on a reduced and irreducible curve of
degree $e \ge 4$ and giving sufficient conditions for global generation
and very ampleness.
Let $k$ be a non-negative integer. A line bundle $L$ on a projective
variety $X$ is said to be {\it $k$-very ample} if the restriction map
$$H^0(X,L) \to H^0(X,L\otimes \mathcal{O}_Z)$$
is surjective for all zero-dimensional subschemes $Z \subset X$ of
length $k+1$; in other words, for all zero-dimensional subschemes $Z$
such that dim$(H^0(Z,\mathcal{O}_Z)) = k+1$.
Note that 0-very ampleness is equivalent to global generation and
1-very ampleness is equivalent to very ampleness.
As a result, $k$-very ampleness is considered a more general
positivity property for a line bundle.
See \cite{BFS,BS1,BS2} for more details on the notion of $k$-very
ampleness.
A general theorem for $k$-very ampleness on
blow ups of projective varieties was proved by Beltrametti and Sommese in
\cite{BS3}. De Volder and Tutaj-Gasi\'{n}ska \cite{VT} study $k$-very ampleness
for line bundles on blow ups of $\mathbb{P}^2$ at general points on an
irreducible and reduced cubic. Szemberg and Tutaj-Gasi\'{n}ska \cite{ST} study $k$-very ampleness
for line bundles on blow ups of $\mathbb{P}^2$ at general points.
The property of $k$-very ampleness is also studied for other
classes of surfaces as well as higher-dimensional varieties. See \cite{BS,F}, for instance.
Our primary motivation comes from \cite{H1,H2,VT}. These papers study
positivity questions when $X$ is the blow up of $\mathbb{P}^2$ at points on
a plane cubic. In this paper we generalize some of these results by considering
blow ups of
$\mathbb{P}^2$ at $r$ {\it distinct} and {\it smooth} points on an irreducible and reduced plane
curve of degree $e$.
More precisely, let $C \subset \mathbb{P}^2$ be an irreducible and reduced curve of degree
$e$. Let $p_1,\ldots,p_r$ be distinct smooth points on $C$. We consider the
blow up $\pi:X \to \mathbb{P}^2$ of $\mathbb{P}^2$ at $p_1,\ldots,p_r$. Let $H$
denote the pull-back of $\mathcal{O}_{\mathbb{P}^2}(1)$ and let $E_1,\ldots,E_r$
be the inverse images of $p_1,\ldots,p_r$ respectively. Given a line
bundle $L=dH-m_1E_1-m_2E_2-\ldots-m_rE_r$ on $X$,
we are concerned with conditions on $d,e,r,m_1,\ldots,m_r$ which ensure
ampleness and $k$-very ampleness of $L$.
In Section \ref{ample}, we study ampleness and prove our main result
(Theorem \ref{main}) in this case.
In Section
\ref{k-very-ample}, we study
$k$-very ampleness.
Here our main result is Theorem \ref{main1} which gives conditions for
$k$-very ampleness.
We work throughout over the complex number field $\mathbb C$.
\section{Ampleness}\label{ample}
The following is our main theorem on ampleness.
\begin{theorem} \label{main} Let $C$ be an irreducible and reduced plane curve of
degree $e$. Let $X \to \mathbb{P}^2$ be the blow up of $\mathbb{P}^2$ at $r$
distinct smooth points $p_1,\ldots,p_r \in C$. Let $H$
denote the pull-back of $\mathcal{O}_{\mathbb{P}^2}(1)$ and let $E_1,\ldots,E_r$
be the inverse images of $p_1,\ldots,p_r$ respectively.
Let $L$ be a line bundle on $X$ with $L\cdot E_i > 0$ for all
$1\le i \le r$.
Let $C_1$ denote the proper transform of $C$ on $X$.
If $L\cdot C_1 > 0$ and
$L \cdot H > L\cdot (E_{i_1}+\ldots+E_{i_e})$ for any $e$ distinct indices
$i_1,\ldots,i_e \in \{1,\ldots,r\}$, then $L$ is ample.
\end{theorem}
\begin{proof}
Let $d: = L\cdot H >0$ and
$m_i := L \cdot
E_i > 0$ for every $1\le i \le r$. By permuting the points, if necessary, assume that
$m_1 \ge \ldots \ge m_r$. So $L = dH-m_1E_1-\ldots-m_rE_r$ and $d >
m_1+\ldots+m_e$, by hypothesis. Also, we have $C_1 = eH-E_1-\ldots-E_r$.
We use the Nakai-Moishezon criterion for ampleness.
We first verify that $L$
meets all irreducible curves $D$ on $X$ positively.
If
$D=C_1$, then $L \cdot D >0$ by hypothesis.
So assume
that $D \ne C_1$. Write $D = fH-n_1E_1-\ldots-n_rE_r$
with $f \ge 0$. If $f=0$, then $D = -n_1E_1-\ldots-n_r E_r$ is
effective and this implies that $D=E_i$ for some $i$. Indeed, first
note that not all $n_i$ can be non-negative, because in that case $D$
is negative of an effective curve. If $n_i < 0$ for some $i$, then $D
\cdot E_i = n_i < 0$. So $E_i$ is a component of $D$ and hence
$D=E_i$. By hypothesis, $L \cdot E_i = m_i > 0$.
Assume now that $f > 0$. Since $D \ne E_i$ for any $i$, it follows
that $n_i \ge 0$ for all $i$.
Since $D \ne C_1$, we have $ef \ge
n_1+\ldots+n_r$.
By hypothesis, $d > m_1+\ldots+m_e$. Hence we have
\begin{eqnarray*}
df &>& m_1f+\ldots+m_ef \\
&=& m_1n_1+\ldots+m_en_e+m_1(f-n_1)+\ldots+m_e(f-n_e)\\
&\ge& m_1n_1+\ldots+m_en_e+m_{e+1}(ef-n_1-n_2-\ldots-n_e) \\
&\ge& m_1n_1+\ldots+m_en_e+m_{e+1}(n_{e+1}+\ldots+n_r) \\
&\ge&
m_1n_1+\ldots+m_en_e+m_{e+1}n_{e+1}+m_{e+2}n_{e+2}+\ldots+m_rn_r.
\end{eqnarray*}
Hence $L\cdot D > 0$.
The condition $L^2 > 0$ follows now because of
Proposition \ref{standard-is-effective}, which says that $L$ is
effective if $L\cdot H \ge L\cdot (E_1+\ldots+E_e)$.
\end{proof}
\begin{proposition}\label{standard-is-effective}
With $X$ as in Theorem \ref{main}, let $L = dH-m_1E_1-\ldots-m_rE_r$
be a line bundle on $X$ with $m_1\ge m_2 \ge \ldots \ge m_r$.
If $d \ge m_1+\ldots+m_e$, then $L$ is effective.
\end{proposition}
\begin{proof}
This proposition is completely analogous to \cite[Lemma 1.4]{H1} and has
the same proof. We give a proof for completeness.
We show that $L$ can be written as a non-negative linear combination
of line bundles of the form $H$, $H-E_1$, $2H-E_1-E_2,\ldots,
(e-1)H-E_1-\ldots-E_{e-1}$, $eH-E_1-\ldots-E_i$ for
$i=e,e+1,\ldots,r$. Since all these are effective, so is $L$.
Let $s$ be the largest index in $\{1,2,\ldots,r\}$ such that $m_s \ne 0$.
If $s \le e-1$, we
have
$L =
m_s(sH-E_1-\ldots-E_s)+(m_{s-1}-m_s)\left((s-1)H-E_1-\ldots-E_{s-1}\right)+
\ldots+(m_1-m_2)(H-E_1)+(d-m_1-m_2-\ldots-m_s)H$.
If $s \ge e$, let $L' =
L-m_s(eH-E_1-\ldots-E_s)$. Then it is easy to see that $L'$ satisfies
the hypotheses of the proposition and the value of $s$ is smaller for it. So we are done
by induction on $s$.
\end{proof}
\begin{remark}\label{collinear}
In Theorem \ref{main}, the only hypothesis which is not in general
necessary for ampleness is the condition that $L\cdot H >
L\cdot(E_1+\ldots+E_e)$. But this condition is necessary if $e$ points
among $p_1,\ldots,p_r$ are collinear (assume that $e < r$). For instance, consider a line
$l$ in $\mathbb{P}^2$
that meets $C$ in $e$ distinct smooth points, say $p_1,\ldots,p_e$.
Then choose any other $r-e$
points. If $L$ is ample, then $L$ meets the proper transform of $l$
positively. So we have $L\cdot H >
L\cdot(E_1+\ldots+E_e)$ after permuting the exceptional divisors, if
necessary. Thus in order to be ample for {\it all} choices of $r$
distinct smooth points on $C$, $L$ must satisfy $L\cdot H >
L\cdot(E_1+\ldots+E_e)$ after a suitable permutation of
$E_1,\ldots,E_r$.
\end{remark}
In Corollary \ref{big-r}, we address the {\it
uniform} case (i.e., $m_1=\ldots=m_r=m$), where we can make more
precise statements.
\begin{corollary}\label{big-r}
Let $X$ be as in Theorem \ref{main}. Let $L
= dH-m\sum_{i=1}^r E_i$ be a
line bundle on $X$. If $r \ge e^2$, $L$ is ample if and only if $L \cdot C_1 >
0$. If $r < e^2$ then $L$ is ample if $d> em$.
\end{corollary}
\begin{proof}
The proof is immediate from Theorem \ref{main}.
Indeed, when $r \ge e^2$, the hypothesis gives $de>rm$ which implies
$d> \frac{r}{e}m \ge m$. When $r<e^2$, the hypothesis gives $de>e^2m>
rm$.
\end{proof}
We note that when $r < e^2$ the condition $d>em$ is also {\em
necessary} if $e$ of the points are collinear. See Remark
\ref{collinear}.
\begin{corollary}\label{nef}
With the set-up as in Theorem \ref{main}, $L$ is nef if $L\cdot C_1
\ge 0$ and $L\cdot H \ge L\cdot (E_1+\ldots+E_e)$. Further, if $r \ge
e^2$ and $L = dH-m\sum_{i=1}^r E_i$, then $L$ is nef if and only if
$L\cdot C_1 \ge 0$.
\end{corollary}
\begin{proof}
For
nefness, we only need to check that $L \cdot D \ge 0$ for all
effective curves. This is immediate from the proof of Theorem
\ref{main} and Corollary \ref{big-r}.
\end{proof}
\begin{remark}
In \cite{H1,H2}, Harbourne considers the case $e=3$. He defines a line
bundle $L=dH-m_1E_1-\ldots-m_rE_r$ to be {\it standard} with respect to
the {\it exceptional configuration} $\{H,E_1,\ldots,E_r\}$ if $d \ge
m_1+m_2+m_3$. Our hypothesis that $d\ge m_1+\ldots+m_e$ may be
considered as a generalization of the notion of standardness to the
case of arbitrary $e$. Harbourne defines $L$ to be {\it excellent}
if it is standard and $L\cdot C_1 > 0$. Suppose that $m_i > 0$ for
every $i$. One of the main results in
\cite{H1,H2} says that $L$ is ample {\em if and only if} it is excellent
with respect to some exceptional configuration. See \cite{H1,H2} for
more details.
Our main Theorem \ref{main} may be considered as a generalization of
one direction of this result to the case of arbitrary $e$. The converse is not true
when $e \ge 4$. See Example \ref{e=4}.
\end{remark}
\begin{example}\label{e=3}
Take $e=3$ and $r=10$. Let $L =
8H-3(E_1+E_2+E_3)-2(E_4+\ldots+E_{10})$. Consider the following transformation:
$H\mapsto 2H-E_1-E_2-E_3$,
$E_1\mapsto H-E_2-E_3$,
$E_2\mapsto H-E_1-E_3$,
$E_3\mapsto H-E_1-E_2$, and
$E_i \mapsto E_i$ for $i = 4,\ldots,10$.
Under this, $L$ is transformed to $7H-2(E_1+\ldots+E_{10})$. This is standard, in
fact excellent, in
the sense of \cite{H1,H2}. So $L$ is ample. One can also check that
$L$ is ample as in Example \ref{e=4}. This example illustrates one of
the main theorems in \cite{H1} which says that a line bundle $L$ is ample if and
only if $L\cdot C_1 > 0$ and $L$ is standard with respect to {\it some}
exceptional configuration.
\end{example}
\begin{example}\label{optimal}
Let $e=4$ and $r=17$. So $X$ is the blow up of $\mathbb{P}^2$ at 17
distinct smooth points on an irreducible,
reduced plane quartic. Let $L =
11H-3(E_1+E_2+\ldots +E_{13})-(E_{14}+\ldots+E_{17})$. Then $L\cdot
C_1 = 44-39-4=1$, but $L^2=121-117-4=0$. So $L$ is not ample.
Here note that $d=11 < m_1+m_2+m_3+m_4=12$.
So the
hypotheses in Theorem \ref{main} can not be weakened.
\end{example}
\begin{example}\label{e=4}
In this example, we show that the hypotheses of Theorem \ref{main} are
not always necessary for ampleness.
Let $C$ be an irreducible and reduced plane quartic and let
$p_1,\ldots,p_{18}$ be distinct smooth points on $C$ such that no four are
collinear.
Let $L =
10H-3(E_1+E_2+E_3)-2(E_4+\ldots+E_{18})$. Since $d= 10< 3+3+3+2=11$, the
hypotheses of Theorem \ref{main} are not
satisfied. However, we claim that $L$ is ample.
It is easy to check that $L^2 = 13$ and $L\cdot C_1 = 1$. So let $D
\ne C_1$ be an irreducible and reduced curve. Write $D =
fH-\sum_{i=1}^{18} n_iE_i$.
If $f=0$, we may assume that
$n_i \ne 0$ for some $i$. We claim in fact that $n_i \le 0$ for all $i$.
Since $D=-(\sum_i n_iE_i)$ is effective,
$n_i$ can not all be non-negative, as in that case $D$ is negative
of an effective divisor. So $n_i < 0$ for some $i$. Then $D\cdot E_i =
n_i < 0$, so that $E_i$ is a component of $D$. Subtracting $E_i$ from
$D$ for all $i$ with $n_i < 0$, we obtain an effective divisor of the
form $\sum\limits_{n_j > 0}(-n_j) E_j$, which must be the zero divisor.
Hence $L \cdot D = \sum_i (-n_i) > 0$.
Now let $f \ge 1$ and $n_1,
\ldots, n_{18} \ge 0$. In fact, if $f=1$, since no four points are
collinear, we have $L
\cdot D > 0$. So let $f \ge 2$.
Since $C_1$ and $D$ are distinct irreducible curves, $C_1 \cdot D = 4f
-\sum_{i=1}^{18} n_i \ge 0$.
Then
$L\cdot D =
10f-3(n_1+n_2+n_3)-2(n_4+\ldots+n_{18}) = \left(8f-2\sum_{i=1}^{18}n_i\right)
+2f-(n_1+n_2+n_3) \ge 2f-(n_1+n_2+n_3).$
If $n_1+n_2+n_3 = 0$, it
follows that $L \cdot D > 0$. Otherwise, without loss of generality,
let $n_1 > 0$.
Intersecting $D$ with
the proper transforms of the line through $p_1,p_2$ and the line
through $p_1,p_3$, we get $f \ge n_1+n_2$ and $f\ge n_1+n_3$. Hence
$2f\ge 2n_1+n_2+n_3 > n_1+n_2+n_3$. The last inequality holds because $n_1
> 0$. Thus $L \cdot D > 0$.
Though $L = 10H-3(E_1+E_2+E_3)-2(E_4+\ldots+E_{18})$ in
this example is ample, it is easy to see that $L$ does not satisfy the condition $d \ge
m_1+m_2+m_3+m_4$ with respect to {\em any} exceptional configuration. This
is easy to check by direct calculation. Note also that $L$ is already
standard in the sense of Harbourne \cite{H1,H2}.
\end{example}
\section{$k$-very ampleness}\label{k-very-ample}
\iffalse
Let $k$ be a non-negative integer. A line bundle $L$ on a projective
variety is said to be {\it $k$-very ample} if the restriction map
$$H^0(X,L) \to H^0(X,L\otimes \mathcal{O}_Z)$$
is surjective for all zero-dimensional subschemes $Z \subset X$ of
length $k+1$, i.e., dim$(H^0(Z,\mathcal{O}_Z)) = k+1$.
Note that 0-very ampleness is equivalent to global generation and
1-very ampleness is equivalent to very ampleness.
This notion has been introduced in a series of papers by Beltrametti,
Francia and Sommese. See \cite{BFS,BS1,BS2} for more details.
general theorem is proved in \cite{BS3} for $k$-very ampleness on
blown up projective varieties. \cite{VT} studies $k$-very ampleness
for line bundles on blow ups of $\mathbb{P}^2$ at general points on an
irreducible and reduced cubic. \cite{ST} studies $k$-very ampleness
for line bundles on blow ups of $\mathbb{P}^2$ at general points.
\fi
Let $C, p_1,\ldots,p_r, X$ be as in Section \ref{ample}. In this
section we consider a
line bundle $L =dH-m\sum_{i=1}^r E_i$ on
$X$ and
investigate $k$-very ampleness of $L$ for a non-negative integer
$k$. We make the assumption that the number of points we blow up is
large compared to $e$. Specifically, we assume that $r \ge e^2+k+1$.
First, we consider the question of global generation.
In other words, we assume $k=0$.
Our arguments for $k=0$ give a flavour of our arguments in the
case $k \ge 1$, which we consider later.
We will use Reider's theorem \cite[Theorem 1]{Re} which gives conditions for global
generation and very ampleness. We only state the conditions for global
generation below.
\begin{theorem}[Reider]\label{reider}
Let $X$ be a smooth complex surface and let $N$ be a nef line bundle on $X$ with $N^2
\ge 5$. If $K_X+N$ is not
globally generated (here $K_X$ is the canonical line bundle of $X$), then there exists an effective divisor $D$ on $X$ such
that
$$D\cdot N = 0, D^2=-1, {\rm ~or~}
D\cdot N = 1, D^2 = 0.$$
\end{theorem}
The following is our theorem on global generation.
\begin{theorem}\label{bpf}
Let $C$ be an irreducible and reduced plane curve of
degree $e$ and let $X \to \mathbb{P}^2$ be the blow up of $\mathbb{P}^2$ at $r$
distinct smooth points $p_1,\ldots,p_r \in C$. Let $H$
denote the pull-back of $\mathcal{O}_{\mathbb{P}^2}(1)$ and $E_1,\ldots,E_r$
the inverse images of $p_1,\ldots,p_r$ respectively.
Let $L =dH-m\sum_{i=1}^r E_i$ be a line bundle on $X$ with $m\ge 0$.
If $(d+3)e>r(m+1)$ and $r \ge e^2+1$, then
$L$ is globally generated.
\end{theorem}
\begin{proof}
Since the conclusion holds if $m=0$, we
assume $m \ge 1$.
Let $N = L-K = (d+3)H-(m+1)\sum_{i=1}^r E_i$.
If $C_1$ denotes the proper transform of $C$ on
$X$, then $N\cdot C_1 = (d+3)e-r(m+1) > 0$, by hypothesis. Moreover,
$d+3 > \frac{r}{e}(m+1) > e(m+1)$, since $r> e^2$. So the
hypotheses of Theorem \ref{main} hold and $N$ is
ample.
Further, $N^2 = (d+3)^2-r(m+1)^2 > \left(\frac{r^2}{e^2}-r\right)(m+1)^2 >
(m+1)^2 \ge 4$. This is because $\frac{r^2}{e^2}-r\ge
r(\frac{e^2+1}{e^2})-r = r(1+\frac{1}{e^2})-r = \frac{r}{e^2} >
1$. So we can apply Reider's
Theorem \ref{reider}.
Suppose that $L$ is not globally generated.
By Theorem \ref{reider}, there is an effective divisor $D$
such that $D\cdot N = 1, D^2 = 0$.
Since $N$ is ample, this is the only
possibility. Writing $D = fH-\sum_{i=1}^r n_iE_i$ and setting $n = \sum_{i=1}^r n_i$,
we have $$D \cdot N = (d+3)f - n(m+1) = 1 \text{~and~} D^2 = f^2 -\sum_{i=1}^r
n_i^2 = 0.$$
Note that if $n_i < 0$ for some $i$, then $D \cdot E_i < 0$. Thus $E_i$ is a
component of $D$ and $D-E_i$ is effective. But then we get a
contradiction because $1 = D\cdot N =
N\cdot (D-E_i) + N\cdot E_i \ge m+1 \ge 2$ (since $N$ is ample). Hence
$n_i \ge 0$ for all $i$ and $n =\sum_i n_i \ge 0$. Since $D^2 =0$, in fact $n > 0$.
We consider two different cases: $n \ge r$ or $n < r$.
First, suppose that $n \ge r$. Since $f^2
= \sum_i n_i^2 \ge \frac{n^2}{r}$ and by hypothesis, $d+3 >
\frac{r}{e}(m+1)$, we have
\begin{eqnarray}\label{3.1-a}
1 = D\cdot N = (d+3)f-n(m+1) >
\frac{n\sqrt{r}}{e}(m+1)-n(m+1).
\end{eqnarray}
We claim that $\frac{n\sqrt{r}}{e}-n \ge 1/2$ for any fixed $e$, all
$r\ge e^2+1$
and for all $n \ge r$. Since this is a linear function in $n$ of
positive slope, it suffices to show that the function is non-negative
when $n=r$. That is, we only have to show that $\frac{(\sqrt{r})r}{e}-r
\ge 1/2$ for $r \ge e^2+1$. For a fixed $e$, this function is
increasing for $r > 0$. So it suffices to show that
$\frac{\sqrt{e^2+1}(e^2+1)}{e}-(e^2+1) \ge 1/2$. It is easy to
see that this inequality holds for $e \ge 1$, for example by
clearing the denominator and squaring.
Thus, by \eqref{3.1-a}, $ 1>\left(\frac{n\sqrt{r}}{e}-n\right)(m+1)
\ge \frac{m+1}{2} \ge 1$, which is a contradiction.
Finally, we consider the case $n < r$. We have $f^2 = \sum_{i=1}^r n_i^2 \ge
n$. So $1 = N\cdot D = (d+3)f-n(m+1) \ge \left(\frac{r\sqrt{n}}{e}
-n\right)(m+1)$. We claim that $\frac{r\sqrt{n}}{e}
-n \ge 1$, which as above leads to a contradiction.
We view $\frac{r\sqrt{n}}{e} -n$ as a quadratic in $\sqrt{n}$. Since
the leading coefficient is -1, it is a downward sloping parabola. If
we show that the value of this function is at least 1 for $n=1$ and
$n=r-1$, then it follows that the value of the function is at least 1
for all $1 \le n \le r-1$. This can be easily verified.
We conclude that $L$ is globally generated.
\end{proof}
Now we will consider the case $k \ge 1$.
Recall the criterion \cite[Theorem 2.1]{BFS} of Beltrametti-Francia-Sommese for $k$-very
ampleness of $L$, which generalizes Reider's criterion in Theorem \ref{reider}.
\begin{theorem}[Beltrametti-Francia-Sommese]\label{bfs-theorem}
Let $N$ be a line bundle on a surface $X$. Let $k \ge 0$ be an integer.
Suppose that $N$ is nef and $N^2 \ge 4k+5$.
If $K_X+N$ is not
$k$-very ample, then there exists an
effective divisor $D$ on $X$ such that
\begin{eqnarray}\label{bfs}
N\cdot D - k-1 \le D^2 < \frac{N\cdot D}{2} < k+1.
\end{eqnarray}
\end{theorem}
If \eqref{bfs} holds for an effective divisor $D$, then our next two
results give some conditions that $D$ must satisfy.
\begin{proposition} \label{n-greater-than-r}
Let $X$ be as in Theorem \ref{bpf}. Let $L =dH-m\sum_{i=1}^r E_i$
and $N = L-K_X =(d+3)H-(m+1)\sum_{i=1}^r E_i$ be line bundles on $X$.
Let $k\ge 1$ be an integer.
Suppose that $r \ge e^2+k+1$, $(d+3)e > r(m+1)$ and $m\ge k$. If an
effective
divisor $D=fH-\sum_{i=1}^r
n_iE_i$ on $X$ satisfies \eqref{bfs}, then $\sum_{i=1}^r n_i < r$.
\end{proposition}
\begin{proof}
Set $n=\sum_{i=1}^r n_i$.
Let $\alpha = N\cdot D = (d+3)f-(m+1)n$ and $\beta = D^2 = f^2 -
\sum_{i=1}^r n_i^2$. If $D$ satisfies \eqref{bfs}, then we have
\begin{eqnarray}\label{bfs-1}
\alpha - k-1 \le \beta < \frac{\alpha}{2} < k+1.
\end{eqnarray}
By hypothesis,
$(d+3)^2 > \frac{r^2}{e^2}(m+1)^2$. Since $(d+3)f=\alpha+(m+1)n$, we have
$$(d+3)^2f^2 = (m+1)^2n^2+2(m+1)n\alpha+\alpha^2 >
\frac{r^2}{e^2}(m+1)^2\left(\beta+\sum_{i=1}^r n_i^2\right).$$
Since $r\sum_{i=1}^r n_i^2 \ge
n^2$, we have
\begin{eqnarray*}
(m+1)^2n^2+2(m+1)n\alpha+\alpha^2 > \frac{r^2}{e^2}(m+1)^2\beta+\frac{r}{e^2}(m+1)^2n^2.
\end{eqnarray*}
We now show that the above inequality is impossible for $n \ge r$.
Specifically, we make the following claim.
{\bf Claim:} Let $r,e,k,\alpha, \beta$ be as in the proposition and
suppose that \eqref{bfs-1} holds. Then for $n\ge r$, we have
$$\frac{r^2}{e^2}(m+1)^2\beta+\frac{r}{e^2}(m+1)^2n^2 \ge
(m+1)^2n^2+2(m+1)n\alpha+\alpha^2.$$
{\bf Proof of Claim:} We consider the difference of the two terms in
the required inequality as a quadratic function in $n$.
Define
$$\lambda(n) :=
\left(\frac{r}{e^2}-1\right)(m+1)^2n^2-2(m+1)n\alpha+\frac{r^2\beta}{e^2}(m+1)^2-\alpha^2.$$
This is quadratic in $n$ with
the leading coefficient $\left(\frac{r}{e^2}-1\right)(m+1)^2 > 0$.
We will show that $\lambda(r) \ge 0$ and $\lambda'(r) \ge 0$, which will
prove the claim and the proposition.
\begin{eqnarray*}
\lambda(r) &=& r^2(m+1)^2\left(\frac{r+\beta}{e^2}-1\right)
-2(m+1)\alpha r-\alpha^2\\
&\ge& r^2(m+1)^2\left(\frac{e^2+k+1+\beta}{e^2}-1\right)
-2(m+1)\alpha r-\alpha^2 \hspace{.3in} ({\rm since~} r \ge e^2+k+1)\\
&=& r^2(m+1)^2\left(\frac{k+1+\beta}{e^2}\right)
-2(m+1)\alpha r-\alpha^2\\
&\ge& r(e^2+k+1)(m+1)^2\left(\frac{k+1+\beta}{e^2}\right)
-2(m+1)\alpha r-\alpha^2\\
&=& r(m+1)^2(k+\beta+1)+ \frac{r(k+1)(m+1)^2(k+\beta+1)}{e^2}-2(m+1)\alpha r-\alpha^2\\
&\ge& r(m+1)^2(k+\beta+1)+ (k+1)(m+1)^2(k+\beta+1)-2(m+1)\alpha r-\alpha^2\\
&\ge& 0.
\end{eqnarray*}
The last inequality follows when we compare the first term with the
third term and the second term with the fourth term. We use the
inequalities
$\alpha \le \beta+k+1$ and $\alpha < 2(k+1)$ which hold by \eqref{bfs-1},
and $m\ge k \ge 1$, which holds by hypothesis.
Next we show that $\lambda'(r)\ge 0$.
$\lambda'(n) = 2\left(\frac{r}{e^2}-1\right)
(m+1)^2n-2(m+1)\alpha$. Thus
\begin{eqnarray*}
\lambda'(r) &=& \left(\frac{2r^2}{e^2}-2r\right)
(m+1)^2-2(m+1)\alpha\\
&\ge& \frac{2r(e^2+k+1)}{e^2}(m+1)^2-2r(m+1)^2-2(m+1)\alpha\\
&=& 2r(m+1)^2+2r(m+1)^2\frac{k+1}{e^2}-2r(m+1)^2-2(m+1)\alpha\\
&=&2r(m+1)^2\frac{k+1}{e^2}-2(m+1)\alpha\\
&\ge& 2(m+1)^2(k+1)-2(m+1)\alpha\\
&\ge& 0 \hspace{.2in}(\text{by \eqref{bfs-1} and the hypothesis that $m \ge k \ge 1$}).
\end{eqnarray*}
This completes the proof of the proposition.
\end{proof}
\begin{proposition}\label{proper-with-c}
Let $L =dH-m\sum_{i=1}^r E_i$ and $N = L-K_X =(d+3)H-(m+1)\sum_{i=1}^r E_i$.
Let $k$ be a positive integer.
Suppose that $r \ge e^2+k+1$, $(d+3)e > r(m+1)$ and $m\ge k$. Let $D$
be an effective divisor on $X$ such that $D \cdot C_1 \ge 0$. Then $D$
does not satisfy \eqref{bfs}.
\end{proposition}
\begin{proof}
Let $D = fH-\sum_{i=1}^r n_iE_i$. Then $f \ge
0$ and if $f=0$, then $n_i \le 0$ for all $i = 1,\ldots, r$.
Let $n = \sum_{i=1}^r n_i$. Then $D^2 = f^2 -\sum_{i=1}^r n_i^2 \le
f^2-n$. Indeed, this follows because $\sum_{i=1}^r n_i^2 \ge n$.
Moreover, the assumption
$D\cdot C_1 \ge 0$ implies that $ef \ge n$.
Suppose that $D$ satisfies
\eqref{bfs}. We will obtain a contradiction.
First let $f=0$. We have $N \cdot D = -(m+1)n < 2(k+1) \le 2(m+1)$. So
$n > -2$. On the other hand, $0 = ef \ge n$. So $n=0$ or $n=-1$. Since
each $n_i$ is non-positive, if $n=0$ then $D=0$. If $n=-1$, then
$D=E_i$ for some $i$. But then $N\cdot D = m+1 \ge k+1$, hence
$N\cdot D -k-1 \ge 0$, while $D^2 = -1$. This violates \eqref{bfs}.
Let $f=1$. In this case, we have $e \ge n$. We may also
assume $n > 0$.
Then $2(k+2) > N \cdot D = (d+3)-n(m+1) >
\left(\frac{r}{e}-n\right)(m+1) \ge (e-n)(m+1)+\left(\frac{k+1}{e}\right)(m+1) \ge
(e-n)(k+1)+\left(\frac{k+1}{e}\right)(k+1)$. Thus $0\le e-n< 2$. So $e=n$ or $e=n+1$.
Let $e=n+1$. Then $0 \ge 1-n \ge D^2 \ge N\cdot D -k-1 >
k+1+\left(\frac{k+1}{e}\right)(k+1)-k-1> 0$, which is a contradiction. If $e=n$,
then $0 \ge 1-n \ge D^2 \ge N\cdot D -k-1 >
\left(\frac{k+1}{e}\right)(k+1)-k-1$. Hence $k+1< e$. But then $D^2 \le 1-n = 1-e
< -k \le N\cdot D -k-1$. The last inequality holds because $N$ is
ample and hence $N \cdot D > 0$. Again we have a
contradiction, because this violates \eqref{bfs}.
Now suppose that $f\ge 2$. As above, we have
$2(k+1) > N\cdot D > \frac{rf(m+1)}{e}-n(m+1) \ge \frac{rf(m+1)}{e}-ef(m+1) =
\left(\frac{rf}{e}-ef\right)(m+1) \ge
\left(\frac{(k+1)f}{e}\right)(m+1).$
\begin{eqnarray}\label{3.4-a}
{\rm Thus~} 2 > \frac{(m+1)f}{e}, {\rm ~or~ equivalently,~} f < \frac{2e}{m+1}.
\end{eqnarray}
If $e \le k+1$, then $f < \frac{2e}{m+1} \le \frac{2e}{k+1} \le 2$,
which contradicts the hypothesis $f \ge 2$. So we may assume $e >
k+1$.
We now make the following claim:
{\bf Claim:} $f^2 - n \le 2-2k$.
{\bf Proof of Claim:}
Note that $ef-n \le 1$. Indeed,
$(d+3)f>\frac{rf}{e}(m+1)>ef(m+1)$. Hence $2(k+1) > N\cdot D =
(d+3)f-n(m+1) > (ef-n)(m+1)\ge (ef-n)(k+1)$. Thus $ef-n < 2$. On the
other hand, by the hypothesis in the proposition $ef-n \ge 0$. Hence
we have:
\begin{eqnarray}\label{3.4-b}
ef-n=0 {\text ~or~} ef-n=1.
\end{eqnarray}
On the other hand,
\begin{eqnarray}
f^2 -n &<& f\left(\frac{2e}{m+1}\right)-n \hspace{.2in} ({\rm by~}
\eqref{3.4-a}) \nonumber \\
&=& ef+\frac{ef(1-m)}{m+1}-n \nonumber\\
&\le& \frac{ef(1-k)}{k+1}+ef-n \hspace{.2in}({\rm ~since~} m \ge k)
\nonumber \\
&\le& \frac{ef(1-k)}{k+1}+1 \hspace{.2in}({\rm ~by~} \eqref{3.4-b})
\nonumber \\
&\le& f(1-k)+1 \hspace{.2in} ({\rm ~since ~}e > k+1 {\rm
~and ~} 1-k \le 0) \nonumber \\
&\le& 2(1-k) +1 = 3-2k \hspace{.2in} ({\rm ~since ~} f\ge 2 {\rm
~and ~} 1-k \le 0). \nonumber
\end{eqnarray}
This completes the proof of the claim. Now we consider three cases:
$\underline{k \ge 3}$: In this case, $f^2-n \le 2-2k \le -k-1$. So
$D^2 = f^2 - \sum_{i=1}^rn_i^2 \le f^2-n \le -k-1$. But
by \eqref{bfs}, we have $N\cdot D -k-1 \le D^2$. Since $N$ is ample,
$-k \le N\cdot D -k-1$, contradicting the inequality $D^2 \le -k-1$.
$\underline{k=1}$: Then by \eqref{bfs}, the claim, and the fact that $N \cdot D \ge
1$, we have
$-1 \le N\cdot D - 2 \le f^2-n \le 0$.
By \eqref{3.4-b}, we have either $ef=n$ or
$ef-1=n$.
If $ef=n$, then $0 \ge f^2-n = f(f-e) \ge -1$. Since
$f \ge 2$, the only possibility $f=e$. But this violates
\eqref{3.4-a}, because $m \ge k = 1$.
On other hand, if $ef-1=n$,
then $0 \ge f^2-n = f^2-ef+1=f(f-e)+1\ge -1$. Since $f \ge 2$, $f^2-n$
must be -1.
But then the only possibility is $f=2$ and $e=1$ and again we have a contradiction to \eqref{3.4-a}.
$\underline{k=2}$: By \eqref{bfs}, the claim, and the fact that $N \cdot D \ge
1$, we have
$-2 \le N\cdot D - 3 \le f^2-n \le -2$.
Hence $f^2-n=-2$.
If $ef=n$, then $-2=f^2-ef=f(f-e)$. This contradicts \eqref{3.4-a}.
If $ef-1 = n$, then $-2=f(f-e)+1$.
Again we obtain a contradiction to
\eqref{3.4-a}.
This completes the proof of the proposition.
\end{proof}
Now we are ready to prove our main result on $k$-very ampleness.
\begin{theorem}\label{main1}
Let $C$ be an irreducible and reduced plane curve of
degree $e$. Let $X \to \mathbb{P}^2$ be the blow up of $\mathbb{P}^2$ at $r$
distinct smooth points $p_1,\ldots,p_r \in C$. Let $H$
denote the pull-back of $\mathcal{O}_{\mathbb{P}^2}(1)$ and let $E_1,\ldots,E_r$
be the inverse images of $p_1,\ldots,p_r$ respectively. Let $k$ be a
non-negative integer.
Let $L=dH-m\sum_{i=1}^r E_i$ be a line bundle on $X$ with $m
\ge k$.
If $(d+3)e>r(m+1)$ and $r \ge e^2+k+1$, then $L$ is $k$-very ample.
\end{theorem}
\begin{proof}
When $k=0$, this is the same as Theorem \ref{bpf}. So we will assume that
$k \ge 1$ and use the criterion of Beltrametti-Francia-Sommese.
Let $N = L-K = (d+3)H-(m+1)\sum_{i=1}^r E_i$. Just as in the proof of
Theorem \ref{bpf}, we conclude that $N$ is ample.
Next we claim that $N^2 \ge 4k+5$. Indeed, we have
\begin{eqnarray*}
N^2 &=& (d+3)^2-r(m+1)^2 \\
&>& \frac{r^2}{e^2}(m+1)^2-r(m+1)^2\\
&=&(m+1)^2r\left( \frac{r}{e^2}-1\right)\\
&\ge& (m+1)^2r\left(\frac{k+1}{e^2}\right)\\
&>& (m+1)^2(k+1) \\
&\ge& 4(k+1) \hspace{.5in} ({\rm since~} m\ge k \ge 1).
\end{eqnarray*}
Since $N^2$ is an integer and $N^2 > 4k+4$, we conclude that $N^2 \ge
4k+5$. Hence we can apply the criterion of Beltrametti-Francia-Sommese.
Suppose that $L$ is not $k$-very ample. Then there exists an effective
divisor $D$ on $X$ such that \eqref{bfs} holds. We will obtain a
contradiction.
Write $D = fH - n_1E_1-\ldots-n_rE_r$
and let $n= \sum_i n_i$. By Proposition \ref{n-greater-than-r}, we have
$n<r$. By Proposition \ref{proper-with-c},
$D \cdot C_1 < 0$. This implies that $C_1$ is a component of $D$ and we
may write $D = aC_1 + D'$ for a positive integer $a$ and an effective
divisor $D'$.
Write $D' = bH - \sum_{i=1}^r l_i E_i$ with $b \ge 0$. Then we have
$f=ae+b$. So
\begin{eqnarray*}
2(k+1) &>& N\cdot D \\
&=& (d+3)f-(m+1)n \\
&>& \frac{r}{e}(ae+b)(m+1)-(r-1)(m+1) \hspace{.2in} ({\rm since~}
n\le r-1)\\
&=& (r(a-1)+1)(m+1)+\frac{rb}{e}(m+1)\\
&\ge& (r(a-1)+1)(k+1)+\frac{rb}{e}(k+1).
\end{eqnarray*}
Thus $a=1$ and $b=0$. In particular, $f=e$.
We have $2(k+1) > N\cdot D = (d+3)e-n(m+1) > (r-n)(m+1)$. Hence $r-n <
2$. On the other hand, $r-n \ge 1$. So $r-n=1$.
Since $N$ is ample, $N\cdot D > 0$. Thus $-k \le N\cdot D -k - 1\le
D^2 = e^2 -\sum_{i=1}^r n_i^2 \le e^2 -n = e^2-r+1 \le -k$. The last
inequality holds because $r \ge e^2+k+1$. Thus we have $D^2 = -k $
and $N\cdot D =1$. But $N\cdot D > m+1 \ge k+1 > 1$. This is a
contradiction.
The proof of the theorem is complete.
\end{proof}
\begin{remark}
\cite[Theorem 4.1]{VT} gives conditions for $k$-very ampleness for any
line bundle on the blow up of $\mathbb{P}^2$ at general points on an
irreducible and reduced cubic. In our context, this is the case
$e=3$. If the line bundle is uniform, that is if $L = dH-m\sum_{i=1}^r
E_i$ and if the number of points $r$ is at least $10+k$, then our Theorem
\ref{main1} is comparable to this result. However, we note that \cite[Theorem
4.1]{VT} deals with any (not just uniform) line bundle $L = dH-\sum_{i=1}^r m_iE_i$
and any $r
\ge 3$.
\end{remark}
The next two examples show that
the hypothesis of
Theorem \ref{main1} can not be weakened for $e=3$.
\begin{example}\label{strict-inequality}
Let $C$ be a smooth plane cubic. Let $X$ be the blow up of 10
distinct points on $C$. Consider the line bundle $L =
7H-2\sum_{i=1}^{10} E_i$ on $X$. We
have $C_1 = 3H-\sum_{i=1}^{10} E_i$. By Corollary \ref{big-r}, $L$ is
ample. We use \cite[Corollary 1.4]{BS1} to show that $L$ is not
globally generated. According to this result, if a line bundle on
a curve of positive genus is $k$-very ample, then the degree of the
line bundle is
at least $k+2$. In our example, if $L$ is globally generated (that is,
if it is 0-very ample), then $L_{|C}$ is also $0$-very ample on $C$.
But deg$(L_{|C}) = L\cdot C =1 < 2$.
So the strict inequality, $e(d+3) > r(m+1)$, in the hypothesis of
Theorem \ref{main1} can not be relaxed.
\end{example}
\begin{example} \label{points}
This is a small variation on Example \ref{strict-inequality}.
Again let $e=3, r= 10$, but now let $m =7$. Consider $L =
24H-7\sum_{i=1}^{10} E_i$. It is easy to check that $L$ is ample (by
Corollary \ref{big-r}) and globally generated (by Theorem
\ref{main1}). But $L$ is not very ample because $L\cdot C_1 = 2 <
1+2$ (again by \cite[Corollary 1.4]{BS1}). Here the hypothesis in
Theorem \ref{main1} on the number of points (namely, $r \ge e^2+k+1=10+k)$
does not hold.
\end{example}
\begin{example} If $e > 3$, our hypotheses in Theorem \ref{main1} are
not likely to be
optimal. We will illustrate this with just one
example.
Let $e=5, k = 5, r=31$ and consider $L_d =
dH-5(E_1+\ldots+E_{31})$. By Theorem \ref{main1}, $L_{35}$ is 5-very
ample. On the other hand, \cite[Corollary 1.4]{BS1} shows that $L_{32}$ is {\it
not} $5$-very ample. We do not know if $L_{33}$ or $L_{34}$ is 5-very
ample. Note that the criterion of Beltrametti-Francia-Sommese
(Theorem \ref{bfs-theorem}) can not be applied here, since $N_d=L_d-K_X$ is not
nef for $d < 35$. Indeed, $N_d\cdot C_1 = 5(d+3)-186 < 0$, for $d <
35$. In other words, our method (which is to use Theorem \ref{bfs-theorem}
to show $k$-very ampleness) itself is
not applicable to $L_{33}$ and $L_{34}$.
\end{example}
{\bf Acknowledgements:} We sincerely thank Tomasz Szemberg
for carefully reading this paper, pointing out some
mistakes and making useful suggestions.
We also thank the referee for pointing out some mistakes and making
numerous suggestions which improved the exposition of the paper.
\bibliographystyle{plain}
|
\section{Introduction}
It is now well established that at least two of the active neutrinos are massive.
New physics beyond the Standard Model(SM) is required to give
small but nonzero neutrino masses. A straightforward remedy is adding the right-handed neutrino(s) to the SM so that the active neutrinos can acquire Dirac masses after the SM electroweak symmetry breaking as what other charged fermions do. However, additional mass suppression mechanisms or very tiny Yukawa couplings, $\lesssim 10^{-12}\times Y_{top}$, are required to bring down the resulting Dirac neutrino masses to the sub-eV level. Alternatively, Majorana neutrino masses are sought to alleviate the problem of huge hierarchy among the Yukawa couplings in the Dirac neutrino cases.
Whatever the UV origin of Majorana neutrino mass is, the key is to generate
the dimension-5 Weinberg effective operator\cite{Weinberg:1979sa}, $(LH)^2$, where $L$ and $H$ are the SM lepton doublet and the Higgs doublet, respectively, at the low energy.
The Weinberg operator conserves the baryon number but violates lepton number by two units.
Since all the SM interactions at the low energies conserve both baryon and lepton numbers,
the new interactions responsible for generating the Majorana neutrino masses must break the lepton number and the relevant new degree(s) of freedom must carry lepton number.
If the relevant new fields also carry nonzero baryon number, there are no tree-level contributions to the Weinberg operator leading to a nature loop suppression to bring down the resulting Majorana neutrino masses.
Therefore, leptoquark, a boson which carries both lepton number and baryon number, is one of the well-motivated candidates to generate small Majorana neutrino masses without excessive fine tuning.
Moreover, since leptoquark participates strong interaction, it would be interesting that the new particles relevant to the neutrino mass generation mechanism could be directly probed at the hadron colliders.
However, it is impossible to generate the desired Weinberg operator by using only one leptoquark because the new interaction vertices always come in conjugated pairs.
Something else in the loop(s) which carries baryon number must also be utilized to have zero net baryon number and non-vanishing lepton number at the end.
The di-quark, a boson which carries $2/3$ of baryon number, is one of the candidates to work with leptoquark
for generating the Weinberg operator\footnote{It is also perfectly possible to generate nonzero neutrino masses with two leptoquarks with different lepton and baryon numbers, see for example\cite{LQ_1_loop_Wise, LQ_1_loop_2}. }.
Neutrino masses aside, the leptoquark and di-quark are common in many new physics models where the lepton number or the baryon number is not conserved\cite{LQ:Buchmuller,LQ:Dimopoulos,LQ:Dimopoulos2,LQ:Eichten,LQ:Angelopoulos,DQ:Hewett}, such as the grand-unified theories, technicolor and composite models.
Yet without positive results, leptoquark and di-quark had been eagerly searched for since the 1980's.
At the colliders, the leptoquark and di-quark could be produced and studied directly. However, the decay rates strongly depend on the unknown couplings between the leptoquark/di-quark and the SM fermions. Thus, the bounds are usually given with specific assumptions on their couplings to the SM fermions, see \cite{ATLAS-1st,ATLAS-2nd,ATLAS-3rd,CMS-1st,CMS-2nd,CMS-3rd,HERA-1st,LQ-CMS-ATLAS,DQ-mass}.
On the other hand, flavor changing processes could be mediated by the leptoquark or di-quark at the tree-level. Strong constraint
can be indirectly derived from the low energy flavor changing experiments\cite{2L2Q}.
Recently, an interesting application of utilizing the scalar leptoquark and scalar di-quark to generate the neutrino masses was discussed by \cite{Masaya}.
In \cite{Masaya}, one scalar leptoquark, $\Delta$ with SM $SU(3)_c\times SU(2)_L\times U(1)_Y$ quantum number $(3,1,-1/3)$, and one scalar di-quark, $S$ with SM quantum number $(6,1,-2/3)$, were augmented to the SM particle content and the neutrino masses can be generated through the two-loop radiative corrections\footnote{See \cite{Queiroz:2014pra} for a recent discussion on the potential connection between $\Delta$ and the dark matter. }.
This two-loop mechanism is very similar to that in the Zee-Babu model\cite{Zee-Babu} except
that $S/\Delta$ replaces the role of the doubly/singly charged scalar in Zee-Babu model. From now on this model is referred as the colored Zee-Babu Model(cZBM). In the cZBM, the resulting neutrino mass matrix pattern and the mixing angles are determined by $Y_L$ and $Y_S$, the Yukawa couplings between leptoquark and di-quark and the SM fermions, see Eq.(\ref{eq:Yukawa}).
Again, $Y_S$ and $Y_L$ are arbitrary and a priori unknown. To proceed, we consider the case that the symmetric $Y_S$'s are democratic and the magnitudes of the six $(Y_S)_{ij}$, where $i,j=1,2,3$ are the flavor indices, are of the same order. This could be realized in the extra-dimensional models with the right-handed down-type quark bulk wave functions cluster together in the extra spatial dimension(s), for applying the geometric setup to generate a special 4-dimensional Yukawa pattern see for example \cite{Chang:RS,Chang:SF}. With this working assumption and the fact that $m_b\gg m_s \gg m_d$, the $Y_L$ can be determined with some reasonable requirements which will be discussed later.
To accommodate all the neutrino data, the tree-level flavor violating processes will be inevitably mediated by $\Delta$ with the realistic $Y_L$ Yukawa couplings.
Moreover, the rates of these resulting tree-level and also those flavor violating processes induced at the loop level must comply with the current experimental bounds.
In addition to $Y_L$, $\Delta$ also admits Yukawa couplings, $Y_R$, which couple $\Delta$ to the right-handed leptons and quarks, see Eq.(\ref{eq:Yukawa}).
Since both $Y_L$ and $Y_R$ contribute to the tree-level flavor violating processes incoherently, $Y_R=0$ is assumed to minimize those rates.
A comprehensive numerical study is performed to search for the realistic configurations.
We find that sizable portion of the realistic solutions overlap with the designed sensitivities of the forthcoming lepton flavor violation experiments.
Moreover, for the neutrino masses in both the normal hierarchy and inverted hierarchy, there are interesting and definite lower bounds on $B(Z\rightarrow\overline{ l} l')$ and $B(l\rightarrow l' \gamma)$ which could be falsified in the future.
Also, the type of neutrino mass hierarchy can be determined if the charged lepton flavor violating double ratios are measured to be within some specific ranges.
If $Y_R=0$, the model has concrete predictions for the scalar leptoquark decay branching ratios for both neutrino mass hierarchies. This will help refine the collider search limit on the scalar leptoquark mass for the $\beta=1/2$ case.
The paper is organized as follows. A more detailed discussion on the model is given in section 2.
In section 3, we study the connection between the neutrino masses and $Y_L$, and the tree-level flavor violating processes as well. The loop-induced flavor violating processes are discussed in section 5. The numerical study are dealt with and discussed in section 5. Finally, the conclusions are summarized in section 6.
\section{Model and neutrino mass}
As mentioned in the previous section, the SM is extended by adding $S$ and $\Delta$.
After rotating the lepton fields into their weak basis and the quarks into their mass basis, the most general gauge invariant Yukawa interaction associated with $S$ and $\Delta$ is
\begin{equation}
\mathscr{L}_Y=-\left[\overline{L_{i}^C}(Y_L)_{ij}i\sigma_2 Q_j
+\overline{(\ell_{Ri})^C}(Y_R)_{ij}u_{R j}\right] \Delta^{*}
-\overline{(d_{Ri})^{C}} (Y_s)_{ij} d_{Rj} S^{*}
+ y^\Delta_{ij} \overline{(u_{R i})^C} d_{R j} \Delta
+h.c.\label{eq:Yukawa}
\end{equation}
where $i,j$ are the flavor indices and the $SU(3)$ indices are suppressed.
Apparently $Y_S$ is symmetric in the flavor space while there is no such constraints on $Y_L$, $Y_R$, and $y^\Delta$.
Moreover, the lagrangian admits a gauge invariant triple coupling term: $ (\mu \Delta^* \Delta^* S +h.c.)$.
As shown in Fig. \ref{neutrino-mass}, the neutrino masses will receive nonzero contributions through 2-loop quantum corrections if both $Y_L$ and $Y_S$ present. If one writes the effective Lagrangian for neutrino masses as $-\frac12 \overline{\nu_{Li}^C}(M_\nu)_{ij} \nu_{Lj}$, the neutrino mass matrix can be calculated to be
\begin{align}
\label{eq:nu_mass_elements}
&(M_\nu)_{ii'}= 24\mu (Y_L)_{ij}m_{dj} I_{jj'} (Y_s^\dagger)_{jj'}m_{dj'}(Y^T_{L})_{j'i'}\,,\\
&I_{jj'}=\int \frac{d^4k_1}{(2\pi)^4}\frac{d^4k_2}{(2\pi)^4}\frac{1}
{(k_1^2-m_{dj}^2)}\frac{1}
{(k_1^2-m_{\Delta}^2)}\frac{1}
{(k_2^2-m_{dj'}^2)}\frac{1}
{(k_2^2-m_{\Delta}^2)}\frac{1}{(k_1+k_2)^2-m_S^2}\,.
\end{align}
Note that the two-loop integral is similar to the one in the Zee-Babu model\cite{ZBM,Analytical-ZBM}.
When the down-type quark mass is much lighter than colored scalars, the integral is flavor independent and it can be simplified to
\begin{align}
&I_{jj'}\simeq I_\nu \equiv
\frac{1}{(4\pi)^4}\frac{1}{M^2}\frac{\pi^2}{3}\tilde{I}\left(\frac{m_S^2}{m_\Delta^2}\right),\qquad
M\equiv \mathrm{max}(m_\Delta, m_S)\,,\\
&\tilde{I}(x)=\left\{\begin{array}{ll}
1+\frac{3}{\pi^2}(\ln^2 x-1) & \mathrm{for}~ x\gg 1\,,\\
1& \mathrm{for}~ x\to 0\,.
\end{array}\right.
\end{align}
For the later use, it is convenient to write the neutrino mass matrix in a compact form
\begin{equation}\label{eq_seesaw_relation}
M_{\nu}=Y_L\omega Y_L^T,
\end{equation}
with the matrix $\omega_{jj'}\equiv 24\mu I_{\nu}m_{j}m_{j'}(Y_s^\dagger)_{jj'}$. Qualitatively speaking, the resulting
neutrino mass is about
\beq
m_\nu \sim {\mu m_b^2 Y_L^2 Y_S \over 32 \pi^2 M^2} \sim 0.06 \mbox{eV} \times\left( {Y_L^2 Y_S \over 10^{-6} }\right)
\times\left( { \mbox{TeV}\over M^2/ \mu }\right)\,.
\label{eq:BOE_nu_mass}
\eeq
One sees that, due to the 2-loop suppression, with a typical values $Y_L,Y_S \sim 0.01$ and $\mu, M \sim 1$ TeV, the sub-eV neutrino mass can be easily achieved without excessively fine tuning.
\begin{figure*}[h]
\begin{center}
\includegraphics[width=8cm]{nu-mass.eps}
\caption{The 2-loop neutrino mass generated from colored scalar. Where $a,b,c$ are the $SU(3)$ indices.} \label{neutrino-mass}
\end{center}
\end{figure*}
However, the simultaneous presence of $Y_{L/R}$ and $y^\Delta$ leads to tree-level proton decay as pointed out in \cite{LQ-Wise}.
A very small $y^\Delta_{11}$ is enough to avoid the rapid proton decay problem. Alternatively, the $y^\Delta$ term can be
eliminated by imposing some ad hoc symmetry. For example, this term can be turned off without upsetting all other interactions if some $Z_2$ parities $\{-,-,+, +, +, -,+\}$ are assigned to $\{ L, l_R, Q, u_R, d_R, \Delta, S \}$, respectively.
Hence, we leave the proton decay problem aside and simply set $y^\Delta=0$ in this study.
The most general renormalizable scalar potential including $S$ and $\Delta$ is
\begin{eqnarray}
\mathscr{V}&=&
- \mu_H^2 (H^\dagger H)
+m_\Delta^{2}\Delta^\dagger \Delta +
m_S^{2} \mbox{Tr} S^\dagger S
+\lambda(H^\dagger H)^2 + \lambda_\Delta (\Delta^\dagger \Delta)^2
+\lambda_S (\mbox{Tr} S^\dagger S)^2
\label{potential}
\\
&& +\lambda_1 (\Delta^\dagger \Delta)(H^\dagger H)
+\lambda_2 \mbox{Tr}(S^\dagger S)(H^\dagger H)
+\lambda_3 \mbox{Tr}(S^\dagger S)(\Delta^\dagger \Delta)
+ \left(\mu \Delta^{*}
\Delta^{*}S + h.c.\right)\nonumber
\end{eqnarray}
where the trace is over the color indices.
The details of the scalar potential are not relevant for the later discussion. We note by passing that only the SM Higgs doublet can acquire a nonzero vacuum expectation value, $\langle H \rangle =v/\sqrt{2} $, and being solely responsible for the electroweak symmetry breaking(EWSB). The tree-level masses of $\Delta$ and $S$ are shifted after EWSB with $ m_\Delta^{2} \rightarrow m_\Delta^{2} + \lambda_1 v^2/2$ and $ m_S^{2} \rightarrow m_S^{2} + \lambda_2 v^2/2$.
To proceed, we need $m_\Delta$ and $m_S$ after EWSB as input.
Since $S$ and $\Delta$ participate strong interactions, they are best searched for at the hadron colliders but so far none has been found yet.
Depending on their couplings to the SM fields, some lower bounds on $m_S$ and $m_\Delta$ were obtained from the null result of collider searches. The current lower bounds on $m_\Delta$ are summarized in Table.\ref{tab:LQ-mass}.
\begin{table*}[h]
\caption{
Summary of leptoquark mass lower bound (in unit of $\mathrm{GeV}$) from direct search with $95\%$ CL.
The values in parentheses are for $\beta=0.5$, and $\beta=1$ otherwise. The leptoquark decays branching ratios into $lq$ and $\nu q$ are denoted as $\beta$ and $(1-\beta)$, respectively,
and $\lambda$ is the Yukawa coupling for $l q \Delta$. The leptoquark is assumed to decay into leptons within only one specific generation.
}
\begin{center}
\begin{tabular}{c|c|c|c}
\hline\hline
& First generation & Second generation & Third generation \\
\hline
CMS & $1005(845)$ \cite{CMS-1st} & $1070 (785)$ \cite{CMS-2nd} & $634$ \cite{CMS-3rd} \\
ATLAS & $ 660 (607)$ \cite{ATLAS-1st} & $685(594)$ \cite{ATLAS-2nd} & $534$ \cite{ATLAS-3rd} \\
ZEUS & $699 (\lambda=0.3)$\cite{HERA-1st} & & \\
\hline\hline
\end{tabular}
\end{center}
\label{tab:LQ-mass}
\end{table*}
For an $E_6$-type diquark, CMS study gives $m_{S}>6\mathrm{TeV}$ \cite{DQ-mass}.
These limits are very sensitive to the assumptions of decay branching fraction as well as the flavor dependant coupling strengthes.
Hence, in the following numerically analysis, we take $m_S=7\,\mathrm{TeV}$ and
$m_\Delta=1\,\mathrm{TeV}$ as the benchmark values\footnote{
Since $S$ and $\Delta$ are also charged under SM $SU(3)$ and $U(1)_Y$, their 1-loop contributions alter the SM $h VV'$ couplings where $VV'=\{\gamma\gamma, \gamma Z, gg\}$.
Following the analysis in \cite{Chang-Ng-Wu,CPX,CGHT} also the data form\cite{ATLAS-hdiphoton,ATLAS-hgammaZ,CMS-hdiphoton,CMS-hgammaZ}, we find that the corrections to the signal strengths
are not significant, $0.96<\mu_{\gamma\gamma, \gamma Z}<1.2$, for $ m_\Delta\in[1,3]$ TeV and $m_S\in[6,8]$ TeV.}.
The triple $\Delta\Delta S$ coupling generates 1-loop correction $\sim \frac{\mu^2}{16\pi^2}\log(\mu^2/m_X^2)$ to $m_X^2$ where $X=S, \Delta$. For these quantum corrections to be perturbative, one needs roughly $ |\mu ^2 \log(\mu^2/m_X^2)|\leq 16\pi^2 m_X^2$
\footnote{ From the Eq.(\ref{eq:BOE_nu_mass}), $\mu$ also has a weak lower bound $|\mu| \gtrsim 10^{-6}\mbox{TeV}\times (M/\mbox{TeV})^2 $ if $Y_L$ and $Y_S$ are required to be less than $1.0$. }.
On the other hand, the dimensionful parameters in the same scalar potential are expected to be around the same order.
These considerations led to similar estimations and $\mu=(0.1- 1)\,\tev$ is assumed in this study.
At the tree-level, the decay channels for leptoquark are $\Delta \rightarrow \ell_i u_j$ and $\Delta \rightarrow \nu_i d_j$. For di-quark, it decays into
$d_i d_j$, and $\Delta \Delta$ if kinematically allowed.
Given that $m_S, m_\Delta \gg m_t$, all the SM final states can be treated massless and the decay widthes
can be calculated to be
\begin{eqnarray}
\Gamma_{\Delta} &=& \sum_{i,j}
\left[ \Gamma(\Delta \rightarrow \ell_i u_j) + \Gamma(\Delta \rightarrow \nu_i d_j) \right]
\sim \frac{ m_{\Delta}}{16 \pi} \sum_{i,j}
\left( 2 \vert (Y_L)_{ij} \vert^2 + \vert (Y_R)_{ij} \vert^2 \right)\,,\\
\Gamma_{S} &=& \Gamma(S \rightarrow \Delta \Delta)\times \theta(m_S-2 m_\Delta) + \sum_{ij } \Gamma(S \rightarrow d_i d_j) \nonumber \\
&\sim & \frac{m_S}{8 \pi} \left\lbrace
\left( 1-4 \frac{m^2_{\Delta}}{m_S^2} \right)^{1/2} \left( \frac{\mu}{m_S}\right)^2\times \theta(m_S-2 m_\Delta)
+ \sum_{i,j} \vert (Y_S)_{ij} \vert^2 \right\rbrace.
\end{eqnarray}
\section{Neutrino Masses and the Tree-level Flavor Violation}
As discussed before, it is assumed that there is no hierarchy among the $Y_S$'s.
Since $m_b \gg m_s \gg m_d$, the matrix $\omega$ can be broken into the leading and sub-leading parts and $\omega =\omega^{(0)}+\omega^{(1)}$, where
\beq
\omega^{(0)}= 24 \mu I_\nu \times \left( \begin{array}{ccc}
0 & 0 & 0 \\
0 & 0 & m_b m_s (Y_S)_{23}^*\\
0 & m_b m_s (Y_S)_{23}^* & m_b^2 (Y_S)_{33}^*
\end{array}
\right)\,,
\eeq
and ${\cal O} \left(\frac{\omega^{(1)}}{\omega^{(0)} }\right) \sim {\cal O}\left( \frac{m_d}{m_b} \right)$.
It is easy to check that the leading neutrino mass matrix $M_\nu^{(0)} = Y_L \omega^{(0)} Y_L^{T}$ is of rank-2 and $\det M_\nu^{(0)}=0$. Hence, at least one of the active neutrinos is nearly massless, $\sim (m_d/m_b)\times \mbox{max}(m_\nu)$, and the scenario of quasi-degenerate neutrinos is disfavored in the cZBM.
At leading order, $(Y_L)_{11,21,31}$ do not enter $M_\nu^{(0)}$ at all.
Therefor, for either normal hierarchy (NH) or the inverted hierarchy(IH) type of the neutrino masses, the eigenmasses are
\begin{itemize}
\item NH:
\begin{equation}
m_1\simeq 0\,, m_2 \simeq \sqrt{\Delta m_{21}^2}\,, m_3 \simeq \sqrt{\Delta m^2 + \frac{\Delta m_{21}^2}{2} }\,,
\end{equation}
and
\item IH:
\begin{equation}
m_3\simeq 0\,, m_1 \simeq \sqrt{|\Delta m^2| - \frac{\Delta m_{21}^2}{2} }\,, m_2 \simeq \sqrt{|\Delta m^2| + \frac{\Delta m_{21}^2}{2} }\,,
\end{equation}
\end{itemize}
where $\Delta m^2 \equiv m_3^2 - ( m_1^2 + m_2^2 )/2$.
Moreover, the absolute values of neutrino mass can be obtained by plugging in the well determined neutrino data\cite{PDG} listed in Table \ref{tab:nu_exp}.
For NH, $ m_2 \sim 0.00868$ eV and $m_3\sim 0.0496$ eV, and for IH, $m_1\sim 0.0483$ eV and $m_2\sim 0.0492$ eV. For both cases, the total sum of neutrino masses automatically agrees with the limit that $\sum m_{\nu}<0.23$ eV at $95\%$ C.L.
from the cosmological observation\cite{nu-mass}.
\begin{table*}[h]
\caption{
The global-fit neutrino data with $1\sigma$ deviation\cite{PDG}. }
\begin{center}
\begin{tabular}{c|c}
\hline\hline
$\sin^2 \theta_{23}$
& $0.437^{+0.033}_{-0.023}$ \quad (NH) \\
& $0.455^{+0.039}_{-0.031}$ \quad (IH) \\
$\sin^2 \theta_{13}$ & $0.0234^{+0.0020}_{-0.0019}$\quad (NH) \\
& $0.0240^{+0.0019}_{-0.0022}$ \quad (IH) \\
$\sin^2 \theta_{12}$ & $0.308^{+0.017}_{-0.017}$ \\
$\Delta m_{21}^2$ & $ \left(7.54^{+0.26}_{-0.22}\right) \times 10^{-5} \mathrm{eV}^2$ \\
$|\Delta m^2|$
& $\left(2.43^{+0.06}_{-0.06}\right)\times 10^{-3}\mathrm{eV}^2$ \quad (NH) \\
& $\left(2.38^{+0.06}_{-0.06}\right)\times 10^{-3}\mathrm{eV}^2 $ \quad (IH) \\
$\delta/\pi$ & $1.39^{+0.38}_{-0.27}$ \quad (NH) \\
& $1.31^{+0.29}_{-0.33}$ \quad (IH) \\
\hline\hline
\end{tabular}
\end{center}
\label{tab:nu_exp}
\end{table*}
Once $m_{1,2,3}$ are fixed, the neutrino mass matrix can be worked out reversely by
\beq
\label{eq:mnu_diag}
M_\nu = U_{PMNS}^* \left( \begin{array}{ccc}
m_1 & 0 & 0 \\
0 & m_2 &0\\
0 & 0& m_3
\end{array}
\right) U_{PMNS}^\dag\,.
\eeq
The standard parametrization is adopted that
\beq
\label{eq:definition_of_pmns}
U_{\mathrm{PMNS}}=\left(\begin{array}{ccc} 1 & 0 & 0\\
0 & c_{23} & s_{23}\\
0 & -s_{23} & c_{23}\end{array}\right)\left(\begin{array}{ccc}
c_{13} & 0 & s_{13} e^{-i\delta} \\
0 & 1 & 0 \\
-s_{13} e^{i\delta} & 0 & c_{13}\end{array}
\right)\left(\begin{array}{ccc}
c_{12} & s_{12} & 0 \\
-s_{12} & c_{12} & 0 \\
0 & 0 & 1\end{array}\right)
\begin{pmatrix}
1 & 0 & 0
\\
0 & e^{ i \alpha_{21}/2} & 0
\\
0 & 0 & e^{ i \alpha_{31}/2}
\end{pmatrix}
\eeq
where $c_{ij}$ and $s_{ij}$ represent $\cos\theta_{ij}$ and $\sin\theta_{ij}$, respectively.
In the case of Majorana neutrinos, $\alpha_{21}$ and $\alpha_{31}$ are the extra CP phases that cannot be determined from the oscillation experiments.
For simplicity, all $Y_S$'s are assumed to be real and the 2 Majorana CP phases will not be discussed in this paper.
The leading order neutrino mass matrix has 5(=6-1) independent entries\footnote{ The symmetric neutrino matrix $M_\nu^{(0)}$ has 6 elements minus 1 constraint that its determinant is zero. }.
With the democratic $Y_S$ assumption, the effective Majorana mass for $(\beta\beta)_{0\nu}$-decay $m_{ee} \sim 0.0018$ eV for the NH case. For the IH case, $m_{ee} \sim 0.0479$ eV which is
within the sensitivity of the planned $(0\nu\beta\beta)$ detectors with $\sim 1$ ton of isotope\cite{Bilenky:2012qi}.
Furthermore, the lightest neutrino mass is $\sim {\cal O}(10^{-5} \mbox{eV})$ for both IH and NH cases. For a given set of parameters, $\{\mu, m_S, m_\Delta, (Y_S)^{(0)}_{23}, (Y_S)^{(0)}_{33}, (Y_L)_{13} \}$, all the other 5 complex Yukawa couplings $(Y_L)_{ij} (j\neq 1)$ can be completely determined up to two signs by the leading $M_\nu^{(0)}$.
For a real $(Y_L)_{13}$, one has
\beqa
\label{eq:YL_leading}
(Y_L)_{23}^{(0)} &=& \frac{(Y_L)_{13}}{(M_\nu)_{11}}\left[(M_\nu)_{12}\pm \sqrt{(M_\nu)_{12}^2-(M_\nu)_{11}(M_\nu)_{22}}\right]\,,\nonr\\
(Y_L)_{33}^{(0)} &=& \frac{(Y_L)_{13}}{(M_\nu)_{11}}\left[(M_\nu)_{13}\pm \sqrt{(M_\nu)_{13}^2-(M_\nu)_{11}(M_\nu)_{33}}\right]\,,\nonr\\
(Y_L)_{12}^{(0)} &=& { (M_\nu)_{11} -B_\nu m_b^2 (Y_L)_{13}^2 (Y_S)_{33}^{(0)} \over 2 B_\nu m_b m_s (Y_L)_{13} (Y_S)_{23}^{(0)} }\,,\nonr\\
(Y_L)_{22}^{(0)} &=& { (M_\nu)_{22} -B_\nu m_b^2 [(Y_L)_{23}^{(0)}]^2 (Y_S)_{33}^{(0)} \over 2 B_\nu m_b m_s (Y_L)_{23}^{(0)} (Y_S)_{23}^{(0)} }\,,\nonr\\
(Y_L)_{32}^{(0)} &=& { (M_\nu)_{33} -B_\nu m_b^2 [(Y_L)_{33}^{(0)}]^2 (Y_S)_{33}^{(0)} \over 2 B_\nu m_b m_s (Y_L)_{33}^{(0)} (Y_S)_{23}^{(0)} }\,,
\eeqa
where $B_\nu=24 \mu I_\nu$. Again, $(Y_L)_{11,21,31}$ do not enter $M_\nu^{(0)}$ at all; they are arbitrary at this level and will be determined in the next order perturbation.
This approximation largely saves the work of numerical study and lays out the base for higher order perturbations beyond $\omega^{(0)}$.
The most important next to leading contribution to $M_\nu$ comes from $(Y_S)_{13}$. If one also perturbs $(Y_S)_{23,33}$ around
$(Y_S)^{(0)}_{23,33}$ with $(Y_S)_{23,33}^{(1)}=(Y_S)^{(0)}_{23,33}+\delta_{23,33}$, the consistent solution to $(Y_L)_{i1}$ for a given set $\{(Y_S)_{13}, \delta_{23},\delta_{33}\}$ are:
\beqa
\label{eq:YL_next}
(Y_L)_{11}^{(1)} &=& -\frac{m_s}{m_d}\frac{\delta_{23}}{(Y_S)_{13}}(Y_L)_{12}^{(0)}-\frac{m_b}{2m_d}\frac{\delta_{33}}{(Y_S)_{13}}(Y_L)_{13}\,,\nonr\\
(Y_L)_{21}^{(1)} &=& -\frac{m_s}{m_d}\frac{\delta_{23}}{(Y_S)_{13}}(Y_L)_{22}^{(0)}-\frac{m_b}{2m_d}\frac{\delta_{33}}{(Y_S)_{13}}(Y_L)_{23}^{(0)}\,,\nonr\\
(Y_L)_{31}^{(1)} &=& -\frac{m_s}{m_d}\frac{\delta_{23}}{(Y_S)_{13}}(Y_L)_{32}^{(0)}-\frac{m_b}{2m_d}\frac{\delta_{33}}{(Y_S)_{13}}(Y_L)_{33}^{(0)}\,.
\eeqa
With only a handful of free parameters, all the leptoquark left-handed Yukawa can be reasonably determined solely by the neutrino data.
However, further checks are needed to determine whether the above solution is phenomenologically viable.
Next, the tree-level flavor violation will be discussed.
\begin{figure*}[h]
\begin{center}
\includegraphics[width=8cm]{TL_FV.eps}
\caption{Tree-level flavor violation mediated by leptoquark and diquark. }
\label{fig:TL_FV}
\end{center}
\end{figure*}
The leptonic and quark flavor violating processes will be generated by exchanging $S$ and $\Delta$ at the tree-level, see Fig.\ref{fig:TL_FV}.
Since $S,\Delta$ are heavy, they can be integrated out below the EWSB scale. After Fierz transformation, we obtain
\begin{eqnarray}
\begin{aligned}
\triangle \mathcal{L}_{\mbox{eff}} =& \left[\frac{ (Y_{L}^{*})_{ml} (Y_{R})_{ij} }{2 m_{\Delta}^2}
\left(
-\overline{\nu_{m}} \PP_{R} \ell_{i} \cdot \overline{d_{l}^{a}} \PP_{R} u_{j}^{a}
+\overline{\ell_{m}} \PP_{R} \ell_{i} \cdot \overline{u_{l}^{a}} \PP_{R} u_{j}^{a}
\right)
+h.c. \right] \\
& -\left[\frac{ (Y_{L}^{*})_{ml} (Y_{L})_{ij} }{2 m_{\Delta}^2} \overline{\nu_{m}} \gamma^{\mu} \PP_{L} \ell_{i} \cdot \overline{d_{l}^{a}} \gamma_{\mu} \PP_{L} u_{j}^{a}
+ h.c. \right]
\\
& + \frac{ (Y_{L}^{*})_{ml} (Y_{L})_{ij} }{2 m_{\Delta}^2}
\left(
\overline{\nu_{m}} \gamma^{\mu} \PP_{L} \nu_{i} \cdot \overline{d_{l}^{a}} \gamma_{\mu} \PP_{L} d_{j}^{a}
+ \overline{\ell_{m}} \gamma^{\mu} \PP_{L} \ell_{i} \cdot \overline{u_{l}^{a}} \gamma_{\mu} \PP_{L} u_{j}^{a}
\right) \\
& +\frac{ (Y_{R}^{*})_{ml} (Y_{R})_{ij} }{2 m_{\Delta}^2} \overline{\ell_{m}} \gamma^{\mu} \PP_{R} \ell_{i} \cdot \overline{u_{l}^{a}} \gamma_{\mu} \PP_{R} u_{j}^{a}
\\
& +
\left[
\frac{ (Y_{L}^{*})_{ml} (Y_{R})_{ij} }{8 m_{\Delta}^2}
\left(
\overline{\nu_{Lm}} \sigma^{\mu\nu} \PP_{R} \ell_{i} \cdot \overline{d_{l}^{a}} \sigma_{\mu\nu} \PP_{R} u_{j}^{a}
-\overline{\ell_{Lm}} \sigma^{\mu\nu} \PP_{R} \ell_{i} \cdot \overline{u_{l}^{a}} \sigma_{\mu\nu} \PP_{R} u_{j}^{a}
\right) + h.c.\right]\\
& + \frac{(Y_s)_{ij}(Y_s^\dagger)_{lm} }{2m_S^2}\left[\overline{d_m^a}\gamma^\mu\PR d_i^a\right]
\left[\overline{d_l^b}\gamma_\mu\PR d_j^b\right]
\label{eq:TLFV}
\end{aligned}
\end{eqnarray}
where $a,b$ are the color indices.
There are way too many new free parameters and rich phenomenology in the most general model.
To simplify the discussion and to extract the essential physics, we consider the case that the new physics has minimal tree-level flavor violation(TLFV). Note that the TLFV contributions from different chiral structures always add incoherently. To minimize the total TLFV we need to suppress the TLFV from each chiral structure as much as possible. Let's concentrate on the
purely left-handed operators first. Observe that (1) A trivial flavor violation free solution is that with $(Y_{L})_{ij}(Y_{L}^\dagger)_{lm} \propto \delta_{im} \delta_{j l}$. It is obvious that these kind of solutions allow only one nonzero entry of $Y_{L}$, as can be easily
seen by looking at Fig.\ref{fig:TL_FV}(a). It always leads to 2 massless neutrinos which has been excluded by the current neutrino oscillation data.
(2) If the requirement is relaxed to $(Y_{L})_{ij}(Y_{L}^\dagger)_{lm} \propto \delta_{im}$ (no leptonic TLFV) or $ (Y_{L})_{ij}(Y_{L}^\dagger)_{lm} \propto \delta_{j l}$ ( no quark TLFV), only one row or one column of $Y_{L}$ can be nonzero\footnote{Neutrino masses aside, similar conclusions also apply to the $Y_R$ matrix for the purely right-handed operators.}
and the resulting neutrino masses have two zeros again. However, only $Y_L$'s are relevant to the neutrino masses.
One can set $Y_R =0$ to minimize the TLFV, and use the 9 remaining $Y_L$'s to accommodate the neutrino masses. Then the lower bound on each flavor violation process can be found since any nonzero $Y_R$ will add to it.
It is very easy to build a model with $Y_R\ll 1$ or $Y_R=0$ and we supplement with two examples.
Example one is to introduce an extra $U(1)_x$ with two SM like Higgs doublets, $H_1$ and $H_2$.
Then the $U(1)_x$ charge assignment $\{\alpha_1,\alpha_2,\alpha_2,\alpha_3,\alpha_4 ,\alpha_5, \alpha_6, \alpha_7 \}$ for $\{ Q, u_R, d_R, L, e_R, H_1, H_2, \Delta\}$ with $ \alpha_5= \alpha_1-\alpha_2$, $\alpha_6=\alpha_3-\alpha_4$, $\alpha_7=-(\alpha_1+\alpha_3)$, $\alpha_7 \neq 2 \alpha_2$ and $\alpha_7 \neq (\alpha_2+\alpha_4)$ will kill $Y_R$ (and also $y^\Delta$) but still allow the charged fermions to acquire the Dirac masses from their Yukawa couplings with $H_1$ or $H_2$. There are other issues needed to be considered in this setup. For example whether the $U(1)_x$ is global or local and whether it is free of anomaly. But these issues do not concern us since they are not relevant to this study and there are well-known model-building machineries available to deal with these problems.
The second example is promoting the 4-dimensional model into a higher-dimensional version. If the wave functions of $l_R$ and $u_R$ in the extra spacial dimension(s) are well separated, like in \cite{Chang:SF}, or have very little overlapping, like in\cite{Chang:RS}, the resulting $Y_R$ is negligible. Anyway, here $Y_R=0$ is taken as a phenomenology assumption which minimizes the TLFV. Some remarks on the case of $Y_R\neq 0$ will be discussed in next section.
A model independent analysis of the effective four-fermion operators was done by\cite{2L2Q}.
The $90\%$ C.L. upper limits on the normalized Wilson coefficient $\epsilon_{ijkl}$ (it is not the totally anti-symmetric tensor),
\begin{equation}
\epsilon_{ijkn} \equiv \frac{ (Y_L)_{ik} (Y_L)_{jn}}{4 \sqrt{2} G_F m_\Delta^2}\,,
\end{equation}
for each 4-fermi operator are extracted and listed in Table \ref{tab:2L2Q-YL}.
We have
\begin{equation}
(Y_L)_{ik} (Y_L)_{jn} < 4 \sqrt{2} G_F m_\Delta^2 \epsilon_{ijkn}^\text{max} \sim 65.98 \times \epsilon_{ijkn}^\text{max} \times\left({ m_\Delta \over 1 \mbox{TeV}}\right)^2\,.
\end{equation}
\begin{table*}[t]
\caption{
The 90\%C.L. upper limits on $\epsilon_{ijkn}$ from \cite{2L2Q}. Here, the dimensionless quantities $\epsilon_{ijkn} \equiv (Y_L)_{ik} (Y_L)_{jn} / ( 4 \sqrt{2} G_F m_\Delta^2 )$, where $i,j$($k,n$)are lepton(quark) flavor indices.
}
\begin{center}
\begin{tabular}{|c|c||c|c||c|c|}
\hline
$\epsilon_{ e e 1 1 } $ & $ 10^{-3}$
& $ \epsilon_{ e e 1 2 } $ & $ 9.4 \times 10^{-6} $ & $ \epsilon_{ e e 1 3 } $ & $ 3.9 \times 10^{-3} $ \\
\hline
$\epsilon_{ e e 2 2 } $ & $ 10^{-2} $ & $ \epsilon_{ e e 2 3 } $ & $ 10^{-3} $ & $ \epsilon_{ e e 3 3 } $ & $ 9.2 \times 10^{-2} $ \\
\hline
\hline
$\epsilon_{ \mu \mu 1 1 } $ & $ 7.3 \times 10^{-3} $ & $ \epsilon_{ \mu \mu 1 2 } $ & $ 9.4 \times 10^{-6} $ & $ \epsilon_{ \mu \mu 1 3 } $ & $ 3.9 \times 10^{-3} $ \\
\hline
$\epsilon_{ \mu \mu 2 2 } $ & $ 1.2 \times 10^{-1} $ & $ \epsilon_{ \mu \mu 2 3 } $ & $ 10^{-3} $ & $ \epsilon_{ \mu \mu 3 3 } $ & $ 6.1 \times 10^{-2} $ \\
\hline
\hline
$ \epsilon_{ \tau \tau 1 1 } $ & $ 10^{-2} $ & $ \epsilon_{ \tau \tau 1 2 } $ & $ 9.4 \times 10^{-6} $ & $ \epsilon_{ \tau \tau 1 3 } $ & $ 3.9 \times 10^{-3} $ \\
\hline
$ \epsilon_{ \tau \tau 2 2 } $ & $ 1.2 \times 10^{-1} $ & $ \epsilon_{ \tau \tau 2 3 } $ & $ 10^{-3} $ & $ \epsilon_{ \tau \tau 3 3 } $ & $ 8.6 \times 10^{-2} $ \\
\hline
\hline
$\epsilon_{ e \mu 1 1} $ & $ 8.5 \times 10^{-7} $ & $\epsilon_{ e \mu 1 2 } $ & $ 9.4 \times 10^{-6} $ & $\epsilon_{ e \mu 1 3 } $ & $ 3.9 \times 10^{-3} $ \\
\hline
$\epsilon_{ e \mu 2 1} $ & $ 9.4 \times 10^{-6} $ & $\epsilon_{ e \mu 2 2 } $ & $ 0.24 $ & $\epsilon_{ e \mu 2 3 } $ & $ 10^{-3} $ \\
\hline
$\epsilon_{ e \mu 3 1 } $ & $ 3.9 \times 10^{-3} $ & $\epsilon_{ e \mu 3 2 } $ & $ 10^{-3} $ & $\epsilon_{ e \mu 3 3 } $ & $ 6.6 \times 10^{-2} $ \\
\hline
\hline
$\epsilon_{ e \tau 1 1 } $ & $ 8.4 \times 10^{-4} $ & $\epsilon_{ e \tau 1 2 } $ & $ 9.4 \times 10^{-6} $ & $\epsilon_{ e \tau 1 3 } $ & $ 3.9 \times 10^{-3} $ \\
\hline
$\epsilon_{ e \tau 2 1 } $ & $ 9.4 \times 10^{-6} $ & $\epsilon_{ e \tau 2 2 } $ & $ 0.24 $ & $\epsilon_{ e \tau 2 3 } $ & $ 10^{-3} $ \\
\hline
$\epsilon_{ e \tau 3 1 } $ & $ 3.9 \times 10^{-3} $ & $\epsilon_{ e \tau 3 2 } $ & $ 10^{-3} $ & $\epsilon_{ e \tau 3 3 } $ & $ 0.2 $ \\
\hline
\hline
$\epsilon_{ \mu \tau 1 1 } $ & $ 9.4 \times 10^{-4} $ & $\epsilon_{ \mu \tau 1 2 } $ & $ 9.4 \times 10^{-6} $ & $\epsilon_{ \mu \tau 1 3 } $ & $ 3.9 \times 10^{-3} $ \\
\hline
$\epsilon_{ \mu \tau 2 1 } $ & $ 9.4 \times 10^{-6} $ & $\epsilon_{ \mu \tau 2 2 } $ & $ 0.24 $ & $\epsilon_{ \mu \tau 2 3 } $ & $ 10^{-3} $ \\
\hline
$\epsilon_{ \mu \tau 3 1 } $ & $ 3.9 \times 10^{-3} $ & $\epsilon_{ \mu \tau 3 2 } $ & $ 10^{-3} $ & $\epsilon_{ \mu \tau 3 3 } $ & $ 1 $ \\
\hline
\end{tabular}
\end{center}
\label{tab:2L2Q-YL}
\end{table*}
For the TLFV mediated by $S$, the last term in Eq.(\ref{eq:TLFV}), it is best constrained by the neutral meson mixings.
Following the convention in \cite{UTfit},
the corresponding Wilson coefficients and the 4-fermi operators for $K$-$\bar{K}$, $B_d$-$\bar{B}_d$ and $B_s$-$\bar{B}_s$ mixing are
\begin{align}
&\tilde{C}^1_K=-\frac{1}{2m_S^2}(Y_s)_{11}(Y_s^\dagger)_{22},\quad \tilde{Q}^1_K=(\bar{s}\gamma^\mu\PR d)(\bar{s}\gamma_\mu\PR d)\,,\nonumber\\
&\tilde C^1_B=-\frac{1}{2m_S^2}(Y_s)_{11}(Y_s^\dagger)_{33},\quad \tilde{Q}^1_B=(\bar{b}\gamma^\mu\PR d)(\bar{b}\gamma_\mu\PR d)\,,\nonumber\\
&\tilde C^1_{B_s}=-\frac{1}{2m_S^2}(Y_s)_{22}(Y_s^\dagger)_{33},\quad \tilde{Q}^1_{B_s}=(\bar{b}\gamma^\mu\PR s)(\bar{b}\gamma_\mu\PR s)\,.
\end{align}
A global analysis with $95\%$ C.L. gave\cite{UTfit}
\begin{align}
& |\mathrm{Re}(\tilde C_K^1)|<9.6\times 10^{-13}, \quad
-4.4 \times 10^{-15}<\mathrm{Im}(\tilde C_K^1)<2.8 \times 10^{-15}\,,\nonumber\\
& | \tilde C_{B_d}^1 |<2.3 \times 10^{-11}, \quad
-\pi<\mathrm{Arg}(\tilde C_{B_d}^1)<\pi\,,\nonumber\\
& | \tilde C_{B_s}^1 |<1.1\times 10^{-9}, \quad
-\pi<\mathrm{Arg}(\tilde C_{B_s}^1)<\pi\,,
\end{align}
in the unit of $\mathrm{GeV}^{-2}$. Or equivalently,
\begin{align}\label{eq:constraint_on_yS}
&|(Y_s)_{11} (Y_s)^\dagger_{22}|< 1.92 \times 10^{-6} \times \left( \frac{m_S}{\text{TeV}} \right)^2 ,\nonumber\\
&|(Y_s)_{11} (Y_s)^\dagger_{33}|< 4.6 \times 10^{-5} \times \left( \frac{m_S}{\text{TeV}} \right)^2,\nonumber\\
&|(Y_s)_{22} (Y_s)^\dagger_{33}|< 2.2 \times 10^{-3} \times \left( \frac{m_S}{\text{TeV}} \right)^2.
\end{align}
For the democratic $Y_S$, the above constraints imply $|Y_S| \lesssim 9 \times 10^{-3} \times(m_S /7 \mbox{TeV})$.
Before ending this section, we recap the assumptions and discussion so far:
\begin{itemize}
\item $Y_S$'s are assumed to be democratic and there is no outstanding hierarchy among these Yukawa couplings.
This leads to one nearly massless active neutrino and $|Y_S| \lesssim 9 \times 10^{-3} \times(m_S /7 \mbox{TeV})$
from the constrains of neutral meson mixings.
\item The Yukawa couplings $Y_R$ are turned off to minimize the TLFV.
\item For a given set of $\{\mu, m_S, m_\Delta, (Y_S)_{13,23,33}\}$ and any one of the $Y_L$'s, all the remaining 8 $Y_L$ can be iteratively determined from the absolute neutrino masses and the $U_{PMNS}$ matrix.
\end{itemize}
\section{Charged lepton flavor violating process at one-loop}
In this section, we shall study the charged lepton flavor violating (cLFV) processes $\ell\to\ell' \gamma$, $Z\to \ell'\bar{\ell}$ and the like which are induced at the 1-loop level with the leptoquark running in the loop, see Fig.\ref{fig:meg}.
\begin{figure*}[h]
\begin{center}
\includegraphics[width=16cm]{meg.eps}
\caption{The Feynman diagrams for 1-loop LFV $\mu\to e\gamma$.} \label{fig:meg}
\end{center}
\end{figure*}
\subsection{$\ell \to \ell'\gamma$}
The effective Lagrangian responsible for the cLFV process $\ell\to \ell'\gamma$\cite{Kuno,WFC-LFV} is parameterized as
\begin{equation}\label{meg-operator}
\mathscr{L} \supset \frac12 \bar{\ell}'\left( d_L^{ll'}\PL + d_R^{ll'} \PR\right)\sigma^{\mu\nu} \ell F_{\mu\nu} +h.c.
\end{equation}
For $m_\ell' \ll m_\ell$, the partial decay width is given as
\beq
\Gamma(\ell \to \ell' \gamma) \simeq { m_\ell^3 \over 16 \pi} (|d^{ll'}_L|^2 + |d^{ll'}_R|^2)\,.
\eeq
A straightforward calculation yields
\beq
\label{eq:dR_MEG}
d_R^{ll'} =- {N_c e \over 16\pi^2 m_\Delta^2}\left[ \left( m_{l'}(Y^*_R)_{l'q}(Y^T_R)_{q l} +m_l (Y^*_L)_{l'q}(Y^T_L)_{q l}\right) {\cal F}_1(r_q) +m_q (Y^*_L)_{l'q}(Y^T_R)_{q l}{\cal F}_2(r_q)\right]\,,
\eeq
where the index $q$ sums over $q=u,c,t$ and $r_q\equiv m_q^2 /m_\Delta^2$. $d_L^{ll'}$ can be obtained by simply switching $Y_L \leftrightarrow Y_R$ in the above expression for $d_R^{ll'}$.
The loop functions are
\beqa
{\cal F}_1(x) &=&{1+4x-5x^2+2x(2+x)\ln x \over 12(1-x)^4}\,,\nonr\\
{\cal F}_2(x)&=& {7-8x+ x^2 + 2(2+x)\ln x \over 6(1-x)^3}\,,
\eeqa
and they take the limits ${\cal F}_1\rightarrow 1/12$ and ${\cal F}_2\rightarrow 7/6+(2\ln x )/3$ when $x\rightarrow 0$.
Unlike at the tree-level, the contributions to the cLFV processes from $Y_L$ and $Y_R$ entangle with each other at the loop-level.
Since $m_q {\cal F}_2(r_q) \gg m_l {\cal F}_1(r_q)$ (for $q=c,t$), generally speaking, the last term in Eq.(\ref{eq:dR_MEG}) which involves both $Y_L$ and $Y_R$ gives the most important contribution to $d_R^{ll'}$\footnote{Barring the cases of fine-tuned cancelations and the hierarchical Yukawa couplings}. Therefore, it is expected that by setting $Y_R=0$ to minimize the TLFV
will also reduce the 1-loop cLFV processes in general. In the $Y_R=0$ case, $d_R^{ll'}$ dominates the cLFV processes because $m_\ell\gg m_{\ell'}$.
With $\tau_\mu^{-1}\approx \Gamma(\mu\to e\bar{\nu}_e\nu_\mu) = \frac{1}{192\pi^3}G_F^2 m_\mu^5$ and $\Gamma_\tau = 0.002265 \gev$,
the branching ratios for $\ell \to \ell' \gamma$ are
\beq
\mathcal{B}(\mu \to e\gamma)
\simeq \frac{12\pi^2 |d_R^{\mu e}|^2 }{ G_F^2 m_\mu^2}\, ,\,\,
\mathcal{B}(\tau\to \ell' \gamma) \simeq \frac{ m_\tau^3 |d_R^{\tau l'}|^2 }{16\pi \Gamma_\tau}\,.
\eeq
and
\beq
d_R^{ll'} = -\frac{e N_c m_\ell}{12 (4\pi)^2 m_\Delta^2}\sum_{q=u,c,t} \left[ a_q^\gamma (Y_L)^*_{\ell' q} (Y_L)_{\ell q} \right]\, ,
\eeq
where $a^{\gamma}_q = 1 + 4 r_q (\ln r_q +1) +{\cal O}(r_q^2)$, $a^{\gamma}_u\sim a^{\gamma}_c\sim 1.0$ and $a^{\gamma}_t\sim 0.82$.
Numerically, we have
\begin{eqnarray}
\mathcal{B}(\mu \to e\gamma) &\simeq & 2.1 \times 10^{-7}\times \left|\sum_{q=u,c,t} a_q^{\gamma} (Y_L^*)_{eq} (Y_L)_{\mu q} \right|^2\times\left( { 1\mbox{TeV} \over m_\Delta }\right)^4\, , \nonumber \\
\mathcal{B}(\tau \to \ell' \gamma) &\simeq & 3.8 \times 10^{-8}\times \left|\sum_{q=u,c,t} a_q^{\gamma} (Y_L^*)_{\ell' q} (Y_L)_{\tau q} \right|^2 \times\left( { 1\mbox{TeV} \over m_\Delta }\right)^4\,.
\end{eqnarray}
\subsection{ Remark on other photon dipole induced processes}
\subsubsection{Anomalous magnetic dipole moment}
Similar calculation with little modification can be carried over for the flavor diagonal cases. For the charged lepton, the anomalous magnetic dipole moment is
\beq
\triangle a_\ell = \frac{N_c }{6 (4\pi)^2} \frac{m_\ell^2}{m_\Delta^2} \left| \sum_{q=u,c,t} a_q^\gamma (Y_L)_{\ell q} \right|^2\, .
\eeq
Assuming that $|(Y_L)_{l q}|^2\sim{\cal O}(1)$ and $m_\Delta=1$ TeV, one has $\triangle a_e\sim 8.0\times 10^{-16}$, $\triangle a_\tau\sim 1.0\times 10^{-8}$, and $\triangle a_\mu\sim 3.0\times 10^{-11}$. Unless $m_\Delta\ll 1$TeV and all 3 $(Y_L)_{\mu q}$ are sizable and in phase, $\triangle a_\mu$ in this model is too small to accommodate the observed discrepancy
$ a_\mu^{exp}- a_\mu^{th} =(2.39\pm0.79)\times 10^{-9}$ at $1\sigma$ C.L.\cite{mu:g-2}.
Moreover, the model predicts a tiny positive $\triangle a_e$ which goes against the direction of the observed value that
$a_e^{exp}- a_e^{th} =-10.6(8.1)\times 10^{-13}$ at $1\sigma$ C.L.\cite{e:g-2}. Of course, a much larger $\triangle a_\mu$ is possible
to explain to observed discrepancy between the experimental measured value and the theoretical prediction if $Y_R\neq 0.$
\subsubsection{Electric dipole moments}
If $Y_R\neq 0$, the 1-loop charged lepton electric dipole moment(EDM), $d_\ell \sim \frac{N_c}{16\pi^2} \frac{m_t}{m_\Delta^2} \mathbf{Im}[Y_L Y_R^*]$, could be large. For $m_\Delta=1 \tev$, $|Y_L|\sim |Y_R|\sim 0.01$, and the CP phase is of order one, the typical electron EDM is around $10^{-24}\, e$-cm which is already 4 orders of magnitude larger than the current limit $|d_e|<8.7\times 10^{-29} e$-cm \cite{e:edm}.
Then, how to suppress the EDM's in this model will be a pressing theoretical issue. A plain solution is setting $m_\Delta \gtrsim 100 \tev$ to avoid the too large EDMs but the phenomenology at the low energies are strongly suppressed as well.
On the other hand, if $Y_R=0$ there is no EDM at the 1-loop level.
In fact, the first non-zero EDM contribution we can construct begins at the 3-loop level and it involves both $V^{CKM}$ and $U^{PMNS}$. An order of magnitude estimate gives:
\beq
d_\ell \sim \frac{\alpha N_c}{(16\pi)^3} \frac{m_\ell}{m_\Delta^2} \mathbf{Im}\left[(Y_L)_{\ell k} V^{CKM}_{kj}(Y_L^\dag)_{ji}U^{PMNS}_{i\ell}\right]\,.
\eeq
If $Y_L$ takes a typical value $\sim 0.01$, $m_\Delta=1 \tev$, and the combined CP phase is $\sim{\cal O}(1)$, this 3-loop electron EDM is expected to be $|d_e|\lesssim 10^{-37}\,e$-cm.
This upper bound is slightly larger than the estimated SM upper bound for $d_e$ but way below the sensitivity of any EDM measurement in the foreseeable future.
Consequently, $d_e$ is a useful handle to test the $Y_R=0$ assumption in the cZBM:
once the electron EDM was measured to be greater than $10^{-37}\, e$-cm, either the $Y_R=0$ assumption with $m_\Delta \sim {\cal O}(\tev)$ must be abandoned or more new physics is needed to go beyond the cZBM.
\subsubsection{$\mu-e$ conversion}
The $\mu-e$ conversion(MEC) will be mediated by the leptoquark at the tree-level as shown in Fig.\ref{fig:TL_FV}. For $Y_R=0$, the relevant 4-fermi operator is
\beq
\frac{(Y_L)^*_{e u} (Y_L)_{\mu u} }{2 m_{\Delta}^2} \overline{e} \gamma^{\mu} \PP_{L} \mu \cdot \overline{u^{a}} \gamma_{\mu} \PP_{L} u^{a}+h.c.
\eeq
The cLFV photon dipole operator discussed in the previous section will also contribute to MEC with an expected relative magnitude $\sim (\alpha/16\pi^2)^2$ comparing to the tree-level one.
Following the analysis of \cite{WFC-LFV}, a more quantitative estimate for the MEC rate is:
\begin{eqnarray}
\mathcal{B}_{conv} &\simeq & \mathcal{C}_{conv} \left\{ \left( \frac{\alpha Z}{16\sqrt{2}\pi G_F m_\Delta^2} \right)^2 |\sum_q a^\gamma_{q}(Y_L)^*_{eq} (Y_L)_{\mu q}|^2 \right. \nonumber \\
&& \left. + \left( \frac{(2Z+N)}{4 \sqrt{2} G_F m_\Delta^2} \right)^2 |(Y_L)^*_{eu} (Y_L)_{\mu u}|^2 \right\} \nonumber \\
&\simeq & \mathcal{C}_{conv} \left\{ 8.9 \times 10^{-11} Z^2 |\sum_q a^\gamma_{q}(Y_L)^*_{eq} (Y_L)_{\mu q}|^2 \right. \nonumber \\
&& \left. + 2.3 \times 10^{-4} (2Z+N)^2 |(Y_L)^*_{eu} (Y_L)_{\mu u}|^2 \right\} \times \left[ \frac{\tev}{m_\Delta} \right]^4 \ ,
\end{eqnarray}
where Z is the atomic number and N is the neutron number for a certain nucleus.
The overall factor $\mathcal{C}_{conv}$ depends on the form factors of the nuclei and the momentum of the muon.
For instance, $\mathcal{C}_{conv} ({}^{48}_{22}Ti) = 1.2\times 10^{-3}$\cite{2L2Q}.
As can be seen, the LFV photon dipole indeed has much smaller contribution to the MEC than the tree-level one.
\subsubsection{$\mu\rightarrow 3e$ }
In this model, there are no tree-level contributions to the cLFV $\mu\rightarrow 3e$ decay.
The $\mu\rightarrow 3e$ process is dominated by the cLFV photon dipole transition and its rate is much smaller than ${\cal B}(\mu\rightarrow e \gamma)$.
As pointed out in \cite{Kuno,WFC-LFV}, the ratio of ${\cal B}(\mu\rightarrow 3e)$ to ${\cal B}(\mu\rightarrow e \gamma)$ is basically model-independent:
\beq
{{\cal B}(\mu\rightarrow 3e) \over {\cal B}(\mu\rightarrow e \gamma)}\sim \frac{2\alpha}{3\pi} \left[\ln \frac{m_\mu}{m_e}-\frac{11}{8}\right]\simeq 0.019\,.
\eeq
Similarly, with replacing the charged lepton masses, the ratios in the rare tau decays are
\beqa
{{\cal B}(\tau\rightarrow 3e) \over {\cal B}(\tau\rightarrow e \gamma)}\sim \frac{2\alpha}{3\pi} \left[\ln \frac{m_\tau}{m_e}-\frac{11}{8}\right]\simeq 0.011\,,\nonr\\
{{\cal B}(\tau\rightarrow 3\mu) \over {\cal B}(\tau\rightarrow \mu \gamma)}\sim \frac{2\alpha}{3\pi} \left[\ln \frac{m_\tau}{m_\mu}-\frac{11}{8}\right]\simeq 0.002\,.
\eeqa
For the decay channels with different flavor final sates, one has\cite{WFC-LFV}
\beqa
{{\cal B}(\tau\rightarrow \mu e e^+ ) \over {\cal B}(\tau \rightarrow e \gamma)}\sim \frac{2\alpha}{3\pi} \left[\ln \frac{m_\tau}{m_e}-\frac{3}{2}\right]\simeq 0.032\,,\\
{{\cal B}(\tau\rightarrow e \mu \mu^+) \over {\cal B}(\tau\rightarrow e \gamma)}\sim \frac{2\alpha}{3\pi} \left[\ln \frac{m_\tau}{m_\mu}-\frac{3}{2}\right]\simeq 0.0064\,.
\eeqa
The decay branching ratios $\tau\rightarrow \mu^+ e e $ and $\tau\rightarrow \mu \mu e^+ $ are negligible because they are doubly suppressed by two cLFV transition vertices.
\subsection{$Z \to \bar{\ell} \ell'$}
The same Feynman diagrams in Fig.\ref{fig:meg} with photon replaced by $Z$ boson lead to cLFV $Z\rightarrow \bar{l}l'$ decays.
Since $Z$ is massive, it can also admit the vector or axial-vector couplings other than the dipole transition couplings as in the $l\rightarrow l'\gamma$ cases.
The most general gauge invariant $Z\rightarrow \bar{l} l'$ amplitude is parameterized as:
\begin{equation}
i\mathscr{M}=ie\overline{u}(p') \left[ (c_R^Z\PR+c_L^Z\PL)\left(-g_{\mu\nu}+\frac{q_\mu q_\nu}{m_Z^2}\right)\gamma^\nu
+\frac{1}{m_Z}\left(d_L^Z\PL+d_R^Z\PR\right)(i\sigma_{\mu\nu}q^\nu)\right]v(-p)\epsilon^\mu(q)\,,
\end{equation}
where the 4-momentums are labeled as in Fig.\ref{fig:meg}.
From the above parametrization, the branching ratio can be easily calculated to be
\begin{eqnarray}
\mathcal{B}(Z\to\overline{\ell}{\ell}')=\frac{\alpha}{6}\frac{m_Z}{\Gamma_Z} \left[(|c_L^Z|^2+|c_R^Z|^2)+\frac{1}{2}\left(|d_L^Z|^2
+|d_R^Z|^2\right)\right] \ ,
\end{eqnarray}
and the experimentally measured value $\Gamma_Z=2.4952\pm0.0023\mathrm{GeV}$\cite{PDG} is used in our study.
The 4 dimensionless coefficients $c^Z_{R,L}, d^Z_{R,L}$ can be obtained through a lengthy but straightforward calculation. The physics is rather simple and can be understood qualitatively. However, the full analytic results are not very illustrating and will not be presented here\footnote{The details will be given in other place.}.
Let's focus on the $Y_R=0$ case to simplify the physics discussion.
First of all, the masses of external charged leptons are much smaller than $m_Z$ and they can be treated massless\footnote{Since $m_u, m_c \ll m_Z$, they can also be treated as massless particles in this process.}.
For $c_R^Z$, the coupling connects both left-handed fermions and there is no need to flip their chiralities. In the loop calculation, $m_Z$ and $m_t$ are the only two dimensionful quantities other than $m_\Delta$. So, by dimensional analysis we know that $c_R^Z \sim {\cal O}(\frac{N_c}{16\pi^2}\frac{m_t^2}{m_\Delta^2})$ (for top quark running in the loop) or $c_R^Z \sim {\cal O}(\frac{N_c}{16\pi^2}\frac{m_Z^2}{m_\Delta^2})$( for light quarks running in the loop).
For the dipole couplings which connect fermions with different handiness, one external charged lepton mass insertion is needed to flip its chirality.
Also, $m_Z$ sets the nature scale of the
momentum transfer in this process. Therefore it is expected that
in general $d^Z/c^Z_R \sim {\cal O}(\frac{m_l}{m_Z})$ or ${\cal O}(\frac{m_l'}{m_Z})$. Thus, the contributions from $d_{L,R}^Z$ can be safely ignored.
On the other hand, both of the two external charged leptons need to flip their chiralities for having a nonzero $c_L^Z$ if $Y_R=0$. Therefore, $c_L^Z \sim {\cal O}(\frac{m_l m_l'}{m_Z^2}) c_R^Z$ and its contribution is totally negligible in this process. The above qualitative understandings agree very well with our full calculation.
Hence, only the leading contribution from $c^Z_R$ is kept in the study.
It is more useful to express the final result in the numerical form:
\beq
\label{eq:BZllp}
\mathcal{B}(Z\to \bar{\ell} \ell') \simeq 1.46 \times 10^{-7} \left|\sum_{q=u,c,t}a^Z_q (Y_L)^*_{\ell' q} (Y_L)_{\ell q}\right|^2
\times \left( \frac{\tev}{m_\Delta} \right)^4\, ,
\eeq
where $a_u^{Z} = a_c^{Z} \simeq -0.125-0.077 \mathbf{i} =-0.1468 e^{ i 31.63^\circ}$ and $a_t^{Z} = 1$.
The imaginary part of $a^Z_{u,c}$ comes from the pole of light-quark propagators in the loop when the light quarks are going on-shell in the $Z$ decay.
Also note that this cLFV decay branching ratio is around $10^{-7}$ if the absolute square in Eq.(\ref{eq:BZllp}) is of order unit.
The ballpark estimate is below but close to the current experimental limits\cite{PDG,ATLAS-Zdecay-2014}.
The interference between the sub-diagrams with $u(c)$ and $t$ running in the loop makes the relative phases between $a_{u,c}^Z$ and $a_t^Z$
observable. This physical phase leads to CP violation and in general $\mathcal{B}(Z \to \bar{\ell} \ell') \neq \mathcal{B}( Z \to \bar{\ell'} \ell)$. Following\cite{ZCP_Bernabeu,ZCP_Rius}, the CP asymmetries are quantified as:
\beq
\eta_{\ell \ell'} \equiv \mathcal{B}(Z \to \bar{\ell} \ell') - \mathcal{B}(Z \to \bar{\ell'} \ell)\,.
\eeq
In this model, we have numerically
\beq
\eta_{\ell \ell'}\simeq (4.53 \times 10^{-8} ) \times {\bf Im}\left[ \left( Y^{\ell'\ell}_u + Y^{\ell' \ell}_c \right) (Y^{\ell' \ell}_t)^* \right] \times \left( \frac{\tev}{m_\Delta} \right)^4\,,
\eeq
where the shorthand notation $Y^{\ell' \ell}_q \equiv (Y_L)^*_{\ell' q} (Y_L)_{\ell q}$. Interestingly, due to the sizable CP phase, the CP asymmetries and the cLFV decay branching ratios are of the same order. Also, for the later convenience, we define
$ \mathcal{B}^Z_{ \ell \ell'} \equiv \mathcal{B}(Z\to \bar{\ell} \ell') + \mathcal{B}(Z\to \ell \bar{\ell'})$.
Before closing this section, we should point out a simple but useful scaling relationship between $Y_S$ and $Y_L$ in this model. Recall that the neutrino mass is proportional to $Y_S Y_L^2$. Therefore, if $Y_S$ is re-scaled by $Y_S \rightarrow \lambda^{-2} Y_S$, then $Y_L$ must goes like $Y_L \rightarrow \lambda Y_L$ to keep the neutrino mass unchanged. After such rescaling, ${\cal B}(\ell\rightarrow \ell'\gamma)$, MEC, $\mathcal{B}^Z_{ \ell \ell'}$ and $\eta_{\ell\ell'}$ go like $\lambda^{4}$ while $\epsilon_{ll'qq'}$, $\Delta a_l$, and EDM go like $\lambda^{2}$ due to their amplitude nature. This scaling relationship largely helps reduce the computer time in finding the realistic configurations.
Now we have everything needed for the numerical and phenomenological study.
\section{Numerical Study}\label{sec: numerical study}
\subsection{Scanning strategy}
As discussed in Sec.3, once the set $\{\mu, m_S, m_\Delta, (Y_S)_{13,23,33}\}$ plus any one out of the 9 $Y_L$'s are fixed, all the remaining 8 $Y_L$'s can be iteratively determined from the absolute neutrino masses and the $U_{PMNS}$ matrix.
In our numerical search, we take $m_\Delta = 1\tev$ and $m_S = 7 \tev$ as the benchmark.
Moreover, for each configuration, $\mu$ is randomly produced within $[0.1 ,1]\tev$.
For each search, the neutrino mixings $sin^2\theta_{12,13,23}$, and the Dirac phase $\delta_{cp}$ are randomly generated within the 1 sigma allowed range from the global fit, Tab.\ref{tab:nu_exp}. For simplicity the two Majorana phases are set to be zero. Then the $U_{PMNS}$ matrix can be determined via Eq.(\ref{eq:definition_of_pmns}).
For a given $U_{PMNS}$, we still need to know the absolute neutrino eigen-masses in order to obtain the neutrino mass matrix, see Eq.(\ref{eq:mnu_diag}).
As has been discussed, we assume the lightest neutrino mass is zero. Depending on the neutrino mass hierarchy, the other 2 absolute neutrino masses can be determined from the given $\Delta m_{21}^2$ and $\Delta m^2$.
These 2 mass squared differences are also randomly generated within the 1 sigma allowed range from the global fit. Then the absolute neutrino mass matrix $M_\nu^{IH}(M_\nu^{NH})$ for the inverted(normal) hierarchy is ready for use.
Next, $|(Y_L)_{13}|$ is randomly generated as a real number between $10^{-10}$ and $1.0$.
Because of the scaling relationship discussed in the previous section, we fix $|(Y_S)_{33}|=0.0097$\footnote{We have $|(Y_S)_{11}|\sim|(Y_S)_{22}|\sim|(Y_S)_{33}|$ and the most stringent bound is $|(Y_S)_{11} (Y_S)_{22}| < 9.408 \times 10^{-5}$, hence $|(Y_S)_{33}| \lesssim 0.0097 $, for $m_S = 7 \tev$.} without losing any generality.
Then, $|(Y_S)_{13,23}|$ are generated within $[0.1,10]\times 0.0097$ and they must obey $ 0.1<|(Y_S)_{13}/(Y_S)_{23}| <10.0$ to be consistent with our working assumption.
The signs of $(Y_L)_{13}$ and $(Y_S)_{13,23,33}$ are also randomly assigned with equal probabilities being positive or negative.
With the above mentioned values, $(Y_L)^{(0)}_{23,33,12,22,32}$ can be fixed via Eq.(\ref{eq:YL_leading}).
Finally, $(Y_L)^{(1)}_{11,21,31}$ can be derived from the next order perturbation, Eq.(\ref{eq:YL_next}).
For that, we put in a small random perturbation within the range that $ 10^{-7}<|\delta_{23}/(Y_S)_{23}|, |\delta_{33}/(Y_S)_{33}| <10^{-2}$.
With all $Y_L$'s ready, the randomly generated configuration is further checked to see whether it is viable.
A configuration will be accepted if it pass all the following criteria:
\begin{itemize}
\item All $|Y_L|$'s are less than one so that the model can be calculated perturbatively.
\item All the TLFV satisfy the current experimental limits listed in Tab.\ref{tab:2L2Q-YL}.
\item All the loop-level cLFV processes must comply with the latest experimental limits\footnote{While we are wrapping up this article, the MEG Collaboration has updated the $\mathcal{B}(\mu\to e\gamma)$ limit with a slightly better value $<4.2\times 10^{-13},\; 90\%$ C.L.\cite{MEG2016}. } summarized in Tab.\ref{tab:loopLFV_exp}.
\begin{table*}[h]
\caption{
Summary of the latest experimental limits we used in the numerical scan.}
\begin{center}
\begin{tabular}{c|c}
\hline\hline
$\mathcal{B}(\mu^+\to e^+\gamma)$ & $<5.7 \times 10^{-13},\; 90\%$ C.L. \cite{MEG}\\
$\mathcal{B}(\tau\to \mu\gamma)$ & $<4.4 \times 10^{-8},\; 90\%$ C.L. \cite{BaBar:tau-LFV} \\
$\mathcal{B}(\tau\to e\gamma)$ & $<3.3 \times 10^{-8},\; 90\%$ C.L. \cite{BaBar:tau-LFV} \\
\hline\hline
$\mathcal{B}^Z_{\tau\mu}$ & $<1.2\times 10^{-5}$,\; $95\%$ C.L. \cite{PDG} \\
$\mathcal{B}^Z_{\tau e}$ & $<9.8\times 10^{-6}$,\; $95\%$ C.L. \cite{PDG} \\
$\mathcal{B}^Z_{\mu e}$ & $<7.5\times 10^{-7}$,\; $95\%$ C.L. \cite{ATLAS-Zdecay-2014} \\
\hline\hline
\end{tabular}
\end{center}
\label{tab:loopLFV_exp}
\end{table*}
\end{itemize}
The phenomenologically viable configurations are collected and then used to calculate the resulting cLFV.
\subsection{Numerical Result}
\subsubsection{$\ell \to \ell' \gamma$ and $Z\to \overline{\ell} \ell'$}
The correlations between these cLFV processes are displayed in Fig.\ref{fig:IH_LFV}(Fig,\ref{fig:NH_LFV}) for IH(NH).
The sign of $(Y_S)_{33}$ is responsible for the two prominent clusters in each scatter plot.
However, the origin of the notable difference is mere technical and it can be traced back to Eq.(\ref{eq:YL_leading}):
The two terms in the numerator of $(Y_L)^{(0)}_{i2}$ have compatible magnitudes.
So when the sign of $(Y_S)_{33}$ is right, the two terms almost cancel out with each other yielding a relatively small $|(Y_L)^{(0)}_{i2}|$.
The opposite happens when the sign of $(Y_S)_{33}$ is wrong.
\begin{center}
\begin{figure}[h]
\subfigure[]{\includegraphics[width=0.3\textwidth]{IH_mu_e_g_tau_e_g.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{IH_mu_e_g_tau_mu_g.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{IH_tau_e_g_tau_mu_g.eps} }\\
\subfigure[]{\includegraphics[width=0.3\textwidth]{IH_mu_e_g_Z_e_mu.eps} } \subfigure[]{\includegraphics[width=0.3\textwidth]{IH_mu_e_g_Z_e_tau.eps} }
\subfigure[]{\includegraphics[width=0.3\textwidth]{IH_mu_e_g_Z_mu_tau.eps}}
\caption{Correlations among the charged lepton flavor violating branching ratios for neutrino masses are of the inverted hierarchy.
In these plots, $m_\Delta=1 \tev$, $m_S=7\tev$ and $|(Y_S)_{33}| =0.0097$. The dashed lines represent the current experimental limits at 90\%C.L. }
\label{fig:IH_LFV}
\end{figure}
\end{center}
\begin{center}
\begin{figure}[h]
\subfigure[]{\includegraphics[width=0.3\textwidth]{NH_mu_e_g_tau_e_g.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{NH_mu_e_g_tau_mu_g.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{NH_tau_e_g_tau_mu_g.eps}}\\
\subfigure[]{\includegraphics[width=0.3\textwidth]{NH_mu_e_g_Z_e_mu.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{NH_mu_e_g_Z_e_tau.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{NH_mu_e_g_Z_mu_tau.eps}}
\caption{Correlations among the charged lepton flavor violating branching ratios for neutrino masses are of the normal hierarchy.
In these plots, $m_\Delta=1 \tev$, $m_S=7\tev$ and $|(Y_S)_{33}| =0.0097$. The dashed lines represent the current experimental limits at 90\%C.L. }
\label{fig:NH_LFV}
\end{figure}
\end{center}
All these plots have $|(Y_S)_{33}|$ fixed at its maximally allowed value $0.0097$.
From here, other configurations can be obtained by simply scaling down $Y_S$ by $Y_S \rightarrow \lambda^{-2} Y_S$ ( with $\lambda>1$ ). In response, all the LFV processes branching ratios move up as $\lambda^4$\footnote{During the scaling, one needs to recheck the configuration is still phenomenologically viable. }. In some plots, the dashed arrows are put in to guide the reader's eyes and show the drifting direction of the
branching ratios during the re-scaling. As the $Y_S$ is dialed down, all the points of ${\cal B}^Z_{ll'}$ go up along the indicated direction until the $B(\mu\rightarrow e\gamma)$ hits the current experimental limit. Interestingly, we can predict the upper limits on ${\cal B}^Z_{ll'}$, ranging from $10^{-11}$ to $10^{-9}$, for the $Y_R=0$ case.
We stress that these upper limits are tied with the $Y_R=0$ assumption; they could be much larger if $Y_R\neq 0$.
This part of parameter space of cZBM could be probed at the planned TeraZ collider where about $10^{12}$ Z bosons will be produced per year with a few $ab^{-1}$ luminosity\cite{TLEP,CEPC}. Moreover, this particular assumption will be ruled out if any excess was measured in the future experiment.
On the other hand, the lower limits on these LFV processes are rather robust and insensitive to the $Y_R=0$ assumption.
Similarly, interesting upper and lower bounds on ${\cal B}(\ell \rightarrow \ell' \gamma)$ and the CP asymmetries $\eta_{\ell \ell'}$ can be predicted in cZBM, see Tab.\ref{tab:range_Z_eta}. Note that the upper bounds on all three ${\cal B}(\ell \rightarrow \ell' \gamma)$ are just below the current experimental limits. Any improvement in these measurements will cut across the interesting parameter space of cZBM.
On the other hand, the cZBM with democratic $Y_S$ and $m_\Delta\sim{\cal O}(\tev)$ can be falsified if no such cLFV processes had been detected above the predicted lower bounds in the future experiments.
\begin{table*}[h]
\begin{center}
\begin{tabular}{c|c|c}
\hline\hline
&lower bounds & upper bounds (for $Y_R=0$) \\
\hline
$\mathcal{B}( \mu\rightarrow e\gamma ) $ & $3.05 \times 10^{-16}$ ($3.98 \times 10^{-18}$) & $ 5.7(5.7) \times 10^{-13} $ \\
$\mathcal{B}( \tau\rightarrow e \gamma ) $ & $3.16 \times 10^{-16}$ ($2.03 \times 10^{-18}$) & $ 2.3 (0.51)\times 10^{-9} $ \\
$\mathcal{B}( \tau \rightarrow \mu \gamma ) $ & $4.67 \times 10^{-17}$ ($1.68 \times 10^{-16}$) & $ 3.4 (2.8) \times 10^{-8} $ \\
\hline\hline
$\mathcal{B}^Z_{e \mu} $ & $2.5 \times 10^{-16}$ ($4.9 \times 10^{-14}$) & $2.2 (8.7) \times 10^{-11}$ \\
$\mathcal{B}^Z_{e \tau} $ & $2.9 \times 10^{-16}$ ($4.6 \times 10^{-14}$) & $ 3.6 (1.0 ) \times 10^{-10} $ \\
$\mathcal{B}^Z_{\mu \tau} $ & $2.5 \times 10^{-14}$ ($7.8 \times 10^{-15}$) & $ 5.5 (4.5 )\times 10^{-9} $ \\
\hline\hline
$\eta_{\mu e} $ & ${}^{+.68}_{-.67}({}^{+2.1}_{-.97}) \times 10^{-13}$ &$ {}^{+2.6}_{-5.4} ({}^{+9.3}_{-8.1}) \times 10^{-13} $ \\
$\eta_{\tau e} $ & ${}^{+2.4}_{-.20}({}^{+.20}_{-1.2}) \times 10^{-12}$ &$ {}^{+2.3}_{-.56} ({}^{+.22}_{-.10}) \times 10^{-11} $ \\
$\eta_{\tau \mu}$ & ${}^{+2.3}_{-.78}({}^{+1.3}_{-1.3}) \times 10^{-11}$ & $ {}^{+3.7}_{-8.1} ({}^{+3.0}_{-3.1}) \times 10^{-11}$\\
\hline\hline
\end{tabular}
\end{center}
\caption{Range of $\mathcal{B}(\ell \to \ell' \gamma)$, $\mathcal{B}( Z \rightarrow \ell \ell' )$, and $\eta_{\ell \ell'}$ for $m_\Delta =1 \tev$ and $m_S=7\tev$. The numbers( in the parentheses ) are for NH(IH) neutrino masses.
The lower bounds are the lowest values found in the numerical search with $|(Y_S)_{33}| = 0.0097$.
The upper bounds are found by rescaling $Y_S$, see text. Note that the sign for $\eta_{\ell \ell'}$ could be either ways.
For a different leptoquark/di-quark mass, all the values should be multiplied by a factor of $(1\tev\cdot m_S/7 m_\Delta^2)^2$.
}
\label{tab:range_Z_eta}
\end{table*}
Note that the double ratio of any pair of cLFV process branching ratios is invariant under the $Y_S$ re-scaling and independent of $m_\Delta$. Our numerical also has concrete
predictions for $Y_R=0$ and these double ratios depend on the neutrino mass pattern, see Fig.\ref{fig:double_R}. Therefore they could provide an intriguing mean to determine the neutrino mass hierarchy.
In particular, the neutrino mass hierarchy can be unambiguously determined if the measured values fell into any of the decisive windows listed in Tab.\ref{tab:sweetspots}. Most of these interesting double ratio windows are plagued by either small cLFV branching ratios or very limited parameter space. However, $R_5\equiv {\cal B}(\tau\rightarrow \mu \gamma )/ {\cal B}(\tau\rightarrow e \gamma )$ and $R_7\equiv {\cal B}^Z_{\mu\tau}/ {\cal B}^Z_{\mu e}$ look quite promising. In the cZBM, if $R_5$ is measured in the future rare tau decay experiment to be within $0.03$ and $30$, the neutrino masses are of NH. If $R_7<1.0$ is measured in the future $Z$-factory, the neutrino masses are of IH in the cZBM. Even in the worst scenario that none of the measured double ratios overlap with these stated windows, one could still tell which neutrino mass hierarchy is more probable by simple statistics and probability theory. For example, if both $R_4$ and $R_5$ were measured to be $\sim 10^3$, then the IH is roughly 4 times more probable than the NH in the cZBM.
The above discussion clearly demonstrates that the neutrino oscillation experiments and the cLFV measurements are complimentary to one another to better understand the origin of the neutrino masses.
\begin{center}
\begin{figure}[h]
\subfigure[]{\includegraphics[width=0.3\textwidth]{R1.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{R2.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{R3.eps}}\\
\subfigure[]{\includegraphics[width=0.3\textwidth]{R4.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{R5.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{R6.eps}}\\
\subfigure[]{\includegraphics[width=0.3\textwidth]{R7.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{R8.eps}}
\subfigure[]{\includegraphics[width=0.3\textwidth]{R9.eps}}\\
\caption{Side by side comparison of the double ratios in NH and IH. See Tab.\ref{tab:sweetspots} for the double ratio definitions. }
\label{fig:double_R}
\end{figure}
\end{center}
\begin{table*}[h]
\begin{center}
\begin{tabular}{c|c|c}
\hline\hline
Double Ratio & IH & NH \\
\hline
$R_1 \equiv {\cal B}^Z_{\mu\tau}/ {\cal B}(\mu\rightarrow e \gamma )$ & $R_1>10^4$ or $R_1<0.1$ & N.A. \\
$R_2 \equiv {\cal B}^Z_{e \tau}/ {\cal B}(\mu\rightarrow e \gamma )$ & $R_2>10^3$ & $R_2<0.1$ \\
$R_3 \equiv {\cal B}^Z_{e \mu}/ {\cal B}(\mu\rightarrow e \gamma )$& $R_3>10^2$ & $R_3<0.1$ \\
$R_4 \equiv {\cal B}(\tau\rightarrow \mu \gamma )/ {\cal B}(\mu\rightarrow e \gamma )$ & $R_4>10^6$ & $R_4<0.003$\\
$R_5 \equiv {\cal B}(\tau\rightarrow \mu \gamma )/ {\cal B}(\tau\rightarrow e \gamma )$ & N.A. & $0.03 <R_5 <30$ \\
$R_6 \equiv {\cal B}(\tau\rightarrow e \gamma )/ {\cal B}(\mu\rightarrow e \gamma )$ & $R_6<0.03$ & N.A. \\
$R_7 \equiv {\cal B}^Z_{\mu\tau}/ {\cal B}^Z_{\mu e}$ & $R_7<1.0 $ & $R_7>3\times 10^4$ \\
$R_8 \equiv {\cal B}^Z_{e \tau}/ {\cal B}^Z_{e \mu }$ & N.A. & $R_8>10^2$ \\
$R_9 \equiv {\cal B}^Z_{\tau \mu}/{\cal B}^Z_{\tau e}$& $R_9<0.01$ & $R_9>3\times 10^4$ \\
\hline\hline
\end{tabular}
\end{center}
\caption{ The definitions of the cLFV double ratios and the ranges which can be used to determine neutrino mass hierarchy.
}
\label{tab:sweetspots}
\end{table*}
\subsubsection{Leptoquark decay branching ratios}
First, some shorthand notations are introduced:
\beq
\Gamma^{\ell^-_i}_\Delta \equiv \sum_{j} \Gamma(\Delta \rightarrow \ell^-_i u_j)\,,
\Gamma^{\nu_i}_\Delta \equiv \sum_{j} \Gamma(\Delta \rightarrow \nu_i d_j)\,,
\eeq
where the quark flavors are summed over.
For $Y_R=0$, the $SU(2)_L$ symmetry ensures that $\Gamma^{\ell^-_i}_\Delta =\Gamma^{\nu_i}_\Delta\propto |\sum_j(Y_L)_{\ell_i j}|^2$.
This corresponds to the $\beta=1/2$ case that 50\% of the leptoquark decays into a neutrino and a down-type quark.
Since the neutrino is hard to be tracked in the detector, we focus on the decay channels with a high energy charged lepton as the primordial final state and define
\beq
B^\Delta_{\ell_i} \equiv { \Gamma^{\ell^-_i}_\Delta \over \Gamma^{e^-}_\Delta+\Gamma^{\mu^-}_\Delta+\Gamma^{\tau^-}_\Delta}\,.
\eeq
The above defined quantity is clearly independent of $m_\Delta$ and $Y_S$ re-scaling.
\begin{center}
\begin{figure}[h]
\subfigure[]{ \includegraphics[width=0.44\textwidth]{IH_LQ_BR.eps}}\,
\subfigure[]{ \includegraphics[width=0.43\textwidth]{NH_LQ_BR.eps}}
\caption{leptoquark decay branching ratios for (a) IH, (b) NH. }
\label{fig:LQ_BR}
\end{figure}
\end{center}
The decay branching ratios for leptoquark from our numerical study are displayed in Fig.\ref{fig:LQ_BR}.
It can be clearly seen in Fig.\ref{fig:LQ_BR} that, depending on the neutrino mass hierarchy, there are special patterns in the leptoquark decay branching ratios. Roughly speaking, for the IH case, the leptoquark decays are either (1) $B^\Delta_{e}\sim 1.0$ or (2) $B^\Delta_{\mu}\sim 55\%$ and $B^\Delta_{\tau}\sim 45\%$. On the other hand, for the NH case, the $B^\Delta_{\mu}$ and $B^\Delta_{\tau}$ are concentrated in the region roughly enclosed by $0.7 \lesssim B^\Delta_{\mu}+B^\Delta_{\tau}\lesssim 1.0$ and $ 0.2\lesssim B^\Delta_{\tau}\lesssim 0.8$. In other words, $ B^\Delta_{e} \lesssim 0.3$ if the neutrino masses are in the NH.
Surprisingly, our numerical study has not found any configuration which has either $B^\Delta_{\mu}\sim 100\%$ or $B^\Delta_{\tau}\sim 100\%$.
Dictated by the neutrino oscillation data, the model predicts that the leptoquark can NOT decays purely into the 2nd or the 3rd generation charged leptons. These concrete branching ratios could be used to provide the new benchmark leptoquark mass limits with a better motivation.
\section{Conclusion}
We have studied the cZBM which exploits a scalar leptoquark $\Delta(3,1,-1/3)$ and a scalar di-quark $S(6,1,-2/3)$ to generate neutrino masses at the 2-loop level. The neutrino mass matrix element $M^\nu_{ij}$ ( i,j=1,2,3) is proportional to the product of $(Y_L)_{ik} m^d_k (Y_S^\dag)_{k k'}m^d_{k'} (Y_L^T)_{k' j}$, see Eq.(\ref{eq:nu_mass_elements}). The Yukawa couplings $Y_L$ and $Y_S$ are a priori unknown and arbitrary. To proceed, we have adopted a modest working assumption that all six $|Y_S|$ are of the same order. Then the $Y_L$ can be iteratively determined owing to the fact that $m_\Delta, m_S\gg m_b\gg m_s\gg m_d$. Moreover, the mass of the lightest neutrino is of order $10^{-5}$ eV and the model disfavors the case of nearly degenerate neutrinos. The tree-level flavor violating processes will be inevitably mediated by $\Delta$ or $S$ with the realistic $Y_L$ which accommodates the neutrino data. Due to the different chiral structures, the contributions to the flavor violating processes from $Y_L$ and $Y_R$ do not interfere with each other at the tree-level. We have considered the case that $Y_R=0$ to minimized the tree-level flavor violating processes( and expect the same to happen for the loop induced cLFV). We also have argued that $Y_R=0$ is actually favored by the fact that there is no electron EDM has been observed yet.
A comprehensive numerical study has been performed to look for the realistic $Y_L$ and $Y_S$ configurations which pass all the known experimental constraints on the flavor violating processes. The viable configurations were collected and have been used to calculate the resulting 1-loop charged lepton flavor violating $Z \rightarrow \overline{l} l'$ and $l \rightarrow l' \gamma$.
Some of the realistic configurations could be probed in the forthcoming cLFV experiments. Interesting and robust lower bounds have been found for these cLFV, see Tab.\ref{tab:range_Z_eta}. Moreover, the neutrino mass hierarchy could be determined if the measured cLFV double ratio(s) is/are in some specific range(s), see Tab.\ref{tab:sweetspots}.
For $Y_R=0$, $\Delta$ has $50\%$ of chance decaying into a charged lepton and an up-type quark. Specific ratios
$\sum_j B(\Delta\rightarrow l_i u_j)$ for each generation charged lepton $l_i$ have been predicted in this model.
Given the potential link between the neutrino masses generation and $\Delta$, it seems well-motivated using the predicted leptoquark branching ratios as a benchmark scenario for the future scalar leptoquark search limits.
\begin{acknowledgments}
WFC is supported by the Taiwan Minister of Science and Technology under
Grant Nos. 102-2112-M-007-014-MY3 and 105-2112-M-007-029.
FX is supported partially by NSFC (National Natural Science Foundation of China ) under Grant No. 11605076,
as well as the FRFCU (Fundamental Research Funds for the Central Universities in China) under the Grant No. 21616309. FX especially acknowledges the hospitality of Institute of Physics,
Academia Sinica, at which part of the work was done.
\end{acknowledgments}
|
\section{Introduction}
Coupled forward-backward stochastic differential equations (SDEs)
are encountered when one applies the classical stochastic maximum
principle to optimal control or differential game problems (see Ma
and Yong \cite{MaYong1999}, Yong and Zhou \cite{YongZhou1999}). The
existence and uniqueness of solutions to such kind of equations are
closely linked to that of optimal controls or Nash equilibrium
points. Forward-backward SDEs are also used to give a probabilistic
interpretation for quasilinear second order partial differential
equations (PDEs) of elliptic or parabolic type (see Pardoux and Tang
\cite{PardouxTang1999}, Wu and Yu \cite{WuYu2014}), which
generalized the classical Feynman-Kac formula for linear PDEs.
Moreover, in mathematical finance forward-backward SDEs are often
adopted to describe the models involving large investors (see for
example Cvitani\'{c} and Ma \cite{CvitanicMa1996}).
Finite horizon forward-backward SDEs were first investigated by
Antonelli \cite{Antonelli1993} and a local existence and uniqueness
result was obtained. He also constructed a counterexample showing
that, a large time duration might lead to non-solvability just under
the Lipschitz assumption. For the global solvability results, three
fundamental methods are available. The first one is the {\it method
of contraction mapping} used by Pardoux and Tang
\cite{PardouxTang1999}. The second one concerns a kind of {\it
four-step scheme} approach introduced by Ma, Protter and Yong
\cite{MaProtterYong1994} which can be regarded as a combination of
the methods of PDEs and probability theory. This method requires the
non-degeneracy of the forward diffusion, and is only effective in
Markovian frameworks. The third one called {\it method of
continuation} is probabilistic. This method gets rid of the
restriction of non-degeneracy, and can deal with non-Markovian
forward-backward SDEs. As a trade-off, a kind of monotonicity
conditions on the coefficients is introduced to ensure the
solvability, which is restrictive in a different way. This method is
initiated by Hu and Peng \cite{HuPeng1995}, Peng and Wu
\cite{PengWu1999}. Later, Yong \cite{Yong1997,Yong2010} made
improvements and made the method more systematic. For some recent
developments on finite horizon forward-backward SDEs, one can refer
to Ma et al. \cite{MWZZ2011}.
In 2000, Peng and Shi \cite{PengShi2000}, for the first time,
studied infinite horizon forward-backward SDEs driven by Brownian
motions employing the method of continuation. Later, Wu
\cite{Wu2003} studied this problem in some different monotonicity
framework from \cite{PengShi2000}. Yin \cite{Yin2008,Yin2011}
investigates the same issue by the method of contraction mapping.
Some existence and uniqueness results and comparison theorems were
obtained. In this paper, we consider a kind of generalized infinite
horizon forward-backward SDEs driven by both Brownian motions and
Poisson processes as follows:
\begin{equation}\label{Sec1_FBSDE}
\left\{
\begin{aligned}
dx(t) =\ & b(t,x(t),y(t),z(t),r(t,\cdot)) dt
+\sigma(t,x(t),y(t),z(t),r(t,\cdot)) dW(t)\\
& +\int_{\mathcal E} \gamma(t,e,x(t-),y(t-),z(t),r(t,e)) \tilde
N(dt,de),\\
-dy(t) =\ & g(t,x(t),y(t),z(t),r(t,\cdot)) dt -z(t) dW(t)
-\int_{\mathcal E} r(t,e) \tilde N(dt, de),\\
x(0) =\ & \Phi(y(0)),
\end{aligned}
\right.
\end{equation}
where the notations and mappings will be given in Section \ref{Sec2}
and Section \ref{Sec3}. We adopt the model with random jumps, since
jump-diffusion processes characterize stochastic phenomena more
often and accurate than just diffusion processes, which provide us
with more realistic models in practice. For example, in finance,
stock prices often exhibit some jump behaviors. Moreover, financial
markets with jump stock prices provide a rich family of incomplete
financial models. For more information about jump diffusion models,
the interested readers may be referred to Cont and Tankov
\cite{ContTankov2004}, {\O}ksendal and Sulem \cite{OS2007}, Shen,
Meng and Shi \cite{ShenMengShi2014}.
In this paper we study the solvability of forward-backward SDEs by
virtue of the method of continuation. The idea is to introduce a
family of infinite horizon forward-backward SDEs parameterized by
$\alpha\in [0,1]$ such that, when $\alpha=1$ the forward-backward
SDE coincides with \eqref{Sec1_FBSDE} and when $\alpha=0$ the
forward-backward SDE is uniquely solvable. We will show that there
exists a fixed step-length $\delta_0>0$, such that, if, for some
$\alpha_0\in [0,1)$, the parameterized forward-backward SDE is
uniquely solvable, then the same conclusion holds for $\alpha_0$
being replaced by $\alpha_0+\delta\leq 1$ with $\delta\in
[0,\delta_0]$. Once this has been proved, we can increase the
parameter $\alpha$ step by step and finally reach $\alpha=1$.
It is worth noting that, we study a kind of more general coupled
forward-backward SDEs in comparison with
\cite{PengShi2000,Wu2003,Yin2008,Yin2011}. Besides the coupling in
$b, \sigma, \gamma$ and $g$, in this paper the initial values are
also in a coupled form: $x(0) = \Phi(y(0))$ (see
\eqref{Sec1_FBSDE}). The traditional technique treating the coupling
in the initial values (or terminal values) when the horizon is
finite (see for example \cite{HuPeng1995,Yong1997,PengWu1999}) is to
parameterize and analyze the initial coefficient $\Phi$ as the same
as other coefficients $(b,\sigma,\gamma,g)$. When this traditional
technique is used to the case of infinite horizon, we can solve two
special cases: (i) the decoupled case: $\Phi(y(0))=x_0$; (ii) the
strong monotonicity case: there exists a constant $\nu>0$ such that,
for any $y_1,y_2\in\mathbb R^n$, $\langle \Phi(y_1)-\Phi(y_2),\
y_1-y_2 \rangle \leq -\nu|y_1-y_2|^2$. However, in many practical
stochastic optimization problems, these two kinds of conditions are
too strong to satisfy naturally. In the present paper, instead of
the traditional parameterization technique, we employ the classical
mean value theorem of continuous functions and some delicate
techniques to handle the coupling between the two initial values.
This technique was introduced for the first time by Wu and Yu
\cite{WuYu2014} to analyze some algebraic equations. By virtue of
the new technique, the conclusion is improved to a general
monotonicity case: $\langle \Phi(y_1)-\Phi(y_2),\ y_1-y_2 \rangle
\leq 0$ (see (H3.3)-(ii)) which is natural in the viewpoint of
practical optimization problems. A potential application of the new
technique is to deal with the associated finite horizon problems and
hope to improve the corresponding results.
Since backward SDEs on an infinite time horizon are well defined
dynamic systems (see Theorem \ref{Sec2_THM}), it is natural and
appealing to study the corresponding optimal control and game
problems arising from various fields. For example, in mathematical
finance, the first process $y$ of solution to some backward SDE is
used to represent the price of some European contingent claim, and
the other processes $(z,r)$ of solution are used to characterize the
corresponding portfolio. Then, in an incomplete security market, the
minimum price of some contingent claim can be given by
$\inf_{v(\cdot)\in\mathcal V} y^v(0)$, where $(y^v,z^v,r^v)$ is the
solution of some controlled backward SDEs and $v(\cdot)\in\mathcal
V$ is the related control process.
As an application of theoretical results, we study a kind of
backward stochastic linear-quadratic (LQ) optimal control problems,
and then the general differential game problems. The LQ problems
constitute an extremely important category of optimization problems,
because many problems arising from practice can be modeled by them,
and more importantly, many non-LQ problems can be approximated
reasonably by LQ problems. On the other hand, LQ problems tend to
have elegant and complete solutions due to their simple and nice
structures, which also provide some understanding and preliminaries
for the general nonlinear problems. By virtue of the unique
solvability of forward-backward SDEs, we obtain unique optimal
controls for control problems and unique Nash equilibrium points for
game problems in closed forms. To our best knowledge, it is the
first time to study this kind of infinite horizon backward LQ
problems. The theoretical results of forward-backward SDEs can also
be applied to nonlinear infinite horizon backward optimization
problems. This subject will be detailed in our future works.
The present paper has the following improvements. (i) Compared with
Peng and Shi \cite{PengShi2000}, we clarify many ambiguous
arguments, supplement some necessary details and improve some
proofs. (ii) A general case in which the two initial values are in a
coupled form is studied in this paper, and to deal with it we
introduce a new technique which also can be applied to analyze other
problems. (iii) We provide an important application of the
theoretical results to infinite horizon backward LQ problems. (iv)
In order to match practical problems more accurately, we adopt a
wider jump-diffusion model.
The rest of this paper is organized as follows. In Section
\ref{Sec2}, we introduce some notations and some preliminary results
on infinite horizon (forward) SDEs and backward SDEs, especially an
existence and uniqueness result of backward SDEs. In Section
\ref{Sec3}, we devote ourselves to investigating the infinite
horizon jump-diffusion coupled forward-backward SDEs. We establish
an existence and uniqueness theorem, two stability results and a
comparison theorem for solutions to forward-backward SDEs. In
Section \ref{Sec4}, we apply the existence and uniqueness theorem to
study a kind of infinite horizon backward stochastic LQ optimal
control and differential game problems. We obtain the unique optimal
control for the control problem, and the unique Nash equilibrium
point for the game problem in closed forms.
\section{Notations and preliminaries on SDEs and backward
SDEs}\label{Sec2}
Let $\mathbb R^n$ be the $n$-dimensional Euclidean space with the
usual Euclidean norm $|\cdot|$ and the usual Euclidean inner product
$\langle \cdot,\ \cdot \rangle$. Let $\mathbb R^{n\times m}$ be the
space consisting of all $(n\times m)$ matrices with the inner
product:
\[
\langle A,\ B \rangle = \tr\{ AB^\top \},\quad \mbox{for any } A, B
\in \mathbb R^{n\times m},
\]
where $\top$ denotes the transpose of matrices. Thus the norm $|A|$
of $A$ induced by the inner product is given by $|A| =
\sqrt{\mbox{tr}{AA^\top}}$.
Let $(\Omega,\mathcal F,\mathbb F,\mathbb P)$ be a complete filtered
probability space. The filtration $\mathbb F = \{ \mathcal F_t;\
0\leq t<\infty \}$ is generated by two mutually independent
stochastic sources augmented by all $\mathbb P$-null sets. One is a
$d$-dimensional standard Brownian motion $W =
(W_1,W_2,\dots,W_d)^\top$, and the other one consists of $l$
independent Poisson random measures $N = (N_1,N_2,\dots,N_l)^\top$
defined on $\mathbb R_{+}\times\mathcal E$, where $\mathcal E
\subset\mathbb R^{\bar l}\setminus \{0\}$ is a nonempty Borel subset
of some Euclidean space. The compensators of $N$ are $\bar N(dt,de)
= (\pi_1(de)dt, \pi_2(de)dt,\dots,\pi_l(de)dt)$ which make $\{
\tilde N((0,t]\times A) =(N-\bar N)((0,t]\times A);\ 0\leq t<\infty
\}$ a martingale for any $A$ belonging to the Borel field $\mathcal
B(\mathcal E)$ with $\pi_i(A)<\infty$, $i=1,2,\dots,l$. Here, for
each $i=1,2,\dots,l$, $\pi_i$ is a given $\sigma$-finite measure on
the measurable space $(\mathcal E, \mathcal B(\mathcal E))$
satisfying $\int_{\mathcal E} (1\wedge |e|^2) \pi_i(de)<\infty$.
We introduce some spaces:
\begin{itemize}
\item $L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^n)$ where $K\in\mathbb R$, the
space of $\mathbb R^n$-valued $\mathbb F$-progressively measurable
processes $f$ defined on $[0,\infty)$ such that
\[
\norm{f(\cdot)}_{L^{2,K}_{\mathbb F}} := \left(\mathbb
E\int_0^\infty |f(t)|^2 e^{Kt} dt\right)^{\frac 1 2} < \infty,
\]
and for simplicity we denote $L^2_{\mathbb F}(0,\infty;\mathbb R^n)
:= L^{2,0}_{\mathbb F}(0,\infty;\mathbb R^n)$;
\item $S^2_{\mathbb F}(0,T;\mathbb R^n)$ where $T>0$, the space of $\mathbb
F$-progressively measurable processes $f$ which have
right-continuous paths with left limits such that
\[
\norm{f(\cdot)}_{S^2_{\mathbb F}}:= \left(\mathbb E\left[ \sup_{t\in
[0,T]} |f(t)|^2 \right]\right)^{\frac 1 2 } < \infty,
\]
and $S^{2,loc}_{\mathbb F}(0,\infty;\mathbb R^n) := \bigcap_{T>0}
S^2_{\mathbb F}(0,T;\mathbb R^n)$;
\item $\mathcal X^K(0,\infty;\mathbb R^n) := L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^n)\cap S^{2,loc}_{\mathbb
F}(0,\infty;\mathbb R^n)$; similarly, we denote $\mathcal
X(0,\infty;\mathbb R^n) := \mathcal X^0(0,\infty;\mathbb R^n) =
L^2_{\mathbb F}(0,\infty;\mathbb R^n) \cap S^{2,loc}_{\mathbb
F}(0,\infty;\mathbb R^n)$;
\item $L^2(\mathcal E,\mathcal B(\mathcal E),\pi;\mathbb R^{n\times
l})$, the space of $\pi$-almost sure equivalence classes $r(\cdot) =
(r_1(\cdot),\cdots,r_l(\cdot))$ formed by the mappings from
$\mathcal E$ to the space of $\mathbb R^{n\times l}$-valued matrices
such that
\[
\norm{r(\cdot)} := \left( \int_{\mathcal E} \tr \left\{
r(e)\diag(\pi(de))r(e)^\top \right\} \right)^{\frac 1 2} =\left(
\sum_{i=1}^l \int_{\mathcal E} |r_i(e)|^2 \pi_i(de) \right)^{\frac 1 2}
<\infty.
\]
This space is equipped with the following inner product:
\[
\langle r(\cdot),\ \bar r(\cdot) \rangle := \int_{\mathcal E} \tr
\left\{ r(e) \diag(\pi(de)) \bar r(e)^\top \right\},\quad\forall\
r(\cdot),\bar r(\cdot)\in L^2(\mathcal E,\mathcal B(\mathcal
E),\pi;\mathbb R^{n\times l});
\]
\item $M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times l})$ where $K\in\mathbb R$, the space of $\mathbb R^{n\times
l}$-valued, $\mathcal P\otimes \mathcal B(\mathcal E)$-measurable
processes $r$ such that
\[
\begin{aligned}
\norm{r(\cdot,\cdot)}_{M^{2,K}_{\mathbb F}}:=\ & \left(\mathbb
E\int_0^\infty\int_{\mathcal E} \tr \left\{
r(t,e)\diag(\pi(de))r(t,e)^\top \right\} e^{Kt} dt\right)^{\frac 1 2
}\\
=\ & \left( \mathbb E\int_0^\infty \norm{r(t,\cdot)}^2 e^{Kt} dt
\right)^{\frac 1 2} <\infty,
\end{aligned}
\]
where $\mathcal P$ is the $\sigma$-algebra generated by the $\mathbb
F$-progressively measurable processes on $[0,\infty)\times \Omega$,
and we denote $M^{2}_{\mathbb F}(0,\infty;\mathbb R^{n\times l}) : =
M^{2,0}_{\mathbb F}(0,\infty;\mathbb R^{n\times l})$;
\end{itemize}
Clearly, for any $K_1 < K_2$, $L^{2,K_2}_{\mathbb
F}(0,\infty;\mathbb R^n)\subset L^{2,K_1}_{\mathbb
F}(0,\infty;\mathbb R^n)$ and $M^{2,K_2}_{\mathbb
F}(0,\infty;\mathbb R^{n\times l}) \subset M^{2,K_1}_{\mathbb
F}(0,\infty;\mathbb R^{n\times l})$, i.e. the sequences of spaces
$\{L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^n)\}_{K\in\mathbb R}$ and
$\{M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times
l})\}_{K\in\mathbb R}$ are decreasing in $K$.
Further, we define the space $\mathcal R := \mathbb R^n\times
\mathbb R^n\times\mathbb R^{n\times d} \times L^2(\mathcal
E,\mathcal B(\mathcal E),\pi;\mathbb R^{n\times l})$. For any
$\theta_1 = (x_1,y_1,z_1,r_1(\cdot))$, $\theta_2 =
(x_2,y_2,z_2,r_2(\cdot)) \in \mathcal R$, the inner product is
defined by
\[
\langle \theta_1,\ \theta_2 \rangle := \langle x_1,\ x_2 \rangle
+\langle y_1,\ y_2 \rangle +\langle z_1,\ z_2 \rangle +\langle
r_1(\cdot),\ r_2(\cdot)\rangle.
\]
Then the norm of $\mathcal R$ is deduced by $|\theta| :=
\sqrt{\langle \theta,\ \theta \rangle}$. We also define
\[
\mathcal L^{2,K}_{\mathbb F}(0,\infty) := L^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^n) \times L^{2,K}_{\mathbb F}(0,\infty;\mathbb
R^n) \times L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times d})
\times M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times l})
\]
with the norm
\[
\norm{\theta(\cdot)}_{\mathcal L^{2,K}_\mathbb F} = \left\{ \mathbb
E\int_0^\infty |\theta(t)|^2 e^{Kt} dt \right\}^{\frac 1 2} =\left\{
\mathbb E\int_0^\infty \Big[ |x(t)|^2 +|y(t)|^2 +|z(t)|^2
+\norm{r(t,\cdot)}^2 \Big] e^{Kt} dt \right\}^{\frac 1 2}.
\]
Moreover, $\mathcal L^{2}_{\mathbb F}(0,\infty) := \mathcal
L^{2,0}_{\mathbb F}(0,\infty)$.
Now, let us consider an infinite horizon (forward) stochastic
differential equation (SDE):
\[
\begin{aligned}
x(t)=\ & x_0 +\int_0^t b(s,x(s))ds +\int_0^t \sigma(s,x(s)) dW(s)\\
& +\int_0^t\int_{\mathcal E} \gamma(s,e,x(s-)) \tilde N(ds,de),
\quad t\in [0,\infty),
\end{aligned}
\]
which is also expressed in a differential form:
\begin{equation}\label{Sec2_SDE}
\left\{
\begin{aligned}
dx(t) =\ & b(t,x(t)) dt +\sigma(t,x(t)) dW(t) +\int_{\mathcal E}
\gamma(t,e,x(t-)) \tilde N(dt,de),\quad t\in
[0,\infty),\\
x(0) =\ & x_0,
\end{aligned}
\right.
\end{equation}
where $x_0\in\mathbb R^n$, $b: \Omega\times [0,\infty) \times\mathbb
R^n\rightarrow\mathbb R^n$, $\sigma: \Omega\times [0,\infty)
\times\mathbb R^n\rightarrow\mathbb R^{n\times d}$ and
$\gamma:\Omega\times [0,\infty) \times\mathcal E\times\mathbb R^n
\rightarrow \mathbb R^{n\times l}$. Moreover, we introduce the
following assumptions:
\begin{enumerate}[(\mbox{H1}.1)]
\item For any $x\in\mathbb R^n$, $b(\cdot,x)$, $\sigma(\cdot,x)$ are
$\mathbb F$-progressively measurable and $\gamma(\cdot,\cdot,x)$ is
$\mathcal P\otimes\mathcal B(\mathcal E)$-measurable. Moreover,
there exists a constant $K\in\mathbb R$ such that $b(\cdot,0)\in
L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^n)$, $\sigma(\cdot,0)\in
L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times d})$ and
$\gamma(\cdot,\cdot,0) \in M^{2,K}_{\mathbb F}(0,\infty;\mathbb
R^{n\times l})$.
\item $b$, $\sigma$ and $\gamma$ are Lipschitz continuous with
respect to $x$, i.e. there exists a constant $C>0$ such that for any
$t\in [0,\infty)$, any $x_1$, $x_2\in\mathbb R^n$,
\[
|b(t,x_1)-b(t,x_2)| +|\sigma(t,x_1)-\sigma(t,x_2)|
+\norm{\gamma(t,\cdot,x_1)-\gamma(t,\cdot,x_2)} \leq C|x_1-x_2|.
\]
\end{enumerate}
By the classical theory of SDEs, under Assumptions (H1.1)-(H1.2),
SDE \eqref{Sec2_SDE} admits a unique strong solution. Precisely, for
any $T\in [0,\infty)$,
\[
\mathbb E\left[ \sup_{x\in [0,T]} |x(t)|^2 \right] <\infty,
\]
i.e. $x\in S^2_{\mathbb F}(0,T;\mathbb R^n)$, and then $x\in
S^{2,loc}_{\mathbb F}(0,\infty;\mathbb R^n)$.
\begin{proposition}\label{Sec2_Prop}
Let Assumptions (H1.1)-(H1.2) hold. We further assume the unique
solution $x$ of SDE \eqref{Sec2_SDE} belongs to $L^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^n)$ where the constant $K$ is given by (H1.1).
Then, we have
\begin{enumerate}[(i)]
\item $x\in\mathcal X^K(0,\infty;\mathbb R^n)$;
\item $\mathbb E[|x(t)|^2 e^{Kt}]$ is bounded and continuous;
\item $\lim_{t\rightarrow\infty} \mathbb E [|x(t)|^2 e^{Kt}] =0$.
\end{enumerate}
\end{proposition}
\begin{proof}
The assertion (i) is obvious since $x$ belongs to both
$S^{2,loc}_{\mathbb F} (0,\infty;\mathbb R^n)$ and $L^{2,K}_{\mathbb
F} (0,\infty;\mathbb R^n)$. For any $t\in [0,\infty)$, we apply
It\^{o}'s formula to $|x(s)|^2 e^{Ks}$ on the interval $[0,t]$:
\[
\begin{aligned}
\mathbb E \left[ |x(t)|^2 e^{Kt} \right] =\ & |x_0|^2 +\mathbb
E\int_0^t 2\Big\langle x(s),\ b(s,x(s)) \Big\rangle e^{Ks} ds\\
& +\mathbb E\int_0^t \Big[ K|x(s)|^2 +|\sigma(s,x(s))|^2
+\norm{\gamma(s,\cdot,x(s))}^2 \Big] e^{Ks} ds.
\end{aligned}
\]
By the Lipschitz condition (H1.2), we have
\[
\begin{aligned}
\mathbb E \left[ |x(t)|^2 e^{Kt} \right] \leq\ & |x_0|^2 +
(1+|K|+2C+4C^2) \mathbb E \int_0^t |x(s)|^2
e^{Ks} ds\\
& +\mathbb E\int_0^t \left[ |b(s,0)|^2 +2|\sigma(s,0)|^2
+2\norm{\gamma(s,\cdot,0)}^2 \right] e^{Ks} ds,
\end{aligned}
\]
where $C$ is the Lipschitz constant. Due to the fact that
$x(\cdot)$, $b(\cdot,0)\in L^{2,K}_{\mathbb F}(0,\infty;\mathbb
R^n)$, $\sigma(\cdot,0)\in L^{2,K}_{\mathbb F} (0,\infty;\mathbb
R^{n\times d})$ and $\gamma(\cdot,\cdot,0)\in M^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^{n\times l})$ (see (H1.1)), the above
inequality implies the deterministic process $\{\mathbb E[|x(t)|^2
e^{Kt}];\ t\geq 0\}$ is bounded. Moreover, in the same way, applying
It\^{o}'s formula to $|x(s)|^2 e^{Ks}$ on the interval $[t_1,t_2]$
leads to
\[
\begin{aligned}
\Big| \mathbb E\left[ |x(t_2)|^2 e^{Kt_2}\right] -\mathbb E \left[
|x(t_1)|^2 e^{Kt_1} \right] \Big| \leq\ & (1+|K|+2C+4C^2) \mathbb E
\int_{t_1}^{t_2} |x(s)|^2
e^{Ks} ds\\
& +\mathbb E\int_{t_1}^{t_2} \left[ |b(s,0)|^2 +2|\sigma(s,0)|^2
+2\norm{\gamma(s,\cdot,0)}^2 \right] e^{Ks} ds,
\end{aligned}
\]
which implies that the process $\{\mathbb E[|x(t)|^2 e^{Kt}];\ t\geq
0\}$ is continuous. We have proved (ii). The above inequality also
shows that, $\mathbb E\left[ |x(t_2)|^2 e^{Kt_2}\right] -\mathbb E
\left[ |x(t_1)|^2 e^{Kt_1} \right] \rightarrow 0$ as
$t_1,t_2\rightarrow\infty$, then $\lim_{t\rightarrow\infty} \mathbb
E\left[ |x(t)|^2 e^{Kt} \right]$ exists. Furthermore, due to
\[
\int_0^\infty \mathbb E\left[ |x(t)|^2 e^{Kt} \right] dt <\infty,
\]
we get the desired conclusion
\[
\lim_{t\rightarrow\infty} \mathbb E\left[ |x(t)|^2 e^{Kt} \right]
=0.
\]
\end{proof}
In the rest of this section, we shall consider an infinite horizon
backward SDE as follows:
\[
\begin{aligned}
y(t) =\ & \int_t^\infty \Big[ G(s,y(s),z(s),r(s,\cdot)) +\varphi(s)
\Big] ds -\int_t^\infty z(s) dW(s)\\
& -\int_t^\infty\int_{\mathcal E} r(s,e) \tilde N(ds,de), \quad t\in
[0,\infty),
\end{aligned}
\]
which is denoted also in the following differential form:
\begin{equation}\label{Sec2_BSDE}
-dy(t) = \Big[ G(t,y(t),z(t),r(t,\cdot)) +\varphi(t) \Big]dt
-z(t)dW(t) -\int_{\mathcal E} r(t,e) \tilde N(dt, de),\quad t\in
[0,\infty),
\end{equation}
where $G: \Omega\times [0,\infty) \times \mathbb R^m \times \mathbb
R^{m\times d} \times L^2(\mathcal E,\mathcal B(\mathcal
E),\pi;\mathbb R^{m\times l}) \rightarrow \mathbb R^m$ and $\varphi:
\Omega\times [0,\infty) \rightarrow \mathbb R^m$. We assume the
following assumptions on the coefficients $(G,\varphi)$:
\begin{enumerate}[(\mbox{H2}.1)]
\item For any $(y,z,r(\cdot))\in \mathbb R^m \times \mathbb R^{m\times d} \times L^2(\mathcal E, \mathcal B(\mathcal E), \pi; \mathbb R^{m\times
l})$, $G(\cdot,y,z,r(\cdot))$ is $\mathbb F$-progressively
measurable, and satisfies $G(t,0,0,0) = 0$ for any $t\in
[0,\infty)$. Moreover, there exists a constant $K\in\mathbb R$ such
that $\varphi\in L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^m)$.
\item $G$ is Lipschitz continuous with respect to $(y,z,r(\cdot))$,
i.e. there exist constants $C_0\geq 0$, $C_1\geq 0$ and $C_2\geq 0$
such that for any $t\in [0,\infty)$, any $y_1, y_2 \in\mathbb R^m$,
any $z_1, z_2 \in\mathbb R^{m\times d}$, any $r_1(\cdot), r_2(\cdot)
\in L^2(\mathcal E,\mathcal B(\mathcal E),\pi;\mathbb R^{m\times
l})$,
\[
|G(t,y_1,z_1,r_1(\cdot)) -G(t,y_2,z_2,r_2(\cdot))| \leq C_0
|y_1-y_2| +C_1|z_1-z_2| +C_2\norm{r_1(\cdot)-r_2(\cdot)}.
\]
\item $G$ satisfies some `weak monotonicity' conditions in the
following sense: there exists a constant $\rho\in\mathbb R$ such
that for any $t\in [0,\infty)$, any $y_1,y_2\in\mathbb R^m$, any
$z\in\mathbb R^{m\times d}$, any $r(\cdot)\in L^2(\mathcal
E,\mathcal B(\mathcal E),\pi;\mathbb R^{m\times l})$,
\[
\langle G(t,y_1,z,r(\cdot))-G(t,y_2,z,r(\cdot)), y_1-y_2 \rangle
\leq -\rho |y_1-y_2|^2.
\]
\item $K+2\rho-2C_1^2-2C_2^2 >0$.
\end{enumerate}
\begin{remark}
\begin{enumerate}[(i)]
\item In the classical form of backward SDEs, the coefficient is
usually denoted by $g(t,y,z,r(\cdot))$. While in the present paper,
for convenience, we set
\[
G(t,y,z,r(\cdot)) = g(t,y,z,r(\cdot))-g(t,0,0,0),\quad \varphi(t) =
g(t,0,0,0),
\]
and denote the coefficient $g$ by $G+\varphi$.
\item From the density of real numbers, it is easy to see that (H2.4) is
equivalent to the following statement: (H2.4') There exists a
constant $\delta>0$ such that
\begin{equation}\label{Sec2_delta}
K+2\rho-2C_1^2-2C_2^2-\delta>0.
\end{equation}
\end{enumerate}
\end{remark}
Naturally, a triple of mappings $(y(\cdot),z(\cdot),r(\cdot,\cdot))$
is called an adapted solution to backward SDE \eqref{Sec2_BSDE} if
and only if $y$ is an $\mathbb R^m$-valued $\mathbb F$-progressively
measurable process, $z$ is an $\mathbb R^{m\times d}$-valued
$\mathbb F$-progressively measurable process, $r$ is an $\mathbb
R^{m\times l}$-valued $\mathcal P\otimes\mathcal B(\mathcal
E)$-measurable process, and $(y,z,r)$ satisfies \eqref{Sec2_BSDE}.
Similar to forward SDEs, we have the following
\begin{corollary}\label{Sec2_Corollary}
Let Assumptions (H2.1) and (H2.2) hold. We further assume
$(y,z,r)\in L^{2,K}_{\mathbb F} (0,\infty;\mathbb R^m) \times
L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{m\times d}) \times
M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{m\times l})$ is a solution
to backward SDE \eqref{Sec2_BSDE} where the constant $K$ is given by
(H2.1). Then we have
\begin{enumerate}[(i)]
\item $y\in\mathcal X^K(0,\infty;\mathbb R^m)$;
\item $\mathbb E[|y(t)|^2 e^{Kt}]$ is bounded and continuous;
\item $\lim_{t\rightarrow\infty} \mathbb E [|y(t)|^2 e^{Kt}] =0$.
\end{enumerate}
\end{corollary}
\begin{proof}
Since $(y,z,r)$ is a solution to backward SDE \eqref{Sec2_BSDE},
then for any $t\in [0,\infty)$,
\[
\begin{aligned}
y(t) =\ & \int_t^\infty \Big[ G(s,y(s),z(s),r(s,\cdot)) +\varphi(s)
\Big] ds -\int_t^\infty z(s) dW(s) -\int_t^\infty\int_{\mathcal E}
r(s,e) \tilde N(ds,de)\\
=\ & \int_0^\infty \Big[ G(s,y(s),z(s),r(s,\cdot)) +\varphi(s) \Big]
ds -\int_0^\infty z(s) dW(s) -\int_0^\infty\int_{\mathcal E} r(s,e)
\tilde N(ds,de)\\
& -\int_0^t \Big[ G(s,y(s),z(s),r(s,\cdot)) +\varphi(s) \Big] ds
+\int_0^t z(s) dW(s) +\int_0^t\int_{\mathcal E} r(s,e) \tilde
N(ds,de)\\
=\ & y(0) -\int_0^t \Big[ G(s,y(s),z(s),r(s,\cdot)) +\varphi(s)
\Big] ds +\int_0^t z(s) dW(s) +\int_0^t\int_{\mathcal E} r(s,e)
\tilde N(ds,de).
\end{aligned}
\]
So the process $y\in L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^m)$ can
be regarded as an adapted solution to a forward SDE with
\[
b(t,y) = -\Big[ G(t,y,z(t),r(t,\cdot)) +\varphi(t) \Big],\quad
\sigma(t,y) = z(t),\quad \gamma(t,e,y) =r(t,e).
\]
Under Assumptions (H2.1) and (H2.2), it is easy to check that the
above coefficients $(b,\sigma,\gamma)$ satisfy Assumptions
(H1.1)-(H1.2). By Proposition \ref{Sec2_Prop}, we get the
conclusions.
\end{proof}
In order to obtain an existence and uniqueness result for backward
SDE \eqref{Sec2_BSDE}, we first establish the following a priori
estimate.
\begin{lemma}\label{Sec2_Lemma}
Let Assumptions (H2.1)-(H2.4) hold. Let $(y_1,z_1,r_1)$ and
$(y_2,z_2,r_2)\in L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^m) \times
L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{m\times d}) \times
M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{m\times l})$ be solutions to
backward SDEs \eqref{Sec2_BSDE} with $\varphi=\varphi_1$ and
$\varphi=\varphi_2$ respectively. Then we have
\begin{equation}\label{Sec2_A_Priori}
\begin{aligned}
& \mathbb E\int_0^\infty \Big[ (K+2\rho-2C_1^2 -2C_2^2 -\delta)
|y_1(t)-y_2(t)|^2\\
& +\frac 1 2 |z_1(t)-z_2(t)|^2 +\frac 1 2 \norm{r_1(t,\cdot)-r_2(t,\cdot)}^2 \Big] e^{Kt} dt\\
\leq\ & \frac 1 \delta \mathbb E\int_0^\infty
|\varphi_1(t)-\varphi_2(t)|^2 e^{Kt} dt,
\end{aligned}
\end{equation}
where $\delta>0$ is defined in \eqref{Sec2_delta}.
\end{lemma}
\begin{proof}
We denote
\[
\begin{aligned}
& \hat \varphi(t) := \varphi_1(t)-\varphi_2(t), \quad \hat y(t) :=
y_1(t)-y_2(t),\\
& \hat z(t) := z_1(t)-z_2(t), \quad \hat r(t,e) :=
r_1(t,e)-r_2(t,e),
\end{aligned}
\]
for any $(\omega,t,e)\in \Omega \times [0,\infty)\times\mathcal E$.
For any given $T>0$, we apply It\^{o}'s formula to $|\hat y(t)|^2
e^{Kt}$ on the interval $[0,T]$:
\[
\begin{aligned}
& \mathbb E\Big[ |\hat y(T)|^2 e^{KT} \Big] -|\hat y(0)|^2\\
=\ & \mathbb E\int_0^T \left[ K|\hat y(t)|^2 +|\hat z(t)|^2
+\norm{\hat r(t,\cdot)}^2 \right]
e^{Kt} dt\\
& -2 \mathbb E\int_0^T \Big\langle \hat y(t),\
G(t,y_1(t),z_1(t),r_1(t,\cdot)) -G(t,y_2(t),z_2(t),r_2(t,\cdot))
+\hat\varphi(t) \Big\rangle e^{Kt} dt.
\end{aligned}
\]
Then
\[
\begin{aligned}
& |\hat y(0)|^2 +\mathbb E\int_0^T \left[ K|\hat y(t)|^2 +|\hat
z(t)|^2 +\norm{\hat r(t,\cdot)}^2 \right] e^{Kt} dt\\
=\ & \mathbb E \Big[ |\hat y(T)|^2 e^{KT} \Big] +2\mathbb E\int_0^T
\left\langle \hat y(t),\ \hat\varphi(t) \right\rangle e^{Kt} dt\\
& +2\mathbb E\int_0^T \Big\langle \hat y(t),\
G(t,y_1(t),z_1(t),r_1(t,\cdot)) -G(t,y_2(t),z_1(t),r_1(t,\cdot))
\Big\rangle e^{Kt} dt\\
& +2\mathbb E\int_0^T \Big\langle \hat y(t),\
G(t,y_2(t),z_1(t),r_1(t,\cdot)) -G(t,y_2(t),z_2(t),r_2(t,\cdot))
\Big\rangle e^{Kt} dt.
\end{aligned}
\]
By Assumptions (H2.2) and (H2.3), we have
\[
\begin{aligned}
& |\hat y(0)|^2 +\mathbb E\int_0^T \left[ K|\hat y(t)|^2 +|\hat
z(t)|^2 +\norm{\hat r(t,\cdot)}^2 \right] e^{Kt} dt\\
\leq\ & \mathbb E \Big[ |\hat y(T)|^2 e^{KT} \Big] +\mathbb
E\int_0^T \Big[ 2|\hat y(t)| |\hat\varphi(t)| -2\rho |\hat y(t)|^2
+2|\hat y(t)|(C_1|\hat z(t)| +C_2\norm{\hat r(t,\cdot)}) \Big]
e^{Kt} dt\\
\leq\ & \mathbb E \Big[ |\hat y(T)|^2 e^{KT} \Big] +\mathbb
E\int_0^T \Big[ (\delta-2\rho+2C_1^2+2C_2^2) |\hat y(t)|^2 +\frac 1
2 |\hat z(t)|^2 +\frac 1 2 \norm{\hat r(t,\cdot)}^2
+\frac{1}{\delta} |\hat\varphi(t)|^2 \Big] e^{Kt} dt.
\end{aligned}
\]
Therefore,
\[
\begin{aligned}
& |\hat y(0)|^2 +\mathbb E\int_0^T
\Big[(K+2\rho-2C_1^2-2C_2^2-\delta) |\hat y(t)|^2 +\frac 1 2 |\hat
z(t)|^2 +\frac 1 2 \norm{\hat r(t,\cdot)}^2\Big] e^{Kt} dt\\
\leq\ & \mathbb E \Big[ |\hat y(T)|^2 e^{KT} \Big] +\frac{1}{\delta}
\mathbb E\int_0^T |\hat\varphi(t)|^2 e^{Kt} dt.
\end{aligned}
\]
Let $T\rightarrow\infty$. Thanks to Corollary \ref{Sec2_Corollary},
we have
\[
\begin{aligned}
& |\hat y(0)|^2 +\mathbb E\int_0^\infty
\Big[(K+2\rho-2C_1^2-2C_2^2-\delta) |\hat y(t)|^2 +\frac 1 2 |\hat
z(t)|^2 +\frac 1 2 \norm{\hat r(t,\cdot)}^2\Big] e^{Kt} dt\\
\leq\ & \frac{1}{\delta} \mathbb E\int_0^\infty |\hat\varphi(t)|^2
e^{Kt} dt,
\end{aligned}
\]
which implies \eqref{Sec2_A_Priori}.
\end{proof}
\begin{theorem}\label{Sec2_THM}
Let Assumptions (H2.1)-(H2.4) hold. When $K>0$, the backward SDE
\eqref{Sec2_BSDE} admits a unique solution $(y,z,r)\in
L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^m) \times L^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^{m\times d}) \times M^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^{m\times l})$.
\end{theorem}
\begin{proof}
Clearly, the a priori estimate \eqref{Sec2_A_Priori} implies the
uniqueness. For the existence, we employ the method used in
\cite{PengShi2000} to construct an adapted solution. In detail, for
$n=1,2,\dots$, we define
\[
\varphi_n(t) = \mathbbm{1}_{[0,n]}(t) \varphi(t),\quad t\in
[0,\infty).
\]
Obviously, the sequence $\{\varphi_n\}_{n=1}^{\infty}$ converges to
$\varphi$ in $L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^m)$. For each
$n$, let $(\bar y_n,\bar z_n,\bar r_n)$ be the unique adapted
solution of the following finite horizon backward SDE:
\[
\begin{aligned}
\bar y_n(t) =\ & \int_t^n \Big[ G(s,\bar y_n(s),\bar z_n(s),\bar
r_n(s,\cdot)) +\varphi_n(s) \Big] ds -\int_t^n \bar z_n(s) dW(s)\\
& -\int_t^n \int_{\mathcal E} \bar r_n(s,e) \tilde N(ds,de),\quad
t\in [0,n].
\end{aligned}
\]
Furthermore, we define
\[
(y_n(t),z_n(t),r_n(t,\cdot)) := \left\{
\begin{aligned}
& (\bar y_n(t), \bar z_n(t), \bar r_n(t,\cdot)), \quad && t\in [0,n],\\
& (0,0,0),\quad && t\in (n,\infty).
\end{aligned}
\right.
\]
Obviously, $(y_n,z_n,r_n)\in L^{2,K}_{\mathbb F}(0,\infty;\mathbb
R^m) \times L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{m\times d})
\times M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{m\times l})$. Since
$G(s,0,0,0)=0$ (see Assumption (H2.1)), then $(y_n,z_n,r_n)$ solves
the following infinite horizon backward SDE:
\[
\begin{aligned}
y_n(t) =\ & \int_t^\infty \Big[ G(s, y_n(s), z_n(s),
r_n(s,\cdot)) +\varphi_n(s) \Big] ds -\int_t^\infty z_n(s) dW(s)\\
& -\int_t^\infty \int_{\mathcal E} r_n(s,e) \tilde N(ds,de),\quad
t\in [0,\infty).
\end{aligned}
\]
Lemma \ref{Sec2_Lemma} implies that $\{(y_n,z_n,r_n)\}_{n=1}^\infty$
is a Cauchy sequence in $L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^m)
\times L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{m\times d}) \times
M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{m\times l})$. We denote by
$(y,z,r)$ the limit of $\{(y_n,z_n,r_n)\}_{n=1}^\infty$, and shall
show that $(y,z,r)$ solves the backward SDE \eqref{Sec2_BSDE}.
First, when $K>0$ we deduce that
\[
\begin{aligned}
\mathbb E\left[\int_t^\infty (z_n(s)-z(s)) dW(s)\right]^2 =\ &
\mathbb E\int_t^\infty |z_n(s)-z(s)|^2 ds\\
\leq\ & \mathbb E \int_0^\infty |z_n(s)-z(s)|^2 e^{Ks} ds
\rightarrow 0,\quad \mbox{as } n\rightarrow\infty,
\end{aligned}
\]
i.e. the item $\int_t^\infty z_n(s) dW(s)$ converges to
$\int_t^\infty z(s) dW(s)$ in $L^2(\Omega,\mathcal F,\mathbb
P;\mathbb R^m)$ which is the space of $\mathcal F$-measurable square
integrable random variables. The same argument also leads to a
similar conclusion: the item $\int_t^\infty \int_{\mathcal E}
r_n(s,e) \tilde N(ds, de)$ converges to $\int_t^\infty\int_{\mathcal
E} r(s,e) \tilde N(ds, de)$ in $L^2(\Omega,\mathcal F,\mathbb
P;\mathbb R^m)$. Second, for any $K>0$, we have
\[
\begin{aligned}
& \mathbb E\left[ \int_t^\infty \Big(
G(s,y_n(s),z_n(s),r_n(s,\cdot)) -G(s,y(s),z(s),r(s,\cdot)) \Big) ds
\right]^2\\
\leq\ & \mathbb E\left[ \int_0^\infty \Big|
G(s,y_n(s),z_n(s),r_n(s,\cdot)) -G(s,y(s),z(s),r(s,\cdot)) \Big|
e^{\frac K 2 s} e^{-\frac K 2 s} ds \right]^2\\
\leq\ & \mathbb E\left[ \left( \int_0^\infty \Big|
G(s,y_n(s),z_n(s),r_n(s,\cdot)) -G(s,y(s),z(s),r(s,\cdot)) \Big|^2
e^{Ks} ds \right) \left(\int_0^\infty e^{-Ks} ds \right) \right]\\
\leq\ & C \mathbb E\int_0^\infty \Big[ |y_n(s)-y(s)|^2
+|z_n(s)-z(s)|^2 +\norm{r_n(s,\cdot)-r(s,\cdot)}^2 \Big] e^{Ks} ds\\
& \rightarrow 0,\quad \mbox{as } n\rightarrow\infty,
\end{aligned}
\]
i.e. the item $\int_t^\infty G(s,y_n(s),z_n(s),r_n(s,\cdot)) ds$
converges to $\int_t^\infty G(s,y(s),z(s),r(s,\cdot)) ds$ in
$L^2(\Omega,\mathcal F,\mathbb P;\mathbb R^m)$. We notice that, here
in order to dominate the $L^1$-norm by the $L^2$-norm, we have to
restrict $K>0$. This is different from the case of finite time
intervals. The same argument also leads to $\int_t^\infty
\varphi_n(s) ds$ converges to $\int_t^\infty \varphi(s) ds$ in
$L^2(\Omega,\mathcal F,\mathbb P;\mathbb R^m)$. At last, since
$\lim_{n\rightarrow\infty} \mathbb E \int_0^\infty |y_n(t)-y(t)|^2
e^{Kt} dt = 0$, there exists a subsequence of $\{y_n\}$ such that
\[
\lim_{n\rightarrow\infty} \mathbb E\Big[ |y_n(t)-y(t)|^2 \Big] = 0,
\quad \mbox{for almost everywhere } t \in [0,\infty).
\]
The proof is completed.
\end{proof}
\section{Coupled forward-backward SDEs}\label{Sec3}
In this section, we study the following kind of coupled
forward-backward SDEs driven by both Brownian motions and Poisson
processes on the infinite interval $[0,\infty)$:
\begin{equation}\label{Sec3_FBSDE}
\left\{
\begin{aligned}
dx(t) =\ & b(t,x(t),y(t),z(t),r(t,\cdot)) dt
+\sigma(t,x(t),y(t),z(t),r(t,\cdot)) dW(t)\\
& +\int_{\mathcal E} \gamma(t,e,x(t-),y(t-),z(t),r(t,e)) \tilde
N(dt,de),\\
-dy(t) =\ & g(t,x(t),y(t),z(t),r(t,\cdot)) dt -z(t) dW(t)
-\int_{\mathcal E} r(t,e) \tilde N(dt, de),\\
x(0) =\ & \Phi(y(0)),
\end{aligned}
\right.
\end{equation}
where $\Phi: \mathbb R^n\rightarrow \mathbb R^n$, $(b,\sigma,g):
\Omega\times [0,\infty)\times \mathcal R \rightarrow \mathbb R^n
\times \mathbb R^{n\times d} \times \mathbb R^n$ and $\gamma:
\Omega\times [0,\infty)\times\mathcal E\times \mathcal R \rightarrow
\mathbb R^{n\times l}$. Similar to Hu and Peng \cite{HuPeng1995},
for any $\theta = (x,y,z,r(\cdot)) \in\mathcal R$, we use the
notation $A(t,\theta) := (-g(t,\theta), b(t,\theta),
\sigma(t,\theta), \gamma(t,\cdot,\theta))$. Now we give the
following assumptions:
\begin{enumerate}[(\mbox{H3}.1)]
\item For any $\theta\in \mathcal R$, $b(\cdot,\theta)$, $\sigma(\cdot,\theta)$, $g(\cdot,\theta)$ are
$\mathbb F$-progressively measurable and
$\gamma(\cdot,\cdot,\theta)$ is $\mathcal P\otimes \mathcal
B(\mathcal E)$-measurable. Moreover, there exists a constant $K>0$
such that $A(\cdot,0)\in \mathcal L^{2,K}_{\mathbb F}(0,\infty)$.
\item $A$ and $\Phi$ are Lipschitz continuous with respect to $\theta$ and $y$ respectively, i.e.
there exists a constant $C>0$ such that for any $t\in [0,\infty)$,
any $\theta_1$, $\theta_2 \in \mathcal R$, any $y_1$, $y_2\in\mathbb
R^n$,
\[
\begin{aligned}
\mbox{(i)} & \quad |A(t,\theta_1)-A(t,\theta_2)| \leq C
|\theta_1-\theta_2|,\\
\mbox{(ii)} & \quad |\Phi(y_1)-\Phi(y_2)| \leq C |y_1-y_2|.
\end{aligned}
\]
\item $A$ and $\Phi$ satisfy the monotonicity conditions in the sense: there exists a constant $\mu>0$ such that for any $t\in
[0,\infty)$, any $\theta_1$, $\theta_2\in\mathcal R$, any $y_1$,
$y_2\in \mathbb R^n$,
\[
\begin{aligned}
\mbox{(i)} & \quad \langle A(t,\theta_1)-A(t,\theta_2),\
\theta_1-\theta_2 \rangle \leq -\mu |\theta_1-\theta_2|^2,\\
\mbox{(ii)} & \quad \langle \Phi(y_1)-\Phi(y_2),\ y_1-y_2 \rangle
\leq 0.
\end{aligned}
\]
\item $2\mu -K >0$.
\end{enumerate}
\begin{remark}\label{Sec3_Remark_Artificial}
Assumption (H3.4) is artificial. In fact, if it does not hold true,
then we can find a $\bar K\in (0,K)$ such that $2\mu-\bar K >0$. Due
to the decreasing property of $\{ L^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^n) \}_{K\in\mathbb R}$ and $\{
M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times l}) \}_{K\in\mathbb
R}$, Assumption (H3.1) implies $A(\cdot,0)$ also belongs to
$\mathcal L^{2,\bar K}_{\mathbb F}(0,\infty)$. So we can deal with
the corresponding problems in a larger space. However, for
convenience, we would like to keep (H3.4) in this paper.
\end{remark}
Next we employ the method of continuation originally introduced by
Hu and Peng \cite{HuPeng1995} to obtain the existence and uniqueness
of the forward-backward SDE \eqref{Sec3_FBSDE}. For this purpose, we
introduce a family of infinite horizon forward-backward SDEs
parametrized by $\alpha\in [0,1]$:
\begin{equation}\label{Sec3_FBSDE_Param}
\left\{
\begin{aligned}
dx^\alpha(t) =\ & \left[\alpha b(t,\theta^\alpha(t))
-\mu(1-\alpha)y^\alpha(t) +\phi(t)\right]dt\\
& +\left[ \alpha\sigma(t,\theta^\alpha(t)) -\mu(1-\alpha)z^\alpha(t)
+\psi(t) \right] dW(t)\\
& +\int_{\mathcal E} \left[\alpha \gamma(t,e,\theta^\alpha(t-))
-\mu(1-\alpha)r^\alpha(t,e)+\xi(t,e)\right] \tilde N(dt, de),\\
-dy^\alpha(t) =\ & \left[ \alpha g(t,\theta^\alpha(t))
+\mu(1-\alpha) x^\alpha(t) +\eta(t) \right] dt -z^\alpha(t) dW(t)\\
& -\int_{\mathcal E} r^\alpha(t,e) \tilde N(dt, de),\\
x^\alpha(0) =\ & \Phi(y^\alpha(0)),
\end{aligned}
\right.
\end{equation}
where $(\eta,\phi,\psi,\xi) \in \mathcal L^{2,K}_{\mathbb
F}(0,\infty)$ and we denote $\theta^\alpha(t) :=
(x^\alpha(t),y^\alpha(t),z^\alpha(t),r^\alpha(t,\cdot))$,
$\theta^\alpha(t-) :=
(x^\alpha(t-),y^\alpha(t-),z^\alpha(t),r^\alpha(t,e))$. We notice
that, the coefficient $\Phi$ is not parameterized as the same as
other coefficients $(b,\sigma,\gamma,g)$. This is a difference from
the traditional parameterization technique used in
\cite{HuPeng1995,PengShi2000,PengWu1999,Yong1997,Yong2010}.
When $\alpha =0$, the forward-backward SDE \eqref{Sec3_FBSDE_Param}
is reduced to
\begin{equation}\label{Sec3_FBSDE_Param_0}
\left\{
\begin{aligned}
dx^0(t) =\ & \left[ -\mu y^0(t) +\phi(t) \right] dt +\left[ -\mu
z^0(t) +\psi(t) \right] dW(t)\\
& +\int_{\mathcal E} \left[ -\mu r^0(t,e) +\xi(t,e) \right] \tilde
N(dt, de),\\
-dy^0(t) =\ & \left[ \mu x^0(t) +\eta(t) \right]dt -z^0(t) dW(t)
-\int_{\mathcal E} r^0(t,e) \tilde N(dt, de),\\
x^0(0) =\ & \Phi(y^0(0)).
\end{aligned}
\right.
\end{equation}
Before proving the unique solvability result for
\eqref{Sec3_FBSDE_Param_0}, we need to consider an algebraic
equation related to the coupling of initial conditions. The
technique dealing with the algebraic equation is similar to the
proof of Lemma 3.4 in \cite{WuYu2014}.
\begin{lemma}\label{Sec3_Lemma_AE}
Let Assumptions (H3.1)-(H3.4) hold. For any $p\in\mathbb R^n$, the
following algebraic equation
\begin{equation}\label{Sec3_AE}
x = \Phi(x+p)
\end{equation}
admits a unique solution $x\in\mathbb R^n$.
\end{lemma}
\begin{proof}
Firstly, we show the uniqueness. If both $x_1$ and $x_2$ satisfy the
algebraic equation \eqref{Sec3_AE}, then
\[
x_1-x_2 = \Phi(x_1+p) -\Phi(x_2+p).
\]
Making inner product with $x_1-x_2$, by the monotonicity condition
on $\Phi$ (see (H3.3)), we get
\[
|x_1-x_2|^2 = \langle \Phi(x_1+p) -\Phi(x_2+p),\ x_1-x_2 \rangle
\leq 0.
\]
We proved the uniqueness.
Secondly, we prove the existence. We define a new function
\begin{equation}
\lambda(x,p) = \Phi(x+p)-x,\quad (x,p)\in\mathbb R^n\times\mathbb
R^n.
\end{equation}
Making inner product of $\lambda$ and $x$, by the monotonicity
condition of $\Phi$, we get
\[
\begin{aligned}
\langle \lambda(x,p),\ x \rangle =\ & \langle \Phi(x+p),\ x \rangle
-|x|^2\\
=\ & \langle \Phi(x+p)-\Phi(p),\ x \rangle +\langle \Phi(p),\ x
\rangle -|x|^2\\
\leq\ & -|x|^2 +\langle \Phi(p),\ x \rangle.
\end{aligned}
\]
From the inequality: $\langle a,\ b \rangle \leq (1/2)
(|a|^2+|b|^2)$, we have
\begin{equation}\label{Sec3_Temp1_Lemma_AE}
\langle \lambda(x,p),\ x \rangle \leq -\frac 1 2 |x|^2 +\frac 1 2
|\Phi(p)|^2.
\end{equation}
We assert that the above inequality implies that, for any
$p\in\mathbb R^n$, there exists an $x(p)$ such that $\lambda(x(p),p)
=0$. This is equivalent to the existence of the algebraic equation
\eqref{Sec3_AE}. In order to highlight the idea of the proof, here
we only prove this conclusion in a simple case where $n=1$. For the
general case $n>1$, the proof is a bit complicated and technical, we
would like to omit it. The interested readers can be referred to
Appendix in \cite{WuYu2014}. When $n=1$, the inequality
\eqref{Sec3_Temp1_Lemma_AE} is rewritten as
\begin{equation}\label{Sec3_Temp2_Lemma_AE}
\lambda(x,p) x \leq -\frac 1 2 x^2 +\frac{\Phi^2(p)}{2}.
\end{equation}
\begin{enumerate}[(i)]
\item When $x<0$, dividing $x$ on both sides of the inequality
\eqref{Sec3_Temp2_Lemma_AE}, we have
\[
\lambda(x,p) \geq -\frac 1 2 x +\frac{\Phi^2(p)}{2x}.
\]
Letting $x\rightarrow -\infty$, we have $\lambda(x,p)\rightarrow
+\infty$.
\item When $x>0$, dividing $x$ on both sides of the inequality
\eqref{Sec3_Temp2_Lemma_AE}, we have
\[
\lambda(x,p) \leq -\frac 1 2 x +\frac{\Phi^2(p)}{2x}.
\]
Letting $x\rightarrow +\infty$, we have $\lambda(x,p)\rightarrow
-\infty$.
\end{enumerate}
Obviously $\lambda$ is a continuous function. From the classical
mean value theorem of continuous functions, we know that, for any
$p$, there exists a real number $x(p)$ such that $\lambda(x(p),p)
=0$. We finish the proof of existence.
\end{proof}
\begin{remark}
One conventional approach to prove the solvability of some algebraic
equations is by virtue of a monotone operator with coerciveness (see
for example Zeidler \cite[Theorem 26.A]{Zeidler1990}). Due to
Assumption (H3.3), $\Phi$ is a monotone operator. However, there is
no coercive condition imposed. So this conventional method cannot be
applied for our problem.
\end{remark}
\begin{lemma}\label{Sec3_Lemma_0}
Let Assumptions (H3.1)-(H3.4) hold. For any $(\eta,\phi,\psi,\xi)\in
\mathcal L^{2,K}_{\mathbb F}(0,\infty)$, the forward-backward SDE
\eqref{Sec3_FBSDE_Param_0} admits a unique solution in $\mathcal
L^{2,K}_{\mathbb F}(0,\infty)$.
\end{lemma}
\begin{proof}
Let us consider a linear infinite horizon backward SDE:
\begin{equation}\label{Sec3_FBSDE_Param_0_BSDE}
\begin{aligned}
-dp(t) =\ & \left[ -\mu p(t) +\phi(t) +\eta(t)\right] dt -\left[
(1+\mu)q(t) -\psi(t) \right] dW(t)\\
& -\int_{\mathcal E} \left[ (1+\mu)k(t,e) -\xi(t,e) \right] \tilde
N(dt, de),
\end{aligned}
\end{equation}
and a linear infinite horizon (forward) SDE combined with an
algebraic equation:
\begin{equation}\label{Sec3_FBSDE_Param_0_SDE}
\left\{
\begin{aligned}
dx(t) =\ & \left[ -\mu(x(t)+p(t)) +\phi(t) \right] dt +\left[ -\mu
q(t) +\psi(t) \right] dW(t)\\
& +\int_{\mathcal E} \left[ -\mu k(t,e) +\xi(t,e) \right] \tilde
N(dt, de),\\
x(0) =\ & \Phi(x(0)+p(0)).
\end{aligned}
\right.
\end{equation}
Due to Theorem \ref{Sec2_THM} with $C_1=C_2=0$ and $\rho=\mu$, the
backward SDE \eqref{Sec3_FBSDE_Param_0_BSDE} admits a unique
solution $(p,q,k)\in L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^n)
\times L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times d}) \times
M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times l})$. Once $(p,q,k)$
is solved, by Lemma \ref{Sec3_Lemma_AE}, we can uniquely solve
$x(0)$ from the initial condition $x(0) = \Phi(x(0)+p(0))$. then we
solve SDE \eqref{Sec3_FBSDE_Param_0_SDE}. It admits a unique
solution $x$. Next we shall show that $x\in L^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^n)$. For any constant $T>0$, we apply
It\^{o}'s formula to $|x(t)|^2 e^{K t}$ on the finite interval
$[0,T]$:
\[
\begin{aligned}
& \mathbb E\left[ |x(T)|^2 e^{K T} \right] +(2\mu-
K)\mathbb E\int_0^T |x(t)|^2 e^{K t} dt\\
=\ & |x(0)|^2 +\mathbb E\int_0^T \Big[ 2 \langle x(t),\ \phi(t)-\mu
p(t) \rangle +|\psi(t)-\mu q(t)|^2 +\norm{\xi(t,\cdot)-\mu
k(t,\cdot)}^2 \Big] e^{K t} dt.
\end{aligned}
\]
Since $K <2\mu$, then there exists a constant $\varepsilon>0$ such
that $2\mu- K-\varepsilon >0$. By the inequality $2\langle a,\ b
\rangle \leq \varepsilon |a|^2 +(1/\varepsilon)|b|^2$, we have
\[
\begin{aligned}
& \mathbb E\left[ |x(T)|^2 e^{K T} \right] +(2\mu-
K-\varepsilon) \mathbb E\int_0^T |x(t)|^2 e^{ K t} dt\\
\leq\ & |x(0)|^2 +\mathbb E\int_0^T \left[ \frac{1}{\varepsilon}
|\phi(t)-\mu p(t)|^2 +|\psi(t)-\mu q(t)|^2 +\norm{\xi(t,\cdot)-\mu
k(t,\cdot)}^2 \right] e^{ K t} dt.
\end{aligned}
\]
Letting $T\rightarrow\infty$, we have
\[
\begin{aligned}
&(2\mu-
K-\varepsilon) \mathbb E\int_0^\infty |x(t)|^2 e^{K t} dt\\
\leq\ & |x(0)|^2 +\mathbb E\int_0^\infty \left[
\frac{1}{\varepsilon} |\phi(t)-\mu p(t)|^2 +|\psi(t)-\mu q(t)|^2
+\norm{\xi(t,\cdot)-\mu k(t,\cdot)}^2 \right] e^{ K t} dt.
\end{aligned}
\]
We have proved $x\in L^{2, K}_{\mathbb F}(0,\infty;\mathbb R^n)$.
It is easy to verify that $(x^0,y^0,z^0,r^0) = (x,x+p,q,k)$ is a
solution to the forward-backward SDE \eqref{Sec3_FBSDE_Param_0}. We
proved the existence.
We would like to prove the uniqueness in a bigger space: $\mathcal
L^{2}_{\mathbb F}(0,\infty)$. Let $\theta_1(\cdot) =
(x_1(\cdot),y_1(\cdot),z_1(\cdot),r_1(\cdot,\cdot))$ and
$\theta_2(\cdot) =
(x_2(\cdot),y_2(\cdot),z_2(\cdot),r_2(\cdot,\cdot))$ belonging to
$\mathcal L^{2}_{\mathbb F}(0,\infty)$ be two solutions to the
forward-backward SDE \eqref{Sec3_FBSDE_Param_0}. We denote
$\hat\theta(\cdot) = (\hat x(\cdot),\hat y(\cdot),\hat z(\cdot),\hat
r(\cdot,\cdot)) = (x_1(\cdot)-x_2(\cdot), y_1(\cdot)-y_2(\cdot),
z_1(\cdot)-z_2(\cdot), r_1(\cdot,\cdot)-r_2(\cdot,\cdot))$ and apply
It\^{o}'s formula to $\langle \hat x(t),\ \hat y(t) \rangle$ on the
interval $[0,T]$ to get
\[
\mathbb E\Big[ \langle \hat x(T),\ \hat y(T) \rangle \Big] +\mu
\mathbb E\int_0^T |\hat \theta(t)|^2 dt =\Big\langle
\Phi(y_1(0))-\Phi(y_2(0)), \hat y(0) \Big\rangle.
\]
By the monotonicity condition of $\Phi$, we get
\[
\mathbb E\Big[ \langle \hat x(T),\ \hat y(T) \rangle \Big] +\mu
\mathbb E\int_0^T |\hat \theta(t)|^2 dt \leq 0.
\]
Letting $T\rightarrow\infty$, thanks to Proposition \ref{Sec2_Prop}
and Corollary \ref{Sec2_Corollary}, we have
\[
\mu \mathbb E\int_0^\infty |\hat \theta(t)|^2 dt \leq 0.
\]
The uniqueness is proved.
\end{proof}
The above Lemma \ref{Sec3_Lemma_0} shows that, when $\alpha=0$, the
forward-backward SDE \eqref{Sec3_FBSDE_Param} is in a simple form
and then is uniquely solvable. It is clear that, when $\alpha =1$
and $(\eta, \phi, \psi, \xi)$ vanish, the forward-backward SDE
\eqref{Sec3_FBSDE_Param} coincides with \eqref{Sec3_FBSDE}. We will
show that there exists a fixed step-length $\delta_0>0$, such that,
if, for some $\alpha_0\in [0,1)$, \eqref{Sec3_FBSDE_Param} is
uniquely solvable for any $(\eta, \phi, \psi, \xi) \in \mathcal
L^{2,K}_{\mathbb F}(0,\infty)$, then the same conclusion holds for
$\alpha_0$ being replaced by $\alpha_0+\delta \leq 1$ with
$\delta\in [0,\delta_0]$. Once this has been proved, we can increase
the parameter $\alpha$ step by step and finally reach $\alpha =1$,
which gives the unique solvability of the forward-backward SDE
\eqref{Sec3_FBSDE}. This idea is adopted from
\cite{HuPeng1995,PengShi2000,PengWu1999,Yong1997,Yong2010}, and this
method is called the method of continuation.
Now, we prove the following continuation lemma.
\begin{lemma}\label{Sec3_Lemma_Continuation}
Under Assumptions (H3.1)-(H3.4), there exists an absolute constant
$\delta_0>0$ such that, if, for some $\alpha_0\in [0,1)$, the
forward-backward SDE \eqref{Sec3_FBSDE_Param} is uniquely solvable
in $\mathcal L^{2,K}_{\mathbb F}(0,\infty)$ for any $(\eta, \phi,
\psi, \xi) \in \mathcal L^{2,K}_{\mathbb F}(0,\infty)$, then the
same is true for $\alpha = \alpha_0 +\delta$ with $\delta\in
[0,\delta_0]$ and $\alpha_0+\delta\leq 1$.
\end{lemma}
\begin{proof}
Let $\delta_0$ be determined as follows. Let $\delta\in
[0,\delta_0]$. For each $\theta(\cdot) =
(x(\cdot),y(\cdot),z(\cdot),r(\cdot,\cdot)) \in \mathcal
L^{2,K}_{\mathbb F}(0,\infty)$, we consider the following
forward-backward SDE (compared to \eqref{Sec3_FBSDE_Param} with
$\alpha =\alpha_0 +\delta$):
\begin{equation}
\left\{
\begin{aligned}
dX(t) =\ & \Big[ \alpha_0 b(t,\Theta(t)) -\mu(1-\alpha_0) Y(t)
+\delta\big( b(t,\theta(t)) +\mu y(t) \big) +\phi(t) \Big]dt\\
& +\Big[ \alpha_0 \sigma(t,\Theta(t)) -\mu(1-\alpha_0) Z(t)
+\delta\big( \sigma(t,\theta(t)) +\mu z(t) \big) +\psi(t)
\Big]dW(t)\\
& +\int_{\mathcal E} \Big[ \alpha_0 \gamma(t,e,\Theta(t-))
-\mu(1-\alpha_0) R(t,e)\\
& +\delta\big( \gamma(t,e,\theta(t-)) +\mu r(t,e) \big) +\xi(t,e)
\Big] \tilde N(dt,de),\\
-dY(t) =\ & \Big[ \alpha_0 g(t,\Theta(t)) +\mu(1-\alpha_0)X(t)
+\delta\big( g(t,\theta(t))-\mu x(t) \big) +\eta(t) \Big] dt\\
& -Z(t) dW(t) -\int_{\mathcal E} R(t,e) \tilde N(dt,de),\\
X(0) =\ & \Phi(Y(0)).
\end{aligned}
\right.
\end{equation}
It is easy to check that $\big( \delta\big(
g(\cdot,\theta(\cdot))-\mu x(\cdot) \big) +\eta(\cdot),\ \delta\big(
b(\cdot,\theta(\cdot))+\mu y(\cdot) \big) +\phi(\cdot),\ \delta\big(
\sigma(\cdot,\theta(\cdot))+\mu z(\cdot) \big) +\psi(\cdot),\
\delta\big( \gamma(\cdot,\cdot,\theta(\cdot-)) +\mu r(\cdot,\cdot)
\big) +\xi(\cdot,\cdot) \big) \in \mathcal L^{2,K}_{\mathbb
F}(0,\infty)$. Then, by our assumptions, the above forward-backward
SDE is uniquely solvable in the space $\mathcal L^{2,K}_{\mathbb F}
(0,\infty)$. We denote the unique solution by $\Theta(\cdot) =
(X(\cdot),Y(\cdot),Z(\cdot),R(\cdot,\cdot))$. We have established a
mapping
\[
\Theta = I_{\alpha_0+\delta}(\theta) : \mathcal L^{2,K}_{\mathbb
F}(0,\infty) \rightarrow \mathcal L^{2,K}_{\mathbb F}(0,\infty).
\]
Next we shall prove it is a contraction.
Let $\theta_1 = (x_1,y_1,z_1,r_1)$, $\theta_2 = (x_2,y_2,z_2,r_2)
\in \mathcal L^{2,K}_{\mathbb F}(0,\infty)$ and $\Theta_1 =(X_1,
Y_1, Z_1,R_1) = I_{\alpha_0+\delta}(\theta_1)$, $\Theta_2 =(X_2,
Y_2, Z_2,R_2) = I_{\alpha_0+\delta}(\theta_2)$. Let
\[
\begin{aligned}
\hat \theta =\ & (\hat x,\hat y,\hat z,\hat r) = (x_1-x_2, y_1-
y_2, z_1- z_2, r_1- r_2),\\
\hat \Theta =\ & (\hat X,\hat Y,\hat Z,\hat R) = (X_1-X_2, Y_2- Y_2,
Z_1-Z_2, R_1- R_2).
\end{aligned}
\]
For any $T>0$, applying It\^{o}'s formula to $\langle \hat X(t),\
\hat Y(t) \rangle e^{Kt}$ on the interval $[0,T]$, we have
\[
\begin{aligned}
& \mathbb E\Big[ \left\langle \hat X(T),\ \hat Y(T) \right\rangle
e^{KT} \Big] -\left\langle \Phi(Y_1(0))-\Phi(Y_2(0)),\ \hat Y(0)
\right\rangle\\
=\ & \alpha_0 \mathbb E\int_0^T \left\langle
A(t,\Theta_1(t))-A(t,\Theta_2(t)),\ \hat\Theta(t) \right\rangle
e^{Kt} dt -\mu(1-\alpha_0) \mathbb E\int_0^T |\hat \Theta(t)|^2
e^{Kt} dt\\
& +\delta\mathbb E\int_0^T \left\langle
A(t,\theta_1(t))-A(t,\theta_2(t)),\ \hat\Theta(t) \right\rangle
e^{Kt} dt +\delta\mu \mathbb E\int_0^T \left\langle \hat\theta(t),\
\hat\Theta(t) \right\rangle e^{Kt} dt\\
& + K\mathbb E\int_0^T \left\langle \hat X(t),\ \hat Y(t)
\right\rangle e^{Kt} dt.
\end{aligned}
\]
By Assumptions (H3.2) and (H3.3), we deduce
\[
\begin{aligned}
& \mathbb E\Big[ \left\langle \hat X(T),\ \hat Y(T) \right\rangle
e^{KT} \Big] +(\mu -\frac 1 2 K) \mathbb E\int_0^T |\hat\Theta(t)|^2 e^{Kt} dt\\
\leq\ & \delta(C+\mu) \mathbb E\int_0^T |\hat\theta(t)|
|\hat\Theta(t)| e^{Kt} dt.
\end{aligned}
\]
Since $\frac{1}{2}K <\mu$ (see Assumption (H3.4)), then there exists
a constant $\varepsilon>0$ such that $\mu- \frac 1 2 K-\varepsilon
>0$. Then we have
\[
\begin{aligned}
& \mathbb E\Big[ \left\langle \hat X(T),\ \hat Y(T) \right\rangle
e^{KT} \Big] +(\mu -\frac 1 2 K -\varepsilon) \mathbb E\int_0^T |\hat\Theta(t)|^2 e^{Kt} dt\\
\leq\ & \delta^2 \cdot \frac{(C+\mu)^2}{4\varepsilon} \mathbb
E\int_0^T |\hat\theta(t)|^2 e^{Kt} dt.
\end{aligned}
\]
Letting $T\rightarrow\infty$, we have
\[
\mathbb E\int_0^\infty |\hat\Theta(t)|^2 e^{Kt} dt \leq \delta^2
\cdot \frac{(C+\mu)^2}{2\varepsilon(2\mu-K-2\varepsilon)} \mathbb
E\int_0^T |\hat\theta(t)|^2 e^{Kt} dt.
\]
Now we choose $\delta_0^2 =
\frac{\varepsilon(2\mu-K-2\varepsilon)}{2(C+\mu)^2}$, then for any
$\delta\in [0,\delta_0]$, we have the following estimate:
\[
\norm{\hat\Theta(\cdot)}_{\mathcal L^{2,K}_{\mathbb F}} \leq \frac 1
2 \norm{\hat\theta(\cdot)}_{\mathcal L^{2,K}_{\mathbb F}}.
\]
This implies that the mapping $I_{\alpha_0+\delta}$ is a
contraction. Hence, it has a unique fixed point, which is the unique
solution of \eqref{Sec3_FBSDE_Param} for $\alpha =\alpha_0+\delta$.
We complete the proof.
\end{proof}
Now, we give an existence and uniqueness result for the
forward-backward SDE \eqref{Sec3_FBSDE}.
\begin{theorem}\label{Sec3_THM}
Under Assumptions (H3.1)-(H3.4), the forward-backward SDE
\eqref{Sec3_FBSDE} admits a unique solution
$(x(\cdot),y(\cdot),z(\cdot),r(\cdot,\cdot)) \in \mathcal
L^{2,K}_{\mathbb F}(0,\infty)$.
\end{theorem}
\begin{proof}
By Lemma \ref{Sec3_Lemma_0} and Lemma \ref{Sec3_Lemma_Continuation},
we can solve the forward-backward SDE \eqref{Sec3_FBSDE_Param}
uniquely for any $\alpha\in [0,1]$ and $(\eta, \phi, \psi, \xi)\in
\mathcal L^{2,K}_{\mathbb F}(0,\infty)$. Particularly,
\eqref{Sec3_FBSDE_Param} with $\alpha=1$ and $(\eta, \phi, \psi,
\xi) =0$, which is \eqref{Sec3_FBSDE}, admits a unique solution. We
finish the proof.
\end{proof}
\begin{remark}
\begin{enumerate}[(i)]
\item By a similar proof as that of Lemma \ref{Sec3_Lemma_0}, the
uniqueness holds true in the bigger space $\mathcal L^2_{\mathbb
F}(0,\infty)$.
\item By Proposition \ref{Sec2_Prop} and Corollary
\ref{Sec2_Corollary}, the unique solution (x,y,z,r) of the
forward-backward SDE \eqref{Sec3_FBSDE} belongs to $\mathcal
X^K(0,\infty;\mathbb R^n) \times \mathcal X^K(0,\infty;\mathbb R^n)
\times L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times d}) \times
M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times l})$ exactly.
\end{enumerate}
\end{remark}
In the rest of this section, we would like to establish some
properties of the solutions to forward-backward SDEs including two
stability results and a comparison theorem. First we establish the
stability results.
\begin{proposition}\label{Sec3.1_Prop_Stability}
Let $(b_1,\sigma_1,\gamma_1,g_1,\Phi_1)$ and
$(b_2,\sigma_2,\gamma_2,g_2,\Phi_2)$ be two sets of coefficients of
forward-backward SDEs satisfying Assumptions (H3.1)-(H3.4). Let
$\theta_1=(x_1,y_1,z_1,r_1)$ and $\theta_2=(x_2,y_2,z_2,r_2)$ be the
corresponding solutions.
\begin{enumerate}[(i)]
\item If we assume that $\Phi_1=\Phi_2$, then
\begin{equation}\label{Sec3.1_L2_1}
\mathbb E\int_0^\infty |\theta_1(t)-\theta_2(t)|^2 e^{Kt} dt \leq C
\mathbb E\int_0^\infty | A_1(t,\theta_2(t)) -A_2(t,\theta_2(t)) |^2
e^{Kt} dt,
\end{equation}
where $C$ is a constant depending on $\mu$ and $K$.
\item If we strengthen the monotonicity condition on $\Phi_1$ as
follows: there exists a constant $\nu>0$ such that for any $y_1$,
$y_2\in\mathbb R^n$,
\begin{equation}\label{Sec3.1_StrongMonot}
\langle \Phi_1(y_1)-\Phi_1(y_2),\ y_1-y_2 \rangle \leq -\nu
|y_1-y_2|^2,
\end{equation}
then
\begin{equation}\label{Sec3.1_L2_2}
\begin{aligned}
& |y_1(0)-y_2(0)|^2 +\mathbb E\int_0^\infty
|\theta_1(t)-\theta_2(t)|^2 e^{Kt} dt\\
& \leq C \left\{ |\Phi_1(y_2(0))-\Phi_2(y_2(0))|^2 + \mathbb
E\int_0^\infty | A_1(t,\theta_2(t)) -A_2(t,\theta_2(t)) |^2 e^{Kt}
dt \right\},
\end{aligned}
\end{equation}
where $C$ is a constant depending on $\mu$, $\nu$ and $K$.
\end{enumerate}
\end{proposition}
\begin{proof}
We apply It\^{o}'s formula to $\langle \hat x(t),\ \hat y(t) \rangle
e^{Kt}$ on the interval $[0,T]$:
\[
\begin{aligned}
& \mathbb E\Big[ \langle\hat x(T),\ \hat y(T)\rangle e^{KT} \Big]
-\langle \Phi_1(y_1(0))-\Phi_1(y_2(0)),\ \hat y(0) \rangle\\
& -\mathbb E\int_0^T \Big\langle A_1(t,\theta_1(t))
-A_1(t,\theta_2(t)),\ \hat \theta(t) \Big\rangle e^{Kt} dt\\
=\ & \langle \Phi_1(y_2(0))-\Phi_2(y_2(0)),\ \hat y(0) \rangle\\
& +\mathbb E\int_0^T \Big\{ \Big\langle
A_1(t,\theta_2(t))-A_2(t,\theta_2(t)), \hat \theta(t) \Big\rangle
+K\Big\langle \hat x(t),\ \hat y(t) \Big\rangle \Big\} e^{Kt} dt,
\end{aligned}
\]
where the notations $\hat x := x_1-x_2$ etc. By the monotonicity
condition on $A_1$, we have
\begin{equation}\label{Sec3.1_Temp}
\begin{aligned}
& \mathbb E\Big[ \langle\hat x(T),\ \hat y(T)\rangle e^{KT} \Big]
-\langle \Phi_1(y_1(0))-\Phi_1(y_2(0)),\ \hat y(0) \rangle
+(\mu-\frac K 2) \mathbb E\int_0^T |\hat\theta(t)|^2 e^{Kt} dt\\
\leq\ & \langle \Phi_1(y_2(0))-\Phi_2(y_2(0)),\ \hat y(0) \rangle
+\mathbb E\int_0^T \Big\langle
A_1(t,\theta_2(t))-A_2(t,\theta_2(t)), \hat \theta(t) \Big\rangle
e^{Kt} dt.
\end{aligned}
\end{equation}
(i) When $\Phi_1=\Phi_2$, considering the monotonicity condition on
$\Phi_1$ (see Assumption (H3.3)), \eqref{Sec3.1_Temp} is reduced to
\[
\begin{aligned}
& \mathbb E\Big[ \langle\hat x(T),\ \hat y(T)\rangle e^{KT} \Big]
+(\mu-\frac K 2) \mathbb E\int_0^T |\hat\theta(t)|^2 e^{Kt} dt\\
\leq\ & \mathbb E\int_0^T \Big\langle
A_1(t,\theta_2(t))-A_2(t,\theta_2(t)), \hat \theta(t) \Big\rangle
e^{Kt} dt.
\end{aligned}
\]
By a similar technique as that in the proof of Lemma
\ref{Sec3_Lemma_Continuation}, we obtain the estimate
\eqref{Sec3.1_L2_1}.
(ii) When $\Phi_1$ satisfies the strong monotonicity condition
\eqref{Sec3.1_StrongMonot}, the inequality \eqref{Sec3.1_Temp} is
reduced to
\[
\begin{aligned}
& \mathbb E\Big[ \langle\hat x(T),\ \hat y(T)\rangle e^{KT} \Big]
+\nu |\hat y(0)|^2
+(\mu-\frac K 2) \mathbb E\int_0^T |\hat\theta(t)|^2 e^{Kt} dt\\
\leq\ & \langle \Phi_1(y_2(0))-\Phi_2(y_2(0)),\ \hat y(0) \rangle
+\mathbb E\int_0^T \Big\langle
A_1(t,\theta_2(t))-A_2(t,\theta_2(t)), \hat \theta(t) \Big\rangle
e^{Kt} dt.
\end{aligned}
\]
Similar to the proof of Lemma \ref{Sec3_Lemma_Continuation}, we get
the estimate \eqref{Sec3.1_L2_2}.
\end{proof}
\begin{remark}
In the proof of the above proposition, it is easy to see, even if
the Lipschitz condition and the monotonicity conditions are not
satisfied by $(b_2,\sigma_2,\gamma_2,g,\Phi)$, the estimates
\eqref{Sec3.1_L2_1} and \eqref{Sec3.1_L2_2} still hold.
\end{remark}
We would like to point out that the $L^2$-estimate (see
\eqref{Sec3.1_L2_1} and \eqref{Sec3.1_L2_2}) for solutions of
forward-backward SDEs will play a key role in studying Pontryagin's
maximum principle and Bellman's dynamic programming principle for
stochastic optimal control and stochastic differential game problems
of infinite horizon forward-backward SDEs.
Next we prove a comparison theorem. As same as before, let $\theta_1
= (x_1,y_1,z_1,r_1)$ and $\theta_2 = (x_2,y_2,z_2,r_2)$ be the
solutions of \eqref{Sec3_FBSDE} with coefficients
$(b,\sigma,\gamma,g,\Phi_1)$ and $(b,\sigma,\gamma,g,\Phi_2)$
respectively. Denote
\[
\hat\theta := \theta_1-\theta_2 = (x_1-x_2,y_1-y_2,z_1-z_2,r_1-z_2)
=: (\hat x,\hat y,\hat z,\hat r).
\]
Similar to Lemma 7 in \cite{PengShi2000}, one can easily prove the
following lemma.
\begin{lemma}\label{Sec3.1_CT_Lemma}
Let Assumptions (H3.1)-(H3.4) holds for $(b,\sigma,\gamma,g,\Phi_1)$
and $(b,\sigma,\gamma,g,\Phi_2)$.
\begin{enumerate}[(i)]
\item For any $t\in [0,\infty)$, we have
\[
\langle \hat x(t),\ \hat y(t) \rangle \geq 0.
\]
\item If we define an $\mathbb F$-stopping time: $\tau = \inf \{ t\geq
0;\ \langle \hat x(t),\ \hat y(t) \rangle =0 \}$, then we further
have
\[
\hat\theta(t)\mathbbm{1}_{[\tau,\infty)}(t) = (\hat x(t),\hat
y(t),\hat z(t),\hat r(t,\cdot))\mathbbm{1}_{[\tau,\infty)}(t) =0.
\]
\end{enumerate}
\end{lemma}
\begin{proof}
(i) Clearly, for any $t\in [0,\infty)$, there exists a sequence of
times $\{T_i\}_{i=1}^\infty$, which increases and diverges as
$i\rightarrow\infty$, such that
\[
\lim_{i\rightarrow\infty} \mathbb E\Big[ \langle \hat x(T_i),\ \hat
y(T_i) \rangle\ \Big|\ \mathcal F_t \Big] =0.
\]
We apply It\^{o}'s formula to $\langle \hat x(s),\ \hat y(s)
\rangle$ on the interval $[t,T_i]$ to have
\[
\begin{aligned}
\mathbb E\Big[ \langle \hat x(T_i),\ \hat y(T_i) \rangle\ \Big|\
\mathcal F_t \Big] -\langle \hat x(t),\ \hat y(t) \rangle =\ &
\mathbb E\bigg[ \int_t^{T_i} \Big\langle A(s,\theta_1(s))
-A(s,\theta_2(s)),\ \hat\theta(s) \Big\rangle ds\ \bigg|\ \mathcal
F_t \bigg]\\
\leq\ & -\mu \mathbb E \bigg[ \int_t^{T_i} |\hat \theta(s)|^2 ds\
\bigg|\ \mathcal F_t \bigg].
\end{aligned}
\]
Then, letting $i\rightarrow\infty$, we have
\[
\langle \hat x(t),\ \hat y(t) \rangle \geq
\lim_{i\rightarrow\infty}\mu \mathbb E \bigg[ \int_t^{T_i} |\hat
\theta(s)|^2 ds\ \bigg|\ \mathcal F_t \bigg] \geq 0.
\]
(ii) For the given $\mathbb F$-stopping time $\tau$, it is easy to
see, for any $T\in [0,\infty)$,
\[
\langle \hat x(T),\ \hat y(T) \rangle -\langle \hat x(\tau\wedge
T),\ \hat y(\tau\wedge T) \rangle \geq 0.
\]
On the other hand, from It\^{o}'s formula,
\[
\begin{aligned}
\mathbb E\Big[\langle \hat x(T),\ \hat y(T) \rangle -\langle \hat
x(\tau\wedge T),\ \hat y(\tau\wedge T) \rangle\Big] =\ & \mathbb
E\int_{\tau\wedge T}^T \Big\langle
A(s,\theta_1(s))-A(s,\theta_2(s)),\ \hat \theta(s) \Big\rangle ds\\
\leq\ & -\mu \mathbb E\int_{\tau\wedge T}^T |\hat\theta(s)|^2 ds
\leq 0.
\end{aligned}
\]
Then, by the above two inequalities, we obtain
\[
\hat\theta(s)\mathbbm{1}_{[\tau\wedge T,T]}(t) =0.
\]
Due to the arbitrariness of $T$, we get the desired conclusion
$\hat\theta(s)\mathbbm{1}_{[\tau,\infty)}(t) =0$.
\end{proof}
\begin{theorem}\label{Sec3.1_THM_Comparison}
Let $n=1$. Let Assumptions (H3.1)-(H3.4) hold for
$(b,\sigma,\gamma,g,\Phi_1)$ and $(b,\sigma,\gamma,g,\Phi_2)$.
\begin{enumerate}[(i)]
\item If $\Phi_1(y_2(0))>\Phi_2(y_2(0))$, then $\hat y(0)>0$.
\item If $\Phi_1(y_2(0))=\Phi_2(y_2(0))$, then $\hat y(0)=0$.
\end{enumerate}
\end{theorem}
\begin{proof}
When $n=1$, Lemma \ref{Sec3.1_CT_Lemma}-(i) is read as $\hat
x(t)\hat y(t)\geq 0$ for any $t\in [0,\infty)$. Especially, taking
$t=0$, we have $\hat x(0)\hat y(0) \geq 0$. Then from the
monotonicity condition on $\Phi_1$,
\begin{equation}\label{Sec3.1_CT_Lemma_Temp}
\begin{aligned}
0 \leq\ & \Big( \Phi_1(y_1(0))-\Phi_2(y_2(0)) \Big)\hat y(0)\\
=\ & \Big[ \Big( \Phi_1(y_1(0))-\Phi_1(y_2(0)) \Big) +\Big(
\Phi_1(y_2(0))-\Phi_2(y_2(0)) \Big) \Big] \hat y(0)\\
\leq\ & \Big( \Phi_1(y_2(0))-\Phi_2(y_2(0)) \Big) \hat y(0).
\end{aligned}
\end{equation}
(i) If $\Phi_1(y_2(0))>\Phi_2(y_2(0))$, then
\eqref{Sec3.1_CT_Lemma_Temp} implies $\hat y(0)\geq 0$. Moreover, if
$\hat y(0) =0$, then the stopping time $\tau$ defined in Lemma
\ref{Sec3.1_CT_Lemma}-(ii) is equal to $0$. By Lemma
\ref{Sec3.1_CT_Lemma}-(ii), we have $\hat x(0) =0$. Since $y_1(0) =
y_2(0)$, then
\[
\Phi_1(y_2(0)) -\Phi_2(y_2(0)) = \Phi_1(y_1(0)) -\Phi_2(y_2(0)) =
\hat x(0) =0.
\]
This is a contradiction. Therefore, in this case we must have $\hat
y(0) > 0$.
(ii) When $\Phi_1(y_2(0)) = \Phi_2(y_2(0))$, we also use a framework
of reduction to absurdity to show $\hat y(0)=0$. We assume that
$\hat y(0)\neq 0$. From \eqref{Sec3.1_CT_Lemma_Temp}, we deduce that
\[
\hat x(0) = \Phi_1(y_1(0)) -\Phi_2(y_2(0))= 0.
\]
By Lemma \ref{Sec3.1_CT_Lemma}-(ii) once again, we have $\hat y(0)
=0$. We obtain a contradiction, and then finish the proof.
\end{proof}
\section{Application to backward stochastic LQ problems}\label{Sec4}
In this section, we apply the solvability result of forward-backward
SDEs studied in the above section to deal with two kinds of backward
stochastic linear-quadratic (LQ) problems with jumps, including an
LQ stochastic optimal control (SOC) problem and an LQ nonzero-sum
stochastic differential game (NZSSDG) problem.
Firstly, for the control problem, the system is given by the
following controlled linear backward SDE on the infinite interval
$[0,\infty)$:
\begin{equation}\label{Sec4.1_Sys}
\begin{aligned}
-dy(t) =\ & \left[ A(t)y(t) +\sum_{i=1}^d B_i(t)z_i(t) +\sum_{i=1}^l
\int_{\mathcal E} C_i(t,e) r_i(t,e) \pi_i(de) +D(t)v(t) +\alpha(t)
\right]dt\\
& -\sum_{i=1}^d z_i(t) dW_i(t) -\sum_{i=1}^l \int_{\mathcal E}
r_i(t,e) \tilde N_i(dt,de),
\end{aligned}
\end{equation}
where $A$, $B_i$ ($i=1,2,\dots,d$), $D$ are $\mathbb
F$-progressively measurable, matrix-valued, bounded processes with
appropriate dimensions; $C_i$ ($i=1,2,\dots, l$) is a $\mathcal
P\otimes \mathcal B(\mathcal E)$-measurable, ($n\times n$)
matrix-valued process such that $\int_{\mathcal E} |C_i(t,e)|^2
\pi_i(de)$ is uniformly bounded for any $(\omega,t)\in \Omega\times
[0,\infty)$; and the nonhomogeneous term $\alpha\in L^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^n)$ where $K>0$ is a constant. The admissible
control set is defined by
\[
\begin{aligned}
\mathcal V := \Big\{ & v \in L^{2,K}_{\mathbb F}(0,\infty;\mathbb
R^k)\ \Big|\ \mbox{with respect to $v$, \eqref{Sec4.1_Sys} admits a unique solution}\\
& (y^v,z^v,r^v)\in L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^n) \times
L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times d}) \times
M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times l}) \Big\},
\end{aligned}
\]
in which each element $v$ is called an admissible control, and
$(y^v,z^v,r^v)$ is called the state trajectory corresponding to $v$.
In what follows, we will show the admissible control set $\mathcal
V$ is nonempty under some suitable conditions. In addition, we are
given a cost functional associated with $v$ in a quadratic form:
\begin{equation}\label{Sec4.1_Cost}
\begin{aligned}
J(v(\cdot)) =\ & \frac 1 2 \langle Qy(0),\ y(0) \rangle + \frac 1 2
\mathbb E\int_0^\infty \bigg[ \langle L(t)y(t),\ y(t) \rangle
+\sum_{i=1}^d \langle M_i(t)z_i(t),\ z_i(t)
\rangle\\
& +\sum_{i=1}^l \int_{\mathcal E} \langle S_i(t,e)r_i(t,e),\
r_i(t,e) \rangle \pi_i(de) +\langle R(t)v(t),\ v(t) \rangle
\bigg]dt,
\end{aligned}
\end{equation}
where $Q$ is an ($n\times n$) symmetric and positive semi-definite
matrix; $L$, $M_i$ ($i=1,2,\dots,d$) are $\mathbb F$-progressively
measurable, ($n\times n$) symmetric and positive semi-definite
matrix-valued, bounded processes; $S_i$ ($i=1,2,\dots,l$) is a
$\mathcal P\otimes\mathcal B(\mathcal E)$-measurable, ($n\times n$)
symmetric and positive semi-definite matrix-valued, bounded process;
$R$ is an $\mathbb F$-progressively measurable, ($k\times k$)
symmetric and positive definite matrix-valued, bounded process.
Moreover, $R^{-1}$ is also bounded.\\
\noindent{\bf Problem (SOC).} The problem is to find an admissible
control $u\in\mathcal V$ such that
\begin{equation}
J(u(\cdot)) = \inf_{v(\cdot)\in\mathcal V} J(v(\cdot)).
\end{equation}
Such an admissible control $u$ is called an optimal control, and
$(y,z,r) := (y^u,z^u,r^u)$ is called the corresponding optimal state
trajectory.\\
The following result links Problem (SOC) to a forward-backward SDE.
\begin{lemma}\label{Sec4.1_Lemma}
If the following forward-backward SDE:
\begin{equation}\label{Sec4.1_FBSDE_SOC}
\left\{
\begin{aligned}
dx(t) =\ & \Big[ A^\top(t)x(t) -L(t)y(t) \Big] dt +\sum_{i=1}^d
\Big[ B_i^\top(t)x(t) -M_i(t)z_i(t) \Big]dW_i(t)\\
& +\sum_{i=1}^l \int_{\mathcal E} \Big[ C^\top_i(t,e)x(t-) -S_i(t,e)
r_i(t,e) \Big] \tilde N_i(dt,de),\\
-dy(t) =\ & \bigg[ D(t)R^{-1}(t)D^\top(t)x(t) +A(t)y(t)
+\sum_{i=1}^d B_i(t)z_i(t)\\
& +\sum_{i=1}^l\int_{\mathcal E} C_i(t,e)r_i(t,e) \pi_i(de)
+\alpha(t) \bigg] dt -\sum_{i=1}^d z_i(t) dW_i(t)\\
& -\sum_{i=1}^l\int_{\mathcal E} r_i(t,e) \tilde N_i(dt,de),\\
x(0) =\ & -Q y(0)
\end{aligned}
\right.
\end{equation}
admits a solution $(x,y,z,r)\in \mathcal L^{2,K}_{\mathbb
F}(0,\infty)$, then
\begin{equation}\label{Sec4.1_OptimalControl}
u(t) = R^{-1}(t) D^\top(t) x(t),\quad t\in [0,\infty),
\end{equation}
provides an optimal control of Problem (SOC). Moreover the optimal
control is unique.
\end{lemma}
\begin{proof}
First, we prove that $u$ defined by \eqref{Sec4.1_OptimalControl} is
an optimal control for Problem (SOC). For each $v\in\mathcal V$, the
corresponding state trajectory is denoted by $(y^v,z^v,r^v)$. Let us
consider the difference of $J(v(\cdot))$ and $J(u(\cdot))$ (the
argument $(t,e)$ is suppressed):
\begin{equation}\label{Sec4.1_Temp}
\begin{aligned}
J(v(\cdot))-J(u(\cdot)) =\ & \frac 1 2 \Big[ \langle Qy^v(0),\
y^v(0) \rangle -\langle Qy(0),\ y(0) \rangle \Big]\\
& +\frac 1 2 \mathbb E\int_0^\infty \bigg\{ \Big[ \langle Ly^v,\ y^v
\rangle -\langle Ly,\ y \rangle \Big] +\sum_{i=1}^d \Big[ \langle
M_i z^v_i,\ z^v_i \rangle -\langle M_iz_i,\ z_i \rangle \Big]\\
& +\sum_{i=1}^l \int_{\mathcal E} \Big[ \langle S_ir_i^v,\
r_i^v\rangle -\langle S_ir_i,\ r_i \rangle \Big] \pi_i(de) +\Big[
\langle Rv,\ v \rangle -\langle Ru,\ u \rangle \Big] \bigg\} dt\\
=\ & \frac 1 2 \langle Q(y^v(0)-y(0)),\ y^v(0)-y(0) \rangle\\
& +\frac 1 2 \mathbb E\int_0^\infty \bigg\{ \langle L(y^v-y),\ y^v-y
\rangle +\sum_{i=1}^d \langle M_i(z^v_i-z_i),\ z^v_i-z_i \rangle\\
& +\sum_{i=1}^l \int_{\mathcal E} \langle S_i(r_i^v-r_i),\ r_i^v-r_i
\rangle \pi_i(de) +\langle R(v-u),\ v-u \rangle \bigg\} dt +\Lambda,
\end{aligned}
\end{equation}
where
\[
\begin{aligned}
\Lambda =\ & \langle Qy(0),\ y^v(0)-y(0) \rangle +\mathbb
E\int_0^\infty \bigg\{ \langle Ly,\ y^v-y \rangle +\sum_{i=1}^d
\langle M_iz_i,\ z_i^v-z_i \rangle\\
& +\sum_{i=1}^l \int_{\mathcal E} \langle S_ir_i,\ r_i^v-r_i \rangle
\pi_i(de) +\langle Ru,\ v-u \rangle \bigg\} dt.
\end{aligned}
\]
Applying It\^{o}'s formula to $\langle x(t),\ y^v(t)-y(t) \rangle$
on the interval $[0,T]$, by the initial condition of $x$ (see
\eqref{Sec4.1_FBSDE_SOC}) and the definition of $u$ (see
\eqref{Sec4.1_OptimalControl}), we have
\[
\begin{aligned}
& \mathbb E\Big[\langle x(T),\ y^v(T)-y(T) \rangle\Big] +\langle
Qy(0),\ y^v(0)-y(0) \rangle\\
=\ & -\mathbb E\int_0^T \bigg\{ \langle Ly,\ y^v-y \rangle
+\sum_{i=1}^d \langle M_iz_i,\ z_i^v-z_i \rangle +\sum_{i=1}^l
\int_{\mathcal E}\langle S_ir_i,\ r_i^v-r_i \rangle\pi_i(de)
+\langle Ru,\ v-u \rangle \bigg\} dt.
\end{aligned}
\]
Letting $T\rightarrow\infty$, we get $\Lambda =0$. Then, since $Q$,
$L$, $M_i$ ($i=1,2,\dots,d$), $S_i$ ($i=1,2,\dots,l$) are positive
semi-definite, and $R$ is positive definite, we have
$J(v(\cdot))-J(u(\cdot))\geq 0$. Due to the arbitrariness of $v$, we
prove that $u$ defined by \eqref{Sec4.1_OptimalControl} is an
optimal control.
For the uniqueness, besides $u$ given by
\eqref{Sec4.1_OptimalControl}, let $\bar u\in\mathcal V$ be another
optimal control, and denote by $(y^{\bar u},z^{\bar u},r^{\bar u})$
the corresponding optimal state trajectory. Obviously $J(\bar
u(\cdot)) = J(u(\cdot))$. Coming back to \eqref{Sec4.1_Temp}, we
have
\[
\begin{aligned}
0 =\ & \frac 1 2 \langle Q(y^{\bar u}(0)-y(0)),\ y^{\bar u}(0)-y(0) \rangle\\
& +\frac 1 2 \mathbb E\int_0^\infty \bigg\{ \langle L(y^{\bar
u}-y),\ y^{\bar u}-y
\rangle +\sum_{i=1}^d \langle M_i(z^{\bar u}_i-z_i),\ z^{\bar u}_i-z_i \rangle\\
& +\sum_{i=1}^l \int_{\mathcal E} \langle S_i(r_i^{\bar u}-r_i),\
r_i^{\bar u}-r_i \rangle \pi_i(de) +\langle R(\bar u-u),\ \bar u-u
\rangle \bigg\} dt \\
\geq\ & \frac 1 2 \mathbb E\int_0^\infty \langle R(\bar u-u),\ \bar
u-u \rangle dt.
\end{aligned}
\]
Because $R$ is positive definite, we get $\bar u(\cdot) = u(\cdot)$.
We have proved the uniqueness of the optimal control.
\end{proof}
In order to obtain the solvability of \eqref{Sec4.1_FBSDE_SOC}, we
assume the following assumptions:
\begin{enumerate}[(\mbox{A1}.1)]
\item There exists a constant $\mu>0$ such that for any $(\omega,t,e)\in \Omega\times
[0,\infty)\times\mathcal E$, any $i=1,2,\dots,d$, any
$j=1,2,\dots,l$,
\[
D(t)R^{-1}(t)D^\top(t) \geq \mu I,\quad L(t) \geq \mu I,\quad
M_i(t)\geq \mu I, \quad S_j(t,e)\geq \mu I.
\]
\item $2\mu-K\geq 0$.
\end{enumerate}
\noindent Here, $I$ denotes the ($n\times n$) identity matrix and
the expression $A\geq B$ means $A-B$ is positive semi-definite as
usual. We notice that, in the viewpoint of Remark
\ref{Sec3_Remark_Artificial}, Assumption (A1.2) is artificial. If it
does not hold true, then we can consider Problem (SOC) in some
larger space. However, for the convenience of presentation, we keep
it here.
\begin{theorem}\label{Sec4.1_THM}
Under Assumptions (A1.1)-(A1.2), the forward-backward SDE
\eqref{Sec4.1_FBSDE_SOC} admits a unique solution $(x,y,z,r) \in
\mathcal L^{2,K}_{\mathbb F}(0,\infty)$. Moreover, $u$ defined by
\eqref{Sec4.1_OptimalControl} is the unique optimal control for
Problem (SOC).
\end{theorem}
\begin{proof}
It is easy to check that Assumptions (A1.1)-(A1.2) imply Assumptions
(H3.1)-(H3.4). Then by Theorem \ref{Sec3_THM}, the forward-backward
SDE \eqref{Sec4.1_FBSDE_SOC} is uniquely solvable. Moreover, thanks
to Lemma \ref{Sec4.1_Lemma}, we can finish the proof.
\end{proof}
Next we extend the LQ SOC problem to an LQ nonzero-sum stochastic
differential game (NZSSDG) problem. Without loss of generality, we
only consider the case of two players in this paper. The case of
$n(\geq 3)$ players can be treated in the same way. In detail, the
game system is described by the following controlled linear backward
SDE on $[0,\infty)$:
\begin{equation}\label{Sec4.2_Sys}
\begin{aligned}
-dy(t) =\ & \bigg[ A(t)y(t) +\sum_{i=1}^d B_i(t)z_i(t) +\sum_{i=1}^l
\int_{\mathcal E} C_i(t,e) r_i(t,e) \pi_i(de) +D_1(t)v_1(t)\\
& +D_2(t)v_2(t) +\alpha(t) \bigg]dt -\sum_{i=1}^d z_i(t) dW_i(t)
-\sum_{i=1}^l \int_{\mathcal E} r_i(t,e) \tilde N_i(dt,de),
\end{aligned}
\end{equation}
where $A$, $B_i$ ($i=1,2,\dots,d$), $D_1$, $D_2$ are $\mathbb
F$-progressively measurable, matrix-valued, bounded processes with
appropriate dimensions; $C_i$ ($i=1,2,\dots, l$) is a $\mathcal
P\otimes \mathcal B(\mathcal E)$-measurable, ($n\times n$)
matrix-valued process such that $\int_{\mathcal E} |C_i(t,e)|^2
\pi_i(de)$ is uniformly bounded for any $(\omega,t)\in \Omega\times
[0,\infty)$; and the nonhomogeneous term $\alpha\in L^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^n)$ where $K>0$ is a constant. $v_1$ and $v_2$
are the control processes of Player 1 and Player 2, respectively. We
introduce the admissible control set for the two players:
\[
\begin{aligned}
\mathcal V := \Big\{ & (v_1,v_2) \in L^{2,K}_{\mathbb
F}(0,\infty;\mathbb
R^{k_1}) \times L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{k_2})\ \Big|\ \mbox{with respect to $(v_1,v_2)$, \eqref{Sec4.2_Sys} admits}\\
& \mbox{a unique solution } (y^{v_1,v_2},z^{v_1,v_2},r^{v_1,v_2})\in
L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^n) \times L^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^{n\times d})\\
& \times M^{2,K}_{\mathbb F}(0,\infty;\mathbb R^{n\times l})
\Big\},
\end{aligned}
\]
in which each element $(v_1,v_2)$ is called an admissible control
pair, and $(y^{v_1,v_2},z^{v_1,v_2},r^{v_1,v_2})$ is called the
state trajectory corresponding to $(v_1,v_2)$. As same as Problem
(SOC), we will show the admissible control set $\mathcal V$ is
nonempty under suitable conditions. For any $(v_1,v_2)\in \mathcal
V$, let
\[
\begin{aligned}
\mathcal V_1(v_2) =\ & \{ \bar v_1 \ |\ (\bar v_1, v_2) \in \mathcal
V \},\\
\mathcal V_2(v_1) =\ & \{ \bar v_2 \ |\ (v_1, \bar v_2) \in \mathcal
V \}.
\end{aligned}
\]
Additionally, the cost functionals of the two players are given as
follows: for $i=1,2$,
\begin{equation}\label{Sec4.2_Cost}
\begin{aligned}
J_i(v_1(\cdot),v_2(\cdot)) =\ & \frac 1 2 \langle Q_iy(0),\ y(0)
\rangle + \frac 1 2 \mathbb E\int_0^\infty \bigg[ \langle
L_i(t)y(t),\ y(t) \rangle +\sum_{j=1}^d \langle M_{ij}(t)z_j(t),\
z_j(t) \rangle\\
& +\sum_{j=1}^l \int_{\mathcal E} \langle S_{ij}(t,e)r_j(t,e),\
r_j(t,e) \rangle \pi_j(de) +\langle R_i(t)v_i(t),\ v_i(t) \rangle
\bigg]dt,
\end{aligned}
\end{equation}
where $Q_i$ is an ($n\times n$) symmetric and positive semi-definite
matrix; $L_i$, $M_{ij}$ ($j=1,2,\dots,d$) are $\mathbb
F$-progressively measurable, ($n\times n$) symmetric and positive
semi-definite matrix-valued, bounded processes; $S_{ij}$
($j=1,2,\dots,l$) is a $\mathcal P\otimes\mathcal B(\mathcal
E)$-measurable, ($n\times n$) symmetric and positive semi-definite
matrix-valued, bounded process; $R_i$ is an $\mathbb
F$-progressively measurable, ($k\times k$) symmetric and positive
definite matrix-valued, bounded process. Moreover, $R_i^{-1}$ is
also bounded.
Suppose each player hopes to minimize his/her cost functional
$J_i(v_1(\cdot),v_2(\cdot))$ by selecting an appropriate admissilbe
control $v_i$ ($i=1,2$), then the game problem is formulated
as follows.\\
\noindent{\bf Problem (NZSSDG).} The problem is to find a pair of
admissible controls $(u_1,u_2)\in\mathcal V$ such that
\begin{equation}\label{Sec4.2_Nash}
\begin{aligned}
& J_1(u_1(\cdot),u_2(\cdot)) = \inf_{v_1(\cdot)\in\mathcal V_1(u_2)}
J_1(v_1(\cdot),u_2(\cdot)),\\
& J_2(u_1(\cdot),u_2(\cdot)) = \inf_{v_2(\cdot)\in\mathcal V_2(u_1)}
J_2(u_1(\cdot),v_2(\cdot)).
\end{aligned}
\end{equation}
Such a pair of admissible controls $(u_1,u_2)$ is called a Nash
equilibrium point. For the sake of notations, we denote the state
trajectory corresponding to $(u_1,u_2)$ by $(y,z,r) :=
(y^{u_1,u_2},z^{u_1,u_2},r^{u_1,u_2})$.
\begin{lemma}\label{Sec4.2_Lemma}
If the following forward-backward SDE:
\begin{equation}\label{Sec4.2_FBSDE_NZSSDG}
\left\{
\begin{aligned}
dx_1(t) =\ & \Big[ A^\top(t)x_1(t) -L_1(t)y(t) \Big]dt +\sum_{j=1}^d
\Big[ B_j^\top(t)x_1(t) -M_{1j}(t)z_j(t) \Big] dW_j(t)\\
& +\sum_{j=1}^l \int_{\mathcal E} \Big[ C_j^\top(t,e)x_1(t-)
-S_{1j}(t,e) r_j(t,e) \Big] \tilde N_j(dt,de),\\
dx_2(t) =\ & \Big[ A^\top(t)x_2(t) -L_2(t)y(t) \Big]dt +\sum_{j=1}^d
\Big[ B_j^\top(t)x_2(t) -M_{2j}(t)z_j(t) \Big] dW_j(t)\\
& +\sum_{j=1}^l \int_{\mathcal E} \Big[ C_j^\top(t,e)x_2(t-)
-S_{2j}(t,e) r_j(t,e) \Big] \tilde N_j(dt,de),\\
-dy(t) =\ & \bigg[ D_1(t)R_1^{-1}(t)D_1^\top(t)x_1(t)
+D_2(t)R_2^{-1}(t)D_2^\top(t)x_2(t) +A(t)y(t)\\
& +\sum_{j=1}^d B_j(t)z_j(t) +\sum_{j=1}^l \int_{\mathcal E}
C_j(t,e)r_j(t,e)\pi_j(de) +\alpha(t) \bigg] dt\\
& -\sum_{j=1}^d z_j(t)dW_j(t) -\sum_{j=1}^l \int_{\mathcal E}
r_j(t,e) \tilde N_j(dt,de),\\
x_1(0) =\ & -Q_1 y(0), \quad x_2(0) = -Q_2 y(0)
\end{aligned}
\right.
\end{equation}
admits a solution $(x_1,x_2,y,z,r)\in L^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^n)\times \mathcal L^{2,K}_{\mathbb
F}(0,\infty)$, then
\begin{equation}\label{Sec4.2_NashPoint}
\left(\begin{array}{ccc} u_1(t) \\ u_2(t) \end{array}\right) =
\left(\begin{array}{ccc} R_1^{-1}(t)D_1^\top(t) x_1(t)\\
R_2^{-1}(t)D_2^\top(t) x_2(t)
\end{array}\right),\quad t\in [0,\infty),
\end{equation}
provides a Nash equilibrium point for Problem (NZSSDG).
\end{lemma}
\begin{proof}
We shall link Problem (NZSSDG) with two LQ SOC problems. Precisely,
for $i=1,2$, we fix $u_{3-i}(\cdot)$ which is defined in
\eqref{Sec4.2_NashPoint}. To minimize (the argument $(t,e)$ is
suppressed)
\begin{equation}\label{Sec4.2_Cost_i}
\begin{aligned}
J_i(v_i(\cdot),u_{3-i}(\cdot)) =\ & \frac{1}{2} \langle Q_i
y^{v_i}(0),\ y^{v_i}(0) \rangle +\frac 1 2 \mathbb E\int_0^\infty
\bigg[ \langle L_iy^{v_i},\
y^{v_i} \rangle +\sum_{j=1}^d \langle M_{ij}z_j^{v_i},\ z_j^{v_i} \rangle\\
& +\sum_{j=1}^l\int_{\mathcal E} \langle S_{ij} r_j^{v_i},\
r_j^{v_i} \rangle \pi_j(de) +\langle R_iv_i,\ v_i \rangle \bigg]dt
\end{aligned}
\end{equation}
subject to
\begin{equation}\label{Sec4.2_Sys_i}
\begin{aligned}
-dy^{v_i} =\ & \bigg[ Ay^{v_i} +\sum_{j=1}^d B_jz_j^{v_i}
+\sum_{j=1}^l\int_{\mathcal E} C_jr_j^{v_i} \pi_j(de) +D_iv_i +\Big(
D_{3-i}u_{3-i} +\alpha \Big)
\bigg]dt\\
& -\sum_{j=1}^d z_j^{v_i} dW_j -\sum_{j=1}^l \int_{\mathcal E}
r_j^{v_i} \tilde N_j(dt,de)
\end{aligned}
\end{equation}
over $\mathcal V_i(u_{3-i})$ is an LQ SOC problem. Since
\eqref{Sec4.2_FBSDE_NZSSDG} admits a solution $(x_1,x_2,y,z,r)$,
then $(x_i,y,z,r)$ solves the following forward-backward SDE:
\begin{equation}\label{Sec4.2_FBSDE_SOC_i}
\left\{
\begin{aligned}
dx_i =\ & \Big[ A^\top x_i - L_iy \Big] dt +\sum_{j=1}^d \Big[
B_j^\top x_i -M_{ij}z_j \Big]dW_j\\
& +\sum_{j=1}^l \int_{\mathcal E}
\Big[ C_j^\top x_i -S_{ij}r_j \Big] \tilde N_j(dt,de),\\
-dy =\ & \bigg[ D_iR_i^{-1}D_i^\top x_i +Ay +\sum_{j=1}^d B_jz_j
+\sum_{j=1}^l\int_{\mathcal E} c_jr_j\pi_j(de) +\Big( D_{3-i}u_{3-i}
+\alpha \Big) \bigg]dt\\
& -\sum_{j=1}^d z_j dW_j -\sum_{j=1}^l\int_{\mathcal E} r_j\tilde
N_j(dt,de),\\
x_i(0) =\ & -Q_i y(0).
\end{aligned}
\right.
\end{equation}
By Lemma \ref{Sec4.1_Lemma}, the LQ SOC problem
\eqref{Sec4.2_Cost_i}-\eqref{Sec4.2_Sys_i} admits a unique optimal
control with the form
\[
u_i(t) = R_i^{-1}(t)D_i^\top(t)x_i(t),\quad t\in [0,\infty),
\]
which is coincided with \eqref{Sec4.2_NashPoint}. In other words,
the following equation holds:
\[
J_i(u_i(\cdot),u_{3-i}(\cdot)) = \inf_{v_i(\cdot)\in\mathcal
V_i(u_{3-i})} J_i(v_i(\cdot),u_{3-i}(\cdot)).
\]
Since $i=1,2$, from the definition of the Nash equilibrium point
(see \eqref{Sec4.2_Nash}), $(u_1,u_2)$ defined by
\eqref{Sec4.2_NashPoint} provides a Nash equilibrium point for
Problem (NZSSDG).
\end{proof}
The forward-backward SDE \eqref{Sec4.2_FBSDE_NZSSDG} is more
complicated. In order to obtain the solvability of
\eqref{Sec4.2_FBSDE_NZSSDG}, we would like to employ a linear
transform which is originally introduced by Hamad\`{e}ne
\cite{Hamadene1998} (see also Yu \cite{Yu2012}). For this transform,
we need to introduce the following assumptions:
\begin{enumerate}[(\mbox{A2}.1)]
\item The matrix-valued processes $D_iR_i^{-1}D_i^\top$, $i=1,2$,
are independent of $t$.
\item The following commutation relations among matrices hold true:
\[
D_i(t)R_i^{-1}(t)D_i^\top(t) H(t) =H(t)
D_i(t)R_i^{-1}(t)D_i^\top(t),\quad t\in [0,\infty),\quad i=1,2,
\]
where $H(t) = A^\top(t)$, $B^\top_j(t)$ ($j=1,2,\dots,d$),
$C^\top_j(t)$ ($j=1,2,\dots,l$).
\item There exists a constant $\delta>0$ such that, for any
$(\omega,t)\in \Omega\times [0,\infty)$,
\[
2A(t) +\sum_{j=1}^d B_j(t)B_j^\top(t) +\sum_{j=1}^l\int_{\mathcal E}
C_j(t,e)C_j^\top(t,e) \pi_j(de) +KI \leq -\delta I.
\]
\end{enumerate}
We notice that, Assumption (A2.3) is not necessary when the
corresponding finite horizon game problems were studied (see for
example \cite{Hamadene1998,Yu2012}). Here we assume it due to the
infinite time horizon.
Now we introduce another forward-backward SDE (the argument $(t,e)$
is suppressed):
\begin{equation}\label{Sec4.2_FBSDE_Hama}
\left\{
\begin{aligned}
d\bar x =\ & \Big[ A^\top \bar x -\left( D_1R_1^{-1}D_1^\top L_1 +
D_2R_2^{-1}D_2^\top L_2 \right)\bar y \Big]dt\\
& +\sum_{j=1}^d \Big[ B_j^\top \bar x -\left( D_1R_1^{-1}D_1^\top
M_{1j} + D_2R_2^{-1}D_2^\top M_{2j} \right)\bar z_j \Big] dW_j\\
& +\sum_{j=1}^l \int_{\mathcal E} \Big[ C_j^\top \bar x -\left(
D_1R_1^{-1}D_1^\top S_{1j} + D_2R_2^{-1}D_2^\top S_{2j} \right)\bar
r_j \Big] \tilde N_j(dt,de),\\
-d\bar y =\ & \bigg[ \bar x +A\bar y +\sum_{j=1}^d B_j\bar z_j
+\sum_{j=1}^l \int_{\mathcal E} C_j\bar r_j\pi_j(de) +\alpha \bigg]
dt\\
& -\sum_{j=1}^d \bar z_jdW_j -\sum_{j=1}^l \bar r_j \tilde
N_j(dt,de),\\
\bar x(0) =\ & -\left( D_1(0)R_1^{-1}(0)D_1^\top(0)Q_1 +
D_2(0)R_2^{-1}(0)D_2^\top(0)Q_2 \right) \bar y(0),
\end{aligned}
\right.
\end{equation}
and give the following result.
\begin{lemma}\label{Sec4.2_Lemma_Hama}
Under Assumptions (A2.1)-(A2.3), the existence and uniqueness of
\eqref{Sec4.2_FBSDE_NZSSDG} are equivalent to that of
\eqref{Sec4.2_FBSDE_Hama}.
\end{lemma}
\begin{proof}
On the one hand, by Assumptions (A2.1) and (A2.2), if
$(x_1,x_2,y,z,r)\in L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^n)\times
\mathcal L^{2,K}_{\mathbb F}(0,\infty)$ is a solution of
\eqref{Sec4.2_FBSDE_NZSSDG}, then
\[
\left\{
\begin{aligned}
\bar x(t) =\ & D_1(t)R_1^{-1}(t)D_1^\top(t)x_1(t) +
D_2(t)R_2^{-1}(t)D_2^\top(t)x_2(t),\\
\bar y(t) =\ & y(t),\quad \bar z_j(t) =z_j(t)\ (j=1,2,\dots,d),\quad
\bar r_j(t,e) =r_j(t,e)\ (j=1,2,\dots,l),
\end{aligned}
\right.\quad t\in [0,\infty)
\]
belonging to $\mathcal L^{2,K}_{\mathbb F}(0,\infty)$ solves
\eqref{Sec4.2_FBSDE_Hama}.
On the other hand, if $(\bar x, \bar y, \bar z, \bar r)\in\mathcal
L^{2,K}_{\mathcal F}(0,\infty)$ is a solution of
\eqref{Sec4.2_FBSDE_Hama}, we let $y=\bar y$, $z=\bar z$, $r=\bar r$
and $(x_1, x_2)$ be the unique solution of the following SDE:
\[
\left\{
\begin{aligned}
dx_1 =\ & \Big[ A^\top x_1 -L_1y \Big]dt +\sum_{j=1}^d \Big[
B_j^\top x_1 -M_{1j}z_j \Big] dW_j +\sum_{j=1}^l \int_{\mathcal E}
\Big[ C_j^\top x_1 -S_{1j}r_j \Big] \tilde N_j(dt,de),\\
dx_2 =\ & \Big[ A^\top x_2 -L_2y \Big]dt +\sum_{j=1}^d \Big[
B_j^\top x_2 -M_{2j}z_j \Big] dW_j +\sum_{j=1}^l \int_{\mathcal E}
\Big[ C_j^\top x_2 -S_{2j}r_j \Big] \tilde N_j(dt,de),\\
x_1(0) =\ & -Q_1y(0),\quad x_1(0) = -Q_2 y(0).
\end{aligned}
\right.
\]
Obviously $(x_1,x_2)\in S^{2,loc}_{\mathbb F}(0,\infty;\mathbb
R^n)\times S^{2,loc}_{\mathbb F}(0,\infty;\mathbb R^n)$. Moreover,
Assumption (A2.3) ensures it also belongs to $L^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^n) \times L^{2,K}_{\mathbb F}(0,\infty;\mathbb
R^n)$. In fact, for $i=1,2$, we apply It\^{o}'s formula to
$|x_i(t)|^2 e^{Kt}$ on the interval $[0,T]$ to have
\[
\begin{aligned}
& \mathbb E\Big[ |x_i(T)|^2 e^{KT} \Big] -|x_i(0)|^2\\
=\ & \mathbb E\int_0^T \bigg[ \Big\langle \Big( 2A +\sum_{j=1}^d
B_jB_j^\top +\sum_{j=1}^l\int_{\mathcal E} C_jC_j^\top\pi_j(de) +KI
\Big)x_i,\ x_i \Big\rangle\\
& +2\Big\langle \Big( -L_iy +\sum_{j=1}^d B_jM_{ij}z_j
+\sum_{j=1}^l\int_{\mathcal E} C_jS_{ij}r_j \pi_j(de) \Big),\ x_i
\Big\rangle\\
& +\sum_{j=1}^d |M_{ij}z_j|^2 +\sum_{j=1}^l\int_{\mathcal E}
|S_{ij}r_j|^2 \pi_j(de) \bigg] e^{Kt} dt.
\end{aligned}
\]
By Assumption (A2.3) and the inequality: $2\langle y,\ x \rangle
\leq (2/\delta)|y|^2 +(\delta/2)|x|^2$,
\[
\begin{aligned}
& \frac{\delta}{2} \mathbb E\int_0^T |x_i|^2 e^{Kt} dt\\
\leq\ & |x_i(0)|^2 +\mathbb E\int_0^T \bigg[ \frac{2}{\delta}\bigg|
-L_iy +\sum_{j=1}^d B_jM_{ij}z_j +\sum_{j=1}^l\int_{\mathcal E}
C_jS_{ij}r_j \pi_j(de) \bigg|^2\\
& +\sum_{j=1}^d |M_{ij}z_j|^2 +\sum_{j=1}^l\int_{\mathcal E}
|S_{ij}r_j|^2 \pi_j(de) \bigg] e^{Kt} dt.
\end{aligned}
\]
Letting $T\rightarrow\infty$, we get $x_i\in L^{2,K}_{\mathbb
F}(0,\infty;\mathbb R^n)$. In what follows, we shall show that
$(x_1,x_2,y,z,r)\in L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^n)\times
\mathcal L^{2,K}_{\mathbb F}(0,\infty)$ defined above is a solution
of the forward-backward SDE \eqref{Sec4.2_FBSDE_NZSSDG}. Actually,
the remaining thing is to show $(x_1,x_2,y,z,r)$ satisfies the
backward equation in \eqref{Sec4.2_FBSDE_NZSSDG}. Compared with the
backward equation in \eqref{Sec4.2_FBSDE_Hama}, we only need to show
$D_1R_1^{-1} D_1^\top x_1 +D_2R_2^{-1} D_2^\top x_2 = \bar x$. For
the convenience, we let
\[
\tilde x(t) = D_1(t)R_1^{-1}(t) D_1^\top(t) x_1(t)
+D_2(t)R_2^{-1}(t) D_2^\top(t) x_2(t).
\]
By Assumptions (A2.1) and (A2.2), we know $\tilde x$ satisfies
\[
\left\{
\begin{aligned}
d\tilde x =\ & \Big[ A^\top \tilde x -\left( D_1R_1^{-1}D_1^\top L_1
+ D_2R_2^{-1}D_2^\top L_2 \right) y \Big]dt\\
& +\sum_{j=1}^d \Big[ B_j^\top \tilde x -\left( D_1R_1^{-1}D_1^\top
M_{1j} + D_2R_2^{-1}D_2^\top M_{2j} \right) z_j \Big] dW_j\\
& +\sum_{j=1}^l \int_{\mathcal E} \Big[ C_j^\top \tilde x -\left(
D_1R_1^{-1}D_1^\top S_{1j} + D_2R_2^{-1}D_2^\top S_{2j} \right)
r_j \Big] \tilde N_j(dt,de),\\
\tilde x(0) =\ & -\left( D_1(0)R_1^{-1}(0)D_1^\top(0)Q_1 +
D_2(0)R_2^{-1}(0)D_2^\top(0)Q_2 \right) y(0),
\end{aligned}
\right.
\]
which coincides with the forward equation in
\eqref{Sec4.2_FBSDE_Hama}. Regarding $(y,z,r)$ as fixed processes,
from the uniqueness of SDE, we have $\bar x =\tilde x$ and we proved
that $(x_1,x_2,y,z,r)$ solves \eqref{Sec4.2_FBSDE_NZSSDG}.
In a similar way, one can prove that the uniqueness of
\eqref{Sec4.2_FBSDE_NZSSDG} is equivalent to that of
\eqref{Sec4.2_FBSDE_Hama}.
\end{proof}
For the solvability of \eqref{Sec4.2_FBSDE_Hama}, and then
\eqref{Sec4.2_FBSDE_NZSSDG}, we impose the following assumptions:
\begin{enumerate}[(\mbox{A3}.1)]
\item The matrix $D_1(0)R_1^{-1}(0)D_1^\top(0) Q_1 +D_2(0)R_2^{-1}(0)D_2^\top (0)
Q_2$ is positive semi-definite.
\item There exists a constant $\mu>0$ such that
\[
D_1(t)R_1^{-1}(t)D_1^\top(t) H_1(t) + D_2(t)R_2^{-1}(t)D_2^\top(t)
H_2(t) \geq \mu I, \quad t\in [0,\infty),
\]
where $H_i(t) = L_i$, $M_{ij}$ ($j=1,2,\dots,d$), $S_{ij}$
($j=1,2,\dots,l$), $i=1,2$.
\item $2\min\{\mu, 1\}-K \geq 0$.
\end{enumerate}
Once again, in the viewpoint of Remark \ref{Sec3_Remark_Artificial},
Assumption (A3.3) is artificial. We keep it here for the convenience
of presentation.
\begin{remark}
\begin{enumerate}[(i)]
\item For the symmetric matrices $D_i(0)R_i^{-1}(0)D_i^\top(0)$ and
$Q_i$ ($i=1,2$), if they are commutative, then there exists an
orthogonal matrix $P$ such that both
$P^{-1}D_i(0)R_i^{-1}(0)D_i^\top(0)P$ and $P^{-1}Q_iP$ are diagonal
matrices. Moreover, due to the semi-definiteness of
$D_i(0)R_i^{-1}(0)D_i^\top(0)$ and $Q_i$, we have the following
statement: if $D_i(0)R_i^{-1}(0)D_i^\top(0) Q_i = Q_i
D_i(0)R_i^{-1}(0)D_i^\top(0)$ ($i=1,2$), then (A3.1) holds true.
\item Similarly, if there exist two constants $\beta_1\geq 0$ and $\beta_2\geq
0$ satisfying $\beta_1+\beta_2>0$ such that
\[
\begin{aligned}
& H_i(t)\geq \beta_i I,\quad D_i(t)R_i^{-1}(t)D_i^\top(t) \geq
\beta_i I,\\
& D_i(t)R_i^{-1}(t)D_i^\top(t) H_i(t) = H_i(t)
D_i(t)R_i^{-1}(t)D_i^\top(t),
\end{aligned}
\]
where $H_i(t) = L_i$, $M_{ij}$ ($j=1,2,\dots,d$), $S_{ij}$
($j=1,2,\dots,l$), $i=1,2$, then (A3.2) holds true.
\end{enumerate}
\end{remark}
\begin{theorem}\label{Sec4.2_THM}
Let Assumptions (A2.1)-(A2.3) and (A3.1)-(A3.3) hold.
\begin{enumerate}[(i)]
\item The forward-backward SDE
\eqref{Sec4.2_FBSDE_NZSSDG} admits a unique solution
$(x_1,x_2,y,z,r)\in L^{2,K}_{\mathbb F}(0,\infty;\mathbb R^n) \times
\mathcal L^{2,K}_{\mathbb F}(0,\infty)$. Moreover, $(u_1,u_2)$
defined by \eqref{Sec4.2_NashPoint} is a Nash equilibrium point for
Problem (NZSSDG).
\item If we further assume that, for any $(\omega, t, e)\in \Omega
\times [0,\infty)\times\mathcal E$, any $i=1,2$, $j=1,2,\dots,d$,
$k=1,2,\dots,l$
\begin{equation}\label{Sec4.2_Assumption_Last}
D_i(t)R_i^{-1}(t)D_i^\top(t) \geq \mu I,\quad L_i(t)\geq \mu I,
\quad M_{ij}(t)\geq \mu I,\quad S_{ik}(t,e)\geq \mu I,
\end{equation}
then $(u_1,u_2)$ defined by \eqref{Sec4.2_NashPoint} is the unique
Nash equilibrium point for Problem (NZSSDG).
\end{enumerate}
\end{theorem}
\begin{proof}
(i) Under Assumptions (A3.1)-(A3.3), by Theorem \ref{Sec3_THM}, the
forward-backward SDE \eqref{Sec4.2_FBSDE_Hama} admits a unique
solution in $\mathcal L^{2,K}_{\mathbb F}(0,\infty)$. With the help
of Assumptions (A2.1)-(A2.3) and Lemma \ref{Sec4.2_Lemma_Hama}, the
forward-backward SDE \eqref{Sec4.2_FBSDE_NZSSDG} admits also a
unique solution in the space $L^{2,K}_{\mathbb F}(0,\infty;\mathbb
R^n)\times \mathcal L^{2,K}_{\mathbb F}(0,\infty)$. Moreover, by
Lemma \ref{Sec4.2_Lemma}, $(u_1,u_2)$ defined by
\eqref{Sec4.2_NashPoint} provides a Nash equilibrium point for
Problem (NZSSDG).
(ii) Let $(\bar u_1, \bar u_2)$ be another Nash equilibrium point
for Problem (NZSSDG). From the viewpoint of Lemma
\ref{Sec4.2_Lemma}, for any $i=1,2$, fix $\bar u_{3-i}$, then $\bar
u_i$ is an optimal control of the LQ SOC problem
\eqref{Sec4.2_Cost_i}-\eqref{Sec4.2_Sys_i}. Thanks to
\eqref{Sec4.2_Assumption_Last} and Theorem \ref{Sec4.1_THM}, $\bar
u_i$ must have the form:
\[
\bar u_i(t) = R_i^{-1}(t)D_i^\top(t) x_i(t),\quad t\in [0,\infty),
\]
where $(x_i,y,z)$ satisfies \eqref{Sec4.2_FBSDE_SOC_i}. Combining
the two cases: $i=1$ and $i=2$, we get the conclusion:
\[
\left(\begin{array}{ccc} \bar u_1(t) \\ \bar u_2(t)
\end{array}\right) = \left(\begin{array}{ccc} u_1(t) \\ u_2(t)
\end{array}\right) =
\left(\begin{array}{ccc} R_1^{-1}(t)D_1^\top(t) x_1(t)\\
R_2^{-1}(t)D_2^\top(t) x_2(t)
\end{array}\right),\quad t\in [0,\infty),
\]
where $(x_1,x_2,y,z,r)$ is the unique solution of
\eqref{Sec4.2_FBSDE_NZSSDG}. We complete the proof.
\end{proof}
\begin{example}
In this example, we would like to focus ourselves on a special case
to illustrate the results obtained in this section. Let the
dimension of the state process $y$, the dimension of Brownian motion
and the number of Poisson random measures are $1$. Under this `$1$
dimension' setting, all the involving matrix-valued processes are
real-valued indeed. Moreover, let $A$, $B$, $C$, $D$, $Q$, $L$, $M$,
$S$, $R$ in Problem (SOC) and $A$, $B$, $C$, $D_i$, $Q_i$, $L_i$,
$M_i$, $S_i$, $R_i$ ($i=1,2$) in Problem (NZSSDG) are independent of
the time variable $t$ (and then they are independent of $\omega$
also due to the $\mathbb F$-adaptedness of processes). Furthermore,
let the nonhomogeneous term $\alpha$ in \eqref{Sec4.1_Sys} and
\eqref{Sec4.2_Sys} vanish.
(1). Let $D\neq 0$, $Q\geq 0$, $L>0$, $M>0$, $R>0$, and there exists
a constant $\kappa>0$ such that $S(e)\geq \kappa$ for all
$e\in\mathcal E$. Define
\[
\mu = \min\left\{ \frac{D^2}{R}, L, M, \kappa \right\}.
\]
For any $K\in (0,2\mu)$, Assumptions (A1.1)-(A1.2) are satisfied. By
Theorem \ref{Sec4.1_THM}, the forward-backward SDE
\eqref{Sec4.1_FBSDE_SOC} admits a unique solution, and Problem (SOC)
has a unique optimal control which is given by
\eqref{Sec4.1_OptimalControl}.
(ii). For $i=1,2$, let $D_i\neq 0$, $Q_i\geq 0$, $L_i>0$, $M_i>0$,
$R_i>0$, and there exists a constant $\kappa>0$ such that
$S(e)\geq\kappa$ for all $e\in\mathcal E$. Moreover, we assume
\begin{equation}
A< -\frac 1 2 B^2 -\frac 1 2 \int_{\mathcal E} |C(e)|^2 \pi(de).
\end{equation}
Define
\[
\begin{aligned}
& \mu = \min \bigg\{ \frac{D_1^2}{R_1}, \frac{D_2^2}{R_2}, L_1, L_2,
M_1, M_2, \kappa, \frac{D_1^2}{R_1}L_1 +\frac{D_2^2}{R_2}L_2,
\frac{D_1^2}{R_1}M_1 +\frac{D_2^2}{R_2}M_2, \bigg(\frac{D_1^2}{R_1}
+\frac{D_2^2}{R_2}\bigg)\kappa \bigg\},\\
& \rho = \min\bigg\{-\frac 1 2 B^2 -\frac 1 2 \int_{\mathcal E}
|C(e)|^2 \pi(de)-A, \mu, 1 \bigg\}.
\end{aligned}
\]
For any $K\in (0,2\rho)$, it is easy to check Assumptions
(A2.1)-(A2.3), (A3.1)-(A3.3), and \eqref{Sec4.2_Assumption_Last} are
satisfied. By Theorem \ref{Sec4.2_THM}, the forward-backward SDE
\eqref{Sec4.2_FBSDE_NZSSDG} admits a unique solution, and Problem
(NZSSDG) has a unique Nash equilibrium point which is given by
\eqref{Sec4.2_NashPoint}.
\end{example}
\section{Conclusion}
In this paper, we investigate a kind of forward-backward stochastic
differential equations (SDEs) driven by both Brownian motions and
Poisson processes on an infinite horizon. In our setting, besides
the coupling of mappings $b$, $\sigma$, $\gamma$ and $g$, the two
initial values are also coupled. We employ a new technique to treat
the coupling between the initial values. For this kind of
forward-backward SDEs, we establish an existence and uniqueness
theorem by virtue of the method of continuation under some
monotonicity conditions. Some important properties including
stability and comparison of solutions are also addressed. These
results generalize that of Peng and Shi \cite{PengShi2000}.
The theoretical results are applied to solve an infinite horizon
linear-quadratic (LQ) backward stochastic optimal control problem
and an LQ nonzero-sum backward stochastic differential game. Under
suitable conditions, we get the solvability of the Hamiltonian
systems related to the LQ control problem and LQ game problem, which
are linear forward-backward SDEs of the type studied previously.
Then the unique optimal control and the unique Nash equilibrium
point are obtained in closed forms, respectively.
|
\section{Introduction}
\label{sec:intro}
Auto-callable structures are quite popular in the world of \
structured products. On top of the auto-callable structure it is
common to add features related to interest payments. Hence,
combining range accrual instruments and auto-call options not only
leads to interesting conditional dynamics, but gives an
illustrative example of a typical structured product
ref.~\cite{Bouzoubaa}. In addition to the strong path dependence
of the coupons the instrument's final redemption becomes path
dependent too. Intriguingly, within the Black--Scholes world one
can obtain a closed form expression for the payoff of such a
derivative. On the other side one can also rely on a
straightforward Monte Carlo (MC) approach ref.~\cite{Glasserman}.
Often the interest payment features embedded in the instrument
accrue a fixed amount daily, related to some trigger levels of the
underlyings. The standard approach for valuation of such
instruments is a daily MC simulation. The goal of this paper is to
propose an alternative semi-analytic approach (SA), which in some
cases performs significantly better than the brute force day to
day MC evaluation ref.~\cite{Korn}. As we are going to show, the
complexity of the evaluation of the auto-call probabilities grows
linearly with the number of observation times of the instrument
and one may expect that at some point the MC approach would become
more efficient. However, even in these higher dimensional cases
the semi-analytic approach provides a better control of the
sensitivities of the instrument, since contrary to the MC approach
it does not rely on a numerical differentiation. A relevant question
is what are the pros and cons of the above methods - i.e SA vs MC.
We address this question performing a thorough numerical
investigation.
Technically our work is heavily based on ref.~\cite{SKIPPER}, where a valuation formula for multi-asset,
multi-period binaries is provided. In addition to applying theses studies to auto-callable
and range accrual structures, we extend the main result of ref.~\cite{SKIPPER} to the case of
time-dependent parameters: volatilities, interest rates and dividend yields.\footnote{To the best
of our knowledge, a closed formula for time-dependant parameters have not been presented in the literature.}
The paper is structured as follows: In section two we begin with a brief description of the type of derivative
instrument that we are studying.
In section three we develop the quasi-analytic approach, extending
the results of ref. \cite{SKIPPER} to the case of time dependent
deterministic parameters obtaining an expression for the
probability of an early redemption in terms of the multivariate
cumulative normal distribution. Building on this approach we
obtain similar expression for the payoff at maturity, subject to
elaborate conditions. In addition we calculate the payoff of the
coupons as a sum over multivariate barrier options ref.~\cite{Hull},
using the developed SA approach to represent the pay-off of the
latter in terms of multivariate cumulative normal distribution
ref.~\cite{Zhang}.
Finally, in section four we apply our approach to concrete examples. We implement numerically both the SA and MC approaches
and demonstrate the advantage of applying the SA approach to lower dimensional systems, especially when a high precision valuation is required.
\section{The instrument}
\label{sec:inst}
In this paper we analyse a type of instrument which combines the features of
range accrual coupons with auto-call options.
--The instrument is linked to the performance of two correlated assets $S_1$
and $S_2$.
--The instrument has a finite number $M$ of observation times $%
T_1,T_2,\dots,T_M$. If at the observation time $T_k$ both assets $S_i$ are
simultaneously above certain barriers $b_{i,\,k}$ the instrument redeems at
100\%. This is the auto-call condition. To shift the valuation time at zero
we define $\tau = T-t$ and discuss the observation times $%
\tau_1\,,\tau_2\,,\dots,\tau_M$.
--At the observation times the instrument pays coupons proportional to the
number of days, in the period between the previous observation time and the
present,\footnote{%
Or the valuation day for the first observation time.} in which both assets $%
S_i$ were above certain barriers $c_i$.
--If the instrument reaches maturity, it redeems at 100\% if both assets $S_i
$ are above certain percentage $\kappa$ of their spot prices at issue time $%
\bar S_i$. If at least one of the assets is bellow $\kappa\,\bar S_i$ the
instrument pays only a part proportional to the minimum of the ratios $%
S_i/\bar S_i$.
\section{Semi-analytic approach}
\label{sec:anal}
In this section we outline our semi-analytic approach. We begin by providing
a formula for the auto-call probability.
\subsection{Indicator functions and common notations}
\label{sec:indic}
Without loss of generality, it is assumed that the auto-callable structure
has two underlyings. On the set of dates are imposed trigger conditions
related to the auto-call feature. If the auto-call triggers have never been
breached at the observation dates the auto-callable structure matures at its
final maturity date. On the opposite case, if one of the auto-call triggers
have been breached the instrument auto-calls at this particular date and has
its maturity.
Let us denote with $P_{k}$ the probability to auto-call at observation time $%
\tau _{k}$. Note that this implies that at previous observation times the
spot prices of the two assets where never simultaneously above the barriers $%
b_{i}$. We introduce the following notations: $X_{i,\,k}$ labels the spot
value of the assets $S_{i}$ at observation time $\tau _{k}$.
Using the standard notations, if probability space $\left( \Omega ,\digamma ,%
\mathbf{P}\right) $ is given, and $A\in \digamma ,$ than the indicator
function is defined as $\mathbf{E}\left( \mathbf{1}_{A}\right) =P(A).$
Using the above definition, the auto-call probability at time $\tau _{k},$%
for the general case with $n$ underlying indices is then given by the
expectation related to some probability measure $Q$ of the indicator
function:
\begin{equation}\label{ind-1}
P_{k}=\mathbf{E}_{Q}\left( \mathbf{1}_{(X_{1,\,1}<\,b_{1,\,1})\cup
(X_{2,\,1}<\,b_{2,\,1}),\,(X_{1,\,2}<\,b_{1,\,2})\cup
(X_{2,\,2}<\,b_{2,\,2}),\,\dots ,\,(X_{1,\,k}<\,b_{1,\,k})\cap
(X_{2,\,k}<\,b_{2,\,k})}\right) \ .
\end{equation}%
In order to simplify the notation, hereafter we will omit the probability
measure $Q.$ For the case of two underlyings, we can also define also the
probability that the instrument will not auto-call after the first $k$
observation times:
\begin{equation}
\bar{P}_{k}=\mathbf{E}\left( \mathbf{1}_{(X_{1,\,1}<\,b_{1,\,1})\cup
(X_{2,\,1}<\,b_{2,\,1}),\,(X_{1,\,2}<\,b_{1,\,2})\cup
(X_{2,\,2}<\,b_{2,\,2}),\,\dots ,\,(X_{1,\,k}<\,b_{1,\,k})\cup
(X_{2,\,k}<\,b_{2,\,k})}\right) \ .
\end{equation}%
Note that at each observation time we have more than one possibilities
reflected in the $\cup $ operation.
For example the event $(X_{1,\,1}<\,b_{1})\cup (X_{2,\,1}<\,b_{2})$ can be
split into the three scenarios $(X_{1,\,1}<\,b_{1})\cap (X_{2,\,1}<\,b_{2})$%
, $(X_{1,\,1}<\,b_{1})\cap (X_{2,\,1}>\,b_{2})$, $(X_{1,\,1}>\,b_{1})\cap
(X_{2,\,1}<\,b_{2})$.
We could do a bit better if we define $\tilde{X}_{1,\,s}=X_{1,\,s}/b_{1,\,s}$
and $\tilde{X}_{2,\,s}=X_{2,\,s}/b_{2,\,s}$. Then the condition $%
(X_{1,\,1}<\,b_{1,\,1})\cup (X_{2,\,1}<\,b_{2,\,1})$ can be split into the
two conditions $(\tilde{X}_{1,\,1}<1)\cap (\tilde{X}_{1,\,1}\,{\tilde{X}%
_{2,\,1}}^{-1}<1)$, $(\tilde{X}_{2,\,1}\,{\tilde{X}_{1,\,1}}^{-1}<1)\cap (%
\tilde{X}_{2,\,1}<1)$. Therefore to evaluate $\bar{P}_{k}$ we need to sum
over $2^{k}$ possible scenarios, each scenario containing $2k$ conditions.
This requires summing over $2^{k}$ different $2k$-dimensional cumulative
multivariate normal distributions~\cite{SKIPPER}, which is computationally
overwhelming for large values of $k$. Fortunately, using de Morgan rules we
can substantially reduce the computational cost.
Let us denote by ${\cal E}_{i}$ the event $(X_{1,\,i}<\,b_{1})\cup
(X_{2,\,i}<\,b_{2})$, then the event $\bar{\cal E}_{i}$ is written as the single
scenario $(X_{1,\,i}>\,b_{1})\cap (X_{2,\,i}>\,b_{2})$.
Using the well known probability relation
\begin{eqnarray*}
P\left( \bigcap\nolimits_{i=1}^{n}{\cal E}_{i}\right) &=&\sum\limits_{i}P\left(
{\cal E}_{i}\right) -\sum\limits_{i,j}P\left( {\cal E}_{i}\cup {\cal E}_{j}\right)
+\sum\limits_{i,j,k}P\left( {\cal E}_{i}\cup {\cal E}_{j}\cup {\cal E}_{k}\right) + \\
&&...+\left( -1\right) ^{n}P\left( \bigcup\nolimits_{i=1}^{n}{\cal E}_{i}\right)
\end{eqnarray*}
and DeMorgan's law
\[
\overline{\left( \bigcup\nolimits_{i=1}^{n}{\cal E}_{i}\right) }=\bigcap%
\nolimits_{i=1}^{n}\overline{{\cal E}_{i}}
\]
can be shown that
\begin{equation}\label{barPk}
\bar{P}_{k}=P\left( \bigcap_{s=1}^{k}{\cal E}_{s}\right)
=1+\sum_{s=1}^{k}\sum_{\sigma _{s}\in C_{s}^{k}}(-1)^{s}P\left(
\bigcap_{j=1}^{s}\bar{{\cal E}}_{\sigma _{s}(j)}\right) \ .
\end{equation}
where the second sum is over all (sorted in ascending order) combinations of
$k$ elements $s-$th class, $C_{s}^{k}$. Note that there are again $2^{k}$
different terms, however only the last term is $2k$-dimensional.\footnote{%
In general the number of $2s$-dimensional terms is $\binom{k}{s}$.}
In the same spirit we can obtain a formula for the auto-call probabilities $%
P_{k}$:
\begin{equation}
P_{k}=\sum_{s=0}^{k-1}\sum_{\sigma _{s}\in C_{s}^{k-1}}(-1)^{s}P\left(
\bigcap_{j=1}^{s}\bar{{\cal E}}_{\sigma _{s}(j)}\cap \bar{{\cal E}}_{k}\right) \ ,
\label{APk}
\end{equation}%
where we have used a convention: $\cap _{j=1}^{0}\bar{{\cal E}}_{\sigma
_{0}(j)}\cap \bar{{\cal E}}_{k}=\bar{{\cal E}}_{k}$. Equations (\ref{barPk}) and (\ref{APk}%
) can be rewritten in terms of indicator functions. For compactness it is
convenient to adopt the notations of ref.~\cite{SKIPPER}. We introduce a
multi-index notation denoting by $X_{I}$ the element $X_{i,\,s}$, where $%
I=1\,,\dots \,,n$ and $n$ is the number of all observed assets' prices. In
the case considered in equation (\ref{ind-1}) we have $n=2\,k$. Using
lexicographical order we can make the map explicit:
\begin{equation}
(i,s)\rightarrow I=I[i,\,s]=2\ast (s-1)+i\, \label{ldxic}
\end{equation}%
where we have used that $i=1,2$. Next we define the following notation:
\begin{equation}\label{actionA}
(X^{A})_{j}={X}_{1}^{A_{j1}}\,\dots \,{X}_{n}^{A_{jn}}~~~~j=1\,,\dots \,,m\ ,
\end{equation}%
where $m$ is the number of barrier conditions and A is an $n\times m$
matrix. With these notations a general indicator function can be written as:
\begin{equation}
\mathbf{1}_{m}(S\,\mathbf{X}^{A}>S\,\mathbf{a})\, \label{ind_func}
\end{equation}%
where $\mathbf{a}$ is a vector of barriers and to allow for different types
of inequalities we have introduced the $m\times m$ diagonal matrix $S$ whose
diagonal elements take the values $\pm 1$ ('$+$' for '$>$' and '$-$' for '$<$%
'). Equations (\ref{barPk}),(\ref{APk}) now become:
\begin{eqnarray}\label{indicator barP}
&&\bar{P}_{k}=\mathbf{E}\left( 1+\sum_{s=1}^{k}\sum_{\sigma _{s}\in
C_{s}^{k}}(-1)^{s}\mathbf{1}_{2s}(\mathbf{X}^{A(\sigma _{s})}>\mathbf{b}%
(\sigma _{s}))\right) \ , \\
\label{indicator P}
&&P_{k}=\mathbf{E}\left( \sum_{s=0}^{k-1}\sum_{\sigma _{s}\in
C_{s}^{k-1}}(-1)^{s}\mathbf{1}_{2s+2}(\mathbf{X}^{\tilde{A}(\sigma _{s})}>%
\mathbf{\tilde{b}}(\sigma _{s}))\right) \ ,
\end{eqnarray}%
where $\mathbf{A}(\sigma _{s})$, $\mathbf{b}(\sigma _{s})$, $\mathbf{\tilde{A%
}}(\sigma _{s})$, $\mathbf{\tilde{b}}(\sigma _{s})$, are $2k\times 2s$, $%
1\times 2s$, $2k\times 2(s+1)$, $1\times 2(s+1)$ matrices, respectively.
Their non-zero entries are:
\begin{eqnarray}
&&A(\sigma _{s})_{I[i,\,\sigma _{s}(j)],\,I[i,\,j]}=1,~~~(\mathbf{b}(\sigma
_{s}))_{I[i,j]}=b_{i,\,\sigma _{s}(j)}\ , \label{Asigma} \\
&&\tilde{A}(\sigma _{s})_{I[i,\,\sigma _{s}(j)],\,I[i,\,j]}=1,~~~(\mathbf{%
\tilde{b}}(\sigma _{s}))_{I[i,j]}=b_{i,\,\sigma _{s}(j)}\ , \\
&&~~~\text{for}~~~i=1,2~\text{and}~j=1,\dots ,s\ . \nonumber \\
&&\tilde{A}(\sigma _{0})_{I[i,\,k],\,i}=1,~~~(\mathbf{\tilde{b}}(\sigma
_{0}))_{i}=b_{i,\,k}\ ,~~~\text{for}~~i=1,2\ .
\end{eqnarray}%
In equation (\ref{Asigma}) we have used the map (\ref{ldxic}). Note that it
is crucial that the combinations $\sigma _{s}$ are sorted in ascending order.
\subsection{A time-dependant valuation formula}
If we restrict ourselves to time independent deterministic parameters
(interest rate, dividend yield, volatility) we can directly apply the
formula derived in ref. \cite{SKIPPER} to calculate the indicator functions
in equations (\ref{indicator barP}) and (\ref{indicator P}). However, this is a
very crude approximation when dealing with long instruments this is why we
extend the results of ref.~\cite{SKIPPER} to the time dependent case. The
starting point is to model the dynamics of the asset $S_{i}$ with a
geometric Brownian motion:
\begin{equation}
\frac{dS_{i}}{S_{i}}=\left( r(s)-q_{i}(s)\right) ds+\sigma
_{i}(s)\,dW_{i}(s)\ , \label{dyn-1}
\end{equation}%
where $W_{i}$ are correlated Brownian motions with correlation coefficient $%
\rho _{ij}$. Indeed the integrated form of equation (\ref{dyn-1}) is:
\begin{equation}
S_{i}(\tau )=S_{i}^{(0)}\,\exp \left\{ \int\limits_{0}^{\tau }\left(
r(s)-q_{i}(s)-\frac{1}{2}\sigma _{i}(s)^{2}\right) ds+\int\limits_{0}^{\tau
}\sigma _{i}\,dW_{i}(s)\right\}
\end{equation}%
For the asset $i$ at time $T_{k}$ we can write:
\begin{equation}
\log \tilde{X}_{i,k}=\log x_{i}+\left( \bar{r}_{i,k}-\bar{q}_{i,k}-\frac{1}{2%
}\,\bar{\sigma}_{i,k}^{2}\right) \tau _{k}+\bar{\sigma}_{i,k}\,\sqrt{\tau
_{k}}\,Z_{i,k}\ , \label{Xik}
\end{equation}%
where $Z_{i,k}$ is given by:
\begin{equation}
Z_{i,k}=\frac{1}{\bar{\sigma}_{i,k}\sqrt{\tau _{k}}}\int\limits_{0}^{\tau
_{k}}\,\sigma _{i}(s)\,dW_{i}(s)\ \label{Zik}
\end{equation}%
and
\begin{eqnarray}
\bar{r}_{i,k} &=&\frac{1}{\tau _{k}}\int\limits_{0}^{\tau _{k}}ds\,r_{i}(s)\
, \nonumber \label{aver} \\
\bar{q}_{i,k} &=&\frac{1}{\tau _{k}}\int\limits_{0}^{\tau _{k}}ds\,q_{i}(s)\
, \\
{\bar{\sigma}}_{i,k}^{2} &=&\frac{1}{\tau _{k}}\int\limits_{0}^{\tau
_{k}}ds\,\sigma _{i}(s)^{2}\ . \nonumber
\end{eqnarray}%
Following ref.~\cite{SKIPPER} we define the quantities:
\begin{eqnarray}
\mu &=&\left( \bar{r}_{i,k}-\bar{q}_{i,k}-\frac{1}{2}\,\bar{\sigma}%
_{i,k}^{2}\right) \tau _{k}\ , \nonumber \label{mu and Sigma} \\
\Sigma &=&\text{diag\thinspace }(\bar{\sigma}_{i,k}\,\sqrt{\tau _{k}})\ .
\end{eqnarray}%
which are straightforward generalisations of the corresponding definitions
in the time independent case~\cite{SKIPPER}. A bit more involved is the
expression for the correlation matrix $R$ defined as:
\begin{equation}
R_{(i,k)(j,l)}\equiv \langle Z_{i,k}\,,Z_{j,l}\rangle \ .
\end{equation}%
Using equation (\ref{Zik}) and the formula:
\begin{equation}
\left\langle \int\limits_{0}^{\tau _{1}}\sigma
_{i}(s)\,dW_{i}(s)\,,\,\int\limits_{0}^{\tau _{2}}\sigma
_{j}(r)\,dW_{j}(r)\right\rangle =\rho _{ij}\int\limits_{0}^{\min (\tau
_{1},\tau _{2})}\sigma _{i}(\tau )\,\sigma _{j}(\tau )\,d\tau \ ,
\end{equation}%
we obtain:
\begin{equation}
R_{(i,k)(j,l)}=\frac{\rho _{ij}}{\sqrt{\tau _{k}\tau _{l}}\,\bar{\sigma}%
_{i,k}\bar{\sigma}_{j,l}}\int\limits_{0}^{\min (\tau _{k},\tau _{l})}\sigma
_{i}(\tau )\,\sigma _{j}(\tau )\,d\tau \ .
\end{equation}%
Next following ref.~\cite{SKIPPER} we define:
\begin{eqnarray}
\Gamma &=&\Sigma \,R\,\Sigma ^{\prime }\ , \label{Gamma,D,C,d} \\
D &=&\sqrt{\text{diag}\left( A\,\Gamma \,A^{\prime }\right) }\ , \nonumber
\\
C &=&D^{-1}\left( A\,\Gamma \,A^{\prime }\right) D^{-1}\ , \nonumber \\
\mathbf{d} &=&D^{-1}\left[ \log (\mathbf{x}^{A}/\mathbf{a})+A\,\mu \right] \
. \nonumber
\end{eqnarray}%
Here it is used that $x_{i,k}=x_{i}$ for all $k=1,\dots ,M$. In therms of
these quantities the indicator function is given by the same expression as
in ref.~\cite{SKIPPER}, but the underlying variables are given in eq. (\ref%
{Gamma,D,C,d}) and due to the time-dependence thay are different from those
given in the work ref.~\cite{SKIPPER},
\begin{equation}
\mathbf{1}_{m}(S\,\tilde{\mathbf{X}}^{A(\omega )}>S\,\mathbf{a})=\mathcal{N}%
_{m}(S\,\mathbf{d(\omega )},\,S\,C(\omega )\,S)\,\ , \label{formula}
\end{equation}%
where $\mathcal{N}_{m}$ is the cumulative multivariate normal distribution
(centred around zero).
Note that equations (\ref{mu and Sigma})--(\ref{formula}) are valid for any $%
n\times m$ matrix $A$ and any positive barrier vector $\mathbf{a}$.
\subsection{Auto-call probability and final payoff}
Applying equation (\ref{formula}) to calculate the auto-call probability $P_k
$ we obtain:
\begin{equation} \label{Pknew}
P_k =\sum_{s=0}^{k-1}\sum_{\sigma_s \in C^{k-1}_s}(-1)^s\mathcal{N}_{2s+2}(%
\mathbf{d}(\sigma_s)\, ,C(\sigma_s))\ ,~~~k=1\,\dots M-1\ ,
\end{equation}
where $\mathbf{d}(\sigma_s)$ and $C(\sigma_s)$ are obtained by substituting $%
\tilde A(\sigma_s)$ and $\mathbf{b(\sigma_s)}$ from equation (\ref{Asigma})
into equation (\ref{Gamma,D,C,d}). Note that the index $k$ in equation (\ref%
{Pknew}) runs from one to $M-1$. The reason is that the last observation
time is the maturity.
Let us denote by $P_{\mathrm{mat}}$ the probability to reach maturity%
\footnote{%
Note also that $P_{\mathrm{mat}} = \bar P_{M-1}$}. Clearly we have:
\begin{equation}
P_{\mathrm{mat}}=1-\sum_{k=1}^{M-1}P_k\ ,
\end{equation}
The probability $P_{\mathrm{mat}}$ can be split into two contributions:
\begin{equation} \label{Pmat}
P_{\mathrm{mat}}=P_{\mathrm{up}} + P_{\mathrm{down}}\,
\end{equation}
Where $P_{\mathrm{up}}$ is the probability to reach maturity with both
assets simultaneously above the barrier $\kappa\,\bar S_i$, and $P_{\mathrm{%
down}}$ is the probability at least one fo the assets to be bellow the
barrier. In fact the probability $P_{\mathrm{up}}$ is exactly $P_M$, hence
we can write:
\begin{equation}
P_{\mathrm{up}}=\sum_{s=0}^{M-1}\sum_{\sigma_s \in C^{k-1}_s}(-1)^s\mathcal{N%
}_{2s+2}(\mathbf{d}(\sigma_s)\, ,C(\sigma_s))\ .
\end{equation}
Clearly this also determines $P_{\mathrm{down}}$ as $P_{\mathrm{down}}=P_{%
\mathrm{mat}}-P_{\mathrm{up}}$. To calculate the payoff at maturity we also
need the average performance of the assets subject to the condition that the
worst performing asset is bellow the barrier $\kappa\,\bar{S_i}$. The
probability for this to happen is exactly $P_{\mathrm{down}}$, which is a
function of the parameter $\kappa$.
Let us denote $\hat X_i =S_i/\bar S_i$ and define $\hat X = \mathrm{min}%
(\hat X_1,\hat X_2)$, the probability $P_{\mathrm{down}}$ can be written as:
\begin{equation}
P_{\mathrm{down}} = P(\hat X < \kappa)\ .
\end{equation}
The average performance of the assets provided that at least one of the
assets is bellow the barrier $\kappa$ is then proportional to the
conditional expectation value $\langle \hat X\rangle |_{\hat X < \kappa}$:
\begin{equation}
\left\langle \mathrm{min}\left(\frac {S_1} {\bar S_1},\,\frac {S_2 }{\bar S_2%
}\right)\right\rangle \Big |_{\hat X < \kappa} =-\frac{1}{P_{\mathrm{down}}}%
\int\limits_0^\kappa d\kappa\, \kappa\,\frac {d P_{\mathrm{up}}}{d \kappa}\ ,
\end{equation}
where we have used that $dP_{\mathrm{mat}}/d\kappa = 0$. Therefore, the
payoff at maturity is given by:
\begin{equation} \label{Vmat}
V_{\mathrm{maturty}}=P_{\mathrm{up}} +P_{\mathrm{down}}\left\langle \mathrm{%
min}\left(\frac {S_1} {\bar S_1},\,\frac {S_2 }{\bar S_2}\right)\right%
\rangle \Big |_{\hat X < \kappa} =P_{\mathrm{up}} - \int\limits_0^\kappa
d\kappa\, \kappa\,\frac {d P_{\mathrm{up}}}{d \kappa}\ ,
\end{equation}
In the next subsection we calculate the contribution of the coupons.
\subsection{Coupon contribution}
\label{section:coupons}
To obtain the total payoff we have to evaluate the contribution of the
coupons. This can be done by summing over a type of two-asset binary (cash-or-nothing)
options, conditional on the survival of the instrument to the appropriate accrual period. Indeed the probability at time $\tau $ both assets to be above the
barrier is given by the probability for such an option to pay. In the case of the first accrual period this reduced to the standard two-asset binary option~\cite{Heynen}.
To write down a closed form expression for this probability we need to add
one more observation time $\tau _{a}$, which will iterate over the accrual
dates. Clearly the simplest case is when $0\leq t_{a}\leq \tau _{1}$, that
is the first accrual period. In this case we apply formula (\ref{formula}),
for just one observation time $\tau _{a}$, with $A=S=\mathbf{1}_{2}$ and $%
\mathbf{a}=\mathbf{c}$. In more details the probability the coupons to pay
at time $\tau _{a}<T_{1}$, $P_{01}(\tau _{a})$ is given by:
\begin{eqnarray}
P_{01}(\tau _{a}) &=&\mathcal{N}_{2}(\mathbf{d}_{2}(\tau _{a}),C_{2})\ ,
\nonumber \label{P01} \\
d_{i} &=&\frac{\log (\bar{S}_{i}/c_{i})+(\bar{r}-\bar{q}_{i}(\tau _{a})-\bar{%
\sigma}_{i}(\tau _{a})^{2}/2)\,\tau _{a}}{{\bar{\sigma}_{i}}(\tau _{a})\sqrt{%
\tau _{a}}}\ , ~~~i=1,2\ ,\\
C_{2} &=&\left(
\begin{array}{cc}
1 & \rho _{12} \\
\rho _{21} & 1%
\end{array}%
\right)
\end{eqnarray}%
where $\bar{r},\bar{q}_{i}(\tau _{a})$ and $\bar{\sigma _{i}}(\tau _{a})$
are given by equations (\ref{aver}) with $\tau _{k}=\tau _{a}$. The total
number of days in which coupons have been payed in the period $0$ to $\tau
_{1}$, $N_{1}$ is then given by:
\begin{equation}
N_{1}=\sum_{\tau _{a}\,=\,1}^{\tau _{\,1}}P_{01}(\tau _{a})\ . \label{N1}
\end{equation}%
In the same way we can obtain a formula for the number of coupon days in the
second accrual period. The only difference is that now in addition to the
condition both assets to be above the accrual barrier we also have the
condition that the instrument did not auto-call at time $\tau _{1}$. In
general the probability the coupons to pay at time $\tau _{a}$ in the $k$-th
accrual period is the joint probability that the instrument did not
auto-call at the first $k-1$ observation times and both assets are above the
accrual barrier at time $\tau _{a}$. Denoting by ${\cal E}^C_{\tau _{a}}$ the event
that the assets are above the accrual barrier at time $\tau _{a}$ and using
the notations from section \ref{sec:indic}, one can show that\footnote{%
The derivation is analogous to that of equation (\ref{APk}).}:
\begin{equation}
P_{k-1,k}(\tau _{a})=\sum_{s=0}^{k-1}\sum_{\sigma _{s}\in
C_{s}^{k-1}}(-1)^{s}P\left( \bigcap_{j=1}^{s}\bar{{\cal E}}_{\sigma _{s}(j)}\cap
{\cal E}^C_{\tau _{a}}\right) \ , \label{APk-1k}
\end{equation}%
where again we have used the convention: $\cap _{j=1}^{0}\bar{{\cal E}}_{\sigma
_{0}(j)}\cap {\cal E}^C_{\tau _{a}}= {\cal E}^C_{\tau _{a}}$. Equation (\ref{APk-1k}) can be
rewritten in analogy to equation (\ref{indicator P}) as:
\begin{equation}
P_{k-1,k}(\tau _{a})=\mathbf{E}\left( \sum_{s=0}^{k-1}\sum_{\sigma _{s}\in
C_{s}^{k-1}}(-1)^{s}\mathbf{1}_{2s+2}(\mathbf{\tilde{X}_{\sigma _{s}}}>%
\mathbf{bc_{\sigma _{s}}})\ \right) ,
\end{equation}%
where $\mathbf{\tilde{X}_{\sigma _{s}}}$ is the vector: $[X_{1,\sigma
_{s}(1)},\,X_{2,\sigma _{s}(1)},\,\dots \,,X_{1,\sigma
_{s}(s)},\,X_{2,\sigma _{s}(s)},\,S_{1}(\tau _{a}),\,S_{2}(\tau _{a})]$ and $%
\mathbf{bc_{\sigma _{s}}}$ is the vector: $[b_{1,\sigma
_{s}(1)},\,b_{2,\sigma _{s}(1)},\,\dots \,,b_{1,\sigma
_{s}(s)},\,b_{2,\sigma _{s}(s)},\,c_{1},\,c_{2}]$. Denoting by $\tilde{C}%
_{\sigma _{s}}(\tau _{a})$ the covariant matrix constructed using equations (%
\ref{aver})-(\ref{Gamma,D,C,d}) with times $\tau _{\sigma _{s}(1)},\,\dots
,\,\tau _{\sigma _{s}(s)},\,\tau _{a}$ and denoting by $\mathbf{\tilde{d}%
_{\sigma _{s}}}$ the corresponding quantity in equation (\ref{Gamma,D,C,d})
constructed using the barrier vector $\mathbf{bc_{\sigma _{s}}}$, we can
write:
\begin{equation}
P_{k-1,k}(\tau _{a})=\sum_{s=0}^{k-1}\sum_{\sigma _{s}\in
C_{s}^{k-1}}(-1)^{s}\mathcal{N}_{2s+2}(\mathbf{\tilde{d}_{\sigma _{s}}}\,,%
\tilde{C}_{\sigma _{s}}(\tau _{a}))\ .
\end{equation}%
For the number of coupon paying days in the $k$-th accrual period we obtain:
\begin{equation}
N_{k}=\sum_{\tau _{a}\,=\,\tau _{k-1}\,+\,1}^{\tau _{k}}\,P_{k-1,k}(\tau
_{a})\ .
\end{equation}%
To calculate the contribution of the coupons to the total payoff we need to
take into account the discount factors, since we have assumed that the
coupons are payed at the observation times\footnote{%
Note that in practise there are a separate payment dates shortly after the
corresponding observation date.}. Note that the probability the coupons to pay
already include the probability to reach that accrual period. Therefore, the total
coupon contribution is given by:
\begin{equation}
VC_{M}=\gamma \,\sum_{s=1}^{M}\,e^{-\bar{r}_{s}\,\tau _{s}}\,N_{s}\ ,
\end{equation}%
where $\gamma $ is the daily rate of the coupon.
\subsection{Total payoff}
Assuming for simplicity that the instrument redeems at 100 \% in the event
of an auto-call (which in reality is quite common), for the total payoff we
obtain:
\begin{eqnarray}
V_{\mathrm{tot}}&=&V_{\mathrm{maturity}}+\sum_{k=1}^{M-1}e^{-\bar{r}_{k}\,\tau _{k}}\,P_k+VC_M
\end{eqnarray}
where we have substituted $P_{\mathrm{mat}}$ from equation (\ref{Pmat}).
\section{Applications}
In this section we outline some of the applications of the formalism
developed above. We begin with the simplest case of a pure accrual
instrument.
\subsection{Pure accrual instrument}
The pure accrual instrument that we consider in this subsection has the
following characteristics:
--It pays a daily coupon at rate $\gamma$ if at closing time both assets $S_i
$ are above the accrual barriers $c_i$
--At maturity (time $\tau_m$), it redeems at 100\% if both assets $S_i$ are
above certain percentage $\kappa$ of their spot prices at issue time $\bar
S_i$. If at least one of the assets is bellow $\kappa\,\bar S_i$ the
instrument pays only a part proportional to the minimum of the ratios $%
S_i/\bar S_i$.
Clearly this is the general instrument that we considered with the auto-call
option removed. In this simple case the semi-analytic approach of section~%
\ref{sec:anal} is particularly efficient. The coupons are calculated by the
first period formulas in equations (\ref{P01}), (\ref{N1}) with $\tau_a =
\tau_m$, while the payoff at maturity is calculated using equation (\ref%
{Vmat}), with $P_{\mathrm{up}}$ given by:
\begin{equation}
P_{\mathrm{up}} = \mathcal{N}_2(\mathbf{\tilde d}_2(\tau_m), C_2) \ ,
\end{equation}
where $C_2$ is given in equation (\ref{P01}) and $\mathbf{\tilde d}_2(\tau_m)
$ is given by:
\begin{equation}
d_{i} =\frac{\log (\bar{S}_{i}/c_{i})+(\bar{r}-\bar{q}_{i}(\tau _m)-\bar{%
\sigma}_{i}(\tau_m)^{2}/2)\,\tau_m}{{\bar{\sigma}_{i}}(\tau_m)\sqrt{%
\tau_m}}\ , ~~~i = 1,2\ ,
\end{equation}
where $\bar{r},\bar{q}_{i}(\tau_m)$ and $\bar{\sigma _{i}}(\tau_m)$
are given by equations (\ref{aver}) with $\tau _{k}=\tau_m$.
\subsection{Dual index range accrual autocallable instrument}
In this section we compare the efficiency of our semi-analytic (SA) approach and that of a standard Monte Carlo (MC) approach. Since the dimensionality of the SA problem increases linearly with the number of
the auto-call dates, we consider the case of one auto-call date and two range accrual periods. Therefore, our problem is four dimensional and we would still need to rely on numerical methods to estimate the cumulative distributions.
To simplify the analysis even further and facilitate the comparison, we simplify the pay-off at maturity. The instrument pays 100\% if both underlyings perform above the final barrier $\kappa$ (as before), but if this condition is not satisfied, instead of redeeming a worse performance: $\mathrm{min}\left({S_1} /{\bar S_1},\, {S_2 }/{\bar S_2}\right)$ fraction, the instrument redeems at $\kappa \times 100\, \%$. Equation (\ref{Vmat}) then simplifies to:
\begin{equation}
V_{\mathrm{maturty}}=P_{\mathrm{up}} +\kappa\,P_{\mathrm{down}}\ .
\end{equation}
The description of the coupon payments remains the same as in section \ref{section:coupons}. The volatilities $\sigma_i$, dividend yields $q_i$ , interest rate $r$ and correlation correlation coefficient $\rho$ used in the numerical example are presented in table 1. In addition the final barrier was set at 60\% ($\kappa = 0.60$) and the daily accrual rate used was (15/365)\% ($\gamma = 0.15/365$). The length of each accrual period was one year so that: $\tau_1 =1$ and $\tau_2 =2$.
\begin{table}[h]
\begin{center}
\begin{tabular}{c|c|c|c}
$\sigma_i$ & $q_i$ & $r$ & $\rho$ \\
\hline
0.25 & 0.005 & 0.01 & 0.78 \\
0.20 & 0.007 & 0.01 & 0.78
\end{tabular}
\end{center}
\caption{\small Volatilities $\sigma_i$, dividend yields $q_i$, interest rate $r$ and correlation $\rho$ used in the numerical example. }
\end{table}%
To compare the efficiency of the algorithms we compared the running times $T_{\epsilon}$ as functions of the absolute error $\epsilon$. The resulting plot is presented in figure \ref{fig:1}.
\begin{figure}[h]
\centering
\includegraphics[width=5.5in]{TvsErr.pdf}
\caption{\small A plot of the running time $T_\epsilon$ in seconds as a function of the absolute error $\epsilon$. The round red dost represent the SA results, while the square blue dost correspond to the MC data.}
\label{fig:1}
\end{figure}
The round dots correspond to the SA approach, while the square points represent the MC data. As one may expect, the running time $T_\epsilon$ for the MC algorithm increases as $\sim1/\epsilon^2$ and while negligible for $\epsilon < 0.01$, it increases rapidly to $\sim 10 s$, for $\epsilon = 5.0 \times 10^{-4}$. On the other side the SA method has a steady computation time $T_\epsilon \sim 4 s$, for $\epsilon < 2.0\times 10^{-4}$. The SA and MC curves intersect at $\epsilon \approx 0.7 \times 10^{-3}$. The advantage of using the SA method for higher precision $\epsilon < 0.7 \times 10^{-3}$ is evident. For example a calculation with $\epsilon = 2.0 \times 10^{-4}$ would require running the MC simulation for roughly $\sim 60\, s$, while the same accuracy can be achieved by the SA method for $\sim 5\, s$, which is a factor of twelve. Clearly the comparison depends on the implementation and the choice of parameters. To make the comparison fair we used MatLab for both methods. Using a vectorised MC algorithm for the Monte Carlo part and the built in MatLab cumulative distribution functions for the SA approach.
Another obvious advantage of the SA approach is the higher precision in the estimation of the sensitivities of the instrument. Semi-analytic expressions could be derided for most of the greeks, which enables their calculation with a limited numerical effort. This is clearly not the case in the MC approach, where one usually relies on a numerical differentiation.
Finally, as we pointed out at the beginning of this section the dimensionality of the problem increases linearly with the number of auto-call times. It is therefore expected that at some point the MC approach would become more efficient. Nevertheless, the SA approach could still be more efficient if the sensitivities are difficult to analyse in the MC approach.
\section{Conclusion}
This paper makes several contributions to the related literature.
Our main result is the development of a semi-analytical valuation method for
auto-callable instruments embedded with range accrual structures. Our approach
includes time-dependent parameters, and hence greatly facilitates practitioners.
In the process we extend the valuation formula for multi-asset, multi-period binaries
of ref.~\cite{SKIPPER} to the case of time-dependent parameters,
which the best of our knowledge is a novel result.
Another merit of this work is the comparison between the straightforward Monte Carlo
and the semi-analytical approaches.Our comparison shows that the semi-analytical
approach becomes more advantageous at higher precisions and is potentially order
of magnitude faster than the brute force Monte Carlo method. The semi-analytical
approach is also particularly useful when calculating the sensitivities of the instrument.
It is widely accepted that the sensitivity calculations are often more important than the
instrument price itself, due to their contribution for the correct instrument hedging.
Finally, our work can be used as a starting point for modelling
more complex structures related to range accrual auto-callable
instruments. Furthermore, although the numerical
examples and the presented formulas are given for the
two-dimension cases, multi-asset and multi-period generalisation
of the formulas can be easily written using the key formulas
presented here.
\paragraph{Acknowledgements:}
We would like to thank Bojidar Ibrishimov for critically reading
the manuscript.
|
\section{Introduction}
In the field of regenerative medicine, an emerging way to treat patients is to
reprogram cells, leading, for instance, to tissue or neuron regeneration. Such a
challenge has become realistic after first experiments have shown that some of
the cell fate decisions can be reversed \cite{Takahashi2016}. Whereas the cells
go through several multipotent states before reaching a differentiated state,
the differentiation process can be inversed, producing induced pluripotent stem
cells (iPSCs) from an already differentiated cell. By using a distinct
differentiation path, this allows to "transform" the type of a cell.
Alternatively, it is also possible to directly perform a trans-differentiation
without necessarily going (back) through a multipotent state
\cite{Graf2009,delSol14}.
In the aforementioned work, the de- and trans-differentiation has been achieved
by targeting specific genes, that we refer to as \emph{Reprogramming
Determinants} (RDs), through the mediation of their transcription factors
\cite{Takahashi2016,Crespo2013}.
The computational prediction of RDs requires to assess multiple features of the cell dynamics and the reprogramming strategy, such as the impact of the kind of perturbations (persistent versus temporary) and of their order; the nature of targeted cell type (differentiated/pluripotent), and the desired inevitability of their reachability (fidelity); the nature and duration of the triggered cascade of regulations (efficiency); and finally, the RD robustness with respect to initial state heterogeneity among cell population, and with respect to uncertainties in the computational model.
So far, no general framework allows to efficiently encompass those features to systematically predict best combinations of RDs in distinct cellular reprogramming events.
In this paper, we address the identification of RDs from \emph{Boolean Networks} (BNs) which model the dynamics of gene regulation and signalling networks. The state of the components (or nodes) of the networks are represented by Boolean variables, and the state changes are specified by Boolean functions which associate the next state of nodes, given the (binary) state of their regulators
\cite{Thomas73,Aracena2008}.
BNs are well suited for an automatic reasoning on large biological networks where the available knowledge is mostly about activation and inhibition relations\cite{Abou-Jaoude2015}. Such activation/inhibition relations between components form a signed directed graph, that we refer to as the \emph{Interaction Graph}.
In this work, we make the assumption that the differentiated cellular states correspond to the \emph{attractors} of the dynamics of the computational model, i.e., the long-run behaviours.
In the scope of BNs, those attractors can be of two kinds: either a single state (referred to as a fixed point), or a terminal cyclic behaviour.
The relationship between the IG of BNs and the number of their attractor has been extensively studied
\cite{Aracena2008,Richard09-MaxFP,Richard10-AAM}.
However, little work exists on the characterization of the perturbations which trigger a change of attractor.
Currently, most of RDs prediction are performed using statistical analysis on
expression data in order to rank candidate transcription factors
\cite{Chang2011,Rackham2016,Jo2016}.
Whereas based on network models, those approaches do not allow to derive a
complete set of solution for the reprogramming problem.
In \cite{Crespo2013}, the authors developed a heuristic to derive candidate RDs
from a pure topological analysis of the interaction graph:
the RDs are selected only in positive cycles that have different values in the
started and target fixed points.
However, there is no guarantee that the derived RDs can actually lead to a
change of attractor in the asynchronous dynamics of the Boolean networks, and
neither that the target fixed point is the only one reachable.
Finally, \cite{Gao2016} gives a formal characterization of RDs subject to
temporal mutations which trigger a change of attractor in the synchronous
semantics of conjunctive Boolean networks.
\paragraph{Contribution}
This work relies on model checking and reachability analysis, that have been proved useful and successful in previous studies\cite{Abou-Jaoude2015,P16-CMSB}.
Given a BN, all of whose attractors are fixed points, given an initial fixed point and a target fixed point, we provide a characterization of the candidate RDs (set of nodes) with respect to the interaction graph and for two settings of cellular reprogramming:
\begin{itemize}
\item
with a permanent perturbation of RDs, the target fixed point becomes reachable in the asynchronous dynamics of the BN;
\item
with a permanent perturbation of RDs, the target fixed point is the sole reachable attractor in the asynchronous dynamics of the BN.
\end{itemize}
For the first case, we prove that all the RDs are distributed among particular strongly connected components of the interaction graph, and we give algorithms to determine them in both settings.
In the second case, we prove that only some of them are distributed among
strongly connected components of the interaction graph.
We provide an algorithm to identify possible combination of permanent
perturbations leading to inevitable reachability of the target fixed point.
Whereas the algorithm may miss some solutions, all returned solutions are
correct.
\paragraph{Outline}
Section ~\ref{sec:background} gives the definitions and basic properties of BNs and of their asynchronous dynamics. The formalization of the BN reprogramming problem with permanent perturbations of nodes is established in Sect.~\ref{sec:bn-reprogramming}. Section ~\ref{sec:rd-scc} states the main results on the characterization of RDs with respect to the interaction graph of BNs. An algorithm to enumerate all RDs by exploiting this characterization is given in Sect.~\ref{sec:enumeration}. Finally, \ref{sec:discussion} discusses the results and sketches future work.
\subsection*{Notations}
Given a finite set $I$, $\powerset{I}$ is the power set of $I$, $\card{I}$ the
cardinality.
Given a positive integer $n$, $[n] = \{1,\dots,n\}$.
Given a Boolean state $x\in\{0,1\}^n$ and set of indexes $I\subset [n]$,
$\bar{x}^I$ is the state where $\bar{x_i}^I = x_i$ if $i \notin I$ and $\bar{x_i}^I = 1-x_i$ if $i \in I$.
Similarly, given $x,y\in\{0,1\}^n$, $x_{[x_{I} = y_{I}]}$ denotes the state where
for all $i \in I,\, (x_{[x_{I} = y_{I}]})_i = y_i$ and for all $i \notin I,\,
(x_{[x_{I} = y_{I}]})_i = x_i $\\
\section{Background}
\label{sec:background}
In this section, we give the formal definition of Boolean networks, their
interaction graph and transition graph in the asynchronous semantics.
Finally, we recall the main link between their attractors and the positive
cycles in their interaction graph.
\subsection{Definitions}
\paragraph{Boolean Network (BN):}
A BN is a finite set of Boolean variables, each of them having a Boolean function. This function is a logical Boolean function depending from the network's variables and determining the next state of the variable.
\begin{definition}[Boolean Network (BN)]
A Boolean Network is a function $f$ such that:
\begin{align*}
f&: &\{0,1\}^n & \rightarrow \{0,1\}^n \\
& &x=(x_1,...,x_n) & \mapsto f(x) = (f_1(x),...,f_n(x))
\end{align*}
\label{def:bn}
\end{definition}
\begin{example}
An example of BN of dimension $3$ ($n = 3$) is
\begin{align*}
f_1(x) &=x_3 \vee (\neg x_1 \wedge x_2) \\
f_2(x) &= \neg x_1 \vee x_2 \\
f_3(x) &= x_3 \vee (x_1 \wedge \neg x_2)
\end{align*}
\label{ex:baseExample}
\end{example}
\paragraph{Interaction Graph:}
To determine the RDs, we rely on a simplification of the interactions between
the genes, and of the concentrations. A gene will either be active or inhibited.
Gene interactions are simplified likewise, a gene either activates or inhibits
another gene, and we ignore time scales. With this in mind, an \emph{interaction graph} (Def.\ref{def:ig}) can be build: genes are the vertices, and the interactions are the oriented arcs, labelled either $+$ or $-$, if it is an activation or an inhibition.
\begin{definition}[Interaction Graph]
An interaction graph is noted as $G = (V, E)$, with $V$ being the vertex set, and $E$ being the directed, signed edge set, $E \subset (V \times V \times \{-,+\})$
\label{def:ig}
\end{definition}
A cycle between a set of nodes $C\subseteq V$ is said positive (resp. negative) if and only if
there is an even (odd) number of negative edges between those nodes.
An interaction graph can also be defined as an abstraction of a Boolean network: the functions are not given and not always known, but if a vertex $u$ is used in the function $f_v$, there is an edge from $u$ to $v$, negative if $f_v(x)$ contains $\neg x_u$ and positive if it contains $x_u$.
\begin{definition}[Interaction Graph of a Boolean network ($G(f)$)]
An interaction graph can be obtained from the Boolean network $f$:
the vertex set is $[n]$, and for all $u,v \in [n]$ there is a positive (resp. negative) arc from $u$ to $v$ if $f_{vu}(x)$ is positive (resp. negative) for at least one $x \in \{0,1\}^n$ (For every $u,\, v \in \{1, . . . , n\}$, the function $f_{vu}$ is the discrete derivative of $f_v$ considering $u$, defined on $\{0, 1\}^n$ by : $f_{vu}(x):=f_v(x_1,..,x_{u-1},1,x_{u+1},..,x_n) - f_v(x_1,..,x_{u-1},0,x_{u+1},..,x_n)$).
\end{definition}
Given an interaction graph $G=(V,E)$, and one of its vertex $u \in V$, $P_u$
denotes the set of ancestors of $u$, i.e., the vertices $v$ for which there exists a path in $E$ from $v$ to $u$.
Similarly, $p_u$ is the set of the parents of $u$, i.e., $v \in p_u \Rightarrow (v,u,s) \in E$.
Furthermore, $\gp{u}$ is the induced subgraph of $G$ with $P_u$ as vertex set.
Fig.~\ref{fig:example-ig} gives an example of an interaction graph, which is also
equal to $G(f)$, where $f$ is the Boolean network of Ex.\ref{ex:baseExample}.
\begin{figure}[t]
\begin{center}
\begin{tikzpicture}[node distance=2cm, every loop/.style={}]
\node[main node] (2) {2} ;
\node[main node] (1) [below left of=2] {1} ;
\node[main node] (3) [below right of=2] {3} ;
\path (1) edge[-|, bend left, red] node[left] {$-$} (2)
(1) edge[-|, loop left, red, ] node {$-$} (1)
(1) edge[->, bend right, blue] node[below] {$+$} (3)
(2) edge[->, bend left, blue] node[right] {$+$} (1)
(2) edge[->, loop above, blue] node {$+$} (2)
(2) edge[-|, red] node[right] {$-$} (3)
(3) edge[->, blue] node[above] {$+$} (1)
(3) edge[->, loop right, blue] node {$+$} (3) ;
\end{tikzpicture}
\end{center}
\caption{Interaction graph of Ex.\ref{ex:baseExample}
A "normal" blue arrow means an activation, and a "flattened" red arrow means an inhibition.
\label{fig:example-ig}
}
\end{figure}
\paragraph{Transition Graph:}
We model the dynamics of a Boolean network $f$ by \emph{transitions} between its
states $x\in\{0,1\}^n$.
In the scope of this paper, we consider the \emph{asynchronous semantics} of
Boolean networks:
a transition updates the value of only one vertex $u\in[n]$.
From a single $x\in\{0,1\}^n$, one has different transitions for each vertex $u$
such that $f_u(x) \neq x_u$.
This leads to the definition of the \emph{transition graph}
(Def.~\ref{def:transition-graph}) where vertices are
all the possible states $\{0,1\}^n$, and edges correspond to asynchronous
transitions.
\begin{definition}[Transition graph]
\label{def:transition-graph}
The transition graph is the graph having $\{0,1\}^n$ as vertex set and the edges
set $\{x \to \bar{x}^{\{u\}} \mid x\in\{0,1\}^n, u\in[n], x_u \neq (f(x))_u\}$. An existing path from $x$ to $y$ is noted $x \to^* y$.
\end{definition}
Fig.\ref{fig:ex-tg} gives the transition graph of the asynchronous dynamics of
Boolean network of Ex.\ref{ex:baseExample}.
\begin{figure}
\begin{center}
\begin{tikzpicture}[node distance=1.5cm]
\node[draw] (000) {000} ;
\node[draw] (010) [above of=000] {010} ;
\node[draw] (100) [right of=000] {100} ;
\node[draw] (110) [right of=010] {110} ;
\node[draw] (001) [above right of=100] {001} ;
\node[draw] (011) [above of=001] {011} ;
\node[draw] (101) [right of=001] {101} ;
\node[draw] (111) [right of=011] {111} ;
\node[draw, align=left, minimum width=2.5cm, minimum height=0.8cm] at (0.8,1.5) [magenta] {} ;
\node[draw, align=left, minimum width=0.9cm, minimum height=0.7cm] at (4.05,1.05) [magenta] {} ;
\node[draw, align=left, minimum width=0.9cm, minimum height=0.7cm] at (4.05,2.55) [magenta] {} ;
\path (000) edge[->] (010)
(001) edge[->] (011)
(001) edge[->] (101)
(010) edge[bend left, ->] (110)
(011) edge[->] (111)
(100) edge[->] (000)
(100) edge[->] (101)
(110) edge[bend left, ->] (010) ;
\end{tikzpicture}
\end{center}
\caption{Transition graph of the Boolean network defined in Ex.\ref{ex:baseExample}.
We use shorter notations, $010$ meaning that the node $1$ has $0$ as value, the node $2$ has $1$ as value, and the node $3$ has $0$ as value.
The attractors are boxed in magenta.
\label{fig:ex-tg}}
\end{figure}
\paragraph{Attractors, Fixed point :}
BN's \emph{Attractors} are the terminal strongly connected components of the transition graph, and can be seen as the long-term dynamics of the system. Note that an attractor is always a set of states, but it can contain either multiple distinct nodes, that is the system oscillate between multiple states (\emph{cyclic attractor}) or a unique point, i.e the system stays in the same state (\emph{fixed point}).
\begin{definition}[Attractor]
\begin{align}
S \subseteq \{0,1\}^n \mbox{ is an attractor} \Leftrightarrow & \, S \neq \emptyset \\
&\mbox{ and } \forall x \in S,\, \forall y \in \{0,1\}^n \setminus S,\, x \not \to y \\
&\mbox{ and } \forall x \in S,\, S \setminus x \mbox{ does not verify (2) }
\end{align}
If $ \card S=1$ then $S$ is a fixed point. Otherwise $S$ is a cyclic attractor.
\end{definition}
Given a BN $f$, $\operatorname{FP}(f) \subseteq \{0,1\}^n$ denotes the set of its fixed points
($\forall x \in \operatorname{FP}(f), f(x)=x$).
\begin{example}
The BN of Ex.\ref{ex:baseExample} has 3 attractors that correspond to the 3
terminal strongly connected components of Fig.\ref{fig:ex-tg}:
$\{010,110\}$ (cyclic attractor), $\{101\}$ and $\{111\}$ (fixed points).
\end{example}
\subsection{On the link between attractors and the interaction graph}
Theorem~\ref{thm:thomas} is a conjecture by Ren\'e Thomas \cite{Thomas73} that
has been since demonstrated for Boolean and discrete networks
\cite{Aracena2008,REMY2008335}:
if a Boolean network has multiple attractors then its interaction graph
necessarily contains a positive cycle.
In the case of multiple fixed points, any pair of fixed point differ at least on
a set of nodes forming a positive cycle.
\begin{theorem}[Thomas' first rule]
If $G=(V,E)$ has no positive cycles, then $f$ has at most one attractor. Moreover,
if $f$ has two distinct fixed points $x$ and $y$, then $G$ has a positive
cycle between vertices $C\subseteq V$ such that $x_v \neq y_v$ for every vertex $v$ in $C$.
\label{thm:thomas}
\end{theorem}
We can also remark that for a vertex to stay at a value $y_v$ where $y$ is a fixed point, it only needs its ancestors to have the same values as in $y$.
\begin{remark}
$\forall y \in \operatorname{FP}(f),\,\forall u \in [n], \forall z \in \{0,1\}^n \mbox{, z verifying } \forall j \in P_u,\, z_j = y_j, \mbox{ we have } f_u(y) = y_u =f_u(z)$.
\end{remark}
\begin{proof}
Let $u$ be a vertex in $[n]$. $f(u)$ only depends of the incoming arcs in $u$, so it only depends of $p_u$, which in turn depends on its parents. By induction, $f_u(y)$ only depends of $P_u$, and so, if $f_u(y) = y_u$ in $G$, then $f_u(y) = y_u$ in $\gp{u}$. \qed
\end{proof}
\section{Formalisation of the BN Reprogramming with Permanent Perturbations}
\label{sec:bn-reprogramming}
Given two fixed points $x$ and $y$ of Boolean network $f$, we want to identify
sets of nodes, referred to as Reprogramming Determinants (RDs), that when changed in $x$ enable to switch to $y$. As our theorems rely on the differences between the fixed points, we chose to focus on fixed points solely. Further work will extend, if possible, these theorems and algorithms to all kind of attractors.
In the scope of this paper, by "change" we mean permanently set the vertex to a new fixed value. If we "change" $u$ to $1$ (resp. $0$), then $f_u(x) =1$ (resp. $0$) for all $x$.
When switching to $y$ (by changing $I$) is possible, we have two cases : it
either means that $y$ is reachable from $x_{[x_I = y_I]}$ (existential
reachability, Def.~\ref{def:er}), or that $y$ is the only reachable fixed point
from $x_{[x_I = y_I]}$ (inevitable reachability, Def.~\ref{def:ir}).
These are two different approaches that we will both consider. To remove the temporal aspect, we make all the changes at the same time (hence $x_{[x_I = y_I]}$, otherwise an order should be visible), and only watch if $y$ is reachable. This also means that there is no indication of how long it takes for $y$ to be reached.
\begin{definition}[Existential Reachability]
\label{def:er}
With the boolean network F, a function $ER_F$ can be defined as $ER_F :
\powerset{\powerset{[n]}}$, with $ER_F(x,y) \mapsto v$ where $v$ is the set of all minimal vertex sets $I$ such that $x_{[x_I=y_I]} \rightarrow^* y$.
\end{definition}
\begin{definition}[Inevitable Reachability]
\label{def:ir}
Similarly, a function
$IR_F : \powerset{\powerset{[n]}}$ can be defined as $IR_F(x,y) \mapsto w$ where $w$ is the set of all minimal vertices sets $I$ such as $\forall z \in \{0,1\}^n,\, x_{[x_I=y_I]} \rightarrow^* z \Rightarrow z \rightarrow^* y$.
\end{definition}
These two functions will give different results, and have different meanings, as shown in the examble below.
\begin{example}
\label{ex:reprog}
Let us consider the BN $f$ of Fig.\ref{fig:sbn} and its transition graph reproduced
in Fig.\ref{fig:sbn-tg}.
$f$ has 4 fixed points: $0000, 0001, 1100$ and $1101$.
Let $x=0000$ and $y=1100$.
Fixing the node $\{1\}$ to $1$ in $x$ makes $y$ reachable :
$1100$ (=$y$) is reachable from $x_{[x_1=1]}=1000$ with the Boolean network $f'$ defined by
$f'_1(x) = 1$ and $f'_2=f_2$, $f'_3=f_3$, $f'_4=f_4$.
The transition graph of $f'$, considering the first node being active,
corresponds to the left part of the transition graph in Fig.\ref{fig:sbn-tg}.
One can then remark that $y$ is not the only fixed point reachable: from $1000$, $1101$ is also reachable.
If we also fix the node $\{4\}$ to $0$, $y$ is the only reachable fixed point from $x_{[x_1=1, x_4=0]}$
in the Boolean network $f''$ such that $f''_1(x) = 1$, $f''_2=f_2$, $f''_3=f_3$,
and $f''_4(x)=0$.
Therefore, with the previous definitions,
$\{1\} \in ER_F(0000,1100)$ but
$\{1\} \notin IR_F(0000,1100)$;
and
$\{1,4\}\in IR_F(0000,1100)$ but
$\{1,4\}\notin ER_F(0000,1100)$.
Moreover, we also have $\{1,2\}$ and $\{1,3\}\in IR_F(0000,1100)$.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\node[draw, align=left] (text) at (-4,-1) {$f_1(x) = x_1 $ \\ $ f_2(x) = x_1 $ \\ $ f_3(x) = x_1 \wedge \neg x_3 $ \\ $ f_4(x) = x_3 \vee x_4$} ;
\node[main node] (a) {1} ;
\node[main node] (b) [below right of=a] {2} ;
\node[main node] (c) [below left of=a] {3} ;
\node[main node] (d) [below of=c] {4} ;
\path (a) edge[blue, loop above] (a)
(a) edge [blue, ->] (b)
(a) edge [blue, ->] (c)
(b) edge[red, -|] (c)
(c) edge[blue, ->] (d)
(d) edge[blue, loop below] (d) ;
\end{tikzpicture}
\end{center}
\caption{A BN of dimension $4$} \label{fig:sbn}
\end{figure}
\begin{figure}
\begin{center}
\begin{tikzpicture}[node distance=1.3cm]
\node (1010) {1010} ;
\node (1110) [right of=1010] {1110} ;
\node (1000) [below of=1010, label={[red]left:$\bar{x}^{\{1\}} =$}] {1000} ;
\node[draw, align=center] (1100) [red, right of=1000, label={[red]below:$y$}] {1100} ;
\node (phantom0) [above left of=1010] {} ;
\node (1011) [above left of=phantom0] {1011} ;
\node (1111) [right of=1011] {1111} ;
\node (1001) [below of=1011] {1001} ;
\node (1101) [red, right of=1001] {1101} ;
\node (phantom1) [right of=1110] {};
\node (0010) [right of=phantom1] {0010} ;
\node (0000) [red, below of=0010, label={[red]below:$x$}] {0000} ;
\node (0110) [right of=0010] {0110} ;
\node (0100) [right of=0000] {0100} ;
\node (phantom2) [above left of=0010] {} ;
\node (0011) [above left of=phantom2] {0011} ;
\node (0111) [right of=0011] {0111} ;
\node (0001) [red, below of=0011] {0001} ;
\node (0101) [right of=0001] {0101} ;
\path (0010) edge[->] (0000)
(0010) edge[bend right, ->] (0011)
(0011) edge[->] (0001)
(0100) edge[->] (0000)
(0101) edge[->] (0001)
(0110) edge[->] (0100)
(0110) edge[->] (0010)
(0110) edge[bend right, ->] (0111)
(0111) edge[->] (0011)
(0111) edge[->] (0101)
(1000) edge[->] (1010)
(1000) edge[->] (1100)
(1001) edge[->] (1101)
(1001) edge[->] (1011)
(1010) edge[bend right, ->] (1011)
(1010) edge[->] (1110)
(1011) edge[->] (1111)
(1110) edge[->] (1100)
(1110) edge[bend right, ->] (1111)
(1111) edge[->] (1101) ;
\end{tikzpicture}
\end{center}
\caption{Transition graph of the BN in Fig.\ref{fig:sbn}
\label{fig:sbn-tg}}
\end{figure}
\end{example}
\section{Reprogramming Determinants and the SCCs of the Interaction Graph}
\label{sec:rd-scc}
In this section, we show the link between the RDs and the Strongly Connected
Components (SCCs) of the interaction graph of the Boolean network $f$.
Our results make the assumption that all the attractors of $f$ are fixed points
(no cyclic attractors).
\subsection{SCC Ordering}
To switch from $x$ to $y$, we want to change the value of each vertex $u$ that has different values for $x$ and $y$ ($x_u \neq y_u$) and to prevent each vertex $v$ that verifies $x_v = y_v$ from changing value. We know that changing the value of a vertex can have an impact on other vertices, but we also know that it will only impact its descendants.
So, if a vertex has a different value in $x$ and $y$ but none of its ancestors do, then it is necessary to change this vertex. So, to know which vertices need to be changed first, the best way is to order them, with a topological order for example.
Of course, if there are loops, an order is impossible to determine, we have to reduce all SCCs to single "super-vertices" to achieve it.
In the remaining of this paper, we will consider SCCs which contain at least
one positive cycle, because they are known to change between fixed points (Theor.\ref{thm:thomas}), we call $\mathcal{O}$ the SCC set that contains all such SCCs.
Reducing the graph to its SCCs makes possible to rank them from $1$ to $k$ with
any topological order, noted $\prec$:
for all $i,j\in[k], j > i\Rightarrow \mathcal{O}_j \not\prec \mathcal{O}_i$.
Let $C_0$ be the set $\{ \mathcal{O}_i \in \mathcal{O} \mid \nexists \mathcal{O}_j,\, \mathcal{O}_j \prec \mathcal{O}_i\}$, and recursively define slices $C_K = \{ \mathcal{O}_i \in (\mathcal{O} \setminus \bigcup_{l \in \{1,..,K-1\}}C_l) \mid \nexists \mathcal{O}_j,\, \mathcal{O}_j \prec \mathcal{O}_i\}$. Given the definition of the slices, for all topological orders, the slice set will be the same. The slices are numbered from $1$ to $c$.
From this order, we know which SCCs need to be impacted, still, SCCs ranked lower in the hierarchy
need not be impacted by the change in their ancestors (see ex.\ref{ex:blockswitch}) The relation $\prec$ only gives an order to make the changes, from which one can determine if further changes are needed.
\begin{example}
Showing that only using the topological order is not sufficient.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\node[draw, align=left] (text) at (-4,0) {$f_1(x) = \neg x_2$ \\ $f_2(x) = \neg x_1$ \\ $f_3(x) = x_1 \vee x_2$ \\ $f_4(x) = x_2 \wedge \neg x_3 $ \\ $f_5(x) = x_4 \vee x_5$} ;
\node[main node] (c) {3} ;
\node[main node] (a) [below left of=c] {1} ;
\node[main node] (b) [below right of=c] {2} ;
\node[main node] (d) [right of=c] {4} ;
\node[main node] (e) [right of=d] {5} ;
\path (a) edge[red, bend left, -|] (b)
(b) edge [red, bend left, -|] (a)
(a) edge [blue, ->] (c)
(b) edge[blue, ->] (c)
(c) edge[red, -|] (d)
(b) edge[blue, ->] (d)
(d) edge[blue, ->] (e)
(e) edge[blue, loop right] (e) ;
\end{tikzpicture}
\end{center}
\caption{BN preventing changes in the lower SCC}
\label{ex:blockswitch}
\end{figure}
Any algorithm that only used the topological order without computing the reachable fixed points would not suffice, as the example from Fig.\ref{ex:blockswitch} shows :
the switch from the fixed point $01100$ to $10101$ would be computed by just modifying $\{1\}$, but in fact $\{4\}$ will always be fixed at $0$, because $\{4\}$ is always inhibited by $\{3\}$, so $\{5\}$ needs to be changed too.
\label{ex:blockswitcht}
\end{example}
\subsection{SCC Filtering}
Whether we want $y$ to be the only reachable attractor, or merely to be one of potential several such attractors, the ordering from the previous part is the same, but the filtering will differ.
\begin{theorem}
\label{theor:ancestrality}
If a vertex $u$ such as $x_u \neq y_u$ and $u$ is not in a positive cycle, then modifying $u$'s ancestors is sufficient to modify $u$. \\ More generally, to switch from $x$ to $y$, modifying only those strongly connected components that contain at least a positive cycle is sufficient.
\end{theorem}
\begin{proof}
Let $u$ be a vertex such that $x_u \neq y_u$ and $u$ does not lie in a positive cycle. If $u$ is in a negative cycle, the incoming arc from the cycle is irrelevant : given that $x$ and $y$ are fixed points and that $u$ has a distinct value in each, the negative cycle does not change $u$'s value. Given that $u$ is not in a positive cycle, $u$ is not in a SCC (or not relevant if it is in a negative cycle). That means that none of the ancestors are descendants of $u$. Let $z$ be the state where all of $P_u$ ($u$'s ancestors) have the same value that in $y$. By the remark from Sect.\ref{sec:background}, for all $v \in \gp{u}$, we have $f_v(z)=z_v=y_v$. So, either $f_u(z)=y_u$, and the theorem is proven, either $f_u(z) \neq y_u$, then, by Theor.\ref{thm:thomas}, $u$ is in a positive cycle, contradiction.
\qed
By recursion over the first part, modifying all the SCCs that contain positive cycles so their vertices have the same value as in $y$ modifies all their children, and then all the children of their children, and so on, until the whole graph has the same values as $y$. \qed
\end{proof}
Selecting the SCCs will differ with the two methods. It relies on the same base, searching the higher SCC that should have its values modified and that is not already selected. "Modified" means that all the values of the SCC are fixed to their values in $y$. The set of the selected SCCs is $\mathcal{S}$.
\subsection{SCC Filtering for Existential Reachability}
We consider the RDs for the BN reprogramming with Existential Reachability.
We give an algorithm to identify different sets of SCCs for which the mutation
in the initial fixed point ensure the reachability of the target fixed point.
We will prove that the identified combination of SCCs is complete and minimal.
Basically, the algorithm reviews linearly the SCC slices according to $\prec$
and adds the minimal combinations of SCCs to $\mathcal{S}$ that are different in $y$ and
the fixed points reachable from $x_{[x_{\mathcal{S}} = y_{\mathcal{S}}]}$:
\begin{enumerate}
\item $\mathcal{S} := \emptyset$
\item For $i$ ranging from $1$ to $c$:
\begin{itemize}
\item $T := \emptyset$
\item $\forall s \in P(C_i)$ such that $s$ minimal\\ $\exists z \in \{0,1\}^n,\, z_{C_i \setminus s} = y_{C_i \setminus s},\, x_{[x_I = y_I | I \in \mathcal{S}]} \to^*z$, $T := T \cup s.$
\item $\mathcal{S} := \mathcal{S} \bar{\times} T$.
\end{itemize}
\end{enumerate}
With $\bar{\times}$ being a product and union : for a set $I$ of subsets $I_1,..,I_k$ and a set $J_1,..,J_l$, this product $\bar{\times}$ is defined by : $I \bar{\times} J = \{I_1 \cup J_1, .., I_1 \cup J_l, I_2 \cup J_1, ...., I_k \cup J_l\}$
\paragraph*{Complexity :}
In the worst case, the above algorithm perform $c\times 2^l$ reachability checks
(PSPACE-complete \cite{ChengEP95}), where $l$ is the size of the largest slice.
\paragraph*{Existence of a solution and proof of correctness :}
Forcing all SCCs of such problem that differ on $x$ and $y$ to have the same value as in $y$ is one solution. In the worst case, that is what the algorithm will find. Since the algorithm tests reachability, and a solution exists, it will find one.
\begin{example}
We apply the algorithm on the BN of Fig.\ref{fig:bn-slices}
with $x=00000$ and $y=11011$.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\node[draw, align=left] (text) at (-4,-1) {$f_1(x) = x_1 $ \\ $ f_2(x) = x_1 $ \\ $ f_3(x) = x_1 \wedge \neg x_2 $ \\ $ f_4(x) = x_3 \vee x_4$\\ $f_5(x) = x_2 \vee x_5 $} ;
\node[main node] (a) {1} ;
\node[main node] (b) [below right of=a] {2} ;
\node[main node] (c) [below left of=a] {3} ;
\node[main node] (d) [below of=c] {4} ;
\node[main node] (e) [below of=b] {5} ;
\node[draw, align=left, minimum width=2.5cm, minimum height=0.5cm] at (0,0) [magenta, label={[magenta]right:$C_1$}] {} ;
\node[draw, align=left, minimum width=2.5cm, minimum height=0.5cm] at (0,-1.7) [magenta, label={[magenta]right:$C_2$}] {} ;
\path (a) edge[blue, loop above] (a)
(a) edge [blue, ->] (b)
(a) edge [blue, ->] (c)
(b) edge[red, -|] (c)
(c) edge[blue, ->] (d)
(b) edge[blue, ->] (e)
(e) edge[blue, loop below] (e)
(d) edge[blue, loop below] (d) ;
\end{tikzpicture}
\end{center}
\caption{BN of dimension $5$ (left) with its interaction graph (right).
Slices are enclosed in boxes. $C_1=\{\{1\}\}$, $C_2=\{\{4\},\{5\}\}$.
\label{fig:bn-slices}}
\end{figure}
\begin{enumerate}
\item
$\mathcal{S} := \emptyset$
\item
$C_1$: $s$ minimal $\Leftrightarrow \, s = \{1\}$
\item
$\mathcal{S} := \mathcal{S} \bar{\times} \{1\} = \{\{1\}\}$
\item
$C_2$: $s$ minimal $\Leftrightarrow \, s = \emptyset$ \footnote{with the path $10000 \rightarrow 10100 \rightarrow 10110 \rightarrow 11110 \rightarrow 11111$ (and the fixed point is the next step, $\rightarrow 11011$ but there is no need to go further than $11111$.)}
\item
$\mathcal{S} := \mathcal{S} \bar{\times} \emptyset = \{\{1\}\}$.
\end{enumerate}
\end{example}
We now prove the completeness of the algorithm and the minimality of the
returned sets of SCCs (Theorem~\ref{thm:er-algo})
and that any RDs in $ER(x,y)$ is spans only and necessarily in one of the set of
SCCs identified by the algorithm (Theorem~\ref{thm:er-scc}).
\begin{theorem}
$\mathcal{S}$ only contains minimal SCC sets, and $\mathcal{S}$ is complete.
\label{thm:er-algo}
\end{theorem}
\begin{proof}
\textbf{Minimality :}
Inside every slice, the SCCs are totally independant one another. Moreover, given the order exploiting, we can deduce that the sum of the minima on each slice is the minimum on the whole graph.
\qed
\textbf{Completeness :}
Let $I$ be a minimal SCC set such as $x_{[x_J = y_J | J \in I]} \to^* y$, then, for every slice $C_i$, $I \cap C_i$ is minimal, since once all the SCCs in a slice can be changed to the way they are in $y$, we can always choose the path that allows this change. Hence $I \in \mathcal{S}$.
\qed
\end{proof}
\begin{theorem}
$\forall c \in ER(x,y),\, \exists I \in \mathcal{S},\, \forall u \in c,\, \exists scc \in I,\, u \in scc$.
\label{thm:er-scc}
\end{theorem}
\begin{proof}
Let $c$ be a vertex set in $ER(x,y)$ and $u$ one of the vertices. If $u \not \in \mathcal{O}$, then $c$ is not minimal, by Theorem \ref{theor:ancestrality}. If for all $I \in \mathcal{S}$, $u$ is in $o \in (\mathcal{O} \setminus I)$ then there exists a path such that changing $o$'s ancestors makes $o$'s change possible, and the ancestors need to be changed as well, by construction of $I$. So $c \setminus u$ would have the same effect, and $c$ would not be minimal. If $u \not \in o$, then there exists $I \in \mathcal{S}$ and $scc \in I$, such as $u \in scc$. \qed
\end{proof}
\subsection{SCC Filtering for Inevitable Reachability}
We now give an algorithm to identify a set of SCCs for which the mutation in the
initial fixed point is sufficient to ensure the \emph{Inevitable Reachability}
of the target fixed point.
The algorithm computes all reachable fixed points from $x$ with the SCCs in $\mathcal{S}$ modified, and find the one, $z$, that has the lower SCC (in the ranking given by $\prec$) in which a vertex $u$ is such that $z_u \neq y_u$. As we are looking for all reachable fixed points, this will always return the same SCC (even if the order is only partial), thus allowing the algorithm to be deterministic. We add this SCC to $\mathcal{S}$, and repeat until $y$ is the only reachable fixed point.
\begin{enumerate}
\item $\mathcal{S} := \emptyset$
\item While $\exists z \in \operatorname{FP}(f),\,z \neq y,\, x_{[x_I = y_I | I \in \mathcal{S}]} \to^* z$
\begin{itemize}
\item $\mathcal{S} := \mathcal{S} \cup \{\mathcal{O}_i\}$, with\\
$i = min_{a \in \{1,..,k\}}(a \mid \exists z \in \operatorname{FP}(f),\, z_{\mathcal{O}_a} \neq y_{\mathcal{O}_a},\, x_{[x_I = y_I | I \in \mathcal{S}]} \to^* z) $
\end{itemize}
\end{enumerate}
If two (or more) SCCs $A$ and $B$ are such that they are differently ordered in two distinct orders, then $A$ has no influence on $B$ and neither has $B$ on $A$. Then, the algorithm will select both SCCs if neither are impacted by the previous changes, so the order does not matter.
\paragraph*{Existence of a solution and proof of correctness :}
A solution is to fix all the SCCs of the graph to their value in $y$. Since there exists a solution and the algorithm tests if $y$ is the only reachable point, and follows the order given by $\prec$, it will end and find a solution.
\paragraph*{Complexity :}
Computing all fixed points reachable is PSPACE-complete \cite{CHJPS14-CMSB}.
It is used $k$ times (number of SCCs) in the worst case.
\begin{example}
We apply the algorithm on the BN of Fig.\ref{fig:bn-ir}.
with $x=00000$ and $y=11011$.
Starting from $\mathcal{S} := \emptyset$, the
only reachable fixed point is $(0)00(0)(0)$ (the SCCs from $\mathcal{O}$ are parenthesized).
The smallest SCC $o$ such as $x_{[x_{\mathcal{S}}=y_{\mathcal{S}}], o} \neq y_o$ is $\mathcal{O}_1$, so
$\mathcal{S} := \emptyset \cup \{\mathcal{O}_1\} = \{\mathcal{O}_1\}$.
The reachable fixed points from $x_{[x_I = y_I | I \in \mathcal{S}]}=10000$ are now: $(1)10(1)(1)$ and $(1)10(0)(1)$.
The smallest SCC $o$ such that $x_{[x_{\mathcal{S}}=y_{\mathcal{S}}], o} \neq y_o$ is $\mathcal{O}_3$.
We set $\mathcal{S} := \mathcal{S} \cup \mathcal{O}_3 = \{\mathcal{O}_1, \mathcal{O}_3\}$
and we obtain that the only reachable fixed point from $x_{[x_I = y_I | I \in \mathcal{S}]}=10010$ is $(1)10(1)(1)$ which is $y$.
So the algorithm stops.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\node[draw, align=left] (text) at (-4,-1) {$f_1(x) = x_1 $ \\ $ f_2(x) = x_1 $ \\ $ f_3(x) = x_1 \wedge \neg x_2 $ \\ $ f_4(x) = x_3 \vee x_4$\\ $f_5(x) = x_2 \vee x_5 $} ;
\node[main node] (a) {1} ;
\node[main node] (b) [below right of=a] {2} ;
\node[main node] (c) [below left of=a] {3} ;
\node[main node] (d) [below of=c] {4} ;
\node[main node] (e) [below of=b] {5} ;
\node[draw, align=left, minimum width=0.7cm, minimum height=0.7cm] at (0,0) [magenta, label={[magenta]right:$\mathcal{O}_1$}] {} ;
\node[draw, align=left, minimum width=0.7cm, minimum height=0.7cm] at (-0.7,-1.7) [magenta, label={[magenta]left:$\mathcal{O}_3$}] {} ;
\node[draw, align=left, minimum width=0.7cm, minimum height=0.7cm] at (0.7,-1.7) [magenta, label={[magenta]right:$\mathcal{O}_2$}] {} ;
\path (a) edge[blue, loop above] (a)
(a) edge [blue, ->] (b)
(a) edge [blue, ->] (c)
(b) edge[red, -|] (c)
(c) edge[blue, ->] (d)
(b) edge[blue, ->] (e)
(e) edge[blue, loop below] (e)
(d) edge[blue, loop below] (d) ;
\end{tikzpicture}
\end{center}
\caption{BN of dimension 5 (left) with its interaction graph (right) on which
the SCCs containing positive cycles ($\mathcal{O}$) are boxed.
\label{fig:bn-ir}}
\end{figure}
\end{example}
\begin{theorem}
$\mathcal{S}$ is minimal.
\end{theorem}
\begin{proof}
If a set $S_1$ exists such that $S_1$ has a lower cardinal than $\mathcal{S}$ and modifying $S_1$ makes $y$ the only reachable point, then we can reduce $S_1$ to a subset of $\mathcal{S}$. Let $s$ be a SCC in $\mathcal{S} \setminus S_1$, thus there exists a fixed point $z$ such that $z_s \neq y_s$ and by construction of $\mathcal{S}$, $z$ is reachable from $x$ modified by $S_1$. \qed
\end{proof}
We remark that, contrary to the case of Existential Reachability, the RDs for
Inevitable Reachability of the target fixed point are not necessarily in SCCs
containing positive cycles.
Indeed, in Ex.\ref{ex:reprog}, we showed that $IR_F(x,y)$ can refer to nodes
that do not belong to $\mathcal{O}$ (such as the node $2$ for the BN of
Fig.\ref{fig:sbn}). But we can also remark that if a RD $v$ is not in a SCC containing a positive cycle, then $x_v = y_v$.
\begin{theorem}
$\forall v \in IR(x,y),\, x_v \neq y_v \Rightarrow \exists scc \in \mathcal{O},\, v \in scc$
\end{theorem}
\begin{proof}
Let $v \in IR(x,y)$ such that for all $scc \in \mathcal{O}$, $v \not \in scc$. By Theor.\ref{thm:thomas}, if $v$ is such that $x_v \neq y_v$, then modifying the SCC in $\mathcal{O}$ is enough to modify $v$. But $v \in IR(x,y)$ and $IR(x,y)$ is minimal, thus $x_v = y_v$.
\qed
\end{proof}
\section{Identifying Determinants within SCCs}
\label{sec:enumeration}
We know that modifying all the SCCs selected is enough to switch from $x$ to $y$, but to reduce the genes selected, we could try to modify only some of the vertices to achieve the same result. But, as dynamics are involved, there could be unwanted changes (or wanted and unpredicted changes, in the case where we want $y$ to be reachable) in the descendants.
An idea could be to select the feedback vertex set of the SCC : by fixing the
vertices from this set, we effectively destroy every circle, thus the only
reachable state of the SCC is the one having the same values as $y$. This, however, does not solve the problem : in Ex.\ref{ex:dynamicsPb}, $\{1\}$ is the feedback vertex set, and we still have the same issue. Moreover, it miss some of the possible solutions (modifying $\{2\}$ or $\{3\}$ could work to change the whole SCC in Ex.\ref{ex:dynamicsPb}) or even dismiss the best solution (in Ex.\ref{ex:dynamicsPb}, changing $\{3\}$ makes $y$ the only reachable fixed point and solves the issue).
\begin{example}
Illustration of the problem with dynamics.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\node[draw, align=left] (text) at (-4,0) {$f_1(x) = \neg x_3 \wedge \neg x_2$ \\ $f_2(x)= \neg x_1$ \\ $f_3(x) = \neg x_1$ \\ $f_4(x) = x_2 \wedge \neg x_1 \wedge \neg x_3$ \\ $f_5(x) = x_4 \vee x_5$} ;
\node[main node] (a) {1} ;
\node[main node] (b) [above right of=a] {2} ;
\node[main node] (c) [below right of=a] {3} ;
\node[main node] (d) [right of=a] {4} ;
\node[main node] (e) [right of=d] {5} ;
\path (a) edge[red, bend left, -|] (b)
(b) edge [red, bend left, -|] (a)
(a) edge[red, bend left, -|] (c)
(c) edge [red, bend left, -|] (a)
(a) edge [red, -|] (d)
(b) edge[blue, ->] (d)
(c) edge[red, -|] (d)
(d) edge[blue, ->] (e)
(e) edge[blue, loop right] (e) ;
\end{tikzpicture}
\end{center}
\caption{BN of dimension 5 (left) and its interaction graph (right)
\label{fig:bn-complex}}
\end{figure}
We decide that $x=10000$ and $y=01100$, and $01101=z$, those are all fixed points. Let's suppose we want $y$ to be the only fixed point reachable. The algorithm will see that if the whole first SCC is modified $\{1,2,3\}$, $y$ is the only reachable fixed point.
It could pick $\{1\}$ to be modified, but instead of $00000 \rightarrow 00100 \rightarrow 01100$, we can have $$00000 \rightarrow 01000 \rightarrow 01010 \rightarrow 01011 \rightarrow 01111 \rightarrow 01101$$
This leads to $z$ being reachable by only modifying $\{1\}$, and so the algorithm would be wrong.
\label{ex:dynamicsPb}
\end{example}
This leaves to two kinds of approaches : either a way to modify the SCC so that it does not impact its descendants can be found, either we need to select the vertices to be modified in the SCC as a intermediary step in the process, and redesign $\mathcal{S}$ as a list of vertices instead of a list of SCCs.
By exploiting the results of the preceding section, we show an algorithm to
compute a set of RDs which guarantees the Inevitable Reachability of the target
fixed point.
The algorithm recursively picks a vertex $u$ in the lowest SCC in the order given by $\prec$ in $\mathcal{O}$, and
modify its associated function to become the constant value $y_u$.
The interaction graph of the resulting Boolean network is a sub-graph of the
initial interaction graph, where all the input edges of the node $u$ have been
removed.
Hence, the SCC $\mathcal{O}_1$ is split in the new interaction graph.
If necessary, another vertex can be picked in the lowest SCC in the new
interaction graph:
\medskip
\noindent
RecursiveAlgorithm($f$, $rd$) :
\begin{itemize}
\item If $\exists z \in \operatorname{FP}(f),\, x_{[x_{rd} = y_{rd}]} \to^* z$ then :
\begin{itemize}
\item $res=\emptyset$
\item
$i = min_{a \in \{1,..,k\}}(a \mid \exists z \in \operatorname{FP}(f),\, z_{\mathcal{O}_a} \neq y_{\mathcal{O}_a},\, x_{[x_I = y_I | I \in \mathcal{S}]} \to^* z) $
\item For all $u \in \mathcal{O}_i$ :
\begin{itemize}
\item $g := f$ with $g_u := y_u$
\item $res := res \cup$ RecursiveAlgorithm($g$, $rd \bar{\times} \{u\}$)
\end{itemize}
\item return $res$
\end{itemize}
\item else :
\begin{itemize}
\item return $rd$
\end{itemize}
\end{itemize}
Remark that the algorithm always find at least one solution: if the target fixed
point is not the only reachable fixed point, then there is at
least one positive cycle (and hence a SCC) which has a different state (and
hence will be selected by our algorithm).
\begin{example}
Applied to the BN of Fig.\ref{fig:bn-complex} with $x=10000$ and $y=01100$, the
above algorithm returns, for instance, the RD $\{2,5\}$:
indeed, $\{2\}$ belongs to $\mathcal{O}_1$.
When fixing $f_2=1$, the new interaction graph has two SCCs with positive
cycles: $\{1,3\}$ and $\{5\}$.
From the state $11000$, two fixed points are reachable:
$01100$, $01101$.
Hence, because the SCC $\{1,3\}$ has the same values than in $y$ in those two fixed
points,
the next vertex in picked in the SCC $\{5\}$.
Finally, from the state $11001$, $y$ is the only reachable fixed point.
\end{example}
\section{Discussion}
\label{sec:discussion}
This paper provides the first formal characterization of the Reprogramming Determinants (RDs) for
switching from one fixed point to another in the scope of the asynchronous dynamics of Boolean
networks.
In the case of reprogramming with existential reachability of the target fixed point, we
prove that all the possible minimal RDs modify nodes in particular combinations
of SCCs of the interaction graph of the Boolean network.
We give an algorithm to determine exactly those combinations of set of nodes.
Our characterizations rely on the verification of reachability properties.
In the case of reprogramming with inevitable reachability of the target fixed
point, we show that the RDs are not necessarily in SCCs.
However, we provide an algorithm which identifies RDs that guarantee the
inevitable reachability by picking nodes in appropriate SCCs.
The algorithm relies on the enumeration of reachable fixed points.
One of the main limitation of our algorithms is the numerous reachability checks it needs to perform.
Future work will consider methods and data structures for factorizing the exploration of the Boolean
network dynamics.
The present work considered only permanent mutations: when a node is mutated, it is assumed it
keeps its mutated value forever (its local Boolean function becomes a constant function).
Considering temporary mutations, i.e., where the local Boolean function of mutated nodes is restored
after some time, is a challenging research direction:
one should determine the ordering and the duration of mutations, and the set of candidate mutations
is \emph{a priori} no longer restricted to connected components, as it is the case for permanent
mutations.
\bibliographystyle{plain}
|
\section{Introduction}
Automatic vanishing point (VP) and horizon line detection are two of
the most fundamental problems in geometric computer
vision~\cite{barnard1983interpreting,magee1984determining}. Knowledge
of these quantities is the foundation for many higher level
tasks, including image mensuration~\cite{criminisi2000single},
facade detection~\cite{liulocal2014},
geolocalization~\cite{baatz2010handling,workman2014rainbow}, and
camera
calibration~\cite{autorecovery2000,grammatikopoulos2007automatic,
jacobs13cloudcalibration,videocampass2002}. Recent work in this area
\cite{global2013,wildenauer2012,kitware2013} has explored novel
problem formulations that significantly increase robustness to noise.
\begin{figure}
\centering
\includegraphics[width=.48\linewidth]{cover_hybrid_sampling}
\includegraphics[width=.48\linewidth]{cover_segment_clusters}
\caption{An example result of our method. (left) Horizon line
candidates, colored by their scores (red means high score), and
the true horizon line (green dash). (right) The horizon line
(magenta) estimated by our algorithm is very close to the true
horizon line (green dash). Line segments are color coded based on
the most consistent detected vanishing point.}
\label{fig:cover}
\end{figure}
A vanishing point results from the intersection of projections of a
set of parallel lines in the world. In man-made environments, such
sets of lines are often caused by the edges of buildings, roads, and
signs. VPs can typically be classified as either vertical, there is
one such VP, and horizontal, there are often many such VPs. Given a
set of horizontal VPs, there are numerous methods to estimate the
horizon line. Therefore, previous approaches to this problem focus on
first detecting the vanishing points, which is a challenging problem
in many images due to line segment intersections that are not true
VPs.
Our approach is to propose candidate horizon lines, score them, and
keep the best (\figref{cover}). We use a deep convolutional neural network to extract
global image context and guide the generation of a set of horizon line
candidates. For each candidate, we identify vanishing points by
solving a discrete-continuous optimization problem. The final score for
each candidate line is based on the consistency of the lines in the
image with the selected vanishing points.
This seemingly simple shift in approach leads to the need for novel
algorithms and has excellent performance. We evaluated the proposed
approach on two standard benchmark datasets, the Eurasian Cities
Dataset~\cite{geoparser2010} and the York Urban
Dataset~\cite{edgebased2008}. To our knowledge, our approach has the
current best performance on both datasets. To evaluate our algorithm
further, we also compare with the previous state-of-the-art method
(Lezama et al.~\cite{alignment2014}) on a recently introduced
dataset~\cite{authors2016deephorizon}; the results shows that our
method is more accurate and much faster.
\begin{figure*}
\centering
\includegraphics[width=0.95\linewidth]{algorithm_pipeline}
\caption{Algorithm overview: 1) use global image context to estimate
a prior over horizon lines (\secref{deeplearning}); 2) extract
line segments; 3) identify the zenith VP
(\secref{zenithdetection}); 4) sample horizon line candidates
consistent with the zenith VP (\secref{horizondetection}); 5) find
VPs on horizon line candidates (\secref{horizondetection}); and 6)
select the best horizon line based on the VPs it contains
(\secref{horizonscoring}).}
\label{fig:pipeline}
\end{figure*}
The main contributions of this work
are: 1) a novel method for horizon line/vanishing point detection,
which uses global image context to guide precise geometric analysis;
2) a strategy for quickly extracting this context, in the form of
constraints on possible horizon lines, using a deep convolutional
neural network; 3) a discrete-continuous method for scoring horizon
line candidates; and 4) an evaluation of the proposed approach on
three benchmark datasets, which highlights that our method is both
fast and accurate.
\subsection{Related Work}
\label{sec:relatedwork}
Vanishing points and the horizon line provide a strong
characterization of geometric scene structure and as such have been
intensely studied for
decades~\cite{barnard1983interpreting,magee1984determining}. For
example, Hoiem et al.~\cite{hoiem2008putting} show how the horizon
line improves the accuracy of object detection. A wide variety of
methods have been introduced to estimate these quantities. We provide
a brief overview of the main approaches, refer
to~\cite{szeliski2010computer} for a comprehensive review.
Two distinct categories of methods exist, distinguished by the
features they use. The first group of methods~\cite{geoparser2010,
manhattanbayesian1999, edgebased2008,atlanta2004} operate directly
on lower-level features, such as edge pixels or image gradients. The
second group of methods~\cite{nopriori2003, edgebased2008,
alignment2014, houghvp1994, selfsketch2012, wildenauer2012,
kitware2013} build on top of the closely related problem of line
segment detection. Our work is most closely related to the latter
category, so we focus our discussion towards them.
The dominant approach to vanishing point detection from line segments
is to cluster the line segments that pass through the same location.
Various methods of clustering have been explored, including
RANSAC~\cite{ransac1981}, J-linkage~\cite{tardif2009}, and the Hough
transform~\cite{hough1959}. Once the line segments have been
clustered, vanishing points can be estimated using one of many
refinement procedures~\cite{alignment2014,atlanta2004,tardif2009,
wildenauer2012,kitware2013}.
These procedures typically minimize a
nonlinear objective function. An important distinction between such
methods is the choice of point and line representation and error
metric. Collins and Weiss~\cite{unitsphere1990} formulate vanishing
point detection as a statistical estimation problem on the Gaussian
Sphere, which is similar to the geometry we use. More recent work
has explored the use of dual space~\cite{alignment2014,dualspace2013}
representations. Among the clustering-based approaches, Xu et
al.~\cite{kitware2013} improve this pipeline by introducing a new
point-line consistency function that models errors in the line segment
extraction step.
Alternatives to clustering-based approaches have been explored.
For example, vanishing point detection from line segments has been
modeled as an Uncapacitated Facility Location (UFL)
problem~\cite{global2013, selfsketch2012}. To avoid error accumulation
issues encountered by a step-by-step pipeline method, Barinova et
al.~\cite{geoparser2010} solve the problem in a unified framework,
where edges, lines, and vanishing points fit into a single graphical
model.
Our approach is motivated by the fact that properties of the scene,
including objects, can provide additional cues for vanishing point and
horizon line placement than line segments alone. Unlike existing
methods that use J-linkage~\cite{tardif2009, kitware2013} or similar
techniques to find an initial set of VPs by clustering detected lines
followed by a refinement step, our approach first proposes candidate
horizon lines using global image context.
\subsection{Approach Overview}
Our approach is motivated by two observations: 1) traditional purely
geometric approaches to vanishing point detection often fail in
seemingly nonsensical ways and 2) identifying the true vanishing
points for many scenes is challenging and computationally expensive
due to the large number of outlier line segments. Driven by these
observations, we propose a two part strategy. First, we use global
image context to estimate priors over the horizon line and the zenith
vanishing point (\secref{deeplearning}). Using these priors, we
introduce a novel VP detection method (\secref{precise}) that samples
horizon lines from the prior and performs a fast one-dimensional
search for high-quality vanishing points in each. Both steps are
essential for accurate results: the prior helps ensure a good
initialization such that our horizon-first detection method may obtain
very precise estimates that are necessary for many scene understanding
tasks. See \figref{pipeline} for an overview of our algorithm.
\section{Problem Formulation}
\label{sec:problem}
The goal of this work is to detect the horizon line, the zenith
vanishing point, and any horizontal vanishing points from a single
image. The remainder of this section defines the notation and basic
geometric facts that we will use throughout. For clarity we use
unbolded letters for points in world coordinates or the image plane
and bolded letters for points or lines in homogeneous coordinates. We
primarily follow the notation convention of Vedaldi and
Zisserman~\cite{selfsketch2012}.
Given a point $(u,v)$ in the image plane, its homogeneous coordinate
with respect to the calibrated image plane is denoted by:
\begin{displaymath}
\mathbf{p} = [\rho(u-c_u), \rho(v-c_v), 1]^\mathsf{T} / \Sigma \; ,
\end{displaymath}
where $\rho$ is a scale constant, $(c_u, c_v)$ is the camera principal
point in the image frame, which we assume to be the center of the
image, and $\Sigma$ is the constant that makes $\mathbf{p}$ a unit
vector.
In homogeneous coordinates, both lines and points are represented as
three-dimensional vectors (\figref{homoIntersection}). Computing the
line, $\mathbf{l}$, that passes through two points, $(\mathbf{p}_1, \mathbf{p}_2)$, and
the point, $\mathbf{p}$, at the intersection of two lines, $(\mathbf{l}_1,
\mathbf{l}_2)$, are defined as follows:
\begin{align}
\mathbf{l} = \frac{\mathbf{p}_1 \times \mathbf{p}_2}{\|\mathbf{p}_1 \times
\mathbf{p}_2\|}
&&
\mathbf{p} = \frac{\mathbf{l}_1 \times \mathbf{l}_2}{\|\mathbf{l}_1 \times \mathbf{l}_2\|} \; .
\end{align}
We denote the smallest angle between two vectors $\mathbf{x}$ and
$\mathbf{y}$ with $\Theta_{\mathbf{x},\mathbf{y}}
= |cos^{-1}(\mathbf{x}^\mathsf{T} \mathbf{y})|$.
We use this to define the consistency between a line, $\mathbf{l}$, and a
point,
$\mathbf{p}$, as: $f_c(\mathbf{p}, \mathbf{l}) =
\operatorname*{max}(\theta_{con} - \Theta_{\mathbf{p},\mathbf{l}}, 0)$.
The maximum value of consistency between a vanishing point and a line
segment is $\theta_{con}$. This will occur if it is possible to extend
the line segment to contain the vanishing point.
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{homoIntersection}
\caption{In homogeneous coordinates, lines (red lines) are defined
by the normal (red arrow) of the plane (red triangle) they form
with the origin (green dot). Two lines form a great circle (blue
circle), whose normal (blue arrow) is their common point (blue
dot) in homogeneous coordinates.}
\label{fig:homoIntersection}
\end{figure}
\section{Horizon Priors from Global Image Context}
\label{sec:deeplearning}
Recent studies show that deep convolutional neural networks (CNNs) are
adaptable for a wide variety of tasks~\cite{yosinski2014transferable},
and are quite fast in practice. We propose to use a CNN to extract
global image context from a single image.
We parameterize the horizon line by its slope angle, $\alpha \in
[-\pi, \pi)$, and offset, $o \in [0, \inf)$, which is the shortest
distance between the horizon line and the principal point. In order to
span the entire horizon line parameter space, we ``squash'' $o$ from
pixel coordinates to the interval $[0, \pi/2)$, through a one-to-one
function, $w = tan^{-1}(o / \kappa)$, in which $\kappa$ is a
scaling factor that affects how dense the sampling is near the center
of the image.
\begin{figure*}
\hfill
\includegraphics[width=0.19\linewidth]{dataset/im1}
\hfill
\includegraphics[width=0.19\linewidth]{dataset/im2}
\hfill
\includegraphics[width=0.19\linewidth]{dataset/im3}
\hfill
\includegraphics[width=0.19\linewidth]{dataset/im4}
\hfill
\includegraphics[width=0.19\linewidth]{dataset/im5}
\hfill
\caption{Example images from our training dataset
(\secref{dataset}), each overlaid with the ground-truth horizon line.
}
\label{fig:dataset}
\end{figure*}
\subsection{Network Architecture}
For our task, we adapt the popular
AlexNet~\cite{krizhevsky2012imagenet} architecture, which was designed
for object recognition as part of the ImageNet ILSVRC-2012
challenge~\cite{russakovsky2015imagenet}. It consists of five
convolutional layers, each followed by a non-linearity (rectified
linear unit), and occasionally interspersed with pooling and local
response normalization. This is followed by three fully connected
layers (referred to as `fc6', `fc7', and `fc8'). A softmax is applied
to the final output layer to produce a categorical distribution over
1000 object classes. We use this as a foundation to create a CNN that
simultaneously generates a categorical distribution for each
horizon-line parameter.
We modify the original AlexNet architecture in the following way: The
first five convolutional layers are left unmodified. These layers are
initialized with weights from a network trained for object detection
and scene classification~\cite{zhou2014places}. We remove the original
fully connected layers (`fc6'--`fc8') and add two disjoint sets of
fully connected layers (`fc6$\alpha$'--`fc8$\alpha$' and
`fc6$w$'--`fc8$w$'), one for each target label, $\alpha$ and $w$. We
convert the slope, $\alpha$, and the squashed offset, $w$, into
independent categorical labels by uniformly dividing their respective
domains into 500 bins. We randomly initialize the weights for these
new layers.
We train our network using stochastic gradient descent, with a
multinomial logistic loss function. The learning rates for the
convolutional layers are progressively increased such that the latter
layers change more. The new fully connected layers are given full
learning rate.
\subsection{Training Database}
\label{sec:dataset}
To support training our model of global image context, we construct a
large dataset of images with known horizon lines. We make use of
equirectangular panoramas downloaded from Google Street View in large
metropolitan cities around the world. We identified a set of cities
based on population and Street View coverage. From each city, we
downloaded panoramas randomly sampled in a $5km \times 5km$ region
around the city center. This resulted in $11\,001$ panoramas from 93
cities. Example cities include New York, Rio de Janeiro, London, and
Melbourne.
We extracted 10 perspective images from each panorama with randomly
sampled horizontal field-of-view (FOV), yaw, pitch, and roll. Here yaw
is relative to the Google Street View capture vehicle. We sampled
horizontal FOV from a normal distribution with $\mu=60^\circ$ and
$\sigma=10^\circ$. Similarly, pitch and roll are sampled from normal
distributions with $\mu=0^\circ$ and $\sigma=10^\circ$ and
$\sigma=5^\circ$, respectively. Yaw is sampled uniformly. We truncate
these distributions such that horizontal FOV $\in [40^\circ,
80^\circ]$, pitch $\in [-30^\circ, 30^\circ]$, and roll $\in
[-20^\circ, 20^\circ]$. These settings were selected empirically to
match the distribution of images captured by casual photographers in
the wild.
Given the FOV, pitch, and roll of a generated perspective image, it is
straightforward to compute the horizon line position in image space.
In total, our training database contains $110\,010$ images with known
horizon line. \figref{dataset} shows several example images from
our dataset annotated with the ground-truth horizon line.
\subsection{Making the Output Continuous}
Given an image, $I$, the network outputs a categorical probability
distribution for the slope, $\alpha$, and squashed offset, $w$. We
make these distributions continuous by approximating them with a
Gaussian distribution. For each, we estimate the mean and variance
from $5\,000$ samples generated from the categorical probability
distribution. Since the relationship between $w$ and $o$ is
one-to-one, this also results in a continuous distribution over
$o$. The resulting distributions, $p(\alpha \vert I)$ and $p(o \vert
I)$, are used in the next step of our approach to aid in detecting the
zenith VP and as a prior for sampling candidate horizon
lines. To visualize this distribution we observe that the horizon line
can be uniquely defined by the point on the line closest to the
principal point. Therefore, we can visualize a horizon line
distribution as a distribution over points in the image.
\figref{horpdf} shows this distribution for two images.
\begin{figure}
\centering
\includegraphics[width=.48\linewidth]{4horpdf}
\includegraphics[width=.48\linewidth]{68horpdf}
\caption{Global image context imposes a strong prior on horizon line
location. The output of our CNN is visualized as an overlaid
heatmap, with red indicating more likely locations. For each image,
the ground-truth horizon line (dash green) and the line that
maximizes the prior (red) are shown.}
\label{fig:horpdf}
\end{figure}
\section{Horizon-First Vanishing Point Detection}
\label{sec:precise}
We propose an approach to obtain accurate estimates of the horizon
line, the zenith vanishing point, and one or more horizontal vanishing
points. Given an image, our approach makes use of the distributions
estimated from global image context (\secref{deeplearning}) and
line segments extracted with LSD~\cite{lsd2010}. The algorithm
consists of the following major steps:
\begin{compactenum}
\item detect the zenith vanishing point
(\secref{zenithdetection})
\item \label{alg:findhorvp} detect horizontal vanishing points on
horizon line candidates
(\secref{horizondetection})
\item \label{alg:scorehorizon} score horizon line candidates with
horizontal vanishing points (\secref{horizonscoring})
\end{compactenum}
The remainder of this section provides details for each of these steps.
\subsection{Detecting the Zenith Vanishing Point}
\label{sec:zenithdetection}
To detect the zenith vanishing point, we first select an initial set
of line segments using the zenith direction,
$\mathbf{l}_\mathbf{z}$, from the global image context, then use
the RANSAC~\cite{ransac1981} algorithm to refine it. The zenith direction
is the line connecting the principal point and the zenith vanishing
point, which is uniquely determined by the horizon line slope (see
supplemental material for a proof).
We compute our initial estimate of $\mathbf{l}_{\mathbf{z}}$ using the global
image context by choosing the value that maximizes the posterior:
$\hat{\alpha} = \operatorname*{arg\,max}_\alpha p(\alpha \vert I)$.
To handle the presence of outlier line segments, we first select a set
of candidate vertical line segments as the RANSAC inputs by
thresholding the angle between each line segment and the estimated
zenith direction, $\Theta_{\mathbf{l},\mathbf{l}_{\mathbf{z}}} < \theta_{ver}$. For a
randomly sampled pair of line segments with intersection, $\mathbf{p}$, we
compute the set of inlier line segments, $\{\mathbf{l} \mid f_c(\mathbf{p}, \mathbf{l})
> 0\}$. If the largest set of inliers has a sufficient portion (more
than 2\% of candidate line segments), we obtain the final estimate of
the zenith vanishing point, $\mathbf{z}$, by minimizing the algebraic
distance, $\|\mathbf{l}^\mathsf{T} \mathbf{p}\|$ using singular value decomposition
(SVD), and update the zenith direction, $\mathbf{l}_{\mathbf{z}}$. Otherwise, we
keep the zenith direction estimated from the global image context.
\subsection{Detecting Horizontal Vanishing Points}
\label{sec:horizondetection}
We start with sampling a set of horizon line candidates,
$\{\mathbf{h}_i\}_1^S$, that are perpendicular to $\mathbf{l}_{\mathbf{z}}$ in the image
space, under the distribution of horizon line offsets, $p(o | I)$.
See \figref{horsampling} for examples of horizon line sampling with
and without global context.
For each horizon line candidate, we identify a set of horizontal VPs
by selecting points along the horizon line where many line segments
intersect. We assume that for the true horizon line the identified
horizontal VPs will be close to many intersection points and that
these intersections will be more tightly clustered than for
non-horizon lines. We use this intuition to define a scoring function
for horizon line candidates.
As a preprocessing step, given the zenith direction, $\mathbf{l}_{\mathbf{z}}$,
and a horizon line candidate, $\mathbf{h}$, we filter out nearly vertical
line segments ($\Theta_{\mathbf{l},\mathbf{l}_{\mathbf{z}}} < \theta_{ver}$), which are
likely associated with the zenith vanishing point, and nearly
horizontal line segments ($\Theta_{\mathbf{l},\mathbf{h}} < \theta_{hor}$), which
result in noisy horizon line intersection points. We remove such
lines from consideration because they lead to spurious, or
uninformative, vanishing points, which decreases accuracy.
Given a horizon line candidate, $\mathbf{h}$, and the filtered line segments
in homogeneous coordinates, $\mathcal{L} = \{\mathbf{l}_i\}$, we select
a set of horizontal VPs, $\mathcal{P} = \{\mathbf{p}_i\}$, by minimizing
the following objective function:
\begin{align}
g(\mathcal{P} \vert \mathbf{h}, \mathcal{L}) & = -\sum_{\mathbf{p}_i
\in \mathcal{P}} {\sum_{\mathbf{l}_j \in
\mathcal{L}}f_c(\mathbf{p}_i, \mathbf{l}_j)}
\label{eq:obj} \\
\text{subject to:} \nonumber \\
& \Theta_{\mathbf{p}_i,\mathbf{p}_j} > \theta_{dist}
\; \text{and} \; \left< \mathbf{p}_i,\mathbf{h} \right>=0,
\; \forall (i,j) \nonumber \; .
\end{align}
The constraint prevents two vanishing points from being too close
together, which eliminates the possibility of selecting multiple
vanishing points in the same location.
We propose the following combinatorial optimization process for
obtaining an initial set of vanishing points, followed by a
constrained nonlinear optimization to refine the vanishing points.
\begin{figure}
\centering
\includegraphics[width=.4075\linewidth]{horizon_sampling_uniform}
\includegraphics[width=.4075\linewidth]{horizon_sampling}
\includegraphics[width=.1630\linewidth]{offset_distribution}
\caption{Our method samples more horizon line candidates (red) near
the ground truth (green dash) with (middle) global image context
than without (left). In the case of sampling with global image
context, the offset PDF, $p(o|I)$ (blue curve), is fit from the
CNN categorical probability distribution outputs (hollow
bins). For clarity, we only show a reduced number of horizon line
candidates and bins.}
\label{fig:horsampling}
\end{figure}
\subsubsection{Initialization by Random Sampling and Discrete
Optimization}
\label{sec:initialization}
To choose an initial set of candidate vanishing points,
$\{\mathbf{p}_i\}_1^M$, we randomly select a subset of line segments,
$\{\mathbf{l}_i\}_1^M$, and compute their intersection with the horizon
line. We then construct a graph with a node for each vanishing point,
$\mathbf{p}_i$, each with weight $\sum_{\mathbf{l}_j \in \mathcal{L}}f_c(\mathbf{p}_i,
\mathbf{l}_j)$, which is larger if there are many line segments in the image
that are consistent with $\mathbf{p}_i$. Pairs of nodes, $(i,j)$, are
connected if the corresponding vanishing points, $\mathbf{p}_i, \mathbf{p}_j$, are
sufficiently close in homogeneous space ($\Theta_{\mathbf{p}_i,\mathbf{p}_j} \le
\theta_{dist}$).
From this randomly sampled set, we select an optimal subset of VPs by
maximizing the sum of weights, while ensuring no VPs in the final set
are too close. Therefore, the problem of choosing the initial set of
VPs reduces to a maximum weighted independent set problem, which is
NP-hard in general. Due to the nature of the constraints, the
resulting graph has a ring-like structure which means that, in
practice, the problem can be quickly solved. Our solver exploits this
sparse ring-like structure by finding a set of VPs that when removed
convert the ring-like graph into a set of nearly linear sub-graphs
(\figref{subset}). We solve each subproblem using dynamic
programming. The set of VPs with maximum weight, $\{\mathbf{p}_i\}_{opt}$,
is used as initialization for local refinement. Usually, 2--4 such
vanishing points are found near the horizon line ground truth.
\begin{figure}
\centering
\includegraphics[width=.92\linewidth]{independent_subset}
\caption{A ring-like graph (left) is converted into three nearly
linear subgraphs (right) by partitioning around a node with
minimal degree. For the subgraphs, the red node is mandatory, the
dashed nodes are excluded, and a subset of the solid nodes are
selected using dynamic programming.}
\label{fig:subset}
\end{figure}
\subsubsection{Vanishing Points Refinement}
\label{sec:refinement}
Since they were randomly sampled, the set of vanishing points
selected during initialization, $\{\mathbf{p}_i\}_{opt}$, may not be at the
optimal locations. We optimize their locations to further minimize
the objective function \equref{obj}. We perform an EM-like algorithm
to refine the vanishing point locations, subject to the constraint
that they lie on the horizon line:
\begin{compactitem}
\item \emph{E-step}: Given a vanishing point, $\mathbf{p}$, assign
line segments that have positive consistency with $\mathbf{p}$:
$\{\mathbf{l}|f_c(\mathbf{p},\mathbf{l}) > 0\}$.
\item \emph{M-step}: Given the assigned line segments as a matrix,
$\mathbf{L} = [\mathbf{l}_1, \mathbf{l}_2,\dots ,\mathbf{l}_n]$, and the horizon line,
$\mathbf{h}$, both represented in homogeneous coordinates, we
solve for a refined vanishing point, $\mathbf{p}^*$, by minimizing the algebraic distance, $\|
\mathbf{L}^\mathsf{T} \mathbf{p} \|$ such that $\mathbf{h}^\mathsf{T} \mathbf{p} = 0$. We define a
basis, $\mathbf{B}_\mathbf{h}$, for the null space of $\mathbf{h}$, and reformulate
the problem as $\mathbf{\lambda}^* = \operatorname*{arg\,min} \|
\mathbf{L}^\top \mathbf{B}_\mathbf{h} \mathbf{\lambda} \|$, which we solve using SVD.
Given the optimal coefficients, $\mathbf{\lambda}^*$, we reconstruct
the optimal vanishing point as: $\mathbf{p}^* =
\frac{\mathbf{B}_\mathbf{h}\mathbf{\lambda}^*}
{\|\mathbf{B}_\mathbf{h}\mathbf{\lambda}^*\|}$.
\end{compactitem}
We run this refinement iteration until convergence. In practice, this
converges quickly; we run at most three iterations for all the
experiments. The final set of optimized VPs is then used to assign a
score to the current horizon line candidate.
\subsection{Optimal Horizon Line Selection}
\label{sec:horizonscoring}
For each horizon line candidate, we assign a score based on the total
consistency of lines in the image with the VPs selected in the
previous section. The score of a horizon line candidate, $\mathbf{h}$, is
defined as:
\begin{equation}
score(\mathbf{h}) = \sum_{\{\tilde{\mathbf{p}}_i\}} \sum_{\mathbf{l}_j \in
\mathcal{L}}f_c(\tilde{\mathbf{p}}_i, \mathbf{l}_j) \; .
\end{equation}
To reduce the impact of false positive vanishing points, we select
from $\{\mathbf{p}_i\}_{opt}$ the two highest weighted vanishing points (or
one if $\{\mathbf{p}_i\}_{opt}$ contains only one element),
$\{\tilde{\mathbf{p}}_i\}$, for horizon line scoring.
\section{Evaluation}
\label{sec:evaluation}
We perform an extensive evaluation of our methods, both quantitatively
and qualitatively, on three benchmark datasets. The results show that
our method achieves state-of-the-art performance based on horizon-line
detection error, the standard criteria in recent work on VP
detection~\cite{geoparser2010,alignment2014,selfsketch2012,kitware2013}.
Horizon detection error is defined as the maximum distance from the
detected horizon line to the ground-truth horizon line, normalized by
the image height. Following tradition, we show the cumulative
histogram of these errors and report the area under the curve (AUC).
Our method is implemented using MATLAB, with the exception of
detecting line segments, which uses an existing C++
library~\cite{lsd2010}, and extracting global image context, which we
implemented using Caffe~\cite{jia2014caffe}. We use the parameters
defined in \tabref{parameters} for all experiments. This differs from
other methods which usually use different parameters for different
datasets.
\begin{table}[h]
\centering
\caption{
Algorithm parameters (given an $H\times W$ image).
}
\vspace{-.5em}
\begin{tabular}{|c|l|l|}
\hline
Name & Usage(s) & Value\\
\hline
$\theta_{con}$ & \secref{problem} & $2^\circ$\\
$\rho$ & \secref{problem} & $2/\operatorname{max}(H,W)$ \\
$\kappa$ & \secref{deeplearning} & $ 1/5 \times H$ \\
$\theta_{ver}$ & \secref{zenithdetection}, \secref{horizondetection} & $ \Theta_{\mathbf{l},\mathbf{l}_{\mathbf{z}}} < 10^\circ$ \\
$\theta_{hor}$ & \secref{horizondetection} & $\Theta_{\mathbf{l},\mathbf{h}} <
1.5^\circ$ \\
$S$ & \secref{horizondetection} & 300 candidates \\
$M$ & \secref{initialization} & 20 line segments \\
$\theta_{dist}$ & \secref{horizondetection}, \secref{initialization} & $
\Theta_{\mathbf{p}_i,\mathbf{p}_j} > 33^\circ$\\
\hline
\end{tabular}
\label{tab:parameters}
\end{table}
\begin{figure*}[]
\centering
\begin{subfigure}{.30\linewidth}
\centering
\includegraphics[width=\linewidth]{YUD_methods}
\subcaption{YUD}
\label{fig:horError_YUD}
\end{subfigure}
\hfill
\begin{subfigure}{.3\linewidth}
\centering
\includegraphics[width=\linewidth]{ECD_methods}
\subcaption{ECD}
\label{fig:horError_ECD}
\end{subfigure}
\hfill
\begin{subfigure}{.3\linewidth}
\centering
\includegraphics[width=\linewidth]{HLW_methods}
\subcaption{HLW}
\label{fig:hlw_err}
\end{subfigure}
\caption{For three benchmark datasets, the fraction of images
(y-axis) with a horizon error less than a threshold (x-axis). The
AUC for each curve is shown in the legend. For additional details see \secref{evaluation}.}
\label{fig:horError}
\end{figure*}
\subsection{Quantitative Evaluation}
The York Urban Dataset (YUD)~\cite{edgebased2008} is a commonly used dataset
for evaluating horizon line estimation methods. It contains 102 images
and ground-truth vanishing points. The scenes obey the Manhattan-world
assumption, however we do not take advantage of this assumption.
\figref{horError_YUD} shows the performance of our methods relative to
previous work on YUD. These results demonstrate that our method
achieves state-of-the-art AUC, improving upon the previous best of
Lezama et al.~\cite{alignment2014} by 0.28\%, a relative
improvement\footnote{We define the relative improvement as
$\frac{\text{AUC}_{new} - \text{AUC}_{old}}{1 - \text{AUC}_{old}}$.} of
5\%. This is especially impressive given that our method only requires
an average of 1 second per image, while Lezama et al.\ requires
approximately 30 seconds per image.
The Eurasian Cities Dataset (ECD)~\cite{geoparser2010} is another
commonly used benchmark dataset, which is considered challenging due
to the large number of outlier line segments and complex scene
geometries. It contains 103 images captured in urban areas and, unlike
the YUD dataset, not all images satisfy the Manhattan-world
assumption. It provides reliable horizon line ground truth and is
widely considered difficult for horizon line detection. To our
knowledge, the previous state-of-the-art performance in terms of the
AUC metric on this dataset was achieved by Lezama et
al.~\cite{alignment2014}. Our algorithm improves upon their
performance, increasing the state of the art to 90.8\%. This is a
significant relative improvement of 14.8\%, especially considering
their improvement relative to the state of the art was 0.5\%. On ECD,
our method takes an average of 3 seconds per image, while Lezama et
al.\ requires approximately 60 seconds per image. We present the
performance comparison with other methods in \figref{horError_ECD}.
The Horizon Lines in the Wild (HLW)
dataset~\cite{authors2016deephorizon} is a new, very challenging
benchmark dataset. We use the provided test set, which contains
approximately $2\,000$ images from diverse locations, with many images
not adhering to the Manhattan-world assumption. \figref{hlw_err}
compares our method with the method of Lezama et
al.~\cite{alignment2014} (the only publicly available implementation
from a recent method). Our method is significantly better, achieving
58.24\% versus 52.59\% AUC.
\begin{table}[t] \centering
\caption{Component error analysis (AUC).}
\vspace{-.5em}
\begin{tabular}{|l|c|c|c|}
\hline
Method & YUD & ECD & HLW \\
\hline
\hline
Lezama et al.~\cite{alignment2014} & 94.51\% & 89.20\% & 52.59\% \\
\hline
NONE+FULL & 93.87\% & 87.94\% & 53.04\% \\
\hline
GISTRF+EMPTY & 53.36\% & 32.69\% & 31.08\% \\
GISTRF+FULL & 94.66\% & 87.60\% & 54.95\% \\
\hline
CNN+EMPTY & 73.67\% & 67.64\% & 49.03\% \\
CNN+FULL (Ours) & {\bf 94.78\%} & {\bf 90.80\%} & {\bf 58.24\%} \\
\hline
\end{tabular}
\label{tab:contribution}
\end{table}
\subsection{Component Error Analysis}
\label{sec:analysis}
Our method consists of two major components: global context extraction
(\secref{deeplearning}) and horizon-first vanishing point detection
(\secref{precise}). This section provides an analysis of the impact
each component has on accuracy.
To evaluate the impact of global context extraction, we considered
three alternatives: our proposed approach (CNN), replacing the CNN
with a random forest (using the Python ``sklearn'' library with 25 trees) applied to a GIST~\cite{oliva2001modeling}
descriptor (GISTRF), and omitting context entirely
(NONE). When omitting the global context, we assume no camera
roll (horizon lines are horizontal in the image) and sample horizon
lines uniformly between $[-2H,2H]$ ($H$ is the image height). To
evaluate the impact of vanishing point detection, we considered two
alternatives: our proposed approach (FULL) and omitting the vanishing
point detection step (EMPTY). When omitting vanishing point detection,
we directly estimate the horizon line, $(\alpha, o)$, by maximizing
the posterior estimated by our global-context CNN, $p(\alpha, o | I)$.
Quantitative results presented in \tabref{contribution} show that both
components play important roles in the algorithm and that CNN provides
better global context information than GISTRF. Though our vanishing
point detection performs well by itself (see column NONE+FULL), global
image context helps improve the accuracy further. \figref{hlw_err}
visualizes these results as a cumulative histogram of horizon error
on HLW. To illustrate the impact of global image context, we present
two examples in \figref{iscontext} that compare horizon line
estimates obtained using global context (CNN+FULL) and without
(NONE+FULL). When using global context, the estimated horizon lines
are very close to the ground truth. Without, the estimates obtained
are implausible, even resulting in an estimate that is off the image.
\begin{figure}
\centering
\includegraphics[width=.48\linewidth]{cnn_vs_nocnn_ecd_hybrid}
\includegraphics[width=.48\linewidth]{cnn_vs_nocnn_ecd_uniform}
%
\includegraphics[width=.48\linewidth]{cnn_vs_nocnn_yud_hybrid}
\includegraphics[width=.48\linewidth]{cnn_vs_nocnn_yud_uniform}
\caption{Two images where horizon line estimates are much better with global context (left) than without (right).}
\label{fig:iscontext}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.1571\linewidth]{montage_ecd5}
\includegraphics[width=0.0884\linewidth]{montage_ecd1}
\includegraphics[width=0.1572\linewidth]{montage_yud2}
\includegraphics[width=0.1571\linewidth]{montage_ecd4}
\includegraphics[width=0.1572\linewidth]{montage_yud3}
\includegraphics[width=0.0884\linewidth]{montage_ecd6}
\includegraphics[width=0.1571\linewidth]{montage_ecd8}
\includegraphics[width=0.1375\linewidth]{montage_ecd5_homo}
\includegraphics[width=0.1375\linewidth]{montage_ecd1_homo}
\includegraphics[width=0.1375\linewidth]{montage_yud2_homo}
\includegraphics[width=0.1375\linewidth]{montage_ecd4_homo}
\includegraphics[width=0.1375\linewidth]{montage_yud3_homo}
\includegraphics[width=0.1375\linewidth]{montage_ecd6_homo}
\includegraphics[width=0.1375\linewidth]{montage_ecd8_homo}
\includegraphics[width=0.1540\linewidth]{montage_yud1}
\includegraphics[width=0.0867\linewidth]{montage_ecd3}
\includegraphics[width=0.1540\linewidth]{montage_yud5}
\includegraphics[width=0.1540\linewidth]{montage_yud4}
\includegraphics[width=0.1731\linewidth]{montage_ecd2}
\includegraphics[width=0.1540\linewidth]{montage_yud6}
\includegraphics[width=0.0867\linewidth]{montage_ecd7}
\includegraphics[width=0.1375\linewidth]{montage_yud1_homo}
\includegraphics[width=0.1375\linewidth]{montage_ecd3_homo}
\includegraphics[width=0.1375\linewidth]{montage_yud5_homo}
\includegraphics[width=0.1375\linewidth]{montage_yud4_homo}
\includegraphics[width=0.1375\linewidth]{montage_ecd2_homo}
\includegraphics[width=0.1375\linewidth]{montage_yud6_homo}
\includegraphics[width=0.1375\linewidth]{montage_ecd7_homo}
\caption{Example results produced by our method. (rows 1 and 3)
Line segments color coded based on the most consistent VP, the
ground-truth (green dash), and detected horizon lines (magenta).
For clarity only the top two horizontal VPs are shown. (rows 2
and 4) The line segments (dots) and their VPs (rings) represented
in homogeneous coordinates. (last column) Two failure cases of our
method, caused by irregularly shaped objects (bottom) and short
edges (top).}
\label{fig:examples}
\end{figure*}
\subsection{Failure Cases}
We highlight two representative failure cases in the last column of
\figref{examples}. The top image fails due to the propagation of
measurement errors from the short line segments. The bottom image is
challenging because the curved structures lead to indistinct VPs.
Despite this, global context helps our method produce plausible
results, while other methods (\eg, \cite{geoparser2010}) fail
dramatically.
\section{Conclusion}
We presented a novel vanishing point detection algorithm that obtains
state-of-the-art performance on three benchmark datasets. The main
innovation in our method is the use of global image context to sample
possible horizon lines, followed by a novel discrete-continuous
procedure to score each horizon line by choosing the optimal vanishing
points for the line. Our method is both more accurate and more
efficient than the previous state-of-the-art algorithm, requiring no
parameter tuning for a new testing dataset, which is common in other
methods.
\ifcvprfinal
\section*{Acknowledgements}
We gratefully acknowledge the support of DARPA (contract CSSG
D11AP00255). The U.S.\ Government is authorized to reproduce and
distribute reprints for Governmental purposes notwithstanding any
copyright annotation thereon. Disclaimer: The views and conclusions
contained herein are those of the authors and should not be
interpreted as necessarily representing the official policies or
endorsements, either expressed or implied, of DARPA or the U.S.\
Government.
\fi
{\small
\bibliographystyle{ieee}
|
\section{Introduction}
The purpose of this paper is to unravel some of the algebraic structure underlying combinatorial games. We show that each simplicial complex is the legal complex of some invariant strong placement game (iSP-game) and board. One implication is that in most situations when studying strong placement games (SP-games) it is enough to consider those with invariance. These results will for example make it easier to study whether each game value under normal play can be achieved by an SP-game, which would affect the study of combinatorial games in general.
In \cite{FHN14a} we initiated the idea of using simplicial complexes to algebraically describe SP-games, a class of combinatorial games. To each SP-game we can assign two simplicial complexes, one representing all legal positions, the so called \textbf{legal complex}, and one representing the minimal illegal positions, the \textbf{illegal complex}. One of the main questions is what complexes appear as game complexes. In Proposition \ref{thm:legalillegal} we show that every simplicial complex is both a legal and an illegal complex of some SP-game and board. The rulesets of these games can be quite complex though and depend highly on the board on which the game is being played. Thus we introduce \textbf{invariance} for SP-games, which, in a sense, forces rulesets to be uniform. Invariance is a concept that was introduced for subtraction games (see for example \cite{Larsson12a}, \cite{LHF11}), where it is defined slightly differently due to the different class of games, but has the same intent, namely that the ruleset does not depend on the board. Similar to the previous question, we are interested in which simplicial complexes come from invariant SP-games (iSP games). \cref{lem:illegal} shows that every simplicial complex without an isolated vertex is the illegal complex of some iSP-game and board, and also that every simplicial complex is a legal complex of an iSP-game. The constructions given in all cases prove the stronger result that such SP-games and boards exist given \textit{any} bipartition of the vertices of the simplicial complex (see \cref{thm:Gammagameandgraph,thm:legalinv}) into Left and Right positions. This construction then allows us to show that for every SP-game there exists an iSP-game such that their game trees are isomorphic. This in turn implies that their game values are the same under both normal and mis\`ere winning conditions. Thus it is enough to only consider iSP-games.
Finally, we discuss another isomorphism for a game $G$ on a board $B$ for which the illegal complex is a graph to an \textbf{independence game}, a special class of SP-games whose legal complexes are flag complexes. This class includes many games actually played. \\
In the next two subsections, we give the background in combinatorial game theory and algebra needed for the paper.
Please see any of \cite{ANW07,BCG04,Siegel13} for further information in combinatorial game theory and
\cite{HH11} for the algebra involved. We then show that each simplicial complex is a game complex, and finally in \cref{sec:iSP} we consider invariant SP-games, and independence games in \cref{sec:independence}.
\subsection{Combinatorial Game Theory}
A \textbf{combinatorial game} is a 2-player game with perfect information and no chance, where the two players are \textbf{Left} and \textbf{Right} (denoted by $L$ and $R$ respectively) and they do not move simultaneously. For the purposes of this paper, the winning condition is irrelevant as long as it does not contradict the other conditions for the games.
We denote a combinatorial game by its name in \textsc{Small Caps}.
In this paper, a \textbf{board} will be a graph. For a game $G$ played on a board $B$ we will use the notation $(G,B)$. The \textbf{pieces}, which can be thought of as tokens or as subgraphs of $B$, will be placed on a non-empty collection of vertices - exactly how is given by the \textbf{ruleset}. A \textbf{position} is a configuration of pieces on the board. A position that can be reached through a sequence of legal moves is called a \textbf{legal position}, otherwise we call it an \textbf{illegal position}. A \textbf{basic position} is a board with only one piece placed.
Given a game $G$ and a board $B$, the \textbf{game tree} of $(G,B)$ is a directed graph tree with the edges labelled $L$ or $R$. The vertices of the tree correspond to positions and $X\stackrel{L}{\rightarrow} Y$ if there is a legal move for Left from position $X$ to $Y$. Similarly for edges labelled with an $R$. The games we consider all have finite game trees.
Brown {et al}\onedot \cite{BCHMMNS14} introduced a subclass of combinatorial games, which they called placement games. Their conditions are slightly weaker than what is required for this work.
\begin{definition}\label{def:placementgame}
A \textbf{strong placement game (SP-game)} is a combinatorial game which satisfies the following:
\begin{itemize}
\item[(i)] The board is empty at the beginning of the game.
\item[(ii)] Players place pieces on empty spaces of the board according to the rules.
\item[(iii)] Pieces are not moved or removed once placed.
\item[(iv)] The rules are such that if it is possible to reach a position through a sequence of legal moves,
then any sequence of moves leading to this position consists of legal moves.
\end{itemize}
\end{definition}
Note that condition (iv) in the above definition is necessary for each position to be independent of the order of moves, which results in commutativity when representing positions by monomials (see \cref{sec:gamecomplexes}) and for the hypergraphs representing the game being simplicial complexes.
This condition also implies that any position, whether legal or illegal, in an SP-game can be decomposed into basic positions.
The games in \cref{def:games} are examples of SP-games and will be used throughout the document. The first
two are games that have been introduced early in the development of game theory (see \cite{BCG04}),
but, surprisingly, not much is known about them. As is the case with many games considered in combinatorial game
theory, the board is not specified since the games can be played on any graph. In this paper, when we refer
to a game we
will specify a board.
\begin{example}\label{def:games}
In \textsc{Snort}, players place a piece on a single vertex which is not adjacent to a vertex containing a piece from their opponent.
In \textsc{Col}, players place a piece on a single vertex which is not adjacent to a vertex containing one of their own pieces.
In \textsc{NoGo}, players place a piece on a single unoccupied vertex. At every point in the game, for each maximal group of connected vertices of the board that contain pieces placed by the same player, one of these needs to be adjacent to an empty vertex.
In \textsc{Domineering } (see \cite{Berlekamp88} and \cite{LMR02}), which is played on grids, both players place dominoes. Left may only place vertically, and Right only horizontally. The vertices of the board are the squares of the grid, and each piece occupies two vertices.
In \textsc{Partizan Octal} games, the board is a strip and the players have `dominoes' of different lengths \cite{FK87,Mesda2009}.
\end{example}
Other examples of SP-games are \textsc{Node-Kayles} and \textsc{Arc-Kayles} (see for example \cite{Bodlaender93}, \cite{FT06}, \cite{Schaefer78}).
Depending on a fixed winning condition, combinatorial games can be divided into equivalence classes. The simplest game (essentially smallest game tree) is called the \textbf{game value} of that equivalence class. Game value form a partially ordered semi-group. When one game consists of two sub-games on different boards, then to find the game value it is sufficient to calculate the game values of the sub-games and taking advantage of the additive structure. This is a very useful concept in combinatorial game theory. For more details see \cite{Siegel13}.
\subsection{Combinatorial Commutative Algebra}
Simplicial complexes are one of the main constructs we use to study SP-games. We begin by introducing the required concepts.
\begin{definition}
An (abstract) \textbf{simplicial complex} $\Delta$ on a finite vertex set $V$ is a set of subsets (called \textbf{faces}) of $V$ with the conditions that if $A\in \Delta$ and $B\subseteq A$, then $B\in \Delta$. The \textbf{facets} of a simplicial complex $\Delta$ are the maximal faces of $\Delta$ with respect to inclusion. A \textbf{non-face} of a simplicial complex $\Delta$ is a subset of its vertices that is not a face.
\end{definition}
Note that a simplicial complex is uniquely determined by its facets. Thus a simplicial complex $\Delta$ with facets $F_1,\ldots, F_k$ is denoted by $\Delta=\langle F_1,\ldots, F_k\rangle$.
A simplicial complex of the form $\Delta=\langle \{i_1,i_2,\ldots, i_r\}\rangle$, where $\{i_1,i_2,\ldots, i_r\}$ is the vertex set of $\Delta$, is called a \textbf{simplex}.
\begin{definition}
Given a face $F$ of a simplicial complex $\Delta$, its \textbf{dimension} $dim(F)$ is $|F|-1$. The dimension of the simplicial complex $\Delta$ is the maximum dimension of any of its faces. A simplicial complex $\Delta$ is called \textbf{pure} if all its facets are of the same dimension. The $\mathbf{k}$\textbf{-skeleton} $\Delta^{[k]}$ of a simplicial complex $\Delta$ is the simplicial complex whose facets are the $k$-dimensional faces of $\Delta$.
\end{definition}
The other structures used to study SP-games are square-free monomial ideals, which we introduce now.
\begin{definition}
Let $k$ be a field and $R$ the polynomial ring $k[x_1,\ldots,x_n]$. A product $x_{1}^{a_{1}}\ldots x_{n}^{a_{n}}\in R$, where the $a_i$ are non-negative integers, is called a \textbf{monomial}. Such a monomial is called \textbf{square-free} if each $a_i$ is either 0 or 1.
\end{definition}
\begin{definition}
Let $k$ be a field and $R$ the polynomial ring $k[x_1,\ldots,x_n]$. A \textbf{monomial ideal} of $R$ is an ideal generated by monomials in $R$. A monomial ideal is called a \textbf{square-free monomial ideal} if it is generated by square-free monomials.
\end{definition}
Let $k$ be a field and $R=k[x_1,\ldots,x_n]$ a polynomial ring. There is a one-to-one correspondence between subsets $\{i_1,\ldots, i_r\}$ of $[n]$ and square-free monomials $x_{i_1}\cdots x_{i_r}$ of $R$. Using this observation we can associate to a square-free monomial ideal two unique simplicial complexes: the facet complex and the Stanley-Reisner complex.
\begin{definition}\label{def:SRFcomplex}
The \textbf{facet complex} of a square-free monomial ideal $I$ of $R$, denoted by $\facetsc{}$, is the simplicial complex whose facets correspond to the square-free monomials in the minimal generating set of $I$. The \textbf{Stanley-Reisner complex} of a square-free monomial ideal $I$ of $R$, denoted by $\SRsc{}$, is the simplicial complex whose faces correspond to the square-free monomials not in $I$. In other words,
\begin{align*}
\facetsc{}&=\langle \{i_1,\ldots, i_r\}\mid x_{i_1}\cdots x_{i_r}\text{ minimal generator of }I\rangle\text{ and}\\
\SRideal{}&=\langle \{i_1, \ldots, i_r\}\mid x_{i_1},\cdots,x_{i_r}\not\in I\rangle.
\end{align*}
\end{definition}
This correspondence works in the opposite direction as well.
\begin{definition}\label{def:SRFideal}
The \textbf{facet ideal} of a simplicial complex $\Delta$, denoted by $\facetI{}$, is the ideal of $R$ generated by the monomials corresponding to the facets of $\Delta$. The \textbf{Stanley-Reisner ideal} of a simplicial complex $\Delta$, denoted by $\SRideal{}$, is the ideal of $R$ generated by the monomials corresponding to the minimal non-faces of $\Delta$. In other words,
\begin{align*}
\facetI{}&=\left(x_{i_1}\cdots x_{i_r}\mid \{i_1,\ldots,i_r\}\text{ facet of }\Delta\right)\text{ and}\\
\SRideal{}&=\left(x_{i_1}\cdots x_{i_r}\mid \{i_1,\ldots,i_r\}\not\in\Delta\right).
\end{align*}
\end{definition}
\subsection{Game Complexes and Ideals}\label{sec:gamecomplexes}
We now introduce the construction of simplicial complexes and square-free monomial ideals which are related to SP-games. Unless otherwise specified, let the underlying ring be $R=k[x_1,\ldots,x_m,y_1,\ldots, y_n]$, where $k$ is a field, $m$ the number of basic positions with a Left piece, and $n$ the number of basic positions with a Right piece.
A square-free monomial $z$ of $R$ represents a position $P$ in the game if it is the product over those $x_i$ and $y_j$ such that Left has played in the basic position $i$ and Right has played in the basic position $j$ in order to reach $P$. By condition (iv) in \cref{def:placementgame}, the order of moves to reach $P$ does not matter, thus we have commutativity.
\begin{example}\label{ex:domineeringI}
Consider \textsc{Domineering } played on the board $B$ given in \cref{fig:domex1}.
\begin{figure}[!ht]
\begin{center}
\begin{tikzpicture}
\draw (0,0)--(2,0)--(2,1)--(1,1)--(1,3)--(0,3)--(0,0);
\draw (1,0)--(1,1);
\draw (0,1)--(1,1);
\draw (0,2)--(1,2);
\draw (0.5,0.5) node {$c$};
\draw (0.5,1.5) node {$b$};
\draw (0.5,2.5) node {$a$};
\draw (1.5,0.5) node {$d$};
\end{tikzpicture}
\end{center}
\caption{Board $B$ with squares labelled}
\label{fig:domex1}
\end{figure}
Since Left and Right both play dominoes, the basic positions are to place a domino on vertices $a,b$ (basic position 1), on $b,c$ (basic position 2), or on $c,d$ (basic position 3). Thus the underlying ring in this case is $R=k[x_1, x_2, x_3, y_1, y_2, y_3]$.
Since Left may only place a domino vertically, the basic position represented by $x_1$ and $x_2$ are legal, while $x_3$ is illegal. Similarly, for Right $y_1$ and $y_2$ are illegal, while $y_3$ is legal.
The monomial $x_1y_3$ represents the position in which Left has placed a domino on vertices $a$ and $b$, and Right has played on $c$ and $d$, which is a legal position. Similarly, $x_2y_3$ represents the position where Left has played on vertices $b$ and $c$, while Right has played on $c$ and $d$, which is illegal since the two dominoes overlap.
\end{example}
A legal position is called a \textbf{maximal legal position} if placing any further piece is illegal, {i.e}\onedot} \def\Ie{{I.e}\onedot it is not properly contained in any other legal position.
If we sort the monomials representing illegal positions by divisibility, the positions corresponding to the minimal elements are called \textbf{minimal illegal positions}. Equivalently, an illegal position is a minimal illegal position if any proper subset of the pieces placed forms a legal position.
\begin{definition}\cite{FHN14a}\label{def:gameComplexIdeal}
If $(G,B)$ is an SP-game, then
\begin{itemize}
\item The \textbf{legal ideal}, $\legalI{G,B}$, is the ideal of $R$ generated by the monomials representing maximal legal positions.
\item The \textbf{illegal ideal}, $\illegalI{G,B}$, is the ideal generated by the monomials representing minimal illegal positions.
\item The \textbf{legal complex}, $\Delta_{G,B}$, is the facet complex of the legal ideal.
\item The \textbf{illegal complex}, $\Gamma_{G,B}$, is the facet complex of the illegal ideal.
\end{itemize}
\end{definition}
If a given simplicial complex is the legal or illegal complex of some game and board, we also call it a \textbf{game complex}.
\begin{remark}
Note that the faces of the legal complex $\Delta_{G,B}$ represent the legal positions of $(G,B)$, while the facets of $\Gamma_{G,B}$ represent the minimal illegal positions. In short we have
\begin{itemize}
\item[(1)] $\legalI{G,B}=\facetI{G,B},$
\item[(2)] $\illegalI{G,B}=\facet{\illegalcomp{G,B}}=\SRideal{G,B},$
\end{itemize}
or equivalently
\begin{itemize}
\item[(1)] $\Delta_{G,B}=\facet{\legalI{G,B}}=\SR{\illegalI{G,B}},$
\item[(2)] $\Gamma_{G,B}=\facet{\illegalI{G,B}}.$
\end{itemize}
This will be used throughout this paper.
\end{remark}
Note that condition (iv) in \cref{def:placementgame} implies that the order of moves does not matter, which gives us commutativity when representing positions by monomials, thus the legal and illegal ideal are indeed commutative ideals. The condition also implies that given any legal position, any subset of the pieces played gives a legal position as well, and thus the hypergraphs representing the game are indeed simplicial complexes.
We will continue \cref{ex:domineeringI} to demonstrates these concepts. This also illustrates again that the vertices of the complexes are the basic positions, not the vertices of the graph/board.
\begin{example}
Consider \textsc{Domineering } played on the board $B$ given in \cref{fig:domex1}. Our underlying ring is $R=k[x_{1}, x_{2}, x_{3}, y_{1}, y_{2}, y_{3}]$.
The maximal legal positions are represented by the monomials $x_{1}y_{3}$ and $x_{2}$. Thus we have the legal ideal \[\legalI{\textsc{Domineering}, B}=\langle x_{1}y_{3}, x_{2}\rangle.\] The legal complex is given in \cref{fig:domex2}.
\begin{figure}[!ht]
\begin{center}
\begin{tikzpicture}
\draw (0,0)--(1,0);
\filldraw (0,0) circle (0.1cm);
\filldraw (1,0) circle (0.1cm);
\filldraw (0.5,1) circle (0.1cm);
\draw (0,0.25) node {$x_{1}$};
\draw (1,0.25) node {$y_{3}$};
\draw (0.5,0.75) node {$x_{2}$};
\end{tikzpicture}
\end{center}
\caption{The legal complex $\Delta_{\textsc{Domineering}, B}$}
\label{fig:domex2}
\end{figure}
The minimal illegal positions are represented by the monomials $x_{1}x_{2}$, $x_{2}y_{3}$, $x_{3}$, $y_{1}$, and $y_{2}$. Thus we have the illegal ideal
\[\illegalI{\textsc{Domineering}, B}=\langle x_{1}x_{2}, x_{2}y_{3}, x_{3}, y_{1}, y_{2}\rangle.\] The illegal complex is given in \cref{fig:domex3}.
\begin{figure}[!ht]
\begin{center}
\begin{tikzpicture}
\draw (0,0)--(1,0)--(2,0);
\filldraw (0,0) circle (0.1cm);
\filldraw (1,0) circle (0.1cm);
\filldraw (2,0) circle (0.1cm);
\filldraw (3,0) circle (0.1cm);
\filldraw (4,0) circle (0.1cm);
\filldraw (5,0) circle (0.1cm);
\draw (0,0.25) node {$x_{1}$};
\draw (1,0.25) node {$x_{2}$};
\draw (2,0.25) node {$y_{3}$};
\draw (3,0.25) node {$x_{3}$};
\draw (4,0.25) node {$y_{1}$};
\draw (5,0.25) node {$y_{2}$};
\end{tikzpicture}
\end{center}
\caption{The illegal complex $\Gamma_{\textsc{Domineering}, B}$}
\label{fig:domex3}
\end{figure}
\end{example}
It is important to note that the legal and illegal complexes and corresponding ideals have an extra layer of structure. The monomials have elements $\{x_1,x_2,\ldots,x_n\}$ and $\{y_1,y_2,\ldots,y_n\}$ and the complexes have their elements partitioned into those corresponding to the Left and Right basic positions. Isomorphisms between game complexes and ideals must also preserve these partitions. Thus, when showing a complex is isomorphic to a game complex, we must also specify the partition.
In general, we call a simplicial complex whose vertex set is bipartioned into sets $\L$ and $\mathfrak{R}$ an $(\L,\mathfrak{R})$-labelled simplicial complex.
The following proposition shows that two games with isomorphic legal complexes have isomorphic game trees, and as a consequence the same game value under most winning conditions (such as normal play and mis\`ere, see \cite{Siegel13}).
\begin{proposition}[\textbf{Isomorphic Game Trees of SP-Games}]\label{thm:isomorphicgametrees}
Two SP-games $G_1$ and $G_2$ played on boards $B_1$ and $B_2$ respectively have isomorphic legal complexes
if and only if the game trees of $G_1$ and $G_2$ are isomorphic.
\end{proposition}
\begin{proof} We prove that isomorphic legal complexes imply isomorphic game trees by induction on the size of the faces ({i.e}\onedot} \def\Ie{{I.e}\onedot the number of pieces in a position). The empty face ({i.e}\onedot} \def\Ie{{I.e}\onedot empty board) corresponds to the root of the game tree, thus is trivially the same for both games.
Now assume that the game trees are isomorphic up to positions with $k$ pieces played.
Consider a position $P_1$ in the game $G_1$ played on $B_1$ with $k$ pieces played. Let $F_1$ be the face of $\Delta_{G_1,B_1}$ (of dimension $k-1$) corresponding to $P_1$. Since $\Delta_{G_1,B_1}$ and $\Delta_{G_2,B_2}$ are isomorphic, there exists a face $F_2\in \Delta_{G_2,B_2}$ (of dimension $k-1$) isomorphic to $F_1$, corresponding to a position $P_2$ of $G_2$, which also has $k$ pieces placed.
Now let $P_1'$ be any option of $P_1$ and $F_1'$ be the corresponding face in $\Delta_{G_1,B_1}$. Then there exists a vertex $v$ such that $F_1'=F_1\cup\{v\}$. Let $F_2'$ be the face of $\Delta_{G_2,B_2}$ isomorphic to $F_1'$. Then there exists a vertex $w$ (corresponding to $v$) such that $F_2'=F_2\cup\{w\}$. Thus the position $P_2'$ corresponding to $F_2'$ is an option of $P_2$.
Further, since the legal complexes have the same bipartition, we have that the following are equivalent:
\begin{enumerate}
\item The position $P_1'$ is a Left- (Right-)option of $P_1$.
\item The vertex $v$ belongs to $\L$ ($\mathfrak{R}$).
\item The vertex $w$ belongs to $\L$ ($\mathfrak{R}$).
\item The position $P_2'$ is a Left- (Right-)option of $P_2$.
\end{enumerate}
Thus for any option of $P_1$ there exists an option of $P_2$ and vice-versa, which shows that the game trees of $G_1,B_1$ and $G_2,B_2$ are isomorphic up to positions of $k+1$, and by induction they are entirely isomorphic.
The proof of the converse is similar and so we omit it.
\end{proof}
If the illegal complexes are isomorphic, it is not always true that the game trees are isomorphic. For example, consider a game $G$ in which neither player can place on a vertex of degree 1. We then have \[\Gamma_{G,P_2}=\langle x_1,x_2,y_1,y_2\rangle\cong\Gamma_{G,P_3}=\langle x_1, x_3, y_1, y_3\rangle.\] The legal complexes $\Delta_{G,P_2}=\emptyset$ and $\Delta_{G, P_3}=\langle x_2, y_2\rangle$ are not isomorphic, and thus by \cref{thm:isomorphicgametrees} their game trees are not either. Another occurrence of this is if there are moves that are always playable in one game, but these moves do not occur at all in the second game. This problem appears and is dealt with in the proof of \cref{thm:legalinv}.\\
A natural and important question is whether any given simplicial complex $\Delta$ is the legal or illegal complex of some game. We will answer this question positively in both cases. This will allow us to view properties of games as properties of simplicial complexes and vice-versa. We are able to show this for any bipartition of the vertices into Left $\L$ and Right $\mathfrak{R}$, where $\L$ or $\mathfrak{R}$ could even be the empty set.
\begin{proposition}[\textbf{Games from Simplicial Complexes}]\label{thm:legalillegal}
Given an $(\L,\mathfrak{R})$-labelled simplicial complex $\Delta$, there exist SP-games $G_1, G_2$ and a board $B$ such that
\begin{itemize}
\item[(a)] $\Delta=\Delta_{G_1,B}$ and
\item[(b)] $\Delta=\Gamma_{G_2,B}$
\end{itemize}
and the sets of Left (respectively Right) positions is $\L$ (respectively $\mathfrak{R}$).
\end{proposition}
\begin{proof}
Let $m=|\L|$ and $n=|\mathfrak{R}|$. Let $B$ be the board consisting of $m$ disjoint $3$-cycles and $n$ disjoint $4$-cycles. In the games $G_1$ and $G_2$, Left will be playing $3$-cycles, while Right will be playing $4$-cycles.
In $\Delta$, label the vertices belonging to $\L$ as $1, \ldots, m$, and the vertices in $\mathfrak{R}$ as $m+1, \ldots, n+m$. Similarly, label the $3$-cycles of $B$ as $1, \ldots, m$, and the $4$-cycles as $m+1, \ldots, n+m$.
(a) In $G_1$, playing on a set of cycles of $B$ is legal if and only if the corresponding set of vertices in $\Delta$ forms a face.
(b) In $G_2$, playing on a set of cycles of $B$ is legal if and only if the corresponding set of vertices in $\Delta$ does not contain a facet.
It is now easy to see that $\Delta=\Delta_{G_1,B}$ and $\Delta=\Gamma_{G_2,B}$.
\end{proof}
As seen above, it is rather simple to construct games on fixed boards from simplicial complexes by restricting the legal moves to certain parts of the board. We now move on to look at games where such restrictions can be relaxed. We call these invariant games.
\section{Invariant Games}\label{sec:iSP}
As we have shown in the previous section, every simplicial complex is the legal or illegal complex of some SP-game and board. The rules created as part of this construction, however, depend heavily on the board. We now define the concept of invariance for SP-games, which in a sense forces the ruleset to be ``uniform'' across the board.
\begin{definition}
The ruleset of an SP-game is \textbf{invariant} if the following conditions hold:
\begin{itemize}
\item Every basic position is legal.
\item The ruleset does not depend on the board, {i.e}\onedot} \def\Ie{{I.e}\onedot if $B_1$ and $B_2$ are isomorphic subgraphs of any board $B$, then a position on $B_1$ is legal if and only if its isomorphic image on $B_2$ is legal.
\end{itemize}
\end{definition}
If the ruleset of an SP-game is invariant, we also say that the game is an \textbf{invariant strong placement game (iSP-game)}.
\textsc{Col } and \textsc{Snort } are examples of games that are invariant, while \textsc{Domineering } and \textsc{NoGo } are not. In \textsc{Domineering } half of the basic positions are illegal (Right cannot play vertically, while Left cannot play horizontally). That \textsc{NoGo } is not invariant is not as obvious. Indeed on most boards both conditions hold, but whenever the board has an isolated vertex, playing on it is illegal (thus the basic position corresponding to that vertex is illegal).
Similar to the question of the previous section, we are interested in which simplicial complexes appear as the legal or illegal complex of an iSP-game.
We will show below that the illegal complex of an iSP-game cannot contain an isolated vertex.
\begin{proposition}\label{thm:illegalinvariantnovertices}
Let $\Gamma$ be a simplicial complex. If $\Gamma$ is the illegal complex of some iSP-game and some board then $\Gamma$ has no facets that are one-element sets.
\end{proposition}
\begin{proof}
Assume that $\Gamma$ has a facet that is a one-element set, {i.e}\onedot} \def\Ie{{I.e}\onedot an isolated vertex, and label this vertex $a$. If $\Gamma$ is the illegal complex of some SP-game $G$ and board $B$, then since $\{a\}$ is a facet of $\Gamma$, there exists a basic position (corresponding to the vertex $a$) which is illegal. Thus $G$ does not satisfy the first condition of invariance.
\end{proof}
Other than the isolated vertex situation, there is no obstruction for a simplicial complex $\Gamma$ being an illegal complex. We set out to prove this by constructing a $\Gamma$-board and a $\Gamma$-game.
\begin{construction}[\textbf{$\mathbf{\Gamma}$-Board}]
Given an $(\L,\mathfrak{R})$-labelled simplicial complex $\Gamma$ with no isolated vertices we can construct a graph $B_\Gamma$ (called the \textbf{$\Gamma$-Board}) as follows:
If $\Gamma$ is empty, then let $B_\Gamma$ be empty.
If $\Gamma$ is non-empty, then let $H=\Gamma^{[1]}$, {i.e}\onedot} \def\Ie{{I.e}\onedot the underlying graph of $\Gamma$. Let $n$ be the number of vertices in the graph $H$ and (re)label the vertices of $H$ as $1,\ldots, n$. Begin constructing the board $B_\Gamma$ by using $n$ cycles of sizes $n^4+4$ and $n^4+5$ and label these $1,\ldots, n$ so that cycle $i$ will have size $n^4+4$ if the vertex $i$ in $H$ belongs to $\L$, and size $n^4+5$ if the vertex $i$ belongs to $\mathfrak{R}$. For each cycle, designate $n-1$ consecutive vertices for joining, called \textbf{connection vertices} (see \cref{fig:cyclei}).
\begin{figure}[!ht]
\begin{center}
\includegraphics[scale=1]{GraphsToGames1bwa.eps}
\end{center}
\caption{Cycle $i$ in the board $B_\Gamma$}
\label{fig:cyclei}
\end{figure}
Call the remaining vertices \textbf{outer vertices}. To each connection vertex, join a cycle of length $n^3$ (called \textbf{inner cycles}). In cycle $i$ label the connection vertices as $i,j$ where $j=1,\ldots, n$ and $j\neq i$.
Label the edges in $H$ as $1,\ldots, k$. If the endpoints of the edge $l$ are the vertices $i$ and $j$, then add a path of $2+l$ vertices to $B_\Gamma$, whose end vertices are $i,j$ and $j,i$ (see \cref{fig:edgetopath}). The $l$ vertices between $i,j$ and $j,i$ are called \textbf{centre vertices}.
\begin{figure}[!ht]
\begin{center}
\includegraphics[scale=0.45]{GraphsToGames2a.eps}
\end{center}
\caption{Effect of an edge in $H$ on the board $B_\Gamma$}
\label{fig:edgetopath}
\end{figure}
\end{construction}
As an example for this construction, consider the following:
\begin{example}\label{ex:graph1}
Let $\Gamma$ be a path of three vertices so that $H=\Gamma$. Let the two end vertices belong to $\L$, and the centre vertex to $\mathfrak{R}$. Since $\Gamma$ consists of three vertices, {i.e}\onedot} \def\Ie{{I.e}\onedot $n=3$, the cycle $i$ (where $i\in\L$) is of length $3^4+4=85$ with two cycles of length $3^3=27$ joined to two adjacent vertices, and the cycle $j$ (where $j\in\mathfrak{R}$) is of length 86 with two cycles of length 27 joined to two adjacent vertices.
Label the edge between vertex 1 (an end vertex) and vertex 2 (the centre vertex) as 1, and the edge between vertex 2 and vertex 3 (the other end vertex) as 2.
The board $B_\Gamma$ is given in \cref{fig:exampleP3}. Dashed cycles consist of 85 vertices, and dotted cycles of 86 vertices, with the two labelled vertices adjacent in both cases. The smaller solid cycles consist of 27 vertices.
\begin{figure}[!ht]
\begin{center}
\includegraphics[scale=0.4]{GraphsToGames3abw.eps}
\end{center}
\caption{Constructing $B_{P_3}$}
\label{fig:exampleP3}
\end{figure}
\end{example}
For the next construction, we will have to specify what is meant with distance between pieces.
\begin{definition}
Let two pieces $P_1$ and $P_2$ be placed on a board $B$ and let $V_1$ and $V_2$ be the set of vertices on which $P_1$, respectively $P_2$, was placed. We then define the \textbf{distance} $d(P_1, P_2)$ between $P_1$ and $P_2$ by
\[d(P_1,P_2)=\min\{d(v_1,v_2): v_1\in V_1, v_2\in V_2\},\]
where $d(v_1,v_2)$ is the graph theoretic distance between $v_1$ and $v_2$, {i.e}\onedot} \def\Ie{{I.e}\onedot the minimum number of edges of a path in $B$ with endpoints $v_1$ and $v_2$.
\end{definition}
\begin{construction}[\textbf{$\Gamma$-Game}]
Given an $(\L,\mathfrak{R})$-labelled simplicial complex $\Gamma$ with no isolated vertices we construct an SP-game $G_\Gamma$ (called the \textbf{$\Gamma$-Game}).
If $\Gamma$ is empty, then let $G_\Gamma$ be the game in which Left and Right place pieces on a single vertex with no restrictions.
If $\Gamma$ is non-empty, then construct $G_\Gamma$ as follows:
\begin{itemize}
\item Let $n$ be the number of vertices of $\Gamma$. Label the edges (the 1-dimensional faces) of $\Gamma$ as $\{1,\ldots,k\}$.
\item Left plays cycles of length $n^4+4$ with cycles of length $n^3$ joined to $n-1$ consecutive vertices,
\item Right plays cycles of length $n^4+5$ with cycles of length $n^3$ joined to $n-1$ consecutive vertices ({i.e}\onedot} \def\Ie{{I.e}\onedot the pieces are as the structure given in \cref{fig:cyclei}), and
\item Let $F$ be a facet of $\Gamma$ of dimension $f-1$, whose 1-dimensional faces are labelled $k_1,\ldots, k_l$, where $l=\binom{f}{2}$. We call the set $\{k_1+1,\ldots, k_l+1\}$ the \textbf{id-set} of $F$. Then no sets of $f$ pieces are allowed such that the set of distances between any two pieces is exactly the id-set of $F$.
\end{itemize}
\end{construction}
\begin{example}\label{ex:graph2}
Let $\Gamma$ be a path of three vertices so that $n=3$. Left's pieces are cycles of length $3^4+4=85$ with two cycles of length $3^3=27$ joined to two adjacent vertices, and Rights pieces are cycles of length 86 with two cycles of length 27 joined to two adjacent vertices.
Since the facets of $\Gamma$ are the two edges (thus of size 2), we have that no two pieces in $G_\Gamma$ are allowed to have distance 2 or distance 3, {i.e}\onedot} \def\Ie{{I.e}\onedot the id-sets are $\{2\}$ and $\{3\}$.
\end{example}
\begin{example}
Consider $\Gamma=\langle abc, ad\rangle$. Label the edge between $a$ and $b$ as 1, between $b$ and $c$ as 2, between $c$ and $a$ as 3, and between $a$ and $d$ as 4.
For the facet $abc$ we have the id-set $\{1+1, 2+1, 3+1\}=\{2,3,4\}$. Thus in the $\Gamma$-game $G_\Gamma$ we cannot have three pieces where the distances between pairs are $\{2,3,4\}$, while two with any one of these distance are allowed.
For the facet $ad$ we have the id-set $\{4+1\}=\{5\}$. Thus in $G_\Gamma$ we cannot have any two pieces with distance 5.
\end{example}
\begin{lemma}\label{lem:illegal}
Given an $(\L,\mathfrak{R})$-labelled simplicial complex $\Gamma$ with no isolated vertices, the $\Gamma$-game $G_\Gamma$ is an iSP-game.
\end{lemma}
\begin{proof}
If $\Gamma$ is empty, then $G_\Gamma$ has no illegal positions, thus is trivially invariant.
If $\Gamma$ is non-empty, then since $\Gamma$ has no isolated vertices,
all facets have at least one edge and therefore all id-sets are non-empty.
In particular, this means that every illegal position of $G_\Gamma$ has at least two pieces,
so there are no illegal basic positions.
Now suppose that we are playing $G_\Gamma$ on a board $B$, and let $B_1$ and $B_2$ be isomorphic subgraphs of $B$. A position $P$ is legal on $B_1$ if and only if there is no id-set which is contained in the set of distances between pieces of $P$, which holds if and only if $P$ is legal on $B_2$.
Thus $G_\Gamma$ is invariant.
\end{proof}
The following statement will prove that every simplicial complex without isolated vertices can appear as the illegal
complex of (many!) iSP-games.
\begin{theorem}[\textbf{Invariant Game from Illegal Complex}]\label{thm:Gammagameandgraph}
Given an $(\L,\mathfrak{R})$-labelled simplicial complex $\Gamma$ with no isolated vertices, fix labellings of the vertices and of the edges. Then $\Gamma$ is the illegal complex of the $\Gamma$-game $G_\Gamma$ played on the $\Gamma$-board $B_\Gamma$, {i.e}\onedot} \def\Ie{{I.e}\onedot $\Gamma_{G_\Gamma,B_\Gamma}=\Gamma$.
\end{theorem}
\begin{proof}
Let $B=B_\Gamma$ and $G=G_\Gamma$ be the $\Gamma$-board and $\Gamma$-game respectively, with the same labelling of the edges of $\Gamma$ if $\Gamma$ is nonempty.
If $\Gamma$ is empty, then $G_\Gamma$ has no illegal positions, thus $\Gamma_{G,B}$ is also empty.
To show that indeed $\Gamma_{G,B}=\Gamma$ for $\Gamma$ nonempty, we will begin by showing that their vertex sets have the same size.
Let $H=\Gamma^{[1]}$. Clearly Left can place one of her pieces on the cycle labelled $i$ in $B$ if the vertex $i$ of $H$ belongs to $\L$. Similarly Right can place on cycles labelled $j$ where $j\in\mathfrak{R}$. Thus each vertex in $H$ corresponds to a position in the game $G$ played on $B$.
We now need to show that there are no other ways for Left or Right to place pieces than what was previously mentioned, {i.e}\onedot} \def\Ie{{I.e}\onedot that the positions of $G$ played on $B$ correspond exactly to the vertices of $H$.
Let $n$ be the number of vertices of $H$ and $k$ be the number of edges. The cycles in $B$ which only use connection and centre vertices have size at most $n(n-1)+\frac{k(k+1)}{2}$ (there are $n(n-1)$ connection vertices and $1+\ldots+k$ centre vertices). Since there are at most $\binom{n}{2}$ edges in $H$, we have
\begin{align*}
n(n+1)+\frac{k(k+1)}{2}&\le n(n+1)+\frac{\frac{n(n+1)}{2}\left(\frac{n(n+1)}{2}+1\right)}{2}\\
&=\frac{1}{8}n^4+\frac{1}{4}n^3+\frac{11}{8}n^2+\frac{5}{4}n
\end{align*}
which is less than $n^4+4$ for all whole numbers.
Thus such cycles are shorter than $n^4+4$, and Left and Right will not be able to play on those.
Furthermore, any cycle of length $n^4+4$ or $n^4+5$ in $B$ needs to include the outer vertices of some cycle $i$ (since as above cycles using only connection and centre vertices are shorter, and the inner cycles are shorter). To then construct a cycle of that length without using all connection vertices of cycle $i$, the cycle would have to include at least one centre vertex. Since centre vertices do not have cycles of length $n^3$ added, this implies that neither Left or Right could play there.
Thus Left and Right are only able to play on the labelled cycles.
Further, since the pieces consist of cycles with a differing number of vertices, either player will only be able to play on the cycles of $B$ that are designated to them. Thus there are $n$ positions, in each of which only one player can play, all corresponding to vertices of $\Gamma$. The vertices of $\Gamma_{G,B}$ are thus a subset of the vertices of $\Gamma$ and $\Gamma_{G,B}$ has less vertices than $\Gamma$ if and only if there exists at least one position in which it is never illegal to play, which we will show cannot happen as part of the rest of the proof.
It remains to show that the facets of $\Gamma_{G,B}$ and $\Gamma$ correspond.
Consider a facet consisting of the vertices $i_1,\ldots, i_k$ in $\Gamma$, thus any two vertices have an edge between them in $H$, and let these edges be $j_1,\ldots, j_l$. Then the positions $i_a$ and $i_b$, $a,b\in \{1,\ldots, k\}$, in $B$ have distance $j_c+1$, where $j_c$ is the edge between $i_a$ and $i_b$ in $H$, (since we joined a path of length $j_c+2$ to their connection vertices). Thus it is illegal to play in all $k$ positions (and this is a minimal illegal position), and thus there is a facet consisting of the vertices $i_1,\ldots, i_k$ in $\Gamma_{G,B}$.
Now let the vertices $i_1,\ldots, i_k$ form a facet in $\Gamma_{G,B}$. Assume that $i_1,\ldots, i_k$ do not form a facet in $\Gamma$. If some subset $S$ of these vertices forms a facet, then by construction of $G$ it would be illegal to play pieces on all of the cycles in $B$ corresponding to vertices in $S$. Thus $i_1,\ldots, i_k$ is not a \textit{minimal} illegal position, a contradiction to those vertices forming a facet in $\Gamma_{G,B}$. If on the other hand $i_1,\ldots, i_k$ is strictly contained in some facet $F$ of $\Gamma$, then by construction of $G$ it is legal to play on cycles $i_1,\ldots, i_k$ in $B$. Thus $i_1,\ldots, i_k$ is not an \textit{illegal} position, a contradiction to those vertices forming a facet in $\Gamma_{G,B}$. Therefore $i_1,\ldots, i_k$ is a facet of $\Gamma$.
Finally, since $H$ has no isolated vertices (by $\Gamma$ not having such), the vertex set of $\Gamma_{G,B}$ is a subset of the vertex set of $H$, {i.e}\onedot} \def\Ie{{I.e}\onedot the vertex set of $\Gamma$. Since furthermore the facets of $\Gamma_{G,B}$ and $\Gamma$ correspond, we have that the vertex set of $\Gamma_{G,B}$ is equal to that of $\Gamma$.
Consequently, the simplicial complexes $\Gamma$ and $\Gamma_{G,B}$ have the same vertex and facet sets, which proves $\Gamma=\Gamma_{G,B}$.
\end{proof}
\begin{example}\label{ex:graph3}
Let $\Gamma$ be a path of three vertices. Let $B=B_\Gamma$ (see \cref{ex:graph1}) and $G=G_\Gamma$ (see \cref{ex:graph2}).
Then $\Gamma_{G,B}=\Gamma$.
\end{example}
Note: Simpler constructions with smaller cycles and pieces are often possible (as shown in the next example), but the above construction is guaranteed to work.
\begin{example}
Let $\Gamma$ be as in \cref{ex:graph3}. Let Left play cycles of length $3$, and Right cycles of length $4$. For the board $B'$ given in \cref{fig:exampleP3simple}, it is easy to check that $\Gamma_{G',B'}=\Gamma$, where $G'$ is the SP-game with no additional rules (this in particular means that pieces are not allowed to overlap).
\begin{figure}[!ht]
\begin{center}
\includegraphics[scale=0.6]{GraphsToGames4.eps}
\end{center}
\caption{Smaller board $B'$}
\label{fig:exampleP3simple}
\end{figure}
\end{example}
The following theorem summarizes our results about illegal complexes of iSP-games.
\begin{theorem}\label{thm:illegalcomplexinvariant}
A given simplicial complex $\Gamma$ is the illegal complex of some iSP-game $G$ played on a board $B$ if and only if $\Gamma$ has no isolated vertices.
\end{theorem}
\begin{proof}
By \cref{thm:illegalinvariantnovertices} we have that if $\Gamma$ is the illegal complex of an iSP-game, then $\Gamma$ has no isolated vertices.
Conversely, if $\Gamma$ has no isolated vertices, then by \cref{thm:Gammagameandgraph}, we have that $\Gamma$ is the illegal complex of some iSP-game and board.
\end{proof}
We will now consider legal complexes. The first result shows that \textit{every} simplicial complex is the legal complex of some iSP-game and board:
\begin{theorem}[\textbf{Invariant Game from Legal Complex}]\label{thm:legalinv}
Given any $(\L,\mathfrak{R})$-labelled simplicial complex $\Delta$, we can construct an iSP-game $G$ and a board $B$ such that $\Delta=\Delta_{G,B}$ and the sets of Left, respectively Right, positions is $\L$, respectively $\mathfrak{R}$.
\end{theorem}
\begin{proof}
We will prove this separately for the case in which the simplicial complex $\Delta$ is not a simplex (corresponding to $\facet{\SR{\Delta}}$ has at least one 1-dimensional face), and when it is a simplex (corresponding to $\facet{\SR{\Delta}}$ has only isolated vertices or is empty).
\textit{Case 1:} If $\Delta$ is not a simplex, the construction is as follows:
Given $\Delta$, let $\Gamma=\facet{\SR{\Delta}}$, {i.e}\onedot} \def\Ie{{I.e}\onedot the simplicial complex whose facets correspond to the minimal non-faces of $\Delta$.
Let $i$ be a vertex in $\Delta$. If $\Delta$ has at least one facet that does not contain $i$, then $i$ will also be a vertex of $\Gamma$. Otherwise it is not a vertex of $\Gamma$.
Let the vertex set of $\Gamma$ be bipartitioned into $\L$ and $\mathfrak{R}$ the same way that the vertex set of $\Delta$ is. Let $n$ be the number of vertices in $\Gamma$ and let $G$ be the $\Gamma$-game and $B_0$ be the $\Gamma$-board, so that $\Gamma_{G,B_0}=\Gamma$. If $\Delta$ has a vertex that is contained in every facet, then the underlying rings of $\Gamma_{G,B_0}$ and $\Gamma$ are not the same, and we thus have to adjust the board as follows:
Without loss of generality, let $1,\ldots, k$ be the vertices of $\Delta$ that are contained in every facet. Then for $l=1,\ldots, k$ let $B_l=B_{l-1}\cup C^l$ where $C^l$ is a cycle of length $n^4+4$ (if the vertex $l$ belongs to $\L$) or length $n^4+5$ (if it belongs to $\mathfrak{R}$) with $n-1$ cycles of length $n^3$ joined to $n-1$ consecutive vertices. Let $B=B_k$. When playing the game $G$ on $B$, it is always legal to play on the disjoint $C^l$ for either Left or Right, thus these positions are never part of a minimal illegal position, which shows that $\Gamma_{G,B_0}=\Gamma_{G, B}$. Furthermore, the underlying rings of $\Gamma$ and $\Gamma_{G,B}$ are the same.
It immediately follows that \[\Delta_{G,B}=\SR{\facet{\Gamma_{G,B}}}=\SR{\facet{\Gamma}}=\Delta.\]
\textit{Case 2:} If $\Delta$ is a simplex, we can construct $G$ and $B$ as follows:
Let $n$ be the number of vertices in $\Delta$ and (re)label the vertices $1,\ldots, n$. Let the board $B$ be a disjoint union of $n$ cycles of size $3$ and $4$ and label these $1,\ldots, n$ so that cycle $i$ will have size $3$ if the vertex $i$ in $\Delta$ belongs to $\L$, and size $4$ if the vertex $i$ belongs to $\mathfrak{R}$.
Let $G$ be the SP-game in which Left plays cycles of length $3$, and Right plays cycles of length $4$. Note that $G$ is invariant.
It is easy to see that $\Delta=\Delta_{G,B}$.
\end{proof}
The following two examples demonstrate this construction in both the case where $\Delta$ is not a simplex and when it is.
\begin{example}
Consider the complex $\Delta=\langle ab, bc\rangle$, where the vertices are partitioned as $\L=\{a,b\}$ and $\mathfrak{R}=\{c\}$. Since $\Delta$ is not a simplex, we will follow the construction given in the first case of the proof of \cref{thm:legalinv}.
The only minimal nonface of $\Delta$ is $ac$, thus the graph $H$ is $P_2$. Since $n=2$, in the SP-game $G$ Left will play cycles of length $n^4+4=20$ with one cycle of length $n^3=8$ added to a vertex, while Right plays cycles of length $n^4+5=21$ with a cycle of length 8 added to a vertex.
The board $B$ is given in \cref{fig:exampleFlag3}. Dashed cycles consist of 20 vertices, and dotted cycles of 21 vertices. The smaller solid cycles consist of 8 vertices.
\begin{figure}[!ht]
\begin{center}
\includegraphics[scale=0.75]{GraphsToGames7.eps}
\end{center}
\caption{Constructing $B$ from $\Delta=\langle ab, bc\rangle$}
\label{fig:exampleFlag3}
\end{figure}
It is now easy to check that $\Delta_{G,B}=\Delta$.
\end{example}
\begin{example}
Consider the simplex $\Delta=\langle abc\rangle$, where the vertices are partitioned as $\L=\{a\}$ and $\mathfrak{R}=\{b,c\}$. Since $\Delta$ is a simplex, we will follow the second construction given in the proof of \cref{thm:legalinv}. Since $n=3$, in the SP-game $G$ Left will play cycles of length 3, while Right plays cycles of length 4.
The board $B$ is given in \cref{fig:exampleFlag2}.
\begin{figure}[!ht]
\begin{center}
\includegraphics[scale=0.75]{GraphsToGames6.eps}
\end{center}
\caption{Constructing $B$ from $\Delta=\langle abc\rangle$}
\label{fig:exampleFlag2}
\end{figure}
It is now easy to check that $\Delta_{G,B}=\Delta$ and that $\Gamma_{G,B}$ is empty.
\end{example}
Concluding our discussion of iSP-games, we have the following result.
\begin{theorem}[\textbf{Every SP-Game Tree Belongs To An iSP-Game}]
Given an SP-game $G$ played on a board $B$, there exists an iSP-game $G'$ played on a board $B'$ so that their game trees are isomorphic.
\end{theorem}
\begin{proof}
Let $\Delta=\Delta_{G,B}$ with $\L$ the vertices corresponding to Left basic positions, and similarly $\mathfrak{R}$. Then by \cref{thm:legalinv} we know that there exists an iSP-game $G'$ and a board $B'$ such that $\Delta=\Delta_{G',B'}$ with the same bipartition. Since $\Delta_{G,B}=\Delta_{G',B'}$, we have by \cref{thm:isomorphicgametrees} that the game trees of $G$ played on $B$ and $G'$ played on $B'$ are isomorphic.
\end{proof}
This in particular implies that under most winning conditions (such as normal play or mis\`ere play) the game values of $G$ played on $B$ and $G'$ played on $B'$ are the same, implying that we can replace one by the other.
\section{Independence Games}\label{sec:independence}
Many of the games we have previously considered have illegal complexes that are graphs. This special class of SP-games is of further interest to us. For example, this class corresponds to flag complexes (see below for more).
\begin{definition}
An SP-game $G$ is called an \textbf{independence game} if for \textit{any} board $B$ the illegal complex $\Gamma_{G,B}$ is a graph without isolated vertices ({i.e}\onedot} \def\Ie{{I.e}\onedot a pure one-dimensional simplicial complex).
\end{definition}
Consider the illegal complex $\Gamma_{G,B}$ of an independence game $G$ on a board $B$. Let $\Gamma_{G,B}'$ be the graph on the vertex set $x_1,x_2,\ldots,x_m, y_1, y_2,\ldots, y_n$ (corresponding to the basic positions of $G$ played on $B$) with edges those of $\Gamma_{G,B}$. Thus the difference between $\Gamma_{G,B}$ and $\Gamma_{G,B}'$ are isolated vertices corresponding to basic positions that are always legal. For many independence games we have $\Gamma_{G,B}'=\Gamma_{G,B}$.
The \textbf{independence complex} of a graph $H$ is a simplicial complex with vertex set that of the graph and faces those sets of vertices that are independent in $H$, {i.e}\onedot} \def\Ie{{I.e}\onedot no two vertices are adjacent. The term `independence game' was chosen for this class of games since the independence sets of $\Gamma_{G,B}'$ correspond to the legal positions of $G$ played on $B$, {i.e}\onedot} \def\Ie{{I.e}\onedot the faces of $\Delta_{G,B}$. Thus in this case $\Delta_{G,B}$ is the independence complex of the graph $\Gamma_{G,B}'$.
Many SP-games, such as \textsc{Col}, \textsc{Snort}, and all \textsc{Partizan Octal} games are independence games. \textsc{NoGo } is an example of an SP-game that is not an independence game. Even though $\Gamma_{\textsc{NoGo}, B}$ is a graph for some boards (for example when $B$ is the graph on two vertices connected by an edge, {i.e}\onedot} \def\Ie{{I.e}\onedot the path of length one, $P_2$), there are many others for which this is not the case. For example, $\Gamma_{\textsc{NoGo}, P_3}$, given in \cref{fig:NogoP3aux}, has two-dimensional faces.
\begin{figure}[!ht]
\begin{center}
\begin{tikzpicture}
\filldraw[fill=gray!70, draw=black] (0, 1)--(1.73, 0)--(0, -1)-- cycle;
\filldraw[fill=gray!70, draw=black] (5.46, 1)--(3.73,0)--(5.46, -1)-- cycle;
\draw (0,1)--(5.46,1)--(1.73,0)--(5.46,-1)--(0,-1)--(3.73,0)--(0,1);
\draw (1.73,0)--(3.73,0);
\filldraw (0,1) circle (0.1cm);
\filldraw (1.73,0) circle (0.1cm);
\filldraw (0,-1) circle (0.1cm);
\filldraw (5.46,1) circle (0.1cm);
\filldraw (3.73,0) circle (0.1cm);
\filldraw (5.46,-1) circle (0.1cm);
\draw (-0.5,1) node {$x_1$};
\draw (1.2,0) node {$x_2$};
\draw (-0.5,-1) node {$x_3$};
\draw (5.96,1) node {$y_1$};
\draw (4.26,0) node {$y_2$};
\draw (5.96,-1) node {$y_3$};
\end{tikzpicture}
\end{center}
\caption{The Illegal Complex $\Gamma_{\textsc{NoGo}, P_3}$}
\label{fig:NogoP3aux}
\end{figure}
One nice property of independence games is that playing an independence game $G$ on a board $B$ is equivalent to forming independence sets of the graph $\Gamma_{G,B}'$ while Left picks vertices in $\L$ and Right in $\mathfrak{R}$.
A \textbf{flag complex} $\Delta$ is a complex whose minimal non-faces all have size 2 (see for example \cite{HH11}). In the case of independence games, since $\Gamma_{G,B}$ is a graph without isolated vertices, we have that $\Delta_{G,B}$ is flag.
Further note that the $\Gamma$-game in the case of $\Gamma$ being a graph is always an independence game (since minimal illegal positions are always pairs of pieces played). Using \cref{thm:legalinv} this implies the following.
\begin{proposition}[\textbf{iSP-Games of Flag Complexes}]\label{cor:independenceSP}
Given any SP-game $G$ and board $B$ such that $\Gamma_{G, B}$ is a non-empty graph, there exists an independence game $G'$ and board $B'$ such that $\Delta_{G,B}= \Delta_{G', B'}$. In the case that $\Gamma_{G,B}$ has no isolated vertices, we also have $\Gamma_{G,B}=\Gamma_{G',B'}$.
\end{proposition}
\begin{proof}
By \cref{thm:legalinv} there exists an iSP-game $G'$ and board $B'$ such that $\Delta_{G,B}=\Delta_{G',B'}$. In the case that $\Delta_{G,B}$ is not a simplex (if $\Gamma_{G,B}$ has at least one edge), the game $G'$ is the $\Gamma$-game $G_{\Gamma_{G,B}}$. As mentioned above, this is an independence game. In the case that $\Delta_{G,B}$ is a simplex, the game $G'$ has no illegal positions, and thus is an independence game trivially.
If $\Gamma_{G,B}$ has no isolated vertices, then the underlying rings of $\Delta_{G,B}$ and $\Delta_{G',B'}$ are the same, thus
\[\Gamma_{G',B'}=\facet{\SR{\Delta_{G',B'}}}=\facet{\SR{\Delta_{G,B}}}=\Gamma_{G,B}.\qedhere\]
\end{proof}
Equivalently, this proposition also states that given an SP-game $G$ and board $B$ such that the minimal non-faces of $\Delta_{G,B}$ are all 1- and 2-element sets, there exists an SP-game $G'$ whose legal complex is always flag and a board $B'$ such that $\Delta_{G,B}=\Delta_{G',B'}$.
As a direct consequence of \cref{cor:independenceSP}, applying \cref{thm:isomorphicgametrees}, we have that these games also have isomorphic game trees.
\begin{corollary}
Given any SP-game $G$ and board $B$ such that $\Gamma_{G, B}$ is a non-empty graph, there exists an independence game $G'$ and board $B'$ such the game trees of $(G,B)$ and $(G',B')$ are isomorphic.
\end{corollary}
\section{Further Questions and Work}\label{sec:discussion}
In this section, we will be discussing some potential further questions and avenues to explore.
The $\Gamma$-board and pieces of the $\Gamma$-game have many more vertices than $\Gamma$ itself. Thus we are interested in whether constructions of a game $G$ and board $B$ are possible for every simplicial complex $\Gamma$ without isolated vertices in which the pieces that Left and Right play occupy only one vertex so that $\Gamma=\Gamma_{G,B}$. This seems unlikely though, thus an interesting question is for which class of simplicial complexes such a construction is possible.
Similarly, we are also interested in for which simplicial complexes $\Delta$ we can find a game $G$ and board $B$ with pieces only a single vertex so that $\Delta=\Delta_{G,B}$.
A flag complex $\Delta$ is a complex whose minimal non-faces all have size 2. Thus if $\Delta$ is flag, then $\Gamma=\facet{\SR{\Delta}}$ is a graph without isolated vertices. Simplicial trees and forests, which are generalizations of graph trees and forests, are flag complexes (see \cite[Lemma 9.2.7]{HH11}), and thus game complexes. Since many properties of simplicial trees are known (see for example \cite{Faridi04} and \cite{Faridi05}) it seems that this class of flag complexes provides a good start to studying whether simpler constructions are possible.
Finally, it is of interest if each game value possible under normal play conditions is also the game value of some SP-game. This problem has received attention for specific SP-games (for \textsc{Domineering} see for example \cite{Kim96,UB15}, for \textsc{Col } and \textsc{Snort } see \cite{BCG04}), and was recently positively answered for a non-SP-game (see \cite{CS16}). Since SP-games are much easier to understand than many other combinatorial games, if the answer to this question is positive, it would provide an excellent new tool for studying combinatorial games. Whether or not this is the case, a similar, but stronger, question is if the simplest game ({i.e}\onedot} \def\Ie{{I.e}\onedot the game value) in each equivalence class containing an SP-game is itself an SP-game. Knowing that each simplicial complex is the legal complex of some SP-game has been indispensable in the exploration of those two questions (see \cite{Huntemann16}).
\bibliographystyle{plain}
|
\section{Introduction}
Nowadays, self-organizing is tremendously becoming a key feature for current and future mobile networking.
Moreover, numerous new applications and some special circumstances require the nodes/network to be self-organizing, self-configuring and self-healing. In order to overcome extreme circumstances (earthquakes, disasters, ...), massive access to the network (sport events, festival, ...) or lack of infrastructure in general, a class of self-organizing networks called Delay Tolerant Networks (DTN) \cite{fall2003delay} has been proposed and are continuously gaining interest. A DTN is a class of an infrastructure-less and fully distributed wireless networks. Such a network is designed to operate over arbitrary distances, including very small scale (e.g., cells communications) to ultra large scale (e.g., interplanetary communications). Intermittent connectivity and an excessively large delay may occur very often in such an environment, which makes the end-to-end connectivity a very challenging issue. Thus, the use of store-carry-forward paradigm seems to be an attractive solution. The main idea here is to exploit the opportunistic inter-contacts between relay nodes to cache-and-forward given data to the final destination.\\
\par DTNs-based applications are very various \cite{gao2015delay}, including digital communication for rural areas (e.g., DarkNet, TrainNet, KioskNet, etc.), personal/wildlife communications (e.g., Pollen, Body Area Networks, ZebraNet, etc.), battlefield communications (e.g., Military Missions and Airborne Networks) disaster rescues and environmental monitoring communications. Moreover, delay tolerant networks could be an attractive/efficient solution to offload legacy networks. They may help out to control the congestion caused by the exponential growth of traffic, heterogeneity in infrastructure/technologies (Device-to-Device, RFID, Drone-based backhaul, etc.), see Fig.\ref{5g}. Moreover, they are expected to be a part of the next generation radio communication system such as 4G LTE-Advanced networks, 5G and Internet of Things (IoT) \cite{prasad2015efficient, 10.4108/eai.26-10-2015.150598}. We recall also the ``Any Time, Any Where, Any Device'' (ATAWAD) paradigm which has been fueled by the prevalence of devices, exploiting collaboration between devices, enhancing functionalities and speeding up the access to data. Technically, DTNs can be helpful to assist data transmission to/from isolated regions where connectivity is hard or even impossible to establish. Moreover, the DTNs devices could be efficient to enhance Quality of Service (QoS) and to reduce energy consumption by strategically offloading the traffic from the network backbone. However, getting nodes in the network to cooperate and act as relays is a still fundamental challenge, due to limited bandwidth, limited battery energy and limited storage capacity. Indeed, participating in data transmission and content caching incurs energy consumption. Therefore, the relays may exhibit a selfish behavior, which significantly degrades the DTNs performance. The latter is a key motivation to develop and deploy efficient and distributed mechanisms to solve the inherent cooperation issue.\\
\begin{figure}[!htbp]
\centerline{\includegraphics[height=5cm,width=9cm]{./Figures/5G.eps}}
\caption{\footnotesize{DTN communications over Device-to-Device communications.}}\label{5g}
\end{figure}
\par
In order to evaluate the network performance, several metrics can be utilized. For instance, one can use the delivery rate, the content loss rate, the protocols overhead, the end-to-end delay, the expected number of transmissions and the energy consumed. Here we consider optimizing the energy consumption subject to the average delivery rate being above some given threshold. In other words, this work is devoted to presenting a new fully distributed framework for QoS support in DTN-like networks.\\
\par
Self-configuring capability and the distributed nature of DTNs are proven to induce selfish behaviour at nodes level, see \cite{karaliopoulos2009assessing,li2010routing,li2010evaluating}. Several incentive mechanisms have been designed to sustain cooperation among selfish relay nodes. Game theory seems to be the perfect tool to design such mechanisms. It is mainly viewed as a tool to investigate the decision-making by the system through an equilibrium analysis instead of optimality analysis. The authors in \cite{jiang2013survey} presented an interesting survey on incentive mechanisms for DTNs. They compared four schemes: 1) virtual currency based incentive mechanism, 2) credit-based incentive mechanism, 3) game-theory-based incentive mechanism and 4) combined incentive mechanism. In \cite{hulke2015game}, the authors discussed different game theoretical-inspired incentive mechanisms and analyzed them while pointing out their advantages and drawbacks. Interestingly, many works have proposed curious schemes to encourage nodes to cooperate. For example, the authors in \cite{el2013evolutionary} used evolutionary games theory and addressed how a reward mechanism could efficiently induce cooperation among relay nodes in delay tolerant networks. The authors in \cite{brun2014modeling} suggested a simple reward-based mechanism scheme, and show how the source could optimally set the reward value based on the relays information.\\
\par
In \cite{long2007non}, a repeated game is constructed in order to capture the interaction between mobile nodes in terms of non-cooperative power control. The authors in \cite{perlaza2012quality} introduced the framework of satisfaction form of game to model the problem of QoS provisioning in decentralized networks, and they provided a comparison between the concept of Generalized Nash Equilibrium (GNE) and the concept of Satisfaction Equilibrium (SE). To the best of our knowledge, this is the first work conducted to analyze distributed caching in DTNs under quality of service constraints using satisfaction game approach.\\
\par
The main contribution of this paper is the design of a new distributed framework for QoS provisioning in DTN-like networks using the concept of satisfaction equilibrium and a reward-based incentive mechanism. By offering some reward (virtual coins) to relay nodes, the source node can efficiently encourage them to use a part of their battery energy and participate to forward a file to some given destination, i.e, caching a given file and wait to find a persistent connection with the interested destination. Within this framework, the resulting distributed caching problem is investigated using the powerful tool of non-cooperative game theory. Notice that in this paper we use the concept of SE \cite{tian1992existence} as a solution concept instead of the well-known Nash equilibrium concept. Now, one needs not only a stable state of the game but also providing certain performance requirements. More precisely, the source has an specific problem which is offering a minimum reward ensuring that the relays delivery probability does not go bellow a threshold value, while the relays have the choice to "accept" or "reject" this offer, depending on the reward value whether it is beneficial or not. Next, we exhibit sufficient conditions for existence of an SE for both pure strategies and mixed strategies. Moreover, aiming to understand the behavior of the DTN source-relay nodes during the interaction and the eventual convergence to the SE, we propose two stochastic algorithms for both the source and the relay nodes. \\
\par
The remaining sections of the paper are organized as follows. In section \ref{MDPF}, we describe the problem, its formulation and our solution design. In Section \ref{GTM}, we present the game theoretical model including, utility functions and the SE formal definition. Section \ref{SEA} exhibits the satisfaction equilibria computation and an analysis of their efficiency. Next, we describe the stochastic learning algorithms adopted in Section \ref{LA}. Section \ref{MI} provides some numerical investigations and simulation runs to claim our work and a conclusion is drawn in Section \ref{Con}.
\section{Problem Formulation}\label{MDPF}
We consider a delay tolerant network including a pair of source-destination, and $n$ relay nodes. When a contact between the source node and a relay node takes place, the source transmits a data file to the relay node. Relay nodes, moving independently in the network area, store the file, carry it and wait until having a direct link opportunity with the destination node to forward the file. We next list the assumptions considered in this paper:
\begin{itemize}
\item The file to forward has a finite lifetime $\tau$ (called also horizon) during which the destination is interested in its content;
\item For simplicity and without loss of generality, we assume that all nodes are identical and equipped with the same wireless interface;
\item We consider that the relay nodes use a two-hop routing policy \cite{panagakis2007study}, which works as follows: when a relay node receives a copy of the file from the source, it stores it and forwards it to the destination node when met within the file's lifetime. The choice of this routing protocol is motivated by an energy efficiency purpose. Indeed such a routing scheme has a good delivery/energy efficiency trade-off;
\item Occurrence of the contacts between any two nodes follows Poisson distribution. Thus, the time interval between two successive contacts (inter-contact time) for each pair of relay nodes is exponentially distributed with a pairwise meeting rate $\lambda \geq 0$. A comprehensive discussion of this modeling can be found in \cite{chaintreau2007impact} and \cite{groenevelt2005message}. We further consider that the contact time is large enough for the complete transmission of the source's file.
\end{itemize}
The source-relay contact probability within the file's life time $\tau$ can be expressed as,
\begin{eqnarray}
p_c =p( t \leq \tau)= \int_0^{\tau} \lambda e^{-\lambda t} dt= 1-e^{-\lambda \tau}.
\end{eqnarray}
Each relar node can be in one of the three states: 1) listening (sending beacons for discovery purposes), 2) transmitting (when in contact with the destination), or 3) receiving data from the source node. In this paper, we neglect the energy consumption related to the discovery/listening state. Thus, the energy consumption per node includes the energy consumed during the receiving state (denoted $e_r$), the energy consumed while transmitting the file to the destination (denoted $e_t$) and the storing energy per time slot (denoted $e$). It follows that the mean energy $e_s$ dissipated while caching a file with lifetime $\tau$ can be written as
\begin{eqnarray}
e_s &=& \int_0^{\tau} e \lambda t e^{-\lambda t} dt\nonumber \\
&=& e\left[\frac{1- (1+\lambda \tau) e^{-\lambda \tau}}{\lambda}\right] \nonumber \\
&=& \frac{e}{\lambda}\left(1-Q_{\tau}\right).
\end{eqnarray}
where $Q_{\tau}$ denotes the probability that a given relay fails in relaying a given file to the destination \cite{altman2009competition}. Hence, the total energy consumption per node becomes $\eta =e_r+e_t+e_s$.
\subsection{Problem Formulation and Solution Design}\label{DOCP}
It is desirable that the nodes behave in a fully cooperative fashion in order to maximize the overall delivery rate. However, in real deployments, the relay nodes may not cooperate due to energy constraints and conflicting interests. Hence, a degradation of the network performance may be observed. To deal with this problem and encourage/force the relay nodes to participate in the file forwarding game, we next develop a reward-based incentive mechanism.\\
\begin{figure}[!htbp]
\centerline{\includegraphics[height=3.5cm,width=8cm]{./Figures/Reward.eps}}
\caption{\footnotesize{The Source-Relay interaction.}}\label{Reward}
\end{figure}
\par Fig.\ref{Reward} describes the interaction between the source node and the relay nodes under the proposed rewarding mechanism. On one hand, the source has the objective of making its file arrive to destination during the file's lifetime, so it generates a copy of this file and attempts to convince the relays encountered to forward it. On the other hand, the encountered relays can either accept (strategy `a') or reject (strategy `r') caching and relaying this file. In order for the long term average delivery rate of the source files to be above some given threshold (minimum QoS), the source needs to determine the appropriate reward to offer to the cooperating relays. When a source-relay contact occurs, the rational relay nodes seek to pick a strategy that maximizes its own payoff. In the meantime, the source offers the reward value merely to satisfy an individual constraint which consists of reaching a delivery rate higher that some threshold $\delta$. Moreover, when a relay node declines the forwarding offer or fails to reach the destination within the file's lifetime, it incurs a penalty in the form of a regret of declining or failing. In the next section, we construct a game theoretic framework to capture the performance of the proposed reward-based mechanism.
\section{Game Theoretical Model}\label{GTM}
Game theory has been used to solve problems in ad hoc, fixed and cellular networks. It is a powerful tool for the analysis of distributed networks. Its equilibrium concept and formulation of the utility function under constraints permit to study the system behavior and its decision strategies. It is mainly used to study the decision-making by the system through an equilibrium analysis. Indeed, in this section we first formulate our case of DTN as a homogeneous One-Shot caching Game, where the source and relays are selfish players playing independently and simultaneously. The source's strategy is the choice of the continuous-valued reward $\alpha$, taken from the interval $[0,\alpha_{max}]$, and the relays have two discrete strategies accept 'a' or reject 'r'. Then, we present the utility function which considers the proposed reward mechanism, the energy consumption, the delivery probability and the regret values. Next, we study the existence and uniqueness of equilibria in pure game when the players choose to play their pure strategies and in mixed game when they independently and randomly select their strategies; in the mixed game, each player (i.e. each relay in contact with the source) accepts with probability $p \in [0,1]$ or reject with $1-p$. Later, we give the conditions driving the system to an operating/stable point, namely a Satisfaction Equilibrium \cite{nash1950equilibrium}. \par
In fact, according to the game's concept, each player's action/strategy corresponds to a certain utility. Rationally, when a relay receives a file from the source, its utility depends on the choice of other relays' strategies. As mentioned above, the relays 'meet' the source with probability $p_c$ and all arrivals are independent. Therefore, the average number of relays among $n$ which are in contact with the source is $\tilde{n}=np_c$.
\par
Further, according to \cite{altman2009competition}, we define $\phi^i(\tilde{n})$ the delivery probability that a given relay $i$ among $\tilde{n}$, plays pure strategy 'a' and succeeds to deliver a given file ( i.e., the first to deliver the file to the destination) as
\begin{eqnarray}
\label{Psucc}
\phi^i(\tilde{n})&=& (1-q_{\tau}) \sum_{j=1}^{\tilde{n}} {\tilde{n}-1 \choose j-1} \frac{(1-q_{\tau})^{\tilde{n}-1} q_{\tau}^{\tilde{n}-j}}{j}\nonumber\\&=&\frac{1-q_{\tau}^{\tilde{n}} }{\tilde{n}}.
\end{eqnarray}
Next, we will formulate the utility function for each player.
\subsection{Utilities Formulation}
We define the utility function of each relay as the difference between the reward that it can win from the source and the energy consumed to cooperate with the source. Thus, we denote $U_i(\mbox{'a'},\alpha,,n_a)$ the utility function of a given relay when it plays its pure strategy accept 'a', and $U_i(\mbox{'r'},\alpha,,n_a)$ when it plays its pure strategy reject 'r'.
\begin{equation}
\label{ru}
\begin{cases}
& \mathcal{U}_i(\mbox{'a'},\alpha,n_a)= \alpha \phi^i(n_a) -\sigma(1- \phi^i(n_a))-\eta\\ \\
&\mathcal{U}_i(\mbox{'r'},\alpha,n_a)=-\alpha \phi^i(n_a)-\gamma \\
\end{cases}
\end{equation}\par
where $\sigma$ and $\gamma$ are appropriate constants (regrets) that the relay incurs when it accepts to cache a given file but it does not succeed to deliver it during its lifetime and when it declines the source offer respectively. $n_a$ denotes the number of relays out of $\tilde{n}$, that are have been in contact with the source and that accepted to cache .\par
Furthermore, because of the selfish behavior, each relay decision is based on self-optimisation; the relay's objective is then to maximize its own utility in a distributed fashion, i.e.
\begin{equation*}
\max_{v \in \{a,r\}} \mathcal{U}_i(\mbox{'v'},\alpha,n_a), \quad \forall i =0, 1, \cdots, n_a.
\end{equation*}\par
The utility function of source $\mathcal{U}_s(\mbox{'a'},,n_a)$ is merely the delivery probability of all relays that have accepted caching.
\begin{equation*}
\mathcal{U}_s(\mbox{'a'},n_a)=n_a \phi^i(n_a), \quad \forall i =0, 1, \cdots, n_a.
\end{equation*}
\subsection{Satisfaction Equilibrium}
Now, we present a definition for the Nash Equilibrium \cite{nash1950equilibrium}, which is the point where no player can improve his payoff by making individual changes in his decisions. Precisely, in our source-relays game we introduce an equilibrium called satisfactory equilibrium \cite{tian1992existence}: the source seeks to satisfy the constraint that its 'well-being' should be greater than a fixed value (threshold), given the strategies adopted by the relays, which behave rationally to maximize their ``well-being'' by playing their optimal strategies.
\begin{definition}
At satisfaction Nash equilibrium $(p^*,\alpha^*)$ we have,
\begin{equation*}
E[\mathcal{U}_s(\mbox{'a'},n)] \geq \delta, \quad \forall \alpha \in [0,\alpha_{max}],
\end{equation*}
and
\begin{equation*}
\displaystyle p^* \in \arg\!\max_{p \in [0,1]}E[ \mathcal{U}_i (p,\alpha^*,n)], \quad \forall i =0, 1, \cdots, n_a.
\end{equation*}
\end{definition}\par
Next, we will perform a thorough analysis of SE, its existence and uniqueness in pure an mixed strategies. Moreover, we discuss the full conditions that drive our distributed system to this steady point, if there exists one.
\section{Satisfaction equilibrium analysis}\label{SEA}
We analyze here the structure of the content caching game solutions and we derive many sufficient conditions for the existence of a SE.
\begin{theorem}
(Nash's Theorem \cite{nash1950equilibrium}) Every finite game in strategic form (i.e., with finite number of players and finite number of pure strategies for each player) has at least one Nash
equilibrium (NE) (involving pure or mixed strategies).
\end{theorem}
Since the condition $\alpha(\delta)$ is imposed on the game, the NE is not a suitable solution. Along this paper we replace the Nash equilibrium by the Satisfaction Equilibrium which is more natural.
\begin{remark}
We highlight that the existence of pure / mixed SE does not necessarily imply its uniqueness. In fact, the existence of the constraint $\alpha(\delta)$ yields the fact to provide conditions to have a unique SE very difficult.
\end{remark}
\subsection{Pure Satisfaction Equilibria (PSE)}
We turn now to derive the Satisfaction Equilibria for pure strategy case. The players act with their pure strategies.
\begin{lemma}The content caching game may have numerous PSE. Satisfaction Equilibria are any ($\alpha^*$, $n_a ^*$) solutions of the following two conditions:
\begin{equation}
n_a ^*\geq \frac{\log(1-\delta)}{log(q_{\tau})}
\end{equation}
and,
\begin{align}
\alpha^*=\frac{\lambda\sigma(n-1+q_{\tau}^{n_a})-n_a(\lambda(\gamma-e_r-e_t)-e (1-q_{\tau}))}{2\lambda(1-q_{\tau}^{n_a})}.
\end{align}
\end{lemma}
\begin{proof}
Assume that the profile $ \mathcal{J}=(\overbrace{\mbox{'a'},\mbox{'a'},...,\mbox{'a'}}^{n_a},\overbrace{\mbox{'r'},\mbox{'r'}....,\mbox{'r'}}^{n-n_a})$ is a satisfaction equilibrium. Then
\begin{equation*}
\begin{cases}
& \mathcal{U}_i(\mbox{'a'},\alpha,n_a) \geq \mathcal{U}_i(\mbox{'r'},\alpha,n_a) ,\quad \forall i =0, 1, \cdots, n_a \\ \\
&\mathcal{U}_i(\mbox{'a'},\alpha,n_a) \leq \mathcal{U}_i(\mbox{'r'},\alpha,n_a), \quad \forall i =0, 1, \cdots, n-n_a \\
\end{cases}
\end{equation*}
Hence, the relays are indifferent between the two strategies \mbox{'a'}, \mbox{'r'}, then
\begin{equation*}
U_i(\mbox{'a'},\alpha^*,n_a) = U_i(\mbox{'r'},\alpha^*,n_a),
\end{equation*}
\begin{equation*}
\phi^i(n_a) -\sigma(1- \phi^i(n_a))-\eta= -\alpha \phi^i(n_a)-\gamma.
\end{equation*}
After a few lines of algebra, we find
\begin{align*}
\alpha^*=\frac{\lambda\sigma(n-1+q_{\tau}^{n_a})-n_a(\lambda(\gamma-e_r-e_t)-e (1-q_{\tau}))}{2\lambda(1-q_{\tau}^{n_a})}.
\end{align*}
The source problem consists of assuring a succeed relays' probability value
\begin{equation*}
n_a\phi^i(n_a) \geq \delta,
\end{equation*}
then,
\begin{equation*}
1- q_{\tau}^{n_a} \geq \delta, \quad \Longrightarrow \quad n_a ^*\geq \frac{\log(1-\delta)}{log(q_{\tau})}
\end{equation*}
\end{proof}\par
The pure equilibrium could fail to achieve a certain lucidity between relay nodes since only a part of relays may accept to cache the file. To solve this problem, we use another concept of equilibrium, named Mixed Satisfaction Equilibrium, in which the relay will accept to cache the file with some probability.
\subsection{Mixed Satisfaction Equilibria (MSE)}
When mixed strategy is allowed, the relays randomize between accepting and rejecting the source offer according to common probability distribution, accepting with $p$, rejecting with $1-p$.
\begin{lemma}The satisfactory caching game has infinitely many Mixed Satisfaction Equilibria ($\alpha^*$, $p^*$). They are solutions of the following system:
\begin{equation}
\begin{cases}
& \alpha^*=\frac{\lambda\sigma(n-1+(1-(1-q_{\tau})p_cp^*)^{n})-n(\lambda(\gamma-e_r-e_t)-e (1-q_{\tau}))}{2\lambda(1-(1-(1-q_{\tau})p_cp^*)^{n})}.\\ \\
& p^* \geq \frac{1- \sqrt[n]{1- \delta}}{(1-q_{\tau})p_c} \\
\end{cases}
\end{equation}
\end{lemma}
\begin{proof}
At the equilibrium each relay is indifferent about which strategy to choose. Namely
\begin{equation*}
\mathcal{U}_i(\mbox{'a'},\alpha^*,n)=\mathcal{U}_i(\mbox{'r'},\alpha^*,n),
\end{equation*}
\begin{equation}
\label{eq}
\alpha^* \phi^i(n) -\sigma(1- \phi^i(n))-\eta= -\alpha^* \phi^i(n)-\gamma,
\end{equation}
where
\begin{eqnarray}
\phi^i(n)&=& z \sum_{j=1}^{n} {n-1 \choose j-1} \frac{z^{n-1} (1-z)^{n-j}}{j}\nonumber\\&=&\frac{1-(1-z)^n }{n},\nonumber
\end{eqnarray}
with $z=p_cp(1-q_{\tau})$. Next, after some algebras from (\ref{eq}) we obtain
\begin{align}
\alpha^*=\frac{\lambda\sigma(n-1+(1-z^*)^{n})-n(\lambda(\gamma-e_r-e_t)-e (1-q_{\tau}))}{2\lambda(1-(1-Z^*)^{n})}.
\end{align}
where $z^*=(1-q_{\tau})p_cp^*$.\\
At Nash equilibrium, The source's objective is
\begin{equation}
n\phi^i(n) \geq \delta,
\end{equation}
then
\begin{equation}
n\frac{1-(1-(1-q_{\tau})p_cp^ *)^n }{n}\geq \delta,
\end{equation}
after some algebras we obtain,
\begin{equation}
p^* \geq \frac{1- \sqrt[n]{1- \delta}}{(1-q_{\tau})p_c}
\end{equation}
\end{proof}
From now on, we refer to the case of strict equality as the Efficient Satisfaction Equilibrium (ESE). Indeed this point correspond to the minimum satisfaction level (QoS threshold) of the source node which means there will be no incentive for the source node to deviate unilaterally. This also correspond to a minimum stable accepting probability such that the whole source and relay nodes have no incentive to change their decision.
\subsection{Efficient Satisfaction Equilibria (ESE)}\label{ESE}
This section exhibits some properties of the efficient satisfaction equilibrium defined in the previous subsection. We consider and extend the Pareto-efficiency \cite{webb2007game},\cite{fudenberg1991game} as criterion to discuss and investigate its efficiency. Notice that the ESE is not always Pareto optimal.
\begin{definition}
A equilibrium is said to be strongly Pareto-optimal, if no player's payoff can be increased without decreasing the payoff of another player, i.e. $\nexists (\mbox{p'},\mbox{$\alpha$'}) $, such
\begin{eqnarray*}
\forall i \in n, E(\mathcal{U}_i(\mbox{p'},\mbox{$\alpha$'},n)] \geq E[\mathcal{U}_i( p^*,\alpha^*,n)] \nonumber \\
\text{and} \quad \exists j \in n, E(\mathcal{U}_j(\mbox{p'},\mbox{$\alpha$'},n)] > E[\mathcal{U}_j(p^*,\alpha^*,,n)] \nonumber \\
\end{eqnarray*}
\end{definition}
\begin{proposition}
The ESE of the induced non-cooperative, symmetric One-shot caching game is Pareto optimal.
\end{proposition}
\begin{proof} In order to prove the strong Pareto-optimality, it is enough to show that for any couple $(\mbox{p'},\mbox{$\alpha$'})$ no strictly higher payoff can be obtained, without decreasing the payoff of other players. Let us, assume that $\exists (\mbox{p'},\mbox{$\alpha$'}) $ such as,
\begin{equation*}
\begin{cases}
& \mbox{p'}=\beta p^*, \quad \beta >1 \\
&\mbox{$\alpha$'}=\psi \alpha^*, \quad 0<\psi < 1
\end{cases}
\end{equation*}
The meaning of this configuration: The source could increase its utility, if reward value decreases and the acceptance probability increases so,
\begin{eqnarray*}
E(\mathcal{U}_s( \mbox{p'},n)]&=&1-(1-(1-q_{\tau})p_c \mbox{p'})^n \nonumber \\
&=&1-(1-(1-q_{\tau})p_c \beta p^* )^n \nonumber \\&&
> 1-(1-(1-q_{\tau})p_c p*)^n \nonumber \\
\iff E(\mathcal{U}_s( \mbox{p'},n)] &> &E(\mathcal{U}_s( p^*,n)]
\end{eqnarray*}
, i.e., the configuration increases the well-being of the source. However, this configuration can also decrease the well-being of the relays
\begin{eqnarray*}
&& E(\mathcal{U}_i(\mbox{p'},\mbox{$\alpha$'},n))=\mbox{p'} D(\mbox{p'},n_a) \mathcal{U}_i(\mbox{'a'},\mbox{$\alpha$'},n_a) \nonumber \\ & & + (1-\mbox{p'})D(\mbox{p'},n_a) \mathcal{U}_i(\mbox{'r'},\mbox{$\alpha$'},n_a) < E(\mathcal{U}_i(p^*,\alpha^*,n)) ,
\end{eqnarray*}
with $D(\mbox{p'},n_a)=\sum_{n_a=0}^{n-1} {n-1 \choose n_a}\mbox{p'}^{n_a}(1-\mbox{p'})^{n-n_a-1}$\\
We cannot improve the utility of the source without decrease the relays' utility, which contradicts the Pareto optimal definition. Consequently, the ESE is strong Pareto optimality.
\end{proof}
\section{Learning Algorithms}\label{LA}
In this, we present our proposal learning algorithms to attend the ESE discussed in previous section. In fact, we will give a formally description of the two Stochastic learning algorithm. Indeed, the stochastic learning technique has been successfully used in Distributed system, particularly in wireless networks. Briefly, At each iteration, the automatons uses only the estimated value of their payoff to update their action value till converge to their unique best response. Indeed, the source's strategy is a independent, continue action which is select a reward value $\alpha \in [0,\alpha_{max}]$, based on local observations, satisfying that the file-caring-relays succeed with probability greater then $\delta$. Precisely, the source does not need the strategies of the relays, it observes only its own payoff and its reward value assigned. Consequently, we propose a satisfactory stochastic learning algorithm based on \cite{sabir2009stochastic} which will lead the source to its optimal decision.
\begin{algorithm}
\KwResult{Satisfaction equilibrium reward value $ \alpha^*$ }
\textbf{Initialization}\; \par
Assign a value for $\alpha \in [0,\alpha_{max}]$.\\
Expected payoff $\mu^*$.\\
\textbf{ Learning pattern:} For each iteration k\\
Observe the estimate value of payoff $\hat{Us}$
\[
\hat{U_s}^{k+1}=\hat{U_s}^{k}+\epsilon_{k+1}(\mu^*-\hat{U_s}^{k})
\]
\[
\alpha_{k+1}= \max(\min(\alpha_{max},\alpha+\epsilon_{k+1} (\mu^*- \hat{U_s}^{k+1})),0)
\]
\caption{Source satisfactory equilibrium stochastic learning algorithm}
\end{algorithm}\par
As regards the relays which will take the decision locally and independently based their probability distribution, according to reward value $ \hat{\alpha}$ estimated by source, so we propose imitative \textbf{CO}mbined fully \textbf{DI}stributed \textbf{PA}yoff and \textbf{S}trategy (CODIPAS) \cite{tembine2012distributed}. This choice of CODIPAS is justified by its functionality such as, the player only need to observe the realization of their utility during previous iterations. They play independently their strategies based on the outdated observation.
\begin{algorithm}
\textbf{Initialization:}\; for each relay $i \in n$ do \par
$\hat{U}_{i,0}^a,\hat{U}_{i,0}^r, p_{i,0}$\\
Define the sequence up to T : $M_{i,k}^a, M_{i,k}^r, L_{i,k}^a, L_{i,k}^r$ for $ k \in \{1,....,T\}$. \\
\textbf{ Learning pattern:} for each relay $i \in n$ do
\[
\hat{U}_{i,k+1}^a=\hat{U}_{i,k+1}^a+M_{i,k}^ae_{i,k}(U_{i,k}-\hat{U}_{i,k}^a)
\]
\[
\hat{U}_{i,k+1}^r=\hat{U}_{i,k+1}^r+M_{i,k}^a(1-e_{i,k})(U_{i,k}-\hat{U}_{i,k}^r)
\]
\[
p_{i,k+1}= \frac{p_{i,k}(1+L_{i,k}^a)^{\hat{U}_{i,k}^a}}{p_{i,k}(1+L_{i,k}^a)^{\hat{U}_{i,k}^a}+(1-p_{i,k})(1+L_{i,k}^r)^{\hat{U}_{i,k}^r}}
\]
\caption{Imitative CODIPAS}
\end{algorithm}
where $e_k$ is the unit vector with the $ i^{th}$ component unity corresponding to the action selected at $k$
\section{Numerical Investigations}\label{MI}
We present here some numerical examples evaluating the performance of our contributions. We consider the following setting: $\delta=0.21$, $\sigma=0.2$ , $\gamma=0.15$ , $e=3.8\times10^{-5}$, $ e_r=2\times10^{-5}$ $e_t=2\times10^{-5}$, $\alpha_{max}=5, n=7$, and we depict the behaviour of the source node and the
relay nodes while varying the horizon $\tau$ (file lifetime), the parameter $\lambda$ that stands for the contact rate and the number of the relay nodes participating in transmission.
\begin{figure}[!htbp]
\centerline{\includegraphics[height=5cm,width=9cm]{./Figures/alpha_h.eps}}
\centerline{(a)}
\centerline{\includegraphics[height=5cm,width=9cm]{./Figures/p_h.eps}}
\centerline{(b)}
\caption{\footnotesize{Reward value and acceptance probability as function of file lifetime and contact rate}}\label{fig1}
\end{figure}
\par
We depict in Fig.\ref{fig1} (a) and Fig.\ref{fig1} (b) the acceptance probability and the reward value while varying the file lifetime for several values of the contact rate at SE. We notice that the relay nodes are cooperative and assist the file forwarding while the source offers high value of reward. Namely, the relay nodes tend to accept with probability $1$. However, when the source decreases the reward the relay nodes decrease automatically their probability to cooperate, which is quite intuitive. This can be explained by the behaviour of the source and relays nodes, such as the source objective is offering minimum value of reward and ensuring its constraint to attain desired delivery probability $\delta$ , while the relay nodes have a problem of trade-off between the reward offered, energy and file lifetime. Indeed, the relay has benefit to accept caching the file as its expected lifetime and the probability to contact the destination are low, because the source is willing to give high value of reward. \par
Fig.\ref{fig2} (a) and Fig.\ref{fig2} (b) show the impact of the number of relay nodes on the acceptance probability for different value of file lifetime and contact rate at SE, this probability decreases as the number of relays increases because the existence of several opponent can decrease the delivery probability of each relay nodes, hence their acceptance probability decreases afraid of to accept and fail to delivery and incur a punishment. Moreover, the influence of $\lambda$ and $\tau$ is illustrated. Acceptance probability decreases as long as $\lambda$, $\tau$ increase, this can be easily explained by the reward offered by the source node which tries to minimum the reward value, then as $\lambda$, $\tau$ increase this value of reward is not more beneficial, so it does not cover the transmission cost of the relays. Consequently, they relays tend to not cooperate.
\begin{figure}[!htbp]
\centerline{\includegraphics[height=5cm,width=9cm]{./Figures/phisto_h.eps}}
\centerline{(a)}
\centerline{\includegraphics[height=5cm,width=9cm]{./Figures/phisto_n.eps}}
\centerline{(b)}
\caption{\footnotesize{ Acceptance probability as function of number of relays, file lifetime and contact rate}}\label{fig2}
\end{figure}
We depict in Fig.\ref{fig4} (a) the delivery probability as function of the file lifetime and contact rate at SE, the figures illustrate the satisfaction regions where the source attains its objective, so while increasing the file lifetime the delivery probability increases till converges to the desired value $\delta$. Hence, for each value of file lifetime $\lambda(\tau)$ corresponds a value of contact rate $\lambda(\tau^*)$ where the satisfaction regions of source begin. The same remark for the contact rate in Fig.\ref{fig4} (b) for each value of file lifetime $\tau(\lambda)$ corresponds a value of contact rate $\tau(\lambda^*)$ where the satisfaction regions of source begin. \par
\begin{figure}[!htbp]
\centering
\centerline{\includegraphics[height=5cm,width=9cm]{./Figures/Psucc.eps}}
\centerline{(a)}
\centerline{\includegraphics[height=5cm,width=9cm]{./Figures/Psucc_lam.eps}}
\centerline{(b)}
\caption{\footnotesize{ Delivery probability as function of file lifetime and contact rate}}\label{fig4}
\end{figure}
\begin{figure}[!htbp]
\centering
\centerline{\includegraphics[height=3cm,width=10cm]{./Figures/alpha.eps}}
\centerline{(a)}
\centerline{\includegraphics[height=3cm,width=10cm]{./Figures/p.eps}}
\centerline{(b)}
\centerline{\includegraphics[height=3cm,width=10cm]{./Figures/n_a.eps}}
\centerline{(c)}
\caption{\footnotesize{Seeking the pure Satisfaction Equilibrium}}\label{fig5}
\end{figure}
In Fig.\ref{fig5} we consider a scenario involving seven relays with symmetric contact rate and file lifetime $ \lambda=0.015, \tau=100$. In fact, this figure depicts the behavior of the proposed learning algorithms over time and how they converge to the pure SE, such as Fig.\ref{fig5} (a) shows the convergence of the source satisfactory equilibrium stochastic learning algorithm to the optimal reward value $\alpha^*$. The Fig.\ref{fig5} (b) stands for the convergence of the relays algorithm (Imitative CODIPAS) to the pure optimal acceptance probability $p^* \in \{0,1\}$ and the Fig.\ref{fig5} (c) shows the number of relays that accept to cache, i.e the number of relay that play the pure action "accept". \par
\begin{figure}[!htbp]
\centerline{\includegraphics[height=3.2cm,width=10cm]{./Figures/alphaMixte.eps}}
\centerline{(a)}
\centerline{\includegraphics[height=3.2cm,width=10cm]{./Figures/pmixte.eps}}
\centerline{(b)}
\caption{\footnotesize{Seeking the mixed Satisfaction Equilibrium}}\label{fig7}
\end{figure}
Moreover, in Fig.\ref{fig7} we considered a scenario including three relays with symmetric contact rate and file lifetime $ \lambda=0.015, \tau=100$ and source with with different target value of delivery probability $\delta \in \{0.02, 0.48, 0.65, 0.85\}$. Hence, the Fig.\ref{fig7} (a) illustrates the convergence of the source satisfactory equilibrium stochastic learning algorithm to the optimal reward value $\alpha^*$ that it is willing to offer for the purpose that the relays accept to cooperate with high probability for different value of $\delta$. Whereas, the Fig.\ref{fig7} (b) shows the different acceptance probability where the relays algorithm converge for different value of reward. Precisely, the two algorithms converge independently and simultaneously to the symmetric pair vector $(\alpha^*,p^*)$ for each value of target delivery probability desired by the source.
\section{Conclusion}\label{Con}
We investigated the support of QoS in DTN-like networks under energy/reward trade-off. We formulated the interaction between a source node and a set of relay nodes as a non-cooperative satisfaction game. Full satisfaction equilibria characterization for both pure and mixed strategies were provided. Then, we proposed two fully distributed learning algorithms to guarantee discovery of the source/relays satisfactory strategies without knowledge of any external information. The SE ensures to the source node to meet its target delivery probability while maximizing the payoff functions of the relay nodes.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
This paper deals with the effective dynamics of a two dimensional condensate
of $N$ interacting bosons.
Fundamentally, the evolution of the system is described by a time-dependent wave-function $\Psi_t \in L^2_{s}(\mathbb{R}^{2N}, \mathbb{C}),
\| \Psi_t \|=1$. Here and below norms without index $\|\cdot\|$ always denote the $L^2$-norm on the appropriate Hilbert space.
$L^2_{s}(\mathbb{R}^{2N}, \mathbb{C})$ denotes the set of all $\Psi \in L^2(\mathbb{R}^{2N}, \mathbb{C})$
which are symmetric under pairwise permutations of the variables
$x_1, \dots, x_N \in \mathbb{R}^2$. If in addition $\Psi_t \in H^2 (\mathbb{R}^{2N},\mathbb{C})$ holds, then $\Psi_t$ solves the $N$-particle
Schr\"odinger equation
\begin{equation}\label{schroe} \text{i}\partial_t \Psi_t = H_U\Psi_t \end{equation}
where the (non-relativistic) Hamiltonian $H_U:H^2 (\mathbb{R}^{2N},\mathbb{C}) \rightarrow L^2 (\mathbb{R}^{2N},\mathbb{C})$ is given by
\begin{equation}\label{hamiltonian}
H_U=-\sum_{j=1}^N \Delta_j+\sum_{1\leq j< k\leq N} U(x_j-x_k) +\sum _{j=1}^N A_t(x_j)
\; .
\end{equation}
In this paper, we will consider two choices for the interaction potential $U$.
\begin{itemize}
\item
Let $U(x) = V_N(x)=e^{2N} V(e^N x) $ for a compactly supported, spherically symmetric, positive potential $V\in
L^\infty_c (\mathbb{R}^2, \mathbb{R}) $.
Below, we will motivate this choice of $V_N$.
Note that, in contrast to existing dynamical mean-field results, $\| V_N \|_1= \mathcal{O}(1)$ does not decay like $1/N$.
\item
Let, for any $\beta>0$,
$U(x)= W_\beta(x)= N^{-1+2 \beta} W(N^\beta x)$ for a compactly supported, spherically symmetric, positive potential $W \in L^\infty_c (\mathbb{R}^2, \mathbb{R}) $.
This scaling can be motivated by formally imposing that that the total potential energy is of the same order as the total kinetic energy, namely of order $N$, if $\Psi_0$ is close to the ground state.
\end{itemize}
The physical situation we are interested in is the description of a Bose-Einstein condensate.
Therefore, one has to start with an initial wave function $\Psi_0$ where the correlations between particles are small.
Define the one particle reduced density matrix $\gamma^{(1)}_{\Psi_0}$ of $\Psi_0$ with integral kernel
$$\gamma^{(1)}_{\Psi_0}(x,x')=\int_{\mathbb{R}^{2N-2}} \Psi_{0}^*(x,x_2,\ldots,x_N)\Psi_{0}(x',x_2,\ldots,x_N)d^2x_2\ldots d^2x_N \;. $$
We assume complete condensation in the limit of large particle number $N$, which amounts to
assume that, for $N \rightarrow \infty$,
$\gamma^{(1)}_{\Psi_0} \rightarrow |\phi_0\rangle\langle\phi_0|$ in trace norm for some $\phi_0
\in L^2(\mathbb{R}^2,\mathbb{C})
, \|\phi_0\|=1$.
Our main goal is to prove that the time evolution of the gas can in good approximation be described by an effective one particle time evolution which describes the the bulk condensate.
This is of particular interest in experiments if one switches off the trapping potential $A_t$ and monitors the dynamics of the condensate.
In particular, we show that the time evolved reduced density matrix $\gamma^{(1)}_{\Psi_t}$
converges to $ |\phi_t\rangle\langle\phi_t|$ in trace norm as $N \rightarrow \infty$ with convergence rate of order $N^{-\eta}$ for some $\eta>0$.
Here, $\phi_t$ is the unique solution of the nonlinear equation
\begin{equation}\label{GP} \text{i} \partial_t
\phi_t=\left(-\Delta +A_t\right) \phi_t+
b_U
|\phi_t|^2\phi_t=:h_{ b_U}^{GP}\phi_t
\end{equation}
with initial datum $\phi_0$. Depending on the interaction potential, we obtain the effective coupling constant $b_U= N\|W_\beta\|_1$ in the case $U= W_\beta$ and $b_U= 4\pi$ in the case $U= V_N$, respectively.
Note that $N\|W_\beta\|_1= \|W\|_1$, hence equation \eqref{GP} does not depend on $N$.
We like to remark that it is well known that convergence of $\gamma^{(1)}_{\Psi_t}$
to $ |\phi_t\rangle\langle\phi_t|$ in trace norm is equivalent to the respective convergence in operator norm since $|\phi_t\rangle\langle\phi_t|$ is a rank-1-projection, see
Remark 1.4. in \cite{rodnianskischlein}. Furthermore, the convergence of the one-particle reduced density matrix
$\gamma^{(1)}_{\Psi_t} \rightarrow| \phi \rangle \langle \phi |$ in trace norm
implies convergence of any $k$-particle reduced density matrix $\gamma^{(k)}_{\Psi_t}$ against $| \phi^{\otimes k} \rangle \langle \phi^{\otimes k} |$ in trace norm as $N \rightarrow \infty$ and $k$ fixed, see for example \cite{knowles}.
We explain our result further. For this, we first like to recall known results from the three dimensional Bose gas. There, one considers the interaction potential to be given by $V_{\beta}(x)
=N^{-1+3\beta} V(N^{\beta} x)$ for $0 \leq \beta \leq 1$.
For $0<\beta<1$, one obtains the cubic nonlinear Schr\"odinger equation with coupling constant $\|V\|_1$. This can be seen as a singular mean-field limit, where the full interaction is replaced by its corresponding mean value
$\int_{\mathbb{R}^3} d^3 y N^{3\beta} V(N^\beta (x-y)) | \phi_t (y) |^2 \rightarrow \|V\|_1 | \phi_t (x) |^2$. For $\beta=1 $, however, the system develops correlations between the particles which cannot be neglected. In this regime, the correct mean field coupling is then given by
$ 8 \pi a$, where $a$ denotes
the scattering length of the potential $\frac{1}{2}V$.
We now want to sketch out the differences of a two dimensional condensate.
Let us first explain, why the short scale correlation structure can be neglected in the case that the potential is given
by $W_{\beta}(x)
=N^{-1+2\beta} W(N^{\beta} x)$ for any $\beta>0$.
Assuming that the energy of
$\Psi_t$ is comparable to the ground state energy, the wave function will develop short scale correlations between the particles.
Heuristically, one may think of $\Psi_t$ of Jastrow-type, i.e.
$\Psi_t (x_1, \dots,x_N) \approx \prod_{i<j} F(x_i-x_j) \prod_{k=1}^ N \phi_t (x_k) $.
\footnote{
One should however note that $\Psi_t$ will not be close to a full product $ \prod_{k=1}^ N \phi_t (x_k) $ in norm. Rather, it can be approximated by a quasifree state satisfying a Bogoloubov-type dynamics, see \cite{SchleinNorm}, \cite{marcin1} and \cite{marcin2} for precise statements. }
The function $F$ accounts for the pair correlations between the particles at short scales of order $N^{-\beta}$.
It is well known that the correlation function $F$ should be described by the zero energy scattering
state $j_{N,R}$ of the potential $\frac{1}{2}W_{\beta}$, where $j_{N,R}$ satisfies
\begin{equation*}
\begin{cases}
\left( - \Delta_x + \frac{1}{2} W_\beta(x) \right) j_{N,R}(x)=0,
\\
j_{N,R}(x)=1 \; \text{for } |x| = R.
\end{cases}
\end{equation*}
Here, the boundary radius $R$ is chosen of order $N^{-\beta}$. That is, $F(x_i-x_j) \approx j_{N,R}(x_i-x_j)$ for $|x_i-x_j| =\mathcal{O}
(N^{-\beta})$ and $F(x_i-x_j) =1$ for $|x_i-x_j| \gg \mathcal{O}
(N^{-\beta})$.
Rescaling
to coordinates $y=N^\beta x$, the zero energy scattering state satisfies
\begin{equation}
\label{scaling}
\left(
-\Delta_y+\frac{1}{2}N^{-1}W(y)
\right)
j_{N,N^\beta R}(y)=0\;.
\end{equation}
Due to the factor $N^{-1}$ in front of $W$, the zero energy scattering equation is almost constant, that is $j_{N,R}(x) \approx 1$, for all $|x| \leq R$.
As a consequence, the microscopic structure $F$, induced by the zero energy scattering state, vanishes for any $\beta>0$ and does not effect the dynamics of the reduced density matrix $\gamma^{(1)}_{\Psi_t}$.
Assuming $\gamma^{(1)}_{\Psi_0} \approx|\phi_0\rangle\langle\phi_0|$, one may thus conclude that the interaction on each particle is then approximately given by its mean value
\begin{align*}
\int_{\mathbb{R}^2} d^2y N W_\beta( x-y) |\phi_t|^2(y)
\rightarrow\|W\|_1 | \phi_t|^2(x)
\;,
\end{align*}
and one obtains the correct coupling in the effective equation \eqref{GP} in the case $U(x)=W_\beta(x)$.
Let us now consider the case for which the dynamics of $\Psi_t$ is generated by the Hamiltonian $H_{V_N}$.
If one would guess the effective coupling of $\phi_t$ to be also given by its mean value, one would end up with the $N$-dependent equation
$\text{i} \partial_t
\phi_t=\left(-\Delta +A_t\right) \phi_t+
N \|V \|_1
|\phi_t|^2\phi_t$. That is, the naive guess of the mean field to be $ N \|V_N \|_1=N \|V \|_1$ yields to a dynamics where the coupling constant of the self interaction differs from its correct value $4 \pi$ by a factor of $\mathcal{O}(N)$.
The reason is that now it is important to take the correlations into account, as
in the case $\beta=1$ for the three dimensional system.
Namely, the scaling of the potential yields to
$j_{N,R}(x)=j_{0,e^{N}R}(e^{N}x)$. Therefore, the microscopic structure needs to be taken into account explicitly. One obtains as an effective coupling
\begin{align*}
\int_{\mathbb{R}^2} d^2y N V_N( x-y)j_{N,R}(x-y) |\phi_t|^2(y)
\rightarrow 4 \pi | \phi_t|^2(x)
\end{align*}
regardless of the shape of the potential $V$.
This observation follows from
\begin{align*}
\int_{\mathbb{R}^2} d^2x V_N( x)j_{N,R}(x)= \frac{4\pi}{ \ln \left(\frac{R}{ae^{-N}} \right)}
\approx 4 \pi N^{-1}
\; ,
\end{align*}
where $a$ is the scattering length of the potential $V$ and $a e^{-N}$ denotes the scattering length of the scaled potential $V_N$ (see Section \ref{secmic} for a detailed discussion).
The logarithmic dependence of the integral above on $a$ is special in two dimensions and this explains the exponential scaling of $V_N$.
We like to remark that it is easy to verify that, for any $s>0$, the potential
$V^{(s)}_N(x) \equiv V_{sN}(x)= e^{2Ns} V(e^{Ns}x)$ yields to an effective coupling $4 \pi / s$. For the sake of simplicity, we will not consider this slight generalization, although our proof is also valid in this case.
The derivation of effective evolution equations is well known in the literature, see e.g. \cite{SchleinNorm, benedikter, erdos1, erdos2, erdos3, erdos4, knowles, marcin1, marcin2,
pickl2, picklgp3d, pickl1, rodnianskischlein}
and references therein.
For the two-dimensional case we consider, it has been proven, for $0 <\beta <3/4$ and $W$ nonnegative, that $\gamma^{(1)}_{\Psi_t}$
converges to $ |\phi_t\rangle\langle\phi_t|$ as $N \rightarrow \infty$ \cite{schlein2d}.
For $0< \beta < 1/6$, it has been proven in \cite{chen2d} that the reduced density matrices converge, assuming that the potential $W$ is attractive, i.e. $W \leq 0$. This result was later extended to $0<\beta <3/4$, using stability properties of the ground state energy \cite{lewin}.
One approach which relates more closely to the experimental setup is to consider a three-dimensional gas of Bosons which is strongly confined in one spatial dimension. Then, one obtains an effective two dimensional system in the unconfined directions. We remark that in this dimensional reduction two limits appear, the length scale in the confined direction and the scaling of the interaction in the unconfined directions.
Results in this direction can be found in \cite{abdallah} and \cite{keler}, see also \cite{teufel}.
It is of course also important to study the problems aforementioned from a static perspective.
For known results regarding the ground state properties, we refer to the monograph \cite{lssy}.
Our proof is based on \cite{picklgp3d}, where the emergence of the Gross-Pitaevskii equation was proven by one of us (P.P.) in three dimensions for $\beta=1$. In particular, we adapt some crucial ideas which allow us to control the microscopic structure of $\Psi_t$.
We shall shortly discuss the physical relevance of the different scalings. On the first view, the interactions discussed above do look rather unphysical. It is questionable to assume that the coupling constant and/or the range of the interaction change as the particle number increases.
Nevertheless, one can think of situations,
where for example the support of the interaction is small and the particle number of the system is adjusted accordingly.
The exponential scaling $V_N(x)= e^{2N} V(e^N x)$ is special. In this case it is possible to rescale space- and time-coordinates in such a way that in the new coordinates the interaction is {\it not} $N$ dependent. Choosing $y=e^Nx$ and $\tau= e^{2N}t$ the Schr\"odinger equation reads
$$
\text{i}\frac{d}{d\tau} \Psi_{e^{-2N}\tau}=\left(-\sum_{j=1}^N \Delta_{y_j}+\sum_{1\leq j< k\leq N} V(y_j-y_k) +\sum _{j=1}^N A_{e^{-2N}\tau}(e^{-N}y_j)\right)\Psi_{e^{-2N}\tau}\;.
$$
The latter equation thus corresponds to an extremely dilute gas of bosons with density $\sim e^{-2N}$.
Since the trapping potential is adjusted according to the density of the gas in the experiment, the $N$ dependence of $A_{e^{-2N}\tau}(e^{-N} \cdot)$ is reasonable.
\section{Formulation of the Problem}
For the sake of simplicity we will bound expressions which are uniformly bounded in $N$ and $t$ by some constant $C$.
We will not distinguish constants
appearing in a sequence of estimates, i.e. in $X\leq CY\leq CZ$ the constants may differ.
For $U \in \lbrace W_\beta, V_N \rbrace$, define the energy functional $\mathcal{E}_{U}:H^1(\mathbb{R}^ {2N}, \mathbb{C})\to \mathbb{R}$
$$
\mathcal{E}_{U}(\Psi)=N^{-1}\langle\hspace{-0.08cm}\langle\Psi,H_{U}\Psi\rangle\hspace{-0.08cm}\rangle\;,$$
where $\langle\hspace{-0.08cm}\langle\cdot,\cdot\rangle\hspace{-0.08cm}\rangle$ denotes the scalar product on $L^2(\mathbb{R}^{2N},\mathbb{C})$.
Furthermore, we define the
Gross-Pitaevskii energy functional $\mathcal{E}_{b_U}^{GP}:H^1(\mathbb{R}^ {2}, \mathbb{C})\to \mathbb{R}$ \begin{align}\label{energyfunct}
\mathcal{E}_{b_U}^{GP}(\phi)=&\langle\nabla\phi,\nabla\phi\rangle+\langle\phi,(A_t+\frac{1}{2} b_U|\phi|^2)\phi\rangle
=\langle\phi,
(h_{b_U}^{GP}-\frac{1}{2} b_U |\phi|^2)\phi\rangle
\end{align}
where $\langle\cdot,\cdot\rangle$ denotes the scalar product on $L^2(\mathbb{R}^2,\mathbb{C})$.
Note that both $\mathcal{E}_{U}(\Psi) $ and $\mathcal{E}_{b_U}^{GP}(\phi)$ depend on $t$, due to the time varying external potential $A_t$.
For the sake of readability, we will not indicate this time dependence explicitly.
We now state our main Theorem:
\begin{theorem}\label{theo}
Let $\Psi_0 \in L^2_{s}(\mathbb{R}^{2N}, \mathbb{C}) \cap H^2(\mathbb{R}^{2N}, \mathbb{C})$ with $\|\Psi_0\|=1$. Let $\phi_0 \in L^{2}(\mathbb{R}^2,\mathbb{C})$ with $\|\phi_0\|=1$ and assume $\lim_{N\to\infty}\gamma^{(1)}_{\Psi_0}=|\phi_0\rangle\langle\phi_0|$ in trace norm. Let the external potential $A_t$, which is defined in \eqref{hamiltonian}, statisfy
$ A_t
\in L^\infty (\mathbb{R}^2, \mathbb{R})$,
$ \dot{A}_t
\in L^\infty (\mathbb{R}^2, \mathbb{R})
$, for all $t \in \mathbb{R}$.
\begin{itemize}
\item[(a)]
For any $\beta>0$, let $W_\beta$ be given by $W_\beta(x)= N^{-1+2 \beta} W(N^{\beta}x)$, for $W \in
L^\infty_c (\mathbb{R}^2,\mathbb{R}) \;, W\geq0$ and $W$ spherically symmetric.
Let $\Psi_t$ the unique solution to $i \partial_t \Psi_t
= H_{W_\beta} \Psi_t$ with initial datum $\Psi_0$.
Let $\phi_t$ the unique solution to $i \partial_t \phi_t
= h^{GP}_{ \|W\|_1} \phi_t$ with initial datum $\phi_0$ and assume that $\phi_t \in H^3(\mathbb{R}^2, \mathbb{C})$.
Let $
\lim_{N \rightarrow \infty}
\left(
\mathcal{E}_{W_\beta}(\Psi_0)
-
\mathcal{E}_{ \|W\|_1}^ {GP}(\phi_0)
\right)
=0
$ and let $\mathcal{E}_{W_\beta}(\Psi_0) \leq C$.
Then, for any $\beta>0$ and for any $t>0$
\begin{equation}
\label{convergenls}
\lim_{N\to\infty}\gamma^{(1)}_{\Psi_t}=|\phi_t\rangle\langle\phi_t|
\end{equation} in trace norm.
\item[(b)]
Let $V_N$ be given by $V_N(x)= e^ {2N} V(e^ {N}x)$, for $V \in
L^\infty_c (\mathbb{R}^2,\mathbb{R}) \;, V\geq0$ and $V$ spherically symmetric.
Let $\Psi_t$ the unique solution to $i \partial_t \Psi_t
= H_{V_\beta} \Psi_t$ with initial datum $\Psi_0$.
Let $\phi_t$ the unique solution to $i \partial_t \phi_t
= h^{GP}_{4 \pi} \phi_t$ with initial datum $\phi_0$ assume that $\phi_t \in H^3(\mathbb{R}^2, \mathbb{C})$.
Let $
\lim_{N \rightarrow \infty}
\left(
\mathcal{E}_{V_N}(\Psi_0)
-
\mathcal{E}_{4 \pi}^ {GP}(\phi_0)
\right)
=0
$ and let $\mathcal{E}_{V_N}(\Psi_0) \leq C$.
Then, for any $t>0$
\begin{equation}
\label{converge}
\lim_{N\to\infty}\gamma^{(1)}_{\Psi_t}=|\phi_t\rangle\langle\phi_t|
\end{equation} in trace norm.
\end{itemize}
\end{theorem}
\begin{remark}
\begin{enumerate}
\item
We expect that for regular enough external potentials $A_t$, the regularity assumption $\phi_t \in H^3(\mathbb{R}^2, \mathbb{C})$ to follow from
regularity assumptions on the initial datum
$\phi_0$. In particular,
if $\phi_0 \in
\Sigma^3(\mathbb{R}^2,\mathbb{C})
=
\lbrace
f \in L^2 (\mathbb{R}^2, \mathbb{C})|
\sum_{ \alpha+ \beta \leq 3} \|x^\alpha \partial_x^\beta f \| < \infty
\rbrace
$ holds,
the bound $\| \phi_t \|_{H^3} < \infty$ has been proven for external potentials which are at most quadratic in space, see \cite{carles} and Lemma \ref{regularityLemma}. In particular, $\| \phi_t \|_{H^3} \leq C$ holds if the external potential is not present, i.e. $A_t=0$.
\item
As already mentioned, the convergence of $\gamma^{(1)}_{\Psi_t}$ to $|\phi_t\rangle\langle\phi_t|$ in trace norm is equivalent to convergence in operator norm, since $|\phi_t\rangle\langle\phi_t|$ is a rank
one projection \cite{rodnianskischlein}.
Other equivalent definitions of asymptotic 100\% condensation can be found in \cite{michelangeli}.
\item
It has been shown that in the limit $N\to\infty$ the energy-difference $\mathcal{E}_{V_N}(\Psi^{gs})-\mathcal{E}_{4 \pi}^{GP}(\phi^{gs})\to0$, where $\Psi^{gs}$ is the ground state of a trapped Bose gas and $\phi^{gs}$ the ground state of the
respective Gross-Pitaevskii energy functional, see \cite{ls}, \cite{lsy}.
\item
In our proof we will give explicit error estimates in terms of the particle number $N$. We shall show that the rate of convergence is of order
$N^{-\delta}$ for some $\delta>0$, assuming that also initially $\gamma^{(1)}_{\Psi_0} \rightarrow|\phi_0\rangle\langle\phi_0|$ converges in trace norm with rate of at least $N^{-\delta}$.
\item One can relax the conditions on the initial condition and only require $\Psi_0 \in L^2_{s}(\mathbb{R}^{2N}, \mathbb{C}) $ using a standard density argument.
\item
Our Theorem can easily be formulated for more general Hamiltonians of the form
$$
H_U= \sum_{k=1}^N h_k + \sum_{i<j} U(x_i-x_j)
\;,
$$
where $h_k$ is a one-particle operator.
In particular, one may think of $h_k$ as the magnetic Hamiltonian, or of $h_k$ as the pseudo-relativistic one-body Hamiltonian.
Our Theorem is then valid under the condition that $H_U$ is self-adjoint on a domain $\mathcal{D}(L^2(\mathbb{R}^{2N}, \mathbb{C}))$ and
$ \phi_t \in H^3(\mathbb{R}^2, \mathbb{C})$ holds.
Here, $\phi_t$ solves
$
i \partial_t \phi_t
=
(h+b_U |\phi_t|^2)
\phi_t
$.
See \cite{knowles} for a detailed discussion on the conditions on $h$.
\item
The condition $\mathcal{E}_{W_{\beta}}(\Psi_0) \leq C$, resp. $\mathcal{E}_{V_N}(\Psi_0) \leq C$ is needed to ensure that $\|\nabla \Psi_t\|$ is uniformly bounded in $N$ for any $t \geq 0$, see Lemma \ref{kinenergyboundedlemma}.
\end{enumerate}
\end{remark}
\section{Organization of the proof}\label{secskel}
The method we use in this paper is introduced in detail in \cite{pickl1} and was used to derive various mean-field equations.
As we have mentioned, our proof is based on \cite{picklgp3d}, which covers the three-dimensional counterpart of our system.
Heuristically speaking, the method we are going to employ is based on the idea of counting for each time $t$ the relative number of those particles which are
not in the state $\phi_t$.
It is then possible to show that the rate of particles which leave the condensate is small, if initially almost all particles are in the state $\phi_0$.
In order to compare the exact dynamic, generated by $H_U$, with the effective dynamic, generated by $h^{GP}_{b_U}$, we define the projectors $p^\phi_j$ and $q^\phi_j$.
\begin{definition}\label{defpro}
Let $\phi\in L^{2}(\mathbb{R}^2,\mathbb{C})$ with $\| \phi \|=1$.
\begin{enumerate}
\item For any $1\leq j\leq N$ the
projectors $p_j^\phi:L^2(\mathbb{R}^{2N},\mathbb{C})\toL^2(\mathbb{R}^{2N},\mathbb{C})$ and $q_j^\phi:L^2(\mathbb{R}^{2N},\mathbb{C})\toL^2(\mathbb{R}^{2N},\mathbb{C})$ are given by
\begin{align*} p_j^\phi\Psi=\phi(x_j)\int\phi^*(\tilde{x}_j)\Psi(x_1,\ldots, \tilde{x}_j,\dots,x_N)d^2\tilde{x}_j\;\;\;\forall\;\Psi\inL^2(\mathbb{R}^{2N},\mathbb{C})
\end{align*}
and $q_j^\phi=1-p_j^\phi$.
We shall also use, with a slight abuse of notation, the bra-ket notation
$p_j^\phi=|\phi(x_j)\rangle\langle\phi(x_j)|$.
\item
For any $0\leq k\leq N$ we define the set
$$
\mathcal{S}_k=\left\lbrace (s_1,s_2,\ldots,s_N)\in\{0,1\}^N\;;\;
\sum_{j=1}^N s_j=k\right\rbrace
$$
and the orthogonal projector $P_{k}^\phi:L^2(\mathbb{R}^{2N},\mathbb{C}) \rightarrow L^2(\mathbb{R}^{2N},\mathbb{C})$ as
$$P_{k}^\phi=\sum_{\vec a\in\mathcal{S}_k}\prod_{j=1}^N\big(p_{j}^{\phi}\big)^{1-s_j} \big(q_{j}^{\phi}\big)^{s_j}\;.$$
For
negative $k$ and $k>N$ we set $P_{k}^\phi=0$.
\item
For any function $m:\mathbb{N}_0\to\mathbb{R}^+_0$ we define the operator $\widehat{m}^{\phi}:L^2(\mathbb{R}^{2N},\mathbb{C})\toL^2(\mathbb{R}^{2N},\mathbb{C})$ as
\begin{equation}\label{hut}\widehat{m}^{\phi}=\sum_{j=0}^N m(j)P_j^\phi\;.\end{equation}
We also need the shifted operators
$\widehat{m}^{\phi}_d:L^2(\mathbb{R}^{2N},\mathbb{C})\toL^2(\mathbb{R}^{2N},\mathbb{C})$ given by
$$\widehat{m}^{\phi}_d=\sum_{j=-d}^{N-d} m(j+d)P_j^\phi\;.$$
\end{enumerate}
\end{definition}
Following a general strategy, which is described in detail in \cite{pickl1}, we define a functional $\alpha:L^2(\mathbb{R}^{2N},\mathbb{C})\timesL^2(\mathbb{R}^2,\mathbb{C})\to\mathbb{R}^+_0$ such that
\begin{enumerate}
\item
$\frac{\text{d}}{\text{d}t}\alpha(\Psi_t,\phi_t)$ can be estimated by $\alpha(\Psi_t,\phi_t)+\mbox{\begin{scriptsize}$\mathcal{O}$\end{scriptsize}}(1)$.
Using a Gr\"onwall type estimate, it then follows that
$\alpha(\Psi_t,\phi_t) \leq C_1 e^{C_2 t}(\alpha(\Psi_0,\phi_0) + \mbox{\begin{scriptsize}$\mathcal{O}$\end{scriptsize}}(1))$, for some constants $C_1,C_2>0$.
\item $\alpha(\Psi,\phi)\to0$
implies convergence of the reduced one particle density matrix of $\Psi$ to
$|\phi\rangle\langle\phi|$ in trace norm.
\end{enumerate}
In the case $\beta=0$ it was shown that the choice $$\alpha(\Psi,\phi)=\left\langle\hspace{-0.14cm}\left\langle\Psi,
\left(\widehat{n}^{\phi}\right)^j\Psi\right\rangle\hspace{-0.14cm}\right\rangle \;,$$
where
$n(k)=\sqrt{k/N}$ and $\langle\hspace{-0.08cm}\langle\cdot\rangle\hspace{-0.08cm}\rangle$ is scalar product on $L^2(\mathbb{R}^{2N},\mathbb{C})$ fulfills these requirements, for arbitrary $j>0$, see for example \cite{pickl1} and \cite{knowles}. For the more involved scaling we consider, it is however necessary to adjust this definition in order to obtain a Gr\"onwall type estimate.
Our proof is organized as follows:
\begin{enumerate}
\item Section \ref{secpre} gives some preliminary Lemmata on the algebra of the operators defined above.
\item For the most difficult scaling we consider here , i.e. $V_N$, we have to take the interaction-induced correlations between the particles into account. In Section \ref{secmic} we
provide some estimates on the zero-energy scattering state. Furthermore, we explain how the effective coupling parameter $b_{V_N}=4 \pi$ can be derived from the microscopic structure.
\item
In Section \ref{secpro} we prove our main Theorem stated above. We first deal with the potential $W_\beta$ and define a counting measure which makes it possible to establish a Gr\"onwall estimate for all $\beta>0$. We will explain in detail how one arrives at this Gr\"onwall estimate.
Later in this section the counting measure is adjusted to the case $V_N$, taking the microscopic structure $j_{N,R}$ of the wave function into account. We then establish a Gr\"onwall estimate and finally prove the Theorem for $V_N$.
The estimates needed in Section \ref{secpro} can then be found in Section \ref{rigestimates}.
\end{enumerate}
\section{Preliminaries}\label{secpre}
We will first fix the notation we are going to employ during the rest of the paper.
\begin{notation}
\begin{enumerate}
\item
Throughout the paper hats $\;\widehat{\cdot}\;$ will always be
used in the sense of Definition \ref{defpro} (c). The label $n$ will always be used for the function $n(k)=\sqrt{k/N}$.
\item
For better readability, we will omit the upper index $\phi$ on $p_j$,
$q_j$, $P_j$, $P_{j,k}$ and $\widehat{\cdot}$. It will be placed exclusively in a few formulas where their $\phi$-dependence plays an important role.
\item
In our estimates below we shall need the operator norm $\|\cdot\|_{\text{op}}$ defined for any linear operator $f:L^2(\mathbb{R}^{2N},\mathbb{C})\toL^2(\mathbb{R}^{2N},\mathbb{C})$ by
$$\|f\|_{\text{op}}=\sup_{\psi \in L^2(\mathbb{R}^{2N},\mathbb{C}), \|\Psi\|=1}\|f\Psi\|\;.$$
\item
We will
denote by $\mathcal{K}(\phi_t, A_t)$ a generic polynomial with finite degree in \\
$\|\phi_t\|_\infty, \|\nabla \phi_t\|_\infty, \|\nabla \phi_t\|, \| \Delta \phi_t \|, \|A_t\|_\infty,
\int_0^t ds \| \dot A_s \|_\infty
$ and $ \| \dot A_t \|_\infty$. Note, in particular, that for a generic constant $C$ the inequality
$C \leq \mathcal{K}(\phi_t, A_t)$ holds.
The exact form of $\mathcal{K}(\phi_t, A_t)$ which appears in the final bounds can be reconstructed easily, collecting all contributions from the different estimates.
\item
We will denote for any multiplication operator
$ F:L^2(\mathbb{R}^2, \mathbb{C}) \rightarrow L^2(\mathbb{R}^2, \mathbb{C}) $
the corresponding operator
$$
\mathds{1}^{\otimes (k-1)} \otimes F \otimes \mathds{1}^{\otimes (N-k)} :
L^2(\mathbb{R}^{2N}, \mathbb{C}) \rightarrow L^2(\mathbb{R}^{2N}, \mathbb{C})
$$
acting on the $N$-particle Hilbert space
by $F(x_k)$. In particular, we will use, for any $ \Psi,\Omega \in L^2(\mathbb{R}^{2N}, \mathbb{C})$ the notation
$$
\langle\hspace{-0.08cm}\langle \Omega, \mathds{1}^{\otimes (k-1)} \otimes F \otimes \mathds{1}^{\otimes (N-k)}\Psi \rangle\hspace{-0.08cm}\rangle
=
\langle\hspace{-0.08cm}\langle \Omega, F(x_k) \Psi \rangle\hspace{-0.08cm}\rangle
\;.
$$
In analogy, for any two-particle multiplication operator $K:L^2(\mathbb{R}^2, \mathbb{C}) ^{\otimes 2}\rightarrow L^2(\mathbb{R}^2, \mathbb{C})^{\otimes 2} $, we denote the operator acting on any $ \Psi \in L^2(\mathbb{R}^{2N}, \mathbb{C})$ by multiplication
in the variable $x_i$ and $x_j$ by $K(x_i,x_j)$. In particular, we denote
$$
\langle\hspace{-0.08cm}\langle \Omega, K(x_i,x_j) \Psi \rangle\hspace{-0.08cm}\rangle
=
\int_{\mathbb{R}^{2N}}
K(x_i, x_j)
\Omega^*(x_1,\ldots,x_N) \Psi (x_1,\ldots,x_N) d^2x_1 \dots d^2x_N \;.
$$
\end{enumerate}
\end{notation}
First we prove some properties of the projectors $p_j,q_j$, which were defined in Definition \ref{defpro}.
\begin{lemma}\label{kombinatorik}
\begin{enumerate}
\item For any weights $m,r:\mathbb{N}_0\to\mathbb{R}^+_0$ we have
that
$$\widehat{m}\widehat{r}\,=\widehat{mr}=\widehat{r}\,\widehat{m}\;\;\;\;\;\;\;\widehat{m}p_j=p_j\widehat{m}\;\;\;\;\;\;\;\widehat{m}q_j=q_j\widehat{m}\;\;\;\;\;\;\;\widehat{m}P_{k}=P_{k}\widehat{m}\;.$$
\item Let $n:\mathbb{N}_0\to\mathbb{R}^+_0$ be given by $n(k)=\sqrt{k/N}$.
Then, the square of $\widehat{n}$
equals the relative particle number operator of particles not in the
state $\phi$, i.e.
\begin{equation}\label{partnumber}\left(\widehat{n}\right)^2=N^{-1}\sum_{j=1}^Nq_j\;.\end{equation}
\item For any weight $m:\mathbb{N}_0\to\mathbb{R}^+_0$ and any function $f \in L^\infty\left(\mathbb{R}^4,\mathbb{R}\right)$ and any
$j,k=0,1,2$ $$\widehat{m} Q_j f(x_1,x_2)Q_k=
Q_j f(x_1,x_2)\widehat{m}_{j-k}Q_k\;,$$ where
$Q_0=p_1 p_2$, $Q_1\in\{p_1q_2,q_1p_2\}$ and
$Q_2=q_1q_2$.
Furthermore, for $j,k\in\{0,1\}$
$$\widehat{m} \widetilde Q_j \nabla_1 \widetilde Q_k=
\widetilde Q_j \nabla_1\widehat{m}_{j-k}\widetilde Q_k\;,$$
where
$\widetilde Q_0=p_1$ and $\widetilde Q_1=q_1$.
\item For any weight $m:\mathbb{N}_0\to\mathbb{R}^+_0$ and any function $f \in L^\infty\left(\mathbb{R}^4,\mathbb{C}\right)$
$$[f(x_1,x_2),\widehat{m}]=\left[f(x_1,x_2),p_1p_2(\widehat{m}-\widehat{m}_2)+(p_1q_2+q_1p_2)(\widehat{m}-\widehat{m}_1)\right]$$
\item Let $f\in L^1\left(\mathbb{R}^2,\mathbb{C}\right)$, $g\in L^2\left(\mathbb{R}^2,\mathbb{C}\right)$. Then,
\begin{align}\label{kombeqa}\|p_j f(x_j-x_k)p_j\|_{\text{op}
\leq& \|f\|_1\|\phi\|_\infty^2\;,
\\ \label{kombeqb}
\|p_jg^*(x_j-x_k)\|_{\text{op}}=&
\|g(x_j-x_k)p_j\|_{\text{op}}\leq \|g\|\;\|\phi\|_\infty
\\\label{kombeqc}
\| |\phi(x_j) \rangle \langle \nabla_j \phi(x_j)| h^* (x_j-x_k)\|_{\text{op}}=&
\|h(x_j-x_k)\nabla_j p_j\|_{\text{op}}\leq \|h\|\|\nabla\phi\|_{\infty}
\;.\end{align}
\end{enumerate}
\end{lemma}
\begin{proof}\begin{enumerate}
\item
follows immediately from Definition \ref{defpro}, using that $p_j$
and $q_j$ are orthogonal projectors.
\item Note that $\cup_{k=0}^N\mathcal{S}_k=\{0,1\}^N$, so $1=\sum_{k=0}^N P_k$. Using also
$(q_j)^2=q_j$ and $q_j p_j=0$ we get \begin{align*}
\sum_{j=1}^Nq_j=\sum_{j=1}^Nq_j\sum_{k=0}^N
P_k= \sum_{k=0}^N\sum_{j=1}^Nq_j
P_k=\sum_{k=0}^Nk P_k=N\widehat{n^2}=N\widehat{n}^2\;.\end{align*}
\item
Using the definitions above we have \begin{align*} \widehat{m}
Q_j f(x_1,x_2)Q_k
=&\sum_{l=0}^N m(l)P_l Q_jf(x_1,x_2)Q_k
\end{align*}
The number of projectors $q_j$ in $P_l Q_j$ in the coordinates $j=3,\ldots,N$ is equal to $l-j$. The $p_j$ and $q_j$ with $j=3,\ldots,N$ commute with $Q_jf(x_1,x_2)Q_k$. Thus
$P_l Q_jf(x_1,x_2)Q_k= Q_jf(x_1,x_2)Q_kP_{l-j+k}$ and
\begin{align*}
\widehat{m}
Q_j f(x_1,x_2)Q_k=& \sum_{l=0}^N m(l) Q_jf(x_1,x_2)Q_kP_{l-j+k}
\\&\hspace{-3cm}= \sum_{\widetilde l=k-j}^{N+k-j} Q_jf(x_1,x_2)m(\widetilde l+j-k)P_{\widetilde l} Q_k
=Q_j f(x_1,x_2)\widehat{m}_{j-k}Q_k\;.
\end{align*}
Similarly one gets the second formula.
\item First note that
\begin{align}\label{multsev}
&\hspace{-1cm}[f(x_1,x_2),\widehat{m}]-\left[f(x_1,x_2),p_1p_2(\widehat{m}-\widehat{m}_2)+p_1q_2(\widehat{m}-\widehat{m}_1)
+q_1p_2(\widehat{m}-\widehat{m}_1)\right]%
\nonumber\\=&[f(x_1,x_2),q_1q_2\widehat{m}]+\left[f(x_1,x_2),p_1p_2\widehat{m}_2+p_1q_2\widehat{m}_1
+q_1p_2\widehat{m}_1\right]%
\;.
\end{align}
We will show that the right hand side is zero.
Multiplying the right hand side with $p_1p_2$ from the left and using (c) one gets
\begin{align*}
&p_1p_2f(x_1,x_2)q_1q_2\widehat{m}+p_1p_2f(x_1,x_2)p_1p_2\widehat{m}_2-p_1p_2\widehat{m}_2f(x_1,x_2)
\\&+p_1p_2f(x_1,x_2)p_1q_2\widehat{m}_1+p_1p_2f(x_1,x_2)q_1p_2\widehat{m}_1%
\\=&p_1p_2\widehat{m}_2f(x_1,x_2)q_1q_2+p_1p_2\widehat{m}_2 f(x_1,x_2)p_1p_2-p_1p_2\widehat{m}_2f(x_1,x_2)
\\&+p_1p_2\widehat{m}_2 f(x_1,x_2)p_1q_2+p_1p_2\widehat{m}_2 f(x_1,x_2)q_1p_2%
\\&=0\;.
\end{align*}
Multiplying (\ref{multsev}) with $p_1q_2$ from the left one gets
\begin{align*}
&p_1q_2f(x_1,x_2)q_1q_2\widehat{m}+p_1q_2f(x_1,x_2)p_1p_2\widehat{m}_2+
p_1q_2f(x_1,x_2)p_1q_2\widehat{m}_1\\&+p_1q_2f(x_1,x_2)q_1p_2\widehat{m}_1-p_1q_2\widehat{m}_1f(x_1,x_2)
\end{align*}
Using (c) the latter is zero. Also multiplying with $q_1p_2$ yields zero due to symmetry in interchanging
$x_1$ with $x_2$.
Multiplying (\ref{multsev}) with $q_1q_2$ from the left one gets
\begin{align*}
&q_1q_2f(x_1,x_2)\widehat{m}q_1q_2-q_1q_2\widehat{m}f(x_1,x_2)+
q_1q_2f(x_1,x_2)p_1p_2\widehat{m}_2+\\&q_1q_2f(x_1,x_2)p_1q_2\widehat{m}_1+q_1q_2f(x_1,x_2)q_1p_2\widehat{m}_1
\end{align*}
which is again zero and so is (\ref{multsev}).
\item
First note that, for bounded operators $A, B$,
$\| A B \|_{\text{op}}=\| B^* A^* \|_{\text{op}}$ holds, where $A^*$ is the adjoint operator of $A$.
To show (\ref{kombeqa}), first note that
\begin{align} p_j f(x_j-x_k)p_j=p_j (f\star|\phi|^2)(x_k)\;.
\label{faltungorigin}
\end{align}
It follows that $$\|p_j f(x_j-x_k)p_j\|_{\text{op}}\leq \|f\|_1\|\phi\|_\infty^2\;.$$
For (\ref{kombeqb}) we write
\begin{align*}
\|g(x_j-x_k)p_j\|_{\text{op}}^2=&\sup_{\|\Psi\|=1}\|g(x_j-x_k)p_j\Psi\|^2=
\\=&\sup_{\|\Psi\|=1}\langle\hspace{-0.08cm}\langle \Psi,p_j |g(x_j-x_k)|^2 p_j\Psi\rangle\hspace{-0.08cm}\rangle\\\leq& \|p_j |g(x_j-x_k)|^2p_j\|_{\text{op}}\;.
\end{align*}
With (\ref{kombeqa}) we get (\ref{kombeqb}).
For \eqref{kombeqc} we use
\begin{align*}
\| g(x_j-x_k) \nabla_j p_j \|_{\text{op}}^ 2
=&
\sup_{\|\Psi\|=1}
\langle\hspace{-0.08cm}\langle \Psi,
p_j
(|g|^ 2 * |\nabla \phi|^2)(x_k)
\Psi \rangle\hspace{-0.08cm}\rangle
\leq
\||g|^ 2 * |\nabla \phi|^2\|_{\infty}
\\
\leq
&
\|g\|^2 \| \nabla \phi \|_\infty^2
\end{align*}
\end{enumerate}
\end{proof}
Within our estimates we will encounter wave functions where some of the symmetry is broken (at this point the
reader should exemplarily think of the wave function
$V_\beta(x_1-x_2)\Psi$ which is not symmetric under exchange of the variables $x_1$ and $x_3$, for example). This leads to the following
definition
\begin{definition}
For any finite set $\mathcal{M}\subset\{1,2,\ldots,N\}$, define the space $\mathcal{H}_{\mathcal{M}}\subsetL^2(\mathbb{R}^{2N},\mathbb{C})$ as the set of
functions which are symmetric in all variables in $\mathcal{M}$
\begin{align*}\Psi\in \mathcal{H}_{\mathcal{M}}\Leftrightarrow& \Psi(x_1,\ldots,x_j,\ldots,x_k,\ldots,x_N)=\Psi(x_1,\ldots,x_k,\ldots,x_j,\ldots,x_N)\\&\text{ for all } j,k\in\mathcal{M}\;.
\end{align*}
\end{definition}
Based on the combinatorics of the $p_j$ and $q_j$, we obtain the following
\begin{lemma}\label{kombinatorikb}
For any $f:\mathbb{N}_0\to\mathbb{R}^+_0$ and any finite set $\mathcal{M}_a\subset \{1,2,\ldots,N\}$ with $1\in\mathcal{M}_a$
and any finite set $\mathcal{M}_b\subset\{1,2,\ldots,N\}$ with $1,2\in\mathcal{M}_b$
\begin{align}\label{komb1}
\left\| \widehat{f} q_1\Psi\right\|^2\leq& \frac{N}{|\mathcal{M}_a|}
\|\widehat{f}\widehat{n}\Psi\|^2\;\;\;\;\;\;\;\;\;\;\text{ for any }\Psi\in\mathcal{H}_{\mathcal{M}_a}
,
\\
\label{komb2} \left\| \widehat{f}
q_1 q_2\Psi\right\|^2\leq& \frac{N^2}{|\mathcal{M}_b|(|\mathcal{M}_b|-1)}
\|\widehat{f}(\widehat{n})^2\Psi\|^2\;\;\;\;\;\text{ for any }\Psi\in\mathcal{H}_{\mathcal{M}_b}\;.\end{align}
\end{lemma}
\begin{proof}
Let $\Psi\in\mathcal{H}_{\mathcal{M}_a}$ for some finite set $1\in\mathcal{M}_a\subset\{1,2,\ldots,N\}$.
By Lemma \ref{kombinatorik} (b), (\ref{komb1}) can be estimated as
\begin{align*}\|\widehat{f}\widehat{n}\Psi\|^2
=&\langle\hspace{-0.08cm}\langle\Psi,(\widehat{f})^{2}(\widehat{n})^2\Psi\rangle\hspace{-0.08cm}\rangle
=N^{-1}\sum_{k=1}^N\langle\hspace{-0.08cm}\langle\Psi,(\widehat{f})^{2}q_k\Psi\rangle\hspace{-0.08cm}\rangle
\\\geq& N^{-1}\sum_{k\in\mathcal{M}_a}\langle\hspace{-0.08cm}\langle\Psi,(\widehat{f})^{2}q_k\Psi\rangle\hspace{-0.08cm}\rangle
=\frac{|\mathcal{M}_a|}{N}\langle\hspace{-0.08cm}\langle\Psi,(\widehat{f})^{2}q_1\Psi\rangle\hspace{-0.08cm}\rangle
\\=&\frac{|\mathcal{M}_a|}{N}\|\widehat{f}
q_1\Psi\|^2\;.
\end{align*}
Similarly, we obtain for $\Psi\in\mathcal{H}_{\mathcal M_b}$
\begin{align*} \|\widehat{f}
(\widehat{n})^2\Psi\|^2 =&\langle\hspace{-0.08cm}\langle\Psi,(\widehat{f}
)^2(\widehat{n})^4\Psi\rangle\hspace{-0.08cm}\rangle
\geq N^{-2}\sum_{j,k\in\mathcal{M}_b}\langle\hspace{-0.08cm}\langle\Psi,(\widehat{f} )^2q_j
q_k\Psi\rangle\hspace{-0.08cm}\rangle
\\=&\frac{|\mathcal{M}_b|(|\mathcal{M}_b|-1)}{N^2}\langle\hspace{-0.08cm}\langle\Psi,(\widehat{f} )^2q_1
q_2\Psi\rangle\hspace{-0.08cm}\rangle+\frac{|\mathcal{M}_b|}{N^2}\langle\hspace{-0.08cm}\langle\Psi,(\widehat{f}
)^2q_1\Psi\rangle\hspace{-0.08cm}\rangle
\\\geq& \frac{|\mathcal{M}_b|(|\mathcal{M}_b|-1)}{N^2} \|\widehat{f} q_1q_2\Psi\|^2
\end{align*}
which concludes the Lemma.
\end{proof}
\begin{corollary}\label{kombinatorikc}
For any weight $m:\mathbb{N}_0\to\mathbb{R}^+_0$ \begin{align}\label{coreins}
\|\nabla_2 \widehat{m}q_2\Psi\|&\leq 2\|\widehat{m}\|_{\text{op}}\|\nabla_2q_2\Psi\| ,
\\\label{corzwei}
\|\nabla_2 \widehat{m}q_1q_2\Psi\|&\leq C\|\widehat{m}\widehat{n}\|_{\text{op}}\|\nabla_2q_2\Psi\|\;.
\end{align}
\end{corollary}
\begin{proof}
Using $p_2+q_2=1$ and triangle inequality,
\begin{align}
\label{azeile}\|\nabla_2 \widehat{m}q_2\Psi\|&\leq\|p_2\nabla_2 \widehat{m}q_2\Psi\|+\|q_2\nabla_2 \widehat{m}q_2\Psi\| ,\\
\label{bzeile}\|\nabla_2 \widehat{m}q_1q_2\Psi\|&\leq\|p_2\nabla_2 \widehat{m}q_1q_2\Psi\|+\|q_2\nabla_2 \widehat{m}q_1q_2\Psi\| \;.
\end{align}
With Lemma \ref{kombinatorik} (c) we get
\begin{align*}
(\ref{azeile})&=\|\widehat{m}_{1}p_2\nabla_2q_2\Psi\|+\|\widehat{m}q_2\nabla_2q_2\Psi\|
\leq(\|\widehat{m}_{1}\|_{\text{op}}+\|\widehat{m}\|_{\text{op}})\|\nabla_2q_2\Psi\|\;.
\end{align*}
Note that the wave function $p_2\nabla_2q_2\Psi$ is symmetric under the exchange of any two variables but $x_2$. Thus we can use Lemma \ref{kombinatorikb} to get
\begin{align*}
(\ref{bzeile})&=\|q_1\widehat{m}_{1}p_2\nabla_2q_2\Psi\|+\|q_1\widehat{m}q_2\nabla_2q_2\Psi\|
\\&\leq\frac{N}{N-1}(\|\widehat{m}_{1}\widehat{n}\|_{\text{op}}+\|\widehat{m}\widehat{n}\|_{\text{op}})\|\nabla_2q_2\Psi\|\;.
\end{align*}
Since $\sqrt{k}\leq\sqrt{k+1}$ for $k\geq 0$ it follows that the latter is bounded by
$$ C(\|\widehat{m}_{1}\widehat{n}_{1}\|_{\text{op}}+\|\widehat{m}\widehat{n}\|_{\text{op}})\|\nabla_2q_2\Psi\|\;.$$
Using that $\|\widehat{r}\|_{\text{op}}=\sup_{0\leq k \leq N}\{r(k)\}=\|\widehat{r}_d\|_{\text{op}}$
for any $d\in\mathbb{N}$ and any weight $r$, the Corollary follows.
\end{proof}
\begin{lemma}\label{trick}
Let $\Omega,\chi\in \mathcal{H}_{\mathcal{M}}$ for some $\mathcal{M}$, let $1\notin\mathcal{M}$ and $2,3\in\mathcal{M}$.
Let $O_{j,k}$ be an operator acting on the $j^{th}$ and $k^{th}$ coordinate. Then
\begin{align*}
|\langle\hspace{-0.08cm}\langle\Omega,O_{1,2}\chi\rangle\hspace{-0.08cm}\rangle|&\leq \|\Omega\|^2+\left|\langle\hspace{-0.08cm}\langle O_{1,2}\chi,O_{1,3}\chi\rangle\hspace{-0.08cm}\rangle\right|+(|\mathcal{M}|)^{-1}\|O_{1,2}\chi\|^2
\;.
\end{align*}
\end{lemma}
\begin{proof}
Using symmetry and Cauchy Schwarz
\begin{align*}
|\langle\hspace{-0.08cm}\langle\Omega,O_{1,2}\chi\rangle\hspace{-0.08cm}\rangle|=&|\mathcal{M}|^{-1}|\langle\hspace{-0.08cm}\langle\Omega,\sum_{j\in\mathcal{M}} O_{1,j}\chi\rangle\hspace{-0.08cm}\rangle|\leq|\mathcal{M}|^{-1}\|\Omega\|\;\|\sum_{j\in\mathcal{M}} O_{1,j}\chi\|
\end{align*}
For the second factor we can write
\begin{align*}
\|\sum_{j\in\mathcal{M}} O_{1,j}\chi\|^2&=\langle\hspace{-0.08cm}\langle\sum_{j\in\mathcal{M}} O_{1,j}\chi,\sum_{k\in\mathcal{M}} O_{1,k}\chi\rangle\hspace{-0.08cm}\rangle
\\\leq&\sum_{j\in\mathcal{M}}|\langle\hspace{-0.08cm}\langle O_{1,j}\chi,O_{1,j}\chi\rangle\hspace{-0.08cm}\rangle|+|\sum_{j\neq k\in\mathcal{M}}\langle\hspace{-0.08cm}\langle O_{1,j}\chi,O_{1,k}\chi\rangle\hspace{-0.08cm}\rangle|
\\\leq&|\mathcal{M}||\langle\hspace{-0.08cm}\langle O_{1,2}\chi,O_{1,2}\chi\rangle\hspace{-0.08cm}\rangle|+|\mathcal{M}|(|\mathcal{M}|-1)|\langle\hspace{-0.08cm}\langle O_{1,2}\chi,O_{1,3}\chi\rangle\hspace{-0.08cm}\rangle|
\end{align*}
Since $ab\leq 1/2a^2+1/2b^2$ and $(a+b)^2\leq 2a^2+2b^2$ holds for any real numbers $a,b$, the Lemma follows.
\end{proof}
In our estimates, we need the regularity conditions
\begin{align*}
\| \nabla \phi_t \|_\infty < \infty
\qquad
\| \phi_t \|_\infty < \infty
\qquad
\| \nabla \phi_t \| < \infty
\qquad
\| \Delta\phi_t \| < \infty
\end{align*}
That is, we need $\phi_t \in H^2(\mathbb{R}^2,\mathbb{C}) \cap W^{1,\infty}(\mathbb{R}^2,\mathbb{C})$.
Then, $ \| \Delta|\phi_t| ^2 \|, \| \Delta|\phi_t| ^2 \|_1$ and $ \| \phi_t^ 2 \| $, which also appear in our estimates, can be bounded by
\begin{align*}
\Delta|\phi_t|^ 2
=&
\phi^*_t \Delta \phi_t
+
\phi_t \Delta\phi^*_t
+
2 (\nabla \phi^*_t) \cdot (\nabla \phi_t)
\\
\| \Delta|\phi_t| ^2 \|
\leq &
2
\| \Delta \phi_t \| \| \phi_t \| _\infty
+
2
\| \nabla \phi_t \| \| \nabla \phi_t \| _\infty
\\
\| \Delta|\phi_t| ^2 \|_1 \leq& 4 \| \Delta \phi_t \|
\\
\| \phi_t^ 2 \| \leq &\| \phi_t \|_\infty \| \phi_t \|
\;.
\end{align*}
Recall the Sobolev embedding Theorem, which implies in particular
$H^k(\mathbb{R}^2,\mathbb{C})= W^{k,2}(\mathbb{R}^ 2,\mathbb{C}) \subset C^{k-2}(\mathbb{R}^2,\mathbb{C})$.
If $\phi \in C^1(\mathbb{R}^2,\mathbb{C}) \cap H^1(\mathbb{R}^2,\mathbb{C})$, then $\phi \in W^{1,\infty}(\mathbb{R}^2,\mathbb{C})$ follows since both $\phi$ and $\nabla \phi$ have to decay at infinity.
Thus, $\phi_t \in H^3(\mathbb{R}^2,\mathbb{C})$ implies $\phi_t \in H^2(\mathbb{R}^2,\mathbb{C}) \cap W^{1,\infty}(\mathbb{R}^2, \mathbb{C})$, which suffices for our estimates. Since $\phi_t$ obeys a defocusing nonlinear Schr\"odinger equation, we expect the regularity of the solution $\phi_t$ to follow from the regularity of the initial datum $\phi_0$. This has been proven for a class of external potentials $A_t$.
\begin{lemma} \label{regularityLemma}
Let $\phi_0 \in
\Sigma^k(\mathbb{R}^2,\mathbb{C})
=
\lbrace
f \in L^2 (\mathbb{R}^2, \mathbb{C})|
\sum_{ \alpha+ \beta \leq k} \|x^\alpha \partial_x^\beta f \| < \infty
\rbrace
$, for $k\geq2$.
Let, for $b>0$,
$\phi_t$ the unique solution to
\begin{align*}
i \partial_t \phi_t = (-\Delta+A_t+ b |\phi_t|^2) \phi_t \;.
\end{align*}
Let $A_{\cdot} \in L^\infty_{\text{loc}}(\mathbb{R}_t \times \mathbb{R}^2_x, \mathbb{C})$ real valued and smooth with respect to the space variable: for (almost) all $t \in \mathbb{R}$,
the map $x \mapsto A_t(x)$ is $C^\infty$. Moreover, $A_t$ is at most quadratic in space, uniformly w.r.t. time $t$:
\begin{align*}
\forall \alpha \in \mathbb{N}^2, | \alpha | \geq 2, \qquad \partial_x^\alpha A_{\cdot}
\in L^\infty(\mathbb{R}_t \times \mathbb{R}^d_x, \mathbb{C}).
\end{align*}
In addition, $t \mapsto \sup_{|x| \leq 1} |A_t(x)|$ belongs to $L^{\infty}(\mathbb{R}, \mathbb{C})$.
Then
\begin{itemize}
\item[(a)]
$ \phi_t \in \Sigma^k(\mathbb{R}^2,\mathbb{C})$, which implies $ \phi_t \in H^k(\mathbb{R}^2,\mathbb{C})$.
\item[(b)]
$\| \phi_t\|=\| \phi_0 \|$.
\item[(c)]
Let $ \phi_0 \in \Sigma^3(\mathbb{R}^2,\mathbb{C})$. Assume in addition that
$\| A_t \|_\infty < \infty$ and $\| \dot{A}_t \|_\infty < \infty$.
Then, for any fixed $t \geq 0$,
$ \mathcal{K}(\phi_t, A_t) < \infty$ follows.
\end{itemize}
\end{lemma}
\begin{proof}
Part (a) is Corollary 1.4. in \cite{carles}. We like to remark that $\| \phi_t \| _{H^k} \leq C$ holds, if $A_t=0$, see Section 1.2. in \cite{carles}.
The conditions on $A_t$ are for example satisfied if
$A_t \in C^\infty_c (\mathbb{R}^2, \mathbb{R})$ for all $t \in \mathbb{R}$ and $A_t(x)=0$, for all $|t| \geq T$.
Part (b) can be verified directly, using the existence of global in time solutions.
Part (c) follows from (a) and the embedding
$H^3(\mathbb{R}^2,\mathbb{C}) \subset H^2(\mathbb{R}^2,\mathbb{C}) \cap W^{1,\infty}(\mathbb{R}^2, \mathbb{C}) $.
\end{proof}
\section{Microscopic structure in 2 dimensions}
\label{secmic}
\subsection{Exponential scaling in 2 dimensions}
In this section we analyze the microscopic structure which is induced by $V_N$. In particular, we explain why the dynamical properties of the system are determined by the low energy scattering regime.
\begin{definition}
Let $V \in L^\infty_c(\mathbb{R}^2, \mathbb{R}), \; V(x) \geq 0$, $V$ spherically symmetric and let $V_N$ be given by $V_N(x) = e^{2N} V(e^N x)$. We define the zero energy scattering state $j_{N,R}$ for any $R \geq \text{diam}(\text{supp} (V_N))$ by
\begin{equation}
\label{eq: defj}
\begin{cases}
\left( - \Delta_x + \frac{1}{2}e^{2N} V(e^N x) \right) j_{N,R}(x)=0
\\
j_{N,R}(x)=1 \; \text{for } |x| = R
\end{cases}
\end{equation}
\end{definition}
One may think of $R$ as the mean interparticle distance of the condensate, i.e. $R= \mathcal{O}(N^{-1/2})$. However, one is quite free in choosing $R$, since the dependence of $j_{N,R}$ on $R$ is only logarithmic (see below).
Rescaling $x \rightarrow e^{N} x =y$, we obtain, setting $\tilde{R}=e^{N} R$ and $s_{\tilde{R}}(y)= j_{0, e^{N}R}(y)$
\begin{equation}
\begin{cases}
\left( - \Delta_y + \frac{1}{2}V(y) \right) s_{\tilde{R}}(y)=0
\\
s_{\tilde{R}}(y)=1 \; \text{for } |y| =\tilde{R}
\end{cases}
\end{equation}
Next, we want to recall some important properties of the scattering state $s_{\tilde{R}}$.
\begin{lemma}
Define $I_{\tilde{R}} = \int_{\mathbb{R}^2} d^2x V(x) s_{\tilde{R}}(x)$.
For the scattering state defined previously the following relations hold:
\begin{enumerate}
\item[(a)] $1 \geq s_{\tilde{R}} (x)\geq 1 + \frac{I_{\tilde{R}}}{4 \pi} \ln \left( \frac{|x|}{\tilde{R}} \right) $
holds for all $|x| \leq \tilde{R}$.
\item[(b)] There exists a positive number $a$, called scattering length of the potential $V$, such that
\begin{align*}
I_{\tilde{R}}=
4 \pi \frac{1}{\ln \left(\frac{\tilde{R}}{a}\right)}
\;.
\end{align*}
Furthermore, $a$ does not depend on $\tilde{R}$.
\item[(c)]
For all $|x| \leq R$
$$1 \geq j_{N,R} (x)\geq 1 + \frac{1}{N+ \ln \left(\frac{R}{a}\right)} \ln \left( \frac{|x|}{R} \right). $$
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}
\item[(a)]
Since $V$ is a nonnegative and compactly supported potential, one can define the scattering state $s_{\tilde{R}}$ by a variational principle. Theorem C.1 in \cite{lssy} then shows that $s_{\tilde{R}}$ is a nonnegative, spherically symmetric function which is monotone nondecreasing in $|x|$.
Denote by $D = \text{diam (supp} (V))$.
It is then easy to see that for $D \leq |x| $ there exists a $I_{\tilde{R}} \in \mathbb{R}$ such that
\begin{equation}
\label{jasymp}
s_{\tilde{R}}(x) = 1 + \frac{I_{\tilde{R}}}{4 \pi} \ln \left( \frac{|x|}{\tilde{R}} \right)
\;.
\end{equation}
It remains to show that $I_{\tilde{R}}= \int_{\mathbb{R}^2}d^2x V(x) s_{\tilde{R}} (x)$. This can be seen by noting that, for $r > D $,
\begin{align*}
\int_{\mathbb{R}^2}d^2x V(x) s_{\tilde{R}}(x)
=&
2 \int_{B_r(0)}d^2x \Delta s_{\tilde{R}}(x)
=
2 \int_{\partial B_r(0)} \nabla s_{\tilde{R}}(x) \cdot ds
\nonumber \\
=&
\frac{I_{\tilde{R}}}{2 \pi} \int_{\partial B_r(0)} \nabla \ln (|x|)\cdot ds
=
\frac{I_{\tilde{R}}}{2 \pi} \int_{0}^{2 \pi} \frac{1}{r} r d\varphi
\\
=&
I_{\tilde{R}} \;.
\end{align*}
By Lemma C.2 in \cite{lssy}, $ s_{\tilde{R}}(x) \geq 1 + \frac{I_{\tilde{R}}}{4 \pi} \ln \left( \frac{|x|}{\tilde{R}} \right)$ follows. Since $s_{\tilde{R}}(x)$ is a monotone nondecreasing function of $|x|$, it also follows that
$s_{\tilde{R}}(x)\leq s_{\tilde{R}}(\tilde{R})=1, \forall |x| \leq \tilde{R}$, which proves (a).
\item[(b)]
By Theorem C.1 in \cite{lssy}, there exists a number $a \in \mathbb{R}^+$, not depending on $\tilde{R}$, such that for all $|x| \geq C$
\begin{align*}
s_{\tilde{R}}(x)= \frac{\ln(|x|/a) }{\ln(\tilde{R}/a)}
\;.
\end{align*}
Comparing this with \eqref{jasymp}, we obtain
\begin{align*}
\int_{\mathbb{R}^2} V(x) s_{\tilde{R}}(\mathbf{x}) dx^2=
4 \pi \frac{1}{\ln \left(\frac{\tilde{R}}{a}\right)}
\;.
\end{align*}
\item[(c)] follows directly from (a) and (b) noting that
\begin{align*}
\int_{\mathbb{R}^2} V(x) s_{\tilde{R}}(\mathbf{x}) dx^2=
\int_{\mathbb{R}^2} V_N(x) j_{N,R}(\mathbf{x}) dx^2
= 4 \pi \frac{1}{N+ \ln \left(\frac{R}{a}\right)}
\;.
\end{align*}
\end{enumerate}
\end{proof}
Assuming that the energy per particle $\mathcal{E}_{V_N}(\Psi)$ is of order one, the wave function $\Psi$ will have a microscopic structure near the interactions $V_N$, given by $j_{N,R}$. The interaction among two particles is then determined by
$ \frac{4 \pi}{N+ \ln \left(\frac{R}{a}\right)} \approx \frac{4 \pi}{N}$. Keeping in mind that each particle interacts with all other $N-1$ particles, we obtain the effective Gross-Pitaevskii equation, for $\varphi_t \in H^2( \mathbb{R}^2, \mathbb{C})$
\begin{align*}
i \partial_t \varphi_t(x) = (- \Delta+ 4 \pi | \varphi_t(x)|^2) \varphi_t(x)
\end{align*}
Thus, choosing $V_N(x)= e^{2N} V(e^Nx)$ leads in our setting to an effective one-particle equation which is determined by the low energy scattering behavior of the particles. We remark that, for any $s>0$, the potential
$e^{2Ns} V(e^{Ns}x)$ yields to the coupling $4 \pi/ s$.
\subsection{Properties of the scattering state}
Note that the potential $V_N$ is strongly peaked within an exponentially small region.
In order to control the dynamics of such a system, we define, with a slight abuse of notation, a
potential $M_{\beta}$ with softer scaling behavior in such a way that the potential $1/2(V_N -M_\beta)$
has scattering length zero. This allows us to ``replace'' $V_N$ by $M_\beta$, which has nicer scaling behavior and is better to control.
In particular, $\|M_\beta\| \leq C N^{-1+\beta}$ can be controlled for $\beta$ sufficiently small, while $\|V_N\|=\mathcal{O}(e^N)$ cannot be bounded by any finite polynomial in $N$.
The potential $M_\beta$ is \textit{not} of the exact scaling $N^{-1+2\beta} M(N^{\beta} x)$.
However, it is in the set $\mathcal{V}_{\beta}$, which we will define now.
\begin{definition}
For any $\beta>0$, we define the set of potentials $\mathcal{V}_{\beta}$ as
\begin{align*}
&
\mathcal{V}_{\beta}=
\Big\lbrace
U \in L^2(\mathbb{R}^2,\mathbb{R}) |
U(x) \geq 0 \; \forall x \in \mathbb{R}^2
,
\|U\|_1 \leq CN^{-1}
,
\|U\| \leq CN^{-1+ \beta},
\\
&
\|U\| _\infty \leq C N^{-1+2 \beta}
,
U(x)=0 \; \forall |x| \geq C N^{-\beta}
,\;
U\text{ is spherically symmetric}
\Big\rbrace .
\end{align*}
Note that $N^{-1+2\beta}W(N^\beta x) \in \mathcal{V}_\beta$ holds, if $W$ is positive, spherically symmetric and compactly supported.
\end{definition}
We like to remark that all relevant estimates below are formulated for $W_\beta \in \mathcal{V}_\beta$.
\begin{definition}\label{microscopic}
For any $\beta >0$ and any $R_\beta > N^{- \beta}$ we define the potential $M_\beta$ via
\begin{align}
\label{eq: defW}
M_\beta(x)
=
\begin{cases}
4 \pi N^{-1+2 \beta} & \text{if } N^{- \beta} < |x| \leq R_\beta \\
0 & \text{else }
\end{cases}
\end{align}
Furthermore, we define the zero energy state $f_\beta$ of the potential
$ \frac{1}{2} (V_N(x)-M_\beta(x) )$, that is
\begin{align}
\label{eq: deff}
\begin{cases}
\left( - \Delta_x + \frac{1}{2} \left(V_N(x)-M_\beta(x) \right) \right) f_{\beta}(x)=0
\\
f_{\beta}(x)=1 \; \text{for } |x| = R_\beta
\end{cases}
\end{align}
\end{definition}
Note that $M_\beta$ and $f_\beta$ depend on $R_\beta$.
We choose $R_\beta$ such that the scattering length of the potential
$\frac{1}{2} \left( V_N-M_\beta(x) \right)$ is zero. This is equivalent to the condition
$\int_{\mathbb{R}} d^2x (V_N(x)-M_\beta(x)) f_\beta(x) =0$.
\begin{lemma}\label{defAlemma}
For the scattering state $f_\beta$, defined by \eqref{eq: deff}, the following relations hold:
\begin{enumerate}
\item[(a)] There exists a minimal value $R_\beta <\infty$ such that $\int_{\mathbb{R}^2} d^2x (V_N(x)-M_\beta(x)) f_\beta(x) =0$.
\end{enumerate}
For the rest of the paper we assume that $R_\beta$ is chosen such that (a) holds.
\begin{enumerate}
\item[(b)] There exists $K_{\beta} \in \mathbb{R}, \; K_{\beta}> 0$ such that
$K_{\beta} f_{\beta}(x) = j_{N,R_\beta}(x) \; \forall |x| \leq N^{-\beta}$.
\item[(c)] For $N$ sufficiently large the supports of $V_N$ and $M_{\beta}$ do not overlap.
\item[(d)]
$f_{\beta}$ is a nonnegative, monotone nondecreasing function in $|x|$.
\item[(e)]
\begin{align}
f_{\beta}(x)=1 \; \text{for } |x| \geq R_\beta \;.
\end{align}
\item[(f)]
\begin{align}
1 \geq K_\beta \ge
1+ \frac{1}{N+ \ln \left(\frac{R_\beta}{a}\right)} \ln \left( \frac{N^{-\beta}}{R_\beta} \right)
\; .
\end{align}
\item[(g)]
$ R_\beta \leq C N^{-\beta}$.
\end{enumerate}
For any fixed $0<\beta$, $N$ sufficiently large such that $V_N$ and $M_{\beta}$ do not overlap, we obtain
\begin{enumerate}
\item[(h)]
\begin{align*}
&| N \| V_{N}f_{\beta} \|_1 - 4 \pi |
=
| N \| M_{\beta}f_{\beta} \|_1 - 4 \pi | \leq C \frac{\ln(N)}{N}
\;.
\end{align*}
\item[(i)]
Define
\begin{align*}
g_\beta(x) = 1 - f_\beta(x)
\;.
\end{align*}
Then,
\begin{align*}
\|g_{\beta}\|_1&\leq C N^{-1-2\beta}\ln N \;,\hspace{1cm}\|g_{\beta}\|\leq C N^{-1-\beta}\ln N
\;,\hspace{1cm}
\|g_{\beta}\|_\infty \leq 1
\;.
\end{align*}
\item[(j)]
\begin{align*}
| N \| M_{\beta}\|_1 - 4 \pi | \leq C \frac{\ln(N)}{N}
\;.
\end{align*}
\item[(k)]
\begin{align*}
M_{\beta} \in \mathcal{V}_\beta
\;,
M_{\beta} f_\beta \in \mathcal{V}_\beta
\;.
\end{align*}
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}
\item[(a)]
In the following, we will sometimes denote,
with a slight abuse of notation, $f_\beta(x)=f_\beta(r)$ for $r=|x|$ (for this, recall that $f_\beta$ is radial symmetric).
We first show that $f_\beta(N^{-\beta}) \neq 0$.
Assume first that there exists a $x_0, |x_0| \leq N^{-\beta}$ such that $f_\beta (x_0) \neq 0$.
We may assume that $f_\beta (x_0) >0$ (otherwise take $-f_\beta$). By continuity of
$f_\beta$, there exists a maximal interval $ I = [a,b] \subset [0, N^{-\beta}]$ on which $f_\beta (x) \geq 0$ for all $|x| \in I$ holds.
Since $M_\beta(x)=0$ for all $x$ with $ |x| \leq N^{- \beta}$, it follows that
$ \Delta f_\beta= \frac{1}{2} V_{N} f_\beta $ for all $x$ with $|x| \leq N^{- \beta}$.
Using, for $|x| \in I$, $ \frac{1}{2} V_{N}(x) f_\beta(x) \geq 0$, we obtain $ \Delta_x f_\beta(x) \geq 0$ for all $|x| \in I$, which implies that
$f_\beta$ is subharmonic on $A= \lbrace x \in \mathbb{R}^2| a \leq |x| \leq b \rbrace$.
By the maximum principle,
$
\max_{x \in A} (f_\beta)
=
\max_{x \in \partial A} (f_\beta)
$ holds.
If it were now that $\max_{x \in \partial A} (f_\beta)=f_{\beta}(a) \geq f_\beta(x_0)>0$, we could conclude for $a>0$, using continuity of $f_\beta$, that
there exists an $\epsilon >0$, such that
$ f_\beta (x) \geq 0$ for all $a-\epsilon \leq |x| \leq b$. This, however contradicts the assumptions on $I$. Note that for $a=0$, we obtain $\partial A =
\lbrace x \in \mathbb{R}^2| |x| = b \rbrace$.
Thus, we may conclude that
$
\max_{x \in A} (f_\beta)
=
f_\beta (b)
$. Since $I$ is the maximal interval on which $f_\beta$ is positive, it then follows that $b=N^{-\beta}$. This shows that $f_\beta(N^{-\beta} )\neq
0$, assuming that there exists a $x_0, |x_0| \leq N^{-\beta}$ such that $f_\beta (x_0) \neq 0$.
Assume now that $f_\beta (x) =0$ for all $|x| \leq N^{-\beta}$. Then, \eqref{eq: deff} is equivalent to
\begin{align*}
\begin{cases}
\left( - \Delta_x - \frac{1}{2} M_\beta(x) \right) f_{\beta}(x)=0
\\
f_{\beta}(x)=1 \; \text{for } |x| = R_\beta
\\
f_{\beta}(x)=0 \; \text{for } |x| \leq N^{-\beta}
\;.
\end{cases}
\end{align*}
Since $f_\beta$ is continuous, $R_\beta>N^{-\beta}$ and $f_\beta (R_\beta)=1$ holds, there exists a minimal value $x_0$ and some $b>x_0$, such that $f_\beta (x) = 0$ for all $|x| \in [0,x_0]$, but $f_\beta (x) > 0$, for $|x| \in ]x_0,b[$ (otherwise take $-f_\beta$).
Using that $f_\beta$ is radially symmetric, it follows from Gauss-theorem and the scattering equation above that
\begin{align*}
\frac{\partial f_\beta}{\partial r}|_{|x|=r} \equiv
\nabla f_\beta |_{|x|=r} \cdot e_r=
-\frac{1}{ 4 \pi r }
\int_{B_r(0)} d^2 x M_\beta(x) f_\beta(x)
<0 \text{ for all } r \leq b
\;,
\end{align*}
which yields to a contradiction.
Therefore, we have shown that $f_\beta(N^{-\beta}) \neq 0$.
Applying Theorem C.1 in \cite{lssy} once more, it then follows that either $f_\beta$ or $-f_\beta$ is a nonnegative, monotone nondecrasing function in $|x|$
for all $|x| \leq N^{-\beta}$.
Using \eqref{eq: deff} and Gauss-theorem,
\begin{align}
\label{integral v-w scattstate}
\frac{\partial f_\beta}{\partial r}|_{|x|=r} =
\frac{1}{ 4 \pi r }
\int_{B_r(0)} d^2 x (V_N(x)-M_\beta(x)) f_\beta(x)
\;.
\end{align}
Thus, $R_\beta$ is the minimal value such that $\frac{\partial f_\beta}{\partial r}|_{R_\beta}=0$.
Therefore, $f_\beta$ or $-f_\beta$ is a nonnegative, monotone nondecreasing continuous function for all $|x| \leq R_\beta$.
Next, we show that $R_\beta< \infty$ by contradiction:
\\
Assume $R_\beta = \infty$. Since $f(N^{-\beta}) \neq 0$, we obtain $ |\int_{\mathbb{R}^2} d^2x W_\beta(x) f_\beta (x)| \geq
|f(N^{-\beta}) \int_{\mathbb{R}^2} d^2x W_\beta(x)| = \infty$, which yields to a contradiction since\\
$ |\int_{\mathbb{R}^2} d^2x W_\beta(x) f_\beta(x)|= |\int_{\mathbb{R}^2} d^2x V_N(x) f_\beta(x)| < \infty$.
\item[b)]
Since $f_\beta(N^{-\beta}) \neq 0$, we can define
\begin{equation*}
h(x) = f_\beta(x)\frac{j_{N,R_\beta}(N^{-\beta})}{f_\beta(N^{-\beta})}
\end{equation*}
on the compact set $\overline{B_{N^{-\beta}}(0)} = \{x \in \mathbb{R}^2 : |x| \leq N^{- \beta} \}$.
It is useful to note that the scattering equations \eqref{jasymp} and \eqref{eq: deff} have a unique solution on a compact set $\overline{B_{N^{-\beta}}(0)}$.
One easily sees that $h(x)= j_{N,R_\beta}(x)$ on $\partial \overline{B_{N^{-\beta}}(0)}$ and satisfies the zero energy scattering equation~\eqref{eq: defj}. By uniqueness it follows that
$h(x) = j_{N,R_\beta}(x) \; \forall x \in \overline{B_{N^{-\beta}}(0)}$.
We can conclude that
\begin{equation}
\label{defKbeta}
K_{\beta} = \frac{j_{N,R_\beta}(N^{-\beta})}{f_\beta(N^{-\beta})}.
\end{equation}
Next, we show that the constant $K_\beta$ is positive.
Since $j_{N,R_\beta}(N^{-\beta})$ is positive, it follows from Eq.~\eqref{defKbeta} that $K_\beta$ and $f_\beta (N^ {-\beta})$ have equal sign. By (a), the sign of $f_\beta$ is constant for $|x|\leq R_\beta$.
Furthermore, from Gauss-theorem and the scattering equation \eqref{eq: deff} we have
\begin{align}
\label{Gauss}
\frac{\partial f_\beta}{\partial r}=
\frac{1}{4 \pi r K_\beta}
\int_{B_r(0)}d^2x V_N(x) j_{N,R_\beta}(x)
\end{align}
for all $r \leq N^{- \beta}$. Since $j_{N,R_\beta}$ and $V_N$ are nonnegative functions,
\begin{align}
\label{eq: sign of derivative of f}
\text{sgn} \left(\frac{\partial f_\beta}{\partial r}|_{r= N^{- \beta}}\right) = \text{sgn}(K_\beta).
\end{align}
Recall that $R_\beta$ is the smallest value such that
$\frac{\partial f_\beta}{\partial r} \big|_{r=R_\beta} = 0$.
If it were now that $K_\beta$ is negative, we could conclude from \eqref{defKbeta} and \eqref{eq: sign of derivative of f} that
$\frac{\partial f_\beta}{\partial r}|_{r= N^{- \beta}} <0$ and
$ f_\beta ( N^{- \beta}) <0$.
Since $R_\beta$ is by definition the smallest value where $\frac{\partial f_\beta}{\partial r} = 0$, we were able to conclude from the continuity of the derivative that $\frac{\partial f_\beta}{\partial r} <0$ for all $r < R_\beta$ and hence $f(R_\beta) <0$. However, this were in contradiction to the boundary condition of the zero energy scattering state (see \eqref{eq: deff}) and thus $K_\beta > 0$ follows.
\item[(c)] This directly follows from $e^{-N} < C N^{- \beta}$ for $N$ sufficiently large.
\item[(d)]
From the proof of property b), we see that $f_\beta$ and its derivative is positive at $N^{- \beta}$.
From \eqref{integral v-w scattstate}, we obtain $\partial_r f_\beta(x) = 0 $ for all $x$ with $|x| > R_\beta$. Due to continuity $\partial_r f_{\beta}(x) >0 $ for all $x$ with $|x| < R_\beta$. Since $f_\beta$ is continuous, positive at $N^{- \beta}$, and its derivative is a nonnegative function, it follows that $f_\beta$ is a nonnegative, monotone nondecreasing function in $|x|$.
\item[(e)]
By definition of $R_\beta$, it follows that $\tilde{I}=\int_{\mathbb{R}^2}d^2x (V_N(x)-W_\beta(x)) f_\beta(x) =0$. Therefore, for all $|x|\geq R_\beta$, $f_\beta$ solves $-\Delta f_\beta(x)=0$, which has the solution
\begin{align*}
f_\beta(x)= 1 + \frac{\tilde{I}}{4 \pi} \ln \left( \frac{|x|}{R_\beta} \right)=1
\end{align*}
\item[(f)]
Since $f_\beta$ is a positive monotone nondecreasing function in $|x|$, we obtain
\begin{align*}
1 \geq f_\beta( N^{- \beta}) =j_{N,R_\beta}( N^{- \beta}) /K_\beta
=
\left(
1+ \frac{1}{N+ \ln \left(\frac{R_\beta}{a}\right)} \ln \left( \frac{N^{-\beta}}{R_\beta} \right)
\right)
/ K_\beta
\end{align*}
We obtain the lower bound
\begin{align*}
K_\beta \ge
1+ \frac{1}{N+ \ln \left(\frac{R_\beta}{a}\right)}\ln \left( \frac{N^{-\beta}}{R_\beta} \right).
\end{align*}
For the upper bound we first prove that $f_{\beta}(x) \geq j_{N,R_\beta}(x)$ holds for all $|x | \leq R_\beta$.
Using the scatting equations \eqref{jasymp} and \eqref{eq: deff} we obtain
\begin{align*}
\Delta_x
(
f_\beta(x)-j_{N,R_\beta}(x)
)
=
\frac{1}{2}
V_N(x)
(
f_\beta(x)-j_{N,R_\beta}(x)
)
-
W_\beta(x) f_{\beta}(x)
\end{align*}
as well as
$f_{\beta}(R_\beta)- j_{N,R_\beta}(R_\beta)=0$.
Since $W_\beta(x) f_{\beta}(x) \geq 0$, we obtain that
$\Delta_x
(
f_\beta(x)-j_{N,R_\beta}(x)
) \leq 0$ for $ N^{-\beta} \leq |x|\leq R_\beta$. That is, $f_\beta(x)-j_{N,R_\beta}(x)$ is superharmonic
for $ N^{-\beta} < |x|< R_\beta$. Using the minimum principle, we obtain, using that
$f_\beta-j_{N,R_\beta}$ is spherically symmetric
\begin{align}
\label{minimum}
\min_{N^{-\beta} \leq |x|\leq R_\beta} ( f_\beta-j_{N,R_\beta})
=
\min_{|x| \in \lbrace N^{-\beta} , R_\beta \rbrace} ( f_\beta-j_{N,R_\beta})
\end{align}
If it were now that $\min_{|x| \in \lbrace N^{-\beta} , R_\beta \rbrace} ( f_\beta-j_{N,R_\beta})
=f_{\beta}(N^{-\beta})- j_{N,R_\beta}(N^{-\beta})
\leq
f_{\beta}(R_\beta)- j_{N,R_\beta}(R_\beta)=0
$, we could conclude that $ f_{\beta}(x)- j_{N,R_\beta}(x) \leq 0$ for all
$ N^{-\beta} \leq |x|\leq R_\beta$. Since $f_\beta(x)-j_{N,R_\beta}(x)$ then obeys
\begin{align*}
\begin{cases}
-\Delta
(
f_\beta(x)-j_{N,R_\beta}(x)
)
+
\frac{1}{2}
V_N(x)
(
f_\beta(x)-j_{N,R_\beta}(x)
)
=
0
\;
&\text{for } |x| \leq N^{-\beta} ,
\\
f_\beta(x)-j_{N,R_\beta}(x) \leq 0 \; &\text{for } |x| = N^{-\beta},
\end{cases}
\end{align*}
we could then conclude that $f_\beta(x)-j_{N,R_\beta}(x) \leq 0$
for all $|x| \leq R_\beta$. From this, we obtain that
$\Delta
(
f_\beta(x)-j_{N,R_\beta}(x)
) \leq 0$ for $ |x|\leq R_\beta$. That is, $f_\beta(x)-j_{N,R_\beta}(x)$ is superharmonic for all
$ |x|\leq R_\beta$.
Using the minimum principle once again, we then obtain
\begin{align*}
\min_{\overline{B_{R_\beta}(0)}} ( f_\beta-j_{N,R_\beta})
=
f_\beta(R_\beta)-j_{N,R_\beta}(R_\beta)=0
\end{align*}
which contradicts $ f_{\beta}(x)- j_{N,R_\beta}(x) \leq 0$ for $|x| \leq R_\beta$.
Therefore, we can conclude in \eqref{minimum} that
$
\min_{N^{-\beta} \leq |x|\leq R_\beta} ( f_\beta-j_{N,R_\beta})
=
f_\beta(R_\beta)-j_{N,R_\beta}(R_\beta)=0
$ holds. Then, it follows that $f_\beta(x)-j_{N,R_\beta}(x) \geq 0$ for all $N^{-\beta} \leq |x| \leq R_\beta$. Using the zero energy scattering equation
$-\Delta
(
f_\beta(x)-j_{N,R_\beta}(x)
)
+
\frac{1}{2}
V_N(x)
(
f_\beta(x)-j_{N,R_\beta}(x)
)
=
0$ for $|x| \leq N^{-\beta}$, we can, together with $ f_\beta(N^{-\beta})-j_{N,R_\beta}(N^{-\beta}) \geq 0$, conclude that $ f_{\beta}(x)- j_{N,R_\beta}(x) \geq 0$ for all $|x| \leq R_\beta$.
As a consequence, we obtain the desired bound
$
K_\beta=
\frac{j_{N,R_\beta}(N^{-\beta})}{f_\beta(N^{-\beta})}\leq 1$.
\item[(g)]
Since $f_\beta$ is a nonnegative, monotone nondecreasing function in $|x|$ with $f_\beta (x)=1$ $\forall |x| \geq R_\beta$, it follows that
\begin{align*}
C f_\beta(N^{-\beta})
=&
f_\beta(N^{-\beta})
\int_{\mathbb{R}^2} d^2 x V_N(x)
\geq
\int_{\mathbb{R}^2} d^2 x V_N(x) f_\beta(x)
\\
=&
\int_{\mathbb{R}^2} d^2 x M_\beta(x) f_\beta(x)
\geq
f_{\beta} (N^{-\beta})
\int_{\mathbb{R}^2} d^2 x M_\beta (x)
\;.
\end{align*}
Therfore, $\int_{\mathbb{R}^2} d^2 x M_\beta (x) \leq C$ holds, which implies that $R_\beta \leq C N^{1/2- \beta}$.
From
\begin{align*}
\frac{1}{K_\beta} \frac{4 \pi}{N+ \ln \left(\frac{R_\beta}{a}\right)}
=&
\frac{1}{K_\beta}
\int_{\mathbb{R}^2}d^2x V_N(x) j_{N,R_\beta}(x)
=
\int_{\mathbb{R}^2}d^2x V_N(x) f_\beta(x)
\\
=&
\int_{\mathbb{R}^2}d^2x M_\beta(x) f_\beta(x)
=
8 \pi^2 N^{-1+2 \beta}
\int_{N^{-\beta}}^{R_\beta}
dr r f_\beta(r)
\end{align*}
we conclude that
\begin{align*}
\int_{N^{-\beta}}^{R_\beta}
dr r f_\beta(r)
=
\frac{N^{1- 2 \beta}}{2 \pi K_\beta
\left(
N+ \ln \left(\frac{R_\beta}{a} \right)
\right)
}
\;.
\end{align*}
Since $f_\beta$ is a nonegative, monotone nondecreasing function in $|x|$,
\begin{align*}
\frac{1}{2}( R_\beta^2- N^{-2 \beta})
\frac{j_{N,R_\beta}(N^{-\beta})}{K_\beta}
=
\frac{1}{2}( R_\beta^2- N^{-2 \beta})
f_\beta(N^{-\beta})
\leq
\int_{N^{-\beta}}^{R_\beta}
dr r f_\beta(r)
\end{align*}
which implies
\begin{align*}
R_\beta^2 N^{2 \beta}
\leq
\frac{N}{\pi
\left(N+ \ln \left(\frac{R_\beta}{a} \right) \right)
j_{N,R_\beta}(N^{-\beta})}+1
\end{align*}
Using $R_\beta\leq C N^{1/2- \beta}$, it then follows
\begin{align*}
j_{N,R_\beta}(N^{-\beta})
=
1+ \frac{1}{N+ \ln \left(\frac{R_\beta}{a}\right)} \ln \left( \frac{N^{-\beta}}{R_\beta} \right)
\geq 1- \frac{C}{N}
\;,
\end{align*}
which implies $R_\beta \leq C N^{-\beta}$.
\item[(h)]
Using
\begin{align*}
\| M_\beta f_\beta\|_1=&
\| V_N f_\beta \|_1 = K_\beta^ {-1} \|V_N j_{N,R_\beta} \|_1
=
K_\beta^{-1} \frac{4 \pi}{N+ \ln \left(\frac{R_\beta}{a}\right)}
\;,
\end{align*}
we obtain
\begin{align*}
|N\| V_N f_\beta\|_1- 4\pi|
=&
|N\| M_\beta f_\beta\|_1- 4\pi|
=
4 \pi
\left| K_\beta^{-1} \frac{N}{N+ \ln \left(\frac{R_\beta}{a}\right)}-1 \right|
\\
=&
\frac{4 \pi}{K_\beta}
\left| \frac{
N-NK_\beta+K_\beta
\ln \left(\frac{R_\beta}{a}\right)}{N+ \ln \left(\frac{R_\beta}{a}\right)} \right|
\leq C \frac{\ln(N)}{N}
\;.
\end{align*}
\item[(i)]
Using for $|x| \leq R_\beta$ the inequalities
$j_{N,R_\beta}(x) \geq 1+
\frac{1}{N+ \ln \left(\frac{R_\beta}{a}\right)}
\ln \left( \frac{|x|}{R_\beta} \right) $
as well as $1\ge f_{\beta}(x) \geq j_{N,R_\beta}(x)$, it follows for $|x| \leq R_\beta$
\begin{align*}
0\leq& g_{\beta}(x) = 1- f_{\beta}(x) \le 1- j_{N,R_\beta}(x)
\leq -
\frac{1}{N+ \ln \left(\frac{R_\beta}{a}\right)}
\ln \left( \frac{|x|}{R_\beta} \right)
\nonumber
\\
\label{gbound}
\leq&
C N^ {-1}
|
\ln \left(N |x| \right)
|
\;.
\end{align*}
Since $g_{\beta}(x)=0$ for $\vert x\vert > R_{\beta}$, we conclude with $R_\beta \leq C N^{-\beta}$
that
\begin{align*}
\|g_{\beta}\|_1
\leq &
\frac{C}{N}
\int_{0}^{R_{\beta}}dr r
|
\ln \left( N r\right)
|
\leq C N^{-1-2\beta} \ln N \;,
\end{align*}
as well as
\begin{align*}
\| g_{\beta}\|^2\leq&
\frac{C}{N^2}
\int_{0}^{R_{\beta}} dr r
\left( \ln \left( Nr \right) \right)^2 \\
= & CN^{-4}
\Big[ r ^2 ( 2 (\ln(r))^2 - 2\ln (r)+1)\Big]_0^{N R_{\beta}}\\
\le & C N^{-2-2\beta} \left(\ln (N)\right) ^2.
\end{align*}
$ \| g_{\beta}\|_\infty = \|1- f_{\beta}\|_\infty \leq 1$, since $f_\beta$ is a nonnegative, monotone nondecreasing function with $f_\beta(x) \leq 1$.
\item[(j)] Using (h) and (i), we obtain with $\| M_{\beta}\|_1 \leq CN^{-1}$
\begin{align*}
&| N \| M_{\beta}\|_1 - 4 \pi |
\leq
| N \| M_{\beta}f_\beta\|_1 - 4 \pi |
+
N \| M_{\beta}g_\beta\|_1
\\
& \leq
C
\left(
\frac{\ln(N)}{N} +
\|\mathds{1}_{|\cdot| \geq N^{-\beta}} g_\beta\|_\infty
\right)
\;.
\end{align*}
Since $g_\beta(x)$ is a nonnegative, monotone nonincreasing function, it follows with $K_\beta \leq 1$
\begin{align*}
& \|\mathds{1}_{|\cdot| \geq N^{-\beta}} g_\beta\|_\infty
=
g_\beta(N^{-\beta})=1-f_\beta(N^{-\beta})
=
1- \frac{j_{N,R_\beta}(N^{-\beta})}{K_\beta}
\\
\leq&
1-\left( 1 + \frac{1}{N+ \ln \left(\frac{R_\beta}{a}\right)} \ln \left( \frac{N^{- \beta}}{R_ \beta} \right) \right)
\;.
\end{align*}
and (j) follows.
\item[(k)]
$M_\beta \in \mathcal{V}_\beta$ follows directly from $R_\beta \leq C N^{-\beta}$.
Furthermore, $ 0 \leq M_\beta (x) f_\beta (x) \leq M_\beta(x)$ implies
$M_\beta f_\beta \in \mathcal{V}_\beta$.
\end{enumerate}
\end{proof}
\section{Proof of the Theorem}\label{secpro}
\subsection{Proof for the potential $W_\beta$}\label{secthe}
\subsubsection{Choosing the weight}
As we have already mentioned, we define a functional $\alpha:L^2(\mathbb{R}^{2N},\mathbb{C})\timesL^2(\mathbb{R}^2,\mathbb{C})\to\mathbb{R}^+_0$ such that
\begin{enumerate}
\item[(I)]
$\frac{\text{d}}{\text{d}t}\alpha(\Psi_t,\phi_t)$ can be estimated by $\alpha(\Psi_t,\phi_t)+\mbox{\begin{scriptsize}$\mathcal{O}$\end{scriptsize}}(1)$,
yielding to a bound of $\alpha(\Psi_t,\phi_t)$ via a Gr\"onwall estimate.
\item[(II)] $\alpha(\Psi,\phi)\to0$
implies convergence of the reduced one particle density matrix $ \gamma^{(1)}_\psi$ to
$|\phi\rangle\langle\phi|$ in trace norm.
\end{enumerate}
For $\beta>0$, the interaction gets peaked as $N\to\infty$ and one has to use smoothness properties of $\Psi_t$ to be able to control the dynamics of the condensate.
For small $\beta$ and many different choices of the weight, one obtains
$$
\frac{d}{dt}\alpha(\Psi_t,\phi_t)\leq \mathcal{K}(\phi_t, A_t)\left(\alpha(\Psi_t,\phi_t)+\mbox{\begin{scriptsize}$\mathcal{O}$\end{scriptsize}}(1)+\langle\hspace{-0.08cm}\langle\Psi_t,\widehat{n}^{\phi_t}\Psi_t\rangle\hspace{-0.08cm}\rangle+
\left|\mathcal{E}_{W_\beta}(\Psi_t)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi_t)\right|\right)\;.$$
This enables us to perform a Gr\"onwall-type estimate if we choose
$$\alpha(\Psi_t,\phi_t)=\left\langle\hspace{-0.14cm}\left\langle\Psi_t,\widehat{n}^{\phi_t}\Psi_t\right\rangle\hspace{-0.14cm}\right\rangle+\left|\mathcal{E}_{W_\beta}(\Psi_t)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi_t)\right|\;.$$
For large $\beta$, however, it is necessary to adjust the weight function for the following reason:
Taking the time derivative of $\alpha(\Psi_t,\phi_t)$, terms of the form $\widehat{n}-\widehat{n}_1$ and $\widehat{n}-\widehat{n}_2$ appear.
For $i=1,2$, the bound
$N\|\widehat{n}-\widehat{n}_i\|_{\text{op}}=\mathcal{O}(N^{1/2})$ can be easily verified.
For $\beta >1/2$ it is not possible to obtain a sufficient decay of $\frac{\text{d}}{\text{d}t} \alpha(\Psi_t,\phi_t)$, see
Lemma \ref{hnorms}, part (b).
For this reason, it is necessary to choose another weight function $\widehat{m}$ in such a way that
$N\|\widehat{m}-\widehat{m}_i\|_{\text{op}}$ is better to control.
\begin{definition}\label{defm}
For $0<\xi <\frac{1}{2}$ define
$$m(k)=\left\{
\begin{array}{ll}
\sqrt{k/N}, & \hbox{for $k\geq N^{1-2\xi}$;} \\
1/2(N^{-1+\xi}k+N^{-\xi}), & \hbox{else.}
\end{array}
\right.
$$
and
$$\alpha^<(\Psi,\phi)=\langle\hspace{-0.08cm}\langle\Psi,\widehat{m}^\phi\Psi\rangle\hspace{-0.08cm}\rangle+\left|\mathcal{E}_{W_\beta}(\Psi)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi)\right|\;.$$
\end{definition}
With this definition, we obtain
$N \|\widehat{m}-\widehat{m}_1\|_{\text{op}} \leq C N^{\xi}$, see \eqref{estab}.
\begin{lemma}\label{equiv}
Let $\Psi \in L^2_s(\mathbb{R}^{2N},\mathbb{C})$ and let $\phi \in L^2(\mathbb{R}^2,\mathbb{C})$.
Let $\alpha^<(\Psi,\phi)$ be defined as above. Then,
\begin{align*}\lim_{N\to\infty}\alpha^<(\Psi,\phi)=0 \;\;\Leftrightarrow \hspace{1cm}&\lim_{N\to\infty}\gamma^{(1)}_{\Psi}=|\phi\rangle\langle\phi|\text{ in trace norm}\\\text{ and }&\lim_{N\to\infty}(\mathcal{E}_{W_\beta}(\Psi)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi))=0\;.
\end{align*}
\end{lemma}
A proof of this Lemma can be found in \cite{picklgp3d}.
Thus, $\alpha(\Psi_t,\phi_t)$ satisfies condition (II).
Next, we calculate the time derivative of $\alpha(\Psi_t,\phi_t)$ in order to obtain the desired Gr\"onwall estimate.
For this, define
\begin{definition}\label{alphasplit}
Let $W_\beta \in \mathcal{V}_\beta$. Define
\begin{equation}\label{defz}Z^\phi_\beta(x_j,x_k)=W_{\beta}(x_j-x_k)-\frac{N\|W_{\beta}\|_1}{N-1}|\phi|^2(x_j)-\frac{N\|W_{\beta}\|_1}{N-1}|\phi|^2(x_k)
\;.
\end{equation}
Note, for $W_\beta(x)= N^{-1+2 \beta} W(N^\beta x)$, we have
$N \|W_\beta\|_1= \|W\|_1$.
With
$$
m^a(k)=m(k)-m(k+1) , \;\;\;\;\; m^b(k)=m(k)-m(k+2)
$$
and
$$\widehat{r}=\widehat{m}^bp_{1}p_{2}+\widehat{m}^a(p_{1}q_{2}+q_1p_2) \;,$$
we define the functionals $\gamma^<_{a,b}:L^2(\mathbb{R}^{2N},\mathbb{C}) \times L^2(\mathbb{R}^2, \mathbb{C})
\to\mathbb{R}^+_0$ by
\begin{align}\label{fnochdao} \gamma^<_a(\Psi,\phi)%
=&\left|\langle\hspace{-0.08cm}\langle\Psi,\dot A_t(x_1)\Psi\rangle\hspace{-0.08cm}\rangle-\langle\phi,\dot A_t\phi\rangle\right|\\
\label{fnochda}
\gamma^<_b(\Psi,\phi)%
=&N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,Z^\phi_\beta(x_1,x_2) \widehat{r}
\Psi\rangle\hspace{-0.08cm}\rangle\right)
\\
\label{asaaltsplit}
=-&2N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,p_{1}q_{2}\widehat{m}^a_{-1}Z^\phi_\beta(x_1,x_2)p_{1}p_{2}
\Psi\rangle\hspace{-0.08cm}\rangle\right)
\\\nonumber
-&N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,q_{1}q_{2}\widehat{m}^b_{-2}W_\beta(x_1-x_2) p_{1}p_{2}
\Psi\rangle\hspace{-0.08cm}\rangle\right)
\\\nonumber
-&2N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,q_{1}q_{2}\widehat{m}^a_{-1}Z^\phi_\beta(x_1,x_2)p_{1}q_{2}
\Psi\rangle\hspace{-0.08cm}\rangle\right)\;.
\end{align}
\end{definition}
\begin{lemma}\label{ableitung}
Let $W_\beta \in \mathcal{V}_\beta$.
Let $\Psi_t$ the unique solution to $i \partial_t \Psi_t
= H_{W_\beta} \Psi_t$ with initial datum $\Psi_0 \in L^2_{s}(\mathbb{R}^{2N}, \mathbb{C}) \cap H^2(\mathbb{R}^{2N}, \mathbb{C}) ,\;
\|\Psi_0\|=1$.
Let $\phi_t$ the unique solution to $i \partial_t \phi_t
= h^{GP}_{N \|W_\beta\|_1} \phi_t$ with initial datum $\phi_0 \in H^2(\mathbb{R}^2,\mathbb{C}) \;,\|\phi_0\|=1$.
Let $\alpha^<(\Psi_t,\phi_t)$ be defined as in Definition \ref{defm}. Then
\begin{equation}\label{lemmaableitungeq}\left|\frac{d}{dt} \alpha^<(\Psi_t,\phi_t)\right|\leq\gamma_a^<(\Psi_t,\phi_t)+\gamma_b^<(\Psi_t,\phi_t)\;.\end{equation}
\end{lemma}
\begin{proof} For the proof of the Lemma we restore the upper index $\phi_t$ in order to pay respect to the time dependence of
$\widehat{m}^{\phi_t}$.
The time derivative of $\phi_t$ is given by (\ref{GP}), i.e. $i\partial_t \phi_t(x_j)=h_{N \|W_\beta\|_1,j}^{GP}\phi_t(x_j)$. Here, $h_{N \|W_\beta\|_1,j}^{GP}$ denotes the operator $h_{N \|W_\beta\|_1}^{GP}$ acting on the $j^{\text{th}}$ coordinate $x_j$.
We then obtain
\begin{align*}&
\frac{d}{dt}\langle\hspace{-0.08cm}\langle\Psi_t,\widehat{m}^{\phi_t}\Psi_t\rangle\hspace{-0.08cm}\rangle
\\=&
i\langle\hspace{-0.08cm}\langle H_{W_\beta}\Psi_t
,\widehat{m}^{\phi_t}\;\Psi_t\rangle\hspace{-0.08cm}\rangle
-
i\langle\hspace{-0.08cm}\langle\Psi_t
,\widehat{m}^{\phi_t}\;H_{W_\beta}\Psi_t\rangle\hspace{-0.08cm}\rangle
-
i\langle\hspace{-0.08cm}\langle \Psi_t
,[\sum_{j=1}^N h_{N \|W_\beta\|_1,j}^{GP},\widehat{m}^{\phi_t}\;]\Psi_t\rangle\hspace{-0.08cm}\rangle
\nonumber
\\=&i\langle\hspace{-0.08cm}\langle\Psi_t
,[H_{W_\beta}-\sum_{j=1}^N h_{N \|W_\beta\|_1,j}^{GP},\widehat{m}^{\phi_t}\;]\Psi_t\rangle\hspace{-0.08cm}\rangle=i\frac{N(N-1)}{2}\langle\hspace{-0.08cm}\langle\Psi_t
,[Z^{\phi_t}_\beta(x_1,x_2),\widehat{m}^{\phi_t}\;]\Psi_t\rangle\hspace{-0.08cm}\rangle
\;,
\end{align*}
where we used symmetry of $\Psi_t$ in the last step.
Using Lemma \ref{kombinatorik} (d), it follows that the latter equals (dropping the explicit dependence on $\phi_t$ from now on)
\begin{align*}&i\frac{N(N-1)}{2}\langle\hspace{-0.08cm}\langle\Psi_t
,[Z^{\phi_t}_\beta(x_1,x_2),p_1p_2(\widehat{m}-\widehat{m}_2)]\Psi_t\rangle\hspace{-0.08cm}\rangle
\\
+&i\frac{N(N-1)}{2}\langle\hspace{-0.08cm}\langle\Psi_t
,[Z^{\phi_t}_\beta(x_1,x_2),(p_1q_2+q_1p_2)(\widehat{m}-\widehat{m}_1)]\Psi_t\rangle\hspace{-0.08cm}\rangle\;.
\end{align*}
Since $Z^{\phi_t}_\beta$ and $p_1p_2(\widehat{m}-\widehat{m}_2)$ as well as $p_1q_2(\widehat{m}-\widehat{m}_1)$ are selfadjoint, we obtain
\begin{align*}
&\frac{d}{dt}\langle\hspace{-0.08cm}\langle\Psi_t,\widehat{m}^{\phi_t}\Psi_t\rangle\hspace{-0.08cm}\rangle
=-N(N-1)\\&\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,(p_{1}p_{2}+p_{1}q_{2}+q_{1}p_{2}+q_{1}q_{2})Z^{\phi_t}_\beta(x_1,x_2)( \widehat{m}^bp_{1}p_{2}+\widehat{m}^a(p_{1}q_{2}+q_1p_2))
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)\;.
\end{align*}
Note that in view of Lemma \ref{kombinatorik} (c) $\widehat{r}Q_jZ^{\phi_t}_\beta(x_1,x_2) Q_j= Q_jZ^{\phi_t}_\beta(x_1,x_2) Q_j\widehat{r}$ for any $j\in\{0,1,2\}$ and any weight $r$. Therefore,
\begin{align*}
\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,p_{1}p_{2}Z^{\phi_t}_\beta(x_1,x_2) \widehat{m}^bp_{1}p_{2}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)&=0\\
\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,(p_{1}q_{2}+q_{1}p_{2})Z^{\phi_t}_\beta(x_1,x_2)\widehat{m}^a(p_{1}q_{2}+q_1p_2)
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)&=0\;.
\end{align*}
Using Symmetry and Lemma \ref{kombinatorik} (c), we obtain the first line \eqref{fnochda}.
Furthermore,
\begin{align*}
\frac{d}{dt}\langle\hspace{-0.08cm}\langle\Psi_t,\widehat{m}^{\phi_t}\Psi_t\rangle\hspace{-0.08cm}\rangle%
=&-2N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t , \widehat{m}^b_{-1}p_{1}q_{2}Z^{\phi_t}_\beta(x_1,x_2)p_{1}p_{2}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\
-&N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,\widehat{m}^b_{-2}q_{1}q_{2}Z^{\phi_t}_\beta(x_1,x_2)p_{1}p_{2}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\
-&2N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,p_{1}p_{2}Z^{\phi_t}_\beta(x_1,x_2) \widehat{m}^ap_{1}q_{2}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\
-&2N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,\widehat{m}^a_{-1}q_{1}q_{2}Z^{\phi_t}_\beta(x_1,x_2)p_{1}q_{2}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)\;.
\end{align*}
Since $p_1p_2|\phi_t^2|(x_1)q_1q_2=p_1p_2q_2|\phi_t^2|(x_1)q_1=0=p_1p_2|\phi_t^2|(x_2)q_1q_2$, we can replace $Z^{\phi_t}_\beta(x_1,x_2)$ in the second line by $W_\beta(x_1-x_2)$.
The third line equals $2N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,\widehat{m}^a p_{1}q_{2}Z^{\phi_t}_\beta(x_1,x_2) p_{1}p_{2}
\Psi\rangle\hspace{-0.08cm}\rangle\right)$. Since
$$m(k-1)-m(k+1)-\left(m(k)-m(k+1)\right)=m(k-1)-m(k)$$
it follows that $\widehat{m}^b_{-1}-\widehat{m}^a=\widehat{m}^a_{-1}$ and we get
\begin{align*}
\frac{d}{dt}\langle\hspace{-0.08cm}\langle\Psi_t,\widehat{m}^{\phi_t}\Psi_t\rangle\hspace{-0.08cm}\rangle%
=&-2N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,p_{1}q_{2}\widehat{m}^a_{-1}Z^{\phi_t}_\beta(x_1,x_2)p_{1}p_{2}
\Psi\rangle\hspace{-0.08cm}\rangle\right)
\\\nonumber
-&N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,q_{1}q_{2}\widehat{m}^b_{-2}W_\beta(x_1-x_2) p_{1}p_{2}
\Psi\rangle\hspace{-0.08cm}\rangle\right)
\\\nonumber
-&2N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,q_{1}q_{2}\widehat{m}^a_{-1}Z^{\phi_t}_\beta(x_1,x_2)p_{1}q_{2}
\Psi\rangle\hspace{-0.08cm}\rangle\right)\;.
\end{align*}
For the second summand of $\alpha^<(\Psi_t,\phi_t)$ we have
\begin{align*}&
\nonumber\frac{d}{dt}\left(\mathcal{E}_{W_\beta}(\Psi_t)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi_t)\right)
=\langle\hspace{-0.08cm}\langle\Psi_t,\dot A_t(x_1)\Psi_t\rangle\hspace{-0.08cm}\rangle-\langle\phi_t,\dot A_t\phi_t\rangle
\\\nonumber&i
\left\langle\phi_t,\left[h_{N \|W_\beta\|_1}^{GP},\left(h_{N \|W_\beta\|_1}^{GP}-\frac{N \|W_\beta\|_1}{2}|\phi_t|^2\right)\right]\phi_t\right\rangle
+
\left\langle\phi_t,\frac{N \|W_\beta\|_1}{2}\left(\frac{d}{dt}|\phi_t|^2\right)\phi_t\right\rangle
\\\nonumber=&\langle\hspace{-0.08cm}\langle\Psi_t,\dot A_t(x_1)\Psi_t\rangle\hspace{-0.08cm}\rangle-\langle\phi_t,\dot A_t\phi_t\rangle
+i
\left\langle\phi_t,\left[h_{N \|W_\beta\|_1}^{GP},\frac{N \|W_\beta\|_1}{2}|\phi_t|^2\right]\phi_t\right\rangle
\\&
-
i
\left\langle\phi_t,\left[h_{N \|W_\beta\|_1}^{GP},\frac{N \|W_\beta\|_1}{2}|\phi_t|^2\right]\phi_t\right\rangle\;.
\end{align*}
Hence $\frac{\text{d}}{\text{d}t}\left|\mathcal{E}_{W_\beta}(\Psi_t)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi_t)\right|\leq\gamma_a^<(\Psi_t,\phi_t)$
which proves the Lemma.
\end{proof}
\subsubsection{Establishing the Gr\"onwall estimate}\label{secest}
\begin{lemma} \label{gammalemma}
Let $W_\beta \in \mathcal{V}_\beta$.
Let $\Psi_t$ the unique solution to $i \partial_t \Psi_t
= H_{W_\beta} \Psi_t$ with initial datum
$\Psi_0 \in L^2_{s}(\mathbb{R}^{2N}, \mathbb{C}) \cap H^2(\mathbb{R}^{2N}, \mathbb{C}) \;, \|\Psi_0\|=1$.
Let $\phi_t$ the unique solution to $i \partial_t \phi_t
= h^{GP}_{N \|W_\beta\|_1} \phi_t$ with initial datum $\phi_0 \in H^{3}(\mathbb{R}^2,\mathbb{C})$.
Let $\mathcal{E}_{W_\beta}(\Psi_0) \leq C$.
Let $\gamma^<_{a}(\Psi_t,\phi_t)$ and $\gamma^<_{b}(\Psi_t,\phi_t)$ be defined as in Definition \eqref{alphasplit}.
Then, there exists an $\eta>0$ such that
\begin{align}
\label{abc}
\gamma_a^<(\Psi_t,\phi_t) \leq &
C \|\dot A_t\|_\infty ( \langle\hspace{-0.08cm}\langle\Psi_t,\widehat{n}^{\phi_t}\Psi_t\rangle\hspace{-0.08cm}\rangle + N^{-\frac{1}{2}} )
\\
\label{theoremeq1}
\gamma_b^<(\Psi_t,\phi_t) \leq &
\mathcal{K}(\phi_t, A_t)
\left(
\langle\hspace{-0.08cm}\langle\Psi_t,\widehat{n}^{\phi_t}\Psi_t\rangle\hspace{-0.08cm}\rangle + N^{-\eta}+
\left|
\mathcal{E}_{W_\beta}(\Psi_0)
-
\mathcal{E}_{N \|W_\beta\|_1}^ {GP}(\phi_0)
\right|
\right)
\end{align}
\end{lemma}
The prove of this Lemma can be found in Section \ref{contgamma1}.
Note that
$$| \langle\hspace{-0.08cm}\langle\Psi_t,\widehat{n}^{\phi_t}\Psi_t\rangle\hspace{-0.08cm}\rangle- \langle\hspace{-0.08cm}\langle\Psi_t,\widehat{m}^{\phi_t}\Psi_t\rangle\hspace{-0.08cm}\rangle|
\leq
\|\widehat{n}^{\phi_t}-\widehat{m}^{\phi_t}\|_{\text{op}}
=
N^{-\xi}
$$
Once we have proven (\ref{theoremeq1}) we get with Lemma \ref{ableitung}, Gr\"onwall's Lemma and the estimate above that
\begin{align*}
&\alpha^<(\Psi_t,\phi_t)\leq
e^{
\int_0^t ds \mathcal{K}(\phi_s, A_s)}
\Big(
\alpha^<(\Psi_0,\phi_0)
\\
&+
\int_0^t ds
\mathcal{K}(\phi_s, A_s)
e^{-\int_0^s d \tau \mathcal{K}(\phi_\tau, A_\tau)}
N^{-\eta}
\Big) \;.
\end{align*}
Note that under the assumptions $\phi_t \in H^{3}(\mathbb{R}^2,\mathbb{C})$ and $
A_t \in L^\infty(\mathbb{R}^2, \mathbb{C}),\; \dot{A}_t \in L^\infty(\mathbb{R}^2, \mathbb{C})$
there exists a constant $C_t < \infty$, depending on $t$, $\phi_0$ and $A_t$, such that
$\int_0^t ds
\mathcal{K}(\phi_s, A_s) \leq C_t$, see Lemma \ref{regularityLemma}.
This proves, using Lemma \ref{equiv}, part (a) of Theorem \ref{theo}.
If the potential is switched off, one expects that $C_t$ is of order $t$ since in this case $\|\phi_t\|_\infty$
and $\|\nabla \phi_t\|_\infty$ are expected to decay like $t^{-1}$.
We want to explain on a heuristic level why $\gamma^<_{b}(\Psi_t,\phi_t)$ is small.
The principle argument follows the ideas and estimates of \cite{picklgp3d}.
The first line in (\ref{asaaltsplit}) is the most important one. This expression is only small if the correct coupling parameter $N \|W_\beta\|_1$ is used in the mean-field equation \eqref{GP}. Then,
$$N p_1 W_\beta(x_1-x_2)p_1=N p_1W_\beta\star|\phi|^2(x_2)p_1
\rightarrow p_1 |\phi|^2(x_2) \|W\|_1 p_1 $$
converges against the mean-field potential, and hence the first expression of (\ref{asaaltsplit}) is small.
In order to bound the second and third line of (\ref{asaaltsplit}), one tries to bound \\$N^2\langle\hspace{-0.08cm}\langle\Psi ,q_{1}q_{2}\widehat{m}^b_{-2}W_\beta(x_1-x_2) p_{1}p_{2}
\Psi\rangle\hspace{-0.08cm}\rangle$ and $N^2\langle\hspace{-0.08cm}\langle\Psi ,q_{1}q_{2}\widehat{m}^a_{-1}W_\beta(x_1-x_2) p_{1}q_{2}
\Psi\rangle\hspace{-0.08cm}\rangle$ in terms of $\langle\hspace{-0.08cm}\langle \Psi, \widehat{n} \Psi \rangle\hspace{-0.08cm}\rangle + \mathcal{O}(N^{-\eta})$ for some $\eta>0$.
For large $\beta$, one needs to use additional smoothness properties of $\Psi_t$. This explains the appearance on $\left|
\mathcal{E}_{W_\beta}(\Psi_0)
-
\mathcal{E}_{N \|W_\beta\|_1}^ {GP}(\phi_0)
\right|$ on the right hand side of \eqref{theoremeq1}. The concise estimates are quite involved and can be found in Section \ref{contgamma1}.
\subsection{Proof for the exponential scaling $V_N$}
\subsubsection{Adapting the weight}
For the most involved scaling $V_N$ it is necessary to modify the counting functional $\alpha^ <(\Psi, \varphi)$ in order to obtain the desired Gr\"onwall estimate. $\gamma^<_b (\Psi, \phi)$, which was defined in \eqref{asaaltsplit}, will not be small if we were to replace $W_\beta$ by $V_N$.
In particular, $ \|V_N\|=\mathcal{O}(e^{N})$ cannot be bounded by any finite polynomial in $1/N$.
In order to control the dynamics of the condensate, one needs to account for the
microscopic structure which is induced by $V_N$, as explained in Section \ref{secmic}.
The idea we will employ is the following:
For the moment, think of the most simple counting functional, namely
$\langle\hspace{-0.08cm}\langle\Psi_t,q^{\phi_t}_1\Psi_t\rangle\hspace{-0.08cm}\rangle=1-\langle\hspace{-0.08cm}\langle\Psi_t,p^{\phi_t}_1\Psi_t\rangle\hspace{-0.08cm}\rangle$. $1-\langle\hspace{-0.08cm}\langle\Psi_t,p^{\phi_t}_1\Psi_t\rangle\hspace{-0.08cm}\rangle$ counts the relative number of particles which are not in the state $\phi_t$. Instead of projecting onto $\phi_t$, we now consider the functional
$$1-\langle\hspace{-0.08cm}\langle\Psi_t,\prod_{j=2}^Nf_{\beta}(x_1-x_j)p^{\phi_t}_1\prod_{j=2}^Nf_{\beta}(x_1-x_j)\Psi_t\rangle\hspace{-0.08cm}\rangle\;,$$ which takes the short scale correlation structure into account.
Neglecting all but two-particle interactions, this can be approximated by
\begin{align*}&1-\langle\hspace{-0.08cm}\langle\Psi_t,\left(1-\sum_{j=2}^Ng_{\beta}(x_1-x_j)\right)p^{\phi_t}_1\left(1-\sum_{j=2}^Ng_{\beta}(x_1-x_j)\right)\Psi_t\rangle\hspace{-0.08cm}\rangle
\\& \approx\langle\hspace{-0.08cm}\langle\Psi_t,q^{\phi_t}_1\Psi_t\rangle\hspace{-0.08cm}\rangle+2(N-1)\Re\left(\langle\hspace{-0.08cm}\langle\Psi_t
, g_{\beta}(x_{1}-x_{2}) p^{\phi_t}_1
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)\;.
\end{align*}
If we now take the time derivative of this new functional, one gets, among other terms,
$ 2
(N-1)
\Im
\langle\hspace{-0.08cm}\langle \Psi_t,
[H_{V_N},f_\beta(x_1-x_2) ]p^{\phi_t}_1
\Psi_t \rangle\hspace{-0.08cm}\rangle$. The commutator equals $f_{\beta}(x_1-x_2)(V_{N}(x_1-x_2)-M_{\beta}(x_1-x_2)$ plus mixed derivatives and one
sees, that $V_N$ is ``replaced''
by $M_{\beta}$ for the price of new terms that have to be estimated.
The strategy we are going to employ is thus to estimate the time derivative of the modified functional and to show that we obtain a Gr\"onwall estimate.
Note, that, using Lemma \ref{kombinatorik} (e) with Lemma \ref{defAlemma} (i)
\begin{equation*}
2(N-1)\Re\left(\left\langle\hspace{-0.14cm}\left\langle\Psi_t
, g_{\beta}(x_{1}-x_{2}) p^{\phi_t}_1
\Psi_t\right\rangle\hspace{-0.14cm}\right\rangle\right)
\leq CN
\|\phi\|_\infty
\|g_\beta\|
\leq C
\|\phi_t\|_\infty
N^{-\beta} \ln(N)
\end{equation*}
holds.
Hence, we obtain the a priori estimate
$$
\langle\hspace{-0.08cm}\langle\Psi_t,q^{\phi_t}_1\Psi_t\rangle\hspace{-0.08cm}\rangle
\leq
\langle\hspace{-0.08cm}\langle\Psi_t,q^{\phi_t}_1\Psi_t\rangle\hspace{-0.08cm}\rangle+2(N-1)\Re\left(\langle\hspace{-0.08cm}\langle\Psi_t
, g_{\beta}(x_{1}-x_{2}) p^{\phi_t}_1
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)
+
C
\|\phi_t\|_\infty
N^{-\beta} \ln(N)
\;.
$$
which explains why the new defined functional implies convergence of the reduced density matrix
$\gamma^{(1)}_{\Psi_t}$ to $|\phi_t \rangle \langle \phi_t|$ in trace norm.
We now adapt the strategy explained above to modify the counting functional
$\alpha^< (\Psi, \phi)$.
\begin{definition}\label{lambda1}
Let $\widehat{r}=\widehat{m}^b p_{1}p_{2}+\widehat{m}^a(p_{1}q_{2}+q_1p_2)$.
We define the functional $\alpha:L^2(\mathbb{R}^{2N},\mathbb{C})\timesL^2(\mathbb{R}^2,\mathbb{C})\to\mathbb{R}^+_0$ by
\begin{align} \label{alphafuerV}
\alpha(\Psi,\phi) =&
\langle\hspace{-0.08cm}\langle \Psi, \widehat{m} \Psi \rangle\hspace{-0.08cm}\rangle
+
\left|
\mathcal{E}_{V_N} (\Psi)
-
\mathcal{E}^{GP}_{4 \pi} (\phi)
\right|
-N(N-1)\Re\left(\langle\hspace{-0.08cm}\langle\Psi
, g_{\beta}(x_{1}-x_{2}) \widehat{r}
\Psi\rangle\hspace{-0.08cm}\rangle\right)
\end{align}
and the functional $\gamma:L^2(\mathbb{R}^{2N},\mathbb{C})\timesL^2(\mathbb{R}^2,\mathbb{C})\to\mathbb{R} $ by
\begin{align} \label{gamma fuer V}
\gamma(\Psi,\phi)=\asa (\Psi,\phi)+\gamma_b (\Psi,\phi)+\gamma_c (\Psi,\phi)+\gamma_d (\Psi,\phi)+\gamma_e (\Psi,\phi)
+
\gamma_f (\Psi, \phi)
\;,
\end{align}
where the different summands are
\begin{enumerate}
\item[(a)] The change in the energy-difference
$$\asa (\Psi,\phi)=\left|\langle\hspace{-0.08cm}\langle\Psi,\dot A_t(x_1)\Psi\rangle\hspace{-0.08cm}\rangle-\langle\phi,\dot A_t\phi\rangle\right|\;.$$
\item [(b)] The new interaction term
\begin{align*}\gamma_b (\Psi,\phi)=&-N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,
\widetilde{Z}_{\beta}^\phi(x_1,x_2)\widehat{r}\,
\Psi\rangle\hspace{-0.08cm}\rangle\right)\\&
-N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
\widehat{r}\,\mathcal{Z}^\phi(x_1,x_2) \Psi\rangle\hspace{-0.08cm}\rangle\right)
\end{align*}
where, using $M_\beta$ as in Definition \ref{microscopic},
\begin{align}
\label{defzz}
&\widetilde{Z}_{\beta}^\phi(x_1,x_2)=\left(M_{\beta}(x_1-x_2)-4 \pi\frac{|\phi|^2(x_1)+|\phi|^2(x_2)}{N-1}\right)f_{\beta}(x_{1}-x_{2})
\\
&
\nonumber
\mathcal{Z}^\phi(x_1,x_2)=
V_N(x_1-x_2)-
\frac{4 \pi}{N-1} |\phi|^ 2(x_1)
-
\frac{4 \pi}{N-1} |\phi|^ 2(x_2)
\;.
\end{align}
\item [(c)] The mixed derivative term
\begin{align*} \gamma_c (\Psi,\phi)=&-4N(N-1)\langle\hspace{-0.08cm}\langle\Psi
, (\nabla_1g_{\beta}(x_1-x_2))\nabla_1
\widehat{r}\Psi\rangle\hspace{-0.08cm}\rangle\;.\end{align*}
\item [(d)] Three particle interactions
\begin{align*}\gamma_d (\Psi,\phi)=&2N(N-1)(N-2)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
\left[V_N(x_1-x_3),
\widehat{r}\right] \Psi\rangle\hspace{-0.08cm}\rangle\right)
\\&N(N-1)(N-2)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
\left[4 \pi |\phi|^2(x_3),
\widehat{r}\right] \Psi\rangle\hspace{-0.08cm}\rangle\right)\;.\end{align*}
\item [(e)] Interaction terms of the correction
\begin{align*}&\gamma_e (\Psi,\phi)=\frac{1}{2}N(N-1)(N-2)(N-3
\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,
g_{\beta}(x_{1}-x_{2})\left[V_N(x_3-x_4),
\widehat{r}\right] \Psi\rangle\hspace{-0.08cm}\rangle\right)\;.
\end{align*}
\item[(f)] Correction terms of the mean field
\begin{align*}
\gamma_f(\Psi,\phi)=
2
N(N-1)\frac{N-2}{N-1}\Im\left(\langle\hspace{-0.08cm}\langle\Psi,g_{\beta}(x_{1}-x_{2})
\left[4 \pi |\phi|^2(x_1),
\widehat{r}\right] \Psi\rangle\hspace{-0.08cm}\rangle\right)
\end{align*}
\end{enumerate}
\end{definition}
\begin{lemma}\label{secondadjlemma}
Let $\Psi_t$ the unique solution to $i \partial_t \Psi_t
= H_{V_N} \Psi_t$ with initial datum $\Psi_0 \in L^2_{s}(\mathbb{R}^{2N}, \mathbb{C}) \cap H^2(\mathbb{R}^{2N}, \mathbb{C}) \;,
\|\Psi_0\|=1$.
Let $\phi_t$ the unique solution to $i \partial_t \phi_t
= h^{GP}_{4 \pi} \phi_t$ with initial datum $\phi_0 \in H^2(\mathbb{R}^2,\mathbb{C}) \;,\|\phi_0\|=1$.
Let $\alpha(\Psi_t,\phi_t)$ and $\gamma(\Psi_t,\phi_t)$ be defined as in \eqref{alphafuerV} and \eqref{gamma fuer V}. Then
$$\frac{\text{d}}{\text{d}t}\alpha(\Psi_t,\phi_t)\leq |\gamma(\Psi_t,\phi_t) |\;.$$
\end{lemma}
\begin{proof}
\begin{align*}
\frac{\text{d}}{\text{d}t}\alpha(\Psi_t,\phi_t)=&\frac{\text{d}}{\text{d}t}
\left|
\mathcal{E}_{V_N} (\Psi)
-
\mathcal{E}^{GP}_{4 \pi} (\phi)
\right|
-N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,\mathcal{Z}^{\phi_t}(x_1,x_2) \widehat{r}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\&-N(N-1)\Re\left(i\langle\hspace{-0.08cm}\langle\Psi_t , g_{\beta}(x_{1}-x_{2})\left[H_{V_N}-\sum_{i=1}^N h^{GP}_{4 \pi,i},
\widehat{r}\right] \Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\&-N(N-1)\Re\left(i\langle\hspace{-0.08cm}\langle\Psi_t ,\left[H_{V_N},
g_{\beta}(x_{1}-x_{2})
\right]\widehat{r}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\end{align*}
Using symmetry and $\Re(iz)=-\Im(z)$
\begin{align*}
=&\frac{\text{d}}{\text{d}t}
\left|
\mathcal{E}_{V_N} (\Psi)
-
\mathcal{E}^{GP}_{4 \pi} (\phi)
\right|
-N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,\mathcal{Z}^{\phi_t}(x_1,x_2) \widehat{r}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\&+N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,g_{\beta}(x_{1}-x_{2})\left[\mathcal{Z}^{\phi_t}(x_1,x_2),
\widehat{r}\right] \Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\&+2N(N-1)(N-2)\Im\left(\langle\hspace{-0.08cm}\langle\Psi _t,g_{\beta}(x_{1}-x_{2})
\left[V_N(x_1-x_3),
\widehat{r}\right] \Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\&+N(N-1)(N-2)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t,g_{\beta}(x_{1}-x_{2})
\left[4 \pi |\phi_t|^2(x_3),
\widehat{r}\right] \Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\&+\frac{1}{2}N(N-1)(N-2)(N-3)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,
g_{\beta}(x_{1}-x_{2})\left[V_N(x_3-x_4),
\widehat{r}\right] \Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\&+N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,\left[H_{V_N},
g_{\beta}(x_{1}-x_{2})
\right]\widehat{r}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)\;.
\\
&+
2
N(N-1)\frac{N-2}{N-1}\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t,g_{\beta}(x_{1}-x_{2})
\left[4 \pi |\phi_t|^2(x_1),
\widehat{r}\right] \Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\end{align*}
The first term is bounded by $\gamma_a$, the third and fourth lines equal $\gamma_d$ (recall that $\Psi$ is symmetric), the fifth equals $\gamma_e$ and the seventh line equals $\gamma_f$.
Using that $(1-g_{\beta}(x_{1}-x_{2}))\mathcal{Z}^\phi(x_1,x_2)=\widetilde{Z}_{\beta}^\phi(x_1,x_2)+(V_{1}(x_1-x_2)-M_{\beta}(x_1-x_2))f_{\beta}(x_{1}-x_{2})$
we get
\begin{align}\label{commu}
\frac{\text{d}}{\text{d}t}\alpha(\Psi_t,\phi_t)
\leq&\gamma_a(\Psi_t,\phi_t)+\gamma_d(\Psi_t,\phi_t)+\gamma_e(\Psi_t,\phi_t)+\gamma_f(\Psi_t,\phi_t)
\nonumber
\\
&-N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,\widetilde{Z}_{\beta}^{\phi_t}(x_1,x_2) \widehat{r}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\\nonumber&-N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,
(V_{N}(x_1-x_2)-M_{\beta_1}(x_1-x_2))f_{\beta}(x_1-x_2)\widehat{r}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\\nonumber&-N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,g_{\beta}(x_{1}-x_{2})
\widehat{r}\mathcal{Z}^{\phi_t}(x_1,x_2) \Psi_t\rangle\hspace{-0.08cm}\rangle\right)
\\\nonumber&+N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi_t ,\left[H_{V_N},
g_{\beta}(x_{1}-x_{2})
\right]\widehat{r}
\Psi_t\rangle\hspace{-0.08cm}\rangle\right)\;.
\end{align}
The first, second and the fourth line give $\gamma_a+\gamma_b+\gamma_d+\gamma_e+ \gamma_f$. Using (\ref{microscopic}) the commutator in the fifth line equals
\begin{align*}[H_{V_N},g_{\beta}(x_1-x_2)]=&-[H_{V_N},f_{\beta}(x_1-x_2)]\\\nonumber=&[\Delta_1+\Delta_2,f_{\beta}(x_1-x_2)]
\\\nonumber=&(\Delta_1+\Delta_2)f_{\beta}(x_1-x_2)\\\nonumber&+(2\nabla_1f_{\beta}(x_1-x_2))\nabla_1+(2\nabla_2f_{\beta}(x_1-x_2))\nabla_2
\\\nonumber=&(V_{N}(x_1-x_2)-M_{\beta}(x_1-x_2))f_{\beta}(x_1-x_2)\\\nonumber&-(2\nabla_1g_{\beta}(x_1-x_2))\nabla_1-(2\nabla_2g_{\beta}(x_1-x_2))\nabla_2\;.
\end{align*}
Using symmetry the third and fifth line in (\ref{commu}) give
$$-4N(N-1)\langle\hspace{-0.08cm}\langle\Psi_t
, (\nabla_1g_{\beta}(x_1-x_2))\nabla_1
\widehat{r}\Psi_t\rangle\hspace{-0.08cm}\rangle=\gamma_c (\Psi_t,\phi_t)\;.$$
\end{proof}
\subsubsection{Establishing the Gr\"onwall estimate}
Again, we will bound the time derivative of $\alpha(\Psi_t,\phi_t)$ such that we can employ a Gr\"onwall estimate.
\begin{lemma} \label{gammalemma fuer V}
Let $\Psi_t$ the unique solution to $i \partial_t \Psi_t
= H_{V_N} \Psi_t$ with initial datum
$\Psi_0 \in L^2_{s}(\mathbb{R}^{2N}, \mathbb{C}) \cap H^2(\mathbb{R}^{2N}, \mathbb{C}) ,\; \|\Psi_0\|=1$.
Let $\phi_t$ the unique solution to $i \partial_t \phi_t
= h^{GP}_{4 \pi} \phi_t$ with initial datum $\phi_0 \in H^{3}(\mathbb{R}^2,\mathbb{C})$.
Let $\mathcal{E}_{V_N}(\Psi_0) \leq C$.
Let $\gamma(\Psi_t,\phi_t)$ be defined as in \eqref{gamma fuer V}.
Then, there exists an $\eta>0$ such that
\begin{align}
\label{theoremeq1 fuer V}
\gamma(\Psi_t,\phi_t) \leq &
\mathcal{K}(\phi_t, A_t)
\left(
\langle\hspace{-0.08cm}\langle\Psi_t,\widehat{n}\Psi_t\rangle\hspace{-0.08cm}\rangle + N^{-\eta}+
\left|
\mathcal{E}_{V_N(\Psi_0)}
-
\mathcal{E}_{4 \pi}^ {GP}(\phi_0)
\right|
\right)
\end{align}
\end{lemma}
A prove of the Lemma can be found in Section \ref{gammacontrolsection}.
The most important estimate is $\gamma_b$, which can be estimated in the same way as $\gamma_b^<$.
All other estimates are based on the smallness of the $L^p$-norms of $g_{\beta}$, see Lemma \ref{defAlemma} (i).
We now show that Lemma \ref{gammalemma fuer V} implies convergence of the reduced density matrix
$\gamma^{(1)}_{\Psi_t}$ to $|\phi_t \rangle \langle \phi_t|$ in trace norm.
Using $\| \widehat{m}^a\|_{\text{op}}
+
\| \widehat{m}^b\|_{\text{op}} \leq C N^{-1+\xi}$, see \eqref{estab},together with Equation \eqref{kombeqb} and Lemma \ref{defAlemma} (i), we obtain
\begin{align*}
& \|g_{\beta}(x_{1}-x_{2}) \widehat{r}\|_{\text{op}}
\leq
\|g_{\beta}(x_{1}-x_{2})p_1(
\widehat{m}^b p_2+
\widehat{m}^a q_2)
\|_{\text{op}}
+
\|g_{\beta}(x_{1}-x_{2}) p_2
q_1
\widehat{m}^a\|_{\text{op}}
\\
\leq&
\mathcal{K}(\phi, A_t)
\|g_{\beta} \|
(
\| \widehat{m}^a\|_{\text{op}}
+
\| \widehat{m}^b\|_{\text{op}}
)
\leq
\mathcal{K}(\phi, A_t)
N^{\xi-2- \beta}
\ln(N) .
\end{align*}
Therefore, we bound
\begin{equation}
\label{zweimal}
N(N-1)\Re\left(\left\langle\hspace{-0.14cm}\left\langle\Psi
, g_{\beta}(x_{1}-x_{2}) \widehat{r}
\Psi\right\rangle\hspace{-0.14cm}\right\rangle\right)
\leq \mathcal{K}(\phi, A_t)
N^{-\beta+\xi} \ln(N) .
\end{equation}
For $\beta$ large enough, (\ref{theoremeq1 fuer V}) implies together with (\ref{zweimal}) that
$$\gamma(\Psi_t,\phi_t)\leq \mathcal{K}(\phi_t, A_t)
\left(\alpha(\Psi_t,\phi_t)
+ N^{-\eta}\right)\;,
$$
for some $\eta>0$.
We get with Lemma \ref{ableitung} and Gr\"onwall's Lemma,
using (\ref{zweimal}) again,
that
\begin{align*}
&\alpha^<(\Psi_t,\phi_t)\leq
e^{
\int_0^t ds \mathcal{K}(\phi_s, A_s)}
\Big(
\alpha^<(\Psi_0,\phi_0)
\\
&+
\int_0^t ds
\mathcal{K}(\phi_s, A_s)
e^{-\int_0^s d \tau \mathcal{K}(\phi_\tau, A_\tau)}
N^{-\eta}
\Big) \;.
\end{align*}
Therefore, we obtain part (b) of Theorem \ref{theo} under the assumption that
$ \phi_t \in H^3 (\mathbb{R}^2, \mathbb{C})$.
\section{Rigorous estimates}
\label{rigestimates}
\subsection{Smearing out the potential $W_\beta$}
Recall that in Section \ref{secmic} we have defined the potential $M_\beta$ in order to control the strongly peaked potential $V_N$.
In order to control $W_\beta$ for large $\beta$, we will employ a similar strategy. We therefore ''smear out'' the potential $W_\beta$, using a technique which is known from standard electrodynamics. For this, we define, for $\beta_1<\beta$, a potential $U_{\beta_1,\beta} \in \mathcal{V}_{\beta_1}$ such that
$\|W_\beta\|_1=\|U_{\beta_1,\beta} \|_1$. Furthermore, define $h_{\beta_1,\beta}$ by
$ \Delta h_{\beta_1,\beta}=W_{\beta}-U_{\beta_1,\beta}$.
The function $h_{\beta_1,\beta}$ can be thought as an electrostatic potential which is caused by the charge $ W_{\beta}-U_{\beta_1,\beta}$.
This construction can then be used to bounds terms like
\begin{align*}
&\langle\hspace{-0.08cm}\langle \chi, W_\beta(x_1-x_2) \Omega \rangle\hspace{-0.08cm}\rangle =
\langle\hspace{-0.08cm}\langle \chi, U_{\beta_1,\beta}(x_1-x_2) \Omega \rangle\hspace{-0.08cm}\rangle
\\
-&
\langle\hspace{-0.08cm}\langle \nabla_1 \chi, (\nabla_1 h_{\beta_1,\beta}) (x_1-x_2) \Omega \rangle\hspace{-0.08cm}\rangle
-
\langle\hspace{-0.08cm}\langle \chi, (\nabla_1 h_{\beta_1,\beta}) (x_1-x_2) \nabla_1 \Omega \rangle\hspace{-0.08cm}\rangle
\; ,
\end{align*}
for some $\chi, \omega \in L^2_s(\mathbb{R}^{2N},\mathbb{C})$.
It is then easy to verify that $h_{\beta_1,\beta}$ and $ \nabla h_{\beta_1,\beta}$ are faster decaying than the potential $W_\beta$.
The right hand side of the equation above is hence better to control, if one has additional control of $\nabla_1 \Omega$ and $\nabla_1 \chi$.
\begin{definition}\label{udef
For any $0\leq\beta_1\leq \beta$ and any $W_{\beta}\in\mathcal{V}_\beta$ we define
$$U_{\beta_1,\beta}(x)=\left\{
\begin{array}{ll}
\frac{1}{\pi}\|W_{\beta}\|_1N^{2\beta_1} & \hbox{for \ $|x|<N^{-\beta_1}$,} \\
0 & \hbox{else.}
\end{array}
\right.
$$
and
\begin{align} \label{defh} h_{\beta_1,\beta}(x)= -\frac{1}{2\pi} \int_{\mathbb R^2} \ln|x-y|
(W_{\beta}(y)-U_{\beta_1,\beta}(y))d^2y
\; .
\end{align}
\end{definition}
\begin{lemma}\label{ulemma}
For any $0\leq \beta_1\leq \beta$ and any $W_{\beta}\in\mathcal{V}_\beta$, we obtain with the above definition
\begin{enumerate}
\item
\begin{align*}
& U_{\beta_1,\beta} \in \mathcal{V}_{\beta_1}\;,
\\
&\Delta h_{\beta_1,\beta}=W_{\beta}-U_{\beta_1,\beta}
.
\end{align*}
\item Pointwise estimates
\begin{align}\label{extraz}
|h_{\beta_1,\beta}(x)|\leq& CN^{-1} \ln (N) , \hspace{1cm}
h_{\beta_1,\beta}(x) =0 \text{ for } |x| \geq N^{- \beta_1} ,
\\|\nabla h_{\beta_1,\beta}(x)|\leq& CN^{-1}\left(\vert x\vert ^2+N^{-2\beta}\right)^{-\frac{1}{2}}.\label{extrazb}
\end{align}
\item Norm estimates
\begin{align*}
\|h_{\beta_1,\beta}\|_\infty & \leq CN^{-1} \ln (N),\\
\hspace{1cm}\|h_{\beta_1,\beta}\|_{\lambda}& \leq CN^{-1-\frac{2}{\lambda}\beta_1} \ln (N)\;\text{ for }1\le \lambda \leq\infty ,
\\\|\nabla h_{\beta_1,\beta}\|_{\lambda}&\leq CN^{-1+\beta-\frac{2}{\lambda}\beta_1} \hspace{0.63cm} \text{ for }1\le \lambda \leq\infty.
\end{align*}
Furthermore, for $\lambda=2$, we obtain the improved bounds
\begin{align}
\|h_{0,\beta}\| \leq &C N^{-1} \text{ for } \beta>0
\; ,
\\
\|\nabla h_{\beta_1,\beta}\|\leq& CN^{-1} (\ln (N))^{1/2}
\;.
\end{align}
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}
\item
$U_{\beta_1,\beta} \in \mathcal{V}_{\beta_1} $ follows directly from the definition of $U_{\beta_1,\beta}$.
\\
The second statement is a well known result of standard electrostatics (therefore recall that the radially symmetric Greens function of the Laplace operator in two dimensions is given by $-\frac{1}{2\pi} \ln \vert x-y \vert$). $W_\beta$ can be understood as a
given charge density. $-U_{\beta_1,\beta}$ then corresponds to a smeared out charge density of opposite sign such that the ``total charge'' is zero. Hence, the
``potential'' $h_{\beta,\beta_1}$ can be chosen to be zero outside the support of the total charge density.\footnote{To see this, recall that the solution of $\Delta h(r) = \rho(r)$ for radially symmetric and regular enough charge density $\rho$ is given by
\begin{align*}
h(r) = \ln (r) \int_0^r r'\rho(r') dr' + \int_r^{\infty} \ln (r') \rho(r') r' dr' + C,
\end{align*}
where $C\in \mathbb R$. The r.h.s. is zero for $r \not\in \text{supp}(\rho)$ when the total charge vanishes $\int_0^{\infty} r \rho(r) dr=0$ and $C$ is chosen equal to zero.}
\item $\vert h_{\beta_1,\beta}(x) \vert=0$ as well as $W_{\beta}(x)-U_{\beta_1,\beta}(x) = 0$ for $\vert x \vert \ge N^{-\beta_1}$ implies that, whenever $\vert h_{\beta_1,\beta}(x) \vert$ is nonzero, $\vert x - y \vert \le 1$ and therefore $-\ln |x-y|\geq 0$ in \eqref{defh}.
Let $2R N^{-\beta} \le \vert x \vert$. Since $-\ln |x-y| > 0$ in the support of $U_{\beta}$ together with the support properties of $W_{\beta}$, one finds that for $ R N^{-\beta} < \vert x \vert$
\begin{align*}
\vert h_{\beta_1,\beta}(x)| & \leq \frac{1}{2\pi} \int -\ln |x-y|
W_{\beta}(y)d^2y \le C \|W_\beta \|_1 \vert \ln (\vert x \vert - RN^{-\beta} )\vert,
\end{align*}
which in turn implies
\begin{align*}
\vert h_{\beta_1,\beta}(x)| \le C \|W_\beta \|_1 \ln N^{\beta}\le C N^{-1} \ln N^{\beta}
\end{align*}
for all $2R N^{-\beta} \le \vert x \vert$.\\
\\
Let next $|x|\leq 2RN^{-\beta}$. Here, one finds
\begin{align*}
\vert h_{\beta_1,\beta}(x)| & \leq \frac{1}{2\pi} \int -\ln |x-y|
W_{\beta}(y)d^2y \nonumber\\
& \le C \| W_{\beta} \|_{\infty} \int_{ B_{RN^{-\beta}(0)}} - \ln ( \vert x-y \vert ) d^2y \nonumber\\
& \le C N^{-1 + 2 \beta} \int_{ B_{RN^{-\beta}(x)}} - \ln \vert y\vert d^2y \nonumber\\
& \le C N^{-1 + 2 \beta} \int_{B_{4RN^{-\beta}(0)}} - \ln \vert y\vert d^2y \nonumber\\
& = C N^{-1 + 2 \beta}\Big[- \vert y\vert^2 (2\ln \vert y \vert -1) \Big]_0^{4RN^{-\beta}} \le C N^{-1} \ln N^{\beta},\nonumber
\end{align*}
which proves the first statement.\\
For the gradient, we estimate the two terms on the r.h.s. of
\begin{align*}|\nabla h_{\beta_1,\beta}(x)|\leq \frac{1}{2\pi }\int \frac{1}{|x-y|} W_{\beta}(y)d^2y + \frac{1}{2\pi }\int \frac{1}{|x-y|} U_{\beta,\beta_1}(y)d^2y
\end{align*}
separately. Let first $2R N^{-\beta} \le \vert x \vert$. Similarly as in the previous argument, one finds
\begin{align*}
\int \frac{1}{|x-y|} W_{\beta}(y)d^2y & \leq \int_{B_{R N^{-\beta}}(0)} \frac{1}{\vert x-y\vert}
W_{\beta}(y)d^2y \le \frac{\|W_\beta \|_1}{\vert x \vert - RN^{-\beta}}
\end{align*}
for $R N^{-\beta} \le \vert x \vert$, which implies that
\begin{align*}
\int \frac{1}{|x-y|} W_{\beta}(y)d^2y \le \frac{C \| W_{\beta}\|_1}{( \vert x \vert^2 + N^{-2\beta})^{\frac{1}{2}}} \le \frac{C N^{-1}}{( \vert x \vert^2 + N^{-2\beta})^{\frac{1}{2}}}
\end{align*}
for all $2R N^{-\beta} \le \vert x \vert$.
For $\vert x \vert \le 2R N^{-\beta}$, we make use of
\begin{align*}
N^{\beta} \le \frac{C}{\left( \vert x \vert^2 + N^{- 2 \beta} \right)^{1/2}}
\end{align*}
and estimate
\begin{align*}
\int \frac{1}{|x-y|} W_{\beta}(y) d^2y & \le \| W_{\beta} \|_{\infty} \int_{ B_{RN^{-\beta}(0)}} \frac{1}{\vert x-y \vert } d^2y \\
& \le C N^{2\beta-1}\int_{0}^{RN^{-\beta }} d\vert y\vert = C N^{-1+\beta}
\le \frac{C N^{-1}}{\left( \vert x \vert^2 + N^{- 2 \beta} \right)^{1/2}} .
\end{align*}
Equivalently, we obtain
\begin{align*}
\int \frac{1}{|x-y|} U_{\beta_1,\beta}(y) d^2y & \le \| U_{\beta_1, \beta} \|_{\infty} \int_{ B_{N^{-\beta_1}(0)}} \frac{1}{\vert x-y \vert } d^2y \\
& = C N^{-1+\beta_1}
\le \frac{C N^{-1}}{\left( \vert x \vert^2 + N^{- 2 \beta_1} \right)^{1/2}}
\le \frac{C N^{-1}}{( \vert x \vert^2 + N^{-2\beta})^{\frac{1}{2}}} ,
\end{align*}
for $\vert x \vert \leq N^{-\beta_1}$.
Since $\nabla h_{\beta_1,\beta}(x) = 0 $ for $\vert x \vert \geq N^{-\beta_1}$, the second statement of (b) follows.
\item The first part of (c) follows from (b) and the fact that the support of $h_{\beta1,\beta}$ and $\nabla h_{\beta_1,\beta}$ has radius $\leq CN^{-\beta_1}$.
The bounds on the $L^2$-norm can be improved by
\begin{align*}
\|\nabla h_{\beta_1,\beta}\|_{2}^2\leq&
C \int_0^{C N^{-\beta_1}} dr r |\nabla h_{\beta_1,\beta}(r)|^2
=
\frac{C}{N^2} \int_0^{C N^{-\beta_1}} dr \frac{r}{r^2+N^{-2\beta } }
\\
=&
\frac{C}{N^2} \ln \left(\frac{N^{-2\beta_1 }+N^{-2\beta }}{N^{-2\beta }} \right)
\leq \frac{C}{N^ 2} \ln(N)
\end{align*}
Using, for $|x|\geq 2 RN^{-\beta}$, the inequality
\begin{align*}
\vert h_{0,\beta}(x)| &\leq C N^{-1} \vert \ln (\vert x \vert - RN^{-\beta} )\vert,
\end{align*}
we obtain
\begin{align*}
\|h_{0,\beta}\|_2^2
=&
\int d^2x
\mathds{1}_{B_{2RN^{-\beta}(0)}}(x)
|h_{0,\beta}(x)|^2
+
\int d^2x
\mathds{1}_{B^c_{2RN^{-\beta}(0)}}(x)
|h_{0,\beta}(x)|^2
\\
\leq &
\|h_{0\beta}\|_\infty^2
|B_{2RN^{-\beta}(0)}|
+
C N^{-2}
\int_{2 RN^{-\beta}}^1 dr
r
\vert \ln (r - RN^{-\beta} )\vert^2
\\
\leq &
C
\left(
N^{-2-2\beta} (\ln(N))^2
+
N^{-2}
\int_{ RN^{-\beta}}^1 dr
(r+ RN^{-\beta} )
( \ln (r))^2
\right)
\;.
\end{align*}
Using
\begin{align*}
&\int_{ RN^{-\beta}}^1 dr
(r+ RN^{-\beta} )
( \ln (r))^2
\\
=&
\left(
\frac{1}{4}r^2
(2 (\ln(r))^2-2\ln(r)+1)
+
R N^{-\beta}
r( (\ln(r))^2-2 \ln(r)+2)
\right)\Big|^1_{RN^{-\beta}}
\\
\leq&
C \left(1+ N^{-\beta}+ N^{-2 \beta} (\ln(N))^2 \right) \;,
\end{align*}
we obtain, for any $\beta>0$,
\begin{align*}
\|h_{0,\beta}\|_2^2
\leq
C N^{-2}
\left(
1+
N^{-\beta}
+
N^{-2\beta}
(\ln(N))^2
\right)
\leq
C N^{-2} \;.
\end{align*}
\end{enumerate}
\end{proof}
\subsection{Estimates on the cutoff}
In order to smear out the singular potentials as explained in the previous section, it seems at first necessary to control $\|\nabla_1 q_1\Psi_t\|$. However,
this term will in fact {\it not} be small for the dynamic generated by $V_N$. There, we rather expect that $\|\nabla_1q_1\Psi_t\| = \mathcal{O}(1)$ holds.
The reason for this is that a relevant part of the kinetic energy is concentrated around the scattering centers.
We must thus cutoff the part of the kinetic energy which is used to form the microscopic structure.
To control this structure, we first cutoff three particle interactions, i.e. we define a cutoff function which
does not depend on $x_1$ and cuts off all parts of the wave function
where two particles $x_j, x_k$ with $j\neq k$, $j,k\neq 1$ come very
close (see $\mathcal{B}^{(d)}_1$ in Definition \ref{hdetail}).
After that we estimate that part of the kinetic energy which
is used to form the microscopic structure. The latter is
concentrated around the scattering centers (i.e. on the sets
$\overline{\mathcal{A}}_j^{(d)}$ given by Definition \ref{hdetail}).
Then we show that the kinetic energy concentrated on the complement of $\overline{\mathcal{A}}_j^{(d)}$, i.e. $\|\mathds{1}_{\mathcal{A}^{(d)}_1}\nabla_1q_1\Psi\|$, is small (see Lemma \ref{energylemma}).
\begin{definition}
\label{hdetail}
For any $j,k=1, \dots,N$ and $ d>0$
let
\begin{equation}
\label{defkleins}a^{(d)}_{j,k}=\{(x_1,x_2,\ldots,x_N)\in
\mathbb{R}^{2N}: |x_j-x_k|<N^{-d}\}
\subseteq \mathbb{R}^{2N}
\end{equation}
$$\overline{\mathcal{A}}^{(d)}_j=\bigcup_{k\neq j}a^{(d)}_{j,k}\;\;\;\;\;\;\;\mathcal{A}^{(d)}_j=\mathbb{R}^{2N}\backslash \overline{\mathcal{A}}_j^{(d)}
\;\;\;\;\;\;\;\overline{\mathcal{B}}^{(d)}_{j}=\bigcup_{k\neq l\neq j}a^{(d)}_{k,l}\;\;\;\;\;\;\;\mathcal{B}^{(d)}_{j}=\mathbb{R}^{2N}\backslash
\overline{\mathcal{B}}^{(d)}_{j}\;.$$
\end{definition}
\begin{lemma}
\label{propo}
Let $\Psi \in L^2_{s}(\mathbb{R}^{2N}, \mathbb{C}) \cap H^1(\mathbb{R}^{2N}, \mathbb{C})$ $\| \Psi\|=1$ and let $\| \nabla _1\Psi \|$ be uniformly bounded in $N$.
Then, for all $j\neq k$ with $1 \leq j,k \leq N$,
\begin{enumerate}
\item
\begin{align*}
\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{j}}p_j\|_{\text{op}}&\leq C \| \phi\|_{\infty}N^{1/2-d}\;,
\\
\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{j}} \nabla_j p_j\|_{\text{op}}&\leq C\| \nabla\phi\|_{\infty} N^{1/2-d}\;,
\\
\|\mathds{1}_{a^{(d)}_{j,k}}p_j\|_{\text{op}}&\leq C \|\phi\|_\infty N^{-d}
\;.
\end{align*}
\item
For any $1<p< \infty$
\begin{align*}\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{j}}\Psi\|
\leq &
C
N^{\frac{1-2d}{2} \frac{p-1}{p}}
\;,
\end{align*}
which implies that
$$
\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{j}}\Psi\| \leq
C
N^{\frac{1}{2}-d+ \epsilon}
$$
for any $\epsilon>0$.
\item
\begin{align*}
\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{j}}\Psi\|\leq &
C
N^{1-d+ \epsilon}
\end{align*}
for any $\epsilon>0$.
\item For any $k\neq j$
$$\|[\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{j}},p_k]\|_{\text{op}}=\|[\mathds{1}_{a^{(d)}_{j,k}},p_k]\|_{\text{op}}=\|[\mathds{1}_{\mathcal{A}^{(d)}_{j}},p_k]\|_{\text{op}}\leq C \|\phi\|_\infty N^{-d}
\;.
$$
\end{enumerate}\end{lemma}
\begin{proof}
\begin{enumerate}
\item First note that the volume of the sets $a^{(d)}_{j,k}$ introduced in Definition \ref{hdetail} are
$|a^{(d)}_{j,k}|=\pi N^{-2d}$.
\begin{align*}
\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{j}}p_j\|_{\text{op}}=&
\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}p_1\|_{\text{op}}=
\|p_1\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}p_1\|_{\text{op}}^{\frac{1}{2}}
\leq \left(\|\phi\|_\infty^2\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\|_{1,\infty}\right)^{1/2}
\end{align*}
where we defined
\begin{align*}
\|f \|_{p, \infty} =
\sup_{x_2, \dots, x_N \in \mathbb{R}^2}
\left(
\int dx_1 |f(x_1, \dots, x_N)|^{p}
\right)^{\frac{1}{p}}
\;.
\end{align*}
Using $\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}} \leq \sum_{k=2}^N \mathds{1}_{a_{1,k}^{(d)}} $
as well as $\left(\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}} \right)^p =\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}$,
we obtain
\begin{align*}
\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}} \|_{p, \infty}
\leq&
\sup_{x_2, \dots, x_N \in \mathbb{R}^2}
\left(
\int dx_1 \sum_{k=2}^N \mathds{1}_{a_{1,k}^{(d)}}
\right)^{\frac{1}{p}}
\leq
\left(
N |a_{1,k}|
\right)^{\frac{1}{p}}
\leq
C
N^{(1-2d) \frac{1}{p}}
\;.
\end{align*}
This implies
\begin{align*}
\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{j}}p_j\|_{\text{op}}
\leq&
C \|\phi\|_\infty N^{\frac{1}{2}-d}
\;.
\end{align*}
The second statement of a) can be proven similarly. Analogously, we obtain
\begin{align*}
\|\mathds{1}_{a^{(d)}_{j,k}}p_j\|_{\text{op}}\leq&\|\phi\|_\infty |a^{(d)}_{j,k}|^{1/2} \leq C \|\phi\|_\infty N^{-d}
\end{align*}
\item
Without loss of generality, we can set $j=1$.
Recall the two-dimensional Sobolev inequality, for $\varrho \in H^1(\mathbb{R}^2,\mathbb{C})$,
$
\| \varrho \|_m \leq
C \| \nabla \varrho\|^{\frac{m-2}{m}} \| \varrho \|^{\frac{2}{m}}
$ holds for any $2<m<\infty$.
Using H\"older and Sobolev for the $x_1$-integration, we get, for $p>1$
\begin{align*} \|
&\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\Psi\|^2
=\langle\hspace{-0.08cm}\langle\Psi,\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle
=
\int d^2 x_2\dots d^2x_N \int d^2x_1 |\Psi(x_1, \dots,x_N)|^2\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}(x_1,\dots,x_N)
\\
\leq &
\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}} \|_{\frac{p}{p-1}, \infty}
\int d^2 x_2\dots d^2x_N
\left(
\int d^2x_1 |\Psi(x_1, \dots,x_N)|^{2p}
\right)^{1/p}
\\
\leq &
C N^{(1-2d)\frac{p-1}{p}}
\int d^2 x_2\dots d^2x_N
\left(
\int d^2x_1 |\nabla_1 \Psi(x_1, \dots,x_N)|^{2}
\right)^{\frac{p-1}{p}}
\left(
\int d^2 \tilde{x}_1 | \Psi(\tilde{x}_1, \dots,x_N)|^{2}
\right)^{\frac{1}{p}}
\end{align*}
Using H\"older for the $x_2,\dots x_N$-integration with the conjugate pair $r= \frac{p}{p-1} $ and $s =p$, we obtain
\begin{align*}
\|
\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\Psi\|^2
\leq &
C N^{(1-2d)\frac{p-1}{p}}
\| \nabla_1 \Psi \|^{2\frac{p-1}{p}}
\| \Psi \|^{\frac{2}{p}} \;.
\end{align*}
Using $\|\nabla_1\Psi \|<C$, (b) follows.
\item We use that $\overline{\mathcal{B}}^{(d)}_{j}\subset\bigcup_{k=1}\overline{\mathcal{A}}^{(d)}_{k}$.
Hence one can find pairwise disjoint sets $ \mathcal{C}_k\subset\overline{\mathcal{A}}^{(d)}_{k}$, $k=1,\ldots,N$
such that $\overline{\mathcal{B}}^{(d)}_{j}\subset\bigcup_{k=1} \mathcal{C}_{k}$. Since the sets $ \mathcal{C}_k$ are pairwise disjoint,
the $\mathds{1}_{ \mathcal{C}_{k}}\Psi $ are pairwise orthogonal and we get
\begin{align*}
\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{j}}\Psi \|^2=\sum_{k=1}\|\mathds{1}_{ \mathcal{C}_{k}}\Psi \|^2
\leq \sum_{k=1}^N\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{k}}\Psi \|^2 \;.
\end{align*}
\item
\begin{align*}
\|[\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}},p_2]\|_{\text{op}} \leq &\|[\mathds{1}_{a_{1,2}},p_2]\|_{\text{op}}
\leq\|\mathds{1}_{a_{1,2}}p_2\|_{\text{op}}+\|p_2\mathds{1}_{a_{1,2}}\|_{\text{op}}
\\\leq &2\|\phi\|_\infty|a_{1,2}|^{\frac{1}{2}}\leq C \|\phi\|_\infty N^{-d}\;.
\end{align*}
\end{enumerate}
\end{proof}
\subsection{Estimates for the functionals $\gamma_a$, $\gamma_a^<$ and $\gamma_b^<$}\label{contgamma1}
\paragraph{Control of $\asa$ and $\gamma^<_a$}
\begin{lemma}\label{opdiff}
For any multiplication operator $B: L^2(\mathbb{R}^2,\mathbb{C}) \rightarrow L^2(\mathbb{R}^2,\mathbb{C})$ and any $\phi\in L^2(\mathbb{R}^ 2, \mathbb{C}) $ and any $\Psi\in L_s^2(\mathbb{R}^{2N},\mathbb{C})$ we have
$$|\langle\hspace{-0.08cm}\langle\Psi , B(x_1) \Psi\rangle\hspace{-0.08cm}\rangle-\langle\phi,B\phi\rangle|\leq C \|B\|_\infty( \langle\hspace{-0.08cm}\langle\Psi , \widehat{n}^\phi\Psi\rangle\hspace{-0.08cm}\rangle + N^{-\frac{1}{2}} ) \;.$$
\end{lemma}
\begin{proof}
Using $1=p_1+q_1$,
\begin{align*}
&\langle\hspace{-0.08cm}\langle\Psi, B(x_1) \Psi\rangle\hspace{-0.08cm}\rangle-\langle\phi,B\phi\rangle
\\
=&
\langle\hspace{-0.08cm}\langle\Psi, p_1B (x_1) p_1\Psi\rangle\hspace{-0.08cm}\rangle+2\Re\langle\hspace{-0.08cm}\langle\Psi , q_1B(x_1) p_1\Psi\rangle\hspace{-0.08cm}\rangle+\langle\hspace{-0.08cm}\langle\Psi , q_1B (x_1) q_1\Psi\rangle\hspace{-0.08cm}\rangle-\langle\phi , B\phi\rangle
%
\\
\leq&
\langle\phi , B\phi\rangle(\|p_1\Psi\|^2-1)+2\Re\langle\hspace{-0.08cm}\langle\Psi, \widehat{n}^{-1/2}q_1B(x_1) p_1\widehat{n}_1^{1/2}\Psi\rangle\hspace{-0.08cm}\rangle \nonumber\\
+ &\langle\hspace{-0.08cm}\langle\Psi, q_1B(x_1) q_1\Psi\rangle\hspace{-0.08cm}\rangle
\end{align*}
where we used Lemma \ref{kombinatorik} (c). Since $\|p_1\Psi\|^2-1 =\|q_1\Psi\|^2$
it follows that
\begin{align}\label{neededtocomplete}
|\langle\hspace{-0.08cm}\langle\Psi, B(x_1) \Psi\rangle\hspace{-0.08cm}\rangle-\langle\phi,B\phi\rangle|&\leq C\| B\|_\infty \left(\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}^2\Psi\rangle\hspace{-0.08cm}\rangle+
\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}_1\Psi\rangle\hspace{-0.08cm}\rangle + \langle\hspace{-0.08cm}\langle\Psi,\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle \right)\nonumber\\
& \le C \|B\|_\infty( \langle\hspace{-0.08cm}\langle\Psi,\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle + N^{-\frac{1}{2}}) \;.
\end{align}
\end{proof}
Using Lemma \ref{opdiff}, setting $B=\dot A$, we get
$$\gamma_a^<(\Psi_t,\phi_t) =\gamma_a(\Psi_t,\phi_t) \leq C \|\dot A_t\|_\infty ( \langle\hspace{-0.08cm}\langle\Psi_t,\widehat{n}^{\phi_t}\Psi_t\rangle\hspace{-0.08cm}\rangle + N^{-\frac{1}{2}} ) \;,$$
which yields the first bound \eqref{abc} in Lemma \ref{gammalemma}.
\paragraph{Control of $\gamma^<_b$}
To control $\gamma_b^<$ we will first prove that $\| \nabla_1 \Psi_t \| $ is uniformly bounded in $N$, if initially the energy per particle $\mathcal{E}_{U}(\Psi_0)$ is of order one.
\begin{lemma} \label{kinenergyboundedlemma}
Let $\Psi_0 \in L^2_{s}(\mathbb{R}^{2N}, \mathbb{C}) \cap H^2(\mathbb{R}^{2N}, \mathbb{C})$ with $\|\Psi_0\|=1$.
For any $U \in L^2(\mathbb{R}^2,\mathbb{R}) ,\; U(x) \geq 0$,
let $\Psi_t$ the unique solution to $i \partial_t \Psi_t
= H_{U} \Psi_t$ with initial datum $\Psi_0$.
Let $\mathcal{E}_{U}(\Psi_0) \leq C$.
Then
\begin{align*}
\| \nabla_1 \Psi_t \| \leq \mathcal{K}(\phi_t, A_t)
\end{align*}
\end{lemma}
\begin{proof}
Using
$
\frac{d}{dt} \mathcal{E}_{U}(\Psi_t) \leq \| \dot A_t\|_\infty
$, we obtain $\mathcal{E}_{U}(\Psi_t) \leq \mathcal{K}(\phi_t, A_t)$.
This yields
\begin{align*}
\| \nabla_1 \Psi_t \|^2
\leq
\mathcal{K}(\phi_t, A_t)- (N-1) \| \sqrt{U}\Psi_t \|^2
+
\| A_t \|_\infty
\leq \mathcal{K}(\phi_t, A_t)
\end{align*}
\end{proof}
Next, we control $\widehat{m}^a$ and $\widehat{m}^b$ which were defined in Definition \ref{defm}.
The difference $m(k)-m(k+1)$ and
$m(k)-m(k+2)$ is of leading order given by the derivative of the function $m(k)$ -- $k$ understood as real variable --
with respect to $k$.
The $k$-derivative of $m(k)$ equals
\begin{equation}\label{mprime}m(k)^\prime=\left\{
\begin{array}{ll}
1/(2\sqrt{kN}), & \hbox{for $k\geq N^{1-2\xi}$;} \\
1/2(N^{-1+\xi}), & \hbox{else.}
\end{array}
\right.\end{equation}
It is then easy to show
that, for any $j\in\mathbb{Z}$, there exists a $C_j<\infty$ such that
\begin{align}\label{estab1}\widehat{m}_j^x&\leq C_jN^{-1} \widehat{n}^{-1}\text{ for }x\in\{a,b\}\\
\label{estab}\|\widehat{m}_j^x\|_{\text{op}}&\leq C_jN^{-1+\xi} \text{ for }x\in\{a,b\}
\\\label{estabn}\|\widehat{n}\widehat{m}_j^x\|_{\text{op}}&\leq C_jN^{-1} \text{ for }x\in\{a,b\}
\\\label{rop}\|\widehat{r}\|_{\text{op}}&\leq \|\widehat{m}^a\|_{\text{op}}+\|\widehat{m}^b\|_{\text{op}}
\leq CN^{-1+\xi}\;.
\end{align}
The different terms we have to estimate for $\gamma^<_b$ are found in (\ref{asaaltsplit}). In order to facilitate the notation, let $ \widehat{w} \in \lbrace N \widehat{m}^a_{-1},N \widehat{m}^b_{-2} \rbrace$. Then $w(k) < n(k)^{-1}$ and $\|\widehat{w}\|_{\text{op}}\leq C N^{\xi}$ follows.
\begin{lemma}\label{hnorms
Let $\beta >0$ and $W_\beta\in\mathcal{V}_\beta$.
Let $\Psi \in L^2_{s}(\mathbb{R}^{2N}, \mathbb{C}) \cap H^2(\mathbb{R}^{2N}, \mathbb{C}) \; , \| \Psi \|=1$
and let $\|\nabla_1\Psi\|
\leq \mathcal{K}(\phi, A_t)$.
Let $w(k)<n(k)^{-1}$ and $\| \widehat{w}\|_{\text{op}} \leq C N^{ \xi}$ for some $\xi \geq 0$. Then,
\begin{enumerate}
\item $$N\left|\langle\hspace{-0.08cm}\langle\Psi p_1p_2
Z^\phi_\beta(x_1,x_2) q_1p_2\widehat{w}\Psi\rangle\hspace{-0.08cm}\rangle\right|\leq
\mathcal{K}(\phi, A_t)
\left( N^{-1}+ N^{- 2 \beta } \ln(N) \right)
\;.
$$
\item
\begin{align*}
& N|\langle\hspace{-0.08cm}\langle\Psi, p_1p_2
W_{\beta}(x_1-x_2)
\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle|
\nonumber
\\
& \leq
\mathcal{K}(\phi, A_t)
\left(
\langle\hspace{-0.08cm}\langle \Psi, \widehat{n} \Psi \rangle\hspace{-0.08cm}\rangle
+
\inf_{\eta>0}
\inf_{\beta_1 >0}
\left(
N^{ \eta- 2 \beta_1} \ln(N)^2
+\| \widehat{w}\|_{\text{op}} N^{-1+2 \beta_1}
+
\| \widehat{w}\|_{\text{op}}^2 N^{ - \eta}
\right)
\right)
\;.
\end{align*}
\item
\begin{align*} &N|\langle\hspace{-0.08cm}\langle\Psi p_1q_2Z^\phi_\beta(x_1,x_2)\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle|
\leq
\mathcal{K}(\phi, A_t)
\Big(
\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle+ N^{-1/6} \ln(N)
\\
+&
\inf \left\lbrace
\left|\mathcal{E}_{V_N}(\Psi)-\mathcal{E}_{4 \pi}^{GP}(\phi)\right|,
\left|\mathcal{E}_{W_\beta}(\Psi)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi)\right|+ N^{-2 \beta} \ln(N)
\right\rbrace
\Big)
\;.
\end{align*}
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}
\item
In view of Lemma \ref{kombinatorikb}, we obtain
\begin{align*}
N \left|\langle\hspace{-0.08cm}\langle\Psi, p_1p_2
Z^\phi_\beta(x_1,x_2) q_1p_2\widehat{w}\Psi\rangle\hspace{-0.08cm}\rangle\right|
\leq& N\|p_1p_2
Z^\phi_\beta(x_1,x_2) q_1p_2\|_{\text{op}}\|\widehat{n}\widehat{w}\Psi\|
\\\leq& CN\|p_1p_2
Z^\phi_\beta(x_1,x_2) q_1p_2\|_{\text{op}}\;.
\end{align*}
$\|p_1p_2
Z^\phi_\beta(x_1,x_2) q_1p_2\|_{\text{op}}$ can be estimated using $p_1q_1=0$ and (\ref{faltungorigin}):
\begin{align*}
& N
\left\|
p_1p_2
\left(
W_\beta(x_1-x_2)-\frac{N\|W_\beta\|_1}{N-1}|\phi(x_1)|^2-\frac{N\|W_\beta\|_1}{N-1}|
\phi(x_2)|^2
\right) q_1p_2
\right\|_{\text{op}}
\\
&\leq\|p_1p_2
(NW_\beta(x_1-x_2)-N\|W_\beta\|_1|\phi(x_1)|^2)
p_2\|_{\text{op}}
+C \|\phi\|_\infty^2 N^{-1}
\\
&\leq \| \phi \|_{\infty} \ \|N(W_\beta\star|\phi|^2)-\|NW_\beta\|_1|\phi|^2\|+C \|\phi\|_\infty^2 N^{-1}\;.
\end{align*}
Let $h$ be given by
$$h(x)=- \frac{1}{2 \pi}\int_{\mathbb{R}^2}d^2y \ln |x-y| NW_\beta(y) + \frac{1}{2\pi}\|NW_\beta\|_1\ln |x|
\;,
$$
which implies
$$\Delta h(x)=N W_\beta(x) - \|NW_\beta\|_1 \delta(x)\;.$$
As above (see Lemma \ref{ulemma}), we obtain
$h(x)=0$ for $x \notin B_{RN^{- \beta}}(0)$, where $R N^{-\beta}$ is the radius of the support of $W_\beta$.
Thus,
\begin{align} \label{roemisch I potetnialestimate}
\|h\|_1
\leq&
\frac{1}{2 \pi}
\int_{\mathbb{R}^2}d^2x \int_{\mathbb{R}^2} d^2y|\ln |x-y|| \mathds{1}_{B_{RN^{- \beta}}(0)}(x) NW_\beta(y)
\\
-&
\frac{1}{2 \pi}
N\|W_\beta\|_1\int_{\mathbb{R}^2} d^2x \ln(|x|) \mathds{1}_{B_{RN^{- \beta}}(0)}(x)
\leq
C N^{-2 \beta} \ln(N)
\end{align}
Integration by parts and Young's inequality give that
\begin{align*}
&\| N(W_\beta\star|\phi|^2)-\|NW_\beta\|_1|\phi|^2\|=\|(\Delta h)\star|\phi|^2\|
\\\leq& \| h\|_1 \|\Delta|\phi|^2\|_2
\leq \mathcal{K}(\varphi, A_t) N^{-2\beta} \ln(N)
\;.
\end{align*}
Thus, we obtain the bound
\begin{align}
N \left|\langle\hspace{-0.08cm}\langle\Psi, p_1p_2
Z^\phi_\beta(x_1,x_2) q_1p_2\widehat{w}\Psi\rangle\hspace{-0.08cm}\rangle\right|
\leq
\mathcal{K}(\phi, A_t)
\left( N^{-1}+ N^{- 2 \beta } \ln(N) \right)
\;,
\end{align}
which then proves (a).
\item
We will first consider $\beta<1/2$.\\
Using Lemma \ref{kombinatorik} (c) and Lemma \ref{trick} with $O_{1,2}=q_2
W_\beta(x_1-x_2) p_2$, $\Omega=N^{-1/2}(\widehat{w})^{1/2}q_1\Psi$ and $\chi=N^{1/2}p_1(\widehat{w}_{2})^{1/2}\Psi$ we get
\begin{align*}
&|\langle\hspace{-0.08cm}\langle\Psi ,p_1 p_2 W_\beta(x_1-x_2) q_1 q_2
\widehat{w}\Psi\rangle\hspace{-0.08cm}\rangle|
\\=&|\langle\hspace{-0.08cm}\langle\Psi ,(\widehat{w})^{1/2}q_1 q_2
W_\beta(x_1-x_2)p_1 p_2(\widehat{w}_{2})^{1/2}\Psi\rangle\hspace{-0.08cm}\rangle|
\\&\leq N^{-1}\left\|(\widehat{w})^{1/2}q_1\Psi\right\|^2+N\big|\langle\hspace{-0.08cm}\langle q_2(\widehat{w}_{2})^{1/2}\,\Psi,p_1\sqrt{W_\beta}(x_1-x_2)
p_3
\sqrt{W_\beta}(x_1-x_3)\\&\hspace{3.7cm}\sqrt{W_\beta}(x_1-x_2)p_2\sqrt{W_\beta}(x_1-x_3)p_1
q_3(\widehat{w}_{2})^{1/2}\,\Psi\rangle\hspace{-0.08cm}\rangle\big|
\\&+N(N-1)^{-1}\|q_2
W_\beta(x_1-x_2) p_2p_1(\widehat{w}_{2})^{1/2}\Psi\|^2&
\\&\leq N^{-1}\left\|(\widehat{w})^{1/2}q_1\Psi\right\|^2+N\|\sqrt{W_\beta}(x_1-x_2)p_1\|_{\text{op}}^4\;\|
q_2(\widehat{w}_{2})^{1/2}\,\Psi\|^2
\\&+2 N(N-1)^{-1}\|p_1q_2(\widehat{w}_{1})^{1/2}
W_\beta(x_1-x_2) p_2p_1\Psi\|^2
\\&+2 N(N-1)^{-1}\|q_1q_2 (\widehat{w})^{1/2}
W_\beta(x_1-x_2) p_2p_1\Psi\|^2
\;.
\end{align*}
With Lemma \ref{kombinatorik} (e) we get the bound
\begin{align*}
\leq& N^{-1}\|(\widehat{w})^{1/2}\widehat{n}\Psi\|^2+N\|\phi \|_\infty^4\|W_\beta\|_1^2\;\|
\widehat{n}(\widehat{w}_{2})^{1/2}\,\Psi\|^2
\\&+2 N(N-1)^{-1}\|W_\beta\|^2\|\phi \|_\infty^2
\left(
\| \widehat{w}_{1}\|_{\text{op}}
+
\| \widehat{w}\|_{\text{op}}
\right)
\;.
\end{align*}
Note, that $\|W_\beta\|_1\leq CN^{-1}$, $\|W_\beta\|^2\leq CN^{-2+2\beta}$.
Furthermore, using $\widehat n<\widehat n_2$, we have under the conditions on $\widehat{w}$
\begin{equation}
\label{einfuegen}
\|(\widehat{w})^{1/2}\widehat{n}_2\Psi\|
\leq
\|(\widehat{w}_2)^{1/2}\widehat{n}_2\Psi\|\leq
\|(\widehat{n}_2)^{1/2}\Psi\|
\leq
\sqrt{\langle\Psi,\widehat{n}\Psi\rangle}+2N^{-\frac{1}{2}}
\;.
\end{equation}
In total, we obtain
\begin{align*}
N|\langle\hspace{-0.08cm}\langle\Psi ,p_1 p_2 W_\beta(x_1-x_2) q_1 q_2
\widehat{w}\Psi\rangle\hspace{-0.08cm}\rangle|\leq
\mathcal{K}(\phi, A_t)
\left(\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle+ \| \widehat{w}\|_{\text{op}} N^{-1+2 \beta}\right)
\end{align*}
and we get (b) for the case $\beta<1/2$.
\item[b)] for $1/2\leq \beta$:
We use $U_{\beta_1,\beta}$ from Definition \ref{udef} for some $0<\beta_1<1/2$.
We then obtain
\begin{align}
\nonumber
& N\langle\hspace{-0.08cm}\langle\Psi, p_1p_2
W_{\beta}(x_1-x_2)
\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle
\\
\label{eins}
=&
N\langle\hspace{-0.08cm}\langle\Psi, p_1p_2
U_{\beta_1,\beta}(x_1-x_2)
\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle
\\
\label{zwei}
+&
N\langle\hspace{-0.08cm}\langle\Psi, p_1p_2
\left(W_{\beta}(x_1-x_2)-U_{\beta_1,\beta}(x_1-x_2)\right)
\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle
\end{align}
Term \eqref{eins} has been controlled above.
So we are left to control \eqref{zwei}.
Let $\Delta h_{\beta_1,\beta}=W_{\beta}-U_{\beta_1,\beta}$. Integrating by parts and using that\\ $\nabla_1 h_{\beta_1,\beta}(x_1-x_2)=-\nabla_2 h_{\beta_1,\beta}(x_1-x_2)$ gives
\begin{align}
\nonumber
&N\left|\langle\hspace{-0.08cm}\langle\Psi, p_1p_2\left(W_{\beta}(x_1-x_2)-U_{\beta_1,\beta}(x_1-x_2)\right)\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\label{csplit1}&
\leq N\left|\langle\hspace{-0.08cm}\langle\nabla_1p_1\Psi, p_2\nabla_2 h_{\beta_1,\beta}(x_1-x_2)\widehat{w} q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\label{csplit2}
&+N\left|\langle\hspace{-0.08cm}\langle \Psi, p_1p_2\nabla_2 h_{\beta_1,\beta}(x_1-x_2) \nabla_1\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle\right|
\;.
\end{align}
Let $ t_1 \in \lbrace p_1, \nabla_1 p_1 \rbrace$ and let
$\Gamma \in \lbrace \widehat{w}q_1\Psi, \nabla_1 \widehat{w}q_1\Psi \rbrace$.
For both \eqref{csplit1} and \eqref{csplit2}, we use
Lemma \ref{trick} with
$O_{1,2}=N^{1+ \eta/2} q_2\nabla_2 h_{\beta_1,\beta}(x_1-x_2)p_2$, $\chi=t_1\Psi$ and
$\Omega= N^{- \eta/2} \Gamma$. This yields
\begin{align}
\label{gammaterm}
&\eqref{csplit1}+\eqref{csplit2}
\leq
2
\sup_{t_1 \in \lbrace p_1, \nabla_1 p_1 \rbrace, \Gamma \in \lbrace \widehat{w}q_1\Psi, \nabla_1 \widehat{w}q_1\Psi \rbrace }
\Big(
N^{-\eta} \| \Gamma \|^2
\\
\label{diagonalterm}
+&
\frac{N^{2+\eta}}{N-1}\|q_2\nabla_2 h_{\beta_1,\beta}(x_1-x_2)t_1p_2\Psi\|^2
\\
\label{nebendiagonale}
+&
N^{2+\eta}\left|\langle\hspace{-0.08cm}\langle \Psi,t_1p_2 q_3 \nabla_2 h_{\beta_1,\beta}(x_1-x_2)\nabla_3 h_{\beta_1,\beta}
(x_1-x_3)t_1q_2 p_3\Psi\rangle\hspace{-0.08cm}\rangle\right|
\Big)
\;.
\end{align}
The first term can be bounded using Corrolary \ref{kombinatorikc} by
\begin{align*}
N^{-\eta}\|\nabla_1 \widehat{w}q_1\Psi\|^2
\leq &
N^{-\eta} \| \widehat{w} \|^2_{\text{op}} \|\nabla_1 q_1 \Psi\|^2
\\
N^{-\eta}\| \widehat{w}q_1\Psi\|^2
\leq&
C N^{-\eta}
\;.
\end{align*}
Thus $\eqref{gammaterm} \leq \mathcal{K}(\phi, A_t)
N^{-\eta} \| \widehat{w} \|^2_{\text{op}} $ using that $\|\nabla_1 q_1 \Psi\| \leq \mathcal{K}(\phi, A_t)$.
By $\|t_1\Psi\|^2 \leq \mathcal{K}(\phi, A_t)$, we obtain
\begin{align*}
(\ref{diagonalterm})
\leq&
\mathcal{K}(\phi, A_t)
\frac{N^{2+\eta}}{N-1}\|\nabla_2 h_{\beta_1,\beta}(x_1-x_2)p_2\|_{\text{op}}^2
%
\leq
\mathcal{K}(\phi, A_t)
\frac{N^{2+\eta}}{N-1}
\|\phi\|_\infty ^2
\|\nabla h_{\beta_1,\beta}\|^2
\\
\leq&
\mathcal{K}(\phi, A_t)
N^{\eta-1}\ln(N)
%
\;,
\end{align*}
where we used Lemma \ref{ulemma} in the last step.
Next, we estimate
\begin{align*}
(\ref{nebendiagonale})\leq& N^{2+\eta}\| p_2
\nabla_2 h_{\beta_1,\beta}(x_1-x_2)t_1 q_2
\Psi\|^2\
\nonumber
\\
\leq&
2N^{2+\eta}\| p_2
h_{\beta_1,\beta}(x_1-x_2)t_1 \nabla_2q_2
\Psi\|^2
\nonumber
\\+&
2N^{2+\eta}\||\varphi(x_2)\rangle \langle \nabla \varphi(x_2)|
h_{\beta_1,\beta}(x_1-x_2)t_1 q_2
\Psi\|^2\
\nonumber
\\
\leq &
2N^{2+\eta}\| p_2
h_{\beta_1,\beta}(x_1-x_2)
\|_{\text{op}}^2
\|
t_1
\nabla_2 q_2
\Psi\|^2\
\nonumber
\\
+&
2N^{2+\eta}
\| |\varphi(x_2)\rangle \langle \nabla \varphi(x_2)|
h_{\beta_1,\beta}(x_1-x_2) \|_{\text{op}}^2
\| t_1 q_2
\Psi\|^2\
\nonumber
\\
\leq &
\mathcal{K}(\phi, A_t)
N^{2+\eta}
\| h_{\beta_1,\beta} \|^2
\nonumber
\\ \leq&
\mathcal{K}(\phi, A_t)
N^{\eta- 2\beta_1} \ln(N)^2
\;.
\end{align*}
Thus, for all $\eta \in \mathbb{R}$
\begin{align*}
& N\langle\hspace{-0.08cm}\langle\Psi, p_1p_2
\left(W_{\beta}(x_1-x_2)-U_{\beta_1,\beta}(x_1-x_2)\right)
\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle
\nonumber
\\
\leq &
\mathcal{K}(\phi,A_t)
\left(\| \widehat{w}\|_{\text{op}}^2 N^{ - \eta} + N^{ \eta-1} \ln(N) +N^{ \eta- 2 \beta_1} \ln(N)^2 \right).
\end{align*}
Combining both estimates for $\beta<1/2$ and $\beta \geq 1/2$, we obtain,
using $ N^{ \eta-1} \ln(N) < N^{ \eta- 2 \beta_1} \ln(N)$,
\begin{align*}
& N\langle\hspace{-0.08cm}\langle\Psi, p_1p_2
W_{\beta}(x_1-x_2)
\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle
\nonumber
\\
& \leq
\mathcal{K}(\phi,A_t)
\left(
\langle\hspace{-0.08cm}\langle \Psi, \widehat{n} \Psi \rangle\hspace{-0.08cm}\rangle
+
\inf_{\eta>0}
\inf_{\beta_1 >0}
\left(
N^{ \eta- 2 \beta_1} \ln(N)^2
+N^{-1+2 \beta_1}
+
\| \widehat{w}\|_{\text{op}}^2 N^{ - \eta}
\right)
\right)
\;.
\end{align*}
and we get (b) in full generality.
\item
We first estimate, noting that $q_1 p_2 |\phi|^2(x_1) q_1q_2=0$,
\begin{align*}
&N
\left|\langle\hspace{-0.08cm}\langle\Psi, q_1p_2
\frac{N \|W_\beta\|_1}{N-1} |\phi|^2(x_2)
\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle\right|
\leq
C
\|\phi\|_\infty^2
\|\widehat{w} q_2\|_{\text{op}}
\|q_1 \Psi \|^2
\\
&\leq
\mathcal{K}(\phi, A_t)
\langle\hspace{-0.08cm}\langle \Psi, \widehat{n} \Psi \rangle\hspace{-0.08cm}\rangle
\;.
\end{align*}
Is is left to estimate $N
\left|\langle\hspace{-0.08cm}\langle\Psi, q_1p_2W_\beta(x_1-x_2)\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle\right|$.
Let $U_{0,\beta}$ be given as in Definition \ref{udef}. Using Lemma \ref{kombinatorik} (c) and integrating by parts we get
\begin{align}\nonumber
N&\left|\langle\hspace{-0.08cm}\langle\Psi, q_1p_2V_\beta(x_1-x_2)\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle\right|\\\leq&\nonumber
N\left|\langle\hspace{-0.08cm}\langle\Psi,q_1p_2U_{0,\beta}(x_1-x_2)q_1q_2 \widehat{w}\Psi\rangle\hspace{-0.08cm}\rangle\right|+
N\left|\langle\hspace{-0.08cm}\langle\Psi,q_1p_2(\Delta_1
h_{0,\beta}(x_1-x_2))q_1q_2\widehat{w}\Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\leq&\nonumber
\|U_{0,\beta}\|_\infty
N\|q_1\Psi\|\;\|\widehat{w}q_1q_2\Psi\|
\\\nonumber%
&+N\left|\langle\hspace{-0.08cm}\langle\nabla_1q_1 p_2\Psi,(\nabla_1
h_{0,\beta}(x_1-x_2))\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle\right|
\\&+\nonumber
N\left|\langle\hspace{-0.08cm}\langle\Psi, \widehat{w}_1q_1p_2(\nabla_1
h_{0,\beta}(x_1-x_2))\nabla_1q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\leq&\label{s1}
N
\|U_{0,\beta}\|_\infty
\|q_1\Psi\|\;\|\widehat{w}q_1q_2\Psi\|
\\\label{s2}
&+N\left|\langle\hspace{-0.08cm}\langle\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi,p_2(\nabla_1
h_{0,\beta}(x_1-x_2))\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\label{s3}
&+N\left|\langle\hspace{-0.08cm}\langle\nabla_1q_1\Psi,\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}p_2(\nabla_1
h_{0,\beta}(x_1-x_2))q_1q_2\widehat{w}\Psi\rangle\hspace{-0.08cm}\rangle\right|
\\&+\label{s4}
N\left|\langle\hspace{-0.08cm}\langle\Psi, \widehat{w}_1q_1p_2(\nabla_1
h_{0,\beta}(x_1-x_2))q_2\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi\rangle\hspace{-0.08cm}\rangle\right|
\\&+\label{s5}
N\left|\langle\hspace{-0.08cm}\langle\Psi, \widehat{w}_1q_1p_2(\nabla_1
h_{0,\beta}(x_1-x_2))q_2\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1q_1\Psi\rangle\hspace{-0.08cm}\rangle\right|
\;.
\end{align}
Lemma \ref{kombinatorikb} and Lemma \ref{ulemma} (a) yields the bound
$$(\ref{s1})\leq
C
\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle
\;.
$$
For (\ref{s3}) and (\ref{s5}) we use Cauchy Schwarz and then Sobolev inequality as in Lemma \ref{propo} to get, for any $p > 1$,
\begin{align*}
&(\ref{s3})+(\ref{s5})\leq
N\left\|\nabla_1q_1\Psi\right\|\left\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}p_2(\nabla_1
h_{0,\beta}(x_1-x_2))q_1q_2\widehat{w}\Psi\right\|
\\&+N\left\|\nabla_1q_1\Psi\right\|\left\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}q_2(\nabla_1
h_{0,\beta}(x_1-x_2))q_1p_2\widehat{w}_1\Psi\right\|
\\\leq& CN\|\nabla_1q_1\Psi\|
\;
N^{\frac{1-2d}{2} \frac{p-1}{p}}
\|\nabla_1 p_2(\nabla_1
h_{0,\beta}(x_1-x_2))q_1q_2\widehat{w}\Psi\|^{\frac{p-1}{p} }
\|p_2(\nabla_1
h_{0,\beta}(x_1-x_2))q_1q_2\widehat{w}\Psi\|^{1/p}
\\+&CN\|\nabla_1q_1\Psi\|\;
N^{\frac{1-2d}{2} \frac{p-1}{p}}
\|\nabla_1 q_2(\nabla_1
h_{0,\beta}(x_1-x_2))q_1p_2\widehat{w}_1\Psi\|^{ \frac{p-1}{p} }
\| q_2(\nabla_1
h_{0,\beta}(x_1-x_2))q_1p_2\widehat{w}_1\Psi\|^{1/p}
\;.
\end{align*}
Using Lemma \ref{kombinatorik}, Lemma \ref{kombinatorikb}, Corollary \ref{kombinatorikc} and Lemma \ref{ulemma}, we obtain
\begin{align*}\|\nabla_1 p_2(\nabla_1
h_{0,\beta}(x_1-x_2))q_1q_2\widehat{w}\Psi\|\leq&
\|p_2(\Delta_1
h_{0,\beta}(x_1-x_2))q_1q_2\widehat{w}\Psi\|
\\&+
\| p_2(\nabla_1
h_{0,\beta}(x_1-x_2))\nabla_1q_1q_2\widehat{w}\Psi\|
\\&\hspace{-3cm}\leq C \left(\|p_2(W_\beta-U_{0,\beta})(x_1-x_2)\|_{\text{op}}+\|p_2\nabla_1
h_{0,\beta}(x_1-x_2))\|_{\text{op}}\right)
\\&\hspace{-3cm}\leq C
\|\phi\|_\infty
\left(
N^{-1+\beta}
+
N^{-1} (\ln(N))^{1/2}
\right)
\; ,
\end{align*}
and similarly
\begin{align*}\|\nabla_1q_2(\nabla_1
h_{0,\beta}(x_1-x_2))q_1p_2\widehat{w}_1\Psi\|\leq&
\|q_2(\Delta_1
h_{0,\beta}(x_1-x_2))q_1p_2\widehat{w}_1\Psi\|
\\&+
\| q_2(\nabla_1
h_{0,\beta}(x_1-x_2))\nabla_1q_1p_2\widehat{w}_1\Psi\|
\\&\hspace{-3cm}\leq C \left(\|p_2(W_\beta-U_{0,\beta})(x_1-x_2)\|_{\text{op}}+
\| \widehat{w}_1\|_{\text{op}}
\|p_2\nabla_1
h_{0,\beta}(x_1-x_2))\|_{\text{op}}\right)
\\&\hspace{-3cm}\leq
C
\|\phi\|_\infty
\left(
N^{-1+\beta}
+
\| \widehat{w}\|_{\text{op}}
N^{-1} (\ln(N))^{1/2}
\right)
\;.
\end{align*}
Moreover, we estimate
\begin{align*}
&\|
p_2
(\nabla_1
h_{0,\beta}(x_1-x_2))q_1q_2\widehat{w}\Psi\|
\leq
C\|\phi\|_\infty \|\nabla_1
h_{0,\beta}\|_2 \leq C\|\phi\|_\infty N^{-1} (\ln(N))^{1/2}
\\
&
\|
q_2
(\nabla_1
h_{0,\beta}(x_1-x_2))q_1p_2\widehat{w}\Psi\|
\leq
C\|\phi\|_\infty \|\nabla_1
h_{0,\beta}\|_2 \leq C \|\phi\|_\infty N^{-1} (\ln(N))^{1/2}
\;.
\end{align*}
Hence, we obtain, for any $p >1$,
\begin{align*}
(\ref{s3})+(\ref{s5})&
\leq
C \|\phi\|_\infty
N^{1+\frac{1-2d}{2} \frac{p-1}{p}}
\left(
N^{-1+\beta}
+
\| \widehat{w}\|_{\text{op}}
N^{-1} (\ln(N))^{1/2}
\right)^{\frac{p-1}{p}}
\left(
N^{-1} (\ln(N))^{1/2}
\right)^{1/p}
\;.
\end{align*}
For $d$ large enough, the right hand side can be bounded by $N^{-1}$, that is
\begin{align*}
\eqref{s3}+\eqref{s5}&
\leq C
\|\phi\|_\infty
N^{-1}
\;.
\end{align*}
For (\ref{s2}) we use that $\nabla_2
h_{0,\beta}(x_1-x_2)=-\nabla_1
h_{0,\beta}(x_1-x_2)$, Cauchy Schwarz and $ab\leq a^2+b^2$ and get
\begin{equation}\label{s2ref}(\ref{s2})\leq \|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi\|^2+N^2\|p_2(\nabla_2
h_{0,\beta}(x_1-x_2))\widehat{w}q_1q_2\Psi\|^2\;.
\end{equation}
$\|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1
\Psi \|^2 $ can be bounded using Lemma \ref{energylemma}.
Integration by parts and Lemma \ref{kombinatorik} (c) as well as $(a+b)^2\leq 2a^2+2b^2$ gives for the second summand
\begin{align}\nonumber
&N^2\|p_1(\nabla_1h_{0,\beta}(x_1-x_2))q_1q_2\widehat{w}\Psi\|^2\leq 2N^2\|p_1h_{0,\beta}(x_1-x_2)\nabla_1q_1q_2\widehat{w}\Psi\|^2
\\
\nonumber+&2N^2\||\phi(x_1)\rangle\langle\nabla_1\phi(x_1)|h_{0,\beta}(x_1-x_2)q_1q_2\widehat{w}\Psi\|^2
\\\label{linea}\leq&2N^2\|p_1h_{0,\beta}(x_1-x_2)q_2(p_1\widehat{w}_1+q_1\widehat{w})\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi\|^2
\\\label{lineb}+&2N^2\|p_1h_{0,\beta}(x_1-x_2)q_2p_1\widehat{w}_1\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1q_1\Psi\|^2
\\\label{linec}+&2N^2\|p_1h_{0,\beta}(x_1-x_2)q_2q_1\widehat{w}\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1q_1\Psi\|^2
\\\label{lined}+&2N^2\||\phi(x_1)\rangle\langle\nabla_1\phi(x_1)|h_{0,\beta}(x_1-x_2)q_1q_2\widehat{w}\Psi\|^2
\;.
\end{align}
For (\ref{linea}) we use Lemma \ref{kombinatorikb}, Lemma \ref{kombinatorik} (e) with Lemma \ref{ulemma} (c) and then Lemma \ref{energylemma}.
\begin{align*}
(\ref{linea})&\leq CN ^2\|p_1h_{0,\beta}(x_1-x_2)\|_{\text{op}}^2\|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi\|^2
%
\\& \leq
\mathcal{K}(\phi,A_t)
\Big( \left\langle\hspace{-0.14cm}\left\langle\Psi,\widehat{n}^{\phi}\Psi\right\rangle\hspace{-0.14cm}\right\rangle+N^{-1/6} \ln(N)
\\
&+
\inf \left\lbrace
\left|\mathcal{E}_{V_N}(\Psi)-\mathcal{E}_{4 \pi}^{GP}(\phi)\right|,
\left|\mathcal{E}_{W_\beta}(\Psi)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi)\right|+ N^{-2 \beta} \ln(N)
\right\rbrace
\Big)
\;.
\end{align*}
Let $s_1 \in \lbrace p_1, q_1 \rbrace$ and let $
\widehat{d} \in \lbrace \widehat{w}, \widehat{w}_1 \rbrace$.
Note that $ \| \widehat{d} \|_{\text{op}}= \| \widehat{w} \|_{\text{op}}$.
Then,
\eqref{lineb} and (\ref{linec}) can be estimated as
\begin{align*}
&
\eqref{lineb},
\eqref{linec}
\leq
2N^2
\| \nabla_1 q_1 \Psi \|^2
\|
\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}} \widehat{d} s_1 q_2 h_{0,\beta}(x_1-x_2)
p_1h_{0,\beta}(x_1-x_2)q_2s_1\widehat{d} \mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1q_1\Psi\|^2
\\
&
\leq
C
N^{2+ 2\frac{1-2d}{2}\frac{p-1}{p}}
\| \nabla_1 q_1 \Psi \|^2
\|
\nabla_1\widehat{d} s_1 q_2 h_{0,\beta}(x_1-x_2)
p_1h_{0,\beta}(x_1-x_2)q_2s_1\widehat{d}\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1q_1\Psi
\|^{2\frac{p-1}{p}}
\\
&\times
\|
\widehat{d} s_1 q_2 h_{0,\beta}(x_1-x_2)
p_1h_{0,\beta}(x_1-x_2)q_2s_1\widehat{d}\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1q_1\Psi\|
^{\frac{2}{p}}
\\
&
\leq
C N^{2+ 2\frac{1-2d}{2}\frac{p-1}{p}}
\| \nabla_1 q_1 \Psi \|^2
\| \widehat{w}\|_{\text{op}}^2
\|p_1 h_{0,\beta}(x_1-x_2)\|_{\text{op}}^2
\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1q_1\Psi \|^2
\\
&\times
\|\nabla_1\widehat{d} s_1 q_2 h_{0,\beta}(x_1-x_2)
p_1\|_{\text{op}}^{2\frac{p-1}{p}}
\|\widehat{d} s_1 q_2 h_{0,\beta}(x_1-x_2)
p_1\|_{\text{op}}^{\frac{2}{p}}
\\
& \leq
\mathcal{K}(\phi,A_t)
N^{2\frac{1-2d}{2}\frac{p-1}{p}} \| \widehat{w}\|_{\text{op}}^4
\|\nabla_1 s_1 h_{0,\beta}(x_1-x_2)
p_1\|_{\text{op}}^{2\frac{p-1}{p}}
\|h_{0,\beta}(x_1-x_2)
p_1\|_{\text{op}}^{\frac{2}{p}}
\\
&
\leq
\mathcal{K}(\phi,A_t)
N^{2\frac{1-2d}{2}\frac{p-1}{p}}
\| \widehat{w}\|_{\text{op}}^4
\left(
\|\nabla \phi\|
\| \nabla_1 h_{0,\beta}\|+ \|h_{0,\beta}\|
\right)^{2\frac{p-1}{p}}
\|h_{0,\beta}\| ^{\frac{2}{p}}
\\
& \leq \mathcal{K}(\phi,A_t) \| \widehat{w}\|_{\text{op}}^4
(\|\nabla \phi\|^2 +\ln(N))^{ \frac{p-1}{p}}
N^{2\frac{1-2d}{2}\frac{p-1}{p}-2}
\;.
\end{align*}
Here, we used, for $s_1 \in \lbrace p_1, 1- p_1 \rbrace$,
\begin{align*}
&\|\nabla_1 s_1 h_{0,\beta}(x_1-x_2)
p_1\|_{\text{op}}
\leq
\|\nabla_1 p_1 h_{0,\beta}(x_1-x_2)
p_1\|_{\text{op}}
+\|\nabla_1 h_{0,\beta}(x_1-x_2)
p_1\|_{\text{op}}
\\
\leq &
\| \varphi \|_\infty
\left(
\| \nabla \varphi \|
\| h_{0,\beta}\|
+
\| \nabla h_{0,\beta}\|
\right)
\end{align*}
and then applied Lemma \ref{kombinatorik} (e).
For $d$ large enough, we obtain
\begin{align*}
\eqref{lineb}+\eqref{linec}
\leq
\mathcal{K}(\phi, A_t)
N^{-2}
\;.
\end{align*}
Line (\ref{lined}) can be bounded by
\begin{align*}
(\ref{lined})\leq&N^2\|h_{0,\beta}(x_1-x_2)\nabla_1p_1\|_{\text{op}}^2\;\|q_1q_2\widehat{w}\Psi\|^2
\leq
N^2\|h_{0,\beta}\|^2 \| \nabla \phi \|_\infty^2 \| q_1 \widehat{w} \|_{\text{op}}^2
\|q_1 \Psi \|^2
\\
\leq& C
\| \nabla \phi \|_\infty^2
\langle\hspace{-0.08cm}\langle\Psi\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle
\;.
\end{align*}
For (\ref{s4}) we use Lemma \ref{trick} with $\Omega=\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi$,\\
$O_{1,2}=Nq_2(\nabla_2
h_{0,\beta}(x_1-x_2))p_2$ and $\chi=\widehat{w}q_1\Psi$.
\begin{align}
(\ref{s4})&\leq \|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi\|^2\label{ehklar}\\
&+
\label{ding3}
2 N\|q_2(\nabla_2
h_{0,\beta}(x_1-x_2))\widehat{w}q_1p_2\Psi\|^2
\\\label{ding4}&+N^2\big|\langle\hspace{-0.08cm}\langle \Psi,q_1q_3\widehat{w}(\nabla_2
h_{0,\beta}(x_1-x_2))
p_2p_3(\nabla_3
h_{0,\beta}(x_1-x_3))\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle\big|
\;.
\end{align}
Line \eqref{ding3} is bounded by
\begin{align*}
\eqref{ding3}
\leq &
C
\| \phi \|_\infty^2
N
\|(\nabla_2 h_{0,\beta}(x_1-x_2))p_2 \|_{\text{op}}^2
\|\widehat{w}q_1\|_{\text{op}}^2
\\
\leq &
C \| \phi \|_\infty^2 N
\|\nabla_2 h_{0,\beta}(x_1-x_2) \|^2
\leq
C \| \phi \|_\infty^2 N^{-1} \ln(N)
\;.
\end{align*}
(\ref{ehklar})+(\ref{ding4}) is bounded by $$
\|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi\|^2+N^2\| p_2 (\nabla_2
h_{0,\beta}(x_1-x_2))\widehat{w}q_1 q_2\Psi\|^2\;.$$
Both terms have been controlled above (see (\ref{s2ref})). In total, we obtain
\begin{align*} &N|\langle\hspace{-0.08cm}\langle\Psi p_1q_2Z^\phi_\beta(x_1,x_2)\widehat{w}q_1q_2\Psi\rangle\hspace{-0.08cm}\rangle|
\leq
\mathcal{K}(\phi,A_t)
\Big(
\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle+ N^{-1/6} \ln(N)
\\
+&
\inf \left\lbrace
\left|\mathcal{E}_{V_N}(\Psi)-\mathcal{E}_{4 \pi}^{GP}(\phi)\right|,
\left|\mathcal{E}_{W_\beta}(\Psi)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi)\right|+ N^{-2 \beta} \ln(N)
\right\rbrace
\Big)
\;.
\end{align*}
\end{enumerate}
\end{proof}
Using this Lemma, it follows that there exists an $\eta>0$ such that
\begin{equation*}
\gamma_b^< (\Psi_t,\phi_t) \leq
\mathcal{K}(\phi_t,A_t)
\left(
\langle\hspace{-0.08cm}\langle\Psi_t,\widehat{n}^{\phi_t}\Psi_t\rangle\hspace{-0.08cm}\rangle + N^{-\eta}+
\left|
\mathcal{E}_{W_\beta}(\Psi_0)
-
\mathcal{E}_{N \|W_\beta\|_1}^ {GP}(\phi_0)
\right|
\right)\;.
\end{equation*}
This proves Lemma \ref{gammalemma}.
\subsection{Estimates for the functional $\gamma$}
\label{gammacontrolsection}
For the most involved scaling which is induced by $V_N$, we need to control $ \|p_1 V_N \Psi\|$.
\begin{lemma}
Let $\Psi \in L_s^2 (\mathbb{R}^{2N}, \mathbb{C})$ and let
$\mathcal{E}_{V_N}(\Psi) \leq C$. Then
\begin{align} \label{p1 VN abschaetzung}
\|p_1 V_N \Psi\| \leq \mathcal{K}(\phi, A_t) N^{-\frac{1}{2}}
\;.
\end{align}
\end{lemma}
\begin{proof}
We estimate
\begin{align*}
&\|p_1 V_N(x_1-x_2) \Psi\|= \|p_1 \mathds{1}_{\text{supp}(V_N)}(x_1-x_2) V_N(x_1-x_2) \Psi\|
\\
\leq & \|p_1 \mathds{1}_{\text{supp}(V_N)}(x_1-x_2) \|_{\text{op}} \| V_N(x_1-x_2) \Psi\|
\;.
\end{align*}
We have
\begin{align*}
\|p_1 \mathds{1}_{\text{supp}(V_N)} (x_1-x_2)\|_{\text{op}}^2\leq \| \varphi \|_\infty^2
\|\mathds{1}_{\text{supp}(V_N)} \|_1 \leq C \| \varphi \|_\infty^2 e^{-2N}
\;.
\end{align*}
Using
\begin{align*}
C \geq \mathcal{E}_{V_N}(\Psi)= \| \nabla \Psi \|^2 + (N-1) \| \sqrt{V_N}(x_1-x_2) \Psi \|^2
+
\langle\hspace{-0.08cm}\langle \Psi, A_t(x_1) \Psi \rangle\hspace{-0.08cm}\rangle
\end{align*}
as well as
\begin{align*}
\| V_N(x_1-x_2) \Psi\|^2 =& \| \sqrt{V_N} (x_1-x_2)\sqrt{ V_N}(x_1-x_2) \Psi\|^2
\leq
\| \sqrt{V_N}\|_{\infty}^2 \| \sqrt{V_N}(x_1-x_2) \Psi \|^2
\nonumber
\\
\leq&
C e^{2N} \frac{\mathcal{E}_{V_N}(\Psi)+ \| A_t \|_\infty }{N}
\leq
C (1+ \| A_t \|_\infty) \frac{e^{2N}}{N}
\;,
\end{align*}
we obtain
\begin{align*}
\|p_1 V_N \Psi\| \leq
\mathcal{K}(\phi, A_t) N^{-\frac{1}{2}} \;.
\end{align*}
\end{proof}
\paragraph{Control of $\gamma_b$}
Recall that
\begin{align*}\gamma_b (\Psi,\phi)&=-N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,
\widetilde{Z}_{\beta}^{\phi}(x_1,x_2)\widehat{r}\,
\Psi\rangle\hspace{-0.08cm}\rangle\right)\\&
-N(N-1)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
\widehat{r}\,\mathcal{Z}^\phi (x_1,x_2) \Psi\rangle\hspace{-0.08cm}\rangle\right)\;.
\end{align*}
Estimate \eqref{p1 VN abschaetzung} yields to the bound
$\|p_1 \mathcal{Z}^\phi (x_1,x_2)\Psi\| \leq \mathcal{K}(\phi, A_t) N^{-1/2}$.
Therefore, the second line of $\gamma_b$ is controlled by
\begin{align*}
&N^2\|g_{\beta}(x_{1}-x_{2})
p_{1}\|_{\text{op}}\|\widehat{r}\|_{\text{op}}\|p_1\mathcal{Z}^\phi (x_1,x_2)\Psi\|
\\
\leq & \mathcal{K}(\phi, A_t)N^{3/2}\|g_{\beta}\| \|\widehat{r}\|_{\text{op}}
\leq \mathcal{K}(\phi, A_t) N^{\xi-1/2-\beta/2} \ln(N) \;.
\end{align*}
The first line of $\gamma_b$ can be bounded with (\ref{defzz}) and $f_{\beta}=1-g_{\beta}$ by
\begin{align}
\nonumber
&N(N-1)|\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,
\widetilde{Z}_{\beta}^{\phi}(x_1,x_2)\widehat{r}\,
\Psi\rangle\hspace{-0.08cm}\rangle\right)|
\\
\leq&
\label{oans}
N^2 |
\Im
\left(
\langle\hspace{-0.08cm}\langle\Psi ,
\left(
M_{\beta}(x_1-x_2)f_{\beta}(x_{1}-x_{2})-
\frac{N}{N-1}
\left(
\| M_{\beta}f_{\beta} \|_1
|\phi(x_1)|^2
+\| M_{\beta}f_{\beta} \|_1
|\phi(x_2)|^2
\right)
\right)\widehat{r}
\Psi\rangle\hspace{-0.08cm}\rangle
\right)
|
\\&+
\label{zwoa}
\frac{N^2}{N-1} |\langle\hspace{-0.08cm}\langle\Psi ,
\left(
\| N M_{\beta}f_{\beta} \|_1- 4\pi\right)\left( |\phi(x_1)|^2+|\phi(x_2)|^2\right)\widehat{r}
\Psi\rangle\hspace{-0.08cm}\rangle|
\\&+
\label{gsuffa}
\frac{N^2}{N-1}|\langle\hspace{-0.08cm}\langle\Psi ,
\left(4\pi|\phi(x_1)|^2+4\pi|\phi(x_2)|^2\right)g_{\beta}(x_{1}-x_{2})
\widehat{r}\Psi\rangle\hspace{-0.08cm}\rangle|
\;.
\end{align}
Since $M_{\beta}f_{\beta} \in \mathcal{V}_\beta$,
\eqref{oans} is of the same form as $\gamma^<_b (\Psi,\varphi)$.
Using Lemma \ref{defAlemma} (h), the second term is controlled by
\begin{align*}
\eqref{zwoa}
\leq
C
\| \varphi \|_\infty^2
N
\left(
N\|M_\beta f_\beta\|_1-4 \pi
\right)
\|\hat{r}\|_{\text{op}}
\leq
C \| \varphi \|_\infty^2
N^{-1+ \xi} \ln(N)
\;.
\end{align*}
The last term is controlled by
$$
\eqref{gsuffa} \leq
C N\|\phi\|_\infty^2\|g_{\beta}(x_{1}-x_{2})
p_{1}\|_{\text{op}}\|\widehat{r}\|_{\text{op}}\leq
C\|\phi\|_\infty^3N^{-1-\beta+\xi} \ln(N)
\;.
$$
and we get
$$|\gamma_b (\Psi,\phi)|\leq
\mathcal{K}(\phi, A_t)
\left(
\langle\hspace{-0.08cm}\langle \Psi, \widehat{m} \Psi \rangle\hspace{-0.08cm}\rangle + |\mathcal{E}_{V_N}(\Psi)
- \mathcal{E}^{GP}_{4 \pi}(\phi) |
+N^{-\eta}\right)$$
for some $\eta>0$.
\paragraph{Control of $\gamma_c$}
Recall that \begin{align*} \gamma_c (\Psi,\phi)=&-4N(N-1)\langle\hspace{-0.08cm}\langle\Psi
, (\nabla_1g_{\beta}(x_1-x_2))\nabla_1
\widehat{r}\Psi\rangle\hspace{-0.08cm}\rangle\;.\end{align*}
Using $\widehat{r}=(p_2+q_2)\widehat{r}=p_2\widehat{r}+p_1q_2\widehat{m}^a$ and $\nabla_1g_{\beta}(x_1-x_2)=-\nabla_2g_{\beta}(x_1-x_2)$, integration by parts yields to
\begin{align}\label{extra1}
|\gamma_c (\Psi,\phi)|\leq&
4N^2|\langle\hspace{-0.08cm}\langle\Psi
, g_{\beta}(x_1-x_2)\nabla_1\nabla_2
(
p_2\widehat{r}+p_1q_2\widehat{m}^a
)
\Psi\rangle\hspace{-0.08cm}\rangle|
\\\label{extra2}&+
4N^2|\langle\hspace{-0.08cm}\langle\nabla_2\Psi
, g_{\beta}(x_1-x_2)\nabla_1
p_2\widehat{r}\Psi\rangle\hspace{-0.08cm}\rangle|
\\\label{extra3}&+4N^2|\langle\hspace{-0.08cm}\langle\nabla_2\Psi
, g_{\beta}(x_1-x_2)\nabla_1
p_1q_2\widehat{m}^a\Psi\rangle\hspace{-0.08cm}\rangle|
\;.\end{align}
We begin with
\begin{align*}
(\ref{extra1})\leq &C N^2
\|g_\beta\|
\| \nabla \varphi \|_\infty
\left(
\| \nabla_1 \widehat{r} \psi \|
+
\| \nabla_ 2 q_2 \widehat{m}^a\Psi \|
\right)
\\
\leq &
C
N^{1-\beta} \ln(N)
\| \nabla \varphi \|_\infty
\left(
\| \nabla_1 \widehat{r} \psi \|
+
\| \nabla_ 2 q_2 \widehat{m}^a\Psi \|
\right)
\;.
\end{align*}
Let $s_1,t_1\in \lbrace p_1,q_1 \rbrace , \; \;
s_2,t_2\in \lbrace p_2,q_2 \rbrace$. Inserting the identity $1=(p_1+q_1)(p_2+q_2)$, we obtain,
for $a \in \lbrace -2,-1,0,1,2 \rbrace$,
\begin{align*}
\|\nabla_1\widehat{r}\Psi\|
\leq& C
\sup_{s_1,s_2,t_1,t_2, a}
\|\widehat{r}_a s_1 s_2 \nabla_1 t_1 t_2 \Psi \|
\leq C
\sup_{t_1, a}
\|\widehat{r}_a\|_{\text{op}} \|\nabla_1 t_1 \Psi \|
\\
\leq &
C N^{-1 + \xi}
\;.
\end{align*}
In analogy $ \| \nabla_ 2 q_2 \widehat{m}^a\Psi \| \leq C \| \widehat{m}^a \|_{\text{op}} \leq C N^{-1 + \xi}$.
This yields the bound
\begin{align*}
(\ref{extra1})\leq
\mathcal{K}(\phi, A_t)
N^{-\beta+ \xi} \ln( N)
\;.
\end{align*}
Furthermore, (\ref{extra2}) is bounded by
\begin{align}
(\ref{extra2})\leq&4N^2\|\nabla_2\Psi\|\;
\|g_{\beta}\|\;\| \nabla \phi\|_\infty\|\nabla_1\widehat{r}\Psi\|
\leq C
\|\nabla \phi\|_\infty\
N^{\xi-\beta} \ln(N) \;.
\end{align}
Similarly, we obtain
\begin{align*}(\ref{extra3})\leq&
4N^2\|\nabla_2\Psi\|\;
\|g_{\beta}\|\;\|\nabla \phi\|_\infty\| q_2 \widehat{m}^a \Psi\|
\leq C \| \nabla \phi\|_\infty\ N^{ \xi-\beta} \ln(N) \;.
\end{align*}
It follows that $ |\gamma_c (\Psi,\phi)|\leq
\mathcal{K}(\phi, A_t)
N^{ \xi-\beta}
\ln(N)
$.
\paragraph{Control of $\gamma_d$}
To control $\gamma_d$ and $\gamma_e$ we will use the notation
\begin{equation}\label{notation} \begin{array}{cc}
m^c(k)=m^a(k)-m^a(k+1) & m^d(k)=m^a(k)-m^a(k+2) \\
m^e(k)=m^b(k)-m^b(k+1) & m^f(k)=m^b(k)-m^b(k+2) \;. \\
\end{array}
\end{equation}
Since the second $k$-derivative of $m$ is given by (see (\ref{mprime}) for the first derivative)
$$m(k)^{\prime\prime}=\left\{
\begin{array}{ll}
-1/(4\sqrt{k^3N}), & \hbox{for $k\geq N^{1-2\xi}$;} \\
0, & \hbox{else.}
\end{array}
\right.$$
it is easy to verify that
\begin{equation}\label{estcdef}\|\widehat{m}_j^x\|_{\text{op}} \leq C N^{-2+3\xi} \text{ for }x\in\{c,d,e,f\}\;.\end{equation}
Recall that \begin{align*}\gamma_d (\Psi,\phi)=&2N(N-1)(N-2)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
\left[V_N(x_1-x_3),
\widehat{r}\right] \Psi\rangle\hspace{-0.08cm}\rangle\right)
\\&N(N-1)(N-2)\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
\left[4 \pi|\phi|^2(x_3),
\widehat{r}\right] \Psi\rangle\hspace{-0.08cm}\rangle\right)\;.\end{align*}
Since $p_j+q_j=1$, we can rewrite $\widehat{r}$ as
$$
\widehat{r}=\widehat{m}^bp_{1}p_{2}+\widehat{m}^a(p_{1}q_{2}+q_1p_2)=(\widehat{m}^b-2\widehat{m}^a)p_{1}p_{2}+\widehat{m}^a(p_{1}+p_2)\;.$$
Thus,
\begin{align}\nonumber|\gamma_d (\Psi,\phi)|\leq &C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
\left[V_N(x_1-x_3),
(\widehat{m}^b-2\widehat{m}^a)p_{1}p_{2}+\widehat{m}^a(p_{1}+p_2)\right] \Psi\rangle\hspace{-0.08cm}\rangle\right|
\\&\nonumber+C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
\left[4\pi|\phi|^2(x_3),
\widehat{r}\right] \Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\label{gleichung3}\leq&C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})p_2
\left[V_N(x_1-x_3),
\widehat{m}^a\right] \Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\label{gleichung1}&+C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
V_N(x_1-x_3)
(\widehat{m}^b-2\widehat{m}^a)p_{1}p_{2}\Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\label{gleichung2}&+C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
(\widehat{m}^b-2\widehat{m}^a)p_{1}p_{2}V_N(x_1-x_3) \Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\label{gleichung4}&+C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
\widehat{m}^ap_{1} V_N(x_1-x_3)\Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\label{gleichung5}&+C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})V_N(x_1-x_3)
\widehat{m}^ap_{1} \Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\label{gleichung6}&+C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
\left[4 \pi |\phi|^2(x_3),
\widehat{r}\right] \Psi\rangle\hspace{-0.08cm}\rangle\right|
\;.
\end{align}
Using Lemma \ref{kombinatorik} (d),
we obtain the following estimate:
\begin{align*}
(\ref{gleichung3})=&
C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})p_2
\left[V_N(x_1-x_3),
p_1p_3\widehat{m}^d+p_1q_3\widehat{m}^c+q_1p_3\widehat{m}^c\right] \Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\leq&
C N^3\big|\langle\hspace{-0.08cm}\langle\Psi ,V_N(x_1-x_3)g_{\beta}(x_{1}-x_{2})p_2\mathds{1}_{\text{supp}(V_N)}(x_1-x_3)
\\&\hspace{4cm}\left(p_1p_3\widehat{m}^d+p_1q_3\widehat{m}^c+q_1p_3\widehat{m}^c\right) \Psi\rangle\hspace{-0.08cm}\rangle\big|
\\&+
C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})p_2\left(
p_1p_3\widehat{m}^d+p_1q_3\widehat{m}^c+q_1p_3\widehat{m}^c\right)V_N(x_1-x_3)\Psi\rangle\hspace{-0.08cm}\rangle\right|
\;.
\end{align*}
Both lines are bounded by
\begin{align*}
&C N^3\|V_N(x_1-x_3)\Psi\|\;\|g_{\beta}(x_{1}-x_{2})p_2\|_{\text{op}}
\\&\left(2\|\mathds{1}_{\text{supp}(V_N)}(x_1-x_3)p_1\|_{\text{op}}+\|\mathds{1}_{\text{supp}(V_N)}(x_1-x_3)p_3\|_{\text{op}}\right)
\left(\|\widehat{m}^d\|_{\text{op}}+\|\widehat{m}^c\|_{\text{op}}\right)
\;.
\end{align*}
In view of Lemma \ref{kombinatorik} (e) with Lemma \ref{defAlemma} (i), $\|g_{\beta}(x_{1}-x_{2})p_2\|_{\text{op}} \leq \|\phi\|_\infty \| g_\beta \| \leq C \|\phi\|_\infty N^{-1-\beta} \ln(N) $.
Using (\ref{estcdef}), together with
$ \|\mathds{1}_{\text{supp}(V_N)}(x_1-x_3)p_1\|_{\text{op}} \|V_N(x_1-x_3)\Psi\|\leq \mathcal{K}(\phi, A_t) N^{-1/2}$,
we obtain, using $\xi <1/2$,
$$(\ref{gleichung3})\leq \mathcal{K}(\phi,A_t) N^{-1/2+3\xi-\beta}\ln(N)
\leq \mathcal{K}(\phi,A_t) N^{1/2+\xi-\beta}\ln(N)
\;.$$
We continue with
\begin{align*}
&(\ref{gleichung1})+(\ref{gleichung2})+(\ref{gleichung4})\\\leq&
C N^3\|V_N(x_1-x_3)\Psi\|\|g_{\beta}(x_{1}-x_{2})p_2\|_{\text{op}}
\\ & \times \|\mathds{1}_{\text{supp}(V_N)}(x_1-x_3)p_1\|_{\text{op}}\|
(\widehat{m}^b-2\widehat{m}^a)\|_{\text{op}}
\\&+C N^3\|g_{\beta}(x_{1}-x_{2})p_2\|_{\text{op}}
\|\widehat{m}^b-2\widehat{m}^a\|_{\text{op}}\|p_{1}V_N(x_1-x_3) \Psi\|
\\&+C N^3\|g_{\beta}(x_{1}-x_{2})p_1\|_{\text{op}}\|
\widehat{m}^a\|_{\text{op}}\|p_{1} V_N(x_1-x_3)\Psi\|
\\\leq & \mathcal{K}(\phi,A_t) N^{1/2+\xi-\beta} \ln(N)
\;.
\end{align*}
Next, we estimate \eqref{gleichung5}.
The support of the function $g_{\beta}(x_{1}-x_{2})V_N(x_1-x_3)$ is such that
$|x_1-x_2| \leq C N^ {-\beta}$, as well as $|x_1-x_3| \leq C e^ {-N}$. Therefore,
$g_{\beta}(x_{1}-x_{2})V_N(x_1-x_3) \neq 0$ implies $ |x_2-x_3| \leq C N^ {- \beta} $. We estimate
\begin{align*}
\eqref{gleichung5} =&C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})V_N(x_1-x_3)p_{1}
\mathds{1}_{ B_{C N^{- \beta}}(0)} (x_2-x_3)
\widehat{m}^a \Psi\rangle\hspace{-0.08cm}\rangle\right|
\\
\leq &C N^3 \| p_{1} V_N(x_1-x_3) g_{\beta}(x_{1}-x_{2}) \Psi \|
\|\mathds{1}_{ B_{C N^{- \beta}}(0)} (x_2-x_3)
\widehat{m}^a \Psi \|
\\
\leq &C N^3
\|p_1 \mathds{1}_{\text{supp}(V_N)} (x_1-x_3)\|_{\text{op}}
\|g_{\beta}(x_{1}-x_{2}) V_N(x_1-x_3) \Psi \|
\| \mathds{1}_{ B_{C N^{- \beta}}(0)} (x_2-x_3)
\widehat{m}^a \Psi \|
\\
\leq & C N^{5/2} \| g_{\beta}\|_\infty
\|\mathds{1}_{ B_{C N^{- \beta}}(0)} \|^{\frac{1}{2}}_{\frac{p}{p-1}}
\| \nabla_1 \widehat{m}^a \Psi \|^{ \frac{p-1}{ p}} \| \widehat{m}^a \Psi \|^{\frac{1}{p}}
\\
\leq &
C N^{5/2} \| g_{\beta}\|_\infty N^{- \beta/2} \| \nabla_1 \widehat{m}^a \Psi \|^{1/2} \| \widehat{m}^a \Psi \|^{1/2}
\\
\leq &
C N^{3/2+\xi- \beta/2}
\;.
\end{align*}
In the fourth line, we applied Sobolev inequality as in the proof of Lemma \ref{propo}, then setting $p=2$.
Furthermore, we used
$ \| \nabla_1 \widehat{m}^a \Psi \|^{1/2} \| \widehat{m}^a \Psi \|^{1/2} \leq C N^{-1 + \xi}$,
as well as $ \| g_{\beta}\|_\infty \leq C$, see Lemma \ref{defAlemma}.
Using Lemma \ref{kombinatorik} (d),
(\ref{gleichung6}) can be bounded by
\begin{align*}
&C N^3\left|\langle\hspace{-0.08cm}\langle\Psi ,g_{\beta}(x_{1}-x_{2})
\left[4 \pi|\phi|^2(x_3),
p_1p_2(\widehat{r}-\widehat{r}_2)+(p_1q_2+q_1p_2)(\widehat{r}-\widehat{r}_1)\right] \Psi\rangle\hspace{-0.08cm}\rangle\right|
\\&\leq CN^3\|\phi\|_\infty^2 \left(\|\widehat{r}-\widehat{r}_2\|_{\text{op}}+\|\widehat{r}-\widehat{r}_1\|_{\text{op}}\right)
\|g_{\beta}(x_{1}-x_{2})p_2\|_{\text{op}}
\;.
\end{align*}
Note that $\|\widehat{r}-\widehat{r}_2\|_{\text{op}}+\|\widehat{r}-\widehat{r}_1\|_{\text{op}}\leq\sum_{j\in\{c,d,e,f\}}\|\widehat{m}^j\|_{\text{op}}
\leq C N^{-2+ 3 \xi}$ holds.
With $\|g_{\beta}(x_{1}-x_{2})p_2\|_{\text{op}}\leq C N^{-1-\beta} \ln(N) $, it then follows that
$$|(\ref{gleichung6})|\leq C
\| \phi \|_\infty^2
N^{3\xi-\beta} \ln(N)\;.$$
In total, we obtain
\begin{align*}
|\gamma_d (\Psi,\phi)| \leq
\mathcal{K}(\phi,A_t)
\left(
N^{3/2+\xi-\beta/2}+N^{1/2+3\xi-\beta}\ln(N)
\right) \;.
\end{align*}
\paragraph{Control of $\gamma_e$}
Recall that \begin{align*}&\gamma_e (\Psi,\phi)=-\frac{1}{2}N(N-1)(N-2)(N-3)\\&\;\;\;\;\;\;\;\;\;\;\Im\left(\langle\hspace{-0.08cm}\langle\Psi ,
g_{\beta}(x_{1}-x_{2})\left[V_N(x_3-x_4),
\widehat{r}\right] \Psi\rangle\hspace{-0.08cm}\rangle\right)
\;.
\end{align*}
Using symmetry, Lemma \ref{kombinatorik} (d) and notation (\ref{notation}),
$\gamma_e$ is bounded by
\begin{align*}\gamma_e (\Psi,\phi)\leq&N^4\big|\langle\hspace{-0.08cm}\langle\Psi ,
g_{\beta}(x_{1}-x_{2})\big[V_N(x_3-x_4),
\widehat{m}^cp_1p_2p_3p_4+2\widehat{m}^dp_1p_2p_3q_4\\&\hspace{5cm}+
2\widehat{m}^ep_1q_2p_3p_4+4\widehat{m}^fp_1q_2p_3q_4\big] \Psi\rangle\hspace{-0.08cm}\rangle\big|
\\\leq&4N^4\|V_N(x_3-x_4)\Psi\|\|\mathds{1}_{\text{supp}(V_N)}(x_3-x_4)p_3\|_{\text{op}}\|g_{\beta}(x_{1}-x_{2})p_1\|_{\text{op}}
\\&\hspace{3cm}(\|\widehat{m}^c\|_{\text{op}}+\|\widehat{m}^d\|_{\text{op}}+\|\widehat{m}^e\|_{\text{op}}+\|\widehat{m}^f\|_{\text{op}})
\;.
\end{align*}
We get with (\ref{estcdef}), Lemma \ref{defAlemma} and Lemma \ref{kombinatorik} that
$$ |\gamma_e (\Psi,\phi)|\leq \mathcal{K}(\phi,A_t) N^{1/2+3\xi- \beta} \ln(N)\;.$$
\paragraph{Control of $\gamma_f$}
Recall that
\begin{align*}
\gamma_f(\Psi,\phi)
=
2
N(N-1)\frac{N-2}{N-1}\Im\left(\langle\hspace{-0.08cm}\langle\Psi,g_{\beta}(x_{1}-x_{2})
\left[4 \pi |\phi|^2(x_1),
\widehat{r}\right] \Psi\rangle\hspace{-0.08cm}\rangle\right)
\;.
\end{align*}
We obtain the estimate
\begin{align*}
|\gamma_f(\Psi,\phi)|
\leq
\mathcal{K}(\phi,A_t)
N^2
\|g_\beta\|
\| \widehat{r} \|_{\text{op}}
\leq
\mathcal{K}(\phi,A_t)
N^{\xi-\beta} \ln(N)
\;.
\end{align*}
Collecting all estimates, we get with $\xi <1/2$
\begin{align*}
|\gamma_c (\Psi,\phi)|+|\gamma_d (\Psi,\phi)|+|\gamma_e (\Psi,\phi)|+|\gamma_f(\Psi,\phi)|
\leq
\mathcal{K}(\phi,A_t)
N^{2-\beta/2} \ln(N)
\;.
\end{align*}
Choosing $\beta$ sufficiently large, we obtain the desired decay and hence Lemma \ref{gammalemma fuer V}.
\subsection{Energy estimates}
\begin{lemma} \label{energylemma}
Let $\Psi \in L^2_s( \mathbb{R}^{2N}, \mathbb{C}) \cap H^1( \mathbb{R}^{2N}, \mathbb{C}) ,\; \| \Psi \|=1$ with $ \|\nabla_1\Psi \| \leq \mathcal{K}(\phi, A_t)$. Let
$ \phi \in H^3(\mathbb{R}^2,\mathbb{C}) , \; \| \phi \|=1$. Define the sets
$ \mathcal{A}^{(d)}_{1},\overline{\mathcal{B}}^{(d)}_{1} $ as in Definition \ref{hdetail}.
Then, for $d$ large enough,
\begin{align*}
&\|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1
\Psi \|^2
+
\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\nabla_1q_1
\Psi \|^2
\leq
\mathcal{K}(\phi, A_t)
\Big( \left\langle\hspace{-0.14cm}\left\langle\Psi,\widehat{n}^{\phi}\Psi\right\rangle\hspace{-0.14cm}\right\rangle+N^{-1/6} \ln(N)
\\
&+
\inf \left\lbrace
\left|\mathcal{E}_{V_N}(\Psi)-\mathcal{E}_{4 \pi}^{GP}(\phi)\right|,
\left|\mathcal{E}_{W_\beta}(\Psi)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi)\right|+ N^{-2 \beta} \ln(N)
\right\rbrace
\Big)
\;.
\end{align*}
\end{lemma}
\begin{proof}
We start with expanding $\mathcal{E}_{W_\beta}(\Psi)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi)$. This yields
\begin{align*}
&\mathcal{E}_{W_\beta}(\Psi)-\mathcal{E}_{N \|W_\beta\|_1}^{GP}(\phi)
=
\|\nabla_1\Psi \|^2+\frac{N-1}{2}\|\sqrt{W_\beta}(x_1-x_2)\Psi\|^2
\nonumber
\\&
-\|\nabla\phi \|^2-\frac{1}{2} N\|W_\beta\|_1 \|\phi^2\|^2+
\langle\hspace{-0.08cm}\langle \Psi, A_t (x_1) \Psi \rangle\hspace{-0.08cm}\rangle - \langle \phi, A_t \phi \rangle
\nonumber
\\=&
\|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi \|^2+\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\mathds{1}_{\overline{
\mathcal{A}^{(d)}}_{1}}\nabla_1\Psi \|^2
+
M( \Psi, \phi)+Q_\beta( \Psi, \phi)
\;,
\end{align*}
where we have defined
\begin{align}
M( \Psi, \phi)=
\label{1 term in M}
&2 \Re\left(\langle\hspace{-0.08cm}\langle\nabla_1q_1\Psi ,
\mathds{1}_{\mathcal{A}^{(d)}_{1}}
\nabla_1p_1\Psi \rangle\hspace{-0.08cm}\rangle\right)
\\&
\label{2 term in M}
+ \|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1p_1\Psi \|^2-\|\nabla\phi \|^2
\\&
\label{3 term in M}
+\langle\hspace{-0.08cm}\langle\Psi , A_t(x_1)\Psi\rangle\hspace{-0.08cm}\rangle-\langle\phi , A_t\phi\rangle
\;,
\\
\nonumber
Q_\beta( \Psi, \phi)=&
\|\mathds{1}_{\mathcal{B}^{(d)}_{1}}\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1\Psi \|^2
\\
\nonumber
+&
\frac{N-1}{2}\langle\hspace{-0.08cm}\langle\Psi ,(1-p_1p_2)W_{\beta}(x_1-x_2)(1-p_1p_2)\Psi\rangle\hspace{-0.08cm}\rangle
\\
\nonumber
+&\frac{N-1}{2}\langle\hspace{-0.08cm}\langle\Psi ,p_1p_2W_{\beta}(x_1-x_2)p_1p_2\Psi\rangle\hspace{-0.08cm}\rangle -\frac{1}{2}N\|W_\beta\|_1\|\phi^2\|^2
\\
\nonumber
+&(N-1)\Re\langle\hspace{-0.08cm}\langle\Psi ,(1-p_1p_2)W_{\beta}(x_1-x_2)p_1p_2\Psi\rangle\hspace{-0.08cm}\rangle
\;.
\end{align}
Notice that the first two terms in $Q_\beta( \Psi, \phi)$ are nonnegative. This yields to the bound
\begin{align}
S_\beta( \Psi, \phi)=&
\label{1-pp term in Sbeta}
(N-1)|\langle\hspace{-0.08cm}\langle\Psi ,(1-p_1p_2)W_{\beta}(x_1-x_2)p_1p_2\Psi\rangle\hspace{-0.08cm}\rangle |
\\
+&
\label{pp pq term in Sbeta}
\left|
\frac{N-1}{2}\langle\hspace{-0.08cm}\langle\Psi ,p_1p_2W_{\beta}(x_1-x_2)p_1p_2\Psi\rangle\hspace{-0.08cm}\rangle -\frac{1}{2}N\|W_\beta\|_1\|\phi^2\|^2
\right|
\\
\nonumber
\geq&
-Q_\beta( \Psi, \phi)
\;.
\end{align}
In total, we obtain
\begin{align*}
&|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi \|^2
+
\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\nabla_1q_1
\Psi \|^2
\leq
M( \Psi, \phi)+S_\beta( \Psi, \phi)+\left|\mathcal{E}_{W_\beta}(\Psi)-\mathcal{E}_{N \|W_{\beta}\|_1}^{GP}(\phi)\right|
\;.
\end{align*}
Next, we split up the energy difference $\mathcal{E}_{V_N}(\Psi)-\mathcal{E}_{4 \pi}^{GP}(\phi)$, using $4 \pi= 4 \pi$,
\begin{align*}
\mathcal{E}_{V_N}(\Psi)-\mathcal{E}_{4 \pi}^{GP}(\phi)&=
\|\nabla_1\Psi \|^2+\frac{N-1}{2}\|\sqrt{V_N}(x_1-x_2)\Psi\|^2
-\|\nabla\phi \|^2
\\
&-2\pi\|\phi^2\|^2+
\langle\hspace{-0.08cm}\langle \Psi, A_t (x_1) \Psi \rangle\hspace{-0.08cm}\rangle - \langle \phi, A_t \phi \rangle
\;.
\end{align*}
In order to better estimate the terms corresponding to the two-particle interactions, we introduce, for $\mu>d$, the potential $M_{\mu} (x)$,
defined in Definition \ref{microscopic}, and continue with
\begin{align*}
\mathcal{E}_{V_N}(\Psi)-\mathcal{E}_{4 \pi}^{GP}(\phi)=&\|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1\Psi\|^2
+ \|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\mathds{1}_{\overline{
\mathcal{A}^{(d)}}_{1}}\nabla_1\Psi \|^2+\|\mathds{1}_{\mathcal{B}^{(d)}_{1}}\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1\Psi \|^2 \nonumber\\&
+\frac{N-1}{2}\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\sqrt{V_N}(x_1-x_2)\Psi \|^2
\\&+
\frac{1}{2}
\langle\hspace{-0.08cm}\langle\Psi ,\sum_{j\neq
1}\mathds{1}_{\mathcal{B}^{(d)}_{1}}\left(V_N-M_{\mu}\right)(x_1-x_j)\Psi \rangle\hspace{-0.08cm}\rangle
\\& +
\frac{1}{2}
\langle\hspace{-0.08cm}\langle\Psi ,\sum_{j\neq
1}\mathds{1}_{\mathcal{B}^{(d)}_{1}}M_{\mu}(x_1-x_j)\Psi \rangle\hspace{-0.08cm}\rangle
-\|\nabla\phi \|^2-2\pi\|\phi^2\|^2
\\&+\langle\hspace{-0.08cm}\langle\Psi , A_t(x_1)\Psi\rangle\hspace{-0.08cm}\rangle-\langle\phi , A_t\phi\rangle
\;.
\end{align*}
Using that $q_1=1-p_1$ and symmetry gives (after reordering)
\begin{align*} &
\mathcal{E}_{V_N}(\Psi)-\mathcal{E}_{4 \pi}^{GP}(\phi)
\\
=&
\|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi \|^2+\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\mathds{1}_{\overline{
\mathcal{A}^{(d)}}_{1}}\nabla_1\Psi \|^2
+
\frac{N-1}{2}\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\sqrt{V_N}(x_1-x_2)\Psi \|^2
\\&
+\frac{N-1}{2}\langle\hspace{-0.08cm}\langle\Psi ,\mathds{1}_{\mathcal{B}^{(d)}_{1}}(1-p_1p_2)M_{\mu}(x_1-x_2)(1-p_1p_2)\mathds{1}_{\mathcal{B}^{(d)}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle
\\&+ \|\mathds{1}_{\mathcal{B}^{(d)}_{1}}\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1\Psi \|^2
+
\frac{1}{2}
\langle\hspace{-0.08cm}\langle\Psi ,\sum_{j\neq
1}\mathds{1}_{\mathcal{B}^{(d)}_{1}}\left(V_N-M_{\mu}\right)(x_1-x_j)\Psi \rangle\hspace{-0.08cm}\rangle
\\&+\frac{N-1}{2}\langle\hspace{-0.08cm}\langle\Psi ,\mathds{1}_{\mathcal{B}^{(d)}_{1}}p_1p_2M_{\mu}(x_1-x_2)p_1p_2\mathds{1}_{\mathcal{B}^{(d)}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle -2\pi\|\phi^2\|^2
\\&+2 \Re\left(\langle\hspace{-0.08cm}\langle\nabla_1q_1\Psi ,
\mathds{1}_{\mathcal{A}^{(d)}_{1}}
\nabla_1p_1\Psi \rangle\hspace{-0.08cm}\rangle\right)
\\&
+(N-1)\Re\langle\hspace{-0.08cm}\langle\Psi ,\mathds{1}_{\mathcal{B}^{(d)}_{1}}(1-p_1p_2)M_{\mu}(x_1-x_2)p_1p_2\mathds{1}_{\mathcal{B}^{(d)}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle
\\&
+ \|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1p_1\Psi \|^2-\|\nabla\phi \|^2
\\&+\langle\hspace{-0.08cm}\langle\Psi, A_t(x_1)\Psi\rangle\hspace{-0.08cm}\rangle-\langle\phi , A_t\phi\rangle
\\
=&
\|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi \|^2+\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\nabla_1\Psi \|^2
+
M( \Psi, \phi)+\tilde{Q}_\mu( \Psi, \phi)
\;.
\end{align*}
with
\begin{align}
&\tilde{Q}_\mu( \Psi, \phi)=
\frac{N-1}{2}\langle\hspace{-0.08cm}\langle\Psi ,\mathds{1}_{\mathcal{B}^{(d)}_{1}}(1-p_1p_2)M_{\mu}(x_1-x_2)(1-p_1p_2)\mathds{1}_{\mathcal{B}^{(d)}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle
\nonumber
\\
&+
\frac{N-1}{2}\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\sqrt{V_N}(x_1-x_2)\Psi \|^2
\nonumber
\\
\label{vielezeilenb}&
+ \|\mathds{1}_{\mathcal{B}^{(d)}_{1}}\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1\Psi \|^2
+
\frac{1}{2}
\langle\hspace{-0.08cm}\langle\Psi ,\sum_{j\neq
1}\mathds{1}_{\mathcal{B}^{(d)}_{1}}\left(V_N-M_{\mu}\right)(x_1-x_j)\Psi \rangle\hspace{-0.08cm}\rangle
\\&
\nonumber
+(N-1)\Re\langle\hspace{-0.08cm}\langle\Psi ,\mathds{1}_{\mathcal{B}^{(d)}_{1}}(1-p_1p_2)M_{\mu}(x_1-x_2)p_1p_2\mathds{1}_{\mathcal{B}^{(d)}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle
\\
&+\frac{N-1}{2}\langle\hspace{-0.08cm}\langle\Psi ,\mathds{1}_{\mathcal{B}^{(d)}_{1}}p_1p_2M_{\mu}(x_1-x_2)p_1p_2\mathds{1}_{\mathcal{B}^{(d)}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle -2\pi\|\phi^2\|^2
\;.
\nonumber
\end{align}
The first two terms in $\tilde{Q}_\mu( \Psi, \phi)$ are nonnegative.
For $\mu >d$ Lemma \ref{positiv V-w} below shows that \eqref{vielezeilenb} is also nonnegative. Thus, for $\mu >d$, we obtain the bound
\begin{align}
\label{vielezeilenc2}
\tilde{S}_\mu( \Psi, \phi)=
&(N-1)\left|
\langle\hspace{-0.08cm}\langle\Psi ,\mathds{1}_{\mathcal{B}^{(d)}_{1}}(1-p_1p_2)M_{\mu}(x_1-x_2)p_1p_2\mathds{1}_{\mathcal{B}^{(d)}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle
\right|
\\
\label{vielezeilenc}
+&
\left|
\frac{N-1}{2}\langle\hspace{-0.08cm}\langle\Psi ,\mathds{1}_{\mathcal{B}^{(d)}_{1}}p_1p_2M_{\mu}(x_1-x_2)p_1p_2\mathds{1}_{\mathcal{B}^{(d)}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle -2\pi\|\phi^2\|^2
\right|
\\
\nonumber
\geq& -\tilde{Q}_\mu (\Psi, \phi)
\;.
\end{align}
In total, we obtain
\begin{align*}\
&|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1q_1\Psi \|^2
+
\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\nabla_1q_1
\Psi \|^2
\leq
M( \Psi, \phi)+\tilde{S}_\mu( \Psi, \phi)+ \left|\mathcal{E}_{V_N}(\Psi)-\mathcal{E}_{4 \pi}^{GP}(\phi)\right|
\;.
\end{align*}
Next, we will estimate $M( \Psi, \phi),S_\beta( \Psi, \phi)$ and $\tilde{S}_\mu( \Psi, \phi)$.
\begin{itemize}
\item Estimate of $S_\beta( \Psi, \phi)$ and $\tilde{S}_\mu( \Psi, \phi)$. \\
We first estimate \eqref{vielezeilenc}, using the same estimate as in \eqref{roemisch I potetnialestimate}.
Note that
$$
\langle\hspace{-0.08cm}\langle\Psi ,\mathds{1}_{\mathcal{B}^{(d)}_{1}}p_1p_2M_{\mu}(x_1-x_2)p_1p_2\mathds{1}_{\mathcal{B}^{(d)}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle=
\langle\phi, M_{\mu} \star|\phi|^2\phi\rangle\langle\hspace{-0.08cm}\langle\Psi ,\mathds{1}_{\mathcal{B}^{(d)}_{1}}p_1p_2\mathds{1}_{\mathcal{B}^{(d)}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle
\;.
$$
Using $\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}} \Psi\|\leq C N^{1-d+ \epsilon}$, for any $\epsilon>0$, (see Lemma \ref{hdetail})
we obtain, together with $\|p_1p_2\Psi\|^2= 1+ 2 \|p_1 q_2\Psi\|^2+ \|q_1 q_2\Psi\|^2$
\begin{align*}
|\eqref{vielezeilenc}|
\leq &
3\| q_1 \Psi \|^2
+
C
\left(
N^{1-d+ \epsilon}
+
N^{2-2d+ 2 \epsilon}
\right)
+
\frac{1}{2}
|N\langle\phi, M_{\mu} \star|\phi|^2\phi\rangle
- N \|M_\mu \|_1 \| \varphi^2 \|^2
|
\\
+&
\frac{1}{2}
| 4 \pi- N \|M_\mu \|_1 |\| \varphi^2 \|^2
+
\frac{1}{2}
\langle\phi, M_{\mu} \star|\phi|^2\phi\rangle
\;.
\end{align*}
Note that, using Note that, using Young's inequality and
\eqref{roemisch I potetnialestimate}
\begin{align*}
&|
\langle\phi, N M_{\mu} \star|\phi|^2\phi\rangle
- N \|M_\mu \|_1\| \varphi^2 \|^2|
=
\left|
\int_{\mathbb{R}^2} d^2 x
|\phi(x)|^2
\left(
N( M_{\mu} \star|\phi|^2)(x)
-N \|M_\mu \|_1|\phi(x)|^2
\right)
\right|
\\
\leq &
\| \phi \|_\infty^2
\|N(M_\mu\star|\phi|^2)-\|NM_\mu\|_1|\phi|^2\|_1
\leq
C
\| \phi \|_\infty^2
\|\Delta|\phi|^2\|_1
N^{-2\mu} \ln(N)
\\
\leq &
\mathcal{K}(\phi, A_t)
N^{-2\mu} \ln(N)
\;.
\end{align*}
Since $|N \|M_{\mu}\|_1- 4 \pi | \leq C \frac{\ln(N)}{N} $ (see Lemma \ref{propo})
and
$\langle\phi, M_{\mu} \star|\phi|^2\phi\rangle \leq \|\phi\|_\infty^4 \|M_\mu \|_1 \leq
C \|\phi\|_\infty^4 N^{-1}
$,
it follows that
\begin{align}
\left|(\ref{vielezeilenc})\right|\leq & \mathcal{K}(\phi,A_t) \left(
\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}^{\phi}\Psi\rangle\hspace{-0.08cm}\rangle+
N^{1-d+ \epsilon}+N^{2-2d+2 \epsilon}+
N^{-2\mu} \ln(N) + N^{-1} \ln(N)
\right)
\nonumber
\\
\leq & \mathcal{K}(\phi,A_t)
\left(\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}^{\phi}\Psi\rangle\hspace{-0.08cm}\rangle+
N^{-1} \ln(N)
\right)
\;,
\end{align}
where the last inequality holds for $d$ large enough (recall that we chose $\mu>d$).
\\
Using the same estimates, we obtain
\begin{align*}
\eqref{pp pq term in Sbeta}
\leq
\mathcal{K}(\phi,A_t)
\left(\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}^{\phi}\Psi\rangle\hspace{-0.08cm}\rangle+
N^{-2\beta} \ln(N)
+
N^{-1} \ln(N)
\right)
\;.
\end{align*}
Line~\eqref{vielezeilenc2} and line \eqref{1-pp term in Sbeta} are controlled by Lemma \ref{energie(1-pp)pp}, which is stated below.
\begin{align*}
\eqref{1-pp term in Sbeta},
\eqref{vielezeilenc2}
\leq
\mathcal{K}(\phi,A_t)( \langle\hspace{-0.08cm}\langle \Psi, \widehat{n} \Psi \rangle\hspace{-0.08cm}\rangle + N^{-1/6} \ln(N) )
\;.
\end{align*}
In total, we obtain, for any $\mu> d \geq 1$, the bound
\begin{align*}
S_\beta (\Psi,\phi)
\leq &
\mathcal{K}(\phi,A_t) \left(\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle+N^{-2\beta } \ln(N) +N^{-1/6 } \ln(N)
\right)
\\
\tilde{S}_\mu (\Psi,\phi)
\leq &
\mathcal{K}(\phi,A_t)\left( \langle\hspace{-0.08cm}\langle\Psi,\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle+N^{-1/6} \ln(N) \right)
\;.
\end{align*}
\item Estimate of $M(\Psi, \phi)$.\\
First, we estimate \eqref{1 term in M}.
\begin{align*}
|\eqref{1 term in M}|\leq& 2|\langle\hspace{-0.08cm}\langle\nabla_1q_1\Psi ,
\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}
\nabla_1p_1\Psi \rangle\hspace{-0.08cm}\rangle|+2|\langle\hspace{-0.08cm}\langle\nabla_1q_1\Psi ,
\nabla_1p_1\Psi \rangle\hspace{-0.08cm}\rangle|
\\&\leq2\|\nabla_1q_1\Psi\|\;\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1p_1\|_{\text{op}}
+2|\langle\hspace{-0.08cm}\langle\widehat{n}^{-1/2}q_1\Psi ,
\Delta_1 p_1\widehat{n}_{1}^{1/2}\Psi \rangle\hspace{-0.08cm}\rangle|
\;.
\end{align*}
By Lemma \ref{hdetail}, we obtain
$\|\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1p_1\|_{\text{op}} \leq
C
\| \nabla \phi \|_\infty
N^{1/2-d}$.
Furthermore, we use
$
\| \nabla_1 q_1\Psi\| \leq \| \nabla_1 \Psi\|+\| \nabla_1 p_1\Psi\|\leq \mathcal{K}(\phi, A_t)
$ (see also Lemma \ref{kinenergyboundedlemma}) and
$
|\langle\hspace{-0.08cm}\langle\widehat{n}^{-1/2}q_1\Psi ,
\Delta_1 p_1\widehat{n}_{1}^{1/2}\Psi \rangle\hspace{-0.08cm}\rangle|
\leq
\mathcal{K}(\phi, A_t)
\|\widehat{n}_{1}^{1/2}\Psi \|
\|\widehat{n}^{1/2}\Psi \|
\leq
\mathcal{K}(\phi, A_t)
(
\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle+ N^{-1}
)
$.
Hence, for $d$ large enough,
\begin{align*}
|\eqref{1 term in M}|&\leq \mathcal{K}(\phi, A_t) (\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle+N^{\frac{1}{2} -d} + N^{-1})
\leq
\mathcal{K}(\phi,A_t) (\langle\hspace{-0.08cm}\langle\Psi,\widehat{n}\Psi\rangle\hspace{-0.08cm}\rangle+ N^{-1})
\;.
\end{align*}
Line~\eqref{2 term in M} is estimated for $d$ large enough, noting that
$ \| \nabla_1 p_1 \Psi \|^2=
\|\nabla\phi \|^2 \| p_1 \Psi \|^2$, by
\begin{align*}\eqref{2 term in M}=&
\|\mathds{1}_{\mathcal{A}^{(d)}_{1}}\nabla_1p_1\Psi \|^2- \|\nabla\phi \|^2 \\
\leq& | \| \nabla_1 p_1 \Psi \|^2- \|\nabla\phi \|^2 |
+ \|\mathds{1}_{\overline{\mathcal{A}^{(d)}_{1}}}\nabla_1p_1\Psi \|^2 \\
\leq& C \left( \| \nabla \varphi \|^2 \langle\hspace{-0.08cm}\langle \Psi , q_1\Psi \rangle\hspace{-0.08cm}\rangle +
\| \nabla \phi \|_\infty^2 N^{1 - 2d} \right)
\\
\leq &
\mathcal{K}(\phi,A_t) \langle\hspace{-0.08cm}\langle \Psi \widehat{n} , \Psi \rangle\hspace{-0.08cm}\rangle
%
\;.
\end{align*}
For line \eqref{3 term in M}, we use Lemma \ref{opdiff} to obtain
\begin{align*}
\eqref{3 term in M} \leq C\|\dot A_t\|_{\infty}
\left(
\langle\hspace{-0.08cm}\langle \Psi, \widehat{n} \Psi \rangle\hspace{-0.08cm}\rangle + N^{-1/2}
\right)
\;.
\end{align*}
In total, we obtain
\begin{align*}
M(\Psi, \phi)
\leq
\mathcal{K}(\phi,A_t)
\left(
\langle\hspace{-0.08cm}\langle \Psi, \widehat{n} \Psi \rangle\hspace{-0.08cm}\rangle + N^{-1/2}
\right)
\;.
\end{align*}
\end{itemize}
\end{proof}
\begin{lemma}\label{positiv V-w}
$ $\\
\begin{itemize}
\item[(a)]
Let $R_\beta$ and $M_\beta$ be defined as in Lemma \ref{microscopic}.
Then, for any $\Psi\in H^1(\mathbb{R}^{2N}, \mathbb{C})$
$$\|\mathds{1}_{|x_1-x_2|\leq R_{\beta}}\nabla_1\Psi\|^2+\frac{1}{2}\langle\hspace{-0.08cm}\langle\Psi,
(V_{N}-M_{\beta})(x_1-x_2)\Psi\rangle\hspace{-0.08cm}\rangle\geq0\;.$$
\item[(b)]
Let $M_\beta$ be defined as in Lemma \ref{microscopic}.
Let $\Psi\in
L_s^2(\mathbb{R}^{2N}, \mathbb{C}) \cap H^1(\mathbb{R}^{2N}, \mathbb{C})$.
Then, for sufficiently large $N$ and for $\beta>d$,
\begin{align*}
\|\mathds{1}_{\mathcal{B}^{(d)}_{1}}\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}}\nabla_1\Psi \|^2
+
\frac{1}{2}
\langle\hspace{-0.08cm}\langle\Psi ,\sum_{j\neq
1}\mathds{1}_{\mathcal{B}^{(d)}_{1}}\left(V_N-M_{\beta}\right)(x_1-x_j)\Psi \rangle\hspace{-0.08cm}\rangle
\geq 0
\;.
\end{align*}
\end{itemize}
\end{lemma}
\begin{proof}
\begin{itemize}
\item[(a)]
We first show nonnegativity of the one-particle operator
$H^{Z_n} : H^2 (\mathbb{R}^2, \mathbb{C}) \rightarrow L^ 2( \mathbb{R}^2, \mathbb{C})$ given by
$$H^{Z_n}=-\Delta+ \frac{1}{2} \sum_{z_k\in Z_n} (V_{N}(\cdot-z_k)-M_{\beta}(\cdot-z_k))$$
for any $n\in \mathbb{N}$ and any $n$-elemental subset $Z_n\subset\mathbb{R}^2$ which is
such that the supports of the potentials
$M_{\beta}(\cdot-z_k)$ are pairwise disjoint for any two $z_k\in
Z_n$.
Since $f_{\beta}(\cdot-z_k)$ is the the zero energy scattering state of the potential
$1/2 V_{N}(\cdot-z_k)- 1/2 W_{\beta}(\cdot-z_k)$, it follows that
$$F^{Z_n}_{\beta}=\prod_{z_k\in Z_n}f_{\beta}(\cdot-z_k)\;.$$ fulfills
$H^{Z_n} F^{Z_n}_{\beta}=0$
for any such $Z_n$.
By construction $f_{\beta}$ is a positive function, so is $F^{Z_n}_{\beta}$.
Since $\frac{1}{2} \sum_{z_k\in Z_n} (V_{N}(\cdot-z_k)-M_{\beta}(\cdot-z_k)) \in L^\infty (\mathbb{R}^2,\mathbb{C})$, this potential is a infinitesimal perturbation of $-\Delta$, thus
$\sigma_{\text{ess}}( H^{Z_n} ) =[0, \infty )$.
Assume now that $H^{Z_n}$ is not nonnegative. Then, there exists a ground state $\Psi_G\in H^2( \mathbb{R}^2,\mathbb{C})$ of $H^{Z_n}$ of
negative energy $E<0$. The phase of the ground state can be chosen such that the ground state is real and positive (see e.g. \cite{teschl}, Theorem 10.12.).
Since such a ground state of negative energy decays exponentially, that is
$\Psi_G (x) \leq C_1 e^{-C_2 |x|}, C_1,C_2 >0$ ,
the following scalar product is well defined (although $F^{Z_n}_{\beta} \notin L^2( \mathbb{R}^2, \mathbb{C})$).
\begin{align}
\label{contra}\langle F^{Z_n}_{\beta},H^{Z_n}\Psi_G\rangle=\langle F^{Z_n}_{\beta},E\Psi_G\rangle<0
\;.
\end{align}
On the other hand we have since $F^{X_n}_{\beta_1,\beta}$ is the zero energy scattering state
$$\langle
F^{Z_n}_{\beta},H^{Z_n}\Psi_G\rangle=\langle H^{Z_n}F^{Z_n}_{\beta},\Psi_G\rangle=0\;.$$
This contradicts (\ref{contra}) and the nonnegativity of $H^{Z_n}$ follows.
Now, assume that there exists a $\psi\in H^2 (\mathbb{R}^2,\mathbb{C}) $ such that the quadratic form
$$
Q (\psi)=
\|\mathds{1}_{|\cdot|\leq R_{\beta}}\nabla\psi\|^2+
\frac{1}{2}\langle\psi,(V_{N}(\cdot)-M_{\beta}(\cdot))\psi\rangle<0\;.$$
Since $V_{\beta_1}$ and $M_{\beta_1}$ are spherically symmetric we can assume that $\psi$ is spherically symmetric. Subsituting $ \psi \rightarrow a \psi ,\; a \in \mathbb{R}$ , we can furthermore assume that, for all $|x| = R_\beta$, $\psi(x) =
1- \epsilon$ for $\epsilon>0$.
Define $\tilde{\psi}$ such that
$ \tilde{\psi}(x)=\psi(x) \text{ for } |x| \leq R_{\beta}$ and $\tilde{\psi}(x)=1$ for $|x|> R_{\beta}+ \epsilon$ and $\epsilon>0$. Furthermore, $\tilde{\psi}$ can be constructed such that
$ \| \mathds{1}_{|x| \geq R_\beta} \nabla \tilde{\psi}\|^2 \leq C (\epsilon+ \epsilon^ 2)$.
Then $Q ( \tilde{\psi})= Q (\psi)<0$ holds, because the operator associated with the quadratic form is supported inside the ball $B_0 (R_{\beta})$.
Using $\tilde{\psi}$, we can construct a set of points $Z_n$ and a $\chi\in H^2(\mathbb{R}^2, \mathbb{C})$ such that $\langle\chi,H^{Z_n}\chi\rangle<0$, contradicting to nonnegativity of $H^{Z_n}$.
For $R>1$ let
$$\xi_R(x)=\left\{
\begin{array}{ll}
R^2/x^2, & \hbox{for $x>R$;} \\
1, & \hbox{else.}
\end{array}
\right.
$$
Let now $Z_n$ be a subset $Z_n\subset\mathbb{R}^3$ with $|Z_n|=n$ which is such that the supports of the potentials
$W_{\beta}(\cdot-z_k)$ lie within the Ball around zero with radius $R$ and are pairwise disjoint for any two $z_k\in Z_n$. Since we are in two dimensions we can choose a $n$ which is of order $R^2$.
Let now $\chi_R(x)=\xi_R (x) \prod_{z_k\in Z_n} \tilde{\psi}(x-z_k)$. By construction, there exists a $D = \mathcal{O}(1)$ such that
$ \chi_R(x)= \tilde{\psi}(x-z_k)$ for $ |x-z_k| \leq D $.
From this, we obtain
\begin{align*}
\langle\chi_R,H^{Z_n}\chi_R\rangle
=&
\| \nabla \chi_R \|^2+
n\frac{1}{2}\langle\psi,(V_{N}(\cdot)-M_{\beta}(\cdot))\psi\rangle
\\
=&
n Q(\psi)+
\sum_{z_k \in Z_n}
\| \mathds{1}_{|x-z_k|\geq R_\beta} \nabla \chi_R \|^2
\\
\leq &
n Q(\psi) +C n (\epsilon +\epsilon^2) + \| \nabla \xi_R \|^2
\\
=&
n Q(\psi) +C n (\epsilon+ \epsilon^2) +C
\;.
\end{align*}
Choosing $R$ and hence $n$ large enough and $\epsilon$ small, we can find a $Z_n$
such that
$\langle \chi_R,H^{Z_n}\chi_R\rangle$ is negative, contradicting nonnegativity of $H^{Z_n}$.
\\
Now, we can prove that
\begin{align}
\label{poslemmac}
\|\mathds{1}_{|x_1-x_2|\leq R_{\beta_1}}\nabla_1\Psi\|^2+\frac{1}{2}\langle\hspace{-0.08cm}\langle\Psi,
(V_{N}-M_{\beta})(x_1-x_2)\Psi\rangle\hspace{-0.08cm}\rangle\geq 0
\;.
\end{align}
holds for any $\Psi\in H^2(\mathbb{R}^{2N},\mathbb{C})$.
Using the coordinate transformation $\tilde{x}_1= x_1-x_2 \; , \tilde{x}_i=x_i \; \forall i \geq 2$, we have $\nabla_{x_1}= \nabla_{\tilde{x}_1}$. Thus \eqref{poslemmac} is equivalent to $
\|\mathds{1}_{|x_1|\leq R_{\beta_1}}\nabla_1\Psi\|^2+\frac{1}{2}\langle\hspace{-0.08cm}\langle\Psi,
(V_{N}-M_{\beta})(x_1)\Psi\rangle\hspace{-0.08cm}\rangle
\geq 0$ $\forall \Psi \in H^2(\mathbb{R}^{2N},\mathbb{C})$
which follows directly from $Q(\psi) \geq 0$ for all $\psi\in H^2(\mathbb{R}^{2},\mathbb{C})$.
By a standard density argument, we can conclude that $Q(\Psi) \geq 0$ $\forall \Psi \in H^1(\mathbb{R}^{2N},\mathbb{C})$.
\item[(b)]
Define
$c_k= \lbrace
(x_1,\dots,x_N) \in \mathbb{R}^{2N}| |x_1-x_k|\leq R_\beta \rbrace$ and
$\mathcal{C}_1= \cup_{k=2}^N c_k$.
For $(x_1, \dots, x_N) \in \mathcal{B}^{(d)}_{1}$ it holds that
$|x_i-x_j| \geq N^{-d}$ for $2 \leq i,j \leq N$.
Let $ \beta >d$. Assume that $N^{-d} > 2 R_\beta$, which hold for $N$ sufficiently large, since $R_\beta \leq CN^{-\beta}$.
Then, it follows that, for $i \neq j$,
$ \left( c_i \cap \mathcal{B}^{(d)}_{1} \right)
\cap
\left( c_j \cap \mathcal{B}^{(d)}_{1} \right)= \emptyset$.
Under the same conditions, we also have $\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}} \geq \mathds{1}_{\mathcal{C}_{1}}$.
Therefore
\begin{align*}
&
\mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}} \mathds{1}_{\mathcal{B}^{(d)}_{1}}
\geq
\mathds{1}_{\mathcal{C}_{1}} \mathds{1}_{\mathcal{B}^{(d)}_{1}}
= \mathds{1}_{\mathcal{C}_{1} \cap \mathcal{B}^{(d)}_{1} }
= \mathds{1}_{ \cup_{k=2}^N \left( c_k \cap \mathcal{B}^{(d)}_{1} \right) }
= \sum_{k=2}^N \mathds{1}_{c_k \cap \mathcal{B}^{(d)}_{1} }
=\mathds{1}_{ \mathcal{B}^{(d)}_{1} }
\sum_{k=2}^N \mathds{1}_{c_k }
\;.
\end{align*}
Note that $\mathds{1}_{\mathcal{B}^{(d)}_{1}}$ depends only on $x_2, \dots, x_N$. By this
\begin{align*}
\| \mathds{1}_{\overline{\mathcal{A}}^{(d)}_{1}} \mathds{1}_{\mathcal{B}^{(d)}_{1}}
\nabla_1 \Psi \|^2
\geq
\sum_{k=2}^N
\|
\mathds{1}_{c_k }
\nabla_1 \mathds{1}_{ \mathcal{B}^{(d)}_{1} } \Psi \|^2
=
(N-1)
\|
\mathds{1}_{|x_1-x_2| \leq R_\beta }
\nabla_1 \mathds{1}_{ \mathcal{B}^{(d)}_{1} } \Psi \|^2
\;.
\end{align*}
This yields
\begin{align*}
\eqref{vielezeilenb}
\geq
(N-1)
\left(
\|
\mathds{1}_{|x_1-x_2| \leq R_\beta }
\nabla_1 \mathds{1}_{ \mathcal{B}^{(d)}_{1} } \Psi \|^2
+
\frac{1}{2}\langle\hspace{-0.08cm}\langle \mathds{1}_{ \mathcal{B}^{(d)}_{1} } \Psi,
(V_{N}-M_{\beta})(x_1-x_2)\mathds{1}_{ \mathcal{B}^{(d)}_{1} } \Psi\rangle\hspace{-0.08cm}\rangle
\right)
\geq 0
\;.
\end{align*}
where the last inequality follows from (a)
\end{itemize}
\end{proof}
\begin{lemma}\label{energie(1-pp)pp
Let $W_\beta \in \mathcal{V}_\beta$.
Let
$
\Psi \in
L^2_{s}(\mathbb{R}^{2N}, \mathbb{C})
\cap
H^1(\mathbb{R}^{2N}, \mathbb{C})
$
and $\|\nabla_1\Psi\|$ be bounded uniformly in $N$.
Let $d $ in Definition \ref{hdetail} of $\mathds{1}_{{\mathcal{B}_1^{(d)}}} $ sufficiently large.
Let $\Gamma \in \lbrace \Psi, \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi \rbrace$.
Then, for all $\beta >0$,
\begin{enumerate}
\item
\begin{align*}
&N\left|\langle\hspace{-0.08cm}\langle \Gamma, q_1p_2
W_\beta(x_1-x_2) p_1p_2 \Gamma\rangle\hspace{-0.08cm}\rangle\right|\leq C \|\phi \|_\infty^2
\langle\hspace{-0.08cm}\langle \Psi, \hat{n} \Psi \rangle\hspace{-0.08cm}\rangle \;.
\end{align*}
\item \begin{align*}
&N|\langle\hspace{-0.08cm}\langle \Gamma ,
p_1p_2 W_\beta(x_1-x_2) q_1q_2 \Gamma\rangle\hspace{-0.08cm}\rangle| \leq
\mathcal{K}(\phi,A_t)
\left(
\langle\hspace{-0.08cm}\langle\Psi, \widehat{n} \Psi \rangle\hspace{-0.08cm}\rangle
+
N^{-1/6} \ln(N)
\right)
\;.
\end{align*}
\item[(c)]
\begin{align*}
&N|\langle\hspace{-0.08cm}\langle \Gamma ,
(1-p_1p_2) W_\beta(x_1-x_2) p_1p_2 \Gamma\rangle\hspace{-0.08cm}\rangle|
\leq
\mathcal{K}(\phi,A_t)
\left(
\langle\hspace{-0.08cm}\langle\Psi, \widehat{n} \Psi \rangle\hspace{-0.08cm}\rangle
+
N^{-1/6} \ln(N)
\right)
\;.
\end{align*}
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}
\item
Let first $\Gamma= \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi$.
Then,
\begin{align}
&N\left|\langle\hspace{-0.08cm}\langle\mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi, q_1p_2
W_\beta(x_1-x_2) p_1p_2 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\rangle\hspace{-0.08cm}\rangle\right|
\nonumber
\\
\leq&
\label{muchopequeno}
N\left|\langle\hspace{-0.08cm}\langle \mathds{1}_{{\overline{\mathcal{B}}^{(d)}_{1}}} \Psi, q_1p_2
W_\beta(x_1-x_2) p_1p_2 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\rangle\hspace{-0.08cm}\rangle\right| \\
\label{pequeno}
+& N\left|\langle\hspace{-0.08cm}\langle \Psi, q_1p_2
W_\beta(x_1-x_2) p_1p_2 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\rangle\hspace{-0.08cm}\rangle\right|
\;.
\end{align}
Using Lemma \ref{propo}
together with $\|p_2 W_\beta (x_1-x_2) p_2 \|_{\text{op}} \leq \|\phi\|_\infty^2 \|W_\beta\|_1$,
the first line can be bounded, for any $\epsilon>0$, by
\begin{align}
&
\eqref{muchopequeno}
\leq \mathcal{K}(\phi, A_t) N \| \mathds{1}_{{\overline{\mathcal{B}}^{(d)}_{1}}} \Psi \|
\| W_{\beta} \|_{1} \leq \mathcal{K}(\phi, A_t) N^{1- d + \epsilon}.
\end{align}
The second term is bounded by
\begin{align*}
\eqref{pequeno}
=& N\left|\langle\hspace{-0.08cm}\langle \sqrt{W_\beta(x_1-x_2)} q_1p_2 (\hat{n})^{-\frac{1}{2}} \Psi,
\sqrt{W_\beta(x_1-x_2)} p_1p_2 \hat{n}_{1}^{\frac{1}{2}} \mathds{1}_{{\mathcal{B}^{(d)}}_{1}} \Psi\rangle\hspace{-0.08cm}\rangle\right| \\
\leq& CN \|\sqrt{W_\beta(x_1-x_2)} p_2 \|_{\text{op}}^2
\left( \|q_1 (\hat{n})^{-\frac{1}{2}} \Psi \|^2
+ \| \hat{n}_1^{\frac{1}{2}} \mathds{1}_{{\mathcal{B}^{(d)}}_{1}} \Psi \|^2
\right) \\
\leq& CN \|\sqrt{W_\beta(x_1-x_2)} p_2 \|_{\text{op}}^2
\left( \langle\hspace{-0.08cm}\langle \Psi, \hat{n} \Psi \rangle\hspace{-0.08cm}\rangle
+ \| \hat{n}_1^{\frac{1}{2}} \Psi \|^2
+ \| \hat{n}_1^{\frac{1}{2}} \mathds{1}_{{\overline{\mathcal{B}}^{(d)}_{1}}} \Psi \|^2
\right) \\
\leq& CN \| W_\beta \|_1 \| \phi \|_\infty^2
\left( \langle\hspace{-0.08cm}\langle \Psi, \hat{n} \Psi \rangle\hspace{-0.08cm}\rangle + \| \mathds{1}_{{\overline{\mathcal{B}}^{(d)}_{1}}} \Psi \|^2
\right) \\
\leq& C \| \phi \|_\infty^2 \left( \langle\hspace{-0.08cm}\langle \Psi, \hat{n} \Psi \rangle\hspace{-0.08cm}\rangle + N^{1 - d + \epsilon} \right)
.
\end{align*}
Choosing $d$ large enough, $N^{1 - d + \epsilon}$ is smaller than $\langle\hspace{-0.08cm}\langle \Psi, \hat{n} \Psi \rangle\hspace{-0.08cm}\rangle$. This yields (a) in the case
$\Gamma= \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi$.
The inequality (a) can be proven analogously for $\Gamma= \Psi$.
\item
Let $\Gamma= \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi$.
We first consider (b) for potentials with $\beta< 1/4$.
We have to estimate
\begin{align}
&N|\langle\hspace{-0.08cm}\langle \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi ,p_1 p_2 W_\beta(x_1-x_2) q_1 q_2
\mathds{1}_{{\mathcal{B}^{(d)}_{1}}}\Psi
\rangle\hspace{-0.08cm}\rangle|
\leq
N|\langle\hspace{-0.08cm}\langle\Psi ,p_1 p_2 W_\beta(x_1-x_2) q_1 q_2
\Psi\rangle\hspace{-0.08cm}\rangle|
\nonumber \\
+&
N|\langle\hspace{-0.08cm}\langle \mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\Psi ,p_1 p_2 W_\beta(x_1-x_2) q_1 q_2
\Psi\rangle\hspace{-0.08cm}\rangle|
+
N|\langle\hspace{-0.08cm}\langle\Psi ,p_1 p_2 W_\beta(x_1-x_2) q_1 q_2
\mathds{1}_{\overline{\mathcal{B}^{(d)}_{1}}}\Psi\rangle\hspace{-0.08cm}\rangle|
\nonumber \\
+&
\label{betaklein}
N|\langle\hspace{-0.08cm}\langle \mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\Psi ,p_1 p_2 W_\beta(x_1-x_2) q_1 q_2
\mathds{1}_{\overline{\mathcal{B}^{(d)}_{1}}}\Psi\rangle\hspace{-0.08cm}\rangle|
\nonumber
\\
\leq &
N|\langle\hspace{-0.08cm}\langle \Psi ,p_1 p_2 W_\beta(x_1-x_2) q_1 q_2
\Psi\rangle\hspace{-0.08cm}\rangle|
\\
+&
\label{muchomuchoklein}
C N \|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\Psi\| \|W_\beta \|_\infty
\;.
\end{align}
The last term is bounded, for any $\epsilon>0$, by
\begin{align*}
\eqref{muchomuchoklein} \leq C N N^{1-d+ \epsilon} N^{-1+2 \beta}
\leq
N^{-2} \; ,
\end{align*}
where the last inequality holds choosing $d$ large enough.
Using Lemma \ref{kombinatorik} (c) and Lemma \ref{trick} with $O_{1,2}=q_2
W_\beta(x_1-x_2) p_2$, $\Omega=N^{-1/2}q_1\Psi$ and $\chi=N^{1/2}p_1\Psi$ we get
\begin{align*}
\eqref{betaklein}
&\leq \left\|q_1\Psi\right\|^2+
N^2\big|\langle\hspace{-0.08cm}\langle q_2\,\Psi,p_1\sqrt{W_\beta}(x_1-x_2)
p_3
\sqrt{W_\beta}(x_1-x_3)\\&\hspace{3.7cm}\sqrt{W_\beta}(x_1-x_2)p_2\sqrt{W_\beta}(x_1-x_3)p_1
q_3\,\Psi\rangle\hspace{-0.08cm}\rangle\big|
\\&+N^2(N-1)^{-1}\|q_2
W_\beta(x_1-x_2) p_2p_1\Psi\|^2&
\\&\leq \left\|q_1\Psi\right\|^2+N^2\|\sqrt{W_\beta}(x_1-x_2)p_1\|_{\text{op}}^4\;\|
q_2\,\Psi\|^2
\\&+C N\|
W_\beta(x_1-x_2) p_2\|_{\text{op}}^2
\;.
\end{align*}
With Lemma \ref{kombinatorik} (e) we get the bound
\begin{align*}
&
\eqref{betaklein}
\leq \|q_1\Psi\|^2+
N^2\|\phi \|_\infty^4\|W_\beta\|_1^2\;\|
q_1\Psi\|^2
\\&+C N\|W_\beta\|^2\|\phi \|_\infty^2
\;.
\end{align*}
Note, that $\|W_\beta\|_1\leq CN^{-1}$, $\|W_\beta\|^2\leq CN^{-2+2\beta}$
Hence
\begin{align*}
\eqref{betaklein}\leq C\left(\langle\hspace{-0.08cm}\langle\Psi,q_1\Psi\rangle\hspace{-0.08cm}\rangle+
\mathcal{K}(\phi)
N^{-1+2\beta}\right)
\;.
\end{align*}
Note that, for $\beta<1/4$, $ N^{-1+2\beta} \leq N^ {-1/6} \ln(N)$.
Using the same bounds for $\Gamma= \Psi$, we obtain (b) for the case $\beta<1/4$.
\item[b)] for $1/4\leq \beta$:\\
We use $U_{\beta_1,\beta}$ from Definition \ref{udef} for some $0<\beta_1<1/4$.
$Z^\phi_\beta(x_1,x_2)-W_{\beta}+U_{\beta_1,\beta}$ has the form of $Z^\phi_{\beta_1}(x_1,x_2)$ which has been controlled above. It is left to control $$
N\left|\langle\hspace{-0.08cm}\langle \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi, p_1p_2\left(W_{\beta}(x_1-x_2)-U_{\beta_1,\beta}(x_1-x_2)\right)q_1q_2 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\rangle\hspace{-0.08cm}\rangle\right|\;.
$$
Let $\Delta h_{\beta_1,\beta}=W_{\beta}-U_{\beta_1,\beta}$. Integrating by parts and using that\\ $\nabla_1 h_{\beta_1,\beta}(x_1-x_2)=-\nabla_2 h_{\beta_1,\beta}(x_1-x_2)$ gives
\begin{align}
\nonumber
&N\left|\langle\hspace{-0.08cm}\langle \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi, p_1p_2\left(W_{\beta}(x_1-x_2)-U_{\beta_1,\beta}(x_1-x_2)\right)q_1q_2 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\label{split1}&\hspace{2cm}=N\left|\langle\hspace{-0.08cm}\langle\nabla_1p_1 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi, p_2\nabla_2 h_{\beta_1,\beta}(x_1-x_2)q_1q_2 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\rangle\hspace{-0.08cm}\rangle\right|
\\\label{split2}
&\hspace{2.4cm}+N\left|\langle\hspace{-0.08cm}\langle \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi, p_1p_2\nabla_2 h_{\beta_1,\beta}(x_1-x_2)\nabla_1q_1q_2 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\rangle\hspace{-0.08cm}\rangle\right|
\;.
\end{align}
Let $ (a_1,b_1)=(q_1,\nabla p_1)$ or $ (a_1,b_1)=( \nabla q_1,p_1)$. Then, both terms can be estimated as follows:
\\
We use Lemma \ref{trick} with $\Omega=N^{-\eta/2} a_1\mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi$,
$O_{1,2}=N^{1+\eta/2}q_2\nabla_2 h_{\beta_1,\beta}(x_1-x_2)p_2$ and $\chi= b_1 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi$. We choose $ \eta < 2 \beta_1$.
\begin{align}\label{nulltesumme}
&N\left|\langle\hspace{-0.08cm}\langle \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi, a_1p_2\nabla_2 h_{\beta_1,\beta}(x_1-x_2)b_1q_2 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\rangle\hspace{-0.08cm}\rangle\right|
\nonumber
\\
&\leq N^{-\eta}\|a_1 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\|^2
\\&+
\label{erstesumme}\frac{N^{2+\eta}}{N-1}\|q_2\nabla_2 h_{\beta_1,\beta}(x_1-x_2)b_1 p_2 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\|^2
%
\\\label{zweitesumme}&+N^{2+\eta}\left|\langle\hspace{-0.08cm}\langle \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi, b_1 p_2 q_3\nabla_2 h_{\beta_1,\beta}(x_1-x_2)\nabla_3 h_{\beta_1,\beta}
(x_1-x_3)b_1 q_2p_3\mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\rangle\hspace{-0.08cm}\rangle\right|^{1/2}
\;.
\end{align}
We obtain (note that $\mathds{1}_{{\mathcal{B}_1^{(d)}}} $ does not depend on $x_1$)
$$(\ref{nulltesumme})\leq N^{-\eta}\|a_1 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\|^2
=
N^{-\eta}\| \mathds{1}_{{\mathcal{B}_1^{(d)}}} a_1\Psi\|^2
\leq
\mathcal{K}(\phi, A_t)
N^{-\eta}\;.$$
since both $ \| \nabla q_1 \Psi\|$ and $ \| q_1 \Psi\|$ are bounded uniformly in $N$.
Since $q_2$ is a projector it follows that
\begin{align*}
(\ref{erstesumme})
\leq&\frac{N^{2+\eta}}{N-1}\|\nabla_2 h_{\beta_1,\beta}(x_1-x_2)p_2\|_{\text{op}}^2
\|b_1 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\|^2
%
\leq
C
\frac{N^{2+\eta}}{N-1}
\|\phi\|_\infty ^2
\|\nabla h_{\beta_1,\beta}\|^2
\|b_1 \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\|^2
\\
\leq&
\mathcal{K}(\phi, A_t)
N^{\eta-1}\ln(N)
\|\phi\|_\infty ^2
\;,
\end{align*}
where we used Lemma \ref{ulemma} in the last step.
Next, we estimate
\begin{align}
(\ref{zweitesumme})\leq&
N^{2+\eta}\| p_2
\nabla_2 h_{\beta_1,\beta}(x_1-x_2)b_1 q_2
\mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi\|^2\
\nonumber
\\
\leq&
\label{ersterterm}
2 N^{2+\eta}\| p_2
\nabla_2 h_{\beta_1,\beta}(x_1-x_2)b_1 q_2
\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\Psi\|^2\
\\
+&
\label{zweiterterm}
2 N^{2+\eta}\| p_2
\nabla_2 h_{\beta_1,\beta}(x_1-x_2)b_1 q_2
\Psi\|^2
\;.
\end{align}
The first term can be estimated as
\begin{align*}
\eqref{ersterterm}
\leq &
C N^{2+ \eta} \|\nabla_2 h_{\beta_1,\beta}(x_1-x_2)b_1\|_{\text{op}}^2
\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\Psi\|^2
\\
\leq &
C N^{2+ \eta} \|\nabla_2 h_{\beta_1,\beta}\|^2 (\| \phi \|^2_\infty+\| \nabla \phi \|^2_\infty)
\|\mathds{1}_{\overline{\mathcal{B}}^{(d)}_{1}}\Psi\|^2
\\
\leq &
\mathcal{K}(\phi, A_t)
N^{2+ \eta} N^{-2} \ln(N) N^{2-2d+ 2 \epsilon}
=
\mathcal{K}(\phi, A_t)
N^{2-2d+2 \epsilon + \eta} \ln(N)
\;,
\end{align*}
for any $\epsilon >0$.
For $d$ large enough, this term is subleading.
The last term can be estimated as
\begin{align*}
\eqref{zweiterterm}
\leq&
2N^{2+\eta}\| p_2
h_{\beta_1,\beta}(x_1-x_2)b_1 \nabla_2q_2
\Psi\|^2
\nonumber
\\+&
2N^{2+\eta}\| |\varphi(x_2)\rangle \langle \nabla \varphi (x_2)|
h_{\beta_1,\beta}(x_1-x_2)b_1 q_2
\Psi\|^2\
\nonumber
\\
\leq &
C N^{2+\eta}\| p_2
h_{\beta_1,\beta}(x_1-x_2)
\|_{\text{op}}^2
\| b_1\nabla_2q_2
\Psi\|^2\
\nonumber
\\
+&
C N^{2+\eta}
\| |\varphi(x_2)\rangle \langle \nabla \varphi(x_2)|
h_{\beta_1,\beta}(x_1-x_2) \|_{\text{op}}^2
\| b_1 q_2
\Psi\|^2\
\nonumber
\\
\leq &
C N^{2+\eta} \left( \|\nabla \varphi \|_\infty^2+ \| \varphi \|_\infty^2 \right)
\| h_{\beta_1,\beta} \|^2 (1+ \| \nabla \phi \|^2)
\nonumber
\\ \leq&
\mathcal{K}(\phi, A_t)
N^{\eta- 2\beta_1} \ln(N)^2
\;.
\end{align*}
Combining both estimates we obtain, for any $\beta>1$,
\begin{align*}
&N\left|\langle\hspace{-0.08cm}\langle \mathds{1}_{{\mathcal{B}_1^{(d)}}} \Psi, p_1p_2W_{\beta}(x_1-x_2)q_1q_2 \mathds{1}_{{\mathcal{B}^{(d)}}_{1}}\Psi\rangle\hspace{-0.08cm}\rangle\right|
\\
&\leq
\inf_{ \eta>0}
\inf_{ 0<\mu <1/4}
\left(
\mathcal{K}(\phi, A_t)
\left(
\langle\hspace{-0.08cm}\langle\Psi, \widehat{n} \Psi \rangle\hspace{-0.08cm}\rangle
+
N^{-1 + 2 \mu}
+
N^{- \eta}
+
N^{\eta-1} \ln(N)
+
N^{\eta -2 \mu} \ln(N)
\right)
\right)
\\
&\leq
\mathcal{K}(\phi, A_t)
\left(
\langle\hspace{-0.08cm}\langle\Psi, \widehat{n} \Psi \rangle\hspace{-0.08cm}\rangle
+
N^{-1/6} \ln(N)
\right)
\;.
\end{align*}
where the last inequality comes from choosing $\eta=1/3$ and $\mu=1/4$.
For $\Gamma= \Psi$, (b) can be estimated the same way, yielding the same bound.
\item[(c)]
This follows from (a) and (b), using that $1-p_1p_2= q_1q_2+p_1q_2+q_1p_2$.
\end{enumerate}
\end{proof}
\section*{Acknowledgments}
We are grateful to David Mitrouskas for many valuable discussions and would like to thank Serena Cenatiempo for helpful discussions.
M.J. gratefully acknowledges financial support by the German National Academic Foundation.
N.L. gratefully acknowledges financial support from Cusanuswerk.
The article is part of one of the author’s (M.J.) forthcoming PhD thesis.
|
\section{Introduction}
After the discovery of the top quark at the Tevatron in 1995, its properties have been extensively measured ever since.
The LHC then overtook the top-quark physics program after its commissoning in 2009.
In most cases the LHC Run 1 results obtained at $\sqrt{s}=7,8\text{~TeV}$ already improve the Tevatron results. However, with the
start of Run 2 at the center-of-mass energy $\sqrt{s}=13\text{~TeV}$, the precision era for top-quark physics has begun. Important measurements are
the determination of the top-quark coupling to gauge bosons and the Higgs boson, spin-correlations of the top-quark decay products as well as
fiducial cross sections and differential distributions.
For now, the most advanced calculations for stable top-quark pair production are the NNLO+NNLL prediction~\cite{Czakon:2013goa} of the total
cross section and NNLO accurate prediction for differential distributions~\cite{Czakon:2015owf,Czakon:2016dgf}. However, top-quarks are extremely
unstable particles, and therefore decay before they can hadronize. Thus, top-quarks can be studied directly from its decay products, where in the
Standard Model they nearly in all cases decay into a bottom quark and a $W$ boson. Depending on the further decay of the $W$ boson one can
classify the decay modes into leptonic, $W\to \ell\nu_\ell$, or hadronic, $W\to q\bar{q}^\prime$, decays. The cleanest signature is obtained if
both top quarks decay leptonically and therefore the final-state consist out of two bottom-quark induced jets, two oppositely charged leptons and
missing energy. Top-quark decays were included in the NLO calculations within the narrow width approximation in the works of Refs.~\cite
{Melnikov:2009dn,Campbell:2012uf}. Another approach was taken in Refs~\cite{Denner:2010jp,Bevilacqua:2010qb,Denner:2012yc,Cascioli:2013wga,Frederix:2013gra,Heinrich:2013qaa}
where top-quarks are not treated as on-shell particles anymore. Instead the leptonic decays have been included and NLO QCD corrections were
calculated for the complete $pp\to e^+\nu_e \mu^- \bar{\nu}_\mu b\bar{b} +X$ process. Just recently also electroweak corrections were presented
in Ref.~\cite{Denner:2016jyo} and a consistent matching to parton showers has been presented in Ref.~\cite{Jezo:2016ujg}.
At the very large energies of the Large Hadron Collider top-quark pairs are often observed accompanied by additional jets and gauge bosons.
In Fig.~\ref{Fig:propaganda} the total cross sections for the associated production processes of top-quark pairs are shown as a function of the
center-of-mass energy $\sqrt{s}$.
\begin{figure}
\begin{center}
\includegraphics[width=0.65\textwidth]{propaganda.eps}
\captionit{The total cross sections for $t\bar{t}X$ processes as a function of $\sqrt{s}$.}
\label{Fig:propaganda}
\end{center}
\end{figure}
It is clear, that the $t\bar{t}j$ production is the largest contribution to the inclusive $t\bar{t}$ sample for moderate values of the $p_T$ cut
on the hardest jet.
For a better estimate of this contribution, we show in Table~\ref{Tab:Xsecs} the NLO QCD cross sections for several different
choices of the $p_T$ cut and the ratio to the NNLO $t\bar{t}$ cross section.
\begin{table}[h]
\begin{center}
\begin{tabular} {|c|r|c|}
\hline
$p_T(j_1)~[\text{~GeV}]$ & $\sigma_{t\bar{t}j}^{NLO}~[pb]$ & $\sigma_{t\bar{t}j}^{NLO}/\sigma_{t\bar{t}}^{NNLO}~[\%]$ \\
\hline
$40$ & $296.97 \pm 0.29$ & $37$ \\
$60$ & $207.88 \pm 0.19$ & $26$ \\
$80$ & $152.89 \pm 0.13$ & $19$ \\
$100$ & $115.60 \pm 0.14$ & $14$ \\
$120$ & $89.05 \pm 0.10$ & $11$ \\
\hline
\end{tabular}
\captionit{The total cross section of $t\bar{t}j$ at NLO as a function of the jet $p_T$ cut.}
\label{Tab:Xsecs}
\end{center}
\end{table}
The values were obtained for $m_t=173.2\text{~GeV}$ for the LHC at $\sqrt{s}=13\text{~TeV}$ using the CT14 PDF sets~\cite{Dulat:2015mca}. NLO results were
calculated using the \textsc{Helac-Nlo}{} framework~\cite{Bevilacqua:2011xh} while the NNLO prediction were obtained using \textsc{Top++}{}~\cite{Czakon:2011xx}.
For a cut of $p_T(j_1)>40\text{~GeV}$ nearly $40\%$ of all $t\bar{t}$ pairs are accompanied by an additional hard jet. Consequently, precise theoretical
predictions for the $t\bar{t}j$ production are necessary.
The process $pp\to t\bar{t}j + X$ has several important applications. First of all, the process can be utilitized in the determination of the
top-quark mass parameter using an alternative method~\cite{Alioli:2013mxa} which has been already used by both collaborations~\cite{Aad:2015waa,CMS:2016khu}.
On the other hand, $t\bar{t}j$ can be used to test physics beyond the Standard Model, for example, models with anomalous couplings to the top
quark~\cite{Gresham:2011dg} would lead to resonances in the invariant mass spectrum of the top-quark and the additional jet. Finally, $t\bar{t}j$
constitutes the dominant background for Higgs boson production in the vector boson fusion process (VBF), $pp\to Hjj \to W^+W^-jj$, if the Higgs
decays into a pair of $W$ bosons~\cite{Rainwater:1999sd,Kauer:2000hi}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{hist_yj1.eps}
\includegraphics[width=0.48\textwidth]{hist_yb1.eps} \\
\captionit{Differential distributions for the rapidity of the hardest light jet (left) and b-jet (right).}
\label{Fig:rapidity}
\end{center}
\end{figure}
For the VBF process the dominant kinematic signature are two forward-backward jets, which are required to be well separated in rapidity, i.e. $|y_1 - y_2| > 4$
and at the same time lie in opposite hemispheres of the detector, $y_1\times y_2 < 0$. In Fig.~\ref{Fig:rapidity} the rapidity of the hardest light and
bottom-quark induced jet are depicted. In the case of top-quark pair production the two leading jets are the b-jets from the decay and as can be
seen from Fig.~\ref{Fig:rapidity} these are produced very centrally in rapidity. Thus, top-quark pair production is strongly suppressed by the
aforementioned selection cuts. On the other hand, if the top-quark pair is accompanied by an additional jet then the suppression due to phase space
cuts are less severe than in the case of $t\bar{t}$ production. The reason is that the light jet can be one of the leading two jets and in
addition it is produced very broad in rapidity as can be seen from Fig.~\ref{Fig:rapidity}. Thus this light jet circumvents the strong
suppression and makes $t\bar{t}j$ production the dominant background.
\begin{figure}
\begin{center}
\includegraphics[width=\textwidth]{fd.eps}
\captionit{Differential distributions for the rapidity of the hardest light jet (left) and b-jet (right).}
\label{Fig:fd}
\end{center}
\end{figure}
In conclusion, precise theoretical predictions for $pp\to t\bar{t}j+X$ are mandatory. For stable top-quarks the NLO QCD corrections have been
presented in Refs.~\cite{Dittmaier:2007wz,Dittmaier:2008uj} and were later supplemented by LO decays~\cite{Melnikov:2010iu} in the narrow width
approximation. Later the top-quark decays have been extended to NLO accuracy~\cite{Melnikov:2011qx}. Top-quark pair production in association
with one hard jet has also been matched to parton showers by several groups~\cite{Kardos:2011qa,Alioli:2011as,Czakon:2015cla}.
Recently, also the approximation of on-shell top-quark decays has been dropped and the calculation of the full process $pp\to e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}j + X${} has been presented
in Ref.~\cite{Bevilacqua:2015qha}. Representative Feynman diagrams for the leading-order production at $\mathcal{O}(\alpha_s^3\alpha^4)$ are
shown in Fig.~\ref{Fig:fd}. This calculation takes into account all double, single and non-resonant top-quark contributions as well as
interferences between them. Additionally, top quarks and $W$, and $Z$ bosons are treated as unstable intermediate particles. Off-shell effects
for the top quark are small for inclusive observables and of the order of $\Gamma_t/m_t \approx 1\%$, as has been found in various calculations~
\cite{Denner:2010jp,Bevilacqua:2010qb,Denner:2012yc,Frederix:2013gra,Cascioli:2013wga,Heinrich:2013qaa,Denner:2015yca,Bevilacqua:2015qha}.
However, for certain observables off-shell effects can have a larger impact and can affect searches for the Higgs boson~\cite{AlcarazMaestre:2012vp} .
\section{Calculational framework}
The calculation of the next-to-leading order QCD corrections for the process\\ $pp\to e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}j + X${} is performed using the \textsc{Helac-Nlo}{} framework, which
has been already successfully utilized for a wide range of top-quark studies~\cite{Bevilacqua:2009zn,Bevilacqua:2010ve,Bevilacqua:2011aa,Worek:2011rd,
Bevilacqua:2012em,Bevilacqua:2014qfa}. The \textsc{Helac-Nlo}{} framework extends the \textsc{Helac-Phegas}{} Monte-Carlo generator~\cite{Kanaki:2000ey,Cafarella:2007pc} by
the capability of calculating NLO QCD corrections. It consists of two main packages:
\textsc{Helac-Dipoles}{}~\cite{Czakon:2009ss} for the evaluation of the real
corrections, where currently two subtraction schemes are available, namely the Catani-Seymour~\cite{Catani:1996vz,Catani:2002hc} and the
Nagy-Soper~\cite{Bevilacqua:2013iha} subtraction scheme. The complicated phase space integrations are performed using \textsc{Kaleu}{}~\cite{vanHameren:2010gg}.
The \textsc{Helac-1Loop}{} package~\cite{vanHameren:2009dr} on the other side deals with the automated computation of the virtual corrections. The program
is based on the so-called OPP reduction method~\cite{Ossola:2006us,Ossola:2008xq,Draggiotis:2009yb} as implemented in \textsc{CutTools}{}~\cite{Ossola:2007ax}
together with Dyson-Schwinger recursions for the computation of tree-level diagrams. Scalar one-loop integrals are computed using the \textsc{OneLOop}{}~\cite{vanHameren:2010cp}
library.
Treating the top quark as an intermediate particle gives rise to singularities due to on-shell propagators. To regularize resonant propagators the
complex-mass scheme~\cite{Denner:2005fg} is employed. Here, the mass of the top quark is replaced by $\mu_t^2 = m_t^2 -im_t\Gamma_t$, where $\Gamma_t$ is
the top-quark decay width.
\section{Numerical results}
The following Standard Model input parameters are used to obtain the numerical results presented in this section
\begin{equation}
\begin{split}
G_F &= 1.16637\cdot 10^{-5}\text{~GeV}^{-2}\;, \qquad m_t = 173.3\text{~GeV}\;,\\
m_W &= 80.399\text{~GeV}\;, \qquad \qquad\quad \Gamma_W = 2.09774\text{~GeV}\;, \\
m_Z &= 91.1876\text{~GeV}\;, \qquad\qquad\quad \Gamma_Z = 2.50966\;, \\
\Gamma_t^{LO} &= 1.48132\text{~GeV}\;, \qquad\qquad \Gamma_t^{NLO} = 1.3542\text{~GeV}\;.
\end{split}
\end{equation}
Predictions are given for the LHC working at a center-of-mass energy of $\sqrt{s}=8\text{~TeV}$. The parton distribution functions togehter with the
corresponding one or two-loop running of $\alpha_s$ are taken from the \textsc{MSTW2008LO68cl} and \textsc{MSTW2008NLO68cl} PDF sets~\cite{Martin:2009iq} repsectively.
The factorization and renormalization scales are fixed to a common value of $\mu_R=\mu_F=m_t$. Bottom-quark induced scattering channels are neglected due to the strong
suppression of the PDFs. Jets are defined using the anti-$k_T$ jet-algorithm~\cite{Cacciari:2008gp} with a separation parameter of $R=0.5$, where only partons with pseudorapidity
$|\eta| < 5$ are considered by the jet-algorithm.
The final state is required to consist of two charged leptons, exactly two bottom-quark induced jets, at least one light jet and missing transverse momentum, where
additional phase space selection cuts are also imposed. For both, bottom-quark induced and light jets, commonly denoted with $j$ in the following, we require
\begin{equation}
p_T(j) > 40\text{~GeV}\;, \quad |y(j)| < 2.5\;, \quad \Delta R_{jj} > 0.5\;.
\end{equation}
While for the leptonic final-state particles we require
\begin{equation}
p_T(\ell) > 30\text{~GeV}\;, \quad |y(\ell)| < 2.5\;, \quad \Delta R_{\ell j} > 0.4\;, \quad \Delta R_{\ell\ell} > 0.4\;, \quad p_T^{miss}>40\text{~GeV}\;.
\end{equation}
The cross section corresponding to the aforementioned setup is depicted in Fig.~\ref{Fig:scaledep} as a function of the renormalization and factorization scales.
The central prediction including the theoretical uncertainty, obtained by scale variation by
a factor of two, is given by
\begin{equation}
\sigma^{LO} = 183.1^{+112.2~(+61\%)}_{-64.2~(-35\%)}~\text{fb}\;, \qquad \sigma^{NLO} = 159.7^{-33.1~(-21\%)}_{-7.9~(-5\%)}~\text{fb}\;.
\end{equation}
The NLO corrections induce a strong reduction of the residual scale dependence which can be seen in Fig.~\ref{Fig:scaledep}. For the total cross section the theoretical uncertainty
is reduced from $61\%$ at LO down to $21\%$ at NLO, by nearly a factor of three. At the same time, the higher-order corrections amount to a reduction of the total cross section
by $-13\%$.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{scale_dep.eps}
\captionit{The total cross section as a function of the common renormalization and factorization scale $\mu_R=\mu_F=m_t$.}
\label{Fig:scaledep}
\end{center}
\end{figure}
Now turning to differential distributions; in Fig.~\ref{Fig:diffs} three pairs of differential distributions are depicted which pairwise share common features. In each plot the
leading-order (blue-dashed) and the next-to-leading order (red-solid) prediction is shown together with their corresponding theory uncertainties. The latter are obtained by varrying the
scale and choosing bin-by-bin the minimum and maximum value from the following set $\{m_t/2, m_t, 2m_t\}$. The lower panel depicts the differential K-factor.
The first pair of distributions to be discussed are the transverse momentum of the hardest jet $p_{T, j_1}$, and of the hardest bottom-quark induced jet $p_{T, b_1}$.
Both observables are dimensionful, where for example the beginning of the spectra are sensitive to the $t\bar{t}j$ threshold, while at the end of the plotted range the dominant
phase space regions are very hard jet configurations. For the transverse momentum of the hardest jet corrections are non uniform over the whole range and the LO shape can not be
rescaled by a constant factor. Furthermore, corrections to the shape of the order of $-50\%$ are found towards the end of the plotted range, while they are moderate at the beginning
of the spectrum.
For the transverse momentum of the hardest b-jet $p_{T, b_1}$ similar corrections have been found. Also here the LO shape is not rescaled by an overall factor and
shape distortions up to $-50\%$ in the tail are found. The common feature of both observables is that the
perturbative convergence is bad, which can be seen from the fact that in the tails the NLO predictions lie outside of the LO uncertainty band. Thus the LO uncertainty gives an
unreliable estimation of the missing higher-order corrections. To resolve this issue dynamic factorization and renormalization scales have to be chosen, which take into account
the transverse momentum of the leading jet.
The next two distributions are the rapidity of the hardest jet, $y_{j_1}$ and the angular separation of the leptons, $\Delta R_{\ell\ell}$.
These two observables represent angular distributions, which should converge faster than dimensionful observables, since the threshold region contributes to the whole range.
For the rapidity we find moderate and negative corrections to the overall normalization of the distribution. Additionally, the convergence in this particular distribution is
excellent.
The angular separation of the lepton pair is given by
\begin{equation}
\Delta R_{\ell\ell} = \sqrt{\Delta\Phi_{\ell\ell}^2 + \Delta y_{\ell\ell}^2}\;,
\end{equation}
where $\Delta\Phi_{\ell\ell}$ is the difference in the azimuthal angles and $\Delta y_{\ell\ell}$ the difference in rapidities of the leptons.
The depicted distribution peaks approximately at $\Delta R_{\ell\ell} = 3$, which means that the leptons are produced mostly back-to-back.
The QCD corrections are quite flat over the whole region with the exception of the peak-region, where moderate negative corrections are found.
The remaining two observables to be discussed are the reconstructed invariant top-quark mass and the invariant mass of the positron and the b-jet.
Both observables highlight the included off-shell effects in this calculation. For instance, the reconstruction of the invariant top-quark mass
by using $M_t = \sqrt{p_t^2}$, where the top four-momentum is given by $p_t=p_b+p_\ell+p_\nu$, shows a clear Breit-Wigner shape. The NLO
corrections introduce large radiative corrections below the peak, which are generated by further radiation from the bottom quark which is not
recombined with its mother particle by the jet-algorithm.
The last observable, $M_{be^+}$, is phenomenologically important for top-quark mass parameter extraction due to its high sensitivity on the top-quark mass.
At the same time, the $M_{be^+}$ spectrum shows features from off-shell effects. The observable is defined by using the $b$-jet that minimizes the
invariant mass with the positron. If the same spectrum would be calculated at LO using the narrow width approximation then the phase space of the
top decay products is more constrained. For example, for on-shell top quarks and $W$ bosons, the invariant mass $M_{be^+}$ exhibits an upper boundary given by
\begin{equation}
M_{be^+} = \sqrt{m_t^2-m_W^2}\approx 153\text{~GeV}\;.
\end{equation}
Only off-shell effects and further radiation give access to the phase space region above $153\text{~GeV}$. For the presented calculation we see that already
the LO prediction exceeds this boundary and NLO corrections are large in this region. The off-shell contributions to the total cross section are small.
However, they are important in the determination of the endpoint of the distribution because they smear this sharp boundary. The knowledge of the endpoint
allows to infer the top-quark mass~\cite{Chatrchyan:2013boa,Aad:2015nba}.
\begin{figure}
\begin{center}
\includegraphics[width=0.48\textwidth]{hist_pTj1.eps}
\includegraphics[width=0.48\textwidth]{hist_pTb1.eps} \\
\includegraphics[width=0.48\textwidth]{hist_yj1.eps}
\includegraphics[width=0.48\textwidth]{hist_dRll.eps} \\
\includegraphics[width=0.48\textwidth]{hist_Mt.eps}
\includegraphics[width=0.48\textwidth]{hist_Mbl.eps}
\captionit{Differential distributions as a function of the transverse momentum of the hardest jet $p_{T\;j_1}$, and of the hardest bottom-quark induced jet $p_{T\;,b_1}$, the rapidity
of the hardest jet $y_{j_1}$, the angular separation $\Delta R_{\ell\ell}$ between the leptons, the averaged reconstructed invariant mass of the top quark and the invariant mass
of the positron and the b-jet for $pp\to e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}j + X${} for the LHC at $\sqrt{s}=8\text{~TeV}$. The uncertainty bands are obtained by scale variation by a factor of two. The lower panel shows the
ratio of NLO to LO prediction.}
\label{Fig:diffs}
\end{center}
\end{figure}
\section{Conclusions}
In this proceeding we reported the first calculation of NLO QCD corrections for\\ $pp\to e^+\nu_e \mu^-\bar{\nu}_\mu b\bar{b}j + X${}, where $W, Z$ and top quark off-shell effects have been
taken into account. The NLO QCD corrections turn out to be moderate for the total cross section as well as for differential distributions.
For instance, the total cross section is reduced by $-13\%$, while the remaining scale uncertainty is at the level of $20\%$.
The study of this process will be extended in the future by a detailed comparison of fixed and dynamic renormalization and factorization scales.
Also a comparison between the four and the five-flavour scheme is desirable, since the uncertainties connected to the chosen scheme can reach up
to $10\%$, as reported in Ref.~\cite{Bevilacqua:2013taa}. Finally, the calculation can be used to improve predictions for the top-quark mass parameter
extraction using an alternative method~\cite{Alioli:2013mxa,Aad:2015waa}.
\textbf{Acknowledgement:} This work has been supported by the DFG under the Grant No. WO 1900/1-1 --
\textit{Signals and Backgrounds Beyond Leading Order. Phenomenological studies for the LHC}.
|
\chapter*{Abstract}
Let $S=K[x_1,...,x_n]$ or $S=K[\![x_1,...,x_n]\!]$ be either a
polynomial or a formal power series ring in a finite number of variables
over a field $K$ of characteristic $p > 0$ with $[K:K^p] < \infty$. Let $R$ be the hypersurface $S/fS$ where $f$ is a nonzero nonunit element of $S$. If $e$ is a positive integer, $F_*^e(R)$ denotes the $R$-algebra structure induced on $R$ via the $e$-times iterated Frobenius map ( $r\rightarrow r^{p^e}$ ). We describe a matrix factorizations of $f$ whose cokernel is isomorphic to $F_*^e(R)$ as $R$-module. The presentation of $F_*^e(R)$ as the cokernel of a matrix factorization of $f$ enables us to find a characterization from which we can decide when the ring $S[\![u,v]\!]/(f+uv)$ has finite F-representation type (FFRT) where $S=K[\![x_1,...,x_n]\!]$. This allows us to create a class of rings that have finite F-representation type but not finite CM type. For $S=K[\![x_1,...,x_n]\!]$, we use this presentation to show that the ring $S[\![y]\!]/(y^{p^d} +f)$ has finite F-representation type for any $f$ in $S$. Furthermore, we prove that $S/I$ has finite F-representation type when $I$ is a monomial ideal in either $S=K[x_1,...,x_n]$ or $S=K[\![x_1,...,x_n]\!]$. Finally, this presentation enables us to compute the F-signature of the rings $S[\![u,v]\!]/(f+uv)$ and $S[\![z]\!]/(f+z^2)$ where $S=K[\![x_1,...,x_n]\!]$ and $f$ is a monomial in the ring $S$. When $R$ is a Noetherian ring of prime characteristic that has FFRT, we prove that $R[x_1,...,x_n]$ and $R[\![x_1,...,x_n]\!]$ have FFRT.
We prove also that over local ring of prime characteristic a module has FFRT if and only it has FFRT by a FFRT system. This enables us to show that if $M$ is a finitely generated module over Noetherian ring $R$ of prime characteristic $p$, then the set of all prime ideals $Q$ such that $M_Q$ has FFRT over $R_Q$ is an open set in the Zariski topology on $\Spec(R)$.
\newpage
\chapter*{Acknowledgments}
First of all, I praise Allah (God), the almighty, merciful and passionate, for granting me the capability and providing me this opportunity to complete my PhD studies.
I owe my deepest gratitude to my PhD supervisor, Dr Moty Katzman. Without
his inspiring guidance, his unselfish help, his encouragements, his valuable advice, this thesis would
not have been possible. I have been extremely blessed to have
a supervisor who has tended to my work greatly and responded to my questions and enquiries so
promptly. I would like to express my gratefulness and appreciation for his patience and kindness during my research which has made my PhD experience an enjoyable one. He is an example that I will strive to follow in my
career.
I would like to thank the members of my viva examination Prof. Holger Brenner and Prof. Tom Bridgland for offering their time reading and reviewing this thesis and for their suggestions and helpful discussions during and after my viva.
It is beyond words to express my deep sense of gratitude to my father Abdullah, and my mother Obaidah, for their immeasurable love, continuous support and countless prayers that stand behind me in my life throughout. I am also so thankful for my brothers and sisters for their encouragements and help. Their love, assistance and confidence in me have sustained and inspired me throughout.
I would like to express my sincerest gratitude to my beloved wife, Hanan Alqaisi, for her gracious sacrifices and unconditional love. Her strong encouragement and unselfish support have been the essential reasons and the powerful source of energy and inspiration behind most of the successful achievements I have had so far. I truly believe that without her, I would not be able to achieve or even plan for my goals. There is no way to express properly how much I appreciate what she has done for me and I will forever be indebted to her for her support and effort. I thank God for her presence in my life. I also thank my wonderful children: Abdullah, Abdulaziz, Faisal, and Renad, for always making me smile and for understanding on those days when I was studying instead of playing with them. Thanks for putting up with me.
A special thanks goes to Prof. Salah-Eddine Kabbaj, Prof. Abdeslam Mimouni and Prof. Jawad Abuihlail for the courses in commutative algebra I took with them at King Fahd University of Petroleum and Minerals of Saudi Arabia. Thanks for their encouragement for me to pursue my PhD. I would like to also thank Prof. Craig Huneke for the other courses in commutative algebra that I took with him at the University of Kansas while I was studying my master degree.
I am very grateful to King Khalid University of Saudi Arabia for funding my studies abroad. Without their financial support, this would not have been accomplished.
Last but not least, I am so grateful for my best friend Nabil Alhakamy for his constant support and encouragement throughout my journey. I would always remember his supportive friendship. I must also thank my classmate Mehmet Yesil for being a nice friend. I would like to thank my friends in Sheffield for the treasuring moments we shared together. I find myself lucky of having them during my stay in Sheffield.
\markright{}
\tableofcontents
\newpage
\chapter{Introduction}
In this chapter, we give an overview of the results in this thesis. The next three chapters will introduce the necessary concepts and any unexplained terminology in this chapter will be explained there.
\section{Notation}
\label{section:Notation}
Throughout this thesis, we shall assume all rings are commutative with a unit unless otherwise is stated. $\mathbb{N}$ denotes the set of the positive integers and $\mathbb{Z}_{+}$ denotes the set of non-negative integers. Let $R$ be a ring. If $I$ is an ideal of $R$ generated by the elements $r_1,...,r_n$, we write $I=(r_1,...,r_n)$. Similarly, if $M$ is a finitely generated $R$-module generated by $m_1,...,m_n$, we write $M=(m_1,...,m_n)$. $R$ is called a Noetherian ring if every ideal of $R$ is finitely generated. The set of all prime ideals of a $R$ is denoted by $\Spec(R)$. $R$ is a local ring if $R$ has only one maximal ideal, and we write $(R,\mathfrak{m})$ to mean that $R$ is a local ring with the unique maximal ideal $\mathfrak{m}$. If $I$ is an ideal of $R$, we define $V(I)=\{ Q \in \Spec(R) \; | \; I \subseteq Q \}$ and the radical of $I$, denoted $\sqrt{I}$, is intersection of all prime ideals in $V(I)$. If $\sqrt{I}=Q$ for some prime ideal $Q$, $I$ is said to be $Q$-primary. The collection $\{ V(I) \; | \; I \text{ is an ideal of } R \}$ defines a topology on $\Spec(R)$ that is called the Zariski topology on $\Spec(R)$ in which $V(I)$ is a closed set for any ideal $I$ of $R$. A chain in $R$ or $\Spec(R)$ is a sequence $Q_0 \subseteq Q_1 \subseteq ... \subseteq Q_n$ of
prime ideals of $R$. If $Q$ is a prime ideal, the height of $Q$ denoted $\height(Q)$ is defined as
$$\height(Q)= \Sup \{ n | \text{ there exists a chain } Q_0 \subsetneqq Q_1 \subsetneqq ... \subsetneqq Q_n=Q \text{ in } \Spec(R)\} $$
and consequently the Krull dimension or simply the dimension of $R$, denoted $\dim(R)$, is given by
$$\dim(R)= \Sup\{ \height(\mathfrak{m}) | \mathfrak{m} \text{ is a maximal ideal in } R \} $$
Furthermore,the dimension of an $R$-module $M$, denoted $\dim_RM$ or $\dim M$, is the dimension of the ring $R/\Ann M$ where $\Ann M$ is the ideal $\Ann M = \{r \in R | rM=0 \} $.
The characteristic of $R$, denoted $\charact(R)$, is the smallest positive integer $n$ such that $\sum_{i=1}^n 1 = 0$; if no such $n$ exists, then the characteristic is zero.
\section{Outline of Thesis}
\label{section:Outline of Thesis}
In Chapter \ref{chapter:Preliminaries}, we gather in section \ref{section:General Background} the necessary concepts making the general background of this thesis. This includes the definitions and the theorems that we need in the subsequent chapters. We state in this section the theorems, propositions and lemmas that have references without proof. In section \ref{section:Technical Lemmas} of this chapter, we prove technical lemmas that are essential for some of the main results in the chapters \ref{chapter:The finite F-representation type over hypersurfaces} and \ref{chapter:F-signature of specific hypersurfaces}.
Chapter \ref{chapter:Matrix Factorization} is devoted to the concept of matrix factorization. In section \ref{section:Definitions and Properties}, we provide the definition and the main properties of this concept and fix notations for specific hypersurfaces. Section \ref{section:Matrix factorization and Maximal Cohen Maculay modules} explains how the concept of matrix factorization can be related to the subject of maximal Cohen Maculay (MCM) modules over specific hypersurfaces.
In 1997, K. Smith and M. Van den Bergh \cite{SV} introduced the notion of finite F-representation type (FFRT) over a class of rings for which the Krull-Remak-Schmidit Theorem holds as a characteristic $p$ extension of the notion of finite Cohen-Macaulay representation
type. They then showed that rings with finite
F-representation type play a role in developing the theory of differential operators
on the rings. However, Y.Yao in his paper \cite{Y} studied the notion of FFRT in more general settings.
T.Shibuta summarized in \cite{TS} several nice properties satisfied for rings of finite F-representation type.
For example, the Hilbert-Kunz multiplicities of
such rings are proved to be rational numbers by G.Seibert \cite{S}. Y.Yao in his paper \cite{Y} proved that tight closure commutes
with localization in such rings.
Chapter \ref{Chapter: Modules of finite Frepresentation type} provides in section \ref{section:Definition and examples} the definition and examples of the notion of finite
F-representation type (FFRT). Y.Yao in \cite{Y} observed that the localization and the completion of modules with FFRT have also FFRT. In section \ref{section:Several FFRT extensions of FFRT rings} we prove this observation and that both of $R[x]$ and $R[\![x]\!]$ have FFRT whenever $R$ has FFRT (Theorem \ref{C4.6}). The rings that have FFRT by FFRT system were introduced by Y.Yao in \cite{Y}. We prove in section \ref{section: Rings of FFRT by FFRT system} that over local ring of prime characteristic a module has FFRT if and only if it has FFRT by a FFRT system (Theorem \ref{P1}). This result enables us in section \ref{section: FFRT locus of a module is an open set} to prove that the FFRT locus of a module is an open set in the Zariski topology on $\Spec(R)$ (Theorem \ref{P4.11}).
We fix in chapter \ref{chapter:The finite F-representation type over hypersurfaces} the notation that $S=K[x_1, \ldots , x_n]$ or $S=K[\![x_1, \ldots , x_n]\!]$ where $K$ is a filed of prime characteristic $p$ with $[K:K^p] < \infty $, and $f$ is a nonzero nonunit element in $S$. In section \ref{section:The presentation of Fe as a cokernel of a Matrix Factorization of f}, we describe a presentation of $F_*^e(S/fS)$ as a cokernel of a matrix factorization of $f$ (Theorem \ref{P5.11}). This presentation of $F_*^e(S/fS)$ as a cokernel of a matrix factorization of $f$ allows us in section \ref{section:When does the ring fuv have finite F-representation type?} to find a characterization of when the ring $S[\![u,v]\!]/(f+uv)$, where $K$ is an algebraically closed field of prime characteristic $p>2$, has FFRT (Theorem \ref{P30}) and hence in section \ref{section:Class of rings that have FFRT but not finite CM type} we provide a class of rings that have finite F- representation type but it does not have finite Cohen-Macaulay type (Theorem \ref{P7.10}). In section \ref{section:Qutient by a monomial ideal}, we use this presentation to prove that $S/I$ has FFRT for any monomial ideal $I$ in $S$ (Theorem \ref{P5.25}). Furthermore, if $S=K[\![x_1,...,x_n]\!]$, we use this presentation to show in section \ref{section:The ring pd has finite Frepresentation type} that the ring $S[\![y]\!]/(y^{p^d} +f)$ has finite F-representation type for any $f$ in $S$ (Theorem \ref{P5.1}).
The notion of the $F$-signature (Definition \ref{D6.1}) was introduced and defined by C. Huneke and G. Leuschke in \cite{HL} on an F-finite local ring of prime
characteristic with a perfect residue field. Y. Yao in his work \cite{Y2} has defined the F-signature to arbitrary local rings $R$ without the assumptions that the residue field is perfect. K.Tucker in his paper \cite{KT} proved that this limit exists. The $F$-signature seems to give subtle information on the singularities of $R$. For example I. Aberbach and G.
Leuschke \cite{AL} have proved that the $F$-signature is positive if and only if $R$ is strongly F-regular. Furthermore, We have $\mathbb{S}(R) \leq 1$ with equality if and only if R is regular \cite[Theorem 4.16.]{KT} or \cite[Corollary 16]{HL}.
In Chapter \ref{chapter:F-signature of specific hypersurfaces}, we compute the $F$-signature of specific hypersurfaces. Indeed, if $K$ is a field of prime characteristic $p$ with $[K:K^p] < \infty$ and $S=K[\![x_1,...,x_n]\!]$, we compute in the sections \ref{section:The F-signature of uv is a monomial} and \ref{section:The Fsignature of z when is a monomial} the $F$-signatures of the rings $S[\![u,v]\!]/(f+uv)$ and $S[\![z]\!]/(f+z^2)$ where $f$ is a monomial and $u,v,z$ are variables over $S$. The presentation of $F_*^e(S/fS)$ as a cokernel of a matrix factorization of $f$ playes a role in those computations. We proved also in section \ref{section:The Fsignatures are the same} that, for any $f\in S=K[\![x_1,...,x_n]\!]$, the rings $S/fS$ and $S[\![y_1,...,y_n]\!]/fS[\![y_1,...,y_n]\!]$ have the same $F$-signature and we give in section \ref{section:The Fsignature of pd} a characterization of the F-signature of the ring $S[\![y]\!]/(y^{p^d}+f)$ for any $f \in S$ and $d \in \mathbb{N}$ .
\chapter{Preliminaries}
\label{chapter:Preliminaries}
In the first section of this chapter, we review and provide the basic background from commutative algebra that is needed for this thesis and the theorems, propositions and lemmas that have references are stated in this section without proof. The second section contains some technical lemmas that are essential for obtaining the main results of this thesis.
\section{General Background}
\label{section:General Background}
\subsection{Graded modules and Rings}
\label{subsection:Graded modules and Rings}
Let $G$ be an abelian group with an additive operation $+$ and identity element $e \in G$ and let $R$ be a ring.
\begin{definition}\cite[Section 1.1]{NV}
\emph{ $R$ is a $G$-graded ring if there exists a family $\{ R_{\alpha} \, | \, \alpha \in G \}$ of additive subgroups $R_{\alpha}$ of $R$ such that $R=\bigoplus_{\alpha \in G}R_{\alpha}$ (as groups) and $R_{\alpha}R_{\beta} \subseteq R_{\alpha + \beta }$ for all $\alpha , \beta \in G$. A nonzero element $x \in R$ is a homogeneous of degree $\alpha\in G$ and we write $\deg(x)=\alpha$ if $x \in R_{\alpha}$}.
\end{definition}
\begin{definition}\cite[Section 2.1]{NV}
\emph{Let $M$ be an $R$-module. If $R=\bigoplus_{\alpha \in G}R_{\alpha}$ is a $G$-graded ring , then $M$ is a $G$-graded module if there exists a family $\{ M_{\alpha} \, | \, \alpha \in G \}$ of additive subgroups $M_{\alpha}$ of $M$ such that $M=\bigoplus_{\alpha \in G}M_{\alpha}$ (as groups) and $R_{\alpha}M_{\beta} \subseteq M_{\alpha + \beta }$ for all $\alpha , \beta \in G$. A nonzero element $x \in M$ is a homogeneous of degree $\alpha\in G$ and we write $\deg(x)=\alpha$ if $x \in M_{\alpha}$}.
\end{definition}
\begin{remark}\cite[Section 2.1]{NV}\label{R2.3}
Let $R=\bigoplus_{\alpha \in G}R_{\alpha}$ be a $G$-graded ring and let $M=\bigoplus_{\alpha \in G}M_{\alpha}$ be a $G$-graded $R$-module. If $N$ is a submodule of $M$ and $ N_{\alpha}=N \cap M_{\alpha}$ for each $\alpha \in G$, then $N$ is a graded submodule of $M$ if $N=\bigoplus_{\alpha \in G}N_{\alpha}$. Furthermore, if $N$ is a graded submodule of the $G$-graded module $M=\bigoplus_{\alpha \in G}M_{\alpha}$, then the quotient module $M/N$ is also a $G$-graded $R$-module as follows: $M/N=\bigoplus_{\alpha \in G}(M/N)_{\alpha} $ where $(M/N)_{\alpha}=(M_{\alpha}+N)/N= M_{\alpha}/N_{\alpha}$ for all $\alpha \in G$.
\end{remark}
\subsection{Projective and flat Modules}
\label{subsection:Projective Modules}
Let $R$ be a ring and $M$ be an $R$-module throughout this subsection.
\begin{definition}
\emph{$M$ is said to be projective if for every surjective module homomorphism $g: L \twoheadrightarrow N$ and every module homomorphism $h: M\rightarrow N$, there exists a module homomorphism $f:M\rightarrow L $ such that $ gf=h $}
\end{definition}
\begin{definition}\cite[Section 2.4]{BCA}\label{Def2.5}
\emph{ A system $\mathfrak{B}=\{m_i\}_{i\in I}$ of elements of $M$ is said to be linearly independent
(over $R$) if the condition $\sum_{i\in I}r_im_i=0$ with $r_i \in R$ for every $i$ and $r_i=0$ for
almost all $i$ implies that $r_i=0$ for every $i$. We say that $\mathfrak{B}$ is a basis of $M$ (over
$R$) if $\mathfrak{B}$ generates $M$ as an $R$-module and is linearly independent over $R$. An
$R$-module is said to be free (or $R$-free) if it has a basis. Furthermore, if an $R$-module $M$ has a finite free basis, $M$ is said to be a free module of finite rank}.
\end{definition}
\begin{remark}\cite[Proposition 4.7.6]{BCA}
Every free module is a projective module.
\end{remark}
\begin{definition}\cite[Section 17.4]{BCA}
\emph{ A projective resolution of $M$ is an exact sequence
$$ ... \rightarrow P_{n+1}\xrightarrow{d_{n+1}}P_n \rightarrow \cdots \rightarrow P_1 \xrightarrow{d_1}P_0 \xrightarrow{\epsilon}M \rightarrow 0 $$
where $P_j$ is projective $R$-module for every $j$. If there exists $n$ such that $P_j=0$ for all $j \geq n+1 $, then we say that $M$ has a finite resolution of length $\leq n$ }.
\end{definition}
It is well known \cite[Proposition 17.4.4]{BCA} that every module has a projective resolution not necessarily finite.
\begin{definition}\cite[Section 18.2]{BCA}
\emph{The projective dimension of $M$, denoted $\pd M$ or $\pd_RM$, is defined by
\begin{equation*}
\pd M=\inf \{ n \, |\, M \text{ has a projective resolution of length } \leq n \}
\end{equation*}
when $M$ has
no finite projective resolution, we write $\pd M= \infty$}.
\end{definition}
\begin{definition}\cite[Section 4.7]{BCA}\label{D2.7}
\emph{ $M$ is said to be flat if for every injective module homomorphism $g: L \longrightarrow N$, the induced module homomorphism $1\otimes g: M\otimes_RL \longrightarrow M\otimes_RN$ is injective. If $M$ is a flat $R$-module, $M$ is said to be faithfully flat if for every
nonzero $R$-module $N$, $ M\otimes_RN \neq 0$. Furthermore, If $S$ is an $R$-algebra,i.e. $S$ is a ring and there exists a ring homomorphism $\phi : R \rightarrow S$, then $S$ is flat (respectively faithfully flat) algebra over $R$ if $S$ is flat (respectively faithfully flat) as $R$-module}.
\end{definition}
\subsection{ Cokernel of Matrix}
\label{subsection:The Cokernl of a Matrix}
\begin{notation}
\emph{If $m,n \in \mathbb{N}$, then $M_{m \times n}(R)$ (and $M_{ n}(R)$) denote the set of all $m \times n$ (and $n \times n$) matrices over a ring $R$ (where $R$ is not necessarily commutative in this notation). If $A \in M_{n\times m}(R)$ is the matrix representing the $R$-linear map $\phi : R^{n}\longrightarrow R^{m}$ given by $ \phi (X)= AX$ for all $ X \in R^{ ^n}$, then we write $A:R^{n}\longrightarrow R^{m}$ to denote the $R$-linear map $\phi$ and $\Cok_R(A)$ denotes the cokernl of $\phi$ while $ \Ima_R(A)$ denotes the image of $\phi $. We write $\Cok(A)$ and $\Ima(A)$ if $R$ is known from the context}.
\end{notation}
\begin{definition}
\emph{If $A,B \in M_n(R)$ where $R$ is a commutative ring, we say that $A$ is equivalent to $B$ and we write $A\sim B$ if there exist invertible matrices $U,V \in M_n(R)$ such that $A =UBV$}.
\end{definition}
We can observe the following remark.
\begin{remark}\label{R18}
\begin{enumerate}
\item [(a)] If $A \in M_{n\times m}(R)$ and $B \in M_{s \times t}(R)$, then we define $A\oplus B$ to be the matrix in $M_{(n+s) \times (m+t)}(R)$ that is given by $A\oplus B = \left[
\begin{array}{cc}
A & \\
& B \\
\end{array}
\right]$. In this case, $ \Cok_R(A\oplus B)=\Cok_R(A)\oplus \Cok_R(B).$
\item [(b)] If $A, B \in M_n(R)$ are equivalent matrices, then $\Cok_R(A)$ is isomorphic to $\Cok_R(B)$ as $R$-modules.
\end{enumerate}
\end{remark}
If $A$ and $B$ are matrices in $M_n(R)$ such that $\Cok_R(A)$ is isomorphic to $\Cok_R(B)$, it is not true in general that this implies that $A$ is equivalent to $B$ \cite[Section 4]{LR}. However, the following Proposition gives a partial converse of \ref{R18}(b).
\begin{proposition}\label{L3.5}
Let $(R,\mathfrak{m})$ be a Noetherian local ring and let $A\in M_s(R)$ and $B\in M_t(R)$ be two matrices determining the $R$-linear maps $A:R^s \rightarrow R^s$ and $B:R^t \rightarrow R^t$ such that $A$ and $B$ are injective and all entries of $A$ and $B$ are in $\mathfrak{m}$. If $\Cok_R(A)$ is isomorphic to $\Cok_R(B)$ as $R$-modules, then $s=t$ and $A$ is equivalent to $B$.
\end{proposition}
\begin{proof}
Consider the following diagram with exact rows\\
\begin{center}
$\begin{CD}
0 @>>> R^t @>B>> R^t @>\delta>> \Cok_R(B) @>>> 0 \\
@. @V \beta VV @V\alpha VV @V\mu V \approx V \\
0 @>>> R^s@>A>> R^s @> \pi >> \Cok_R(A) @>>> 0
\end{CD} $
\end{center}
The projectivity of $R^t$ induces the $R$-linear map $\alpha$ that makes the right square of the diagram commute. Since $\Ima( \alpha B ) \subseteq \Ima(A)$, the projectivity of $R^t$ induces $\beta$ that makes the diagram commutative. The fact that all entries of $A$ are in $\mathfrak{m}$ yields that $\Ima(A) \subseteq \mathfrak{m}R^s $. Now, if $y \in R^s$, there exists $x \in R^t$ such that $\pi(y)= \mu \delta (x) =\pi \alpha (x)$. Therefore, $y- \alpha(x) \in \Ker (\pi) = \Ima( A )$ and then $y \in \Ima (\alpha) +\Ima( A) \subseteq \Ima (\alpha) + \mathfrak{m}R^{s} $. This implies that $R^s= \Ima (\alpha) + \mathfrak{m}R^{s} $ and hence by Nakayamma lemma \cite[Lemma 2.1.7]{BCA} it follows that $\alpha$ is surjective. The surjectivity of $\alpha$ shows that $t \geq s$. By similar argument we show that $s \geq t$ and hence $s=t$. Since $\alpha:R^s \rightarrow R^s $ is surjective, $\alpha$ is isomorphism \cite[Theorem 2.4]{Mat}. The five lemma \cite[Proposition 2.72]{JR} confirms that $\beta$ is isomorphism too. If $U$ and $V$ are the invertible matrices defining $\alpha$ and $\beta$ (see Remark \ref{Rem 2.9}), then $UB=AV$ as desired.
\end{proof}
\subsection{Presentation of finitely generated modules}
\label{subsection:Presentation of a module}
In this subsection, $R$ is a ring and $M$ is an $R$-module.
\begin{definition}\cite[Section 4.2]{BCA}
\emph{$M$ is said to be finitely presented if there exists an exact sequence
$G \xrightarrow{\psi} F \xrightarrow{\phi} M \xrightarrow{} 0$ where $G$ and $F$ are free modules of finite rank}.
\end{definition}
\begin{remark}\label{Rem 2.9}
If $R$ is Noetherain and $M$ is a finitely generated $R$-module, there exists an exact sequence
$G \xrightarrow{\psi} F \xrightarrow{\phi} M \xrightarrow{} 0$ where $G$ and $F$ are free modules of finite rank. If $\{g_1,...,g_m\}$ and $\{f_1,...,f_n\}$ are bases for $G$ and $F$ respectively and $ \psi(g_j)= \sum_{i=1}^n a_{ij}f_i$ for all $1 \leq j \leq m$, then the matrix $A=[a_{ij}]$ is said to represent $\psi$ with $\Cok A = M$ and we write $\Mat(\psi)=A$. If $G \xrightarrow{\theta} G $ and $G \xrightarrow{\tau} G $ are $R$-linear maps where $G$ is a free module of finite rank, then $\Mat(\theta + \tau)= \Mat(\theta) + \Mat(\tau)$ and $\Mat(\theta \tau)= \Mat(\theta) \Mat(\tau).$
\end{remark}
\subsection{Localization of modules and rings}
\label{subsection:Localization of modules and rings}
Let $R$ be a ring and $M$ be an $R$-module throughout this subsection.
\begin{definition}\cite[Section 2.7]{BCA}
\emph{If $W$ is a multiplicative closed set, i.e $1\in W$ and $st \in W$ for all $s,t \in W$, define the relation $\sim $ on $W \times M$ by $(s,m) \sim (t,n)$, where $(s,m) , (t,n) \in W \times M$, if and only if there exists $w \in W$ such that $wtm=wsn$. The relation $\sim$ is an equivalence relation and for every $(s,m) \in W \times M$ let $ \frac{m}{s}$ denote the equivalence class of $(s,m)$. Let $W^{-1}M$ denote the set of all equivalence classes $ \frac{m}{s}$ for all $(s,m) \in W \times M$}.
\end{definition}
It is straightforward to verify the following well known fact.
\begin{proposition}
If $W$ is a multiplicative closed set, then:
\begin{enumerate}
\item [(a)]$W^{-1}R$ is a commutative ring for which the addition and the multiplication are given by $\frac{a}{s} + \frac{b}{t}=\frac{ta+sb}{st}$ and $\frac{a}{s} . \frac{b}{t}=\frac{ab}{st}$ for all $\frac{a}{s} , \frac{b}{t} \in W^{-1}R$.
\item [(b)]$W^{-1}M$ is an $W^{-1}R$-module for which the addition and the scalar multiplication are given by $\frac{m}{s} + \frac{n}{t}=\frac{tm+sn}{st}$ and $\frac{a}{u} \frac{m}{s}=\frac{am}{us}$ for all $\frac{a}{u}\in W^{-1}R$ and $\frac{m}{s} , \frac{n}{t} \in W^{-1}M$.
\end{enumerate}
\end{proposition}
\begin{remark}
\begin{enumerate}
\item [(a)] If $P$ is a prime ideal in $R$, then $W= R \setminus P$ is a multiplicative closed set and we write $M_P$ (or $R_P$) to denote $W^{-1}M$ (or $W^{-1}R$).
\item [(b)] If $u \in R\setminus \{0\}$, then $W=\{1,u,u^2,...,u^n,...\}$ is a multiplicative closed set and we write $M_u$ (or $R_u$) to denote $W^{-1}M$ (or $W^{-1}R)$.
\end{enumerate}
\end{remark}
Recall from Corollary4.20, Rule4.22, and Example 4.18 in ChapterIII of \cite{Kunz} and \cite[Theorem 4.4]{Mat} the following property of localizations.
\begin{proposition}\label{LLL2}
Let $S$ and $T$ be a multiplicative closed sets of a ring $R$ and let $\hat{T}$ be the image of $T$ in $S^{-1}R$. For every $R$-module $M$, it follows that:
\begin{enumerate}
\item [(a)] If $S \subseteq T$, then $\hat{T}^{-1}(S^{-1}R) \cong T^{-1}R$ as rings and $\hat{T}^{-1}(S^{-1}M) \cong T^{-1}M$ as $T^{-1}R$-modules.
\item [(b)] If $f,g \in R$ and $\frac{g}{1}$ is the image of $g$ in $R_f$, then $(R_f)_{\frac{g}{1}}\cong R_{fg}$ as rings and $(M_f)_{\frac{g}{1}}\cong M_{fg}$ as $R_{fg}$-modules.
\item [(c)] Let $I$ be an ideal of $R$, let $P \in \Spec(R)$ contain $I$, and let $\bar{P}$ be the image of $P$ in $R/I$. If $\rho(\frac{r+I}{s+I})=\frac{r}{s}+I_P$ for all $\frac{r+I}{s+I} \in(R/I)_{\bar{P}}$, then the map $\rho:(R/I)_{\bar{P}}\rightarrow R_P/I_P$ is a ring isomorphism.
\item [(d)] $S^{-1}M$ is isomorphic to $S^{-1}R\otimes_{R}M$ as $S^{-1}R$-modules.
\end{enumerate}
\end{proposition}
Let $W$ be a multiplicative closed set of a ring $R$. If $M$ and $N$ are $R$-modules, there exists an $W^{-1}R$-linear map $W^{-1}\Hom_R(M,N) \rightarrow \Hom_{W^{-1}R}(W^{-1}M,W^{-1}N)$ given by $\frac{f}{w}\mapsto \frac{1}{w}W^{-1}f$ for every $f \in \Hom_R(M,N)$ and $w \in W$ where $W^{-1}f \in \Hom_{W^{-1}R}(W^{-1}M,W^{-1}N)$ is given by $W^{-1}f(\frac{m}{w})= \frac{f(m)}{w}$ for all $\frac{m}{w} \in W^{-1}M$. However, this $W^{-1}R$-linear map is isomorphism in the following case.
\begin{lemma}\cite[Lemma 4.87.]{JR}\label{LL1}
Let $W$ be a multiplicative closed set of a ring $R$. If $M$ is finitely presented, then $W^{-1}\Hom_R(M,N)$ is isomorphic to $\Hom_{W^{-1}R}(W^{-1}M,W^{-1}N)$ as $W^{-1}R$-module for every $R$-module $N$. In particular case, if $R$ is Noetherian and $M$ is finitely generated, then $W^{-1}\Hom_R(M,N)$ is isomorphic to $\Hom_{W^{-1}R}(W^{-1}M,W^{-1}N)$ as $W^{-1}R$-module for every $R$-module $N$.
\end{lemma}
We need the following lemma later in section \ref{section: FFRT locus of a module is an open set}.
\begin{lemma}\label{LL2}
Let $M$ and $N$ be finitely generated modules over a Noetherian ring $R$. If $W$ is a multiplicative closed set of $R$ such that $W^{-1}M$ is isomorphic to $W^{-1}N$ as $W^{-1}R$-module, then there exists $u \in W$ such that $M_u$ is isomorphic to $N_u$ as $R_u$-module.
\end{lemma}
\begin{proof}
Let $ \phi \in \Hom_{W^{-1}R}(W^{-1}M,W^{-1}N)$ be an isomorphism. By lemma \ref{LL1}, there exists $f \in \Hom_R(M,N)$ and $t\in W$ such that $\phi = \frac{1}{t}W^{-1}f$. Therefore, $f$ induces the following exact sequence resulted from the exactness of localizations
\begin{equation*}
0 \rightarrow W^{-1}(\Ker f ) \rightarrow W^{-1}M \xrightarrow{W^{-1}f} W^{-1}N \rightarrow W^{-1}(\Cok f ) \rightarrow 0.
\end{equation*}
Since $W^{-1}(\Cok f ) = \Cok W^{-1}f = \Cok \frac{1}{t}W^{-1}f= \Cok \phi $ and $W^{-1}(\Ker f ) = \Ker W^{-1}f = \Ker \frac{1}{t}W^{-1}f= \Ker \phi $, it follows $ W^{-1}(\Cok f ) = 0 = W^{-1}( \Ker f )$ and hence there exists $u \in W$ such that $ u \Cok f = 0 = u \Ker f$. Such $u$ exists as $\Cok f$ and $\Ker f$ are finitely generated $R$-modules. As a result, we get that $ \Cok f_u= (\Cok f )_u = 0$ and $ \Ker f_u = ( \Ker f )_u= 0 $ which proves that $M_u$ is isomorphic to $N_u$ as $R_u$-module.
\end{proof}
\subsection{Completion of modules and rings}
\label{subsection:The Completion of modules and rings}
In this subsection, $R$ is a ring and $M$ is an $R$-module.
\begin{definition}\cite[Chapter 8]{BCA}
\emph{A filtration on $R$ is a sequence $\{ I_n \}_{n\geq 0}$ of ideals of $R$ such that $I_0=R$, $I_n \supseteq I_{n+1}$
and $I_nI_m \subseteq I_{n+m}$ for all $m, n$. A ring with a filtration is
called a filtered ring.
If $R$ is a filtered ring with filtration $\{ I_n \}_{n\geq 0}$, a filtration on the $R$-module $M$ is a sequence $F=\{M_n\}_{n \geq 0}$ of submodules of $M$ such that $M_0=M$, $M_n \supseteq M_{n+1}$
and $I_mM_n \subseteq M_{m+n}$ for all m, n. In this case $M$ is called a
filtered $R$-module. The condition $I_mM_n \subseteq M_{m+n}$ is sometimes expressed by
saying that the filtration $\{M_n\}_{n \geq 0}$ is compatible with the filtration $\{ I_n \}_{n\geq 0}$.
An example of a filtration is the $I$-adic filtration
corresponding to an ideal $I$ of $R$. This is the filtration given by $I_n = I^n$
and $M_n = I^nM$ for all $n\geq 1$ , $M_0= M$ and $I_0=R$}.
\end{definition}
\begin{definition}
\emph{Let $R$ be a filtered ring with filtration $A=\{ I_n \}_{n\geq 0}$ and let $M$ be a
filtered $R$-module with (compatible) filtration $F=\{ M_n \}_{n\geq 0}$.
\begin{enumerate}
\item [(a)] We use $\widehat{R}_A$ (or $\widehat{R}$ if there is no ambiguity) to denote the completion of $R$ with respect to the filtration $A$ that is defined as $$ \widehat{R}_A= \varprojlim R/I_n=\{(x_n+I_n)_{n\geq 1}\in \prod_{n\geq 1}R/I_n \mid x_{n+1}-x_n \in I_n, \forall n\geq 1 \}.$$
\item [(b)] We use $\widehat{M}_F$ (or $\widehat{M}$ if there is no ambiguity) to denote the completion of $M$ with respect to the filtration $F$ that is defined as $$ \widehat{M}_F= \varprojlim M/M_n=\{(x_n+M_n)_{n\geq 1}\in \prod_{n\geq 1}M/M_n \mid x_{n+1}-x_n \in M_n, \forall n\geq 1 \}.$$
\item [(c)]If $I$ is an ideal of $R$, the $I$-adic completion of $R$ (and $M$) is denoted by $\widehat{R}_I$ (and $\widehat{M}_I$) and consequently are defined as $$ \widehat{R}_I= \varprojlim R/I^n=\{(x_n+I^n)_{n\geq 1}\in \prod_{n\geq 1}R/I^n \mid x_{n+1}-x_n \in I^n, \forall n\geq 1 \}$$ and $$ \widehat{M}_I=\varprojlim M/I^nM=\{(x_n+I^nM)_{n\geq 1}\in \prod_{n\geq 1}M/I^nM \mid x_{n+1}-x_n \in I^nM \forall, n\geq 1 \}.$$
\end{enumerate}
}
\end{definition}
We know from \cite[Section 8]{Mat} , \cite[Chapter 10]{AM}, or \cite[Section 8.2]{BCA} that $\widehat{R}$ is a ring and $\widehat{M}$ is $\widehat{R}$-module as follows:
\begin{proposition}
If $R$ is a filtered ring with filtration $A=\{ I_n \}_{n\geq 0}$ and $M$ is a
filtered $R$-module with (compatible) filtration $F=\{ M_n \}_{n\geq 0}$, then:
\begin{enumerate}
\item [(a)]$\widehat{R}_A$ is a ring with addition and multiplication given by $(r_n+I_n)_{n\geq 1}+(s_n+I_n)_{n\geq 1}= ((r_n + s_n)+I_n)_{n\geq 1}$ and $(r_n+I_n)_{n\geq 1}(s_n+I_n)_{n\geq 1}= (r_n s_n+I_n)_{n\geq 1}$ for all $(r_n+I_n)_{n\geq 1},(s_n+I_n)_{n\geq 1}\in \hat{R}$
\item [(b)]$\widehat{M}_F$ is an $\widehat{R}_A$-module with addition and scalar multiplication given by $(x_n+M_n)_{n\geq 1}+(y_n+M_n)_{n\geq 1}= ((x_n + y_n)+M_n)_{n\geq 1}$ and $(r_n+I_n)_{n\geq 1}(y_n+M_n)_{n\geq 1}= (r_n y_n + M_n)_{n\geq 1}$ for all $(x_n+M_n)_{n\geq 1},(y_n+M_n)_{n\geq 1}\in \widehat{M}_F$ and $(r_n+I_n)_{n\geq 1}\in \widehat{R}_A$
\end{enumerate}
\end{proposition}
An example of the completion is the following:
\begin{proposition}\cite[Section 7.1]{E}\label{EX1}
Let $A=R[x_1,..,x_t]$ be a polynomial ring over the ring $A$. If $I=(x_1,...,x_t)$, then $\widehat{A}_I=R[\![x_1,...,x_t]\!]$.
\end{proposition}
\begin{definition}
\emph{Let $R$ be a filtered ring with filtration $A=\{ I_n \}_{n\geq 0}$, and let $F=\{ M_n \}_{n\geq 0}$ and $\tilde{F}=\{ \tilde{M}_n \}_{n\geq 0}$
be two compatible filtrations on $M$. The filterations $F=\{ M_n \}_{n\geq 0}$ and $\tilde{F}=\{ \tilde{M}_n \}_{n\geq 0}$ are equivalent if the following holds: Given $r \geq 0$, there exist $n(r) \geq 0$ and
$\tilde{n}(r) \geq 0$ such that $M_{n(r)} \subseteq \tilde{M}_r$
and $\tilde{M}_{\tilde{n}(r)} \subseteq M_r$}.
\end{definition}
We need the following Proposition later in the subsection \ref{subsection:Modules over rings of prime characteristic p}.
\begin{proposition}\cite[Lemma 8.2.1]{BCA}\label{P2.20}
Let $R$ be a filtered ring with filtration $A=\{ I_n \}_{n\geq 0}$. If $F=\{ M_n \}_{n\geq 0}$ and $\tilde{F}=\{ \tilde{M}_n \}_{n\geq 0}$
are two compatible equivalent filtrations on $M$, then
$\varprojlim M/M_n$ is isomorphic $\varprojlim M/\tilde{M}_n$ as $\hat{R}$-modules.
\end{proposition}
When $R$ is Noetherian and $M$ is a finitely generated $R$-module, we have the following useful theorem
\begin{theorem}\cite[Theorems 8.7, 8.8, and 8.14]{Mat}\label{thm 2.21}
Let $R$ be a Noetherian ring and let $M$ be a finitely generated $R$-module. If $I$ is an ideal of $R$, then
\begin{enumerate}
\item [(a)] $\widehat{M}_I$ is isomorphic to $M\otimes_R \widehat{R}_I$ as $\widehat{R}_I$-module.
\item [(b)] $\widehat{R}_I$ is flat $R$-algebra. Furthermore, if $I$ is contained in the intersection of all maximal
ideals of $R$, then $\widehat{R}_I$ is faithfully flat $R$-algebra.
\end{enumerate}
\end{theorem}
If $M$ and $N$ are $R$-modules, recall that $M$ is said to be a direct summand of $N$ if there exists an $R$-module $L$ such that $N= M \oplus L$.
\begin{proposition}\cite[Theorem 3.6 and Corollary 3.7]{NW}\label{Pro2.22}
Let $(R,\mathfrak{m})$ be a Noetherian local ring. If $M$ and $N$ are finitely generated $R$-modules, then:
\begin{enumerate}
\item [(a)] $M$ is a direct summand of $N$ if and only if $\widehat{M}_{\mathfrak{m}}$ is a direct summand of $\widehat{N}_{\mathfrak{m}}$.
\item [(b)] $M \cong N$ as $R$-modules if and only if $\widehat{M}_{\mathfrak{m}}\cong \widehat{N}_{\mathfrak{m}}$ as $\widehat{R}_{\mathfrak{m}}$-modules.
\end{enumerate}
\end{proposition}
\subsection{Modules over rings of prime characteristic p }
\label{subsection:Modules over rings of prime characteristic p}
Throughout this section, all rings are of prime characteristic $p$ unless otherwise stated, $e \in \mathbb{N}$, and $q=p^e$. Let $R$ be a ring of prime characteristic $p$. For every $e\in \mathbb{N}$ , and $a,b \in R$, it follows that $(a+b)^{p^e}=a^{p^e}+ b^{p^e}$ and consequently the map $F^e: R \rightarrow R$ that is given by $F^e(a)= a^{p^e}$ is a ring homomorphism. This homomorphism is called the $e$-th Frobenius iterated map on the ring $R$.
\begin{definition}
\emph{If $M$ is an $R$ module, $F_*^e(M)$ denotes the $R$-module obtained via the restriction under the Frobenius homomorphism $F^e:R\rightarrow R$. Thus, $F_*^e(M)$ is the $R$-module that is the same as
$M$ as an abelian group but for every $m\in M$ we set $F_*^e(m)$ to represent the corresponding element in $F_*^e(M)$ and the $R$-module structure of $F_*^e(M)$ is given by
\begin{center}
$rF_*^{e}(m)=F_*^{e}(r^{p^e}m)$ for all $m \in M$ and $r\in R$
\end{center}
In a particular case, $F_*^e(R)$ is the abelian group $R$ that has an $R$-module structure via
\begin{center}
$rF_*^{e}(a)=F_*^{e}(r^{p^e}a) $ for all $a ,r\in R$.
\end{center}
}
\end{definition}
If $I$ is an ideal of $R$, then $I^{[q]}$ denotes the ideal generated by the set $\{r^q | r \in I \}$. As a result, if $I= (r_1,...,r_n)$, then $I^{[q]}= (r_1^q,...,r_n^q)$.
We can observe the following:
\begin{remark}\label{R2.29}
Let $M$ and $N$ be modules over a ring $R$. If $e\in \mathbb{N}$, it follows that:
\begin{enumerate}
\item [(a)] $F_*^{e+d}(M)=F_*^e(F_*^d(M))$ for all $d \in \mathbb{N}$.
\item [(b)] $F_*^e(R)$ is a ring itself, isomorphic to $R$, with an addition and a multiplication given by $F_*^e(a)+F_*^e(b)= F_*^e(a+b)$ and $F_*^e(a) F_*^e(b)= F_*^e(a b)$ for all $a$ and $b$ in $R$.
\item [(c)] $F_*^e(M)$ is also $F_*^e(R)$-module via $F_*^e(r)F_*^{e}(m)=F_*^{e}(rm)$ for all $m \in M$ and $r\in R$.
\item [(d)] If $I$ is an ideal, then $IF_*^e(M)=F_*^e(I^{[q]}M)$.
\item [(e)] If $N$ is a submodule of $M$, then $F_*^e(N)$ is a submodule of $F_*^e(M)$.
\item [(f)] If $N$ is a submodule of $M$, then $F_*^e(M)/F_*^e(N)$ is isomorphic to $F_*^e(M/N)$ as $R$-module.
\item [(g)] If $\phi:N\rightarrow M$ is an $R$-module homomorphism, then so is the map $F_*^e(\phi):F_*^e(N)\rightarrow F_*^e(M)$ that is given by $F_*^e(\phi)(F_*^e(m))=F_*^e(\phi(m))$ for each $m\in M$.
\item [(h)] $F_*^e(-)$ is an exact functor on the category of $R$-modules.
\item [(i)] If $R^q=\{r^q \, | \, r \in R \}$, then $R^q$ is a subring of $R$ and consequently $R$ is an $R^q$-module.
\item [(j)] If $\{M_i\}_{i \in I}$ is a family of $R$-modules, then $F_*^e(\prod_{i\in I}M_i)$ is isomorphic to $ \prod_{i\in I}F_*^e(M_i)$ and $F_*^e(\bigoplus_{i\in I}M_i)$ is isomorphic to $ \bigoplus_{i\in I}F_*^e(M_i)$ as $R$-modules.
\end{enumerate}
\end{remark}
\begin{proposition}\label{L2.3}
Let $M$ be an $R$-module. If $W$ is a multiplicative closed set of $R$, then $F_*^e(W^{-1}M)$ is isomorphic to $W^{-1}F_*^e(M)$ as $W^{-1}R$-module.
\end{proposition}
\begin{proof}
For every $m \in M$ and $w \in W$, define $\phi(F_*^e(\frac{m}{w}))=\frac{F_*^e(w^{q-1}m)}{w}$. If $n \in M$ and $y \in W$ satisfy that $F_*^e(\frac{m}{w})=F_*^e(\frac{n}{y})$, then $\frac{m}{w}=\frac{n}{y}$ and hence $uym=uwn$ for some $u\in W$. Therefore, $(uwy)^{q-1}uym=(uwy)^{q-1}uwn$ implies that $F_*^e(u^qy^qw^{q-1}m)=F_*^e(u^qw^qy^{q-1}n)$ and hence $uyF_*^e(w^{q-1}m)= uwF_*^e(y^{q-1}n)$. This shows that $\frac{F_*^e(w^{q-1}m)}{w}=\frac{F_*^e(y^{q-1}n)}{y}$ and hence $\phi:F_*^e(W^{-1}M)\rightarrow W^{-1}F_*^e(M) $ is well defined. Notice for any $m\in M$ and $w\in W$ that
\begin{equation*}
\frac{F_*^e(m)}{w}=\frac{w^{q-1}F_*^e(m)}{w^q}=\frac{F_*^e((w^q)^{q-1}m)}{w^q}=\phi(F_*^e(\frac{m}{w^q}))
\end{equation*}
This shows that $\phi$ is surjective. One can verify that $\phi$ is also injective module homomorphism over $W^{-1}R$ and hence
$\phi:F_*^e(W^{-1}M)\rightarrow W^{-1}F_*^e(M) $ is an isomorphism as $W^{-1}R$-module isomorphism.
\end{proof}
\begin{proposition}\label{L2.4}
Let $M$ be an $R$-module and let $I$ be a finitely generated ideal of $R$. If $\widehat{M}_I$ is the $I$-adic completion of $M$, then $F_*^e(\widehat{M}_I)$ is isomorphic to $\widehat{F_*^e(M)}_I$ as $\hat{R}_I$-modules.
\end{proposition}
\begin{proof}
For each $n \in \mathbb{N} $, let $J_n=I^n $. It follows that
\begin{eqnarray*}
\widehat{F_*^e(M)}_I &=& \{ (F_*^e(x_n)+I^nF_*^e(M))_{n=1}^{\infty}\in \prod_{n=1}^{\infty}\frac{F_*^e(M)}{I^nF_*^e(M)}\, |\, F_*^e(x_{n+1})- F_*^e(x_n) \in I^nF_*^e(M) \} \\
&=& \{ (F_*^e(x_n)+F_*^e(J_n^{[q]}M))_{n=1}^{\infty}\in \prod_{n=1}^{\infty}\frac{F_*^e(M)}{F_*^e(J_n^{[q]}M)} \, | \, F_*^e(x_{n+1})- F_*^e(x_n) \in F_*^e(J_n^{[q]}M) \} \\
&=& \varprojlim\frac{F_*^e(M)}{F_*^e(J_n^{[q]}M)}.
\end{eqnarray*}
On the other hand, we have
\begin{eqnarray*}
F_*^e(\widehat{M}_I) &=& \{ F_*^e((x_n +I^nM)_{n=1}^{\infty})\in F_*^e(\prod_{n=1}^{\infty}\frac{M}{I^nM})\, | \, x_{n+1} - x_n \in I^nM \} \\
&\cong& \{ (F_*^e(x_n)+F_*^e(I^nM))_{n=1}^{\infty}\in \prod_{n=1}^{\infty}\frac{F_*^e(M)}{F_*^e(I^nM)}\,|\, F_*^e(x_{n+1})- F_*^e(x_n) \in F_*^e(I^nM) \} \\
&=&\varprojlim \frac{F_*^e(M)}{F_*^e(I^nM)}.
\end{eqnarray*}
For $m$ sufficiently larger than $n$, we find that $F_*^e(I^mM) \subseteq F_*^e(J_n^{[q]}M)$ and it is obvious that $F_*^e(J_n^{[q]}M)\subseteq F_*^e(I^nM)$. This shows by Proposition \ref{P2.20} that $\varprojlim \frac{F_*^e(M)}{F_*^e(I^nM)}$ is isomorphic to $\varprojlim \frac{F_*^e(M)}{F_*^e(J_n^{[q]}M)}$ and hence $F_*^e(\widehat{M}_I)$ is isomorphic to $\widehat{F_*^e(M)}_I$ as $\hat{R}_I$-modules.
\end{proof}
\begin{definition}
\emph{ Let $M$ be an $R$-module where $R$ is a ring not necessarily of prime characteristic in this definition. $M[x]$ denotes the set of all polynomials in $x$ with coefficients in $M$, i.e. every element in $M[x]$ has the form $\sum_{j=0}^tm_jx^j $ where $t \in \mathbb{Z}_{+}$ and $m_j \in M$ for each $0\leq j \leq t$. The zero polynomial and the addition between two polynomials in $M[x]$ can be defined similarly as in the case of polynomial rings and for $f= \sum_{i=0}^sr_ix^i \in R[x]$ and $m=\sum_{j=0}^tm_j x^j \in M[x]$ we define
$$fm=\sum_{i=0}^s\sum_{j=0}^tr_im_jx^{i+j}.$$}
\end{definition}
One can check the following remark.
\begin{remark}\label{R4.3}
If $M$ is an $R$-module where $R$ is a ring not necessarily of prime characteristic in this remak, then
\begin{enumerate}
\item [(a)] $M[x]$ is an $R[x]$-module(\cite[Chapter 2]{AM}).
\item [(b)] $M[x]$ is isomorphic to $M \otimes_RR[x]$ as $R[x]$-modules (\cite[Chapter 2]{AM}).
\item [(c)] If $M$ is a finitely generated $R$-module that is generated by $\{m_1,...,m_n\}$, then $M[x]$ is a finitely generated $R[x]$ -module that is generated by $\{m_1\otimes_R1,...,m_n\otimes_R1\}$.
\end{enumerate}
\end{remark}
\begin{remark}
If $R$ is a ring, $M$ is an $R$-module, $e \in \mathbb{N} $ and $q=p^e$, notice that:
\begin{enumerate}
\item [(a)] $F_*^e(M)[x]$ is $R[x]$-module with scalar multiplication given as follows:\\
If $f= \sum_{i=0}^sr_ix^i \in R[x]$and $m=\sum_{j=0}^tF_*^e(m_j)x^j \in F_*^e(M)[x]$ and , then
\begin{equation}\label{Equat1}
fm=\sum_{i=0}^{s}\sum_{i=0}^tF^e(r_i^qm_j)x^{i+j}.
\end{equation}
\item [(b)]$F_*^e(M[x])$ is $R[x]$-module with scalar multiplication given as follows:
If $f= \sum_{i=0}^sr_ix^i \in R[x]$ and $F_*^e(m)=\sum_{j=0}^tF_*^e(m_jx^j) \in F_*^e(M[x])$ and , then
\begin{equation}\label{Equat2}
fF_*^e(m)=\sum_{i=0}^{s}\sum_{j=0}^tF_*^e(r_i^qm_jx^{qi+j}).
\end{equation}
\end{enumerate}
\end{remark}
\begin{proposition}\label{P2.41}
Let $R$ be a ring and let $M$ be an $R$-module. \\ If $M_k=\{\sum_{j=0}^{t}F_*^e(m_jx^{qj+k})\,|\,m_j \in M \text{ and } t\in \mathbb{Z}_{+}\}$ for each $0\leq k \leq q-1$, then:
\begin{enumerate}
\item [(a)] $M_k$ is $R[x]$-submodule of $F_*^e(M[x])$.
\item [(b)] $F_*^e(M[x])=\bigoplus_{k=0}^{q-1}M_k$.
\item [(c)] $F_*^e(M)[x]$ is isomorphic to $M_k$ as $R[x]$-modules.
\item [(d)] $F_*^e(M[x])$ is isomorphic to $ (F_*^e(M)[x])^{\oplus q}$ as $R[x]$-modules.
\end{enumerate}
\end{proposition}
\begin{proof}
(a) If $r=\sum_{i=0}^{s}r_ix^i \in R[x]$ , $m=\sum_{j=0}^{t}F_*^e(m_jx^{qj+k})\in M_k$ and $n=\sum_{j=0}^{t}F_*^e(n_jx^{qj+k})\in M_k$, then $m+n=\sum_{j=0}^{t}F_*^e((m_j+n_j)x^{qj+k})\in M_k$ and, by the scalar multiplication given by equation \ref{Equat2}, $$rm=(\sum_{i=0}^{s}r_ix^i)(\sum_{j=0}^{t}F_*^e(m_jx^{qj+k}))=\sum_{i=0}^{s}\sum_{j=0}^{t}F_*^e(r_i^qm_jx^{q(i+j)+k}))\in M_k.$$
This shows that $M_k$ is $R[x]$-submodule of $F_*^e(M[x])$.
(b) If $m = \sum_{j=0}^t m_j x^j\in M[x]$, then $F_*^e(m) = \sum_{j=0}^t F_*^e(m_jx^j)$.
For each $j\in \mathbb{Z}_{+}$, there exist unique $ 0\leq c_j \leq q-1 $ and $b_j \in \mathbb{Z}_{+}$ such that $j=qb_j + c_j$. Therefore, $F_*^e(m) = \sum_{j=0}^t F_*^e(m_jx^{qb_j + c_j})$ and accordingly we can write $F_*^e(m)= \sum_{k=0}^{q-1}f_k $ where $f_k\in M_k$ for each $0\leq k \leq q-1 $. Now let $f_k=\sum_{j=0}^{u}F_*^e(m_{k,j}x^{qj+k})\in M_k$ for all $0\leq k \leq q-1$. Notice that $\sum_{k=0}^{q-1}f_k=0$ if and only if $\sum_{k=0}^{q-1}\sum_{j=0}^{u}F_*^e(m_{k,j}x^{qj+k})=0$ if and only if $\sum_{k=0}^{q-1}\sum_{j=0}^{u}m_{k,j}x^{qj+k}=0$. For all $0 \leq k, l \leq q-1$ and $j,i\in \mathbb{Z}_{+}$ notice that $qj+k=qi+l$ if and only if $k=l$ and $j=i$. This makes $\sum_{k=0}^{q-1}\sum_{j=0}^{u}m_{k,j}x^{qj+k}=0$ if and only if $m_{k,j}=0$ for all $0 \leq k \leq q-1$ and $0 \leq j \leq u$. Therefore $\sum_{k=0}^{q-1}f_k=0$ if and only if $f_k=0$ for all $0 \leq k \leq q-1$. This shows that $F_*^e(M[x])=\bigoplus_{k=0}^{q-1}M_k$.
(c) If $\sum_{j=0}^{t}F_*^e(m_j)x^{j} \in F_*^e(M)[x]$, define $\phi(\sum_{j=0}^{t}F_*^e(m_j)x^{j}) =\sum_{j=0}^{u}F_*^e(m_jx^{qj+k})$. One can check that $\phi(m+n)=\phi(m)+\phi(n)$ for all $m,n\in F_*^e(M)[x]$ and $\phi:F_*^e(M)[x]\rightarrow M_k$ is a group isomorphism. Furthermore, if $r=\sum_{i=0}^{s}r_ix^i \in R[x]$ and $m=\sum_{j=0}^{t}F_*^e(m_j)x^{j}\in F_*^e(M)[x]$, we get by the scalar multiplications given by equations \ref{Equat1} and \ref{Equat2} that
\begin{equation*}
\phi(rf)=\phi(\sum_{i=0}^{s}\sum_{j=0}^{t}F_*^e(r_i^qm_j)x^{i+j})=\sum_{i=0}^{s}\sum_{j=0}^{t}F_*^e(r_i^qm_jx^{q(i+j)+k}) \text{ and }
\end{equation*}
\begin{equation*}
r\phi(f)= (\sum_{i=0}^{s}r_ix^i)(\sum_{j=0}^{t}F_*^e(m_jx^{qj+k}))=\sum_{i=0}^{s}\sum_{j=0}^{t}F_*^e(r_i^qm_jx^{q(i+j)+k}).
\end{equation*}
This shows that $\phi:F_*^e(M)[x]\rightarrow M_k$ is an isomorphism of $R[x]$-modules.
(d) follows from (b) and (c).
\end{proof}
\begin{proposition}\label{Le2.34}
Let $R$ be a ring. If $f=\sum_{n=0}^{\infty}f_n\in R[\![x_1,...,x_t]\!]$ where $f_n$ is a homogeneous polynomial in $R[x_1,...,x_t]$ of degree $n$ for all $n \geq 0$, then $f^q=\sum_{n=0}^{\infty}f_n^q$.
\end{proposition}
\begin{proof}
let $\mathfrak{m}$ be the maximal ideal of $R[\![x_1,...,x_t]\!]$. If $g=\sum_{n=0}^{\infty}g_n\in R[\![x_1,...,x_t]\!]$ where $g_n$ is a homogeneous polynomial in $R[x_1,...,x_t]$ of degree $n$, notice that $f=g$ if and only if $f+\mathfrak{m}^n=g+\mathfrak{m}^n$ for all $n\geq 1$. For every $n\geq 1$, we have
\begin{eqnarray*}
f^q+\mathfrak{m}^n &=& (f+ \mathfrak{m}^n)^q \\
&=& (\sum_{j=0}^{n-1}f_j+ \mathfrak{m}^n)^q\\
&=& (\sum_{j=0}^{n-1}f_j)^q+ \mathfrak{m}^n\\
&=& \sum_{j=0}^{n-1}f^q_j+ \mathfrak{m}^n \\
&=& h_n+\mathfrak{m}^n
\end{eqnarray*}
where
\begin{equation*}
h_n= \left\{
\begin{array}{ll}
\sum_{j=0}^{r_n-1}f_j^q, & \hbox{if } n=r_nq \hbox{ for some }r_n \in \mathbb{Z}_{+} \\
\sum_{j=0}^{r_n}f_j^q, & \hbox{if } n=r_nq+s_n \hbox{ for some }r_n \in \mathbb{Z}_{+} \hbox{ and } 1 \leq s_n \leq q-1.
\end{array}
\right.
\end{equation*}
On the other hand, if $g=\sum_{n=0}^{\infty}f_n^q$, then for all $n\geq 1$ we get
\begin{equation*}
g+\mathfrak{m}^n = g_n+\mathfrak{m}^n
\end{equation*}
where
\begin{equation*}
g_n= \left\{
\begin{array}{ll}
\sum_{j=0}^{r_n-1}f_j^q, & \hbox{if } n=r_nq \hbox{ for some }r_n \in \mathbb{Z}_{+} \\
\sum_{j=0}^{r_n}f_j^q, & \hbox{if } n=r_nq+s_n \hbox{ for some }r_n \in \mathbb{Z}_{+} \hbox{ and } 1 \leq s_n \leq q-1.
\end{array}
\right.
\end{equation*}
This shows that $f^q=\sum_{n=0}^{\infty}f_n^q$.
\end{proof}
\begin{proposition}\label{L2.35}
Let $R$ be a ring. If $S=R[\![x_1,...,x_t]\!]$, then $F_*^e(S)$ is isomorphic to $\prod_{n=0}^{\infty}F_*^e(R_n)$ as $S$-modules where $R_n$ is the group of all homogeneous polynomials in $R[x_1,...,x_t]$ of degree $n$ for all $n\in \mathbb{Z}_{+}$ with $R_0=R$. Furthermore, if $f=\sum_{j=0}^{\infty}f_n\in R[\![x_1,...,x_n]\!]$ where $f_n$ is a homogeneous polynomial in $R[x_1,...,x_n]$ of degree $n$ for all $n \geq 0$, we can write $F_*^e(f)= \sum_{n=0}^{\infty}F_*^e(f_n)$.
\end{proposition}
\begin{proof}
Notice that $R[\![x_1,...,x_t]\!]=\prod_{n=0}^{\infty}R_n$ where $R_n$ is the group of all homogeneous polynomials in $R[x_1,...,x_t]$ of degree $n$ for all $n\in \mathbb{Z}_{+}$ with $R_0=R$. If we define $\phi (F_*^e(\sum_{n=0}^{\infty}f_n))=\sum_{n=0}^{\infty}F_*^e(f_n) $, for every $f= \sum_{n=0}^{\infty}f_n \in R[\![x_1,...,x_n]\!] $, then $\phi: F_*^e(R[\![x_1,...,x_t]\!])\rightarrow \prod_{n=0}^{\infty}F_*^e(R_n)$ is a group isomorphism. Let $f=\sum_{n=0}^{\infty}f_n$ and $g=\sum_{n=0}^{\infty}g_n$ be elements in $R[\![x_1,...,x_n]\!]$. Recall from Proposition \ref{Le2.34} that $f^q=\sum_{n=0}^{\infty}f_n^q$ and hence
\begin{equation*}
fF_*^e(g)=F_*^e(f^qg)=F_*^e( \sum_{n=0}^{\infty}h_n)
\end{equation*}
where $h_n = \sum_{j=0}^{r_n}f_{r_n-j}^qg_{jq+s_n}$ whenever $n=r_nq+s_n$ for $r_n,s_n \in \mathbb{Z}_{+}$ with $0 \leq s_n \leq q-1$. Notice that $ \prod_{n=0}^{\infty}F_*^e(R_n)$ can be considered as $S$-module as follows:
If $f=\sum_{n=0}^{\infty}f_n \in S$ and $ \sum_{n=0}^{\infty}F_*^e(g_n)\in \prod_{n=0}^{\infty}F_*^e(R_n)$, then
\begin{equation*}
f\sum_{n=0}^{\infty}F_*^e(g_n)= \sum_{n=0}^{\infty}w_n
\end{equation*}
where $w_n= \sum_{j=0}^{r_n}f_{r_n-j}F_*^e(g_{jq+s_n})$ whenever $n=r_nq+s_n$ for $r_n,s_n \in \mathbb{Z}_{+}$ with $0 \leq s_n \leq q-1$. Therefore
\begin{equation*}
\phi(fF_*^e(g))=\phi(F_*^e(f^qg)) = f\sum_{n=0}^{\infty}F_*^e(g_n)=f\phi(F_*^e(g)).
\end{equation*}
This proves that $\phi: F_*^e(S)\rightarrow \prod_{n=0}^{\infty}F_*^e(R_n)$ is an isomorphism as $S$-modules and hence we can write $F_*^e(\sum_{j=0}^{\infty}f_n)=\sum_{n=0}^{\infty}F_*^e(f_n)$.
\end{proof}
\begin{proposition}\label{L2.34}
If $\Lambda_e$ is a subset of the ring $R$, then
\begin{enumerate}
\item [(a)] $\Lambda_e$ is a free basis of $R$ as a free $R^q$-module if and only if $\{ F_*^e(\lambda) \, | \, \lambda \in \Lambda_e\}$ is a free basis of $F_*^e(R)$ as a free $R$-module.
\item [(b)] If $\Lambda_e$ is a free basis of $R$ as a free $R^q$-module and $x$ is a variable on $R$, then $\{ \lambda x^j \, | \, \lambda \in \Lambda_e \text{ and } 0 \leq j \leq q-1 \}$ is a free basis of $R[x]$ as a free $R^q[x^q]$-module.
\item [(c)] If $\Lambda_e$ is a free basis of $R$ as a free $R^q$-module and $x_1,...,x_n$ are variables on $R$, then
\begin{equation*}
\{ \lambda x_1^{k_1}x_2^{k_2}...x_{n}^{k_{n}} \, | \, \lambda \in \Lambda_e , 0\leq k_j\leq q-1 , 1\leq j\leq n \}
\end{equation*}
is a free basis of $R[x_1,...,x_n]$ as a free $R^q[x_1^q,...,x_n^q]$-module.
\item [(d)] If $\Lambda_e$ is a free basis of $R$ as a free $R^q$-module and $x_1,...,x_n$ are variables on $R$, then
\begin{equation*}
\{ F_*^e(\lambda x_1^{k_1}x_2^{k_2}...x_{n}^{k_{n}}) \, | \, \lambda \in \Lambda_e , 0\leq k_j\leq q-1 , 1\leq j\leq n \}
\end{equation*}
is a free basis of $F_*^e(R[x_1,...,x_n])$ as a free $R[x_1,...,x_n]$-module.
\item [(e)] If $\Lambda_e$ is a finite free basis of $R$ as a free $R^q$-module of finite rank and $x$ is a variable on $R$, then $\{ \lambda x^j \, | \, \lambda \in \Lambda_e \text{ and } 0 \leq j \leq q-1 \}$ is a free basis of $R[\![x]\!]$ as a free $R^q[\![x^q]\!]$-module.
\item [(f)] If $\Lambda_e$ is a finite free basis of $R$ as a free $R^q$-module of finite rank and $x_1,...,x_n$ are variables on $R$, then
\begin{equation*}
\{ \lambda x_1^{k_1}x_2^{k_2}...x_{n}^{k_{n}} \, | \, \lambda \in \Lambda_e , 0\leq k_j\leq q-1 , 1\leq j\leq n \}
\end{equation*}
is a finite free basis of $R[\![x_1,...,x_n]\!]$ as a free $R^q[\![x_1^q,...,x_n^q]\!]$-module.
\item [(g)] If $\Lambda_e$ is a finite free basis of $R$ as a free $R^q$-module of finite rank and $x_1,...,x_n$ are variables on $R$, then
\begin{equation*}
\{ F_*^e(\lambda x_1^{k_1}x_2^{k_2}...x_{n}^{k_{n}}) \, | \, \lambda \in \Lambda_e , 0\leq k_j\leq q-1 , 1\leq j\leq n \}
\end{equation*}
is a finite free basis of $F_*^e(R[\![x_1,...,x_n]\!])$ as a free $R[\![x_1,...,x_n]\!]$-module.
\end{enumerate}
\end{proposition}
\begin{proof}
(a) For any finite subset $\Lambda \subseteq \Lambda_e$ notice that $ r= \sum_{\lambda\in \Lambda} r_{\lambda}^q\lambda$ if and only if $ F_*^e(r)= \sum_{\lambda\in \Lambda} r_{\lambda}F_*^e(\lambda)$ where $r_{\lambda} \in R$ for all $\lambda \in \lambda_e$. This proves the result.
(b) If $r \in R$, there exists a finite set $\Lambda \subseteq \Lambda_e$ such that $r= \sum_{\lambda \in \Lambda}r_{\lambda}^q\lambda $ where $r_{\lambda}\in R$ for all $\lambda \in \Lambda$. Since every $n \in \mathbb{Z}_{+}$ can be written as $n=uq+t$ where $u,t \in \mathbb{Z}_{+}$ and $0 \leq t \leq q-1$, it follows that $rx^n= \sum_{\lambda \in \Lambda}r_{\lambda}^q(x^{u})^q\lambda x^t$. This shows that $\{ \lambda x^j \, | \, \lambda \in \Lambda_e \text{ and } 0 \leq j \leq q-1 \}$ is a generating set of $R[x]$ as $R^q[x^q]$-module. Our task now is to show that $\{ \lambda x^j \, | \, \lambda \in \Lambda_e \text{ and } 0 \leq j \leq q-1 \}$ is linearly independent set (See Definition \ref{Def2.5}). It is enough to show that every finite set $\Omega$ on the following form is linearly independent where
\begin{equation*}
\Omega= \{\lambda_{(i,j)}x^{j} \, | \,\lambda_{(i,j)}\in \Lambda_e , 0 \leq j \leq q-1, 1 \leq i \leq n_j \} \text{ where } n_j \in \mathbb{N} \text{ for all } j.
\end{equation*}
For every $f \in R[x]$ and $n \in \mathbb{Z}_{+}$ , let $[f]_n$ denote the coefficient of $x^n$ in $f$. Now let $f=\sum_{j=0}^{q-1}\sum_{i=1}^{n_j}f_{(i,j)}^q\lambda_{(i,j)}x^{j}$ where $f_{(i,j)}\in R[x]$ for all $0 \leq j \leq q-1$ and $1 \leq i \leq n_i$ and we aim to show that $f=0$ implies that $f_{(i,j)}=0$ for all $i$ and $j$. This can be achieved by proving that $[ f_{(i,j)}]_s=0$ for every $s \in \mathbb{Z}_{+}$. Let $s \in \mathbb{Z}_{+}$ and $0 \leq t \leq q-1$. If $\alpha,\beta \in \mathbb{Z}_{+}$ with $0 \leq \beta \leq q-1$ notice that $sq+t= \alpha q +\beta$ if and only if $s=\alpha$ and $t=\beta$ and consequently we get
\begin{equation}\label{Equat3}
[f]_{sq+t}=\sum_{i=0}^{n_t}([f_{(i,t)}]_s)^q\lambda_{(i,t)}.
\end{equation}
Now if $f=0$, we get $[f]_{sq+t}=0$. Since $\lambda_{(i,t)} \in \Lambda_e$ for all $0 \leq i \leq n_t$, it follows from \ref{Equat3} that $[f_{(i,t)}]_s=0$ for all $0 \leq i \leq n_t$. This shows that $[f_{(i,j)}]_s=0$ for all $0 \leq j \leq q-1$, $0 \leq i \leq n_j$, and $s \in \mathbb{Z}_{+}$ and consequently $\{ \lambda x^j \, | \, \lambda \in \Lambda_e \text{ and } 0 \leq j \leq q-1 \}$ is a basis of $R[x]$ as $R^q[x^q]$-module.
(c) Use the result (b) above and the induction on $n$.
(d) Use the results above ((c) and (a)).
(e) Let $\Lambda_e= \{ \lambda_1, \ldots, \lambda_m \}$ and let $f=\Sigma_{n=0}^{\infty}r_nx^n$. For every $n \in \mathbb{Z}_{+}$, there exist $r_n,t_n \in \mathbb{Z}_{+}$ with $0 \leq t_n \leq q-1$ such that $n=qr_n+t_n$. This enables us to write
\begin{eqnarray*}
f &=& \sum_{n=0}^{\infty}r_nx^n=\sum_{k=0}^{\infty}r_{qk}x^{qk}+ \sum_{k=0}^{\infty}r_{qk+1}x^{qk+1}+ \ldots +\sum_{k=0}^{\infty}r_{qk+q-1}x^{qk+q-1} \\
&=& \sum_{j=0}^{q-1}\sum_{k=0}^{\infty}r_{qk+j}x^{qk+j}.
\end{eqnarray*}
For every $k,j \in \mathbb{Z}_{+}$ with $0 \leq j \leq q-1$, we can write $r_{qk+j}= \sum_{i=1}^{m}u_{(i,j,k)}^q\lambda_i$ where $u_{(i,j,k)}\in R$ for all $i,j$ and $k$. Therefore, for each $0 \leq j \leq q-1$ we get
\begin{eqnarray*}
\sum_{k=0}^{\infty}r_{qk+j}x^{qk+j} &=& \sum_{k=0}^{\infty}[\sum_{i=1}^{m}u_{(i,j,k)}^q\lambda_i]x^{qk+j} \\
&=& \sum_{i=1}^{m}\sum_{k=0}^{\infty}u_{(i,j,k)}^q\lambda_ix^{qk+j} \\
&=& \sum_{i=1}^{m}[\sum_{k=0}^{\infty}(u_{(i,j,k)}x^{k})^q]\lambda_ix^{j} \\
&=& \sum_{i=1}^{m}[\sum_{k=0}^{\infty}u_{(i,j,k)}x^{k}]^q\lambda_ix^{j} \text{ (Proposition \ref{Le2.34}) }.
\end{eqnarray*}
As a result,
\begin{eqnarray*}
f &=& \sum_{j=0}^{q-1}\sum_{k=0}^{\infty}r_{qk+j}x^{qk+j} \\
&=& \sum_{j=0}^{q-1}\sum_{i=1}^{m}[\sum_{k=0}^{\infty}u_{(i,j,k)}x^{k}]^q\lambda_ix^{j} \\
&=& \sum_{j=0}^{q-1}\sum_{i=1}^{m}[f_{(i,j)}]^q\lambda_ix^{j}
\end{eqnarray*}
where $f_{(i,j)}= \sum_{k=0}^{\infty}u_{(i,j,k)}x^{k}$ for all $i$ and $j$. This shows that $\{ \lambda x^j \, | \, \lambda \in \Lambda_e \text{ and } 0 \leq j \leq q-1 \}$ is a generating set of $R[\![x]\!]$ as a $R^q[\![x^q]\!]$-module. For every $f \in R[\![x]\!]$ and $n \in \mathbb{Z}_{+}$ , let $[f]_n$ denote the coefficient of $x^n$ in $f$. Let $f=\sum_{j=0}^{q-1}\sum_{i=1}^{m}f_{(i,j)}^q\lambda_{i}x^{j}$ where $f_{(i,j)}\in R[\![x]\!]$ for all $1 \leq i \leq m$ and $0 \leq j \leq q-1$. We aim to show that if $f=0$ , we get that $f_{(i,j)}=0$ for all $i$ and $j$. This can be achieved by proving that $[ f_{(i,j)}]_s=0$ for every $s \in \mathbb{Z}_{+}$. Let $s \in \mathbb{Z}_{+}$ and $0 \leq t \leq q-1$. If $\alpha,\beta \in \mathbb{Z}_{+}$ with $0 \leq \beta \leq q-1$ notice that $sq+t= \alpha q +\beta$ if and only if $s=\alpha$ and $t=\beta$ and consequently we get
\begin{equation}\label{Equat4}
[f]_{sq+t}=\sum_{i=0}^{m}([f_{(i,t)}]_s)^q\lambda_{i}.
\end{equation}
Now if $f=0$, we get $[f]_{sq+t}=0$. Since $\lambda_{i} \in \Lambda_e$ for all $0 \leq i \leq m$, it follows from \ref{Equat4} that $[f_{(i,t)}]_s=0$ for all $0 \leq i \leq m$. This shows that $[f_{(i,j)}]_s=0$ for all $0 \leq j \leq q-1$, $0 \leq i \leq m$, and $s \in \mathbb{Z}_{+}$ and consequently $\{ \lambda x^j \, | \, \lambda \in \Lambda_e \text{ and } 0 \leq j \leq q-1 \}$ is a basis of $R[\![x]\!]$ as $R^q[\![x^q]\!]$-module.
(f) Use the result (e) above and the induction on $n$.
(g) Use the results above ((f) and (a)).
\end{proof}
\begin{corollary}\label{L2.5}
Let $K$ be a field of positive prime characteristic $p$ and $q=p^e$ for some $e \in \mathbb{N} $. Let $\Lambda_e$ be the basis of $K$ as $K^q$ vector space.
\begin{enumerate}
\item [(a)] If $S:= K[x_1,...,x_{n}]$, then $F_*^{e}(S)$ is a free $S$-module with the basis
\begin{center}
$\Delta_n^e:= \{ F_*^{e}(\lambda x_1^{k_1}x_2^{k_2}...x_{n}^{k_{n}}) \, | \, \lambda \in \Lambda_e , 0\leq k_j\leq q-1 , 1\leq j\leq n \}$.
\end{center}
Furthermore, if $K(x_1,...,x_{n})$ is the fraction field of $K[x_1,...,x_{n}]$ and
\begin{equation*}
\Omega_n^e:= \{ F_*^{e}(\frac{\lambda x_1^{k_1}x_2^{k_2}...x_{n}^{k_{n}}}{1}) \, | \, \lambda \in \Lambda_e , 0\leq k_j\leq q-1 , 1\leq j\leq n \},
\end{equation*}
then $\Omega_n^e$ is a basis of $F_*^e(K(x_1,...,x_{n}))$ as $K(x_1,...,x_{n})$-vector space.
\item [(b)]If $S:= K[x_1,...,x_{n},....]$, then $F_*^{e}(S)$ is a free $S$-module with the basis $\Delta^e = \cup_{n \geq 1}\Delta_n^e$ where $\Delta_n^e$ as above. Furthermore, if $K(x_1,...,x_{n},...)$ is the fraction field of $K[x_1,...,x_{n},...]$ and $\Omega^e= \cup_{n \geq 1}\Omega_n^e$ where $\Omega_n^e$ as above, then we get $F_*^e(K(x_1,...,x_{n},...))$ is a $K(x_1,...,x_{n},...)$-vector space with the infinite basis $\Omega^e$.
\item [(c)]If $S:= K[\![x_1,...,x_{n}]\!]$ and $K$ is finite $K^q$-vector space, then $F_*^{e}(S)$ is a finitely generated free $S$-module with the basis
\begin{center}
$\Delta_n^e:= \{ F_*^{e}(\lambda x_1^{k_1}x_2^{k_2}...x_{n}^{k_{n}}) \, | \, \lambda \in \Lambda_e , 0\leq k_j\leq q-1 , 1\leq j\leq n \}$.
\end{center}
\end{enumerate}
\end{corollary}
\begin{proof}
(a) It is obvious from Proposition \ref{L2.34} (d) that $\Delta_n^e$ is a basis for $F_*^{e}(S)$ is a free $S$-module. Notice that $\{ \frac{\lambda x_1^{k_1}x_2^{k_2}...x_{n}^{k_{n}}}{1} \, | \, \lambda \in \Lambda_e , 0\leq k_j\leq q-1 , 1\leq j\leq n \}$ is a basis for $K(x_1,...,x_{n})$ as $K^q(x^q_1,...,x^q_{n})$-vector space. It follows from Proposition \ref{L2.34} (a) that $\Omega_n^e$ is a basis of $F_*^e(K(x_1,...,x_{n}))$ as $K(x_1,...,x_{n})$-vector space.
(b) Notice that $ K[x_1,...,x_{n},....]= \cup_{n \geq 1}K[x_1,...,x_{n}] $ and consequently we get $ K(x_1,...,x_{n},...)= \cup_{n \geq 1}K(x_1,...,x_{n}) $. It follows obviously that $F_*^e(K[x_1,...,x_{n},...])= \cup_{n \geq 1}F_*^e(K[x_1,...,x_{n}])$ and $F_*^e(K(x_1,...,x_{n},...))= \cup_{n \geq 1}F_*^e(K(x_1,...,x_{n}))$. Therefore, we obtain from (a) that $\Delta^e = \cup_{n \geq 1}\Delta_n^e$ is a basis for $F_*^e(S)$ as $S$-module and hence $\Omega^e$ is an infinite basis of $F_*^e(K(x_1,...,x_{n},...))$ as $K(x_1,...,x_{n},...)$-vector space.
(c)It is obvious from Proposition \ref{L2.34} (g).
\end{proof}
The following example explains that we can not remove the finiteness condition in Corollary \ref{L2.5} (c).
\begin{example}
Let $K$ be a field of positive prime characteristic $p$ and $S=K[\![x]\!]$. Suppose that $\Lambda_e$ is an infinite basis of $K$ as $K^q$ vector space and let $\Delta_e=\{\lambda x^j \, |\, \lambda \in \Lambda_e, 0 \leq j \leq q-1 \}$. We aim to show that $\{F_*^e(\lambda x^j) \, |\, \lambda \in \Lambda_e, 0 \leq j \leq q-1 \}$ is not a basis for $F_*^e(S)$ as $S$-module. It is enough by Proposition \ref{L2.34} (a) to show that $\Delta_e$ is not a free basis for $S$ as $S^q$-module. Assume the contrary that $\Delta_e$ is a free basis for $S$ as $S^q$-module. For every $f \in S$ and $n \in \mathbb{Z}_{+}$ , let $[f]_n$ denote the coefficient of $x^n$ in $f$. Let $\{\lambda_n\}_{n \geq 0}$ be an infinite subset of $\Lambda_e$ such that $\lambda_i \neq \lambda_j$ whenever $i\neq j$ and let $f= \sum_{n=0}^{\infty}\lambda_nx^{n}\in S$. Therefore, there exist nonnegative integers $n_0,...,n_{q-1}$ such that $f=\sum_{j=0}^{q-1}\sum_{i=1}^{n_j}f_{(i,j)}^q\lambda_{(i,j)}x^{i}$ where $ \lambda_{(i,j)} \in \Lambda_e $ and $f_{(i,j)} \in S $ for all $0 \leq j \leq q-1$ and $1 \leq i \leq n_j$. For every $s, \alpha \in \mathbb{Z}_{+}$ and $0 \leq t,\beta \leq q-1$, we notice that $sq+t= \alpha q +\beta$ if and only if $s=\alpha$ and $t=\beta$ and consequently
\begin{equation*}
\lambda_{sq+t}=[f]_{sq+t}=\sum_{i=0}^{n_t}([f_{(i,t)}]_s)^q\lambda_{(i,t)}.
\end{equation*}
The above equation implies that $\lambda_{sq+t} \in \{\lambda_{(i,j)} \, | \, 0 \leq j \leq q-1, 1 \leq i \leq n_j \}\cup \{0\}$ and consequently $\{\lambda_n\}_{n \geq 0} \subseteq \{\lambda_{(i,j)} \, | \, 0 \leq j \leq q-1, 1 \leq i\leq n_j \}\cup \{0\}$ which is a contradiction as $\{\lambda_n\}_{n \geq 0}$ is infinite set of distinct elements.
\end{example}
\begin{definition}\label{D2.6}
\emph{A Noetherian ring $R$ is said to be F-finite if $F_*^1(R)$ is finitely generated $R$ module (or equivalently that $F_*^e(R)$ is finitely generated $R$-module for all $e\in \mathbb{N}$)}
\end{definition}
\begin{remark}
If $R$ is $F$-finite ring, then it follows that:
\begin{enumerate}
\item [(a)]$R/I$ is $F$-finite ring for any ideal $I$ of $R$.
\item [(b)]$R[x]$ is $F$-finite ring.
\item [(c)] $R[\![x]\!]$ is $F$-finite ring.
\end{enumerate}
\end{remark}
\begin{proof}
Let $\Delta$ be a set of $R$ satisfying that $\{F_*^1(\delta) | \delta \in \Delta \}$ is a generating set of $F_*^1(R)$ as $R$-module.\\
(a) Notice that $\{F_*^1(\delta +I ) | \delta \in \Delta \}$ is a generating set of $F_*^1(R/I)$ as $R/I$-module.\\
(b) Let $f=\sum_{j=0}^nr_jx^j \in R[x]$ and hence $F_*^1(f)=\sum_{j=0}^nF_*^1(r_j)F_*^1(x^j)$. For each $0\leq j \leq n$, $j=u_jp+t_j$ for some $u_j, t_j \in \mathbb{Z}$ with $0 \leq t_j < p$ and we can write $F_*^1(r_j)= \sum_{\delta \in \Delta }r_{(j,\delta)} F_*^1(\delta)$ where $r_{(j,\delta)} \in R$ for all $ \delta \in \Delta$. Therefore,
\begin{equation*}
F_*^1(f)=\sum_{j=0}^nF_*^1(r_j)F_*^1(x^j)= \sum_{j=0}^n\sum_{\delta \in \Delta }r_{(j,\delta)}
x^{u_j} F_*^1(\delta x^{t_j}).
\end{equation*} This shows that $\{F_*^1(\delta x^t ) | \delta \in \Delta , 0 \leq t \leq p-1 \}$ is a generating set of $F_*^1(R[x])$ as $R[x]$-module.\\
(c) Let $I$ be the ideal generated by $x$ in $R[x]$ and let $A=R[x]$. Since $F_*^1(A)$ is finitely generated $A$-module, it follows from Theorem \ref{thm 2.21} that
$F_*^1(A)\otimes_A\widehat{A}_I$ is isomorphic to $ \widehat{F_*^1(A)}_I$ as $\widehat{A}_I$-module and consequently $ \widehat{F_*^1(A)}_I$ is finitely generated $\widehat{A}_I$-module. Now apply Proposition \ref{L2.4}.
\end{proof}
\begin{remark}\label{C2.9}
Let $M$ be a finitely generated $R$-module. If $R$ is F-finite, then $F_*^e(M)$ is a finitely generated $R$-module for all $e \in \mathbb{N}$.
\end{remark}
\begin{proof}
Let $B$ be a generating set of $M$ as $R$-module. If $A$ is a generating set of $R$ as an $R^{p^e}$-module, notice that $\{F_*^e(ab) | a \in A \text{ and } b\in B \}$ is a generating set of $F_*^e(M)$ as a finitely generated $R$-module.
\end{proof}
If $R$ is any ring (not necessarily of prime characteristic $p$), a non-zero $R$-module $M$ is said to be decomposable provided that there exist non-zero R-modules $M_1, M_2$ such that $M = M_1 \oplus M_2$; otherwise $M$ is indecomposable.
\begin{discussion}\label{disc2.33}
\begin{enumerate}
\item [(a)] Recall that if $M$ is a non-zero Noetherian module, then $M$ is a direct sum (not necessarily unique) of finitely many indecomposable
modules \cite[Proposition 2.1, Example 2.3]{NW}. However, if $M$ is a finitely generated module over a complete Noetherian local ring $R$ (where $R$ is not necessarily of prime characteristic $p$), by the Krull-Remak-Schmidt theorem \cite[Corollary 1.10]{CMR} or \cite[Theorem 2.13.]{NW}, $M$ can be written uniquely up to isomorphism as a direct sum of finitely many indecomposable
$R$-modules. In other words, if $M$ is a finitely generated module over a complete Noetherian local ring $R$, and If $ M \cong M_1\oplus ... \oplus M_s \approx N_1\oplus ... \oplus N_t$, where the $M_i$ and $N_j$ are finitely generated indecomposable $R$-modules, then
$s=t$ and, after renumbering, $ M_i \cong N_i$ for each $i$.
\item [(b)]As a result, if $(R, \mathfrak{m})$ is a Noetherian local ring not necessarily of prime characteristic $p$ and $M$ is a finitely generated $R$-module, then $M$ can be decomposed as $ M \cong R^a \oplus N$ where $a$ is a nonnegative integer and $N$ is an $R$-module that has no free direct summand. The number $a$ is unique and independent of the decomposition as when we take the $\mathfrak{m}$-adic completion, by the Krull-Remak-Schmidt theorem we stated above and Proposition \ref{Pro2.22} $a$ is uniquely determined.
\end{enumerate}
\end{discussion}
The following notion was introduced in \cite[Section 0]{Y}.
\begin{definition}\label{D2.11}
\emph{Let $(R, \mathfrak{m})$ be a Noetherian local ring. If $M$ is a finitely generated $R$-module, the maximal rank of
free direct summand of $M$ is denoted by $\sharp(M,R)$.}
\end{definition}
If $(R, \mathfrak{m})$ is an F-finite Noetherian local ring and $M$ is a finitely generated $R$-module, it follows from Remark \ref{C2.9} that $F_*^e(M)$ is a finitely generated $R$-module for all $e \in \mathbb{N}$ and consequently $F_*^e(M)$ is Noetherian $R$-module for all $e \in \mathbb{N}$. From the above discussion, there exists a unique non-negative integer $a_e$ and an $R$-module $M_e$ that has no free direct summand such that $F_*^e(M) \cong R^{a_e}\oplus M_e$.
This proves the following Remark.
\begin{remark}\label{C2.10}
Let $(R, \mathfrak{m})$ be an F-finite Noetherian local ring. If $M$ is a finitely generated $R$-module, for every $e \in \mathbb{N}$ there exists a unique nonnegative integer $a_e$ such that $F_*^e(M)\cong R^{a_e}\oplus M_e$ where $M_e$ has no non-zero free direct summand with the convention that $R^0=\{0\}$. The number $a_e$ is the maximal rank of
free direct summand of the $R$-module $F_*^e(M)$ and we write $\sharp(F_*^e(M),R)=a_e$.
\end{remark}
By \cite[Section 1]{RF} we can define $F$-pure ring as follows.
\begin{definition}
\emph{If $(R, \mathfrak{m})$ is an F-finite Noetherian local ring, then $R$ is $F$-pure if $\sharp(F_*^1(R),R) > 0$ (equivalently, $\sharp(F_*^e(R),R) > 0$ for all $e\in \mathbb{N}$).}
\end{definition}
Recall that if $I$ and $J$ are two ideals of a ring $R$, then $(I:J)$ is the following ideal
\begin{equation*}
(I:J)=\{r \in R | rJ \subseteq I \}.
\end{equation*}
R.Fedder in his paper \cite{RF} established the following criterion for the $F$-purity of a quotient of regular local ring of characteristic $p$.
\begin{proposition}\cite[Proposition 1.7]{RF}\label{P2.13}
Let $(S,\mathfrak{m})$ be a regular local ring (see Definition \ref{Def2.64}) of prime characteristic $p$ and let
$R = S/I$ where $I$ is an ideal of $S$. Then $R$ is $F$-pure if and only if $(I^{[p]}:I)\nsubseteq \mathfrak{m}^{[p]}.$
\end{proposition}
\subsection{Maximal Cohen Maculay Modules}
\label{subsection:Maximal Cohen Maculay Modules}
Throughout this section, $R$ is a Noetherian ring, and $M$ is a finitely generated $R$-module.
Recall that an element $r \in R$ is called a zerodivisor on $M$ if there exists a nonzero element $m \in M$ such that $rm=0$. An element $r\in R$ is said to be nonzerodivisor on M if $r$ is not a zerodivisor on
$M$, i.e. for every $m \in M\setminus\{0\}$ it follows that $rm \neq 0$ .
\begin{definition}\cite[Section 19.1]{BCA}
\emph{An element $r$ of $R$ is said to be $M$-regular
if $rM \neq M$ and $r$ is a nonzerodivisor on $M$. The sequence $r_1,...,r_n$ is an $M$-regular sequence or simply an $M$-sequence if $M\neq (r_1,...,r_n)M$, $r_1$ is a nonzerodivisor on $M$, and $r_i$ a nonzerodivisor on $M/(r_1,...,r_{i-1})M $
for every $ 2 \leq i\leq n$. An $M$-sequence $r_1,...,r_n$ in an ideal $I \subseteq R$, is called a maximal $M$-sequence in $I$ if $r_1,...,r_n, r$ is not a sequence on $M$ for any $r \in I$.}
\end{definition}
we can observe the following remark.
\begin{remark}\label{R 2.40}
If $r_1,...,r_n$ is a sequence on $M$, then $r_1^{k_1},...,r_n^{k_n}$ is a sequence on $M$ for every positive integers $k_1,...,k_n$. In particular case, if $R$ has a prime characteristic $p$ and $r_1,...,r_n$ is a sequence on $M$, then $r_1,...,r_n$ is a sequence on $F_*^e(M)$ for every $e\in \mathbb{N}$.
\end{remark}
It is well known \cite[ Corollary 19.1.4]{BCA} that any two maximal $M$-regular sequences in an ideal $I \subseteq R$ have the same length. This enables us to provide this definition.
\begin{definition}\cite[Section 19.1]{BCA}
\emph{Let $I$ be an ideal of $R$ with $IM\neq M$. The $I$-depth of $M$, denoted $\depth_IM$, is the length of any maximal $M$-regular sequence in $I$.
If $(R,\mathfrak{m})$ is a Noetherian local ring and $M$ is a nonzero finitely generated
$R$-module (so that $\mathfrak{m}M \neq M$ by Nakayama Lemma \cite{BCA}) then the $\mathfrak{m}$-depth of $M$ is called
simply the depth of $M$, and in this case it is also denoted $\depth M$.}
\end{definition}
\begin{definition}\cite[Definition 2.1.1]{BH}
\emph{Let $(R, \mathfrak{m})$ be a Noetherian local ring and let $M$ be a finitely generated $R$-module. $M$ is a Cohen Macaulay module if $M = 0$ or $ \depth M =
\dim M $. If $ \depth M = \dim R $, then $M$ is called a Maximal Cohen Macaulay module (henceforth abbreviated MCM). If $R$ is an arbitrary Noetherian ring and $M$ is a finitely generated $R$-module, then $M$ is Cohen Macaulay (respectively MCM) if $M_{\mathfrak{m}}$ is a Cohen Macaulay $R_{\mathfrak{m}}$-module (respectively a MCM $R_{\mathfrak{m}}$-module) for all maximal ideals $\mathfrak{m}$ of $R$.}
\end{definition}
\begin{lemma}\label{L2.55}
Let $M$ be an $R$-module where $R$ is a ring (not necessarily Noetherian). Suppose that $I$ is an ideal of $R$ such that $IM=0$. Let $P$ be a prime ideal of $R$ containing $I$. If $\bar{P}= P/I$, then $M_P$ is isomorphic to $M_{\bar{P}}$ as $(R/I)_{\bar{P}}$-modules.
\end{lemma}
\begin{proof}
First notice that $M$ is an $R/I$-module via the scalar multiplication $(r+I)m=rm$ for all $r+I \in R/I$ and $m\in M$ and consequently $I_PM_P=0$ makes $M_P$ an $R_P/I_P$-module with the scalar multiplication $(\frac{r}{s}+I_P)\frac{m}{t}=\frac{rm}{st}$ for all $\frac{r}{s} \in R_P$ and $\frac{m}{t} \in M_P$. As a result, it follows from Proposition \ref{LLL2}(c) that $M_P$ is an $(R/I)_{\bar{P}}$-modules via the scalar multiplication $\frac{r+I}{s+I}\frac{m}{t}=\frac{rm}{st}$ for all $\frac{r+I}{s+I} \in(R/I)_{\bar{P}}$ and $\frac{m}{t} \in M_P$. Furthermore, $M_{\bar{P}}$ is an $(R/I)_{\bar{P}}$-module via the scalar multiplication $\frac{r+I}{s+I}\frac{m}{t+I}=\frac{rm}{st+I}$ for all $\frac{r+I}{s+I} \in(R/I)_{\bar{P}}$ and $\frac{m}{t+I} \in M_{\bar{P}}$. For every $\frac{m}{t} \in M_P$, define $ \phi(\frac{m}{t})= \frac{m}{t+I}$. One can check that $\phi$ defines an isomorphism $ \phi: M_P \rightarrow M_{\bar{P}}$ of $(R/I)_{\bar{P}}$-modules.
\end{proof}
\begin{proposition}\label{P2.56}
Let $I$ be an ideal of $R$ such that $IM=0$. If $M$ is a Cohen Macaulay (respectively MCM) $R$-module, then $M$ is a Cohen Macaulay (respectively MCM) $R/I$-module.
\end{proposition}
\begin{proof}
First recall that $M$ is an $R/I$-module via the scalar multiplication given by $(r+I)m=rm$ for all $r \in R$ and $m\in M$. As a result, if $r_1,...,r_n \in R$ , then $r_1,...,r_n$ is $M$-sequence on $R$ if and only if $r_1+I,...,r_n+I$ is $M$-sequence on $R/I$. Therefore, if $R$ is local, then $M$ is a Cohen Macaulay (respectively MCM) $R/I$-module whenever $M$ is a Cohen Macaulay (respectively MCM) $R$-module. Now suppose that $R$ is non local and $M$ is a Cohen Macaulay (respectively MCM) $R$-module. This means that $M_P$ is a Cohen Macaulay (respectively MCM) $R_P$-module for every maximal ideal $P$ of $R$. As a result, if $P$ is a maximal ideal of $R$ containing $I$ and $\bar{P}=P/I$, then $M_P$ is a Cohen Macaulay (respectively MCM) $R_P/R_P$-module and consequently $M_P$ is a Cohen Macaulay (respectively MCM) $(R/I)_{\bar{P}}$-module. Since $M_P$ is isomorphic to $M_{\bar{P}}$ as $(R/I)_{\bar{P}}$-modules (Lemma \ref{L2.55}), it follows that $M_{\bar{P}}$ is a Cohen Macaulay (respectively MCM) $(R/I)_{\bar{P}}$-module. This shows that $M$ is a Cohen Macaulay (respectively MCM) $R/I$-module.
\end{proof}
\begin{proposition}\cite[Theorem 2.1.3]{BH}\label{P2.43}
Let $R$ be a Noetherian ring and $M$ a finitely generated $R$-module. Suppose that $r_1,...,r_t$ is an $M$-sequence and let $I=(r_1,...,r_t)$. If $M$ is a Cohen Macaulay $R$-module, then $M/IM$ is a Cohen Macaulay $R/I$-module.
\end{proposition}
The following Proposition describes the behaviour of the depth along exact sequences.
\begin{proposition}\label{P2.29}\cite[Proposition 1.2.9]{BH}
Let $ 0 \rightarrow U \rightarrow M \rightarrow N \rightarrow 0$ be an exact sequence of finite $R$-modules. If $ I \subseteq R$ is an ideal, then: \\
\begin{enumerate}
\item [(a)]$ \depth_I(M) \geq \min \{ \depth_I(U) , \depth_I(N) \}$ .
\item [(b)]$ \depth_I(U) \geq \min \{ \depth_I(M) , \depth_I(N)+1 \}$.
\item [(c)]$ \depth_I(N) \geq \min \{ \depth_I(U)-1 , \depth_I(M) \}$.
\end{enumerate}
\end{proposition}
This proposition leads to the following corollary.
\begin{corollary}
Let $(R, \mathfrak{m})$ be a Noetherian local ring. If $U$ and $N$ are finitely generated $R$-modules, then the $R$-module $U\oplus N$ is MCM if and only if $U$ and $N$ are both MCM $R$-modules.
\end{corollary}
\begin{proof}
Let $M$ be the $R$-module $U\oplus N$. Then we have the following short exact sequence
\begin{equation*}
0 \rightarrow U \rightarrow M \rightarrow N \rightarrow 0.
\end{equation*}
If $U$ and $N$ are both MCM $R$-modules, it follows that $\depth N= \depth U = \dim R$. Since $\depth M\leq \dim M$ \cite[Theorem 19.2.1]{BCA}, it follows from Proposition \ref{P2.29} (a) that
\begin{equation*}
\dim R = \min \{ \depth U , \depth N \} \leq \depth M \leq \dim M \leq \dim R.
\end{equation*}
Therefore, $\depth M =\dim R$ and consequently $M$ is MCM $R$-module. \\
Now assume that $M = U\oplus N$ is MCM $R$-module.
First we will show that $\depth U=\depth N$. Assume that $\depth U<\depth N$. Since $$ 0 \rightarrow U \rightarrow M \rightarrow N \rightarrow 0$$ is a short exact sequence, it follows from Proposition \ref{P2.29} (b) that
$$\min \{ \depth M , \depth N+1 \} \leq \depth U.$$ If $\depth N+1 =\min \{ \depth M , \depth N+1 \}$, then $\depth N+1 \leq \depth U < \depth N$ which is absurd. This makes
$$\dim R= \depth M = \min \{ \depth M , \depth N+1 \} \leq \depth U < \depth N $$ which is absurd too (as $\depth N \leq \dim N \leq \dim R$). Therefore, the assumption that $\depth U<\depth N$ is impossible and we conclude that $\depth N \leq \depth U$. Using the fact that $ 0 \rightarrow N \rightarrow M \rightarrow U \rightarrow 0$ is a short exact sequence and similar argument as above we conclude that $\depth U \leq \depth N$ and consequently $\depth U = \depth N$. Now the fact that $ 0 \rightarrow U \rightarrow M \rightarrow N \rightarrow 0$ is a short exact sequence and Proposition \ref{P2.29}(b) imply that
$$\min \{ \depth M , \depth N+1 \} \leq \depth U$$ and consequently $$\dim R=\depth M =\min \{ \depth M , \depth N+1 \} \leq \depth U=\depth N\leq \dim R.$$ This shows that $\dim R=\depth M =\depth U=\depth N$ as desired.
\end{proof}
An easy induction yields the following corollary.
\begin{corollary}\label{C2.30}
Let $(R, \mathfrak{m})$ be a Noetherian local ring. If $M_1,...,M_n$ are finitely generated $R$-modules, then the $R$-module $ \bigoplus_{i=1}^nM_i$ is MCM if and only if $M_i$ is MCM for every $1 \leq i \leq n$.
\end{corollary}
\begin{definition}\cite[Background]{RW}
\emph{The ring $(R,\mathfrak{m})$ is said to have finite Cohen-Macaulay type (or finite CM type) if there are, up to isomorphism, only finitely many indecomposable MCM R-modules.}
\end{definition}
In 1957, M.Auslander and D.Buchsbaum introduced a formula that relates the projective dimension of an $R$-module $M$ with $\depth M$ and $\depth R$ as follows
\begin{proposition}\cite[Section 19.2]{BCA}\label{P2.23}
Let $(R,\mathfrak{m})$ be a Noetherian local ring and $M$ a nonzero finitely generated $R$-module. If $\pd_RM < \infty$, then
\begin{equation*}
\pd M+\depth M=\depth R.
\end{equation*}
\end{proposition}
\subsection{Multiplicity and simple singularities}
\label{subsection:The multiplicity and simple singularity}
$R$ is a ring and $M$ is an $R$-module throughout this subsection (which provides the required material for the main result in section \ref{section:Class of rings that have FFRT but not finite CM type}).
\begin{definition}\cite[Section 2]{Mat}
\emph{ A chain $M=M_0 \supset M_1 \supset ... \supset M_n = 0 $ of submodules of $M$ is called a composition series if each $M_j/M_{j+1}$ is simple,i.e $M_j/M_{j+1} \cong R/ \mathfrak{m}_j$ for some maximal ideal $\mathfrak{m}_j$ in $R$. In this case, $n$ is called the length of the composition series.}
\end{definition}
Any two composition series of an $R$-module $M$, by Jordan--H\"{o}lder theorem \cite[Theorem 6.1.4]{BCA}, have the same length. This yields the following definition.
\begin{definition}\cite[Section 2]{Mat}
\emph{ An $R$-module $M$ is called of finite length if it has a composition series. The length of this composition series, denoted $\ell_R(M)$, is called the length of $M$.}
\end{definition}
\begin{definition}\cite[Section 20.1]{BCA} \label{Def2.64}
\emph{ Let $(R,\mathfrak{m})$ be a Noetherian local ring. If $\dim(R)=n$, we say that $R$ is a regular local ring or (RLR) if $\mathfrak{m}$ can be generated by $n$ elements.}
\end{definition}
\begin{definition}\label{D7.5} \cite[Definition A.19 ]{CMR} \\
\emph{Let $(R,\mathfrak{m},k)$ be a local ring of dimension $d$, let $I$ be an
$\mathfrak{m}$-primary ideal of $R$, and let $M$ be a finitely generated $R$-module.
The multiplicity of $I$ on $M$ is defined by
$$e_R(I,M) = \lim_{n\rightarrow \infty}\frac{d!}{n^d}\ell_R(M/I^nM) $$
where $\ell_R(-)$ denotes length as an $R$-module. In particular we set $e_R(M)=e_R(\mathfrak{m},M)$ and call it the multiplicity of $M$. Finally, we denote $e(R)=e_R(R)$
and call it the multiplicity of the ring $R$.}
\end{definition}
\begin{proposition}\label{F1} \cite[Corollary A.24]{CMR}\\
Let $(S,\mathfrak{n})$ be a regular local ring and $f \in S$ a non-zero nonunit.
Then the multiplicity of the hypersurface ring $R = S/( f )$ is the largest
integer $t$ such that $f \in \mathfrak{n}^t$.
\end{proposition}
\begin{definition} \cite[Definition 9.1]{CMR}
\emph{Let $(S,\mathfrak{n})$ be a regular local ring, and let $R = S/( g )$, where
$0 \neq g \in \mathfrak{n}^2$. We call R a simple singularity provided there are only finitely
many ideals $L$ of $S$ such that $g \in L^2$}.
\end{definition}
\begin{proposition}\label{F2} \cite[ Lemma 9.3]{CMR} \\
Let $(S,\mathfrak{n},k)$ be a regular local ring, $0 \neq f \in \mathfrak{n}^2$, and $R =S/( f )$
with $d = \dim(R)>1$.
If $R$ is a simple singularity and $k$ is an infinite field, then $e(R) \leq 3$.
\end{proposition}
\begin{proposition}\label{F3} \cite[Theorem 9.2]{CMR} \\
Let $(S,\mathfrak{n})$ be a regular local ring, $0 \neq f \in \mathfrak{n}^2$, and $R =S/( f )$.
If $R$ has finite CM type, then $R$ is a simple singularity.
\end{proposition}
\section{Technical Lemmas}
\label{section:Technical Lemmas}
Throughout this section, we adopt the following notation
\begin{notation}
\emph{Let $\mathfrak{P}$ denote a ring with identity that is not necessarily commutative.
Let $m$ and $ n$ be positive integers. If $\lambda \in \mathfrak{P}$, $ 1 \leq i \leq m $ and $ 1 \leq j \leq n $, then $L_{i,j}^{m \times n}(\lambda)$ (and $L_{i,j}^{ n}(\lambda)$) denotes the $m \times n$ (and $n\times n$) matrix whose $(i,j)$ entry is $\lambda$ and the rest are all zeros. When $i\neq j$, we write $E_{i,j}^n(\lambda):= I_n+L_{i,j}^n(\lambda)$ where $I_n$ is the identity matrix in $M_n(\mathfrak{P})$. If there is no ambiguity, we write $E_{i,j}(\lambda)$ (and $L_{i,j}(\lambda)$) instead of $E_{i,j}^n(\lambda)$ (and $L_{i,j}^{ n}(\lambda)$).}
\end{notation}
It is easy to observe the following remark
\begin{remark}\label{r2.1}
Let $m,n$ and $k$ be positive integers such that $1 \leq k,m \leq n$ with $k\neq m$. If $ \lambda \in \mathfrak{P}$ and $A \in M_n(\mathfrak{P})$, then :
\begin{enumerate}
\item [(a)] $E_{k,m}(\lambda)A$ is the matrix obtained from $A$ by adding $\lambda$ times row $m$ to row $k$.
\item [(b)] $AE_{k,m}(\lambda)$ is the matrix obtained from $A$ by adding $\lambda$ times column $k$ to column $m$.
\end{enumerate}
\end{remark}
\begin{lemma}\label{L2.9}
Let $m$ be an integer with $m \geq 2$ and $n=2m$. If $A$ is a matrix in $M_n(\mathfrak{P})$ that is given by
\begin{equation*}
A= \left[
\begin{array}{ccccccc}
b & & & & & & x \\
0 & b & & & & & \\
1 & 0 & b & & & & \\
& 1 & 0 & b & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & 1 & 0 & b & \\
& & & & 1 & 0 & b \\
\end{array}
\right]
\end{equation*}
then there exist two invertible matrices $M,N \in M_n(\mathfrak{P})$ such that:
\begin{enumerate}
\item[(a)]
$M$ has the form
\begin{equation*}
M= \left[
\begin{array}{ccccccc}
1 & 0 & a_{1,3} & & \ldots & & a_{1,n} \\
& 1 & 0 & a_{2,4} & & & \\
& & 1 & 0 & & & \vdots \\
& & & \ddots & \ddots & \ddots & \\
& & & & \ddots & \ddots & a_{n-2,n} \\
& & & & & 1 & 0 \\
& & & & & & 1 \\
\end{array}
\right].
\end{equation*}
\item[(b)]$N$ has the form
\begin{equation*}
N= \left[
\begin{array}{ccccccc}
1 & 0 & b_{1,3} & & \ldots & & b_{1,n} \\
& 1 & 0 & b_{2,4} & & & \\
& & 1 & 0 & & & \vdots \\
& & & \ddots & \ddots & \ddots & \\
& & & & \ddots & \ddots & b_{n-2,n} \\
& & & & & 1 & 0 \\
& & & & & & 1 \\
\end{array}
\right].
\end{equation*}
\item[(c)]
\begin{equation*}
MAN=\left[
\begin{array}{ccccccc}
0 & & & & & (-1)^{m-1}b^m & x \\
0 & 0 & & & & &(-1)^{m-1}b^m \\
1 & 0 & 0 & & & & \\
& 1 & 0 & 0 & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & 1 & 0 & 0 & \\
& & & & 1 & 0 & 0 \\
\end{array}
\right].
\end{equation*}
\end{enumerate}
\end{lemma}
\begin{proof}
We will prove the result by induction on $m\geq 2$. Let
\begin{equation*}
A=\left[
\begin{array}{cccc}
b & 0 & 0 & x \\
0 & b & 0 & 0 \\
1 & 0 & b & 0 \\
0 & 1 & 0 & b \\
\end{array}
\right]
\end{equation*}
It follows from remark \ref{r2.1} that
\begin{equation*}
E_{2,4}(-b)E_{1,3}(-b)A E_{1,3}(-b) E_{2,4}(-b)=\left[
\begin{array}{cccc}
0 & 0 & -b^2 & x \\
0 & 0 & 0 & -b^2 \\
1 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 \\
\end{array}
\right].
\end{equation*}
Taking $M=E_{2,4}(-b)E_{1,3}(-b)$ and $N=E_{1,3}(-b) E_{2,4}(-b)$ yields that
\begin{equation*}
M=N=\left[
\begin{array}{cccc}
1 & 0 & -b & 0 \\
0 & 1 & 0 & -b \\
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 \\
\end{array}
\right].
\end{equation*}
Now let $n=2(m+1)$ and let $A$ be the $n \times n$ matrix that is given by
\begin{equation*}
A= \left[
\begin{array}{ccccccc}
b & & & & & & x \\
0 & b & & & & & \\
1 & 0 & b & & & & \\
& 1 & 0 & b & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & 1 & 0 & b & \\
& & & & 1 & 0 & b \\
\end{array}
\right].
\end{equation*}
Let $\hat{A}$ be the $2m \times 2m$ matrix, obtained from $A$ by deleting the last two rows and the last tow columns of $A$, that is given by
\begin{equation*}
\hat{A}= \left[
\begin{array}{ccccccc}
b & & & & & & \\
0 & b & & & & & \\
1 & 0 & b & & & & \\
& 1 & 0 & b & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & 1 & 0 & b & \\
& & & & 1 & 0 & b \\
\end{array}
\right].
\end{equation*}
We can write
\begin{equation*}
A= \left[
\begin{array}{cccc|cc}
& & & & 0 & x \\
& \hat{A} & & & 0 & 0 \\
& & & & \vdots & \vdots \\
& & & & & \\\hline
0 & \ldots & 1 & 0 & b & 0 \\
0 & \ldots & 0 & 1 & 0 & b \\
\end{array}
\right].
\end{equation*}
By the induction hypothesis (where $x=0$), there exist two matrices $\hat{M},\hat{N} \in M_{2m}(\mathfrak{P})$ such that
\begin{equation*}
\hat{M}\hat{A}\hat{N}=\left[
\begin{array}{ccccccc}
0 & & & & & (-1)^{m-1}b^m & 0 \\
0 & 0 & & & & &(-1)^{m-1}b^m \\
1 & 0 & 0 & & & & \\
& 1 & 0 & 0 & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & 1 & 0 & 0 & \\
& & & & 1 & 0 & 0 \\
\end{array}
\right].
\end{equation*}
let $B$ and $C$ be $n \times n$ matrices, where $n=2(m+1)$,that are given by
$ B = \left[
\begin{array}{cc}
\hat{M} & \\
& I_2 \\
\end{array}
\right]$ and $ C = \left[
\begin{array}{cc}
\hat{N} & \\
& I_2 \\
\end{array}
\right]$
where $I_2$ is the identity matrix in $M_2(\mathfrak{P})$.
As a result, it follows that
\begin{equation*}
BAC =\left[
\begin{array}{ccccccc|cc}
0 & & & & & (-1)^{m-1}b^m & 0 & 0 &x \\
0 & 0 & & & & &(-1)^{m-1}b^m & 0 & 0 \\
1 & 0 & 0 & & & & & & \\
& 1 & 0 & 0 & & & & & \\
& & \ddots & \ddots & \ddots & & & & \\
& & & 1 & 0 & 0 & & & \\
& & & & 1 & 0 & 0 & & \\ \hline
& & & & & 1 & 0 & b & \\
& & & & & & 1 & 0 & b \\
\end{array}
\right].
\end{equation*}
Now multiply the $(n-1)$-th row by $(-1)^{m}b^m$ and add it to the first row and then multiply the $(n-3)$-th column by $-b$ and add it to the $(n-1)$-th column. After that, multiply the $n$-th row by $(-1)^{m}b^m$ and add it to the second row and then multiply the $(n-2)$-th column by $-b$ and add it to the $n$-th column. It follows that
\begin{equation*}
MAN
=\left[
\begin{array}{ccccccc}
0 & & & & & (-1)^{m}b^{m+1} & x \\
0 & 0 & & & & &(-1)^{m}b^{m+1} \\
1 & 0 & 0 & & & & \\
& 1 & 0 & 0 & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & 1 & 0 & 0 & \\
& & & & 1 & 0 & 0 \\
\end{array}
\right]
\end{equation*}
where $M=E_{2,n}((-1)^{m}b^m) E_{1,n-1}((-1)^{m}b^m)B $ and $N= C E_{n-3,n-1}(-b) E_{n-2,n}(-b)$. It is clear from the construction of the matrices $B$ and $C$ and from remark \ref{r2.1} that $M$ and $N$ have the right form.
\end{proof}
\begin{corollary}\label{C2.14}
Let $n=2m+1$ where $m$ is an integer with $m \geq 2$ and let $A$ be a matrix in $M_n(\mathfrak{P})$ given by
\begin{equation*}
A= \left[
\begin{array}{ccccccc}
b & & & & & x & 0 \\
0 & b & & & & & y \\
1 & 0 & b & & & & \\
& 1 & 0 & b & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & 1 & 0 & b & \\
& & & & 1 & 0 & b \\
\end{array}
\right].
\end{equation*}
Then there exist invertible matrices $M$ and $N$ in $M_n(\mathfrak{P})$ such that
\begin{equation*}
MAN =\left[
\begin{array}{ccccccc}
0 & & & & & x & (-1)^{m}b^{\frac{n+1}{2}} \\
0 & 0 & & & & (-1)^{m-1}b^{\frac{n-1}{2}} & y \\
1 & 0 & 0 & & & & \\
& 1 & 0 & 0 & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & 1 & 0 & 0 & \\
& & & & 1 & 0 & 0 \\
\end{array}
\right].
\end{equation*}
\end{corollary}
\begin{proof}
Let $ \hat{A}$ be the $2m \times 2m $ matrix obtained from $A$ be deleting the last row and the last column of $A$
\begin{equation*}
\hat{A}=\left[
\begin{array}{ccccccc}
b & & & & & & x \\
0 & b & & & & & \\
1 & 0 & b & & & & \\
& 1 & 0 & b & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & 1 & 0 & b & \\
& & & & 1 & 0 & b \\
\end{array}
\right].
\end{equation*}
It follows that
\begin{equation*}
A= \left[
\begin{array}{ccccc|c}
& & & & & 0 \\
& & & & & y \\
& & \hat{A} & & & 0 \\
& & & & & \vdots \\
& & & & & 0 \\ \hline
0 & \ldots & 0 & 1 & 0 & b \\
\end{array}
\right].
\end{equation*}
By Lemma \ref{L2.9} there exist $2m \times 2m$ matrices $\hat{M}$ and $ \hat{N}$ such that
\begin{equation*}
\hat{M}\hat{A}\hat{N}=\left[
\begin{array}{ccccccc}
0 & & & & & (-1)^{m-1}b^m & x \\
0 & 0 & & & & &(-1)^{m-1}b^m \\
1 & 0 & 0 & & & & \\
& 1 & 0 & 0 & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & 1 & 0 & 0 & \\
& & & & 1 & 0 & 0 \\
\end{array}
\right].
\end{equation*}
Let $B$ and $C$ be the matrices in $M_n(\mathfrak{P})$ given by
\begin{equation*}
B= \left[
\begin{array}{ccc|c}
& & & 0 \\
& \hat{M} & & \vdots \\
& & & 0 \\ \hline
0 & \ldots & 0 & 1 \\
\end{array}
\right] \text{ and } C= \left[
\begin{array}{ccc|c}
& & & 0 \\
& \hat{N} & & \vdots \\
& & & 0 \\ \hline
0 & \ldots & 0 & 1 \\
\end{array}
\right].
\end{equation*}
As a result, it follows that
\begin{equation*}
BAC= \left[
\begin{array}{ccccc|c}
0 & & & (-1)^{m-1}b^m & x & 0 \\
0 & 0 & & & (-1)^{m-1}b^m & y \\
1 & 0 & 0 & & & 0 \\
& \ddots & \ddots & \ddots & & \vdots \\
& & 1& 0 & 0 & 0 \\ \hline
0 & \ldots & 0 & 1 & 0 & b \\
\end{array}
\right].
\end{equation*}
Now multiply the last row of $BAC$ by $(-1)^{m}b^m$ and add it to the first row after that multiply the $(n-2)$-th column of $BAC$ by $-b$ and add it to the last column. This produces the required result.
Indeed, if $M = E_{1,n}((-1)^{m}b^m)B$ and $N= C E_{n-2,n}(-b)$, we get that $M$ and $N$ are invertible matrices satisfying that
\begin{equation*}
MAN =\left[
\begin{array}{ccccccc}
0 & & & & & x & (-1)^{m}b^{m+1} \\
0 & 0 & & & & (-1)^{m-1}b^{m} & y \\
1 & 0 & 0 & & & & \\
& 1 & 0 & 0 & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & 1 & 0 & 0 & \\
& & & & 1 & 0 & 0 \\
\end{array}
\right].
\end{equation*}
\end{proof}
\begin{lemma}\label{L.16}
Let $n$ be an integer with $n \geq 2 $. If $ A \in M_n(\mathfrak{P})$ is given by
$$ A = \begin{bmatrix} b & & & & & \\
1 & b & & & & \\
& 1 & b & & & \\
& & \ddots & \ddots & & \\
& & &1& b \end{bmatrix} , $$
there exist invertible upper triangular matrices $B, C \in M_n(\mathfrak{P})$ such that the $(i,i)$ entries of $B$ and $C$ are the identity element of $\mathfrak{P}$ for all $i=1, ..., n $ and
\[ BAC = \left[\begin{array}{ccccc} 0 & & & & (-1)^{n+1}b^n \\
1 & 0 & & & \\
& 1 &0 & & \\
& & \ddots &\ddots & \\
& & &1& 0 \end{array} \right]. \]
\end{lemma}
\begin{proof}
We will prove this lemma by induction on $n \geq 2$. If $ A= \begin{bmatrix} b & 0 \\ 1 & b \end{bmatrix} $, then $ E_{1,2}(-b) A E_{1,2}(-b)= \begin{bmatrix} 0 & (-1)^3b^2 \\ 1 & 0 \end{bmatrix}$ as required. Let $ A$ be a matrix in $ M_{n+1}(\mathfrak{P}) $ that is given by
\begin{equation*}
A = \begin{bmatrix} b & & & & & \\
1 & b & & & & \\
& 1 & b & & & \\
& & \ddots & \ddots & & \\
& & &1& b \end{bmatrix}.
\end{equation*}
Let $\hat{A} $ be the $n \times n$ matrix over $\mathfrak{P}$ obtained from $A$ by deleting the last row and last column of $A$
\begin{equation*}
\hat{A} = \begin{bmatrix} b & & & & & \\
1 & b & & & & \\
& 1 & b & & & \\
& & \ddots & \ddots & & \\
& & &1& b \end{bmatrix}.
\end{equation*}
It follows that $ A$ can be written as
\begin{equation*}
A=
\left[
\begin{array}{c|c}
& 0 \\ \hat{A} & \vdots \\ & 0 \\ \hline
0 \ldots 1 & b\\
\end{array}
\right].
\end{equation*}
By the induction hypothesis, there exist invertible upper triangular matrices $\hat{B}, \hat{C} \in M_n(\mathfrak{P})$ such that the $(i,i)$ entry of $\mathfrak{P}$ and $C$ is the identity element of $A$ for all $i=1, ..., n $ and
$$\hat{B} \hat{A}\hat{C} = \begin{bmatrix} 0 & & & & (-1)^{n+1}b^n \\
1 & 0 & & & \\
& 1 &0 & & \\
& & \ddots & \ddots & & \\
& & &1& 0 \end{bmatrix}. $$
Let
\begin{equation*}
M=\left[
\begin{array}{cccc|c}
& & & & 0 \\ & \hat{B} & & &\vdots \\ &&&&0 \\\hline
0& \ldots &0& 0 & 1\\
\end{array}
\right]
\end{equation*}
and
\begin{equation*}
N= \left[
\begin{array}{cccc|c}
& & & & 0 \\ &\hat{C} & & &\vdots \\ &&&&0 \\\hline
0& \ldots &0& 0 & 1\\
\end{array}
\right].
\end{equation*}
As a result, it follows that
\begin{equation*}
MAN= \left[\begin{array}{ccccc|c} 0 & & & & (-1)^{n+1}b^n &0 \\
1 & 0 & & & & \\
& 1 &0 & & & \\
& & \ddots &\ddots & &\\
& & &1& 0 & 0 \\ \hline
0 & & &0 & 1 & b \end{array} \right].
\end{equation*}
Now if $B=E_{1,n+1}((-1)^{n+2}b^n)M$ and $ C=NE_{ n,n+1}(-b) $, then $B$ and $C$ are invertible upper triangular matrices in $M_{n+1}(\mathfrak{P})$ such that the $(i,i)$ entry of $B$ and $C$ is the identity element of $\mathfrak{P}$ for all $i=1, ..., n+1 $ and
\begin{equation*}
BAC=\begin{bmatrix} 0 & & & & (-1)^{n+2}b^{n+1} \\
1 & 0 & & & \\
& 1 &0 & & \\
& & \ddots & \ddots & & \\
& & &1& 0 \end{bmatrix}.
\end{equation*}
\end{proof}
\begin{lemma}\label{1}
Let $n$ be an element in $ \mathbb{N} $ with $n \geq 2 $. If $ B \in M_n(\mathfrak{P})$ is given by
$$ B = \begin{bmatrix} b & & & & y \\
1 & b & & & & \\
& 1 & b & & & \\
& & \ddots & \ddots & & \\
& & &1& b \end{bmatrix} , $$
then the matrix $B$ is equivalent to the matrix
\begin{equation*}
\begin{bmatrix} 1 & & & & & \\
& 1 & & & & \\
& &1 & & & \\
& & & \ddots & & \\
& & & & y+(-1)^{n+1}b^n \end{bmatrix}.
\end{equation*}
\end{lemma}
\begin{proof}
Let $B$ be a matrix in $M_n(\mathfrak{P})$ with $n \geq 2$ . The result is obvious when $n=2$. Now assume that $n > 2$ and let $\hat{B}$ be the $(n-1) \times (n-1)$ matrix over $\mathfrak{P}$ obtained from $B$ by deleting the last row and last column of $B$
\begin{equation*}
\hat{B}= \begin{bmatrix} b & & & & \\
1 & b & & & & \\
& 1 & b & & & \\
& & \ddots & \ddots & & \\
& & &1& b \end{bmatrix}.
\end{equation*}
Then $B$ has the following form
\begin{equation*}
B =\left[
\begin{array}{cccc|c}
& & & & y \\ & \hat{B}& & \\ &&&&0 \\\hline
0& \ldots &0& 1 & b\\
\end{array}
\right].
\end{equation*}
Now use \ref{L.16} and appropriate row and column operations to get the result.
\end{proof}
\begin{corollary}\label{L.20}
Let $n $ be a positive integer such that $ n \geq 3 $ , $ 1 \leq k \leq n-1 $ and let $m= n-k$. Suppose that $u$ and $ v$ are two variables on $\mathfrak{P}$ and let $A_1^{(k)} \in M_k(\mathfrak{P})$ and $A_2^{(k)} \in M_m(\mathfrak{P})$ be given by $$A_1^{(k)} = \begin{bmatrix} b & & & & & \\
1 & b & & & & \\
& 1 & b & & & \\
& & \ddots & \ddots & & \\
& & &1& b \end{bmatrix} \text{ and }A_2^{(k)} = \begin{bmatrix} b & & & & & \\
1 & b & & & & \\
& 1 & b & & & \\
& & \ddots & \ddots & & \\
& & &1& b \end{bmatrix} . $$
If $B_k = \left[ \begin{array}{c|c} A_1^{(k)} & L_{1,m}^{k \times m}(v) \\ \hline L_{1,k
}^{m \times k}(u) & A_2^{(k)} \end{array}\right] =
\left[ \begin{array}{cccc|cccc} b & & & & & & & v \\
1 & b & & & & & & \\
& \ddots & \ddots & & & & & \\
& & 1& b & & & & \\ \hline
& & &u & b & & & \\
& & & & 1 & b & & \\
& & & & & \ddots & \ddots & \\
& & & & & &1 & b
\end{array}\right] $, then $ B_k$ is equivalent to the matrix $C_k= I_{n-2} \oplus \left[
\begin{array}{cc}
(-1)^{k+1}b^k & v \\
u & (-1)^{m+1}b^m \\
\end{array}
\right]\in M_n(\mathfrak{P})$ where $I_{n-2}$ is the identity matrix in $M_{n-2}(\mathfrak{P})$.
Moreover, if $D \in M_n(\mathfrak{P})$ is given by $ D = \begin{bmatrix} b & & & &uv \\
1 & b & & & \\
& 1 & b & & \\
& & \ddots & \ddots & \\
& & &1& b \end{bmatrix}$ , then $D$ is equivalent to the matrix $\tilde{D}= I_{n-2}\oplus \left[
\begin{array}{cc}
(-1)^{n}b^{n-1} & uv \\
1 & b \\
\end{array}
\right]\in M_n(\mathfrak{P})$ where $I_{n-2}$ is the identity matrix in $M_{n-2}(\mathfrak{P})$.
\end{corollary}
\begin{proof}
By Lemma \ref{L.16}, there exist upper triangular matrices $B_1, C_1 \in M_k(\mathfrak{P})$ and $B_2, C_2 \in M_{n-k}(\mathfrak{P})$ with $1$ along their diagonal such that
$$ B_1 A_1^{(k)}C_1 = \left[ \begin{array}{cccc} 0 & & & (-1)^{k+1}b^k \\
1 & 0 & & \\
& \ddots & \ddots & \\
& &1 & 0 \end{array} \right] ,
B_2 A_2^{(k)}C_2 = \left[ \begin{array}{cccc} 0 & & & (-1)^{m+1}b^m \\
1 & 0 & & \\
& \ddots & \ddots & \\
& &1 & 0 \end{array} \right] $$
where $m=n-k$.
Define $B,C \in M_n(\mathfrak{P})$ to be
$B = \begin{bmatrix} B_1 & 0 \\ 0 & B_2 \end{bmatrix} $ and $C = \begin{bmatrix} C_1 & 0 \\ 0 & C_2 \end{bmatrix} $.
Therefore,
\begin{align*}
B B_kC & = \left[ \begin{array}{cc} B_1 A_1^{(k)}C_1 & L_{1,m}^{k \times m}(v) \\ L_{1,k}^{m \times k}(u) & B_2 A_2^{(k)}C_2 \end{array}\right] \\
& = \left[ \begin{array}{cccc|cccc} 0 & & &(-1)^{k+1}b^k & & & & v \\
1 & 0 & & & & & & \\
& \ddots & \ddots & & & & & \\
& & 1&0 & & & & \\ \hline
& & &u & 0 & & & (-1)^{m+1}b^m \\
& & & & 1 &0 & & \\
& & & & & \ddots & \ddots & \\
& & & & & &1 & 0
\end{array}\right].
\end{align*}
Switching columns and rows of $ R B_kC $ yields the desired equivalent matrix.
Now by induction on $n \geq 3$ we prove the result related to $D$.
If $D= \left[
\begin{array}{ccc}
b & 0 & uv \\
1 & b & 0 \\
0 & 1 & b \\
\end{array}
\right]$, we get $E_{1,2}(-b)DE_{1,2}(-b)=\left[
\begin{array}{ccc}
0 & -b^2 & uv \\
1 & 0 & 0 \\
0 & 1 & b \\
\end{array}
\right].$
Switch the rows to get the desired result. Now assume that $D$ is $(n+1) \times (n+1)$ matrix. Then $D$ can be written as
\begin{equation*}
D=
\left[
\begin{array}{c|c}
& uv \\ \hat{D} & \vdots \\ & 0 \\ \hline
0 \ldots 1 & b\\
\end{array}
\right]
\end{equation*}
where $\hat{D}$ is the $n\times n$ matrix over $\mathfrak{P}$ that is given by
\begin{equation*}
\hat{D}= \begin{bmatrix} b & & & & \\
1 & b & & & & \\
& 1 & b & & & \\
& & \ddots & \ddots & & \\
& & &1& b \end{bmatrix}.
\end{equation*}
By Lemma \ref{L.16}, there exist upper triangular matrices $\hat{B}, \hat{C }\in M_n(\mathfrak{P})$ such that the $(i,i)$ entries of $\hat{B}$ and $\hat{C}$ are the identity element of $\mathfrak{P}$ for all $i=1, ..., n $ and
\[ \hat{B}\hat{D}\hat{C} = \left[\begin{array}{ccccc} 0 & & & & (-1)^{n+1}b^n \\
1 & 0 & & & \\
& 1 &0 & & \\
& & \ddots &\ddots & \\
& & &1& 0 \end{array} \right]. \]
Let
\begin{equation*}
M=\left[
\begin{array}{cccc|c}
& & & & 0 \\ & \hat{B} & & &\vdots \\ &&&&0 \\\hline
0& \ldots &0& 0 & 1\\
\end{array}
\right]
\end{equation*}
and
\begin{equation*}
N= \left[
\begin{array}{cccc|c}
& & & & 0 \\ &\hat{C} & & &\vdots \\ &&&&0 \\\hline
0& \ldots &0& 0 & 1\\
\end{array}
\right].
\end{equation*}
As a result, it follows that
\begin{equation*}
MDN= \left[\begin{array}{ccccc|c} 0 & & & & (-1)^{n+1}b^n &uv \\
1 & 0 & & & & \\
& 1 &0 & & & \\
& & \ddots &\ddots & &\\
& & &1& 0 & 0 \\ \hline
0 & & &0 & 1 & b \end{array} \right].
\end{equation*}
Switching the columns of $MDN $ yields the desired equivalent matrix.
\end{proof}
\chapter{Matrix Factorization}
\label{chapter:Matrix Factorization}
In this chapter, we discus the concept of a matrix factorization and their basic properties needed later in the rest of this thesis.
Matrix factorizations were introduced by David Eisenbud in \cite{ED} who proved that the MCM modules over hypersurfaces have a periodic resolutions.
\section{Definitions and Properties}
\label{section:Definitions and Properties}
\begin{definition}\cite[Definition 1.2.1]{DI}
\emph{Let $f$ be a nonzero element of a ring $S$. A matrix factorization of $f$ is a pair $(\phi,\psi)$ of homomorphisms
between finitely generated free $S$-modules $ \phi: G \rightarrow F$ and $\psi: F \rightarrow G $, such
that $\psi\phi = f I_G$ and $\phi\psi = f I_F$}.
\end{definition}
\begin{remark}\label{C3.2}
Let $f$ be a nonzero element of a commutative ring $S$. If $( \phi: G \rightarrow F,\psi: F \rightarrow G )$ is a matrix factorization of $f$, then:
\begin{enumerate}
\item [(a)] $f\Cok( \phi )= f \Cok( \psi )=0$.
\item [(b)] If $f$ is a non-zerodivisor, then $\phi$ and $\psi$ are injective.
\item [(c)] If $S$ is a domain, then $G$ and $F$ are finitely generated free modules having the same rank.
\end{enumerate}
\end{remark}
\begin{proof}
(a) Since $\psi\phi = f I_G$ and $\phi\psi = f I_F$, it follows that $fG \subseteq \Ima(\psi)$ and $fF\subseteq \Ima( \phi)$ which proves the result. \\
(b) Let $x \in G$ such that $\phi(x)=0$. Thus $ \psi(\phi(x))=0$ and hence $fx=0$. Since $f$ is a non-zerodivisor, it follows that $x=0$ and hence $\phi$ is injective. By similar argument, we prove that $\psi$ is injective.\\
(c) If $M= \Cok( \phi )$, then the following short sequence is exact.
\begin{equation}\label{E111}
0 \longrightarrow G\xrightarrow{\phi}F \rightarrow M \longrightarrow 0
\end{equation}
Recall that the rank of the finitely generated module $M$ over the domain $S$, denoted $rank_SM$, is the dimension of the vector space $K\otimes_SM$ over $K$ \cite[Section 11.6]{E} where $K$ is the quotient field of the integral domain $S$. We know from Proposition \ref{LLL2} (d) that $K\otimes_SM $ is isomorphic to $W^{-1}M$ as $S$-module where $W=S\setminus\{0\}$. Since $fM=0$, it follows that $K\otimes_SM=0$. Therefore, tensoring the short exact sequence (\ref{E111}) with $K$ over $S$ yields that $K\otimes_SG \simeq K\otimes_SF$ and thus $G$ and $F$ have the same rank as free $S$-modules.
\end{proof}
As a result, we can define the matrix factorization of a nonzero element $f$ in a domain as the following.
\begin{definition}\label{D23}
\emph{Let $S$ be a domain and let $f\in S$ be a nonzero element. A matrix factorization (of size $n$) is a pair $(\phi,\psi)$ of $n \times n$ matrices with
coefficients in $S$ such that $\psi\phi = \phi\psi = fI_n$ where $I_n$ is the identity matrix in $M_n(S)$. By $\Cok_S(\phi,\psi)$ and $\Cok_S(\psi,\phi)$, we mean $\Cok_S(\phi)$ and $\Cok_S(\psi)$ respectively. There are two distinguished trivial matrix factorizations of any element
$f$, namely $( f ,1)$ and $(1, f )$. Note that $\Cok_S(1, f ) = 0$, while $\Cok_S(f ,1)=S/fS$. Two matrix factorizations $(\phi,\psi)$ and $(\alpha,\beta)$ of $f$ are said to be equivalent (and we write $(\phi,\psi)\sim(\alpha,\beta)$) if $\phi,\psi,\alpha,\beta \in M_n(S)$ for some positive integer $n$ and there exist invertible matrices $V,W \in M_n(S)$ such that $V\phi = \alpha W$ and $ W \psi = \beta V$. If $(S,\mathfrak{m})$ is a local domain, a matrix factorization $(\phi,\psi)$ of an element $f \in \mathfrak{m}\setminus\{0\}$ is reduced if all entries of $\phi$ and $\psi$ are in $\mathfrak{m}$.}
\end{definition}
We can notice the following remark:
\begin{remark}\label{R3.8}
Let $(S,\mathfrak{m})$ be a local domain, $f \in \mathfrak{m} \setminus \{0\}$, $R=S/fS$ and let $ u$, $v$ and $z$ be variables on $S$.
Suppose that $(\phi,\psi)$ and $(\alpha,\beta)$ are two $n \times n$ matrix factorizations of $f$. Then
\begin{enumerate}
\item[(a)] $\Cok_S(\phi)$ and $\Cok_S(\psi)$ are both modules over the ring $R$ as $f\Cok_S(\phi)=0$ and $f\Cok_S(\psi) =0$.
\item[(b)] The $S$-linear maps $ \phi: S^n \rightarrow S^n$ and $\psi: S^n \rightarrow S^n $ are both injective.
\item[(c)] If $(\phi,\psi)\sim(\alpha,\beta)$, then $\Cok_S(\phi,\psi)$ is isomorphic to $\Cok_S(\alpha, \beta)$ over $S$ (and consequently over $R$), likewise,
$\Cok_S(\psi, \phi)$ is isomorphic to $\Cok_S(\beta, \alpha)$ over $S$ (and consequently over $R$).
\item[(d)] If $(\phi,\psi)$ and $(\alpha,\beta)$ are reduced matrix factorizations of $f$ such that $\Cok_S(\phi,\psi)$ is isomorphic to $\Cok_S(\alpha,\beta)$,
then $(\phi,\psi)\sim(\alpha,\beta)$.
\item[(e)] We define $(\phi,\psi)\oplus(\alpha,\beta):=(\phi \oplus \alpha, \psi \oplus \beta)$ and hence $(\phi,\psi)\oplus(\alpha,\beta)$ is a matrix factorization of $f$.
\item[(f)] We define $ (\phi, \psi)^{\maltese}:= ( \begin{bmatrix} \phi & -vI \\ uI & \psi \end{bmatrix} , \begin{bmatrix} \psi & vI \\ -uI & \phi \end{bmatrix})$ and hence $(\phi, \psi)^{\maltese}$
is a matrix factorization for $f+uv$ in $S[\![u,v]\!]$ (and in $S[u,v]$). Furthermore, if $(\phi,\psi)\sim(\alpha,\beta)$, then $(\phi,\psi)^{\maltese}\sim(\alpha,\beta)^{\maltese}$.
\item[(g)] $[(\phi, \psi)\oplus(\alpha,\beta)]^{\maltese}$ is equivalent to $ (\phi, \psi)^{\maltese}\oplus(\alpha,\beta)^{\maltese}$.
\item[(h)] We define $ [\Cok_S(\phi, \psi)]^{\maltese}=\Cok_{S[[u,v]]}(\phi, \psi)^{\maltese}$ and hence if $(\phi_j,\psi_j)$ is a matrix factorization of $f$ for all $1 \leq j \leq n$, then $$[ \bigoplus\limits_{j=1}^{n}\Cok_S(\phi_j,\psi_j)]^{\maltese} =\bigoplus \limits_{j=1}^{n}\Cok_{S[\![u,v]\!]}(\phi_j,\psi_j)^{\maltese}.$$
\item[(i)] If $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$, then $\Cok_{S[\![u,v]\!]}(f, 1)^{\maltese}=R^{\bigstar}= \Cok_{S[\![u,v]\!]}(1, f)^{\maltese}$ and hence we can write $(R)^{\maltese}= R^{\bigstar}$ as $R=\Cok_S(f,1)$.
\item[(j)] We define $ (\phi, \psi)^{\sharp}:= ( \begin{bmatrix} \phi & -zI \\ zI & \psi \end{bmatrix} ,
\begin{bmatrix} \psi & zI \\ -zI & \phi \end{bmatrix})$, and hence $ (\phi, \psi)^{\sharp}$ is a matrix factorization of $f+z^2$ in $S[\![z]\!]$ (and in $S[z]$).
\item[(k)] If $(\phi,\psi)\sim(\alpha,\beta)$, then $(\phi,\psi)^{\sharp}\sim(\alpha,\beta)^{\sharp}$.
\item[(l)] $ [(\phi, \psi)\oplus(\alpha,\beta)]^{\sharp}$ is equivalent to $ (\phi, \psi)^{\sharp}\oplus(\alpha,\beta)^{\sharp}$.
\item[(m)] If $R^{\sharp}=S[\![z]\!]/(f+z^2)$, then $$R^{\sharp}=S[\![z]\!]/(f+z^2)= \Cok_{S[\![z]\!]}(f, 1)^{\sharp} = \Cok_{S[\![z]\!]}(1,f)^{\sharp}.$$
\end{enumerate}
\end{remark}
\begin{proof}
We will just prove the result (d) and (f) as the rest follows from the definitions. \\
(d) Assume that $(\phi,\psi)$ and $(\alpha,\beta)$ are reduced matrix factorizations of $f$ such that $\Cok_S(\phi,\psi)$ is isomorphic to $\Cok_S(\alpha,\beta)$. This means that $\Cok_S(\phi)$ is isomorphic to $\Cok_S(\alpha)$ and consequently from Proposition \ref{L3.5} it follows that $V\phi = \alpha W$ for invertible matrices $V$ and $W$. Therefore, we have $V\phi \psi = \alpha W \psi$ and thus $V (fI) = \alpha W \psi$. As a result, we get $\beta V (fI) = \beta \alpha W \psi = f I W \psi$ and consequently $ f \beta V= f W \psi$. Since $f$ is an element of the integral domain $S$, we conclude that $ \beta V= W \psi$. This proves that $(\phi,\psi)$ is equivalent to $(\alpha,\beta)$.
(f) Assume that $V\phi = \alpha W$ and $ W \psi = \beta V$ for invertible matrices $V$ and $W$. It follows that
\begin{equation*}
\left[
\begin{array}{cc}
V & \\
& W \\
\end{array}
\right] \left[
\begin{array}{cc}
\phi & -vI \\
uI & \psi \\
\end{array}
\right] = \left[
\begin{array}{cc}
\alpha & -vI \\
uI & \beta \\
\end{array}
\right]\left[
\begin{array}{cc}
W & \\
& V \\
\end{array}
\right]
\end{equation*}
and
\begin{equation*}
\left[
\begin{array}{cc}
W & \\
& V \\
\end{array}
\right] \left[
\begin{array}{cc}
\psi & vI \\
-uI & \phi \\
\end{array}
\right] = \left[
\begin{array}{cc}
\beta & vI \\
-uI & \alpha \\
\end{array}
\right]\left[
\begin{array}{cc}
V & \\
& W \\
\end{array}
\right].
\end{equation*}
This proves that the matrix factorization $ ( \left[
\begin{array}{cc}
\phi & -vI \\
uI & \psi \\
\end{array} \right], \left[
\begin{array}{cc}
\psi & vI \\
-uI & \phi \\
\end{array}
\right] )$ is equivalent to the matrix factorization $$ ( \left[
\begin{array}{cc}
\alpha & -vI \\
uI & \beta \\
\end{array} \right] , \left[
\begin{array}{cc}
\beta & vI \\
-uI & \alpha \\
\end{array}
\right] )$$ and therefore $(\phi,\psi)^{\maltese}\sim(\alpha,\beta)^{\maltese}.$
\end{proof}
\begin{notation}
\emph{ If $(\phi,\psi)$ is a matrix factorization, we write $$(\phi,\psi)^n=\underbrace{(\phi,\psi)\oplus...\oplus(\phi,\psi)}_{\text{$n$ times}}.$$}
\end{notation}
\begin{definition}
\emph{Let $f$ be a nonzero element of a domain. A matrix factorization $(\phi,\psi)$ of $f$ is called trivial if it is equivalent to one of the following forms:
\begin{equation*}
(f,1)^n, (1,f)^n, \text{ or } (f,1)^r \oplus (1,f)^t
\end{equation*}
where $n$ is the size of $(\phi,\psi)$ and $0 < r,s < n$ with $ r+s=n$.}
\end{definition}
\section{Matrix factorization and Maximal Cohen Maculay modules}
\label{section:Matrix factorization and Maximal Cohen Maculay modules}
The importance of the concept of the matrix factorization in the subject of MCM modules appears clearly in the following proposition.
\begin{proposition}\label{P25} \cite[Proposition 8.3]{CMR}
Let $(S,\mathfrak{m})$ be a regular local ring and let $f$ be a non-zero
element of $\mathfrak{m}$ and $R=S/fS$ .
\begin{enumerate}
\item[(a)] For every MCM $R$-module $M$, there is a matrix factorization $(\phi , \psi)$ of $f$
with $\Cok\phi \cong M$. \\
\item[(b)] If $(\phi,\psi)$ is a matrix factorization of $f$, then $\Cok\phi$ and $\Cok\psi$ are MCM
$R$-modules.
\end{enumerate}
\end{proposition}
\begin{proof}
Let $\dim R= d$. \\
(a) By \cite[Corollary 11.18]{AM} it follows that $\dim S=d+1$. Notice that $M$ can be viewed as $S$-module such that every $M$-regular sequence on $R$ is $M$-regular sequence on $S$ and every $M$-regular sequence on $S$ is also $M$-regular sequence on $R$. As a result, $\depth_SM=\depth_RM$. Since $M$ is finitely generated module over the regular local ring $S$, it follows from Auslander-Buchsbaum-Serre Theorem \cite[Theorem2.2.7]{BH} that $\pd_SM < \infty $. Auslander-Buchsbaum formula \ref{P2.23} implies that
\begin{equation*}
\pd_SM=\depth S-\depth_SM=\dim S-\depth_RM=\dim S-\dim R=1.
\end{equation*}
Therefore, there exist projective $S$-modules $F$ and $G$, and consequently they are free $S$-modules \cite[Proposition 18.4.1]{BCA}, such that the following sequence is exact
\begin{equation*}
0 \longrightarrow G\xrightarrow{\phi}F \xrightarrow{\lambda} M \longrightarrow 0.
\end{equation*}
For each $x \in F$, we notice that $\lambda(fx)=0$ and consequently $fF \subseteq \Ker \lambda = \Ima \phi$. As $\phi$ is injective, for each $x \in F$ there exists a unique element $y \in G$ such that $ \phi( y )=fx$. This enables us to define a map $ \psi: F \rightarrow G$ via $ \psi(x)= y$ if and only if $ \phi( y )=fx$. It is easy to check that $\psi$ is an injective homomorphism over $S$ satisfying that $\psi \phi = fI_G$ and $\psi\phi=fI_F$ where $I_F$ and $I_G$ denote the identity maps on $F$ and $G$ respectively. Therefore, $M=\Cok(\phi,\psi)$. \\
(b) If $(\phi,\psi)$ is a matrix factorization of $f$, then there exist two free $S$-modules $F$ and $G$ having the same rank and making the following short sequences exact over $S$
\begin{equation*}
0 \longrightarrow G\xrightarrow{\phi}F \rightarrow \Cok \phi \longrightarrow 0
\end{equation*}
and
\begin{equation*}
0 \longrightarrow G\xrightarrow{\psi}F \rightarrow \Cok \psi \longrightarrow 0.
\end{equation*}
This means that $\pd_S ( \Cok \phi )= 1 = \pd_S ( \Cok \psi )$. Since $\depth_S S=\dim S $, it follows from Auslander-Buchsbaum formula \ref{P2.23} that
\begin{equation*}
\depth_R \Cok \phi= \depth_S \Cok \phi = \depth_SS-\pd_S ( \Cok \phi )=\dim S-1=\dim R
\end{equation*}
and
\begin{equation*}
\depth_R \Cok \psi= \depth_S \Cok \psi = \depth_SS-\pd_S ( \Cok \psi )=\dim S-1=\dim R.
\end{equation*}
This proves that $\Cok \phi$ and $\Cok \psi$ are MCM $R$-modules.
\end{proof}
\begin{definition} \emph{Let $R$ be a ring, a non-zero $R$-module $M$ is called a stable $R$-module if $M$ does not have
a direct summand isomorphic to $R$.}
\end{definition}
D. Eisenbud has established a relationship between reduced matrix factorizations and stable MCM modules as follows.
\begin{proposition}\label{A}\cite[ Corollary 7.6]{YY} \cite[Theorem 8.7]{CMR}
Let $(S,\mathfrak{m})$ be a regular local ring and let $f$ be a non-zero
element of $\mathfrak{m}$ and $R=S/fS$. Then
the association $ (\phi, \psi)\mapsto \Cok(\phi, \psi)$ yields a bijective correspondence between the
set of equivalence classes of reduced matrix factorizations of f and the set of isomorphism
classes of stable MCM modules over R.
\end{proposition}
\begin{remark}
Let $(S,\mathfrak{m})$ be a regular local ring. If $f$ is a non-zero
element of $\mathfrak{m}$, let $R=S/fS$ and $R^{\bigstar}:=S[\![u,v]\!]/(f+uv)$. If $M$ is a stable MCM $R$-module, then $M=\Cok_S(\phi,\psi)$ where $(\phi,\psi)$ is a reduced matrix factorization of $f$. This enables us to define $M^{\maltese}=\Cok_{S[\![u,v]\!]}(\phi,\psi)^{\maltese}$. Indeed, if $M=\Cok_S(\alpha,\beta)$ for some matrix factorization $(\alpha,\beta)$ of $f$, then by Proposition \ref{A} $(\alpha,\beta)$ is reduced matrix factorization of $f$ and $(\alpha,\beta) \sim (\phi,\psi)$. According to Remark \ref{R3.8} (f), $(\alpha,\beta)^{\maltese} \sim (\phi,\psi)^{\maltese}$. This shows that $\Cok_{S[\![u,v]\!]}(\alpha,\beta)^{\maltese} \cong \Cok_{S[\![u,v]\!]}(\phi,\psi)^{\maltese}$. Therefore, the association $M\rightarrow M^{\maltese}$ is well defined on the class of stable MCM $R$-modules.
\end{remark}
If $R$ and $R^{\bigstar}$ are as in Remark \ref{R3.8} , the indecomposable non-free MCM modules over $R$ and $R^{\bigstar}$ can be related in the following situation.
\begin{proposition}\label{P28} \cite[Theorem 8.30]{CMR}
Let $(S, \mathfrak{m}, K)$ be a complete regular local ring such that $K$ is algebraically closed of characteristic not $2$ and $f\in \mathfrak{m}^2 \smallsetminus \{0\}$. If $R=S/fS$ and $R^{\bigstar}:=S[\![u,v]\!]/(f+uv)$, then the association $M\rightarrow M^{\maltese}$ defines a bijection between the isomorphisms classes of indecomposable non-free MCM modules over $R$ and $R^{\bigstar}$.
\end{proposition}
Any matrix factorization has a decomposition as follows.
\begin{proposition}\label{P.24}(cf.\cite[Result 7.5.2]{YY})
Let $(S,\mathfrak{m})$ be a local domain and $f \in \mathfrak{m} \smallsetminus \{0\}$.
If $(\phi,\psi)$ is a nontrivial matrix factorization of $f$ of size $n$, then $(\phi,\psi)$ can be written as
\begin{equation*}\label{E4}
(\phi,\psi) = (\alpha,\beta)\oplus (f,1)^t\oplus (1,f)^r
\end{equation*}
where $(\alpha,\beta)$ is a reduced matrix factorization of $f$ and $0\leq t, r < n$. Furthermore, if $(S,\mathfrak{m})$ is a regular local ring, the above decomposition is unique up to equivalence.
\end{proposition}
\begin{proof}
By the induction on the size of the matrix factorization we will prove the result.
The case when the size is one is obvious. Suppose that $(\phi,\psi)$ is a matrix factorization of $f$ of size $(n+1) $. If $(\phi,\psi)$ is reduced, we are done. Without lose of generality, assume that one entry of $\phi$ is a unite. Using row and column operations, there exist invertible matrices $U,V$ in $M_n(S)$ such that
\begin{equation*}
U \phi V= \left[
\begin{array}{cc}
1 & \\
& \tilde{\phi} \\
\end{array}
\right]
\end{equation*}
where $\tilde{\phi}$ is $n \times n$ matrix. Set $\hat{\psi}= V^{-1} \psi U^{-1}$ and notice that $ \hat{\psi}\left[
\begin{array}{cc}
1 & \\
& \tilde{\phi} \\
\end{array}
\right]=fI$ and $ \left[
\begin{array}{cc}
1 & \\
& \tilde{\phi} \\
\end{array}
\right]\hat{\psi}=fI$. This makes $\hat{\psi}= \left[ \begin{array}{cc}
f & \\
& \tilde{\psi} \\
\end{array}\right]$ where $(\tilde{\phi},\tilde{\psi})$ is a matrix factorization of $f$. Therefore, $(\phi,\psi) \sim (\tilde{\phi},\tilde{\psi}) \oplus (1,f)$. If $(\tilde{\phi},\tilde{\psi})$ is reduced, we get the desired result. Otherwise, apply the induction hypothesis on $(\tilde{\phi},\tilde{\psi})$ to completes the proof.
Now assume that $(S,\mathfrak{m})$ is a regular local ring, $R=S/fS$, and let $(\phi,\psi)$ be a nontrivial matrix factorization of $f$ of size $n$. Suppose
that $(\phi,\psi)$ can be written as $ (\alpha_j,\beta_j)\oplus (f,1)^{t_j}\oplus (1,f)^{r_j}$ where $(\alpha_j,\beta_j)$ is a reduced matrix factorization of $f$ for $j=1,2$. This makes $\widehat{(M_1)}_{\mathfrak{m}}\oplus (\widehat{R}_{\mathfrak{m}})^{\oplus t_1}=\widehat{(M_2)}_{\mathfrak{m}}\oplus (\widehat{R}_{\mathfrak{m}})^{\oplus t_2}$ where $M_j=\Cok_j(\alpha_j,\beta_j)$ for $j=1,2$. Since $M_j$ has no free direct summands (Proposition \ref{A}), it follows from Proposition \ref{Pro2.22}(a) that $\widehat{(M_j)}_{\mathfrak{m}}$ has no free direct summands where $j=1,2$. Therefore, by Krull-Remak-Schmidt theorem (see discussion \ref{disc2.33}) $t_1=t_2$. Since $(\psi,\phi)$ can be written as $ (\beta_j,\alpha_j) \oplus (1,f)^{r_j}\oplus (f,1)^{t_j}$, it follows from a similar argument that $r_1=r_2$. Furthermore, Krull-Remak-Schmidt theorem (see discussion \ref{disc2.33}) implies that $\widehat{(M_1)}_{\mathfrak{m}} \cong \widehat{(M_2)}_{\mathfrak{m}}$ and consequently
\begin{equation*}
\Cok_S(\alpha_1,\beta_1) =M_1 \cong M_2= \Cok_S(\alpha_2,\beta_2) \text{ (see Proposition \ref{Pro2.22} (b))}.
\end{equation*}
Since $(\alpha_1,\beta_1)$ and $(\alpha_2,\beta_2)$ are reduced with $\Cok_S(\alpha_1,\beta_1) \cong \Cok_S(\alpha_2,\beta_2)$, it follows from Remark \ref{R3.8} (d) that $(\alpha_1,\beta_1)$ and $(\alpha_2,\beta_2)$ are equivalent. This finishes the proof.
\end{proof}
One can use Proposition \ref{A} and Remark \ref{R3.8} to show the following corollary.
\begin{corollary}\label{C3.8}
Let $(S,\mathfrak{m})$ be a regular local ring , $f \in\mathfrak{ m} \smallsetminus \{0\}$, $R=S/fS$, $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$ , and $R^{\sharp}=S[\![z]\!]/(f+z^2)$ where $u,v$ and $z$ are variables over $S$ .
If $(\phi,\psi)$ is a matrix factorization of $f$ having the decomposition
\begin{equation*}\label{E44}
(\phi,\psi) = (\alpha,\beta)\oplus (f,1)^t \oplus (1,f)^r
\end{equation*}
where $(\alpha,\beta)$ is reduced and $t, r$ are non-negative integers, then :
\begin{enumerate}
\item[(a)] $\sharp (\Cok_S(\phi,\psi),R)=t$ and $\sharp (\Cok_S(\psi,\phi),R)=r$.
\item[(b)] $\sharp (\Cok_{S[\![u,v]\!]}(\phi,\psi)^{\maltese},R^{\bigstar})=\sharp (\Cok_S(\phi,\psi),R)+\sharp (\Cok_S(\psi,\phi),R)$.
\item[(c)] $\sharp (\Cok_{S[\![z]\!]}(\phi,\psi)^{\sharp},R^{\sharp})=\sharp (\Cok_S(\phi,\psi),R)+\sharp (\Cok_S(\psi,\phi),R)$.
\end{enumerate}
\end{corollary}
\begin{proof}
(a) It is obvious that $ \Cok( f,1)=R$ but $\Cok(1,f)=\{0\}$ and hence $\Cok(\phi,\psi)=\Cok(\alpha,\beta)\oplus R^t$. Since $(\alpha,\beta)$ is reduced, it follows from Proposition \ref{A} that $\Cok(\alpha,\beta)$ is stable. This implies by Discussion \ref{disc2.33}(b) that $\sharp (\Cok(\phi,\psi),R)=t$. Now, since $(\psi,\phi) = (\beta, \alpha)\oplus (f,1)^r \oplus (1,f)^t$, it follows from what we have proved that $\sharp (\Cok(\psi,\phi),R)=r$.\\
(b) Notice by Remark \ref{R3.8} (i) that $\Cok(f,1)^{\maltese} = R^{\bigstar} =\Cok(1,f)^{\maltese}$. This makes
\begin{equation*}
\Cok(\phi,\psi)^{\maltese}= \Cok(\alpha,\beta)^{\maltese} \oplus (R^{\bigstar})^t \oplus (R^{\bigstar})^r.
\end{equation*}
Since $(\alpha,\beta)^{\maltese}$ is a reduced matrix factorization of $f+uv$, it follows from Proposition \ref{A} that $\Cok(\alpha,\beta)^{\maltese}$ is stable $R^{\bigstar}$-module. This proves, by Discussion \ref{disc2.33}(b), that
$$\sharp (\Cok(\phi,\psi)^{\maltese},R^{\bigstar})=t+r=\sharp (\Cok(\phi,\psi),R)+\sharp (\Cok(\psi,\phi),R).$$
(c) Similar argument to the above argument.
\end{proof}
Krull-Remak-Schmidt theorem (see discussion \ref{disc2.33}) and Proposition \ref{Pro2.22} enable us to establish the following Proposition.
\begin{proposition}\label{C3.14}
Let $(S,\mathfrak{m})$ be a regular local ring, $f \in \mathfrak{m} \smallsetminus \{0\}$, and $R=S/fS$. If $(\phi, \psi)$ is a reduced matrix factorization of $f$, then $$(\phi, \psi)\sim [(\phi_1,\psi_1)\oplus (\phi_2,\psi_2)\oplus \dots\oplus(\phi_n,\psi_n)]$$ where $(\phi_i,\psi_i)$ is a reduced matrix factorization of $f$ with $\Cok_S(\phi_i,\psi_i)$ is non-free indecomposable MCM $R$-module for all $ 1 \leq i \leq n$. Furthermore, the above representation of $(\phi, \psi)$ is unique up to equivalence when $S$ is also complete.
\end{proposition}
\begin{proof}
By Proposition \ref{A}, $M:=\Cok_S(\phi , \psi )$ is stable MCM $R$-module. We may assume by Discussion \ref{disc2.33}(a) that $M=M_1\oplus \dots \oplus M_n$ where $M_j$ is a non-free indecomposable MCM $R$-module for each $ 1 \leq j \leq n$ (Corollary \ref{C2.30}). Again by Proposition \ref{A}, we have $M_j=\Cok_S(\phi_j,\psi_j)$ for some reduced matrix factorization $(\phi_j,\psi_j)$ of $f$ for all $ 1 \leq j \leq n$. As a result, $\Cok_S(\phi , \psi )$ is isomorphic to $\Cok_S[(\phi_1,\psi_1)\oplus (\phi_2,\psi_2)\oplus \dots\oplus(\phi_n,\psi_n)]$ and hence by Remark \ref{R3.8}(d), $(\phi, \psi)\sim [(\phi_1,\psi_1)\oplus (\phi_2,\psi_2)\oplus \dots\oplus(\phi_n,\psi_n)]$. Now if $(\phi, \psi)\sim [(\alpha_1,\beta_1)\oplus (\alpha_2,\beta_2)\oplus \dots\oplus(\alpha_m,\beta_m)]$ is another representation of $(\phi, \psi)$ and $N_j= \Cok_S(\alpha_j,\beta_j)$ where $(\alpha_j,\beta_j)$ is a matrix factorization of $f$ for all $1 \leq j \leq m$, then $ \bigoplus_{i=1}^n M_i$ is isomorphic to $ \bigoplus_{j=1}^m N_j$ as $R$-modules. By Krull-Remak-Schmidt theorem (Discussion \ref{disc2.33}(a)), $n=m$ and, after renumbering, $M_i\cong N_i$ for each $i$. Therefore, by Proposition \ref{Pro2.22} (b) and hence by Remark \ref{R3.8}(d) $(\phi_i,\psi_i) \sim (\alpha_i, \beta_i)$ for each $i$.
\end{proof}
\begin{proposition}\label{P3.15}
Let $(S, \mathfrak{m}, K)$ be a complete regular local ring such that $K$ is algebraically closed of characteristic not $2$, and $f\in \mathfrak{m}^2 \smallsetminus \{0\}$. let $R=S/fS$ and $R^{\bigstar}:=S[\![u,v]\!]/(f+uv)$. If $(\phi, \psi)$ and $(\alpha, \beta)$ are reduced matrix factorizations of $f$, then $\Cok_S(\phi, \psi)$ is isomorphic to $\Cok_S(\alpha, \beta)$ over $R$ if and only if $\Cok_{S[\![u,v]\!]}(\phi, \psi)^{\maltese}$ is isomorphic to $\Cok_{S[\![u,v]\!]}(\alpha,\beta)^{\maltese}$ over $R^{\bigstar}$.
\end{proposition}
\begin{proof}
If $\Cok_S(\phi, \psi)$ is isomorphic to $\Cok_S(\alpha, \beta)$, by Remark \ref{R3.8} (d), $(\phi, \psi)\sim(\alpha, \beta)$ and consequently $(\phi, \psi)^{\maltese}\sim(\alpha, \beta)^{\maltese}$. This shows that $\Cok_{S[[u,v]]}(\phi, \psi)^{\maltese}$ is isomorphic to $\Cok_{S[[u,v]]}(\alpha,\beta)^{\maltese}$ over $R^{\bigstar}$.
Assume now that $\Cok_{S[\![u,v]\!]}(\phi, \psi)^{\maltese}$ is isomorphic to $\Cok_{S[\![u,v]\!]}(\alpha,\beta)^{\maltese}$ over $R^{\bigstar}$.
Using Proposition \ref{C3.14} we see that $(\phi, \psi)\sim [(\phi_1,\psi_1)\oplus (\phi_2,\psi_2)\oplus \dots\oplus(\phi_n,\psi_n)]$
and $(\alpha, \beta)\sim [(\alpha_1,\beta_1)\oplus (\alpha_2,\beta_2)\oplus \dots\oplus(\alpha_t,\beta_t)]$ where $(\phi_i,\psi_i)$
and $(\alpha_j,\beta_j)$ are reduced matrix factorizations of $f$ satisfying that $\Cok_S(\phi_i,\psi_i)$ and $\Cok_S(\alpha_j,\beta_j)$ are
non-free indecomposable MCM $R$-module for all $ i $ and $ j $.
Remark \ref{R3.8} (f),(g) and (c) imply that
$$\Cok_{S[\![u,v]\!]}(\phi, \psi)^{\maltese}= \bigoplus _{j=1}^n \Cok_{S[\![u,v]\!]}(\phi_j, \psi_j)^{\maltese} $$
and $$\Cok_{S[\![u,v]\!]}(\alpha,\beta)^{\maltese}= \bigoplus_{i=1}^t \Cok_{S[\![u,v]\!]}(\alpha_i,\beta_i)^{\maltese}.$$
Notice that Proposition \ref{P28} implies that $\Cok_{S[\![u,v]\!]}(\phi_j, \psi_j)^{\maltese}$ and $\Cok_{S[\![u,v]\!]}(\alpha_i,\beta_i)^{\maltese}$ are indecomposable non-free MCM modules over $R^{\bigstar}$ for all $i$ and $j$. Now Krull-Remak-Schmidit Theorem (Discussion \ref{disc2.33}(a)) gives that $n=t$ and after renumbering we get $\Cok_{S[\![u,v]\!]}(\phi_i, \psi_i)^{\maltese}$ is isomorphic to $\Cok_{S[\![u,v]\!]}(\alpha_i,\beta_i)^{\maltese}$ over $R^{\bigstar}$ and consequently
Proposition \ref{P28} implies that $\Cok_S(\phi_i, \psi_i)$ is isomorphic to $\Cok_S(\alpha_i,\beta_i)$ over $R$ for all $1 \leq i \leq n$. Therefore, by Remark \ref{R3.8}(d) $(\phi_i, \psi_i)\sim (\alpha_i,\beta_i)$. As a result, it follows that
$ (\phi , \psi ) \sim \bigoplus_{j=1}^n (\phi_j , \psi_j )\sim \bigoplus_{j=1}^n (\alpha_j ,\beta_j ) \sim (\alpha , \beta) $ and hence $\Cok_S(\phi , \psi )$
is isomorphic to $\Cok_S(\alpha, \beta )$ as desired.
\end{proof}
\chapter{Modules of finite F-representation type }
\label{Chapter: Modules of finite Frepresentation type}
In this chapter, all rings are Noetherian of prime characteristic $p$ unless otherwise stated.
\section{Definition and examples}
\label{section:Definition and examples}
The notion of finite F-representation type was introduced by K. Smith
and M. Van den Bergh in \cite{SV} for $F$-finite rings over which the Krull-Remak-Schmidit Theorem is satisfied, i.e, they defined the notion of finite F-representation type for an $F$-finite ring $R$ with the property that every finitely generated $R$-module can be written uniquely up to isomorphism as a direct sum of finitely many indecomposable
$R$-modules. However, Y.Yao in \cite{Y} generalized this notion to be defined for finitely generated modules over Noetherian rings of prime characteristic $p$. After that, S. Takagi and R. Takahashi in \cite{TT} and T.Shibuta in \cite{TS} studied this notion under the general assumption made by Y.Yao that the ring is just Noetherian of prime characteristic $p$. In this thesis, we also adapt the same definition of this notion under the same general assumptions for the ring as they appear in \cite[Definition1.1]{Y} and \cite[Definition 2.1]{TS}.
\begin{definition}\label{D1}
\emph{Let $R$ be a ring.
If $M$ , $M_1$, ... , $M_s$ are finitely generated $R$-modules, then $M$ is said to have finite F-representation type (henceforth abbreviated FFRT) by the $R$-modules $M_1$, ... , $M_s$ if for every positive integer $e$, the $R$-module $F_*^e(M)$ is isomorphic
to a finite direct sum of the $R$-modules $M_1,\ldots ,M_s$, that is, there exist
nonnegative integers $t_{(e,1)}, \ldots , t_{(e,s)}$ such that
$$F_*^e(M)= \bigoplus_{j=1}^s M_j^{\oplus t_{(e,j)}}.$$
In particular, $R$ is said to have finite F-representation type if there exist finitely
generated $R$-modules $M_1,\ldots,M_s$ by which $R$ has finite F-representation
type.}
\end{definition}
S. Takagi and R. Takahashi exhibit in \cite{TT} examples of rings with finite F-representation type.
\begin{example}\label{Ex FFRT}
A ring $R$ has finite F-representation type in the following cases:
\begin{enumerate}
\item [(a)] $R=K[x_1,...,x]$ or $R=K[\![x_1,...,x]\!]$ where $K$ is a field of prime characteristic $p$ with $[K:K^p]< \infty$ (Corollary \ref{L2.5}).
\item [(b)](\cite[Observation 3.1.2]{SV}) $R$ is an complete $F$-finite regular local
ring of prime characteristic $p > 0$ with $[K:K^p]< \infty$ where $K$ is the residue field of $R$.
\item [(c)](\cite[Observation 3.1.3]{SV}) $R$ is a Cohen-Macaulay $F$-finite local ring of prime characteristic $p$ with finite Cohen-Macaulay type.
\item [(d)]\cite[Example1.3(ii) ]{TT} $R$ is an Artinian $F$-finite local ring of prime characteristic $p$ with $[K:K^p]< \infty$ where $K$ is the residue field of $R$.
\item [(e)]\cite[Theorem 1]{TS} $R$ is a complete local one-dimensional domain
of prime characteristic such that its residue field is algebraically closed or finite.
\end{enumerate}
\end{example}
T.Shibuta in his paper \cite{TS} presents examples \cite[Example 3.3 and Example 3.4]{TS} of a complete local
one-dimensional domains which do not have finite F-representation type with a
perfect residue field.
\begin{remark}
Notice from the definitions \ref{D1} and \ref{D2.6} that a Noetherian ring that has FFRT is F-finite. As a result, if $R$ is a Noetherian ring that is not F-finite, then $R$ does not have FFRT. For example, if $K=\mathbb{Z}/p\mathbb{Z}$ where $p$ is a prime integer and $R=K(x_1,x_2,...,x_n,...)$, then $F_*^e(R)$ is $R$-vector space with infinite basis (Corollary \ref{L2.5}(b)). This proves that $R$ is not F-finite and hence $R$ does not have $FFRT$ (K.Schwede and W.Zhang observed in \cite[Section 2]{SZ} that $R$ is not F-finite).
\end{remark}
\section{Several FFRT extensions of FFRT rings}
\label{section:Several FFRT extensions of FFRT rings}
If $R$ is a ring that has FFRT, Y.Yao in \cite{Y} observed that localizations and completions of $R$ both have FFRT. In this section, we prove this observation and that each of $R[x]$ and $R[\![x]\!]$ has FFRT.
\begin{proposition}\label{P4.2}
Let $R$ be ring and $M$ be a finitely generated $R$-module. Assume that $M$ has FFRT on $R$. If $W$ is a multiplicative closed set and $I$ is an ideal of $R$, then:
\begin{enumerate}
\item [(a)] $W^{-1}M$ has FFRT on $W^{-1}R$.
\item [(b)] $\widehat{M}_I$ has FFRT on $\widehat{R}_I$.
\end{enumerate}
\end{proposition}
\begin{proof}
Suppose that $M$ has FFRT by finitely generated
$R$-modules $M_1$, ... , $M_s$. If $e \in \mathbb{Z}^{+}$, there exist
nonnegative integers $n_{(e,1)}, \ldots , n_{(e,s)}$ such that
$$F_*^e(M)= \bigoplus_{j=1}^s M_j^{\oplus n_{(e,j)}}. $$
It follows from Proposition \ref{L2.3} and Proposition \ref{LLL2} that
\begin{equation*}
F_*^e(W^{-1}M) = W^{-1}F_*^e(M)=\bigoplus_{j=1}^s (W^{-1}M_j)^{\oplus n_{(e,j)}}.
\end{equation*}
As a result, $W^{-1}M$ has FFRT on $W^{-1}R$ by the finitely generated
$W^{-1}R$-modules $W^{-1}M_1$, ... , $W^{-1}M_s$. Furthermore, by Proposition \ref{L2.4} and Theorem \ref{thm 2.21}, we get that
\begin{equation*}
F_*^e(\widehat{M}_I) = \widehat{F_*^e(M)}_I= \bigoplus_{j=1}^s (\widehat{(M_j)}_I)^{\oplus n_{(e,j)}}.
\end{equation*}
Therefore, we conclude that $\widehat{M}_I$ has FFRT on $\widehat{R}_I$ by the finitely generated
$\widehat{R}_I$-modules $\widehat{M_1}_I$, ... , $\widehat{M_s}_I$.
\end{proof}
\begin{proposition}\label{P4.5}
Let $R$ be a ring and let $M$ be a finitely generated $R$-module.
If $M$ has FFRT over $R$, then $M[x]$ has FFRT over $R[x]$.
\end{proposition}
\begin{proof}
If $e \in \mathbb{Z}^{+}$ and $q=p^e$, recall from Proposition \ref{P2.41} that $$ F_*^e(M[x])= (F_*^e(M)[x])^{\oplus^{q}}.$$
Since $M$ has FFRT, there exist finitely generated $R$-modules $M_1,...,M_s$ and nonnegative integers $n_{(e,1)},...,n_{(e,s)}$ such that $$F_*^e(M)=\bigoplus_{j=1}^s M_j^{\oplus n_{(e,j)}}.$$
Tensoring with $R[x]$ both sides of the above equality and using the Remark \ref{R4.3} yield that $$F_*^e(M)[x] =\bigoplus_{j=1}^s (M_j[x])^{\oplus n_{(e,j)}}.$$ As a result, it follows that $$F_*^e(M[x]) =\bigoplus_{j=1}^s (M_j[x])^{\oplus qn_{(e,j)}}.$$
Therefore, we conclude that $ M[x]$ has FFRT by the finitely generated $R[x]$-modules $M_1[x],...,M_s[x]$.
\end{proof}
\begin{theorem}\label{C4.6}
Let $R$ be a ring of prime characteristic $p$. If $R$ has FFRT, then:
\begin{enumerate}
\item [(a)] $R[x]$ has FFRT over $R[x]$.
\item [(b)] $R[\![x]\!]$ has FFRT over $R[\![x]\!]$.
\item[(c)] $R[x_1,\ldots,x_n]$ has FFRT over $R[x_1,\ldots,x_n]$.
\item[(d)] $R[\![x_1,\ldots,x_n]\!]$ has FFRT over $R[\![x_1,\ldots,x_n]\!]$.
\end{enumerate}
\end{theorem}
\begin{proof}
(a) This is a particular case of Proposition \ref{P4.5}.\\
(b) Recall from Proposition \ref{EX1} that if $I=(x)R[x]$, then $\widehat{R[x]}_I= R[\![x]\!]$. Now apply the above result and Proposition \ref{P4.2}.\\
(c) and (d) follows from (a) and (b).
\end{proof}
\section{Modules of FFRT by FFRT system}
\label{section: Rings of FFRT by FFRT system}
The notion of FFRT system was introduced by Yao in his paper \cite{Y} as follows:
\begin{definition}
\emph{A finite set $\Gamma$ of finitely generated $R$-modules is said to be a finite $F$-representation type System (or FFRT system) if for every $N \in \Gamma$, $F_*^1(N)$ can be written as a finite direct sum whose direct summands are all taken from $\Gamma$.
We say that $M$ has FFRT by a FFRT system if there exists a FFRT system $\Gamma $ such that $M$ has FFRT by $\Gamma $.}
\end{definition}
Let $R$ be a Noetherian ring not necessarily of prime characteristic. A class $\textit{C}(R)$ of finitely generated $R$-modules is called reasonable if it satisfies that every $R$-module
that is isomorphic to a direct summand of a module in $\textit{C}(R)$ is in $\textit{C}(R)$. The notion of reasonable class was introduced by R.Wiegand in his paper \cite[Section 1]{RW} who proved the following useful theorem.
\begin{proposition}\label{T1}\cite[Theorem 1.4.]{RW}
Let $R$ be a Noetherian semilocal ring not necessarily of prime characteristic,i.e. $R$ has finitely many maximal ideals, and let $S$ be a faithfully flat $R$-algebra (Definition \ref{D2.7}). Let $\textit{C}(R)$ and $\textit{C}(S)$ be reasonable classes of modules such that
$S\otimes_{R}M \in \textit{C}(S)$ for all $M \in \textit{C}(R)$. If $\textit{C}(S)$ contains only finitely many
indecomposable modules up to isomorphism, the same holds for $\textit{C}(R)$.
\end{proposition}
\begin{theorem}\label{P1}
Let $R$ be a local ring. If a finitely generated $R$-module $M$ has FFRT, then $M$ has FFRT by a FFRT system.
\end{theorem}
\begin{proof}
Let $\mathfrak{m}$ be the maximal ideal of $R$ and let $\hat{R}$ (respectively$\hat{M}$) denote the $\mathfrak{m}$-adic completion of $R$ (respectively $M$). Recall from Proposition \ref{thm 2.21} that $\hat{R}$ is a faithfully
flat $R$-algebra. Assume that $\Gamma$ is a finite set of finitely generated $R$-modules such that $M$ has FFRT by $\Gamma$. If $\hat{\Gamma}= \{ \hat{N} \, | \, N \in \Gamma \}$, then $\hat{M}$ has FFRT by $\hat{\Gamma}$ (Proposition \ref{P4.2}). Now let $ L_{\hat{R}}^0(\hat{\Gamma})= \hat{\Gamma}$ and let $ L_{\hat{R}}^e (\hat{\Gamma})$ denote the set of all direct summands of $F_*^e(N)$ for all $N \in L_{\hat{R}}^{e-1} (\hat{\Gamma)}$. Set $ L_{\hat{R}} (\hat{\Gamma})= \bigcup_{e \in \mathbb{N}}L_{\hat{R}}^e (\hat{\Gamma})$. Similarly, we define $ L_R (\Gamma)= \bigcup_{e \in \mathbb{N}}L_R^e (\Gamma)$ where $ L_R^e (\Gamma)$ is the set of all direct summands of $F_*^e(N)$ for all $N \in L_R^{e-1} (\Gamma)$ and $ L_R^0(\Gamma)= \Gamma$. Therefore, $L_{\hat{R}} (\hat{\Gamma})$ and $ L_R (\Gamma)$ are reasonable classes of modules. We aim now to show that
$ \hat{R}\otimes_{R}N \in L_{\hat{R}} (\hat{\Gamma})$ for all $N \in L_{\hat{R}} (\hat{\Gamma})$. For this purpose, we show that $ \hat{R}\otimes_{R}N \in L_{\hat{R}}^e(\hat{\Gamma})$ whenever $N \in L_{R}^e(\Gamma)$ for all $e \in \mathbb{Z}_{+}$. The case when $e=0$ is obvious. If $N \in L_{R}^e(\Gamma)$, then $N$ is a direct summand of $F_*^e(H)$ for some $H \in L_{R}^{e-1}(\Gamma)$ and hence $\hat{R}\otimes_{R}N$ is a direct summand of $\hat{R}\otimes_{R}F_*^e(H)$. Notice from Proposition \ref{thm 2.21} and Proposition \ref{L2.4} that $\hat{R}\otimes_{R}F_*^e(H)= \widehat{F_*^e(H)}=F_*^e(\widehat{H})=F_*^e(\hat{R}\otimes_{R}H)$. The induction hypothesis implies that $\hat{R}\otimes_{R}H \in L_{\hat{R}}^{e-1}(\hat{\Gamma})$ and consequently $ \hat{R}\otimes_{R}N \in L_{\hat{R}}^e(\hat{\Gamma})$. Since $\widehat{M}$ has FFRT and $\hat{R}$ satisfies Krull-Schmidt theorem on the class of all finitely generated $R$-modules (Discussion \ref{disc2.33}), it follows that $\widehat{M}$ has FFRT by FFRT system and hence $L_{\hat{R}} (\hat{\Gamma})$ contains only finitely many indecomposable modules up to isomorphism and consequently by Proposition \ref{T1} the same holds for $ L_R (\Gamma)$. Now let $\{ M_1, ... , M_s \}$ be the set of representatives for the isomorphism classes of those indecomposable modules in $ L_R (\Gamma)$. Therefore, every $R$-module $N \in \Gamma$ can be written as a finite direct sum whose direct summands are all taken from $\{ M_1, ... , M_s \}$ and hence $M$ has FFRT by $\{ M_1, ... , M_s \}$. Furthermore, $F_*^1(M_j)$ can be written as a finite direct sum whose direct summands are all taken from $\{M_1, ... , M_s\}$ which makes $\{ M_1, ... , M_s \}$ a FFRT system. This proves that $M$ has FFRT by the FFRT system $\{ M_1, ... , M_s \}$.
\end{proof}
\section{FFRT locus of a module is an open set}
\label{section: FFRT locus of a module is an open set}
Let $R$ be a ring and $M$ be a finitely generated $R$-module.
The FFRT locus of $M$ is the set
$$FFRT(M):=\{Q \in \Spec(R) \;|\; M_Q \text{ has } FFRT \text{ over } R_Q\}.$$ In this section, we will prove that $FFRT(M)$ is an open set in the Zariski topology on $\Spec(R)$.
The following Lemma is essential to prove the main result in this section.
\begin{lemma}\label{L4}
Let $R$ be a ring and $M$ a finitely generated $R$-module. If $Q$ is a prime ideal such that $M_Q$ has Finite F-representation type over $R_Q$, then $M_u$ has Finite F-representation type over $R_u$ for some $u \in R \setminus Q$.
\end{lemma}
\begin{proof}
Let $Q $ be a prime ideal for which $M_Q$ has Finite F-representation type. It follows from Theorem \ref{P1} that $M_Q$ has FFRT by a FFRT $\{{M_1}_Q,...,{M_t}_Q \}$ where $M_1,...,M_t$ are finitely generated $R$-modules. As a result, for every positive integer $e$ and every $j \in \{ 1,...,t \}$ there exist nonnegative integers $\alpha(1,j)$ and $\beta(i,j)$ for all $ 1 \leq i,j \leq t$ such that
\begin{equation*}
F_*^1(M)_Q=F_*^1(M_Q)= \bigoplus_{j=1}^t[(M_j)_Q]^{\oplus\alpha(1,j)} = [\bigoplus_{j=1}^tM_j^{\oplus\alpha(1,j)}]_Q
\end{equation*}
and
\begin{equation*}
F_*^1(M_i)_Q=F_*^1(({M_i})_Q)= \bigoplus_{j=1}^t[(M_j)_Q]^{\oplus\beta(i,j)} = [\bigoplus_{j=1}^tM_j^{\beta(i,j)}]_Q \text{ for all } 1 \leq i \leq t.
\end{equation*}
By lemma \ref{LL2}, there exist $s, s_1,...,s_t \in R \setminus Q$ such that
\begin{equation*}
F_*^1(M)_s=F_*^1(M_s)= \bigoplus_{j=1}^t[(M_j)_s]^{\oplus\alpha(1,j)} = [\bigoplus_{j=1}^tM_j^{\oplus\alpha(1,j)}]_s
\end{equation*}
and
\begin{equation*}
F_*^1(M_i)_{s_i}=F_*^1(({M_i})_{s_i})= \bigoplus_{j=1}^t[(M_j)_{s_i}]^{\oplus\beta(i,j)} = [\bigoplus_{j=1}^tM_j^{\oplus\beta(i,j)}]_{s_i} \text{ for all } 1 \leq i \leq t.
\end{equation*}
Let $u=ss_1...s_t$. We will prove by the induction on $e\geq 1$ that $F_*^e(M_u)$ can be written as a direct sum with direct summand taken from $\{(M_1)_u,\ldots,(M_t)_u \}$.
It follows from Proposition \ref{LLL2} and the above equations that
\begin{equation} \label{Equ1}
F_*^1(M)_u=F_*^1(M_u)= \bigoplus_{j=1}^t[(M_j)_u]^{\oplus\alpha(1,j)} = [\bigoplus_{j=1}^tM_j^{\oplus\alpha(1,j)}]_u
\end{equation}
and
\begin{equation*}
F_*^1(M_i)_u=F_*^1(({M_i})_u)= \bigoplus_{j=1}^t[(M_j)_u]^{\oplus\beta(i,j)} = [\bigoplus_{j=1}^tM_j^{\oplus\beta(i,j)}]_u \text{ for all } 1 \leq i \leq t.
\end{equation*}
Now assume that $ F_*^e(M_u)= \bigoplus_{i=1}^t[(M_i)_u]^{\oplus\alpha(e,i)}$ where $\alpha(e,i)$ is nonnegative for all $ 1 \leq i \leq t$. By Equation \ref{Equ1}, it follows that
\begin{eqnarray*}
F_*^{e+1}(M_u) &=& F_*^1[\bigoplus_{i=1}^t[(M_i)_u]^{\oplus\alpha(e,i)}] \\
&=& \bigoplus_{i=1}^tF_*^1[({M_i})_u]^{\oplus\alpha(e,i)} \\
&=& \bigoplus_{i=1}^t[\bigoplus_{j=1}^t[(M_j)_u]^{\oplus\beta(i,j)}]^{\oplus\alpha(e,i)}
\end{eqnarray*}
This induction on $e$ proves that $M_u$ has Finite F-representation type over $R_u$.
\end{proof}
For every $u \in R$ where $R$ is a ring, recall that $V(u)$ denote the set of all prime ideals $P$ containing $u$ and let $D(u)= \Spec(R)\backslash V(u)$. Recall that the collection $\{D(u) \, | \, u \in R \}$ forms a basis of open
sets for the Zariski topology on $\Spec(R)$ (cf \cite[Section 4]{Mat}).
\begin{theorem}\label{P4.11}
If $R$ is a ring and $M$ is a finitely generated $R$-module, then the FFRT locus of $M$ is an open set in the Zariski topology on $\Spec(R)$.
\end{theorem}
\begin{proof}
If $FFRT(M)$ is empty, $FFRT(M)$ is open. Assume now that $FFRT(M)$ is not empty. If $Q \in FFRT(M) $, there exists by lemma \ref{L4} an element $u \in R \setminus Q$ such that $M_u$ has FFRT over $R_u$ and hence by Proposition \ref{LLL2}(a) and Proposition \ref{P4.2} $M_P$ has FFRT over $R_P$ for all $P \in \textit{D}(u)$. This proves that $ \textit{D}(u) \subseteq FFRT(M)$ and hence $FFRT(M)$ is open set in the Zariski topology on $\Spec(R)$.
\end{proof}
\chapter{On the FFRT over hypersurfaces}
\label{chapter:The finite F-representation type over hypersurfaces}
Let $S=K[x_1,...,x_n]$ or $S=K[\![x_1,...,x_n]\!]$ where $K$ is a field of prime characteristic $p$ with $[K:K^p]< \infty$ and $R=S/fS$ for some $f\in S$, in this chapter we introduce a presentation of $F_*^e(R)$ as a cokernel of a matrix factorization of $f$ that is denoted $M_S(f,e)$. The properties of this presentation and its applications to the concept of finite F-representation type will be considered in this chapter.
\section{The presentation of $F_*^e(S/fS)$ as a cokernel of a Matrix Factorization of $f$}
\label{section:The presentation of Fe as a cokernel of a Matrix Factorization of f}
Throughout the rest of this thesis, unless otherwise mentioned, we will adopt the following notation:
\begin{notation}\label{N4.1}
\emph{$K$ will denote a field of prime characteristic $p$ with $[K:K^p] < \infty $, and we set $q=p^e$ for some $e \geq 1$. $S$ will denote the ring $K[x_1,\dots,x_n]$ or $K[\![x_1,\dots,x_n]\!]$. Let $\Lambda_e$ be a basis of $K$ as $K^{p^e}$-vector space. We set $$ \Delta_e := \{\lambda x_1^{a_1}\dots x_n^{a_n}\,|\, 0 \leq a_i \leq p^e-1 \text{ for all } 1\leq i \leq n \text{ and } \lambda \in \Lambda_e \} $$
and set $r_e := | \Delta_e| = [K:K^p]^eq^n .$}
\end{notation}
\begin{discussion}\label{D4.2}
Recall from Corollary \ref{L2.5} that $ \{ F_*^e(j)\, |\, j \in \Delta_e \} $ is a basis of $F_*^e(S)$ as free $S$-module. Let $f \in S$. If $S \xrightarrow{f} S $ is the $S$-linear map given by $s \longmapsto fs$, let $F_*^e(S) \xrightarrow{F_*^e(f)} F_*^e(S) $ be the $S$-linear map that is given by $F_*^e(s) \longmapsto F_*^e(fs)$ for all $s\in S$. We write $M_S(f,e)$ (or $M(f,e)$ if $S$ is known) to denote the matrix $\Mat(F_*^e(f))$ which is the $r_e \times r_e$ matrix representing the $S$-linear map $F_*^e(S) \xrightarrow{F_*^e(f)} F_*^e(S) $ with respect to the basis $ \{ F_*^e(j)\, |\, j \in \Delta_e \} $ (see \ref{Rem 2.9}). Indeed, if $j \in \Delta_e $, there exists a unique set $\{f_{(i,j)} \in S \,|\, i \in \Delta_e \} $ such that $F_*^e(jf)= \bigoplus_{i\in\Delta_e}f_{(i,j)}F^e(i)$ and consequently $M_S(f,e)=[f_{(i,j)}]_{(i,j) \in \Delta^2_e}$. The matrix $M_S(f,e)$ is called the matrix of relations of $f$ over $S$ with respect to $e$.
\end{discussion}
\begin{example}
Let $K$ be a perfect field of prime characteristic 3 , $S=K[x,y]$ or $S=k[\![x,y]\!]$ and let $ f =x^2 + xy$. We aim to construct $M_S(f,1)$. Since the set $\{ F_*^1(1) , F_*^1(x) , F_*^1(x^2) , F_*^1(y) , F_*^1(yx) , F_*^1(yx^2),F_*^1(y^2) , F_*^1(y^2x) , F_*^1(y^2x^2) \}$ is the basis of $F_*^1(S)$ as $S$-module, we get that \\
$ F_*^1(f) = F_*^1(x^2 + xy) = F_*^1(x^2) + F_*^1(yx) $ \\
$ F_*^1(xf)= F_*^1(x^3 + x^2y) = xF_*^1(1) + F_*^1(x^2y)$ \\
$ F_*^1(x^2f)= F_*^1(x^4 + x^3y) = xF_*^1(x) + xF_*^1(y)$ \\
$ F_*^1(yf) = F_*^1(yx^2 + xy^2) = F_*^1(yx^2) + F_*^1(y^2x) $ \\
$ F_*^1(yxf) = F_*^1(yx^3 + x^2y^2) = xF_*^1(y) + F_*^1(y^2x^2) $ \\
$ F_*^1(yx^2f) = F_*^1(yx^4 + x^3y^2) = xF_*^1(yx) + xF_*^1(y^2) $ \\
$ F_*^1(y^2f) = F_*^1(y^2x^2 + xy^3) = F_*^1(y^2x^2) + yF_*^1(x) $ \\
$ F_*^1(y^2xf) = F_*^1(y^2x^3 + x^2y^3) = xF_*^1(y^2) + yF_*^1(x^2) $ \\
$ F_*^1(y^2x^2f) = F_*^1(y^2x^4 + x^3y^3) = xF_*^1(y^2x) + xyF_*^1(1) $ \\
Therefore,
$$ M_S(f,1)= \begin{bmatrix} 0 & x & 0 & 0 & 0 & 0 & 0 & 0 & yx \\ 0 & 0 & x & 0 & 0 & 0 & y & 0 & 0 \\ 1 & 0 & 0 & 0 & 0 & 0 & 0& y & 0 \\ 0 & 0 & x & 0 & x & 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 & 0 & x & 0 & 0 & 0 \\ 0 & 1 & 0 & 1 & 0 & 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 & x & 0 & x & 0 \\ 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & x \\ 0 & 0 & 0 & 0 & 1 & 0 & 1 & 0 & 0 \end{bmatrix} $$
\end{example}
\begin{remark}\label{Ex4.3}
If $m \in \mathbb{N}$, then $F_*^e(f^{mq}j)=f^mF_*^e(j)$ for all $j \in \Delta_e$. This makes $M_S(f^{mq},e)=f^mI$ where $I$ is the identity matrix of size $r_e \times r_e$.
\end{remark}
\begin{proposition}\label{P4.3}
If $f,g \in S$, then
\begin{enumerate}
\item[(a)] $ M_S(f+g,e)=M_S(f,e)+M_S(g,e)$,
\item[(b)] $ M_S(fg,e)=M_S(f,e)M_S(g,e)$ and consequently $ M_S(f,e)M_S(g,e)=M_S(g,e)M_S(f,e)$, and
\item[(c)] $ M_S(f^m,e)= [M_S(f,e)]^m$ for all $m \geq 1$
\end{enumerate}
\end{proposition}
\begin{proof}
The proof follows immediately from Discussion \ref{D4.2} and Remark \ref{Rem 2.9}.
\end{proof}
According to Remark \ref{Ex4.3} and Proposition \ref{P4.3}, we get that $$M_S(f^k,e)M_S(f^{q-k},e)=M_S(f^{q-k},e)M_S(f^k,e)= M_S(f^q,e)= fI_{r_e}$$ for all $0 \leq k \leq q-1$. This shows the following result.
\begin{proposition}\label{P5.5}
For every $f \in S$ and $0 \leq k \leq q-1$, the pair $ ( M_S(f^k,e),M_S(f^{q-k},e))$ is a matrix factorization of $f$.
\end{proposition}
\begin{discussion}
Let $x_{n+1}$ be a new variable and let $ L=S[x_{n+1}]$ if $S=K[x_1,\ldots,x_n]$ or ($ L=S[\![x_{n+1}]\!]$ if $S=K[\![x_1,\ldots,x_n]\!]$). We aim to describe $M_L(g,e)$ for some $g \in L$ by describing the columns of $M_L(g,e)$. First we will construct a basis of the free $L$-module $F_*^e(L)$ using the basis $ \{ F_*^e(j) | j \in\Delta_e \} $ of the free $S$-module $F_*^e(S)$. For each $0\leq v \leq q-1 $, let $ \mathfrak{B}_v= \{F_*^{e}(j x_{n+1}^{v})\, |\, j \in\Delta_e \}$ and set $ \mathfrak{B} = \mathfrak{B}_0 \cup \mathfrak{B}_1 \cup \mathfrak{B}_2\cup \dots \cup \mathfrak{B}_{q-1} $.
Therefore $\mathfrak{B}$ is a basis for $F_*^e(L)$ as free $L$-module and if $g \in L$, we write
\begin{eqnarray*}F_*^e(g)& = & \bigoplus\limits_{i\in\Delta_e}g_{i}^{(0)}F^{e}(i) \oplus\bigoplus\limits_{i\in\Delta_e}g_{i}^{(1)}F^{e}(ix_{n+1}^1)\oplus\dots \oplus \bigoplus\limits_{i\in\Delta_e}g_{i}^{(q-1)}F^{e}(ix_{n+1}^{q-1})
\end{eqnarray*}
where $ g_{i}^{(s)} \in L $ for all $0 \leq s \leq q-1$ and $i\in\Delta_e $.
For each $ 0 \leq s \leq q-1 $ let $ [F_*^e(g)]_{\mathfrak{B}_s} $ denote the column whose entries are the coordinates $ \{g_{i}^{(s)} \in L \,|\, i\in\Delta_e\} $ of $F_*^e(g)$ with respect to $\mathfrak{B}_s$. Let $[F_*^e(g)]_\mathfrak{B}$ be the $r_eq \times 1$ column that is composed of the columns $[F_*^e(g)]_{\mathfrak{B}_0},\dots,[F_*^e(g)]_{\mathfrak{B}_{q-1}}$ respectively.
Therefore $M_L(g,e)$ is the $r_eq \times r_eq$ matrix over $L$ whose columns are all the columns $ [F_*^e(jx_{n+1}^s g)]_\mathfrak{B}$ where $ 0 \leq s \leq q-1$ and $ j \in \Delta_e$. This means that
$ M_L(g,e) = \left[\begin{array}{ccc}
C_{0} & \ldots & C_{q-1} \\
\end{array}
\right]$ where $C_{m}$ is the $r_eq \times r_e$ matrix over $L$ whose columns are the columns $ [F_*^e(jx_{n+1}^m g)]_\mathfrak{B}$ for all $j \in \Delta_e$. If we define $C_{(k,m)}$ to be the $r_e \times r_e$ matrix over $L$ whose columns are $ [F_*^e(jx_{n+1}^m g)]_{\mathfrak{B}_k}$ for all $j\in\Delta_e$, then $C_{m}$ consists of $ C_{(0,m)},\dots,C_{(q-1,m)} $ respectively
and hence the matrix
$M_L(g,e)$ is given by :
\begin{equation}
M_L(g,e) = \left[
\begin{array}{ccc}
C_{0} & \ldots & C_{q-1} \\
\end{array}
\right]=\left[
\begin{array}{ccc}
C_{(0,0)} & \ldots & C_{(0,q-1)} \\
\vdots& & \vdots \\
C_{(q-1,0)} & \ldots & C_{(q-1,q-1)} \\
\end{array}
\right]
\end{equation}
\end{discussion}
Using the above discussion we can prove the following lemma
\begin{lemma}\label{L4.7}
Let $f \in S$ with $A = M_S(f,e)$ and let $ L=S[x_{n+1}]$ if $S=K[x_1,\ldots,x_n]$ or ($ L=S[\![x_{n+1}]\!]$ if $S=K[\![x_1,\ldots x_n]\!]$). If $0\leq d \leq q-1 $, then
\begin{equation}
M_L(fx_{n+1}^d,e) = \left[
\begin{array}{ccc}
C_{(0,0)} & \ldots & C_{(0,q-1)} \\
\vdots& & \vdots \\
C_{(q-1,0)} & \ldots & C_{(q-1,q-1)} \\
\end{array}
\right]
\end{equation}
where
\begin{equation*}
C_{(k,m)} = \begin{cases}
A &\text{if } (m,k) \in \{ (d,0),(d+1,1),\ldots ,(q-1,q-1-d)\} \\
x_{n+1}A &\text{if } (m,k) \in \{ (0,q-d),(1,q-1-d),\ldots ,(d,q-1)\}\\
0 & \text{otherwise}
\end{cases}
\end{equation*}
\end{lemma}
\begin{proof}
If $A = M_S(f,e)=[f_{(i,j)}]$, for each $j \in \Delta_e$ we can write $F_*^e(jf)= \bigoplus_{i\in \Delta_e}f_{(i,j)}F^e(i)$. If $g=fx_{n+1}^d$, for every $1 \leq m \leq q-1 $ and $j \in \Delta_e$, it follows that $F_*^e(jx_{n+1}^mg) = \bigoplus_{i\in \Delta_e}f_{(i,j)}F^e(ix_{n+1}^{d+m})$. Therefore,
\begin{equation*}
F_*^e(jx_{n+1}^mg) = \begin{cases}
\bigoplus_{i\in \Delta_e}f_{(i,j)}F^e(ix_{n+1}^{d+m}) &\text{if } d+m \leq q-1 \\
\bigoplus_{i\in \Delta_e}x_{n+1}f_{(i,j)}F^e(ix_{n+1}^{d+m -q }) &\text{if } d+m > q-1
\end{cases}
\end{equation*}
Accordingly, if $m \leq q-1-d $, then
\begin{equation*}
C_{(k,m)} = \begin{cases}
A &\text{ if } k=d+m \\
0 &\text{ if } k \neq d+m
\end{cases}
\end{equation*}
However, if $m > q-1-d $, it follows that
\begin{equation*}
C_{(k,m)} = \begin{cases}
x_{n+1}A &\text{if } k=d+m -q \\
0 &\text{if } k \neq d+m -q
\end{cases}
\end{equation*}
This shows the required result.
\end{proof}
\begin{proposition}\label{EP}
Let $ L=S[x_{n+1}]$ (if $S=K[x_1,\ldots,x_n]$)or $ L=S[\![x_{n+1}]\!]$ (if $S=K[\![x_1,\ldots,x_n]\!]$). Suppose that $g \in L$ is given by $$ g = g_0 + g_1x_{n+1} + g_2x_{n+1}^2 + \dots +g_dx_{n+1}^d$$ where $d < q$ and $g_k \in S$ for all $ 0 \leq k \leq d $ . If $A_k=M_S(g_k,e)$ for each $0 \leq k \leq d $ then \[ M_L(g,e) = \begin{bmatrix}
A_0& & & & x_{n+1}A_d&x_{n+1}A_{d-1} &\ldots &x_{n+1}A_1 \\
A_1&A_0& & & &x_{n+1}A_d & &\vdots \\
& & & & & &\ddots & \\
& & & & & & &x_{n+1}A_d \\
\vdots & \vdots & \ddots & \ddots& & & & \\
A_d&\vdots& & & & & & \\
& A_d & & & & & &\\
& & & & & & & \\
& & &A_d &A_{d-1} &A_{d-2}& \ldots &A_0
\end{bmatrix}.\]
\end{proposition}
\begin{proof}
Recall from Lemma \ref{L4.7} that
$$ M_A(g_0,e)= \left[
\begin{array}{cccc}
A_0 & & & \\
& A_0 & & \\
& & \ddots & \\
& & & A_0 \\
\end{array}
\right]$$
$$ M_A(g_1x_{n+1},e)=\left[
\begin{array}{cccccc}
& & & & & x_{n+1}A_1 \\
A_1 & & & & & \\
& A_1 & & & & \\
& & \ddots & & & \\
& & & & A_1 & \\
& & & & & \\
\end{array}
\right]$$
$$M_A(g_2x_{n+1}^2,e)=\left[
\begin{array}{ccccc}
& & & x_{n+1}A_2 & \\
& & & & x_{n+1}A_2 \\
A_2 & & & & \\
&\ddots & & & \\
& &A_2 & & \\
\end{array}
\right]
$$
and finally we get
$$ M_A(g_dx_{n+1}^d,e)=\begin{bmatrix}
& & & & x_{n+1}A_d& & & \\
& & & & &x_{n+1}A_d & & \\
& & & & & &\ddots & \\
& & & & & & &x_{n+1}A_d \\
& & & & & & & \\
A_d& & & & & & & \\
& A_d & & & & & &\\
& & \ddots & & & & & \\
& & &A_d & & & & \\
\end{bmatrix}. $$
Proposition \ref{P4.3}(a) implies that $$ M_A(g,e) = M_A(g_0,e) + M_A(g_1x_{n+1},e) + M_A(g_{2}x_{n+1}^2,e) + ... +M_A(g_dx_{n+1}^d,e).$$
This proves the result.
\end{proof}
\begin{example}
Let $K$ be a perfect field of prime characteristic 3 and let $S=K[x]$ or $S=K[\![x]\!]$ . Assume $ L=S[y]$ (if $S=K[x]$) or $L=S[\![y]\!]$ (if $S=K[\![x]\!]$ ). Let $ f =x^2 + xy$, $f_0=x^2$, and $f_1=x$.
By Proposition \ref{EP}, it follows that
$$ M_L(f,1)=\begin{bmatrix} M_S(f_0,1) & & yM_S(f_1,1) \\ M_S(f_1,1) & M_S(f_0,1) & \\ & M_S(f_1,1) & M_S(f_0,1) \end{bmatrix} = \left[\begin{array}{ccc|ccc|ccc} 0 & x & 0 & 0 & 0 & 0 & 0 & 0 & yx \\ 0 & 0 & x & 0 & 0 & 0 & y & 0 & 0 \\ 1 & 0 & 0 & 0 & 0 & 0 & 0& y & 0 \\ \hline 0 & 0 & x & 0 & x & 0 & 0 & 0 & 0 \\ 1 & 0 & 0 & 0 & 0 & x & 0 & 0 & 0 \\ 0 & 1 & 0 & 1 & 0 & 0 & 0 & 0 & 0 \\ \hline 0 & 0 & 0 & 0 & 0 & x & 0 & x & 0 \\ 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & x \\ 0 & 0 & 0 & 0 & 1 & 0 & 1 & 0 & 0 \end{array} \right]. $$
\end{example}
\begin{theorem}\label{P5.11}
Let $f \in S$ be a non-zero non-unit element. If $R=S/fS$, then
\begin{enumerate}
\item[(a)]$F_*^e(R)$ is a MCM $R$-module.
\item[(b)] $F_*^e(R)$ is isomorphic to $\Cok_S(M_S(f,e))$ as $S$-modules (and as $R$-modules).
\end{enumerate}
\end{theorem}
\begin{proof}
(a) First, if $S=K[\![x_1,\dots,x_n]\!]$, then Proposition \ref{P2.43} implies that $R$ is Cohen Maculay and consequently $F_*^e(R)$ is a MCM $R$-module (see Remark \ref{R 2.40}).
Now, if $S=K[x_1,\dots,x_n]$, then Proposition \ref{P2.43} implies that $R$ is Cohen Maculay,i.e. $R_{\mathfrak{n}}$ is Cohen Maculay for every maximal ideal $\mathfrak{n}$ of $R$. It follows from Remark \ref{R 2.40} that $F_*^e(R_{\mathfrak{n}})$ is a MCM $R_{\mathfrak{n}}$-module and hence by Proposition \ref{L2.3} $F_*^e(R)_{\mathfrak{n}}$ is a MCM $R_{\mathfrak{n}}$-module. This shows that $F_*^e(R)$ is a MCM $R$-module.
(b) Write $I=f S$. Since $ \{ F_*^e(j) \, | \, j \in \Delta_e \} $ is a basis of $F_*^e(S)$ as free $S$-module,
the module $F_*^e(R)$ is generated as $S$-module by the set $ \{ F_*^e(j + I) \,| \,j \in \Delta_e \} $. For every $g \in S$, define
$ \phi( F_*^e(g)) = F_*^e(g + I)$. It is clear that $\phi: F_*^e(S) \longrightarrow F_*^e(R)$ is a surjective homomorphism of $S$-modules whose kernel is the $S$-module
$F_*^e(I)$ that is generated by the set $ \{ F_*^e(jf) \, | \, j \in \Delta_e \} $.
Now, define the $S$-linear map $ \psi : F_*^e(S) \rightarrow F_*^e(S)$ by $ \psi( F_*^e(h)) = F_*^e(hf)$ for all $h\in S$.
We have an exact sequence $F_*^e(S)\xrightarrow{\psi}F_*^e(S)\xrightarrow{\phi}F_*^e(R) \xrightarrow{} 0$. Notice for each $j \in \Delta_e$ that $\psi(F_*^e(j))= F_*^e(jf)= \bigoplus_{i\in\Delta_e}f_{(i,j)}F^e(i)$
and hence $M_S(f,e)$
represents the map $\psi$ on the given free-bases (Remark \ref{Rem 2.9}).
By Proposition \ref{P5.5} and Remark \ref{C3.2}(a), it follows that $F_*^e(R)$ is isomorphic to $\Cok_S(M_S(f,e))$ as $R$ -modules.
\end{proof}
\begin{corollary}\label{C4.12}
Let $f\in S$ be a non-zero non-unit element. If $1 \leq k \leq q-1$ and $R=S/fS$, then
\begin{enumerate}
\item[(a)] $F_*^e(S/f^kS)$ is a MCM $S/f^kS$ -modules isomorphic to $\Cok_S(M_S(f^k,e))$ as $S$-modules (and as $S/f^kS$-modules), and
\item[(b)] $F_*^e(S/f^kS)$ is a MCM $R$-module isomorphic to $\Cok_S(M_S(f^k,e))$ as $S$-modules (and as $R$-modules).
\end{enumerate}
\end{corollary}
\begin{proof}
(a) can be proved by applying Proposition \ref{P5.11} to $f^k$ instead of $f$.
(b) Since the pair $(M_S(f^k,e),M_S(f^{q-k},e))$ is a matrix factorization of $f$, it follows that $f \Cok_S(M_S(f^k,e))=0$. Notice that $fF_*^e(S/f^kS)=0$
This makes $F_*^e(S/f^kS)$ and $\Cok_S(M_S(f^k,e))$ $R$-modules and consequently $F_*^e(S/f^kS)$ is isomorphic to $\Cok_S(M_S(f^k,e))$ as $R$ -modules. Since $F_*^e(S/f^kS)$ is a MCM $S/f^kS$ -modules and $(f+f^kS)F_*^e(S/f^kS)=0$, it follows by Proposition \ref{P2.56} that $F_*^e(S/f^kS)$ is MCM module over the ring $\frac{S/f^kS}{(f+f^kS)(S/f^kS)}= S/fS$.
\end{proof}
\begin{proposition}\label{L4.13}
Let $K$ be a field of prime characteristic $p>2 $ with $[K:K^p] <\infty$ and let $T=S[z]$ if $S=K[x_1,\dots,x_n]$ (or $T=S[\![z]\!]$ if $S=K[\![x_1,\dots,x_n]\!]$). If $A=M_S(f,e)$ for some $f \in S$, then
\begin{equation*}
F_*^e(T/(f+z^2))= \Cok_{T}\left[
\begin{array}{cc}
A^{\frac{q-1}{2}} & -zI \\
zI & A^{\frac{q+1}{2}} \\
\end{array}
\right].
\end{equation*}
\end{proposition}
\begin{proof}
Let $I$ be the identity matrix in $M_{r_e}(S)$. It follows from Proposition \ref{EP} that $ M_{T}(f+z^2, e)$ is a $q \times q$ matrix over the ring $M_{r_e}(S)$ that is given by
\begin{equation*}
M_{T}(f+z^2, e) = \left[
\begin{array}{ccccccc}
A & & & & & zI & 0 \\
0 & A & & & & & zI \\
I & 0 & A & & & & \\
& I & 0 & A & & & \\
& & \ddots & \ddots & \ddots & & \\
& & & I & 0 & A & \\
& & & & I & 0 & A \\
\end{array}
\right].
\end{equation*}
It follows from Corollary \ref{C2.14} and Theorem \ref{P5.11} that
\begin{equation*}
F_*^e(T/(f+z^2)) = \Cok_{T}\left[
\begin{array}{cc}
zI & (-1)^mA^{\frac{q+1}{2}} \\
(-1)^{m-1} A^{\frac{q-1}{2}} & zI \\
\end{array}
\right] \text{ where } m=\frac{q-1}{2}.
\end{equation*}
If $m$ is odd integer, we get
\begin{eqnarray*}
F_*^e(T/(f+z^2)) &=&\Cok_{T}\left[
\begin{array}{cc}
A^{\frac{q-1}{2}} & zI \\
zI & -A^{\frac{q+1}{2}} \\
\end{array}
\right]\\
&=& \Cok_{T}(
\left[
\begin{array}{cc}
A^{\frac{q-1}{2}} & zI \\
zI & -A^{\frac{q+1}{2}} \\
\end{array}
\right]\left[
\begin{array}{cc}
I & 0 \\
0 & -I \\
\end{array}
\right]) \\
&=& \Cok_{T}\left[
\begin{array}{cc}
A^{\frac{q-1}{2}} & -zI \\
zI & A^{\frac{q+1}{2}} \\
\end{array}
\right].
\end{eqnarray*}
However, if $m$ is even, it follows that
\begin{eqnarray*}
F_*^e(T/(f+z^2)) &=&\Cok_{T}\left[
\begin{array}{cc}
-A^{\frac{q-1}{2}} & zI \\
zI & A^{\frac{q+1}{2}} \\
\end{array}
\right]\\
&=& \Cok_{T}(\left[
\begin{array}{cc}
-I & 0 \\
0 & I \\
\end{array}
\right]\left[
\begin{array}{cc}
-A^{\frac{q-1}{2}} & zI \\
zI & A^{\frac{q+1}{2}} \\
\end{array}
\right])
\\
&=& \Cok_{T}\left[
\begin{array}{cc}
A^{\frac{q-1}{2}} & -zI \\
zI & A^{\frac{q+1}{2}} \\
\end{array}
\right].
\end{eqnarray*}
\end{proof}
\begin{proposition}\label{P21}
Let $u$ and $v$ be new variables on $S$ and let $L=S[u,v]$ if $S=K[x_1,\dots,x_n]$ (or $L=S[\![u,v]\!]$ if $S=K[\![x_1,\dots,x_n]\!]$). Let $R^{\bigstar}=L/(f+uv)$ where $f \in S$. If $A$ is the matrix $M_S(f,e)$ and $I$ is the identity matrix in the ring $M_{r_e}(S)$, then $$ F_*^e(L/(f+uv))= (R^{\bigstar})^{r_e} \bigoplus \bigoplus_{k=1}^{q-1}\Cok_LB_k $$ where $ B_k = \begin{bmatrix} A^k & -vI \\ uI & A^{q-k} \end{bmatrix}$ for all $ 1\leq k \leq q-1$.
\end{proposition}
\begin{proof}
Recall that $\mathfrak{D}=\{ F_*^e(ju^sv^t) \,|\, j\in \Delta_e , 0 \leq s,t \leq q-1 \}$ is a free basis of $F_*^e(L)$ as $L$-module.
We introduce a $\mathbb{Z}/q\mathbb{Z}$-grading on both $L$ and $F_*^e(L)$ as follows:
$L$ is concentrated in degree 0, while $\deg(F_*^e(x_i))=0$ for each $1 \leq i \leq n$, $\deg(F_*^e(u))=1$ and $\deg(F_*^e(v))=-1$.
We can now write
$F_*^e(L)= \bigoplus_{k=0}^{q-1}M_k$ where $M_k$ is the free $L$-submodule of $F_*^e(L)$
of homogeneous elements of degree $k$, i.e. $M_k$ is the $L$-submodule of $F_*^e(L)$ that is generated by the subset
$$\mathfrak{D}_k = \{F_*^e(ju^sv^t)\,|\, \deg(F_*^e(ju^sv^t))=k \}$$ of the basis $\mathfrak{D}$.
Note that
$\mathfrak{D}_0= \{F_*^e(ju^sv^s)\,|\, j\in\Delta_e , 0 \leq s \leq q-1 \} $, and that for all $1 \leq k \leq q-1$
\begin{eqnarray*}
\mathfrak{D}_k&=& \{F_*^e(ju^{k+r}v^{r})\,|\, j\in\Delta_e , 0 \leq r \leq q-k-1 \} \cup\\
& & \{F_*^e(ju^{r}v^{q-k+r})\,|\, j\in\Delta_e , 0 \leq r \leq k-1 \}.
\end{eqnarray*}
Let $J$ be the ideal $(f+uv)L$. Since $\deg(F_*^e(f+uv))=0$, it follows that $F_*^e(J)= \bigoplus_{k=0}^{q-1}M_kF_*^e(f+uv)$ and consequently
\begin{equation}\label{EEEE1}
F_*^e(L/(f+uv))= F_*^e(L)/F_*^e(J)=\bigoplus_{k=0}^{q-1}M_k/M_kF_*^e(f+uv).
\end{equation}
We now show that
$M_k/M_kF_*^e(f+uv) \cong \Cok_LC_k $ where $C_0= \begin{bmatrix} (-1)^{q}A^{q-1} & uvI \\ I & A \end{bmatrix} $ and $ C_k = \begin{bmatrix} (-1)^{q-k+1}A^{q-k} & vI \\ uI & (-1)^{k+1}A^{k} \end{bmatrix}$ for all $1\leq k \leq q-1$.
Recall that if $M_s(f,e)=[f_{(i,j)}]$, then $F_*^e(jf)=\bigoplus_{i\in\Delta_e}f_{(i,j)}F_*^e(i)$ for all $j \in \Delta_e$.
Therefore,
\begin{eqnarray}\label{Eq18}
\nonumber F_*^e(ju^sv^t(f+uv)) &=& F_*^e(jf)F_*^e(u^sv^t) + F_*^e(ju^{s+1}v^{t+1}) \\
&=& \bigoplus_{i\in\Delta_e}f_{(i,j)}F_*^e(iu^sv^t) \oplus F_*^e(ju^{s+1}v^{t+1}).
\end{eqnarray}
Since $\deg(F_*^e(iu^sv^t))=\deg(F_*^e(ju^{s+1}v^{t+1}))$ for all $i,j \in \Delta_e$ and all $0 \leq s,t \leq q-1$, it follows that $ F_*^e(ju^sv^t(f+uv)) \in M_k$ for all $F_*^e(ju^sv^t) \in \mathfrak{D}_k$. This enables us to define the homomorphism $\psi_k :M_k\rightarrow M_k$ that is given by $\psi_k(F_*^e(ju^sv^t))=F_*^e(ju^sv^t(f+uv))$ for all $F_*^e(ju^sv^t) \in \mathfrak{D}_k$ and consequently we have the following short exact sequence $$M_k\xrightarrow{\psi_k}M_k\xrightarrow{\phi_k}M_k/M_kF_*^e(f+uv) \xrightarrow{} 0$$
where $\phi_k :M_k\rightarrow M_k/M_kF_*^e(f+uv)$ is the canonical surjection.
Notice that if $0 \leq s < q-1$, equation (\ref{Eq18}) implies that
\begin{equation}\label{Eq19}
F_*^e(ju^sv^s(f+uv))= \bigoplus_{i\in\Delta_e}f_{(i,j)}F_*^e(iu^sv^s) \oplus F_*^e(ju^{s+1}v^{s+1})
\end{equation} and
\begin{equation}\label{Eq20}
F_*^e( ju^{q-1}v^{q-1}(f+uv))= \bigoplus_{i\in\Delta_e}f_{(i,j)}F_*^e(iu^{q-1}v^{q-1}) \oplus uv F_*^e(j) ,
\end{equation}
therefore $\psi_0$ is represented by the matrix $\begin{bmatrix} A & & &uvI \\ I & A & & \\ & \ddots & \ddots & \\ & & I & A\end{bmatrix}$ which is a $q\times q$ matrix over the ring $M_{r_e}(L).$
Now Corollary \ref{L.20} implies that
\begin{equation}
M_0/M_0F_*^e(f+uv)\cong \Cok_L \begin{bmatrix} (-1)^{q}A^{q-1} & uvI \\ I & A \end{bmatrix}.
\end{equation}
However,
\begin{equation*}
\begin{bmatrix} (-1)^{q}A^{q-1} & uvI \\ I & A \end{bmatrix}\sim \begin{bmatrix} 0 &A(A^{q-1})+ uvI \\ I & A \end{bmatrix}\sim \begin{bmatrix} A(A^{q-1})+ uvI & 0\\ 0 & I \end{bmatrix}.
\end{equation*}
Since $AA^{q-1}=fI$ as $(A,A^{q-1})$ is a matrix factorization of $f$ (Proposition \ref{P5.5}), it follows that
\begin{equation}\label{EEEE2}
M_0/M_0F_*^e(f+uv)\cong \Cok_L \begin{bmatrix} (f+uv)I & 0 \\ 0 & I \end{bmatrix}=(R^{\bigstar})^{r_e}.
\end{equation}
Now let $1 \leq k \leq q-1$. If $0 \leq r < q-k-1$, then it follows from equation (\ref{Eq18}) that
\begin{equation}\label{Eq21}
F_*^e( ju^{k+r}v^r(f+uv))= \bigoplus_{i\in\Delta_e}f_{(i,j)}F_*^e(iu^{k+r}v^{r}) \oplus F_*^e(ju^{k+r+1}v^{r+1})
\end{equation}
and
\begin{equation}\label{Eq22}
F_*^e( ju^{q-1}v^{q-k-1}(f+uv))= \bigoplus_{i\in\Delta_e}f_{(i,j)}F_*^e(iu^{q-1}v^{q-k-1}) \oplus u F_*^e(jv^{q-k}).
\end{equation}
However, if $0 \leq r < k-1$, it follows from (\ref{Eq18}) that
\begin{equation}\label{Eq222}
F_*^e( ju^{r}v^{q-k-r}(f+uv))= \bigoplus_{i\in\Delta_e}f_{(i,j)}F_*^e(iu^{r}v^{q-k-r}) \oplus F_*^e(ju^{r+1}v^{q-k-r+1})
\end{equation}
and
\begin{equation}\label{Eq23}
F_*^e( ju^{k-1}v^{q-1}(f+uv))=\bigoplus_{i\in\Delta_e}f_{(i,j)}F_*^e(iu^{k-1}v^{q-1}) \oplus vF_*^e(ju^{k}).
\end{equation}
As a result, $\psi_k$ is represented by the matrix $ \left[ \begin{array}{cccc|cccc} A & & & & & & & vI \\
I & A & & & & & & \\
& \ddots & \ddots & & & & & \\
& & I& A & & & & \\ \hline
& & &uI & A & & & \\
& & & & I & A & & \\
& & & & & \ddots & \ddots & \\
& & & & & &I & A
\end{array}\right] $ which is a $q\times q$ matrix over the ring $M_{r_e}(L)$ where $uI$ is in the
$(q-k+1,q-k)$ spot of this matrix.
Therefore, Corollary \ref{L.20} implies that
\begin{eqnarray}
\nonumber M_k/M_kF_*^e(f+uv) &\cong & \Cok_L\begin{bmatrix} (-1)^{q-k+1}A^{q-k} & vI \\ uI & (-1)^{k+1} A^{k} \end{bmatrix}.
\end{eqnarray}
If $k$ is odd integer, we notice that
\begin{eqnarray}\label{EEq}
\nonumber M_k/M_kF_*^e(f+uv) &\cong & \Cok_L( \left[
\begin{array}{cc}
-I & 0 \\
0 & I \\
\end{array}
\right] \begin{bmatrix} (-1)^{q-k+1}A^{q-k} & vI \\ uI & (-1)^{k+1} A^{k} \end{bmatrix}) \\
&\cong & \Cok_L \begin{bmatrix} A^{q-k} & -vI \\ uI & A^{k} \end{bmatrix}.
\end{eqnarray}
Similar argument when $k$ is even shows that
\begin{equation}\label{Eq24}
M_k/M_kF_*^e(f+uv) \cong \Cok_L \begin{bmatrix} A^{q-k} & -vI \\ uI & A^{k} \end{bmatrix}.
\end{equation}
Now the proposition follows from (\ref{EEEE1}), (\ref{EEEE2}), (\ref{EEq}) and (\ref{Eq24}).
\end{proof}
If $(\phi,\psi)$ is a matrix factorization of a nonzero nonunit element $f$ in a domain $S$ and $u,v,z$ are variables on $S$, recall from Remark \ref{R3.8} that
\begin{equation*}
(\phi, \psi)^{\maltese}:= ( \begin{bmatrix} \phi & -vI \\ uI & \psi \end{bmatrix} , \begin{bmatrix} \psi & vI \\ -uI & \phi \end{bmatrix}), \text{ and }(\phi, \psi)^{\sharp}:= ( \begin{bmatrix} \phi & -zI \\ zI & \psi \end{bmatrix} ,
\begin{bmatrix} \psi & zI \\ -zI & \phi \end{bmatrix})
\end{equation*}
Furthermore, $(\phi, \psi)^{\maltese}$ is a matrix factorization of $f+uv$ in $S[\![u,v]\!]$ (and in $S[u,v]$)
and $ (\phi, \psi)^{\sharp}$ is a matrix factorization of $f+z^2$ in $S[\![z]\!]$ (and in $S[\![z]\!]$). Using this notation and the notation in Definition \ref{D2.11} we can establish the following proposition.
\begin{proposition}\label{C4.19}
Let $S=K[\![x_1,\dots,x_n]\!]$ and $f \in \mathfrak{m} \setminus \{0\}$ where $\mathfrak{m}$ is the maximal ideal of $S$. Let $R=S/fS$, $R^{\bigstar}= S[\![u,v]\!]/(f+uv)$, and $ R^{\sharp}= S[\![z]\!]/(f+z^2)$. If $A=M_S(f,e)$ and $1\leq k \leq q-1$, then
\begin{enumerate}
\item [(a)] $\sharp(\Cok_{S[\![z]\!]}(A^k, A^{q-k})^{\sharp},R^{\sharp}) = \sharp(\Cok_S(A^k),R)+\sharp(\Cok_S(A^{q-k}),R)$.
\item [(b)] If $K$ be a field of prime characteristic $p > 2$, it follows that\\
$\sharp(F_*^e(R^{\sharp}),R^{\sharp}) = \sharp(\Cok_S(A^{\frac{q-1}{2}}),R)+\sharp(\Cok_S(A^{\frac{q+1}{2}}),R)$ and hence \\
$\sharp(F_*^e(R^{\sharp}),R^{\sharp})=\sharp(F_*^e(S/f^{\frac{q-1}{2}}S),R)+\sharp(F_*^e(S/f^{\frac{q+1}{2}}S),R)$.
\item [(c)] $\sharp(\Cok_{S[\![u,v]\!]}(A^k, A^{q-k})^{\maltese},R^{\maltese})= \sharp(\Cok_S(A^k),R)+\sharp(\Cok_S(A^{q-k}),R)$.
\item [(d)] $\sharp ( F_*^e(R^{\bigstar}),R^{\bigstar})= r_e + 2 \sum_{k=1}^{q-1} \sharp ( \Cok_S(A^k),R)$.
\end{enumerate}
\end{proposition}
\begin{proof}
Notice that $ (A^{k},A^{q-k})$ is a matrix factorization of $f$ (Proposition \ref{P5.5}).
(a) If $ (A^{k},A^{q-k}) \sim (f,1)^{r_e} $ (or $ (A^{k},A^{q-k})\sim (1,f)^{r_e} $) then $(F_*^e( S/f^{q-k}S)=\Cok_S(A^{q-k})= \{0\}$ (or $(F_*^e( S/f^{k}S )=\Cok_S(A^{k})= \{0\}$) which is impossible. As a result, if $ (A^{k},A^{q-k})$ is a trivial matrix factorization of $f$, then the only possible
case is that $ (A^{k},A^{q-k}) \sim (f,1)^u \oplus (1,f)^v$ where $0 < u,v < r_e$ with $u+v = r_e$ , $ \sharp(\Cok_S(A^{k}),R)= u$, and $ \sharp(\Cok_S(A^{q-k}),R)= v$. Remark \ref{R3.8} (j),(k),(e) and (f) implies that
\begin{eqnarray*}
\Cok_{S[\![z]\!]}(A^k, A^{q-k})^{\sharp} &=& [\Cok_{S[\![z]\!]}(f, 1)^{\sharp}]^u\oplus[\Cok_{S[\![z]\!]}(1,f)^{\sharp}]^v \\
&=& [R^{\sharp}]^{u+v}.
\end{eqnarray*}
Therefore, if $ (A^{k},A^{q-k})$ is a trivial matrix factorization of $f$, it follows that
$$\sharp(\Cok_{S[\![z]\!]}(A^k, A^{q-k})^{\sharp},R^{\sharp}) = \sharp(\Cok_S(A^k),R)+\sharp(\Cok_S(A^{q-k}),R).$$
On the other hand, if $ (A^{k},A^{q-k})$ is not trivial matrix factorization of $f$, it follows from Corollary \ref{C3.8} that
\begin{equation*}
\sharp(\Cok_{S[\![z]\!]}(A^k, A^{q-k})^{\sharp},R^{\sharp}) = \sharp(\Cok_S(A^k),R)+\sharp(\Cok_S(A^{q-k}),R).
\end{equation*}
(b) follows from the result (a) above, Proposition \ref{L4.13} and the fact that $\Cok_SA^k=\Cok_SM_S(f^k,e)=F_*^e(S/f^kS)$ for all $1\leq k \leq q-1$ (Corollary\ref{C4.12} and Proposition \ref{P4.3}(c)).
(c) can be proved similarly to (a).
(d) By Proposition \ref{P21}, it follows that
\begin{equation*}
F_*^e(R^{\bigstar})= (R^{\bigstar})^{r_e} \bigoplus \bigoplus_{j=1}^{q-1} \Cok_{S[\![u,v]\!]} \begin{bmatrix} A^k & -vI \\ uI & A^{q-k} \end{bmatrix}.
\end{equation*}
However, For each $1 \leq k \leq q-1 $ , we recall from Remark \ref{R3.8} (e) that
\begin{equation*}
\Cok_{S[\![u,v]\!]}\begin{bmatrix} A^k & -vI \\ uI & A^{q-k} \end{bmatrix}=\Cok_{S[\![u,v]\!]}(A^k,A^{q-k})^{\maltese}.
\end{equation*}
Therefore, by Krull-Remak-Schmidt theorem (Discussion \ref{disc2.33}), the result (c) above and the convention that $\Cok_{S}(A^k,A^{q-k})=\Cok_{S}(A^k)$ it follows that
\begin{eqnarray*}
\sharp ( F_*^e(R^{\bigstar}),R^{\bigstar}) &=& r_e + \sum_{k=1}^{q-1}\sharp (\Cok_{S[[u,v]]}(A^k,A^{q-k})^{\maltese}, R^{\bigstar}) \\
&=& r_e + \sum_{k=1}^{q-1} \left[\sharp(\Cok_{S}(A^k,A^{q-k}),R)+ \sharp(\Cok_{S}(A^{q-k},A^k), R)\right] \\
&=& r_e + 2 \sum_{k=1}^{q-1} \sharp (\Cok_{S}(A^k,A^{q-k}),R)\\
&=& r_e + 2 \sum_{k=1}^{q-1} \sharp (\Cok_{S}(A^k),R).
\end{eqnarray*}
\end{proof}
\section{The ring $S[[y]]/(y^{p^d} +f)$ has finite F-representation type}
\label{section:The ring pd has finite Frepresentation type}
We keep the same notation as in Notation \ref{N4.1}. The following proposition can be obtained as a special case from \cite[Theorem 3.10]{TS}. However, we provide a different proof that is based basically on the Theorem \ref{P5.11}.
\begin{theorem}\label{P5.1}
If $d \in \mathbb{N}$ and $S:= K[\![x_1,...,x_n]\!]$,
then $S[\![y]\!]/(y^{p^d}+f)$ has FFRT for any $f \in S$ and any prime integer $p>0$.
\end{theorem}
\begin{proof}
Let $A$ be the matrix $M_S(f,e)$ in $M_{r_e}(S)$ where $ e > d$ and let $I$ be the identity matrix in $M_{r_e}(S)$.
Let $M$ and $N$ be the diagonal matrices of size $p^d \times p^d$ with entries in $M_{r_e}(S)$ with $A$ and $I$ along the diagonals of $M$ and $N$ as follows: $$M= \begin{bmatrix} A & & & & & \\
& A & & & & \\
& & A & & & \\
& & & \ddots & & \\
& & & & A \end{bmatrix} $$ and $$N= \begin{bmatrix} I & & & & & \\
& I & & & & \\
& & I & & & \\
& & & \ddots & & \\
& & & & I \end{bmatrix}. $$
Using Proposition \ref{EP} we write $M_{S[\![y]\!]}(y^{p^d}+f,e)$ as the following $p^{e-d} \times p^{e-d}$ matrix over the ring $M_{p^d}(M_{r_e}(S[\![y]\!]))$:
$$ M_{S[\![y]\!]}(y^{p^d}+f,e) = \begin{bmatrix} M & & & & yN \\
N & M & & & & \\
& N & M & & & \\
& & \ddots & \ddots & & \\
& & &N& M \end{bmatrix} . $$
Using Lemma \ref{1}, and Theorem \ref{P5.11} we see that
\begin{equation}\label{Eq5.14}
F_*^e(S[\![y]\!]/(y^{p^d}+f)) \cong \Cok_{S[\![y]\!]}( yN+(-1)^{1+p^{e-d}}M^{p^{e-d}})\cong (\Cok_{S[\![y]\!]}(yI+A^{p^{e-d}}))^{\oplus p^d}.
\end{equation}
We aim to prove the existence of finitely generated $S[\![y]\!]/(y^{p^d}+f) $-modules $M_1,...,M_s$ such that $\Cok_{S[\![y]\!]}(yI+A^{p^{e-d}})$ can be written as a direct sum with direct summands taken from $\{M_1,...,M_s\}$ for every $e$. Notice from Remark \ref{R2.29}(a) and Theorem \ref{P5.11} that
\begin{eqnarray*}
F_*^d[ F_*^{e-d}( S[\![y]\!]/(y^{p^d}+f)^{p^{e-d}}))] &\cong & F_*^e( S[\![y]\!]/(y^{p^e}+f^{p^{e-d}})) \\
&\cong &\Cok_{S[\![y]\!]}(M_{S[\![y]\!]}(y^{p^e}+f^{p^{e-d}},e)).
\end{eqnarray*}
However, Proposition \ref{P4.3} and Proposition \ref{EP} imply that
$$ M_{S[\![y]\!]}(y^{p^e}+f^{p^{e-d}},e) = (yI+A^{p^{e-d}})^{\oplus p^e}.$$
As a result, we get
$$ F_*^d[ F_*^{e-d}( S[\![y]\!]/(y^{p^d}+f)^{p^{e-d}}))] \cong [ \Cok_{S[\![y]\!]}(yI+A^{p^{e-d}})]^{\oplus p^e}.$$
If $r_{e-d}=[K:K^p]^{e-d}p^{(e-d)(n+1)}$, let $\tilde{I}$ be the identity matrix in $M_{r_{e-d}}(S[\![y]\!])$.
By Theorem \ref{P5.11} and Remark \ref{Ex4.3}, it follows that
\begin{eqnarray*}
F_*^{e-d}( S[\![y]\!]/(y^{p^d}+f)^{p^{e-d}}) &\cong & \Cok_{S[\![y]\!]}(M_{S[\![y]\!]}((y^{p^d}+f)^{p^{e-d}},e-d) \\
&\cong &\Cok_{S[\![y]\!]}((y^{p^d}+f)\tilde{I}) \\
&\cong & [ S[\![y]\!]/(y^{p^d}+f)]^ {\oplus ^{r_{e-d}}}
\end{eqnarray*}
and hence $$ F_*^e( S[\![y]\!]/(y^{p^e}+f^{p^{e-d}}))= [F_*^d(S[\![y]\!]/(y^{p^d}+f)]^ {\oplus ^{r_{e-d}}}.$$
This makes
\begin{equation}\label{Eq5.15}
[ \Cok_{S[\![y]\!]}(yI+A^{p^{e-d}})]^{\oplus p^e}\cong[F_*^d(S[[y]]/(y^{p^d}+f))]^ {\oplus ^{r_{e-d}}}
\end{equation}
as $S[\![y]\!]/(y^{p^d}+f))$-modules. Since $ F_*^d(S[\![y]\!]/(y^{p^d}+f))$ can be written as a direct sum with direct summands taken from a finite set of indecomposable finitely generated $S[\![y]\!]/(y^{p^d}+f))$-modules, say $M_1 , ... , M_s$, it follows from Krull-Remak-Schmidt theorem (Discussion \ref{disc2.33}) that $\Cok_{S[\![y]\!]}(yI+A^{p^{e-d}})$ is also a direct sum with direct summands taken from $M_1 , ... , M_s$.
\end{proof}
\section{When does the ring $S[[u,v]]/(f+uv)$ have finite F-representation type?}
\label{section:When does the ring fuv have finite F-representation type?}
We keep the same notation as in Notation \ref{N4.1}. The purpose of this section is to provide a characterization of when the ring $S[\![u,v]\!]/(f+uv)$ has finite F-representation type. This characterization enables us to exhibit a class of rings in section \ref{section:Class of rings that have FFRT but not finite CM type} that have FFRT but not finite CM type.
\begin{theorem}\label{P30}
Let $K$ be an algebraically closed field of prime characteristic $p > 2$ and $q=p^e$. Let $S:= K[\![x_1,\dots,x_{n}]\!]$ and let $ \mathfrak{m}$ be the maximal ideal of $S$ and $f \in \mathfrak{m}^2 \setminus \{0\}$. Let $R=S/(f)$ and $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$. Then $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$ has FFRT over $R^{\bigstar}$ if and only if there exist indecomposable $R$-modules $N_1,\dots,N_t$ such that $F_*^e(S/(f^k))$ is a direct sums with direct summands taken from $R, N_1,\dots,N_t $ for every $e\in \mathbb{N}$ and $1 \leq k < p^e $ .
\end{theorem}
\begin{proof}
First, suppose that the ring $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$ has FFRT over $R^{\bigstar}$ by $\{R^{\bigstar} , M_1, \dots , M_t , \}$ where $M_j$ is an indecomposable non-free MCM $R^{\bigstar}$-module for all $j$. Therefore, there exist a nonnegative integers $n_{(e)},n_{(e,1)},\ldots,n_{(e,t)}$ such that
\begin{equation*}
F_*^e(S[\![u,v]\!]/(f+uv))=( R^{\bigstar})^{n_{(e)}}\bigoplus \bigoplus_{j=0}^{t}M_i^{n_{(e,j)}}.
\end{equation*}
By Proposition \ref{P28} and \ref{A}, it follows that $M_j= \Cok_{S[\![u,v]\!]}(\alpha_j,\beta_j)^{\maltese}$ where $(\alpha_j,\beta_j)$ is a reduced matrix factorization of $f$ such that $\Cok_S(\alpha_j,\beta_j)$ is non-free indecomposable MCM $R$-module for all $j$. Let $A=M_S(f,e)$ and consequently by Proposition \ref{P4.3} $A^k= (M_S(f,e))^k=M_S(f^k,e)$. Recall from the Proposition \ref{P21} and the notation in Remark \ref{R3.8} (e) that
\begin{equation}\label{Eq12}
F_*^e(S[\![u,v]\!]/(f+uv))= ( R^{\bigstar})^{r_e} \oplus \bigoplus \limits_{k=1}^{q-1}\Cok_{S[\![u,v]\!]}(A^k,A^{q-k})^{\maltese }.
\end{equation}
This makes
\begin{equation}\label{E13}
( R^{\bigstar})^{n_{(e)}}\bigoplus \bigoplus_{j=0}^{t}M_i^{n_{(e,j)}} \cong( R^{\bigstar})^{r_e} \oplus \bigoplus \limits_{k=1}^{q-1}\Cok_{S[\![u,v]\!]}(A^k,A^{q-k})^{\maltese }.
\end{equation}
If $(A^k,A^{q-k})$ is nontrivial matrix factorization of $f$, by Proposition \ref{P.24}, there exist a reduced matrix factorization $(\phi_k , \psi_k)$ of $f$ and non-negative integers $t_k$ and $r_k$ such that $(A^k,A^{q-k})\sim (\phi_k , \psi_k) \oplus (f,1)^{t_k} \oplus (1,f)^{r_k}$. This gives by Remark \ref{R3.8} (g), (h), and (i) that $\Cok_{S[[u,v]]}(A^k,A^{q-k})^{\maltese }= \Cok_{S[[u,v]]}(\phi_k , \psi_k)^{\maltese} \oplus [R^{\bigstar}]^{t_k+r_k} $.
By the equation \ref{E13} and Krull-Remak-Schmidt theorem (Discussion \ref{disc2.33}), there exist non-negative integers $n_{(e,k,1)},\dots., n_{(e,k,t)}$ such that
\begin{eqnarray*}
\Cok_{S[\![u,v]\!]}(\phi_k , \psi_k)^{\maltese } &\cong& \bigoplus\limits_{j=1}^{t}M_j^{n_{(e,k,j)}} \\
&\cong& \bigoplus\limits_{j=1}^{t}[\Cok_{S[\![u,v]\!]}(\alpha_j,\beta_j)^{\maltese}]^{ \oplus n_{(e,k,j)}}\\
&\cong& \Cok_{S[\![u,v]\!]}[\bigoplus_{j=1}^{t}(\alpha_j,\beta_j)^{ \oplus n_{(e,k,j)}}]^{\maltese}.
\end{eqnarray*}
Now, from Proposition \ref{P3.15} and Remark \ref{R3.8} (d) it follows that $$ \Cok_{S}(\phi_k , \psi_k) \cong \Cok_{S}[\bigoplus_{j=1}^{t}(\alpha_j,\beta_j)^{ \oplus n_{(e,k,j)}}] \cong \bigoplus_{j=1}^{t} N_j^{ \oplus n_{(e,k,j)}}$$ where $N_j$ denotes the non-free indecomposable MCM $R$-module $ \Cok_S(\alpha_j,\beta_j)$ for all $j \in \{1,\dots,t\}$. Therefore,
\begin{eqnarray*}
F_*^e(S/f^kS) &\cong& \Cok_SM_S(f^k,e) \text{ (Theorem \ref{P5.11}) } \\
&=& \Cok_S(A^k) \text{ (Proposition \ref{P4.3} (c)) } \\
&=& \Cok_S(A^k,A^{q-k}) \text{ (Definition \ref{D23})} \\
&=& \Cok_S[(\phi_k , \psi_k) \oplus (f,1)^{r_k} \oplus (1,f)^{t_k}] \\
&=& R^{r_k}\oplus\bigoplus_{j=1}^{t}N_j^{ \oplus n_{(e,k,j)}}.
\end{eqnarray*}
However, if $ (A,A^{q-1})\sim (f,1)^{r_e} $ (or $ (A,A^{q-1})\sim (1,f)^{r_e} $) then $(F_*^e( \frac{S}{f^{q-1}S})=\Cok_SA^{q-1}= \{0\}$ (or $(F_*^e( \frac{S}{fS})=\Cok_SA= \{0\}$) which is impossible. As a result, if $(A^k,A^{q-k})$ is a trivial matrix factorization of $f$, then the only possible case is that $(A^k,A^{q-k})\sim (f,1)^{b}\oplus (1,f)^{c}$ where $0 < b,c < r_e$ with $b+c=r_e$. In this case, $ F_*^e(S/f^kS)= R^b$
This shows that $F_*^e(S/(f^k))$, for every $e\in \mathbb{N}$ and $1 \leq k<p^e $, is a direct sums with direct summands taken from $\{ R , N_1,\dots,N_t\}$.
Now suppose $F_*^e(S/(f^k))$ is a direct sums with direct summands taken from indecomposable $R$-modules $ N_1,\dots,N_t $ for every $e\in \mathbb{N}$ and $1 \leq k<p^e $. Therefore, for each $1 \leq k \leq q-1$, there exist non-negative integers $n_{(e,k)},n_{(e,k,1)},\dots., n_{(e,k,t)}$ such that
\begin{equation}\label{E31}
\Cok_S(A^k,A^{q-k}) \cong F_*^e(S/f^kS)\cong R^{\oplus n_{(e,k)}} \oplus \bigoplus\limits_{j=1}^{t}N_j^{\oplus n_{(e,k,j)}} \text{ over } R.
\end{equation}
Since $F_*^e(S/f^kS)$ is a MCM $R$-module (by Corollary \ref{C4.12}), it follows that $N_j$ is an indecomposable non-free MCM $R$-module for each $j \in \{1,\dots,t\}$ and hence
by Proposition \ref{A} $N_j= \Cok_S(\alpha_j,\beta_j)$ for some reduced matrix factorization $(\alpha_j,\beta_j)$ for all $j $. If $M_j = \Cok_{S[\![u,v]\!]}(\alpha_j,\beta_j)^{\maltese}$, it follows that
$$ \Cok_{S[\![u,v]\!]}(A^k,A^{q-k})^{\maltese}= (R^{\bigstar})^{n_{(e,k)}} \oplus \bigoplus\limits_{j=1}^{t}M_j^{n_{(e,k,j)}} $$ and hence by the Equation \ref{Eq12} we conclude that the ring $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$ has FFRT by $\{ R^{\bigstar},M_1,\dots, M_t\}$.
\end{proof}
The following result is a direct application of the above proposition.
\begin{corollary}
Let $K$ be an algebraically closed field of prime characteristic $p > 2$ and $q=p^e$. Let $S:= K[\![x_1,\dots,x_{n}]\!]$ and let $ \mathfrak{m}$ be the maximal ideal of $S$ and $f \in \mathfrak{m}^2 \setminus \{0\}$. Let $R=S/(f)$ and $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$. If $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$ has FFRT over $R^{\bigstar}$, then $S/f^kS$ has FFRT over $S/f^kS$ for every positive integer $k$.
\end{corollary}
\begin{proof}
Let $k$ be a positive integer and let $e_0$ be a positive integer such that $k < p^{e_0}$. If $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$ has FFRT over $R^{\bigstar}$, from Theorem \ref{P30}, there exist finitely generated $R$-modules $N_1,\dots,N_t$ such that $F_*^e(S/(f^k))$ for each $e \geq e_0$ is a direct sums with direct summands taken from the finite set $\{R,N_1,\dots,N_t\}$. Notice from the proof of Theorem \ref{P30} that $N_j=\Cok_S(\alpha_j,\beta_j)$ for some reduced matrix factorization $(\alpha_j,\beta_j)$ of $f$ for all $1 \leq j \leq t $. As a result, $f\Cok_S(\alpha_j,\beta_j)=0$ and hence $f^k\Cok_S(\alpha_j,\beta_j)=0$ for every positive integer $k$. This makes $N_j$ a module over $S/f^kS$. Therefore,$F_*^e(S/(f^k))$ for each $e \geq e_0$ is a direct sums with direct summands taken from the finite set $\{R,N_1,\dots,N_t\}$ of the $S/f^kS$-modules. This is enough to show that $S/f^kS$ has FFRT over $S/f^kS$ for every positive integer $k$.
\end{proof}
The above corollary implies evidently the following.
\begin{corollary}
Let $K$ be an algebraically closed field of prime characteristic $p > 2$ and $q=p^e$. Let $S:= K[\![x_1,\dots,x_{n}]\!]$ and let $ \mathfrak{m}$ be the maximal ideal of $S$ and $f \in \mathfrak{m}^2 \setminus \{0\}$. Let $R=S/(f)$ and $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$. If $S/f^kS$ does not have FFRT over $S/f^kS$ for some positive integer $k$, then $R^{\bigstar}$ does not have FFRT. In particular, if $R$ does not have FFRT , then $R^{\bigstar}$ does not have FFRT.
\end{corollary}
An easy induction gives the following result.
\begin{corollary}
Let $K$ be an algebraically closed field of prime characteristic $p > 2$ and $q=p^e$. Let $S:= K[\![x_1,\dots,x_{n}]\!]$ and let $ \mathfrak{m}$ be the maximal ideal of $S$, $f \in \mathfrak{m}^2 \setminus \{0\}$ and let $R=S/(f)$. If $R$ does not have FFRT, then the ring $$S[\![u_1,v_1,u_2,v_2,\dots,u_t,v_t]\!]/(f+u_1v_1+u_2v_2+\dots+u_tv_t)$$ does not have FFRT for all $t\in \mathbb{N}$.
\end{corollary}
\section{Class of rings that have FFRT but not finite CM type}
\label{section:Class of rings that have FFRT but not finite CM type}
We keep the same notation as in Notation \ref{N4.1}. Recall that every F-finite local ring $(R,\mathfrak{m})$ of prime characteristic that has finite CM representation type has also FFRT (section \ref{section:Definition and examples}). The main result of this section is to provide a class of rings that have FFRT but not finite CM representation type Theorem \ref{P7.10}.
\begin{lemma}\label{L31}
If $ 0 \leq d,k \leq q-1 $, then
\begin{enumerate}
\item [(a)]$ d k = nq + s$ where $0 \leq n \leq d-1 $ and $ 0 \leq s \leq q-1 $.
\item [(b)]For any $ 0 \leq \beta \leq q-1 $, $ d k + \beta = cq + t $ where $ 0 \leq c \leq d $ and $ 0 \leq t \leq q-1 $.
\item [(c)] Fix $c,d,k \in \mathbb{Z}_{+}$ such that $0 \leq d,k \leq q-1 $ and $ 0 \leq c \leq d $. Then there exists an $ \alpha \in \{0,\ldots, q-1\}$ such that $\alpha = qc-dk+ s$ for some $ s \in \{0,\ldots, q-1\}$ if and only if $|qc-dk|< q$. Furthermore, If $|qc-dk|< q$, there exist $q-|qc-dk|$ values of $\alpha $ such that $\alpha = qc-dk+ s$ for some $ s \in \{0,\dots, q-1\}$.
\end{enumerate}
\end{lemma}
\begin{proof}
(a) By Division Algorithm, $ d k = nq + s $ for some $ n \in \mathbb{N} $ and $ 0 \leq s \leq q-1 $. If $ n \geq d $ , we get $ nq \geq dq > dk = nq + s \geq nq $ which is a contradiction. This shows that $ 0 \leq n \leq d-1 $. \\
(b) From the above result, we get that $ d k = nq + s$ where $ 0 \leq n \leq d-1 $ and $ 0 \leq s \leq q-1 $. Let $ 0 \leq \beta \leq q-1 $. If $ \beta + s < q $, we get $ d k + \beta = cq + t $ where $ c =n \in \{ 0 , ... , d-1 \} $ and $ t = \beta +s \in \{ 0 , ... , q-1 \} $. Now, suppose that $ \beta + s \geq q $. We notice that $ 0 \leq \beta + s -q \leq q -1$. Therefore $ d k + \beta = cq + t $ where $ c = n+1 \in \{ 1 , ... , d \} $ and $ t = \beta + s -q \in \{ 0 , ... , q-1 \} $.\\
(c) First, if there is $ \alpha \in \{0,\ldots, q-1\}$ such that $\alpha = qc-dk+ s$ for some $ s \in \{0,\ldots, q-1\}$, then $|qc-dk|=|\alpha -s|< q$ (as $ \alpha,s \in \{0,\ldots, q-1\}$). Now let $u = qc-dk$ and suppose that $|u| < q$. If $ 0 \leq u < q$, we can choose $\alpha \in \{u, u+1,\dots,q-1 \}$. On the other hand, if $ -q < u < 0$, then $\alpha$ can be taken from $\{q-1+u , q-2+u ,\dots,0\}$. In both cases, $\alpha$ can be chosen by $q-|c q - kd |$ ways.
\end{proof}
If $\alpha = (\alpha_1, \ldots , \alpha_n) \in \mathbb{Z_{+}}^n $, we write $ x^\alpha = x_1^{\alpha_1}\dots x_n^{\alpha_n}$ where $x_1,\dots,x_n$ are different variables.
\begin{proposition}\label{L4.25}
Let $f = x_1^{d_1}x_2^{d_2} \dots x_n^{d_n} $ be a monomial in $S$ where $d_j \in \mathbb{Z_{+}}$ for each $j$. Let $\Gamma = \{(\alpha_1, \ldots , \alpha_n) \in \mathbb{Z_{+}}^n \, | \, 0 \leq \alpha_j\leq d_j \text{ for all } 1\leq j \leq n \}$ , $d=(d_1,\ldots,d_n)$, and let $e$ be a positive integer such that $q=p^e > \max \{ d_1 , \dots , d_n \}+1$. If $A=M_S(f,e)$, then for each $1 \leq k \leq q-1$ the matrix $A^k=M_S(f^k,e)$ is equivalent to diagonal matrix, $D$, of size $r_e \times r_e$ in which the diagonal entries are of the form $x^c$ where $c \in \Gamma$. Furthermore, if $c=(c_1,\dots,c_n) \in \Gamma $ and
\begin{equation*}
\eta_k(c_j)= \begin{cases}q- | c_jq - kd_j | \text { if } | c_jq - kd_j | < q \\
0 \text{ otherwise }
\end{cases}
\end{equation*}
then
\begin{equation*}
\Cok_S(A^q,A^{q-k}) = \bigoplus_{c \in \Gamma } \left[ \Cok_S(x^c, x^{d-c})\right]^{\oplus \eta_k(c)}
\end{equation*}
where $ \eta_k(c) = [K: K^q]\prod_{j=1}^n \eta_k(c_j) $ and $(x^c, x^{d-c})$ is the $1 \times 1$ matrix factorization of $f$ with the convention that $M^{\oplus0}=\{0\}$ for any module $M$.
\end{proposition}
\begin{proof}
Choose $e \in \mathbb{N}$ such that $q=p^e > \max \{ d_1 , \dots , d_n \}+1$ and let $1\leq k \leq q-1 $. If $j=\lambda x_1^{\beta_1}\dots x_n^{\beta_n}\in \Delta_e$, we get $F_*^e(jf^k)=F_*^e(\lambda x_1^{kd_1+\beta_1}\dots x_n^{kd_n+\beta_n})$. Since $d_j , k \in \{ 0 , \dots , q-1 \} $, Lemma \ref{L31} implies that there exist $ 0 \leq c_{i} \leq d_i $ and $ 0 \leq u_{i}\leq q-1 $ for each $1 \leq i \leq n$ such that $ d_i k + \beta_{i} = c_{i}q + u_{i} $ and hence $ F_*^e(jf^k) = x_1^{c_{1}}\dots x_n^{c_{n}}F_*^e(\lambda x_1^{u_{1}}\dots x_n^{u_{n}}) $.
Therefore each column of $M_S(f^k,e)$ contains only one non-zero element of the form $ x_1^{c_{1}}\dots x_n^{c_{n}} $ where $ 0 \leq c_{i}\leq d $ for all $1 \leq i \leq n$. Notice that a row in $M_S(f^k,e)$ will contain two elements of the form $ x_1^{c_{1}}\dots x_n^{c_{n}} $ and $ x_1^{l_{1}}\dots x_n^{l_{n}} $ where $ 0 \leq l_{j} ,c_{i}\leq d $ for all $1 \leq i \leq n$ if there exist $\mu x_1^{\beta_1}\dots x_n^{\beta_n},\gamma x_1^{\sigma_1}\dots x_n^{\sigma_n}, \lambda x_1^{u_{1}}\dots x_n^{u_{n}}\in \Delta_e$ such that
\begin{equation*}
F_*^e((\mu x_1^{\beta_1}\dots x_n^{\beta_n})( x_1^{kd_1}\dots x_n^{kd_n})) = x_1^{c_{1}}\dots x_n^{c_{n}}F_*^e(\lambda x_1^{u_{1}}\dots x_n^{u_{n}}) \text{ and }
\end{equation*}
\begin{equation*}
F_*^e((\gamma x_1^{\sigma_1}\dots x_n^{\sigma_n})( x_1^{kd_1}\dots x_n^{kd_n})) = x_1^{l_{1}}\dots x_n^{l_{n}}F_*^e(\lambda x_1^{u_{1}}\dots x_n^{u_{n}}).
\end{equation*}
This makes $\mu= \lambda= \gamma $, $\beta_i+kd_i= qc_i+u_i$, and $\sigma_i+kd_i= ql_i+u_i$ for all $1 \leq i \leq n$.
Accordingly, $\beta_i- \sigma_i = q(c_i-l_i)$ for all $1 \leq i \leq n$. Since $0\leq \beta_i, \sigma_i\leq q-1$ and $0\leq c_i, l_i\leq d \leq q-1$ for all $1 \leq i \leq n$, it follows that $\beta_i= \sigma_i$ and $c_i=l_i$ and for all $1 \leq i \leq n$. This also shows that each row of $M_S(f^k,e)$ contains only one non-zero element of the form $ x_1^{c_{1}}\dots x_n^{c_{n}} $ where $ 0 \leq c_{i}\leq d $ for all $1 \leq i \leq n$. Since each column and row of $M_S(f^k,e)$ contains only one non-zero element of the form $ x_1^{c_{1}}\dots x_n^{c_{n}} $ where $ 0 \leq c_{i}\leq d $ for all $1 \leq i \leq n$, using the row and column operations, the matrix $M_S(f^k,e)$ is equivalent to a diagonal matrix, $D$, of size $r_e \times r_e$ in which the diagonal entries are of the form $ x_1^{c_1}\dots x_n^{c_n} $ where $ 0 \leq c_{i}\leq d_i $ for all $1 \leq i \leq n$. Now fix $c=(c_1,\dots,c_n) \in \Gamma$ and let $\eta(c)$ stand for how many times $x^c$ appears as an element in the diagonal of $D$. It is obvious that $\eta(c)$ is exactly the same as the number of the $n$-tuples $(\alpha_1,\dots,\alpha_n)$ with $ 0 \leq \alpha_j \leq q-1 $ satisfying that
\begin{equation}\label{EEE2}
F_*^e(\lambda x_1^{kd_1+\alpha_1}\dots x_n^{kd_n+\alpha_n}) = x_1^{c_1}\dots x_n^{c_n}F_*^e(\lambda x_1^{s_1}\dots x_n^{s_n})
\end{equation}
for some $ s_1,..,s_n \in \{0,\dots,q-1 \}$ for all $\lambda \in \Lambda_e$. Notice that an $n$-tuple $(\alpha_1,\dots,\alpha_n)$ with $ 0 \leq \alpha_i \leq q-1 $ will satisfy (\ref{EEE2}) if and only if $\alpha_i= c_iq - kd_i + s_i$ for some $0 \leq s_i \leq q-1 $ for all $1\leq i \leq n$. As a result, by Lemma\ref{L31}, there exists an $n$-tuples $(\alpha_1,\dots,\alpha_n)\in \mathbb{Z}^n$ with $ 0 \leq \alpha_i \leq q-1 $ that satisfies (\ref{EEE2}) if and only if $|c_iq - kd_i| < q$ for all $1\leq i \leq n$. Furthermore, by Lemma\ref{L31}, for each $i\in \{1,\dots,n\}$, if $|c_iq - kd_i| < q$, then there exist $q-|c_iq - kd_i|$ values of $\alpha_i $ such that $\alpha_i = qc_i-d_ik+ s_i$ for some $ s_i \in \{0,\dots, q-1\}$.
Set
\begin{equation*}
\eta_k(c_j)= \begin{cases}q- |c_iq - kd_i | \text { if } | c_iq - kd_i | < q \\
0 \text{ otherwise }
\end{cases}
\end{equation*}
Thus we get that $ \eta_k(c) = [K: K^q]\prod_{j=1}^n \eta_k(c_i) $ and consequently we have
\begin{equation*}
\Cok_S(A^q,A^{q-k}) = \bigoplus_{c \in \Gamma } \left[ \Cok_S(x^c, x^{d-c})\right]^{\oplus \eta_k(c)}
\end{equation*}
where $(x^c, x^{d-c})$ is the $1 \times 1$ matrix factorization of $f$ with the convention that $M^{\oplus0}=\{0\}$ for any module $M$.
\end{proof}
\begin{corollary}\label{C32}
Let $K$ be an algebraically closed field of prime characteristic $p > 2$ and $q=p^e$. Let $S:= K[\![x_1,\dots,x_{n}]\!]$, $f =x_1^{d_1}x_2^{d_2} \dots x_n^{d_n} $ where $d_j \in \mathbb{N}$ for each $j$, and $d=(d_1,\ldots,d_n)$. Then $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$ has FFRT over $R^{\bigstar}$. Furthermore, for every $e \in \mathbf{N}$ with $q=p^e > \max \{ d_1 , \dots , d_n \}+1$, $F_*^e(R^{\bigstar})$ has the following decomposition:
\begin{equation*}
F_*^e(R^{\bigstar})= (R^{\bigstar})^{r_e} \bigoplus \bigoplus_{k=1}^{q-1}\left[\bigoplus_{c \in \Gamma } \left[\Cok_{S[\![u,v]\!]}(x^c,x^{d-c})^{\maltese}\right]^{\oplus \eta_k(c)} \right]
\end{equation*}
where $\eta_k(c)$ and $\Gamma$ as in the above Proposition.
\end{corollary}
\begin{proof}
Let $e \in \mathbb{N}$ with $q=p^e > \max \{ d_1 , \dots , d_n \}+1$ and let $1 \leq k \leq q-1 $. Let $\Gamma $ and $ \eta_k(c)$ be as in the above Proposition. If $A=M_S(f,e)$, it follows that
\begin{equation*}
F_*^e(S/f^k) \cong \Cok_S(A^k,A^{q-k}) \cong \bigoplus_{c \in \Gamma } \left[ \Cok_S(x^c, x^{d-c})\right]^{\oplus \eta_k(c)}.
\end{equation*}
If $ \mathfrak{M}= \{\Cok_S(x^c, x^{d-c})\, | \, c \in \Gamma \}\cup \{ F^j(S/f^i)\, |\, p^j \leq \max \{ d_1 , \dots , d_n \} \text{ and } 0 \leq i \leq p^j\} $,
then $F_*^e(S/(f^k))$ is a direct sums with direct summands taken from the finite set $\mathfrak{M}$ for every $e\in \mathbb{N}$ and $1 \leq k<p^e $. By Theorem \ref{P30} $R^{\bigstar}$ has FFRT.
Furthermore, we can describe explicitly the direct summands of $F_*^e(R^{\bigstar})$. Indeed, if $\hat{\Gamma}:= \{ c \in \Gamma \, | \, \eta_k(c)> 0 \text{ and } c \notin \{d,0\} \} $, it follows that $$(A^k,A^{q-k}) \sim \bigoplus_{c \in \hat{\Gamma} } (x^c,x^{d-c})^{\oplus \eta_k(c)}\bigoplus (x^d,1)^{\oplus \eta_k(d)}\bigoplus (1,x^d)^{\oplus \eta_k(0)} $$ where $ \eta_k(d)$ (respectively $ \eta_k(0)$) denotes how many times $x^d$ (respectively $1$) appears in $A^k$.
Recall by Remark \ref{R3.8} that $( A^k, A^{q-k} )^{\maltese}$ is a matrix factorization of $f+uv$ and
$$(A^k,A^{q-k})^{\maltese} \sim \bigoplus_{c \in \hat{\Gamma} } \left[(x^c,x^{d-c})^{\maltese}\right]^{\oplus \eta_k(c)}\bigoplus \left[(x^d,1)^{\maltese}\right]^{\oplus \eta_k(d)}\bigoplus \left[(1,x^d)^{\maltese}\right]^{\oplus \eta_k(0)} $$
Therefore
\begin{eqnarray*}
\Cok_{S[\![u,v]\!]}(A^k,A^{q-k})^{\maltese} &=& \bigoplus_{c \in \hat{\Gamma} } \left[\Cok_{S[\![u,v]\!]}(x^c,x^{d-c})^{\maltese}\right]^{\oplus \eta_k(c)}\\
& & \bigoplus \left[\Cok_{S[\![u,v]\!]}(x^d,1)^{\maltese}\right]^{\oplus \eta_k(d)} \\
& & \bigoplus \left[\Cok_{S[\![u,v]\!]}(1,x^d)^{\maltese}\right]^{\oplus \eta_k(0)}.
\end{eqnarray*}
By Proposition \ref{P21}, the above equation, and the convention that $M^{\oplus 0}=\{0\}$, we can write
\begin{equation*}
F_*^e(S[\![u,v]\!]/(f+uv)) = (R^{\bigstar})^{q^n} \bigoplus \bigoplus_{k=1}^{q-1}\left[\bigoplus_{c \in \Gamma } \left[\Cok_{S[\![u,v]\!]}(x^c,x^{d-c})^{\maltese}\right]^{\oplus \eta_k(c)} \right].
\end{equation*}
\end{proof}
We benefit from the proof of Proposition \ref{L4.25} above when we compute the $F$-signature in the next chapter.
The following theorem provides an example of rings that have FFRT but not finite CM type.
\begin{theorem}\label{P7.10}
Let $K$ be an infinite algebraically closed field with $\charact (K) > 2$,
and let $S = K[\![x_1, \dots, x_d]\!]$ where $d > 2$.
If $ f \in S$ is a monomial of degree grater than $3$ and $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$, then $R^{\bigstar}$ has FFRT but it does not have finite CM representation type.
\end{theorem}
\begin{proof}
Let $t$ be the degree of the monomial $f$ and let $\mathfrak{m}$ be the maximal ideal of $S$. Clearly, $t$ is the largest natural number satisfying $ f \in \mathfrak{m}^t - \mathfrak{m}^{t+1}$ and consequently the multiplicity $e(R)$ of the ring $R$ is $e(R)=t$ (Proposition \ref{F1}). Since $e(R)=t > 3$, it follows from Proposition \ref{F2} that $R$ is not a simple singularity. Therefore, by Proposition \ref{F3} $R$ does not have finite CM type. Consequently, by Proposition \ref{P28}, $R^{\bigstar}$ does not have finite CM type as well. However, Corollary \ref{C32} implies that $R^{\bigstar}$ has FFRT.
\end{proof}
\section{$S/I$ has FFRT when $I$ is a monomial ideal}
\label{section:Qutient by a monomial ideal}
We keep the same notation as in in Notation \ref{N4.1}.
If $x_1,...,x_n$ are variables, a monomial
in $x_1, . . . , x_n $ is an element of the form $x_1^{r_1}...x_n^{r_n}$ where $r_1,...,r_n\in \mathbb{Z}_{+}$. If $R = A[x_1, . . . , x_n]$ (or $R = A[\![x_1, . . . , x_n]\!]$) where $A$ is a nonzero commutative ring with identity. A monomial ideal in $R$ is an ideal
of $R$ that can be generated by monomials in $x_1, . . . , x_n$.
We need the following Proposition in order to prove Proposition \ref{P5.24}.
\begin{proposition}\label{P4.19}
Let $f_1,...,f_t$ be nonzero and non unite elements in $S$ and let $R= S/( f_1,...,f_t )S$. If $[ M_S(f_1,e)... M_S(f_t,e)]$ is the $r_e \times tr_e$ matrix over $S$ whose columns are the columns of the matrices $M_S(f_1,e),..., M_S(f_t,e)$ respectively, then :\\
1) $F_*^e(R)$ is isomorphic to $\Cok_S[ M_S(f_1,e)... M_S(f_t,e)]$ as $S$ -modules.\\
2) $F_*^e(R)$ is isomorphic to $\Cok_S[ M_S(f_1,e)... M_S(f_t,e)]$ as $R$ -modules.\\
\end{proposition}
\begin{proof}
Let $I$ be the ideal $( f_1,...,f_t )S$. Since $ \{ F_*^e(j) | j \in \Delta_e \} $ is a basis of $F_*^e(S)$ as free $S$-module, $F_*^e(R)$ is generated by $ \{ F_*^e(j + I) \,| \,j \in \Delta_e \} $ as $S$-module. For every $ F_*^e(g) \in F_*^e(S)$, define $ \phi( F_*^e(g)) = F_*^e(g + I)$. Then $\phi: F_*^e(S) \longrightarrow F_*^e(R)$ is a surjective homomorphism of $S$-modules. For every $1 \leq k \leq t$ recall that $M_S(f_k,e)=[f^{(k)}_{(i,j)}]$, where $f^{(k)}_{(i,j)}$, indexed by $i,j \in \Delta_e$, satisfies that $F_*^e(jf_k)= \bigoplus_{i\in\Delta_e}f^{(k)}_{(i,j)}F_*^e(i)$. Now, define the $S$-module homomorphism
$ \psi : F_*^e(S)^{\oplus t} \rightarrow F_*^e(S)$ by
\begin{equation*}
\psi [(F_*^e(g_1) , \ldots , F_*^e(g_t) )] = F_*^e(g_1f_1)+ \ldots +F_*^e(g_tf_t)
\end{equation*}
for all $(F_*^e(g_1) , \ldots , F_*^e(g_t) )\in F_*^e(S)^{\oplus t} $. Since $\Ima \psi = \Ker \phi= F_*^e(I)$, we have an exact sequence $F_*^e(S)^{\oplus t}\xrightarrow{\psi}F_*^e(S)\xrightarrow{\phi}F_*^e(R) \xrightarrow{} 0$. For every $j \in \Delta_e $ and $1 \leq k \leq t$ define $j^{(k)}$ to be the element in $ F_*^e(S)^{\oplus t} $ whose $k$th coordinate is $F_*^e(j)$ and zero elsewhere and let $\Omega_e^{(k)}=\{ j^{(k)} \,|\, j \in \Delta_e\}$. Since $ \{ F_*^e(j)\, |\, j \in \Delta_e \} $ is a basis of $F_*^e(S)$ as free $S$-module, it follows that $\Omega_e=\Omega_e^{(1)}\cup...\cup \Omega_e^{(t)} $ is a basis for $ F_*^e(S)^{\oplus t}$ as free $S$-module. Notice for each $j \in \Delta_e$ and $1 \leq k \leq t$ that $\psi(j^{(k)})= F_*^e(jf_k)= \bigoplus_{i\in\Delta_e}f^{(k)}_{(i,j)}F_*^e(i)$
and hence the matrix $[ M_S(f_1,e)... M_S(f_t,e)]$
represents the map $\psi$ on the given free-bases (Remark \ref{Rem 2.9}). This proves that $F_*^e(R)$ is isomorphic $\Cok_S[ M_S(f_1,e)... M_S(f_t,e)]$ as $S$ -modules.
2) Since $IF_*^e(R) =0$ and $I \Cok_S[ M_S(f_1,e)... M_S(f_t,e)]=I\Cok \psi=0$, it follows that $F_*^e(R)$ is isomorphic $\Cok_S[ M_S(f_1,e)... M_S(f_t,e)]$ as $R$ -modules.
\end{proof}
\begin{proposition}\label{P5.24}
Let $S$ denote the ring $K[x_1,...,x_n]$ or the ring $K[\![x_1,...,x_n]\!]$. For each $1 \leq j \leq t$, let $ f_j = x_1^{d_{(1,j)}} \ldots x_n^{d_{(n,j)}}$ where $d_{(i,j)} \in \mathbb{Z}_{+}$ for all $i$ and $j$ and set
$$ \mathcal{G}_j = \{x_1^{m_{(1,j)}} \ldots x_n^{m_{(n,j)}} \, | \, 0 \leq m_{(i,j)}\leq d_{(i,j)}, 1 \leq i \leq n \} .$$
Let $I$ be the monomial ideal $I=(f_1,...,f_n)$ and let $\mathfrak{J}$ be the set of all ideals $J=(g_1,...,g_t)$ where $g_j \in \mathcal{G}_j $ for all $1 \leq j \leq t$. If $R=S/I$, then
$$F_*^e(R)=\bigoplus_{J \in \mathfrak{J}}[S/J]^{\oplus \alpha_e(J)} \text{ where } \alpha_e(J) \in \mathbb{Z}_{+}$$
with the convention that $ [S/J]^{\oplus \alpha_e(J)} = \{0\}$ when $ \alpha_e(J)=0$.
\end{proposition}
\begin{proof}
For every $1\leq j \leq t$, notice that the proof of Proposition \ref{L4.25} shows that each column and each row of $M_S(f_j,e)$ contains only one none zero element of the set $\mathcal{G}_j $. Therefore, if $A=[ M_S(f_1,e)... M_S(f_t,e)]$, then each row of $A$ contains only $t$ none zero elements and each column contains only one none zero element such that all of them belong to $\mathcal{G}_1 \cup ... \cup \mathcal{G}_t$. Let $\Upsilon$ be the set of all $1 \times t$ matrix of the form $\left[
\begin{array}{ccc}
g_1 & \ldots & g_t \\
\end{array}
\right] $ with $g_j \in \mathcal{G}_j$ for all $1\leq j \leq t$. Using the row and column operations, the matrix $A$ is equivalent to an $r_e \times tr_e$ matrix of the form
\begin{equation}\label{Eq4.39}
A \sim \left[
\begin{array}{ccccc}
A_1 & & & & \\
& & \ddots & & \\
& & & & A_{r_e} \\
\end{array}
\right]
\end{equation}
where $A_i \in \Upsilon$ for all $ i \in \{1,...,r_e\}$.
Notice for every $ i \in \{1,...,r_e\}$ that $\Cok_SA_i= S/J$ for some $J \in \mathfrak{J}$ and that the Proposition \ref{P4.19} implies that $ F_*^e(R)=\Cok_SA=\bigoplus_{i=0}^{r_e}\Cok_SA_i$ where $A_i \in \Upsilon$ .
Therefore, we can write $$F_*^e(R)=\bigoplus_{J \in \mathfrak{J}}[S/J]^{\oplus \alpha_e(J)} \text{ where } \alpha_e(J) \in \mathbb{Z}_{+}$$
with the convention that $ [S/J]^{\oplus \alpha_e(J)} = \{0\}$ when $ \alpha_e(J)=0$. Since $\Omega$ is finite set, the set $\Upsilon$ is also a finite set.
\end{proof}
Proposition \ref{P5.24} implies the following result. However, the following result can be obtained from \cite[Example 1.3 (v)]{TT} but the above proposition provides another proof.
\begin{theorem}\label{P5.25}
Let $S$ denote the ring $K[x_1,...,x_n]$ or the ring $K[\![x_1,...,x_n]\!]$. Let $I$ be a monomial ideal in $S$ generated by the monomials $f_1,...,f_t$. If $R=S/I$, then $R$ has Finite F-representation type on $R$.
\end{theorem}
\chapter{F-signature of specific hypersurfaces}
\label{chapter:F-signature of specific hypersurfaces}
Recall from Remark \ref{C2.10} that if $R$ is an $F$-finite local ring, for every $e \in \mathbb{N}$ there exist a unique nonnegative integer $a_e$ and an $R$-module $M_e$ that has no free direct summand such that $F_*^e(R)=R^{a_e}\oplus M_e$ and $a_e(R)=\sharp (F_*^e(R),R)=a_e$. Now we are ready to define the $F$-signature as it appears in \cite{KT} as follows.
\begin{definition}\label{D6.1}
\emph{Let $(R,\mathfrak{m},K)$ be a $d$-dimensional $F$-finite Noetherian local ring of prime characteristic
$p$. If $[K:K^p] $ is the dimension of $K$ as $K^p$-vector space and $\alpha (R) =\log_p[K:K^p]$, then the $F$-signature of $R$, denoted $\mathbb{S}(R)$, is defined as
\begin{equation*}
\mathbb{S}(R)=\lim_{e\rightarrow \infty} \frac{a_e(R)}{p^{e(d+\alpha (R))}}.
\end{equation*}
}
\end{definition}
\begin{proposition}\label{R6.2}
If $(R,\mathfrak{m},K)$ is as above, then $\mathbb{S}(R)= \mathbb{S}(\hat{R})$ where $\hat{R}$ is the $\mathfrak{m}$-adic completion of $R$.
\end{proposition}
\begin{proof}
If $L$ denotes the residue field of the local ring $(\hat{R}, \hat{\mathfrak{m}})$, then $L$ is isomorphic to $\hat{K}$ where $\hat{K}$ is the $\mathfrak{m}$-adic completion of $K$. Since $\hat{K}$ is isomorphic to $K$, it follows that $\alpha (R) =\alpha (\hat{R})$.
It is well known that $\dim R= \dim \hat{R}$ \cite[Corollary 10.2.2]{BCA}. Now, if $a_e(R)=a_e$, we can write $F_*^e(R)=R^{a_e}\oplus M_e$ where $M_e$ is an $R$-module that has no free direct summand. As a result, we get $F_*^e(\hat{R})=\hat{R}^{a_e}\oplus \hat{M_e}$. However, if $\hat{R}$ is a direct summand of $\hat{M_e}$, it follows from Proposition \ref{Pro2.22} (a) that $R$ is a direct summand of $M_e$ which is a contradiction. Therefore, $a_e(\hat{R})=a_e(R)$ and consequently $\mathbb{S}(R)= \mathbb{S}(\hat{R})$.
\end{proof}
\begin{remark}\label{R6.3}
If $R=\bigoplus_{n=0}^{\infty}R_n$ is a graded ring with $R_0$
equals to a field $K$ of characteristic $p> 0$, then for every $e \in \mathbb{N}$ there exist a unique nonnegative integer $a_e$ and an $R$-module $M_e$ that has no free direct summand such that $F_*^e(R)=R^{a_e}\oplus M_e$. If $a_e(R)=a_e$ and $\alpha (R) =\log_p[K:K^p]$ where $[K:K^p] $ is the dimension of $K$ as $K^p$-vector space, then the limit $\mathbb{S}(R)=\lim_{e\rightarrow \infty} \frac{a_e(R)}{p^{e(d+\alpha (R))}}$ is well-defined \cite[Lemma 6.6]{MVK}. Furthermore, if $ \mathfrak{m}$ is the homogenous maximal ideal of $R$, then $a_e(R)=a_e(R_{\mathfrak{m}})$ and $\mathbb{S}(R)=\mathbb{S}(R_{\mathfrak{m}})$ \cite[Lemma 6.6]{MVK}.
According to Proposition \ref{R6.2}, $\mathbb{S}(R_{\mathfrak{m}})= \mathbb{S}(\widehat{R_{\mathfrak{m}}})$ where $\widehat{R_{\mathfrak{m}}}$ is the $\mathfrak{m}R_{\mathfrak{m}}$-adic completion of the ring $R_{\mathfrak{m}}$. Since $\widehat{R_{\mathfrak{m}}}$ is isomorphic to $\hat{R}$ where $\hat{R}$ is the $\mathfrak{m}$-adic completion of $R$ \cite[Section 22]{AK}, it follows that $\mathbb{S}(R)=\mathbb{S}(\hat{R})$.
\end{remark}
In this chapter, we will compute the $F$-signature of some hypersurfaces.
\section{The F-signature of $ \frac{S[\![u,v]\!]}{f+uv}$ when $f$ is a monomial}
\label{section:The F-signature of uv is a monomial}
We will keep the same notation as in Notation \ref{N4.1} unless otherwise stated. \\
\begin{notation}\label{N1}
\emph{Let $\Delta = \{1,\dots,n \} $ and let $d, d_1,\dots,d_n$ be real numbers. For every $1\leq s \leq n-1$, define
\begin{equation*}
W^{(n)}_s = \sum_{j_1,\dots,j_s \in \Delta } [(d-d_{j_1})\dots(d- d_{j_s})(\prod_{j \in \Delta \setminus \{j_1,\dots,j_s\}}d_j)]
\end{equation*}
\begin{equation*}
W^{(n)}_n= \prod_{i=1}^{n}(d-d_{i}) \text{ and } W^{(n)}_0= \prod_{i=1}^{n}d_{i}.
\end{equation*}
}
\end{notation}
For example, if $d, d_1,d_2,d_3,d_4$ are real numbers, we get that $W^{(4)}_0=d_1 d_2 d_3 d_4$
\begin{eqnarray*}
W^{(4)}_1 &=&(d-d_1)d_2d_3d_4 + (d-d_2)d_1d_3d_4+(d-d_3)d_1d_2d_4 \\
&+& (d-d_4)d_1d_2d_3
\end{eqnarray*}
\begin{eqnarray*}
W^{(4)}_2 &=& (d-d_1)(d-d_2)d_3d_4+(d-d_1)(d-d_3)d_2d_4+ (d-d_1)(d-d_4)d_2d_3 \\
&+& (d-d_2)(d-d_3)d_1d_4 + (d-d_2)(d-d_4)d_1d_3
\end{eqnarray*}
\begin{eqnarray*}
W^{(4)}_3 &=&(d-d_2)(d-d_3)(d-d_4)d_1+ (d-d_1)(d-d_3)(d-d_4)d_2 \\
&+&(d-d_1)(d-d_2)(d-d_4)d_3+(d-d_1)(d-d_2)(d-d_3)d_4
\end{eqnarray*}
and $W^{(4)}_4=(d-d_1)(d- d_2)(d- d_3)(d- d_4).$
According to the above notation, we can observe the following remark.
\begin{remark}\label{R5.3}
Let $d, d_1,\dots,d_n$ be real numbers where $n\geq 1$ and let $W^{(n)}_j$ be defined on $d, d_1,\dots,d_n$ as in \ref{N1}. If $d_{n+1}$ is a real number, then $W^{(n+1)}_j$ is defined on $d, d_1,\dots,d_n,d_{n+1}$ for all $ 1 \leq j \leq n $ as follows: $$W^{(n+1)}_j= d_{n+1} W^{(n)}_j+(d-d_{n+1}) W^{(n)}_{j-1}.$$
Furthermore, $W^{(n+1)}_0=d_{n+1}W^{(n)}_0$ and $W^{(n+1)}_{n+1}=(d-d_{n+1})W^{(n)}_n$.
\end{remark}
The following lemma is needed to prove Proposition \ref{P8.4}.
\begin{lemma}\label{L5.4}
If $r$ , $q$, $d_j$ and $u_j$ are real numbers for all $1 \leq j \leq n$, then
\begin{equation*}
\prod_{j=1}^{n}(d_jr + \frac{q(d-d_j)}{d}+ u_j) = \sum_{j=0}^{n} \frac{q^j}{d^j}W^{(n)}_jr^{n-j}+ \sum_{c=0}^{n-1}g^{(n)}_c(q)r^{c}
\end{equation*}
where $g^{(n)}_c(q)$ is a polynomial in $q$ of degree $n-1-c$ for all $0\leq c \leq n-1$.
\end{lemma}
\begin{proof}
By induction on $n$, we will prove this lemma. It is clear when $n=1$. The induction hypothesis implies that
\begin{eqnarray*}
\prod_{j=1}^{n+1}(d_jr + \frac{q(d-d_j)}{d}+ u_j) &=& (d_{n+1}r + \frac{q(d-d_{n+1})}{d}+ u_{n+1})\prod_{j=1}^{n}(d_jr + \frac{q(d-d_j)}{d}+ u_j) \\
&=& (d_{n+1}r + \frac{q(d-d_{n+1})}{d}+ u_{n+1})(\sum_{j=0}^{n} \frac{q^j}{d^j}W^{(n)}_jr^{n-j}+ \\
& & \sum_{c=0}^{n-1}g^{(n)}_c(q)r^{c})\\
&=& A+B+C
\end{eqnarray*}
where
\begin{eqnarray*}
A &=& d_{n+1}r (\sum_{j=0}^{n} \frac{q^j}{d^j}W^{(n)}_jr^{n-j}+ \sum_{c=0}^{n-1}g^{(n)}_c(q)r^{c})\\
&=& \sum_{j=0}^{n}d_{n+1} \frac{q^j}{d^j}W^{(n)}_jr^{n-j+1}+\sum_{c=0}^{n-1}d_{n+1}g^{(n)}_c(q)r^{c+1}\\
B &=& \frac{q(d-d_{n+1})}{d}(\sum_{j=0}^{n} \frac{q^j}{d^j}W^{(n)}_jr^{n-j}+ \sum_{c=0}^{n-1}g^{(n)}_c(q)r^{c})\\
&=& \sum_{j=0}^{n}(d-d_{n+1}) \frac{q^{j+1}}{d^{j+1}}W^{(n)}_jr^{n-j} +\sum_{c=0}^{n-1}\frac{q(d-d_{n+1})}{d}g^{(n)}_c(q)r^{c}\\
C &=& u_{n+1} (\sum_{j=0}^{n} \frac{q^j}{d^j}W^{(n)}_jr^{n-j}+ \sum_{c=0}^{n-1}g^{(n)}_c(q)r^{c})\\
&=& \sum_{j=0}^{n} u_{n+1}\frac{q^j}{d^j}W^{(n)}_jr^{n-j}+\sum_{c=0}^{n-1}u_{n+1}g^{(n)}_c(q)r^{c}.
\end{eqnarray*}
Write $A=A_1+A_2$ where
\begin{eqnarray*}
A_1 &=& \sum_{j=0}^{n}d_{n+1} \frac{q^j}{d^j}W^{(n)}_jr^{n-j+1} \text{ and } \\
A_2 &=& \sum_{c=0}^{n-1}d_{n+1}g^{(n)}_c(q)r^{c+1}
\end{eqnarray*}
and write $B=B_1+B_2$ where
\begin{eqnarray*}
B_1 &=& \sum_{j=0}^{n}(d-d_{n+1}) \frac{q^{j+1}}{d^{j+1}}W^{(n)}_jr^{n-j} \text{ and } \\
B_2 &=& \sum_{c=0}^{n-1}\frac{q(d-d_{n+1})}{d}g^{(n)}_c(q)r^{c}.
\end{eqnarray*}
Notice that
\begin{eqnarray*}
A_1+B_1&=& \sum_{j=0}^{n}d_{n+1} \frac{q^j}{d^j}W^{(n)}_jr^{n-j+1} + \sum_{j=0}^{n}(d-d_{n+1}) \frac{q^{j+1}}{d^{j+1}}W^{(n)}_jr^{n-j} \\
&=& d_{n+1}W^{(n)}_0r^{n+1}+ \sum_{j=1}^{n}d_{n+1} \frac{q^j}{d^j}W^{(n)}_jr^{n-j+1}\\
& & +\sum_{j=0}^{n-1}(d-d_{n+1}) \frac{q^{j+1}}{d^{j+1}}W^{(n)}_jr^{n-j}+ (d-d_{n+1}) \frac{q^{n+1}}{d^{n+1}}W^{(n)}_n \\
&=& d_{n+1}W^{(n)}_0r^{n+1}+ \sum_{j=1}^{n}d_{n+1} \frac{q^j}{d^j}W^{(n)}_jr^{n-j+1}\\
& & +\sum_{j=1}^{n}(d-d_{n+1}) \frac{q^{j}}{d^{j}}W^{(n)}_{j-1}r^{n-j+1}+ (d-d_{n+1}) \frac{q^{n+1}}{d^{n+1}}W^{(n)}_n \\
&=& d_{n+1}W^{(n)}_0r^{n+1}+ \sum_{j=1}^{n}\frac{q^j}{d^j}[d_{n+1} W^{(n)}_j+(d-d_{n+1}) W^{(n)}_{j-1}]r^{n-j+1} \\
& & +(d-d_{n+1}) \frac{q^{n+1}}{d^{n+1}}W^{(n)}_n.
\end{eqnarray*}
Now apply Remark \ref{R5.3} to get that
\begin{equation}\label{E.8.14}
A_1+B_1= \sum_{j=0}^{n+1} \frac{q^j}{d^j}W^{(n+1)}_jr^{n+1-j}.
\end{equation}
Now define
$$g^{(n+1)}_n(q)=d_{n+1}g^{(n)}_{n-1}(q)+u_{n+1}W^{(n)}_0,$$
$$g^{(n+1)}_0(q)=\frac{q}{d}(d-d_{n+1})g^{(n)}_0(q)+u_{n+1}\frac{q^n}{d^n}W^{(n)}_n+u_{n+1}g^{(n)}_0(q),$$
and
\begin{equation*}
g^{(n+1)}_i(q)= d_{n+1}g^{(n)}_{i-1}(q)+ \frac{q(d-d_{n+1})}{d}g^{(n)}_{i}(q)+ u_{n+1}\frac{q^{n-i}}{d^{n-i}}W^{(n)}_{n-i}+u_{n+1}g^{(n)}_{i}(q)
\end{equation*}
for every $1\leq i \leq n-1 $.
Since $g^{(n)}_i(q)$ is a polynomial in $q$ of degree $n-1-i$ for all $0\leq i \leq n-1$, it follows from the above definitions that $g^{(n+1)}_i(q)$ is a polynomial
in $q$ of degree $n-i$ for all $0\leq i \leq n $.
Notice that
\begin{eqnarray*}
A_2+ B_2 + C &=& \sum_{c=0}^{n-1}d_{n+1}g^{(n)}_c(q)r^{c+1} + \sum_{c=0}^{n-1}\frac{q(d-d_{n+1})}{d}g^{(n)}_c(q)r^{c} \\
& & + \sum_{j=0}^{n} u_{n+1}\frac{q^j}{d^j}W^{(n)}_jr^{n-j}+\sum_{c=0}^{n-1}u_{n+1}g^{(n)}_c(q)r^{c} \\
&=& \sum_{c=0}^{n-2}d_{n+1}g^{(n)}_c(q)r^{c+1}+d_{n+1}g^{(n)}_{n-1}(q)r^{n}\\
& & +\frac{q(d-d_{n+1})}{d}g^{(n)}_0(q)+ \sum_{c=1}^{n-1}\frac{q(d-d_{n+1})}{d}g^{(n)}_c(q)r^{c} \\
& & + u_{n+1}W^{(n)}_0r^{n}+ \sum_{j=1}^{n-1} u_{n+1}\frac{q^j}{d^j}W^{(n)}_jr^{n-j}+ u_{n+1}\frac{q^n}{d^n}W^{(n)}_n\\
& & + u_{n+1}g^{(n)}_0(q)+\sum_{c=1}^{n-1}u_{n+1}g^{(n)}_c(q)r^{c}\\
&=&[\frac{q}{d}(d-d_{n+1})g^{(n)}_0(q)+u_{n+1}\frac{q^n}{d^n}W^{(n)}_n+u_{n+1}g^{(n)}_0(q)] \\
& & +[\sum_{c=0}^{n-2}d_{n+1}g^{(n)}_c(q)r^{c+1} + \sum_{c=1}^{n-1}\frac{q(d-d_{n+1})}{d}g^{(n)}_c(q)r^{c} \\
& & + \sum_{j=1}^{n-1} u_{n+1}\frac{q^j}{d^j}W^{(n)}_jr^{n-j}+\sum_{c=1}^{n-1}u_{n+1}g^{(n)}_c(q)r^{c}] \\
& & + [d_{n+1}g^{(n)}_{n-1}(q)r^n+u_{n+1}W^{(n)}_0r^n].
\end{eqnarray*}
As a result, we can write
\begin{eqnarray*}
A_2+ B_2 + C &=&[\frac{q}{d}(d-d_{n+1})g^{(n)}_0(q)+u_{n+1}\frac{q^n}{d^n}W^{(n)}_n+u_{n+1}g^{(n)}_0(q)] \\
& & +[\sum_{i=1}^{n-1}d_{n+1}g^{(n)}_{i-1}(q)r^{i} + \sum_{i=1}^{n-1}\frac{q(d-d_{n+1})}{d}g^{(n)}_i(q)r^{i} \\
& & + \sum_{i=1}^{n-1} u_{n+1}\frac{q^{n-i}}{d^{n-i}}W^{(n)}_{n-i}r^{i}+\sum_{i=1}^{n-1}u_{n+1}g^{(n)}_i(q)r^{i}] \\
& & + [d_{n+1}g^{(n)}_{n-1}(q)r^n+u_{n+1}W^{(n)}_0r^n]\\
&=&[\frac{q}{d}(d-d_{n+1})g^{(n)}_0(q)+u_{n+1}\frac{q^n}{d^n}W^{(n)}_n+u_{n+1}g^{(n)}_0(q)] \\
& & +\sum_{i=1}^{n-1}[d_{n+1}g^{(n)}_{i-1}(q) + \frac{q(d-d_{n+1})}{d}g^{(n)}_i(q) \\
& & + u_{n+1}\frac{q^{n-i}}{d^{n-i}}W^{(n)}_{n-i} + u_{n+1}g^{(n)}_i(q)]r^{i} \\
& & + [d_{n+1}g^{(n)}_{n-1}(q)+u_{n+1}W^{(n)}_0]r^n\\
&=& \sum_{i=0}^{n}g^{(n+1)}_i(q)r^{i}
\end{eqnarray*}
and hence
\begin{eqnarray*}
\prod_{j=1}^{n+1}(d_jr + \frac{q(d-d_j)}{d}+ u_j) &=& A+B+C = (A_1+B_1)+ (A_2+B_2+C) \\
&=& \sum_{j=0}^{n+1} \frac{q^j}{d^j}W^{(n+1)}_jr^{n+1-j}+ \sum_{c=0}^{n}g^{(n+1)}_c(q)r^{c}.
\end{eqnarray*}
\end{proof}
\begin{theorem}\label{P8.4}
Let $f= x_1^{d_1}\dots x_n^{d_n}$ be a monomial in $S=K[\![x_1,\dots,x_n]\!]$ where $d_j $ is a positive integer for each $1\leq j \leq n $. If $d = \max \{d_1,\dots, d_n \}$ and $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$, then the F-signature of $R^{\bigstar}$ is given by \\
\begin{equation}
\mathbb{S}(R^{\bigstar})=\frac{2}{d^{n+1}}\left[\frac{d_1d_2\dots d_n}{n+1} + \frac{W^{(n)}_1}{n}+\dots+\frac{W^{(n)}_s}{n-s+1}+\dots+\frac{W^{(n)}_{n-1}}{2}\right]
\end{equation}
where $W^{(n)}_1,\dots,W^{(n)}_{n-1}$ are defined as in Notation \ref{N1}.
\end{theorem}
\begin{proof}
Let $R=S/fS$ and $R^{\bigstar}=S[\![u,v]\!]/(f+uv)$. Set $[K:K^p]=b$ and recall from Notation \ref{N4.1} that $\Lambda_e$ is the basis of $K$ as $K^q$-vector space where $q=p^e$. We know from Proposition \ref{C4.19} (d) that
\begin{equation}\label{E1}
\sharp ( F_*^e(R^{\bigstar}),R^{\bigstar})= r_e + 2 \sum_{k=1}^{q-1} \sharp ( \Cok_S(A^k),R)
\end{equation}
where $r_e=b^eq^n$ , $A=M_S(f,e)$ and $A^k=M_S(f^k,e)$. Since $f^k$ is a monomial, it follows from Proposition \ref{L4.25} that the matrix $A^k=M_S(f^k,e)$ is equivalent to a diagonal matrix $D$ whose diagonal entries are taken from the set $\{x_1^{u_1}\dots x_n^{u_n} |0\leq u_j\leq d_j \text{ for all } 1\leq j \leq n \} $.
This makes $\Cok_S(A^k)= \Cok_S(D)$ and consequently the number $\sharp ( \Cok_S(A^k),R) $ is exactly the same as the number of the $n$-tuples $(\alpha_1,\dots,\alpha_n)$ with $0 \leq \alpha_j \leq q-1 $ satisfying that
\begin{equation}\label{EE2}
F_*^e(\lambda x_1^{kd_1+\alpha_1}\dots x_n^{kd_n+\alpha_n}) = x_1^{d_1}\dots x_n^{d_n}F_*^e(\lambda x_1^{s_1}\dots x_n^{s_n})
\end{equation}
where $ s_1,..,s_n \in \{0,\dots,q-1 \}$ for all $\lambda \in \Lambda_e$. However, an $n$-tuple $(\alpha_1,\dots,\alpha_n)$ with $ 0\leq \alpha_j \leq q-1 $ will satisfy (\ref{EE2}) if and only if $\alpha_j= d_j(q-k) + s_j$ for some $ 0\leq s_j \leq q-1 $ for each $ 1 \leq j \leq n $. As a result, there exists $n$-tuples $(\alpha_1,\dots,\alpha_n)\in \mathbb{Z}^n$ with $ 0\leq \alpha_j \leq q-1 $ satisfying (\ref{EE2}) if and only if $d_j(q-k) < q$ for all $ 1 \leq j \leq n $.
Set $ N_j(k):= \{ \alpha_j \in \mathbb{Z} \, | \, d_j(q-k)\leq \alpha_j < q \}$ for all $ 1 \leq j \leq n $. Therefore,
\begin{equation}\label{E18}
\sharp ( \Cok_S(A^k),R)= b^e|N_1(k)||N_2(k)|\dots|N_n(k)|
\end{equation}
where
\begin{equation*}
|N_j(k)|=\left\{
\begin{array}{ll}
q-d_jq+d_jk, & \hbox{ if } d_j(q-k) < q \\
0, & \hbox{ otherwise}
\end{array}
\right.
\end{equation*}
Let $d = \max \{d_1,\dots, d_n \}$. Notice that
\begin{eqnarray*}
\sharp ( \Cok_S(A^k),R)\neq 0 & \Leftrightarrow & |N_j(k)|\neq 0 \text{ for all } j \in \{ 1,\dots,n \} \\
& \Leftrightarrow & d_j(q-k) < q \text{ for all }j \in \{ 1,\dots,n \}\\
& \Leftrightarrow & d(q-k) < q \\
& \Leftrightarrow & \frac{q(d-1)}{d} < k.\\
\end{eqnarray*}
Therefore,
\begin{equation}\label{EE18}
\sharp ( \Cok_S(A^k),R) = b^e\prod_{j=1}^{n}(q-d_jq+d_jk) \text{ whenever } k> \frac{q(d-1)}{d}.
\end{equation}
Let $q=du+t$ where $t\in \{0,..,d-1 \}$. If $t \neq 0$, then one can verify that
\begin{equation}\label{EE19}
\frac{q(d-1)}{d} < q-\frac{q-t}{d} < \frac{q(d-1)}{d}+1.
\end{equation}
Therefore, \\ $\sharp ( \Cok_S(A^k),R)\neq 0$ if and only if $ k \in \{q-\frac{q-t}{d}+r \, | \, r \in \{0,\dots,\frac{q-t}{d}-1\} \}.$
However, if $t=0$, it follows that $\frac{q(d-1)}{d}= q- \frac{q}{d}\in \mathbb{Z}$ and consequently
\begin{equation}\label{EEE19}
\sharp ( \Cok_S(A^k),R)\neq 0 \Leftrightarrow k \in \{q-\frac{q}{d}+r \, | \, r \in \{1,\dots,\frac{q}{d}-1\} \}.
\end{equation}
Assume now that $t\neq 0$. This implies that
\begin{eqnarray*}
\sum_{k=1}^{q-1} \sharp ( \Cok_S(A^k),R) &=& b^e \sum_{k=q-\frac{q-t}{d}}^{q-1}\prod_{j=1}^{n}(q-d_jq+d_jk) \text{ (Use \ref{EE18} and \ref{EEE19} })\\
&=& b^e \sum_{r=0}^{\frac{q-t}{d}-1}\prod_{j=1}^{n}(q-d_jq+d_j(r+q-\frac{q-t}{d})) \\
&=& b^e\sum_{r=0}^{\frac{q-t}{d}-1}\prod_{j=1}^{n}(d_jr + \frac{q(d-d_j)}{d}+ \frac{d_jt}{d}).\\
\end{eqnarray*}
Recall from Lemma \ref{L5.4} that
\begin{equation}\label{E61}
\prod_{j=1}^{n}(d_jr + \frac{q(d-d_j)}{d}+ \frac{d_jt}{d})= \sum_{j=0}^{n} \frac{q^j}{d^j}W^{(n)}_jr^{n-j} +\sum_{c=0}^{n-1}g^{(n)}_c(q)r^{c}
\end{equation}
where $g^{(n)}_c(q)$ is a polynomial in $q$ of degree $n-1-c$ for all $0 \leq c \leq n-1$.
Set $ \delta =\frac{q-t}{d}-1$. By Faulhaber's formula \cite{CG}, if $s$ is a positive integer, we get the following polynomial in $\delta$ of degree $s+1$
\begin{equation}\label{E4}
\sum_{r=1}^{\delta} r^s = \frac{1}{s+1}\sum_{j=0}^s(-1)^j\binom{s+1}{j}B_j\delta^{s+1-j}
\end{equation}
where $B_j$ are Bernoulli numbers, $B_0=1$ and $B_1= \frac{-1}{2}$. This makes
\begin{equation}\label{E62}
\sum_{r=0}^{\delta} r^s=\frac{q^{s+1} }{(s+1)d^{s+1}}+V_s(q)
\end{equation}
where $V_s(q)$ is a polynomial of degree $s$ in $q$.
From Faulhaber's formula and the equations (\ref{E61}), and (\ref{E62}), we get that
\begin{eqnarray*}
\sum_{k=1}^{q-1} \sharp ( \Cok_S(A^k),R) &=& b^e\sum_{r=0}^{\delta}[\sum_{j=0}^{n} \frac{q^j}{d^j}W^{(n)}_jr^{n-j} +\sum_{c=0}^{n-1}g^{(n)}_c(q)r^{c}] \\
&=& b^e[\sum_{j=0}^{n} \frac{q^j}{d^j}W^{(n)}_j\sum_{r=0}^{\delta}r^{n-j} +\sum_{c=0}^{n-1}g^{(n)}_c(q)\sum_{r=0}^{\delta}r^{c}]\\
&=&b^e[\sum_{j=0}^{n} \frac{q^j}{d^j}W^{(n)}_j(\frac{q^{n-j+1} }{(n-j+1)d^{n-j+1}}+V_{n-j}(q)) \\
& & + \sum_{c=0}^{n-1}g^{(n)}_c(q)(\frac{q^{c+1} }{(c+1)d^{c+1}}+V_c(q))] \\
&=& \frac{b^eq^{n+1}}{d^{n+1}}\sum_{j=0}^{n}\frac{W^{(n)}_j}{n-j+1}+ b^e[\sum_{j=0}^{n}\frac{q^j}{d^j}W^{(n)}_jV_{n-j}(q) \\
& & +\sum_{c=0}^{n-1}g^{(n)}_c(q)\frac{q^{c+1} }{(c+1)d^{c+1}}+\sum_{c=0}^{n-1}g^{(n)}_c(q)V_c(q)].
\end{eqnarray*}
Since $ \sum_{j=0}^{n}\frac{q^j}{d^j}W^{(n)}_jV_{n-j}(q)$ and $\sum_{c=0}^{n-1}g^{(n)}_c(q)\frac{q^{c+1} }{(c+1)d^{c+1}}+\sum_{c=0}^{n-1}g^{(n)}_c(q)V_c(q)$ are polynomials in $q=p^e$ of degree $n$ and $n-1$ respectively, it follows that
$$ \lim_{e\rightarrow \infty }\frac{1}{b^ep^{e(n+1)}}b^e[\sum_{j=0}^{n}V_{n-j}(q)+\sum_{c=0}^{n-1}g^{(n)}_c(q)\frac{q^{c+1} }{(c+1)d^{c+1}}+\sum_{c=0}^{n-1}g^{(n)}_c(q)V_c(q)]=0.$$
Therefore
\begin{eqnarray*}
\lim_{e\rightarrow \infty }\frac{1}{b^ep^{e(n+1)}}\sum_{k=1}^{q-1} \sharp ( \Cok_S(A^k),R) &=& \frac{1}{d^{n+1}}\sum_{j=0}^{n}\frac{W^{(n)}_j}{n-j+1}.
\end{eqnarray*}
By the equation (\ref{E1}) and the above equation we conclude that the F-signature of the ring $R^{\bigstar}$ is given by
\begin{equation}\label{E64}
\mathbb{S}(R^{\bigstar})=\frac{2}{d^{n+1}}\Big[\frac{d_1d_2\dots d_n}{n+1} + \frac{W^{(n)}_1}{n}+\dots+\frac{W^{(n)}_s}{n-s+1}+\dots+\frac{W^{(n)}_{n-1}}{2} \Big].
\end{equation}
Second if $q=du $, then $\frac{q(d-1)}{d}= q- \frac{q}{d}\in \mathbb{Z}$ and consequently
\begin{equation*}
\sharp ( \Cok_S(A^k),R)\neq 0 \Leftrightarrow k \in \{q-\frac{q}{d}+r \, | \, r \in \{1,\dots,\frac{q}{d}-1\} \}.
\end{equation*}
Therefore
\begin{eqnarray*}
\sum_{k=1}^{q-1} \sharp ( \Cok_S(A^k),R) &=& b^e\sum_{k=q-\frac{q}{d}+1}^{q-1}\prod_{j=1}^{n}(q-d_jq+d_jk)\\
&=& b^e \sum_{r=0}^{\frac{q}{d}-2}\prod_{j=1}^{n}(q-d_jq+d_j(r+q-\frac{q}{d}+1)) \\
&=& b^e \sum_{r=0}^{\frac{q}{d}-1}\prod_{j=1}^{n}(d_jr + \frac{q(d-d_j)}{d}+d_j).\\
\end{eqnarray*}
By an argument similar to the above argument, we conclude the same result that
\begin{equation}\label{E64}
\mathbb{S}(R^{\bigstar})=\frac{2}{d^{n+1}}\Big[\frac{d_1d_2\dots d_n}{n+1} + \frac{W^{(n)}_1}{n}+\dots+\frac{W^{(n)}_s}{n-s+1}+\dots+\frac{W^{(n)}_{n-1}}{2}\Big].
\end{equation}
\end{proof}
\begin{remark}
Let $K$ be a perfect field of positive characteristic $p$ and let $R=\frac{K[\![x_1,x_2,u,v]\!]}{(x_1x_2-uv)}$.
Applying Theorem \ref{P8.4} gives that $\mathbb{S}(R)=\frac{2}{3}$.
Let $r,s\geq 2$ be integers. If $A$ is the Segre product of the polynomial rings $K[x_1,...,x_r]$ and $K[y_1,...,y_s]$, i.e.,
$R$ is the subring of $ K[x_1,...,x_r, y_1,...,y_s]$ generated over $K$ be the monomials $x_iy_j$
for $1 \leq i \leq r$ and $1 \leq j \leq s$, it is well-known that $A$ is isomorphic to the determinantal ring
obtained by killing the size two minors of an $r \times s$ matrix of indeterminates, and that the
dimension of the ring $A$ is $d =r+s-1$. A.Singh in \cite[Example 7]{AS} shows that
$$ \mathbb{S}(A)= \frac{1}{d!}\sum_{i=0}^{s}(-1)^{i}\binom{d+1}{i}(s-i)^d $$
As a result, the $F$-signature of the determinantal ring $\frac{K[x_1,x_2,u,v]}{(x_1x_2-uv)}$ is $\mathbb{S}(\frac{K[x_1,x_2,u,v]}{(x_1x_2-uv)})=\frac{2}{3}$ and consequently by Remark \ref{R6.3} we get also that $\mathbb{S}(\frac{K[\![x_1,x_2,u,v]\!]}{(x_1x_2-uv)})=\frac{2}{3}$.
\end{remark}
\begin{remark}
Let $K$ be a perfect field of positive characteristic $p$ and let $R=\frac{K[\![x,u,v]\!]}{(x^d-uv)}$.
According to Theorem \ref{P8.4}, we get that $\mathbb{S}(R)=\frac{1}{d}$.
However, we can conclude that $\mathbb{S}(R)=\frac{1}{d}$ from the first case of \cite[Example 8]{HL}.
\end{remark}
\section{The $F$-signature of $S/fS$ and $S[\![y]\!]/fS[\![y]\!]$ are the same}
\label{section:The Fsignatures are the same}
We keep the same notation as in Notation \ref{N4.1}.
\begin{proposition}
Let $S=K[\![x_1,...,x_n]\!]$. If $f$ is a nonunit nonzero element of $S$ and $y$ is a new variable on $S$, then
the $F$-signature of $S/fS$ and $S[\![y]\!]/fS[\![y]\!]$ are the same and consequently the $F$-signature of $S/fS$ and $S[\![y_1,....,y_m]\!]/fS[\![y_1,...,y_m]\!]$ are the same for every positive integer $m$.
\end{proposition}
\begin{proof}
Let $R=S/fS$ and $B=S[\![y]\!]/fS[\![y]\!]$. For any $e$ and $q=p^e$, recall that if $A=M_S(f,e)$, then $F_*^e(S/fS)= \Cok_S(A)$ (Theorem \ref{P5.11}) and $M_{S[[y]]}(f,e)$ is a $q \times q$ matrix over the ring $M_{r_e}(S[\![y]\!])$ (Proposition \ref{EP}) that is given by
\begin{equation}\label{E10}
M_{S[[y]]}(f,e)=\left[
\begin{array}{ccc}
A & & \\
&\ddots & \\
& & A \\
\end{array}
\right].
\end{equation}
Recall from Proposition \ref{P5.5} that $(A,A^{q-1})$ is a matrix factorization of $f$. If $(A,A^{q-1})$ is a nontrivial matrix factorization of $f$, it follows from Proposition \ref{P.24} that $(A,A^{q-1})$ can be represented uniquely up to equivalence as
$$ (A,A^{q-1})\sim (\phi,\psi)\oplus (f,1)^u \oplus (1,f)^v$$
where $(\phi, \psi)$ is a reduced matrix factorization of $f$ over $S$ (and hence over $S[\![y]\!]$), $u = \sharp(\Cok_S(A),R)$ and $v = \sharp(\Cok_S(A^{q-1}),R)$ (see Corollary \ref{C3.8}).
In another words, $A$ is equivalent to the matrix $\left[
\begin{array}{cc}
\phi & \\
& fI_u \\
\end{array}
\right]$ where $I_u$ is the $u \times u$ identity matrix. Since $(\phi, \psi)$ is also a reduced matrix factorization of $f$ in $S[\![y]\!]$, it follows from Proposition \ref{A} that $\Cok_{S[\![y]\!]}(\phi)$ is stable $B$-module and $\Cok_{S[\![y]\!]}\left[\begin{array}{cc}
\phi & \\
& fI_u \\
\end{array}
\right] = B^u \oplus \Cok_{S[\![y]\!]}(\phi)$. Using this result and the relation \ref{E10} we conclude that $ F_*^e(B)=B^{qu}\oplus [\Cok_{S[\![y]\!]}(\phi)]^{\oplus q}$ where $[\Cok_{S[\![y]\!]}(\phi)]^{\oplus q}$ is stable $B$-module. This shows that
\begin{equation}
\sharp(F_*^e(B),B)=q \sharp(F_*^e(R),R).
\end{equation}
However, if $ (A,A^{q-1})\sim (f,1)^{r_e} $ (or $ (A,A^{q-1})\sim (1,f)^{r_e} $) then we obtain that $(F_*^e( \frac{S}{f^{q-1}S})=\Cok_SA^{q-1}= \{0\}$ (or $(F_*^e( \frac{S}{fS})=\Cok_SA= \{0\}$) which is impossible. As a result, if $(A,A^{q-1})$ is a trivial matrix factorization of f, then the only possible
case is that $ (A,A^{q-1})\sim (f,1)^u \oplus (1,f)^v$ where $0 < u,v < r_e$ with $u+v = r_e$ and consequently $\sharp(F_*^e(R),R)= \sharp(\Cok_S(A),R)= u$ . In this case, $A$ is equivalent to the matrix $\left[
\begin{array}{cc}
I_v & \\
& fI_u \\
\end{array}
\right]$ where $I_u$ (respectively $I_v$) is the $u \times u$ (respectively $v \times v$) identity matrix of the ring $M_{r_e}(S)$. It follows from the relation \ref{E10} that
\begin{equation}\label{E11}
\sharp(F_*^e(B),B)=q \sharp(F_*^e(R),R).
\end{equation}
Notice that $\alpha (B) =\log_p[K:K^p]=\alpha (R)$. Therefore,
\begin{equation*}
\mathbb{S}(B)=\lim_{q\rightarrow \infty} \frac{\sharp(F_*^e(B),B)}{p^{e(n+\alpha (B))}}=\lim_{q\rightarrow \infty}\frac{p^e \sharp(F_*^e(R),R)}{p^{e(n+\alpha (B))}}=\lim_{q\rightarrow \infty}\frac{ \sharp(F_*^e(R),R)}{p^{e(n-1+\alpha (R))}}=\mathbb{S}(R).
\end{equation*}
\end{proof}
\section{The F-signature of $\frac{S[\![z]\!]}{(f+z^2)}$ when $f$ is a monomial}
\label{section:The Fsignature of z when is a monomial}
We will keep the same notation as in Notation \ref{N4.1} unless otherwise stated. \\
\begin{theorem} \label{Thm6.9}
Let $f= x_1^{d_1}\dots x_n^{d_n}$ be a monomial in $S=K[\![x_1,\dots,x_n]\!]$ where $d_j $ is a positive integer for each $1 \leq j \leq n$ and $K$ is a field of prime characteristic $p>2$ with $[K:K^p] < \infty $. Let $R=S/fS$ and $R^{\sharp}=S[\![z]\!]/(f+z^2)$. It follows that: \\
1) If $ d_j=1 $ for each $1 \leq j \leq n$, then $ \mathbb{S}(S[\![z]\!]/(f+z^2))= \frac{1}{2^{n-1}}$. \\
2) If $d = \max \{d_1,\dots, d_n \}\geq 2$, then $ \mathbb{S}(S[\![z]\!]/(f+z^2))= 0$.
\end{theorem}
\begin{proof}
Set $[K:K^p]=b$ and recall from Notation \ref{N4.1} that $\Lambda_e$ is the basis of $K$ as $K^q$-vector space. We know by Proposition \ref{L4.13} that
\begin{equation*}
F_*^e(R^{\sharp})= \Cok_{S[\![z]\!]}\left[
\begin{array}{cc}
A^{\frac{q-1}{2}} & -zI \\
zI & A^{\frac{q+1}{2}} \\
\end{array}
\right] \text{ where } A=M_S(f,e).
\end{equation*}
Recall from Proposition \ref{C4.19} (b) that
\begin{equation}\label{E12}
\sharp(F_*^e(R^{\sharp}),R^{\sharp})= \sharp(\Cok_S(A^{\frac{q-1}{2}}),R) + \sharp(\Cok_S(A^{\frac{q+1}{2}}),R).
\end{equation}
Now, let $k\in \{\frac{q-1}{2}, \frac{q+1}{2}\}$ and
set $ N_j(k):= \{ \alpha_j \in \mathbb{Z} \, | \, d_j(q-k)\leq \alpha_j < q \}$ for all $1\leq j \leq n $. Using the same argument that was previously used in the proof of Proposition \ref{P8.4}, it follows that
\begin{equation}\label{Equ17}
\sharp ( \Cok_S(A^k),R)= b^e|N_1(k)||N_2(k)|\dots|N_n(k)|
\end{equation}
and
\begin{equation}\label{Equ18}
\sharp ( \Cok_S(A^k),R)= b^e\prod_{j=1}^{n}(q-d_jq+d_jk) \text{ whenever } k> \frac{q(d-1)}{d}.
\end{equation}
Now if $d_1=d_2=\dots=d_n=1$, it follows from equation (\ref{Equ18}) that $ \sharp ( \Cok_S(A^k),R)= b^ek^n$ for $k\in \{\frac{q-1}{2}, \frac{q+1}{2}\}.$ Therefore, Equation (\ref{E12}) implies that
\begin{equation}\label{EE15}
\sharp(F_*^e(R^{\sharp}),R^{\sharp})=b^e[ ( \frac{q-1}{2})^n + ( \frac{q+1}{2})^n]
\end{equation}
and consequently
\begin{equation*}
\mathbb{S}(R^{\sharp})= \lim_{e\rightarrow \infty } \sharp(F_*^e(R^{\sharp}),R^{\sharp})/b^ep^{en} = \frac{1}{2^{n-1}}.
\end{equation*}
Now let $d_i= \max \{d_1,\dots,d_n \}$ for some $1\leq i \leq n $.
First assume that $d_i=2 $. If $k=\frac{q-1}{2}$ , it follows that $d_i(q-k) > q $ and consequently $|N_i(k)|=0 .$ The equation (\ref{Equ17}) implies $\sharp ( \Cok_S(A^k),R)=0$. When $k=\frac{q+1}{2}$ , we get that $d_i(q-k)=q-1$ and consequently $N_i(k)= \{q-1 \} $ which makes $|N_i(k)|=1 $. Notice that when $k=\frac{q+1}{2}$ and $d_j=1$ , it follows that $|N_j(k)|= \frac{q+1}{2} $. As a result, if $k=\frac{q+1}{2}$, we conclude that
\begin{equation*}
\sharp ( \Cok_S(A^k),R)=b^e |N_1(k)||N_2(k)|\dots|N_n(k)| \leq b^e (\frac{q+1}{2})^{n-1}.
\end{equation*}
Therefore,
\begin{equation*}
\sharp(F_*^e(R^{\sharp}),R^{\sharp})= \sharp(\Cok_S(A^{\frac{q-1}{2}}),R) + \sharp(\Cok_S(A^{\frac{q+1}{2}}),R) \leq b^e(\frac{q+1}{2})^{n-1}.
\end{equation*}
As a result,
\begin{equation*}
\mathbb{S}(R^{\sharp})= \lim_{e\rightarrow \infty } \sharp(F_*^e(R^{\sharp}),R^{\sharp})/b^ep^{en} = 0.
\end{equation*}
Second assume that $d_i > 2 $. In this case, for every $k\in \{\frac{q-1}{2}, \frac{q+1}{2}\}$, it follows that $d_i(q-k)> q$ and consequently $|N_i(k)|= 0 $. Therefore
\begin{equation*}
\sharp(F_*^e(R^{\sharp}),R^{\sharp})= \sharp(\Cok_S(A^{\frac{q-1}{2}}),R) + \sharp(\Cok_S(A^{\frac{q+1}{2}}),R) = 0
\end{equation*}
and consequently
\begin{equation*}
\mathbb{S}(R^{\sharp})= \lim_{e\rightarrow \infty } \sharp(F_*^e(R^{\sharp}),R^{\sharp})/b^ep^{en} = 0.
\end{equation*}
\end{proof}
\begin{remark}\label{R6.10}
Let $A=K[u_1,...,u_n]$ be the polynomial ring in the indeterminates $u_1,...,u_n$ on the field $K$ and let $GL(n,K)$ be the group of all invertible $n \times n$ matrices over $K$. Suppose that $G$ is a finite subgroup of $GL(n,K)$ and let $|G|$ denote the order of $G$ which is the number of elements of $G$. An element $g \in GL(n,K)$ is called a pseudo-reflection if the rank of the matrix $g-I_n$ is one where $I_n$ is the $n\times n$ identity matrix over $K$. For every $g=[g_{ij}] \in G$ and $f=f(u_1,...,u_n)\in A$, let $g(f)=f(v_1,...,v_n)$ where $v_i=\sum_{j=1}^{n}g_{ij}u_j$. A polynomial $f \in A$ is invariant under $G$ if $g(f)=f$ for all $g\in G$. Notice that a polynomial $f$ is invariant under $G$ if and only if its homogeneous components are invariant under $G$ \cite[Chapter 7]{IVA}. The set of all invariant polynomials is denoted $A^G$, this means that
\begin{equation*}
A^G=\{f \in S \,|\, g(f)=f \text{ for all } g \in G \}
\end{equation*}
It is well known that $A^G$ is a graded subring of $A$ that is called the invariant subring of $A$ by $G$.
If $K$ is a field of characteristic $p>0$, $A=K[u_1,...,u_n]$ and $G$ is a finite subgroup of $GL(n,K)$ such that $|G|$ is a unit in $K$ and $G$ has no pseudo-reflection, then the $F$-signature of the invariant subring $A^G$ is given by $ \mathbb{S}(A^G)=\frac{1}{|G|}$ \cite[Remark 2.3]{Y2} and \cite[Theorem 4.2]{WY}.
\end{remark}
I would like to thank H.Brenner and T.Bridgland who suggested using invariant theory for providing another proof of the first result of Theorem \ref{Thm6.9}.
\begin{remark}
We can use Remark \ref{R6.10} to prove the first result of Theorem \ref{Thm6.9} as follows.
Let $K$ be a field of characteristic $p>2$, and let $A=K[u_1,...,u_n]$. Suppose that $G$ is the subgroup of $GL(n,K)$ consisting of all diagonal matrices whose diagonal entries are all taken from $\{1,-1\}$ with determinant equal to one. If $g \in G$, then $g$ has an even number of diagonal elements that are $-1$. This makes $G$ have no pseudo-reflection. Furthermore, if $H$ is the subgroup of $GL(n,K)$ consisting of all diagonal matrices whose diagonal entries are all taken from $\{1,-1\}$, then $|H|=2^{n}$ and $G$ is a finite subgroup of $H$ such that $H\setminus G$ is the subset of $H$ consisting of all diagonal matrices whose diagonal entries are all taken from $\{1,-1\}$ with determinant equal to $-1$. We can define a bijection between $G$ and $H\setminus G$ by sending $g\in G$ to $\tilde{g}\in H \setminus G$ where $\tilde{g}$ is obtained from $g$ by changing the sign of the first diagonal element of $g$. This makes $|G|=|H \setminus G|$. Since $|G|+|H \setminus G|= |H| = 2^n$, we get that $|G|=2^{n-1}$. Now, clearly the monomials $u_1^2,...,u_n^2$ and $u_1...u_n$ are invariant under $G$ and hence $K[u_1^2,...,u_n^2,u_1...u_n]\subseteq A^G$. If $f=u_1^{t_1}...u_n^{t_n}$ is a monomial, then $f=(u_1^{2d_1}...u_n^{2d_n})(u_1^{e_1}...u_n^{e_n})$ where $t_j=2d_j+e_j$ and $e_j \in \{0,1\}$ for all $j=1,...,n$. Therefore, $f$ is invariant under $G$ if and only if $e_j=1$ for all $j=1,...,n$ or $e_j=0$ for all $j=1,...,n$. This shows that a monomial $f=u_1^{t_1}...u_n^{t_n}$ is invariant under $G$ if and only if $f \in K[u_1^2,...,u_n^2,u_1...u_n]$. Now, if $F\in A$ is a homogenous polynomial of degree $d$, then $F=f_1+...+f_r$ where $f_j$ is a monomial of degree $d$ for every $j=1,...,r$. Since $g(F)= b_1f_1+...+b_rf_r$ where $b_j\in \{-1,1\}$ for all $j \in \{1,...,r\}$ and $g\in G$, we get that $F$ is invariant under $G$ if and only if $f_j$ is invariant under $G$ for all $j=1,...,r$. Therefore, $F$ is invariant under $G$ if and only if $F \in K[u_1^2,...,u_n^2,u_1...u_n]$. This shows that $K[u_1^2,...,u_n^2,u_1...u_n]=A^G$ and consequently Remark \ref{R6.10} gives that $$ \mathbb{S}(K[u_1^2,...,u_n^2,u_1...u_n])=\frac{1}{2^{n-1}}.$$ Now, if we define a ring homomorphism $\phi:K[x_1,...,x_n,z]\rightarrow K[u_1,...,u_n]$ by $\phi(x_j)=u_j^2$ for all $j=1,...,n$ and $\phi(z)=u_1...u_n$, we get that
\begin{equation*}
\frac{K[x_1,...,x_n,z]}{(x_1...x_n-z^2)} \cong K[u_1^2,...,u_n^2,u_1...u_n] \text{ as rings. }
\end{equation*}
Therefore,
\begin{equation*}
\mathbb{S}(\frac{K[x_1,...,x_n,z]}{(x_1...x_n-z^2)})=\frac{1}{2^{n-1}}.
\end{equation*}
Now, we can use Remark \ref{R6.3} to conclude that
\begin{equation*}
\mathbb{S}(\frac{K[\![x_1,...,x_n,z]\!]}{(x_1...x_n-z^2)})=\frac{1}{2^{n-1}}.
\end{equation*}
\end{remark}
\begin{remark}
Notice that when $d = \max \{d_1,\dots, d_n \} > 2$, we can prove that $\mathbb{S}(R^{\sharp})=0$ using Fedder's Criteria (Proposition \ref{P2.13}). Indeed, let $\mathfrak{m}$ be the maximal ideal of $S[\![z]\!]$ and let $R^{\sharp}=S[\![z]\!]/(f+z^2)$. If $d = \max \{d_1,\dots, d_n \} > 2$, then $(f+z^2)^{q-1}\in \mathfrak{m}^{[q]}$ which makes, by Fedder's Criteria, $\sharp(F_*^e(R^{\sharp}),R^{\sharp})=0$ for all $e \in \mathbb{Z}^{+}$. This means clearly that
\begin{equation*}
\mathbb{S}(R^{\sharp})= \lim_{e\rightarrow \infty } \sharp(F_*^e(R^{\sharp}),R^{\sharp})/b^ep^{en} = 0.
\end{equation*}
\end{remark}
\section{The F-signature of the ring $S[\![y]\!]/(y^{p^d} +f)$ }
\label{section:The Fsignature of pd}
We will keep the same notation as in Notation \ref{N4.1} unless otherwise stated. \\
\begin{proposition}
Let $\mathfrak{m}$ be the maximal ideal of the ring $S=K[\![x_1,...,x_n]\!]$ where $K$ is a field of positive prime characteristic $p $ with $[K:K^p] = b<\infty $ and let $f$ be a nonzero element in $\mathfrak{m}$ .
If $d \in \mathbb{Z}^{+}$ and $\Re=S[\![y]\!]/(y^{p^d}+f)$ ,
then $\mathbb{S}(\Re)=\frac{\sharp(F_*^d(\Re),\Re)}{b^dp^{nd}}$ .
\end{proposition}
\begin{proof}
Let $[K:K^p] = b $ and let $e \in \mathbb{Z}$ such that $e > d$ . If $A=M_S(f,e)$, equations (\ref{Eq5.14}) and (\ref{Eq5.15}) in the proof of Theorem \ref{P5.1} show that
\begin{equation}\label{E16}
F_*^e(\Re)= \Cok_{S[\![y]\!]}(M_{S[\![y]\!]}(y^{p^d}+f,e))=(\Cok_{S[\![y]\!]}(yI+A^{p^{e-d}}))^{\oplus p^d}
\end{equation}
and
\begin{equation}\label{E17}
[ \Cok_{S[\![y]\!]}(yI+A^{p^{e-d}})]^{\oplus p^e}= [F_*^d( \Re)]^ {\oplus ^{b^{e-d}p^{(n+1)(e-d)}}}.
\end{equation}
If $W=\Cok_{S[\![y]\!]}(yI+A^{p^{e-d}})$, by equations \ref{E16}, \ref{E17} we get that
\begin{equation*}
\sharp(F_*^e(\Re),\Re)=p^d \sharp(W,\Re) \text{ and } p^e \sharp(W,\Re)=b^{e-d}p^{(n+1)(e-d)}\sharp(F_*^d(\Re),\Re).
\end{equation*}
This makes
\begin{equation*}
\sharp(F_*^e(\Re),\Re)=b^{e-d}p^{n(e-d)}\sharp(F_*^d(\Re),\Re)
\end{equation*}
and consequently
\begin{equation*}
\mathbb{S}(\Re)=\lim_{e\rightarrow \infty } \frac{\sharp(F_*^e(\Re),\Re)}{b^ep^{en}} =\lim_{e\rightarrow \infty } \frac{b^{e-d}p^{n(e-d)}\sharp(F_*^d(\Re),\Re)}{b^ep^{en}}=\frac{\sharp(F_*^d(\Re),\Re)}{b^dp^{nd}}.
\end{equation*}
\end{proof}
|
\section{Introduction}
\vspace{1mm}
\noindent
Coupled systems of linear differential equations arise frequently in intermediate calculations
when one tries to tackle massless or massive three-loop Feynman integrals. Namely, using
integration
by parts (IBP) methods\footnote{In our calculations we used heavily the {\tt C++} program
{\tt Reduze~2}~\cite{Reduze2} based on Laporta's algorithm \cite{Laporta:2001dd}.}~\cite{IBP} one
can determine such a coupled system and together with the initial values one obtains a complete
description of the involved Feynman integrals. Various techniques~\cite{DEQ} have been elaborated
to extract relevant information from these systems that sheds light on the involved Feynman
integrals. For heuristic methods to find closed form solutions of such systems, in case the
solutions are given by iterative integrals, see for
instance~\cite{Kotikov:2010gf}.
A completely algorithmic approach has been worked out
recently in~\cite{NewUncouplingMethod,VLadders,UncoupleTHM}
if the unknown functions $\hat{I}_i(x)$ with $1\leq i\leq r$ of the coupled system admit
power series representations\footnote{In addition there may be trailing terms such as $\ln^k(x)$
for each of these power series and one may shift $N$ by a finite number of integer values.}
\begin{equation}\label{Equ:HatIAnsatz}
\hat{I}_i(x)=\sum_{N=0}^{\infty}I_i(N)x^N.
\end{equation}
Then under the assumption that sufficiently many initial values $I_i(\nu)$ at nonnegative integers $\nu\in\mathbb{N}$ are available, we can utilize our summation machinery that decides constructively if the $I_i(N)$ can be written in terms of nested sums over hypergeometric products. This means that they can be composed by elements from the rational function field $\mathbb{K}(N)$, the three operations
($+,-,\cdot$), hypergeometric products of the form $\prod_{k=l}^Nh(k)$ with $l\in\mathbb{N}$ and $h(k)$ being a rational function in $k$ and being free of $N$, and sums of the form $\sum_{k=l}^Nh(k)$ with $l\in\mathbb{N}$ and with
$h(k)$ being a nested hypergeometric sum expression w.r.t.\ $k$ and being free of $N$. This class of special functions covers as special cases harmonic sums~\cite{BlumVerm}, generalized harmonic sums~\cite{Moch:2001zr,Ablinger:2013cf}, cyclotomic
harmonic sums~\cite{Ablinger:2011te} or nested binomial
sums~\cite{BinSums,Ablinger:2014bra}; for surveys on these quantities see e.g. \cite{REF1}.
The central step of our procedure~\cite{NewUncouplingMethod,VLadders} is to translate the problem
of solving a coupled system of differential equations to the problem of solving scalar recurrences
that
depend only on one of the unknown functions $I_i(N)$. To obtain a complete algorithm we utilize
uncoupling algorithms~\cite{UNCOUPL,Zuercher:94}, recurrence solving algorithms~\cite{dAlembert,ParticularSol,BKSS:12}, and simplifiers of the solutions by means of symbolic summation~\cite{Summation} based on difference field~\cite{DFTheory} and difference ring~\cite{DRTheory} theory.
This successful interplay of the underlying summation tools (uncoupling, recurrence solving, and
indefinite summation) has been implemented within the package \texttt{SolveCoupledSystem}~\cite{UncoupleTHM} that relies on the summation packages~\texttt{Sigma} and~\texttt{SumProduction}~\cite{Summation} (concerning the summation tools) and the the package~\texttt{OreSys}~\cite{OreSys} (executing Z\"urcher's uncoupling algorithm~\cite{Zuercher:94}). The package \texttt{SolveCoupledSystem} has been applied successfully for various challenging 3-loop calculations~\cite{DiffCalculations}. These heavy calculations were possible by incorporating besides \texttt{Sigma} also the package~\texttt{HarmonicSums}: whenever possible, we employed its algorithms that are tuned for harmonic sums, cyclotomic sums, generalized harmonic sums and nested binomial sums.
Still, in recent calculations two serious problems arose. First, in some instances we obtained
coupled systems of difference equations that we failed to uncouple due to restricted time and
memory resources. Second, we succeeded in uncoupling the system but derived scalar recurrences
whose orders were very large. We emphasize that \texttt{Sigma} was still capable of solving these
large recurrences (in~\cite{GuessingProject} we solved gigantic recurrences up to order 35) but
finding 16 or more initial values to combine the solutions accordingly was the bottleneck:
extracting such a large amount of initial values from the given Feynman integrals was out of
the scope
of the existing algorithms, like exploiting the $\alpha$-parameterization of the
integrals~\cite{Ablinger:2014uka}.\footnote{Yet one might think of other methods to obtain these
initial values. It is, however, clear that the calculation of finite Mellin moments of complicated
diagrams in general constitutes no simple task either, cf.~\cite{MOM}, for very high moments.}
Based on our existing algorithmic machinery (see Subsection~\ref{Sec:PureREC} below) we will present a new tactic in Subsection~\ref{Sec:DEREC} that has been implemented within the package \texttt{SolveCoupledSystem} and that overcomes the problems described above. A conclusion concerning future calculations that will rely on this new machinery will be given in Section~\ref{Sec:Conclusion}.
The present algorithm also decides, whether or not a system can be solved in terms of iterative
integrals over whatsoever alphabet or not, and allows therefore to single out systems which are
not
first order reducible requiring other techniques of solution, cf.~\cite{CIS}.
\section{Two basic strategies}
\vspace{1mm}
\noindent
Suppose we are given a coupled system of differential equations coming from IBP-methods~\cite{IBP,Laporta:2001dd,Reduze2}. More precisely, we are given a matrix $\hat{A}$ of dimension $r\times r$ with entries from the rational function field $\mathbb{K}(x)$ and we are given a vector $\vect{\hat{b}}(x)=(\hat{b}_1(x),\dots,\hat{b}_r(x))$ of length $r$ where each entry is given in terms of a linear combination of master integrals with coefficients from $\mathbb{K}(x)$.
Then, given sufficiently many initial values we seek for the uniquely determined solution
$\vect{\hat{I}}(x)=(\hat{I}_1(x),\dots,\hat{I}_r(x))$
of the system
\begin{equation}\label{Equ:DEProblem}
D_x\vect{\hat{I}}(x)=\hat{A}\,\vect{\hat{I}}(x)+\vect{\hat{b}}(x)
\end{equation}
with $D_x\vect{\hat{I}}(x)=(D_x\hat{I}_1(x),\dots,D_x\hat{I}_r(x))$, where $D_x$ is the
differential operator w.r.t.\ $x$ acting on functions in $x$.
In our setting we may assume that the master integrals arising in the $\hat{b}_i(x)$ have a power series
representation, where the coefficients are already computed. Within our calculations in~\cite{DiffCalculations} this has been accomplished with
\begin{enumerate}
\item the package \texttt{EvaluateMultiSums}~\cite{Summation} relying on the summation package \texttt{Sigma} and the special functions package \texttt{HarmonicSums}~\cite{HarmonicSums} (see also~\cite{BlumVerm,Moch:2001zr,Ablinger:2013cf,Ablinger:2011te,BinSums,Ablinger:2014bra});
\item the package \texttt{MultiIntegrate}~\cite{Integration} being a specially tuned version of the multivariate Almkvist--Zeilberger algorithm;
\item the package \texttt{SolveCoupledSystem} containing the algorithms that will be explained below -- here the package has been applied to systems which do not depend on the functions given in $\vect{I}(x)$.
\end{enumerate}
In particular, we assume that the coefficients are given in terms of nested sums over hypergeometric products.
This enables one to write also the $\hat{b}_i(x)$ in a power series representation
\begin{equation}\label{Equ:HatbAnsatz}
\hat{b}_i(x)=\sum_{N=0}^{\infty}b_i(N)\,x^N
\end{equation}
whose coefficients $b_i(N)$ are given in terms of nested sums over hypergeometric products.
Furthermore, we assume that also the unknown master integrals $\hat{I}_i$ are analytic, i.e., there exist the power series representations~\eqref{Equ:HatIAnsatz}.\\
Under these assumptions we seek for the coefficients $I_1(N),\dots, I_r(N)$ in closed form. This means that we look for representations in terms of nested sums over hypergeometric products whenever such a representation is possible.
In Subsection~\ref{Sec:PureREC} we will recall our first
algorithm~\cite{NewUncouplingMethod,VLadders,UncoupleTHM} that solves this problem and that has been successfully applied to various challenging massive 3-loop Feynman integrals~\cite{DiffCalculations}. In this approach we will address some efficiency problems that have occurred recently. By a variation of the employed building blocks, we will present a new tactic that overcomes these problems in Subsection~\ref{Sec:DEREC}.
\subsection{Tactic 1: Uncoupling a system of difference equations}\label{Sec:PureREC}
\vspace{1mm}
\noindent
In the first approach the system of differential equations~\eqref{Equ:DEProblem} with the unknown functions $\hat{I}_i(x)$, which have power series representations~\eqref{Equ:HatIAnsatz}, is translated to a system of difference equations for the corresponding coefficients $I_i(N)$. Afterwards one applies computer algebra algorithms in order to find a symbolic representation of the $I_i(N)$ in terms of nested sums over hypergeometric products.
\noindent More precisely, we execute the following steps.
\begin{myEnumerate}
\item We utilize that an analytic function $\sum_{N=0}^{\infty}h(N)x^N$
with a certain convergence disc fulfills
\begin{equation}\label{Equ:DRelation}
D_x\sum_{N=0}^{\infty}h(N)x^N=\sum_{N=1}^{\infty}N\,h(N)x^{N-1};
\end{equation}
similarly, if we argue in the formal power series setting, the operator $D_x$ is just defined in this way.
Plugging the already computed series expansions of $\hat{b}_i(x)$ and the Ansatz~\eqref{Equ:HatIAnsatz} into~\eqref{Equ:DEProblem} (note that in general we can start with a coupled system of differential equations of higher-order; see~\cite{UncoupleTHM})
and doing coefficient comparison w.r.t.\ $x^N$ yields a coupled system of linear difference
equations, say of order $m\geq0$, which is of the form
\begin{equation}\label{Equ:CoupledDifferenceSys}
A_m\,\vect{I}(N+m)+A_{m-1}\,\vect{I}(N+m-1)+\dots+A_0\,\vect{I}(N)=\vect{b}(N),
\end{equation}
with $\vect{I}(N+l)=(I_1(N+l),\dots,I_r(N+l))$ for $l\in\mathbb{N}$ and where the matrices $A_m,\dots,A_0$ have dimension $r\times r$ with entries from $\mathbb{K}(n)$. Ideally, one should solve this coupled system directly. For rational solutions this problem is solved in~\cite{SolveCoupled}; for first steps towards the more general setting of $\Pi\Sigma$-fields we refer to~\cite{SolveCoupledPiSigma}. So far, there are no algorithms available to solve such systems directly in terms of nested sums over hypergeometric products. Therefore we will proceed as follows.
\item We uncouple the system such that some of the unknown coefficients $I_i(N)$ of the power series are solutions of a scalar linear recurrence (which does not depend on other unknown functions) and
where the remaining unknown coefficients can be given as a linear combination of the coefficients
that are determined by the solutions of the scalar recurrences. To accomplish this task, we can
apply the following three sub-steps.
\begin{myEnumerate2}
\item First, we transform the system~\eqref{Equ:CoupledDifferenceSys} of order $m$ to a first-order system as follows. For $1\leq i\leq r$, let $n_i\in\mathbb{N}$ be the maximal value such that $I_i(N+n_i)$ occurs in~\eqref{Equ:CoupledDifferenceSys}. Then one can introduce auxiliary-functions $y_{i,j}(N)$ with $1\leq j< n_{i}$ and $1\leq i\leq r$ where $I_i(N+j)$ is rephrased by $y_{i,j}(N)$ in~\eqref{Equ:CoupledDifferenceSys}. Introducing in addition the relations $y_{i,j}(N+1)=y_{i,j+1}(N)$ and replacing $y_i(N+n_i)$ by $y_{n_i-1}(N+1)$ in~\eqref{Equ:CoupledDifferenceSys} one arrives at a first-order system of difference equations. In other words, solving this first-order system with the unknown $y_{i,j}(N)$ yields immediately the solutions for the $I_i(N)$. For simplicity, we write again $I_i(N)$ instead of $y_i(N)$, and suppose that we are given the system
\begin{equation}\label{Equ:FirstOrderRecN}
\vect{I}(N+1)= A\,\vect{I}(N)+\vect{b}(N),
\end{equation}
where $A$ is an $r\times r$ matrix and $\vect{b}(N)=(b_1(N),\dots,b_r(N))$ is a vector where $b_i(N)$ are expression in terms of nested sums over hypergeometric products.
\item If $A$ is an invertible matrix we skip this step. Otherwise, we proceed as follows.
One can perform simple row and column operations in order get again a system
of this form with dimension $r'<r$ such that the $r'\times r'$ matrix is invertible and
such that
we are given linear combinations of the $r'$ unknown integrals which produce the desired integrals
$I_i(N)$ for $1\leq i\leq r$. Hence, solving this reduced system provides the solution of the original system~\eqref{Equ:FirstOrderRecN}. For further considerations we will continue to work with~\eqref{Equ:FirstOrderRecN} where $A$ is invertible.
\item Finally, we use Z\"urcher's algorithm~\cite{Zuercher:94} that is available in the package~\texttt{OreSys}~\cite{OreSys} in order to uncouple this system. In general, one obtains $n$ ($1\leq n\leq r$) scalar recurrences, namely, for $1\leq i\leq n$ we get the recurrences
\begin{equation}\label{Equ:SingleREC}
a_{i,i}(N) I_i(N+m_i)+a_{i,m_i-1}(N) I_i(N+m_i-1)+\dots+a_{i,0} I_i(N)=f_i(N)
\end{equation}
of orders $m_i$ where the $a_{i,j}(N)$ are from $\mathbb{K}(N)$ and the $f_i(N)$ are given as a linear combination of the $b_1(N),\dots,b_r(N)$ over $\mathbb{K}(N)$ with possible shifts in $N$. Since the $b_i(N)$ are given in terms of nested sums over hypergeometric products, also the $f_i(N)$ can be given in terms of nested sums over hypergeometric products.
In addition, we get explicit linear combinations of the $I_1(N),\dots,I_n(N)$ and the $b_1(N),\dots,b_r(N)$ with possible shifts in $N$ that produce the remaining integrals $I_{n+1}(N),\dots,I_{m}(N)$. More generally, a subset $M$ of $\{I_1,\dots,I_m\}$ can be described by scalar recurrences and the complementary set can be represented by a linear combination of the functions from $M$; after reordering, we may suppose that $M=\{I_1,\dots,I_n\}$.
\end{myEnumerate2}
\item Finally, we aim at deciding algorithmically if the coefficients $I_i(N)$ of the integrals~\eqref{Equ:HatIAnsatz} can be represented in terms of nested sums over hypergeometric products. In this regard, we have to emphasize that the expressions $b_i(N)$ of $\vect{b}(N)$ arising in~\eqref{Equ:FirstOrderRecN} depend usually on an extra parameter $\varepsilon$ coming from the analytic continuation of the space-time dimension $D=\varepsilon+4$. Similarly, the matrix $A$ in~\eqref{Equ:FirstOrderRecN} depends on this dimensional parameter $\varepsilon$. Thus we usually are given the rational function field $\mathbb{K}=\mathbb{K}'(\varepsilon)$ where $\mathbb{K}'$ is a subfield of $\mathbb{K}$ containing $\mathbb{Q}$. Further, there is usually no hope to represent the $b_i(N)$ and also the the integrals $I_i(N)$ in terms of nested sums over hypergeometric products where $\varepsilon$ occurs inside of the sums and products. However, in most applications, there is such a representation, if one considers the functions in its $\varepsilon$-expansion~\cite{BKSS:12,epRepresentation}. More precisely, we assume that the $b_i(N)$ are given in the form
\begin{equation}\label{Equ:biExpansion}
b_i(N)=b_{i,o}(N)\varepsilon^o+b_{i,o+1}(N)\varepsilon^{o+1}+b_{i,o+2}(N)\varepsilon^{o+2}+\dots,
\end{equation}
where the first coefficients $b_{i,j}(N)$ are given in terms of nested sums over hypergeometric
products (which are free of $\varepsilon$). Further, one is not interested in an expression of the $I_i(N)$, where the nested sums and products depend on $\varepsilon$, but one is interested in the first $l_i$ coefficients of the $\varepsilon$-expansion
$$I_i(N)=I_{i,o}(N)\varepsilon^{o}+I_{i,o+1}(N)\varepsilon^{o+1}+\dots I_{i,o}(N)\varepsilon^{o+l_i},$$
where $I_{i,j}(N)$ is given in terms of nested sums over hypergeometric products (which are free of $\varepsilon$). In our 3-loop calculations we usually have $o=-3$. More precisely, suppose that we are given the initial values
\begin{equation}\label{Equ:InitialV}
I_i(\nu)=I_{i,o}(\nu)\varepsilon^{o}+I_{i,o+1}(\nu)\varepsilon^{o+1}+\dots I_{i,o+l_i}(\nu)\varepsilon^{o+l_i}+\dots
\end{equation}
for\footnote{Sometimes we need more initial values; the necessary number can be detected during the recurrence solving.} $\nu=1,\dots,m_i$.
Then we can decide algorithmically if the coefficients $I_{i,j}(N)$ can be represented in terms of nested sums over hypergeometric products by executing the following sub-steps.
\begin{myEnumerate2}
\item Note that the $f_i(N)$ for $1\leq i\leq n$ arising in the inhomogeneous parts of the recurrences~\eqref{Equ:SingleREC} are given as a linear combination of the $b_i(N)$. Since they are given in the form~\eqref{Equ:biExpansion} where the $b_{i,j}$ are nested sum expressions over hypergeometric products, one can collect terms in $\varepsilon$ and gets the $\varepsilon$-expansion
\begin{equation}\label{Equ:InhomExpand}
f_i(N)=f_{i,o}(N)\varepsilon^{o}+f_{i,o+1}(N)\varepsilon^{o+1}+\dots f_{i,o}(N)\varepsilon^{o+l_i}+\dots,
\end{equation}
where the $f_{i,j}(N)$ are expressions in terms of nested sums over hypergeometric products (which are free of $\varepsilon$).
\item Given the initial values~\eqref{Equ:InitialV} and the recurrences~\eqref{Equ:SingleREC} with~\eqref{Equ:InhomExpand} where the $f_{i,j}(N)$ are given in terms of nested sums over hypergeometric products, we can activate our recurrence solver~\cite{BKSS:12} based on~\cite{dAlembert,ParticularSol}.
For all $i$ with $1\leq i\leq n$, we can constructively decide if the integral $I_i(N)$ can be expressed in terms of nested product-sum expressions. If such a representation does not exist for some $i$, we stop and learned that our problem has to be formulated outside of the class of nested sums over hypergeometric products.
\begin{remark}
The solutions in terms of nested sums over hypergeometric products are also called d'Alembertian solutions~\cite{dAlembert} which can be computed by using algorithms from~\cite{dAlembert,ParticularSol}. Further we remark that the sum representations produced by these algorithms are rather complicated: they are highly nested and the summands have denominators which do not factor nicely. Here we utilize symbolic summation algorithms ~\cite{Summation} based on difference field theory~\cite{DFTheory} and difference ring theory~\cite{DRTheory} in order to rewrite the found solutions in terms of special functions like cyclotomic generalized harmonic sums or nested binomial sums.
\end{remark}
\item We can combine the computed $\varepsilon$-expansions of the $I_1(N),\dots,I_n(N)$ and their shifted versions yielding the $\varepsilon$-expansions of the remaining integrals $I_{n+1}(N),\dots,I_m(N)$ whose coefficients are given in terms of nested sums over hypergeometric products.
\end{myEnumerate2}
\begin{remark}\label{Remark:EliRelation}
Within intermediate steps but also for the final result it is important to get compact representations of the found nested sums and products. In general, we can exploit the underlying summation theory~\cite{DRTheory} in order to compute expressions in terms of nested sums and products where no algebraic (i.e., polynomial) relations exist among the occurring sums and products.
Restricting to harmonic sums and cyclotomic sums, one can employ an alternative and very efficient
machinery. Namely, it has been shown in~\cite{AIHarmonicSums} that such an optimal representation can be obtained using the underlying quasi-shuffle relations~\cite{Blumlein:2003gb}. Currently, further investigations along the lines of~\cite{AIHarmonicSums} are carried out to avoid the rather involved difference ring algorithms and to reduce the calculations to quasi-shuffle relations and further properties given by the occurring alphabets of the nested sums.
\end{remark}
\item Optionally, we can rewrite the power series solutions in terms of nested integrals involving the extra parameter $x$ using the package \texttt{HarmonicSums}.
\end{myEnumerate}
\noindent The above tactic can be summarized by the following diagram.
$$\xymatrix@!R=1.cm@C1.1cm{
\boxed{\txt{DE system\\
$\sum_i\hat{A}_i\,D^i \hat{I}(x)=\hat{\vect{b}}(x)$}{\ar[rrr]_{\txt{\footnotesize holonomic closure properties (1)}}}}&&&{\boxed{\txt{REC system\\
$\sum_iA_i\,I(N+i)=\vect{b}(N)$}}{\ar[ddd]|{\txt{\footnotesize uncoupling algorithm (2)}}}}\\
&{\boxed{\txt{iterated\\integrals\\ $\hat{I}_k(x)$}}}&\\
&&&\\
{\boxed{\txt{
closed form solutions of\\
$I_1(N),\dots I_r(N)$\\
in the class of nested\\
sums and products\\
-- if this is possible}}}{\ar[ruu]|<>(0.4){\txt{\footnotesize\texttt{HarmonicSums.m}}}}
&&&{\boxed{\txt{uncoupled REC system\\
$\sum_i a_i(N)I_1(N+i)=f(N)$\\
$I_k(N)=\text{expr}_k(I_1(N)), k>1$}}{\ar[lll]_{\txt{\footnotesize recurrence solving (3)}}
}}}$$
\subsection{Tactic 2: Uncoupling a system of differential equations}\label{Sec:DEREC}
\vspace{1mm}
\noindent
In the above approach we transformed the system of differential equations to a system of recurrences in step~(1), uncoupled the system in step~(2) and finally decided if the unknown integrals can be given in terms of nested sums over hypergeometric products in steps~(3) and~(4). We remark that in step~(1) one obtains usually a coupled system of higher-order difference equations. Uncoupling this system directly, or as proposed in our strategy, bringing it first to a first-order system and uncoupling it afterwards, worked very well for many concrete calculations. However, recently we entered examples where the calculation time of the known uncoupling algorithms blew up dramatically. In contrast to that, the original system provided by IBP methods is usually rather small (and in particular first-order), and the uncoupling algorithms applied to this differential equation system behaves rather tame.
Therefore a different and very promising tactic is to reverse steps~(1) and~(2). More precisely, we can proceed as follows.
\begin{myEnumerate}
\item The uncoupling algorithms available in~\texttt{OreSys}~\cite{OreSys} work for any Ore algebra covering besides the difference case also the differential case. Within \texttt{SolveCoupledSystem} we use again Z\"urcher's algorithm. But this time we uncouple immediately the first-order system~\eqref{Equ:DEProblem} of differential equations. In most instances, the matrix $\hat{A}$ is already invertible. Otherwise, we carry out a similar preprocessing step as worked out in Step~(2b) of Subsection~\ref{Sec:PureREC}. Summarizing, one obtains $n$ ($1\leq n\leq r$) scalar differential equations of the form
\begin{equation}\label{Equ:DESingle}
a_{i,r_i}(x) D^{r_i}_x\hat{I}_i(x)+a_{i,r_i-1}(x) D^{r_i-1}_x\hat{I}_i(x)+\dots+a_{i,0}(x) D^{0}_x\hat{I}_i(x)=\sum_{j=1}^r\sum_k d_{i,j,k}(x)D_x^k\hat{b}_j(x)
\end{equation}
where the sums on the right hand side are finite and the $a_{i,j}(x)$ and $d_{i,j,k}(x)$ are from $\mathbb{K}[x]$. In addition, the remaining integrals $\hat{I}_{n+1}(x),\dots,\hat{I}_{m}(x)$ are related to the $\hat{I}_j(x)$ with $1\leq i\leq n$ and the $\hat{b}_i(x)$ with $1\leq i\leq r$ as follows: for all $i$ with $n<i\leq m$ we get
\begin{equation}\label{Equ:RemainingIx}
\hat{I}_i(x)=\sum_{j=1}^n\sum_k\alpha_{i,j,k}(x)D_x^k\hat{I}_{j}(x)+\sum_{j=1}^r\sum_k\beta_{i,j,k}(x)D_x^k\hat{b}_{j}(x)
\end{equation}
where the sums on the right are finite and where the $\alpha_{i,j,k}(x)$ and $\beta_{i,j,k}(x)$ are
given from $\mathbb{K}(x)$.
\item Next, we transform the scalar equations~\eqref{Equ:DESingle} to recurrences of the form~\eqref{Equ:SingleREC} with $a_{i,j}(N)\in\mathbb{K}[N]$ using the same tactic as used in Step~(1) of Subsection~\ref{Sec:PureREC}. Under the assumption that the $b_i(N)$ have an $\varepsilon$-expansion~\eqref{Equ:biExpansion} where the first coefficients $b_{i,j}(N)$ can be given in terms nested sums over hypergeometric products, we can derive $\varepsilon$-expansions~\eqref{Equ:InhomExpand} of the inhomogeneous parts of the recurrences~\eqref{Equ:SingleREC}.
\item Afterwards we apply Step~(3) of Subsection~\ref{Sec:PureREC}. Namely, together with the initial values~\eqref{Equ:InitialV} with $\nu=1,\dots,m_i$
we can decide algorithmically if the coefficients $I_{i,j}(N)$ of~\eqref{Equ:InitialV} for $1\leq i\leq n$ can be represented in terms of nested sums over hypergeometric products.
\item Finally, we compute the coefficients $I_i(N)$ of the integrals~\eqref{Equ:HatIAnsatz} for $n< i\leq m$ as follows.
We replace the $\hat{I}_i(x)$ and $\hat{b}_i(x)$ on the right hand side of~\eqref{Equ:RemainingIx} by their power series representations~\eqref{Equ:HatIAnsatz} and~\eqref{Equ:HatbAnsatz}. In addition, we replace the $I_i(N)$ and $b_i(N)$ arising in the power series representations further by the
given representation in terms of nested sums over hypergeometric products (possibly in its $\varepsilon$-expansion).
Then we utilize the package \texttt{SumProduction}~\cite{Summation} that contains rather efficient algorithms~\cite{VLadders} to calculate the $N$th coefficient of the corresponding power series representation.
First, we separate the expressions accordingly and get the representation
\begin{equation}\label{Equ:IExpandX}
\hat{I}_i(x)=\varepsilon^{\omega}\hat{I}_{i,\omega}(x)+\varepsilon^{\omega+1}\hat{I}_{i,\omega+1}(x)+\dots+\varepsilon^{\omega+l_i}\hat{I}_{i,o+l_i}(x)
\end{equation}
for some $\omega\in\mathbb{Z}$
where each $\hat{I}_{i,j}(x)$ can be written as a finite sum consisting of summands of the form
\begin{equation}\label{ExtractHCoeff}
\hat{H}(x)=\hat{q}(x)\sum_{N=0}^{\infty}x^N h_1(N)\dots h_l(N).
\end{equation}
Here we have $\hat{q}(x)\in\mathbb{K}'(x)$ and for all $1\leq i\leq l$ we have that $h_i(N)\in\mathbb{K}'(x)$, $h_i(N)$ is a hypergeometric product or $h_i(N)$ is a nested sum over hypergeometric products.
With a brute force approach one can now compute the $N$th coefficient for each such expression as given in~\eqref{ExtractHCoeff}, i.e., we can compute an expression $H(N)$ in terms of nested sums over hypergeometric products such that
$$\hat{H}(x)=\sum_{N=0}^{\infty}H(N)x^N$$
holds. This finally yields the $N$th coefficient $I_{i,j}(N)$ of
$$\hat{I}_{i,j}(x)=\sum_{N=0}^{\infty}I_{i,j}(N)x^N$$
in terms of nested sums over hypergeometric products.\\
Namely, first we expand
$q(x)=\frac{a(x)}{b(x)}$ with $a(x),b(x)\in\mathbb{K}'[x]$ in a series expansion
\begin{equation}\label{Equ:eExp}
\hat{q}(x)=\sum_{N=\mu}^{\infty}q(N)x^i
\end{equation}
with $\mu\in\mathbb{Z}$ as follows~\cite{VLadders}. Consider the complete factorization
\begin{equation}\label{Equ:RootsOfB}
b(x)=c\,x^{\nu_0}(x-\rho_1)^{\nu_1}(x-\rho_2)^{\nu_2}\dots(x-\rho_v)^{\nu_v}
\end{equation}
with $c\in\mathbb{K}'\{0\}$ and $\rho_i\in\bar{\mathbb{K}}\setminus\{0\}$ ($\bar{\mathbb{K}}$ is the algebraic closure of $\mathbb{K}'$) where $\nu_i\in\mathbb{N}$ counts the multiplicity of the roots
$\rho_i$ ($\rho_0=0$). Then, as worked out in~\cite[Thm. 4
1.1]{Stanley:97}, we can calculate the
expansion
$$\frac{1}{b(x)}=\frac{1}{x^{\nu_0}}\sum_{N=0}^{\infty}\beta(N)x^N$$
with
$$\beta(N)=p_1(N)\,\rho_1^N+p_2(N)\,\rho_2^N+\dots+p_v(N)\,\rho_v^N,$$
where the $p_i(N)$ are polynomials in $N$ with degree at most $\nu_i-1$. Now we perform the Cauchy product on $\hat{q}(x)=a(x)\,\frac{1}{x^{\nu_0}}\sum_{N=0}^{\infty}\beta(N)x^N$, and it follows that the coefficient $q(N)$ of the expansion~\eqref{Equ:eExp} can be written again as a linear combination of the $\rho_i^N$ with polynomial coefficients in $N$.
Finally, we obtain
$$H(N)=\sum_{k=\mu}^{N+\nu_0}h_1(k)h_2(k)\dots h_l(k)r(k)q(N-k)$$
by applying once more the Cauchy product.
Since $q(N-k)$ is given as a linear combination of the $\rho_i^{N-k}=\rho_i^{N}\rho_i^{-k}$ where the coefficients are polynomials in $N$ and $k$,
we can pull out all expressions that depend on $N$. Summarizing, we can write $H(N)$ as an expression in terms of nested sums over hypergeometric products. In particular, the summands of the arising sums are built by the objects
$h_1(k)h_2(k)\dots h_l(k)$ given in~(\ref{Equ:IExpandX},\ref{ExtractHCoeff}) and the roots
$\rho_i^k$ from~\eqref{Equ:RootsOfB}.
\end{myEnumerate}
In short, we can summarize the second approach with the following diagram.
$$\xymatrix@!R=1.4cm@C2.cm{
\boxed{\txt{DE system\\
$D_x\hat{I}(x)=\hat{A}\,\hat{I}(x)+\hat{b}(x)$}}{\ar[rr]_{\txt{\footnotesize uncoupling algorithm (1)}}}&&
{\boxed{\txt{uncoupled DE system\\
$\sum_i a_i(x)D_x^i \hat{I}_1(x)=\hat{f}(x)$\\
${\bf}\hat{I}_k(x)=\text{expr}_k(\hat{I}_1(x)), k>1$}}{\ar[dd]|{\txt{\footnotesize holonomic closure prop.\ (2)}}}
{\ar[ddll]|{\txt{\footnotesize extract coefficients (4)}}}}\\
\\
{\boxed{\txt{
closed form solutions of\hspace*{1.4cm}\\
$\overbrace{\text{\small$I_1(N),\dots,I_n(N)$}}^{\text{step (3)}},\overbrace{\text{\small$I_{n+1}(N),\dots, I_m(N)$}}^{\text{step (4)}}$\\
in the class of nested sums\hspace*{1cm}\\
over hypergeometric products\hspace*{0.45cm}\\
-- if this is possible\hspace*{2.2cm}}}}
&&{\boxed{\txt{scalar recurrence\\
$\sum_i a'_i(N)I_1(N)=f(N)$}}{\ar[ll]_{\txt{\footnotesize recurrence solver (3)}}}}\\
}$$
\subsubsection{Improvement 1: compute the $N$th coefficients efficiently (step (4))}\label{Sec:Improvement1}
\vspace{1mm}
\noindent
Tactic~2 has the advantage that one uncouples the system of differential equations that one is given, e.g., by IBP methods. They are usually in a rather nice shape and can be uncoupled in many examples extremely efficiently. However, this big advantage is paid by the challenge to compute the $N$th coefficient of the $\hat{I}_{i,j}(x)$ in~\eqref{Equ:IExpandX}: IBP methods usually calculate expressions with rather complicated denominators, i.e., one obtains subexpressions $\hat{H}(x)$ in~\eqref{ExtractHCoeff} where the rational functions $\hat{q}(x)=\frac{a(x)}{b(x)}$ have denominators $b(x)$ that do not have nice irreducible factors over $\mathbb{K}'$ (mostly over $\mathbb{Q}$). As a consequence, the method proposed in step~(4) yields alien sums whose summands are not expected to appear within the final result.
\medskip
\noindent\textit{Example.} Consider
$$\hat{H}(x)=\frac{1}{1-a x}\sum_{N=0}^{\infty } x^N S_{2,1}({N})=\sum_{N=0}^{\infty} H(N)x^N$$
with $a\in\mathbb{Z}$. Then the $N$th coefficient is
$$
H(N)=\begin{cases}
\frac{a^{N+1}}{a-1}S_{2,1}\big({{\frac{1}{a},1},N}\big)
-\frac{1}{a-1}S_{2,1}({N})&\text{ if }a\in\mathbb{Z}\setminus\{0,1\}\\
S_{2,1}({N})&\text{ if }a=0\\
(N+1)S_{2,1}(N)-\frac12 S_2(N)-\frac12 S_1(N)^2&\text{ if }a=1.
\end{cases}
$$
The arising sums with $a=0,\pm1$ pop up almost everywhere within our
calculations~\cite{DiffCalculations}. So denominators of the form $1-ax$ with $a=0,\pm1$ are no surprise. In some instances, also the generalized sums with $a=\pm 2$ arise, and thus also the underlying denominators will appear. However, sums coming from $a\in\mathbb{Z}$ and $a\neq0,\pm1,\pm 2$
have not occurred in our ongoing calculations, but will arise within intermediate calculations when one executes step~(4). The situation gets worse if non-linear factors are treated. One of the simplest cases is $\hat{q}(x)=\frac1{1-x-x^2}$. In this situation, e.g., the subexpression
$$\frac{1}{1-x-x^2}\sum_{N=0}^{\infty} x^N S_{2,1}({N})=\sum_{N=0}^{\infty} H(N)x^N$$
has the $N$th coefficient
\begin{equation}\label{Equ:AlienSums}
H(N)=\frac{3 \sqrt{5}-5}{10}(-1)^N \Phi^{-N-1} S_{2,1}({{-\Phi ,1},N})
+\frac{5+3 \sqrt{5}}{10} \phi ^{-N-1} S_{2,1}({{\phi,1},N})
-S_{2,1}({N})
\end{equation}
with the golden ration $\Phi=\frac{\sqrt{5}+1}{2}$ and $\phi=\frac{\sqrt{5}-1}{2}$.
\medskip
One option is to produce the final expression with all these artificial sums. Then eliminating all relations among these sums (see Remark~\ref{Remark:EliRelation}) will lead to an expression where all alien sums will collapse and the expected sums will remain. However, in many calculations the number of sums in the summands~\eqref{ExtractHCoeff} can be quite large -- we considered cases with up to 1000 sums. Performing than step~(4) naively the number of sums will explode. Even worse, if algebraic numbers like $\phi$ and $\Phi$ arise, the computation of algebraic relations
turns into a real computer algebra challenge.\\
In order to avoid these troubles, the following alternative option worked out in~\cite{VLadders} has been incorporated into the package~\texttt{SumProduction} that we use heavily as a subroutine within our newly developed package~\texttt{SolveCoupledSystem}. Recall that the desired coefficients $\hat{I}_{i,j}(x)$ of~\eqref{Equ:IExpandX}
can be written as a big expressions summed up by subexpressions of the form~\eqref{ExtractHCoeff}. We truncate now the infinite sums within the given expression of $\hat{I}_{i,j}(x)$, i.e., the arising sums are of the form
$$\sum_{N=0}^{A}x^N h_1(N)\dots h_l(N)$$
(instead of the form~\eqref{ExtractHCoeff}). Now we eliminate all algebraic relations among these
nested sums over hypergeometric products (involving in addition the parameter $x$). Here the
following magic happens: in all our examples, we observed that all sums that would contribute to
alien terms as given in~\eqref{Equ:AlienSums} vanish.
Still, the proposed tactic is rather expensive to treat all truncated sums (all with the same upper bound $A$) simultaneously. As
a compromise we filter out only those subexpressions that might contribute to alien sums, more
precisely, we partition the subexpressions into several parts:
(a) sums with denominators that have only nice irreducible factors (over $\mathbb{Q}$) in the
denominator,
which we keep untouched (this is usually the largest part of the full expression), and (b) sums
with denominators that have only bad irreducible factors in the denominator, i.e., which are of
the
form $(1-ax)$ with $a\neq0,\pm1$ (or even $a\neq\pm2$) and factors which are not linear (over $\mathbb{Q}$).
Even more, we partition the sums with bad denominators further such that sums are collected which
have common bad factors. Then we compute for each such partition of bad sums, which consists of conquerable subexpressions, all algebraic relations. In all our examples all these bad sums within their clusters vanish. Finally, we compute the limit $A\to\infty$ and end up at an alternative expression of $\hat{I}_{i,j}(x)$ where now all the unwanted sums are gone. Computing finally the $N$th coefficient as worked out in step~(4) will lead to an expression where the nested sums over hypergeometric products have nice denominators. In particular, the number of these sums is now manageable, and we can compute an alternative representation in terms of nested sums where all arising sums are algebraically independent among each other (see Remark~\ref{Remark:EliRelation}).
\subsubsection{Improvement 2: compute recurrence relations of smaller order (step
(2))}\label{Sec:Improvement2}
\vspace{1mm}
\noindent
Recall the first steps of our proposed procedure.
Suppose we obtained the scalar differential equation~\eqref{Equ:DESingle} in step~(1) with $a_{i,j}(x)$ and $d_{i,j,k}(x)$ being from $\mathbb{K}[x]$. Then plugging in~\eqref{Equ:HatIAnsatz} into~\eqref{Equ:DESingle}, using the rule~\eqref{Equ:DRelation} and taking the $N$th coefficient, we end up at a linear recurrence of the form~\eqref{Equ:CoupledDifferenceSys} where the
the $f_i(N)$ are given as a linear combination of the $b_1(N),\dots,b_r(N)$ over $\mathbb{K}(N)$ with possible shifts in $N$.
Under the assumption that the $b_i(N)$ have an $\varepsilon$-expansion~\eqref{Equ:biExpansion} where the first coefficients $b_{i,j}(N)$ can be given in terms nested sums over hypergeometric products, we can derive an $\varepsilon$-expansion~\eqref{Equ:InhomExpand} of the inhomogeneous part of the recurrence~\eqref{Equ:SingleREC}.
Observe that the order of the derived recurrence is bounded by the maximum of the degrees of the coefficients $a_{i,j}(x)$.
If the order is not too large, the method proposed above works perfectly fine. However, in recent examples we calculated recurrences of order 16 or higher, and it is then almost impossible to calculate 16 (or more) initial values that are needed for step~(3). Luckily, in all these examples it turns out that $d(x)=\gcd(a_{i,0}(x),\dots,\gcd(a_{i,r_i}(x))\in\mathbb{K}[x]$ has a rather high degree. Hence dividing~\eqref{Equ:DESingle} through $d(x)$ leads to the differential equation
\begin{equation}\label{Equ:DESingleMod}
a'_{i,r_i}(x) D^{r_i}_x\hat{I}_i(x)+a'_{i,r_i-1}(x) D^{r_i-1}_x\hat{I}_i(x)+\dots+a'_{i,0}(x) D^{0}_x\hat{I}_i(x)=\sum_{j=1}^r\sum_k d'_{i,j,k}(x)D_x^k\hat{b}_j(x)
\end{equation}
where the $a'_{i,j}=\frac{a_{i,j}(x)}{d(x)}\in\mathbb{K}[x]$ have substantially smaller degrees and where $d'_{i,j,k}=\frac{d_{i,j,k}}{d(x)}\in\mathbb{K}(x)$. Hence taking the $N$th coefficient on the left hand side of~\eqref{Equ:DESingleMod} will yield a difference operator in $I_i(N)$ whose order is substantially smaller (the order is bounded by the maximum of the degrees of the coefficients $a'_{i,j}(x)$). However, in order to get the $N$th coefficient on the right hand side of~\eqref{Equ:DESingleMod}, i.e., in order to get a recurrence of the from~\eqref{Equ:SingleREC}, further calculations are necessary.
Since $d'_{i,j,k}\in\mathbb{K}(x)$ is usually not a polynomial in $x$, we have to apply again the rather
involved calculations steps as sketched in step~(4). In this regard, the improvements of Section~\ref{Sec:Improvement1} play a central role to carry out these calculations efficiently and to provide a linear recurrence with a substantially smaller recurrence order for step~(2) of our procedure.
\section{Conclusion}\label{Sec:Conclusion}
\vspace{1mm}
\noindent
In both algorithms, presented in Subsections~\ref{Sec:PureREC} and~\ref{Sec:DEREC} respectively,
one ends up at scalar recurrences~\eqref{Equ:SingleREC}: sometimes they are the same, sometimes
one or the other method finds a better recurrence (with smaller coefficient size or with lower
recurrence order). This suggests to apply both tactics (up to a certain point) and to execute the
version in full detail that is more appropriate for the concrete problem. As mentioned already
above, a central advantage of the second tactic is that one can uncouple the system straightforwardly
(without any preprocessing steps as sketched in Subsection~\ref{Sec:PureREC} that might blow up the
system). In some instances this leads to a much better space--time behavior. However, in this approach
one has to compute the $N$th coefficient of the remaining integrals, which again can be rather time
consuming. But using our sophisticated symbolic summation technologies
(see Subsection~\ref{Sec:Improvement1}) this problem turns out to be feasible in many examples. The
second advantage of our new method is that one might find recurrences with smaller orders
(see Subsection~\ref{Sec:Improvement2}). As a consequence, one needs less initial values to
determine the respective master integrals. Since the calculations of such initial values is
rather challenging, we expect that this last feature will support future calculations.
|
\section{Introduction} \label{introduction}
We consider eigenvalue problems
\begin{equation} \label{main}
\begin{aligned}
Hy &:= -y'' + V(x) y = \lambda y, \\
\operatorname{dom} (H) &= H^1 (\mathbb{R}),
\end{aligned}
\end{equation}
and also (for any $s \in \mathbb{R}$)
\begin{equation} \label{main_s}
\begin{aligned}
H_s y &:= -y'' + sy' + V(x) y = \lambda y, \\
\operatorname{dom} (H_s) &= H^1 (\mathbb{R}),
\end{aligned}
\end{equation}
where $y \in \mathbb{R}^n$ and $V \in C(\mathbb{R})$ is a real-valued
symmetric matrix satisfying the following asymptotic conditions:
\medskip
\noindent
{\bf (A1)} The limits $\lim_{x \to \pm \infty} V(x) = V_{\pm}$ exist, and
for all $M \in \mathbb{R}$,
\begin{equation*}
\int_{-M}^{\infty} (1 + |x|) |V(x) - V_+| dx < \infty;
\quad
\int_{-\infty}^M (1+|x|) |V(x) - V_-| dx < \infty.
\end{equation*}
\medskip
\noindent
{\bf (A2)} The eigenvalues of $V_{\pm}$ are all non-negative. We denote
the smallest among all these eigenvalues $\nu_{\min} \ge 0$.
\medskip
Our particular interest lies in counting the number of negative
eigenvalues for $H$ (i.e., the Morse index). We proceed by relating
the Morse index to the Maslov index, which is described in
Section \ref{maslov_section}. In essence, we find that the
Morse index can be computed in terms of the Maslov index, and that
while the Maslov index is less elementary than the Morse index,
it can be computed (numerically) in a relatively straightforward way.
The Maslov index has its origins in the work of V. P. Maslov
\cite{Maslov1965a} and subsequent development by V. I. Arnol'd
\cite{arnold67}. It has now been studied extensively, both
as a fundamental geometric quantity \cite{BF98, CLM, F, P96, rs93}
and as a tool for counting the number of eigenvalues on specified
intervals \cite{BJ1995, BM2013, CDB09, CDB11, Chardard2009,
CJLS2014, DJ11, FJN03, J88, J88a, JM2012}. In this latter context,
there has been a strong resurgence of interest following the
analysis by Deng and Jones (i.e., \cite{DJ11}) for multidimensional
domains. Our aim in the current analysis is to rigorously develop
a relationship between the Maslov index and the Morse index in
the relatively simple setting of (\ref{main}). Our approach is
adapted from \cite{CJLS2014, DJ11, HS}.
As a starting point, we define what we will mean by a {\it Lagrangian
subspace} of $\mathbb{R}^{2n}$.
\begin{definition} \label{lagrangian_subspace}
We say $\ell \subset \mathbb{R}^{2n}$ is a Lagrangian subspace
if $\ell$ has dimension $n$ and
\begin{equation*}
(Jx, y)_{\mathbb{R}^{2n}} = 0,
\end{equation*}
for all $x, y \in \ell$. Here, $(\cdot, \cdot)_{\mathbb{R}^{2n}}$ denotes
Euclidean inner product on $\mathbb{R}^{2n}$, and
\begin{equation*}
J =
\begin{pmatrix}
0 & -I_n \\
I_n & 0
\end{pmatrix},
\end{equation*}
with $I_n$ the $n \times n$ identity matrix. We sometimes adopt standard
notation for symplectic forms, $\omega (x,y) = (Jx, y)_{\mathbb{R}^{2n}}$.
In addition, we denote by $\Lambda (n)$ the collection of all Lagrangian
subspaces of $\mathbb{R}^{2n}$, and we will refer to this as the
{\it Lagrangian Grassmannian}.
\end{definition}
A simple example, important for intuition, is the case $n = 1$, for which
$(Jx, y)_{\mathbb{R}^{2}} = 0$ if and only if $x$ and $y$ are linearly
dependent. In this case, we see that any line through the origin is a
Lagrangian subspace of $\mathbb{R}^2$. More generally, any Lagrangian
subspace of $\mathbb{R}^{2n}$ can be spanned by a choice of $n$ linearly
independent vectors in $\mathbb{R}^{2n}$. We will find it convenient to collect
these $n$ vectors as the columns of a $2n \times n$ matrix $\mathbf{X}$,
which we will refer to as a {\it frame} (sometimes {\it Lagrangian frame})
for $\ell$. Moreover, we will often write $\mathbf{X} = {X \choose Y}$,
where $X$ and $Y$ are $n \times n$ matrices.
Suppose $\ell_1 (\cdot), \ell_2 (\cdot)$ denote paths of Lagrangian
subspaces $\ell_i: I \to \Lambda (n)$, for some parameter interval
$I$. The Maslov index associated with these paths, which we will
denote $\operatorname{Mas} (\ell_1, \ell_2; I)$, is a count of the number of times
the paths $\ell_1 (\cdot)$ and $\ell_2 (\cdot)$ intersect, counted
with both multiplicity and direction. (Precise definitions of what we
mean in this context by {\it multiplicity} and {\it direction} will be
given in Section \ref{maslov_section}.) In some cases, the Lagrangian
subspaces will be defined along some path
\begin{equation*}
\Gamma = \{ (x (t), y(t)): t \in I\},
\end{equation*}
and when it's convenient we'll use the notation
$\operatorname{Mas} (\ell_1, \ell_2; \Gamma)$.
We will verify in Section \ref{ODEsection} that under our assumptions
on $V (x)$, and for $\lambda < \nu_{\min}$, (\ref{main}) will have
$n$ linearly independent solutions
that decay as $x \to -\infty$ and $n$ linearly independent solutions
that decay as $x \to +\infty$. We express these respectively as
\begin{equation*}
\begin{aligned}
\phi_{n+j}^- (x; \lambda) &= e^{\mu_{n+j}^- (\lambda) x} (r_j^- + \mathcal{E}_j^- (x;\lambda)) \\
\phi_j^+ (x; \lambda) &= e^{\mu_j^+ (\lambda) x} (r_{n+1-j}^+ + \mathcal{E}_j^+ (x;\lambda)),
\end{aligned}
\end{equation*}
with also
\begin{equation*}
\begin{aligned}
\partial_x \phi_{n+j}^- (x; \lambda) &= e^{\mu_{n+j}^- (\lambda) x} (\mu_{n+j}^- r_j^- + \tilde{\mathcal{E}}_j^- (x;\lambda)) \\
\partial_x \phi_j^+ (x; \lambda) &= e^{\mu_j^+ (\lambda) x} (\mu_j^+ r_{n+1-j}^+ + \tilde{\mathcal{E}}_j^+ (x;\lambda)),
\end{aligned}
\end{equation*}
for $j = 1,2,\dots,n$, where the nature of the $\mu_j^{\pm}$, $r_j^{\pm}$, and
$\mathcal{E}_j^{\pm} (x; \lambda), \tilde{\mathcal{E}}_j^{\pm} (x; \lambda)$
are developed in Section \ref{ODEsection}. The only details we'll need for
this preliminary discussion are: (1) that we can continuously extend these
functions to $\lambda = \nu_{\min}$ (though they may no longer decay at
one or both endstates), and (2) that under assumptions (A1) and (A2)
\begin{equation} \label{limits}
\lim_{x \to \pm \infty} \mathcal{E}_j^{\pm} (x;\lambda) = 0;
\quad
\lim_{x \to \pm \infty} \tilde{\mathcal{E}}_j^{\pm} (x;\lambda) = 0.
\end{equation}
We will verify in Section \ref{ODEsection} that if we create a frame
$\mathbf{X}^- (x; \lambda) = {X^- (x; \lambda) \choose Y^- (x; \lambda)}$
by taking $\{\phi_{n+j}^-\}_{j=1}^n$ as the columns of $X^-$ and
$\{\phi_{n+j}^{- \, '}\}_{j=1}^n$ as the respective columns of $Y^-$ then
$\mathbf{X}^-$ is a frame for a Lagrangian subspace, which we will
denote $\ell^-$. Likewise, we can create a frame
$\mathbf{X}^+ (x; \lambda) = {X^+ (x; \lambda) \choose Y^+ (x; \lambda)}$
by taking $\{\phi_j^+\}_{j=1}^n$ as the columns of $X^+$ and
$\{{\phi_j^+}'\}_{j=1}^n$ as the respective columns of $Y^+$. Then
$\mathbf{X}^+$ is a frame for a Lagrangian subspace, which we will
denote $\ell^+$.
In constucting our Lagrangian frames, we can view the exponential
multipliers $e^{\mu_j^{\pm} x}$ as expansion coefficients, and if
we drop these off we retain frames for the same spaces. That is, we
can create an alternative frame for
$\ell^-$ by taking the expressions $r_j^- + \mathcal{E}_j^- (x;\lambda)$
as the columns of $X^-$ and the expressions
$\mu_{n+j}^- r_j^- + \tilde{\mathcal{E}}_j^- (x;\lambda)$ as the
corresponding columns for $Y^-$. Using (\ref{limits}) we see that
in the limit as $x$ tends to $-\infty$ we obtain the frame
$\mathbf{R}^- (\lambda) = {R^- \choose S^- (\lambda)}$, where
\begin{equation*}
\begin{aligned}
R^- &=
\begin{pmatrix}
r_1^- & r_2^- & \dots & r_n^-
\end{pmatrix} \\
S^- (\lambda) &=
\begin{pmatrix}
\mu_{n+1}^- (\lambda) r_1^- & \mu_{n+2}^- (\lambda) r_2^- & \dots & \mu_{2n}^- (\lambda) r_n^-
\end{pmatrix}.
\end{aligned}
\end{equation*}
(The dependence on $\lambda$ is specified here to emphasize the fact
that $S^- (\lambda)$ depends on $\lambda$ through the multipliers
$\{\mu_{n+j}^-\}_{j=1}^n$.) We will verify in Section \ref{ODEsection}
that $\mathbf{R}^- (\lambda)$ is the frame for a Lagrangian subspace,
and we denote this space $\ell_{\mathbf{R}}^- (\lambda)$.
Proceeding similarly with $\ell^+$, we obtain the asymptotic Lagrangian
subspace $\ell_{\mathbf{R}}^+ (\lambda)$ with frame
$\mathbf{R}^+ (\lambda) = {R^+ \choose S^+ (\lambda)}$,
where
\begin{equation} \label{RplusSplus}
\begin{aligned}
R^+ &=
\begin{pmatrix}
r_n^+ & r_{n-1}^+ & \dots & r_1^+
\end{pmatrix} \\
S^+ (\lambda) &=
\begin{pmatrix}
\mu_1^+ (\lambda) r_n^+ & \mu_2^+ (\lambda) r_{n-1}^+ & \dots & \mu_n^+ (\lambda) r_1^+
\end{pmatrix}.
\end{aligned}
\end{equation}
(The ordering of the columns of $\mathbf{R}^+$ is simply a convention, which
follows naturally from our convention for
indexing $\{\phi_j^+\}_{j=1}^n$.)
Let $\bar{\Gamma}_0$ denote the contour in the $x$-$\lambda$ plane
obtained by fixing $\lambda = 0$ and letting $x$ run from $-\infty$
to $\infty$.
We are now prepared to state the main theorem of the paper.
\begin{theorem} \label{main_theorem}
Let $V \in C(\mathbb{R})$ be a real-valued symmetric matrix, and suppose (A1)
and (A2) hold. Then
\begin{equation*}
\operatorname{Mor}(H) = - \operatorname{Mas}(\ell^-, \ell^+_{\mathbf{R}}; \bar{\Gamma}_0).
\end{equation*}
\end{theorem}
\begin{remark} The advantage of this theorem resides in the fact
that the Maslov index on the right-hand side is generally
straightforward to compute numerically. See, for example,
\cite{BM2013, CDB06, CDB09, CDB11, Chardard2009}, and the
examples we discuss in Section \ref{applications_section}.
The choice of $\lambda = 0$ for $\bar{\Gamma}_0$ is not necessary
for the analysis, and indeed if we fix any $\lambda_0$ so that
$\sigma_{ess} (H) \subset [\lambda_0, \infty)$ then
$\operatorname{Mas} (\ell^-, \ell_{\mathbf{R}}^+; \bar{\Gamma}_{\lambda_0})$
will be negative the count of eigenvalues of $H$ strictly
less than $\lambda_0$. Since $\bar{\Gamma}_0$ plays a distinguished
role, we refer to $\operatorname{Mas} (\ell^-, \ell_{\mathbf{R}}^+; \bar{\Gamma}_{0})$
as the {\it Principal Maslov Index} (following \cite{HS}).
\end{remark}
\begin{remark} In Section \ref{maslov_section} our definition of
the Maslov index will be for compact intervals $I$. We will see
that we are able to view $\ell^- (x; 0)$ as a continuous path
of Lagrangian subspaces on $[-1,1]$ by virtue of the change
of variables
\begin{equation} \label{change}
x = \ln (\frac{1 + \tau}{1 - \tau}).
\end{equation}
\end{remark}
We will verify in Section \ref{convection_section} that for
$s \in \mathbb{R}$, any eigenvalue of $H_s$ with real part
less than or equal to $\nu_{\min}$ must be real-valued. This
observation will allow us to construct the Lagrangian subspaces
$\ell^-$ and $\ell^+_{\mathbf{R}}$ in that case through a
development that looks identical to the discussion above. We
obtain the following theorem.
\begin{theorem} \label{main_theorem_s}
Let $V \in C(\mathbb{R})$ be a real-valued symmetric matrix, and suppose (A1)
and (A2) hold. Let $s \in \mathbb{R}$, and let $\ell^-$ and $\ell^+_{\mathbf{R}}$
denote Lagrangian subspaces developed for (\ref{main_s}). Then
\begin{equation*}
\operatorname{Mor}(H_s) = - \operatorname{Mas}(\ell^-, \ell^+_{\mathbf{R}}; \bar{\Gamma}_0).
\end{equation*}
\end{theorem}
\begin{remark} As described in more detail in Sections
\ref{convection_section} and \ref{applications_section},
equations of forms (\ref{main}) and (\ref{main_s})
arise naturally when a gradient system
\begin{equation*}
u_t + F'(u) = u_{xx}
\end{equation*}
is linearized about a stationary solution $\bar{u} (x)$
or a traveling wave solution $\bar{u} (x - st)$
(respectively). The case of solitary waves, for
which (without loss of generality)
\begin{equation*}
\lim_{x \to \pm \infty} \bar{u} (x) = 0,
\end{equation*}
has been analyzed in
\cite{BJ1995, CDB09, CDB11, CH2007, Chardard2009}
(with $s \ne 0$ in \cite{BJ1995} and $s = 0$ in the
others). In particular, theorems along the lines of
our Theorem \ref{main_theorem} (though restricted
to the case of solitary waves) appear as Corollary 3.8
in \cite{BJ1995} and Proposition 35 in Appendix C.2
of \cite{Chardard2009}.
\end{remark}
{\it Plan of the paper.} In Section \ref{ODEsection} we develop
several relatively standard results from ODE theory that will be
necessary for our construction and analysis of the Maslov index.
In Section \ref{maslov_section}, we define the Maslov index,
and discuss some of its salient properties, and in Section
\ref{schrodinger_section} we prove Theorem \ref{main_theorem}.
In Section \ref{convection_section}, we verify that the analysis
can be extended to the case of any $s \in \mathbb{R}$, and finally,
in Section \ref{applications_section} we provide some
illustrative applications.
\section{ODE Preliminaries} \label{ODEsection}
In this section, we develop preliminary ODE results that will serve as the
foundation of our analysis. This development is standard, and follows \cite{ZH}, pp.
779-781. We begin by clarifying our terminology.
\begin{definition} \label{spectrum}
We define the point spectrum of $H$, denoted $\sigma_{pt} (H)$, as the set
\begin{equation*}
\sigma_{pt} (H) = \{\lambda \in \mathbb{R}: H \phi = \lambda \phi
\, \, \text{for some} \, \, \phi \in H^1 (\mathbb{R}) \backslash \{0\}\}.
\end{equation*}
We define the essential spectrum of $H$, denoted $\sigma_{ess} (H)$,
as the values in $\mathbb{R}$ that
are not in the resolvent set of $H$ and are not isolated eigenvalues
of finite multiplicity.
\end{definition}
As discussed, for example, in \cite{Henry, KP}, the essential spectrum of $H$ is determined
by the asymptotic equations
\begin{equation} \label{asymptotic}
- y'' + V_{\pm} y = \lambda y.
\end{equation}
In particular, if we look for solutions
of the form $y (x) = e^{i k x} r$, for some scalar constant $k \in \mathbb{R}$
and (non-zero) constant vector $r \in \mathbb{R}^n$ then the essential spectrum
will be confined to the allowable values of $\lambda$. For (\ref{asymptotic}),
we find
\begin{equation*}
(k^2 I + V_{\pm}) r = \lambda r,
\end{equation*}
so that
\begin{equation*}
\lambda(k) \ge \frac{(V_{\pm}r, r)}{\|r\|^2}.
\end{equation*}
Applying the min-max principle, we see that if the eigenvalues of
$V_{\pm}$ are all non-negative then we will have
$\sigma_{ess} (H) \subset [0, \infty)$, and more generally
if $\nu_{\min}$ denotes the smallest eigenvalue of $V_{\pm}$ then we will
have $\sigma_{ess} (H) \subset [\nu_{\min}, \infty)$.
Away from essential spectrum, we begin our construction of asymptotically
decaying solutions to (\ref{main}) by looking for solutions of
(\ref{asymptotic}) of the form $\phi (x;\lambda) = e^{\mu x} r$,
where in this case $\mu$ is a scalar function of $\lambda$, and
$r$ is again a constant vector in $\mathbb{R}^n$. In this case,
we obtain the relation
\begin{equation*}
(-\mu^2 I + V_{\pm} - \lambda I) r = 0,
\end{equation*}
from which we see that the values of $\mu^2 + \lambda$
will correspond with eigenvalues of $V_{\pm}$, and
the vectors $r$ will be eigenvectors of $V_{\pm}$. We denote the
spectrum of $V_{\pm}$ by $\sigma (V_{\pm}) = \{\nu_j^{\pm}\}_{j=1}^n$,
ordered so that $j < k$ implies $\nu_j^{\pm} \le \nu_k^{\pm}$, and
we order the eigenvectors correspondingly so that
$V_{\pm} r_j^{\pm} = \nu_j^{\pm} r_j^{\pm}$ for all
$j \in \{1, 2, \dots, n\}$. Moreover, since $V_{\pm}$ are symmetric
matrices, we can choose the set $\{r_j^{-}\}_{j=1}^n$ to be
orthonormal, and similarly for $\{r_j^{+}\}_{j=1}^n$.
We have
\begin{equation*}
\mu^2 + \lambda = \nu_j^{\pm} \implies
\mu = \pm \sqrt{\nu_j^{\pm} - \lambda}.
\end{equation*}
We will denote the admissible values of $\mu$ by $\{\mu_j^{\pm}\}_{j=1}^{2n}$,
and for consistency we choose our labeling scheme so that
$j < k$ implies $\mu_j^{\pm} \le \mu_k^{\pm}$ (for $\lambda \le \nu_{\min}$).
This leads us to the specifications
\begin{equation*}
\begin{aligned}
\mu_j^{\pm} (\lambda) &= - \sqrt{\nu_{n+1-j}^{\pm} - \lambda} \\
\mu_{n+j}^{\pm} (\lambda) &= \sqrt{\nu_{j}^{\pm} - \lambda},
\end{aligned}
\end{equation*}
for $j = 1, 2, \dots, n$.
We now express (\ref{main}) as a first order system, with
$\mathbf{p} = {p \choose q} = {y \choose y'}$. We find
\begin{equation} \label{first_order}
\frac{d \mathbf{p}}{dx} = \mathbb{A} (x; \lambda) \mathbf{p}; \quad
\mathbb{A} (x; \lambda) =
\begin{pmatrix}
0 & I \\
V(x) - \lambda I & 0
\end{pmatrix},
\end{equation}
and we additionally set
\begin{equation*}
\mathbb{A}_{\pm} (\lambda) := \lim_{x \to \pm \infty} \mathbb{A} (x; \lambda)
=
\begin{pmatrix}
0 & I \\
V_{\pm} - \lambda I & 0
\end{pmatrix}.
\end{equation*}
We note that the eigenvalues of $\mathbb{A}_{\pm}$ are precisely the values
$\{\mu_j^{\pm}\}_{j=1}^{2n}$, and the associated eigenvectors are
$\{\scripty{r}_{\,j}^{\,\pm}\}_{j=1}^n =
\{{ r_{n+1-j}^{\pm} \choose \mu_j^{\pm} {r_{n+1-j}^{\pm}}}\}_{j=1}^n$ and
$\{\scripty{r}_{\,n+j}^{\,\pm}\}_{j=1}^n =
\{{r_{j}^{\pm} \choose {\mu_{n+j}^{\pm} {r_{j}^{\pm}}}}\}_{j=1}^n$.
\begin{lemma} \label{ODElemma}
Let $V \in C(\mathbb{R})$ be a real-valued symmetric matrix, and suppose (A1)
and (A2) hold. Then for any $\lambda < \nu_{\min}$ there exist $n$
linearly independent solutions of
(\ref{first_order}) that decay as $x \to -\infty$ and $n$ linearly independent
solutions of (\ref{first_order}) that decay as $x \to +\infty$. Respectively,
we can choose these so that they can be expressed as
\begin{equation*}
\begin{aligned}
\mathbf{p}_{n+j}^- (x; \lambda) &= e^{\mu_{n+j}^- (\lambda) x}
(\scripty{r}_{\,n+j}^{\,-}
+ \mathbf{E}_{n+j}^-); \quad j = 1, 2, \dots, n, \\
\mathbf{p}_j^+ (x; \lambda) &= e^{\mu_{j}^+ (\lambda) x}
(\scripty{r}_{\,j}^{\,+}
+ \mathbf{E}_j^+); \quad j = 1, 2, \dots, n,
\end{aligned}
\end{equation*}
where for any $\lambda_{\infty} > 0$,
$\mathbf{E}_{n+j}^{-} = \mathbf{O} ((1+|x|)^{-1})$,
uniformly for $\lambda \in [-\lambda_{\infty}, \nu_{\min}]$, and
similarly for $\mathbf{E}_j^+$.
Moreover, there exist $n$ linearly independent solutions of
(\ref{first_order}) that grow as $x \to -\infty$ and $n$ linearly independent
solutions of (\ref{first_order}) that grow as $x \to +\infty$. Respectively,
we can choose these so that they can be expressed as
\begin{equation*}
\begin{aligned}
\mathbf{p}_{j}^- (x; \lambda) &= e^{\mu_{j}^- (\lambda) x}
(\scripty{r}_{\,j}^{\,-}
+ \mathbf{E}_{j}^-); \quad j = 1, 2, \dots, n, \\
\mathbf{p}_{n+j}^+ (x; \lambda) &= e^{\mu_{n+j}^+ (\lambda) x}
(\scripty{r}_{\,n+j}^{\,+}
+ \mathbf{E}_{n+j}^+); \quad j = 1, 2, \dots, n, \\
\end{aligned}
\end{equation*}
where for any $\lambda_{\infty} > 0$,
$\mathbf{E}_j^{-} = \mathbf{O} ((1+|x|)^{-1})$,
uniformly for $\lambda \in [-\lambda_{\infty}, \nu_{\min}]$,
and similarly for $\mathbf{E}_{n+j}^+$.
Finally, the solutions extend continuously as $\lambda \to \nu_{\min}$
(from the left)
to solutions of (\ref{main}) that neither grow nor decay at the
associated endstate.
\end{lemma}
\begin{proof}
Focusing on solutions that decay as $x \to - \infty$, we express
(\ref{first_order}) as
\begin{equation} \label{first_order_minus}
\frac{d \mathbf{p}}{dx} = \mathbb{A}_- (\lambda) \mathbf{p}
+ \mathcal{R}_- (x) \mathbf{p}; \quad
\mathcal{R}_- (x) = \mathbb{A} (x; \lambda) - \mathbb{A}_- (\lambda)
=
\begin{pmatrix}
0 & 0 \\
V(x) - V_- & 0
\end{pmatrix}.
\end{equation}
We have seen that asymptotically decaying solutions to the
asymptotic equation
$\frac{d\mathbf{p}}{dx} = \mathbb{A}_- (\lambda) \mathbf{p}$
have the form $\mathbf{p}_{n+j}^- (x; \lambda) = e^{\mu_{n+j}^- (\lambda)} \scripty{r}_{\,n+j}^{\,-}$,
and so it's natural to look for solutions of the form
\begin{equation*}
\mathbf{p}_{n+j}^- (x; \lambda) = e^{\mu_{n+j}^- (\lambda) x} \mathbf{z}_{n+j}^- (x;\lambda),
\end{equation*}
for which we have
\begin{equation} \label{scaled_equation}
\frac{d \mathbf{z}_{n+j}^- (x;\lambda)}{dx} =
(\mathbb{A}_- (\lambda) - \mu_{n+j}^- (\lambda) I) \mathbf{z}_{n+j}^- (x; \lambda)
+ \mathcal{R}_- (x) \mathbf{z}_{n+j}^- (x; \lambda).
\end{equation}
Let $P_{n+j}^- (\lambda)$ project onto the eigenspace of $\mathbb{A}_- (\lambda)$
associated with eigenvalues $\sigma (\mathbb{A}_- (\lambda)) \ni \mu \le \mu_{n+j}^-$,
and let $Q_{n+j}^- (\lambda)$ likewise project onto the eigenspace of $\mathbb{A}_- (\lambda)$
associated with $\sigma (\mathbb{A}_- (\lambda)) \ni \mu > \mu_{n+j}^-$.
Notice particularly that there exists some $\eta > 0$ so that $\mu - \mu_{n+j}^- \ge \eta$
for all $\mu$ associated with $Q_{n+j}^- (\lambda)$.
For some fixed $M>0$, we will look for a solution to (\ref{scaled_equation}) in
$L^{\infty} (-\infty, -M]$ of the form
\begin{equation} \label{contraction_map}
\begin{aligned}
\mathbf{z}_{n+j}^- (x; \lambda) &= \scripty{r}_{\,n+j}^{\,-}
+ \int_{-\infty}^x e^{(\mathbb{A}_- (\lambda) - \mu_{n+j}^- (\lambda) I) (x - \xi)}
P_{n+j}^- (\lambda) \mathcal{R}_- (\xi) \mathbf{z}_{n+j}^- (\xi; \lambda) d\xi \\
&-
\int_x^{-M} e^{(\mathbb{A}_- (\lambda) - \mu_{n+j}^- (\lambda) I) (x - \xi)}
Q_{n+j}^- (\lambda) \mathcal{R}_- (\xi) \mathbf{z}_{n+j}^- (\xi; \lambda) d\xi.
\end{aligned}
\end{equation}
We proceed by contraction mapping, defining $\mathcal{T} \mathbf{z}_{n+j}^- (x; \lambda)$
to be the right-hand side of (\ref{contraction_map}). Let
$\mathbf{z}_{n+j}^-, \mathbf{w}_{n+j}^- \in L^{\infty} (-\infty, -M]$, so that
\begin{equation}
\begin{aligned}
|\mathcal{T} \mathbf{z}_{n+j}^- - \mathcal{T} \mathbf{w}_{n+j}^-|
& \le
K \|\mathbf{z}_{n+j}^- - \mathbf{w}_{n+j}^- \|_{L^{\infty} (-\infty, -M]}
\Big{\{} \int_{-\infty}^{x} |\mathcal{R}_- (\xi)| d\xi
+ \int_{x}^{-M} e^{\eta (x-\xi)} |\mathcal{R}_- (\xi)| d\xi
\Big{\}} \\
&=: I_1 + I_2,
\end{aligned}
\end{equation}
for some constant $K > 0$.
By assumption (A1) we know
\begin{equation*}
\int_{-\infty}^0 (1+|x|) |\mathcal{R}_- (x)| dx = C < \infty,
\end{equation*}
so that
\begin{equation*}
\int_{-\infty}^{-M} (1+M) |\mathcal{R}_- (\xi)| d\xi
\le
\int_{-\infty}^{-M} (1+|\xi|) |\mathcal{R}_- (\xi)| d\xi
\le C,
\end{equation*}
giving the inequality
\begin{equation*}
\int_{-\infty}^{-M} |\mathcal{R}_- (\xi)| dx \le \frac{C}{1+M}.
\end{equation*}
Likewise, we can check that
\begin{equation*}
\int_{x}^{-M} e^{\eta (x-\xi)} |\mathcal{R}_- (\xi)| dx
\le \frac{C}{1+M}.
\end{equation*}
We see that
\begin{equation*}
|\mathcal{T} \mathbf{z}_{n+j}^- - \mathcal{T} \mathbf{w}_{n+j}^-| \le
\frac{2 K C}{1+M} \|\mathbf{z}_{n+j}^- - \mathbf{w}_{n+j}^- \|_{L^{\infty} (-\infty, -M]},
\end{equation*}
for all $x \in (-\infty, -M]$ so that
\begin{equation*}
\|\mathcal{T} \mathbf{z}_{n+j}^- - \mathcal{T} \mathbf{w}_{n+j}^-\|_{L^{\infty} (-\infty, -M]} \le
\frac{2 K C}{1+M} \|\mathbf{z}_{n+j}^- - \mathbf{w}_{n+j}^- \|_{L^{\infty} (-\infty, -M]},
\end{equation*}
and for $M$ large enough we have the desired contraction. Moreover, the exponential
decay in $I_2$ allows us to see that
\begin{equation*}
\lim_{x \to -\infty} \mathbf{z}_{n+j}^- (x; \lambda) = \scripty{r}_{\,n+j}^{\,-},
\end{equation*}
with the asymptotic rate indicated.
For continuity down to $\lambda = \mu_{\min}$, we notice that in this case some of the
$\mu_{n+j}^-$ may be 0, so $\mathbf{p}_{n+j}^-$ will not decay as $x \to -\infty$.
Nonetheless, our calculation remains valid, and in this case there is simply no
exponential scaling.
Finally, we note that the case $x \to +\infty$ is similar.
\end{proof}
Recall that we denote by $\mathbf{X}^- (x; \lambda)$ the $2n \times n$
matrix obtained by taking each $\mathbf{p}_{n+j}^- (x; \lambda)$ from
Lemma \ref{ODElemma} as a column. In order to check that
$\mathbf{X}^- (x; \lambda)$ is the frame for a Lagrangian subspace, let
$\phi, \psi \in \{\mathbf{p}_{n+j}^- (x; \lambda)\}_{j=1}^n$, and consider
$\omega (\phi, \psi) = (J \phi, \psi)$. First,
\begin{equation*}
\frac{d}{dx} \omega (\phi, \psi) =
(J \frac{d\phi}{dx}, \psi) + (J\phi, \frac{d\psi}{dx})
= (J \mathbb{A} \phi, \psi) + (J\phi, \mathbb{A} \psi).
\end{equation*}
It's important to note at this point that we can express $\mathbb{A}$
as $\mathbb{A} = J \mathbb{B}$, for the symmetric matrix
\begin{equation*}
\mathbb{B} (x; \lambda) =
\begin{pmatrix}
V(x) - \lambda I & 0 \\
0 & -\lambda I
\end{pmatrix}.
\end{equation*}
Consequently
\begin{equation*}
\begin{aligned}
\frac{d}{dx} \omega (\phi, \psi) &=
(J^2 \mathbb{B} \phi, \psi) + (J\phi, J\mathbb{B} \psi)
=
- (\mathbb{B} \phi, \psi) - (J^2 \phi, \mathbb{B} \psi) \\
&=
- (\mathbb{B} \phi, \psi) + (\phi, \mathbb{B} \psi) = 0,
\end{aligned}
\end{equation*}
where the final equality follows from the symmetry of $\mathbb{B}$.
We conclude that $\omega (\phi, \psi)$ is constant in $x$, but since
\begin{equation*}
\lim_{x \to -\infty} \omega (\phi, \psi) = 0,
\end{equation*}
this constant must be 0.
In order to see that this limit holds even if neither
$\phi$ nor $\psi$ decays as $x \to -\infty$ (possible if
$\lambda = \nu_{\min}$), we note that in this case we have
\begin{equation*}
\begin{aligned}
\phi (x;\lambda) &= \scripty{r}_{\,n+i}^{\,-} + \mathbf{E}_{n+i}^- \\
\psi (x;\lambda) &= \scripty{r}_{\,n+j}^{\,-} + \mathbf{E}_{n+j}^-,
\end{aligned}
\end{equation*}
for some $i \ne j$ and with $\mu_{n+i}$ and $\mu_{n+j}$ both 0. Then
\begin{equation*}
\lim_{x \to -\infty} \omega (\phi, \psi) = \omega (\scripty{r}_{\,n+i}^{\,-}, \scripty{r}_{\,n+j}^{\,-})
= (J {r_i^- \choose 0}, {r_j^- \choose 0}) = 0.
\end{equation*}
Proceeding in the same way, we can verify that $\mathbf{X}^+ (x; \lambda)$
is also a frame for a Lagrangian subspace.
We conclude this section by verifying that $\mathbf{R}^- (\lambda)$
(specified in the introduction) is the frame
for a Lagrangian subspace. To see this, we change notation a bit
from the previous calculation and take
${\phi \choose \mu \phi}, {\psi \choose \nu \psi}
\in \{\scripty{r}_{\,n+j}^{\,-}\}_{j=1}^n$. We compute
\begin{equation*}
\omega ({\phi \choose \mu \phi}, {\psi \choose \nu \psi})
= (J {\phi \choose \mu \phi}, {\psi \choose \nu \psi})
= (\nu - \mu) (\phi, \psi) = 0,
\end{equation*}
where the final equality follows from orthogonality of
the eigenvectors of $V_-$. Likewise, we find that
$\mathbf{R}^+ (\lambda)$ is a Lagrangian subspace.
\section{The Maslov Index} \label{maslov_section}
Given any two Lagrangian subspaces $\ell_1$ and $\ell_2$, with associated
frames $\mathbf{X}_1 = {X_1 \choose Y_1}$ and
$\mathbf{X}_2 = {X_2 \choose Y_2}$, we can define the complex $n \times n$
matrix
\begin{equation} \label{tildeW}
\tilde{W} = - (X_1 + i Y_1) (X_1 - i Y_1)^{-1} (X_2 - i Y_2) (X_2 + i Y_2)^{-1}.
\end{equation}
As verified in \cite{HLS}, the matrices $(X_1 - iY_1)$ and $(X_2 + iY_2)$ are both
invertible, and $\tilde{W}$ is unitary. We have the following theorem
from \cite{HLS}.
\begin{theorem} \label{intersection_theorem}
Suppose $\ell_1, \ell_2 \subset \mathbb{R}^{2n}$ are Lagrangian
subspaces, with respective frames $\mathbf{X}_1 = {X_1 \choose Y_1}$ and
$\mathbf{X}_2 = {X_2 \choose Y_2}$, and let $\tilde{W}$ be as defined
in (\ref{tildeW}). Then
\begin{equation*}
\dim \ker (\tilde{W} + I) = \dim (\ell_1 \cap \ell_2).
\end{equation*}
That is, the dimension of the eigenspace of $\tilde{W}$ associated with
the eigenvalue $-1$ is precisely the dimension of the intersection of
the Lagrangian subspaces $\ell_1$ and $\ell_2$.
\end{theorem}
Following \cite{BF98, F}, we use Theorem \ref{intersection_theorem},
along with an approach to spectral flow introduced in \cite{P96},
to define the Maslov index. Given a parameter interval $I = [a,b]$,
which can be normalized to $[0,1]$, we consider maps
$\ell:I \to \Lambda (n)$, which will be
expressed as $\ell (t)$. In order to specify a notion of continuity,
we need to define a metric on $\Lambda (n)$, and following
\cite{F} (p. 274), we do this in terms of orthogonal projections
onto elements $\ell \in \Lambda (n)$. Precisely, let $\mathcal{P}_i$
denote the orthogonal projection matrix onto $\ell_i \in \Lambda (n)$
for $i = 1,2$. I.e., if $\mathbf{X}_i$ denotes a frame for $\ell_i$,
then $\mathcal{P}_i = \mathbf{X}_i (\mathbf{X}_i^t \mathbf{X}_i)^{-1} \mathbf{X}_i^t$.
We take our metric $d$ on $\Lambda (n)$ to be defined
by
\begin{equation*}
d (\ell_1, \ell_2) := \|\mathcal{P}_1 - \mathcal{P}_2 \|,
\end{equation*}
where $\| \cdot \|$ can denote any matrix norm. We will say
that $\ell: I \to \Lambda (n)$ is continuous provided it is
continuous under the metric $d$.
Given two continuous maps $\ell_1 (t), \ell_2 (t)$ on a parameter
interval $I$, we denote by $\mathcal{L}(t)$ the path
\begin{equation*}
\mathcal{L} (t) = (\ell_1 (t), \ell_2 (t)).
\end{equation*}
In what follows, we will define the Maslov index for the path
$\mathcal{L} (t)$, which will be a count, including both multiplicity
and direction, of the number of times the Lagrangian paths
$\ell_1$ and $\ell_2$ intersect. In order to be clear about
what we mean by multiplicty and direction, we observe that
associated with any path $\mathcal{L} (t)$ we will have
a path of unitary complex matrices as described in (\ref{tildeW}).
We have already noted that the Lagrangian subspaces $\ell_1$
and $\ell_2$ intersect at a value $t_0 \in I$ if and only
if $\tilde{W} (t_0)$ has -1 as an eigenvalue. In the event of
such an intersection, we define the multiplicity of the
intersection to be the multiplicity of -1 as an eigenvalue of
$\tilde{W}$ (since $\tilde{W}$ is unitary the algebraic and geometric
multiplicites are the same). When we talk about the direction
of an intersection, we mean the direction the eigenvalues of
$\tilde{W}$ are moving (as $t$ varies) along the unit circle
$S^1$ when they cross $-1$ (we take counterclockwise as the positive direction). We note
that all of the eigenvalues certainly do not all need to be moving in
the same direction, and that we will need to take care with
what we mean by a crossing in the following sense: we must decide
whether to increment the Maslov index upon arrival or
upon departure. Indeed, there are several different approaches
to defining the Maslov index (see, for example, \cite{CLM, rs93}),
and they often disagree on this convention.
Following \cite{BF98, F, P96} (and in particular Definition 1.4
from \cite{BF98}), we proceed by choosing a
partition $a = t_0 < t_1 < \dots < t_n=b$ of $I = [a,b]$, along
with numbers $\epsilon_j \in (0,\pi)$ so that
$\ker\big(\tilde{W} (t) - e^{i (\pi + \epsilon_j)} I\big)=\{0\}$
for $t_{j-1} \le t \le t_j$;
that is, $e^{i(\pi + \epsilon_j)} \in {\mathbb{C}} \setminus \sigma(\tilde{W} (t))$,
for $t_{j-1} \le t \le t_j$ and $j=1,\dots,n$.
Moreover, we notice that for each $j=1,\dots,n$ and any
$t \in [t_{j-1},t_j]$ there are only
finitely many values $\theta \in [0,\epsilon_j)$
for which $e^{i(\pi+\theta)} \in \sigma(\tilde{W} (t))$.
Fix some $j \in \{1, 2, \dots, n\}$ and consider the value
\begin{equation} \label{kdefined}
k (t,\epsilon_j) :=
\sum_{0 \leq \theta < \epsilon_j}
\dim \ker \big(\tilde{W} (t) - e^{i(\pi+\theta)}I \big).
\end{equation}
for $t_{j-1} \leq t \leq t_j$. This is precisely the sum, along with multiplicity,
of the number of eigenvalues of $\tilde{W} (t)$ that lie on the arc
\begin{equation*}
A_j := \{e^{i t}: t \in [\pi, \pi+\epsilon_j)\}.
\end{equation*}
(See Figure \ref{Aj}.)
The stipulation that
$e^{i(\pi\pm\epsilon_j)} \in {\mathbb{C}}\setminus \sigma(\tilde{W} (t))$, for
$t_{j-1} \le t \le t_j$
asserts that no eigenvalue can enter $A_j$ in the clockwise direction
or exit in the counterclockwise direction during the interval $t_{j-1} \le t \le t_j$.
In this way, we see that $k(t_j, \epsilon_j) - k (t_{j-1}, \epsilon_j)$ is a
count of the number of eigenvalues that enter $A_j$ in the counterclockwise
direction (i.e., through $-1$) minus the number that leave in the clockwise direction
(again, through $-1$) during the interval $[t_{j-1}, t_j]$.
\begin{figure}[ht]
\begin{center}\includegraphics[%
width=8cm,
height=8cm]{Maslov_S1a.pdf}\end{center}
\caption{The arc $A_j$. \label{Aj}}
\end{figure}
In dealing with the catenation of paths, it's particularly important to
understand this quantity if an eigenvalue resides at $-1$ at either $t = t_{j-1}$
or $t = t_j$ (i.e., if an eigenvalue begins or ends at a crosssing). If an eigenvalue
moving in the counterclockwise direction
arrives at $-1$ at $t = t_j$, then we increment the difference forward, while if
the eigenvalue arrives at -1 from the clockwise direction we do not (because it
was already in $A_j$ prior to arrival). On
the other hand, suppose an eigenvalue resides at -1 at $t = t_{j-1}$ and moves
in the counterclockwise direction. The eigenvalue remains in $A_j$, and so we do not increment
the difference. However, if the eigenvalue leaves in the clockwise direction
then we decrement the difference. In summary, the difference increments forward upon arrivals
in the counterclockwise direction, but not upon arrivals in the clockwise direction,
and it decrements upon departures in the clockwise direction, but not upon
departures in the counterclockwise direction.
We are now ready to define the Maslov index.
\begin{definition} \label{dfnDef3.6}
Let $\mathcal{L} (t) = (\ell_1 (t), \ell_2 (t))$, where $\ell_1, \ell_2:I \to \Lambda (n)$
are continuous paths in the Lagrangian--Grassmannian.
The Maslov index $\operatorname{Mas}(\mathcal{L};I)$ is defined by
\begin{equation}
\operatorname{Mas}(\mathcal{L};I)=\sum_{j=1}^n(k(t_j,\epsilon_j)-k(t_{j-1},\epsilon_j)).
\end{equation}
\end{definition}
\begin{remark} As discussed in \cite{BF98}, the Maslov index does not depend
on the choices of $\{t_j\}_{j=0}^n$ and $\{\epsilon_j\}_{j=1}^n$, so long as
they follow the specifications above.
\end{remark}
One of the most important features of the Maslov index is homotopy invariance,
for which we need to consider continuously varying families of Lagrangian
paths. To set some notation, we denote by $\mathcal{P} (I)$ the collection
of all paths $\mathcal{L} (t) = (\ell_1 (t), \ell_2 (t))$, where
$\ell_1, \ell_2:I \to \Lambda (n)$ are continuous paths in the
Lagrangian--Grassmannian. We say that two paths
$\mathcal{L}, \mathcal{M} \in \mathcal{P} (I)$ are homotopic provided
there exists a family $\mathcal{H}_s$ so that
$\mathcal{H}_0 = \mathcal{L}$, $\mathcal{H}_1 = \mathcal{M}$,
and $\mathcal{H}_s (t)$ is continuous as a map from $(t,s) \in I \times [0,1]$
into $\Lambda (n)$.
The Maslov index has the following properties (see, for example, \cite{HLS} in
the current setting, or Theorem 3.6 in \cite{F} for a more general result).
\medskip
\noindent
{\bf (P1)} (Path Additivity) If $a < b < c$ then
\begin{equation*}
\operatorname{Mas} (\mathcal{L};[a, c]) = \operatorname{Mas} (\mathcal{L};[a, b]) + \operatorname{Mas} (\mathcal{L}; [b, c]).
\end{equation*}
\medskip
\noindent
{\bf (P2)} (Homotopy Invariance) If $\mathcal{L}, \mathcal{M} \in \mathcal{P} (I)$
are homotopic, with $\mathcal{L} (a) = \mathcal{M} (a)$ and
$\mathcal{L} (b) = \mathcal{M} (b)$ (i.e., if $\mathcal{L}, \mathcal{M}$
are homotopic with fixed endpoints) then
\begin{equation*}
\operatorname{Mas} (\mathcal{L};[a, b]) = \operatorname{Mas} (\mathcal{M};[a, b]).
\end{equation*}
\section{Application to Schr\"odinger Operators} \label{schrodinger_section}
For $H$ in (\ref{main}), a value $\lambda \in \mathbb{R}$
is an eigenvalue (see Definition \ref{spectrum}) if and only if there exist coefficient
vectors $\alpha (\lambda), \beta (\lambda) \in \mathbb{R}^n$
and an eigenfunction $\phi (x; \lambda)$ so that
$\mathbf{p} = {\phi \choose \phi'}$ satisfies
\begin{equation*}
\mathbf{X}^- (x; \lambda) \alpha (\lambda)
= \mathbf{p} (x; \lambda) = \mathbf{X}^+ (x; \lambda) \beta (\lambda).
\end{equation*}
This clearly holds if and only if the Lagrangian subspaces
$\ell^- (x; \lambda)$ and $\ell^+ (x; \lambda)$ have
non-trivial intersection. Moreover, the dimension of intersection
will correspond with the geometric multiplicity of $\lambda$
as an eigenvalue. In this way, we can fix any $x \in \mathbb{R}$ and compute
the number of negative eigenvalues of $H$, including multiplicities,
by counting the intersections of $\ell^- (x; \lambda)$ and
$\ell^+ (x; \lambda)$, including multiplicities. Our approach
will be to choose $x = x_{\infty}$ for a sufficiently large
value $x_{\infty} > 0$. Our tool for counting the number and
multiplicity of intersections will be the Maslov index, and
our two Lagrangian subspaces (in the roles
of $\ell_1$ and $\ell_2$ above) will be $\ell^- (x; \lambda)$
and $\ell^+_{\infty} (\lambda) := \ell^+ (x_{\infty}; \lambda)$.
We will denote the Lagrangian frame associated
with $\ell^+_{\infty}$ by
\begin{equation*}
\mathbf{X}_{\infty}^+ (\lambda)
= {X^+_{\infty} (\lambda) \choose Y^+_{\infty} (\lambda)}.
\end{equation*}
\begin{remark} \label{discontinuous}
We will verify in the appendix that while the limit
\begin{equation*}
\ell^-_{+\infty} (\lambda) := \lim_{x \to +\infty} \ell^- (x; \lambda)
\end{equation*}
is well defined for each $\lambda \le 0$, the resulting
limit is not necessarily continuous as a function of $\lambda$. This
is our primary motivation for working with $x_{\infty}$ rather than
with the asymptotic limit.
\end{remark}
Our analysis will be based on computing the Maslov index along a
closed path in the $x$-$\lambda$ plane, determined by sufficiently large
values $x_{\infty}, \lambda_{\infty} > 0$. First, if we fix $\lambda = 0$ and
let $x$ run from $-\infty$ to $x_{\infty}$, we denote the
resulting path $\Gamma_0$ (the {\it right shelf}). Next,
we fix $x = x_{\infty}$ and let $\Gamma_+$ denote a path in
which $\lambda$ decreases from $0$ to $-\lambda_{\infty}$.
Continuing counterclockwise along our path, we denote by
$\Gamma_{\infty}$ the path obtained by fixing
$\lambda = -\lambda_{\infty}$ and letting $x$ run from $x_{\infty}$
to $-\infty$ (the {\it left shelf}). Finally, we close the path
in an asymptotic sense by taking a final path, $\Gamma_-$,
with $\lambda$ running from $-\lambda_{\infty}$ to $0$
(viewed as the asymptotic limit as $x \to - \infty$; we refer
to this as the {\it bottom shelf}). See Figure \ref{box_figure}.
\begin{figure}[ht]
\begin{center}\includegraphics[%
width=12cm,
height=8cm]{maslov_box.pdf}\end{center}
\caption{Maslov Box. \label{box_figure}}
\end{figure}
We recall that we can take the vectors in our frame $\mathbf{X}^- (x; \lambda)$
to be
\begin{equation*}
\scripty{r}_{\,n+j}^{\,-} + \mathbf{E}_{n+j}^-,
\end{equation*}
from which we see that $\ell^- (x;\lambda)$ approaches the asymptotic frame
$\mathbf{R}^- (\lambda)$ as $x \to -\infty$. Introducing the change of
variables
\begin{equation*}
x = \ln (\frac{1+\tau}{1-\tau}) \iff \tau = \frac{e^x - 1}{e^x +1},
\end{equation*}
we see that $\ell^-$ can be viewed as a continuous map on the compact
domain
\begin{equation*}
[-1, \frac{e^{x_\infty} - 1}{e^{x_\infty}+1}] \times [-\lambda_{\infty}, 0].
\end{equation*}
In the setting of (\ref{main}), our evolving Lagrangian subspaces have
frames $\mathbf{X}^- (x; \lambda)$ and $\mathbf{X}^+_{\infty} (\lambda)$,
so that $\tilde{W}$ from (\ref{tildeW}) becomes
\begin{equation} \label{tildeWmain}
\begin{aligned}
\tilde{W} (x; \lambda)
& = - (X^- (x; \lambda) + i Y^- (x; \lambda)) (X^- (x; \lambda) - i Y^- (x; \lambda))^{-1} \\
&\quad \times (X^+_{\infty}(\lambda) - i Y^+_{\infty} (\lambda)) (X^+_{\infty} (\lambda) + i Y^+_{\infty} (\lambda))^{-1}.
\end{aligned}
\end{equation}
Since $\tilde{W} (x; \lambda)$ is unitary, its eigenvalues are confined to
the until circle in $\mathbb{C}$, $S^1$. In the limit as $x \to -\infty$ we
obtain
\begin{equation} \label{tildeWmainInf}
\begin{aligned}
\tilde{W}^- (\lambda) := \lim_{x \to -\infty} \tilde{W} (x; \lambda)
& = - (R^- + i S^- (\lambda)) (R^- - i S^- (\lambda))^{-1} \\
& \quad \times (X^+_{\infty}(\lambda) - i Y^+_{\infty} (\lambda)) (X^+_{\infty} (\lambda) + i Y^+_{\infty} (\lambda))^{-1}.
\end{aligned}
\end{equation}
\subsection{Monotonicity} \label{monotonicity_section}
Our first result for this section
asserts that the eigenvalues of $\tilde{W} (x;\lambda)$ and $\tilde{W}^- (\lambda)$
rotate monotonically as $\lambda$ varies
along $\mathbb{R}$. In order to prove this, we will use a lemma from
\cite{HS}, which we state as follows (see also Theorem V.6.1 in \cite{At}).
\begin{lemma} [\cite{HS}, Lemma 3.11.]
Let $\tilde{W} (\tau)$ be a $C^1$ family of unitary $n \times n$ matrices on
some interval $I$, satisfying a differential equation
$\frac{d}{d\tau} \tilde{W} (\tau) = i \tilde{W} (\tau) \tilde{\Omega} (\tau)$,
where $\tilde{\Omega} (\tau)$ is a continuous, self-adjoint and negative-definite
$n \times n$ matrix.
Then the eigenvalues of $\tilde{W} (\tau)$ move (strictly) monotonically clockwise on the
unit circle as $\tau$ increases.
\label{HS_monotonicity}
\end{lemma}
We are now prepared to state and prove our monotonicity lemma.
\begin{lemma} \label{monotonicity_lemma}
Let $V \in C(\mathbb{R})$ be a real-valued symmetric matrix, and suppose (A1)
and (A2) hold. Then for each fixed $x \in \mathbb{R}$
the eigenvalues of $\tilde{W} (x; \lambda)$ rotate monotonically clockwise
as $\lambda \in (-\infty, \nu_{\min})$ increases. Moreover, the eigenvalues
of $\tilde{W}^{-} (\lambda)$ rotate (strictly) monotonically clockwise as
$\lambda \in (-\infty, \nu_{\min})$ increases.
\end{lemma}
\begin{remark} \label{monotonicity_remark}
The monotoncity described in Lemma \ref{monotonicity_lemma} seems to be
generic for self-adjoint operators in a broad range of settings (see, for
example, \cite{HS}); monotonicity in $x$ is not generic.
\end{remark}
\begin{proof} Following \cite{HS}, we begin by computing
$\frac{\partial \tilde{W}}{\partial \lambda}$, and for this
calculation it's convenient to write
$\tilde{W} (x; \lambda) = - \tilde{W}_1 (x; \lambda) \tilde{W}_2 (\lambda)$,
where
\begin{equation*}
\begin{aligned}
\tilde{W}_1 (x; \lambda) &=
(X^- (x;\lambda) + iY^- (x;\lambda)) (X^- (x;\lambda) - iY^- (x;\lambda))^{-1} \\
\tilde{W}_2 (\lambda) &=
(X^+_{\infty} (\lambda) - i Y^+_{\infty} (\lambda))
(X^+_{\infty} (\lambda) + i Y^+_{\infty} (\lambda))^{-1}.
\end{aligned}
\end{equation*}
For $\tilde{W}_1$, we have (suppressing independent variables for notational brevity)
\begin{equation*}
\begin{aligned}
\frac{\partial \tilde{W}_1}{\partial \lambda} &=
(X^-_{\lambda} + iY^-_{\lambda}) (X^- - iY^-)^{-1}
- (X^- + iY^-) (X^- - iY^-)^{-1} (X^-_{\lambda} - iY^-_{\lambda}) (X^- - iY^-)^{-1} \\
&= (X^-_{\lambda} + iY^-_{\lambda}) (X^- - iY^-)^{-1}
- \tilde{W}_1 (X^-_{\lambda} - iY^-_{\lambda}) (X^- - iY^-)^{-1}.
\end{aligned}
\end{equation*}
If we multiply by $\tilde{W}_1^*$ we find
\begin{equation*}
\begin{aligned}
\tilde{W}_1^* \frac{\partial \tilde{W}_1}{\partial \lambda}
&= ({X^-}^t + i {Y^-}^t)^{-1} ({X^-}^t - i{Y^-}^t) (X^-_{\lambda} + iY^-_{\lambda}) (X^- - iY^-)^{-1} \\
&- (X^-_{\lambda} - iY^-_{\lambda}) (X^- - iY^-)^{-1} \\
&= ({X^-}^t + i {Y^-}^t)^{-1} \Big{\{} ({X^-}^t - i{Y^-}^t) (X^-_{\lambda} + iY^-_{\lambda}) \\
& \quad \quad - ({X^-}^t + i {Y^-}^t) (X^-_{\lambda} - iY^-_{\lambda}) \Big{\}} (X^- - iY^-)^{-1} \\
&= \Big( ({X^-} - i {Y^-})^{-1} \Big)^*
\Big{\{} 2i {X^-}^t Y^-_{\lambda} - 2i {Y^-}^t X^-_{\lambda} \Big{\}} (X^- - iY^-)^{-1}.
\end{aligned}
\end{equation*}
Multiplying back through by $\tilde{W}_1$, we conclude
\begin{equation*}
\frac{\partial \tilde{W}_1}{\partial \lambda} = i \tilde{W}_1 \tilde{\Omega}_1,
\end{equation*}
where
\begin{equation*}
\tilde{\Omega}_1 = \Big( ({X^-} - i {Y^-})^{-1} \Big)^*
\Big{\{} 2 {X^-}^t Y^-_{\lambda} - 2 {Y^-}^t X^-_{\lambda} \Big{\}} \Big( (X^- - iY^-)^{-1} \Big).
\end{equation*}
Likewise, we find that
\begin{equation*}
\frac{\partial \tilde{W}_2}{\partial \lambda} = i \tilde{W}_2 \tilde{\Omega}_2,
\end{equation*}
where
\begin{equation} \label{tildeomega}
\tilde{\Omega}_2 = \Big( (X^+_{\infty} + i Y^+_{\infty})^{-1} \Big)^*
\Big{\{} 2 {Y^+_{\infty}}^t \partial_{\lambda} X^+_{\infty}
- 2 {X^+_{\infty}}^t \partial_{\lambda} Y^+_{\infty} \Big{\}} \Big( (X^+_{\infty} + iY^+_{\infty})^{-1} \Big).
\end{equation}
Combining these observations, we find
\begin{equation} \label{combined1}
\begin{aligned}
\frac{\partial \tilde{W}}{\partial \lambda} &= - \frac{\partial \tilde{W}_1}{\partial \lambda} \tilde{W}_2
- \tilde{W}_1 \frac{\partial \tilde{W}_2}{\partial \lambda}
= - i \tilde{W}_1 \tilde{\Omega}_1 \tilde{W}_2 - i \tilde{W}_1 \tilde{W}_2 \tilde{\Omega}_2 \\
&= - i \tilde{W}_1 \tilde{W}_2 (\tilde{W}_2^* \tilde{\Omega}_1 \tilde{W}_2) - i \tilde{W}_1 \tilde{W}_2 \tilde{\Omega}_2
= i \tilde{W} \tilde{\Omega},
\end{aligned}
\end{equation}
where (recalling that $\tilde{W} = - \tilde{W}_1 \tilde{W}_2$)
\begin{equation*}
\tilde{\Omega} = \tilde{W}_2^* \tilde{\Omega}_1 \tilde{W}_2 + \tilde{\Omega}_2.
\end{equation*}
We see that the behavior of $\frac{\partial \tilde{W}}{\partial \lambda}$ will be determined by
the quantities ${X^-}^t Y^-_{\lambda} - {Y^-}^t X^-_{\lambda}$ and
${Y^+_{\infty}}^t \partial_{\lambda} X^+_{\infty}
- {X^+_{\infty}}^t \partial_{\lambda} Y^+_{\infty}$.
For the former, we differentiate with respect to $x$ to find
\begin{equation*}
\begin{aligned}
\frac{\partial}{\partial x}
\Big{\{} {X^-}^t Y^-_{\lambda} - {Y^-}^t X^-_{\lambda} \Big{\}}
&=
{X^-}^t_x Y^-_{\lambda} + {X^-}^t Y^-_{\lambda x}
- {Y^-}^t_x X^-_{\lambda} - {Y^-}^t X^-_{\lambda x} \\
&=
{Y^-}^t Y^-_{\lambda} + {X^-}^t (V X^- - \lambda X^-)_{\lambda}
- (VX^- - \lambda X^-)^t X^-_{\lambda} - {Y^-}^t Y^-_{\lambda} \\
&= - {X^-}^t X^-,
\end{aligned}
\end{equation*}
where we've used $X^-_x = Y^-$ and $Y^-_x = V(x) X^- - \lambda X^-$.
Integrating from $- \infty$ to $x$, we find
\begin{equation*}
{X^-}^t Y^-_{\lambda} - {Y^-}^t X^-_{\lambda} =
- \int_{-\infty}^x {X^-}^t (y;\lambda) X^- (y;\lambda) dy,
\end{equation*}
from which it is clear that ${X^-}^t Y^-_{\lambda} - {Y^-}^t X^-_{\lambda}$ is
negative definite, which implies that $\tilde{\Omega}_1$ is negative
definite.
Likewise, even though $x_{\infty}$ is fixed, we can differentiate
\begin{equation*}
{Y^+ (x; \lambda)}^t X^+_{\lambda} (x; \lambda)
- {X^+ (x; \lambda)}^t Y^+_{\lambda} (x; \lambda)
\end{equation*}
with respect to
$x$ and evaluate at $x = x_{\infty}$ to find
\begin{equation*}
{Y^+_{\infty}}^t \partial_{\lambda} X^+_{\infty}
- {X^+_{\infty}}^t \partial_{\lambda} Y^+_{\infty}
= - \int_{x_{\infty}}^{+\infty} {X^-}^t (y;\lambda) X^- (y;\lambda) dy,
\end{equation*}
from which it is clear that ${Y^+_{\infty}}^t \partial_{\lambda} X^+_{\infty}
- {X^+_{\infty}}^t \partial_{\lambda} Y^+_{\infty}$ is
negative definite, which implies that $\tilde{\Omega}_2$ is negative
definite.
We conclude that $\tilde{\Omega}$ is negative definite, at which point we can
employ Lemma 3.11 from \cite{HS} to obtain the claim.
For the case of $\tilde{W}^- (\lambda)$, we have
$\tilde{W}^- (\lambda) = - \tilde{W}_1 (\lambda) \tilde{W}_2 (\lambda)$,
where
\begin{equation*}
\tilde{W}_1 (\lambda) =
({R^-} + i{S^-}) (R^- - iS^-)^{-1},
\end{equation*}
and $\tilde{W}_2 (\lambda)$ is as above. Computing as before, we find
\begin{equation*}
\frac{\partial \tilde{W}_1}{\partial \lambda} = i \tilde{W}_1 \tilde{\Omega}_1,
\end{equation*}
where in this case
\begin{equation*}
\tilde{\Omega}_1 = \Big( ({R^-} - i {S^-})^{-1} \Big)^*
\Big{\{} 2 {R^-}^t S^-_{\lambda} - 2 {S^-}^t R^-_{\lambda} \Big{\}} \Big( (R^- - iS^-)^{-1} \Big).
\end{equation*}
Recalling that $R^-_{\lambda} = 0$, we see that the nature of $\tilde{\Omega}_1$ is
determined by ${R^-}^t S^-_{\lambda}$. Recalling that
\begin{equation*}
S^- (\lambda) =
\begin{pmatrix}
\mu_{n+1}^- (\lambda) r_1^- & \mu_{n+2}^- (\lambda) r_{2}^- & \dots & \mu_{2n}^- (\lambda) r_n^-
\end{pmatrix},
\end{equation*}
we have (recalling $\mu_{n+j}^+ (\lambda) = \sqrt{\nu_{j}^- - \lambda}\,$)
\begin{equation*}
S^-_{\lambda} (\lambda) =
- \frac{1}{2} \begin{pmatrix}
\frac{1}{\mu_{n+1}^- (\lambda)} r_1^- & \frac{1}{\mu_{n+2}^- (\lambda)} r_{2}^+ & \dots & \frac{1}{\mu_{2n}^- (\lambda)} r_n^-
\end{pmatrix}.
\end{equation*}
In this way, orthogonality of the $\{r_j^-\}_{j=1}^n$ leads to the relation
\begin{equation} \label{RSlambda2}
{R^-}^t S^-_{\lambda} = -\frac{1}{2}
\begin{pmatrix}
\frac{1}{\mu_{n+1}^- (\lambda)} & 0 & \dots & 0 \\
0& \frac{1}{\mu_{n+2}^- (\lambda)} & \dots & 0 \\
\vdots & \vdots & \vdots & \vdots \\
0 & 0 & \dots & \frac{1}{\mu_{2n}^- (\lambda)}
\end{pmatrix}.
\end{equation}
Since the $\{\mu_{n+j}^-\}_{j=1}^n$ are all positive (for $\lambda < \nu_{\min}$),
we see that $\tilde{\Omega}_1$ is self-adjoint and negative definite.
The matrix $\tilde{W}_2$ is unchanged, so we can draw the same
conclusion about monotonicity.
\end{proof}
\subsection{Lower Bound on the Spectrum of $H$} \label{bound_section}
We have already seen that if the eigenvalues of $V_{\pm}$ are all non-negative
then the essential spectrum of $H$ is bounded below by 0. In fact, it's
bounded below by the smallest eigenvalue of the two matrices $V_{\pm}$. For
the point spectrum, if $\lambda$ is an eigenvalue of $H$ then there
exists a corresponding eigenfunction $\phi (\cdot; \lambda) \in H^1 (\mathbb{R})$.
If we take an $L^2 (\mathbb{R})$ inner product of (\ref{main}) with
$\phi$ we find
\begin{equation*}
\lambda \|\phi\|_2^2 = \|\phi'\|_2^2 + \langle V\phi, \phi\rangle
\ge - C \|\phi\|_2^2,
\end{equation*}
for some contant $C > 0$ taken so that
$|\langle V\phi,\phi\rangle| \le C \|\phi\|_2^2$ for all
$\phi \in H^1 (\mathbb{R})$.
We conclude that $\sigma_{pt} (H) \subset [-C, \infty)$. For example,
$C = \|V\|_{\infty}$ clearly works. In what follows,
we will take a value $\lambda_{\infty}$ sufficiently large, and in
particular we will take $\lambda_{\infty} > C$ (additional requirements
will be added as well, but they can all be accommodated by taking
$\lambda_{\infty}$ larger, so that this initial restriction continues
to hold).
\subsection{The Top Shelf} \label{top_section}
Along the top shelf $\Gamma_+$, the Maslov index counts intersections of the Lagrangian
subspaces $\ell^- (x_{\infty}; \lambda)$ and
$\ell^+_{\infty} (\lambda) = \ell^+ (x_{\infty}; \lambda)$.
Such intersections will correspond with solutions of (\ref{main}) that
decay at both $\pm \infty$, and hence will correspond with eigenvalues.
Moreover, the dimension of these intersections will correspond with
the dimension of the space of solutions that decay at both $\pm \infty$,
and so will correspond with the geometric multiplicity of the eigenvalues.
Finally, we have seen that the eigenvalues of $\tilde{W} (x; \lambda)$
rotate monotonically counterclockwise as $\lambda$ decreases from $0$
to $-\lambda_{\infty}$ (i.e., as $\Gamma_+$ is traversed), and so the
Maslov index on $\Gamma_+$ is a direct count of the crossings, including
multiplicity (with no cancellations arising from crossings in opposite
directions). We conclude that the Maslov index associated with this path
will be a count, including multiplicity, of the negative eigenvalues of
$H$; i.e., of the Morse index. We can express
these considerations as
\begin{equation*}
\operatorname{Mor}(H) = \operatorname{Mas}(\ell^-, \ell^+_{\infty}; \Gamma_+).
\end{equation*}
\subsection{The Bottom Shelf} \label{bottom_section}
For the bottom shelf, we have
\begin{equation} \label{tildeWbottom1}
\begin{aligned}
\tilde{W}^- (\lambda) &= - (R^- + i S^- (\lambda)) (R^- - i S^- (\lambda))^{-1} \\
&\times (X^+_{\infty} (\lambda) - i Y^+_{\infty} (\lambda))
(X^+_{\infty} (\lambda) + i Y^+_{\infty} (\lambda))^{-1}.
\end{aligned}
\end{equation}
By choosing $x_{\infty}$ suitably large, we can ensure that the frame $\mathbf{X}^+_{\infty} (\lambda)$
is as close as we like to the frame $\mathbf{R}^+ (\lambda)$, where we recall
$\mathbf{R}^- = {R^- \choose S^-}$ and $\mathbf{R}^+ = {R^+ \choose S^+}$. (As noted
in Remark \ref{discontinuous} $\ell^-_{+\infty} (\lambda)$ is not necessarily
continuous in $\lambda$, but $\ell^+_{\mathbf{R}} (\lambda)$ certainly is
continuous in $\lambda$.) We will
proceed by analyzing the matrix
\begin{equation} \label{tildeWbottom}
\tilde{\mathcal{W}}^- (\lambda) := - (R^- + i S^- (\lambda)) (R^- - i S^- (\lambda))^{-1} (R^+ - i S^+ (\lambda)) (R^+ + i S^+ (\lambda))^{-1},
\end{equation}
for which we will be able to conclude that for $\lambda < \nu_{\min}$,
$-1$ is never an eigenvalue. By continuity, we will be able to draw conclusions
about $\tilde{W}^- (\lambda)$ as well.
\begin{lemma} \label{bottom_lemma}
For any $\lambda < \nu_{\min}$ the spectrum of
$\tilde{\mathcal{W}}^- (\lambda)$ does not include $-1$.
\end{lemma}
\begin{proof}
We need only show that for any $\lambda < \nu_{\min}$
the $2n$ vectors comprising the columns of $\mathbf{R}^-$ and $\mathbf{R}^+$
are linearly independent. We proceed by induction, first establishing that
any single column of $\mathbf{R}^-$ is linearly independent of the columns
of $\mathbf{R}^+$. Suppose not. Then there is some $j \in \{1, 2, \dots, n\}$,
along with some collection of constants $\{c_k\}_{k=1}^n$ so that
\begin{equation} \label{setup1}
\scripty{r}_{\,n+j}^{\,-} = \sum_{k=1}^n c_k \scripty{r}_{\,k}^{\,+}.
\end{equation}
Recalling the definitions of $\scripty{r}_{\,n+j}^{\,-}$ and $\scripty{r}_{\,k}^{\,+}$,
we have the two equations
\begin{equation*}
\begin{aligned}
r_j^- &= \sum_{k=1}^n c_k r^+_{n+1-k} \\
\mu_{n+j}^- r_j^- &= \sum_{k=1}^n c_k \mu_k^+ r^+_{n+1-k}.
\end{aligned}
\end{equation*}
Multiplying the first of these equations by $\mu_{n+j}^-$,
and subtracting the second equation from the result, we find
\begin{equation*}
0 = \sum_{k=1}^n (\mu_{n+j}^- - \mu_k^+) c_k r^+_{n+1-k}.
\end{equation*}
Since the collection $\{r^+_{n+1-k}\}_{k=1}^n$ is linearly
independent, and since $\mu_{n+j}^- - \mu_k^+ > 0$ for all
$k \in \{1, 2, \dots, n\}$ (for $\lambda < \nu_{\min}$),
we conclude that the constants
$\{c_k\}_{k=1}^n$ must all be zero, but this contradicts
(\ref{setup1}).
For the induction step, suppose that for some $1 \le m < n$,
any $m$ elements of the collection $\{\scripty{r}_{\,n+j}^{\,-}\}_{j=1}^n$
are linearly independent of the set $\{ \scripty{r}_{\,k}^{\,+} \}_{k=1}^n$.
We want to show that any $m+1$ elements of the collection
$\{\scripty{r}_{\,n+j}^{\,-}\}_{j=1}^n$
are linearly independent of the set $\{ \scripty{r}_{\,k}^{\,+} \}_{k=1}^n$.
If not, then by a change of labeling if necessary there exist constants
$\{c_l^-\}_{l=2}^{m+1}$ and $\{c_k^+\}_{k=1}^n$ so that
\begin{equation} \label{setup2}
\scripty{r}_{\,n+1}^{\,-} = \sum_{l=2}^{m+1} c_l^- \scripty{r}_{\,n+l}^{\,-}
+ \sum_{k=1}^n c_k^+ \scripty{r}_{\,k}^{\,+}.
\end{equation}
Again, we have two equations
\begin{equation*}
\begin{aligned}
r_1^- &= \sum_{l=2}^{m+1} c_l^- r_l^- + \sum_{k=1}^n c_k^+ r_{n+1-k}^+ \\
\mu^-_{n+1} r_1^- &= \sum_{l=2}^{m+1} c_l^- \mu^-_{n+l} r_l^-
+ \sum_{k=1}^n c_k^+ \mu_k^+ r_{n+1-k}^+.
\end{aligned}
\end{equation*}
Multiplying the first of these equations by $\mu^-_{n+1}$, and
subtracting the second equation from the result, we obtain the
relation
\begin{equation*}
0 = \sum_{l=2}^{m+1} c_l^- (\mu^-_{n+1} - \mu^-_{n+l}) r_l^-
+ \sum_{k=1}^n c_k^+ (\mu^-_{n+1} - \mu_k^+) r_{n+1-k}^+.
\end{equation*}
By our induction hypothesis, the vectors on the right-hand side
are all linearly independent, and since $\mu^-_{n+1} - \mu_k^+ > 0$
for all $k \in {1, 2, \dots, n}$, we can conclude that
$c_k^+ = 0$ for all $k \in {1, 2, \dots, n}$. (Notice that we
make no claim about the $c_l^-$.) Returning to (\ref{setup2}),
we obtain a contradiction to the linear independence of
the collection $\{\scripty{r}_{\,n+j}^{\,-}\}_{j=1}^n$.
Continuing the induction up to $m = n-1$ gives the claim for
$\lambda < \nu_{\min}$.
\end{proof}
\begin{remark} It is important to note that we do not include the
case $\lambda = \nu_{\min}$ in our lemma, and indeed the lemma
does not generally hold in this case. For example, consider the
case in which $V(x)$ vanishes identically at both $\pm \infty$
(i.e., $V_- = V_+ = 0$). In this case, we can take $R^- = I$,
$S^- = \sqrt{-\lambda} I$, $R^+ = \check{I}$, and
$S^+ = -\sqrt{-\lambda} \check{I}$, where
\begin{equation*}
\check{I} =
\begin{pmatrix}
0 & 0 & \dots & 1 \\
0 & 0 & 1 & 0 \\
\vdots & \vdots & \vdots &\vdots \\
1 & 0 & \dots & 0
\end{pmatrix}.
\end{equation*}
We easily find
\begin{equation*}
\tilde{\mathcal{W}}^- (\lambda) = - \frac{(1+i\sqrt{-\lambda})^2}{(1-i\sqrt{-\lambda})^2} I,
\end{equation*}
and we see explicitly that $\tilde{W}^- (0) = - I$, so that all $n$ eigenvalues
reside at $-1$. Moreover, as $\lambda$ proceeds from 0 toward $- \infty$ the
eigenvalues of $\tilde{W}^- (\lambda)$ remain coalesced, and move monotonically
counterclockwise around $S^1$, returning to $-1$ in the limit as $\lambda \to - \infty$.
In this case, we can conclude that for the path from $0$ to $-\lambda_{\infty}$,
the Maslov index does not increment.
\end{remark}
We immediately obtain the following lemma.
\begin{lemma} \label{bottom_shelf_case1}
Let $V \in C(\mathbb{R})$ be a real-valued symmetric matrix, and suppose (A1)
and (A2) hold. If
\begin{equation*}
\dim (\ell^-_{\mathbf{R}} (0) \cap \ell^+_{\mathbf{R}} (0)) = 0
\end{equation*}
then we can choose $x_{\infty}$ sufficiently large so that we will have
\begin{equation*}
\dim (\ell^-_{\mathbf{R}} (\lambda) \cap \ell^+_{\infty} (\lambda)) = 0
\end{equation*}
for all $\lambda \in [-\lambda_{\infty}, 0]$. It follows that in
this case
\begin{equation*}
\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_-) = 0.
\end{equation*}
Moreover, if
\begin{equation*}
\dim (\ell^-_{\mathbf{R}} (0) \cap \ell^+_{\mathbf{R}} (0)) \ne 0
\end{equation*}
then given any $\lambda_0$ with $0 < \lambda_0 < \lambda_{\infty}$
we can take $x_{\infty}$ sufficiently large so that
\begin{equation*}
\dim (\ell^-_{\mathbf{R}} (\lambda) \cap \ell^+_{\infty} (\lambda)) = 0
\end{equation*}
for all $\lambda \in [-\lambda_{\infty}, -\lambda_0]$.
\end{lemma}
\begin{proof} First, if
\begin{equation*}
\dim (\ell^-_{\mathbf{R}} (0) \cap \ell^+_{\mathbf{R}} (0)) = 0
\end{equation*}
then none of the eigenvalues of $\tilde{\mathcal{W}}^- (0)$ is $-1$,
and so according to Lemma \ref{bottom_lemma}, none of the
eigenvalues of $\tilde{\mathcal{W}}^- (\lambda)$ is $-1$ for
any $\lambda \in [-\lambda_{\infty}, 0]$. In particular,
since the interval $[-\lambda_{\infty}, 0]$ is compact
there exists some $\epsilon > 0$ so that each eigenvalue
$\tilde{\omega} (\lambda)$ of $\tilde{\mathcal{W}}^- (\lambda)$
satisfies
\begin{equation*}
|\tilde{\omega} (\lambda) + 1| > \epsilon
\end{equation*}
for all $\lambda \in [-\lambda_{\infty}, 0]$.
Similarly as above, we can make the change of variables
\begin{equation*}
x_{\infty} = \ln (\frac{1+\tau_{\infty}}{1-\tau_{\infty}}),
\iff
\tau_{\infty} = \frac{e^{x_{\infty}}-1}{e^{x_{\infty}} + 1}.
\end{equation*}
This allows us to view $\tilde{W}^-$ as a continuous function on
the compact domain $(x_{\infty},\lambda) \in [1-\delta, 1] \times [-\lambda_{\infty}, 0]$,
where $\delta > 0$ is small, indicating that $x_{\infty}$ is taken
to be large.
We see that $\tilde{W}^-$ is uniformly continuous and so by choosing
$\tau_{\infty}$ sufficiently close to 1, we can force the eigenvalues
of $\tilde{W}^-$ to be as close to the eigenvalues of
$\tilde{\mathcal{W}}^- (\lambda)$ as we like. We take $\tau_{\infty}$
sufficiently close to 1 so that for each $\lambda \in [-\lambda_{\infty}, 0]$
and each eigenvalue $\tilde{\omega}$ of $\tilde{\mathcal{W}}^- (\lambda)$
there is a corresponding eigenvalue of $\tilde{W}^-$, which we denote $\omega (\lambda)$
so that $|\tilde{\omega} (\lambda) - \omega (\lambda)| < \epsilon/2$. But
then
\begin{equation*}
\begin{aligned}
\epsilon &< |\tilde{\omega} (\lambda)+1| = |\tilde{\omega} (\lambda) - \omega (\lambda)
+ \omega (\lambda) + 1| \\
&\le
|\tilde{\omega} (\lambda) - \omega (\lambda)| + |\omega (\lambda) + 1|
< \frac{\epsilon}{2} + |\omega (\lambda) + 1|,
\end{aligned}
\end{equation*}
from which we conclude that
\begin{equation*}
|\omega (\lambda) + 1| \ge \frac{\epsilon}{2},
\end{equation*}
for all $\lambda \in [-\lambda_{\infty}, 0]$.
For the {\it Moreover} claim, we simply replace $[-\lambda_{\infty}, 0]$
with $[-\lambda_{\infty}, -\lambda_0]$ in the above argument.
\end{proof}
\subsection{The Left Shelf} \label{left_section}
For the left shelf $\Gamma_{\infty}$, we need to understand the Maslov index associated with
$\tilde{W} (x; - \lambda_{\infty})$ (with $\lambda_{\infty}$ sufficiently large)
as $x$ goes from $-\infty$ to $x_{\infty}$ (keeping in mind that the path $\Gamma_{\infty}$
reverses this flow). In order to accomplish this, we follow the approach of
\cite{GZ, ZH} in developing large-$|\lambda|$ estimates on solutions
of (\ref{main}), uniformly in $x$. For $\lambda < 0$, we set
\begin{equation*}
\xi = \sqrt{-\lambda} x; \quad \phi (\xi) = y(x),
\end{equation*}
so that (\ref{main}) becomes
\begin{equation*}
\phi'' (\xi) + \frac{1}{\lambda} V (\frac{\xi}{\sqrt{-\lambda}}) \phi = \phi.
\end{equation*}
Setting $\Phi_1 = \phi$, $\Phi_2 = \phi'$,
and $\Phi = {\Phi_1 \choose \Phi_2} \in \mathbb{R}^{2n}$, we can
express this equation as
\begin{equation*}
\Phi' = \mathbb{A} (\xi; \lambda) \Phi; \quad
\mathbb{A} (\xi; \lambda) =
\begin{pmatrix}
0 & I \\
I - \frac{1}{\lambda} V (\frac{\xi}{\sqrt{-\lambda}}) & 0
\end{pmatrix}.
\end{equation*}
We begin by looking for solutions that decay as $x \to - \infty$ (and so
as $\xi \to - \infty$); i.e., we begin by constructing the frame
$\mathbf{X}^- (x; - \lambda_{\infty})$. It's convenient to write
\begin{equation*}
\mathbb{A} (\xi; \lambda) = \mathbb{A}_- (\lambda) + \mathbb{E}_- (\xi; \lambda),
\end{equation*}
where
\begin{equation*}
\mathbb{A}_- (\lambda) =
\begin{pmatrix}
0 & I \\
I - \frac{1}{\lambda} V_- & 0
\end{pmatrix};
\quad
\mathbb{E}_- (\xi; \lambda) =
\begin{pmatrix}
0 & 0 \\
\frac{1}{\lambda} (V_- - V(\frac{\xi}{\sqrt{-\lambda}})) & 0
\end{pmatrix}.
\end{equation*}
Fix any $M \gg 0$ and note that according to (A1), we have
\begin{equation*}
\begin{aligned}
\int_{-\infty}^M |\mathbb{E}_- (\xi; \lambda)| d\xi
& \le \frac{1}{|\lambda|} \int_{-\infty}^M |V(\frac{\xi}{\sqrt{-\lambda}}) - V_-| d\xi \\
& = \frac{1}{|\lambda|} \int_{-\infty}^{\frac{M}{\sqrt{-\lambda}}}
|V(x) - V_-| \sqrt{-\lambda} dx \le \frac{K}{\sqrt{-\lambda}},
\end{aligned}
\end{equation*}
for some constant $K = K(M)$. Recalling that we are denoting
the eigenvalues of $V_-$ by $\{\nu_j^-\}_{j=1}^n$, we readily check
that the eigenvalues of $\mathbb{A}_- (\lambda)$ can be
expressed as
\begin{equation*}
\begin{aligned}
\hat{\mu}_j^- (\lambda) &= - \sqrt{1 - \frac{\nu^-_{m+1-j}}{\lambda}} = \frac{1}{\sqrt{-\lambda}} \mu_j^- \\
\hat{\mu}_{n+j}^- (\lambda) &= \sqrt{1 - \frac{\nu_j^-}{\lambda}} = \frac{1}{\sqrt{-\lambda}} \mu_{n+j}^-,
\end{aligned}
\end{equation*}
for $j = 1,2, \dots, n$ (ordered, as usual, so that $j < k$ implies
$\hat{\mu}_j^- \le \hat{\mu}_k^-$). In order to select a solution decaying with
rate $\hat{\mu}_{n+j}^-$ (as $\xi \to -\infty$), we look for solutions
of the form
$\Phi (\xi; \lambda) = e^{\hat{\mu}_{n+j}^- (\lambda) \xi} Z (\xi; \lambda)$,
for which $Z$ satisfies
\begin{equation*}
Z' = (\mathbb{A}_- (\lambda) - \hat{\mu}_{m+j} (\lambda) I) Z
+ \mathbb{E}_- (\xi; \lambda) Z.
\end{equation*}
Proceeding similarly as in the proof of Lemma \ref{ODElemma}, we obtain
a collection of solutions
\begin{equation*}
Z_{n+j}^- (\xi; \lambda) =
\hat{\scripty{r}}_{\,n+j}^{\,-}
+ \mathbf{O} (|\lambda|^{-1/2}),
\end{equation*}
which lead to
\begin{equation*}
\Phi_{n+j}^- (\xi; \lambda) = e^{\hat{\mu}_{n+j}^- (\lambda) \xi}
(\hat{\scripty{r}}_{\,n+j}^{\,-}
+ \mathbf{O} (|\lambda|^{-1/2})),
\end{equation*}
where $\hat{\scripty{r}}$ corresponds with $\scripty{r}$, with
$\mu$ is replaced by $\hat{\mu}$.
Returning to original coordinates, we construct the frame $\mathbf{X}^- (x;\lambda)$
out of basis elements
\begin{equation*}
\begin{pmatrix}
y (x) \\ y' (x)
\end{pmatrix}
= e^{\sqrt{-\lambda} \hat{\mu}_{n+j}^- (\lambda) x}
\Big(
\begin{pmatrix}
r_j^- \\ \sqrt{-\lambda} \hat{\mu}_{n+j}^- r_j^-
\end{pmatrix}
+
\begin{pmatrix}
\mathbf{O} (|\lambda|^{-1/2}) \\ \mathbf{O} (1)
\end{pmatrix}
\Big).
\end{equation*}
Recalling that when specifying a frame for $\ell^-$ we can
view the exponential multipliers as expansion coefficients,
we see that we can take as our frame for $\ell^-$ the
matrices
\begin{equation*}
\begin{aligned}
X^- (x; \lambda) &= R^- + \mathbf{O} (|\lambda|^{-1/2}) \\
Y^- (x; \lambda) &= S^- + \mathbf{O} (1),
\end{aligned}
\end{equation*}
where the $\mathbf{O} (\cdot)$ terms are uniform for
$x \in (-\infty, M]$, and we have observed that
$\mu_j^- = \sqrt{-\lambda} \hat{\mu}_j^-$, for
$j = 1, 2, \dots, 2n$. Likewise, we find that for
$-\lambda > 0$ sufficiently large
\begin{equation*}
\begin{aligned}
X^+_{\infty} (\lambda) &= R^+ + \mathbf{O} (|\lambda|^{-1/2}) \\
Y^+_{\infty} (\lambda) &= S^+ + \mathbf{O} (1).
\end{aligned}
\end{equation*}
Turning to $\tilde{W} (x; \lambda)$, we first observe that $S^- (\lambda)^{-1}$ can easily be identified,
using the orthogonality of $R^-$; in particular, the $i$-th row of $S^- (\lambda)^{-1}$
is $\frac{1}{\mu_{n+i}^-} (r_i^-)^t$, which is $\mathbf{O} (|\lambda|^{-1/2})$. In
this way, we see that
\begin{equation*}
\begin{aligned}
X^- (x;\lambda) - i Y^- (x; \lambda)
& = R^- + \mathbf{O} (|\lambda|^{-1/2}) -i S^- (\lambda) + \mathbf{O} (1) \\
&= - i S^- (\lambda) \Big{\{} i S^- (\lambda)^{-1} ( R^- + \mathbf{O} (1)) + I \Big{\}} \\
&= - i S^- (\lambda) (I + \mathbf{O} (|\lambda|^{-1/2})),
\end{aligned}
\end{equation*}
and so
\begin{equation*}
\begin{aligned}
(X^- (x;\lambda) - i Y^- (x;\lambda))^{-1}
&= i (I + \mathbf{O} (|\lambda|^{-1/2}))^{-1} S^- (\lambda)^{-1} \\
&= i (I + \mathbf{O} (|\lambda|^{-1/2})) S^- (\lambda)^{-1},
\end{aligned}
\end{equation*}
by Neumann approximation.
Likewise,
\begin{equation*}
\begin{aligned}
X^- (x;\lambda) + i Y^- (x;\lambda)
&= R^- + \mathbf{O} (|\lambda|^{-1/2}) + i S^- (\lambda) + \mathbf{O} (1) \\
&= (iI + (R^- + \mathbf{O} (1)) S^- (\lambda)^{-1} ) S^- (\lambda) \\
&= (iI + \mathbf{O} (|\lambda|^{-1/2})) S^- (\lambda).
\end{aligned}
\end{equation*}
In this way, we see that
\begin{equation*}
\begin{aligned}
(X^- (x;\lambda) &+ i Y^- (x;\lambda)) (X^- (x;\lambda) - i Y^- (x;\lambda))^{-1} \\
&= (iI + \mathbf{O} (|\lambda|^{-1/2})) S^- (\lambda) i (I + \mathbf{O} (|\lambda|^{-1/2})) S^- (\lambda)^{-1} \\
&= (iI + \mathbf{O} (|\lambda|^{-1/2})) (iI + \mathbf{O} (|\lambda|^{-1/2})) \\
&= -I + \mathbf{O} (|\lambda|^{-1/2}).
\end{aligned}
\end{equation*}
Proceeding similarly for $\mathbf{X}_{\infty} (\lambda)$, we have
\begin{equation*}
(X^+_{\infty} (\lambda) + i Y^+_{\infty} (\lambda) (X^+_{\infty} (\lambda) - i Y^+_{\infty} (\lambda))^{-1}
= -I + \mathbf{O} (|\lambda|^{-1/2}),
\end{equation*}
and so
\begin{equation*}
\tilde{W} (x; \lambda) = -I + \mathbf{O} (|\lambda|^{-1/2})
\end{equation*}
uniformly in $x$.
We see that for $\lambda_{\infty}$ sufficiently large the eigenvalues
of $\tilde{W} (x; - \lambda_{\infty})$ are near $-1$
uniformly for $x \in (-\infty, M]$.
Turning to the behavior of $\tilde{W} (x; \lambda)$ as
$x$ tends to $+\infty$ (i.e., for $x \ge M$),
we recall from Section \ref{bound_section} that if $\lambda_{\infty}$
is large enough then $- \lambda_{\infty}$ will not be an
eigenvalue of $H$. This means the evolving Lagrangian
subspace $\ell^-$ cannot intersect the space of solutions
asymptotically decaying as $x \to +\infty$, and so the
frame $X^- (x;\lambda)$ must be comprised of solutions that
grow as $x$ tends to $+\infty$. The construction of these
growing solutions is almost identical to our construction of
the decaying solutions $\Phi_j^-$, and we'll be brief.
In this case, it's convenient to write
\begin{equation*}
\mathbb{A} (\xi; \lambda) = \mathbb{A}_+ (\lambda) + \mathbb{E}_+ (\xi; \lambda),
\end{equation*}
where
\begin{equation*}
\mathbb{A}_+ (\lambda) =
\begin{pmatrix}
0 & I \\
I - \frac{1}{\lambda} V_+ & 0
\end{pmatrix};
\quad
\mathbb{E}_+ (\xi; \lambda) =
\begin{pmatrix}
0 & 0 \\
\frac{1}{\lambda} (V_+ - V(\frac{\xi}{\sqrt{-\lambda}})) & 0
\end{pmatrix}.
\end{equation*}
The eigenvalues of $\mathbb{A}_+ (\lambda)$ can be
expressed as
\begin{equation*}
\begin{aligned}
\hat{\mu}_j^+ (\lambda) &= - \sqrt{1 - \frac{\nu^+_{m+1-j}}{\lambda}}
= \frac{1}{\sqrt{-\lambda}} \mu_j^+ \\
\hat{\mu}_{n+j}^+ (\lambda) &= \sqrt{1 - \frac{\nu_j^+}{\lambda}}
= \frac{1}{\sqrt{-\lambda}} \mu_{n+j}^+,
\end{aligned}
\end{equation*}
for $j = 1,2, \dots, n$ (ordered, as usual, so that $j < k$ implies
$\hat{\mu}_j^+ \le \hat{\mu}_k^+$). In order to select a solution growing with
rate $\hat{\mu}_{n+j}^+$ (as $\xi \to +\infty$), we look for solutions
of the form
$\Phi (\xi; \lambda) = e^{\hat{\mu}_{n+j}^+ (\lambda) \xi} Z (\xi; \lambda)$,
for which $Z$ satisfies
\begin{equation*}
Z' = (\mathbb{A}_+ (\lambda) - \hat{\mu}_{n+j}^+ (\lambda) I) Z
+ \mathbb{E}_+ (\xi; \lambda) Z.
\end{equation*}
Proceeding as with the frame of solutions that decay as $x \to -\infty$,
we find that for $M$ sufficiently large (so that asymptotically decaying
solutions become negligible), we can take as our frame for $\ell^-$
\begin{equation*}
\begin{aligned}
X^+ (x; \lambda) &= R^+ + \mathbf{O} (|\lambda|^{-1/2}) \\
Y^+ (x; \lambda) &= \tilde{S}^+ + \mathbf{O} (1),
\end{aligned}
\end{equation*}
where
\begin{equation*}
{\tilde{S}}^+ =
\begin{pmatrix}
\mu_{n+1}^+ r_n^+ & \mu_{n+2}^+ r_{n-1}^+ & \dots & \mu_{2n}^+ r_1^+
\end{pmatrix},
\end{equation*}
and the $\mathbf{O} (\cdot)$ terms are uniform for $x \in [M, \infty)$.
Proceeding now almost exactly as we did for the interval $(-\infty, M]$
we find that for $\lambda_{\infty}$ sufficiently large the eigenvalues
of $\tilde{W} (x; - \lambda_{\infty})$ are near $-1$
uniformly for $x \in [M, \infty)$.
We summarize these considerations in a lemma.
\begin{lemma} Let $V \in C(\mathbb{R})$ be a real-valued symmetric matrix,
and suppose (A1) and (A2) hold. Then given any $\epsilon > 0$ there
exists $\lambda_{\infty} > 0$ sufficiently large so that for all
$x \in \mathbb{R}$ and for any eigenvalue
$\omega (x; -\lambda_{\infty})$ of $\tilde{W} (x; -\lambda_{\infty})$
we have
\begin{equation*}
|\omega (x; -\lambda_{\infty}) + 1| < \epsilon.
\end{equation*}
\end{lemma}
\begin{remark} We note that it would be insufficient to simply take
$M = x_{\infty}$ in our argument. This is because our overall argument
is structured in such a way that we choose $\lambda_{\infty}$ first,
and then choose $x_{\infty}$ sufficiently large, based on this value.
(This if for the bottom shelf argument.) But $\lambda_{\infty}$ must
be chosen based on $M$, so $M$ should not depend on the value of
$x_{\infty}$.
\end{remark}
We now make the following claim.
\begin{lemma} \label{left_lemma}
Let $V \in C(\mathbb{R})$ be a real-valued symmetric matrix, and suppose (A1)
and (A2) hold. Then given any $M > 0$ there exists $\lambda_{\infty} > 0$
sufficiently large so that
\begin{equation*}
\operatorname{Mas} (\ell^-, \ell_{\infty}^+; \Gamma_{\infty}) = 0,
\end{equation*}
for any $x_{\infty} > M$.
\end{lemma}
\begin{proof} We begin by observing that by taking $\lambda_{\infty}$ sufficiently
large, we can ensure that for all $x_{\infty} > M$ the eigenvalues of
$\tilde{W} (x_{\infty}; -\lambda_{\infty})$
are all near $-1$. To make this precise, given any $\epsilon > 0$ we can take
$\lambda_{\infty}$ sufficiently large so that the eigenvalues of
$\tilde{W} (x_{\infty}; -\lambda_{\infty})$ are confined to the arc
$\mathcal{A}_{\epsilon} = \{e^{i \theta}: |\theta - \pi| < \epsilon\}$.
Moreover, we know from Lemma \ref{monotonicity_lemma} that as $\lambda$
decreases toward $-\lambda_{\infty}$ the eigenvalues of $\tilde{W} (x_{\infty}; \lambda)$
will monotonically rotate in the counterclockwise direction, and so the
eigenvalues of $\tilde{W} (x_{\infty}; -\lambda_{\infty})$ will in fact be confined to
the arc $\mathcal{A}_{\epsilon}^+ = \{e^{i \theta}: -\epsilon < \theta - \pi < 0\}$.
(See Figure \ref{Ape}; we emphasize that none of the eigenvalues can cross $-1$, because such a crossing
would correspond with an eigenvalue of $H$, and we have assumed $\lambda_{\infty}$
is large enough so that there are no eigenvalues for $\lambda \le - \lambda_{\infty}$.)
Likewise, by the same monotonicity argument, we see that the eigenvalues of
$\tilde{W}^- (-\lambda_{\infty})$ are also confined to $\mathcal{A}_{\epsilon}^+$.
\begin{figure}[ht]
\begin{center}\includegraphics[%
width=8cm,
height=8cm]{Maslov_S1b.pdf}\end{center}
\caption{Eigenvalues confined to $A^+_{\epsilon}$. \label{Ape}}
\end{figure}
Turning now to the flow of eigenvalues as $x$ proceeds from $x_{\infty}$ to $-\infty$
(i.e., along the reverse direction of $\Gamma_{\infty}$), we note by uniformity
of our large-$|\lambda|$
estimates that we can take $\lambda_{\infty}$ large enough so that the eigenvalues
of $\tilde{W}(x; -\lambda_{\infty})$ are confined to $\mathcal{A}_{\epsilon}$
(not necessarily $\mathcal{A}_{\epsilon}^+$) for all $x \in \mathbb{R}$.
Combining these observations, we conclude that the eigenvalues of
$\tilde{W}(x; -\lambda_{\infty})$ must begin and end in $\mathcal{A}_{\epsilon}^+$,
without completing a loop of $S^1$, and consequently the Maslov index along the
entirety of $\Gamma_{\infty}$ must be 0.
\end{proof}
\subsection{Proof of Theorem \ref{main_theorem}} \label{proof_section}
Let $\Gamma$ denote the contour obtained by proceeding counterclockwise
along the paths $\Gamma_0$, $\Gamma_+$, $\Gamma_{\infty}$, $\Gamma_-$.
By the catenation property of the Maslov index, we have
\begin{equation*}
\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma) =
\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_0) + \operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_+)
+ \operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_{\infty}) + \operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_-).
\end{equation*}
Moreover, by the homotopy property, and by noting that $\Gamma$ is homotopic to
an arbitrarily small cycle attached to any point of $\Gamma$, we can conclude
that $\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma) = 0$. Since
$\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_{\infty}) = 0$, and
$\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_+) = \operatorname{Mor} (H)$, it follows immediately
that
\begin{equation} \label{loop}
\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_0) + \operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_{-}) + \operatorname{Mor} (H)
= 0.
\end{equation}
We will complete the proof with the following claim.
\begin{claim} Under the assumptions of Theorem \ref{main_theorem},
\begin{equation*}
\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_0) + \operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_{-})
= \operatorname{Mas}(\ell^-, \ell^+_{\mathbf{R}}; \bar{\Gamma}_0).
\end{equation*}
\end{claim}
\begin{proof} First, consider the case
\begin{equation*}
\kappa = \dim (\ell^-_{\mathbf{R}} (0) \cap \ell^+_{\mathbf{R}} (0)) = 0,
\end{equation*}
where we introduce the notation $\kappa$ for notational convenience.
In this case, we know from Lemma \ref{bottom_shelf_case1} that for
$x_{\infty}$ sufficiently large we will have
$\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_-) = 0$. It remains to show that
\begin{equation*}
\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_0)
= \operatorname{Mas}(\ell^-, \ell^+_{\mathbf{R}}; \bar{\Gamma}_0).
\end{equation*}
As usual, let $\tilde{W} (x; \lambda)$ denote the unitary matrix
(\ref{tildeWmain}) (which we recall depends on $x_{\infty}$),
and let $\tilde{\mathcal{W}} (x; \lambda)$ denote the unitary
matrix
\begin{equation}
\begin{aligned}
\tilde{\mathcal{W}} (x; \lambda)
&= - (X^- (x; \lambda) + i Y^- (x; \lambda)) (X^- (x; \lambda) - i Y^- (x; \lambda))^{-1} \\
&\times (R^+ - i S^+ (\lambda))
(R^+ + i S^+ (\lambda))^{-1}.
\end{aligned}
\end{equation}
I.e., $\tilde{W} (x; \lambda)$ is the unitary matrix used in the
calculation of $\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_0)$ and
$\tilde{\mathcal{W}} (x; \lambda)$ is the unitary matrix used
in the calculation of $\operatorname{Mas}(\ell^-, \ell^+_{\mathbf{R}}; \bar{\Gamma}_0)$.
Likewise, set
\begin{equation*}
\begin{aligned}
\tilde{W}^- (\lambda) &= \lim_{x \to -\infty} \tilde{W} (x; \lambda) \\
\tilde{\mathcal{W}}^- (\lambda) &= \lim_{x \to -\infty} \tilde{\mathcal{W}} (x; \lambda),
\end{aligned}
\end{equation*}
both of which are well defined. (Notice that while the matrix
$\tilde{\mathcal{W}} (x; \lambda)$ has not previously appeared, the other matrices
here, including $\tilde{\mathcal{W}}^- (\lambda)$, are the same as
before.)
By taking $x_{\infty}$ sufficiently large we can ensure that
the spectrum of $\tilde{W}^- (0)$ is arbitrarily close to the spectrum
of $\tilde{\mathcal{W}}^- (0)$ in the following sense: given
any $\epsilon > 0$ we can take $x_{\infty}$ sufficiently large so
that for any $\omega \in \sigma (\tilde{W}^- (0))$ there exists
$\tilde{\omega} \in \sigma (\tilde{\mathcal{W}}^- (0))$ so
that $|\omega - \tilde{\omega}| < \epsilon$.
Turning to the other end of our contours, we first take the case
$\nu_{\min} > 0$ so that $\lambda = 0$ is not embedded in
essential spectrum. In this case, $\tilde{W} (x_{\infty};0)$
will have $-1$ as an eigenvalue if and only if $\lambda = 0$
is an eigenvalue of $H$, and the multiplicity of $-1$ as an
eigenvalue of $\tilde{W} (x_{\infty};0)$ will correspond with
the geometric multiplicity of $\lambda = 0$ as an eigenvalue of
$H$. For $\tilde{\mathcal{W}} (x; 0)$ set
\begin{equation*}
\tilde{\mathcal{W}}^+ (0) = \lim_{x \to \infty} \tilde{\mathcal{W}} (x; 0),
\end{equation*}
which is well defined by our construction in the appendix.
As with $\tilde{W} (x_{\infty};0)$, $\tilde{\mathcal{W}}^+ (0)$
will have $-1$ as an eigenvalue if and only if $\lambda = 0$
is an eigenvalue of $H$, and the multiplicity of $-1$ as an
eigenvalue of $\tilde{\mathcal{W}}^+ (0)$ will correspond with
the geometric multiplicity of $\lambda = 0$ as an eigenvalue of
$H$. By choosing $x_{\infty}$ sufficiently large, we can ensure
that the eigenvalues of $\tilde{W} (x_{\infty};0)$ are arbitrarily
close to the eigenvalues of $\tilde{\mathcal{W}}^+ (0)$. I.e.,
$-1$ repeats as an eigenvalue the same number of times for these
two matrices, and the eigenvalues aside from $-1$ can be made
arbitrarily close.
We see that the path of matrices $\tilde{W} (x;0)$, as $x$ runs
from $-\infty$ to $x_{\infty}$ can be viewed as a small perturbation
from the path of matrices $\tilde{\mathcal{W}} (x;0)$, as
$x$ runs from $-\infty$ to $+\infty$. In order to clarify this,
we recall that by using the change of variables (\ref{change}) we
can specify our path of Lagrangian subspaces on the compact interval
$[-1, 1]$. Likewise, the interval $(-\infty, x_{\infty}]$ compactifies
to $[-1, (e^{x_{\infty}} - 1)/(e^{x_{\infty}} + 1)]$. For this
latter interval, we can make the further change of variables
\begin{equation*}
\xi = \frac{2}{1+r_{\infty}} \tau + \frac{1 - r_{\infty}}{1 + r_{\infty}},
\end{equation*}
where $r_{\infty} = (e^{x_{\infty}} - 1)/(e^{x_{\infty}} + 1)$, so
that $\tilde{W} (x;0)$ and $\tilde{\mathcal{W}} (x; 0)$ can both be specfied on
the interval $[-1, 1]$. Finally, we see that
\begin{equation*}
|\xi - \tau| = (1+\tau) \frac{1-r_{\infty}}{1+r_{\infty}},
\end{equation*}
so by choosing $x_{\infty}$ sufficiently large (and hence $r_{\infty}$
sufficiently close to $1$), we can take the values of $\xi$ and $\tau$
as close as we like. By uniform continuity the eigenvalues of the
adjusted path will be arbitrarily close to those of the original
path.
Since the endstates associated with these paths are arbitrarily close,
and since the eigenvalues of
one path end at $-1$ if and only if the eigenvalues of the other path
do, the homotopy invariance argument in \cite{HLS} can be employed to show
that the spectral flow must be the same along each of these paths,
and this establishes the claim.
In the event that $\nu_{\min} = 0$ so that $\lambda = 0$ is
embedded in essential spectrum, it may be the case that $\tilde{W} (x_{\infty};0)$
has $-1$ as an eigenvalue even if $\lambda = 0$ is not an eigenvalue.
More generally, the multiplicity of $-1$ as an eigenvalue of
$\tilde{W} (x_{\infty};0)$ may not correspond with the geometric
multiplicity of $\lambda = 0$ as an eigenvalue of $H$. Rather,
in such cases the multiplicity of $-1$ as an eigenvalue of
$\tilde{W} (x_{\infty};0)$ will correspond with the dimension of the
intersection of the space of solutions
that are obtained obtained as $\lambda \to 0^-$ limits of solutions
that decay at $-\infty$ and the space of solutions that are obtained
as $\lambda \to 0^-$ limits of solutions that decay at $+\infty$.
(Here, we are keeping in mind that as $\lambda \to 0^-$ if a decaying
solution ceases to decay then there will be a corresponding growing
solution that ceases to grow.) Once again, $\tilde{\mathcal{W}}^+ (0)$
will have $-1$ as an eigenvalue if and only if $\tilde{W} (x_{\infty};0)$
does, and we will be able to apply the same argument as discussed above
to establish the claim.
We now turn to the case
\begin{equation*}
\kappa = \dim (\ell^-_{\mathbf{R}} (0) \cap \ell^+_{\mathbf{R}} (0)) \ne 0,
\end{equation*}
and as with the case $\kappa = 0$ we begin by assuming $\nu_{\min} > 0$.
The matrix $\tilde{\mathcal{W}}^- (0)$ will have $-1$ as an eigenvalue
with multiplicity $\kappa$. By monotonicity in $\lambda$, and Lemma
\ref{bottom_shelf_case1} we know that for any $\lambda < 0$ the eigenvalues
will have rotated away from $-1$ in the counterclockwise direction. In
particular, given any $\epsilon > 0$ we can find $\lambda_0 > 0$ sufficiently
small so that the eigenvalues of $\tilde{\mathcal{W}}^- (-\lambda_0)$ are on the arc
\begin{equation*}
A_{\epsilon}^- = \{e^{i \theta}: \pi < \theta < \pi + \epsilon\},
\end{equation*}
while no other eigenvalues of $\tilde{\mathcal{W}}^- (0)$ are on
the arc $A_{\epsilon}^-$.
Recalling that $\tilde{W}^- (0)$ can be viewed as a small perturbation
of $\tilde{\mathcal{W}}^- (0)$, we see that for $x_{\infty}$ sufficiently
large there will be a cluster of $\kappa$ eigenvalues of $\tilde{W}^- (0)$ near
$-1$, on an arc $A_{\tilde{\epsilon}}$, where $\tilde{\epsilon}$ can be
made as small as we like by our choice of $x_{\infty}$. Moreover, by monotonicity
in $\lambda$, we can choose $\lambda_0 > 0$ sufficiently small (perhaps smaller
than the previous choice) so that the corresponding eigenvalues of
$\tilde{W}^- (-\lambda_0)$ are confined to the arc $A_{\tilde{\epsilon}}^-$.
See Figure \ref{eigs_close}.
\begin{figure}[ht]
\begin{center}\includegraphics[%
width=8cm,
height=8cm]{Maslov_S1c.pdf}\end{center}
\caption{The eigenvalues of $\tilde{W}^- (-\lambda_0)$ and $\tilde{\mathcal{W}} (-\lambda_0)$. \label{eigs_close}}
\end{figure}
At this point, we can consider the spectral flow of the family of
matrices $\tilde{W} (x;\lambda)$ along the path obtained by first
fixing $x = -\infty$ and letting $\lambda$ run from $-\lambda_0$
to 0, and then fixing $\lambda = 0$ and letting $x$ run from
$-\infty$ to $x_{\infty}$. We denote this path $\Gamma_1$.
Correspondingly, we can consider the spectral flow of the family of
matrices $\tilde{\mathcal{W}} (x;\lambda)$ along the path obtained
by first fixing $x = -\infty$ and letting $\lambda$ run from
$-\lambda_0$ to 0, and then fixing $\lambda = 0$
and letting $x$ run from $-\infty$ to $+\infty$. We denote this
path $\Gamma_2$.
We are now in almost precisely the same case as when $\kappa = 0$,
and again we can use an argument similar to the homotopy argument of
\cite{HLS} to verify that these two spectral flows will give equivalent
values. This means, of course, that the Maslov indices along
these paths for the respective pairs $(\ell^-, \ell^+_{\infty})$
(for $\Gamma_1$)
and $(\ell^-, \ell^+_{\mathbf{R}})$ (for $\Gamma_2$) will be equivalent.
I.e.,
\begin{equation*}
\operatorname{Mas} (\ell^-, \ell^+_{\mathbf{R}}; \Gamma_1)
= \operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_2).
\end{equation*}
However, by Lemma \ref{bottom_shelf_case1} we know that the pair
$(\ell^-, \ell^+_{\mathbf{R}})$ has no intersections along
$\Gamma_-$ (except at $\lambda = 0$). By monotonicity, as
$\lambda \to 0^-$ the eigenvalues of $\tilde{\mathcal{W}}^- (\lambda)$
will rotate in the clockwise direction, so those rotating to $-1$
will not increment the Maslov index. We conclude that
\begin{equation*}
\operatorname{Mas} (\ell^-, \ell^+_{\mathbf{R}}; \Gamma_2) =
\operatorname{Mas} (\ell^-, \ell^+_{\mathbf{R}}; \bar{\Gamma}_0),
\end{equation*}
where we recall from the introduction that $\bar{\Gamma}_0$ is the contour
obtained by fixing $\lambda = 0$ and letting $x$ run from $-\infty$
to $+\infty$.
Likewise, according to Lemma \ref{bottom_shelf_case1} we can take $x_{\infty}$
sufficiently large so that $\tilde{W}^- (\lambda)$ does not have $-1$ as
an eigenvalue for any $\lambda \in [-\lambda_{\infty}, - \lambda_0]$. This
implies that
\begin{equation*}
\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_1) =
\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_-) + \operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_0).
\end{equation*}
Combining, we find
\begin{equation*}
\operatorname{Mas} (\ell^-, \ell^+_{\mathbf{R}}; \bar{\Gamma}_0) =
\operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_-) + \operatorname{Mas} (\ell^-, \ell^+_{\infty}; \Gamma_0),
\end{equation*}
which is the claim. The case $\nu_{\min} = 0$ for $\kappa \ne 0$ follows
similarly as for $\kappa = 0$.
\end{proof}
Upon combining the claim with (\ref{loop}), we obtain Theorem \ref{main_theorem}.
\section{Equations with Constant Convection} \label{convection_section}
For a traveling wave solution $\bar{u} (x - st)$ to the Allen-Cahn equation
\begin{equation} \label{AC}
u_t + F'(u) = u_{xx},
\end{equation}
it's convenient to switch to a shifted coordinate frame in which the wave
is a stationary solution $\bar{u} (x)$ for the equation
\begin{equation} \label{SAC}
u_t - su_x + F'(u) = u_{xx}.
\end{equation}
In this case, linearization about the wave leads to an eigenvalue problem
\begin{equation} \label{Es}
H_s y := - y'' + s y' + V(x) y = \lambda y,
\end{equation}
where $V(x) = F'' (\bar{u} (x))$.
Our goal in this section is to show that our development for
(\ref{main}) can be extended to the case (\ref{Es}) in a
straightforward manner. For this discussion, which is adapted from
\cite{BJ1995}, we take any real number $s \ne 0$, and we continue
to let assumptions (A1) and (A2) hold.
The main issues we need to address are as follows: (1) we need to show
that the point spectrum for $H_s$ is real-valued; (2) we need to show
that the $n$-dimensional subspaces associated with $H_s$ are Lagrangian;
and (3) we need to show that the eigenvalues of the associated
unitary matrix $\tilde{W} (x; \lambda)$ rotate monotonically as
$\lambda$ increases (or decreases). Once these items have been verified,
the remainder of our analysis carries over directly to the case
$s \ne 0$.
\subsection{Essential Spectrum} \label{essential_spectrum_subsection}
As for the case $s = 0$ the essential spectrum for $s \ne 0$ can be
identified from the asymptotic equations
\begin{equation} \label{Asym_Es}
- y'' + s y' + V_{\pm} y = \lambda y.
\end{equation}
Precisely, the essential spectrum will correspond with
values of $\lambda$ for which (\ref{Asym_Es}) admits a solution of the form
$y(x) = e^{ikx} r$ for some constant non-zero vector $r \in \mathbb{C}^n$. Upon
substitution of this ansatz into (\ref{Es}) we obtain the relations
\begin{equation*}
(k^2 I + iskI + V_{\pm}) r = \lambda r.
\end{equation*}
We take a $\mathbb{C}^n$ inner product with $r$ to see that
\begin{equation*}
\lambda (k) |r|^2 = (V_{\pm} r, r)_{\mathbb{C}^n} + (i s k + k^2) |r|^2,
\end{equation*}
or equivalently
\begin{equation*}
\lambda (k) = \frac{(V_{\pm} r, r)_{\mathbb{C}^n}}{|r|^2} + i s k + k^2.
\end{equation*}
We conclude that the essential spectrum is confined on and to the right of
parabolas opening into the real complex half-plane, described by the
relations
\begin{equation*}
\text{Re } \lambda = \frac{(V_{\pm} r, r)_{\mathbb{C}^n}}{|r|^2} + \frac{1}{s^2} (\text{Im }\lambda)^2.
\end{equation*}
For notational convenience, we denote by $\Omega$ this region in $\mathbb{C}$
on or two the right of these parabolas.
\subsection{In $\mathbb{C} \backslash \Omega$ the Point Spectrum of $H_s$ is Real-Valued} \label{real_section}
For any $\lambda \in \mathbb{C} \backslash \Omega$, we can look for ODE solutions with asymptotic
behavior $y(x) = e^{\mu x} r$. Upon substitution into (\ref{Asym_Es}) we obtain
the eigenvalue problem
\begin{equation*}
(-\mu^2 + s \mu + V_{\pm} - \lambda) r = 0.
\end{equation*}
As in Section \ref{ODEsection} we denote the eigenvalues of $V_{\pm}$
by $\{\nu_j^{\pm}\}_{j=1}^n$, with associated eigenvectors $\{r_j^{\pm}\}_{j=1}^n$.
We see that the possible growth/decay rates $\mu$ will
satisfy
\begin{equation*}
\mu^2 -s \mu + \lambda = \nu_j^{\pm}
\implies
\mu = \frac{s \pm \sqrt{s^2 - 4(\lambda - \nu_j^{\pm})}}{2}.
\end{equation*}
We label the $2n$ growth/decay rates similarly as in
Section \ref{ODEsection}
\begin{equation*}
\begin{aligned}
\mu_j^{\pm} (\lambda) &= \frac{s - \sqrt{s^2 - 4(\lambda - \nu_{n+1-j}^{\pm})}}{2} \\
\mu_{n+j}^{\pm} (\lambda) &= \frac{s + \sqrt{s^2 - 4(\lambda - \nu_{j}^{\pm})}}{2},
\end{aligned}
\end{equation*}
for $j = 1, 2, \dots, n$.
Now, suppose $\lambda \in \mathbb{C} \backslash \Omega$ is is an eigenvalue for
$H_s$. For this fixed value, we can obtain asymptotic ODE estimates on solutions of
(\ref{Es}) with precisely the same form as those described in Lemma \ref{ODElemma}
(keeping in mind that the specifications of $\{\mu_j^{\pm}\}_{j=1}^{2n}$ are
different).
Letting $\psi (x; \lambda)$ denote the eigenfunction associated with $\lambda$,
we conclude that $\psi(x; \lambda)$ can be expressed both as a linear combination
of the solutions that decay as $x \to -\infty$ (i.e., those associated with rates
$\{\mu_{n+j}^-\}_{j=1}^n$) and as a linear combination of the solutions that decay
as $x \to +\infty$ (i.e., those associated with rates $\{\mu_{j}^+\}_{j=1}^n$).
Keeping in mind that we are in the case $s \ne 0$, we make the change of variable
$\phi(x) = e^{-\frac{s}{2} x} y(x)$, for which a direct calculation yields
\begin{equation*}
\mathcal{H}_s \phi := e^{-\frac{s}{2} x} H_s e^{\frac{s}{2} x} \phi
= - \phi'' + (\frac{s^2}{4} + V(x)) \phi = \lambda \phi.
\end{equation*}
Moreover, if $y(x)$ is a solution of $H_s y = \lambda y$ that decays with
rate $\mu_{n+j}^- (\lambda)$ as $x \to -\infty$ then the corresponding
$\phi(x)$ will decay as $x \to -\infty$ with rate
\begin{equation}
-\frac{s}{2} + \frac{s + \sqrt{s^2 - 4(\lambda - \nu_{j}^{\pm})}}{2}
=
\frac{1}{2} \sqrt{s^2 - 4(\lambda - \nu_{j}^{\pm})} > 0,
\end{equation}
and likewise if $y(x)$ is a solution of $H_s y = \lambda y$ that decays with
rate $\mu_{j}^+ (\lambda)$ as $x \to +\infty$ then the corresponding
$\phi(x)$ will decay as $x \to +\infty$ with rate
\begin{equation}
-\frac{s}{2} + \frac{s - \sqrt{s^2 - 4(\lambda - \nu_{j}^{\pm})}}{2}
=
- \frac{1}{2} \sqrt{s^2 - 4(\lambda - \nu_{j}^{\pm})} < 0.
\end{equation}
In this way we see that $\varphi(x; \lambda) = e^{-\frac{s}{2} x} \psi(x; \lambda)$
is an eigenfunction for $\mathcal{H}_s$, associated with the eigenvalue
$\lambda$. But $\mathcal{H}_s$ is self-adjoint, and so its spectrum is
confined to $\mathbb{R}$. We conclude that $\lambda \in \mathbb{R}$.
Finally, we observe that although the real value $\lambda = \nu_{\min}$ is
embedded in the essential spectrum, it is already in $\mathbb{R}$. In this
way, we conclude that any eigenvalue $\lambda$ of $H_s$ with
$\text{Re }\lambda \le \nu_{\min}$ must be real-valued.
\subsection{Bound on the Point Spectrum of $H_s$} \label{point_spectrum_subsection}
Suppose $\lambda \in \mathbb{R}$ is an eigenvalue of $H_s$ with associated
eigenvector $\psi (x; \lambda)$. Taking an $L^2 (\mathbb{R})$ inner product
of $H_s \psi = \lambda \psi$ with $\psi$ we obtain the relation
\begin{equation*}
\|\psi'\|^2 + s \langle \psi', \psi \rangle + \langle V \psi, \psi \rangle
= \lambda \|\psi\|^2.
\end{equation*}
We see that
\begin{equation*}
\begin{aligned}
\lambda \|\psi\|^2 &\ge \|\psi'\|^2 - |s| \|\psi'\| \|\psi\| + \langle V \psi, \psi \rangle \\
&\ge \|\psi'\|^2 - |s| (\frac{\epsilon}{2} \|\psi'\|^2 + \frac{1}{2\epsilon} \|\psi\|^2)
- \|V\|_{\infty} \|\psi\|^2 \\
&\ge
- (\frac{1}{2\epsilon} + \|V\|_{\infty}) \|\psi\|^2,
\end{aligned}
\end{equation*}
from which we conclude that $\lambda$ is bounded below. (In this calculation,
$\epsilon > 0$ has been taken sufficiently small.)
\subsection{The Spaces $\ell^- (x; \lambda)$ and $\ell^-_{\mathbf{R}} (\lambda)$ are Lagrangian}
\label{lagrangian_section}
Since $\sigma_p (H) \subset \mathbb{R}$, we can focus on $\lambda \in \mathbb{R}$,
in which case the growth/decay rates $\{\mu_j^{\pm}\}_{j=1}^{2n}$ remain ordered
as $\lambda$ varies. In light of this, the estimates of Lemma \ref{ODElemma} remain
valid precisely as stated, with our revised definitions of these rates. The Lagrangian
property for $\mathbf{R}^- = {R^- \choose S^-}$ can be verified precisely as before,
but for $\mathbf{X}^- (x; \lambda) = {X^- (x; \lambda) \choose Y^- (x; \lambda)}$
the calculation changes slightly. For this, take $\lambda \le \nu_{\min}$ and
temporarily set
\begin{equation*}
A(x; \lambda) := X^- (x; \lambda)^t Y^- (x; \lambda) - Y^- (x; \lambda)^t X^- (x; \lambda),
\end{equation*}
and compute (letting prime denote differentiation with respect to $x$)
\begin{equation*}
\begin{aligned}
A'(x; \lambda) &= X^{- \,\prime} (x; \lambda)^t Y^- (x; \lambda) + X^- (x; \lambda)^t Y^{- \,\prime} (x; \lambda) \\
& - Y^{- \,\prime} (x; \lambda)^t X^- (x; \lambda) - Y^- (x; \lambda)^t X^{- \,\prime} (x; \lambda).
\end{aligned}
\end{equation*}
Using the relations
\begin{equation} \label{more_useful_relations}
X^{- \,\prime} (x; \lambda) = Y^- (x; \lambda);
\quad Y^{- \,\prime} (x; \lambda) = (V(x) - \lambda I) X^- (x; \lambda) + s Y^- (x; \lambda),
\end{equation}
we find that
\begin{equation*}
A' (x; \lambda) = s A (x; \lambda).
\end{equation*}
It follows immediately that $e^{-s x} A (x; \lambda) = c$ for some constant $c$. But that rates of decay
associated with $A(x; \lambda)$ have the form
\begin{equation*}
\mu_{n+j}^- (\lambda) + \mu_{n+k}^- (\lambda)
= s + \frac{1}{2} \sqrt{s^2 - 4 (\lambda - \nu_j^-)} + \frac{1}{2} \sqrt{s^2 - 4 (\lambda - \nu_k^-)},
\end{equation*}
from which we see that the exponents associated with $e^{-s x} A (x; \lambda)$ take the form
\begin{equation*}
\frac{1}{2} \sqrt{s^2 - 4 (\lambda - \nu_j^-)} + \frac{1}{2} \sqrt{s^2 - 4 (\lambda - \nu_k^-)} > 0.
\end{equation*}
It is now clear that by taking $x \to -\infty$ we can conclude that $c = 0$. We conclude that
$A(x; \lambda) = 0$ for all $x \in \mathbb{R}$, and it follows that $\mathbf{X}^- (x; \lambda)$
is the frame for a Lagrangian subspace (see Proposition 2.1 of \cite{HLS}).
\subsection{Monotoncity} \label{monotonicity_subsection}
In this case, according to Lemma 4.2 in \cite{HLS} monotonicity of $\tilde{W} (x; \lambda)$
(in $\lambda$) will be determined by the matrices
\begin{equation} \label{submon}
X^- (x; \lambda) \partial_{\lambda} Y^- (x; \lambda)
- Y^- (x; \lambda) \partial_{\lambda} X^- (x; \lambda)
\end{equation}
and
\begin{equation} \label{submon2}
X^+_{\infty} (\lambda) \partial_{\lambda} Y^+_{\infty} (\lambda)
- Y^+_{\infty} (\lambda) \partial_{\lambda} X^+_{\infty} (\lambda).
\end{equation}
(On the bottom shelf, (\ref{submon}) will be replaced by
$(R^-)^t \partial_{\lambda} S^- (\lambda) - S^- (\lambda)^t \partial_{\lambda} R^-$.)
Let's temporarily set
\begin{equation*}
B (x; \lambda) := X^- (x; \lambda) \partial_{\lambda} Y^- (x; \lambda)
- Y^- (x; \lambda) \partial_{\lambda} X^- (x; \lambda),
\end{equation*}
and compute (letting prime denote differentiation with respect to $x$)
\begin{equation*}
\begin{aligned}
B' (x; \lambda) &:= X^{- \, \prime} (x; \lambda) \partial_{\lambda} Y^- (x; \lambda)
+ X^{-} (x; \lambda) \partial_{\lambda} Y^{- \, \prime} (x; \lambda) \\
& - Y^{- \, \prime} (x; \lambda) \partial_{\lambda} X^- (x; \lambda)
- Y^{-} (x; \lambda) \partial_{\lambda} X^{- \, \prime} (x; \lambda) \\
&= - X^- (x; \lambda)^t X^- (x; \lambda) + s B (x; \lambda),
\end{aligned}
\end{equation*}
where we have used (\ref{more_useful_relations}) to get this final relation.
Integrating this last expression, we find that
\begin{equation*}
B(x; \lambda) = - \int_{-\infty}^x e^{s (x - y)} X^- (y; \lambda)^t X^- (y; \lambda) dy,
\end{equation*}
from which we conclude that $B(x; \lambda)$ is negative definite. We can proceed
similarly to verify that (\ref{submon2}) is also positive definite, and the matrix
associated with the bottom shelf can be analyzed as in the case $s = 0$.
\section{Applications} \label{applications_section}
In this section, we discuss three illustrative examples that we hope
will clarify the analysis. For the first two, which are adapted from
\cite{CDB09}, we will be able to
carry out explicit calculations for a range of values of $\lambda$.
The third example, adapted from \cite{HK1}, will employ Theorem \ref{main_theorem}
more directly, in that we will determine that a certain operator
has no negative eigenvalues by computing only the principal
Maslov index.
\subsection{Example 1.} \label{example1_section}
We consider the Allen-Cahn equation
\begin{equation*}
u_t = u_{xx} - u + u^2,
\end{equation*}
which is known to have a pulse-type stationary solution
\begin{equation*}
\bar{u} (x) = \frac{3}{2} \operatorname{sech}^2 (\frac{x}{2}).
\end{equation*}
(See \cite{CDB09}.) Linearizing about $\bar{u} (x)$
we obtain the eigenvalue problem
\begin{equation*}
-y'' + (1 - 2 \bar{u} (x)) y = \lambda y,
\end{equation*}
which has the form (\ref{main}) with $n=1$
and $V(x) = 1-2\bar{u}(x)$ (for which (A1)-(A2)
are clearly satisfied). Setting $\Phi = {y \choose y'}$,
we can express this equation as a first order system
$\Phi' = \mathbb{A} (x;\lambda) \Phi$, with
\begin{equation} \label{esystem1}
\mathbb{A} (x;\lambda) =
\begin{pmatrix}
0 & 1 \\
1-2\bar{u}(x)-\lambda & 0
\end{pmatrix}; \quad
\mathbb{A}_{\pm} =
\begin{pmatrix}
0 & 1 \\
1-\lambda & 0
\end{pmatrix}.
\end{equation}
As observed in \cite{CDB09}, this equation can be solved
exactly for all $x \in \mathbb{R}$ and $\lambda < 1$
(in this case $\sigma_{ess} (H) \subset [1,\infty)$).
In particular, if we set $s = \frac{x}{2}$,
$\gamma = 2\sqrt{1-\lambda}$, and
\begin{equation*}
H^{\pm} (s,\lambda) = \mp a_0 + a_1 \tanh s \mp \tanh^2 s + \tanh^3 s,
\end{equation*}
with
\begin{equation*}
a_0 = \frac{\gamma}{15} (4-\gamma^2); \quad
a_1 = \frac{1}{5} (2\gamma^2-3); \quad
a_2 = -\gamma,
\end{equation*}
then (\ref{esystem1}) has (up to multiplication by a constant) exactly one solution that decays as $x \to -\infty$,
\begin{equation*}
\Phi^- (x;\lambda) = e^{\gamma s}
\begin{pmatrix}
H^- (s, \lambda) \\
\frac{1}{2} (H_s^- (s,\lambda) + \gamma H^- (s,\lambda))
\end{pmatrix},
\end{equation*}
and
exactly one solution that decays as $x \to +\infty$,
\begin{equation*}
\Phi^+ (x;\lambda) = e^{-\gamma s}
\begin{pmatrix}
H^+ (s, \lambda) \\
\frac{1}{2} (H_s^+ (s,\lambda) - \gamma H^+ (s,\lambda))
\end{pmatrix}.
\end{equation*}
The target space can be obtained either from $\Phi^+ (x;\lambda)$
(by taking $x \to \infty$) or by working with $\mathbb{A}_+ (\lambda)$
directly (as discussed during our analysis), and in either case we
find that a frame for the target space is
$\mathbf{R}^+ = {R^+ \choose S^+} = {1 \choose - \sqrt{1-\lambda}}$.
Computing directly, we see that
\begin{equation*}
(R^+ - iS^+ (\lambda)) (R^+ + i S^+ (\lambda))^{-1}
= \frac{1+i\sqrt{1-\lambda}}{1-i\sqrt{1-\lambda}}.
\end{equation*}
Likewise, the evolving frame in this case can be taken to be
\begin{equation*}
\mathbf{X}^- (x; \lambda) = {X^- (x; \lambda) \choose Y^- (x, \lambda)}
=
\begin{pmatrix}
H^- (s, \lambda) \\
\frac{1}{2} (H_s^- (s,\lambda) + \gamma H^- (s,\lambda))
\end{pmatrix}.
\end{equation*}
We set
\begin{equation*}
\tilde{\mathcal{W}} (x; \lambda) = (X^- (x; \lambda) + i Y^- (x; \lambda)) (X^- (x; \lambda) - i Y^- (x; \lambda))^{-1}
(R^+ - i S^+ (\lambda)) (R^+ + i S^+ (\lambda))^{-1},
\end{equation*}
which in this case we can compute directly. The results of such a
calculation, carried out in MATLAB, are depicted in
Figure \ref{example1_figure}.
\begin{remark} For the Maslov Box, we should properly use $\tilde{W} (x; \lambda)$
as defined in (\ref{tildeWmain}) for some sufficiently large $x_{\infty}$, but for
the purpose of graphical illustration (see Figure \ref{example1_figure}) there
is essentially no difference between working with $\tilde{\mathcal{W}} (x; \lambda)$
and working with $\tilde{W} (x;\lambda)$ defined with $x_{\infty} = 3$.
\end{remark}
Referring Figure \ref{example1_figure}, the curves comprise
$x$-$\lambda$ pairs for which $\tilde{\mathcal{W}} (x; \lambda)$ has
$-1$ as an eigenvalue. The eigenvalues in this case are
known to be $-\frac{5}{4}$, $0$, and $\frac{3}{4}$, and we
see that these are the locations of crossings along the top
shelf, which for plotting purposes we've indicated at
$x = 3$ for this example. We note particularly that the
Principal Maslov Index is -1, because the path
$\Gamma_0$ is only crossed once (the middle curve approaches
$\Gamma_0$ asymptotically, but this does not increment the
Maslov index).
\begin{figure}[ht]
\begin{center}\includegraphics[%
width=10cm,
height=8cm]{example1.jpg}\end{center}
\caption{Figure for Example 1. \label{example1_figure}}
\end{figure}
\subsection{Example 2.} \label{example2_section}
We consider the Allen-Cahn system
\begin{equation} \label{AC_sys}
\begin{aligned}
u_t &= u_{xx} - 4u + 6u^2 - c(u-v) \\
v_t &= v_{xx} - 4v + 6v^2 + c(u-v),
\end{aligned}
\end{equation}
where $c > - 2$, with also $c \ne 0$. System (\ref{AC_sys})
is known to have a stationary solution
\begin{equation*}
\begin{aligned}
\bar{u} (x) &= \operatorname{sech}^2 x \\
\bar{v} (x) &= \operatorname{sech}^2 x
\end{aligned}
\end{equation*}
(see \cite{CDB09}). Linearizing about this vector solution,
we obtain the eigenvalue system
\begin{equation} \label{eg2E}
\begin{aligned}
- \phi'' + (4 - 12\bar{u} (x) + c) \phi - c \psi &= \lambda \phi \\
- \psi'' - c \phi + (4 - 12 \bar{v} (x) + c) \psi &= \lambda \psi,
\end{aligned}
\end{equation}
which can be expressed in form (\ref{main}) with
$y = {\phi \choose \psi}$ and
\begin{equation*}
V (x) =
\begin{pmatrix}
(4 - 12\bar{u} (x) + c) & -c \\
-c & (4 - 12 \bar{v} + c)
\end{pmatrix}.
\end{equation*}
Following \cite{CDB09} we can solve this system explicity in terms
of functions
\begin{equation*}
\begin{aligned}
w^- (x; \kappa) &= e^{\sqrt{\kappa}x} \Big(a_0 + a_1 \tanh x + a_2 \tanh^2 x + \tanh^3 x \Big) \\
w^+ (x; \kappa) &= e^{-\sqrt{\kappa}x} \Big(-a_0 + a_1 \tanh x - a_2 \tanh^2 x + \tanh^3 x \Big),
\end{aligned}
\end{equation*}
where
\begin{equation*}
a_0 = \frac{\kappa}{15} (4 - \kappa); \quad a_1 = \frac{1}{5} (2 \kappa - 3); \quad a_2 = - \sqrt{\kappa},
\end{equation*}
and the values of $\kappa$ will be specified below.
We can now construct a basis for solutions decaying as $x \to -\infty$ as
\begin{equation*}
\mathbf{p}_3^- (x; \lambda) =
\begin{pmatrix}
w^- (x; -\lambda + 4) \\
w^- (x; -\lambda + 4)
\end{pmatrix};
\quad
\mathbf{p}_4^- (x; \lambda) =
\begin{pmatrix}
- w^- (x; -\lambda + 4 + 2c) \\
w^- (x; -\lambda + 4 + 2c)
\end{pmatrix},
\end{equation*}
and a basis for solutions decaying as $x \to +\infty$ as
\begin{equation*}
\mathbf{p}_1^+ (x; \lambda) =
\begin{pmatrix}
w^+ (x; -\lambda + 4) \\
w^+ (x; -\lambda + 4)
\end{pmatrix};
\quad
\mathbf{p}_2^+ (x; \lambda) =
\begin{pmatrix}
- w^+ (x; -\lambda + 4 + 2c) \\
w^+ (x; -\lambda + 4 + 2c)
\end{pmatrix}.
\end{equation*}
These considerations allow us to construct
\begin{equation*}
\begin{aligned}
X^- (x; \lambda) &=
\begin{pmatrix}
w^- (x; -\lambda + 4) & - w^- (x; -\lambda + 4 + 2c) \\
w^- (x; -\lambda + 4) & + w^- (x; -\lambda + 4 + 2c)
\end{pmatrix}; \\
X^+ (x; \lambda) &=
\begin{pmatrix}
w^+ (x; -\lambda + 4) & - w^+ (x; -\lambda + 4 + 2c) \\
w^+ (x; -\lambda + 4) & + w^+ (x; -\lambda + 4 + 2c)
\end{pmatrix},
\end{aligned}
\end{equation*}
with then $Y^- (x; \lambda) = X^-_x (x;\lambda)$ and
$Y^+ (x; \lambda) = X^+_x (x; \lambda)$.
In order to construct the target space, we write (\ref{eg2E})
as a first-order system by setting $\Phi_1 = \phi$,
$\Phi_2 = \psi$, $\Phi_3 = \phi'$, and $\Phi_4 = \psi'$.
This allows us to write
\begin{equation*}
\mathbf{\Phi}' = \mathbb{A} (x; \lambda) \mathbf{\Phi};
\quad
\mathbb{A} (x; \lambda) =
\begin{pmatrix}
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 \\
- f(x;\lambda) & -c & 0 & 0 \\
-c & -f(x;\lambda) & 0 & 0
\end{pmatrix},
\end{equation*}
where $f(x; \lambda) = \lambda - 4 - c + 12 \bar{u}$. We set
\begin{equation*}
\mathbb{A}_+ (\lambda) := \lim_{x \to + \infty} \mathbb{A} (x; \lambda)
= \begin{pmatrix}
0 & 0 & 1 & 0 \\
0 & 0 & 0 & 1 \\
- \lambda + 4 + c & -c & 0 & 0 \\
-c & -\lambda + 4 + c & 0 & 0
\end{pmatrix}.
\end{equation*}
If we follow our usual ordering scheme for indices then for $-2 < c < 0$ we
have $\nu_1^+ = 4 + 2c$ and $\nu_2^+ = 4$, with corresponding
eigenvectors $r_1^+ = {1 \choose -1}$ and $r_2^+ = {1 \choose 1}$.
Accordingly, we have $\mu_1^+ (\lambda) = - \sqrt{-\lambda + 4}$,
$\mu_2^+ (\lambda) = -\sqrt{-\lambda + 4 + 2c}$,
$\mu_3^+ (\lambda) = \sqrt{-\lambda + 4 + 2c}$,
and $\mu_4^+ (\lambda) = \sqrt{-\lambda + 4}$. We
conclude that a frame for $\ell^+_{\mathbf{R}} (\lambda)$
is $\mathbf{R}^+ = {R^+ \choose S^+ (\lambda)}$, where
\begin{equation*}
R^+ =
\begin{pmatrix}
1 & 1 \\
1 & -1
\end{pmatrix}; \quad
S^+ (\lambda) =
\begin{pmatrix}
\mu_1^+ (\lambda) & \mu_2^+ (\lambda) \\
\mu_1^+ (\lambda) & -\mu_2^+ (\lambda)
\end{pmatrix}.
\end{equation*}
The resulting spectral curves are plotted in Figure
\ref{example2_figure} for $c = -1$. In this case, it
is known that $H$ has exactly six eigenvalues:
$-7$, $-5$, $-2$, $0$, $1$ and $3$ (the eigenvalues
$1$ and $3$ are omitted from our window). We see that
the three crossings along the line $\lambda = 0$ correspond
with the count of three negative eigenvalues.
\begin{figure}[ht]
\begin{center}\includegraphics[%
width=12cm,
height=8cm]{hls2example2.jpg}\end{center}
\caption{Figure for Example 2. \label{example2_figure}}
\end{figure}
\subsection{Example 3.} \label{example3_section}
Consider the Allen-Cahn system
\begin{equation} \label{allencahn}
u_t = u_{xx} - D_u F(u),
\end{equation}
where
\begin{equation*}
F(u_1, u_2) = u_1^2 u_2^2 + u_1^2 (1 - u_1 - u_2)^2 + u_2^2 (1 - u_1 - u_2)^2,
\end{equation*}
which is adapted from p. 39 of \cite{HK1}. In this setting,
stationary solutions $\bar{u} (x)$ satisfying endstate
conditions
\begin{equation*}
\lim_{x \to \pm \infty} \bar{u} (x) = u_{\pm},
\end{equation*}
for $u_- \ne u_+$ are called {\it transition waves}. A
transition wave solution for (\ref{allencahn}) is depicted
in Figure \ref{transitionfront}. In this case, we have
$u_1^- = 1$, $u_2^- = 0$, $u_1^+ = 0$, and $u_2^+ = 1$.
\begin{figure} \label{transitionfront}
\begin{center}\includegraphics[%
width=12cm,
height=8cm]{ternary1-eps-converted-to.pdf}\end{center}
\caption{Transition front solution for a ternary Cahn-Hilliard system.}
\end{figure}
Upon linearization of (\ref{allencahn}) about $\bar{u} (x)$, we obtain
the eigenvalue problem
\begin{equation} \label{allencahnE}
- \phi'' + V (x) \phi = \lambda \phi,
\end{equation}
where $V (x) := D_u^2 F(\bar{u})$ denotes the usual Hessian matrix. In
this case,
\begin{equation*}
V_- =
\begin{pmatrix}
2 & 2 \\
2 & 4
\end{pmatrix};
\quad
V_+ =
\begin{pmatrix}
4 & 2 \\
2 & 2
\end{pmatrix}.
\end{equation*}
Using our usual labeling scheme, we have $\nu_1^+ = 3 - \sqrt{5}$
and $\nu_2^+ = 3 + \sqrt{5}$, with respective eigenvectors
\begin{equation*}
r_1^+ =
\begin{pmatrix}
1 \\ -\frac{1+\sqrt{5}}{2} \\
\end{pmatrix}
\quad
r_2^+ =
\begin{pmatrix}
1 \\ -\frac{1-\sqrt{5}}{2}
\end{pmatrix}.
\end{equation*}
The corresponding values $\{\mu_j^+\}_{j=1}^4$ are
$\mu_1^+ = - \sqrt{\nu_2^+ - \lambda}$, $\mu_2^+ = - \sqrt{\nu_1^+ -\lambda}$,
$\mu_3^+ = \sqrt{\nu_1^+ - \lambda}$, $\mu_4^+ = \sqrt{\nu_2^+ -\lambda}$.
For the target space $\ell^+_{\mathbf{R}}$ we use the frame
\begin{equation*}
\mathbf{R}^+ (\lambda) =
\begin{pmatrix}
R^+ \\ S^+ (\lambda)
\end{pmatrix}
=
\begin{pmatrix}
r_2^+ & r_1^+ \\
\mu_1^+ r_2^+ & \mu_2^+ r_1^+
\end{pmatrix}.
\end{equation*}
For the evolving Lagrangian subspace $\ell^- (x; \lambda)$ we need a basis
for the two-dimensional space of solutions that decay as $x \to -\infty$.
Generally, we construct this basis from the solutions
\begin{equation*}
\mathbf{p}_{2+j}^- (x; \lambda) =
e^{\mu_{2+j}^- (\lambda) x}
(\scripty{r}_{\,2+j}^{\,-}
+ \mathbf{E}_{2+j}^-); \quad j = 1, 2,
\end{equation*}
from Lemma \ref{ODElemma}, but computationally it is easier to
note that for $\lambda = 0$, $\bar{u}_x$ is a solution of
(\ref{main}) that decays as $x \to -\infty$. In \cite{HK1}
the authors check that $\bar{u}_x (x)$ decays at the slower
rate (i.e., the rate of $\mathbf{p}_3^-$), so we can take
as our frame
\begin{equation*}
\mathbf{X}^- (x; 0) =
\begin{pmatrix}
p_4^- (x; 0) & \bar{u}_x (x) \\
p_4^{-\,\prime} (x; 0) & \bar{u}_{xx} (x)
\end{pmatrix},
\end{equation*}
which we scale to
\begin{equation*}
\mathbf{X}^- (x; 0) =
\begin{pmatrix}
e^{-\mu_4^- (\lambda) x} p_4^- (x; 0) & e^{-\mu_3^- (\lambda) x} \bar{u}_x (x) \\
e^{-\mu_4^- (\lambda) x} p_4^{-\,\prime} (x; 0) & e^{-\mu_3^- (\lambda) x} \bar{u}_{xx} (x)
\end{pmatrix}.
\end{equation*}
The advantage of this is that $\bar{u} (x)$ is already known, and the
faster-decaying solution $\mathbf{p}_4^- (x; 0)$ can be generated
numerically in a straightforward way (see \cite{HK1}).
In practice, we compute $\tilde{W} (x; 0)$ for $x$ running from $-10$
to $10$, and find that one of its eigenvalues remains confined to
the semicircle with positive real part and that the other rotates
monotonically clockwise, nearing $-1$ as $x$ approaches $10$. In
this case, we know that $\lambda = 0$ is an eigenvalue, from which
we can conclude that (at least) one of the eigenvalues will approach
$-1$ as $x \to +\infty$. We view this as strong numerical evidence
(though certainly not numerical proof) that $\lambda = 0$ is a simple
eigenvalue of $H$, and that there are no negative eigenvalues of
$H$.
\section*{Appendix}
In this short appendix, we construct the asymptotic Lagrangian path
\begin{equation*}
\ell^-_{+\infty} (\lambda) = \lim_{x \to + \infty} \ell^- (x; \lambda),
\end{equation*}
and show that it is not generally continuous in $\lambda$. For a related
discussion from a different point of view, we refer to Lemma 3.7 of
\cite{AGJ}.
As a start, we recall that one choice of frame for $\ell^- (x;\lambda)$
is
\begin{equation*}
\mathbf{X}^- (x; \lambda)
= \begin{pmatrix}
\mathbf{p}^-_{n+1} (x; \lambda) & \mathbf{p}^-_{n+2} (x; \lambda) & \dots
& \mathbf{p}^-_{2n} (x; \lambda)
\end{pmatrix},
\end{equation*}
where we have from Lemma \ref{ODElemma}
\begin{equation*}
\mathbf{p}^-_{n+j} (x; \lambda) = e^{\mu_{n+j}^- (\lambda) x}
(\scripty{r}_{\,n+j}^{\,-}
+ \mathbf{E}_{n+j}^-); \quad j = 1, 2, \dots, n.
\end{equation*}
Each of the $\mathbf{p}^-_{n+j}$ can be expressed as a linear combination of
the basis of solutions $\{\mathbf{p}_k^+\}_{k=1}^{2n}$, where we recall from
Lemma \ref{ODElemma} that the solutions $\{\mathbf{p}_k^+\}_{k=1}^{n}$ decay
as $x \to +\infty$, while the solutions $\{\mathbf{p}_k^+\}_{k=n+1}^{2n}$
grow as $x \to +\infty$. I.e., for each $j = 1,2,\dots,n$, we can write
\begin{equation*}
\mathbf{p}^-_{n+j} (x; \lambda)
= \sum_{k=1}^{2n} c_{j k} (\lambda) \mathbf{p}_k^+ (x;\lambda),
\end{equation*}
and so the collection of vector functions on the right-hand side
provides an alternative way to express the same frame $\mathbf{X}^- (x;\lambda)$.
We note that if $\lambda = \nu_{\min}$ then the modes
$\{\mathbf{p}_k^+ (x;\lambda)\}_{k=n+1}^{2n}$ cannot be obtained as
$\lambda \to 0^-$ limits of solutions that grow as $x \to +\infty$,
because these will coalesce with solutions obtained as $\lambda \to 0^-$
limits of solutions that decay as $x \to +\infty$. However, for such
values of $\lambda$ we can still find $2n$ linearly independent solutions
of (\ref{first_order}), and the $\{\mathbf{p}_k^+ (x;\lambda)\}_{k=n+1}^{2n}$
correspond with the $n$ solutions not obtained as $\lambda \to 0^-$
limits of solutions that decay as $x \to +\infty$. For a direct approach
toward defining such solutions, readers are referred to \cite{H}.
Fix $\lambda \in [-\lambda_{\infty}, \nu_{\min}]$, and suppose the fastest
growth mode $\mathbf{p}^+_{2n} (x; \lambda)$ appears in the expansion of
at least one of the $\mathbf{p}^-_{n+j}$ (i.e., the coefficient associated with
this mode is non-zero). (There may be additional modes that grow at the same
rate $\mu_{2n}^+$, but they will have different, and linearly independent,
associated eigenvectors $\scripty{r}_{\,n+j}^{\,-}$, allowing us to distinguish
them from $\mathbf{p}^+_{2n} (x; \lambda)$.) By taking appropriate linear
combinations, we can identify a new frame for $\ell^- (x; \lambda)$ for which
$\mathbf{p}_{2n}^+$ only appears in one column. If $\mathbf{p}^+_{2n} (x; \lambda)$
does not appear in the sum for any $\mathbf{p}^-_{n+j}$ we can start with
$\mathbf{p}_{2n-1}^+$ and proceed similarly, continuing until we get to the first
mode that appears. Since the $\{\mathbf{p}^-_{n+j}\}_{j=1}^n$ form a basis
for an $n$-dimensional space, we will be able to distinguish $n$ modes in this
way. At the end of this process, we will have created a new frame for $\mathbf{X}^- (x; \lambda)$
with columns $\{\tilde{\mathbf{p}}^-_{j}\}_{j=1}^n$, where
\begin{equation*}
\tilde{\mathbf{p}}_{j}^- (x; \lambda) =
e^{\mu_{k(j)}^+ x} \Big(s_{k(j)}^+ + \tilde{\mathbf{E}}^+_{k(j)} (x; \lambda) \Big),
\end{equation*}
for some appropriate map $j \mapsto k(j)$.
If the rate $\mu_{k (j)}^+$ is distinct as an eigenvalue of $\mathbb{A}_+ (\lambda)$
then we will have $s_{k(j)}^+ = \scripty{r}_{\,k(j)}^{\,+}$, but if $\mu_{k (j)}^+$
is not distinct then $s_{k(j)}^+$ will generally be a linear combination of
eigenvectors of $\mathbb{A}_+ (\lambda)$ (and so, of course, still an eigenvector
of $\mathbb{A}_+ (\lambda)$). This process may also introduce an expansion coefficient
in front of $s_{k(j)}^+$, but this can be factored out in the specification of
the frame.
As usual, we can view the exponential scalings $e^{\mu^+_{k(j)} x}$ as expansion coefficients,
and take as our frame for $\ell^- (x;\lambda)$ the $2n \times n$ matrix with columns
$s_{k(j)}^+ + \tilde{\mathbf{E}}^+_{k(j)} (x; \lambda)$. Taking now the limit
$x \to \infty$ we see that we obtain the asymptotic frame
\begin{equation}
\mathbf{X}^-_{+ \infty} (\lambda)
= \begin{pmatrix}
s_{k(1)}^+ & s_{k(2)}^+ & \dots & s_{k(n)}^+
\end{pmatrix}.
\end{equation}
We can associate $\ell^-_{+\infty} (\lambda)$ as the Lagrangian subspace with
this frame, verifying that this Lagrangian subspace is well-defined.
Last, we verify our comment that $\ell^-_{+\infty} (\lambda)$ is not generally
continuous as a function of $\lambda$. To see this, we begin by noting that
if $\lambda_0 \in [-\lambda_{\infty}, \nu_{\min}]$ is not an eigenvalue of
$H$ then the leading modes selected in our process must all be growth modes,
and we obtain $\mathbf{X}^-_{+ \infty} (\lambda_0) = \mathbf{R}^+ (\lambda_0)$,
in agreement with Lemma 3.7 in \cite{AGJ}. Suppose, however, that
$\lambda_0 \in [-\lambda_{\infty},\nu_{\min})$
is an eigenvalue of $H$, and for simplicity assume $\lambda_0$ has geometric
multiplicity 1. Away from essential spectrum, $\lambda_0$ will be isolated, and
so we know that any $\lambda$ sufficiently close to $\lambda_0$ will not
be in the spectrum of $H$. We conclude that the frame for $\lambda_0$ will
comprise $n-1$ of the eigenvectors $\{\scripty{r}_{\,n+j}^{\,+}\}_{j=1}^n$, along with
one of the $\{\scripty{r}_{\,j}^{\,+}\}_{j=1}^n$. Since the exchanged vectors
will lead to bases of different spaces, we can conclude that
$\ell^-_{+\infty} (\lambda)$ is not continuous at $\lambda_0$.
In order to clarify the discussion, we briefly consider the simple case $n=1$.
In this case, we have (for $\lambda < \nu_{\min}$) a single solution
$\mathbf{p}_2^- (x;\lambda)$ that decays as $x \to - \infty$, and we can write
\begin{equation*}
\mathbf{p}_2^- (x; \lambda) = c_{11} (\lambda) \mathbf{p}_1^+ (x; \lambda)
+ c_{12} (\lambda) \mathbf{p}_2^+ (x; \lambda),
\end{equation*}
where $\mathbf{p}_1^+ (x; \lambda)$ decays as $x \to + \infty$ and
$\mathbf{p}_2^+ (x; \lambda)$ grows as $x \to + \infty$.
If $\lambda_0$ is not an eigenvalue of $H$ we must have $c_{12} (\lambda_0) \ne 0$,
and so
\begin{equation*}
\begin{aligned}
\mathbf{p}_2^- (x; \lambda_0) &=
c_{11} (\lambda_0) e^{\mu_1^+ (\lambda_0) x} (\scripty{r}_{\,1}^{\,+}
+ \mathbf{E}_{1}^+ (x; \lambda_0))
+ c_{12} (\lambda_0) e^{\mu_2^+ (\lambda_0) x} (\scripty{r}_{\,2}^{\,+}
+ \mathbf{E}_{2}^+ (x; \lambda_0)) \\
&=
c_{12} (\lambda_0) e^{\mu_2^+ (\lambda_0) x}
\Big(\scripty{r}_{\,2}^{\,+} + \mathbf{E}_{2}^+ (x; \lambda_0)
+ \frac{c_{11} (\lambda_0)}{c_{12} (\lambda_0)} e^{(\mu_1^+ (\lambda_0) - \mu_2^+ (\lambda_0)) x}
(\scripty{r}_{\,1}^{\,+} + \mathbf{E}_{1}^+ (x; \lambda_0)) \Big) \\
&=
c_{12} (\lambda_0) e^{\mu_2^+ (\lambda_0) x}
(\scripty{r}_{\,2}^{\,+} + \tilde{\mathbf{E}}_{2}^+ (x; \lambda_0)),
\end{aligned}
\end{equation*}
where $\tilde{\mathbf{E}}_{2}^+ (x; \lambda_0)) = \mathbf{O} ((1+|x|)^{-1})$.
We can view $\scripty{r}_{\,2}^{\,+} + \tilde{\mathbf{E}}_{2}^+ (x; \lambda_0)$ as
a frame for $\ell^- (x; \lambda_0)$, and it immediately follows that as $x \to \infty$
the path of Lagrangian subspaces $\ell^- (x; \lambda_0)$ approaches the Lagrangian
subspace with frame $\scripty{r}_{\,2}^{\,+}$ (denoted $\ell^-_{+ \infty} (\lambda_0)$
above). Moreover, since
$\scripty{r}_{\,1}^{\,+}$ serves as a frame for $\ell^+_{\mathbf{R}} (\lambda_0)$
we can construct $\tilde{\mathcal{W}} (x; \lambda_0)$ from this pair. Taking the
limit as $x \to \infty$ we see that
\begin{equation*}
\tilde{\mathcal{W}}^+ (\lambda_0) := \lim_{x \to \infty} \tilde{\mathcal{W}} (x; \lambda_0)
= - \frac{r_1^+ + i \mu_2^+ (\lambda_0) r_1^+}{r_1^+ - i \mu_2^+ (\lambda_0) r_1^+}
\cdot \frac{r_2^+ - i \mu_1^+ (\lambda_0) r_2^+}{r_2^+ + i \mu_1^+ (\lambda_0) r_2^+}.
\end{equation*}
By normalization, we can take both $r_1^+$ and $r_2^+$ to be $1$, and we must have
$\mu_1^+ = - \mu_2^+$, so
\begin{equation*}
\tilde{\mathcal{W}}^+ (\lambda_0)
= - \frac{1 + i \mu_2^+ (\lambda_0)}{1 - i \mu_2^+ (\lambda_0)}
\cdot \frac{1 + i \mu_2^+ (\lambda_0)}{1 - i \mu_2^+ (\lambda_0)}
= - \frac{(1 + i \mu_2^+ (\lambda_0))^2}{(1 - i \mu_2^+ (\lambda_0))^2},
\end{equation*}
which can only be $-1$ if $\mu_2^+ (\lambda_0) = 0$ (a case ruled out
in this calculation).
On the other hand, if $\lambda_0 \in \sigma_{pt} (H)$ we will have $c_{12} (\lambda_0) = 0$
(and $c_{11} (\lambda_0) \ne 0$). In this case, the frame for $\ell^- (x; \lambda_0)$
will be $\scripty{r}_{\,1}^{\,+} + \tilde{\mathbf{E}}_{1}^+ (x; \lambda_0)$, and taking
$x \to +\infty$ we see that $\ell^- (x; \lambda_0)$ will approach the Lagrangian subspace
with frame $\scripty{r}_{\,1}^{\,+}$. Recalling again that $\scripty{r}_{\,1}^{\,+}$
serves as a frame for $\ell^+_{\mathbf{R}} (\lambda_0)$ we see that
\begin{equation*}
\begin{aligned}
\tilde{\mathcal{W}}^+ (\lambda_0) &:= \lim_{x \to \infty} \tilde{\mathcal{W}} (x; \lambda_0)
= - \frac{r_2^+ + i \mu_1^+ (\lambda_0) r_2^+}{r_2^+ - i \mu_1^+ (\lambda_0) r_2^+}
\cdot \frac{r_2^+ - i \mu_1^+ (\lambda_0) r_2^+}{r_2^+ + i \mu_1^+ (\lambda_0) r_2^+} \\
&= -1.
\end{aligned}
\end{equation*}
For Example 1 in Section \ref{applications_section}, we have
$\mu_1^+ (\lambda) = - \sqrt{1-\lambda}$ and
$\mu_2^+ (\lambda) = + \sqrt{1 - \lambda}$. We know that in that example
$\lambda_0 = 0$ is an eigenvalue, so we have $\tilde{\mathcal{W}}^+ (0) = -1$,
but for $\lambda \ne 0$, $|\lambda| < 1$, we have
\begin{equation*}
\tilde{\mathcal{W}}^+ (\lambda)
= - \frac{(1+i\sqrt{1-\lambda})^2}{(1-i\sqrt{1-\lambda})^2}.
\end{equation*}
If we substitute $\lambda = 0$ into this relation, we obtain $+1$,
and so we see that $\tilde{\mathcal{W}}^+ (\lambda)$ is not
continous in $\lambda$ (at $\lambda = 0$ in this case).
|
\section{Introduction}
Monter Carlo (MC) method has long been known to mathematicians and physicians as one of the most versatile and widely used computational algorithms. With the advantage of a dimension-independent convergence rate, it is regarded as the most efficient method to overcome the \emph{the curse of dimensionality} (\cite{Metropolis:2012ft}). However, the slow convergent rate, $\mathcal{O}({N}^{-1/2})$(where $N$ denotes the sample size), often results in unaffordable computational cost to generate high-resolution samples with a large sample size. Specifically, when MC is applied to the complex system models described by differential equations with uncertainties, which usually arise from e.g., data inaccuracies and information loss, the computational cost will dramatically (polynomially) grow due to the larger sample size required as one moves onto high-resolution meshes. To mitigate such growth, many efforts such as quasi-Monte Carlo method (\cite{2013-paper-Sloan-QMC, 2012-paper-Kuo-QMC, 1993-paper-Niederreiter-QMC}), variance reduction method (\cite{1964-book-Hammersley-VR}), importance sampling and stratified sampling method (\cite{2015-paper-Bugallo-IS, 2001-book-Liu-MC}), etc, have been made to speed up the convergence of MC.
Besides these ameliorated methods, the mulit-level Monte Carlo (MLMC) method has attracted much attention for its promising potential in the reduction of computational complexity of uncertainty quantification (UQ) problems (e.g., \cite{Heinrich:2001jl, {Kebaier:2005ce},{Giles:2008gc}} and references therein).
Similar to multi-grid method for iteratively solving large linear deterministic systems (\cite{Wesseling:1992ug}), the MLMC algorithm utilizes a hierarchy of resolutions instead of one. The basic idea, roughly speaking, is to obtain
independent numerical samples on the coarse grids (higher level),
then improve the results on the fine grids (lower level) iteratively.
The variance decays with level at a faster rate than the computational expense increases.
It can be shown that MLMC could strike a balance between the efficiency and accuracy in solving
the UQ problems and obtaining the quantity of interests (QoI).
There is a large body of literature on MLMC,
and some relevant references are listed as follows: Barth et al
(\cite{Barth:2011gz}) couples the MLMC method with the finite element
method (FEM) to solve stochastic elliptic equations, and presents rigorous error analysis. Mishra et al (\cite{Mishra:2012ec, Mishra:2012ea,
Mishra:2012cv, {Mishra:2016hv}}) couple the MLMC method with the
finite volume method for hyperbolic systems. Kornhuber et al
(\cite{Kornhuber:2014iz}) applies the MLMC with FEM to study
stochastic elliptic variational inequalities. Li et al
(\cite{Li:2016ie}) couples the MLMC with the weak Galerkin method to
study the elliptic equations. For a survey of the MLMC and the
literature on its applications, see \cite{Giles:2015dd}.
In this paper we are concerned with the applicability of the MLMC method for
long-term simulations of turbulent geophysical flows with uncertain parameters.
For turbulent flows, pointwise behaviors of the solutions are no longer
relevant. In fact, after the initial spinup period, the difference between solutions
on two different meshes is spatially uncorrelated,
even if all the other settings are the same.
Thus, the usual notion of error convergence, e.g.,
the pointwise error estimates under certain norms, no longer applies.
Due to this unreliable nature of the pointwise solutions,
the research objective of turbulence simulations is
often focused on computing certain aggregated QoI's,
such as the global mean of sea surface temperature,
instead of the pointwise solutions.
Our main motivation for this work is to adapt the analysis of MLMC to QoI under some
verifiable assumptions. Both the assumptions and the conclusions will
be examined using the Antarctic Circumpolar Current
(ACC) model.
Turbulence models often include closures to account for unresolved eddy
activities. These closures in general need to be adjusted according to the
level of mesh resolutions. This is a dramatic departure
from the situation involving steady-state or laminar flows, where the
discrete model is kept the same, and only grid resolutions vary. But
this departure does not automatically invalidate the MLMC for turbulent
flows. Eddy closures are implemented to prevent instability and to
improve qualitative large-scale behaviors of the solution. However, the
accuracy of the estimate of the QoI is
aligned with the grid resolutions, i.e., the estimate will
improve or worsen as the mesh refines or
coarsens.
Based on this premise, the effectiveness and applicability of
MLMC could be expected for simulations of turbulent flows.
The numerical scheme used in this paper is a staggered C-grid
finite difference finite volume scheme (\cite{Ringler:2010io}) based on a Voronoi tessellation (VT, \cite{Du:1999gs, Du:2003ka}). The
VT primarily consists of pentagons and hexagons, and thus nesting
between different levels of meshes is impossible, which implies a
direct comparison between the solutions on two different meshes is also
impossible. This would result in a major hurdle in applying MLMC to steady-state or laminar flows, but for turbulent flows, the pointwise
behaviors of the solution are uncorrelated, and the focus is instead on
QoI. Thus, the issue with mesh matching is irrelevant
here.
The objective of the present paper is two-fold: {\bf(i).} to
explore the effectiveness of the MLMC method in the presence of the
challenges associated with turbulent flows. {\bf(ii).} to explore
an optimal way to set up the MLMC simulations. The rest
of the paper is organized as follows. In Section
\ref{sec:monte-carlo-multi}, we briefly review the MC and the MLMC
methods, and detail the possible strategies for setting up the MLMC
simulations. In Section \ref{sec:numer-exper-using}, we apply the MLMC
method to a turbulent channel flow mimicking the ACC, and examine the effectiveness of the method under various
strategies. The paper ends with some concluding remarks in Section
\ref{sec:discussions}.
\section{The Monte Carlo and the multi-level Monte Carlo
methods}\label{sec:monte-carlo-multi}
We designate the QoI to be calculated by $U$, which
can be e.g., volume transport, mean sea-surface temperature (SST), etc.
We denote the number of levels of grid resolution by $L$, and the
resolution at each level by $r_l$, $1\leq l\leq L$, and the highest
resolution by $r\equiv r_1$. We assume that
\begin{equation}
\label{eq:1}
r_l = 2r_{l-1} = 2^{l-1}r_1 \equiv 2^{l-1} r.
\end{equation}
\begin{assumption}
We assume that, at each level, the computational cost is proportional
to the total number of spatial-temporal degrees of freedom $N_l$. For
simplicity, in the sequel, we identify the computational cost with
$N_l$. We further assume that the total number of degrees of freedom
is proportional to $r_l^{-3}$, that is,
\begin{equation}
\label{eq:2}
N_l = C_1 r_l^{-3}.
\end{equation}
\end{assumption}
We designate the total number of degrees of freedom at the highest
resolution $r_1 \equiv r$ by
\begin{equation}
\label{eq:3}
N \equiv N_1 = C_1 r^{-3},
\end{equation}
where $C_1$ is a constant. The cubic relation between $N$ and $r$ is tailored towards models of large-scale geophysical
flows, where the vertical resolution is often held fixed and the time
step size varies linearly according to the horizontal resolution.
From (\ref{eq:1}) and (\ref{eq:2}) it is derived that
\begin{equation}
\label{eq:4}
\dfrac{N_l}{N} = 8^{-(l-1)}.
\end{equation}
\subsection{The Monte Carlo method}\label{sec:monte-carlo-method}
We recall the classical MC method as it is applied to the
ensemble simulations at a fixed resolution $r$. The numerical
approximation of the QoI $U$ at this resolution is
denoted by $U_r$, and the computational cost of each individual
simulation by $N$, which is related to the grid resolution through
(\ref{eq:3}). We denote each realization by a superscript $m$, as in
$U^m$ and $U_r^m$, $1\leq m\leq M$. The MC mean is defined as
\begin{equation}
\label{eq:5}
\mbox{\rm\sffamily E}_M\left[ U_r\right]: = \dfrac{1}{M} \sum_{m=1}^M U_r^m.
\end{equation}
We now examine the difference between the sample mean and the
expectation $\mbox{\rm\sffamily E}[U]$ of the true solution $U$. We use the standard notations for the $\sigma$-finite probability space $(\Omega,\mathcal{F},P)$, where sample space $\Omega$ is a set of all possible outcomes, $\mathcal{F}$ is a $\sigma$-algebra of events, and $P : \mathcal{F}\rightarrow[0,1]$ is a probability measure.
\begin{equation}
\label{eq:6}
\left\| \mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}_M[U_r]\right\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq
\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}_M[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} + \|\mbox{\rm\sffamily E}_M[U] - \mbox{\rm\sffamily E}_M[U_r]\|_{\small{\mathcal{L}^2(\Omega; P)}}.
\end{equation}
We note that, by the Central Limit Theorem,
\begin{equation}
\label{eq:7}
\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}_M[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} = \dfrac{\delta}{\sqrt{M}},
\end{equation}
where $\delta$ is the standard deviation in the true solution. For the
second term on the right-hand side of (\ref{eq:6}),
\begin{align*}
& \|\mbox{\rm\sffamily E}_M[U] - \mbox{\rm\sffamily E}_M[U_r]\|_{\small{\mathcal{L}^2(\Omega; P)}}= \|\mbox{\rm\sffamily E}_M[U-U_r]\|_{\small{\mathcal{L}^2(\Omega; P)}}
= \|\dfrac{1}{M}\sum_{m=1}^M(U^m - U_r^m)\|_{\small{\mathcal{L}^2(\Omega; P)}} \\ \leq & \dfrac{1}{M}\sum_{m=1}^M \left\|U^m - U^m_r\right\|_{\small{\mathcal{L}^2(\Omega; P)}}= \mbox{\rm\sffamily E}_M\left[\|U-U_r\|_{\small{\mathcal{L}^2(\Omega; P)}}\right]=\left\|U-U_r\right\|_{\small{\mathcal{L}^2(\Omega; P)}}.
\end{align*}
\begin{assumption}
We assume that the $L^2$-norm of the error in the quantity of interest
is proportional to $r^\alpha$, where $\alpha$ designates the rate of
convergence regarding the quantity.
\end{assumption}
That is, designating the $L^2$-norm of the error at the resolution $r$
by $e$, we may write that
\begin{equation}
\label{eq:64}
e \equiv \| U- U_r\|_{\small{\mathcal{L}^2(\Omega; P)}} = C_2r^\alpha,
\end{equation}
where $C_2$ is a constant independent of the grid resolution.
Hence, concerning the MC mean of the true solution and the MC
mean of the approximate solution, we have
\begin{equation}
\label{eq:8}
\|\mbox{\rm\sffamily E}_M[U] - \mbox{\rm\sffamily E}_M[U_r]\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq e
= C_2 r^\alpha.
\end{equation}
Combining (\ref{eq:7}) and (\ref{eq:8}) yields
\begin{equation}
\label{eq:9}
\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}_M[U_r]\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq
\dfrac{\delta}{\sqrt{M}} + e.
\end{equation}
The first term on the right-hand side represents the discretization
error of the probability space, and the second term represents the
discretization error of the temporal-spatial space.
For a given resolution $r$, the the spatial-temporal discretization
error is fixed. The sample size $M$ should be chosen so that the
probability space discretization error is on the same order as the
temporal-spatial discretization error. Thus, we set
\begin{align}
\dfrac{\delta}{\sqrt{M}} &= e,\ M = \dfrac{\delta^2}{e^2}\label{eq:10}.
\end{align}
Given this choice of $M$, now the combined errors in the MC mean can
be given,
\begin{equation}
\label{eq:11}
\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}_M[U_r]\|_{\small{\mathcal{L}^2(\Omega; P)}}
\leq 2 e.
\end{equation}
The total computational cost for the Monte Carlo method, $\mathcal{C}_{\mathrm{\tiny MC}}$, can also be calculated, using
\eqref{eq:3}, \eqref{eq:64}, and \eqref{eq:10},
\begin{equation}
\label{eq:12}
\mathcal{C}_{\mathrm{\tiny MC}} = N\cdot M =
\dfrac{\delta^2}{C_1^{\frac{2}{3}\alpha} C^2_2} N^{1+\frac{2}{3}\alpha}.
\end{equation}
The total computational cost for the ensemble simulation using the
conventional MC method grows polynomially in terms of the
computational cost for each individual simulation, and the degree of
the polynomial is $1+\frac{2}{3}\alpha$. Also as expected, larger
deviation $\delta$ in the true solution would demand more
computational resource.
\subsection{The multi-level Monte Carlo method}
We denote the numerical approximation of $U$ at each level by $U_l$,
$1\le l\le L$, and each realization of $U_l$ by $U^m_l$, $1\le m\le
M_l$.
We note that the numerical approximation $U_1$ at the lowest level
(highest resolution) can be decomposed as
\begin{equation}
\label{eq:13}
U_1 = \sum_{l=1}^{L-1}(U_l - U_{l+1}) + U_L.
\end{equation}
Then clearly,
\begin{equation}
\label{eq:14}
\mbox{\rm\sffamily E}[U_1] = \sum_{l=1}^{L-1}\mbox{\rm\sffamily E}[U_l - U_{l+1}] + \mbox{\rm\sffamily E}[U_L].
\end{equation}
In practice, the mean is approximated by MC mean, and as it has
been shown above, the accuracy of such approximation is determined by
two competing factors, the variance in the random variable $\delta$ and the
sample size $M$. A larger variance requires a larger sample size. The
success of the MLMC method is built on the hypothesis that the
variance of the difference between two solutions at successive levels
is much smaller than the variance of each individual solution, and
thus requires a much smaller sample size. We now define the $L$-level
sample mean of $U$, $ \mbox{\rm\sffamily E}^L[U]$, as
\begin{equation}
\label{eq:15}
\mbox{\rm\sffamily E}^L[U] = \sum_{l=1}^{L-1} \mbox{\rm\sffamily E}_{M_l}[U_l - U_{l+1}] + \mbox{\rm\sffamily E}_{M_L}[U_L],
\end{equation}
where $M_l$ represents the sample size, and the sample mean $ \mbox{\rm\sffamily E}_{M_l}$
at each level is defined in the same way as (\ref{eq:5}).
The relation between $M_l$ and the total sample size $\widetilde M_l$ at
each level is as follows,
\begin{equation}
\label{eq:16}
\left\{
\begin{aligned}
&\widetilde M_1 = M_1,\\
&\widetilde M_l = M_{l-1} + M_l,\qquad 2\le l\le L.
\end{aligned}\right.
\end{equation}
We now examine the theoretical mean $\mbox{\rm\sffamily E}[U]$ and the $L$-level sample
mean $\mbox{\rm\sffamily E}^L[U]$.
\begin{equation}
\label{eq:17}
\| \mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}^L[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq \| \mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}[U_1]\|_{\small{\mathcal{L}^2(\Omega; P)}} + \| \mbox{\rm\sffamily E}[U_1] - \mbox{\rm\sffamily E}^L[U]\|_{\small{\mathcal{L}^2(\Omega; P)}}.
\end{equation}
We note that,
\begin{align*}
\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}[U_1]\|_{\small{\mathcal{L}^2(\Omega; P)}} = \|\mbox{\rm\sffamily E}[U-U_1]\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq & \mbox{\rm\sffamily E}[\|U-U_1\|_{\small{\mathcal{L}^2(\Omega; P)}}] \\ = & \|U-U_1\|_{\small{\mathcal{L}^2(\Omega; P)}}.
\end{align*}
By the standing assumption \eqref{eq:64}, we obtain an estimate of the
first term on the right-hand side of \eqref{eq:17},
\begin{equation}
\label{eq:18}
\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}[U_1]\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq e\equiv C_2\cdot r_1^\alpha.
\end{equation}
For the second term on the right-hand side of \eqref{eq:17}, using the
relation (\ref{eq:14}) and the definition (\ref{eq:15}), we find that
\begin{align*}
&\|\mbox{\rm\sffamily E}[U_1] - \mbox{\rm\sffamily E}^L[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} \\
=& \left\| \sum_{l=1}^{L-1}\left( \mbox{\rm\sffamily E}[U_l - U_{l+1}] -
\mbox{\rm\sffamily E}_{M_l}[U_l-U_{l+1}]\right) + \mbox{\rm\sffamily E}[U_L] - \mbox{\rm\sffamily E}_{M_L}[U_L] \right\|_{\small{\mathcal{L}^2(\Omega; P)}}\\
\leq & \sum_{l=1}^{L-1}\left\| \mbox{\rm\sffamily E}[U_l - U_{l+1}] -
\mbox{\rm\sffamily E}_{M_l}[U_l-U_{l+1}] \right\|_{\small{\mathcal{L}^2(\Omega; P)}} + \|\mbox{\rm\sffamily E}[U_L] - \mbox{\rm\sffamily E}_{M_L}[U_L] \|_{\small{\mathcal{L}^2(\Omega; P)}}\\
\leq & \sum_{l=1}^{L-1} \dfrac{\delta[U_l - U_{l+1}]}{\sqrt{M_l}} +
\dfrac{\delta[U_L]}{\sqrt{M_L}}.
\end{align*}
For $1\leq l\leq L-1$, by the standard definition of variance, we
deduce that
\begin{align*}
& \delta[U_l - U_{l+1}] = \mbox{\rm\sffamily E}[|U_l - U_{l+1}|^2] - \mbox{\rm\sffamily E}[U_l - U_{l+1}]^2\\
\leq & \mbox{\rm\sffamily E}[|U_l - U_{l+1}|^2] \leq 2\left( \mbox{\rm\sffamily E}[|U_l - U|^2] + \mbox{\rm\sffamily E}[|U_{l+1} - U|^2]\right),
\end{align*}
and, again, by
the standing assumption \eqref{eq:64},
\begin{equation}
\delta[U_l - U_{l+1}]
\leq 2C_2^2(1+4^\alpha)\cdot r_l^{2\alpha}.\label{eq:49}
\end{equation}
For the variance at the lowest resolution, $\delta[U_L]$, we again
start from the definition,
\begin{align*}
\delta[U_L] &= \|U_L-\mbox{\rm\sffamily E}[U_L]\|_{\small{\mathcal{L}^2(\Omega; P)}}
= \|U_L - U + U-\mbox{\rm\sffamily E}[U] + \mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}[U_L]\|_{\small{\mathcal{L}^2(\Omega; P)}}\\
&\leq \|U_L-U\|_{\small{\mathcal{L}^2(\Omega; P)}} + \|U-\mbox{\rm\sffamily E}[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} + \|\mbox{\rm\sffamily E}[U-U_L]\|_{\small{\mathcal{L}^2(\Omega; P)}}\\
&\leq 2\|U-U_L\|_{\small{\mathcal{L}^2(\Omega; P)}} + \|U-\mbox{\rm\sffamily E}[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq 2C_2r_L^\alpha + \delta[U].
\end{align*}
Combining the last three estimates, we obtain
\begin{align*}
\|\mbox{\rm\sffamily E}[U_1] - \mbox{\rm\sffamily E}^L[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} &\leq C_2\sqrt{2(1+4^\alpha)}
\sum_{l=1}^{L-1}\dfrac{r_l^\alpha}{\sqrt{M_l}} +
\dfrac{2C_2r_L^\alpha}{\sqrt{M_L}} +
\dfrac{\delta[U]}{\sqrt{M_L}}.
\end{align*}
Assuming that $\alpha\ge 0$, which should be true for all practically
useful numerical
schemes, we may bring the second term on the right-hand side into the
summation, and we thus obtain
\begin{equation}
\label{eq:19}
\|\mbox{\rm\sffamily E}[U_1] - \mbox{\rm\sffamily E}^L[U]\|_{\small{\mathcal{L}^2(\Omega; P)}}
\leq
C_2\sqrt{2(1+4^\alpha)}\sum_{l=1}^L\dfrac{r_L^\alpha}{\sqrt{M_l}} + \dfrac{\delta[U]}{\sqrt{M_L}}.
\end{equation}
Combining \eqref{eq:17}, \eqref{eq:18} and \eqref{eq:19} leads us to
\begin{equation}
\label{eq:20}
\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}^L[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq e +
C_2\sqrt{2(1+4^\alpha)}\sum_{l=1}^L\dfrac{r_l^\alpha}{\sqrt{M_l}} +
\dfrac{\delta[U]}{\sqrt{M_L}}.
\end{equation}
This estimate shows that the error in the
$L$-level sample mean of the quantity $U$ can be attributed to three
components: the temporal-spatial discretization error (first term), the probability
space discretization error (third term), and the error for using a
multi-level structure (the second term).
So far, the sample size at each level, $M_l$ has been left
to be determined. Determining the sample size will be a delicate balancing
act between controlling the computational cost and controlling the
error. The potential of the MLMC method lies in the
fact that, within the probability space discretization
error (the third term), the standard deviation of the analytical
solution is divided by
the sample size at the highest level (lowest resolution), where the
computational cost for an individual simulation is the
lowest. We should also note that the first term, the temporal-spatial
discretization error, is not affected by the sample size at any level.
Hence, a general principle for determining the sample size is to make
sure the probability space discretization error (the third term) and
each term in the
summation (the second term) is roughly on the order of the
temporal-spatial
discretization error (the first term), or smaller.
By the this principle, we know exactly what the sample size at the
lowest resolution should be (see \eqref{eq:10}). But to reach this
sample size starting from the highest resolution can take many
different paths. Here, we explore several different strategies for
determining the sample size at each level.
\\
\noindent{\itshape Strategy \#1}\\
Our first strategy is to choose the sample size for each level so that
each term in the summation of \eqref{eq:20} is equal or smaller than the temporal-spatial
discretization error. Hence we set
\begin{equation*}
e \equiv C_2r_1^\alpha =
C_2\sqrt{2(1+4^\alpha)}\dfrac{r_l^\alpha}{\sqrt{M_l}},
\end{equation*}
which leads to
\begin{equation}
\label{eq:21}
M_l = 2(1+4^\alpha)\cdot 2^{2\alpha(l-1)},\qquad 1\le l\le L.
\end{equation}
The number of $L$ is determined by requiring that the sample size at
the lowest resolution, $M_L$ be sufficiently large to make the
probability discretization error be on the same order as the
temporal-spatial discretization error, that is,
\begin{equation}
\label{eq:22}
\dfrac{\delta[U]}{\sqrt{M_L}} = e,
\end{equation}
from which, and \eqref{eq:21}, we deduce that
\begin{equation}
\label{eq:23}
L = 1 + \dfrac{2\log \delta[U] - 2\log e - \log
2(1+4^\alpha)}{2\alpha\log 2},
\end{equation}
or, using \eqref{eq:18},
\begin{equation}
\label{eq:54}
L = \dfrac{\log\delta[U]}{\alpha\log 2} - \dfrac{\log r_1}{\log
2} +1 - \dfrac{2\log C_2 + \log 2(1+4^\alpha)}{2\alpha\log 2}.
\end{equation}
We note from \eqref{eq:3} that
\begin{equation*}
\log r_1 = \dfrac{\log C_1 - \log N}{3}.
\end{equation*}
Substituting this expression into \eqref{eq:54} yields
\begin{equation}
\label{eq:55}
L = \dfrac{\log\delta[U]}{\alpha\log 2} + \dfrac{\log N}{3\log
2} +1 - \dfrac{\log C_1}{3\log 2} - \dfrac{2\log C_2 + \log
2(1+4^\alpha)}{2\alpha\log 2}.
\end{equation}
The expression on the right-hand size indicates that, generally, a
larger variance in the analytical solution requires more levels. Under
the same order of convergence ($\alpha$), and the same constant
coefficients $C_1$ and $C_2$, a higer number of degrees of freedom ($N$,
or, in other words, a finer mesh) also requires more levels.
Based on this strategy, the total error in the $L$-level sample mean
is
\begin{equation}
\label{eq:24}
\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}^L[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq (L+2) e.
\end{equation}
We denote the computational cost under this strategy as
$\mathcal{C}_{MLMC1}$, which can be calculated as
\begin{align*}
\mathcal{C}_{MLMC1} &= \sum_{l=1}^L \widetilde M_l N_l = \sum_{l=1}^{L-1}
M_l(N_l+N_{l+1}) + M_L N_L\\
&= \sum_{l=1}^{L-1} 2(1+4^\alpha)2^{2\alpha(l-1)}\left(8^{-(l-1)} +
8^{-l}\right)N + 2(1+4^\alpha)2^{2\alpha(L-1)}8^{-(L-1)}N\\
&= 2(1+4^\alpha)N\left(\sum_{l=1}^{L-1}
\dfrac{9}{8}2^{2\alpha(l-1)}8^{-(l-1)} +
2^{2\alpha(L-1)}8^{-(L-1)}\right) \\
&\leq \dfrac{9(1+4^\alpha)}{4}N\sum_{l=1}^L 2^{(2\alpha-3)(l-1)}.
\end{align*}
If $\alpha < 3/2$, then the summation on the right-hand side
increases monotonically as $L$ increases, and converges to a finite
number as $L$ tends to infinity, with the limit depending on the
convergence rate $\alpha$ only. Thus, in this case, the computational
cost grows linearly as $N$ increases.
\begin{equation}
\label{eq:66}
\mathcal{C}_{MLMC1} = O(N).
\end{equation}
If $\alpha = 3/2$, then
\begin{equation*}
\mathcal{C}_{MLMC1} \leq \dfrac{9(1+4^\alpha)}{4} N L.
\end{equation*}
With $L$ as given in \eqref{eq:55}, we conclude that
\begin{equation}
\label{eq:56}
\mathcal{C}_{MLMC1} = O(N(\log\delta[U] + \log N)).
\end{equation}
We note that the case where $\alpha = 3/2$ {\it exactly} is rare in
practice. But the result obtained here, together with the result for
the $\alpha < 3/2$, indicates that, with larger $\alpha$, the
computational cost will increase faster as $N$ increases.
Finally, if $\alpha > 3/2$, then
\begin{equation*}
\mathcal{C}_{MLMC1} = \dfrac{9(1+4^\alpha)}{4}N\cdot\dfrac{2^{(2\alpha - 3)L}
- 1}{2^{2\alpha -3} - 1}.
\end{equation*}
Upon substituting the expression \eqref{eq:55} for $L$ in the above,
we obtain that
\begin{equation}
\label{eq:57}
\mathcal{C}_{MLMC1} = O\left( \delta[U]^\frac{2\alpha-3}{\alpha} \cdot N^{1+\frac{2\alpha-3}{3}}\right).
\end{equation}
In this case, the computational cost grows polynomially in $N$ and
$\delta[U]$, similar to the situation with the classical Monte Carlo
method (see
\eqref{eq:12}), but the exponents on both $\delta[U]$ and $N$ are
lower in the case here, indicating that, even if the convergence rate
$\alpha$ is greater than $3/2$, there still are potential savings in
computational time by choosing the MLMC method.
The problem with this strategy is that the error depends on the number
of levels, which may be large. In the following strategies, we amply
$M_l$ by certain factors so that the summation in \eqref{eq:20}
actually converges even as the number of levels goes to infinity, so
that the final error is actually independent of the number of levels
taken.\\
\noindent{\itshape Strategy \#2}\\
Under this strategy, we make the error term in the summation on the
right-hand side of \eqref{eq:20} decrease exponentially as the level
number $l$ goes up, that is, we set
\begin{equation*}
\dfrac{\sqrt{2(1+4^\alpha)}\cdot r_l^\alpha}{\sqrt{M_l}} =
\left(\dfrac{1}{2}\right)^{l-1} r_1^\alpha,
\end{equation*}
which leads to
\begin{equation}
\label{eq:26}
M_l = 2(1+4^\alpha)\cdot 4^{(l-1)(\alpha+1)}.
\end{equation}
To ensure that the error term due to the inherent variance of the
system be on the same level as the discretization error, we require
that
\begin{equation*}
\dfrac{\delta[U]}{\sqrt{M_L}} = e,
\end{equation*}
from which we infer that
\begin{equation*}
M_L = \left(\dfrac{\delta[U]}{e}\right)^2.
\end{equation*}
Using the formula \eqref{eq:26} for $M_L$, we obtain a lower bound for
the number of levels required,
\begin{equation}
\label{eq:27}
L = \dfrac{2\log\delta[U] - 2\log e - \log
2(1+4^\alpha)}{(\alpha+1)\log 4} + 1,
\end{equation}
or, using \eqref{eq:18},
\begin{equation}
\label{eq:28}
L = \dfrac{2\log\delta[U] - 2\alpha\log r_1 - 2\log C_2- \log
2(1+4^\alpha)}{(\alpha+1)\log 4} + 1.
\end{equation}
We note that, from \eqref{eq:3},
\begin{equation}
\label{eq:29}
\log r_1 = \dfrac{\log C_1 - \log N}{3}.
\end{equation}
Hence, we have
\begin{equation}
\label{eq:30}
L = \dfrac{2}{(\alpha+1)\log 4}\log\delta[U] +
\dfrac{2\alpha}{3(\alpha + 1)\log 4}\log N + 1
-
\dfrac{\frac{2\alpha}{3}\log C_1 + 2\log C_2 + \log
2(1+4^\alpha)}{(\alpha+1)\log 4}.
\end{equation}
This expression indicates that, generally, large variance in the
analytic solution requires more levels. It is also clear from the
expression that, under the same convergence rate
$\alpha$, and the same constants $C_1$ for computational cost and
$C_2$ for the error, finer mesh (larger $N$) will also requires more
levels.
Under this strategy, the error in the $L$-level mean is independent of
the number of levels, for
\begin{equation}
\label{eq:32}
\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}^L[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq e\left\{1 + \sum_{l=1}^L
\left(\dfrac{1}{2}\right)^{l-1} + 1\right\} \leq 4e.
\end{equation}
We denote the computational cost under this strategy by
$\mathcal{C}_{MLMC2}$. It is calculated as follows,
\begin{align*}
\mathcal{C}_{MLMC2} &= \sum_{l=1}^L \widetilde M_l N_l
= M_1 N_1 + \sum_{l=2}^L (M_{l-1}+M_l)N_l\\
&= 2(1+4^\alpha)N + \sum_{l=2}^L 2(1+4^\alpha)\left(
4^{(l-2)(\alpha+1)} + 4^{(l-1)(\alpha+1)}\right)\cdot 8^{-(l-1)}
N\\
& = 2(1+4^\alpha)(1+4^{-(\alpha+1)}) N \sum_{l=1}^L 2^{(l-1)(2\alpha -1)}.
\end{align*}
If $\alpha < 1/2$, then the summation on the right-hand side increases
monotonically as $L$ increases, and converges to a limit as $L$ tends
to infinity. The limit depends on the convergence rate $\alpha$ only.
Thus, in this case, the computational cost $\mathcal{C}_{MLMC2}$ grows linearly
in $N$. Specifically,
\begin{equation}
\label{eq:58}
\mathcal{C}_{MLMC2} \leq \dfrac{2(1+4^\alpha)(1+4^{-(\alpha+1)})}{1-2^{2\alpha
- 1}} N.
\end{equation}
\begin{equation}
\label{eq:67}
\mathcal{C}_{MLMC2} = O(N).
\end{equation}
If $\alpha = 1/2$, then
\begin{equation*}
\mathcal{C}_{MLMC2} = 2(1+4^\alpha)(1+4^{-(\alpha+1)}) NL.
\end{equation*}
With $L$ as given in \eqref{eq:30}, we conclude that
\begin{equation}
\label{eq:59}
\mathcal{C}_{MLMC2} = O(N(\log\delta[U] + \log N)).
\end{equation}
We now consider the more common scenario where $\alpha > 1/2$.
\begin{equation}
\label{eq:31}
\mathcal{C}_{MLMC2} = 2(1+4^\alpha)(1-4^{-(\alpha + 1)}) N\cdot
\dfrac{2^{(2\alpha -1)L}-1}{2^{2\alpha - 1} - 1}.
\end{equation}
Substitute the expression \eqref{eq:30} for $L$ into the above, we
obtain that
\begin{equation}
\label{eq:33}
\mathcal{C}_{MLMC2} = C_3 \cdot
\dfrac{2(1+4^\alpha)\cdot(1-4^{-(\alpha+1)})}{2^{(2\alpha -1 )} -
1}\delta[U]^{\frac{2\alpha -1}{\alpha+1}}\cdot
N^{1+\frac{\alpha(2\alpha -1)}{3(\alpha +1)}}.
\end{equation}
\begin{equation}
\label{eq:25}
\mathcal{C}_{MLMC2} = O\left(\delta[U]^{\frac{2\alpha -1}{\alpha+1}}\cdot
N^{1+\frac{\alpha(2\alpha -1)}{3(\alpha +1)}}\right).
\end{equation}
Comparing with the conventional MC method, the computational
cost for the MLMC method under the current strategy still grows
polynomially as $N$ increases, but at a lower degree, for it is
trivial to verify that
\begin{equation*}
\dfrac{\alpha(2\alpha-1)}{3(\alpha+1)} \leq \dfrac{2}{3}.
\end{equation*}
The impact of the variance in the
analytical solution on the computational cost is also lower, for it is
obvious that
\begin{align*}
\dfrac{2\alpha -1}{\alpha + 1} &\leq 2.
\end{align*}
\noindent{\itshape Strategy \#3}\\
This strategy chooses a sample size so that each term on the right-hand
side of \eqref{eq:20}, including the individual terms in the
summation, contributes equally to the total error, and then amplify
the sample size by a level dependent factor to ensure
convergence. With $\sigma>0$ being a positive parameter, we
let
\begin{equation}
\label{eq:34}
M_l = 2(1+4^\alpha)(L-l+1)^{2(1+\sigma)}\cdot 2^{2\alpha(l-1)}.
\end{equation}
As before, the number of levels is determined by requiring that the sample size
at the highest level satisfies the relation \eqref{eq:22}, which leads
to
\begin{equation}
\label{eq:35}
L = \dfrac{2\log\delta[U] - 2\log e - \log
2(1+4^\alpha)}{2\alpha\log 2} + 1,
\end{equation}
or, using \eqref{eq:18},
\begin{equation}
\label{eq:36}
L = \dfrac{\log\delta[U]}{\alpha\log 2} - \dfrac{\log r_1}{\log
2} +1 - \dfrac{2\log C_2 + \log 2(1+4^\alpha)}{2\alpha\log 2}.
\end{equation}
We note from \eqref{eq:3} that
\begin{equation*}
\log r_1 = \dfrac{\log C_1 - \log N}{3}.
\end{equation*}
Substituting this expression into \eqref{eq:36} yields
\begin{equation}
\label{eq:38}
L = \dfrac{\log\delta[U]}{\alpha\log 2} + \dfrac{\log N}{3\log
2} +1 - \dfrac{\log C_1}{3\log 2} - \dfrac{2\log C_2 + \log
2(1+4^\alpha)}{2\alpha\log 2}.
\end{equation}
The expression on the right-hand size indicates that, generally, a
larger variance in the analytical solution requires more levels. Under
the same order of convergence ($\alpha$), and the same constant
coefficients $C_1$ and $C_2$, a higher number of degrees of freedom ($N$,
or, in other words, a finer mesh) also requires more levels.
Under this strategy, the total error \eqref{eq:20} in the sample mean
can be estimated,
\begin{align*}
&\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}^L[U]\|_{\small{\mathcal{L}^2(\Omega; P)}}
\leq e \left\{ 1 + \sum_{l=1}^L (L-l+1)^{-(1+\sigma)} +
1\right\}
= e \left\{ 2 + \sum_{l=1}^L l^{-(1+\sigma)}\right\}.
\end{align*}
We note that, thanks to the positiveness of the parameter $\sigma$,
the summation converges even as $L$ tends to infinity. We can bound
the summation by an integral, and we have
\begin{equation}
\label{eq:39}
\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}^L[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq e \left(3 + \int_1^\infty
l^{-(1+\sigma)} dl\right) \leq \left(3+\dfrac{1}{\sigma}\right)
e.
\end{equation}
The computational cost $\mathcal{C}_{MLMC3}$ can also be estimated,
\begin{align*}
\mathcal{C}_{MLMC3} =& \sum_{l=1}^L \widetilde M_l N_l= \sum_{l=1}^L(M_{l-1} + M_l) N_l= \sum_{l=1}^{L-1} M_l(N_l + N_{l+1}) + M_L N_L\\
=& 2(1+4^\alpha)N\left\{\sum_{l=1}^{L-1} (L-l +1)^{2(1+\sigma)}\cdot
2^{(2\alpha-3)(l-1)}\cdot (1+2^{-3}) +
2^{2\alpha-3)(L-1)}\right\}.
\end{align*}
We note that the last term in the curly bracket can be rolled over
into the summation, and an inequality follows,
\begin{equation}
\label{eq:40}
\mathcal{C}_{MLMC3} \leq\dfrac{9(1+4^\alpha)}{4} N \sum_{l=1}^L (L-l+1)^{2(1+\sigma)}\cdot
2^{(2\alpha -3)(l-1)}.
\end{equation}
If $\alpha < 3/2$, then
\begin{align*}
\mathcal{C}_{MLMC3} \leq& \dfrac{9(1+4^\alpha)}{4} N L^{2(1+\sigma)}
\sum_{l=1}^L 2^{2(\alpha-3)(l-1)}
\leq& \dfrac{9(1+4^\alpha)}{4} N
L^{2(1+\sigma)}\cdot\dfrac{1}{1-2^{2\alpha -3}}\\
=& \dfrac{9(1+4^\alpha)}{4(1-2^{2\alpha -3})} N
L^{2(1+\sigma)}.
\end{align*}
Substituting \eqref{eq:38} into the expression above, we find that
\begin{equation}
\label{eq:41}
\mathcal{C}_{MLMC3} \leq \dfrac{9(1+4^\alpha)}{4(1-2^{(2\alpha -3)})}\cdot
N\cdot \left(\dfrac{\log\delta[U]}{\alpha\log 2} + \dfrac{\log
N}{3\log 2} + C\right)^{2(1+\sigma)}.
\end{equation}
From the above, we conclude that
\begin{equation}
\label{eq:42}
\mathcal{C}_{MLMC3} \sim N\cdot (\log\delta[U]+\log N )^{2(1+\sigma)}.
\end{equation}
If $\alpha = 3/2$, then
\begin{align*}
\mathcal{C}_{MLMC3} =& \dfrac{9(1+4^\alpha)}{4}\cdot N \cdot \sum_{l=1}^L
l^{2(1+\sigma)}
\leq \dfrac{9(1+4^\alpha)}{4} \cdot N \cdot \int_1^{L+1}
l^{2(1+\sigma)} dl\\
\leq & \dfrac{9(1+4^\alpha)}{4(2\sigma +3)} \cdot N \cdot \left\{
(L+1)^{2\sigma +3} \right\}
= \dfrac{9(1+4^\alpha)}{4(2\sigma +3)} \cdot N \cdot \left(
\dfrac{\log\delta[U]}{\alpha\log 2} + \dfrac{\log N}{3\log 2} +
C\right)^{2\sigma + 3},
\end{align*}
where
\begin{equation*}
C = 2 - \dfrac{\log C_1}{3\log 2} - \dfrac{2\log C_2 + \log
2(1+4^\alpha)}{2\alpha\log 2}.
\end{equation*}
Therefore, for this case,
\begin{equation}
\label{eq:43}
\mathcal{C}_{MLMC3} \sim N\cdot(\log\delta[U] + \log N)^{2\sigma + 3}.
\end{equation}
If $\alpha > 3/2$, then the situation is more complicated.
\begin{align*}
\mathcal{C}_{MLMC3} &\leq\dfrac{9(1+4^\alpha)}{4} N \sum_{l=1}^L
(L-l+1)^{2(1+\sigma)}\cdot
2^{(2\alpha -3)(l-1)}\\
&\leq \dfrac{9(1+4^\alpha)}{4} N L^{2(1+\sigma)} \sum_{l=1}^L
2^{(2\alpha -3)(l-1)}.
\end{align*}
Using the expression \eqref{eq:38}, we determine that
\begin{equation}
\label{eq:60}
\mathcal{C}_{MLMC3} = O\left((\log\delta[U] + \log N)^{2(1+\sigma)}\cdot
\delta[U]^\frac{2\alpha -3}{\alpha}\cdot N^{1+\frac{2\alpha - 3}{3}}\right).
\end{equation}
This resembles the situation under Strategy \#2, and the cost grows
polynomially as $N$ increases.
\\
\noindent{\itshape Strategy \#4}\\
It is similar to Strategy \#3, but the sample size are amplified at
higher levels (lower resolutions). We set
\begin{equation}
\label{eq:45}
M_l = 2(1+4^\alpha)\cdot l^{2(1+\sigma)} \cdot 2^{2\alpha(l-1)}.
\end{equation}
To determine the number of levels $L$, we require $M_L$ to satisfy the
relation \eqref{eq:22},
\begin{align}
2(1+4^\alpha)\cdot L^{2(1+\sigma)}\cdot 2^{2\alpha(L-1)} = &
\dfrac{\delta[U]^2}{e^2},\nonumber\\
L^{2(1+\sigma)}\cdot 2^{2\alpha(L-1)} =&
\dfrac{\delta[U]^2}{e^2\cdot 2(1+4^\alpha)}.\label{eq:46}
\end{align}
The number of levels cannot be solved for explicitly from
\eqref{eq:46}. But it is clear that it is smaller than that of
Strategy
\#3. This strategy leads to the same total error in the $L$-level
sample mean,
\begin{equation}
\label{eq:47}
\|\mbox{\rm\sffamily E}[U] - \mbox{\rm\sffamily E}^L[U]\|_{\small{\mathcal{L}^2(\Omega; P)}} \leq
\left(3 + \dfrac{1}{\sigma}\right) e.
\end{equation}
We denote the computational cost under this strategy by $\mathcal{C}_{MLMC4}$,
\begin{align*}
\mathcal{C}_{MLMC4} =& \sum_{l=1}^L \widetilde M_l \cdot N_l
= \sum_{l=1}^{L-1} M_l\cdot (N_l + N_{l+1}) + M_L\cdot N_L\\
\leq & \dfrac{9(1+4^\alpha)}{4}\cdot N\cdot \sum_{l=1}^L
l^{2(1+\sigma)}\cdot 2^{(2\alpha-3)(l-1)}.
\end{align*}
If $\alpha < 3/2$, then the summation on the right-hand side
converges. The limit, denoted by $C_{\alpha,\sigma}$, depends on the
parameters $\alpha$ and $\sigma$ only. Thus we have the estimate
\begin{equation}
\label{eq:61}
\mathcal{C}_{MLMC4} \leq \dfrac{9(1+4^\alpha)}{4} C_{\alpha,\sigma}N.
\end{equation}
\begin{equation}
\label{eq:44}
\mathcal{C}_{MLMC4} = \mathcal{O}(N).
\end{equation}
If $\alpha = 3/2$, then
\begin{align*}
\mathcal{C}_{MLMC4} &= \dfrac{9(1+4^\alpha)}{4}\cdot N\sum_{l=1}^L
l^{2(1+\sigma)}
\leq \dfrac{9(1+4^\alpha)}{4(2\sigma +3)}\cdot N\cdot
(L+1)^{2\sigma + 3}.
\end{align*}
The cost is the same as $\mathcal{C}_{MLMC3}$ for the same value of $\alpha$.
\begin{equation}
\label{eq:65}
\mathcal{C}_{MLMC4} =\mathcal{O}\left( N\cdot(\log\delta[U] + \log N)^{2\sigma + 3}\right).
\end{equation}
If $\alpha > 3/2$, then
\begin{equation}
\label{eq:48}
\mathcal{C}_{MLMC4} = \dfrac{9(1+4^\alpha)}{4} \cdot N\cdot \sum_{l=1}^L
l^{2(1+\sigma)} \cdot 2^{(2\alpha-3)(l-1)}.
\end{equation}
In this case, $\mathcal{C}_{MLMC4}$ is greater than $\mathcal{C}_{MLMC3}$, but shares the
same estimate, that is,
\begin{equation}
\label{eq:62}
\mathcal{C}_{MLMC4} = \mathcal{O}\left(N\cdot(\log\delta[U] + \log N)^{2(1+\sigma)}\cdot
\delta[U]^\frac{2\alpha -3}{\alpha}\cdot N^\frac{2\alpha - 3}{3}\right).
\end{equation}
\begin{sidewaystable}[H]\small
\centering
\ra{1.3}
\begin{tabular}{@{}c|cccccc@{}}\toprule
{ } & Linear & Quasilinear & Polynomial \\
\midrule
Classical MC & & & $\mathcal{O}(\delta[U]^2 N^{1+\frac{2}{3}\alpha})$ \\
& & & \\
\midrule
Strategy \#1 & $\mathcal{O}(N)$ & $\mathcal{O}\left(N(\log\delta[U]+\log
N)\right)$ & $\mathcal{O}\left(\delta[U]^\frac{2\alpha
-3}{\alpha}\cdot N^{1+\frac{2\alpha-3}{3}}\right)$\\
{ } & $(\alpha <3/2)$ & $(\alpha = 3/2)$ & $(\alpha > 3/2)$ \\
\midrule
Strategy \#2 & $\mathcal{O}(N)$ & $\mathcal{O}\left(N(\log\delta[U]+\log
N)\right)$ & $\mathcal{O}\left(\delta[U]^\frac{2\alpha
-1}{\alpha+1}\cdot N^{1+\frac{\alpha(2\alpha-1)}{3(\alpha+1)}}\right)$\\
{ } & $(\alpha <1/2)$ & $(\alpha = 1/2)$ & $(\alpha > 1/2)$ \\
\midrule
Strategy \#3 & $\mathcal{O}\left(N\cdot(\log\delta[U] + \log
N)^{2(1+\sigma)}\right)$ & $\mathcal{O}\left(N\cdot(\log\delta[U] + \log
N)^{2\sigma+3}\right)$ & $\mathcal{O}\left((\log\delta[U] + \log
N)^{2(1+\sigma)}\cdot\delta[U]^\frac{2\alpha-3}{\alpha}\cdot N^{1+\frac{2\alpha-3}{3}}\right)$\\
{ } & $(\alpha <3/2)$ & $(\alpha = 3/2)$ & $(\alpha > 3/2)$ \\
\midrule
Strategy \#4 & $\mathcal{O}(N)$ & $\mathcal{O}\left(N(\log\delta[U]+\log
N)^{2\sigma + 3}\right)$ & $\mathcal{O}\left((\log\delta[U] + \log
N)^{2(1+\sigma)}\cdot\delta[U]^\frac{2\alpha-3}{\alpha}\cdot
N^{1+\frac{2\alpha-3}{3}}\right)$ \\
{ } & $(\alpha <3/2)$ & $(\alpha = 3/2)$ & $(\alpha > 3/2)$ \\
\bottomrule
\end{tabular}
\caption{Comparison of growth rates for the classical MC
method and the MLMC method under different
strategies. $\delta[U]$ represents the standard deviation in the
true solution, $N$ the computational cost of an individual
simulation at the highest resolution, $\alpha$ the convergence
rate, and $\sigma$ an arbitrary positive parameter chosen by the
user. }
\label{tab:comparison}
\end{sidewaystable}
The computational cost for each strategy, as well as the cost for the
classical MC method, are summarized in Table \ref{tab:comparison}. All
strategies, except Strategy \#3, experience three stages of cost
growth, depending on the convergence rate $\alpha$: linear,
quasi-linear, and polynomial. When the convergence rate is high, the
computational cost for all strategies grow polynomially, similar to
the situation of the classical MC method. But the degrees of the
polynomials are lower, offering potential savings in computing
times. Strategy \#3 appears disadvantage in that it lacks linear
growth for the computational cost, apparently due to the fact that the
sample size at the lowest level (highest resolution) is amplified.
\subsection{Estimates}\label{sec:estimates}
Under each one of the strategies discussed above, the calculation of
the number of levels, the sample size at each level, the error in the
$L$-level sample mean, and the computational cost depend on a few key
parameters, namely $\delta[U]$, the standard deviation in the true
solution, $\alpha$, the convergence rate of the numerical scheme
regarding the QoI, and $e$, the
$L^2$-norm of the error in the first approximation $U_1$. Determining the
true values of these parameters touches upon several fundamental
mathematical and numerical issues that, in many cases involving
real-world applications, are completely open. For example, for many
nonlinear systems, e.g.,~the three-dimensional Navier-Stokes equations
governing fluids, the existence and uniqueness of a global solution is
still an open question. Similarly, the numerical analysis to determine
the convergence rate of numerical schemes for nonlinear systems is
very challenging, even not possible. We leave these theoretical issues
to future endeavors. In the current work, we explore approaches to
estimate these parameters from the discrete simulation data.
The standard deviation $\delta[U]$ in the true solution can be
approximated by the unbiased sample variance (\cite{Kenney:1951ta}),
\begin{equation}
\label{eq:37}
\delta[U] \approx \dfrac{1}{M_l -1} \sum_{m=1}^{M_l} \left( U^m_l
- \mbox{\rm\sffamily E}_{M_l}[U_l]\right)^2.
\end{equation}
The convergence rate $\alpha$ cannot be calculated directly using the
$L^2$-norm of the error in $U_l$ and the
relation \eqref{eq:64}, since the true solution $U$ is not
available. Instead, we use the standard deviation of the difference
between solutions at two consecutive levels, i.e.~$\delta[U_l -
U_{l+1}]$.
Instead
of the coefficient $\sqrt{2(1+4^\alpha)}C_2$ on the right-hand side of
\eqref{eq:49}, we assume that there exists another constant $C_3$ such
that
\begin{equation}
\label{eq:50}
\delta[U_l - U_{l+1}] = C_3 r_l^\alpha.
\end{equation}
The computation of $\alpha$ will not be affected by the value of
$C_3$, since
\begin{equation}
\label{eq:63}
\dfrac{\delta[U_1 - U_2]}{\delta[U_2 - U_3]} =
\left(\dfrac{r_1}{r_2}\right)^\alpha = \left(\dfrac{1}{2}\right)^\alpha.
\end{equation}
Of course, in actual calculations, the standard deviation on the
left-hand side of \eqref{eq:50} will be replaced by the square root of
the unbiased sample variance (formula \eqref{eq:37}).
The $L^2$-norm of the error in the first approximation
$U_1$, $e$, cannot be calculated directly from \eqref{eq:64} either,
due to
the lack of the true solution $U$.
Instead, using \eqref{eq:49}, and the convergence rate just computed,
we can obtain an estimate on $e$,
\begin{equation}
\label{eq:51}
e = \dfrac{\delta[U_1-U_2]}{\sqrt{2(1+4^\alpha)}}.
\end{equation}
\section{Numerical experiments using ACC}\label{sec:numer-exper-using}
The Antarctic Circumpolar Current (ACC) is a circular current
surrounding the Antarctic continent. It is the primary channel through
which the world's oceans (Atlantic, Indian, and Pacific)
communicate. Thanks to the predominant westerly wind in that region,
the current flows from west to east. The ACC is the strongest current
in the world, volume-wise. It is estimated that the volume transport is
about 135 Sv (1 Sv = $10^6$ $\mathrm{m}^3\,\mathrm{s}^{-1}$) through
the Drake passage
(\cite{Gent:2001fg,{Hughes:1999iu},{Warren:1996cx}}), which is about
135 times the total volume transport of all the rivers in the
world. The above estimate is a time average; the actual volume
transport oscillate on seasonal and intradecadal scales.
\begin{figure}[h]
\centerline{\includegraphics[width=4in]{bottom-topo-sample1rubber.pdf}}
\caption{The random bottom topography sample \# 1.}
\label{fig:topo-sample}
\end{figure}
Here, we demonstrate how the MLMC method can be
combined with an ocean circulation model to quantify the volume
transport of the ACC. In order to stay focused on the methodology that
is being explored here, we sharply reduce the physics of this problem
while still retain its essential features. The fluid domain is a
re-entrant rectangle that is $2,000$ km long and $1,733$ km wide; the same size
as \cite{Chen:2016cr}, and also see \cite{McWilliams:1981up}.
The flow
is governed by a three-layer isopycnal model, which reads
\begin{equation}
\label{eq:52}
\left\{\begin{aligned}
& \dfrac{\partial h_i}{\partial t} + \nabla\cdot\left(h_i \mathbf{u}_i\right) = 0,\\
& \dfrac{\partial\mathbf{u}_i}{\partial t} + h_iq_i\bs{\mathrm{k}}\times\mathbf{u}_i =
-\nabla\left(\dfrac{\phi_i}{\rho_0}+K_i\right) + \mathbf{D}_i + \mathbf{F}_i, \\
& \dfrac{\partial}{\partial t} (h_i\sigma_i) + \nabla\cdot\left(h_i\sigma_i \mathbf{u}_i\right) = 0,\\
\end{aligned}\right.
\end{equation}
where $i=1,2,3$ is the layer index starting at the ocean surface. The prognostic variables $h_i$,
$\mathbf{u}_i$ and $\sigma_i$ denote the layer thickness, horizontal
velocity, and some tracer
respectively, and the diagnostic variables $q_i$, $\phi_i$ and $K_i$
denote the potential vorticity, Montgomery potential and kinetic
energy, respectively, and they are defined as
\begin{align*}
& q_i = \dfrac{\nabla\times\mathbf{u}_i + f}{h_i},\qquad i=1,2,3, \\
& K_i = \frac{1}{2}|\mathbf{u}_i|^2,\qquad i=1,2,3,\\
& \phi_1 = p_0 + \rho_1g(h_1+h_2+h_3+b),\\
& \phi_2 = \phi_1 + (\rho_2-\rho_1)g(h_2+h_3+b),\\
& \phi_3 = \phi_2 + (\rho_3-\rho_2)g(h_3+b),
\end{align*}
with $p_0$ denoting the surface pressure and $b$ the bathymetry;
$\mathbf{D}_i$ denotes the horizontal viscous diffusion, which usually takes the form
of harmonic or biharmonic diffusion.
The external forcing term $\mathbf{F}_i$ for
each layer is specified as follows,
\begin{equation}
\label{eq:53}
\mathbf{F}_i = \left\{
\begin{aligned}
&\dfrac{\boldsymbol{\tau}}{\rho_1 h_1} \textrm{ (wind stress) }, & &i = 1,\\
& 0, & &i =2,\\
&-\mathbf{d}\textrm{ (bottom drag) }, & & i=3.
\end{aligned}\right.
\end{equation}
The model is made up of
three isopycnal layers with mean layer thickness of 500 m, 1250 m and
3250 m and with densities of $1010\, \textrm{kg}\,\textrm{m}^{-3}$, $1013\, \textrm{kg}\,\textrm{m}^{-3}$
and $1016\,\textrm{kg}\,\textrm{m}^{-3}$. The system is forced by a zonal wind stress
on the top layer with the form
$$\tau = \tau_0\sin^2 \left(\dfrac{\pi y}{L_y}\right),$$
where $\tau_0 = 0.1\textrm{N}\,\textrm{m}^{-2}$. The uncertainty in the model
is presented by the bottom topography. We assume that the bottom
topography of the domain is largely flat with small but random
features,
\begin{equation*}
b = \sum_{k,l = 4}^{20} \dfrac{H}{k^2 + l^2}\left(a_{kl}(\omega)
\cos\left(\dfrac{2\pi k x}{L_x}\right) +
b_{kl}(\omega)\sin\left(\dfrac{2\pi k
x}{L_x}\right)\right)\sin\left(\dfrac{l\pi y}{Ly}\right),
\end{equation*}
where $a_{kl}(\omega)$ and $b_{kl}(\omega)$ are random variables. Thus the
bottom is controlled by 578 random parameters. One sample of the
topography is shown in Figure \ref{fig:topo-sample}.
Similar types of bottom
topography profiles have been used by \cite{TREGUIER:1990wc}.
The numerical simulations are conducted using the MPAS isopycnal ocean
model (\cite{Ringler:2013vw}). MPAS, which stands for Model Prediction
Across Scales, implements a C-grid finite difference / finite volume
scheme that is detailed in \cite{Thuburn:2009tb,
{Ringler:2010io}}. MPAS utilizes arbitrarily unstructured
Delaunay-Voronoi tessellations (\cite{Du:1999gs, {Du:2003ka}}). For
this experiment, we have four levels of resolutions available: 10 km,
20 km, 40 km, and 80 km. To account for the effect of the unresolved
eddies, the biharmonic hyperviscosity is used. The viscosity
parameters are chosen to minimize the diffusive effect while still
ensure a stable simulation. For the aforementioned resolutions, the
viscosity parameters are $10^9\,\mathrm{m}^4\mathrm{s}^{-1}$,
$10^{10}\,\mathrm{m}^4\mathrm{s}^{-1}$,
$10^{11}\,\mathrm{m}^4\mathrm{s}^{-1}$,
and $10^{12}\,\mathrm{m}^4\mathrm{s}^{-1}$, respectively. At the
coarsest resolution (80 km), the Gent-McWilliams closure
(\cite{Gent:1990jt,{Gent:1995fa}}) is turned on,
with a constant parameter $400\,\mathrm{m}^2\mathrm{s}^{-1}$, to
account for the cross-channel transport and to prevent the top fluid layer
thickness from thinning to zero. GM is not used in any other higher
resolution simulations. The configurations for each mesh resolution
are summarized in Table \ref{tab:resolutions}.
Each simulation is run for 40 years to spin up the current. The output
data are saved every 10 days for the next 10 years.
\begin{table}[h]\small
\centering
\begin{tabular}[h]{@{}c|cccc@{}}\toprule
{ } & Eddy closures & Spatial DOFs & Time step (s) & Processes\\
\midrule
10 km & Hyperviscosity & 480,000 & 45 & 64 \\
\midrule
20 km & Hyperviscosity & 120,000 & 90 & 16 \\
\midrule
40 km & Hyperviscosity & 30,000 & 180 & 4 \\
\midrule
80 km & Hyper. + GM & 7,500 & 360 & 1 \\
\bottomrule
\end{tabular}
\caption{The configurations for each resolution. The spatial degrees
of freedom (DOFs) is calculated as (number of cells + number of
edges) $\times$ number of layers.}
\label{tab:resolutions}
\end{table}
\begin{figure}[h]
\centering
\includegraphics[width=4.5in]{vorticity-y40-sample1.pdf}
\caption{The snapshots of the vorticity field at year 40,
computed with the random bottom topography sample \#1.}
\label{fig:vorticity-y40}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=4in]{trend.png}
\caption{The changes of volume transport across a subset of the sample space.}
\label{fig:trend}
\end{figure}
The interior of large-scale geophysical flows has Reynolds numbers on
the order of $10^{20}$. Thus the large-scale geophysical flows are
turbulent in nature, and mesoscale and submesoscale eddy activities
are important part of the ocean dynamics
(\cite{{McWilliams:1985fu},{Danabasoglu:1994uh},{Levy:2001dx},
{Holland:2010jk}, {Brannigan:2015hz}}). For turbulent flows, the
pointwise instantaneous behavior of the flow is not
reliable anymore. But one can hope that observing the flow long enough
can reveal reliable and useful statistics about the flow. Figure
\ref{fig:vorticity-y40} shows the snapshots of the relative vorticity
field on Year 40 for mesh resolutions with the same bottom topography
profile. The highest resolution, 10km (Panel (a)), depicts a scene of
rapid mixing
by a wide range of mesoscale and submesoscale eddies.
As the mesh gets coarser, the level of eddy activities decrease. The
comparison also makes it clear that these flows are largely
independent of each other, for there appears to be no correlation
between the basic flow patterns of these simulations, other than the
fact that they are all west-to-east flows driven by a common
windstress.
However, a comparison of the volume transport by these simulations
over a common set of topographic profiles tells a different and
reassuring story. In Figure (\ref{fig:trend}), each curve
represents results on one mesh resolution. While on any particular
topography profile, the results from different resolutions do not agree,
the curves across all the 20 samples, especially those
for the 10 km, 20 km, and 40 km, largely follow the same pattern.
The agreement of the patterns of the curves
indicates that a great deal of information in the curve for the
highest resolutions is actually available in the curves of the lower
resolutions, and this is a vindication for the multi-level method
that we are pursuing here.
\begin{figure}[h]
\centering
\includegraphics[width=5in]{convergence_mean_std_20km_rate.pdf}
\caption{Sample mean and variance and the convergence rate. The mean
and variance are estimated using samples from the 20km simulations.}
\label{fig:rate}
\end{figure}
In order to set up the MLMC simulations under the
various strategies proposed before, three key parameters are needed:
the standard deviation $\delta[U]$ in the true solution, the error
$e$ in the
finest solutions, and the convergence rate $\alpha$. To fully
determine these key parameters requires the true solution itself $U$,
which is not available in any practical applications. But they can
easily estimated (see Section \ref{sec:estimates}). Using data from
the 20 km simulations, we compute the MC mean and the standard
deviation according to the formulae \eqref{eq:5} and \eqref{eq:37}. To
probe the sensitivity of these estimates to the sample sizes, we
compute the quantities with several independent sample sets with
varying sizes, and the results are shown in Figure \ref{fig:rate}
(left panel). Based on this figure, we take
\begin{equation*}
\delta[U] \approx 7.36\times 10^7\textrm{ m}^3/\textrm{s}.
\end{equation*}
Using the formula \eqref{eq:63} and data from 10 km, 20 km, 40 km, and
80 km simulations (Figure \ref{fig:rate} (right panel)), the
convergence rate $\alpha$ is estimated to be
\begin{equation*}
\alpha \approx 1.07.
\end{equation*}
This convergence appears slow but expected for long term simulations of
turbulent flows. The underlying numerical scheme, namely a C-grid
finite volume scheme, has been found to be accurate of orders $1\sim
2$ for laminar flows (\cite{Ringler:2010io}, also see
\cite{Chen:2013bl}). Finally the error $e$ in the finest solutions
is calculated using the formula \eqref{eq:51} and data from the 10km
and 20km simulations,
\begin{equation*}
e \approx 9.60\times 10^6\textrm{ m}^3/\textrm{s}.
\end{equation*}
\begin{table}[h]\small
\centering
\ra{1.3}
\begin{tabular}{@{}c|ccccc|c@{}}\toprule
{ } & $L$ & $m_1$ & $m_2$ & $m_3$ & $m_4$ & Comp.~load \\
\midrule
Classical MC & n/a & 59 & & & & 59 \\
\midrule
Strategy \#1 & 3 & 11 & 48 & 210 & & 22.4\\
\midrule
Strategy \#2 & 2 & 11 & 191 & & & 36.3\\
\midrule
Strategy \#3 & 3 & 876 & 763 & 210 & & 1096.1\\
\midrule
Strategy \#4 & 2 & 11 & 763 & & & 107\\
\bottomrule
\end{tabular}
\caption{Multi-level setup under different strategies}
\label{tab:setup}
\end{table}
Using these estimated parameters and the formulae set forth under
various strategies propose in the previous section, we calculate the
number of levels, the sample size at each level, and the computational
load for each strategy for the multi-level method. The sample size and
computational load for
the classical Monte Carlo method are also calculated. The
computational load are calculated in terms of the the computational
load for one single simulation at the highest resolution (lowest
level). The issue of efficiency, overhead, etc.~are neglected. For
example, the classical MC method requires 59 simulations at
the highest resolution, and therefore its computational load is
59. The results are listed in Table \ref{tab:setup}. Several striking
features are present in the results. First of all, under all
strategies, the numbers of required levels are low (2 or 3). This can
be attributed to the fact that the error in the finest solutions are
high compare to the variance in the true solution. Second, the sample
sizes at the lowest level (highest resolution) are identical for
Strategy \#1, 2 and 4 (11 for all three). This is no coincidence. A
careful examination
of the formulae \eqref{eq:21}, \eqref{eq:26}, \eqref{eq:45} reveal
that, at the lowest level $l=1$, the sample sizes for Strategies \#1,
2, and 4 are identical, and depend on the convergence rate $\alpha$
only. Thus, irregardless of the actual highest resolution used, the
sample sizes for this model at the lowest level will remain the same
(=11) and identical for all three strategies. Finally, for Strategy
\#3, the sample size at the highest resolution is too high, and
results in a computational load even higher than that of the classical
Monte Carlo method. The reason is that this strategy requires a error
distribution that is low at the highest resolution, and high at the
lowest resolution. In terms of computational loads, Strategy \#1 is
the optimal choice, and it is followed by Strategy \#2. Both
strategies are better than the classical Monte Carlo method. Strategy
\#4 is actually more costly than the classical method, due to the
bloated sample size at the next level.
\begin{table}[h]\footnotesize
\centering
\ra{1.3}
\begin{tabular}{@{}c|cccccc@{}}\toprule
{ } & Est.~volume & Est.~error & CPU time & CPU efficiency \\
{ } & transport ($\mathrm{m}^3/\mathrm{s}$) &
($\mathrm{m}^3/\mathrm{s}$) &(hours) & (DOFs
/ CPU second)\\
\midrule
Classical MC & $3.58\times 10^8$ & $1.92\times 10^7$ & 158,446 &
343,172 \\\midrule
Strategy \#1 & $3.55\times 10^8$ & $4.80\times 10^7$ & 45,917 &
449,719\\\midrule
Strategy \#2 & $3.57\times 10^8$ & $3.84\times 10^7$ & 68,457
&488,013\\
\bottomrule
\end{tabular}
\caption{Comparison between the classical MC, and the MMC under
strategies \#1 and \#2. The efficiency is calculated as Total DOFs
/ Total CPU time. Total DOFs is calculated as Spatial DOFs
$\times$ Time steps $\times$ Number of samples.}
\label{tab:results}
\end{table}
We proceed to calculate the estimates and errors in the estimates using
the classical MC method and the MLMC method under strategies \#1 and 2. The results from the classical MC
can serve as a reference, since by the analysis of
\ref{sec:monte-carlo-method}, its error should be the
smallest. Strategies \#3 and 4 are not used
due to the shear sizes of their computational loads. The classical MC
and both Strategies \#1 and 2 produce similar estimates, $\sim
3.58\times 10^8\textrm{ m}^3/\mathrm{s}$, for the volume
transport (second column of Table \ref{tab:results}). The error for
each method is listed in the third column. The result of the classical
MC method has an error of about 5.4\%. The
error for Strategy \#1 is the largest, about 13.5\%. This is expected,
because the error for this strategy depends on the number of levels
(see \eqref{eq:24}), which is higher than that for Strategy \#2.
The primary advantage of MLMC is efficiency. The first
indicator for efficiency is of course the computational load that each
method will incur, which has already be listed in Table
\ref{tab:setup}. These numbers are the theoretical computational load,
and takes no consideration of computational overhead, parallelization,
etc. Here, we examine the actual efficiency for each
method. First, we look at
the total CPU hours used by each method (fourth column of Table
\ref{tab:results}). The MLMC with strategy \#1 uses the least amount
of CPU time, 45,917 CPU hours, a saving of $71\%$ compared with the MC
method. Strategy \#2 use 68,457 CPU
hours, a saving of $67\%$.
The savings of MLMC strategies in the actual CPU times are largely in
line with the savings in computational loads (Table \ref{tab:setup}),
but appear more dramatic than what the latter would suggest. This is
due to the increased CPU efficiency under the MLMC methods.
It is well known that MC methods are easy to parallelize,
and therefore highly scalable on supercomputers. The MLMC has the
potential to increase the efficiency over the classical MC even
further, by running more small-sized simulations and fewer large
simulations. Due to the large sizes of the
computations in this project, we are not able to perform a actual
scalability analysis, which involves running the experiment with
different numbers of total available processes. However, we can
indirectly examine the issue of scalability by comparing the efficiency for each CPU
core for the methods considered here (last column of Table
\ref{tab:results}). Compared with the classical MC, Strategy \#1
increases the CPU efficiency by 31\%, and Strategy \#2 increases the
CPU efficiency even more, by 42\%.
\section{Discussions}\label{sec:discussions}
The success of the MLMC method relies on a crucial assumption, namely
that, a lot of the information contained in high-resolution
simulations is also available from low-resolution simulations, under
the identical or similar model configurations. The higher the
correlation, the better the MLMC method will work. For steady-state or
laminar flows, especially when the simulations are backed up by
rigorous error estimates, the correlation between high-resolution and
low-resolution solutions is high and quantifiable, and the MLMC method
works very well (see references cited in Introduction). For long-term
simulations of turbulent flows, the situation is different. It is
known that the pointwise behaviors of high- and low-resolution
solutions of turbulent flows are uncorrelated (Figure
\ref{fig:vorticity-y40}). But pointwise behaviors of turbulent flows
are of little interest. What are important are certain aggregated
quantities such as mean SST. Then, naturally arise
the questions as to whether these aggregated quantities are correlated
across different resolutions, and whether the MLMC method can be used
to save computation times. Through an experiment with the Antarctic
Circumpolar Current, the present work gives affirmative answers to both
of these questions. The conclusions drawn in this work cannot be
generalized universally to all turbulent flows, because, after all,
there is no universal theory for turbulent flows yet. But it is
reasonable to expect that the same results should hold in similar
situations. Specifically, the MLMC method can be effective in saving
computation times when the QoI demonstrates a certain
level of correlation across different resolutions.
Another objective of this paper is to explore how the MLMC simulation
can be set up. Four different strategies are presented, based on the
desired error distributions. One surprising finding is that the
performance of each strategy, with regard to computational cost,
depends on the convergence rate. For all strategies discussed, the
higher the convergence rate is, the faster the computational cost will
grow. This sounds counter-intuitive. Here, the focus is on how fast
the total computational cost will grow in terms of the computational
cost of a single high-resolution simulation (linearly, quadratically,
etc.) Of course, for the same highest resolution, a higher convergence
rate will eventually leads to more accurate results, and the
associated higher computational cost is a price paid for this higher
accuracy.
Among all the four strategies discussed in this work, Strategy \#3,
which amplify the sample sizes at high resolutions, seems to be of
little use, because of the unreasonably high cost. Strategy \#1 is the
most natural choice, but it may lead to bigger margin of errors if the
number of levels is high. In that situation, Strategy \#2 \& \#4 can
be used. In our experiment with the ACC, the highest resolution has
40,000 grid points, and Strategy \#2 outperforms Strategy \#4 with a
lower computational cost. But it should be kept in mind that, even at a
very modest convergence rate ($\alpha < 1/2$), the
computational cost for Strategy \#2 grow polynomially with respect to
the computational cost for a single high-resolution
simulation. Therefore, it is conceivable that, as higher resolution
are taken into use, Strategy \#4 will eventually outperform Strategy
\#2.
\section*{Acknowledgment}
This work was in part supported by Simons
Foundation (\#319070 to Qingshan Chen) and NSF of China (\#91330104 to Ju Ming).
\bibliographystyle{amsplain}
|
\section{Introduction}
Suppose $(\Sigma, g)$ is a compact Riemann surface {without boundary}, $G$ is a compact Lie group {equipped} with a metric, $\mathfrak{g}$ is the Lie algebra
of $G$ and $\mathfrak{g}^\ast$ is the dual of $\mathfrak{g}$, and $P$ is a principal $G$-bundle on $\Sigma$. Let $(M, \omega)$ be a compact symplectic manifold which supports a Hamiltonian action of $G$ with moment map
$\mu : M \to \mathfrak{g}^*$, and $\pi : \mathcal{F} = P\times_{G} M \to \Sigma$ be the associated {fiber} bundle with {fiber} $M$.
Then $G$ extends to an equivariant action on $\mathcal F$, and $\mu$ extends to a map on the bundle $\mu:\mathcal{F}\to P\times_{ad}\mathfrak{g}^\ast$.
Denote the space of smooth connections on $P$ by $\mathscr{A}$, the space of smooth sections on $\mathcal{F}$ by $\S$. Denote the $W^{k,p}$-Sobolev completions of the spaces $\mathscr{A}$ and $\S$ by $\mathscr{A}_{k,p}$ and $\S_{k,p}$ respectively.
Then for a pair $(A,\phi) \in \mathscr{A}_{1,2}\times \S_{1,2}$, the \emph{Yang-Mills-Higgs functional} is defined by
\begin{equation}\label{YMH1}
\mathcal{E}(A,\phi) := \norm{F_A}^2_{L^2(\Sigma)}+ \norm{D_A \phi}^2_{L^2(\Sigma)} + \norm{\mu(\phi) - c}^2_{L^2(\Sigma)},
\end{equation}
where $F_A$ is the curvature of $A$, {$D_A$ is the exterior derivative associated with connection $A$},
and $c \in \mathfrak{g}^\ast$ is a fixed central element.
The \emph{Yang-Mills-Higgs (or YMH) functional} is a composition of the famous \emph{Yang-Mills functional}, the kinetic energy functional and the Higgs potential {energy functional}. Since the YMH functional appears naturally in
{the} classical gauge theory, it {has generated a lot of} interests among {both} physicists and mathematicians during the past decades. For example, the Ginzburg-Landau equation in the superconductivity theory coincides with the variational equation of YMH functional. The critical points of the YMH functional are known as \emph{Yang-Mills-Higgs fields} and the YMH functional is an appropriate Morse function to study the underlying spaces (\cite{AB, JT, P, So, T}). On the other hand, it is also well-known that the minimal YMH fields are the so-called \emph{symplectic vortices} and their moduli space can be used to define invariants on symplectic manifolds with Hamiltonian actions \cite{CGS,M}.
A natural method to study the existence of critical points is the heat flow method. If we denote the formal adjoint operator{s} of the exterior derivative $D_A$ and {the} covariant derivative $\nabla_A$ by $D_A^*$ and $\nabla_A^*$ respectively, then the
equation of heat flow of Yang-Mills-Higgs functional can be written as
\begin{equation}\label{e:heat}
\left\{
\begin{aligned}
\frac{\partial A}{\partial t} &= -D_A^*F_A - \phi^* D_A\phi, \\
\frac{\partial\phi}{\partial t} &= -D_A^*D_A\phi - (\mu(\phi) - c)\cdot \nabla\mu(\phi).
\end{aligned}
\right.
\end{equation}
Here $\nabla$ denotes the connection induced by the metric on $(M,\omega)$ and $\phi^* D_A\phi$
stands for {the} element in the dual space of $\Omega^1(AdP)$ consisting of $AdP$-valued 1-forms,
with $AdP=P\times_{Ad}\mathfrak{g}$, {which is defined as follows:}
\[ \int_\Sigma \langle\phi^* D_A\phi, B\rangle dv_g := \int_\Sigma \langle D_A\phi, B\phi \rangle dv_g,
\ \forall\ B\in \Omega^1(AdP). \]
Here we would like to mention a few relevant references from the vast literatures concerning both the YMH flow and the closely related Yang-Mills flow. The heat flow of Yang-Mills functional was first suggested by Atiyah and Bott~\cite{AB} and has been studied by many people. For example, R{\aa}de~\cite{R} studied the Yang-Mills heat flow in dimensions 2 and 3, and Struwe~\cite{St} has studied the Yang-Mills flow in dimension 4 and showed both the existence and uniqueness of local smooth solutions, while Schlatter~\cite{Sc1, Sc2} has described the blow-up phenomenon and long time behavior of the Yang-Mills flow. Recently, Hong, Tian and Yin \cite{HTY} have applied the Yang-Mills $\alpha$-flow to construct a weak solution of the Yang-Mills flow in dimension 4. For the YMH flow of a vector bundle over 4 dimensional manifold, results similar to \cite{St, Sc1} were obtained by Fang and Hong \cite{FH}. Hong and Tian \cite{HT} have also studied the asymptotic behavior of both the Yang-Mills flow and the YMH flow in higher dimensions. It is worth mentioning that the Yang-Mills flow and the YMH flow have also been employed in the study of the existence of Hermitian-Einstein metrics and the so-called Hitchin-Kobayashi correspondence(cf. \cite{D, UY, Si, H, LZ}). For more related results, we refer to the book by Feehan~\cite{F} and references therein.
The equation of heat flow of YMH functional~(\ref{e:heat}) has its independent interest from the analytic point of view. It is a coupled quasilinear degenerate parabolic system with critical nonlinearities: the first equation in~(\ref{e:heat}) is analogous {to} the Yang-Mills flow, while the second equation {in~(\ref{e:heat})} can be viewed
as the gauged heat flow of harmonic maps. Thus it is {natural to expect} that
the equation of YMH flow \eqref{e:heat} should reflect some features that are common with respect to
both the {Yang-Mills flow} and the heat flow of harmonic maps. It is worthwhile to mention that
(i) since the dimension of the base manifold ($\Sigma, g$) is two, the nonlinearity of first equation of YMH flow \eqref{e:heat} becomes subcritical, while the nonlinearity of second equation of YMH flow \eqref{e:heat} is critical;
(ii) since the fiber is a compact manifold that may support nontrivial harmonic maps from $\mathbb S^2$, the second
equation of YMH flow \eqref{e:heat} may develop finite time singularity.
Recall that the critical dimension of {Yang-Mills
flow} is {four} in which the {Yang-Mills} functional is conformally invariant.
R{\aa}de~\cite{R} {has} showed that the {Yang-Mills flow} admits a global smooth solution in subcritical dimensions 2 and 3, while the existence of global smooth solutions to the Yang-Mills flow in dimension 4 is an outstanding open problem. Yu~\cite{Y} showed the local existence of the YMH flow in dimension 2 with smooth initial data and studied the bubbling analysis at the first singular time.
In this paper, we will show that a weaker version of R{\aa}de's result holds for the YMH flow
\eqref{e:heat}, namely there exists a global weak solution of (\ref{e:heat}) that is smooth away from finitely many points.
Note that finite time singularities do occur for harmonic map heat flows in dimension two~\cite{CDY,CL} (see also \cite{YL} for other related work), which could be regarded as a special case of the YMH flow.
Before stating our results, we first give the definition of weak solutions.
\begin{defn} For $0<T\le +\infty$ and $(A_0,\phi_0)\in \mathscr{A}_{0,2}\times \S_{1,2}$,
a pair of sections $(A,\phi)$
is called a weak solution to the YMH flow equation~(\ref{e:heat}) under the initial condition
\begin{equation} \label{IVP}
(A,\phi)\big|_{t=0}=(A_0,\phi_0),
\end{equation}
on the interval $[0, T)$, if \begin{itemize}
\item[(i)] $(A, \phi)\in C^0\big([0, T), \mathscr{A}_{0,2}\big)\times \Big(C^0([0,T], \S_{0,2})\cap L^2([0,T), \S_{1,2})\Big)$,
and $F_A\in L^2([0,T), L^2)$.
\item [(ii)] the following holds:
\begin{equation}\label{e:weak-sol}
\left\{
\begin{aligned}
&\int_0^T\int_\Sigma\< A, \partial_t B\> \,dv_gdt = \int_0^T\int_\Sigma\big(\< F_A, D_AB\> + \< D_A\phi, B\phi\>\big)\,dv_gdt;\\
&\int_0^T\int_\Sigma\< \phi, \partial_t \psi\> \,dv_gdt = \int_0^T\int_\Sigma\big(\< D_A\phi, D_A\psi\> + \< (\mu(\phi)-c)\nabla\mu(\phi), \psi\>\big) \,dv_gdt,\\
\end{aligned}
\right.
\end{equation}
for any test functions $B\in H_0^1\big([0, T], \mathscr{A}_{0,2}\big)\cap L^2\big([0,T), \mathscr{A}_{1,2}\big)$,
and $\psi\in H^1_0\big([0,T), \S_{0,2})\big)\cap L^2\big([0,T], \S_{1,2}\big)$.
\item [(iii)] $(A, \phi)$ satisfies
(\ref{IVP}) in $L^2$-sense.
\end{itemize}
\end{defn}
Now we state our main theorem on the existence of global weak solutions to the YMH flow.
\begin{thm}\label{main}
Let $(A_0, \phi_0)\in \mathscr{A}_{1,2}\times \S_{1,2}$. There exist a global weak solution $(A, \phi)$ to the YMH flow~(\ref{e:heat})
and (\ref{IVP}) such that
\begin{itemize}
\item[i)] the energy inequality $\mathcal{E}(A(t),\phi(t))\le\mathcal{E}(A_0,\phi_0)$ holds for all $0\le t<+\infty$, and
\[ A\in C^0([0, \infty), \mathscr{A}_{0,2});\ \phi\in C^0([0, \infty), \S_{0,2}); \ F_A\in L^\infty([0, \infty), L^2);
\ D_A\phi\in L^\infty([0,\infty), L^2).\]
\item[ii)] there exist a positive integer $L\le [\frac{\mathcal{E}(0)}{\alpha(M)}]$, and gauge transformations
$\{s_i\}_{i=1}^L\subset\mathscr{G}_{2,2}$, and $0=T_0<T_1<T_2<\cdots<T_L<+\infty$ such that for $1\le i\le L$,
$(s_i^*A,s_i^*\phi)\in C^\infty\big(\Sigma\times (T_{i-1}, T_i]\setminus \{(x_1^i,T_i),\cdots, (x_{j(i)}^i,T_i)\}\big)$ for some $j(i)\le
\frac{\mathcal{E}(0)}{\alpha(M)}$. Here $\alpha(M)>0$ is defined by
$$\alpha(M)=\inf\big\{\int_{\mathbb S^2}|\nabla h|^2\,dv_{g_0}: \ h\in C^\infty(\mathbb S^2, M)
\ \mbox{is a nontrivial harmonic map}\big\}.
$$
\item [iii)] for each $1\le i\le L$, there exist finitely many nontrivial harmonic maps $\omega_{ij}:\mathbb S^2\to M$,
$1\le j\le L_i$ with $L_i\le \frac{\mathcal{E}(0)}{\alpha(M)}$, such that
\begin{equation}\label{EIfinitetime}
\lim_{t\uparrow T_i}\mathcal{E}(A(t), \phi(t))=\mathcal{E}(A(T_i), \phi_i(T_i))+\sum_{j=1}^{L_i}\int_{\mathbb S^2}|\nabla\omega_{ij}|^2\,dv_{g_0}.
\end{equation}
\item[iv)] a) there exist $t_i\uparrow\infty$, a sequence of gauge transformations $\{s_i\}\subset\mathscr{G}_{2,2}$,
a set of finitely many points $S:= \{x_1, \cdots, x_{k_0}\}\subset \Sigma$, with $k_0\le \frac{\mathcal{E}(0)}{\alpha(M)}$,
and a Yang-Mills-Higgs field $(A_\infty,\phi_\infty)\in\mathscr{A}\times\S$ such that
$s_i^*A(t_i)\rightarrow A_\infty$ in $H^1(\Sigma)$, $s_i^*\phi(t_i)\rightharpoonup \phi_\infty$ in $H^1(\Sigma)$,
and $s_i^*\phi(t_i)\rightarrow \phi_\infty$
in $H^1_{loc}(\Sigma\setminus S)$, as $i\rightarrow \infty$; b) there exist finitely many nontrivial harmonic maps $\omega_{p}: \mathbb S^2 \to M, 1\le p\le p_0\le \frac{\mathcal{E}(0)}{\alpha(M)}$ such that
\begin{equation}\label{EIinfinitytime}
\lim_{i\to \infty} \mathcal{E}(A(t_i), \phi(t_i)) = \mathcal{E}(A_\infty, \phi_\infty) + \sum_{p=1}^{p_0}\mathcal{E}(\omega_{p});
\end{equation}
and c) the images of $\{\omega_{p}\}_{p=1}^{p_0}$ and that of $\phi_\infty$ are connected.
\end{itemize}
\end{thm}
It is an interesting question whether the global weak solution given by Theorem \ref{main} is unique in a certain class
of weak solutions, which we plan to investigate in the future.
The paper is organized as follows. In section 2, we review some preliminary facts on connections and curvatures on
G-bundles. In section 3, we utilize the DeTurck's trick to obtain the local existence of unique smooth solutions to
the YMH flow \eqref{e:heat}. In section 4, we establish a priori estimates of smooth solutions to \eqref{e:heat}
under the smallness condition. In section 5, through smooth approximations of $H^1$-initial data and applications of
the results from section 3 and section 4, we show the existence of local weak solutions to \eqref{e:heat} under $H^1$-initial data.
In section 6 and section 7, by extending the bubbling analysis for YMH fields to approximate YMH fields with $L^2$-controlled tension fields, we obtain both the energy identity at finite singular time and asymptotic behavior at time infinity for the global
weak solutions to \eqref{e:heat}. In section 8, by utilizing Uhlenbeck's gauge fixing techniques, we are able to extend
the existence of local weak solutions to global weak solutions to \eqref{e:heat}. The details of proof of Theorem \ref{main} are included in that of Theorem 5.1, Theorem 7.1, Theorem 7.2, and Theorem 8.3.
\section{Preliminaries}
First we recall some basic definitions of connection{\bf s} on bundle{\bf s}. Let $Ad P := P{\times_{Ad}}\ \mathfrak{g}$ be the adjoint bundle of the principal $G$-bundle $P$. The space $\mathscr{A}$ of connections on $P$ is an affine space modelled on $\Omega^1(Ad P)$,
here $\Omega^k(Ad P)$ denotes the space of $Ad P$-valued $k$-forms for $k\ge 1$. Namely, if we fix a reference connection $D_{ref}$
on $P$, then
\[ \mathscr{A} = \Big\{D_A = D_{ref} + a\ |\ a\in \Omega^1(Ad P)\Big\}. \]
Any connection $A\in \mathscr{A}$ gives rise to a covariant derivative $\nabla_A$ and an exterior differential operator $D_A$, which is the anti-symmetric part of $\nabla_A$, on the principal $G$-bundle $P$. The curvature of a connection
$A\in\mathscr{A}$ is defined by
\[ F_A := D_A\circ D_A \in \Omega^2(Ad P). \]
The covariant derivative $\nabla_A$ can be extended to the associated bundle $\mathcal{F}$ as follows. Given a local trivialization of $\mathcal{F}$, let $\psi: \mathcal{F}|_U \to U\times M$ be the coordinate map. Then a section $\phi\in \S$ is locally equivalent to a map $u\in C^\infty(U, \mathcal{F})$ and the connection $A$ can be written as
$\displaystyle D_A \big|_U = d + A$,
where $A := A_\alpha dx^\alpha$ is a $\mathfrak{g}$-valued $1$-form. The covariant derivative $\nabla_A\phi$ is then defined by
$$\nabla_A\phi\big|_U := du + A\cdot u,$$
where the action of $A$ on the map $u$, $A\cdot u$, is induced by the symplectic action of $G$ on $M$.
More precisely, by the action of $G$ on $M$, every $\xi \in \mathfrak{g}$ corresponds to a symplectic vector field $X_\xi$ by
\[ X_\xi(y) := \frac{d}{dt}\Big|_{t=0} exp(\xi t)\cdot y, ~\forall y \in M. \]
Then the action of the $\mathfrak{g}$-valued $1$-form $A$ on $u$ is given by
\[ A\cdot u = A_\alpha\cdot u dx^\alpha = X_{A_\alpha}(u)dx^\alpha. \]
The operation of covariant derivative $\nabla_A$ on the (vertical) tangent bundle $T\mathcal{F}^v$ is slightly different. Denote the Levi-Civita connection induced by the metric $h$ on $M$ by $\nabla$, then we can define $\nabla_A:\Gamma(T\mathcal{F}^v) \to \Gamma(T\mathcal{F}^v\times T^*\Sigma)$ by
\[ \nabla_A V = \nabla V + A\cdot V:= \nabla V + \nabla_V X_{A_\alpha}\cdot dx^{\alpha}, \ V\in\Gamma(T\mathcal{F}^v). \]
The covariant derivative $\nabla_A$ also extends to $T\mathcal{F}^v$-valued $p$-forms. Namely, we can define $\nabla_A: \Omega^p(T\mathcal{F}^v) \to \Gamma(T\mathcal{F}^v\otimes \Omega^p(\Sigma)\otimes T^*\Sigma)$ by
\[ \nabla_A(\eta \otimes \omega) = \nabla_A \eta \otimes \omega + \eta \otimes \nabla \omega,
\ \eta \otimes \omega \in \Omega^p(T\mathcal{F}^v), \]
where $\nabla$ is the Levi-Civita connection on $\Sigma$. The exterior derivative $D_A$ is defined
through the projection $\Omega^p(T\mathcal{F}^v)\otimes T^*\Sigma \to \Omega^p(T\mathcal{F}^v)$ in the standard way.
We will also need the following basic facts. For any connection $A$, we have the first Bianchi identity:
\begin{equation}\label{e:bianchi1}
D_A F_A = 0,
\end{equation}\label{e:bianchi2}
and the second Bianchi identity:
\begin{equation}
D_A^*D_A^* F_A = 0.
\end{equation}
There are two Laplace operators for the connection $A$ on fiber bundles. Namely, the Hodge Laplacian
$$\Delta_A = D_A^*D_A + D_AD_A^*,$$
and the rough Laplacian $\nabla_A^*\nabla_A$. The well-known Weitzenb\"ock formula describes the difference of these two Laplace operators on vector bundle valued forms. For example, the Weitzenb\"ock formula for $\Phi\in \Omega^p(Ad P)$ is
\begin{equation}\label{e:weitzenbock}
\nabla_A^*\nabla_A \Phi = \Delta_A \Phi + F_A\#\Phi + R_\Sigma\#\Phi,
\end{equation}
where $R_\Sigma$ is the Riemannian curvature of the base manifold $\Sigma$ and $\#$ denotes a multi-linear map with smooth coefficients.
Note that for a fiber bundle where the fiber $M$ is a Riemannian manifold with curvature tensor $R_M$, an extra term emerges in the Weitzenb\"ock formula. More precisely, for a section $\phi\in \S$, there is a pull-back bundle $\phi^*T\mathcal{F}$ on $\Sigma$. The curvature on $\phi^*T\mathcal{F}$ contains not only $F_A$ but also the pull-ball curvature $\phi^*R_M$. Therefore, for a section $\Psi \in \Omega^p(\phi^*T\mathcal{F})$, we have
\begin{equation}\label{e:weitzenbock1}
\nabla_A^*\nabla_A \Psi = \Delta_A \Psi + F_A\#\Psi + R_\Sigma\#\Psi + R_M\#d\phi\#d\phi\#\Psi.
\end{equation}
This is the case when we apply this formula on $\Psi = D_A\phi$.
\medskip
\noindent\emph{Notations}: For simplicity, we will omit the subscription $A$ and simply use $D, F, \nabla, \Delta$ instead of $D_A, F_A, \nabla_A, \Delta_A$ if no confusions may occur.
\section{Local smooth solutions}
In this section, we will show both the existence and uniqueness of local smooth solutions to the YMH flow (\ref{e:heat}) for any smooth initial data $(A_0, \phi_0)$.
First, fix a smooth reference connection. For example, we may choose the initial connection $A_0$ as the reference connection. Then any connection $D$ corresponds to a 1-form $a\in \Omega^1(Ad P)$ by
$$D = A_0 + a. $$
The curvature of $A$ is
\[ F_A = D_A\circ D_A = F_{A_0} + D_0a + a\wedge a = F_{A_0} + D_A a - a\wedge a, \]
since $D_A a = D_{A_0}a + [a, a]$. Then the equation (\ref{e:heat}) can be written as
\begin{equation}\label{e:heat01}
\left\{
\begin{aligned}
\frac{\partial a}{\partial t} &= -D_A^*D_Aa - D_A^*F_{A_0} + D_A^*(a\wedge a) - \phi^*D_A\phi, \\
\frac{\partial\phi}{\partial t} &= -D_A^*D_A\phi - (\mu(\phi) - c)\cdot \nabla\mu(\phi).
\end{aligned}
\right.
\end{equation}
It is well-known that that the first equation of $A$ in the system (\ref{e:heat01}) is degenerate,
since the Yang-Mills functional is invariant under gauge transformations.
To overcome this difficulty, we adapt DeTurck's trick and consider a
gauged system equivalent to \eqref{e:heat01} that is parabolic.
Our main result in this section is the following.
\begin{thm}\label{t:smooth}
For any smooth initial data $(A_0, \phi_0)\in \mathscr{A}\times \S$, there exist a $T>0$ and a unique smooth solution
$(A,\phi)$ to the Yang-Mills-Higgs flow equation~(\ref{e:heat}) in $\Sigma\times [0, T)$,
with $\displaystyle (A,\phi)\big|_{t=0}=(A_0,\phi_0)$.
\end{thm}
\begin{proof}
We consider the following perturbed system for $\overline {A}$ and $\bar{\phi}$:
\begin{equation}\label{e:heat-DeTurck}
\left\{
\begin{aligned}
\frac{\partial\bar{a}}{\partial t} &= -\overline{D}^*\overline{F} - \bar{\phi}^*\overline{D}\bar{\phi} -\overline{D}\overline{D}^*\bar{a}, \\
\frac{\partial\bar{\phi}}{\partial t} &= -\overline{\nabla}^*\overline{D}\bar{\phi} - (\mu(\bar{\phi}) - c)\cdot \nabla\mu(\bar{\phi}) + (\overline{D}^*\bar{a})\bar{\phi},
\end{aligned}
\right.
\end{equation}
under the initial condition $(\bar{a}(0), \bar{\phi}(0)) = (0, \phi_0)$, where
$$\overline{D} = D_{\overline{A}}, \ \bar{a} = \overline{D} - A_0,\ \overline{F} = F_{\overline{A}},
\ \overline{\nabla} = \nabla_{\overline{A}}.$$
If we denote the Hodge Laplacian on $\Omega^p(Ad P)$ by $\overline{\Delta} := \overline{D}^*\overline{D} + \overline{D}\overline{D}^*$ and the Laplace-Beltrami operator on $\Gamma(\mathcal{F})$ by $\overline{\Delta}' :={\rm{tr}} \big(\overline{\nabla}^2\big)$, then the above system can be written as follows:
\begin{equation}\label{e:heat-DeTurck1}
\left\{
\begin{aligned}
\frac{\partial{\bar{a}}}{\partial t} + \overline{\Delta}\bar{a} &= -\overline{D}^*F_0 + \overline{D}^*(\bar{a}\wedge \bar{a}) - \bar{\phi}^*\overline{D}\bar{\phi}, \\
\frac{\partial{\bar{\phi}}}{\partial t} - \overline{\Delta}'\bar{\phi} &= - (\mu(\bar{\phi}) - c)\cdot \nabla\mu(\bar{\phi}) + (\overline{D}^*\bar{a})\bar{\phi},
\end{aligned}
\right.
\end{equation}
under the initial condition $(\bar{a}(0), \bar{\phi}(0)) =(0, \phi_0)$.
Note that the Hodge Laplacian $\overline\Delta$ and the Laplace-Beltrami operator $\overline\Delta'$ differs by a sign when applied to $\Omega^0(\mathcal{F})$. Then one can verify that the system (\ref{e:heat-DeTurck1}) is a quasilinear parabolic system. The standard parabolic theory implies that
for any smooth initial data $(A_0, \phi_0)$, there exists a unique smooth solution $(\bar{a}, \bar{\phi})\in C^\infty([0,T), \mathscr{A}\times \S)$ to (\ref{e:heat-DeTurck1})
for some $0<T=T(A_0,\phi_0)\le +\infty$ (see for example~\cite{FH} for a proof).
Next we choose a family of gauge transformations $\{S(t)\}_{0\le t<T}$, which satisfies the following ordinary differential equation:
\begin{equation}\label{e:gauge-DeTurck}
\left\{
\begin{aligned}
\dt{S} &= -S\circ \overline{D}^*\bar{a}, \\
S(0) &= {\rm {id}}.
\end{aligned}
\right.
\end{equation}
We claim that the pair $(A(t), \phi(t)) = S(t)^*(\overline{A}(t), \bar{\phi}(t))$, $0\le t<T$, is a solution to the YMH
flow equation
(\ref{e:heat}). Indeed, since
\[ D = S\circ \overline{D} \circ S^{-1}, ~ \dt{S^{-1}} = -S^{-1}\circ\dt{S}\circ S^{-1},\]
we have, by (\ref{e:gauge-DeTurck}),
\begin{equation*}
\begin{split}
\frac{\partial{a}}{\partial t} &= \dt{S}\circ \overline{D} \circ S^{-1} + S\circ \frac{\partial{\overline{D}}}{\partial t} \circ S^{-1}
+ S\circ \overline{D} \circ \dt{S^{-1}} \\
&= S\circ (-\overline{D}^*\bar{a}\circ \overline{D} -\overline{D}^*\overline{F} - \bar{\phi}^*\overline{D}\bar{\phi}
- \overline{D}\overline{D}^*\bar{a} + \overline{D}\circ \overline{D}^*\bar{a}) \circ S^{-1} \\
&= S\circ (-\overline{D}^*\overline{F} - \bar{\phi}^*\overline{D}\bar{\phi}) \circ S^{-1} \\
&= -D_A^*F_A - \phi^* D_A\phi,
\end{split}
\end{equation*}
where we have used the identity:
\[ \overline{D}\circ \overline{D}^*\bar{a} = \overline{D}\overline{D}^*\bar{a} + \overline{D}^*\bar{a}\circ \overline{D}. \]
Also, since $\phi = S\circ\bar{\phi}$, we have
\begin{equation*}
\begin{split}
\frac{\partial{\phi}}{\partial t} &= \dt{S}\circ\bar{\phi} + S\circ \frac{\partial{\bar{\phi}}}{\partial t} \\
&= -S\circ (\overline{D}^*\bar{a})\circ \bar{\phi} + S\circ\big(-\overline{\nabla}^*\overline{D}\bar{\phi} - (\mu(\bar{\phi}) - c)\cdot \nabla\mu(\bar{\phi})
+ (\overline{D}^*\bar{a})\bar{\phi}\big) \\
&= S\circ \big(-\overline{\nabla}^*\overline{D}\bar{\phi} - (\mu(\bar{\phi}) - c)\cdot \nabla\mu(\bar{\phi})\big) \\
&= -\nabla_A^*D_A\phi - (\mu(\phi) - c)\cdot \nabla\mu(\phi).
\end{split}
\end{equation*}
Moreover, the initial condition is preserved under the gauge transformation. Therefore we obtain a local smooth solution
$(A,\phi)\in C^\infty([0,T), \mathscr{A}\times \S)$ to the original YMH flow equation (\ref{e:heat}) by solving the gauged parabolic system (\ref{e:heat-DeTurck1})
and the gauge transformations (\ref{e:gauge-DeTurck}).
The uniqueness part can be shown from the observation that we can reverse the above process and obtain a local smooth solution to (\ref{e:heat-DeTurck}) from a local smooth solution to the YMH flow equation (\ref{e:heat}) through the gauge transformations. In fact, suppose $(A, \phi)\in
C^\infty([0, T), \mathscr{A}\times \S)$ is a solution to the YMH flow equation (\ref{e:heat}), then we first solve the following equation
to get a family of gauge transformations:
\begin{equation}\label{e:gauge-DeTurck1}
\left\{
\begin{aligned}
\dt{S} &= -D^*D S - D^*\circ Sa, \\
S(0) &= {\rm{id}}.
\end{aligned}
\right.
\end{equation}
Note that this is a parabolic system of $S$. Indeed, write $D = D_0 + a$, we have
\begin{equation*}
\begin{aligned}
D^*D S &= D_0^*D_0S + D_0^*(aS) + a^*(D_0S) + a^*aS\\
&= D_0^*D_0S + (D_0^*a)S + 2a^*(D_0S) + a^*aS.
\end{aligned}
\end{equation*}
Since $S$ can be viewed as a 0-form, we have $D_0^*S = 0$ and hence $\Delta_0 S = D_0^*D_0S$.
Thus the equation (\ref{e:gauge-DeTurck1}) is a parabolic system and has a unique local smooth
solution $S\in C^\infty([0,T))$.
A straightforward calculation shows that the pair $(\overline{A}(t), \bar{\phi}(t)) = (S^{-1})^*(A(t), \phi(t))$,
$0\le t<T$, is a solution to the system (\ref{e:heat-DeTurck}). In fact, the gauge transformation $S$ we get from (\ref{e:gauge-DeTurck1}) is exactly
the same gauge transformation we obtain from (\ref{e:gauge-DeTurck}). To see this, note that since $\overline{D}^* = (S^{-1})^*(D^*) = S^{-1}\circ D^*\circ S$ and
$$\bar{a} = \overline{D} - A_0 = S^{-1}\circ D \circ S - D + a, $$
there holds
\[ S\circ \overline{D}^*\bar{a} = D^*\circ S\bar{a} = D^*\circ(D\circ S - S\circ D + Sa) = D^*D S + D^*\circ Sa. \]
Thus the equation (\ref{e:gauge-DeTurck}) and (\ref{e:gauge-DeTurck1}) are identical.
Now we can complete the proof of uniqueness. Suppose that we have two local smooth solutions $(A_1, \phi_1), (A_2, \phi_2)\in C^\infty([0, T),\mathscr{A}\times\S)$
of the YMH flow equation (\ref{e:heat}), then one solves the equation (\ref{e:gauge-DeTurck1}) to get two smooth gauge transformations $S_1(t), S_2(t)$,
$0\le t<T$, and hence two local smooth solutions $(\bar{a}_1, \bar{\phi}_1), (\bar{a}_2, \bar{\phi}_2)$ to the system (\ref{e:heat-DeTurck}) with same initial value. By the uniqueness theorem of (\ref{e:heat-DeTurck}), we have $(\bar{a}_1, \bar{\phi}_1) = (\bar{a}_2, \bar{\phi}_2)$, $0\le t<T$. On the other hand,
since the parabolic system (\ref{e:gauge-DeTurck1}) is equivalent to the equation (\ref{e:gauge-DeTurck}), it follows from
the uniqueness of ordinary differential equations that $S_1 = S_2$. This in turn implies that $(A_1, \phi_1)=(A_2, \phi_2)$ for
$0\le t<T$. This completes the proof.
\end{proof}
\section{A priori estimates}
In this section, we will derive some a priori estimates of the local smooth solution obtained in the last section.
For $0<T\le +\infty$,
suppose $(A, \phi) \in C^\infty([0, T), \mathscr{A})\times C^\infty([0, T), \S)$ is a smooth solution to the YMH flow
equation (\ref{e:heat}), with the initial value $(A_0, \phi_0) \in \mathscr{A}\times \S$.
We first derive the evolution equation for the curvature $F_A$ of $A$ under the YMH flow
equation (\ref{e:heat}). By the definition of $F$ and a direct calculation, we have
\[ \frac{\partial{F_A}}{\partial t} = \frac{\partial}{\partial t}{}(D_A\circ D_A)
= \frac{\partial D_A}{\partial t}\circ D_A + D_A\circ \frac{\partial{D_A}}{\partial t}
= D_A(\frac{\partial{A}}{\partial t}). \]
Hence, under the YMH flow equation (\ref{e:heat}), $F_A$ satisfies:
\begin{equation}\label{e:curvature}
\frac{\partial{F_A}}{\partial t} = - D_A D_A^* F_A - D_A(\phi^* D_A \phi).
\end{equation}
Note that it is not hard to see
\[ D_A(\phi^* D_A \phi) = D_A\phi\#D_A\phi + \phi^* F_A\phi. \]
The equation (\ref{e:curvature}) is a parabolic equation, since by the first Bianchi identity (2.1), we have
\[ \Delta F_A = D_A D_A^* F_A. \]
Applying the Weitzenb\"ock formula (\ref{e:weitzenbock}), we can rewrite (\ref{e:curvature}) as
\begin{equation}\label{e:curvature1}
\frac{\partial{F_A}}{\partial t} = -\nabla_A^*\nabla_A F_A - R_\Sigma\#F_A - F_A\#F_A - D_A\phi\#D_A\phi - \phi^*F_A\phi.
\end{equation}
Now we set the energy density of $(A(t),\phi(t))$ by
\[ e(A(t), \phi(t)) := |F_A|^2+|D_A\phi(t)|^2 + |\mu(\phi(t))-c|^2, \]
the (total) Yang-Mills-Higgs energy of $(A(t),\phi(t))$ by
$$\mathcal{E}(t) := \mathcal{E}(A(t), \phi(t))=\int_\Sigma e(A(t), \phi(t))\,dv_g,$$
and the local Yang-Mills-Higgs energy of $(A(t),\phi(t))$ by
$$\mathcal{E}(t, B_R(x)) := \int_{B_R(x)}e(A(t),\phi(t))\,dv_g,$$
where $B_R(x):=\big\{y\in\Sigma: \ d(y,x)<R\big\}\subset\Sigma$ denotes the geodesic ball with center $x$ and radius $R$,
and $d(\cdot, \cdot)$ denotes the distance function on ($\Sigma, g$).
Since the YMH flow equation~(\ref{e:heat}) is the negative gradient flow of the YMH functional,
we have the following global
energy inequality.
\begin{lem}\label{l:energy-inequ}
{\rm{(energy inequality)}} For $0\le t<T$, it holds
\begin{equation}\label{e:energy-inequ}
\mathcal{E}(t) + 2\int_0^t\int_\Sigma\Big(\abs{\frac{\partial{A}}{\partial t}}^2 + \abs{\frac{\partial{\phi}}{\partial t}}^2\Big)\, dv_gdt
\le \mathcal{E}(0).
\end{equation}
\end{lem}
\begin{proof}
From the equation (\ref{e:heat}), we have
\begin{equation}\nonumber
\begin{aligned}
\dt{\mathcal{E}} &= 2\int_\Sigma\Big( \< D_A^*F_A + \phi^* D_A\phi, \frac{\partial{A}}{\partial t}\> + \<\nabla_A^*\nabla_A\phi
+ (\mu(\phi) - c) \cdot\nabla\mu(\phi), \frac{\partial{\phi}}{\partial t}\>\Big) \,dv_g \\
&= - 2\int_\Sigma\Big(\abs{\frac{\partial{A}}{\partial t}}^2 + \abs{\frac{\partial{\phi}}{\partial t}}^2\Big) \,dv_g.
\end{aligned}
\end{equation}
Integrating this inequality on $[0, t]$ yields (\ref{e:energy-inequ}).
\end{proof}
We also have the following local energy inequality.
\begin{lem}\label{l:energy-inequ-local}
{\rm{(local energy inequality)}} There exist $R_0>0$ and $C_0>0$ depending only $\Sigma$ such that for $0<t<T$ and $x\in\Sigma$,
it holds
\begin{equation}\label{e:energy-inequ-local}
\mathcal{E}(t, B_R(x)) \le \mathcal{E}(0, B_{2R}(x)) + \frac{Ct}{R^{2}}\mathcal{E}(0).
\end{equation}
\end{lem}
\begin{proof} Let $\eta \in C^\infty_0(B_{2R})$ such that $0\le \eta\le 1$, $\eta = 1$ on $B_R$, and $|d\eta| \le \frac{C}{R}$. Then we
can calculate
\begin{equation*}
\begin{aligned}
\frac12\dt{} \int_\Sigma e(A, \phi)\eta^2 \,dv_g
=& \int_\Sigma \eta^2\< D_A\phi, \frac{\partial A}{\partial t}\phi \> \,dv_g
+ \int_\Sigma \eta^2\< D_A\phi, D_A\frac{\partial\phi}{\partial t} \> \,dv_g \\
&+ \int_\Sigma\eta^2\< F_A, D_A\frac{\partial A}{\partial t}\> \,dv_g
+ \int_\Sigma\eta^2(\mu(\phi) - c)\cdot\nabla\mu(\phi)\cdot
\frac{\partial\phi}{\partial t} \,dv_g.
\end{aligned}
\end{equation*}
Rearranging terms in the right side and integrating by parts, we have
\begin{equation*}
\begin{aligned}
\frac12\dt{} \int_\Sigma e(A, \phi)\eta^2 \,dv_g
=& \int_\Sigma\eta^2\< D_A^*D_A\phi + (\mu(\phi)-c)\cdot\nabla\mu(\phi), \frac{\partial\phi}{\partial t} \> \,dv_g\\
&+ \int_\Sigma \eta^2\< D_A^*F_A + \phi^* D_A\phi, \frac{\partial A}{\partial t} \> \,dv_g \\
&+ \int_\Sigma d(\eta^2)\< D_A\phi, \frac{\partial\phi}{\partial t} \> \,dv_g
+ \int_\Sigma d(\eta^2)\< F_A, \frac{\partial A}{\partial t}\> \,dv_g.
\end{aligned}
\end{equation*}
From the equation (\ref{e:heat}) and Young's inequality, we obtain
\begin{equation*}
\begin{aligned}
\frac12\dt{} \int_\Sigma e(A, \phi)\eta^2 \,dv_g
\le& - \frac12\int_\Sigma\eta^2\big(|\frac{\partial\phi}{\partial t}|^2 + |\frac{\partial A}{\partial t}|^2\big) \,dv_g
+ \frac{C}{R^{2}}\int_\Sigma \big(|D_A\phi|^2 + |F_A|^2\big) \,dv_g.
\end{aligned}
\end{equation*}
Integrating on $[0,t]$, we arrive at
\begin{equation*}
\begin{aligned}
& \mathcal{E}(t, B_{R}(x)) + \int_0^t\int_{B_R(x)}\Big(|\frac{\partial\phi}{\partial t}|^2 + |\frac{\partial A}{\partial t}|^2\Big) \,dv_g \\
& \le \mathcal{E}(0, B_{2R}(x))
+ \frac{C}{R^2}\int_0^t\int_{\Sigma}(|D_A\phi|^2 + |F_A|^2) \,dv_g dt.
\end{aligned}
\end{equation*}
This and the energy inequality (\ref{e:energy-inequ}) now yield the local energy inequality (\ref{e:energy-inequ-local}).
\end{proof}
Next, we will derive a Bochner type formula for smooth solutions to the YMH flow equation (\ref{e:heat}).
In order to do this, we first set for any $\lambda>0$,
\[ \widehat{e}_\lambda(A, \phi) := \sqrt{\lambda+|F_A|^2}+|D_A\phi|^2 . \]
Recall that for any smooth section $\Phi$ of $\mathscr{A}\times\S$, there holds
\begin{equation}\label{e1}
\frac12 \Delta_g|\Phi|^2 = -\< \nabla_A^*\nabla_A \Phi, \Phi\> + |\nabla_A\Phi|^2.
\end{equation}
\begin{lem}\label{l:bochner}
{\rm{(Bochner formula)}} Let $(A,\phi)\in C^\infty([0, T), \mathscr{A}\times \S)$ be a smooth solution of the YMH
flow equation (\ref{e:heat}).
There exists a constant $C>0$ depending only on $\Sigma,\ \mathscr{A},$ and $\S$ such that
\begin{equation}\label{bochner1}
\big(\frac{\partial}{\partial t} - \Delta_g\big)\widehat{e}_1(A, \phi)
\le C\big(1 + |F_A| + |D_A\phi|^2\big)\widehat{e}_1(A, \phi)
\end{equation}
holds on $\Sigma\times (0, T)$.
\end{lem}
\begin{proof} First, applying the identity (\ref{e1}), we have
\begin{equation*}
\frac12 \Delta_g|F_A|^2 = - \< \nabla_A^*\nabla_A F_A, F_A\> + |\nabla_A F_A|^2.
\end{equation*}
Second, applying the equation (\ref{e:curvature1}), we have
\begin{equation*}
\begin{aligned}
\frac12 \frac{\partial}{\partial t} |F_A|^2
&= -\< \nabla^*\nabla F_A + F_A\# F_A + R_\Sigma\# F_A, F_A\> - F_A\#D_A\phi\#D_A\phi - \< F_A\phi, F_A\phi\> \\
&\le -\< \nabla^*\nabla F_A, F_A\> +C(1+|F_A| + |R_\Sigma|)|F_A|^2 + C|F_A||D_A\phi|^2.
\end{aligned}
\end{equation*}
Combining these two inequalities yields
\begin{equation}\label{e3}
\frac12\big(\frac{\partial}{\partial t} - \Delta_g\big)|F_A|^2 + |\nabla_A F_A|^2 \le C(1+ |F_A|)|F_A|^2 + C|F_A||D_A\phi|^2.
\end{equation}
Direct calculations imply that
\begin{eqnarray}\label{bochner-lhs}
\frac12\big(\frac{\partial}{\partial t} - \Delta_g\big)|F_A|^2
&=& \frac12\big(\frac{\partial}{\partial t} - \Delta_g\big)\big(\sqrt{1+|F_A|^2}\big)^2\nonumber\\
&=&\sqrt{1+|F_A|^2}\Big(\frac{\partial}{\partial t} - \Delta_g\Big)\sqrt{1+|F_A|^2}
-\Big|\nabla\sqrt{1+|F_A|^2}\Big|^2.
\end{eqnarray}
Recall that by Kato's inequality, it holds that
$$\big|\nabla |F_A|\big|\le \big|\nabla_A F_A\big|,$$
and hence we have that
\begin{eqnarray}\label{kato1}
\big|\nabla_A F_A\big|^2-\Big|\nabla\sqrt{1+|F_A|^2}\Big|^2
&\ge&\big|\nabla_AF_A\big|^2-\frac{|F_A|^2\big|\nabla|F_A|\big|^2}{1+|F_A|^2}\nonumber\\
&\ge& \big|\nabla_AF_A\big|^2-\big|\nabla|F_A|\big|^2\ge 0.
\end{eqnarray}
It is clear that by substituting (\ref{bochner-lhs}) into (\ref{e3}) and applying (\ref{kato1}), we obtain
\begin{equation}\label{e4}
\big(\frac{\partial}{\partial t} - \Delta_g\big)\sqrt{1+|F_A|^2} \le C(1+ |F_A|)|F_A| + C|D_A\phi|^2.
\end{equation}
Next we will make a similar calculation on $|D_A\phi|^2$. By (\ref{e1}), we have
\begin{equation*}
\frac12 \Delta_g|D_A\phi|^2 = - \< \nabla_A^*\nabla_A D_A\phi, D_A\phi\> + |\nabla_A D_A\phi|^2.
\end{equation*}
Applying the YMH flow equation (\ref{e:heat}), we have
\begin{equation*}
\begin{aligned}
&\frac12 \frac{\partial}{\partial t} |D_A\phi|^2
= \big\< D_A(\frac{\partial\phi}{\partial t}), D_A\phi\big\> + \big\< (\frac{\partial}{\partial t}{D_A})\phi, D_A\phi\big\>\\
&= -\big\< D_A\big(D_A^*D_A\phi + \nabla h(\phi)\big), D_A\phi\big\>
- \big\< \big(D_A^*F_A+\phi^*D_A\phi\big)\phi, D_A\phi \big\>\\
&= -\big\< D_AD_A^*D_A\phi, D_A\phi\big\> - \nabla^2 h(\phi)\big(D_A\phi, D_A\phi\big)
- \big\<(D_A^*F_A)\phi, D_A\phi\big\> - \big|\phi^*D_A\phi\big|^2.
\end{aligned}
\end{equation*}
Here $h(\phi) := \frac12 \abs{\mu(\phi) - c}^2$. Applying the Weiztenb\"ock formula (\ref{e:weitzenbock1}), we have
\begin{equation*}
\begin{aligned}
&D_AD_A^*D_A\phi = \Delta_A D_A\phi - D_A^*D_AD_A\phi\\
&= \nabla_A^*\nabla_A D_A\phi + F_A\#D_A\phi + R_\Sigma\# D_A\phi + R_M\#D_A\phi\#D_A\phi\#D_A\phi - D_A^*(F_A\phi)\\
&= \nabla_A^*\nabla_A D_A\phi + F_A\#D_A\phi + R_\Sigma\# D_A\phi + R_M\#D_A\phi\#D_A\phi\#D_A\phi\\
&\quad - (D_A^*F_A)\phi-F_A\#D_A\phi.
\end{aligned}
\end{equation*}
Combining these two equations together, it is not hard to see that
\begin{equation}\label{e5}
\big(\frac{\partial}{\partial t} - \Delta_g\big) |D_A\phi|^2 \le -|\nabla_A D_A\phi|^2 + C(1+ |F_A|+|D_A\phi|^2)|D_A\phi|^2.
\end{equation}
Putting (\ref{e4}) together with (\ref{e5}) yields Bochner's formula (\ref{bochner1}). The proof is complete.
\end{proof}
Now we will prove an $\epsilon$-gradient estimate for smooth solutions $(A,\phi)$ of
the YMH flow equation (\ref{e:heat}).
For any point $z :=(x,t)\in \Sigma\times(0, T)$, denote the parabolic ball, with center $z$ and radius $R>0$, by
\[ \mathbb P_R(z) := \Big\{(y, s)\in \Sigma\times \mathbb{R} \ \big |\ y\in B_R(x), \ t-R^2<s<t\Big\}. \]
\begin{lem}\label{l:ep-reg}
{\rm{($\epsilon$-regularity)}}
There exist positive constants $\epsilon_0, R_0, C_0$ depending only on $\Sigma,\ \mathscr{A}$, and $\S$ such that for any
$z_0=(x_0, t_0)\in\Sigma\times (0, T)$ and $0<R<\min\big\{R_0, \sqrt{t_0}\big\}$, if
\begin{equation}\label{e:ep-condition}
\sup_{t_0-R^2<t<t_0}\int_{B_{R}(x_0)}\big(|F_A| + |D_A\phi|^2\big)\,dv_g \le \epsilon_0,
\end{equation}
then for any $0<r<R$, it holds that
\begin{equation}\label{ep-reg}
\sup_{z\in \mathbb P_r(z_0)} \big(|F_A| + |D_A\phi|^2\big)(z)\le \frac{C_0}{(R-r)^2}.
\end{equation}
\end{lem}
\begin{proof} Let $i_0=i_0(\Sigma)>0$ denote the injectivity radius of $\Sigma$ and $0<R_0 < \min\{i_0, \sqrt{\epsilon_0}, 1\}$. Note that
$\mathcal{F}\big|_{B_{R_0}(x_0)}$ is a trivial fiber bundle. For $0<R\le R_0$, define $v:[0,R]\to\mathbb R_+$ by
\[ v(r) := (R-r)^2\sup_{z\in \mathbb P_r(z_0)}\big(\sqrt{1+|F_A|^2}+|D_A\phi|^2\big)(z). \]
Assume that $v$ attains its maximum over $[0,R]$ at $r_0\in (0, R)$. Let $z_1 := (x_1, t_1) \in \mathbb P_{r_0}(z_0)$ be
such that
\[ e_0 := \big(\sqrt{1+|F_A|^2}+|D_A \phi|^2\big)(z_1)
= \sup_{z\in \mathbb P_{r_0}(z_0)} \big(\sqrt{1+|F_A|^2}+|D_A\phi|^2\big)(z). \]
It is clear that for any $0\le r\le R$,
\begin{equation}\label{e6}
v(r) \le v(r_0) = (R-r_0)^2\sup_{z\in \mathbb P_{r_0}(z_0)}\big(\sqrt{1+|F_A|^2}+|D_A \phi|^2\big)(z) = (R-r_0)^2 {e}_0.
\end{equation}
Observe that if $(R-r_0)^2 e_0\le 4$, then we would have
$$v(r)\le v(r_0)=(R-r_0)^2 e_0\le 4,\ \forall\ 0\le r\le R, $$
which implies (\ref{ep-reg}). Hence we may assume
\begin{equation}\label{radius_ass}
(R-r_0)^2 e_0>4.
\end{equation}
For $\lambda>0$, define the parabolic dilation $P_\lambda:\mathbb R^2\times \mathbb R\to\mathbb R^2\times\mathbb R$ by
$$P_\lambda(x,t)=(\lambda x, \lambda^2 t).$$
Denote $\rho_0 = {e}_0^{-\frac12}$ and define the rescaled pair $(\phi_{\rho_0}, A_{\rho_0})$ and the metric $g_{\rho_0}$ by
$$\phi_{\rho_0}(z) = \phi\big(z_1+ P_{\rho_0}(z)\big), \ A_{\rho_0}(z) = \rho_0A\big(z_1+ P_{\rho_0} (z)\big),
\ g_{\rho_0}(x)=g(x_1+\rho_0x),
$$
for $z\in \displaystyle \mathbb P_{\frac{r_0}{\rho_0}}\Big(P_{\rho_0^{-1}}(z_0-z_1)\Big)$
and $x\in \displaystyle B_{\frac{r_0}{\rho_0}}\Big(\frac{x_0-x_1}{\rho_0}\Big)$.
It is easy to check that
\[ \big|F_{A_{\rho_0}}\big|^2(z) = \rho_0^4\big|F_A\big|^2(P_{\rho_0}(z)), \ \big|D_{A_{\rho_0}}\phi_{\rho_0}\big|^2(z)
= \rho_0^2\big|D_A\phi\big|^2(P_{\rho_0}(z)). \]
Note that (\ref{radius_ass}) is equivalent to
\begin{equation}\label{e60}
\rho_0< \frac{R-r_0}2.
\end{equation}
Observe that since
\begin{eqnarray*} \widehat{e}_{\rho_0^4}(A_{\rho_0}, \phi_{\rho_0})(z)
&=& \Big(\sqrt{\rho_0^{4}+|F_{A_{\rho_0}}|^2}+\big|D_{A_{\rho_0}}\phi_{\rho_0}\big|^2\Big)(z)\\
&=&\rho_0^2\widehat{e}_1(\phi, A)(z_1+P_{\rho_0}(z)),
\end{eqnarray*}
for all $z\in \mathbb P_{\frac{r_0}{\rho_0}}\Big(P_{\rho_0^{-1}}(z_0-z_1)\Big)$,
it follows that
\[ \widehat{e}_{\rho_0}(A_{\rho_0}, \phi_{\rho_0})(0)
= \rho_0^2\widehat{e}_1(A, \phi)(z_1)=e_0^{-1}e_0= 1. \]
From (\ref{e6}), we have that $\mathbb P_{\rho_0}(z_1) \subset \mathbb P_{\frac{R+r_0}{2}}(z_0)$ and
hence by the definition of $v(r)$, it holds
\begin{equation}\label{grad_bound}
\begin{aligned}
\sup_{z\in\mathbb P_1(0)}\widehat{e}_{\rho_0^4}(A_{\rho_0}, \phi_{\rho_0})(z)&
= \rho_0^2\sup_{z\in \mathbb P_{\rho_0}(z_1)}\widehat{e}_1(A, \phi)(z) \\
&\le e_0^{-1}\sup_{z\in\mathbb P_{\frac{R+r_0}{2}}(z_0)}
\widehat{e}_1(A, \phi)(z) = e_0^{-1}\big(\frac{R-r_0}{2}\big)^{-2} v\big(\frac{R+r_0}{2}\big)\\
&\le e_0^{-1}\cdot \big(\frac{R-r_0}{2}\big)^{-2}(R-r_0)^2 e_0 = 4.
\end{aligned}
\end{equation}
From Bochner's formula (\ref{bochner1}) for $(\phi, A)$ and straightforward calculations, we have
that
\begin{eqnarray} \big(\frac{\partial}{\partial t} - \Delta_{g_{\rho_0}}\big)
\widehat{e}_{\rho_0^4}\big(A_{\rho_0}, \phi_{\rho_0}\big)(z)
&=&\rho_0^4\Big[\big(\frac{\partial}{\partial t} - \Delta_{g}\big)
\widehat{e}_{1}\big(A, \phi\big)\Big](P_{\rho_0}(z))\nonumber\\
&\le& C\rho_0^4 \big(1+|F_A|+|D_A\phi|^2\big)\widehat{e}_1\big(A, \phi\big)(P_{\rho_0}(z))\nonumber\\
&=&C\big(\rho_0^2+|F_{A_{\rho_0}}|+|D_{A_{\rho_0}}\phi_{\rho_0}|^2\big)
\widehat{e}_{\rho_0^4}\big(A_{\rho_0}, \phi_{\rho_0}\big)(z)\nonumber\\
&\le& C\widehat{e}_{\rho_0^4}\big(A_{\rho_0}, \phi_{\rho_0}\big)(z),\label{subsolution}
\end{eqnarray}
where we have used (\ref{grad_bound}) and the fact $\rho_0\le 1$ in the last step.
Thus, by Moser's Harnack inequality for parabolic equations, we obtain
\begin{equation}
\begin{aligned}
1=\widehat{e}_{\rho_0^4}\big(A_{\rho_0}, \phi_{\rho_0}\big)(0)
&\le C \int_{\mathbb P_1(0)} \widehat{e}_{\rho_0^4}\big(A_{\rho_0}, \phi_{\rho_0}\big)= C\rho_0^{-2}\int_{\mathbb P_{\rho_0}(z_1)}\widehat{e}_1(A, \phi)\\
&\le C \sup_{t_1 - \rho_0^2 < t <t_1} \int_{B_{\rho_0}(x_1)}\widehat{e}_1(A, \phi)\\
&\le C\sup_{t_0-R^2<t<t_0}\int_{B_{R}(x_0)}(1+|F_A|+|D_A\phi|^2)\\
& \le C(\epsilon_0+R_0^2)\le C\epsilon_0.
\end{aligned}
\end{equation}
This is clearly impossible if we choose a sufficiently small $\epsilon_0>0$. Thus (\ref{radius_ass}) doesn't hold, which implies
$$e_0^{-\frac12} \ge \frac{R-r_0}{2}$$
and hence $v(r) \le 4$ for all $0\le r\le R$. In particular, we have
\[ \sup_{z\in \mathbb P_r(z_0)}\widehat{e}_1(\phi, A)(z) = \frac{1}{(R-r)^2}v(r)\le \frac{4}{(R-r)^{2}}. \]
This completes the proof.
\end{proof}
Based on the $\epsilon$-regularity Lemma \ref{l:ep-reg}, we can derive point-wise estimates for higher order
derivatives of smooth solutions $(\phi, A)$ of the YMH flow equation (\ref{e:heat}),
under the small energy condition (\ref{e:ep-condition}).
\begin{lem}\label{l:higher-reg} Under the same assumptions as in Lemma \ref{l:ep-reg}, there holds
\begin{equation}\label{e:higher-reg}
\sup_{\mathbb P_{\frac{R}{2}}(z_0)}\big(R^{l+1}|\nabla^{l-1}_A F_A|+R^l|\nabla^l_A\phi|\big) <C\big(l,\epsilon_0\big)
\end{equation}
for any $l\in \mathbb{Z}^+$.
\end{lem}
\begin{proof} Note that when $l=1$, (\ref{e:higher-reg}) follows directly from Lemma~\ref{l:ep-reg}.
By a simple scaling argument, we may assume $R=2$ and $z_0=(0,0)$.
In order to establish (\ref{e:higher-reg}) for all $l\ge 2$, we need to show
\begin{equation}\label{e9}
\int_{\mathbb P_{1+2^{-l}}(0)} \big(|\nabla^{l-1}_AF_A|^2 + |\nabla_A^l\phi|^2\big) \le C(l,\epsilon_0)
\int_{\mathbb P_2(0)} e(\phi, A),
\ \forall\ l\ge 2.
\end{equation}
From Lemma \ref{l:ep-reg}, we have the following estimate
\begin{equation}\label{grad_bound1}
\sup_{\mathbb P_{\frac32}(0)}\big(|F_A|+|D_A\phi|\big)\le C(\epsilon_0).
\end{equation}
We prove (\ref{e9}) by an induction on $l$. Consider the case $l=2$,
we may assume for simplicity that $g|_{B_2(0)}$ is the euclidean metric.
Let $\eta\in C_0^\infty(B_2(0))$ such that $0\le \eta\le 1$, $\eta\equiv 1$ on $B_{\frac54}(0)$,
$\eta\equiv 0$ outside $B_{\frac32}(0)$, and $|\nabla\eta|\le 8$.
Multiplying (\ref{e:curvature1}) by $\eta^2F_A$, integrating the resulting equation over $B_2(0)$,
and applying integration by parts and H\"older's inequality, we obtain
\begin{equation}\label{l=2}
\begin{aligned}
\frac12\dt{}\int_{B_2(0)}|F_A|^2\eta^2 &
= -\int_{B_2(0)} \big(|\nabla_A F_A|^2\eta^2+\<\nabla_A F_A , F_A\nabla \eta^2\>\big)- \int_{B_2(0)}\eta^2\<R_\Sigma\#F_A, F_A\>\\
&\quad - \int_{B_2(0)}\eta^2\<F_A\#F_A, F_A\>
- \int_{B_2(0)} \eta^2\<D_A\phi\#D_A\phi+\phi^*F_A\phi, F_A\>\\
&\le -\frac12\int_{B_2(0)}\eta^2|\nabla_A F_A|^2+
C\int_{B_{\frac32}(0)}\big(|F_A|^2+|F_A|^3 + |D_A\phi|^4\big).
\end{aligned}
\end{equation}
By Fubini's theorem, we can find $t_*\in (-4, -(\frac54)^2)$ such that
\begin{equation}\label{fubini1}
\int_{B_2(0)} (|F_A|^2+|D_A\phi|^2)(t_*)\le C\int_{\mathbb P_2(0)}(|F_A|^2+|D_A\phi|^2).
\end{equation}
Integrating (\ref{l=2}) over $t\in [t_*, 0]$ and applying (\ref{fubini1}) and (\ref{grad_bound1}), we get
\begin{eqnarray}\label{l=2.1}
&&\int_{B_{\frac54}(0)} |F_A|^2(0)+\int_{t_*}^0\int_{B_{2}(0)}\eta^2|\nabla_A F_A|^2\nonumber\\
&&\le\int_{B_{\frac54}(0)} |F_A|^2(t_*)+C\int_{\mathbb P_{\frac32}(0)}(|F_A|^2 +|F_A|^3 + |D_A\phi|^4)\nonumber\\
&&\le C\int_{\mathbb P_2(0)}(|F_A|^2+|D_A\phi|^2)+C\int_{\mathbb P_{\frac32}(0)}(|F_A|^3 + |D_A\phi|^4)\nonumber\\
&&\le C(\epsilon_0)\int_{\mathbb P_2(0)}\big(|F_A|^2+|D_A\phi|^2\big).
\end{eqnarray}
To obtain a similar estimate for $\phi$, we take $\nabla_A$ to both sides of the YMH
flow equation (\ref{e:heat})$_2$ so that
it holds
\begin{equation}\label{heat2}
\begin{aligned}
\frac{\partial}{\partial t}(\nabla_A\phi)
&= -\nabla_A\nabla_A^*\nabla_A\phi - \nabla^2h(\phi)\cdot\nabla_A\phi + \frac{\partial A}{\partial t}\phi,
\end{aligned}
\end{equation}
where $h(\phi) = \frac12|\mu(\phi)-c|^2$.
Multiplying (\ref{heat2}) by $\eta^2\nabla_A\phi$, integrating the resulting equation on
$B_2(0)$, applying integration by parts and H\"older's inequality, and using the equation (\ref{e:heat})$_1$,
we obtain that, for any $0<\delta<1$,
\begin{equation*}
\begin{aligned}
\frac12\dt{}\int_{B_2(0)} |\nabla_A\phi|^2 \eta^2&\le -\int_{B_2(0)} \big(\eta^2|\nabla_A^*\nabla_A\phi|^2
+\<\nabla_A^*\nabla_A\phi, \nabla_A\phi \nabla\eta^2\>\big)\\
&\quad+\int_{B_2(0)} \big|\nabla^2_\phi h(\phi)\big|\big|\nabla_A\phi\big|^2 \eta^2
+ \int_{B_2(0)}\big|\frac{\partial A}{\partial t}\big|\big|\nabla_A\phi\big|\eta^2\\
&\le -\frac12\int_{B_2(0)}\eta^2\big|\nabla_A^*\nabla_A\phi\big|^2
+ \delta\int_{B_2(0)}\eta^2\big|\frac{\partial A}{\partial t}\big|^2
+ C\delta^{-1}\int_{B_2(0)}\eta^2\big|\nabla_A\phi\big|^2\\
&\le -\frac12\int_{B_2(0)}\eta^2\big|\nabla_A^*\nabla_A\phi\big|^2
+ 2\delta\int_{B_2(0)}\eta^2\big|D_A^*F_A\big|^2
+ C\delta^{-1}\int_{B_2(0)}\eta^2\big|\nabla_A\phi\big|^2\\
&\le -\frac12\int_{B_2(0)}\eta^2\big|\nabla_A^*\nabla_A\phi\big|^2
+ C\delta\int_{B_2(0)}\eta^2\big|\nabla_AF_A\big|^2
+ C\delta^{-1}\int_{B_2(0)}\eta^2\big|\nabla_A\phi\big|^2.
\end{aligned}
\end{equation*}
Integrating this inequality over $[t_*,0]$ and applying (\ref{fubini1}), we obtain
\begin{equation}\label{l=2.2}
\int_{B_2(0)} |\nabla_A\phi|^2(0)+ \int_{t_*}^0\int_{B_2(0)} \eta^2\big|\nabla_A^*\nabla_A\phi\big|^2
\le C\delta\int_{t_*}^0\int_{B_2(0)}\eta^2\big|\nabla_A F_A|^2 + C\delta^{-1}\int_{\mathbb P_2(0)}e(A, \phi).
\end{equation}
Adding (\ref{l=2.1}) and (\ref{l=2.2}) and choosing a sufficiently small $\delta>0$, we obtain
\begin{equation}\label{l=2.3}
\int_{B_2(0)} e(A, \phi)(0)+\int_{t_*}^0 \int_{B_2(0)} \eta^2(\big|\nabla_A^*\nabla_A\phi\big|^2+\big|\nabla_AF_A\big|^2\big)
\le C(\epsilon_0)\int_{\mathbb P_2(0)}e(A, \phi).
\end{equation}
Note that by integrating by parts, applying Ricci's identities for interchanging derivatives, and (\ref{grad_bound1}), we have
that
\begin{equation}\label{l=2.4}
\begin{aligned}
\int_{t_*}^0\int_{B_2(0)}\eta^2\big|\nabla^2_A\phi\big|^2 &\le 2\int_{t_*}^0\int_{B_2(0)}\eta^2\big|\nabla_A^*\nabla_A \phi\big|^2
+ C\int_{t_*}^0\int_{B_2(0)}\big(|\nabla\eta|^2|\nabla_A\phi|^2+\eta^2|\nabla_A\phi|^4\big)\\
&\le 2\int_{t_*}^0\int_{B_2(0)}\eta^2\big|\nabla_A^*\nabla_A \phi\big|^2
+ C(\epsilon_0)\int_{\mathbb P_2(0)}e(A, \phi),
\end{aligned}
\end{equation}
where the constant $C>0$ depends on the curvature $F_A$ and
the curvatures of $\Sigma$ and $M$.
Combining (\ref{l=2.3}) with (\ref{l=2.4}) yields (\ref{e9}) for $l=2$.
For any $k\ge 3$, we assume that (\ref{e9}) holds for all $l\le k$. We need to show that it also holds for $l = k+1$.
To do it, first apply the operator $\nabla_A^{k}$ to both sides of the YMH flow equation (\ref{e:heat})$_2$ of $\phi$
the operator $\nabla_A^{k-1}$ to both sides of the YMH flow equation (\ref{e:curvature1}). After interchanging the
order of derivatives, we obtain a system in the following form:
\begin{equation}\label{k-ymh}
\left\{
\begin{aligned}
\frac{\partial}{\partial t}\big(\nabla_A^{k}\phi\big) &= -\nabla_A^*\nabla_A^{k+1}\phi + \mathcal Q_1^k(\nabla_A\phi, F_A)
+\nabla_A^{k}(\nabla h(\phi)), \\
\frac{\partial}{\partial t}\big(\nabla_A^{k-1}F_A\big) &= -\nabla_A^*\nabla_A^{k} F_A + \mathcal Q_2^k(\nabla_A\phi, F_A)
+\nabla_A^{k-1}(g(F_A, \phi)).
\end{aligned}
\right.
\end{equation}
where $\mathcal Q_1^k, \mathcal Q_2^k$ are lower order terms depending on derivatives of $\nabla_A\phi$ and $F_A$ up to order $k-1$, and
\[ g(F_A, \phi) = -R_\Sigma\#F_A -F_A\#F_A-D_A\phi\#D_A\phi-\phi^*F_A\phi. \]
More precisely, $\mathcal Q_1^k$ comes from changing $\nabla_A^k(\frac{\partial\phi}{\partial t})$ to $\frac{\partial}{\partial t}(\nabla_A^k \phi)$,
and $\nabla_A^k\nabla_A^*\nabla_A\phi$ to $\nabla_A^*\nabla_A\nabla_A^k\phi$.
If we denote the total curvature of the fiber bundle by $\widetilde{R}$ which involves $R_\Sigma, R_M$ and $F_A$, we have
\begin{equation}\label{e:Q1}
\mathcal Q_1^k(\nabla_A\phi, F_A) = \sum\nabla^{a}\widetilde{R}\big(\nabla_A^b \nabla_A\phi, \nabla_A^c \nabla_A\phi\big)\nabla_A^d
\nabla_A\phi,
\end{equation}
where the sum is taken over all indices $a, b, c, d\ge 0$ satisfying $a+b+c+d = k-1$. Then one can verify that
\begin{eqnarray}\label{e:Q11}
&&\Big|\sum\nabla_A^{a}\widetilde{R}\big(\nabla_A^b \nabla_A\phi, \nabla_A^c \nabla_A\phi\big)\nabla_A^d \nabla_A\phi\Big| \nonumber\\
&&\le C\sum\big|\nabla_A^{j_1} F_A\big|\cdots\big|\nabla^{j_r}_A F_A\big|\cdot\big|\nabla_A^{j_{r+1}}\nabla_A\phi\big|\cdots
\big|\nabla_A^{j_{r+s}}\nabla_A\phi\big|,
\end{eqnarray}
where $C>0$ is a constant depending on $R_\Sigma$ and $R_M$, and the indices satisfy
\begin{equation}\label{e:indices}
0\le j_1, \cdots, j_{r+s}\le k-1; \quad j_1+\cdots+ j_{r+s} = k-1; \quad s\ge 3.
\end{equation}
Moreover, we have
\begin{equation}\label{e:h}
\big| \nabla_A^k(\nabla h(\phi))\big| \le C\sum |\nabla_A^{i_1}\phi|\cdots|\nabla_A^{i_p}\phi|
\end{equation}
where $C$ is a constant depending on $h$ and the indices $i_1,\cdots, i_p\ge 1$ satisfy $i_1+\cdots i_p = k$.
Thus multiplying (\ref{k-ymh})$_1$ by $\eta^2\nabla_A^k\phi$ and integrating on $B_2(0)$, we get
\begin{equation}\label{e:Q15}
\begin{aligned}
&\frac{d}{dt}\int_{B_2(0)}\eta^2|\nabla_A^k\phi|^2 + \int_{B_2(0)}|\nabla_A^{k+1}\phi|^2\\
&\le C\int_{B_2(0)}\eta^2|\nabla_A^{k}\phi|^2 + C\int_{B_2(0)}\eta^2|\mathcal Q_1^k||\nabla_A^k\phi| + C\int_{B_2(0)}\eta^2|\nabla_A^k(\nabla h(\phi))||\nabla_A^k\phi|\\
&=: I + II + III.
\end{aligned}
\end{equation}
By (\ref{e:Q1}) and (\ref{e:Q11}), we have
\begin{equation}\label{e:Q12}
\begin{aligned}
II \le &C\int_{B_2(0)}\eta^2|F_A|^r|\nabla_A\phi|^{s-1}|\nabla_A^k\phi|^2 + C\int_{B_2(0)}\eta^2|F_A|^{r-1}|\nabla_A\phi|^{s}|\nabla_A^{k-1}F_A||\nabla_A^k\phi|\\
& +C\sum\int_{B_2(0)}\eta^2|\nabla_A^{j_1} F_A|\cdots|\nabla_A^{j_r} F_A|\cdot|\nabla_A^{j_{r+1}}\nabla_A\phi|\cdots|\nabla_A^{j_{r+s}}\nabla_A\phi||\nabla_A^k\phi|\\
\le &C(\epsilon_0)\int_{B_2(0)}\eta^2(|\nabla_A^k\phi|^2+|\nabla_A^{k-1}F_A|^2)\\
& +C\sum\int_{B_2(0)}\eta^2|\nabla_A^{j_1} F_A|^2\cdots|\nabla_A^{j_r} F_A|^2\cdot|\nabla_A^{j_{r+1}}\nabla_A\phi|^2\cdots|\nabla_A^{j_{r+s}}\nabla_A\phi|^2,
\end{aligned}
\end{equation}
where the sum in the last term is taken over all the indices satisfying (\ref{e:indices}) with $j_1,\cdots j_{r+s}\le k-2$.
Recall the Sobolev embedding for a section $\Psi$ in dimension 2
\begin{equation*}
\norm{\Psi}_{L^q}\le C\norm{\Psi}_{W^{1,2}} \le C\norm{\Psi}_{H^{1,2}}, ~\forall q>1
\end{equation*}
where $\norm{\Psi}_{W^{1,2}}$ is the usual Sobolev norms of $|\Phi|$ and
\begin{equation*}
\norm{\Psi}_{H^{1,2}} := \norm{\Psi}_{L^2} + \norm{\nabla_A\Psi}_{L^2}
\end{equation*}
is the Sobolev norms with respect to $\nabla_A$. It follows that that for all $q>1, j\le k-2$, we have
\begin{equation*}
\norm{\nabla_A^{j}F_A}_{L^q}\le C\norm{\nabla_A^j F_A}_{H^{1,2}}\le C\sum_{i=1}^{k-1} \norm{\nabla_A^i F_A}_{L^2},
\end{equation*}
and
\begin{equation*}
\norm{\nabla_A^{j}\nabla_A\phi}_{L^q}\le C\norm{\nabla_A^{j+1}\phi}_{H^{1,2}} \le C\sum_{i=1}^{k-1} \norm{\nabla_A^{i+1}\phi}_{L^2}.
\end{equation*}
Applying this and the H\"older inequality in (\ref{e:Q12}), we get
\begin{equation}\label{e:Q13}
II\le C(\epsilon_0)\sum_{i=1}^{k-1}\left(\int_{B_2(0)}\eta^2(|\nabla_A^i F_A|^2 + |\nabla_A^{i+1}\phi|^2)\right)
\end{equation}
By (\ref{e:h}) and a similar argument, we have
\begin{equation}\label{e:Q14}
III\le C(\epsilon_0)\sum_{i=1}^{k-1}\left(\int_{B_2(0)}\eta^2|\nabla_A^{i+1}\phi|^2\right)
\end{equation}
Now inserting (\ref{e:Q13}) and (\ref{e:Q14}) back into (\ref{e:Q15}), we arrive at
\begin{equation*}
\frac{d}{dt}\int_{B_2(0)}\eta^2|\nabla_A^k\phi|^2 + \int_{B_2(0)}|\nabla_A^{k+1}\phi|^2 \le C(\epsilon_0)\sum_{i=1}^{k-1}\left(\int_{B_2(0)}\eta^2(|\nabla_A^i F_A|^2 + |\nabla_A^{i+1}\phi|^2)\right).
\end{equation*}
Consequently, integrating this inequality on $t$ as we did before and using the induction assumption, we obtain the desired bound
\begin{equation*}
\int_{t_*}^0\int_{B_2(0)}|\nabla_A^{k+1}\phi|^2 \le C(\epsilon_0)\sum_{i=1}^{k-1}\left(\int_{t_*}^0\int_{B_2(0)}\eta^2(|\nabla_A^i F_A|^2 + |\nabla_A^{i+1}\phi|^2)\right).
\end{equation*}
The estimate on $\nabla_A^{k}F_A$ can be achieved in the same way. Namely, $\mathcal Q_2^k$ emerges from interchanging the order of derivatives on $F_A$. Since we have
\begin{equation*}
\Big|\nabla_A^{k-1}\nabla_A^*\nabla_A F_A - \nabla_A^*\nabla_A^k F_A\Big|
\le C \sum_{i=0}^{k} \big|\nabla_A^i F_A\#\nabla_A^{k-i}F_A\big|,
\end{equation*}
it is easy to verify that the following estimate holds:
\begin{equation*}
\begin{aligned}
\Big|\mathcal Q_2^k(\nabla_A\phi, F_A)\Big|+\Big|\nabla^{k-1}_A(g(F_A,\phi))\Big|\le
& C \sum_{i=0}^{k} \left(|\nabla_A^i F_A||\nabla_A^{k-i}F_A| + |\nabla^{i+1}_A \phi||\nabla^{k-i}_A\phi|\right)\\
&+ C\sum_{0\le j_1+j_2+j_3\le k} |\nabla_A^{j_1} F_A||\nabla_A^{j_2}\phi||\nabla_A^{j_3}\phi|.
\end{aligned}
\end{equation*}
Then multiplying the equation (\ref{k-ymh})$_2$ by $\nabla_A^{k-1}F_A$ and integrating by parts, we may obtain the desired bounds (\ref{e9}) for $l = k+1$.
Finally, (\ref{e9}) together with the parabolic Sobolev embedding theorems completes our proof.
\end{proof}
\begin{rem} {\rm In dimension two, the assumption (\ref{e:ep-condition}) in Lemma \ref{l:ep-reg}
can be weaken to that there exists $R_0>0$ depending on $(\phi_0, A_0)$ and $\epsilon_0>0$ such that
\begin{equation}\label{e:ep-condition1}
\sup_{t_0-R_0^2\le t\le t_0}\int_{B_{R_0}(x_0)}|D_A\phi|^2\,dv_g\le \epsilon_0.
\end{equation}
In fact, it follows from H\"older's inequality and the energy inequality (\ref{e:energy-inequ}) that
$$\int_{B_{R_0}(x_0)}|F_A|\,dv_g\le {\rm{Vol}}(B_{R_0}(x_0))\Big(\int_{B_{R_0}(x_0)}|F_A|^2\,dv_g\Big)^\frac12
\le C \big(\mathcal E(t)\big)^\frac12R_0\le C\big(\mathcal E(0)\big)^\frac12 R_0\le \epsilon_0,$$
provided $\displaystyle R_0\le \frac{\epsilon_0}{C\sqrt{\mathcal E(0)}}$. }
\end{rem}
\section{Local weak solutions of the YMH flow}
In this section, we will establish the local existence of
weak solutions to the initial value problem of the YMH flow equation~(\ref{e:heat}),
with any initial data $(A_0, \phi_0)\in \mathscr{A}_{1,2}\times \S_{1,2}$. For simplicity, we will denote, for $0<T\le +\infty$ and $1<p\le \infty, 1<q\le \infty$,
\begin{eqnarray*}
L^p(L^q):=L^p([0,T), L^q(\Sigma)), \
L^p(H^1):=L^p([0,T), H^1(\Sigma)), \
H^1(L^q):=H^1([0,T), L^q(\Sigma)).
\end{eqnarray*}
Now we state the theorem on the local existence of weak solutions.
\begin{thm}\label{t:local}
Assume $(A_0, \phi_0)\in \mathscr{A}_{1,2}\times\S_{1,2}$,
there exist $0<T_0\le +\infty$ and a weak solution $(A, \phi)$ to
the YMH flow equation~(\ref{e:heat}) and the initial condition (\ref{IVP}) on the interval $[0, T_0)$. Moreover, $(A,\phi)$ enjoys the properties that
$F_A\in L^\infty(L^2)$, $\phi\in L^\infty(H^1)$, and there exists a gauge transformation $s\in \mathscr{G}_{1,2}$
such that $(s^*A,s^*\phi)\in C^\infty\big((0,T_0), \mathscr{A}\times \S\big)$. Finally, if $0<T_0<+\infty$ is the maximal time interval for the weak solution,
then
\begin{equation}\label{local-ymh-concen}
\limsup_{t\uparrow T_0} \max_{x\in\Sigma}\int_{B_R(x)} e(A(t), \phi(t))\,dv_g\ge \alpha(M), \ \forall\ R>0,
\end{equation}
where $\alpha(M)>0$ is a positive constant given by
\begin{equation}\label{energy-gap}
\alpha(M):=\inf\Big\{\int_{\mathbb S^2} |\nabla h|^2\,dv_{g_0}: \ h\in C^\infty(\mathbb S^2, M)\ {\rm{is\ a\ nontrivial\ harmonic\ map}} \Big\}.
\end{equation}
\end{thm}
\begin{proof} We divide the proof into five steps:
\smallskip
\noindent{\it Step 1: Smooth approximation of initial data.} Since $P\times_G M$ is a smooth manifold and ${\rm{dim}}(\Sigma)=2$, it follows from Schoen-Uhlenbeck's
density Lemma (cf. \cite{SU}), a local trivialization of the bundle $\mathcal{F}$, and the partition of unity that there exists a sequence of smooth sections
$\phi_0^n\in \S$ such that $\phi_0^n\rightarrow \phi$ in $H^1(\Sigma)$. It is standard that
there is a sequence of smooth sections $A_{0}^n\in \mathscr{A}$ such that $A_{0}^n\rightarrow A_0$ in $H^1(\Sigma)$.
Thus we may assume a uniform energy bound of $(A_0^n,\phi_0^n)$:
\begin{equation}\label{e10}
\mathcal{E}_n(0): = \mathcal{E}\big(A_{0}^n, \phi_{0}^n\big) \le C (=1+\mathcal{E}(0)).
\end{equation}
Applying Theorem~\ref{t:smooth}, we conclude that there exist $T_n>0$ and a unique smooth solution $(A_n, \phi_n)\in C^\infty([0,T_n), \mathscr{A}\times \S)$
to the YMH heat flow equation~(\ref{e:heat}) and the initial condition (\ref{IVP}) with $(A_0,\phi_0)$ replaced by $(A_{0}^n, \phi_{0}^n)$.
We may assume that $T_n>0$ is the maximal time interval for $(A_n,\phi_n)$.
\smallskip
\noindent{\it Step 2: A uniform lower bound of $T_n$}. Let $\epsilon_0$ and $R_0$ be the constants given by Lemma~\ref{l:ep-reg}.
Let $\epsilon_0>0$ and $R_0>0$ be the constants given by lemma~\ref{l:ep-reg}. Since $(A_{0}^n, \phi_{0}^n)\rightarrow (A_0, \phi_0)$ in $H^1(\Sigma)$,
there exists a uniform $R_1\in (0, R_0)$ independent of $n$ such that
\begin{equation}\label{e11}
\mathcal{E}_n(0, 2R_1): =\max_{x\in\Sigma} \int_{B_{2R_1}(x)}e(A_{0}^n, \phi_{0}^n)\,dv_g \le \frac{\epsilon_0}{2}.
\end{equation}
Then by Lemma~\ref{l:energy-inequ-local}, there exists $\theta_0\in (0,1)$, depending only on $\Sigma$, $\mathcal{E}(0)$ and $R_1$, such that
for $T_n^1=\min\big\{T_n, \theta_0 R_1^2\big\}>0$ there holds
\begin{eqnarray}\label{e12}
\mathcal{E}_n(t, {R_1}):=\max_{x\in\Sigma} \int_{B_{R_1}(x)}e\big(A_n(t), \phi_n(t)\big)\,dv_g
&\le& \mathcal{E}_n(0, 2R_1) + CT_n^1R_1^{-2}\L_n(0)\nonumber\\
&\le& \frac{\epsilon_0}2+CT_n^1R_1^{-2}\big(1+\mathcal{E}(0)\big)\nonumber\\
&\le& \epsilon_0,
\end{eqnarray}
for all $0\le t\le T_n^1$.
Now we want to show $T_n\ge \theta_0 R_1^2$. Suppose, otherwise, $T_n<\theta_0R_1^2$. Then we have
$T_n=T_n^1<\theta_0R_1^2$ so that (\ref{e12}) holds for $0\le t\le T_n$. Applying Lemma \ref{l:ep-reg}
and Lemma \ref{l:higher-reg}, we can conclude that for any $0<\delta<T_n$ it holds
\begin{equation}\label{e:uniform-est}
\sup_n\max_{(x,t)\in\Sigma\times [\delta_n, T_n)} \Big(\big|\nabla^{l-1}F_{A_n}\big|+\big|\nabla^l_{A_n}\phi_n\big|\Big)
\le C(l,R_1,\epsilon_0,\delta), \ \forall\ l\ \in\mathbb Z^+.
\end{equation}
Applying (\ref{e:uniform-est}) (with $\delta=\frac{T_n}2$) and taking $l$-order derivatives $\nabla_A^l$ to the equation (\ref{e:heat})
for $l\ge 1$, it is not hard to see that for any $l\in \mathbb Z^+$, $\big(A_n(t), \phi_n(t)\big)$ is uniformly bounded
in $C^l(\Sigma)$ for $0\le t<T_n$. Therefore
there exists $(A_n(T_n), \phi_n(T_n))\in \mathscr{A}\times\S$ such that
$$\lim_{t\uparrow T_n}\big(A_n(t),\phi_n(t)\big)=\big(A_n(T_n), \phi_n(T_n)\big)
\ {\rm{in}}\ C^l(\Sigma), \ \forall \ l\in \mathbb Z^+.$$
This contradicts the maximality of $T_n$. Thus we must have that
$T_n\ge \theta_0 R_1^2$ for all $n\ge 1$.
\smallskip
\noindent{\it Step 3: Weak convergence of $(A_n,\phi_n)$ in $\Sigma\times [0,\theta_0R_1^2]$.}
From the energy inequality (\ref{e:energy-inequ}), we have
\begin{equation}\label{e:t-energy-bd}
\sup_{0\le t\le\theta_0R_1^2}\mathcal{E}_n(t)+ \int_0^{\theta_0R_1^2}\int_\Sigma \Big(\big|\frac{\partial\phi_n}{\partial t}\big|^2+\big|{\frac{\partial A_n}{\partial t}}\big|^2\Big)\,dv_gdt
\le \mathcal{E}_n(0) \le C.
\end{equation}
This and direct calculations imply that $\big\|A_n(t)\big\|_{L^2(\Sigma)}\in H^1\big([0,\theta_0R_1^2]\big)
\subset C^\frac12\big([0,\theta_0R_1^2]\big)$ and satisfies the estimate: for any $0\le t_1\le t_2\le\theta_0R_1^2$,
\begin{equation} \label{e:holder-est}
\Big| \big\|A_n(t_1)\big\|_{L^2(\Sigma)}-\big\|A_n(t_2)\big\|_{L^2(\Sigma)}\Big|\le \Big\|\frac{\partial A_n}{\partial t}\Big\|_{L^2(\Sigma\times
[0,\theta_0R_1^2])} |t_1-t_2|^\frac12\le C|t_1-t_2|^\frac12.
\end{equation}
By (\ref{e10}), (\ref{e:holder-est}), and (\ref{e:t-energy-bd}), we also have that $\phi_n$ is bounded in $L^2([0,\theta_0R_1^2], \S_{1,2})$,
i.e.,
\begin{equation} \label{e:spat-energy-bd}
\sup_{n}\big\|\phi_n\big\|_{L^2([0,\theta_0R_1^2], H^1(\Sigma))}\le C\mathcal{E}_n(0)\le C.
\end{equation}
It follows from (\ref{e:t-energy-bd}), (\ref{e:holder-est}), and (\ref{e:spat-energy-bd}) that we may assume that
there exist a connection $A\in L^\infty([0, \theta_0R_1^2], \mathscr{A}_{0,2})$ and $\phi\in L^2([0,\theta_0R_1^2], \S_{1,2})$
such that after passing to a subsequence,
\begin{equation}\label{weak1}
A_n\rightharpoonup A \ {\rm{weak^*\ in}}\ L^\infty(L^2); \ \phi_n\rightharpoonup \phi \ {\rm{in}}\ L^2(H^1) \ {\rm{and}}\
\phi_n\rightarrow \phi \ {\rm{in}}\ L^2(L^2).
\end{equation}
From (\ref{e10}), we may assume that there exist $F\in L^\infty([0,\theta_0R_1^2], L^2(\Sigma))$,
$\psi\in L^\infty([0,\theta_0R_1^2], L^2(\Sigma))$ such that after passing to a subsequence,
\begin{equation}\label{weak2}
F_{A_n}\rightharpoonup F; \ D_{A_n}\phi_n\rightharpoonup \psi\ \ {\rm{in}}\ L^2(L^2).
\end{equation}
\smallskip
\noindent{\it Step 4: Uhlenbeck's gauge fixings and smooth convergence of $(A_n,\phi_n)$.}
In order to show $(A,\phi)$ is a weak solution of (\ref{e:heat}) and (\ref{IVP}), we need to identify $F$ and $\psi$
with $F_A$ and $D_A\phi$ respectively. To achieve this, we need to control the connections
$A_n$ through Uhlenbeck's gauge fixing techniques. To do it, we adopt the argument by \cite{HTY} Lemma 3.7.
First note that $(A_n, \phi_n)$ satisfies the smallness condition:
\begin{equation}\label{e:uniform-small}
\max_{0\le t\le \theta_0R_1^2}\mathcal{E}_n(t, R_1)\le\epsilon_0.
\end{equation}
Hence, by Lemma \ref{l:ep-reg} and Lemma \ref{l:higher-reg}, there exists $C>0$ such that for any
$n$, $(A_n, \phi_n)$ satisfies,
for any $k\ge 0$,
\begin{equation}\label{e:higher-est1}
\big\|\nabla^k_{A_n}F_{A_n}(t)\big\|_{L^\infty(\Sigma)}\le \frac{C\epsilon_0}{t^k},
\ \big\|\nabla^{k+1}_{A_n}\phi_n(t)\big\|_{L^\infty(\Sigma)}\le \frac{C\epsilon_0}{t^{\frac{k}2}},
\ \forall\ 0<t\le\theta_0R_1^2.
\end{equation}
Set $t_0=\theta_0R_1^2$. Using (\ref{e:higher-est1}) with $k=0$, we can apply Uhlenbeck's gauge fixing (see \cite{Uh}
Theorem 1.5) to obtain a sequence of Coulomb gauge transforms $s_n\in \mathscr{G}_{1,2}$ such that
$\widetilde{A}_n(t_0)=s_n^*(A_n)(t_0)\in \mathscr{A}$ and satisfies
\begin{equation}\label{slice-uniform-est}
\Big\|\widetilde{A}_n(t_0)\Big\|_{C^k(\Sigma)}\le C\Big(\|A_n(t_0)\|_{L^2(\Sigma)}+\sum_{l=0}^{k-1}\big\|\nabla^l_{A_n}F_{A_n}(t_0)\big\|_{L^\infty(\Sigma)}\Big)\le C(k, t_0, \epsilon_0),
\ \forall\ k\ge 1.
\end{equation}
Set $\widetilde{\phi}_n=s_n^*(\phi_n)$. Since (\ref{e:heat}) is invariant under time-independent gauge transforms,
$(\widetilde{A}_n,\widetilde{\phi}_n)$ is also a solution of (\ref{e:heat}) in
$[0, t_0]$. Set
$$\widetilde{\mathcal E}_n(t, R_1):=\max_{x\in\Sigma}\int_{B_{R_1}(x)} e(\widetilde{A}_n(t), \widetilde{\phi}_n(t))\,dv_g.$$
Then, since the local Yang-Mills-Higgs energy $\mathcal{E}_n(t,R_1)$ is invariant under gauge transformations,
(\ref{e:uniform-small}) holds for $(\widetilde{A}_n,\widetilde{\phi}_n)$, i.e.,
$$
\max_{0\le t\le \theta_0R_1^2}\widetilde{\mathcal E}_n(t, R_1)=
\max_{0\le t\le \theta_0R_1^2}\mathcal{E}_n(t, R_1)\le\epsilon_0.
$$
so that by Lemma \ref{l:ep-reg}
and Lemma \ref{l:higher-reg}, (\ref{e:higher-est1}) holds for $(\widetilde{A}_n,\widetilde{\phi}_n)$, i.e.,
\begin{equation}\label{e:higher-est2}
\big\|\nabla^k_{\widetilde{A}_n}F_{\widetilde{A}_n}(t)\big\|_{L^\infty(\Sigma)}\le \frac{C\epsilon_0}{t^k},
\ \big\|\nabla^{k+1}_{\widetilde{A}_n}\widetilde{\phi}_n(t)\big\|_{L^\infty(\Sigma)}\le \frac{C\epsilon_0}{t^{\frac{k}2}},
\ \forall\ 0<t\le\theta_0R_1^2.
\end{equation}
Now we take $\partial_t^k\nabla_{\widetilde{A}_n}^l$ of both sides of the equation (\ref{e:heat}) for
$(\widetilde{A}_n, \widetilde{\phi}_n)$ for any $k,l\ge 1$, and apply (\ref{slice-uniform-est}) and
(\ref{e:higher-est2}) to conclude that
$(\widetilde{A}_n, \widetilde{\phi}_n)\in C^\infty(\Sigma\times (0,t_0])$ satisfies
\begin{equation}\label{e:uniform-est2}
\sup_{n} \Big\|\big(\widetilde{A}_n, \widetilde{\phi}_n\big)\Big\|_{C^k(\Sigma\times [\delta, t_0])}
\le C(k, \delta, t_0,\epsilon_0), \ \forall\ k\ge 0, \ \forall\ 0<\delta<t_0.
\end{equation}
From (\ref{e:uniform-est2}), we may assume there exists $\big(\widetilde{A}, \widetilde{\phi}\big)\in C^\infty((0, t_0], \mathscr{A}\times \S)$ such
that after passing to a subsequence, $(\widetilde{A}_n, \widetilde{\phi}_n)\rightarrow (\widetilde{A},\widetilde{\phi})$ in
$C^k(\Sigma\times [\delta, t_0])$ for any $k\ge 1$ and $0<\delta<t_0$. Since $\big(\widetilde{A}_n,\widetilde{\phi}_n\big)$
are smooth solutions of (\ref{e:heat}) on $\Sigma\times (0,t_0]$, it follows that $\big(\widetilde{A},\widetilde{\phi}\big)$ is also a smooth
solution of (\ref{e:heat}).
Since $\displaystyle\widetilde{A}_n(t_0)=s_n^*(A_n)(t_0)=s_n^{-1} ds_n+s_n^{-1} A_n(t_0)s_n$, it follows from (\ref{weak1}) and
(\ref{slice-uniform-est}) that
\begin{equation}\label{h1-bound}
\|ds_n\|_{L^2(\Sigma)}\le C\big(\|\widetilde{A}_n(t_0)\|_{L^2(\Sigma)}+\|A_n(t_0)\|_{L^2(\Sigma)}\big)\le C.
\end{equation}
Thus we may assume there exists $s\in \mathscr{G}_{1,2}$ such that after passing to a subquence,
$s_n\rightharpoonup s \ {\rm{in}}\ H^1(\Sigma)$ and $s_n\rightarrow s$ in $L^2(\Sigma)$.
This, combined with (\ref{weak1}), implies that $\big(\widetilde{A},\widetilde{\phi}\big)=\big(s^*A, s^*\phi\big)$ or
equivalently $(A,\phi)=\Big((s^{-1})^*\widetilde{A}, (s^{-1})^*\widetilde\phi\Big)$ in $\Sigma\times (0,t_0]$. Since
(\ref{e:heat}) is invariant under time-independent gauge transformations, we conclude that $(F,\psi)=(F_A, D_A\phi)$,
and $(A,\phi)$ is a weak
solution of (\ref{e:heat}) in $\Sigma\times [0,t_0]$ and satisfies (\ref{IVP}).
\smallskip
\noindent{\it Step 5: Characterization of a finite maximal time interval $T_0$}.
Let $\epsilon_0>0$ be given by Lemma \ref{l:ep-reg}.
First we claim if $0<T_0<+\infty$ is the maximal time interval for a weak solution $(A,\phi)$ constructed through step 1 to step 4, then
\begin{equation}\label{local-ymh-concen1}
\limsup_{t\uparrow T_0} \max_{x\in\Sigma}\int_{B_R(x)} e(A(t), \phi(t))\,dv_g\ge \frac{\epsilon_0}2,\ \forall\ R>0.
\end{equation}
For, otherwise, there exists $R_0>0$ such that for any sufficiently small $\delta>0$,
$$\mathcal{E}(T_0-\delta^2, 2R_0):=\max_{x\in\Sigma}\int_{B_{2R_0}(x)\times \{T_0-\delta^2\}} e(A(t), \phi(t))\,dv_g<\frac{\epsilon_0}2.$$
This, combined with lemma \ref{l:energy-inequ-local}, implies that there exists $\theta_0\in (0,1)$ such that for $T_0-\delta^2\le t\le T_0-\delta^2+(\theta_0 R_0)^2$,
$$\mathcal{E}(t, R_0)\le \mathcal{E}(T_0-\delta^2, 2R_0)+\frac{C(t-(T_0-\delta^2))}{R_0^2}\mathcal{E}(0)\le \frac{\epsilon_0}2+ C\theta_0^2 \mathcal{E}(0)\le \epsilon_0.$$
By choosing $\delta=\theta_0R_0$, this implies that
\begin{equation}\label{nonsing-time}
\sup_{T_0-\theta_0^2R_0^2\le t\le T_0}\mathcal{E}(t,R_0)\le\epsilon_0.
\end{equation}
From (\ref{nonsing-time}), we can repeat the argument from step 1 to step 4 to conclude that there exists a $s\in\mathscr{G}_{1,2}$
such that $\big(s^*A(T_0), s^*\phi(T_0)\big)\in \mathscr{A}\times \S$ and hence $T_0$ is not a maximal time interval. This contradicts the definition of $T_0$.
The improvement of $\frac{\epsilon_0}2$ in (\ref{local-ymh-concen1}) to $\alpha(M)$ in (\ref{local-ymh-concen}) follows from the blow-up analysis
performed near $T_0$, see section 6 below.
\end{proof}
For the uniqueness of weak solutions, we have
\begin{thm}
If the initial data belongs to $W^{2,p}$ for $p>2$, then the weak solution is unique.
\end{thm}
\begin{proof}
The proof follows exactly as the smooth case in Theorem~\ref{t:smooth}. Note that when $A\in W^{2,p}, p>2$, the coefficients of equation~(\ref{e:gauge-DeTurck}) belong to $C^0$ and the uniqueness of solutions to the ordinary differential equation
is guaranteed. Thus the arguments in the proof of uniqueness part of Theorem~\ref{t:smooth} holds the same in this case.
\end{proof}
\section{Compactness of approximate YMH fields}
In \cite{So1}, the first author discussed in detail the convergence and blow-up behavior of a sequence of YMH fields. It was shown that a sequence of YMH fields with bounded YMH energy converges to a YMH field along with
possibly finitely many bubbles, i.e., nontrivial harmonic maps from $\mathbb S^2$, which are attached to the limiting YMH field. This phenomenon, referred as {\it bubble tree} convergence, has first been established in the study of compactness of harmonic maps from Riemann surfaces, see for example \cite{P1}.
Extending the arguments in \cite{So1}, we can prove the bubble tree convergence of a sequence of approximate YMH fields, which is needed to describe the asymptotic behavior of the YMH heat flow at both finite singular times and the time infinity. More precisely, for a pair $(A, \phi)\in \mathscr{A}_{1,2}\times \S_{1,2}$, set
\[ \tau_1(A,\phi) = D_A^*F_A+\phi^*D_A\phi;\ ~ \tau_2(A, \phi) = D^*_AD_A\phi+(\mu(\phi)-c)\cdot\nabla\mu(\phi). \]
We will show the bubble tree convergence of a sequence $(A_n, \phi_n)\in \mathscr{A}_{1,2}\times \S_{1,2}$, with bounded YMH energies $\mathcal{E}(A_n,\phi_n)$, satisfying
\begin{equation}\label{e:bounded-tension}
\norm{\tau_1(A_n,\phi_n)}_{L^2(\Sigma)} + \norm{\tau_2(A_n, \phi_n)}_{L^2(\Sigma)} \le C.
\end{equation}
\begin{thm}\label{t:compact}
Suppose $(A_n, \phi_n)\in \mathscr{A}_{1,2}\times \S_{1,2}$ is a sequence of fields, with bounded YMH energies $\mathcal{E}(A_n,\phi_n)$, which satisfies~(\ref{e:bounded-tension}).
Then there exist a subsequence of $(A_n,\phi_n)$, still denoted as $(A_n, \phi_n)$, a set of finitely many points $\mathbf{x} = \{x_1, x_2,\cdots, x_k\}\subset \Sigma$,
and an approximate YMH field $(A_\infty, \phi_\infty)\in \mathscr{A}_{2,2}\times \S_{2,2}$, with $L^2$-tension fields
$\tau_1(A_\infty, \phi_\infty)$ and $\tau_2(A_\infty, \phi_\infty)$, such that the following properties hold:
\begin{enumerate}
\item There exist gauge transformations $\{s_n\}\subset\mathscr{G}_{1,2}$ such that $s_n^*A_n\rightarrow A_\infty$ in $H^1(\Sigma)$ and
$s_n^*\phi_n\rightarrow\phi_\infty$ in $H^1_{loc}(\Sigma\setminus\mathbf{x})$.
\item There exist finitely many nontrivial harmonic maps $\omega_{ij}:\mathbb S^2\to M, 1\le i\le k, 1\le j\le l$, such that
\[\lim_{n\to \infty} \mathcal{E}(A_n, \phi_n) = \mathcal{E}(A_\infty, \phi_\infty) + \sum_{1\le i\le k,1\le j\le l}\mathcal{E}(\omega_{ij}),\]
where $\displaystyle \mathcal{E}(\omega_{ij}) = \int_{\mathbb S^2}\big|\nabla\omega_{ij}\big|^2dv_{g_0}$ is the Dirichlet energy of $\omega_{ij}$.
\item The images of the bubbles $\displaystyle \{\omega_{ij}\}_{1\le i\le k, 1\le j\le l}$ and the limiting map $\phi_\infty$ are connected.
\end{enumerate}
\end{thm}
Theorem \ref{t:compact} has been proved by \cite{So1} for YMH fields $(A_n,\phi_n)$, i.e,
$\tau_1(A_n,\phi_n)=\tau_2(A_n,\phi_n)=0$. It turns out the argument in \cite{So1} can be modified
to prove Theorem \ref{t:compact} on approximate YMH fields satisfying the condition (\ref{e:bounded-tension}).
Here we sketch a proof. First we need the following lemma.
\begin{lem}\label{l:reg} There exists $\epsilon_0>0$ such that for $x_0\in\Sigma$ and $r_0>0$ if $A \in \mathscr{A}_{1,2}\big |_{B_r(x_0)}$
is a connection on the ball $ B_{r_0}(x_0)\subset\Sigma$, with $D_A^*F_A\in L^2(B_{r_0}(x_0))$, which satisfies
\begin{equation}\label{small-curvature}
r_0\big\|F_A\big\|_{L^2(B_{r_0}(x_0))}\le\epsilon_0,
\end{equation}
then there exists a gauge transformation $s\in \mathscr{G}_{2,2}$ such that the following estimate holds:
\begin{equation}\label{h2-estimate}
\norm{s^*A}_{H^2(B_{\frac{r_0}2}(x_0))} \le C(r_0)(\norm{F_A}_{L^2(B_{r_0}(x_0))} + \norm{D_A^*F_A}_{L^2(B_{r_0}(x_0))}).
\end{equation}
\end{lem}
\begin{proof} If we define $\widehat{A}(x)=r_0A(x_0+r_0x)$ and $\widehat{g}(x)=g(x_0+r_0x)$ for $x\in B_1$, then
by simple scaling arguments we have that $\widehat{A} \in \mathscr{A}_{1,2} \big|_{B_1}$, with $D_{\widehat A}^*F_{\widehat{A}}\in L^2(B_1)$,
and satisfies
$$\big\|D_{\widehat A}^*F_{\widehat{A}}\big\|_{L^2(B_1)}
=r_0^2 \big\|D_{A}^*F_A\big\|_{L^2(B_{r_0}(x_0)}\le \big\|D_{A}^*F_A\big\|_{L^2(B_{r_0}(x_0)},$$
and
$$\int_{B_1} |F_{\widehat{A}}|^2\,dv_{g_0}
=\int_{B_{r_0}(x_0)} r_0^2|F_{A}|^2\,dv_g<\epsilon_0^2.$$
Thus we may assume, for simplicity, that $x_0=0$ and $r_0=1$.
Since $A\in\mathscr{A}_{1,2}$ satisfies (\ref{small-curvature}), it follows from \cite{Uh} Theorem 1.3
that there exists a gauge transform $s\in \mathscr{G}_{1,2}$ such that $\widetilde A=s^*A$ satisfies
\begin{equation}\label{coulomb-gauge1}
d^*\widetilde A=0, \ \big\|\widetilde A\big\|_{H^1(B_1)}\le C\big\|F_A\big\|_{L^2(B_1)}.
\end{equation}
Using $d^*\widetilde A=0$, we obtain the following elliptic equation:
\begin{equation}\label{e:ep1}
\Delta \widetilde A + \big[\widetilde A, d\widetilde A\big] + \big[\widetilde A, [\widetilde A,\widetilde A]\big]
=D_{\widetilde A}^*F_{\widetilde A}.
\end{equation}
By Sobolev's embedding $H^1(B_1) \hookrightarrow L^q(B_1)$ for any $1<q<+\infty$, and the standard $W^{2,p}$-estimate
on (\ref{e:ep1}), we have that for any $1<p<2$, $\widetilde A\in W^{2,p}(B_{\frac34})$ and
$$\big\|\widetilde A\big\|_{W^{2,p}(B_{\frac34})}\le C\Big(\big\|\widetilde A\big\|_{H^1(B_1)}
+\big\|D_{\widetilde A}^*F_{\widetilde A}\big\|_{L^2(B_1)}\Big)\le C\big(\norm{F_A}_{L^2(B_1)} + \norm{D_A^*F_A}_{L^2(B_1)}\big).
$$
This, combined with Sobolev's embedding $W^{2,p}(B_{\frac34}) \hookrightarrow C^0(B_{\frac34})$,
yields $\widetilde A\in C^0(B_{\frac34})$ and
$$
\big\|\widetilde A\big\|_{L^\infty(B_{\frac34})}\le C\big(\norm{F_A}_{L^2(B_1)} + \norm{D_A^*F_A}_{L^2(B_1)}\big).
$$
Now we can apply $W^{2,2}$-estimate to (\ref{e:ep1}) to obtain that $\widetilde A\in W^{2,2}(B_{\frac12})$ satisfies the desired
estimate (\ref{h2-estimate}). This, after scaling back to the original scale, completes the proof.
\end{proof}
Following the gluing procedure given by \cite{Uh} and Lemma \ref{l:reg}, we obtain the following proposition.
\begin{prop} \label{l:global-gauge-fix} For any $\Lambda>0$, there exists $C(\Lambda)>0$ such that
if $A\in \mathscr{A}_{1,2}$, with $D_A^*F_A\in L^2(\Sigma)$, has
\begin{equation}
\|F_A\|_{L^2(\Sigma)}\le \Lambda,
\end{equation}
then there is a gauge transform $s\in \mathscr{G}_{2,2}$
such that $\widetilde A=s^*A\in \mathscr{A}_{2,2}$ satisfies
\begin{equation}\label{h2-estimate3}
\big\|\widetilde A\big\|_{H^2(\Sigma)}\le C(\Lambda)\Big(\big\|F_A\big\|_{L^2(\Sigma)}+\big\|D_A^*F_A\big\|_{L^2(\Sigma)}\Big).
\end{equation}
\end{prop}
\begin{proof} Choose $0<r_0\le \frac{\epsilon_0}{\Lambda}$, we have
$$r_0\big\|F_A\big\|_{L^2(B_{r_0}(x_0))}\le \epsilon_0$$
holds for any $x_0\in\Sigma$.
Since $\Sigma$ is compact, there exist a positive integer $k_0 \le Cr_0^{-2}$ and points $\{x_1,\cdots, x_{k_0}\}\subset\Sigma$
such that $\Sigma$ is covered by $\{B_{\frac{r_0}2}(x_i)\}_{1\le i\le k_0}$.
Applying Lemma \ref{l:reg} on each $B_{r_0}(x_i)$, $1\le i\le k_0$, and using the gluing technique by \cite{Uh}, we can find a
gauge transform $s\in \mathscr{G}_{1,2}$ such that $\widetilde A=s^*A\in \mathscr{A}_{2,2}$ satisfies, for $1\le i\le k_0$,
\begin{equation}\label{h2-estimate4}
\big\|\widetilde A\big\|_{H^2(B_{\frac{r_0}2}(x_i))}^2\le C(r_0)\Big(\big\|F_A\big\|_{L^2(B_{r_0}(x_i))}^2
+\big\|D_A^*F_A\big\|_{L^2(B_{r_0}(x_i))}^2\Big).
\end{equation}
It is clear that (\ref{h2-estimate3}) follows by summing up (\ref{h2-estimate4}) over $1\le i\le k_0$.
Since $ds=s A-\widetilde A s$, it is easy to see that $s\in \mathscr{G}_{2,2}$. This completes the proof.
\end{proof}
Before presenting the proof of Theorem \ref{t:compact}, we recall the well-known bubble-tree convergence result of a sequence of approximate harmonic maps,
with bounded Dirichlet energies and $L^2$-tension fields, which was proved in~\cite{DT} and~\cite{LinW}.
\begin{thm}\label{t:bubble-convergence-harmonic}
Let $\{u_n\}\subset H^2(B_1,M)$ have uniformly bounded Dirichlet energies $\mathcal{E}(u_n, B_1)$ and have their tension fields
$\tau(u_n):=\Delta_g u_n+\Pi(u_n)(du_n, du_n)$ uniformly bounded in $L^2(B_1)$.
Then there exist an approximate harmonic map $u_\infty\in H^2(B_1,M)$, with tension field $\tau(u_\infty)\in L^2(B_1)$,
such that after passing to a subsequence, $u_n$ bubble tree converges to $u_\infty$.
More precisely, there exist finitely many points $\mathbf{x} = \{x_1, x_2,\cdots, x_k\}\subset B_1$ such that the following properties hold:
\begin{enumerate}
\item $u_n$ converges to $u_\infty$ strongly in $H^1_{loc}(B_1\setminus\mathbf{x})$.
\item There exist finitely many bubbles (i.e. nontrivial harmonic maps) $\omega_{ij}:\mathbb S^2\to M, 1\le i\le k, 1\le j\le l$, such that
\[\lim_{n\to \infty} \mathcal{E}(u_n, B_r) = \mathcal{E}(u_\infty, B_r) + \sum_{1\le i\le k,1\le j\le l}\mathcal{E}(\omega_{ij}),\]
where $0<r<1$ is such that ${\bf x}\subset B_r$.
\item The images of the bubbles $\displaystyle\{\omega_{ij}\}_{1\le i\le k, 1\le j\le l}$ and the limiting map $u_\infty$ are connected.
\end{enumerate}
\end{thm}
Now we are ready to prove Theorem \ref{t:compact}.
\begin{proof}[Proof of Theorem~\ref{t:compact}]
Since $A_n$ has uniformly bounded $L^2$-curvatures, Proposition \ref{l:global-gauge-fix} implies
that there exists
a sequence of gauge transformations $\{s_n\}\subset \mathscr{G}_{2,2}$ such that $\widetilde{A}_n:=s_n^* A_n\in \mathscr{A}_{2,2}$ satisfies
\begin{eqnarray}\label{h2-estimate1}
\big\|\widetilde{A}_n\big\|_{H^2(\Sigma)}&\le& C\Big(\big\|F_{A_n}\big\|_{L^2(\Sigma)}+\big\|D_{A_n}^*F_{A_n}\big\|_{L^2(\Sigma)}\Big)\nonumber\\
&\le& C\big(\|F_{A_n}\|_{L^2(\Sigma)}+\|\tau_1(A_n,\phi_n)\|_{L^2(\Sigma)}+\|D_{A_n}\phi_n\|_{L^2(\Sigma)}\big)\le C.
\end{eqnarray}
Therefore there exists $\widetilde A\in \mathscr{A}_{2,2}$ such that $\widetilde A_n\rightarrow \widetilde A$ strongly in $H^1(\Sigma)$.
Set $\widetilde{\phi}_n=s_n^*(\phi_n)$. It follows that $\widetilde\phi_n\in \S_{2,2}$. Now we claim that on each local chart
$U$ of $\Sigma$, $\widetilde\phi_n$ can be viewed as a sequence of approximate harmonic maps from $(U, g)$ to $M$, with
uniformly bounded $L^2$-tension fields $\tau(\widetilde\phi_n)$. In fact, since $\mathcal{F}\big|_U$ is a trivial bundle, we can
identify $\widetilde \phi_n$ on $U$ as a map $u_n$ from $U$ to $M$.
To write the equation of $u_n$, we isometrically embed $M$ into some Euclidean space $\mathbb R^L$ and let
$\Pi$ denote its second fundamental form. Write $D_{\widetilde{A}_n}=d+\widetilde{A}_n$ and
$D_{\widetilde{A}_n}^*=d^*+\widetilde{A}_n^*$. It is easy to see that $\widetilde{A}_n^*=-\widetilde{A}_n$. Hence we
have, on $U$,
\begin{eqnarray}\label{e:u-equation}
&&\tau(u_n):=\Delta_g u_n+\Pi(u_n)(du_n, du_n)\nonumber\\
&&=\tau_2(\widetilde{A}_n, \widetilde{\phi}_n)-(\mu(\widetilde\phi_n)-c)\nabla\mu(\widetilde\phi_n)
-d^*\widetilde{A}_n \cdot u_n-2\widetilde{A}_n\cdot du_n -\widetilde{A}_n^2\cdot u_n.
\end{eqnarray}
Here $\Delta_g$ is the Laplace operator with respect to the metric $g$ in $U$.
Applying (\ref{h2-estimate1}) to (\ref{e:u-equation}), we obtain that
\begin{eqnarray*}
&&\norm{\tau(u_n)}_{L^2(U)} \\
&&\le C\Big[\norm{\tau_1(A_n,\phi_n)}_{L^2(U)} + \norm{\tau_2(\widetilde{A}_n, \widetilde{\phi}_n)}_{L^2(U)}
+ \norm{F_{A_n}}_{L^2(U)} + \norm{D_{\widetilde A_n} \widetilde{\phi_n}}_{L^2(U)}\Big] \\
&&\le C\Big[\norm{\tau_1(A_n,\phi_n)}_{L^2(U)} + \norm{\tau_2({A}_n, {\phi}_n)}_{L^2(U)}
+ \norm{F_{A_n}}_{L^2(U)} + \norm{D_{A_n} {\phi_n}}_{L^2(U)}\Big]\\
&&\le C.
\end{eqnarray*}
This implies that $\tau(u_n)$ is uniformly bounded in $L^2(U)$. Thus we may apply Theorem~\ref{t:bubble-convergence-harmonic} on each local chart
$U$ to conclude that $\widetilde\phi_n$ {\it bubble-tree} converges to a limit section $\phi_\infty\in \S_{2,2}$, namely
the properties 2 and 3 in Theorem 6.1 hold. This completes the proof.
\end{proof}
\section{Asymptotic behavior of YMH flow at finite time singularities and time infinity}
In this section, we will discuss the asymptotic behaviors of the global weak solution $(A,\phi)$ to the YHM flow equation
(\ref{e:heat}) under the initial condition (\ref{IVP}), established in Theorem 5.1, at each possible finite singular time $T_i$, $1\le i\le L$, and at time
infinity. The main results of this section are consequences of the compactness properties on
approximate YMH fields obtained in the previous section.
\begin{thm}\label{t:finite-sing} For $(A_0,\phi_0)\in \mathscr{A}_{1,2}\times \S_{1,2}$, assume that $0<T_1<+\infty$ is the maximal time interval for the
local weak solution $(A, \phi)$ to the YMH flow equation (\ref{e:heat}), under the initial condition (\ref{IVP}), constructed
by Theorem 5.1. Then the following properties hold:
\begin{enumerate}
\item There exists a pair $(A(T),\phi(T))\in \mathscr{A}_{0,2}\times\S_{1,2}$
such that $(A(t),\phi(t))\rightarrow (A(T),\phi(T))$ in $L^2(\Sigma)$, and $\phi(t)\rightharpoonup\phi(T)$
in $H^1(\Sigma)$, as $t\rightarrow T^-$.
\item There exist a set of finitely many points $\mathbf{x} := \{x_1, \cdots, x_k\}\subset \Sigma$, with $k\le \frac{\L(0)}{\epsilon_0}$,
and a gauge transformation $s\in \mathscr{G}_{1,2}$ such that $(s^*A(t), s^*\phi(t))$ converges to $(s^*A(T), s^*\phi(T))$
in $C^\infty_{loc}(\Sigma\setminus\mathbf{x})$, as $t\rightarrow T^-$.
\item There exist finitely many nontrivial harmonic maps $\omega_{ij}: \mathbb S^2 \to M, 1\le i \le k, 1\le j\le l$ such that
\begin{equation}
\lim_{t\to T^-} \mathcal{E}(A(t), \phi(t)) = \mathcal{E}(A(T), \phi(T)) + \sum_{1\le i\le k, 1\le j\le l}\mathcal{E}(\omega_{ij}).
\end{equation}
\end{enumerate}
\end{thm}
\begin{proof} We may assume, after a suitable gauge fixing, that
$(A,\phi)\in C^\infty(\Sigma\times (0,T))$. Since $(A,\phi)\in C([0,T], L^2(\Sigma))$ and $\phi\in L^\infty([0,T], H^1(\Sigma))$, the property 1
follows easily.
For the property 2, define the energy concentration set ${\bf x}\subset \Sigma$ by
\begin{equation}\label{sing_set}
{\bf x}:=\bigcap_{r>0}\Big\{x\in\Sigma: \ \limsup_{t\uparrow T^{-}}\int_{B_r(x)} \big(|F_A|^2+|D_A\phi|^2+|\mu(\phi)-c|^2\big)(t)\,dv_g\ge \epsilon_0
\Big\}.
\end{equation}
By (\ref{e:energy-inequ}) and a simple covering argument,
we can show that ${\bf x}=\{x_1,\cdots, x_L\}\subset\Sigma$ is a set of finitely many points, with $L\le \frac{\mathcal E(0)}{\epsilon_0}$.
For any
$x_0\in \Sigma\setminus {\bf x}$, it follows from (\ref{sing_set}) that there exist $r_0>0$ and $\delta_0>0$ such that
$$
\int_{B_{r_0}(x_0)} \big(|F_A|^2+|D_A\phi|^2+|\mu(\phi)-c|^2\big)(t)\,dv_g <\epsilon_0, \ \forall\ T-\delta_0^2\le t<T.
$$
Hence, for any compact set $K\subset \Sigma\setminus {\bf x}$, by a simple covering argument
there exists $r_1=r_1(K)>0$ such that
\begin{equation}\label{uniform-small}
\sup_{x\in K} \int_{B_{r_1}(x)} \big(|F_A|^2+|D_A\phi|^2+|\mu(\phi)-c|^2\big)(t)\,dv_g <\epsilon_0, \ \forall\ T-r_1^2\le t<T.
\end{equation}
Hence by Lemma \ref{l:ep-reg} and Lemma \ref{l:higher-reg} we have that
$$
\max_{x\in K}\Big\|\big|\nabla^l_A F_A\big|+\big|\nabla_A^{l+1}\phi\big|\Big\|_{L^\infty\big(B_{\frac{r_1}2}(x)\times [T-\frac{r_1^2}4, T)\big)}
\le C(l, \epsilon_0, K), \ \forall\ l\ge 0.
$$
Similar to the argument in the proof of Theorem 5.1, this, combined with the YMH flow equation (\ref{e:heat}) and the fact $A(t)\in C^\infty(\Sigma)$ for $t=T-\frac{r_1^2}4$, implies that $(A,\phi)\in C^\infty\big(K_{\frac{r_1}2}\times [T-\frac{r_1^2}4, T]\big)$ and
\begin{equation}\label{higher-reg1}
\big\|(A,\phi)\big\|_{C^l\big(K_{\frac{r_1}2}\times [T-\frac{r_1^2}4, T]\big)}\le C(l,\epsilon_0, K), \ \forall\ l\ge 0.
\end{equation}
Here $K_{r}:=\big\{x\in\Sigma: {\rm{dist}}(x,K)< r\big\}$ denotes $r$-neighborhood of $K$, for $r>0$.
It follows from (\ref{higher-reg1}) and the property 1 that $(A(t), \phi(t))\rightarrow (A(T),\phi(T))$ in $C^\infty(K_{\frac{r_1}4})$
as $t\rightarrow T^-$. This yields property 2, since $K\subset \Sigma\setminus {\bf x}$ is arbitrary.
For the property 3, we may assume, for simplicity, that ${\bf x}=\{x_0\}$ only consists of a single point.
First we need to show\\
\noindent{\it Claim 1}. There exists a constant $m\ge \epsilon_0$ such that
\begin{equation}\label{radon1}
e(A(t),\phi(t))\,dv_g\rightharpoonup e(A(T),\phi(T))\,dv_g +m \delta_{x_0}
\end{equation}
for $t\rightarrow T^{-}$, as convergence of Radon measures on $\Sigma$. Here as in the section 1,
$$e(A(t),\phi(t))=(|F_A|^2+|\nabla_A\phi|^2+|\mu(\phi)-c|^2)(t),$$
and $\delta_{x_0}$ denotes the delta mass centered at $x_0$.
The proof is based on Lemma \ref{l:energy-inequ-local}
and can be carried out similarly to \cite{LinW} Lemma 4.1. For the convenience of readers, we sketch it here.
For ${\bf x}=\{x_0\}$ and $(A(t),\phi(t))\rightarrow (A(T),\phi(T))$ in
$H^1_{\rm{loc}}(\Sigma\setminus\{x_0\})$ by the property 2, we have that $F_{A(t)}\rightharpoonup F_{A(T)}$
in $L^2(\Sigma)$. Hence, for given two sequences $t_i^{j} \rightarrow T^{-}$, there exist
$m_j\ge \epsilon_0$, $j=1,2$, such that
\[
e(A(t_i^{j}),\phi(t_i^j))\,dv_g\rightharpoonup e(A(T),\phi(T))\,dv_g +m_j \delta_{x_0}, \ {\rm{for}}\ i\rightarrow\infty,
\]
as convergence of Radon measures, for $j=1,2$. It suffices to show $m_1=m_2$. For any $\epsilon>0$, choose
$\delta>0$ such that $\displaystyle\int_{B_{2\delta}(x_0)}e(A(T),\phi(T))\,dv_g<\epsilon$. Thus we have
\begin{eqnarray*}
&&m_1\ge \int_{B_{2\delta}(x_0)}e(A(t_i^{1}),\phi(t_i^1))\,dv_g-\epsilon\\
&&\ge \int_{B_{\delta}(x_0)}e(A(t_i^{2}),\phi(t_i^2))\,dv_g-C\frac{|t_i^1-t_i^2|}{\delta^2}\mathcal E(0)
-\Big|\int_{t_i^1}^{t_i^2}\int_\Sigma\big(|\frac{\partial A}{\partial t}|^2+|\frac{\partial \phi}{\partial t}|^2\big)\Big|-\epsilon\\
&&\ge \int_{B_{\delta}(x_0)}e(A(t_i^{2}),\phi(t_i^2))\,dv_g-2\epsilon\ge m_2-2\epsilon.
\end{eqnarray*}
This yields $m_1\ge m_2$. Similarly, we have $m_2\ge m_1.$ Hence $m_1=m_2$ and (\ref{radon1}) follows.
To simplify the presentation, assume further that $(\Sigma,g)=(\mathbb R^2, dx^2)$, $x_0=(0,0)\in\mathbb R^2$, and $T=0$.
From (\ref{radon1}), there exist $t_i\uparrow 0$ and $\lambda_i\downarrow 0$ such
that
\begin{equation}
\lim_{t_i\uparrow 0}\int_{B_{\lambda_i}}e(A,\phi)(x,t_i)\,dx=m.
\end{equation}
Define $A_i(x,t)=\lambda_i A(\lambda_i x, t_i+\lambda_i^2 t), \phi_i(x,t)=\phi(\lambda_i x, t_i+\lambda_i^2 t)$
for $(x,t)\in \mathbb R^2\times [-2, 0]$. Then $(A_i, \phi_i)$ solves the scaled version of
the YMH heat flow equation (\ref{e:heat}):
\begin{equation}\label{e:heat1}
\begin{cases}\displaystyle\frac{\partial A_i}{\partial t}=-D_{A_i}^*F_{A_i}-\lambda_i^2\phi_i^*D_{A_i}\phi_i,\\
\displaystyle\frac{\partial \phi_i}{\partial t}=-D_{A_i}^*D_{A_i}\phi_i-\lambda_i(\mu(\phi_i)-c)\cdot \nabla (\mu(\phi_i)),
\end{cases}
\end{equation}
and satisfies
\begin{equation}
\int_{-2}^0\int_{\mathbb R^2}
\Big(\lambda_i^{-2}\big|\frac{\partial A_i}{\partial t}\big|^2+\big|\frac{\partial \phi_i}{\partial t}\big|^2\Big)
=\int_{t_i-2\lambda_i^2}^{t_i}\int_{\mathbb R^2}
\Big(\big|\frac{\partial A}{\partial t}\big|^2+\big|\frac{\partial \phi}{\partial t}\big|^2\Big)\rightarrow 0,
\ {\rm{as}}\ i\rightarrow\infty.
\end{equation}
Therefore, by Fubini's theorem, there exists $\tau_i\in (-1,-\frac12)$ such that
\begin{equation}\label{good-slice}
\lim_{i\rightarrow\infty}\int_{\mathbb R^2}
\Big(\lambda_i^{-2}\big|\frac{\partial A_i}{\partial t}\big|^2+\big|\frac{\partial \phi_i}{\partial t}\big|^2\Big)(x,\tau_i)
=0.
\end{equation}
From the local energy inequality (\ref{e:energy-inequ-local}), we also have
$$
\int_{B_{R\lambda_i}}e(A,\phi)(x,t_i+\lambda_i^2\tau_i)
\ge \int_{B_{\lambda_i}}e(A,\phi)(x,t_i)-CR^{-2}\mathcal E(0)\ge m+o(1)-CR^{-2}\mathcal E(0).
$$
This and (\ref{radon1}) imply that
\begin{equation}\label{energy-loss1}
\lim_{R\rightarrow +\infty}\lim_{i\rightarrow\infty}\int_{B_{R\lambda_i}}e(A,\phi)(x,t_i+\lambda_i^2\tau_i)=m.
\end{equation}
Define $B_i(x)=A_i(x,\tau_i)$ and $\psi_i(x)=\phi_i(x,\tau_i)$ for $x\in\mathbb R^2$.
It follows from the (\ref{energy-loss1}),
(\ref{good-slice}), and (\ref{e:heat1}) that $(B_i, \psi_i)\in \mathscr{A}\times \S$ satisfies:
\begin{equation}\label{energy-loss2}
\int_{B_R}\big(\lambda_i^{-2}|F_{B_i}|^2+|D_{B_i}\psi_i|^2+\lambda_i^2|\mu(\psi_i)-c|^2\big)(x)
=\int_{B_{R\lambda_i}}e(A,\phi)(x,t_i+\lambda_i^2\tau_i)=m+o(1),
\end{equation}
\begin{equation}\label{l2-bound}
\int_{B_R}|B_i(x)|^2=\int_{B_{R\lambda_i}}|A|^2(x,t_i+\lambda_i^2\tau_i)\le C,
\end{equation}
\begin{equation}\label{small-F}
\int_{B_R}\big|F_{B_i}(x)\big|^2=\lambda_i^2\int_{B_{R\lambda_i}}|F_{A}|^2(x,t_i+\lambda_i^2\tau_i)\le m\lambda_i^2,
\end{equation}
\begin{eqnarray}\label{small-DF}
\int_{B_R}\big|D_{B_i}^*F_{B_i}(x)\big|^2&\le& C\int_{B_R}\big(|\frac{\partial A_i}{\partial t}|^2+\lambda_i^4|D_{A_i}\phi_i|^2\big)(x,t_i+\lambda_i^2\tau_i)\le C(o(1)+\lambda_i^2)\lambda_i^2,
\end{eqnarray}
and
\begin{equation}\label{approx-hm1}
\int_{B_R}\big|D_{B_i}^*D_{B_i}\psi_i(x)\big|^2
\le C\int_{B_R}\big(|\frac{\partial \phi_i}{\partial t}|^2+\lambda_i^2|D_{A_i}\phi_i|^2\big)(x,t_i+\lambda_i^2\tau_i)\le C\lambda_i^2+o(1).
\end{equation}
From (\ref{l2-bound}, (\ref{small-F}), and (\ref{small-DF}), we can apply Lemma \ref{l:reg} to conclude that on $B_R$
there exist a sequence of gauge transformations $\{s_i\}\subset \mathscr{G}_{2,2}$ such that
$\widetilde{B}_i:=s_i^*B_i$ satisfies
\begin{equation}\label{small-H1}
\big\|\widetilde{B}_i\big\|_{H^2(B_R)}\le C(R)\big(\|F_{B_i}\|_{L^2(B_R)}+\|D_{B_i}^*F_{B_i}\|_{L^2(B_R)}\big)\le C(R)\lambda_i^2.
\end{equation}
Thus we may assume that $\widetilde{B}_i$ converges to $0$ weakly in $H^2(B_R)$ and strongly in $H^1(B_R)$. While (\ref{approx-hm1})
implies that $\widetilde{\psi}_i=s_i^*\psi_i$ satisfies
\begin{equation}\label{approx-hm2}
\int_{B_R}\big|D_{\widetilde{B}_i}^*D_{\widetilde{B}_i}\widetilde{\psi}_i(x)\big|^2
=\int_{B_R}\big|D_{B_i}^*D_{B_i}\psi_i(x)\big|^2\le C\lambda_i^2+o(1).
\end{equation}
From (\ref{small-H1}) and (\ref{approx-hm2}), we can verify, similar to that of Theorem 6.1, that
$\big\{\widetilde{\psi}_i\big\}$ is a sequence of approximate harmonic maps, whose tension fields
$\tau(\widetilde{\psi}_i)$ have $\displaystyle\big\|\tau(\widetilde{\psi}_i)\big\|_{L^2(B_R)}\le C\lambda_i$.
Now we need
\smallskip
\noindent{\it Claim 2}. After passing to a subsequence, $\widetilde{\psi}_i$ converges to a constant map in $L^2(B_R)$.
To see this, observe that
\begin{equation}\label{l2-vanish}
\int_{B_R}\big|\widetilde\psi_i(x)-s_i^*\phi_i(x, -\lambda_i^{-2}t_i)\big|^2=\int_{B_R}\big|\phi_i(x, \tau_i)-\phi_i(x, -\lambda_i^{-2}t_i)\big|^2
\le 4\int_{-2}^2\int_{B_R}\big|\frac{\partial\phi_i}{\partial t}\big|^2\rightarrow 0,
\end{equation}
\begin{equation}\label{small-h1norm1}
\int_{B_R}\big|D_{s_i^*A_i} s_i^*\phi_i\big|^2(x,-\lambda_i^{-2}t_i)
=\int_{B_R}\big|D_{A_i}\phi_i|^2(x,-\lambda_i^{-2}t_i)=\int_{B_{R\lambda_i}}\big|D_{A}\phi\big|^2(x,0)\rightarrow 0,
\end{equation}
and
\begin{eqnarray}&&\int_{B_R}\big|s_i^*A_i(x, -\lambda_i^{-2}t_i)\big|^2
\le 2\int_{B_R}\big|s_i^*A_i(x, -\lambda_i^{-2}t_i)-\widetilde{B}_i(x)\big|^2+2\int_{B_R}\big|\widetilde{B}_i(x)\big|^2\nonumber\\
&&\le 2\int_{B_R}\big|s_i^*A_i(x, -\lambda_i^{-2}t_i)-s_i^*A_i(x,\tau_i)\big|^2+2\int_{B_R}\big|\widetilde{B}_i(x)\big|^2\nonumber\\
&&=2\int_{B_R}\big|A_i(x, -\lambda_i^{-2}t_i)-A_i(x,\tau_i)\big|^2+2\int_{B_R}\big|\widetilde{B}_i(x)\big|^2\nonumber\\
&&\le 4\int_{-2}^2\int_{B_R}\big|\frac{\partial A_i}{\partial t}\big|^2+2\int_{B_R}\big|\widetilde{B}_i(x)\big|^2\rightarrow 0.
\label{small-h1norm2}
\end{eqnarray}
Claim 2 now follows from (\ref{l2-vanish}), (\ref{small-h1norm1}), and (\ref{small-h1norm2}).
For each $R>0$, we now can apply Theorem 6.4 to $\widetilde{\psi}_i$ on $B_R$ to conclude that there exist $N_R$ bubbles
$\big\{\omega_{l,R}\big\}_{l=1}^{N_R}$, with $N_R\le \frac{m}{\alpha(M)}$, such that
\begin{equation}\label{energy-id1}
\lim_{i\rightarrow\infty}\int_{B_R} \big|D_{\widetilde{B}_i}\widetilde{\psi}_i\big|^2
=\sum_{l=1}^{N_R}\mathcal E(\omega_{l,R}).
\end{equation}
We may assume that there exists an integer $d\in \big[1, \frac{m}{\alpha(M)}\big]$ such that for $N_R=d$ for $R>>1$.
Note that for $l=1,\cdots, d$, $\big\{\omega_{l,R}\big\}_{R>1}$ is a sequence of harmonic maps from $\mathbb S^2$ to $M$
whose energies are uniformly bounded. Hence we can apply the energy identity result for harmonic maps (cf. \cite{P1})
to conclude that for $l=1,\cdots, d$, there exist $N_l$ bubbles $\big\{\omega_{l,j}\big\}_{j=1}^{N_l}$,
with $N_l\le \frac{m}{\alpha(M)}$, such that
\begin{equation}\label{energy-id2}
\lim_{R\rightarrow\infty}\mathcal E(\omega_{l,R})=\sum_{j=1}^{N_l}\mathcal E(\omega_{l,j}).
\end{equation}
Hence we have
\begin{equation}\label{energy-id3}
\lim_{R\rightarrow\infty}\lim_{i\rightarrow\infty}\int_{B_R} \big|D_{\widetilde{B}_i}\widetilde{\psi}_i\big|^2
= \sum_{l=1}^d\sum_{j=1}^{N_l}\mathcal E(\omega_{l,j}).
\end{equation}
It is readily seen that the proof of property 3 will be complete if we can show
\begin{equation}\label{noconcentration-F}
\lim_{R\rightarrow\infty}\lim_{i\rightarrow\infty}\int_{B_R} \lambda_i^{-2}\big|F_{B_i}\big|^2=0.
\end{equation}
Set $\widehat{t}_i=t_i+\lambda_i^2\tau_i$. Apply Uhlenbeck's gauge fixing for $A(\cdot, \widehat{t}_i)$ on $B_{\delta_0}$ for a small
$\delta_0>0$ (here $\delta_0>0$ is chosen so that for $1<p<2$, $\displaystyle\int_{B_{\delta_0}}|F_{A}|^p(x,\widehat{t}_i)$ is small,
which is possible since $\displaystyle\int_{\mathbb R^2}|F_A|^2(x,\widehat{t}_i)\le \mathcal{E}(0)$), we obtain a sequence of gauge transformations
$\alpha_i\in\mathscr{G}_{2,2}$ such that $\widehat{A}_i(x)=\alpha_i^*A(x,\widehat{t}_i)$ satisfies
\begin{equation}\label{gauge-fix1}
d^*\widehat{A}_i=0 \ {\rm{in}}\ B_{\delta_0}, \ \big\|\widehat{A}_i\big\|_{H^1(B_{\delta_0})}\le C\big\|F_{A}(\widehat{t}_i)\big\|_{L^2(B_{\delta_0})}.
\end{equation}
Set $\widehat{B}_i(x)=\lambda_i\widehat{A}_i(\lambda_i x,\widehat{t}_i)$ for $x\in B_{\delta_0\lambda_i^{-1}}$. Then we
have
\begin{equation}\label{curv-trans}
\int_{B_R} \lambda_i^{-2}\big|F_{B_i}\big|^2=\int_{B_R} \lambda_i^{-2}\big|F_{\widehat{B}_i}\big|^2
=\int_{B_R}\big|dC_i+\lambda_i [C_i,C_i]\big|^2,
\end{equation}
where $C_i(x)=\widehat{A}_i(\lambda_i x,\widehat{t}_i)$ for $x\in B_{\delta_0\lambda_i^{-1}}$.
From (\ref{e:heat1}) and (\ref{gauge-fix1}), we see that $\widehat{A}_i$ solves the elliptic equation:
\begin{equation}\label{e:heat2}
\Delta_0\widehat{A}_i(x)
=-\big(\widehat{A}_i\#d\widehat{A}_i+\widehat{A}_i\#\widehat{A}_i\#\widehat{A}_i
-\alpha_i^*\phi D_{\widehat {A}_i}\alpha_i^*\phi\big)(x,\widehat{t}_i)
-\frac{\partial (\alpha_i^*A)}{\partial t}(x,\widehat{t}_i), \ {\rm{in}}\ B_{\delta_0}.
\end{equation}
Here $\Delta_0$ denotes the standard Laplace operator on $\mathbb R^2$.
It follows from (\ref{e:heat2}) that $C_i$ solves
\begin{eqnarray}\label{e:heat3}
\Delta_0C_i(x)
&=&-\lambda_i^2\big(\widehat{A}_i\#d\widehat{A}_i+\widehat{A}_i\#\widehat{A}_i\#\widehat{A}_i
-\alpha_i^*\phi D_{\widehat {A}_i}\alpha_i^*\phi\big)(\lambda_i x,\widehat{t}_i)
-\lambda_i^2\frac{\partial (\alpha_i^*A)}{\partial t}(\lambda_i x,\widehat{t}_i)\nonumber\\
&=& I_1^i+I_2^i+I_3^i+I_4^i,
\end{eqnarray}
in $B_{\delta_0\lambda_i^{-1}}$.
It is easy to see
\begin{equation}\label{H1-bound}
\int_{B_{\delta_0\lambda_i^{-1}}}|\nabla_0 C_i|^2=\int_{B_{\delta_0}}|\nabla_0 \widehat{A}_i|^2(\widehat{t}_i)
\le C.
\end{equation}
Using (\ref{gauge-fix1}) and (\ref{good-slice}), we can estimate
$$\big\|I_1^i\big\|_{L^{\frac43}(B_{\delta_0\lambda_i^{-1}})}
\le C\lambda_i^{\frac12}\big\|\widehat{A}_i(\widehat{t}_i)\big\|_{L^4(B_{\delta_0})}\big\|d\widehat{A}_i(\widehat{t}_i)\big\|_{L^2(B_{\delta_0})}
\le C\lambda_i^{\frac12}\big\|\widehat{A}_i(\widehat{t}_i)\big\|_{H^1(B_{\delta_0})}^2\le C\lambda_i^{\frac12},$$
$$
\big\|I_2^i\big\|_{L^{\frac43}(B_{\delta_0\lambda_i^{-1}})}
\le C\lambda_i^{\frac12}\big\|\widehat{A}_i(\widehat{t}_i)\big\|_{L^4(B_{\delta_0})}^3\le C\lambda_i^{\frac12},$$
$$\big\|I_3^i\big\|_{L^{2}(B_{\delta_0\lambda_i^{-1}})}
\le C\lambda_i\big\|D_{\widehat{A}_i}\alpha_i^*\phi(\widehat{t}_i)\big\|_{L^2(B_{\delta_0})}
=C\lambda_i\big\|D_A\phi(\widehat{t}_i)\big\|_{L^2(B_{\delta_0})}
\le C\lambda_i,$$
and
$$
\big\|I_4^i\big\|_{L^{2}(B_{\delta_0\lambda_i^{-1}})}^2
=\lambda_i^2\int_{B_{\delta_0}}\big|\frac{\partial A}{\partial t}(\widehat{t}_i)\big|^2
=\lambda_i^{-2}\int_{B_{\delta_0\lambda_i^{-1}}}\big|\frac{\partial A_i}{\partial t}(\widehat{t}_i)\big|^2=o(1)\rightarrow 0.
$$
Applying $W^{2,\frac43}$-estimate to the equation (\ref{e:heat3}),
we conclude that $C_i\in W^{2,\frac43}(B_{\frac{\delta_0}{2\lambda_i}})$,
and
\begin{eqnarray}
\big\|\nabla_0 C_i\big\|_{W^{1,\frac43}(K)}&\le&
C(K)\Big[\sum_{j=1}^4\|I_j^i\|_{L^{\frac43}(B_{\delta_0\lambda_i^{-1}})}+\|\nabla_0 C_i\|_{L^2(B_{\delta_0\lambda_i^{-1}})}\Big]
\nonumber\\
&\le& C(K)\big[1+o(1)+\lambda_i^\frac12\big]
\end{eqnarray}
for any compact subset $K\subset B_{\frac{\delta_0}{2\lambda_i}}$.
Hence we may assume, after passing to a subsequence, that
$C_i\rightarrow C$ in $H^1_{\rm{loc}}(\mathbb R^2)$. From (\ref{e:heat3}) and (\ref{H1-bound}), we see that
$$\Delta_0 C=0 \ \ {\rm{in}}\ \ \mathbb R^2, \ \int_{\mathbb R^2}|\nabla_0 C|^2<+\infty.$$
Thus $C$ must be a constant. This implies
that for any $R>0$,
\begin{equation}\label{H1-convergence}
\lim_{i\rightarrow \infty}\int_{B_R}|\nabla_0 C_i|^2=0.
\end{equation}
Observe that by (\ref{gauge-fix1})
$$\int_{B_{\delta_0\lambda_i^{-1}}}|\lambda_i^{\frac12}C_i|^4
=\int_{B_{\delta_0}}|\widehat{A}_i|^4\le C,
\ {\rm{and}}\ \int_{B_{\delta_0\lambda_i^{-1}}}\big|\nabla_0 (\lambda^{\frac12}C_i)\big|^2
=\lambda_i\int_{B_{\delta_0}}\big|\nabla_0\widehat{A}_i\big|^2\rightarrow 0.
$$
Thus we may assume that there exists a constant $\widetilde C$
such that $\lambda_i^\frac12 C_i\rightarrow \widetilde C$ in $L^4_{\rm{loc}}(\mathbb R^2)$.
Since
$$\int_{B_L}|\widetilde C|^4=\lim_{i\rightarrow \infty}\int_{B_L}|\lambda_i^{\frac12}C_i|^4\le C$$
holds for any $L>0$, we must have $\widetilde C=0$.
Hence for any $R>0$, we have
\begin{equation}\label{L4-convergence}
\lim_{i\rightarrow \infty}\int_{B_R}\big|\lambda_i[C_i, C_i]\big|^2=0.
\end{equation}
It is clear that (\ref{noconcentration-F}) follows from (\ref{curv-trans}), (\ref{H1-convergence}) and
(\ref{L4-convergence}). The proof is now complete.
\end{proof}
For the asymptotic behavior of the global weak solution constructed by Theorem 5.1 at time infinity, we have
\begin{thm}\label{t:infinity-sing} For $(A_0,\phi_0)\in \mathscr{A}_{1,2}\times \S_{1,2}$, assume that $(A, \phi)$ is the global
weak solution to the YMH flow equation (\ref{e:heat}), under the initial condition (\ref{IVP}), constructed
by Theorem 5.1. Then the following properties hold:
\begin{enumerate}
\item There exist $t_i\uparrow +\infty$, a sequence of gauge transformations $\{s_i\}\subset\mathscr{G}_{2,2}$,
a set of finitely many points $\mathbf{x} := \{x_1, \cdots, x_k\}\subset \Sigma$, with $k\le \frac{\mathcal{E}(0)}{\epsilon_0}$,
and a Yang-Mills-Higgs field $(A_\infty,\phi_\infty)\in\mathscr{A}\times\S$ such that
$s_i^*A(t_i)\rightarrow A_\infty$ in $H^1(\Sigma)$, $s_i^*\phi(t_i)\rightharpoonup \phi_\infty$ in $H^1(\Sigma)$,
and $s_i^*\phi(t_i)\rightarrow \phi_\infty$
in $H^1_{loc}(\Sigma\setminus\mathbf{x})$, as $i\rightarrow \infty$.
\item There exist finitely many nontrivial harmonic maps $\omega_{ij}: \mathbb S^2 \to M, 1\le i \le k, 1\le j\le l$ such that
\begin{equation}\label{energy-identity-infinity}
\lim_{i\to \infty} \mathcal{E}(A(t_i), \phi(t_i)) = \mathcal{E}(A_\infty, \phi_\infty) + \sum_{1\le i\le k, 1\le j\le l}\mathcal{E}(\omega_{ij}).
\end{equation}
\item The images of $\{\omega_{ij}\}_{1\le i\le k, 1\le j\le l}$ and that of $\phi_\infty$ are connected.
\end{enumerate}
\end{thm}
\begin{proof} For simplicity, we may assume that $(A,\phi)\in C^\infty((0,+\infty), \mathscr{A}\times\S)$. From (\ref{e:energy-inequ}),
there exist $t_i\uparrow +\infty$ and $0<L_0<+\infty$ such that
\begin{equation}\label{energy-conv}
\lim_{i\rightarrow\infty}\mathcal{E}(A(t_i),\phi(t_i))=L_0,
\end{equation}
and
\begin{equation}\label{good-slice1}
\lim_{i\rightarrow\infty}\int_\Sigma \big(|\frac{\partial A}{\partial t}|^2+|\frac{\partial\phi}{\partial t}|^2\big)(x,t_i)=0.
\end{equation}
Set $(A_i(x),\phi_i(x)):=(A(x,t_i),\phi(x,t_i))$ for $x\in\Sigma$. Then $(A_i,\phi_i)\in\mathscr{A}\times\S$ is a sequence
of approximate YMH fields, with uniformly bounded YMH energies, such that its
tension fields $\tau_1(A_i,\phi_i)=\frac{\partial A}{\partial t}(t_i)$ and $\tau_2(A_i,\phi_i)
=\frac{\partial\phi}{\partial t}(t_i)$. Hence it follows from (\ref{good-slice1}) that
\begin{equation}\label{vanish-tensions}
\lim_{i\rightarrow\infty}\Big[\big\|\tau_1(A_i,\phi_i)\big\|_{L^2(\Sigma)}+\big\|\tau_2(A_i,\phi_i)\big\|_{L^2(\Sigma)}\Big]=0.
\end{equation}
Applying Lemma 6.2 to $(A_i,\phi_i)$, we conclude that there exist gauge transformations $\{s_i\}\subset \mathscr{G}_{2,2}$,
a YMH field $(A_\infty,\phi_\infty)\in \mathscr{A}_{2,2}\times \S_{2,2}$ such that
$$s_i^*A_i\rightarrow A_\infty \ {\rm{in}}\ H^1(\Sigma), \ s_i^*\phi_i\rightharpoonup \phi_\infty
\ {\rm{in}}\ H^1(\Sigma).$$
The remaining parts of Theorem \ref{t:infinity-sing}, except the smoothness of $(A_\infty,\phi_\infty)$, follow directly from Theorem \ref{t:compact}.
While smoothness of YMH fields in $\mathscr{A}_{2,2}\times \S_{2,2}$, after suitable gauge transformations,
can be done by the bootstrap arguments
(see, e.g., \cite{So} Theorem 3.3).
\end{proof}
\section{Existence of global weak solutions of the YMH flow}
In this section, we indicate how to extend the local weak solution $(A,\phi)$ to the YMH flow equation (\ref{e:heat}) and
(\ref{IVP}) beyond the first singular time to a global weak solution eventually.
Note that, different form higher dimension cases, an isolated singular point of a $H^1$ connection $A$ on a Riemann
surface, in general, may not be removed even if the curvature $F$ vanishes. This can be illustrated by the following example.
Let $\mathbb{D}^* = \{x\in \mathbb{R}^2|\ 0<|x|\le 1\}$ be the punctured disk, and $P$ is a principal $U(1)$-bundle over $\mathbb{D}^*$. Let $a\in \mathbb{R}$ be a constant. Then $A=ad\theta$ is a well defined smooth connection on $P$. Obviously, the curvature $F$ is identically zero on the disk. However, $A$ can be extended to the whole disk by a gauge transformation if and only if the holonomy $\text{Hol(A)} = e^{2\pi a}$ equals to identity.
The next theorem shows that the limit connection $A(T)$ obtained in Theorem~\ref{t:local}
belongs to $H^1$ after a gauge transformation, which is needed for an extension of weak solutions
to (\ref{e:heat}) beyond $T$.
\begin{thm}\label{t:limit-pair} Assume $0<T_1<+\infty$ is the first singular time for the local weak solution
$(A, \phi)$ of the YMH flow equation~(\ref{e:heat}) and (\ref{IVP}), constructed by Theorem 5.1.
Then there exists a time-independent gauge transformation
$s\in \mathscr{A}_{2,2}$ such that $\big(\tilde{A}, \tilde{\phi}\big) := \big(s^*A(T_1), s^*\phi(T_1)\big) \in \mathscr{A}_{1,2}\times \S_{1,2}$.
\end{thm}
\begin{proof}
Let $\{t_k\}_{k=1}^\infty \subset [0, T_1)$ be a sequence approaching $T_1$. By Theorem \ref{t:local}, we have
\begin{equation}
(A_k, \phi_k) := (A(t_k), \phi(t_k))\to (A(T_1), \phi(T_1)) \text{~in~} L^2(\Sigma).
\end{equation}
Since $(A_k, \phi_k)$ is smooth and the curvatures $F_{A_k}$ of connection $A_k$ satisfies $\norm{F_{A_k}}_{L^2(\Sigma)}$
is uniformly bounded. Then by Uhlenbeck's compactness theorem(Theorem~1.5 in~\cite{Uh}), there exists a subsequence,
still denoted by $A_k$, and a sequence of gauge transformations $\{s_k\}\subset \mathscr{G}_{2,2}$ such that $\widetilde{A}_k := s_k^*A_k$
converges weakly in $H^1(\Sigma)$ to a limit connection $\widetilde{A}\in \mathscr{A}_{1,2}$. In particular, $\widetilde{A}_k$ is bounded in $H^1(\Sigma)$.
Since
\[ \widetilde{A}_k = s_k^{-1}ds_k + s_k^{-1}A_ks_k, \]
it follows that
\[ \norm{ds_k}_{L^2(\Sigma)} \le \norm{A_k}_{L^2(\Sigma)} + \norm{\widetilde{A}_k}_{L^2(\Sigma)}\]
is bounded. Hence $s_k$ weakly converges in $H^1(\Sigma)$ to a limit gauge transformation $s\in \mathscr{G}_{1,2}$. Then one easily check
that $\widetilde{A}_k$ converges to $s^*A(T_1)$ weakly in $L^1(\Sigma)$. By the uniqueness of limit, we find that
$s^*A(T_1) = \widetilde{A}\in \mathscr{G}_{1,2}$. It is easy to check that $s^*\phi(T_1)\in \S_{1,2}$,
since $s, \phi(T_1)\in H^1(\Sigma)\cap L^\infty(\Sigma)$. This completes the proof.
\end{proof}
\begin{rem}{\rm
It was shown by Struwe~\cite{St} Lemma 3.6 and Schlatter~\cite{Sc2} Lemma 2.4
that for Yang-Mills flow, the connection can be extended to $C^0(H^1)$ as long as the curvature dose not concentrate. This is obtained by considering the evolution equation for the curvature $F$, which turns out to be a well-behaved parabolic equation. One may suspect that for the YMH flow in dimension two, the connection $A$ may belong to $C^0(H^1)$, since the curvature does not concentrate in the subcritical dimension two. However, this may not be true since the section $\phi$ may concentrate and blow up at the singular time $T_1$ so that the equation for the curvature is not well-defined at $T_1$. Thus we have to invoke Uhlenbeck's theorem to ensure the connection $A(T_1)\in \mathscr{A}_{1,2}$ only after a suitable gauge transformation.}
\end{rem}
Applying Theorem~\ref{t:limit-pair}, we can prove the following theorem on the existence of global weak solutions
to (\ref{e:heat}).
\begin{thm}\label{t:global}
Let $(A_0, \phi_0)\in \mathscr{A}_{1,2}\times \S_{1,2}$. There exist a global weak solution $(A, \phi)$ to the YMH flow~(\ref{e:heat})
and (\ref{IVP}) such that
\begin{itemize}
\item[i)] the energy inequality $\mathcal{E}(A(t),\phi(t))\le\mathcal{E}(A_0,\phi_0)$ holds for all $0\le t<+\infty$, and
\[ A\in C^0([0, \infty), \mathscr{A}_{0,2}); \phi\in C^0([0, \infty), \S_{0,2}); F_A\in L^\infty([0, \infty), L^2); D_A\phi\in L^\infty([0,\infty), L^2).\]
\item[ii)] There exist a positive integer $L\le [\frac{\mathcal{E}(0)}{\alpha(M)}]$, and gauge transformations
$\{s_i\}_{i=1}^L\subset\mathscr{G}_{2,2}$, and $0=T_0<T_1<T_2<\cdots<T_L<+\infty$ such that for $1\le i\le L$,
$(s_i^*A,s_i^*\phi)\in C^\infty\big(\Sigma\times (T_{i-1}, T_i]\setminus \{(x_1^i,T_i),\cdots, (x_{j(i)}^i,T_i)\}\big)$ for some $j(i)\le
[\frac{\mathcal{E}(0)}{\alpha(M)}]$.
\end{itemize}
\end{thm}
\begin{proof}
i) By Theorem~\ref{t:local}, there exists a local weak solution $(A(t), \phi(t))$ to the YMH flow equation~(\ref{e:heat}) on the time interval $[0, T_1)$ with initial data $(A_0, \phi_0)$. Assume $0<T_1<+\infty$ is the maximal time interval.
Then $T_1$ can be characterized by Theorem~\ref{t:local}. Let $(A_1, \phi_1)$ be the limit of
$(A(t),\phi(t))$ in $L^2(\Sigma)$ as $t\to T_1$. Then by Theorem~\ref{t:limit-pair}, we may find a gauge transformation
$s_1\in \mathscr{G}_{1,2}$ such that $$(\widetilde{A}_1, \widetilde{\phi}_1) := s_1^*(A_1, \phi_1) \in \mathscr{A}_{1,2}\times \S_{1,2}.$$
Now at time $T_1$, we set $(\widetilde{A}_1, \widetilde{\phi}_1)$ as the initial data and apply Theorem \ref{t:local} again
to obtain a local weak solution $(\widetilde{A}(t), \widetilde{\phi}(t))$ of the YMH flow~(\ref{e:heat}) on some time interval $[T_1, T_2)$.
Since the YMH heat flow equation is invariant under time-independent gauge transformations,
$(A_1(t), \phi_1(t)) := \big(s_1^{-1}\big)^*(\widetilde{A}(t), \widetilde{\phi}(t))$, $T_1\le t<T_2$, is still a solution to the YMH flow~(\ref{e:heat}).
Since
$$(A_1, \phi_1) = (A(T_1), \phi(T_1)) = (A_1(T_1), \phi_1(T_1)),$$
we can patch $(A(t), \phi(t))$, $0\le t\le T_1$, and $(A_1(t), \phi_1(t))$, $T_1\le t<T_2$, at $T_1$ to form a new solution,
still denoted by $(A(t), \phi(t))$, to the YMH flow (\ref{e:heat}) on the time interval $[0, T_2)$ such that
\[ A(t) \in C^0([0, T_2), L^2); F(t) \in L^\infty([0, T_2), L^2); \phi(t)\in C^0([0, T_2), L^2), D_A\phi\in L^\infty([0,T_2], L^2). \]
$T_2$ can again be characterized by Theorem~\ref{t:local}. Now we can repeat the above process inductively.
Since at each singular time $T_i$, there is an energy loss of amount at least $\alpha(M)$ by Theorem \ref{t:local},
the process stops after at most $L$ steps, for some $L\le [\frac{\mathcal{E}(0)}{\alpha(M)}]$.
Therefore, we obtain a global weak solution of the YMH flow (\ref{e:heat}), which satisfies the
properties stated in Theorem \ref{t:global}.
\end{proof}
\bigskip
\noindent{\bf Acknowledgements}. Part of this work was initiated when the first author was visiting University of Kentucky in 2013, which was supported by the AMS Fan Fund China Exchange program. The first author is partially supported by NSFC No.11201387 and Natural Science Foundation of Fujian Province of China No. 2014J01023. The second author is partially supported by NSF 1522869.
|
\section*{References}}
\usepackage{fancyhdr}
\fancyhf{}
\renewcommand{\headrulewidth}{0pt}
\rfoot{\thepage}
\pagestyle{fancy}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\newenvironment{proof}[1][Proof]{\begin{trivlist}
\item[\hskip \labelsep {\bfseries #1}]}{\end{trivlist}}
\newenvironment{definition}[1][Definition]{\begin{trivlist}
\item[\hskip \labelsep {\bfseries #1}]}{\end{trivlist}}
\newenvironment{example}[1][Example]{\begin{trivlist}
\item[\hskip \labelsep {\bfseries #1}]}{\end{trivlist}}
\newenvironment{remark}[1][Remark]{\begin{trivlist}
\item[\hskip \labelsep {\bfseries #1}]}{\end{trivlist}}
\newenvironment{prop}[1][Proposition]{\begin{trivlist}
\item[\hskip \labelsep {\bfseries #1}]}{\end{trivlist}}
\newcommand{\specialcell}[2][c]{%
\begin{tabular}[#1]{@{}c@{}}#2\end{tabular}}
\begin{document}
\begin{frontmatter}
\title{Unsupervised Feature Selection Based on the Morisita Estimator of Intrinsic Dimension}
\author{Jean GOLAY and Mikhail KANEVSKI}
\address{Institute of Earth Surface Dynamics, Faculty of Geosciences and Environment, University of Lausanne, 1015 Lausanne, Switzerland. Email: jean.golay@unil.ch.}
\begin{abstract}
This paper deals with a new filter algorithm for selecting the smallest subset of features carrying all the information content of a data set (i.e. for removing redundant features). It is an advanced version of the fractal dimension reduction technique, and it relies on the recently introduced Morisita estimator of Intrinsic Dimension (ID). Here, the ID is used to quantify dependencies between subsets of features, which allows the effective processing of highly non-linear data. The proposed algorithm is successfully tested on simulated and real world case studies. Different levels of sample size and noise are examined along with the variability of the results. In addition, a comprehensive procedure based on random forests shows that the data dimensionality is significantly reduced by the algorithm without loss of relevant information. And finally, comparisons with benchmark feature selection techniques demonstrate the promising performance of this new filter.
\end{abstract}
\begin{keyword}
Unsupervised feature selection \sep Morisita index \sep Intrinsic dimension \sep Redundancy minimization \sep Data mining
\end{keyword}
\end{frontmatter}
\section{Introduction}
Recent breakthroughs in technology have radically improved our ability to collect and store data. Consequently, more and more variables (or features \footnote{In this paper, the term ``feature'' is used as a synonym for ``variable''.}) are available to perform data mining tasks, but in general, a lot of them are redundant (i.e. they do not carry additional information beyond that subsumed by other features), or partially redundant, and contribute to the emergence of four major issues: (1) the reduction in the accuracy of learning algorithms because of the curse of dimensionality \cite{Bell61}, (2) the computer performance limitations related to memory and processing speed, (3) the difficulty in visualizing large amounts of complex and high-dimensional data and (4) the interpretability of the results which becomes less tractable making it difficult to gain an insight into the mechanisms that generated the data.
Due mainly to these redundant and partially redundant features, data points do not occupy the full $E$-dimensional space $\mathbb{R}^E$ ($E$ is the number of features in a data set) in which they are embedded. Instead, they are often regarded as residing on a lower $M$-dimensional manifold where $M (\leqslant E)$ is the \textit{Intrinsic Dimension} (ID) of data \cite{Cama16}. Dimensionality Reduction (DR) methods \cite{LeeVer07,Liu15} can help remove redundant information by trying to map the original data space coordinates to an intrinsic coordinate system of dimensionality $M$. Depending on the assumptions made about the shape of the manifold, the mapping can be either linear (e.g. PCA \cite{Pear01}) or non-linear (e.g. kernel-PCA \cite{Schol99}), and a great advantage of the DR approach is its potential to capture complex dependencies. On the other hand, DR often leads to a deterioration in the physical interpretability of the data and to difficulties in the understanding of subsequent results. A possible solution to these drawbacks is the implementation of feature selection methods.
The goal of \textit{feature selection} \cite{Guy06,Liu07,Koh97,Fen14,Pen05} is to select the smallest subset of original features which maintains some meaningful characteristics with respect to a chosen criterion. According to the possible use of output information (e.g. class labels), feature selection methods can be broadly classified as either supervised or unsupervised. Advanced supervised methods aim to select features which are both relevant to the prediction (i.e. classification or regression) of some output information and related as little as possible to one another (i.e. select relevant and non-redundant features). In contrast, unsupervised methods do not make use of any a priori knowledge regarding an output, and they can be further divided into two categories: Cluster Recognition (CR) and Redundancy Minimization (RM).
The CR methods aim to find the smallest subset of features that uncovers the most ``interesting'' and ``natural'' groupings (i.e. clusters) of data points \cite{Dy04,Dy07,Ale13,Chen15}. They rely on criteria of relevance that do not involve any output information, and they can be categorized into filters and wrappers \cite{Joh94}. The former (e.g. the Laplacian score method \cite{He06}, SPEC \cite{Zha07,Zha11} and MCFS \cite{Cai10}) do not incorporate the clustering algorithm that will ultimately be applied, while the latter do (e.g. methods introduced in \cite{Dy04,Dy03} or reviewed in \cite{Dy07}). In contrast, the RM methods are often not restricted to clustering problems, and they can be used as preprocessing tools in a wide variety of data mining approaches. Their goal is to select the smallest subset of features in such a way that all the information content of a data set is preserved as much as possible. In other words, they aim to eliminate all the redundant information by selecting the most informative features (i.e. the non-redundant features). To achieve this goal, the RM methods often use criteria based on PCA loading values \cite{Kim11} or on measures of feature dependency, such as the maximal information compression index \cite{Mit02}, mutual information \cite{Fen16,Mart07} and fractal-based measures of ID \cite{Trai00,Sousa07,Mo12}. More recently, Wang et al. \cite{Wan15} proposed a criterion that minimizes the reconstruction error of a linear projection of the original features, while ensuring low redundancy. Further, the RM methods can be thought of as filters, and like many other methods of feature selection, they can rely on greedy (e.g. Sequential Forward Selection (SFS) \cite{Whit71} and Sequential Backward Elimination (SBE) \cite{Ree99,Har98}) or randomized (e.g. simulated annealing \cite{Kirk83}) search strategies if they consider multivariate interactions and aim to find the best subset among the $2^E-1$ combinations of features. Lastly, methods combining the CR and the RM approaches have also been developed. Many of them use a graph Laplacian matrix to preserve the data structure and involve a low redundancy constraint or a more advanced regularization term \cite{Han15}.
More specifically, the use of ID for unsupervised feature selection was introduced by Traina et al. \cite{Trai00,Trai10}. They extended the concept of ID to fractal dimensions and proposed the Fractal Dimension Reduction (FDR) algorithm. FDR is a filter algorithm for non-linear RM that follows a SBE search strategy. It aims to eliminate the features which do not contribute to increasing the value of the data ID (i.e. the ID of the studied data set), and it relies on R\'{e}nyi's dimension of order $2$ \cite{Hent83}, $D_2$, for the ID estimation. An extension to FDR was proposed by De Sousa et al. \cite{Sousa07} to identify subsets of correlated attributes in databases according to user-defined levels of correlation. Finally, Mo and Huang \cite{Mo12} modified FDR by replacing $D_2$ with the correlation dimension $df_{cor}$ \cite{Grass833}.
The present paper deals with a novel ID-based filter algorithm for RM. It relies on the recently introduced Morisita estimator of ID, $M_m$, which was shown to be more effective than $D_2$ and $df_{cor}$ in situations where the data points were sparsely distributed \cite{Go15}. Besides, the proposed algorithm follows a SFS search strategy; it can process large and highly non-linear data, and its implementation is straightforward in \textsf{R} and \textsc{Matlab}. Another advantage is that the number of features to be selected can be determined directly from the results. And it is also worth mentioning that $M_m$ was already used successfully to perform supervised feature selection in regression problems \cite{Go15Esann}.
The remainder of this paper is organized as follows. Section \ref{Mindex} presents the Morisita estimator of ID, and Section \ref{ID_cor} explains the relationship between ID and data redundancy. In Section \ref{ID_RM}, the proposed algorithm for RM is introduced, and Section \ref{exp_res} is devoted to numerical experiments conducted on simulated data and on real world case studies from the UCI machine learning repository. The quality of the results is assessed using a comprehensive methodology based on random forests \cite{Bre01}, and comparisons with benchmark feature selection techniques (including FDR) are also discussed. Finally, conclusions are drawn in the last section with a special emphasis on potentialities and future challenges.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{Mindid.pdf}
\caption{The Morisita estimator of ID, $M_2$, applied to two features of the \textsf{R} data set ``Trees'' \cite{Rcran}. The red square in the left-hand panel indicates the ln value of the multipoint Morisita index computed with the grid displayed on the right.}\label{Fig_Mindid}
\end{figure}
\section{The Morisita Estimator of Intrinsic Dimension}\label{Mindex}
\subsection{Overview}
The Morisita estimator of ID \cite{Go15}, $M_m$, is derived from the multipoint Morisita index $I_{m,\delta}$ \cite{Go14,Hul90,Mori59}. $I_{m,\delta}$ is computed by means of an $E$-dimensional grid of $Q$ cells (or quadrats) of diagonal size $\delta$ superimposed over the data points (see Figure \ref{Fig_Mindid}). It measures how many times more likely it is that $m$ ($m\geq 2$) points selected at random will be from the same cell than it would be if the $N$ points of the studied data set were distributed according to a random distribution generated from a Poisson process (i.e. complete spatial randomness). $I_{m,\delta}$ is given by the following formula: \begin{equation}\label{Eq_mindex}
I_{m,\delta}=Q^{m-1}\frac{\sum_{i=1}^Q n_i(n_i-1)(n_i-2) \dotsm (n_i-m+1)}{N(N-1)(N-2) \dotsm (N-m+1)}
\end{equation}where $n_i$ is the number of data points in the $i^{th}$ cell. In general, $m$ is set to $2$, and the computation of the index is iterated for $R$ different values of $\delta$. These values must be chosen by the user and determine the scales at which the phenomenon will be characterized. Within the range of these scale values, if the data set follows a fractal behaviour (i.e. is self-similar), the functional relationship between $\log{(I_{m,\delta})}$ and $\log{(1/\delta)}$ is linear, its slope, $S_m$, is the Morisita slope, and $M_m$ can be written as:\begin{equation}\label{Eq_mindid}M_m = E - \left( \frac{S_{m}}{m-1}\right).\end{equation}In practice, each feature is rescaled to the $[0,1]$ interval (so is the grid), and $\delta$ is replaced with the edge length, $\ell$, of the cells. In this context, $\ell^{-1}$ is simply the number of cells along each axis of the $E$-dimensional space where the data points are embedded.
\subsection{Detailed Procedure}
In the remainder of this paper, the Morisita estimator of ID will be used only with $m=2$ as advocated in \cite{Go15}. The following steps summarize how to compute the ID of a data set using $M_{m=2}$:
\begin{enumerate}[noitemsep]
\item Rescale each of the $E$ features to the $[0,1]$ interval.
\item \label{item_par} Choose the values of the parameter $\ell^{-1}$ so that the functional relationship of Step \ref{item_lm} can be well approximated by a linear regression model (see Subsection \ref{sub_par}).
\item Superimpose an $E$-dimensional grid over the data points. The size of the grid cells is controlled by the user through the parameter $\ell^{-1}$ which is simply the number of cells along each axis of the grid.
\item Count the number of data points falling into the cells of the grid. This step must be repeated for each value of the parameter $\ell^{-1}$ chosen by the user.
\item Compute the multipoint Morisita index $I_{m=2,\ell^{-1}}$ for each value of the parameter $\ell^{-1}$ using Equation \ref{Eq_mindex}. Notice that the values of $I_{m=2,\ell^{-1}}$ are equal to those of $I_{m=2,\delta}$, since $\delta$ and $\ell^{-1}$ are two different ways of characterizing the size of the same cells.
\item \label{item_lm} Carry out the linear regression of $\log{(I_{m=2,\ell^{-1}})}$ on $\log{(\ell^{-1})}$. Then $S_{m=2}$ is simply the slope of the regression model.
\item Compute $M_{m=2}$ using Equation \ref{Eq_mindid}.
\end{enumerate}
The procedure is illustrated in Figure \ref{Fig_Mindid} for $E=2$. On the right, the two features $F_1$ and $F_2$ have been rescaled to the $[0,1]$ interval and a 2-dimensional grid is superimposed over the data points. The number of cells along each of the two axes of the grid is equal to $4$. This is the value of the parameter $\ell^{-1}$ which allows the user to control the grid resolution. The calculation of $I_{m=2,\ell^{-1}}$ was iterated four times ($R=4$) for $\ell^{-1} \in \lbrace 1,2,3,4\rbrace$, and the results were used to draw the log-log plot shown on the left of the figure. The dashed line represents the linear regression model of Step \ref{item_lm}. Its slope is the Morisita slope $S_2$.
In the next sections, $M_2$ will be computed using the MINDID algorithm \cite{Go15} whose complexity is $\mathcal{O}(N*E*R)$. MINDID is part of the \textsf{R} package ``IDmining'' \cite{GoR16}. It is its high execution speed that allows the efficient implementation of the unsupervised feature selection technique introduced in this paper. For instance, MINDID is able to compute the ID of the butterfly data set (see Section \ref{sim_data}) in $0.07$ seconds (s), $0.17$ s and $0.94$ s for respectively $N=10^3,10^4,10^5$. The experiment was carried out in the \textsf{R} environment using an Intel Core i7-2600 CPU @ 3.40 GHz along with 16.0 GB of RAM under Windows 7.
\subsection{Discussion About the Parameter $\ell^{-1}$}\label{sub_par}
The values of the parameter $\ell^{-1}$ can be chosen by drawing the log-log plot relating $\log{(I_{m=2,\ell^{-1}})}$ to $\log{(\ell^{-1})}$ for a sufficiently large set of values of $\ell^{-1}$. Then the values to be retained for the ID estimation are those corresponding to the linear part of the plot. Here is some guideline for drawing the plot:
\begin{enumerate}[noitemsep]
\item The minimum value that $\ell^{-1}$ can take on is $1$.
\item The maximum value of $\ell^{-1}$ must ensure that one grid cell at least contains two points. If all the occupied cells contain no more than one point, the value of the multipoint Morisita index falls down to zero and does not reflect the scaling behaviour of the data anymore.
\item Between the minimum and the maximum, $\ell^{-1}$ can take on any integer value. In general, it follows a geometric sequence of ratio $1$ or $2$.
\end{enumerate}
Once the plot has been drawn, only the values of $\ell^{-1}$ consistent with a linear regression model must be kept for the ID estimation. Although a linear dependency is possible only for data sets exhibiting self-similarity over several scales, it is hardly ever a limitation, since many real world data sets follow such a behaviour \cite{Sousa07}.
It was also shown that the range of selected scales does not need to be very large in order to achieve good results in terms of redundancy detection and quantification \cite{Go15Esann,Go15}. However, in some extreme cases where $N$ is very low with regard to $E$, it may happen that the maximum value of $\ell^{-1}$ cannot be greater than $1$, making it impossible to compute the data ID. But this limitation does not prevent the application of the Morisita estimator to challenging case studies. And it is also worth mentioning that $M_2$ is more robust to small sample sizes than R\'{e}nyi's dimension of order $2$, $D_2$, that is used in most ID-based methods of data mining \cite{Go15}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{IDRedundancy.pdf}
\caption{Illustration of the three scenarios used to describe the possible redundancies between two variables (or features). Typically, two variables, $F_1$ and $F_2$, can be either (a) not redundant, (b) linearly redundant or (c) non-linearly redundant. In addition, the marginal point distributions are provided for the three scenarios.}\label{Fig_IDRedundancy}
\end{figure}
\section{Intrinsic Dimension and Redundancy}\label{ID_cor}
This section focuses on the use of the data ID to find redundancies among variables (or features). The case with only two variables is first presented and then extended to multivariate interactions.
Three scenarios are used to summarize the possible redundancies that can exist between two random variables. Each scenario considers 50 sampled data points and is illustrated in one of the panels of Figure \ref{Fig_IDRedundancy}. In the left-hand panel, two variables, called $F_1$ and $F_2$, are \textit{not redundant} and, with $ID(\cdot)$ denoting the ID of a data set, one can write that:
\begin{equation}\label{equ_ID_1}
ID(F_1,F_2)\approx ID(F_1)+ID(F_2)\approx 1+1=2.
\end{equation}This result follows from the fact that the data points cover both the 1-dimensional spaces constructed with the two variables taken separately and the 2-dimensional space of the two variables considered together. In contrast, if $F_1$ and $F_2$ are \textit{linearly} or \textit{non-linearly redundant}, as illustrated in the other two panels, their joint ID is approximately equal to $1$:
\begin{equation}\label{equ_ID_2}
ID(F_1,F_2)\approx ID(F_1)\approx ID(F_2)\approx 1.
\end{equation}The resulting point patterns do not cover the 2-dimensional space. Instead, they cover a line in a similar way to what can be observed when $F_1$ and $F_2$ are taken separately. In the third scenario, the non-linear redundancy implies that the marginal distributions of the two features are not identical to the point distribution on the arc. However, it has only a small impact on the ID estimates as reported in \cite{Trai00}. This is the reason why ID-based estimators can effectively deal with non-linear redundancies.
In other words, a feature (or variable) that is redundant does not (or hardly) contribute to increasing the ID of a data set. For instance, in the last two scenarios, one of the features is redundant with the other and can be disregarded without any loss of information. As a consequence, the data ID remains approximately unchanged after the addition of $F_2$ to the data set consisting solely of $F_1$ (and vice versa). In contrast, the first scenario requires that the two features be kept (i.e. none of them is redundant), and the data ID increases from $1$ to $2$ when $F_2$ is added to the data set consisting of $F_1$ (and vice versa).
The same reasoning applies to multivariate redundancy. Let $F_1$, $\ldots$, $F_k$ be $k$ random variables. Then, in the case where $F_k$ is not redundant with the $k-1$ other variables, Equation \ref{equ_ID_1} can be generalized as follows:
\begin{equation}\label{equ_IDRed_1}
ID(F_1,\ldots ,F_{k-1},F_k)\approx ID(F_1,\ldots ,F_{k-1})+ID(F_k)
\end{equation}and, in the case where $F_k$ is completely redundant, the multivariate version of Equation \ref{equ_ID_2} can be written as:
\begin{equation}\label{equ_IDRed_2}
ID(F_1,\ldots ,F_{k-1},F_k)\approx ID(F_1,\ldots ,F_{k-1}).
\end{equation}Finally, in the case where $F_k$ is only partially redundant, and according to the extension of ID to fractal dimensions (see e.g. Traina et al. \cite{Trai10} and De Sousa et al. \cite{Sousa07}), one can write that:
\begin{equation}\label{equ_IDRed_3}
ID(F_1,\ldots ,F_{k-1},F_k)-ID(F_1,\ldots ,F_{k-1})\approx W
\end{equation}where $W \in\left] 0,ID(F_k)\right[$ and where the less redundant $F_k$ is, the greater $W$.
In the next subsection, the proposed algorithm executes a procedure of feature selection which follows directly from Equations \ref{equ_IDRed_1}, \ref{equ_IDRed_2} and \ref{equ_IDRed_3}. However, in real world applications, variable distributions can often greatly depart from the simple ones shown in Figure \ref{Fig_IDRedundancy}. In such situations, Equations \ref{equ_IDRed_1}, \ref{equ_IDRed_2} and \ref{equ_IDRed_3} remain unchanged, but
it must be noticed that:
\begin{equation}
ID(F_i)\leqslant 1 \quad \forall i\in\lbrace 1,2,\ldots,k\rbrace
\end{equation}where the equality holds only if the variables are uniformly distributed. Thus, for Equations \ref{equ_IDRed_1}, \ref{equ_IDRed_2} and \ref{equ_IDRed_3}, one can specify that:
\begin{equation}\label{equ_non_unif}
ID(F_1,\ldots ,F_{k-1},F_k)\leq k.
\end{equation}
\begin{algorithm}[t]
\caption{MBRM}\label{mbrm_algo}
\textbf{INPUT:}
A dataset $A$ with $E$ features $F_{1,\ldots, E}$.
A vector $L$ of values $\ell^{-1}$.
An integer $C$ ($\leq E$) indicating the number of steps of the SFS procedure to be performed (by default $C=E$).
Two empty vectors of length $C$: $SelF$ and $IDF$ for storing, respectively, the names of the selected features and the data ID estimates.
An empty matrix $Z$ for storing the selected features.
Optional: the ID estimate $IDA$ of the full data set $A$.
\textbf{OUTPUT:} $SelF$ and $IDF$.
\begin{algorithmic}[1]
\STATE Rescale each feature to $[0,1]$.
\STATE Unless $IDA$ is given: $IDA=M_2(A)$ (MINDID used with $L$)
\FOR{$i = 1 \ \TO \ C$}
\FOR{$j = 1 \ \TO \ (E+1-i)$}
\STATE $Diff(Z,F_j)=\left| IDA-M_2(Z,F_j)\right|$ (MINDID used with $L$)
\ENDFOR
\STATE Store in $SelF[i]$ the name of the $F_j$ yielding the lowest value of $Diff$.
\STATE Store the corresponding value of $M_2(Z,F_j)$ in $IDF[i]$.
\STATE Remove the corresponding $F_j$ from $A$ and add it into $Z$.
\ENDFOR\\
\end{algorithmic}
\end{algorithm}
\section{The Morisita-Based Filter for Redundancy Minimization}\label{ID_RM}
The Morisita-Based filter for Redundancy Minimization (MBRM) aims to select the smallest subset of features necessary to carry all the information content of a data set. MBRM follows a SFS search strategy and, in each step of the procedure, it searches for the feature which yields the greatest value of $W$ (i.e. the feature which carries the largest amount of new information according to Equation \ref{equ_IDRed_3}). To achieve this goal, a fast solution, robust to sample size and noise, is to reduce the following difference to zero (see Algorithm \ref{mbrm_algo}):
\begin{equation}\label{equ_MBRM}
Diff(F):=\left| M_2(A)-M_2(F)\right|
\end{equation}where $A$ is a data set consisting of $E$ features $F_{1,\ldots, E}$, $F$ is a set of $i$ features of $A$ with $i\leqslant E$ and $M_2(\cdot)$ denotes the estimation of the data ID using the Morisita estimator $M_2$. More precisely, MBRM starts with $i=1$ and searches for the individual feature which contributes the most to reducing $Diff$. Once identified, this feature is retained. Then the algorithm carries on in searching for the second feature ($i=2$) which leads to the largest possible decrease in $Diff$ when combined with the previously retained feature. The operation is iterated until $i=C$ with $C\leqslant E$. Typically, $C$ can be set to a value lower than $E$ (by default $C=E$) if it is assumed that $A$ is highly affected by redundancy. Finally, the set of features contributing to decreasing the value of $Diff$ to approximately $0$ (i.e. $Diff\approx 0$) is the smallest set of features carrying all the information content of $A$.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{MBRM.pdf}
\caption{Illustration of how MBRM works. A data set $A$ is used. It consists of four features $F_1$, $F_2$, $F_3$ and $F_4$. The result shows that $F_1$ is redundant and the three other features are enough to carry all the information content of $A$.}\label{Fig_MBRM}
\end{figure}
Figure \ref{Fig_MBRM} illustrates how MBRM works. In this example, the data set $A$ consists of four features, namely $F_1$, $F_2$, $F_3$ and $F_4$, and the full data ID is equal to $2.20$. In Step 1 ($i=1$), the feature that contributes the most to reducing $Diff$ is $F_4$, and the value of $M_2(F_4)$ is reported in the plot. Then MBRM moves on to Step 2 ($i=2$) where $F_3$ is identified as the feature leading to the largest decrease in the value of $Diff$ when combined with $F_1$. Therefore, the value of $M_2(F_4,F_3)$ is added to the plot in second position. It appears before the values of $M_2(F_4,F_3,F_2)$ and $M_2(F_4,F_3,F_2,F_1)$ which result from Steps 3 ($i=3$) and 4 ($i=4$). Eventually, the plot highlights that only $F_4$, $F_3$ and $F_2$ contribute to increasing the ID estimates of the data. Hence, it can be concluded that the removal of $F_1$ will have no impact on the information content of $A$. In this simple example, if $C$ had been set to $3$, MBRM would have returned only $F_4$, $F_3$ and $F_2$. No information would have been lost as indicated by the equality between $M_2(F_4,F_3,F_2)$ and $M_2(A)$.
MBRM relies on the recent MINDID algorithm \cite{Go15} for the computation of $M_2$ (see Section \ref{Mindex}). Consequently, MBRM takes as input a vector $L$ consisting of $R$ values of $\ell^{-1}$. It will be shown in the next section that these values can be chosen with regard to the full data set $A$ and remain unchanged throughout the steps of the sequential forward search. In the same way as MINDID, MBRM is linear on both the number $N$ of points and the number $R$ of scales. But the complexity of the sequential search is quadratic on the number $E$ of features. In spite of this limitation, the execution time remains competitive thanks to the fast computation speed of MINDID. It can also be significantly reduced by setting $C$ to a low value. For instance, when applied to the butterfly data set (see Section \ref{sim_data}), MBRM runs in $2.14$ seconds (s), $3.06$ s and $12.45$ s for respectively $N=10^3,10^4,10^5$ and $C=E$. And, with $C=3$ (i.e. the number of features necessary to carry all the information content of the data), the execution time of MBRM is reduced to $0.68$ s, $1.04$ s and $4.07$ s for the same numbers of points. The computations were performed in the \textsf{R} environment using an Intel Core i7-2600 CPU @ 3.40 GHz along with 16.0 GB of RAM under Windows 7.
Finally, a possible extension of MBRM would be to allow the user to provide the value of the full data ID in cases where it is known a priori. As a result, MBRM would never need to be applied to spaces with more than $C$ dimensions. However, in this paper, the full data ID will always be estimated using $M_2(\cdot)$ within MBRM, since it is hardly ever available in advance.
\section{Experimental Study}\label{exp_res}
In this section, numerical experiments designed to test the MBRM algorithm are presented. A simulated data set (see Subsection \ref{sim_data}) was used to examine the impact of sample size and noise as well as the ability of the algorithm to capture non-linear relationships between features. Monte-Carlo simulations were also performed in order to quantify the variability of the results. In addition, MBRM was applied to real world case studies (see Subsection \ref{real_data}). In this context, the preserved amount of relevant information was assessed by comparisons with benchmark feature selection techniques through the use of Random Forests \cite{Bre01}. Notice that all the experiments were carried out in the \textsf{R} environment \cite{Rcran,GoR16}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{Sample_size.pdf}
\caption{MBRM applied to the input variables of the butterfly data set: (a) $1$ simulation with $N=1000$, (b) $100$ simulations with $N=1000$, (c) $100$ simulations with $N=2000$, (d) $100$ simulations with $N=10000$. In each panel, the dashed line indicates the ID of the full data set and, when $100$ simulations are used, the mean ID estimates are given along with error bars indicating $\pm$ the standard deviations. Notice that the names of the features were shortened to $F$ because the way they are ordered by the SFS search strategy can change between the simulations.}\label{Fig_sampsize}
\end{figure}
\subsection{Simulated Data Set}\label{sim_data}
The MBRM algorithm was applied to the input space of the butterfly data set \cite{Go15Esann}. It consists of eight variables (or features): two uniformly distributed variables $F_1,F_2 \in \left] -5,5\right[$, $F_3=log_{10}{(F_1+5)}$, $F_4 = F_1^2-F_2^2$, $F_5 = F_1^4-F_2^4$, a third uniformly distributed variable $F_6\in \left]-5,5\right[$, $F_7=log_{10}{(F_6+5)}$ and $F_8=F_6+F_7$. The butterfly data set is then generated by random sampling of $F_1$, $F_2$ and $F_6$. In this paper, the following sample sizes were considered: $N=1000,2000,10000$, and for each of them, $100$ versions of the data set were produced.
The MBRM algorithm was applied to the butterfly data set for each of the above-mentioned sample size. The results are displayed in Figure \ref{Fig_sampsize}. The first panel shows the plot provided by MBRM when applied to one simulation of $1000$ data points. After the addition of $F_8$ to $F_2$ and $F_1$, the ID estimate is equal to $3.02$ (i.e. $M_2(F_2,F_1,F_8)=3.02$) and a clear cut-off point can be observed, since the remaining features hardly contribute to increasing the value of the full data ID (i.e. hardly contribute to reducing the value of $Diff$ defined in Equation \ref{equ_MBRM}). Consequently, the MBRM algorithm has detected that only three features, namely $F_1$, $F_2$ and $F_8$, are enough to carry all the information content of the butterfly data set. This is correct by construction, since the full data set was generated from $F_1$, $F_2$ and $F_6$ and since $F_8$ is redundant with $F_6$. MBRM has thus successfully fulfilled its goal.
The remaining panels focus on the variability of the ID estimates with regard to the number of data points. The plots depict the evolution of the mean ID estimates during the SFS search procedure, and the errors bars represent $\pm$ the standard deviations of the estimates. The mean value and the standard deviation of the full data ID estimates are also given. They are denoted respectively by the dashed and solid black lines. As expected, the variability tends to be higher for lower sample sizes. This is clearly visible for the full data ID estimates, but it never prevents MBRM from detecting that only three features can convey all the information content of the butterfly data set. Notice, however, that the ordering of the features resulting from the SFS search procedure can change from one simulation to the next. Consequently, the names of the features were shortened to $F$ in each panel.
\begin{table}
\centering
\footnotesize
\begin{tabular}{cc}
\hline
$N$ & First Three Features (Occurrences)\\
\hline
\hline
1000 & $F_1$,$F_2$,$F_6$ ($72$); $F_1$,$F_2$,$F_8$ ($19$); $F_1$,$F_2$,$F_7$ ($7$); $F_1$,$F_3$,$F_6$ ($1$); $F_1$,$F_4$,$F_6$ ($1$)\\
2000 & $F_1$,$F_2$,$F_6$ ($82$); $F_1$,$F_2$,$F_8$ ($18$)\\
10000 & $F_1$,$F_2$,$F_6$ ($100$) \\
\hline
\end{tabular}
\caption{The first three features selected by MBRM (not necessarily in that order) when applied successively to $100$ simulations of the butterfly dataset for $3$ different sample sizes. The number of occurrences of each triplet is indicated in brackets.}
\label{Tab_Feat_Order}
\end{table}
Another interesting point is the composition of the set of features picked first by MBRM. By construction, three features should be identified by the algorithm as being necessary to preserve the information content of the butterfly data set. Several triplets of features could work, but MBRM favours $F_1$, $F_2$ and $F_6$ as indicated in Table \ref{Tab_Feat_Order}. This is mainly due to the non-linear construction of $F_3$, $F_4$, $F_5$ and $F_7$ (because non-linearity influences the way the data points are distributed over the data manifold as shown in the right-hand panel of Figure \ref{Fig_IDRedundancy}, which can in turn have a slight effect on the ID estimates) and to the fact that $F_4$ and $F_5$ are fully redundant with $F_1$ and $F_2$ considered jointly rather than independently. For these reasons, when the sample size is large enough, the features constructed to be redundant do not maximize the ID estimates as much as $F_1$, $F_2$ or $F_6$ and, consequently, they are not picked first by MBRM.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{Noise_and_Shuffling.pdf}
\caption{MBRM applied to the input variables of the butterfly data set with $N=10000$: (left) $100$ simulations with different levels of Gaussian noise, (right) $1$ simulation after the shuffling of $F_5$ and $F_7$. The dashed lines indicate the ID of the full data sets, and in the left-hand panel, the mean ID estimates are given along with error bars indicating $\pm$ the standard deviations.}\label{Fig_noise_shuf}
\end{figure}
A good algorithm should also be robust to noise. Consequently, another experiment was carried out for which $F_3$, $F_4$, $F_5$, $F_7$ and $F_8$ were corrupted with a Gaussian noise. For each of these features, the mean of the noise was fixed at $0$, and the standard deviation (sd) was successively set to $1\%$, $10\%$, $20\%$ and $40\%$ of the standard deviation of the original feature. The left-hand panel of Figure \ref{Fig_noise_shuf} shows the results. At $1\%$, the noise hardly affects MBRM, and the cut-off point is visible up to $20\%$. Beyond this threshold, the algorithm has difficulty fulfilling its task, but it is quite normal, since it could be argued that a data set affected by more than $40\%$ of noise should be used with caution. Moreover, it is also clear that the full data ID and the variability of the estimates increase with the level of noise. This result is simply a consequence of the fact that the noise enlarges the portion of space occupied by the data points.
Finally, the right-hand panel of Figure \ref{Fig_noise_shuf} displays the results provided by MBRM after the shuffling of $F_5$ and $F_7$. The shuffling makes the two features completely independent of, respectively, $F_1$ and $F_6$ (i.e non-redundant with $F_1$ and $F_6$) and, consequently, MBRM detects that they carry new information, which is correct. Notice also that the full data ID is lower than $5$, as expected from Equation \ref{equ_non_unif}: since $F_5$ and $F_7$ are not uniformly distributed, the data ID is lower than $k=5$. This means that, in general, the value of the full data ID cannot be used as the number of features to be retained. It must only be considered a lower bound. And, of course, selecting less features than the value of the ID will always lead to a loss of information. For instance, $M_2(F_1,F_2,F_4)\approx 2$, which implies that two features, say $F_1$ and $F_2$, are necessary to correctly synthesize the data, but $F_4$ alone cannot account for all the available information. Differently put, it is possible to reconstruct $F_4$ from $F_1$ and $F_2$, but neither $F_1$ nor $F_2$ can be perfectly retrieved from $F_4$.
\subsection{Real World Case Studies}\label{real_data}
Four real world data sets from the UCI machine learning repository \cite{Lich} were used: PageBlocks, Parkinson, Ionosphere and LIBRAS movements. Duplicate data points were removed, and the main characteristics of the resulting data are briefly summarized in Table \ref{Tab_data}. In addition, the last column of the table provides the values of the parameter $\ell^{-1}$ used to apply the MBRM algorithm to the input space of the data sets. The complete feature selection procedure was performed as follows (for each data set):
\begin{enumerate}[noitemsep]
\item The plot relating $\log{(I_{m=2,\ell^{-1}})}$ and $\log{(\ell^{-1})}$ was drawn for the $E$ features taken together and for values of $\ell^{-1}$ ranging from $1$ to the highest possible value (i.e. the value ensuring the presence of two points in at least one of the $Q$ cells).
\item The range of values of $\ell^{-1}$ corresponding to the linear part of the log-log plot was retained. For each of the four data sets, this final range was that of step $1$.
\item If the upper bound of the range was lower than $30$, each integer value of the range was used in the final set given in Table \ref{Tab_data}. If not, for efficiency purposes, only the values following a geometric sequence of ratio $2$ were retained.
\item MBRM was applied to each data set with the parameter values of Step $3$, and the features necessary to approximately reach the full data ID were selected.
\end{enumerate}
The results are given in Figure \ref{Fig_realdata} and summarized in Table \ref{Tab_res_MBRM}. For each data set, MBRM offers a clear cut-off point allowing an easy distinction between the features to be selected and those considered redundant. Moreover, it leads to a significant reduction in the data dimensionality. In the case of LIBRAS, only $24\%$ of the original features were selected, and the other data sets were reduced by half.
\begin{table}
\centering
\footnotesize
\begin{tabular}{ccccc}
\hline
Data Sets & $N$ & $E$ & \# Classes & Parameter $\ell^{-1}\in$\\
\hline
\hline
PageBlocks & $5393$ & $10$ & $5$ & $\lbrace 2^n\mid n=0,1,2,\ldots ,11\rbrace$\\
Parkinson & $195$ & $22$ & $2$ & $\lbrace 1,2,3,\ldots ,8\rbrace$ \\
Ionosphere & $350$ & $34$ & $2$ & $\lbrace 1,2,3,\ldots ,13\rbrace$ \\
LIBRAS & $330$ & $90$ & $15$ & $\lbrace 1,2,3,4,5\rbrace$ \\
\hline
\end{tabular}
\caption{Data set characteristics and values of the parameter $\ell^{-1}$.}
\label{Tab_data}
\end{table}
\begin{table}
\centering
\footnotesize
\begin{tabular}{cccc}
\hline
Data Sets & \# Sel. Feat.& $M_2(All \ Feat.)$ & Execution Time (s)\\
\hline
\hline
PageBlocks & $5 \ (50\%)$ & $2.13$ & $3.45$ \\
Parkinson & $12 \ (52\%)$ & $4.87$ & $5.32$ \\
Ionosphere & $16 \ (47\%)$ & $3.19$ & $25.68$ \\
LIBRAS & $22 \ (24\%)$ & $6.43$ & $130.81$ \\
\hline
\end{tabular}
\caption{Summary of the results provided by MBRM and execution time (in seconds). In the column ``\# Sel. Feat.'', the values in brackets are the percentages of selected features with regard to the total numbers of features.}
\label{Tab_res_MBRM}
\end{table}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{RealData.pdf}
\caption{Results of MBRM: (a) PageBlocks, (b) Parkinson, (c) Ionoshpere, (d) LIBRAS Movements. The features to be selected are indicated in bold red, and the dashed lines correspond to the ID of the full data sets.}\label{Fig_realdata}
\end{figure}
In terms of efficiency, MBRM is linear on $N$, but its bottleneck is the number of features. Therefore, it takes $3.45$ seconds (s) to run the algorithm on PageBlocks (using \textsf{R} and an Intel Core i7-2600 CPU @ 3.40 GHz along with 16.0 GB of RAM under Windows 7), while $130.81$ s are necessary for LIBRAS. Nevertheless, if it can be assumed that a data set contains many redundant features, the parameter $C$ can be set to a value lower than $E$ to shorten the execution time. For instance, for the LIBRAS data set, if $C$ is set to the number of selected features, the execution time of MBRM is cut down to $17.69$ s.
The effectiveness of MBRM in fulfilling redundancy minimization must also be assessed. In the case of the PageBlocks data set, it is straightforward, since the five features considered redundant by the algorithm are truly redundant. In other words, they can be computed from the selected features as explained on the UCI machine learning repository \cite{Lich}. Regarding the remaining data sets, no a priori knowledge about the redundancy between the features is available, and a more complex procedure is required to evaluate the amount of information kept by the selected features. A possible solution is to use an algorithm, such as Random Forest (RF), to classify the data points according to the output variables. In this paper, the choice of RF was motivated by its recently recognized benchmark performance and its reputation of working well in high-dimensional spaces \cite{Bre01,Has09,Zh16}. Then the fundamental idea consists in performing the classification task twice, once with all the input features and once with only the selected features. If the reduced subset does not lead to a higher error rate than that obtained with the full data set, the relevant information has been preserved, and the feature selection can be considered successful.
The same idea can also be extended to compare different techniques of feature selection. A technique that leads to a significant increase in the classification error does not perform as well as a technique that maintains the accuracy obtained with the full data set. Here, four benchmark algorithms were used: Fractal Dimension Reduction (FDR) \cite{Trai00}, Feature Similarity Feature Selection (FSFS) based on the maximum information compression index \cite{Mit02}, Laplacian Score (LScore) \cite{He06} and Multi-Cluster Feature Selection (MCFS) \cite{Cai10}. For FSFS, LScore and MCFS, the number of selected features was set to the number of features selected by MBRM, and the number of nearest neighbours, $k$, of LScore and MCFS was fixed at $5$. Finally, regarding MCFS, the number of used eigenvectors was set to the number of classes.
The exact evaluation procedure implemented in this research was repeated six times for each data set: once with all the features, once with the features selected by MBRM and four more times with the features selected by FDR, FSFS, LScore and MCFS. This procedure follows four steps and is inspired by similar procedures suggested in \cite{Ja97,Reu03}:\begin{enumerate}[noitemsep]
\item $20\%$ of the data points were randomly assigned to a test set and the remaining ones were passed on to Step 2.
\item 10-fold cross-validation was performed to tune the two parameters of RF (these two parameters are the number of trees and the number of predictors processed at each split of the trees). The Overall Accuracy (OA) was used as evaluation metric.
\item \label{item_tst}A RF model was trained with the values of the two parameters of Step 2 and was then used to classify the data points of the test set. Finally, the test OA, $OA_{t}$, was computed using the following formula:\begin{equation}\label{Eq_OA}
OA_{t}=\frac{1}{N_t} \sum_{i=1}^{N_t}I(y_i = \hat{y_i})
\end{equation}where $I(y_i = \hat{y_i})$ is an indicator variable, $N_t$ is the number of data points in the test set, $y_i$ refers to the true class label of the $i^{th}$ observation and $\hat{y_i}$ is the predicted class label.
\item Steps 1 to 3 were repeated $20$ times, and the mean test OA (in percent) along with the corresponding standard deviation were computed. These two values are given in Table \ref{Tab_res_RF_OA} for each data set.
\end{enumerate}
Table \ref{Tab_res_RF_OA} shows that the mean overall accuracies obtained using the features selected by MBRM are better than or close to those computed from the full data sets. By ``close'', it is meant that they are within half a standard deviation of one another. This suggests that the selected features carry all the information content relevant to the classification tasks and that the MBRM algorithm has fulfilled its purpose. Moreover, compared to the four benchmark techniques used in the experiments, MBRM always provided equal or better results. For instance, in the case of the LIBRAS data set, LScore, FSFS and MCFS led to a significant increase in the mean overall accuracy, while MBRM maintained the value obtained using all the features. Regarding FDR, the algorithm could not be run because the log-log plot necessary to estimate the ID did not exhibit any linear behaviour. The same remark holds for the Ionosphere data.
\begin{landscape}
\begin{table}
\centering
\footnotesize
\begin{tabular}{cccccc|c}
\hline
Data Sets & MBRM & FDR & FSFS & LScore & MCFS & All Feat. \\
\hline
\hline
PageBlocks & {\boldmath$97.37 \ (0.51)$} & $97.14 \ (0.52)$ & $97.09 \ (0.63)$ & $96.79 \ (0.53)$ & \underline{$97.36 \ (0.58)$} & $97.54\ (0.56)$ \\
Parkinson & {\boldmath$92.05 \ (5.32)$} & $91.54 \ (5.46)$ & \underline{$91.92 \ (5.41)$} & $89.49 \ (4.77)$ & $90.13\ (3.93)$ & $90.38\ (5.93)$ \\
Ionosphere & {\boldmath$91.93 \ (3.29)$} & - & $90.43 \ (2.94)$ & \underline{$91.29 \ (3.31)$} & $91.21\ (3.08)$ & $92.57\ (2.92)$ \\
LIBRAS & {\boldmath$76.29\ (3.82)$} & - & $66.74 \ (5.91)$ & $57.80 \ (5.76)$ & \underline{$69.62 \ (5.81)$} & $75.68\ (4.68)$\\
\hline
\end{tabular}
\caption{Mean overall accuracies (over $20$ random splits and in percent) yielded by Random Forests when applied to the features selected by MBRM, FDR, FSFS, LScore, MCFS and to the full data sets. The standard deviations are indicated in brackets. Besides, the best results provided by the feature selection techniques are shown in bold, and the second best results are underlined. The goal is to be higher than or (if smaller) as close as possible to the reference values given in the column ``All Feat.''.}
\label{Tab_res_RF_OA}
\end{table}
\begin{table}
\centering
\footnotesize
\begin{tabular}{cccccc|c}
\hline
Data Sets & MBRM & FDR & FSFS & LScore & MCFS & All Feat. \\
\hline
\hline
PageBlocks &{\boldmath$84.53 \ (2.44)$} & $83.04 \ (2.53)$ & $82.78 \ (3.18)$ & $80.73 \ (2.38)$ & \underline{$84.41 \ (3.01)$} & $85.56\ (2.95)$ \\
Parkinson &{\boldmath$77.06 \ (15.15)$}& $75.40 \ (15.84)$ & \underline{$76.54 \ (15.70)$} & $71.13 \ (13.27)$ & $72.41 \ (10.20)$ & $71.94\ (18.55)$ \\
Ionosphere &{\boldmath$82.36 \ (7.10)$} & - & $78.91 \ (6.64)$ & \underline{$81.02 \ (6.79)$} & $80.68 \ (6.87)$ & $83.84\ (6.16)$ \\
LIBRAS &{\boldmath$74.40 \ (4.11)$} & - & $64.12 \ (6.33)$ & $54.54 \ (6.29)$ & \underline{$67.24 \ (6.16)$} & $73.73\ (5.04)$\\
\hline
\end{tabular}
\caption{Mean Kappa coefficients (over $20$ random splits and multiplied by $100$) yielded by Random Forests when applied to the features selected by MBRM, FDR, FSFS, LScore, MCFS and to the full data sets. The standard deviations are indicated in brackets. Besides, the best results provided by the feature selection techniques are shown in bold, and the second best results are underlined. The goal is to be higher than or (if smaller) as close as possible to the reference values given in the column ``All Feat.''.}
\label{Tab_res_RF_K}
\end{table}
\end{landscape}
In addition to the overall accuracy, Cohen's Kappa coefficient of agreement $\kappa$ \cite{Coh60} was also computed in Step \ref{item_tst} of the evaluation procedure (see Table \ref{Tab_res_RF_K}). Cohen's kappa is an evaluation metric
that takes into account the number of correctly classified data points that may occur by chance. This way, the possible bias of the OA towards large classes is reduced. Cohen's Kappa is commonly used in data mining (see e.g. \cite{Ste16,Vo14}), and it is given by:\begin{equation}\label{Eq_K}
\kappa_{t}=\frac{N_t\sum_{c=1}^{B}T_c-\sum_{c=1}^{B}G_cP_c}{N_t^2-\sum_{c=1}^{B}G_cP_c}
\end{equation}where the subscript $t$ indicates that the coefficient is computed on a test set, $B$ is the number of classes, $T_c$ indicates the number of correctly classified samples for class $c$ and $N_t$ is the number of data points in the test set. Finally, $G_c$ and $P_c$ are the actual number of samples belonging to class $c$ and the number of samples classified in this class. Cohen's Kappa ranges in $[-1,1]$, but negative values are hardly ever met. It is equal to $1$ in case of complete agreement and to zero (or below) if a classifier does not perform better than what would be expected from pure randomness. The results are provided in Table \ref{Tab_res_RF_K} and confirm those of Table \ref{Tab_res_RF_OA}. Notice, however, that they are lower and more prone to variability than the results yielded by the overall accuracy. But this is consistent with the way Cohen's Kappa accounts for the agreement due to chance.
Based on the numerical experiments presented in this section, it can be concluded that MBRM performed well and that it is a new promising tool for redundancy minimization.
\section{Conclusion}
This paper introduces a new algorithm for unsupervised feature selection called Morisita-Based filter for Redundancy Minimization (MBRM). MBRM relies on the Morisita estimator of Intrinsic Dimension (ID) and aims to identify the smallest subset of features containing all the information content of a data set. It was successfully tested on simulated data with different levels of sample size and noise. In addition, real world case studies from the UCI machine learning repository were used. MBRM turned out to be effective in a wide range of situations characterized by different numbers of data points, features and classes. When no information about the redundancy between the features was available in advance, a comprehensive procedure based on random forests was implemented to assess the performance of the feature selection. The classification results demonstrated that MBRM did not lead to any loss of relevant information, while it cuts down the size of every data set by half or more. Comparisons with benchmarks techniques confirmed the promising performance of the proposed algorithm. MBRM has also a couple of practical advantages over more traditional techniques of unsupervised feature selection. First, it is able to determine how many features should be kept. Second, the values of its parameter $\ell^{-1}$ can be set without resorting to any learning machine or a priori knowledge. Finally, it was shown that MBRM was able to outperform the FDR algorithm due to the Morisita estimator of ID.
From a broader perspective, this research contributes to highlighting that the concept of ID can help mitigate issues raised by large data sets. Future research will be devoted to challenging applications in hyperspectral remote sensing. In this context, comparisons between ID-based and traditional methods will be carried out, and the use of ID to perform advanced data mining tasks other than feature selection will be thoroughly explored. In addition, the automatic parallelization of the inner loop of the MBRM algorithm will also be examined.
\section{Acknowledgements}
The authors are grateful to the anonymous reviewers for their helpful and constructive comments that contributed to improving the paper. They also would like to thank Michael Leuenberger and Mohamed Laib for many fruitful discussions about machine learning and statistics.
\bibliographystyle{elsarticle-num}
|
\section{INTRODUCTION}
\label{sec:intro}
\input{blast-tng_model}
The Balloon-borne Large Aperture Submillimeter Telescope - The Next Generation (BLAST-TNG)\cite{Galitzki2014b} consists of a 2.5\,m Cassegrain telescope with three microwave kinetic inductance detector (MKID) arrays operating over 30\% bandwidths centered on 250, 350, and 500\,$\mu$m, which have 1836, 950, 460 detectors with diffraction limited resolution of 25$^{\prime\prime}$, 35$^{\prime\prime}$, and 50$^{\prime\prime}$, respectively. It will be flown on a stratospheric balloon as part of NASA's long duration balloon program. An overview of the instrument design is shown in Figure \ref{fig:model}.
The primary science goal of BLAST-TNG is to probe polarized thermal emission from dust in Galactic star forming regions in order to detect the local magnetic field orientation as projected on to the plane of the sky. The Antarctic flights in 2010\cite{Pascale2012} and 2012\cite{Galitzki2014} of the previous instrument, BLASTPol, produced maps of magnetic field morphology in star forming regions with unprecedented levels of detail\cite{Matthews2014, Gandilo2016, Fissel2016, Santos2016, Soler2016, Galitzki2016, Shariff2016}. The demonstrated scientific potential of submillimeter polarimetry observations prompted the design and construction of BLAST-TNG, which will have 16 times the BLASTPol mapping speed as well as better resolution.
The increased collecting area, sensitivity, and resolution will allow for the mapping of more clouds than the 2012 flight and enable observations of diffuse galactic dust emission. BLAST-TNG will allow us to investigate the role magnetic fields play in the formation and evolution of molecular clouds and their associated sub-structures including filaments, cores, and protostars as well as furthering our understanding of dust grain emission and alignment mechanisms. Additionally, BLAST-TNG provides a unique platform for observations which link the large-scale {\it Planck}\cite{PlanckI2015} polarimetry maps to the small scale, but high resolution, polarimetry maps made by interferometers such as ALMA\cite{ALMA2016}.
\section{Magnetic Fields in Star-Forming Regions}
\label{sec:magfield}
In order to create a complete theory of star formation we must understand the processes that regulate the star formation rate in molecular clouds. Recent progress includes using observations of dust emission and extinction, which shows how core mass distribution correlates with observed stellar mass distribution\cite{Nutter2007}. {\it Herschel} data has yielded many important results relevant to the star formation process such as observations that filamentary structures show a tendency for the densest filaments to be perpendicular to the local magnetic field, while faint, low density sub-filaments tend to be parallel to the local magnetic field\cite{Palmeirim2013}. Additionally, {\it Planck}'s all-sky maps of polarized dust emission have greatly increased our understanding of the role magnetic fields play in our galaxy. From these maps, strong evidence has been found that magnetic field orientation is correlated with molecular cloud structure and the degree of this alignment depends on the magnetic field strength and turbulent properties of the cloud being observed\cite{PlanckXXXV2016}.
However, there are many questions concerning the star-formation process and the evolution of cloud structure that remain to be addressed\cite{McKee2007}. Examples include whether the lifetimes of molecular clouds and their internal structures are equal\cite{Vazquez2006} to or larger\cite{Netterfield2009,Blitz2007,Goldsmith2008} than the turbulent crossing time. To have lifetimes longer than crossing times would require a supporting mechanism to counteract gravity. Magnetic fields could provide such support and numerical simulations have shown magnetic fields in clouds can drastically alter star formation efficiencies and the lifetimes of molecular clouds\cite{Li2010,Hennebelle2011}. However, knowledge of magnetic fields and their interaction with molecular cloud structure is still fairly limited. Zeeman splitting observations have produced measurements of the field strength along the line of sight, but are limited to bright regions, and optical extinction polarization observations have produced a small number of magnetic field pseudo-vectors, but only in areas of low extinction\cite{Crutcher2010,Falgarone2008}.
The most promising method for detecting magnetic fields over large ranges of dust column density is with far-IR and submillimeter polarimetry\cite{Hildebrand2000,WardThompson2000,WardThompson2009}. Spinning dust grains preferentially align with short axes parallel to the local magnetic field through the process of radiative alignment torques\cite{Andersson2015}. The grains emit modified blackbody radiation that peaks in the far-IR/submillimeter and is polarized orthogonally to the local magnetic field. BLAST-TNG will have the unique capability to create degree-scale polarization maps of molecular clouds with sub-arcminute resolution with a mapping speed that will allow it to cover multiple targets during each flight. BLAST-TNG data will build on the results of BLASTPol, enabling additional and more detailed comparisons between polarization maps and numerical simulations\cite{WardThompson2000}.
BLAST-TNG observations will target the following three key questions in star formation:\emph{
}i) \emph{Is core morphology and evolution determined by large-scale magnetic fields?} ii) \emph{Does filamentary structure have
a magnetic origin? iii) What is the field strength, and how does it vary from cloud to cloud?} as further discussed in Ref. \citenum{Fissel2010}.
\section{Instrument}
\label{sec:instrument}
\input{fgr_optic_v1}
BLAST-TNG will continue the legacy of BLASTPol with the construction of a new instrument that incorporates many successful elements from previous ballooning experiments. The optical layout of BLAST-TNG is shown in Figure \ref{fig:optics} with the optical component parameters listed in Table \ref{tab:optica}. A 2.5\,m diameter carbon fiber primary mirror\footnote{Vanguard Space Technologies: 9431 Dowdy Drive, San Diego, CA 92126} provides diffraction-limited observations. The telescope uses a Cassegrain configuration with the primary mirror illuminating a 56\,cm diameter aluminum secondary mirror. The secondary mirror is actuated to allow for refocusing during the flight to adjust to the differential thermal contraction of the support structure. The beam is then re-imaged by the 4\,K cold optics which use a modified Offner relay configuration. The light is split by two dichroic filters\cite{Ade2006} into the science bands at 250, 350, and 500\,$\mu$m\ to allow for simultaneous observations of the same patch of sky. The cold optics have been installed in the flight cryostat and will be characterized concurrently with the detector performance. The primary and secondary mirrors and supporting structure are on schedule for a July 2016 delivery.
Achieving a circular, 22$^{\prime}$\ diameter, field of view (FOV) requires re-imaging optics significantly larger than those of BLASTPol. The size constraints necessitated the construction of a new cryostat which was also designed to have a longer hold time of $\sim$28 days, versus the previous 13 day hold time of BLASTPol. The BLAST-TNG cryostat has a 250 liter liquid helium bath to provide cooling to 4\,K and utilizes two vapor-cooled shields (VCS) to provide additional thermal isolation. The arrays are kept at 270\,mK by a closed-cycle \textsuperscript{3}He refrigerator with a 1\,K intercept stage provided by an open cycle \textsuperscript{4}He pumped pot that is fed off the main Helium tank and vents to atmosphere during flight. The detector arrays and cold optics are contained within an Amuneal\footnote{Amuneal Manufacturing Corp., 4737 Darrah Street, Philadelphia, PA 19124} enclosure which provides shielding from ambient magnetic fields. The light is coupled to the detectors with an aluminum feedhorn block with a three step Potter\cite{Potter1963} style profile to provide a more symmetric beam in both polarization directions than previously achieved with conical feedhorns\cite{Rownd2003}. Each feedhorn pixel contains two orthogonally oriented MKIDs (See Fig. \ref{fig:horn_layouts}). BLAST-TNG will serve as a pathfinder instrument for MKIDs \cite{Day2003}, which have never been flown before. The development of MKID arrays for astronomy is an extremely active area of detector research, and flight testing them will be a significant milestone.
Additional details of the instrument design including the electronics, pointing system, control system, and detector readout can be found in Ref. \citenum{Galitzki2014b}.
\input{HornArrays_v1}
\subsection{Cryogenic Performance}
The BLAST-TNG cryostat\footnote{Precision Cryogenic Systems Inc., 7804 Rockville Rd., Indianapolis, IN 46214} was delivered to the University of Pennsylvania (UPenn) in February 2015. Over the following year the thermometry, refrigerator systems, re-imaging optics, and the 250\,$\mu$m\ detector array were integrated and tested in increments to allow us to characterize the performance of the cryogenic system and debug systems as they were added. The cross section of the cryostat can be seen in Figure \ref{fig:cryostat}. A thermal model was developed to describe the expected steady-state thermal load on each temperature stage. The thermal model took into account radiative loading as well as conduction through the G10 mechanical supports, cabling, motor axles, the helium fill tube, and the 1\,K refrigerator exhaust tube (See Table \ref{tab:therm_model}).
During testing, the VCSs were observed to reach a steady temperature of $\sim$65\,K and $\sim$165\,K. The steady state load on the 4\,K cold plate was measured to be $\sim$340\,mW, as determined by the measured helium boil-off rate of the cryostat. The loading corresponds to a 22.5 day hold time for the 250 liter tank with an approximate boil-off rate of 11 liters of liquid helium per day. The observed loading is approximately 40\% larger than the predicted loading of 240\,mW from the thermal model used in the design of the cryostat. We believe our excess loading is due in part to un-modeled light leaks in the MLI blanket around fixtures and feedthroughs, as suggested by SPIDER\cite{Gudmundsson2015}, but also due to unexpected complexities in the thermal behavior of G10 material, which provides the mechanical support in the BLAST-TNG cryostat.
\input{cryostat}
By adjusting the conductive model of the G10 we can account for most of the additional loading observed. This observation has prompted us to replace the G10 sections between the 165\,K and 65\,K stages and between the 65\,K and 4\,K stages with thinner walled material. The change will decrease the effective loading on the 4\,K stage to bring the performance in line with our target 28 day hold time. It should also be noted from Ref. \citenum{Gudmundsson2015} that the thermal loading at balloon flight altitudes is observed to be less due to the cooler temperature of the cryogenic vessel and the reduced optical loading through the window which can increase the hold times of cryostats during flight.
As a baseline for comparing cryostat performance a figure of merit was developed in Ref. \citenum{Holmes2001} that divides the radiative loading, $H$, by the cryogen depletion rate, $R$. For the SPIDER cryostat along with many satellite cryostats, $H/R \approx 60$\,W\,Days/L whereas the BLAST-TNG cryostat ranks slightly higher with $H/R\approx 230$\,W\,Days/L using the 250\,L cryogenic volume, a hold time of 22.5 days, and the approximately 6\,m$^2$ surface area of the 4\,K shield.
\subsection{Detector Performance}
The 250\,$\mu$m\ MKID detectors have been shown to be photon noise limited in a 7 pixel test array that was feedhorn coupled to a variable blackbody source. Additional details of the tests and results can be found in Ref. \citenum{Hubmayr2015}. The noise equivalent power (NEP) of the detectors was fit using a three component noise model that included recombination noise and the photon noise on top of a flat background noise level. The fit determined that the detectors were limited by photon noise in the range 1\,pW to 20\,pW, which comfortably encompasses the expected BLAST-TNG flight loading on the detectors of $\sim$5 to $\sim$14\,pW. The fit to the noise model also produces an estimate of the optical coupling efficiency which was determined to be $\sim 70\%$.
The 250\,$\mu$m\ (1.2\,THz) array is designed to have a bandpass from 325 to 175\,$\mu$m\ (1.0 to 1.4\,THz) which is defined on the long wavelength end by the feedhorn profile and on the short wavelength end, by a low pass filter mounted in front of the feedhorn block. The bandpass was tested using a hot $1050^\circ$\,C thermal source coupled to a Fourier Transform Spectrometer (FTS) that filled the feedhorn beam. The tests determined that the edges of the bandpass for both polarization directions in a single pixel to be in agreement with the design of the filters and feedhorns\cite{Dober2015}.
The MKID's polarization performance has also been tested and is described in detail in Ref. \citenum{Dober2015}. The tests were performed with a source chopped between $1050^\circ$\,C and $20^\circ$\,C with a wire grid polarizer mounted between the source and the detectors. The polarizing grid is then rotated to produce a sinusoidal response at the detectors with orthogonal detectors 90 degrees out of phase. Fits to these measurements determined the cross-polar signal for both detector directions in a pixel to be $<$3\%, demonstrating the effectiveness of feedhorn coupled MKIDs for submillimeter polarimetry measurements.
Additional details of the MKID design and performance can be found in Ref. \citenum{McKenney2016}.
\subsection{Readout Electronics}
The BLAST-TNG readout is a highly multiplexed 1024 channel digital spectrometer which is the first of its kind to have been developed for the second generation Reconfigurable Open Architecture Computing Hardware (ROACH-2)\cite{Werthimer2011}. The ROACH-2 architecture is based on a field programmable gate array (FPGA), and is a product of the Collaboration for Astronomy Signal Processing and Electronics Research (CASPER). The ROACH is used in conjunction with a DAC/ADC board that was designed for ARCONS\cite{McHugh2012}. Building on the legacy of previous KID readouts created for the ROACH platform, the BLAST-TNG firmware performs coarse and fine channelization of 512\,MHz of RF bandwidth. The resulting channels are $\sim$100\,kHz wide and can be read out at rates of 200\,Hz to 500\,kHz. BLAST-TNG uses five ROACH-2 boards, three for the 250\,$\mu$m\ array and one each for the 350 and 500\,$\mu$m\ arrays, along with a suite of IF components and single board computers. The readout hardware is housed in the `ROACH-2 Motel', a custom enclosure designed for BLAST-TNG. Control software is written in Python, and will be ported to C for integration into the flight software. Details of the system’s overall performance are forthcoming.
\subsection{Polarimetry}
\label{sec:polarimetry}
\input{hwpr_v1}
The primary scanning strategy during flight will be a slow raster scan, which works well for extended sources. A typical BLAST-TNG raster will scan across targets in azimuth at a speed of approximately $ 0.5^\circ{\rm s}^{-1}$ with an elevation scan speed calculated to change the elevation by 1/3 the array FOV in one crossing of the target in azimuth. A Half Wave Plate (HWP) is used to modulate the polarization signal, so that each pixel samples {\it I}, {\it Q}, and {\it U} multiple times during the course of a scan which allows us to control for polarization systematics. The HWP is stepped between four set angles (0$^\circ$, 22.5$^\circ$, 45$^\circ$, and 67.5$^\circ$) after each completed scan of a source in elevation. The HWP used in BLASTPol\cite{Moncelsi2014} had a 10\,cm diameter aperture and was made from five layers of 500\,$\mu$m\ thick sapphire with an anti-reflective coating. However, the technology did not scale easily to larger apertures necessitating a new approach. The HWP developed for BLAST-TNG uses metal mesh filter technology\cite{Ade2006} to make a large diameter, $\sim$18\,cm, HWP (See Fig. \ref{fig:hwpr}). These types of HWP have demonstrated broad-band transmission at THz frequencies\cite{Pisano2012, Pisano2014}.
The detector rows are oriented parallel to the nominal scan direction with the MKID orientation rotated by 45 degrees in adjacent pixels along the scan direction. This alignment allows for sampling of both {\it Q} and {\it U} Stokes parameter on a timescale that is shorter than the array's common-mode $1/f$ noise, creating a redundant polarization modulation feature. The sampling timescale of the Stokes parameter is $\sim$0.025 s, which is determined by the detector separation, $\sim$45$^{\prime\prime}$\ at 250\,$\mu$m, and typical scan speed, $\sim 0.5^\circ{\rm s}^{-1}$.
\section{Conclusion}
\label{sec:conclusion}
BLASTPol successfully demonstrated the viability of balloon-borne telescopes to explore the polarization of the submillimeter sky and has led directly to the development of BLAST-TNG. Progress in the construction of the new instrument is on schedule and is proceeding rapidly. The cryogenics system has been thoroughly tested and is in a near flight ready status. Additionally, the detector technology is well understood and has undergone extensive testing at NIST prior to delivery to UPenn. The 250\,$\mu$m\ array has been installed in the cryostat and has been successfully incorporated into the readout architecture with observed optical sensitivity during cryogenic operations. We expect to demonstrate the detector's target sensitivity in the flight cryostat during the summer of 2016. The 350 and 500\,$\mu$m\ arrays and feedhorn blocks are being fabricated for testing and installation by fall 2016. The new BLAST-TNG instrument will give us magnetic field morphology maps at unprecedented resolution that span entire molecular cloud structures, allowing us to link the full-sky polarimetry maps of {\it Planck} with the high-resolution, small-area polarimetry maps, of telescopes such as ALMA. This will provide strong constraints on models of magnetic fields and turbulent interaction within magnetic fields as well as increasing our understanding of dust grain models.
\section{Acknowledgments}
BLAST is funded by NASA through grant number NNX13AE50G S08. Detector development is supported in part by NASA through NNH13ZDA001N-APRA. Brad Dober was funded by a NASA Earth and Space Science Fellowship NNX12AL58H S02. The BLAST-TNG collaboration would like to acknowledge the Xilinx University Program for their generous donation of five Virtex-6 FPGAs for use in our ROACH-2 readout electronics.
Peter Ashton was supported through Reach for the Stars, a GK-12 program supported by the National Science Foundation under grant DGE-0948017.
We would also like to thank the Columbia Scientic Balloon Facility (CSBF) staff for their continued outstanding work.
|
\section{Introduction}
Small planets are abundant in the galaxy, but the compositional diversity of small planets is not well understood. Theoretical models of planet formation predict that planets intermediate in size between Earth and Neptune could be gaseous ``mini-Neptunes,'' water worlds, or rocky ``Super-Earths'' \citep{kuchner_et_al2003, leger_et_al2004, valencia_et_al2006, seager_et_al2007, fortney_et_al2007, rogers_et_al2011, lopez_et_al2012, zeng+sasselov2013}. Recent studies have explored the compositional diversity of small planets using hierarchical Bayesian modeling of the observed planet radii and measured planet masses \citep{rogers2014} or theoretical models \citep{wolfgang+lopez2014}, but a thorough investigation of planet densities is hindered by the small number of small planets with well-measured masses and radii. There are currently only nine planets smaller than $2.7\rearth$ with masses measured to 20\% precision: 55 Cnc e \citep{gillon_et_al2012, nelson_et_al2014}, CoRoT-7b \citep{barros_et_al2014, haywood_et_al2014}, GJ1214b \citep{charbonneau_et_al2009}, HD97658b \citep{dragomir_et_al2013}, HIP116454b \citep{vanderburg_et_al2014}, Kepler-36b \citep{carter_et_al2012}, Kepler-78b \citep{pepe_et_al2013, howard_et_al2013}, and Kepler-10b and 10c \citep{dumusque_et_al2014}.
The host star Kepler-93 (KIC~3544595, KOI~69) is one of the brightest stars observed by \emph{Kepler} ($V = 10.2$, $Kp =9.93$), enabling very high precision photometry of 17~ppm on six-hour timescales \citep{christiansen_et_al2012}. \emph{Kepler} observed Kepler-93 throughout the baseline mission (Quarters 0--17) and conducted observations at short cadence (exposure time of 58.5~s) beginning in Quarter~2 and extending until the end of the mission. Due to the high photometric precision of the \mbox{Kepler-93} observations, the planet was detected in the first four months of \emph{Kepler} data \citep{borucki_et_al2011b}. \citet{marcy_et_al2014} acquired \nhires~Keck HIRES radial velocity observations of Kepler-93 from July 2009 - September 2012 and provided an estimate of $2.6 \pm 2.0 \mearth$ for the mass of Kepler-93b. \citet{marcy_et_al2014} also noted a large linear RV trend of $11.2 \pm 1.5$ m s$^{-1}$ yr$^{-1}$ and calculated lower limits on the mass and period of the perturbing companion of $M > 3M_{\rm Jup}$ and $P > 5$~yr. Incorporating an additional 14~spectra from the 2013 observing season, the HIRES mass estimate for Kepler-93b increased to $3.8\pm1.5\mearth$ \citep{ballard_et_al2014}. Nonetheless, the 40\% error on the mass measurement allows a wide range of planetary compositions including a rocky body, an ice world, and even a substantial primordial envelope of hydrogen and helium \citep{ballard_et_al2014}.
In contrast, the properties of the host star \mbox{Kepler-93} are well-constrained. Using 37~months of \emph{Kepler} short cadence data, \citet{ballard_et_al2014} conducted an asteroseismic investigation to characterize Kepler-93 in exquisite detail. They estimated an average stellar density of $1.652\pm0.006$~g~cm$^{-3}$, a stellar mass of $0.911\pm0.033\msun$, and a stellar radius of $0.919\pm0.011\rsun$. Adopting priors from their asteroseismic investigation, they fit the \emph{Kepler} photometry to obtain a precise radius estimate of $1.478\pm0.019\rearth$ for Kepler-93b.
In addition to characterizing the host star, \citet{ballard_et_al2014} present a variety of evidence that Kepler-93b is a bona fide planet rather than an astrophysical false positive. First, they report that the steep shape of the ingress and egress portions of the Kepler-93b light curve cannot be reproduced by a non-planetary companion. Second, they note that the infrared transit depth they measured with the \emph{Spitzer Space Telescope} is consistent with the planetary interpretation of Kepler-93b. Third, they place stringent limits on the presence of nearby stars based on Keck AO images \citep{marcy_et_al2014}. Fourth, they state that the stellar density derived from the transit duration \citep{seager+mallen-ornelas2003, nutzman_et_al2011} is consistent with the asteroseismic stellar density constraint, indicating that the planet likely orbits the target star rather than the companion causing the large RV trend.
In this paper we refine the mass measurement of Kepler-93b from $2.5\sigma$ to $6\sigma$ by analyzing two seasons of HARPS-N radial velocities in addition to the publicly available HIRES data. We discuss these observations and our data reduction methods in Section~\ref{sec:obs}. In Section~\ref{sec:rv}, we develop a model to fit the observed radial velocities. Finally, we discuss the implications of the resulting planet mass and present our conclusions in Section~\ref{sec:disc}.
\section{Observations \& Data Reduction}
\label{sec:obs}
We obtained \nharpsn~spectra of Kepler-93 using the HARPS-N spectrograph on the 3.57-m Telescopio Nazionale Galileo (TNG) at the Observatorio del Roque de los Muchachos. HARPS-N is a high-precision, vacuum-stabilized, high-resolution ($R \simeq 115,000$) echelle spectrograph. The design is very similar to the design of the original HARPS instrument at the ESO 3.6-m \citep{mayor_et_al2003}. The main differences are that HARPS-N is fed by octagonal fibers rather than circular fibers to improve the scrambling of the light and features a monolithic 4096 x 4096 CCD instead of the dual CCD configuration used for the HARPS focal plane \citep{consentino_et_al2012}.
We acquired \nthirteen and \nfourteen~HARPS-N observations of Kepler-93 during the 2013 and 2014 observing seasons, respectively. In most cases, we used an exposure time of 30~minutes and achieved a mean S/N per extracted pixel of 103 at 550nm. (Four of the spectra had an exposure time of 15~minutes and one had 27~minutes; these were all gathered in July 2013.) One of the observations collected in 2013 was contaminated by light from a mercury lamp and was therefore removed from the analysis. The final HARPS-N dataset analyzed in this paper consists of \nharpsn~spectra. In most cases (75 of \nharpsn~spectra), we observed Kepler-93 using simultaneous thorium argon (observing mode {\tt HARPN\_ech\_obs\_thosimult}). The remaining eleven observations were obtained without simultaneous thorium argon in observing mode {\tt HARPN\_ech\_obs\_objAB}.
\begin{deluxetable*}{ccccccc}
\tablecolumns{7}
\tablecaption{HARPS-N Radial Velocity Observations of Kepler-93}
\tablehead{
\colhead{BJD$_{\rm UTC}$} &
\multicolumn{2}{c}{RV (m/s)} &
\colhead{Bisector} &
\multicolumn{2}{c}{log($R'_{HK}$) (dex)} &
\colhead{$t_{\rm exp}$} \\[0.2em]
\cline{2-3}
\cline{5-6}\\[-0.8em]
\colhead{-2450000} &
\colhead{Value} &
\colhead{Error } &
\colhead{(m/s)} &
\colhead{Value} &
\colhead{Error } &
\colhead{(s)}
}
2456462.686262 & 27335.24 & 1.02 & -28.39 & -5.01 & 0.01 & 1800 \\
2456463.584483 & 27337.75 & 0.94 & -31.63 & -5.00 & 0.01 & 1800 \\
2456464.609617 & 27331.57 & 1.75 & -27.24 & -5.04 & 0.02 & 1800 \\
2456465.606438 & 27342.34 & 1.00 & -32.74 & -5.02 & 0.01 & 1800 \\
2456466.608850 & 27337.58 & 0.86 & -29.26 & -5.00 & 0.01 & 1800 \\
$...$ & $...$ & $...$ & $...$ & $...$ & $...$ & $...$
\enddata
\tablenotetext{}{(This table is available in its entirety in a machine-readable form in the online journal. A portion is shown here for guidance regarding its form and content.)}
\label{tab:rvs}
\end{deluxetable*}
We reduced the data with the standard HARPS-N pipeline by cross-correlating the observed spectra with a numerical mask based on the spectrum of a G2V~star \citep{baranne_et_al1996, pepe_et_al2002}. We provide the resulting RVs and their $1\sigma$ errors in Table~\ref{tab:rvs} along with the observation BJDs, exposure times, bisector spans, and stellar activity levels as measured by the Ca II $\log(R'_{\rm HK}$) activity indicator \citep{noyes_et_al1984}. The BJDs in Table~\ref{tab:rvs} are provided in UTC, but we converted the times to TDB (the units used by the \emph{Kepler} mission) using the IDL routine {\tt utc2bjd.pro}\footnote{\url{http://astroutils.astronomy.ohio-state.edu/time/pro/utc2bjd.pro}} prior to fitting the RVs. We did not find evidence for a correlation between RV and bisector span or $\log(R'_{\rm HK})$.
\section{Analysis of the Radial Velocity Data}
\label{sec:rv}
Our full data set included RVs from four~seasons of HIRES observations (2009 July -- 2012 September) and two seasons of HARPS-N observations (2013 June -- 2014 October). We fit the combined HARPS-N and HIRES data set by incorporating a single offset $RV_{\rm off}$ between the HIRES and HARPS-N data. We used the following general model:
\begin{equation}
\begin{split}
\mathcal{M}(t_i)& = \gamma + \rm{RV}_{\rm off} + \beta(t_i) \\
& \quad +K \left[\cos(\theta(t_i, T_C, P, e) + \omega) + e \cos{\omega}\right]
\end{split}
\end{equation}
where $\gamma$ is the systemic velocity of Kepler-93, \mbox{$RV_{\rm off} = RV_{\rm HARPS-N}-RV_{\rm HIRES}$} is the offset between the HIRES and {HARPS-N} RVs, $\beta(t_i)$ is a long-term RV trend due to a third component in the system, $K$ is the semi-amplitude due to Kepler-93b, and $\omega$ is the argument of periastron. The function $\theta$ is the true anomaly of Kepler-93b at time $t_i$ and depends on the period $P$, epoch of transit $T_C$, eccentricity $e$. When fitting eccentric orbits, we used IDL routine {\tt keplereq.pro} written by Brian Jackson\footnote{\url{http://www.lpl.arizona.edu/~bjackson/idl_code/keplereq.pro}} to solve Kepler's equation for the eccentric anomaly. The routine uses the method suggested by \citet{mikkola1987} as an initial guess.
We considered linear and quadratic parameterizations of the long-term trend $\beta(t_i)$ and circular and eccentric orbits for Kepler-93b. For all models, we determined an initial solution using the Levenberg-Marquardt minimization algorithm as implemented by {\tt lmfit} in IDL. We then explored the region of parameter-space near the best-fit solution using a Bayesian Markov Chain Monte Carlo analysis with a Metropolis-Hastings acceptance criterion \citep{metropolis_et_al1953}. We initialized $N$ chains, where $N$ was twice the number of free parameters in the chosen model. We selected different initial positions for each chain by perturbing each free parameter of the best-fit solution by a random number drawn from a distribution with a width of five times the step size. We tuned the step sizes such that the acceptance fractions for each parameter were 10--30\%. For the MCMC analysis, we set uniform priors for all parameters except the orbital period and epoch of transit. We allowed only non-negative values for the RV semi-amplitude $K$ and the separate stellar jitter terms $\sigma_{sj}$ for the HIRES and HARPS-N observations (see below).
The \emph{Kepler} photometry places tight constraints on the period and epoch of transit \citep{ballard_et_al2014}. We incorporated this knowledge into our MCMC analysis by including Gaussian priors on period and transit epoch in the likelihood calculation. As shown in \citet{dumusque_et_al2014}, using the tight prior from Kepler photometry when fitting a circular model to RV observations of a planet in a circular orbit yields a result very similar to that from a combined photometric and spectroscopic fit. We also tested fitting the data while allowing the epoch of transit to float and find the epoch of transit at \mbox{BJD = 2454944.29514}. This epoch differs from the value determined by \citet{ballard_et_al2014} by 4 minutes ($0.3\sigma$). The possible shift in the transit center is therefore insignificant. Accordingly, we adopt the photometric ephemeris determined by \citet{ballard_et_al2014}.
In our calculations, we shifted the epoch of transit close to the start of the HARPS-N RVs to reduce error propagation. We increased the efficiency of our model fits by parameterizing eccentric models using $\sqrt{e} \cos(\omega)$ and $\sqrt{e} \sin(\omega)$ rather than varying $e$ and $\omega$ directly \citep{ford2006, eastman_et_al2013}. As in \citet{dumusque_et_al2014}, we accounted for stellar activity by incorporating a stellar jitter term $\sigma_{sj}$ in our adopted likelihood $\mathcal{L}$:
\begin{equation}
\mathcal{L}= \prod_{i=1}^{N}\left({\frac{1}{\sqrt{2 \pi (\sigma_i^2 + \sigma_{sj}^2)}} \exp \left[ - \frac{(RV(t_i) - \mathcal{M}(t_i))^2}{2 (\sigma_i^2+\sigma_{sj}^2)}\right] }\right)
\end{equation}
where $RV(t_i)$ is the measured RV at each time $t_i$ in the set of $N$ observations, $\mathcal{M}$ is the model, $\sigma_i$ is the instrumental noise listed in Table~\ref{tab:rvs}, and the stellar jitter noise $\sigma_{sj}$ is allowed to adopt a different constant value for the HARPS-N and HIRES data.
We ran each chain for a minimum of $10^4$ steps and checked for convergence by computing the Gelman-Rubin potential scale reduction factor $\hat{R}$ for each parameter \citep{gelman_et_al2004}. We stopped the MCMC analysis when $\hat{R}< 1.03$ for all parameters. Next, we accounted for ``burn-in'' by identifying the point in each chain at which the likelihood first became higher than the median likelihood of the chain and removing all earlier steps. After combining all of the chains, we selected the median values of each parameter as the best-fit value and assigned symmetric errors encompassing 68\% of values closest to the adopted best-fit value.
We then used Bayesian statistics to determine which of the models considered best describes the data. We followed the method of \citet{chib+jeliazkov2001} as described in the Appendix of \citet{haywood_et_al2014} to calculate the Bayes factor between pairs of models using the posterior distributions and acceptance probabilities from our MCMC analyses. This method was previously used by \citet{dumusque_et_al2014} to compare RV models of the Kepler-10 system. We found that penalties incurred by the additional complexity of fitting the orbit of Kepler-93b with an eccentric model or fitting the long-term trend with a quadratic model outweighed the improvement in the likelihood. We also compared the models by holding the stellar jitter terms fixed to $\sigma_{sj, {\rm HARPS-N}}=1.56$~m~s$^{-1}$ and $\sigma_{sj, {\rm HIRES}}=2.03$~m~s$^{-1}$
and computing the Bayesian Information Criterion \citep[BIC, ][]{schwarz1978} and finite sample Akaike Information Criterion \citep[AIC$\rm{_C}$,][]{hurvich+tsai1989}. When considering only the HARPS-N data, we found that the model with a quadratic trend and a circular orbit for Kepler-93b was preferred over the models with a linear trend and circular orbit (\mbox{$\Delta$BIC = 6.1}, \mbox{$\Delta$AIC$\rm{_C}$ = 8.3}), linear trend and eccentric orbit (\mbox{$\Delta$BIC = 12.4}, \mbox{$\Delta$AIC$\rm{_C}$ = 10.4}), or quadratic trend and eccentric orbit (\mbox{$\Delta$BIC = 6.6}, \mbox{$\Delta$AIC$\rm{_C}$ = 2.5).}
Nonetheless, when we included the HIRES data we found that the simplest model (linear trend and circular orbit) was preferred over the models with a linear trend and eccentric orbit (\mbox{$\Delta$BIC = 5.6}, \mbox{$\Delta$AIC$\rm{_C}$ = 0.7}), quadratic trend and circular orbit (\mbox{$\Delta$BIC = 4.7}, \mbox{$\Delta$AIC$\rm{_C}$ = 2.1}), or quadratic trend and eccentric orbit (\mbox{$\Delta$BIC = 10.5}, \mbox{$\Delta$AIC$\rm{_C}$ = 3.0}). We therefore treat the perturbation from Kepler-93c as a linear trend and model the orbit of Kepler-93b as circular. (For the eccentric fits, we found a median eccentricity of 0.15 and an upper limit of $e < 0.31$ with 95\% confidence.) We present the resulting system properties including a mass estimate for \mbox{Kepler-93b} of \massest in Table~\ref{tab:params} and display the measured RVs and the best-fit model in Figure~\ref{fig:rv}. As highlighted in Figure~\ref{fig:resid}, the HARPS-N residuals are gaussian with a distribution centered on zero and containing 68\% of the data within a half width 1.6~m~s$^{-1}$. For the HIRES residuals the region encompassing 68\% of the data has a half width of 3.4~m~s$^{-1}$.
\begin{deluxetable}{lcc}
\tablecolumns{3}
\tabletypesize{\footnotesize}
\tablecaption{Parameters for the Kepler-93 System}
\tablehead{
\colhead{} &
\colhead{Value and } &
\colhead{} \\
\colhead{Parameter} &
\colhead{$1\sigma$ Errors} &
\colhead{Ref.}
}
\multicolumn{3}{l}{\textbf{Kepler-93 (star) = KIC 3544595 = KOI 69}}\\
\hline\\[-0.7em]
Right ascension & $19^{\rm h}25^{\rm m}40.^{\rm s}39$ & 1,2 \\[0.15em]
Declination & $+38^{\rm d}40^{\rm m}20.^{\rm s}45$ & 1,2 \\[0.15em]
Kepler magnitude & 9.931 & 3 \\[0.15em]
2MASS K & 8.370 & 3 \\[0.15em]
$T_{\rm eff}$ (K) & $5669 \pm 75$ & 4 \\[0.15em]
$R_*$ (solar radii) & $0.919 \pm 0.011$ & 4 \\[0.15em]
$M_*$ (solar masses) & $0.911\pm 0.033$ & 4 \\[0.15em]
$[Fe/H]$ & $-0.18 \pm 0.10$ & 4 \\[0.15em]
$\log g$ & $4.470 \pm 0.004$ & 4 \\[0.15em]
Age (Gyr) & $6.6 \pm 0.9$ & 4 \\[0.15em]
Systemic Velocity\tablenotemark{a} (m s$^{-1}$) & $27337.89 \pm 0.51$ & 5 \\[0.15em]
HIRES Offset (m s$^{-1}$) & $27304.1 \pm 1.5$ & 5 \\[0.15em]
RV Jitter (HARPS-N) & $1.58 \pm 0.19$ & 5 \\[0.15em]
RV Jitter (HIRES) & $2.09 \pm 0.71$ & 5 \\[0.15em]
\hline\\[-0.7em]
\multicolumn{3}{l}{\textbf{Kepler-93b (planet) = KOI 69.01}}\\
\hline\\[-0.7em]
\multicolumn{3}{l}{\emph{Transit and orbital parameters} }\\
Orbital period $P$ (days) & $ 4.72673978 \pm 9.7 \times 10^{-7}$ & 4 \\[0.15em]
Transit epoch $T_C$ (BJD) & $2454944.29227 \pm 0.00013$ & 4 \\[0.15em]
$R_p/R_*$ & $0.014751 \pm 0.000059$ & 4 \\[0.15em]
$a/R_*$ & $12.496 \pm 0.015$ & 4 \\[0.15em]
Inc (deg) & $89.183 \pm 0.044$ & 4 \\[0.15em]
Impact parameter & $0.1765 \pm 0.0095$ & 4 \\[0.15em]
Orbital eccentricity $e$ & 0 (fixed) & 5\\[0.15em]
RV semi-amplitude $K$ (m s$^{-1}$) & $1.63 \pm 0.27$ & 5\\[0.15em]
\hline \\[-0.7em]
\multicolumn{3}{l}{\emph{Planetary Parameters}} \\
$R_p (\rearth)$ & $1.478 \pm 0.019$ & 4 \\[0.15em]
$M_p (\mearth)$ & $4.02 \pm 0.68$ & 5 \\[0.15em]
$\rho_p$(g cm$^{-3}$) & $6.88 \pm 1.18$ & 5 \\[0.15em]
$\log g_p$ (cgs) & $3.26 \pm 0.07$ & 5 \\[0.15em]
$a$ (AU) & $0.053 \pm 0.002$ & 5 \\[0.15em]
$T_{\rm eq}$ (K)\tablenotemark{b} & $1037 \pm 13$ & 4 \\[0.15em]
\hline \\[-0.7em]
\textbf{Kepler-93c (companion)} & \\
\multicolumn{3}{l}{\emph{Fit Parameters}} \\
Acceleration (m s$^{-1}$ yr$^{-1}$) & $12.0 \pm 0.4$ & 5\\[0.15em]
\hline \\[-0.7em]
\multicolumn{3}{l}{\emph{Companion Limits}}\\
Mass ($M_J$) & $> 8.5$ & 5 \\
Orbital period $P$ (yr) & $ > 10$ & 5
\enddata
\label{tab:params}
\tablenotetext{}{\textbf{References:} (1)~\citet{hog_et_al1998}, (2)~\citet{hog_et_al2000}, (3)~\citet{brown_et_al2011}, (4)~\citet{ballard_et_al2014}, (5)~This Paper.}
\tablenotetext{a}{Systemic velocity at BJD 2456461.57573945.}
\tablenotetext{b}{Assuming a Bond albeo of 0.3.}
\end{deluxetable}
\begin{figure*}[tbp]
\begin{center}
\centering
\includegraphics[width=0.49\textwidth]{best_timeall_lin.eps}
\includegraphics[width=0.49\textwidth]{best_phasedall_lin.eps}
\end{center}
\caption{Best-fit model (black) for the Kepler-93 system and measured HIRES (light blue) and HARPS-N (dark blue) RVs after correcting for the offset between HIRES and HARPS-N. The errors include contributions from both instrumental noise and stellar jitter. \emph{Top Left:} Measured RVs versus time after removing the signal of the planet \mbox{Kepler-93b}. \emph{Bottom Left: } RV residuals versus time after removing the full planet+trend fit. \emph{Top Right: } Phase-folded signal of Kepler-93b after removing the long-term trend due to Kepler-93c. The large red circles with error bars show the weighted mean and corresponding uncertainties of the measured RVs, conveniently binned to equal arbitrary intervals in phase. The points shown in gray are repeated to better reveal the behavior of the data near phase=0. \emph{Bottom Right: } RV residuals versus phase after removing the full planet+trend fit. The red circles are the binned data.}
\label{fig:rv}
\end{figure*}
\begin{figure}[tbp]
\begin{center}
\centering
\includegraphics[width=0.5\textwidth]{best_residall_lin.eps}
\end{center}
\caption{Histogram of the residuals of the HIRES (light blue) and HARPS-N (dark blue) observations. The dashed lines mark the median of each distribution (-0.16 m/s for HIRES, 0.002 m/s for HARPS-N) and the dot-dash lines encompass 68\% of the measurements. The half width of the 68\% interval is 1.6~m~s$^{-1}$ for the HARPS-N data and 3.4~m~s$^{-1}$ for the HIRES data.}
\label{fig:resid}
\end{figure}
\begin{figure}[tbp]
\begin{center}
\centering
\includegraphics[width=0.5\textwidth]{ao_rv_lim.eps}
\end{center}
\caption{Limits on the mass and separation of the companion Kepler-93c. The Keck AO observations exclude a companion within the blue region. The combined HIRES and HARPS-N RVs exclude the teal region due to the amplitude of the trend and the maroon region due to the baseline of the observations. Kepler-93c is therefore constrained to lie within the white region. The dashed purple line divides substellar and stellar companions. These limits assume that the companion has an orbit with $i = 90^\circ$ and $e = 0$. }
\label{fig:93c}
\end{figure}
\begin{figure*}[tbp]
\begin{center}
\centering
\includegraphics[width=1\textwidth]{mrplot.eps}
\end{center}
\caption{Mass-radius diagram for planets smaller than $2.7\rearth$ with masses measured to better than 20\% precision. The shaded gray region in the lower right indicates planets with iron content exceeding the maximum value predicted from models of collisional stripping \citep{marcus_et_al2010}. The solid lines are theoretical mass-radius curves \citep{zeng+sasselov2013} for planets with compositions of 100\% H$_2$O (blue), 25\% MgSiO$_3$ -- 75\% H$_2$O (purple), 50\% MgSiO$_3$ -- 50\% H$_2$O (green), 100\% MgSiO$_3$ (black), 50\% Fe -- 50\% MgSiO$_3$ (red), and 100\% Fe (orange). Our best-fit relation based on the \citet{zeng+sasselov2013} models is the dashed light blue line representing an Earth-like composition (modeled as 17\% iron and 83\% magnesium silicate using a fully-differentiated, two-component model). The shaded region surrounding the line indicates the 2\% dispersion in radius expected from variation in Mg/Si and Fe/Si ratios \citep{grasset_et_al2009}.}
\label{fig:comp}
\end{figure*}
The expected circularization timescale for Kepler-93b is significantly shorter than the $6.6\pm0.9$~Gyr age of the star \citep{ballard_et_al2014}. Following \citet{goldreich+soter1966}, we calculated a tidal circularization timescale of \tcircest for a $4.02~\mearth$, $1.48~\rearth$ planet in an orbit with $a = 0.053~AU$ around a $0.91\msun$~star. We assumed $Q = 100$ based on the tidal quality factors estimated for terrestrial planets in the Solar System \citep{yoder1995, henning_et_al2009}. Obtaining a tidal circularization timescale similar to the age of the system would require $Q = 9000$, comparable to the estimate for Neptune \citep{zhang+hamilton2008}. Although the tidal circularization argument is consistent with the preference for a circular orbit, we caution that the tidal quality factors for exoplanets are largely unknown.
\subsection{Limits on the Properties of Kepler-93c}
The baseline of our RV data is too short to measure the period and minimum mass of the perturber responsible for the long-term trend, but we can place lower limits on the companion properties. \citet{wang_et_al2014c} conducted a similar analysis of the properties of \mbox{Kepler-93c} based on AO observations and Keck/HIRES RVs. They found a linear RV trend of $12.2 \pm 0.2$~m~s$^{-1}$~yr$^{-1}$ and argued that Kepler-93c is most likely to have a mass below $101~M_J$ and a semi major axis $a = 15.5 - 33$~AU if it is a stellar companion. For the substellar case, they found limits of $a = 5.5 - 27.6$~AU and $M = 10 - 80 M_J$.
Our additional two years of HARPS-N observations have allowed us to further restrict the allowed parameter space for Kepler-93c. We measured a linear trend of $12.0 \pm 0.4$ m~s$^{-1}$~yr$^{-1}$ for 5~years, implying that \mbox{Kepler-93c} has $P > 10$~yr and $M > 8.5~M_J$. Assuming the 100~pc distance to Kepler-93 estimated by \citet{ballard_et_al2014}, the resulting semimajor axis $a > 4.5$~AU corresponds to an angular separation of $0\farcs045$. At this separation, the detection limit from Keck AO imaging is 1.7~\emph{Ks} magnitudes fainter than Kepler-93. We can therefore place an upper limit of $Ks >10.1$ on Kepler-93c unless Kepler-93c happened to have an orbital geometry precluding detection at the epoch of the Keck observations. Converting the $Ks$ upper limit into a mass limit via the \citet{delfosse_et_al2000} relation\footnote{The Delfosse relation predicts stellar mass from $Ks_{\rm CIT}$ whereas the Keck observations were acquired in $Ks_{\rm 2MASS}$. We converted between the two systems assuming a color of $J-K = 1$ and using the color-dependent conversions provided at \url{http://www.astro.caltech.edu/~jmc/2mass/v3/transformations/}.} and the distance, we found a mass upper limit of $0.64 \msun$ for angular separations beyond $0\farcs045$. We display the combined limits from the AO and RV data in Figure~\ref{fig:93c}. In the future, astrometric measurements from Gaia \citep{perryman_et_al2001} will likely provide additional constraints on the properties of the Kepler-93 system. We will then be able to investigate the dynamical history of the system and test whether Kepler-93c might be responsible for scattering Kepler-93b inward onto a short-period orbit.
\section{Discussion and Conclusions}
\label{sec:disc}
Combining our estimate of \massest for the mass of Kepler-93 with the radius estimate of $Rp = 1.478 \pm 0.019 \rearth$ from \citet{ballard_et_al2014}, we find a density of \rhoest. In Figure~\ref{fig:comp}, we show \mbox{Kepler-93b} on the mass-radius diagram. In this diagram we plot only those planets smaller than $2.7\rearth$ and with masses determined to a precision better than 20\%. In addition to Venus and the Earth, there are ten such planets. We observe that Kepler-93b falls in a cluster of planets with radii 50\% larger than that of the Earth, all of which have extremely similar densities: \mbox{Kepler-10b} \citep[$\rho = 5.8 \pm 0.8$~g~cm$^{-3}$;][]{dumusque_et_al2014}, \mbox{Kepler-36b} \citep[$\rho = 7.46^{+0.74}_{-0.59}$~g~cm$^{-3}$;][]{carter_et_al2012}, and CoRoT-7b \citep[$\rho = 6.56 \pm 1.40$~g~cm$^{-3}$;][]{barros_et_al2014, haywood_et_al2014}. This cluster falls upon a relation that includes Earth, Venus, and Kepler-78b \citep{howard_et_al2013, pepe_et_al2013}, which is itself only 20\% larger than the Earth. To investigate this further, we used the two-component iron-magnesium silicate models of \citet{zeng+sasselov2013} to see if we could find a single composition that explained these seven worlds.
For the solar system planets, we artificially include mass and radius errors equal to the mean fractional errors for the exoplanets considered so that they do not have undue influence on the resulting fit. We find the lowest $\chi^2$ for a model composition of 83\%~MgSiO$_3$ and 17\%~Fe. We arrive at the same best-fit relation when we exclude Earth and Venus. We caution that the two-component models used in this analysis make two simplifying approximations about the interior structure of planets that cause the core mass fraction to be underestimated: (1) the core contains only iron and the mantle contains only magnesium silicate and (2) the planet is completely dry with no water content. Accordingly, we expect the actual core mass fraction to be slightly higher by 5-8\% to account for incorporation of lighter elements like oxygen, sulfur, and silicon in the core and the inclusion of water in the mantle. In addition, there could be a change of roughly 2\% towards higher or lower core fractions due to uncertainties in the equations of state used in the model calculations. Our purpose in this exercise is to test whether we can find one composition that successfully explains all seven planets, not to place stringent constraints on the abundance of magnesium silicate or iron.
Intriguingly, all of these planets, which are smaller than $1.6\rearth$, have a tight dispersion around this best-fit compositional curve, suggesting that the distribution of small planet compositions has low intrinsic scatter. In the solar system, the strong agreement between abundance ratios of elements in meteorites and those of the solar photosphere \citep{lodders2003} is a key constraint by which we deduce the composition of the interior of the Earth. Therefore, we might look to the bulk abundances of exoplanet host stars for similar constraints on the interior compositions of their terrestrial planets. \citet{grasset_et_al2009} use a set of planetary models to investigate the dependence of planet radii on elemental abundances. Varying the ratios of iron to silicate and magnesium to silicate within the range observed for the photospheric abundances of nearby exoplanet host stars \citep{beirao_et_al2005, gilli_et_al2006}, \citet{grasset_et_al2009} predicted that the radii of terrestrial planets would vary by roughly 2\% at a given mass. Our findings are in agreement with this picture: We measure a mean absolute deviation of 1.9\% between the estimated planet radii and the values predicted by a 83\%~MgSiO$_3$/17\%~Fe model for planets less massive than $6\mearth$. Indeed, rocky planets very close to their host stars seem to obey a well-defined relationship between radius and mass, although with only 5 such examples outside the Solar system, the immediate task is to characterize other terrestrial exoplanets with similar precision. Increasing the sample of small planets with well-constrained masses and radii will allow us to learn whether additional rocky planets could also be explained by a single mass-radius relation and investigate whether the relation found for close-in planets extends to planets in more distant orbits.
Our mass-radius diagram also includes five planets more massive than $6\mearth$: 55 Cnc e, GJ1214b, HD97658b, HIP 116454b, and Kepler-10c. In contrast, none of these more massive planets have a high density consistent with the best-fit magnesium silicate/iron composition described above. In agreement with \citet{rogers2014}, we find that planets larger than approximately $1.6\rearth$ (e.g., more massive than approximately $6\mearth$) contain significant fractions of volatiles or H/He gas. These planets appear to have a diversity of compositions that is not well-explained by a single mass-radius relation \citep{wolfgang+lopez2014}.
The discussion above focused exclusively on planets smaller than $2.7\rearth$ with masses measured to better than 20\%. Some low-mass worlds with very low densities are known, notably the Kepler-11 system \citep{lissauer_et_al2013} and KOI-314c \citep{kipping_et_al2014b}. Thus we are not proposing that all planets less massive than $6\mearth$ obey a single mass-radius relation; rather, we suggest that the rocky analogs of the Earth might do so.
\acknowledgments
The HARPS-N project was funded by the Prodex Program of the Swiss Space Office (SSO), the Harvard- University Origin of Life Initiative (HUOLI), the Scottish Universities Physics Alliance (SUPA), the University of Geneva, the Smithsonian Astrophysical Observatory (SAO), and the Italian National Astrophysical Institute (INAF), University of St. Andrews, Queen's University Belfast and University of Edinburgh. The research leading to these results has received funding from the European Union Seventh Framework Programme (FP7/2007-2013) under Grant Agreement No. 313014 (ETAEARTH).
C. D. is supported by a National Science Foundation Graduate Research Fellowship. X. D. would like to thank the Swiss National Science Foundation (SNSF) for its support through an Early Postdoc Mobility fellowship. P.~F. acknowledges support by Funda\c{c}\~ao para a Ci\^encia e a Tecnologia (FCT) through Investigador FCT contracts of reference IF/01037/2013 and POPH/FSE (EC) by FEDER funding through the program ``Programa Operacional de Factores de Competitividade - COMPETE''. This publication was made possible through the support of a grant from the John Templeton Foundation. The opinions expressed in this publication are those of the authors and do not necessarily reflect the views of the John Templeton Foundation.
|
\section{Introduction}
In this paper we present a solution of the Einstein equations for what we call a {\em "hybrid" black hole}. This solution is obtained by proper gluing of the external Kerr metric and the internal Weyl metric. These metrics are glued along the joint horizon surface, where a lightlike thin massive shell is located. The Weyl metric describes an interior of a non-rotating black hole distorted by the presence of the shell. At the same time the shell has an angular momentum which is the origin of the `dragging into the rotation' of the external spacetime, described the Kerr metric. The event horizon is a special null limit of rigidly rotating ZAMO surfaces in the Kerr geometry, that were discussed in the recent paper by the authors \cite{FroFro}.
To glue the Kerr and Weyl metrics along a common horizon we used a general approach by Barrabes and Israel presented in their remarkable paper \cite{BI}. We shall describe this formalism adopted to our problem later. Here we just make a few remarks which might be useful for better understanding of this approach. We consider two different regions, one located outside the horizon and the other inside it, as two distinct spacetime manifolds $\MC{M}_-$ and $\MC{M}_+$. We call them Kerr and Weyl domains, respectively. Their metrics are ${g}^-_{\mu\nu}$ and ${g}^+_{\mu\nu}$, $(\mu,\nu=0,1,2,3)$, and we denote the independent coordinates in these domains by $x_-^{\mu}$ and $x_+^{\mu}$. The manifolds $\MC{M}_-$ and $\MC{M}_+$ have boundaries $\Sigma_-$ and $\Sigma_+$, respectively. They are null surfaces with respect to the corresponding metrics in the domains. We obtain a single spacetime manifold $\MC{M}=\MC{M}_-\cup\MC{M}_+$ by gluing $\MC{M}_-$ and $\MC{M}_+$ along their boundaries, that is by making the natural identification $\Sigma_-=\Sigma_+=\Sigma$.
The Kerr (Weyl) domain $\MC{M}_-$ ($\MC{M}_+$) lies to the past (future) with respect to the common horizon $\Sigma$.
Denote by $y^a$, $a=1,2,3$, the internal coordinates on $\Sigma$. The induced metrics on $\Sigma_{\pm}$,
\be
h^{\pm}_{ab}={\pa x_{\pm}^{\mu}\over \pa y^a}{\pa x_{\pm}^{\nu}\over \pa y^b} g^{\pm}_{\mu\nu}\, ,
\ee
must be isometric. We denote this metric on $\Sigma$ by $h_{ab}$.
Because the surface $\Sigma$ of the joint horizon is null, the three-metric ${h}_{ab}$ is degenerate. The vector $\BM{n}$ of the normal to $\Sigma$ is null, and hence it is tangent to the horizon, so that $h_{ab}n^a n^b=0$. Integral lines of this vector field are null geodesics which are the generators of $\Sigma$. In a general case, there exists an ambiguity in the choice of the parametrization of these lines, so that when one glues $\Sigma_+$ and $\Sigma_-$ one must guarantee that these parametrizations are chosen to be the same. There exist a pair of commuting Killing vectors in each of the domains, and a special linear combinations of them, with constant coefficients, are null on $\Sigma_{\pm}$. In this paper we use what is called "static" soldering, defined by identification of the corresponding advanced time coordinates over $\Sigma_-$ and $\Sigma_+$. For this choice of the soldering the parameters of the null shell at the horizon are time independent (for more details, see \cite{BI}).
The metric $g_{\mu\nu}$ is continuous at the horizon in properly chosen coordinates. However, it derivatives have jumps. The formalism developed in \cite{BI} allows one to relate these jumps to a specially chosen massive thin null shell. In this paper we demonstrate that the external (Kerr) metric can be glued with the internal (Weyl) metric along the common event horizon and calculate the parameters of the corresponding null shell.
The paper is organized as follows. In Sections~II and III we study the near horizon geometry in the Kerr and Weyl domains, respectively. In Section~IV we describe the procedure of gluing these metrics along the horizon and calculate the parameters of the massive null shell spread over the horizon. Section~V contains discussion of the obtained results and their possible generalizations.
We use units in which $G=c=1$ and the sign convention adopted in the book \cite{MTW}.
\section{Kerr domain}
\subsection{Kerr metric}
The Kerr solution depends on two arbitrary constants, mass $M$ and a rotation parameter $a=J/M\le M$, where $J$ is the angular momentum of the rotating black hole. The metric $ds_-^2$ in the Boyer-Lindquist coordinates $(t,r,\theta,\phi)$ is (see e.g., \cite{MTW,FN})
\ba\n{mds}
&&ds_-^2=d\Gamma^2+d\gamma^2\, \\
&&d\Gamma^2=A dt^2+2B dt d\phi +C d\phi^2\, ,\\
&&d\gamma^2=\Sigma\left( {dr^2\over \Delta}+ d\theta^2\right)\, ,\\
&&A=-\left(1-{2Mr\over \Sigma}\right)\, ,\
B=-{2Mra\sin^2\theta\over \Sigma}\, ,\\
&&C={P\sin^2\theta\over \Sigma}\hh P=(r^2+a^2)^2-\Delta a^2 \sin^2\theta\, ,\\
&&\Delta=r^2-2Mr+a^2\hhh \Sigma=r^2+a^2\cos^2\theta\, .
\ea
The coordinate $t$ takes values in the interval $(-\infty,\infty)$, and $\phi$ is a periodic coordinate with the period $2\pi$. The (commuting) Killing vectors for the metric \eq{mds} are
\be
\BM{\xi}_t=\partial_t\hh \BM{\xi}_{\phi}=\partial_{\phi}\, .
\ee
These vectors are uniquely specified by the following properties: (i) $\BM{\xi}_t$ is timelike and normalized to one at the infinity, and (ii) the integral lines of $\BM{\xi}_{\phi}$ are closed circles.
The Kerr metric is also invariant under the reflections
\be
(t,\phi)\to (-t,-\phi)\, , \mbox{and } \theta\to\pi-\theta\, .
\ee
For $a<1$ the equation $\Delta=0$ has two roofs
\be
r_{\pm}=M\pm\sqrt{M^2-a^2}\, .
\ee
The larger one, $r_+$, determines a position of the event horizon. The angular velocity of the horizon is
\be
\Omega={a\over r_+^2+a^2}={a\over 2Mr_{+}}\, .
\ee
A linear combination of the Killing vectors (which itself is a Killing vector)
\be\n{KIL}
\BM{\eta}=\BM{\xi}_t+\Omega \BM{\xi}_{\phi}
\ee
is a null generator of the event horizon.
\subsection{Horizon metric}
The Boyer-Lindquist coordinates become singular at the horizon. Because we are interested in a near horizon geometry it is convenient to introduce so called {\em Kerr ingoing coordinates} that are regular at the (future) event horizon. For this purpose we make the following coordinate transformation
\be
dv=dt+dr_*\, ,\
d\tilde{\phi}=d\phi+a{dr\over \Delta}\, , \
dr_*=(r^2+a^2){dr\over \Delta}\, .
\ee
The Kerr metric in these coordinates $(v,r,\theta,\tilde{\phi})$ takes the form
\ba\n{KI}
ds_-^2&=&-{\Delta\over \Sigma}\omega_1^2+2 dr\, \omega_1 +\Sigma d\theta^2 +{\sin^2\theta\over \Sigma}\omega_2^2\, ,\\
\omega_1&=&dv -a\sin^2\theta d\tilde{\phi}\, ,\\
\omega_2&=& (r^2+a^2)d\tilde{\phi} - a dv\, .
\ea
The metric on the horizon can be obtained from \eq{KI} if we impose a constraint $r=r_+$. This metric is\footnote{Here and later we use notation $\hor$ to stress that a corresponding relation is valid only on the horizon $H$.}
\be\n{MH}
dh_-^2\hor\Sigma_+ d\theta^2 +{(r_+^2 +a^2)^2 \sin^2\theta\over \Sigma_+}(d\tilde{\phi}-\Omega dv)^2\, ,
\ee
where
\be
\Sigma_+=r_+^2+a^2 \cos^2\theta\, .
\ee
It is convenient to introduce coordinates which is co-rotating with the horizon by making the transformation
\be\n{ppsi}
\psi=\tilde{\phi}-\Omega v\, .
\ee
Then the horizon metric (\ref{MH}) takes the form
\be\n{HOR}
dh_-^2= B^2( \CAL{F} d\theta^2+{\sin^2\theta\over \CAL{F}}d\psi^2)\, .
\ee
Here
\be\n{mplus}
B=\sqrt{r_+^2 +a^2}\, ,\ \CAL{F}={1+\beta^2 \cos^2\theta\over 1+\beta^2}\, ,\
\beta=a/r_+\, .
\ee
The parameter $\beta$ changes in the interval from $0$ (for a non-rotating black hole) till $1$ (for an extremely rotating one). From the form of the induced metric (\ref{HOR}) it is easy to see that the horizon area is
\be
\CAL{A}_H^-=4\pi B^2=4\pi (r_+^2 +a^2)\, .
\ee
The vector $\pa_v$ coincides with the Killing vector $\BM{\eta}$ at the horizon, and $v$ is the advanced Killing time coordinate. This Killing parametrization of the horizon generators is uniquely fixed by these conditions.
It is easy to see that the metric $dh_-^2$ is degenerate. We shall use the coordinates $y^a=(v,\theta,\psi)$ on the horizon. Here and later $a,b=1,2,3$. Let us denote $y^A=(\theta,\psi)$, $A,B=2,3$. Then one has
\be
dh_-^2\hor h_{ab}dy^a dy^b\hor h_{AB} dy^A dy^B\, .
\ee
In other words, the components of the metric $h_{ab}$ in the null direction $v$ vanish.
The surface gravity $\kappa$ at the Killing horizon of the Killing vector field $\BM{\eta}$ is defined as follows
\be\n{surfgr}
\kappa^2=-{1\over 2} \eta_{\mu;\nu}\eta^{\mu;\nu}\, .
\ee
In a general case $\kappa$ depends on the normalization of the Killing vector. For the Kerr spacetime this normalization is fixed by a requirement that $\BM{\xi}_t$ is a unit timelike vector at the infinity. The corresponding value of the surface gravity is
\be
\kappa={\sqrt{M^2-a^2}\over r_+^2+a^2}={r_+^2-a^2\over 2r_+(r_+^2+a^2)}\, .
\ee
\subsection{Null tetrads in the Kerr domain}
Let us consider a null tetrad $(\BM{k},\BM{n},\BM{m},\bar{\BM{m}})$ which in the $(v,r,\theta,\tilde{\phi})$ coordinates has components
\ba\n{tet}
k^{\mu}&=& \left( 0,-{r^2+a^2\over \Sigma},0,0\right)\, ,\nonumber\\
n^{\mu}&=& \left( 1, {\Delta\over 2(r^2+a^2)},0,{a\over r^2+a^2}\right)\, ,\\
m^{\mu}&=& {1\over \sqrt{2}(r+ia\cos\theta)}\left( ia \sin\theta,0,1,{i\over \sin\theta}\right) .\nonumber
\ea
These vectors satisfy the following orthogonality conditions
\be
(\BM{k},\BM{n})=-1\, , \ (\BM{m},\bar{\BM{m}})=1\, , \mbox{other products vanish}\, .
\ee
The directions of the vectors $\BM{k}$ and $\BM{n}$ coincide with the principle null directions of the Kerr metric.
This tetrad is regular at the future event horizon \cite{Teuk}\footnote{We change the standard notations for the real null vectors of the tetrad of the paper \cite{Teuk} $\BM{l}\to \BM{n}$ and $\BM{n}\to \BM{k}$ in order to achieve consistency with notations adopted in \cite{BI}.}. This makes it different from the Kinnersley tetrad, which is usually chosen for the separation of variable in the field equations, as well as from symmetric (Carter) null tetrad. The in-tetrad can be obtained from the latter by suitable "null rotations".
The value of the vectors of the in-tetrad (\ref{tet}) at the horizon in $(v,r,\theta,\tilde{\phi})$ coordinates is
\ba\n{KNn}
{k}^{\mu}&\hor& \left( 0,-{r_+^2+a^2\over \Sigma_+},0,0\right)\, ,\
n^{\mu}\hor \left( 1,0,0,\Omega\right)\, ,\\
m^{\mu}&\hor& {1\over \sqrt{2}(r_+ +ia\cos\theta)}\left( ia \sin\theta,0,1,{i\over \sin\theta}\right).
\ea
It is easy to see that the vector $\BM{n}$ at the horizon coincides with Killing vector $\BM{\eta}$ given by \eq{KIL}, which is the null generator of the horizon.
\subsection{Near horizon geometry}
We have two coordinate systems in the vicinity of the horizon: the ingoing Kerr coordinates $x^{\mu}=(v,r,\theta,\tilde{\phi})$ and the co-rotating ones $x_-^{\mu}=(v,r,\theta,\psi)$, where $\psi$ is defined by \eq{ppsi}. In the co-rotating coordinates the null vector $\BM{n}$ on the horizon has the following form
\be
\BM{n}\hor (1,0,0,0)\, .
\ee
We introduce two more holonomic vectors tangent to the horizon
\be
\BM{e}_2=(0,0,1,0)\hh \BM{e}_3=(0,0,0,1)\, .
\ee
In the coordinates $y_-^i=(v,\theta,\psi)$ on the surface of the horizon one has
\be\n{Khol}
\BM{e}_1=\BM{n}\hor \pa_v\hhh\BM{e}_2\hor\pa_{\theta}\hhh\BM{e}_{3}\hor\pa_{\tp}=\pa_{\psi}\, .
\ee
We need to add to this set a new vector, which is transverse to the horizon. We denote it by $\BM{N}$. Let us notice that for the calculation of the transverse extrinsic curvature one needs only to know this vector on the horizon.
The normalization conditions
\be\n{NNN}
\BM{N}^2\hor 0\hhh (\BM{N},\BM{n})\hor -1\hhh (\BM{N},\BM{e}_{2})\hor (\BM{N},\BM{e}_{3})\hor 0\, ,
\ee
uniquely determine it. Simple calculations give
\be
{N}_{\mu}\hor \left(-1, {a^2\sin^2\theta\over2( r^2_+ +a^2)},0,0\right)\, .
\ee
Let us emphasize that the components of $\BM{N}$ are the same in both of the coordinates $x^{\mu}$ and $x_-^{\mu}$.
A transverse extrinsic curvature $\CAL{K}_{ab}$ is defined as follows \cite{BI}
\be\n{KW}
\CAL{K}_{ab}=-N_{\mu} e_b^{\nu}\nabla_{\nu} e_a^{\mu}\hh (i,j=1,2,3)\, .
\ee
The components of this object are invariant under the coordinate transformations in the bulk space and it transforms as a tensor under the change of the coordinates on the horizon. For the chosen holonomic basis it is symmetric, $\CAL{K}_{ab}=\CAL{K}_{ba}$.
To perform these calculations we used the program GRTensor.The calculations are performed in $(v,r,\theta,\tp)$ coordinates. However, since $\CAL{K}_{ab}$ is invariant under the coordinate transformations in the bulk space, its value remains the same in the `co-rotating' coordinate $(v,r,\theta,\psi)$.
The result is
\ba
&&\CAL{K}^{-}_{11}={r_+^2-a^2\over 2r_+(r_+^2+a^2)}\, ,\
\CAL{K}^{-}_{12}=-{a^2\sin\theta \cos\theta\over \Sigma_+}\, ,\nonumber\\
&&\CAL{K}^{-}_{13}=-{a\sin^2\theta \CAL{A}_{13}\over 2r_+\Sigma_+^2}\, , \
\CAL{K}^{-}_{22}=-{r_+(r_+^2+a^2)\over \Sigma_+}\, ,\\
&&\CAL{K}^{-}_{23}=-{a(r_+^2+a^2)\sin\theta \cos\theta\over \Sigma_+}\, ,\nonumber\\
&&\CAL{K}^{-}_{33}=-{(r_+^2+a^2)\sin^2\theta\CAL{A}_{33}\over 2r_+\Sigma_+^3}\, .\nonumber
\ea
Where
\ba
\CAL{A}_{13}&=&r_+^2(3r_+^2+a^2)+a^2(r_+^2-a^2)\cos^2\theta\, ,\n{a13}\\
\CAL{A}_{33}&=& r_+^2(r_+^2-a^2)(2r_+^2+a^2)\nonumber\\
&+&a^2(5r_+^4+2r_+^2a^2+a^4)\cos^2\theta\\
&+&a^4(r_+^2-a^2)\cos^4\theta\, .\nonumber
\ea
Let us also notice that we use the parameter $r_+$ in the above formulas instead of the black hole mass $M$. It is easy to check that
\be\n{KK}
\CAL{K}^{-}_{vv}=\kappa\, .
\ee
\section{Weyl domain}
\subsection{Metric of a static axisymmetric distorted black hole}
Let us consider the inner domain, $\MC{M}_+$, located to the future of the event horizon. According to our assumption this is a metric of a static vacuum axisymmetric distorted black hole. It has two commuting and orthogonal Killing vectors $\BM{\zeta}_T$ and $\BM{\zeta}_\varphi$. Both of them are spacelike inside the horizon. The vector $\BM{\zeta}_\varphi$ is singled out by the property that its integral lines are closed. The other Killing vector $\BM{\zeta}_T$ is uniquely defined, up to a normalization factor, by the condition that it is orthogonal to $\BM{\zeta}_\varphi$. We fix its normalization later.
The Weyl metric is a static axisymmetric vacuum solution of the Einstein equations \cite{Weyl}. A special sub-class of these metrics, that have a regular horizon, describes {\em distorted black holes} \cite{Chandra,Chan}. Properties of static distorted black holes were discussed, e.g., in \cite{GeHa}. This metric has a Killing horizon, where $\BM{\zeta}_T$ becomes null. It separates exterior and interior domains of the Weyl-distorted black hole. In this work we focus on the interior domain only. The properties of the interior of a distorted black hole were studied in \cite{FrSh} (see also \cite{IB}).
We write the inner metric of Weyl-distorted static black hole in the form
\ba\n{WM}
ds_+^2&=&e^{2U}F dT^2+e^{-2U+2V}\left( -{d\rho ^2\over F}+\rho^2 d\vt^2\right)\nonumber\\
&+&e^{-2U} \rho ^2 \sin^2\vartheta d\varphi^2\, ,
\ea
where $F={2m\over \rho }-1$, $U=U(\rho ,\vartheta )$ and $V=V(\rho ,\vartheta )$. When $U=V=0$ the metric \eq{WM} reduces to the Schwarzschild one for the interior of a non-distorted black hole.
The metric (\ref{WM}) is invariant under the following transformation
\be\n{Winv}
U=\tilde{U}+b\hhh T=e^{-b}\tilde{T}\hhh \rho=e^{b}\tilde{\rho}\hhh m=e^{b}\tilde{m}\, .
\ee
In the general case the functions $U$ and $V$ can be found by solving the Einstein equations. The potential $U$ obeys a simple linear equation, that follows from $R_{TT}=0$ (see, e.g., \cite{FrSh}). For our purpose it is convenient to write its solution in the form \cite{man1}
\be\n{UU}
U(\rho,\vt) =\sum_{k=1}^{\infty} \alpha_k \left({R\over m}\right)^k P_k\, ,
\ee
where $\alpha_k$ are constants and
\ba\n{PZR}
P_k&=&P_k(Z)\hh Z={(\rho -m) \cos\vt\over R }\, \\
R&=&\left[m^2\cos^2\vt-F\rho^2\right]^{1/2}\, .\n{RR}
\ea
Here $P_k(\ldots)$ are Legendre polynomials.
The quantities $R/m$ and $(\rho-m)/R$ are invariant under the transformation (\ref{Winv}). In the black hole interior $R^2$ can become null and negative. However, the combination of the power of $R$ and the Legendre polynomial $P_k$, that enters the series \eq{WM}, remains finite and well defined.
One can include in the expansion (\ref{UU}) an extra term, corresponding to $k=0$. But it is a constant. We fix an ambiguity (\ref{Winv}) putting this term equal to zero.
After the solution for $U$ is given, the function $V$ can be found by integrating the expressions, that are quadratic in partial derivatives of $U$. An elegant form of a solution for $V$ was derived in \cite{man1}. Following this work we present $V$ in the form\footnote{Instead of coordinates $(x,y)$ of this work, we use $(\rho,\vt)$, connected with them as follows $x=(\rho-m)/m$ and $y=\cos\vt$.} (see also \cite{man2})
\ba
V(\rho ,\vartheta)&=&\sum_{i=1}^{\infty}\sum_{j=1}^{\infty} { ij \alpha_i\alpha_j\over i+j}\left({R\over m}\right)^{i+j} (P_i P_j-P_{i-1} P_{j-1})\, \nonumber\\
&-&{1\over m}\sum_{i=1}^{\infty} \alpha_i \sum_{j=0}^{i-1} A_{ij}\left({R\over m}\right)^{j}P_j\, ,\\
A_{ij}&=&(-1)^{i+j}[\rho-m(1-\cos\vt)]+\rho -m(1+\cos\vt)\, .\nonumber
\ea
As one can see from the metric (\ref{WM}), the square of the Killing vector $\BM{\zeta}_T$ is proportional to $F$ and hence this vector becomes null at $\rho=2m$. It is easy to show that this relation determines a position of the Killing horizons. One of the Einstein equations ($R_{\rho\vt}=0$) implies that there is the following relation between the values of $U$ and $V$ on the horizon \cite{GeHa}
\be
V(2m,\vt)\hor 2U(2m,\vt)-2u_0.
\ee
Possible conical singularities at fixed points of the Killing vector $\BM{\zeta}_{\vp}$ are absent when
\be
U|_{\vt=0,\pi}\hor u_0\hh V|_{\vt=0,\pi}\hor 0\, .
\ee
In what follows we assume that these conditions are satisfied. In particular, they imply the following restrictions on the coefficients $\alpha_i$
\be
\sum_{k=1}^{\infty} \alpha_{2k-1}=0\hh \sum_{k=1}^{\infty} \alpha_{2k}=u_0\, .
\ee
It is convenient to use the following notation
\be\n{UU0}
U=\CAL{U}+u_0\, .
\ee
Local analysis of the Weyl solutions near a regular event horizon shows that in its vicinity $\CAL{U}$ and $V$ have the following form (for details, see \cite{FrSh,IB})
\ba\n{decom}
\CAL{U}&=&\CAL{U}_0(\vt)+(1-\rho/2m)\CAL{U}_1(\vt)+\ldots\, ,\\
V&=&2\CAL{U}_0(\vt)+(1-\rho/2m) V_1(\vt)+\ldots \, .
\ea
Here `dots' denote higher in powers of $(1-\rho/2m)$ terms. For known $\CAL{U}_0$, the functions $\CAL{U}_{i\ge 1}$ and $\CAL{V}_{i\ge 1}$, which enter \eq{decom}, are defined by recursive relations. In particular, one has
\ba
\CAL{U}_1&=&\cot\vt \ \CAL{U}_{0,\vt}+\CAL{U}_{0,\vt \vt}\, ,\\
V_1&=&\cot\vt \ \CAL{U}_{0,\vt}-\CAL{U}_{0,\vt \vt}-\CAL{U}_{0,\vt}^2\, .
\ea
Using the definition (\ref{surfgr}) for the Killing vector $\zeta_T=\partial_T$ one obtains the following value of the surface gravity at the horizon
\be\n{surfm}
\hat{\kappa}_+={e^{2u_0}\over 4m}\, .
\ee
In what follows we shall sew the outer (Kerr) and the inner (Weyl) metric.
The Killing vector $\BM{\eta}$ and the corresponding Killing parametrization of the horizon of the Kerr metric are uniquely defined. However, the Killing vector $\zeta_T$ of the Weyl metric can differ on the horizon from $\BM{\eta}$ by a normalization factor $\lambda$. For this reason we define new time coordinate $t_+$ and new Killing vector $\BM{\xi}_+$ as follows
\be
t_+=\lambda T\hh \BM{\xi}_+=\partial_{t_+}=\lambda^{-1}\partial_T\, .
\ee
The surface gravity for $\BM{\xi}_+$ is
\be\n{kap}
\kappa_+=\lambda^{-1}{e^{2u_0}\over 4m}\, .
\ee
We fix the parameter $\lambda$ later.
\subsection{Regular at the future horizon coordinates}
The Weyl metric (\ref{WM}) has a coordinate singularity on the horizon.
Let us construct new coordinates, that are regular at the horizon. Using \eq{UU0} we rewrite the metric (\ref{WM}) as follows
\ba\n{WMM}
ds_+^2&=&e^{-2u_0}\left[ e^{2\CAL{U}}F d\tilde{T}^2+e^{-2\CAL{U}+2V}\left( -{d\rho ^2\over F} +\rho^2 d\vt^2\right)\right. \nonumber\\
&+&e^{-2\CAL{U}} \rho ^2 \sin^2\vartheta d\varphi^2\bigg] \, ,
\ea
where $\tilde{T}=\exp(2u_0)T$. Let us introduce ingoing null coordinate (advanced time) $\tilde{v}$
\be
d\tilde{v}=d\tilde{T}-{d\rho\over F}\, .
\ee
One has
\ba\n{VVV}
&&e^{2\CAL{U}}F d\tilde{T}^2-e^{-2\CAL{U}+2V} {d\rho ^2\over F}=\\
&&e^{2\CAL{U}}F d\tilde{v}^2+2e^{2\CAL{U}} d\tilde{v}\, d\rho +4m e^{2\CAL{U}_0}[(2\CAL{U}_1-V_1)+\ldots] d\rho^2\, .\nonumber
\ea
The dots here denote linear and higher order in $(2m-\rho)$ terms. Relations (\ref{WMM}) and (\ref{VVV}) explicitly demonstrate that the Weyl-distorted metric in $(\tilde{v},\rho,\vt,\vp)$ coordinates (\ref{WMM}) is regular at the horizon.
Putting $\rho=2m$ one obtains a metric on the black hole surface
\be\n{HOR_m}
dh_+^2\hor 4m^2 e^{-2u_0}\left[e^{2\CAL{U}_0} d\vt^2 + e^{-2\CAL{U}_0}\sin^2\vt \ d\vp^2\right]\, .
\ee
As expected, this metric is degenerate and its components in $\hat{v}$-direction vanish.
The surface area of the distorted black hole is
\be\n{m_min}
\CAL{A}_H^+=16\pi m^2 e^{-2u_0}\, .
\ee
We shall also use another advanced time coordinate $\hat{v}$ with slightly different normalization
\be
\hat{v}=\exp(-2u_0)\tilde{v}\hh d\hat{v}=dT-e^{-2u_0}{d\rho\over F}\, .
\ee
If one has a vector $\BM{A}$, which has components $(A^T,A^{\rho},\ldots)$ in $(T,\rho)$ coordinates, then its components in $(\hat{v},\rho)$ coordinates are $(A^{\hat{v}},A^{\rho},\ldots)$, where
\be\n{VT}
A^{\hat{v}}=A^T-e^{-2u_0}{A^{\rho}\over F}\, .
\ee
Let us also notice that the Killing vector $\BM{\zeta}_T$ in the ingoing coordinates $\hat{x}^{\mu}=(\hat{v},\rho,\vt,\vp)$ takes the form $\partial_{\hat{v}}$.
\subsection{Null tetrads}
The sectors $({T},\rho)$ and $(\vt,\vp)$ in the Weys domain are orthogonal to each other. This implies a convenient prescription for a choice of a null tetrad. The vectors $(\hat{\BM{N}},\hat{\BM{n}},\BM{E}_2,\BM{E}_3)$ of such a tetrad have the following contravariant components in $(T,\rho,\vt,\vp)$ coordinates
\ba\n{Nne}
\hat{\BM{N}}&=&\left( -{e^{-U}\over \sqrt{2F}}, -e^{U-V}\sqrt{F}/\sqrt{2},0,0\right) \, ,\nonumber\\
\hat{\BM{n}}&=&\left( {e^{-U}\over \sqrt{2F}}, -e^{U-V}\sqrt{F}/\sqrt{2},0,0\right) \, ,\\
\BM{E}_2&=&\left(0,0, {e^{U-V}\over \rho} ,0\right)\, ,\
\BM{E}_3=\left(0,0, 0, {e^{U-V}\over \rho \sin\vt},0\right)\, .\nonumber
\ea
These vectors obey the relations
\be
(\hat{\BM{N}},\hat{\BM{n}})=-1\hh (\BM{E}_2,\BM{E}_2)=(\BM{E}_3,\BM{E}_3)=1\, ,
\ee
the other scalar products vanish. The signs in the definition of the null vectors $\hat{\BM{N}}$ and $\hat{\BM{n}}$ are chosen so, that both of them are future directed. These vectors are chosen in the symmetric form. However they are not regular at the horizon. Regular at the horizon vectors can be obtained by a boost transformation
\be
\BM{N}_+={\sqrt{2}\over \sqrt{F}} e^{-U} \hat{\BM{N}}\hh
\BM{n}_+={\sqrt{F} e^{U}\over \sqrt{2}} \hat{\BM{n}}\, .
\ee
Using the relation
\be
2\CAL{U}-V=(2\CAL{U}_1-V_1)(1-\rho/2m)+\ldots\, .
\ee
it is easy to check that the vectors
\ba
\BM{N}_+&=&\left( -{e^{-2U}\over F},-e^{-V},0,0\right)\, ,\\
\BM{n}_+&=&\left({1\over 2},-{1\over 2}e^{2U-V} F,0,0\right)\, ,
\ea
are regular at the horizon. In $(v_+,\rho)$ coordinates their values on the horizon are
\ba\n{Nnhor}
\BM{N}_+&\hor& \left( 2m e^{-2(\CAL{U}_0+u_0)}(2\CAL{U}_1-V_1),-e^{-2\CAL{U}_0},0,0\right)\, ,\\
\BM{n}_+&\hor& \left(1,0,0,0\right)\, .
\ea
The limit of the other two vectors of the null tetrad on the horizon is
\be\n{eehor}
\BM{E}_2\hor\left(0,0,{e^{-\CAL{U}_0+u_0}\over 2m},0\right)\, ,\
\BM{E}_3\hor\left(0,0,0,{e^{-\CAL{U}_0+u_0}\over 2m\sin\vt}\right)\, .
\ee
\subsection{Near horizon geometry}
We can use the coordinates $\hat{x}^{\mu}=(\hat{v},\rho,\vt,\vp)$ in the vicinity of the Weyl horizon.
However, as we explained earlier, there is no guarantee that the coordinate $\hat{v}$ on the horizon coincides with the Kerr advanced time $v$. So we write
\be
v=\lambda \hat{v}\hh \BM{n}=\lambda^{-1}\BM{n}_+\hh \BM{N}=\lambda \BM{N}_+\, .
\ee
For the proper choice of a constant $\lambda$ the advanced time $v$ and vectors $\BM{n}$ and $\BM{N}$ are continues at the common horizon surface. We specify $\lambda$ later and keep it in the formulas for a while.
Thus the coordinates on the surface of the horizon $\Sigma_+$ are $y_+^i=(v,\vt,\vp)$. We introduce the following holonomic basis of vectors tangent to the horizon $\Sigma_+$
\be\n{Whol}
\BM{n}=\pa_{v}\hh\BM{e}_{\vt}=\pa_{\vt}\hh \BM{e}_{\vp}=\pa_{\vp}\, ,
\ee
One can see that vectors $\BM{e}_{\vt}$ and $\BM{e}_{\vp}$ are colinear with $\BM{E}_2$ and $\BM{E}_3$ and differ from them only by the normalization.
A transverse extrinsic curvature $\CAL{K}_{ab}$ is defined as follows \cite{BI}
\be\n{KW}
\CAL{K}_{ab}=-N_{\mu} e_b^{\nu}\nabla_{\nu} e_a^{\mu}\, .
\ee
We use the program GRTensor to calculate the components of this objects. The result is
\ba
\CAL{K}^{+}_{11}&=&-{1\over 2}\lambda^{-1} e^{2U-V} (2F U_{,\rho }+F_{,\rho })\, ,\nonumber\\
\CAL{K}^{+}_{12}&=&U_{,\vt}\hh \CAL{K}^{+}_{21}={1\over 2}V_{,\vt}\, ,\\
\CAL{K}^{+}_{22}&=&-\lambda\rho e^{-2U+V} \left[\rho (V_{,\rho }-U_{,\rho })+1\right]\, ,\nonumber\\
\CAL{K}^{+}_{33}&=&\lambda \rho \sin^2\vt e^{-2U-V} \left[\rho U_{,\rho }-1\right]\, ,\nonumber
\ea
Other components are zero.
Using local expansions (\ref{decom}) of $U$ and $V$ near the horizon one obtains the following expressions for the components of $\CAL{K}^{+}_{ab}$ on the horizon
\ba
\CAL{K}^{+}_{11}&=& \kappa_+\equiv {e^{2u0}\over 4\lambda m}\, ,\nonumber\\
\CAL{K}^{+}_{12}&=&\CAL{K}^{+}_{21}=\CAL{U}_{0,\vt}\, , \nonumber \\
\CAL{K}^{+}_{22}&=&-{1\over 2\kappa_+}\left[1+(\CAL{U}_1-V_1)\right]\\
\CAL{K}^{+}_{33}&=& -{1\over 2\kappa_+} e^{-4\CAL{U}_0}\sin^2\vt(1-\CAL{U}_1)\, .\nonumber\\
\ea
\section{Gluing solutions together}
\subsection{Induced geometry}
By comparing the induced metrics $dh_{\pm}^2$ on the horizon surface, given by \eq{HOR} and \eq{HOR_m}, one arrives at the following conclusions:
\begin{itemize}
\item Coordinates $\vt$ and $\vp$ on the Weyl horizon coincide with the coordinates $\theta$ and $\psi$ on the Kerr horizon;
\item The surface areas of the both horizons are identical if
\be\n{c1}
2me^{-u_0}=\sqrt{r_+^2+a^2} ;
\ee
\item The shapes of the both horizons are the same provided
\be\n{c2}
\exp{(2\CAL{U}_0)}=\CAL{F}\equiv{1+\beta^2\cos^2\theta\over 1+\beta^2}\, ;
\ee
\end{itemize}
In what follows we shall use $\theta$ and $\psi$ as common coordinates on the joint horizon. The conditions (\ref{c1}) and (\ref{c2}) guarantee that the induced geometries on the joint horizon coincide.
Using relations (\ref{RR}) one has at the horizon $\rho=2m$
\be
R\hor m|\cos\theta|\, ,\ \varepsilon\equiv \cos\theta/|\cos\theta|=\pm 1\, ,\ P_i\hor P_i(\varepsilon)=\varepsilon^i\, ,\nonumber
\ee
and the relation (\ref{UU}) takes the form
\be\n{UUUU}
U\hor\sum_{i=1}^{\infty} \alpha_i \cos^i\theta\, .
\ee
Using \eq{c2} one can write
\ba\n{uU}
U&\hor& u_0+\CAL{U}_0=u_0+{1\over 2}\ln \CAL{F}\, ,\nonumber\\
&=&u_0+{1\over 2}[\ln(1+\beta^2 \cos^2\theta)-\ln(1+\beta^2)]\, .
\ea
\begin{itemize}
\item By comparing this relation with \eq{UUUU} one concludes that
\be\n{c3}
u_0={1\over 2}\ln(1+\beta^2)\, ,
\ee
\item The relation (\ref{c3}) implies that
\be
e^{2u_0}={r_+^2+a^2\over r_+^2}={2M\over r_+}\, ,
\ee
and hence
\be
m=M={r_+^2+a^2\over 2r_+}\, .
\ee
\item If the surface gravities $\kappa$ and $\kappa_+$ coincide one has
\be\n{lam}
\lambda={2(r_+^2+a^2)\over r_+^2-a^2}\, .
\ee
\item
The holonomic bases $\BM{e}_i$ defined by \eq{Khol} and \eq{Whol} are identical.
\end{itemize}
Using the above relations, one can rewrite $\CAL{K}^{+}_{ab}$ in terms of coordinates and parameters of the external Kerr solution
\ba
\CAL{K}^{+}_{11}&=& \kappa_+\equiv {e^{2u0}\over 4\lambda m}\, ,\
\CAL{K}^{+}_{12}=\CAL{K}^{+}_{21}=-{a^2 \sin\theta \cos\theta\over \Sigma_+}\, , \nonumber \\
\CAL{K}^{+}_{22}&=&(2\kappa_+\Sigma_+^2)^{-1}\times [r_+^2(r_+^2+2a^2)\\
&-&a^2(2r_+^2+3a^2)\cos^2\theta +2a^4\cos^4\theta]\, ,\nonumber\\
\CAL{K}^{+}_{33}&=&-{(r_+^2-a^2\cos^2\theta)\sin^2\theta\over 2\kappa_+ \Sigma_+^2}\, .\nonumber
\ea
\subsection{Potential $U$ in the Weyl domain}
Before considering the jump conditions for transverse extrinsic curvature let us show that the relation (\ref{c2}) determines the potential $U$ everywhere in the Weyl domain.
Using the Teylor decomposition of the depending on the angle $\theta$ part in the right-hand side of \eq{uU} one obtains
\be\n{alpha}
\alpha_{2k-1}=0\hh \alpha_{2k}=(-1)^{k+1}{\beta^{2k}\over 2k}\, .
\ee
It is easy to check that
\be\n{sumH}
\sum_{k=1}^{\infty} \alpha_{2k}={1\over 2}\ln(1+\beta^2)\, .
\ee
as it should be.
We use now the obtained value for the coefficients $\alpha_i$ and substitute them into \eq{UU}. The summation in the resulting series can be performed and one obtains (see Appendix)
\ba\n{UUU}
U&=&{1\over 2}\ln\left( {\MC{X}_+ \MC{X}_-\over 4}\right)\, ,\\
\MC{X}_{\pm}&=&\sqrt{X_{\pm}^2+\lambda}+X_{\pm}\, ,\\
X_{\pm}&=&1\pm i {y} Z\hhh \lambda={y}^2(Z^2-1)\hhh {y}=\beta R/ m\, ,\\
P_k&=&P_k(Z)\hh Z={(\rho -m) \cos\vt\over R }\, \\
R&=&\left[m^2\cos^2\vt-2m\rho +\rho^2\right]^{1/2}\, .
\ea
Let us notice that expression \eq{UUU} for the potential $U$ in the Weyl domain contains complex quantities. However, the function $U$ is real, as it should be.
One can check the obtained answer by substituting it in the equation $R_{TT}=0$. On the horizon $|Z|=1$ and $\lambda=0$, and the potential $U$ takes the form
\be
U\hor {1\over 2} \ln(X_+ X_-)\hor {1\over 2}\ln(1+{y}^2)\hor{1\over 2}\ln(1+\beta\cos^2\theta)\, .
\ee
Thus the boundary conditions are also satisfied.
\subsection{Massive thin null shell}
To satisfy the continuity of the Killing vectors at the horizon
in what follows we impose the condition
\be\n{pm}
\kappa_-=\kappa_+\, ,
\ee
and use relation (\ref{lam})
to fix the normalization of the advance time coordinates $v$. We require the continuity of $v$ in such a chosen parametrization.
Following \cite{BI} we denote by $\gamma_{ab}$ the jump of the transverse extrinsic curvature
\be
{1\over 2}\gamma_{ab}=[\CAL{K}_{ab}]\equiv \CAL{K}^+_{ab}-\CAL{K}^-_{ab}\, .
\ee
The condition (\ref{pm}) implies
\be\n{g11}
n^a n^b \gamma_{ab}=\gamma_{11}=0\hh \gamma_{12}=0\, .
\ee
In order to use the formalism developed in \cite{BI} we introduce an object $g_*^{ab}$ obeying the equation
\be
g_*^{ac}g_{bc}=\delta^a_b+n^a N_b\, .
\ee
For the adopted choice of the null vectors $\BM{n}$ and $\BM{N}$ the metric $g_*^{ac}$ can be chosen as the contravariant two-metric $g^{AB}$, bordered by zeros. Using (\ref{g11}) one can write the relation (31) of \cite{BI} in the form
\be
-16\pi S^{ab}=[g_*^{ac} n^b n^d+g_*^{bd} n^a n^c-n^a n^b g_*^{cd}]\gamma_{cd}\, .
\ee
We put $\epsilon=0$ in this relation since the horizon surface is null, and put $\eta=-1$, since $(\BM{n},\BM{N})=-1$.
Let us denote
\be
q^a=-{1\over 2} g_*^{ac} n^d \gamma_{cd}\hh
\CAL{K}= g_*^{cd}(\CAL{K}^+_{cd}-\CAL{K}^-_{cd})\, .
\ee
Then one has
\be\n{set}
8\pi S^{ab}=\CAL{K} n^a n^b + 2q^{(a} n^{b)}\, .
\ee
It is easy to check that
\be
q^a=q\xi_{\psi}^a \hh q=-{a \CAL{A}_{13}\over 2r_+ \Sigma_+(r_+^2+a^2)^2}\, ,
\ee
Here $\xi_{\psi}=\partial_{\psi}$ is the Killing vector, generating the rotations of the horizon, and $\CAL{A}_{13}$ is given by \eq{a13}.
Let us denote
\be
\mbox{tr}\CAL{K}^{\pm}=g^{CD}\CAL{K}^{\pm}_{CD}\, .
\ee
Then, since $g^{CD}$ is continuous at the horizon, one has
\be
\CAL{K}=\mbox{tr}\CAL{K}^{+}-\mbox{tr}\CAL{K}^{-}\, .
\ee
The calculations give
\ba
\mbox{tr}\CAL{K}^{-}&=-&{ 4r_+^4+3r_+^2 a^2+a^4
+a^2(r_+^2-a^2)\cos^2\theta\over 2r_+ (r_+^2+a^2)\Sigma_+}\, ,\nonumber\\
\mbox{tr}\CAL{K}^{+}&=-&{(r_+^2+a^2)r_+ B^+
\over (r_+^2-a^2)\Sigma_+^3}\, ,\\
B^+&=&r_+^2(2r_+^2+3 a^2)- (3r_+^2+4a^2)a^2\cos^2\theta\nonumber\\
&+&2a^4\cos^4\theta\, ,\nonumber\\
\CAL{K}&=& -{ a^2\sin^2\theta B \over 2r_+ (r_+^4-a^4)\Sigma_+^3}\, ,\\
B&=&r_+^4(15 r_+^4+18 r_+^2 a^2+7 a^4)\nonumber\\
&+&2r_+^2 a^2(r_+^4-6r_+^2 a^2-3a^4)\cos^2\theta\nonumber\\ &+&(r_+^2-a^2)^2 a^4 \cos^4\theta\, .\nonumber
\ea
Suppose $\Phi(x^{\mu})=0$ is the equation of the shell and $\alpha$ is defined by the relation
\be
\alpha^{-1}\partial_{\mu}\Phi\hor n_{\mu}\, ,
\ee
then, considered as a distribution, the stress-energy tensor of the shell is \cite{BI}
\be
T_{\Sigma}^{\mu\nu}=\alpha S^{\mu\nu} \delta(\Phi)\, .
\ee
We choose $\Phi=r_+-r$, so that $\Phi$ increases from $\MC{M}_-$ to $\MC{M}_-$. The vectors $\BM{n}$ and $\Phi_{,\mu}$ (in the Kerr domain) have components
\be
n_{\mu}\hor (0,{\Sigma_+\over r_+^2+a^2},0,0)\hh
\Phi_{,\mu}\hor (0,-1,0,0)\, ,
\ee
so that
\be
\alpha=-{r_+^2+a^2\over \Sigma_+}\, .
\ee
The function $\alpha$ is negative, as it should be for the null shells (see \cite{BI}).
\section{Discussions}
In this paper we constructed a new solution of the Einstein equations, which we called a "hybrid" black hole. The corresponding metric is obtained by gluing of the external Kerr metric and the internal Weyl-distorted metric along the common horizon . To glue these metrics we used the method of the massive thin null shells \cite{BI}. The horizon surface is null, so that the metric, induced on it is degenerate. For proper gluing of these geometries one needs to identify the null geodesics at the both horizons, that are their generators, and require that the transverse two-dimensional metrics are isometric. To exclude the re-parametrization freedom, one needs to fix the parameters along the horizon generators, before the gluing. For this purpose we use the Killing parameters. Namely, we required the continuity of the advanced time coordinate $v$ defined in the external (Kerr) and the internal (Weyl) domain. The only left ambiguity in the choice of $v$ is connected with the ambiguity of the normalization of the corresponding Killing vector. This ambiguity is fixed in the Kerr domain by the condition that the timelike at infinity Killing vector has a unit norm their. To fix the normalization of the corresponding Killing vector in the Weyl domain we imposed the condition that the surface gravities calculated at the horizon of the Kerr and Weyl metrics are the same.
We calculated the stress-energy tensor of the massive thin null shell $S^{ab}$ (formula (\ref{set})). We demonstrated that it is a sum of two terms. The first one, ${1\over 8\pi}\CAL{K} n^a n^b$, is proportional to $a^2$. It describes the null fluid, propagating along the horizon. The second one, ${1\over 4\pi}q^{(a} n^{b)}$, is proportional to $a$ and it describes a current along the horizon, carrying the angular momentum. Such a "rotating" null shell generates theangular momentum of the external (Kerr) metric. At the same time it distorts the "non-rotating" interior. The massive null shell "respects" both of the spacetime symmetries: it is stationary (time independent) and axisymmetric.
The horizon of the "hybrid" black hole is a special (null) case of a rigidly rotating ZAMO surface \cite{FroFro}. An interesting question is whether it is possible to construct a solution of the Einstein equations which is a junction of the Kerr and Weyl metrics, glued at a timelike surface od a rigidly rotating ZAMO surface.
\section*{Acknowledgments}
The authors thank the Natural Sciences and Engineering Research Council of Canada for the financial support. One of the authors (V.F.) is also grateful to the Killam Trust for its financial support.
|
\section{Introduction}
\label{sec.intro}
The last decade brought a dynamic evolution of the computing capabilities of graphics
processing units (GPUs). In that time, the performance of a single card increased
from tens of GFLOPS in NVxx to TFLOPS in the newest Kepler/Maxwell NVIDIA chips~\cite{cuda}.
This raw processing power did not go unnoticed by the engineering and science communities,
which started applying GPUs to accelerate a wide array of calculations in what became
known as GPGPU -- general-purpose computing on GPUs. This led to the development
of special GPU variants optimized for high performance computing (e.g. the NVIDIA
Tesla line), but it should be noted that even commodity graphics cards, such as those
from the NVIDIA GeForce series, still provide enormous computational power and can
be a very economical (both from the monetary and energy consumption point of view)
alternative to large CPU clusters.
The spread of GPGPU techniques was further facilitated by the development of CUDA and OpenCL
-- parallel programming paradigms allowing efficient exploitation of the available
GPU compute power without exposing the programmer to too many low-level details of the
underlying hardware. GPUs were used successfully to accelerate many problems,
e.g. the numerical solution of stochastic differential equations~\cite{januszewski.kostur.09,spiechowicz.15},
fluid simulations with the lattice Boltzmann method~\cite{tolke2009,januszewski.kostur.13},
molecular dynamics simulations~\cite{anderson2008},
classical~\cite{preis2009} and quantum Monte Carlo~\cite{anderson2007} simulations,
exact diagonalization of the Hubbard model ~\cite{siro.harju.12}, {\it etc}.
Parallel computing in general, and its realization in GPUs in particular,
can also be extremely useful in many fields of solid state physics. For
a large number of problems, the ground state of the system and
its free energy are of special interest. For instance, in order to
determine the phase diagram of a model,
free energy has to be calculated for a large number of points in the
parameter space. In this paper, we address this very issue and illustrate
it on a concrete example of a superconducting system with an oscillating
order parameter (OP), specifically an iron-based multi-band superconductor (FeSC).
Our algorithm is not limited to systems of this type and can also be used for systems
in the homogeneous superconducting state (BCS).
The discovery of high temperature superconductivity in FeSC~\cite{kamihara.watanabe.08} began a
period of intense experimental and theoretical research.~\cite{ishida.nakai.09}
All FeSC include a two-dimensional structure which is shown in Fig.
\ref{fig.feas}.a. The Fermi surfaces (FS) in FeSC are composed of hole-like
Fermi pockets (around the $\Gamma = ( 0,0 )$ point) and electron-like Fermi
pockets (around the $M = ( \pi , \pi )$ point) -- Fig. \ref{fig.feas}.b.
Moreover, in FeSC we expect the presence of $s_{\pm}$ symmetry of the
superconducting OP.~\cite{mazin.singh.08} In this case the OP exhibits a sign
reversal between the hole pockets and electron pockets. For one $\mathtt{Fe}$
ion in the unit cell, the OP is proportional to $\cos k_{x} \cdot \cos k_{y} $.
\begin{figure}[!h]
\begin{center}
\includegraphics{feas}
\end{center}
\caption{(Color on-line) (Panel a) $\mathtt{FeAs}$ layers in FeSC are built by
$\mathtt{Fe}$ ions (red dots) forming a square lattice surrounded by
$\mathtt{As}$ ions (green dots) which also form a square lattice. $\mathtt{As}$
ions are placed above or under the centers of the squares formed by
$\mathtt{Fe}$. This leads to two inequivalent positions of $\mathtt{Fe}$ atoms,
so that there are two ions of $\mathtt{Fe}$ and $\mathtt{As}$ in an elementary
cell. (Panel b) True (folded) Fermi surface in the first Brillouin zone for two
$\mathtt{Fe}$ ions in unit cell. The colors blue, red and green correspond to the FS
for the 1st, 2nd, and 3rd band, respectively.} \label{fig.feas}
\end{figure}
FeSC systems show complex low-energy band structures, which have been extensively studied.~\cite{mazin.singh.08,kunes.arita.10,boeri.dolgov.08,
singh.du.08,graser.maier.09,kuroki.onari.08} A consequence of this is a more
sensitive dependence of the FS to doping.~\cite{pan.li.13} In the
superconducting state, the gap is found to be on the order of 10 meV, small
relative to the breadth of the band.~\cite{ding.richard.08} This increases
the required accuracy of calculated physical quantities
needed to determine the phase diagram of the superconducting
state, such as free energy.~\cite{tai.zhu.13,ptok.14}
In this paper we show how the increased computational cost of
obtaining thermodynamically reliable results can be offset by parallelizing the
most demanding routines using CUDA, after a suitable transformation of variables
to decouple the interacting degrees of freedom. In Section \ref{sec.theory_ph}
we discuss the theoretical background of numerical
calculations. In Section \ref{sec.algorithm} we describe the implementation of
the algorithm and compare its performance when executed on the CPU and GPU.
We summarize the results in Section \ref{sec.summary}.
\section{Theoretical background}
\label{sec.theory_ph}
Many theoretical models of FeSC systems have been proposed, with
two~\cite{raghu.qi.08},
three~\cite{daghofer.nicholson.10,daghofer.nicholson.12,korshunov.togushova.13},
four~\cite{korshunov.eremin.08} and five
bands~\cite{graser.maier.09,kuroki.onari.08}. Most of the models mentioned
describe one \texttt{Fe} unit cell and closely approximate the band and FS structure
(Fig~\ref{fig.feas}.b) obtained by LDA
calculations.~\cite{singh.du.08,ding.richard.08,kondo.santander.08,
cvetkovic.tesanovic.2.09} In every model the non-interacting tight-binding
Hamiltonian of FeSC in momentum space can be described by:
\begin{eqnarray}
H_{0} = \sum_{\alpha\beta{\bm k}\sigma} T^{\alpha\beta}_{{\bm k}\sigma} c_{\alpha{\bm k}\sigma}^{\dagger} c_{\beta{\bm k}\sigma} ,
\end{eqnarray}
where $c_{\alpha{\bm k}\sigma}^{\dagger} (c_{\alpha{\bm k}\sigma})$ is the
creation (annihilation) operator for a spin $\sigma$ electron of momentum ${\bm
k}$ in the orbital $\alpha$ (the set of orbitals is model dependent). The hopping
matrix elements $T_{{\bm k}\sigma}^{\alpha\beta} = T_{\bm k}^{\alpha\beta} -
\delta_{\alpha\beta} ( \mu + \sigma h )$ determine the model of FeSC. Here,
$\mu$ is the chemical potential and $h$ is an external magnetic field parallel
to the \texttt{FeAs} layers. For
our analysis we have chosen the minimal two-band model proposed by Raghu {\it
et al.}~\cite{raghu.qi.08} and the three-band model proposed by Daghofer {\it et
al.}~\cite{daghofer.nicholson.10,daghofer.nicholson.12} (described in
\ref{app.twoband} and \ref{app.threeband} respectively). The band structure and FS
of the FeSC system can be reconstructed by diagonalizing the Hamiltonian $H_{0}$:
\begin{eqnarray}
H'_{0} &=& \sum_{\varepsilon{\bm k}\sigma} E_{\varepsilon{\bm k}\sigma} d_{\varepsilon{\bm k}\sigma}^{\dagger} d_{\varepsilon{\bm k}\sigma} ,
\end{eqnarray}
where $d_{\varepsilon{\bm k}\sigma}^{\dagger} (d_{\varepsilon{\bm k}\sigma})$
is the creation (annihilation) operator for a spin $\sigma$ electron of
momentum ${\bm k}$ in the band $\varepsilon$.
\paragraph{Superconductivity in multi-band iron-base systems in high magnetic fields}
FeSC superconductors are
layered~\cite{singh.du.08,ding.richard.08,kondo.santander.08,cvetkovic.tesanovic.2.09,
lioa.kondoa.09,cvetkovic.tesanovic.09}, clean~\cite{kim.tanatar.11,khim.lee.11}
materials with a relatively high Maki parameter $\alpha \sim
1-2$.~\cite{khim.lee.11,cho.kim.11,zhang.liao.11,kurita.kitagawa.11,terashima.kihou.13}
All of the features are shared with heavy fermion systems, in which strong
indications exist to observe the Fulde--Ferrell--Larkin--Ovchinnikov (FFLO)
phase~\cite{FF,LO} -- a superconducting phase with
an oscillating order parameter in real space, caused by the non-zero value of the total
momentum of Cooper pairs.
In contrast to the BCS state where Cooper pairs form a singlet state $({\bm
k}\uparrow,-{\bm k} \downarrow)$, the FFLO phase is formed by pairing states
$({\bm k}\uparrow,-{\bm k}+{\bm q} \downarrow)$. These states can occur between
the Zeeman-split parts of the Fermi surface in a high external magnetic field
(when the paramagnetic pair-breaking effects are smaller than the diamagnetic
pair-breaking effects).~\cite{matsuda.shimahara.07} In one-band materials, the
FFLO can be stabilized by anisotropies of the Fermi-surface and of the
unconventional gap function,~\cite{matsuda.izawa.06} by pair hopping
interaction~\cite{ptok.maska.09} or, in systems with nonstandard quasiparticles, with spin-dependent
mass.~\cite{kaczmarczyk.spalek.09,kaczmarczyk.spalek.10,maska.mierzejewski.10,kaczmarczyk.sadzikowski.11}
This phase can be also realized in inhomogeneous systems in the presence of
impurities~\cite{wang.hu.06,wang.hu.07,ptok.10} or spin density
waves~\cite{ptok.maska.11}. In some situations, the FFLO can be also stable in
the absence of an external magnetic field.~\cite{loder.kampf.10}
In multi-band systems, the experimental~\cite{khim.lee.11,cho.kim.11,
tarantini.gurevich.11,burger.hardy.13,zocco.grube.13} and
theoretical~\cite{ptok.14,gurevich.10,gurevich.11,ptok.crivelli.13,
mizushima.takahashi.14,takahashi.mizushima.14,crivelli.ptok.14} works point to
the existence of the FFLO phase in FeSC.
Through the analysis of the Cooper pair susceptibility in the minimal two-band
model of FeSC, such systems are shown to support the existence of an FFLO phase,
regardless of the exhibited OP symmetry. It should be noted that the state with
nonzero Cooper pair momentum, in FeSC superconductors with the $s_{\pm}$ symmetry,
is the ground state of the system near the Pauli
limit.~\cite{ptok.14,ptok.crivelli.13} This holds true also for the three-band
model (e.g. \ref{app.suscept} and Ref. \citep{crivelli.ptok.14}).
\paragraph{Free energy for intra-band superconducting phase}
In absence of inter-band interactions, the BCS and the FFLO phase (with
Cooper pairs with total momentum ${\bm q}_{\varepsilon}$ equal zero and
non-zero respectively) can be described by the effective Hamiltonian:
\begin{equation}
H_{SC} = \sum_{\varepsilon{\bm k}} \left( \Delta_{\varepsilon{\bm k}} d_{\varepsilon{\bm k}\uparrow} d_{\varepsilon,-{\bm k}+{\bm q}_{\varepsilon} \downarrow} + H.c. \right) ,
\end{equation}
where $\Delta_{\varepsilon{\bm k}} = \Delta_{\varepsilon} \eta( {\bm k} )$ is
the amplitude of the OP for Cooper pairs with total momentum ${\bm
q}_{\varepsilon}$ (in band $\varepsilon$ with symmetry described by the form factor
$\eta ({\bm k})$ -- for more details see Ref. \cite{ptok.crivelli.13}). Using
the Bogoliubov transformation we can find the eigenvalues of the full
Hamiltonian $H = H_{0} + H_{SC}$: \begin{eqnarray}
\label{eq.enequasiparticle} \mathcal{E}_{\varepsilon{\bm k}}^{\pm} &=& \frac{E_{\varepsilon{\bm k}\uparrow} - E_{\varepsilon,-{\bm k}+{\bm q}\downarrow}}{2} \pm \sqrt{ \left( \frac{E_{\varepsilon{\bm k}\uparrow} + E _{\varepsilon,-{\bm k}+{\bm q}\downarrow}}{2} \right)^{2} + | \Delta_{\varepsilon{\bm k}} |^{2} } .
\end{eqnarray}
In this case we formally describe two independent bands. The total free energy for the system is given by $\Omega = \sum_{\varepsilon} \Omega_{\varepsilon}$, where
\begin{eqnarray}
\label{eq.freeene}
\Omega_{\varepsilon} = - \frac{1}{\beta} \sum_{\alpha \in \{+, -\}} \sum_{\bm k} \ln \left( 1 +
\exp ( - \beta \mathcal{E}_{\varepsilon{\bm k}}^{\alpha} ) \right) + \sum_{\bm
k} \left( E_{\varepsilon{\bm k} \downarrow} - \frac{ 2 | \Delta_{\varepsilon}
|^{2} }{ V_{\varepsilon} } \right) ,
\end{eqnarray}
corresponding to the free energy in $\varepsilon$-th band, where $V_{\varepsilon}$ is the respective interaction intensity
and $\beta = 1 / k_B T$.
\begin{figure}[!h]
\begin{center}
\includegraphics[scale=1]{ene}
\caption{(Color on-line) Free energy $\Omega_{\varepsilon}$ for different
parameters $V_{\varepsilon}$ in $h \simeq h_{C}^{BCS}$ -- results for
$h_{C}^{BCS} = 0.025 eV$ (panels a, c and e) and $0.005 eV$ (panels b, d and
f).} \label{fig.minene}
\end{center}
\end{figure}
\paragraph{Historical and technical note}
The historically basic concept of the FFLO phase was simultaneously proposed
by two independent groups, Fulde-Ferrell~\cite{FF} and
Larkin-Ovchinnikov~\cite{LO} in 1964. The first group proposed a
superconducting phase where Cooper pairs have only one non-zero total
momentum ${\bm q}$, and the superconducting order parameter in real space
$\Delta ( {\bm R}_{j} ) \sim \exp ( i {\bm R}_{j} \cdot {\bm q} )$. In the
second case, Cooper pairs have two possible momenta: ${\bm q}$ and the opposite
$- {\bm q}$, with an equal amplitude of the order parameter. Thus in real space
the superconducting order parameter is given by $\Delta ( {\bm R}_{j} ) \sim
\exp ( i {\bm R}_{j} \cdot {\bm q} ) + \exp ( - i {\bm R}_{j} \cdot {\bm q} ) =
\cos ( {\bm R}_{j} \cdot {\bm q} )$. However, the most general case of FFLO is
a superconducting order parameter given by a sum of plane waves, where the
Cooper pairs have all compatible values of the momentum ${\bm q}_{\alpha}$ in
the system:
\begin{equation}
\Delta ( R_{j} ) = \sum_{\alpha = 1}^{M} \Delta_{\alpha} \exp ( i {\bm R}_{j} \cdot {\bm q}_{\alpha} )
\end{equation}
where $M$ is the cardinality of the first Brillouin zone (in the square lattice
it is equal to $N_{x} \times N_{y}$). For the historical reasons described
above, whenever $M = 1$ (${\bm q}_{1} \neq 0$ and $\Delta_{1} \neq 0$) we can
speak about the Fulde--Ferrell (FF) phase, whereas for $M = 2$ (and ${\bm
q}_{1} = -{\bm q}_{2}$, $\Delta_{1} = \Delta_{2}$) about the
Larkin--Ovchinnikov (LO) phase.
Larger $M$ impose a more demanding spatial decomposition of the order
parameter, both in the theoretical and computational sense. However, every time
it can be reduced to the diagonalization of the (block) matrix representation
of the Hamiltonian. Using the translational symmetry of the lattice, the
problem for the FF phase ($M=1$) in one-band systems corresponds to the
independent diagonalization of $2 \times 2$ matrices (with eigenvalues given
like in Eq.~\ref{eq.enequasiparticle} with the number of bands $\varepsilon = 1$)
for each of the $N_{x} \times N_{y}$ different momentum sectors. In case of the
LO phase ($M=2$), the calculation can be similarly decomposed in momentum
space or using other spatial symmetries of the system (an example of this
procedure can be found in Ref.~\cite{ptok.maska.11}), with a much greater
computational effort due to the lower degree of symmetry, leading to $N_{x}$
independent diagonalization problems of size $2 N_{y} \times 2 N_{y}$. In the
{\it full} FFLO phase (i.e. in a system with
impurities~\cite{wang.hu.06,wang.hu.07,ptok.10} or a vortex lattice~\cite{maska.03}), the
spatial decomposition is determined in real space using the self-consistent
Bogoliubov-de Gennes equations, which require the full diagonalization of a
Hamiltonian of maximal rank $2 (N_x
N_y)$~\cite{loder.kampf.08,yanase.09,ptok.12,zhou.zhang.10} at every
self-consistent step. To work around these limitations, iterative
methods~\cite{litak.miller.95,martin.annett.98} or the Kernel Polynomial
Method~\cite{weisse.wellein.06} can be used. These methods are based on the
idea of expressing functions of the energy spectrum in an orthogonal basis,
e.g. Chebyshev polynomial
expansion.~\cite{furukawa.motome.04,covaci.peeters.10,gao.huang.12,nagai.nakai.12,nagai.ota.12,nagai.shinohara.13,he.song.13}
By doing so, it becomes possible to conduct self-consistent calculations in
the superconducting state without performing the diagonalization procedure.
The time expense of iterative methods can also be reduced by a careful GPU
implementation, which is currently a work in progress.
In the present work, we describe how the calculation of the free energy can be accelerated in the FF
phase, which due to its greater symmetry allows optimal parallelization on a
GPU architecture.
\section{Parallel calculation of free energy}
\label{sec.algorithm}
\subsection{Programming models -- OpenMP, OpenACC and CUDA C}
Parallel programing can be realized in CPUs and GPUs in many different ways. In
this section we compare the performance of the same algorithm implemented
using OpenMP~\cite{openmp}, PGI CUDA/OpenACC Fortran~\cite{cloutier}, and directly in CUDA C~\cite{cuda}.
The first two are generic extensions of Fortran/C++ that make it easy to, respectively,
use multiple CPU cores, and compile a subset of existing Fortran/C++ code for a GPU.
They take the form of annotations which can be added to existing code, and as such,
enable the use of additional computational power with very little
overhead by the programmer. Typically, much better efficiency can be achieved
by the third option -- i.e. a specifically optimized implementation targeting
the GPU architecture directly. This requires more work on the part of the
programmer, both in adjusting the algorithms and in rewriting the code, but
it makes it possible to fully utilize the available resources.
\subsection{GPU algorithm}
The global ground state for a fixed magnetic field strength $h$ and temperature
$T$ is found by minimizing the free energy over the set of
$\Delta_{\varepsilon}$ and ${\bm q}_{\varepsilon}$. In case of $n$ independent
bands this corresponds to global minimization of the free energy
$\Omega_{\varepsilon}$ in every band separately, for every ${\bm
q}_{\varepsilon}$ in the first Brillouin zone (FBZ) -- Algorithm~\ref{alg.1}.
For the calculation of the free energy $\Omega_{\varepsilon}$, we must know the
eigenvalues $E_{\varepsilon{\bm k}\sigma}$ reconstructing the band structure of
our systems. In the case of the two-band model, it can simply be found analytically (see
\ref{app.twoband}). However, for models with more bands
(such as the three-band model -- \ref{app.threeband}) the band
structure has to be determined numerically (e.g. using a linear algebra
library, such as Lapack (CPU) or Magma (GPU)~\cite{magma}). With this approach,
the calculation of $E_{\varepsilon{\bm k}\uparrow}$ and $E_{\varepsilon,-{\bm
k}+{\bm q}\downarrow}$ becomes a computationally costly procedure, and if
it were to be repeated inside the inner loop of Algorithm~\ref{alg.1},
it would significantly impact the execution time. For this reason,
we propose to precalculate the eigenvalues for every momentum vector ${\bm k} \in FBZ$
and store them in memory for models with more than two bands.
The main downside of this approach is the large increase in memory usage.
While Algorithm \ref{alg.1} is simple to realize on a CPU, its execution time
is proportional to the system size $N_{x} \times N_{y}$, and as such
scales quadratically with $N_{x}$ for a square lattice ($N_{x}$
and $N_{y}$ are the number of lattice sites in the $x$ and $y$ direction,
respectively).
\begin{algorithm*}
\caption{Finding ${\bm q}_{\varepsilon}$ and $\Delta_{\varepsilon}$ corresponding to a global minimum of free energy in band $\varepsilon$.}
\label{alg.1}
\begin{algorithmic}[1]
\FOR{${\bm q}_{\varepsilon} \in FBZ$}
\STATE generate matrices $E_{\varepsilon{\bm k}\uparrow}$ and $E_{\varepsilon,-{\bm k}+{\bm q}_{i}\downarrow}$ for ${\bm k} \in FBZ$
\FOR{$\Delta_{\varepsilon} = 0$ to $\Delta_{max}$}
\STATE calculate matrices $\mathcal{E}_{\varepsilon{\bm k}}^{\pm}$ for ${\bm k} \in FBZ$ -- Eq.~\ref{eq.enequasiparticle}
\STATE calculate $\Omega_{\varepsilon}$
\STATE find and save $\Delta_{\varepsilon}$ corresponding to a fixed ${\bm q}_{\varepsilon}$ and minimal value $\Omega_{\varepsilon}$
\ENDFOR
\STATE find and save ${\bm q}_{\varepsilon}$ and $\Delta_{\varepsilon}$ corresponding to minimum of $\Omega_{\varepsilon}$
\ENDFOR
\end{algorithmic}
\end{algorithm*}
Sometimes the physical properties of the system make it possible
to reduce the amount of computation -- for instance when it is known that
the minimum of the energy is attained for values of momentum ${\bm q}_{\varepsilon}$ in specific
directions -- Fig. \ref{fig.minene}.~\cite{ptok.14,ptok.maska.09,ptok.10,ptok.maska.11,ptok.crivelli.13,crivelli.ptok.14,
mierzejewski.ptok.10}
In this case, the outer loop of Algorithm \ref{alg.1} can be restricted to
${\bm q}_{\varepsilon} \in \mathcal{Q} \subset FBZ$, where $\mathcal{Q}$ is a set of
$N \ll N_{x} \times N_{y}$ vectors. Such reductions are not unique to linear systems
with translational symmetry but are also the case for systems with rotational symmetry.~\cite{loder.kampf.08,yanase.09,ptok.12}
In the case of BCS-type superconductivity where Cooper pairs have zero total momentum
(${\bm q}_{\varepsilon} = 0$), Algorithm \ref{alg.1} can be further simplified
by taking into account the following property of the dispersion relation:
$E_{\varepsilon, -{\bm k}} = E_{\varepsilon {\bm k}}$ in Eq.~\ref{eq.enequasiparticle}.
This can be particularly useful in determining the system energy in the presence
of the BCS phase -- i.e. either in complete absence of external magnetic fields
or when only weak fields are present.
A more general approach to the reduction of the execution time of our algorithm is to exploit the large degree
of parallelism inherent in the problem. In fact, Algorithm~\ref{alg.1} can be classified as
,,embarrassingly parallel'' since the vast majority of computation can be carried out independently
for all combinations of $\{{\bm q}_{\varepsilon}, {\bm k}, \Delta_{\varepsilon}\}$. For simplicity, in this paper
we concentrate on optimizing the inner loop, as all the presented methods apply to the outer
loop in a similar fashion.
We present two approaches to this problem. The first is to parallelize the execution
of the serial loop over $\Delta_{\varepsilon}$ with OpenMP to fully utilize all available CPU
cores. This has the advantage of simplicity, as the implementation requires minimal
changes to the original (serial) code.
The second approach is to implement Algorithm~\ref{alg.1} on a GPU using the CUDA
environment. Modern GPUs are capable of simultaneously executing thousands
of threads in SIMT (Same Instruction, Multiple Threads) mode. From a programmer's
point of view, all the threads are laid out in a 1-, 2- or 3-dimensional grid and
are executing a \emph{kernel function}. The grid is further subdivided into blocks (groups
of threads), which are handled by a physical computational subunit of the GPU (the so-called
streaming multiprocessor). Threads within a block can exchange data efficiently during execution,
but cross-block communication can only take place through global GPU memory, which is significantly
slower.
\begin{figure}[!h]
\begin{center}
\includegraphics{algor}
\caption{Schematic representation Algorithm \ref{alg.1} mapped to GPU hardware.}
\label{fig.alogr}
\end{center}
\end{figure}
To fully utilize the GPU hardware, we split Algorithm~\ref{alg.1} into three steps. In the
first step, we execute the \texttt{ComputeFreeEnergy} kernel (Algorithm~\ref{alg.2}) on a
3D grid $N_x \times N_y \times \Delta_{max}$. To take advantage of the efficient intra-block
communication, we also carry out partial sums within the block (corresponding to a subset
of values spanning $N_x$) using the parallel sum-reduction algorithm.~\cite{sum}
In the second step, we execute the sum-reduction algorithm again on the partial sums
that were generated by Algorithm~\ref{alg.2}. In the third and last step, we copy
the output of step 2 from GPU memory to host memory, and look for the
value of $\Delta_i$ corresponding to the lowest free energy with a linear search. Depending on the
exact configuration of the kernels in step 1 and 2, the summation might not
be complete at the beginning of step 3. If this is the case, we carry out
the remaining summation within the serial loop computing $\Delta_i$. With
block sizes of 128 and 1024 used for the kernels in steps 1 and 2, we can
sum up to $2^{17}$ terms in parallel on the GPU. We found that the remaining
summation was not worth the overhead of carrying it out on the GPU. Should
this not be the case for some larger problems, further parallel execution
can be trivially achieved by repeating step 2 one more time.
\begin{algorithm*}
\caption{The \texttt{ComputeFreeEnergy} CUDA kernel.}
\label{alg.2}
\begin{algorithmic}[1]
\STATE compute $\Delta_{i}$ and ${\bm k}$ corresponding to the current thread
\STATE load $E_{i{\bm k}\uparrow}$ and $E_{i,-{\bm k}+{\bm q}_{\varepsilon}\downarrow}$ from global memory (precomputed by a separate kernel)
\STATE compute $\mathcal{E}_{i{\bm k}}^{\pm}$ and $\Omega_{i}$
\STATE sum $\Omega_{i}$ for a range of ${\bm k}$ corresponding to one block of threads
\STATE save the partial sum from the previous step in global GPU memory
\end{algorithmic}
\end{algorithm*}
\subsection{Performance evaluation}
To test our approach, we executed Algorithms~\ref{alg.1}~and~\ref{alg.2}
on Linux machines with the following hardware:
\begin{itemize}
\item CPU: Intel(R) Core(TM) i7-3960X CPU @ 3.30GHz -- 6 cores / 12 threads,
\item GPU: NVIDIA Tesla K40 (GK180) with the SM clock set to 875 MHz.
\end{itemize}
The programs were run for a single value of ${\bm q}_i$ and 200 values
of $\Delta_i$. Calculations were done for a square lattice of size $N \times N$
for various values of $N$. The execution times (including only the computation
part of the code, and excluding any time spent on startup or input/output)
are presented in Figure~\ref{fig.scaling}.
\begin{figure}[!h]
\begin{center}
\includegraphics[width=\textwidth]{scaling.ps}
\caption{(Color on-line) Left panel: Execution time of Algorithm~\ref{alg.1} and~\ref{alg.2} for one vector ${\bm q}_{i}$.
Right panel: speedup factors for all configurations at $N = 2000$. The last 3 case
names correspond to runs of the same CUDA C code in double precision (DP), single precision (SP),
and single precision with fast intrinsic functions (SPFM).
All versions of the Fortran code used double precision calculations.}
\label{fig.scaling}
\end{center}
\end{figure}
Comparing the best CPU execution time (with OpenMP) to the GPU Fortran code using OpenACC,
we find a speedup factor of $5.8$ in the limit of large lattices.
The custom GPU code shows slightly better performance, with a $7$x speedup for the double precision
version, and additional speedup factors of $1.7$ for single precision, and $1.6$ for intrinsic
functions. When taken together,
the fastest GPU version is $19.2$ times faster than the OpenMP code and $119.1$ times faster
than the serial CPU code utilizing only a single core.
It is remarkable that the original Fortran code enhanced with OpenACC annotations provides
performance comparable to a manual implementation in CUDA C. This result shows the power
of appropriately used annotations marking parallelizable regions of the code. While still
requiring explicit input from the programmer and a good understanding of the structure of the code,
this approach is in practice significantly faster than writing the program from scratch in CUDA C
and dealing with low level details of GPU programming and resource allocation. This conclusion however only
applies in the limit of large lattices (see the left panel in Figure~\ref{fig.scaling}).
For smaller ones, the CUDA C code can be seen to be noticeably faster than OpenACC, which
is likely caused by the automatically generated GPU code introducing unnecessary overhead.
It should be noted that the last two speedup factors were achieved by trading off precision of
calculations for performance -- e.g. intrinsic functions are faster, but less precise implementations
of transcendental functions.
In our tests, we obtained the same results with all three approaches. This might not be true for
some other systems though, so we advise careful experimentation. With a factor of 2.8x between
the most and least precise method, it might also be worthwhile to run larger parameter scans
at lower precision and then selectively verify with double precision calculations.
\section{Summary}
\label{sec.summary}
The rich phenomenology and the subtle competing and interplaying phenomena of
high-$T_{C}$ materials such as FeSC (Section \ref{sec.intro}), require us to
probe fine regimes and precisely determine possible experimental signatures of
exotic phases such as FFLO (Section \ref{sec.theory_ph}).
By conducting our calculations in momentum space, and by
fully exploiting the symmetries of
the system, we are able to increase the size of the studied system by two
orders of magnitude compared to previously reported results
and practically eliminate finite size effects.
The cost is borne by the increased complexity of the efficient
custom-tailored GPU implementation, described in Section \ref{sec.algorithm}.
Our method shown here on the example of an iron-based multi-band superconductor
exhibiting a FFLO phase, can also be used in calculations of the ground state
in standard BCS-type superconductors.
Overall, we achieved a 19x speedup compared to the CPU implementation (119x
compared a single CPU core). In the spectrum of GPU-accelerated results in physics,
this puts us towards the higher end, with the highest speedups
being $\approx 700$x for compute-bound problems with large inherent parallelism.~\cite{januszewski.kostur.09}
\section*{Acknowledgments}
D.C. is supported by the Forszt PhD fellowship, co-funded by the European Social Fund.
B.G. is supported by the NCN project DEC-2011/01/N/ST3/02473.
The authors would like to thank NVIDIA for providing hardware resources for development and benchmarking.
|
\section{Introduction}
\indent In 2013, Lu and Ren \cite {luren} discussed anticipated backward stochastic differential equations (anticipated BSDEs) on finite state, continuous time Markov chains:
\begin{equation*}
\left \{ \begin{array}{ll}
-dY_t = f(t,Y_t,Z_t,Y_{t+\delta(t)}, Z_{t+\zeta(t)} )dt- Z_t' dM_t,&t\in [ 0,T]; \\
\hskip.57cmY_{t}= \xi_t,&t\in[T,T+K];\\
\hskip.54cmZ_t=\eta_t,&t\in [ T,T+K],\end{array} \right.
\end{equation*}
where $\xi_.,\eta_.$ are called the terminal conditions, $f$ is called the driver and $M=\{M_t\in\mathbb{R}^N,~t\geq 0\}$ is a martingale coming from the semimartingale representation of a continuous time
Markov chain. In the same paper, they established the existence and uniqueness of the solutions to this kind of equation.\\
\indent In this paper, we provide more properties of the solutions to anticipated BSDEs with Markov chain noise. First we study how to bound the solutions by the terminal conditions and the driver. Then we deduce there exists a duality between these equations and stochastic differential delayed equations (SDDEs) on Markov chains. This means anticipated BSDEs with Markov chain noise exist naturally.\\
\indent Lu and Ren \cite{luren} also established a comparison theorem for
1-dimensional anticipated BSDEs on Markov chains
under conditions involving not only the two drivers but also
the two solutions. We shall provide a comparison result involving
conditions only on the two drivers. This means the comparison result is easier to apply. For example, penalization of reflected anticipated BSDEs on Markov chains, the converse comparison theorem for anticipated BSDEs on Markov chains and so on can be established using our comparison result.\\
\indent The paper is organized as follows. In Section 2, we introduce the model and give some preliminary results. Section 3 provides a new proof of the solutions to anticipated BSDEs on Markov chains and an estimate of the solutions. In Section 4 we show the duality between these equations and SDDEs on Markov chains. We establish in Section 5 a comparison result for one-dimensional anticipated BSDEs with Markov chain noise.
\section{The Model and Some Preliminary Results}\label{prelim}
Consider a finite state Markov chain. Following the papers
\cite{RE1} and \cite{RE2} of van der Hoek and Elliott, we assume the
finite state Markov chain $X=\{X_t, t\geq 0 \}$ is defined on the
probability space $(\Omega,\mathcal{F},P)$ and the state space of
$X$ is identified with the set of unit vectors $\{e_1,e_2\cdots,e_N\}$ in
$\mathbb{R}^N$, where $e_i=(0,\cdots,1\cdots,0) ' $ with 1 in the
$i$-th position. Then the Markov chain has the semimartingale
representation:
\begin{equation}\label{semimartingale}
X_t=X_0+\int_{0}^{t}A_sX_sds+M_t.
\end{equation}
Here, $A=\{A_t, t\geq 0 \}$ is the rate matrix of the chain $X$ and
$M$ is a vector martingale (See Elliott, Aggoun and Moore
\cite{RE4}).
We assume the elements $A_{ij}(t)$ of $A=\{A_t, t\geq 0 \}$ are bounded. Then the martingale $M$ is square integrable.\\
\indent Take $\mathcal{F}_t=\sigma\{X_s ; 0\leq s \leq t\}$ to be
the $\sigma$-algebra generated by the Markov process $X=\{X_t\}$
and $\{\mathcal{F}_t\}$ to be its filtration. Since $X$ is right continuous and has left
limits, (written RCLL), the filtration $\{\mathcal{F}_t\}$ is also
right-continuous. The following is given in Elliott \cite{elliott} as Lemma 2.21 :
\begin{lemma}\label{indistinguish}
Suppose $V$ and $Y$ are real valued processes defined on the same probability space $(\Omega, \mathcal{F},P)$ such that for every $t \geq 0$, $V_t = Y_t$, a.s. If both processes are right continuous, then $V$ and $Y$ are indistinguishable, that is:
$$P(V_t = Y_t,~ \text{for any}~ t\geq 0)=1. $$
\end{lemma}
\indent The following product rule for semimartingales can be found in \cite{elliott}.
\begin{lemma}[Product Rule for Semimartingales]\label{ItoPR}
Let $Y$ and $Z$ be two scalar RCLL semimartingales, with no
continuous martingale part. Then
\begin{equation*}
Y_tZ_t = Y_TZ_T - \int_t^T Y_{s_-} dZ_s - \int_t^T Z_{s_-} dY_s -
\sum_{t < s \leq T} \Delta Z_s \Delta Y_s.
\end{equation*}
Here, $\sum\limits_{0< s \leq t} \Delta Z_s \Delta Y_s$ is the optional
covariation of $Y_t$ and $Z_t$ and is also written as $[Z,Y]_t$.
\end{lemma}
For our (vector) Markov chain $X_t \in \{e_1,\cdots,e_N\}$,
note that $X_t X'_t = \text{diag}(X_t)$. Also, $dX_t= A_t X_tdt+ dM_t$. By Lemma \ref{ItoPR}, we know for $t\in[0,T],$
\begin{align}\label{1}
\nonumber
X_tX'_t &= X_0X'_0 + \int_0^t X_{s-} dX'_s + \int_0^t (dX_{s}) X'_{s-} + \sum_{0 < s \leq t} \Delta X_s \Delta X'_s \\
\nonumber
&= \text{diag}(X_0) + \int_0^t X_s (A_sX_s)' ds + \int_0^t X_{s-} dM'_s + \int_0^t A_s X_s X'_{s-} ds\\
\nonumber
& ~~ + \int_0^t (dM_s ) X'_{s-} + [X,X]_t \\
\nonumber
& = \text{diag} (X_0) + \int_0^t X_s X'_s A'_s ds + \int_0^t X_{s-} dM'_s + \int_0^t A_s X_s X'_{s-} ds \\
&~~ + \int_0^t (dM_s) X'_{s-} + [X,X]_t
- \left\langle X,X\right\rangle_t + \left\langle X,X \right\rangle_t.
\end{align}
Here, $\left\langle X, X\right\rangle$ is the unique predictable $N\times N$ matrix process such that
$[X,X]-\left\langle X,X \right\rangle$ is a matrix valued martingale. Write
\begin{equation*}
L_t = [X,X]_t - \left\langle X,X\right\rangle_t, \quad t \in [0,T].
\end{equation*}
However, we also have
\begin{equation}\label{2}
X_tX'_t = \text{diag} (X_t) = \text{diag}(X_0) + \int_0^t \text{diag} (A_s X_s) ds +
\int_0^t \text{diag}(M_s).
\end{equation}
Equating the predictable terms in \eqref{1} and \eqref{2}, we have
\begin{equation}\label{3}
\left\langle X, X\right\rangle_t = \int_0^t \text{diag}(A_sX_s) ds - \int_0^t \text{diag}(X_s)
A'_s ds - \int_0^t A_s \text{diag}(X_s) ds.
\end{equation}
\indent For $n\in\mathbb{N}$, denote for $\phi \in \mathbb{R}^n$, the Euclidean norm $|\phi|_{n}=\sqrt{\phi'\phi}$ and for $\psi \in \mathbb{R}^{n\times n}$, the matrix norm $\|\psi\|_{n\times n}=\sqrt{Tr(\psi'\psi)}$.\\
\indent Let $\Psi_.$ be the matrix
\begin{equation}\label{Psi}\Psi_t = \text{diag}(A_tX_t)- \text{diag}(X_t)A'_t - A_t \text{diag}(X_t).
\end{equation}
Then $d\langle X,X\rangle_t=\Psi_tdt.$ For any $t>0$, Cohen and Elliott \cite{Sam1,Sam3}, define the semi-norm $\|.\|_{X_t}$, for
$C, D \in \mathbb{R}^{N\times K}$ as :
\begin{align*}
\left\langle C, D\right\rangle_{X_t} & = Tr(C' \Psi_tD), \\[2mm]
\|C\|^2_{X_t} & = \left\langle C, C\right\rangle_{X_t}.
\end{align*}
It follows from equation \eqref{3} that
\[\int_t^T \|C\|^2_{X_s} ds = \int_t^T C' d\left\langle X, X\right\rangle_s C.\]
\indent The following lemma comes from Yang, Ramarimbahoaka and Elliott \cite{zhedim}.
\begin{lemma}\label{normbound}
For any $B\in \mathbb{R}^N$,
$$ ~~~~\|B\|_{X_t} \leq \sqrt{3m} |B|_N, ~~\text{ for any }t\in[0,T],$$
where $m>0$ is the bound of $\|A_t\|_{N\times N}$, for any $t\in[0,T]$.
\end{lemma}
\indent Lemma \ref{Z2} can be found in Ramarimbahoaka, Yang and Elliott \cite{RYE}.
\begin{lemma}\label{Z2}
For a predictable process $Z_.$ in $\mathbb{R}^N$ verifying
$E \left[ \int_0^t \|Z_u\|^2_{X_u} du\right]$ $ < \infty,$
we have:
\begin{equation*}
E \left[\left(\int_0^t Z'_{u} dM_u \right)^2\right] = E \left[ \int_0^t \|Z_u\|^2_{X_u} du\right].
\end{equation*}
\end{lemma}
\indent Denote by $\mathcal{P}$, the $\sigma$-field generated by the processes defined on $(\Omega, P, \mathcal{F})$ which are predictable with respect to the filtration $\{\mathcal{F}_t\}_{t \in [0,\infty)}$. For any $t,s,r\in[0,\infty)$, $t\leq r \leq s$, consider the following spaces: \\[2mm]
$ L^2(\mathcal{F}_t;\mathbb{R}): =\{ \mathbb{R} \text{-valued}~ \mathcal{F}_t \text{-measurable random variables such that~} E[|\xi|^2]$\\$~~~~~~~~~~~~~~~~~~< \infty\};$\\[2mm]
$L^2_{\mathcal{F}}(t,s;\mathbb{R}): =\{\phi_.:[t,s]\times\Omega\rightarrow\mathbb{R};~ \phi_.~\text{adapted and RCLL process with}\\~~~~~~~~~~~~~~~~~~~~E[\int^s_t|\phi(t)|^2dt]<\infty\}$;\\[2mm]
$H^2(t,s;\mathbb{R}^N)=\{\phi_.:[t,s]\times\Omega\rightarrow\mathbb{R}^N;~ \phi_.\in\mathcal{P}~\mbox{with }E[\int^s_t\|\phi(t)\|_{X_t}^2dt]<\infty\};$\\[2mm]
$ H^2(\mathcal{F}_r;\mathbb{R}^N): =\{ \varphi_r $ is an $\mathbb{R}^N \text{-valued}~ \mathcal{F}_r \text{-measurable random variables with~} \varphi_.$\\$~~~~~~~~~~~~~~~~~~~~~\in H^2(t,s;\mathbb{R}^N)\}.$\\[2mm]
\indent Consider the following one-dimensional BSDE with the Markov chain noise:
\begin{equation}\label{BSDEMC}
Y_t = \xi + \int_t^T f(s, Y_s, Z_s ) ds -\int_t^T Z'_{s} dM_s
,~~~~~t\in[0,T].
\end{equation}
Here the terminal condition $\xi$ and the coefficient $f$ are known. Lemma \ref{existence} (Theorem 6.2 in Cohen and Elliott \cite{Sam1})
gives the existence and uniqueness result for solutions to the BSDEs
driven by Markov chains:
\begin{lemma}\label{existence}
Assume $\xi \in L^2(\mathcal{F}_T;\mathbb{R})$ and the predictable
function $f: \Omega \times [0, T] \times \mathbb{R} \times
\mathbb{R}^N \rightarrow \mathbb{R}$ satisfies a Lipschitz
condition, in the sense that there exist constants $l_1,l_2>0$ such
that for each $y_1,y_2 \in \mathbb{R}$ and $z_1,z_2 \in
\mathbb{R}^{N}$,
\begin{equation}\label{Lipchl}
|f(t,y_1,z_1) - f(t, y_2, z_2)| \leq l_1 |y_1-y_2| + l_2\|z_1
-z_2\|_{X_t}.
\end{equation}
We also assume $f$ satisfies
\begin{equation}\label{finite}
E [ \int_0^T |f(t,0,0)|^2 dt] <\infty.
\end{equation}
Then there exists a solution $(Y_., Z_.)\in L^2_{\mathcal{F}}(0,T;\mathbb{R})\times H^2(0,T;\mathbb{R}^N)$
to the BSDE (\ref{BSDEMC}). Moreover, this solution is
unique up to indistinguishability for $Y_.$ and equality $d\langle
M,M\rangle_t$ $\times\mathbb{P}$-a.s. for $Z_.$.
\end{lemma}
\indent Campbell and Meyer \cite{campbell} gave the following definition:
\begin{definition}\label{defMoore}
The Moore-Penrose pseudoinverse of a square matrix $Q$ is the matrix $Q^{\dagger}$ satisfying the properties:\\[2mm]
1) $QQ^{\dagger}Q = Q$ \\[2mm]
2) $Q^{\dagger}QQ^{\dagger} = Q^{\dagger}$ \\[2mm]
3) $(QQ^{\dagger})' = QQ^{\dagger}$ \\[2mm]
4) $(Q^{\dagger}Q)'=Q^{\dagger}Q.$
\end{definition}
\indent Recall the matrix $\Psi_.$ given by \eqref{Psi}. We adapt Lemma 3.5 in Cohen and Elliott \cite{Sam3} for our framework as follows:
\begin{lemma}\label{sam35}
For any driver satisfying \eqref{Lipchl} and \eqref{finite}, for any $Y_.$ and $Z_.$,
\[P(f(t,Y_{t-},Z_t) = f(t,Y_{t-}, \Psi_t\Psi_t^{\dagger} Z_t), ~\text{ for all}~ t \in [0,+\infty])=1\]
and
\[\int_0^t Z'_s dM_s = \int_0^t (\Psi_s\Psi^{\dagger}_s Z_s)' dM_s.\]
Therefore, without any loss of generality, we shall assume $Z_.=(\Psi\Psi^{\dagger} Z)_.$.
\end{lemma}
\begin{ass}\label{ass111}
Assume the Lipschitz constant $l_2$ of the driver $f$ given in \eqref{Lipchl} satisfies $$~~~~~l_2\|\Psi_t^{\dagger}\|_{N \times N} \sqrt{6m}\leq 1, ~~~\text{ for any }~t \in [0,T],$$ where $\Psi_.$ is given in \eqref{Psi} and $m>0$ is the bound of $\|A_t\|_{N\times N}$, for any $t\in[0,T]$.
\end{ass}
\indent The following lemma, which is a comparison result for BSDEs driven by a Markov chain, is found in Yang, Ramarimbahoaka and Elliott \cite{zhedim}.
\begin{lemma} \label{CT'} For $i=1,2,$ suppose $(Y_.^{(i)},Z_.^{(i)})$ is the solution of the
BSDE:
$$Y^{(i)}_t = \xi_i + \int_t^T f_i(s, Y^{(i)}_s, Z^{(i)}_s ) ds
- \int_t^T (Z_{s}^{(i)})' dM_s,\hskip.4cmt\in[0,T].$$
Assume $\xi_1,\xi_2\in L^2(\mathcal{F}_T;\mathbb{R})$, and $f_1,f_2:\Omega \times [0,T]\times \mathbb{R}\times \mathbb{R}^N \rightarrow \mathbb{R}$ satisfy conditions such that the above two BSDEs have unique solutions. Moreover assume $f_1$ satisfies \eqref{Lipchl} and Assumption \ref{ass111}.
If $\xi_1 \leq \xi_2 $, a.s. and $f_1(t,Y_t^{(2)}, Z_t^{(2)}) \leq f_2(t,Y_t^{(2)}, Z_t^{(2)})$, a.e., a.s., then
$$P( Y_t^{(1)}\leq Y_t^{(2)},~~\text{ for any } t \in [0,T])=1.$$
\end{lemma}
\indent Lemma \ref{Duality} is the duality between linear stochastic differential equations (SDEs) and linear BSDEs both on on Markov chains in \cite{zhedim}.
\begin{lemma}\label{Duality} Suppose $a_.\in
L^{2}_\mathcal{F}(0,T;\mathbb{R}),$ $f_.\in
L^{2}_\mathcal{F}(0,T;\mathbb{R}),~b_.\in
L^2_{\mathcal{F}}(0,T;\mathbb{R}^{1\times N}),$ and $a_.,b_.$ are uniformly bounded.
For $t\in [0, T]$, consider 1-dimensional SDE on the Markov chain
\begin{equation*}
\begin{cases}
dU_s= U_s a_s ds + U_{s-} b_{s-} (\Psi_s^{\dagger})'dM_s, ~~ s \in [t,T];\\
~~U_t =1.
\end{cases}
\end{equation*}
Then the solution of the one-dimensional linear BSDE on Markov chain
\begin{align*}
Y_t = \xi + \int_t^T (a_sY_s+b_sZ_s+f_s) ds - \int_t^T Z'_s dM_s, ~~~ t \in [0,T]
\end{align*} satisfies
\begin{equation*}
P (Y_t = E [ \xi U_T + \int_t^T f_s U_s ds | \mathcal{F}_t], ~ \text{for any}~ t\in [0,T])=1.
\end{equation*}
\end{lemma}
\section{An estimate of the solutions to anticipated BSDEs with Markov chain model}
\indent In order to make this
paper self-contained, we shall provide a proof of the existence and uniqueness of solutions of anticipated BSDEs with Markov chain noise by using the
fixed point theorem, rather than using Picard iterations as in Lu and Ren \cite{luren}.\\
\indent Consider the following anticipated BSDE on the Markov chain:
\begin{equation}\label{bfewwww10}
\left \{ \begin{array}{ll}
-dY_t = f(t,Y_t,Z_t,Y_{t+\delta(t)}, Z_{t+\zeta(t)} )dt- Z_t' dM_t,&t\in [ 0,T]; \\
\hskip.57cmY_{t}= \xi_t,&t\in[T,T+K];\\
\hskip.54cmZ_t=\eta_t,&t\in [ T,T+K].\end{array} \right.
\end{equation}Here $M$ is defined in (\ref{semimartingale}), $\delta(\cdot)$ and $\zeta(\cdot)$ are two $\mathbb{R}^+$-valued continuous functions defined on $[0,T]$ such that\\
${\bf (i)}$ there exists a constant $K\geq0$ such that
for any $s\in[0,T],$
$$ s+\delta(s)\leq T+K,\hskip2cm s+\zeta(s)\leq T+K;$$
${\bf (ii)}$ there exists a constant $L\geq0$ such that
for any $t\in[0,T]$ and a nonnegative and integrable function $g(\cdot)$,
$$ \int ^T _tg(s+\delta(s))ds\leq L\int ^{T+K} _tg(s)ds;$$
$$\int ^T _tg(s+\zeta(s))ds\leq L\int ^{T+K} _tg(s)ds.$$
\indent Assume that for any $s\in
[0,T],~f(s,\omega,y,z,\xi,\eta):\Omega\times \mathbb{R}\times
\mathbb{R}^{N}\times
L^{2}(\mathcal{F}_r;\mathbb{R})\times
H^{2}(\mathcal{F}_{r'};\mathbb{R}^{N})\longrightarrow
L^{2}(\mathcal{F}_{s},\mathbb{R}),$ where $r,r'\in[s,T+K]$, and $f$ satisfies the following conditions \\
${\bf (H1)}$ There exist two constants $c_1,c_2>0,$ such that for any $s\in
[0,T],~y,y'\in\mathbb{R},$ $z,z'\in\mathbb{R}^{N},~\xi_{\cdot},\xi'_{\cdot}\in L^2_{\mathcal{F}}
(s,T+K;\mathbb{R}),~\eta_{\cdot},\eta'_{\cdot}\in
H^2 (s,T+K;\mathbb{R}^N),~r,\bar{r}\in[s,T+K],$ we have
$$\begin{array}{ll}
\hskip.46cm|f(s,y,z,\xi_r,\eta_{\bar{r}})-f(s,y',z',\xi'_r,\eta'_{\bar{r}})| \\[0.2cm]
\leq
c_1(|y-y'|+E^{\mathcal{F}_s}[|\xi_r-\xi'_r|])+c_2(\|z-z'\|_{X_s}+E^{\mathcal{F}_s}[\|\eta_{\bar{r}}-\eta'_{\bar{r}}\|_{X_s}]).
\end{array}
$$
${\bf (H2)}$ $E[\int_{0} ^{T} |f(s,0,0,0,0)|^2ds] < \infty$.\\[0.2cm]
\indent Lu and Ren \cite{luren} proved the result of Theorem \ref{bft3000000} below. Here, we give an alternative proof.
\begin{theorem}\label{bft3000000} Suppose that $f$ satisfies $(H1)$ and $(H2),$ $\delta,\zeta$ satisfy $(i)$ and $(ii).$ Then for arbitrary given terminal
conditions $\xi_{\cdot}\in L^2_{\mathcal{F}}
(T,T+K;\mathbb{R}),$ $\eta_{\cdot}\in H^2
(T,T+K;\mathbb{R}^{N})$, the anticipated BSDE (\ref{bfewwww10}) has a
unique solution, i.e., there exists a unique pair of
stochastic processes $(Y.,Z.)\in
L^2_\mathcal{F}(0,T+K;\mathbb{R})\times H^2(0,T+K;\mathbb{R}^{N})$ satisfying equation (\ref{bfewwww10}). Moreover, this solution is
unique up to indistinguishability for $Y$ and equality $d\langle
X,X\rangle_t$ $\times\mathbb{P}$-a.s. for $Z$.\end{theorem}
\noindent{\bf Proof.} Set $c:=\max\{c_1,c_2\}$. We fix
$\beta=16c^2(L+1),$ where $L$ is given in $(ii)$. Now we introduce a norm in the Banach space
$L^2_\mathcal{F}(0,T+K;\mathbb{R}):$\\[2mm]
\indent\indent\indent\indent\indent\indent$\|\nu_.\|_{L^2}=(E[\int ^{T+K} _0|\nu_s|^{2}e^{\beta s}ds])^{\frac{1}{2}}.$\\[2mm]
Define an equivalence class of $\varphi_.$ by $[\varphi_.]:=\{\psi_.;~E[\int_0^{T+K} \|\psi_t-\varphi_t\|^2_{X_s} ds]=0 \}$ and denote the factor space of equivalence classes of processes in $H^2(0,T+K;\mathbb{R}^{N})$ by $\hat{H}^2(0,T+K;\mathbb{R}^{N}):=\{[\varphi_.];~\varphi_.\in H^2(0,T+K;\mathbb{R}^{N}) \}.$ Then $\hat{H}^2(0,T+K;\mathbb{R}^{N})$ is a Banach space with the norm \\[2mm]
\indent\indent\indent\indent\indent\indent$\|\mu_.\|_{\hat{H}^2}=(E [ \int_0^{T+K}\|\mu_s\|^2_{X_s}e^{\beta s} ds ])^{\frac{1}{2}}.$\\[2mm]
Set
$$ \left \{ \begin{array}{ll}Y_t =\xi_T + \int ^T _t f(s,y_s,z_s,y_{s+\delta(s)},z_{s+\zeta(s)} )ds - \int^{T} _{t} Z_s'dM_s,&t\in[0,T]; \\
Y_{t} = \xi_t,&t\in[T,T+K];\\
Z_t=\eta_t,&t\in[T,T+K].\end{array} \right.$$ By Lemma \ref{existence}, we know for any
$(y_{\cdot},z_{\cdot}) \in L^2 _\mathcal{F}(0,T+K;\mathbb{R})\times H^2(0,T+K;\mathbb{R}^{N})$,
the above equation has a solution $(Y_{\cdot},Z_{\cdot})\in L^2 _\mathcal{F}(0,T+K;\mathbb{R})\times H^2(0,T+K;$ $\mathbb{R}^{N})$, moreover, this solution is
unique up to indistinguishability for $Y_{\cdot}$ and equality $d\langle
X,X\rangle_t\times\mathbb{P}$-a.s. for $Z_{\cdot}$. That is, this solution is
unique up to indistinguishability for $(Y_{\cdot},Z_{\cdot})\in L^2 _\mathcal{F}(0,T+K;\mathbb{R})\times \hat{H}^2(0,T+K;\mathbb{R}^{N})$. Define a mapping $h
:L^2 _\mathcal{F}(0,T+K;\mathbb{R})\times \hat{H}^2(0,T+K;\mathbb{R}^{N})\longrightarrow L^2 _\mathcal{F}(0,T+K;$ $\mathbb{R})\times \hat{H}^2(0,T+K;\mathbb{R}^{N})$ such
that $h[(y_{\cdot},z_{\cdot})]=(Y_{\cdot},Z_{\cdot})$. Now we prove
that $h$ is a contraction mapping under the norm
$\|\cdot\|_{L^2}+\|\cdot\|_{\hat{H}^2}$. For two arbitrary elements
$(y_{\cdot},z_{\cdot})$ and $(y'_{\cdot},z'_{\cdot})$ in
$L^2 _\mathcal{F}(0,T+K;\mathbb{R})\times \hat{H}^2(0,T+K;\mathbb{R}^{N})$
set $(Y_{\cdot},Z_{\cdot})=h[(y_{\cdot},z_{\cdot})]$ and
$(Y'_{\cdot},Z'_{\cdot})=h[(y'_{\cdot},z'_{\cdot})]$. Denote their
differences by
$(\hat{y}_{\cdot},\hat{z}_{\cdot})=((y-y')_{\cdot},(z-z')_{\cdot})$ and $(\hat{Y}_{\cdot},\hat{Z}_{\cdot})=((Y-Y')_{\cdot},(Z-Z')_{\cdot}).$
Applying the product rule (Lemma \ref{ItoPR}) to $|\hat{Y}_t|$, we have
\begin{align*}
|\hat{Y}_t|^2
& = -2 \int_t^T \hat{Y}_{s-} d\hat{Y}_{s} - \sum_{t \leq s \leq T} \Delta \hat{Y}_s\Delta \hat{Y}_s \\
&= -2 \int_t^T \hat{Y}_s (f(s,y_s,z_s,y_{s+\delta(s)},z_{s+\zeta(s)})-f(s,y'_s,z'_s,y'_{s+\delta(s)},z'_{s+\zeta(s)})) ds \\
& \quad -2 \int_t^T \hat{Y}_{s-} (\hat{Z}_s)' dM_s - \sum_{t \leq s \leq T} \Delta \hat{Y}_{s}\Delta \hat{Y}_{s}.
\end{align*}
Also
\begin{align*}
&\sum_{t\leq s \leq T} \Delta \hat{Y}_{s} \Delta \hat{Y}_{s}
= \sum_{t\leq s \leq T}( (\hat{Z}_s)' \Delta X_s)( ( \hat{Z}_s)'\Delta X_s ) = \sum_{t\leq s \leq T} (\hat{Z}_s )' \Delta X_s \Delta X_s' \hat{Z}_s \\
&= \int_t^T ( \hat{Z}_s )' (dL_s + d\left\langle X,X\right\rangle_s) \hat{Z}_s= \int_t^T (\hat{Z}_s)' dL_s\hat{Z}_s + \int_t^T \|\hat{Z}_s \|_{X_s}^2 ds.
\end{align*}
Applying It\^{o}'s formula to $e^{\beta s} |\hat{Y}_s|^2$ for $s\in[0,T]$ and then taking the expectation:
\begin{align*}
&E [ |\hat{Y}_0|^2 ]+ E [ \int_0^T \beta |\hat{Y}_s|^2 e^{\beta s} ds ] + E [ \int_0^T\|\hat{Z}_s\|^2_{X_s}e^{\beta s} ds ] \\
& = 2 E [ \int_0^T \hat{Y}_s (f(s,y_s,z_s,y_{s+\delta(s)},z_{s+\zeta(s)})-f(s,y'_s,z'_s,y'_{s+\delta(s)},z'_{s+\zeta(s)})) e^{\beta s}ds ]\\
&\leq E [ \int_0^T (\frac{\beta}{2}|\hat{Y}_s|^2+\frac{2}{\beta} |f(s,y_s,z_s,y_{s+\delta(s)},z_{s+\zeta(s)})-f(s,y'_s,z'_s,y'_{s+\delta(s)},z'_{s+\zeta(s)})|^2)e^{\beta s} ds ].
\end{align*}
Since $\delta(s),\zeta(s)$ satisfy $(ii)$ and $f$ satisfies $(H1)$,
by the Fubini Theorem we have
$$\begin{array}{lll}
E[\int^T_0(\dfrac{\beta}{2}|\hat{Y}_s|^{2}+\|\hat{Z}_s \|_{X_s}^{2})e^{\beta
s}ds]\\[0.3cm]
\leq\dfrac{2c^2}{\beta}E[\int^{T}_0(|\hat{y}_s|+\|\hat{z}_s\|_{X_s}+E^{\mathcal{F}_s}[~|\hat{y}_{s+\delta(s)}|+\|\hat{z}_{s+\zeta(s)}\|_{X_s}])^2e^{\beta
s}ds]\\[0.3cm]
\leq\dfrac{8c^2}{\beta}E[\int^{T}_0(|\hat{y}_s|^2+\|\hat{z}_s\|_{X_s}^2+|\hat{y}_{s+\delta(s)}|^2+\|\hat{z}_{s+\zeta(s)}\|_{X_s}^2)e^{\beta
s}ds]\\[0.3cm]
\leq\dfrac{8c^2(L+1)}{\beta}E[\int^{T+K}_0(|\hat{y}_s|^{2}+\|\hat{z}_s\|_{X_s}^{2})e^{\beta
s}ds].
\end{array}$$
Note $\beta=16c^2(L+1),$ therefore
$$E[\int ^{T+K} _0(|\hat{Y}_s|^{2}+\|\hat{Z}_s\|_{X_s}^{2})e^{\beta
s}ds]\leq\frac{1}{2}E[\int
^{T+K}_0(|\hat{y}_s|^{2}+\|\hat{z}_s\|_{X_s})e^{\beta s}ds],
$$
or $$\|\hat{Y}_{\cdot}\|_{L^2}+\|\hat{Z}_{\cdot}\|_{\hat{H}^2}
\leq\frac{1}{\sqrt{2}}(\|\hat{y}_{\cdot}\|_{L^2}+\|\hat{z}_{\cdot}\|_{\hat{H}^2}).$$
Consequently $h$ is a strict contraction mapping on
$L^2_\mathcal{F}(0,T+K;\mathbb{R})\times \hat{H}^2(0,T+K;\mathbb{R}^{N}).$
It follows by the Fixed Point Theorem that the anticipated BSDE (\ref{bfewwww10})
has a unique solution $(Y_{\cdot},Z_{\cdot})\in
L^2_\mathcal{F}(0,T+K;\mathbb{R})\times \hat{H}^2(0,T+K;\mathbb{R}^{N}).$ That is, the solution $(Y_{\cdot},Z_{\cdot})\in
L^2_\mathcal{F}(0,T+K;\mathbb{R})\times H^2(0,T+K;\mathbb{R}^{N})$ is
unique up to indistinguishability for $Y$ and equality $d\langle
X,X\rangle_t\times\mathbb{P}$-a.s. for $Z$.
$\mbox{}\hfill\Box$\\[2mm]
\indent Our method allows us to find an estimate of the solution to equation
(\ref{bfewwww10}).
\begin{prop} Assume that $f$ satisfies $(H1)$ and $(H2),$ $\delta$ and
$\zeta$ satisfy $(i)$ and $(ii)$. Then there exists a
constant $C>0$ depending only on $c_1,c_2$ in $(H1),$ $L$ in $(ii)$, $K$ and
$T $, such that for each $\xi_{\cdot}\in L^2_{\mathcal{F}}
(T,T+K;\mathbb{R}),~\eta_{\cdot}\in H^2
(T,T+K;\mathbb{R}^{N}),$ the solution $(Y.,Z.)$ to the
anticipated BSDE (\ref{bfewwww10}) satisfies
\begin{equation}\label{bfe1000}\begin{array}{lll}
E[\sup \limits_{0\leq s\leq T }|Y_s|^{2}+\int^T_0\|Z_s\|_{X_s}^{2}ds]\\[0.3cm]
\leq
CE[|\xi_{T}|^{2}+\int^{T+K}_T(|\xi_s|^{2}+\|\eta_s\|_{X_s}^{2})ds+(\int^T_0|f(s,0,0,0,0)|ds)^{2}].
\end{array}\end{equation}
\end{prop}
\noindent{\bf Proof.} Set $c=:\max\{c_1,c_2\}$. let $\beta >0$ be an arbitrary constant. Using It$\hat{\text{o}}$'s formula for $e^{\beta t}|Y_t|^2$, we deduce
$$\begin{array}{ll}
E [ |Y_0|^2 ]+ E [ \int_0^T \beta |Y_s|^2 e^{\beta s} ds ] + E[ \int_0^T e^{\beta s}\|Z_s\|^2_{X_s} ds ] \\[3mm]
= E [ e^{\beta T} |\xi_T|^2 ] + 2 E [ \int_0^T e^{\beta s} Y_s f(s, Y_s, Z_s,Y_{s+\delta(s)}, Z_{s+\zeta(s)} ) ds ]\\[3mm]
\leq E [e^{\beta T}|\xi_T|^2] + 2 E[ \int_0^T e^{\beta s} |Y_s|\cdot|f(s,0,0,0,0)|ds]\\[3mm]
+2 E[ \int_0^T e^{\beta s}|Y_s|\cdot |f(s, Y_s, Z_s,Y_{s+\delta(s)}, Z_{s+\zeta(s)} )-f(s,0,0,0,0)|ds] \\[3mm]
\leq E [e^{\beta T}|\xi_T|^2] + 2 E[ \int_0^T e^{\beta s}|Y_s| \cdot|f(s, 0, 0,0,0)| ds]\\[3mm]
+2c E[ \int_0^T e^{\beta s}|Y_s|( |Y_s|+E^{\mathcal{F}_s}[|Y_{s+\delta(s)}|] + \|Z_s\|_{X_s}+E^{\mathcal{F}_s}[\|Z_{s+\zeta(s)} \|_{X_s}])ds] \\[3mm]
\leq E [e^{\beta T}|\xi_T|^2] +2 E [ \sup\limits_{s\in[0,T]}e^{\frac{1}{2}\beta s}|Y_s|\cdot \int_0^Te^{\frac{1}{2}\beta s} |f(s,0,0,0,0)| ds]\\[3mm]
+ (3c+3c^2+3c^2L) E[\int_0^T e^{\beta s}|Y_s|^2 ds] + cE[\int_0^T e^{\beta s}|Y_{s+\delta(s)}|^2 ds]\\[3mm]
+ \dfrac{1}{3} E [ \int_0^T e^{\beta s}\|Z_s\|^2_{X_s} ds]+ \dfrac{1}{3L} E [ \int_0^T e^{\beta s}\|Z_{s+\zeta(s)}\|^2_{X_s} ds]\\[3mm]
\leq E [e^{\beta T}|\xi_T|^2] +\alpha E [ \sup\limits_{s\in[0,T]}e^{\beta s}|Y_s|^2]+\dfrac{1}{\alpha}E[ (\int_0^Te^{\frac{1}{2}\beta s} |f(s, 0,0,0, 0)| ds)^2] \\[3mm]
+ (3c+3c^2+3c^2L+cL) E[\int_0^{T+K} e^{\beta s}|Y_s|^2 ds] + \dfrac{2}{3} E [ \int_0^{T+K} e^{\beta s}\|Z_s\|^2_{X_s} ds],
\end{array}
$$
where $\alpha>0$ is also an arbitrary constant. Set $\beta=3c+3c^2+3c^2L+cL+1$, we obtain
\begin{align}\label{apriory21}
&
E[ \int_0^T |Y_s|^2 e^{\beta s} ds] + \frac{1}{3} E [ \int_0^T e^{\beta s}\|Z_s\|^2_{X_s} ds] \\
\nonumber&\leq E [e^{\beta T}|\xi_T|^2] +\alpha E [ \sup\limits_{s\in[0,T]}e^{\beta s}|Y_s|^2]+\dfrac{1}{\alpha}E[ (\int_0^Te^{\frac{1}{2}\beta s} |f(s, 0,0,0, 0)| ds)^2]\\
\nonumber& + (3c+3c^2+3c^2L+cL) E[\int_T^{T+K} e^{\beta s}|\xi_s|^2 ds] + \dfrac{2}{3} E [ \int_T^{T+K} e^{\beta s}\|\eta_s\|^2_{X_s} ds].
\end{align}
Using Doob's inequality and Lemma \ref{Z2}, we know
\begin{align}\label{marsup}
\nonumber &E [ \sup_{0\leq t \leq T}|\int_t^T Z_s' dM_s|^2 ] =E [ \sup_{0\leq t \leq T}|\int_0^T Z_s' dM_s-\int_0^t Z_s' dM_s|^2]\\
\nonumber&\leq 2 E [|\int_0^T Z_s' dM_s|^2+ \sup_{0\leq t \leq T}|\int_0^t Z_s' dM_s|^2] \leq 10 E [ |\int_0^T Z'_sdM_s|^2]\\
&= 10 E[ \int_0^T \|Z_s\|^2_{X_s}ds]\leq10 E[ \int_0^T e^{\beta s}\|Z_s\|^2_{X_s}ds].
\end{align}
Because $ Y_t = \xi + \int_t^T f(s, Y_s, Z_s,Y_{s+\delta(s)}, Z_{s+\zeta(s)}) ds-\int_t^T Z_s' dM_s,~ 0 \leq t \leq T$, by (\ref{marsup}) we have
\begin{align}\label{supp}
\nonumber&E[\sup\limits_{0\leq t \leq T}|Y_t|^2]\\
\nonumber&\leq E[3|\xi_T|^2 + 3(\int_0^T| f(s, Y_s, Z_s,Y_{s+\delta(s)}, Z_{s+\zeta(s)}) |ds)^2 + 3 \sup\limits_{0\leq t \leq T}|\int_t^T Z_s' dM_s|^2] \\
\nonumber&\leq 3E[|\xi_T|^2]+ 30E[ \int_0^T e^{\beta s}\|Z_s\|^2_{X_s}ds] \\
\nonumber&+ 3E[( \int_0^T (|f(s,0,0,0,0)|+c|Y_s|+c\|Z_s\|_{X_s}+c|Y_{s+\delta(s)}|+c\|Z_{s+\zeta(s)}\|_{X_s})ds )^2 ] \\
\nonumber& \leq 3E[|\xi_T|^2]+ 30E[ \int_0^T e^{\beta s}\|Z_s\|^2_{X_s}ds]+ 15E[( \int_0^T |f(s,0,0,0,0)|ds)^2] \\
\nonumber&+ 15Tc^2E[\int_0^T( |Y_s|^2+\|Z_s\|^2_{X_s}+|Y_{s+\delta(s)}|^2+\|Z_{s+\zeta(s)}\|_{X_s}^2)ds] \\
\nonumber&\leq 3E[e^{\beta T}|\xi_T|^2]+15E[( \int_0^T |f(s,0,0,0,0)|ds)^2] \\
\nonumber&+ 15(2 +Tc^2+Tc^2L)E[\int_0^{T}e^{\beta s} (|Y_s|^2+\|Z_s\|^2_{X_s})ds]\\
&+ 15Tc^2LE[\int_T^{T+K}e^{\beta s} (|\xi_s|^2+\|\eta_s\|^2_{X_s})ds].
\end{align}
Set $\alpha=\dfrac{1}{90(2+Tc^2+Tc^2L)}$. Then by (\ref{apriory21}) and (\ref{supp}), we deduce there exists a constant $C>0$ depending on $T,c,L$ and $K$ such that (\ref{bfe1000}) holds.$\mbox{}\hfill\Box$
\section{Duality between SDDEs and Anticipated BSDEs on Markov chains}
It is well known that there is perfect duality
between SDEs and BSDEs (see El Karoui, Peng, and Quenez \cite{EPQ1997}). Cohen, Elliott \cite{Sam3} and Yang, Ramarimbahoaka, Elliott \cite{zhedim} showed duality between SDEs and BSDEs driven by Markov chains. In \cite{PengYang} Peng and Yang considered duality between SDDEs and anticipated
BSDEs. We now establish duality between SDDEs and anticipated
BSDEs with Markov chain noise.
\begin{lemma}\label{normcon}
For any $B \in \mathbb{R}^{N\times N}$,
$$ ~~~~\|B\|_{X_t}^2 \leq 3m \|B\|_{N\times N}^2, ~~\text{ for any }t\in[0,T],$$
where $m>0$ is the bound of $\|A_t\|_{N\times N}$, for any $t\in[0,T]$.
\end{lemma}
\noindent{\bf Proof.} Write $B=(B_1,B_2,\ldots,B_N)$, where $B_i\in \mathbb{R}^{N}$, for any $1\leq i\leq N$. Then $\|B\|_{N\times N}^2=\sum^{N}_{i=1}|B_i|_N^2.$
Noticing that for any $1\leq i\leq N$, $B_i'\Psi_tB_i\in\mathbb{R}$, we obtain for any $t\in[0,T]$,
\begin{align*}
\|B\|_{X_t}^2 &=Tr((B_1,B_2,\ldots,B_N)'\Psi_t(B_1,B_2,\ldots,B_N))\\[2mm]
& = Tr((B_1'\Psi_tB_1,B_2'\Psi_tB_2,\ldots,B_N'\Psi_tB_N))\\[2mm]
&=\sum^{N}_{i=1}B_i'\Psi_tB_i=\sum^{N}_{i=1}\|B_i\|_{X_t}^2.
\end{align*}
By Lemma \ref{normbound} we have
$\|B\|_{X_t}^2\leq 3m\sum^{N}_{i=1}|B_i|_{N}^2=3m\|B\|_{N\times N}^2,$ for any $t\in[0,T].$$\mbox{}\hfill\Box$
\begin{ass}\label{ass0}
Assume there exists a constant $l>0$ such that for any $t\in[0,T]$, $\|\Psi_t^{\dagger}\|_{N\times N}^2\leq l,$ where $\Psi$ is given in \eqref{Psi}.
\end{ass}
\begin{lemma}\label{Zequality} Suppose that Assumption \ref{ass0} holds, $f$ satisfies $(H1),(H2)$ and $\delta,\zeta$ satisfy $(i)$ and $(ii).$ Then for any $\xi_{\cdot}\in L^2_{\mathcal{F}}
(T,T+K;\mathbb{R}),\eta_{\cdot}\in H^2
(T,T+K;$ $\mathbb{R}^{N})$, the solution $Z.\in H^2(0,T+K;\mathbb{R}^{N})$ of the anticipated BSDE (\ref{bfewwww10}) satisfies $Z.=(\Psi\Psi^{\dag}Z).$, $d\langle
X,X\rangle_t$ $\times\mathbb{P}$-a.s.\end{lemma}
\noindent{\bf Proof.} Set $c:=\max\{c_1,c_2\}$. By the proof of Theorem \ref{bft3000000},We know there exists a sequence of
$\{(y^{(n)}_{\cdot},z^{(n)}_{\cdot});~n\in\mathbb{N}\}\subseteq L^2 _\mathcal{F}(0,T+K;\mathbb{R})\times H^2(0,T+K;$ $\mathbb{R}^{N})$ satisfying for any $n\in\mathbb{N},$
$$ \left \{ \begin{array}{ll}y^{(n+1)}_t =\xi_T + \int ^T _t f(s,y^{(n+1)}_s,z^{(n+1)}_s,y^{(n)}_{s+\delta(s)},z^{(n)}_{s+\zeta(s)} )ds - \int^{T} _{t} (z^{(n+1)}_s)'dM_s,&t\in[0,T]; \\
y^{(n+1)}_{t} = \xi_t,&t\in[T,T+K];\\
z^{(n+1)}_t=\eta_t,&t\in[T,T+K].\end{array} \right.$$
Then $$E[\int ^{T+K} _0(|y^{(n)}_s-Y_s|^{2}+\|z_s^{(n)}-Z_s\|_{X_s}^{2})e^{\beta
s}ds]\rightarrow0, ~~~~~\mbox{as }~~n\rightarrow\infty,
$$
where $(Y_{\cdot},Z_{\cdot})\in L^2 _\mathcal{F}(0,T+K;\mathbb{R})\times H^2(0,T+K;\mathbb{R}^{N})$ is the solution of the anticipated BSDE (\ref{bfewwww10}). Thus, $E[\int ^{T+K} _0\|z_s^{(n)}-Z_s\|_{X_s}^{2}
ds]\rightarrow0$ as $n\rightarrow\infty$. By Lemma \ref{sam35}, we have for any $n\in\mathbb{N}$, $E[\int ^{T+K} _0\|z_s^{(n)}-\Psi_s\Psi_s^\dag z_s^{(n)}\|_{X_s}^{2}ds]=0.$ Since the elements $A_{ij}(t)$ of $A=\{A_t, t\geq 0 \}$ are bounded, there exists a constant $m>0$ such that $\|A_t\|_{N \times N} \leq m$, for any $t\in[0,T]$. Noting $\Psi_t=\text{diag}(A_tX_t)- \text{diag}(X_t)A'_t - A_t \text{diag}(X_t)$ given in (\ref{Psi}), by Lemma \ref{normcon} we obtain for any $t\in[0,T]$,
\begin{align*}
\|\Psi_t\|^2_{X_t}&\leq 3m \|\text{diag}(A_tX_t)- \text{diag}(X_t)A'_t - A_t \text{diag}(X_t)\|_{N\times N}^2\\
& \leq 3m (|A_tX_t|_{ N}+|X_t|_N\cdot\|A_t\|_{N\times N} +\| A_t\|_{N\times N}\cdot|X_t|_N )^2\\
& \leq 3m(\|A_t\|_{N\times N}\cdot|X_t|_{ N}+|X_t|_N\cdot\|A_t\|_{N\times N} +\| A_t\|_{N\times N}\cdot|X_t|_N )^2\\
& \leq 27m\|A_t\|_{N\times N}^2 \leq 27m^3.
\end{align*}
Hence, by Assumption \ref{ass0} and Lemma \ref{normcon}, we deduce
\begin{align*}
&E[\int ^{T+K} _0\|\Psi_s\Psi_s^\dag z_s^{(n)}-\Psi_s\Psi_s^\dag Z_s\|_{X_s}^{2}ds]\\
&\leq E[\int ^{T+K} _0\|\Psi_s\|_{X_s}^2\cdot\|\Psi^\dag_s\|_{X_s}^2 \cdot \|z_s^{(n)}- Z_s\|_{X_s}^{2}ds]\\
&\leq27m^3l E[\int ^{T+K} _0 \|z_s^{(n)}- Z_s\|_{X_s}^{2}ds]\rightarrow0,~~~~~~\mbox{as }~~n\rightarrow\infty.
\end{align*}
Therefore,
\begin{align*}
~~&E[\int ^{T+K} _0\|Z_s-\Psi_s\Psi_s^\dag Z_s\|_{X_s}^{2}ds]=\lim_{n\rightarrow\infty}E[\int ^{T+K} _0\|Z_s-\Psi_s\Psi_s^\dag Z_s\|_{X_s}^{2}ds]\\
&\leq 3\lim_{n\rightarrow\infty}E[\int ^{T+K} _0\|Z_s-z_s^{(n)}\|_{X_s}^{2}ds]+3\lim_{n\rightarrow\infty}E[\int ^{T+K} _0\| z_s^{(n)}-\Psi_s\Psi_s^\dag z^{(n)}_s\|_{X_s}^{2}ds]\\
&+3\lim_{n\rightarrow\infty}E[\int ^{T+K} _0\|\Psi_s\Psi_s^\dag z_s^{(n)}-\Psi_s\Psi_s^\dag Z_s\|_{X_s}^{2}ds]=0.~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\mbox{}\hfill\Box
\end{align*}
\begin{theorem}\label{bft1} Suppose $\theta>0$ is a given constant, $a_.,\mu_.\in
L^{2}_\mathcal{F}(t_0-\theta,T+\theta;\mathbb{R}),$ $\varphi_.\in
L^{2}_\mathcal{F}(t_0,T;\mathbb{R}),~b_.,\sigma_.\in
L^2_\mathcal{F}(t_0-\theta,T+\theta;\mathbb{R}^{1\times N}),$ and $a_.,\mu_.,b_.,\sigma_.$ are uniformly bounded. Then for all $ U_.\in
L^2_{\mathcal{F}} (T,T+\theta;\mathbb{R}),~V_.\in H^2
(T,T+\theta;\mathbb{R}^{N}),$ the solution $Y_.$ to
anticipated BSDE with Markov chain noise
\begin{equation*}
\left \{ \begin{array}{ll}
-dY_{t}=(a_{t}Y_{t}+\mu_{t}E^{\mathcal{F}_{t}}[Y_{t+\theta
}]+b_{t}Z_{t}+\sigma_{t}E^{\mathcal{F}_{t}}[Z_{t+\theta}
]+\varphi_t)dt-Z'_{t}dM_{t},&t\in[t_0,T]; \\
\hskip.56cmY_{t}=U_t,&t\in [T,T+\theta];\\
\hskip.54cmZ_{t}=V_t,&t\in[T,T+\theta].
\end{array}
\right. \end{equation*}
can be given by the closed formula:
$$
Y_{t}
=E^{\mathcal{F}_{t}}[\hat{X}_TU_T+\int^T_{t}\hat{X}_s\varphi_sds+\int^{T+\theta}_{T}(\mu_{s-\theta}\hat{X}_{s-\theta}U_s+\sigma_{s-\theta}\hat{X}_{s-\theta}V_{s})ds],
$$
for any $ t \in [0,T]$, a.s., where $\hat{X}_s$ is the solution to SDDE with Markov chain
\begin{equation*}
\left \{ \begin{array}{ll}
d\hat{X}_{s}=(a_{s}\hat{X}_{s}+\mu_{s-\theta}\hat{X}_{s-\theta})ds+(\hat{X}_{s-}b_{s-}
+\hat{X}_{(s-\theta)-}\sigma_{(s-\theta)-})(\Psi_s^\dagger)'dM_{s},&s\in [ t,T+\theta];\\
\hskip.245cm\hat{X}_{t}
=1,\\
\hskip.245cm\hat{X}_s =0,& s\in [ t-\theta,t).
\end{array}
\right.\end{equation*}\end{theorem}
\noindent {\bf Proof.}
By (\ref{semimartingale}), $[M,M]_t=[X,X]_t=\left\langle X,X \right\rangle_t +L_t$ and $d \left\langle X,X\right\rangle_t = \Psi_t dt$. By Definition \ref{defMoore} and Lemma \ref{Zequality}, $Z_t'=(\Psi_t\Psi_t^{\dagger} Z_t)'=Z_t'(\Psi_t\Psi^{\dagger}_t)'=Z_t'\Psi_t\Psi^{\dagger}_t$ for $t\in[t_0,T]$. Applying It\^{o}'s formula to
$\hat{X}_sY_s$ for $s\in[t,T],$ we derive
\begin{align*}
&d(\hat{X}_sY_s)\\
& = \hat{X}_{s-} dY_s + Y_{s-}d\hat{X}_s + d[\hat{X},Y]_s \\[.5mm]
& = - \hat{X}_sa_{s}Y_{s}ds- \hat{X}_s\mu_{s}E^{\mathcal{F}_{s}}[Y_{s+\theta
}]ds- \hat{X}_sb_{s}Z_{s}ds- \hat{X}_s\sigma_{s}E^{\mathcal{F}_{s}}[Z_{s+\theta}
]ds \\[.5mm]
&- \hat{X}_s\varphi_sds+ \hat{X}_{s-}Z'_{s} dM_s + Y_s \hat{X}_s a_s ds + Y_s\mu_{s-\theta}\hat{X}_{s-\theta}ds\\[.5mm]
&+ Y_{s-}\hat{X}_{s-}b_{s-}(\Psi_s^\dagger)'dM_{s}
+Y_{s-}\hat{X}_{(s-\theta)-}\sigma_{(s-\theta)-}(\Psi_s^\dagger)'dM_{s} + Z'_s \Delta M_s \hat{X}_{s-}b_{s-} (\Psi_s^{\dagger})' \Delta M_s\\[.5mm]
&+ Z'_s \Delta M_s \hat{X}_{(s-\theta)-}\sigma_{(s-\theta)-}(\Psi_s^\dagger)' \Delta M_s\\[.5mm]
& =- \hat{X}_s\mu_{s}E^{\mathcal{F}_{s}}[Y_{s+\theta
}]ds- \hat{X}_sb_{s}Z_{s}ds- \hat{X}_s\sigma_{s}E^{\mathcal{F}_{s}}[Z_{s+\theta}
]ds - \hat{X}_s\varphi_sds\\[.5mm]
&+ \hat{X}_{s-}Z'_{s} dM_s + Y_s\mu_{s-\theta}\hat{X}_{s-\theta}ds+ Y_{s-}\hat{X}_{s-}b_{s-}(\Psi_s^\dagger)'dM_{s}+Y_{s-}\hat{X}_{(s-\theta)-}\sigma_{(s-\theta)-}(\Psi_s^\dagger)'dM_{s} \\[.5mm]
&
+ Z'_s \Delta M_s \Delta M'_s \Psi_s^{\dagger} \hat{X}_{s-}b'_{s-}+ Z'_s \Delta M_s \Delta M'_s\Psi_s^\dagger \hat{X}_{(s-\theta)-}\sigma'_{(s-\theta)-} \\[.5mm]
& =- \hat{X}_s\mu_{s}E^{\mathcal{F}_{s}}[Y_{s+\theta
}]ds- \hat{X}_sb_{s}Z_{s}ds- \hat{X}_s\sigma_{s}E^{\mathcal{F}_{s}}[Z_{s+\theta}
]ds - \hat{X}_s\varphi_sds\\[.5mm]
&+ \hat{X}_{s-}Z'_{s} dM_s + Y_s\mu_{s-\theta}\hat{X}_{s-\theta}ds+ Y_{s-}\hat{X}_{s-}b_{s-}(\Psi_s^\dagger)'dM_{s}+Y_{s-}\hat{X}_{(s-\theta)-}\sigma_{(s-\theta)-}(\Psi_s^\dagger)'dM_{s} \\[.5mm]
&
+ Z'_s d[M ,M]_s\Psi_s^{\dagger} \hat{X}_{s-}b'_{s-}+ Z'_s d[M ,M]_s\Psi_s^\dagger \hat{X}_{(s-\theta)-}\sigma'_{(s-\theta)-} \\[.5mm]
& = - \hat{X}_s\mu_{s}E^{\mathcal{F}_{s}}[Y_{s+\theta
}]ds- \hat{X}_sb_{s}Z_{s}ds- \hat{X}_s\sigma_{s}E^{\mathcal{F}_{s}}[Z_{s+\theta}
]ds - \hat{X}_s\varphi_sds+ \hat{X}_{s-}Z'_{s} dM_s\\[.5mm]
& + Y_s\mu_{s-\theta}\hat{X}_{s-\theta}ds+ Y_{s-}\hat{X}_{s-}b_{s-}(\Psi_s^\dagger)'dM_{s}+Y_{s-}\hat{X}_{(s-\theta)-}\sigma_{(s-\theta)-}(\Psi_s^\dagger)'dM_{s} + Z'_s\Psi_s \Psi_s^{\dagger}\hat{X}_sb'_s ds \\[.5mm]
& + Z'_sdL_s \Psi_s^{\dagger} \hat{X}_{s-}b'_{s-}+ Z'_s \Psi_s\Psi_s^\dagger \hat{X}_{s-\theta}\sigma'_{s-\theta}ds+ Z'_s dL_s\Psi_s^\dagger \hat{X}_{(s-\theta)-}\sigma'_{(s-\theta)-}\\[.5mm]
&= - \hat{X}_s\mu_{s}E^{\mathcal{F}_{s}}[Y_{s+\theta
}]ds- \hat{X}_sb_{s}Z_{s}ds- \hat{X}_s\sigma_{s}E^{\mathcal{F}_{s}}[Z_{s+\theta}
]ds -\hat{X}_s\varphi_sds+ \hat{X}_{s-}Z'_{s} dM_s\\[.5mm]
& + Y_s\mu_{s-\theta}\hat{X}_{s-\theta}ds+ Y_{s-}\hat{X}_{s-}b_{s-}(\Psi_s^\dagger)'dM_{s}+Y_{s-}\hat{X}_{(s-\theta)-}\sigma_{(s-\theta)-}(\Psi_s^\dagger)'dM_{s} \\[.5mm]
&+ Z'_s\hat{X}_sb'_s ds + Z'_sdL_s \Psi_s^{\dagger} \hat{X}_{s-}b'_{s-}+ Z'_s \hat{X}_{s-\theta}\sigma'_{s-\theta}ds+ Z'_s dL_s\Psi_s^\dagger \hat{X}_{(s-\theta)-}\sigma'_{(s-\theta)-}
\end{align*}
\begin{align*}
&= - \hat{X}_s\mu_{s}E^{\mathcal{F}_{s}}[Y_{s+\theta
}]ds- \hat{X}_s\sigma_{s}E^{\mathcal{F}_{s}}[Z_{s+\theta}
]ds - \hat{X}_s\varphi_sds+ \hat{X}_{s-}Z'_{s} dM_s\\[.5mm]
& + Y_s\mu_{s-\theta}\hat{X}_{s-\theta}ds+ Y_{s-}\hat{X}_{s-}b_{s-}(\Psi_s^\dagger)'dM_{s}+Y_{s-}\hat{X}_{(s-\theta)-}\sigma_{(s-\theta)-}(\Psi_s^\dagger)'dM_{s} \\[.5mm]
& + Z'_sdL_s \Psi_s^{\dagger} \hat{X}_{s-}b'_{s-}+ Z'_s \hat{X}_{s-\theta}\sigma'_{s-\theta}ds+ Z'_s dL_s\Psi_s^\dagger \hat{X}_{(s-\theta)-}\sigma'_{(s-\theta)-} .
\end{align*}
Because $\hat{X}_{t}=1$ and $\hat{X}_s=0$, $s\in [ t-\theta,t)$, taking conditional expectations under
$\mathcal{F}_{t},$ we have
$$\begin{array}{ll}
Y_{t}\\[0.2cm]
=E^{\mathcal{F}_{t}}[\hat{X}_TY_T+\int^T_{t} \hat{X}_s\mu_{s}E^{\mathcal{F}_{s}}[Y_{s+\theta
}]ds+ \int^T_{t}\hat{X}_s\sigma_{s}E^{\mathcal{F}_{s}}[Z_{s+\theta}
]ds +\int^T_{t}\hat{X}_s\varphi_sds\\[2mm]
~~~~~~~~-\int^T_{t} Y_s\mu_{s-\theta}\hat{X}_{s-\theta}ds-
\int^T_{t}Z'_s \hat{X}_{s-\theta}\sigma'_{s-\theta}ds]\\[0.2cm]
=E^{\mathcal{F}_{t}}[\hat{X}_TY_T+\int^T_{t}\hat{X}_s\varphi_sds]+E^{\mathcal{F}_{t}}[\int^T_{t}(\hat{X}_s\mu_s Y_{s+\theta
}- \hat{X}_{s-\theta}\mu_{s-\theta}Y_s)ds]\\[0.2cm]
+E^{\mathcal{F}_{t}}[\int^T_{t}(\hat{X}_s\sigma_{s}Z_{s+\theta}- \hat{X}_{s-\theta}\sigma_{s-\theta}Z_s )ds]\\[0.2cm]
=E^{\mathcal{F}_{t}}[\hat{X}_TY_T+\int^T_{t}\hat{X}_s\varphi_sds]+E^{\mathcal{F}_{t}}[\int^{T+\theta}_{t+\theta}\hat{X}_{s-\theta}\mu_{s-\theta}Y_sds
-\int^{T}_{t}\hat{X}_{s-\theta}\mu_{s-\theta}Y_sds]\\[0.2cm]
+E^{\mathcal{F}_{t}}[\int^{T+\theta}_{t+\theta}\hat{X}_{s-\theta}\sigma_{s-\theta}Z_s ds-\int^{T}_{t}\hat{X}_{s-\theta}\sigma_{s-\theta}Z_s ds]\\[0.2cm]
=E^{\mathcal{F}_{t}}[\hat{X}_TY_T+\int^T_{t}\hat{X}_s\varphi_sds+\int^{T+\theta}_{T}(\hat{X}_{s-\theta}\mu_{s-\theta} Y_{s
}+\hat{X}_{s-\theta}\sigma_{s-\theta}Z_{s})ds]\\[0.2cm]
-E^{\mathcal{F}_{t}}[\int^{t+\theta}_{t}(\hat{X}_{s-\theta}\mu_{s-\theta} Y_{s
}+\hat{X}_{s-\theta}\sigma_{s-\theta}Z_{s})ds]\\[0.2cm]
=E^{\mathcal{F}_{t}}[\hat{X}_TU_T+\int^T_{t}\hat{X}_s\varphi_sds+\int^{T+\theta}_{T}(\mu_{s-\theta}\hat{X}_{s-\theta}U_s+\sigma_{s-\theta}\hat{X}_{s-\theta }V_{s})ds],~~~\mbox{a.e.,~a.s.}
\end{array}$$
By Lemma \ref{indistinguish}, we obtain
$Y_t=E^{\mathcal{F}_{t}}[\hat{X}_TU_T+\int^T_{t}\hat{X}_s\varphi_sds+\int^{T+\theta}_{T}\mu_{s-\theta}\hat{X}_{s-\theta}U_s ds$ + $\int^{T+\theta}_{T}\sigma_{s-\theta}\hat{X}_{s-\theta}V_{s}ds]$, for any $t \in [0,T]$, a.s.$\mbox{}\hfill\Box$
\section{Comparison theorem of 1-dimensional anticipated BSDEs with Markov chain model}
\indent The main idea of the proof of our comparison theorem comes from the proof of the comparison theorem for
anticipated BSDEs with Brownian motion noise in Peng and Yang \cite{PengYang}.\\
\indent Let $
(Y^{(1)}_{\cdot},Z^{(1)}_{\cdot}),~(Y^{(2)}_{\cdot},Z^{(2)}_{\cdot})$
be respectively the solutions of the following two 1-dimensional
anticipated BSDEs:
$$
\left \{\begin{array}{ll}
-dY^{(j)}_t =f_j(t,Y^{(j)}_t,Z^{(j)}_t,Y^{(j)}_{t+\delta(t)})dt -Z^{(j)}_tdM_{t},&0 \leq t\leq T;\\[1mm]
\hskip.575cmY^{(j)}_{t}=\xi^{(j)}_t,&T\leq t \leq T+K,
\end{array}
\right.
$$
where $j=1,2.$
\begin{theorem}\label{bft2} Assume $\xi_{\cdot}^{(1)},\xi_{\cdot}^{(2)}\in L^2_{\mathcal{F}}(T,T+K;\mathbb{R})$, $\delta$ satisfies
$(i),(ii),$ and $f_1,f_2$ satisfy conditions such that the above two anticipated BSDEs have unique solutions. Suppose
\begin{enumerate}
\item $f_1$ satisfies (H1), moreover, the Lipschitz constant $c_2$ of $f_1$ satisfies $$c_2\|\Psi_t^{\dagger}\|_{N \times N} \sqrt{6m}\leq 1, ~~~\text{ for any }~t \in [0,T],$$ where $\Psi$ is given in \eqref{Psi} and $m>0$ is the bound of $\|A_t\|_{N\times N}$, for any $t\in[0,T]$.
\item for any $t\in[0,T],$ $y\in\mathbb{R},$ $z\in\mathbb{R}^N$,
$f_1(t,y,z,\cdot)$ is increasing, i.e., $f_1(t,y,z,\theta_r)$ $\geq
f_1(t,y,z,\theta'_r)$, if $\theta_r\geq\theta'_r$,
$\theta_{\cdot},\theta'_{\cdot}\in
L^2_{\mathcal{F}}(t,T+K;\mathbb{R}),r\in[t,T+K]$.\end{enumerate}
If
$\xi^{(1)}_s\leq\xi^{(2)}_s,~s\in[T,T+K],$ and
$f_1(t,Y^{(2)}_t,Z^{(2)}_t,Y^{(2)}_{t+\delta(t)})\leq
f_2(t,Y^{(2)}_t,Z^{(2)}_t,Y^{(2)}_{t+\delta(t)}),$ a.e., a.s., then
$$P( Y_t^{(1)}\leq Y_t^{(2)},~~\text{ for any } t \in [0,T])=1.$$\end{theorem}
\noindent{\bf Proof.} Set
$$ \left \{ \begin{array}{ll} Y^{(3)}_t = \xi_T ^{(1)} +\int ^T_tf_1(s,Y^{(3)}_s,Z^{(3)}_s,Y^{(2)}_{s+\delta(s)})ds -\int ^T_t (Z^{(3)}_s)'d M_{s},&t\in[0,T]; \\[1mm]
Y^{(3)}_{t} =
\xi^{(1)}_t,&t\in[T,T+K].
\end{array} \right. $$
By Lemma \ref{existence}, we know there exists a solution $(Y^{(3)}_{\cdot},Z^{(3)}_{\cdot})\in L^2_{\mathcal{F}}(0,T;\mathbb{R})\times H^2(0,T;\mathbb{R}^N)$
to the above BSDE. Moreover, this solution is
unique up to indistinguishability for $Y.$ and equality $d\langle
M,M\rangle_t\times\mathbb{P}$-a.s. for $Z$. Set
$\tilde{f}_t=f_2(t,Y^{(2)}_t,Z^{(2)}_t,Y^{(2)}_{t+\delta(t)})-f_1(t,Y^{(2)}_t,Z^{(2)}_t,Y^{(2)}_{t+\delta(t)})$
and
$y_.=Y_.^{(2)}-Y_.^{(3)},~z_.=Z_.^{(2)}-Z_.^{(3)},~\tilde{\xi}_.=\xi_.^{(2)}-\xi_.^{(1)}.$
Then the pair $(y,z)$ can be regarded as the solution to the linear
BSDE
$$ \left \{ \begin{array}{ll}
y_t=\tilde{\xi}_T+\int ^T _t(a_sy_s+b_sz_s+\tilde{f}_s)ds-\int ^T
_tz_sdM_s,&t\in[0,T];\\[1mm]
y_t=\tilde{\xi}_t,&t\in[T,T+K],
\end{array} \right.
$$
where
$$\begin{array}{ll}
a_s=\begin{cases}\dfrac{f_1(t,Y^{(2)}_t,Z^{(2)}_t,Y^{(2)}_{t+\delta(t)})-f_1(t,Y^{(3)}_t,Z^{(2)}_t,Y^{(2)}_{t+\delta(t)})}{y_s},&\mbox{if } \,y_s\neq 0;\\
0,& \mbox{if } \,y_s= 0,\end{cases} \\\\
b_s=\begin{cases}\dfrac{f_1(t,Y^{(3)}_t,Z^{(2)}_t,Y^{(2)}_{t+\delta(t)})-f_1(t,Y^{(3)}_t,Z^{(3)}_t,Y^{(2)}_{t+\delta(t)})}{|z_s|_N^2}z_s',& \mbox{if } \,z_s\neq 0;\\
0,& \mbox{if } \,z_s= 0.\end{cases}
\end{array} $$
Since $f_1$ satisfies (H1), we deduce for any $s\in[0,T]$, $|a_s|\leq c_1$ and by Lemma \ref{normbound},
$$|b_s|_N\leq c_2\frac{\|z_s\|_{X_s}\cdot|z_s|_N}{|z_s|_N^2}\leq c_2\sqrt{3m}.$$
By Lemma \ref{Duality}, we know
\begin{equation*}
P (y_t = E [ \tilde{\xi}_T U_T + \int_t^T \tilde{f}_s U_s ds | \mathcal{F}_t], ~ \text{for any}~ t\in [0,T])=1,
\end{equation*}
where $U$ is the solution of a one-dimensional SDE
\begin{equation}\label{sde}
\begin{cases}
dU_s= U_s a_s ds + U_{s-} b_{s-} (\Psi_s^{\dagger})'dM_s, ~~ s \in [t,T];\\
~~U_t =1.
\end{cases}
\end{equation}
Denote $$dV_s = a_sds + b_{s-} (\Psi_s^{\dagger})'dM_s,~~s\in[0,T].$$ The solution to SDE \eqref{sde} is given by the Dol\'{e}an-Dade exponential (See \cite{elliott}):
\[U_s = \exp(V_s - \frac{1}{2}\left\langle V^c,V^c\right\rangle_s)\prod_{0 \leq u\leq s} (1+\Delta V_u )e^{-\Delta V_u},~~s\in[0,T],\]
where $$\Delta V_u = b_{u-} (\Psi_u^{\dagger})' \Delta M_u = b_{u-} (\Psi_u^{\dagger})' \Delta X_u.$$
Since $f_1$ satisfies $c_2\|\Psi_t^{\dagger}\|_{N \times N} \sqrt{6m}\leq 1,$ for any $t \in [0,T],$ where $\Psi$ is given in \eqref{Psi} and $m>0$ is the bound of $\|A_t\|_{N\times N}$, for any $t\in[0,T]$, by Lemma \ref{normbound} we have
\begin{align*}
|\Delta V_u| \leq |b_{u-} |_N \cdot \|(\Psi_u^{\dagger})'\|_{N\times N} \cdot |\Delta X_u|_N
\leq c_2\sqrt{3m}\dfrac{1}{\sqrt{6m}c_2}\sqrt{2}
=1.
\end{align*}
Hence we have $U_s \geq0$, $s\in[0,T]$. As $\tilde{\xi}_T \geq 0$, a.s., and $\tilde{f}_s \geq 0$, a.e., a.s., we know for any $t \in [0,T]$,
$$
y_t = E [ \tilde{\xi}_T U_T + \int_t^T \tilde{f}_s U_s ds |\mathcal{F}_t] \geq0, ~a.s.
$$
Since $y_.$ is RCLL, by Lemma \ref{indistinguish}, we obtain
\[P(Y^{(2)}_t\geq Y^{(3)}_t, ~ \text{for any } t \in [0,T])=P(y_t \geq 0, ~ \text{for any } t \in [0,T])=1.\]
Set
$$ \left \{\begin{array}{ll}
Y^{(4)}_t = \xi_T ^{(1)} +\int^T_tf_1(s,Y^{(4)}_s,Z^{(4)}_s,Y^{(3)}_{s+\delta(s)})ds -\int ^T_t
(Z^{(4)}_s)'dM_{s},&t\in[0,T];\\[1mm]
Y^{(4)}_{t} =\xi^{(1)}_t,&t\in[T,T+K].
\end{array} \right. $$
Recall for any $t\in[0,T],y\in\mathbb{R},z\in\mathbb{R}^N,$
$f_1(t,y,z,\cdot)$ is increasing and $Y^{(2)}_t\geq Y^{(3)}_t,$ for any $ t \in [0,T]$, a.e. Also, $f_1$ satisfies $c_2\|\Psi_t^{\dagger}\|_{N \times N} \sqrt{6m}\leq 1$ for $t \in [0,T]$. So
by Lemma \ref{CT'} we
obtain
$$P(Y^{(3)}_t\geq Y^{(4)}_t,~~\text{ for any } t \in [0,T])=1.$$
For $n=5,6,\cdot\cdot\cdot,$ we consider the following sequence of classical
BSDEs on Markov chain:
$$ \left \{\begin{array}{ll}
Y^{(n)}_t = \xi_T ^{(1)} +\int^T_tf_1(s,Y^{(n)}_s,Z^{(n)}_s,Y^{(n-1)}_{s+\delta(s)})ds -\int ^T_t
(Z^{(n)}_s)'dM_{s},&t\in[0,T];\\[1mm]
Y^{(n)}_{t} =\xi^{(1)}_t,&t\in[T,T+K].
\end{array} \right. $$
Similarly for any $n\in\mathbb{N},n\geq4$, we know the above equation has a unique solution $(Y^{(n)}_.,Z^{(n)}_. )\in L^2_{\mathcal{F}}(0,T;\mathbb{R})\times H^2(0,T;\mathbb{R}^N)$. Moreover, there exists a subset $A_n\subseteq\Omega$ with $P(A_n)=1$ such that for any $\omega\in A_n$, $Y_t^{(n)}(\omega)\geq Y_t^{(n+1)}(\omega),$ for any $t\in[0,T].$ Hence
$$
P(\bigcap_{n=4}^{+\infty}A_n) = 1 - P(\bigcup_{n=4}^{+\infty} A_n^c)
\geq 1 -\sum_{n=4}^{+\infty}P(A^c_n)
=1.
$$
That is,
$$P(Y_t^{(4)}\geq Y_t^{(5)}\geq \ldots\geq Y_t^{(n)}\geq\ldots, ~~\text{for
any }t\in[0,T])=1.$$
So
$$P(Y_t^{(2)}\geq Y_t^{(3)}\geq Y_t^{(4)}\geq Y_t^{(5)}\geq\ldots\geq Y_t^{(n)}\geq\ldots, ~~\text{for
any }t\in[0,T])=1.$$
Let $\beta>0$ be an arbitrary constant and $c=\max\{c_1,c_2\}$. We use
$\parallel\nu(\cdot)\parallel_{L^2}$ and $\parallel\mu(\cdot)\parallel_{\hat{H}^2}$ in the proof of Theorem
\ref{bft3000000} as the norms in the Banach spaces
$L^2_\mathcal{F}(0,T+K;\mathbb{R})$ and $\hat{H}^2(0,T+K;\mathbb{R}^{N})$, respectively. Set
$\hat{Y}^{(n)}_s=Y^{(n)}_s-Y^{(n-1)}_s,~\hat{Z}^{(n)}_s=Z^{(n)}_s-Z^{(n-1)}_s,~n\geq4.$
Then $(\hat{Y}^{(n)}_{\cdot},~\hat{Z}^{(n)}_{\cdot})$ satisfies the
following BSDE
$$ \left \{
\begin{array}{ll} \hat{Y}^{(n)}_t = \int
^T_t(f_1(s,Y^{(n)}_s,Z^{(n)}_s,Y^{(n-1)}_{s+\delta(s)})-f_1(s,Y^{(n-1)}_s,Z^{(n-1)}_s,Y^{(n-2)}_{s+\delta(s)}))ds\\[0.1cm]
\hskip1.2cm-\int^T_t (\hat{Z}^{(n)}_s)'dM_{s},\hskip1cmt\in[0,T];\\[0.1cm]
\hat{Y}^{(n)}_{t}=0,\hskip3.5cmt\in[T,T+K].
\end{array} \right. $$
Apply It\^{o}'s formula to $e^{\beta s} |\hat{Y}_s|^2$ for $s\in[0,T]$ and then take the expectation:
\begin{align*}
&E [ |\hat{Y}_0^{(n)}|^2 ]+ E [ \int_0^T \beta |\hat{Y}_s^{(n)}|^2 e^{\beta s} ds ] + E [ \int_0^T\|\hat{Z}_s^{(n)}\|^2_{X_s}e^{\beta s} ds ] \\
& = 2 E [ \int_0^T \hat{Y}_s^{(n)} (f_1(s,Y^{(n)}_s,Z^{(n)}_s,Y^{(n-1)}_{s+\delta(s)})-f_1(s,Y^{(n-1)}_s,Z^{(n-1)}_s,Y^{(n-2)}_{s+\delta(s)})) e^{\beta s}ds ]\\
&\leq E [ \int_0^T (\frac{\beta}{2}|\hat{Y}^{(n)}_s|^2+\frac{2}{\beta} |f_1(s,Y^{(n)}_s,Z^{(n)}_s,Y^{(n-1)}_{s+\delta(s)})-f_1(s,Y^{(n-1)}_s,Z^{(n-1)}_s,Y^{(n-2)}_{s+\delta(s)})|^2)e^{\beta s} ds ].
\end{align*}
Thus
\begin{align*}
&
E[\int^T_0(\dfrac{\beta}{2}|\hat{Y}^{(n)}_s|^{2}+\|\hat{Z}_s^{(n)}\|^2_{X_s})e^{\beta
s}ds] \\
&\leq
\dfrac{2}{\beta}E[\int^T_0|f_1(s,Y^{(n)}_s,Z^{(n)}_s,Y^{(n-1)}_{s+\delta(s)})-f_1(s,Y^{(n-1)}_s,Z^{(n-1)}_s,Y^{(n-2)}_{s+\delta(s)})|^{2}e^{\beta
s}ds]\\
&
\leq\dfrac{6c^2}{\beta}E[\int^T_0(|\hat{Y}^{(n)}_s|^2+\|\hat{Z}^{(n)}_s\|_{X_s}^2+|\hat{Y}^{(n-1)}_{s+\delta(s)}|^2)e^{\beta
s}ds]\\
&
\leq\dfrac{6c^2}{\beta}E[\int^T_0(|\hat{Y}^{(n)}_s|^2+\|\hat{Z}^{(n)}_s\|_{X_s}^2)e^{\beta
s}ds]+\dfrac{6c^2L}{\beta}E[\int^T_0|\hat{Y}^{(n-1)}_s|^2e^{\beta
s}ds].
\end{align*}
Set $\beta=18c^2L+18c^2+3$. Then
\begin{align*}\dfrac{2}{3}E[\int^T_0(|\hat{Y}^{(n)}_s|^{2}+\|\hat{Z}^{(n)}_s\|_{X_s}^2)e^{\beta
s}ds]&
\leq\dfrac{1}{3}E[\int^T_0|\hat{Y}^{(n-1)}_s|^{2}e^{\beta s}ds]\\
&\leq\dfrac{1}{3}E[\int^T_0(|\hat{Y}^{(n-1)}_s|^{2}+\|\hat{Z}^{(n-1)}_s\|^{2}_{X_s})e^{\beta
s}ds].
\end{align*}
Hence
$$
E[\int^T_0(|\hat{Y}^{(n)}_s|^{2}+\|\hat{Z}^{(n)}_s\|_{X_s}^{2})e^{\beta
s}ds]
\leq(\frac{1}{2})^{n-4}E[\int^T_0(|\hat{Y}^{(4)}_s|^{2}+\|\hat{Z}^{(4)}_s\|^{2}_{X_s})e^{\beta
s}ds].
$$
It follows that $(Y^{(n)}_{\cdot})_{n\geq4}$ and
$(Z^{(n)}_{\cdot})_{n\geq4}$ are, respectively, Cauchy sequences in
$L^2_{\mathcal{F}}(0,T+K;\mathbb{R})$ and in
$\hat{H}^2(0,T+K;\mathbb{R}^{N})$. Denote their limits by
$Y_{\cdot}$ and $Z_{\cdot}$, respectively. Then $$P(Y_t^{(2)}\geq Y_t^{(3)}\geq \ldots\geq Y_t^{(n)}\geq\ldots\geq Y_t, ~~\text{for
any }t\in[0,T])=1.$$ Since
$L^2_{\mathcal{F}}(0,T+K;\mathbb{R})$ and $\hat{H}^2(0,T+K;\mathbb{R}^{N})$
are both Banach spaces, we obtain $(Y_{\cdot},Z_{\cdot})\in
L^2_{\mathcal{F}}(0,T+K;\mathbb{R})\times \hat{H}^2(0,T+K;\mathbb{R}^{N}).$
Note for any $\in[0,T]$,
$$\begin{array}{lll}
\hskip.44cmE[\int^T_t|f_1(s,Y^{(n)}_s,Z^{(n)}_s,Y^{(n-1)}_{s+\delta(s)})-f_1(s,Y_s,Z_s,Y_{s+\delta(s)})|^{2}e^{\beta
s}ds]\\[0.2cm]
\leq3c^2E[\int^T_t(|Y^{(n)}_s-Y_s|^2+\|Z^{(n)}_s-Z_s\|_{X_s}^2+L|Y^{(n-1)}_s-Y_s|^2)e^{\beta
s}ds]\rightarrow0,~n\rightarrow\infty.
\end{array}$$
Therefore, $(Y_{\cdot},Z_{\cdot})$ satisfies the following
anticipated BSDE
$$ \left \{ \begin{array}{ll} Y_t = \xi_T ^{(1)} +\int ^T_tf_1(s,Y_s,Z_s,Y_{s+\delta(s)})ds -\int ^T_t Z'_sdM_{s},&0 \leq t\leq T;\\[1mm]
Y_{t} = \xi^{(1)}_t,&T\leq t \leq
T+K.
\end{array} \right. $$
By Theorem \ref{bft3000000} we know
$$P(Y_t=Y^{(1)}_t, ~ \text{for any } t \in [0,T])=1.$$ Because $P(Y^{(2)}_t\geq
Y^{(3)}_t\geq Y_t,$ for any $t\in[0,T])=1$, it holds immediately that
$$\hskip3.5cm P(Y^{(1)}_t\leq
Y^{(2)}_t, ~ \text{for any } t \in [0,T])=1.\hskip3cm \mbox{}\hfill\Box$$
|
\section{\bf Introduction}\par
With the recent discovery~\cite{atlas-cms-2012}
of a Brout-Englert-Higgs (BEH)~\cite{EBH}
boson after the start-up and successful running
of the LHC for three years, as we investigate its properties
and any hints in the LHC data for physics beyond the Standard Model,
we realize that we have entered the era of precision QCD,
by which we mean
predictions for QCD processes at the total precision tag of $1\%$ or better.
We have argued in Refs.~\cite{radcor2011,radcor2013,herwiri,1305-0023,qced} that
exact, amplitude-based
resummation of large higher order effects is a viable strategy
to achieve such precision tags and we have developed the MC HERWIRI1.031
~\cite{herwiri} as a first platform in this connection, with some emphasis on its MC@NLO/HERWIRI1.031 realization of exact NLO matrix-element matched parton shower calculus.
We have previously illustrated the comparison of the new MC, which carries the
IR-improved~\cite{irdglap1,irdglap2} DGLAP-CS~\cite{dglap,cs} theory
in a HERWIG6510~\cite{hwg} environment, with the data of ATLAS~\cite{atlaspt} and CMS~\cite{cmsrap} at the LHC and with the data of D0~\cite{d0pt} and CDF~\cite{galea} at FNAL. In what follows, we extend these studies to the data of LHCb~\cite{lhcbdata} at the LHC.\par
More precisely, any precision theory platform should be able to cover the entire
observable phase space for hard processes in order to fully exploit the data at the LHC. In the single $Z/\gamma^*$ production and decay to lepton pairs, the LHCb probes the regime in the lepton pseudo-rapidity
$\eta= -\ln(\tan(\theta/2))$ given by
$2.0<\eta<4.5$, where $\theta$ is the polar angle with respect
to a beam direction. This should be compared to the regimes probed by the data we studied in Refs.~\cite{herwiri,1305-0023} which for ATLAS had $|\eta|<2.4$ for $e^+,\; e^-$ and which for CMS had $|\eta|<2.1$ for $\mu^+,\;\mu^-$ and $|\eta|<4.6$
for $e^+,\; e^-$. The cuts on the lepton transverse momenta were similar with $p_T>20$ GeV/c for all the data we discuss here. Thus, the LHCb data provide a different check on the comparison between the theoretical predictions, where we are interested in comparing the IR-improved results with NLO exact ME/shower matching to the corresponding unimproved ones, both with and without the 'ad hocly' hard intrinsic ${\rm PTRMS}\cong 2.2$ GeV/c that we have found was necessary for the unimproved calculations to explain both the rapidity and the $p_T$ spectra from ATLAS and CMS that we discussed in Refs.~\cite{herwiri}. Here, ${\rm PTRMS}$ is
the rms value of a Gaussian intrinsic $p_T$ distribution
for the proton constituents
in HERWIG65~\cite{hwg}. \par
What we are particularly interested to see is how the IR-improvement interplays with the change in the phase space for the accepted lepton pairs in the $Z$ mass region. In all of our work, we rely on the data as given by the experimentalists so that all conclusions we draw must be interpreted with this understanding:
{\it our new IR-improved MC has not been used, as yet, by any of the LHC collaborations in unfolding their data in any way in obtaining their publicly available results, as far as we know, so that this limits the strength of our conclusions in a direct way}. For example, if a bin-to-bin migration effect has been estimated with a parton shower MC that has an ad hoc infrared cut-off, it is unknown what the same effect would be if it were estimated with our new IR-improved MC which does not need such a cut-off. Here, we want to try to put some emphasis on the following point. In the usual parton shower with the ad hoc infrared cutoff, let us call it $k_0$, the region of the shower emission phase space below the energy $k_0$ is dropped; for, the exponentiating virtual correction, $V_{sh}$, is defined to be the negative of the integral of the shower's real emission distribution $R$ so that when one sums over all real emissions one gets the total correction factor $e^{V_{sh}+\int R}=e^{0}=1$ and the shower does not change the normalization. Moreover, as the $k_0$ dependence in the integral over the real emission in the shower exactly cancels against that in the virtual correction, one formally has no dependence on $k_0$. But, this is only formally true. In reality, the actual events the shower makes depend on $k_0$, with softer values generically generating more soft radiation than harder values, so comparison with events in the experiments can illustrate very well the effects of changing $k_0$. In the IR-improved patron shower MC HERWIRI1.031, which realizes IR-improved DGLAP-CS theory~\cite{dglap,cs} in the
HERWIG6.5~\cite{hwg} environment, the regime below $k_0$ is taken into account by amplitude based resummation methods as presented in Refs.~\cite{irdglap1,irdglap2}. This is expected to produce events that more closely resemble those seen in the experiments as we have illustrated in Refs.~\cite{herwiri,1305-0023}.\par
More specifically, we discuss in what follows the comparisons between the IR-improved and unimproved parton shower(PS) MC predictions, with
the exact MC@NLO~\cite{mcatnlo} PS/ME matched exact ${\cal O}(\alpha_s)$ correction, and the LHCb data on the $Z/\gamma*$ rapidity, $p_T$ and $\phi^*_\eta$ distributions, where the variable
$\phi^*_\eta$ was introduced in Refs.~\cite{phietastr} as an attempt to overcome some of the apparent experimental difficulties in measuring finely binned $p_T$ spectra for lower values of $p_T$ for the $Z/\gamma^*$ in the LHC environment.
We have the definition $\phi_\eta^*=\tan(\frac{1}{2}(\pi-\Delta\phi))\sin\theta^* \cong \left|\sum \frac{{p_i}_T \sin\phi_i}{Q}\right| +{\cal O}(\frac{{{p_i}_T}^2}{Q^2})$, where $\Delta\phi=\phi_1-\phi_2$ is the azimuthal angle
between the two leptons which have transverse momenta $\vec{p_i}_T,\; i=1,2,$
and $\theta^*$ is the scattering angle of the dilepton system relative to the beam direction when one boosts to the frame along the beam direction such that the leptons are back to back. We present here the comparisons with
the respective MC@NLO~\cite{mcatnlo} parton shower/matrix element matched exact ${\cal O}(\alpha_s)$ corrections included, as this correction is now established to be important~\cite{herwiri}.\par
Our discussion proceeds as follows. In the next section, we give a brief review
of the theoretical paradigm we are using here, as this approach to precision QCD is still not a familiar one. In the Sect. 3, we turn to comparison with the LHCb data; we also discuss attendant theoretical implications therein.
In Sect. 4 we sum up.
\section{Review of Exact Amplitude-Based Resummation Theory and Its Parton Shower MC Implementation}
Our discussion here starts from the following
fully differential representation of a hard LHC scattering process:
\begin{equation}
d\sigma =\sum_{i,j}\int dx_1dx_2F_i(x_1)F_j(x_2)d\hat\sigma_{\text{res}}(x_1x_2s),
\label{bscfrla}
\end{equation}
where the $\{F_j\}$ and
$d\hat\sigma_{\text{res}}$ are the respective parton densities and
resummed reduced hard differential cross section which
has been resummed
for all large EW and QCD higher order corrections in a manner consistent
with achieving a total precision tag of 1\% or better for the
theoretical precision of (\ref{bscfrla}). As we have explained
in Refs.~\cite{herwiri,1305-0023}, to have
the latter precision tag as a realistic goal,
we have developed
the $\text{QCD}\otimes\text{QED}$ resummation theory in Refs.~\cite{qced}
for the reduced cross section in (\ref{bscfrla}) and for the
resummation of the evolution of the parton densities therein as well.
\par
Specifically,
for both the resummation of the reduced cross section
and that of the evolution of the parton densities,
the defining formula
may be identified as
\begin{eqnarray}
&d\bar\sigma_{\rm res} = e^{\rm SUM_{IR}(QCED)}
\sum_{{n,m}=0}^\infty\frac{1}{n!m!}\int\prod_{j_1=1}^n\frac{d^3k_{j_1}}{k_{j_1}} \cr
&\prod_{j_2=1}^m\frac{d^3{k'}_{j_2}}{{k'}_{j_2}}
\int\frac{d^4y}{(2\pi)^4}e^{iy\cdot(p_1+q_1-p_2-q_2-\sum k_{j_1}-\sum {k'}_{j_2})+
D_{\rm QCED}} \cr
&\tilde{\bar\beta}_{n,m}(k_1,\ldots,k_n;k'_1,\ldots,k'_m)\frac{d^3p_2}{p_2^{\,0}}\frac{d^3q_2}{q_2^{\,0}},
\label{subp15b}
\end{eqnarray}\noindent
where $d\bar\sigma_{\rm res}$ is either the reduced cross section
$d\hat\sigma_{\rm res}$ or the differential rate associated to a
DGLAP-CS~\cite{dglap,cs} kernel involved in the evolution of the $\{F_j\}$ and
where the {\em new} (YFS-style~\cite{yfs,yfs-jw}) {\em non-Abelian} residuals
$\tilde{\bar\beta}_{n,m}(k_1,\ldots,k_n;k'_1,\ldots,k'_m)$ have $n$ hard gluons and $m$ hard photons and we show the final state with two hard final
partons with momenta $p_2,\; q_2$ specified for a generic $2f$ final state for
definiteness. The infrared functions ${\rm SUM_{IR}(QCED)},\; D_{\rm QCED}\; $
are defined in Refs.~\cite{qced,irdglap1,irdglap2} as follows:
\begin{eqnarray}
{\rm SUM_{IR}(QCED)}=2\alpha_s\Re B^{nls}_{QCED}+2\alpha_s{\tilde B}^{nls}_{QCED}\cr
D_{\rm QCED}=\int \frac{d^3k}{k^0}\left(e^{-iky}-\theta(K_{max}-k^0)\right){\tilde S}^{nls}_{QCED}
\label{irfns}
\end{eqnarray}
where the dummy parameter $K_{max}$ is such that nothing depends on it and where we have introduced
\begin{eqnarray}
B^{nls}_{QCED} \equiv B^{nls}_{QCD}+\frac{\alpha}{\alpha_s}B^{nls}_{QED},\cr
{\tilde B}^{nls}_{QCED}\equiv {\tilde B}^{nls}_{QCD}+\frac{\alpha}{\alpha_s}{\tilde B}^{nls}_{QED}, \cr
{\tilde S}^{nls}_{QCED}\equiv {\tilde S}^{nls}_{QCD}+{\tilde S}^{nls}_{QED}.
\label{irfns1}
\end{eqnarray}
Here, the superscript $nls$ denotes that the infrared functions are DGLAP-CS synthesized as explained in Refs.~\cite{dglpsyn,qced,irdglap1,irdglap2} and the infrared functions
$B_A,\; {\tilde B}_A,\; {\tilde S}_A, \; A=QCD,\; QED,$ are given
in Refs.~\cite{yfs,yfs-jw,qced,irdglap1,irdglap2}\footnote{We note that the ratio of QED and QCD couplants on the RHS of (\ref{irfns1}) is suppressed in Ref.~\cite{1305-0023}, for example.}.
The
simultaneous resummation of QED and QCD large IR effects is exact here.
See Refs.~\cite{herwiri,1305-0023} for discussion of the physical meanings of the various components of the master formula as well as for discussion
of our connection with the methods in Ref.~\cite{gatheral}. Here, we do continue to stress that
in our formulation in (\ref{subp15b})
{\it the entire soft gluon phase space is included in the
representation -- no part of it
is dropped}.
\par
The new non-Abelian residuals $\tilde{\bar\beta}_{m,n}$
allow~\cite{qced} rigorous shower/ME matching via their shower subtracted analogs. What this means is that
in (\ref{subp15b}) we make the replacements
\begin{equation}
\tilde{\bar\beta}_{n,m}\rightarrow \hat{\tilde{\bar\beta}}_{n,m}
\end{equation}
where the $\hat{\tilde{\bar\beta}}_{n,m}$ have had all effects in the showers
associated to the $\{F_j\}$ removed from them. Contact between the $\hat{\tilde{\bar\beta}}_{n,m}$ and the
differential distributions in MC@NLO proceeds as
follows. Representing the no-emission probability, the Sudakov from factor, as
$$\Delta_{MC}(p_T)=e^{[-\int d\Phi_R \frac{R_{MC}(\Phi_B,\Phi_R)}{B}\theta(k_T(\Phi_B,\Phi_R)-p_T)]},$$
the MC@NLO differential cross section can be written as~\cite{mcatnlo}
\begin{equation}
\begin{split}
d\sigma_{MC@NLO}&=\left[B+V+\int(R_{MC}-C)d\Phi_R\right]d\Phi_B[\Delta_{MC}(0)+\int(R_{MC}/B)\Delta_{MC}(k_T)d\Phi_R]\\
&\qquad\qquad +(R-R_{MC})\Delta_{MC}(k_T)d\Phi_Bd\Phi_R
\label{mcatnlo1}
\end{split}
\end{equation}
where $B$ is Born distribution, $V$ is the regularized virtual contribution,
$C$ is the corresponding counter-term required at exact NLO, $R$ is the respective
exact real emission distribution for exact NLO and $R_{MC}=R_{MC}(P_{AB})$ is the parton shower real emission distribution, with the obvious notation for the respective phase spaces $\{d\Phi_A, \; A=B,\; R\}$.
From comparison with (\ref{subp15b}) restricted to its QCD aspect we get~\cite{herwiri,1305-0023} the identifications, accurate to ${\cal O}(\alpha_s)$,
\begin{equation}
\begin{split}
\frac{1}{2}\hat{\tilde{\bar\beta}}_{0,0}&= \bar{B}+(\bar{B}/\Delta_{MC}(0))\int(R_{MC}/B)\Delta_{MC}(k_T)d\Phi_R\\
\frac{1}{2}\hat{\tilde{\bar\beta}}_{1,0}&= R-R_{MC}-B\tilde{S}_{QCD}
\label{eq-mcnlo}
\end{split}
\end{equation}
where we defined~\cite{mcatnlo} $$\bar{B}=B(1-2\alpha_s\Re{B_{QCD}})+V+\int(R_{MC}-C)d\Phi_R$$ and we understand here
that the DGLAP-CS kernels in $R_{MC}$ are to be taken as the IR-improved ones
as we exhibit below~\cite{irdglap1,irdglap2}.
Here the QCD virtual and real infrared functions
$B_{QCD}$ and $\tilde{S}_{QCD}$ are understood to be DGLAP-CS synthesized as explained in Refs.~\cite{qced,irdglap1,irdglap2} to
avoid double counting of effects. In view of
(\ref{eq-mcnlo}),
the way to the extension of frameworks such as MC@NLO to exact higher
orders in $\{\alpha_s,\;\alpha\}$ is therefore open via our $\hat{\tilde{\bar\beta}}_{n,m}$
and will be taken up elsewhere~\cite{elswh,1407-7290}.
\par
In Ref.~\cite{1305-0023} we have presented detailed discussion of the relationship between our approach to precision QCD and those presented in Refs.~\cite{stercattrent1,scet1,colsop,colsopster}. We refer the reader interested in this relationship to
Ref.~\cite{1305-0023}. Here, we focus on the observation that strict control on the theoretical precision
in (\ref{bscfrla}) requires both the resummation of the reduced cross section
and that of the attendant evolution of the $\{F_j\}$. We turn now
to the latter.
\par
More precisely, we get an improvement
of the IR limit of the kernels, $P_{AB}$, in the DGLAP-CS theory itself
when we apply the QCD restriction of the formula in (\ref{subp15b}) to the
calculation of these kernels. This is the IR-improved DGLAP-CS
theory~\cite{irdglap1,irdglap2} in which large IR effects are resummed
for the kernels themselves.
The attendant new resummed kernels, $P^{\exp}_{AB}$, whose
implementation in the HERWIG6.5 environment generates the new
MC HERWIRI1.031~\cite{herwiri},
are as follows~\cite{irdglap1,irdglap2,herwiri}:
{\small
\begin{align}
P^{\exp}_{qq}(z)&= C_F F_{\mathrm{YFS}}(\gamma_q)e^{\frac{1}{2}\delta_q}\left[\frac{1+z^2}{1-z}(1-z)^{\gamma_q} -f_q(\gamma_q)\delta(1-z)\right],\nonumber\\
P^{\exp}_{Gq}(z)&= C_F F_{\mathrm{YFS}}(\gamma_q)e^{\frac{1}{2}\delta_q}\frac{1+(1-z)^2}{z} z^{\gamma_q},\nonumber\\
P^{\exp}_{GG}(z)&= 2C_G F_{\mathrm{YFS}}(\gamma_G)e^{\frac{1}{2}\delta_G}\{ \frac{1-z}{z}z^{\gamma_G}+\frac{z}{1-z}(1-z)^{\gamma_G}\nonumber\\
&\qquad +\frac{1}{2}(z^{1+\gamma_G}(1-z)+z(1-z)^{1+\gamma_G}) - f_G(\gamma_G) \delta(1-z)\},\nonumber\\
P^{\exp}_{qG}(z)&= F_{\mathrm{YFS}}(\gamma_G)e^{\frac{1}{2}\delta_G}\frac{1}{2}\{ z^2(1-z)^{\gamma_G}+(1-z)^2z^{\gamma_G}\},
\label{dglap19}
\end{align}}
where the superscript ``$\exp$'' indicates that the kernel has been resummed as
we described.
Here
$C_F$($C_G$) is the quadratic Casimir invariant for the quark(gluon) color representation respectively, and
the YFS~\cite{yfs} infrared factor
is given by $$F_{\mathrm{YFS}}(a)=e^{-C_Ea}/\Gamma(1+a)$$ where $\Gamma(w)$($C_E=0.57721566...$) is Euler's gamma function(constant), respectively.
The respective resummation functions $\gamma_A,\delta_A,f_A, A=q,G$ are given in Refs.~\cite{irdglap1,irdglap2}
\footnote{The improvement in Eq.\ (\ref{dglap19})
should be distinguished from the
resummation in parton density evolution for the ``$z\rightarrow 0$''
Regge regime -- see for example Refs.~\cite{ermlv,guido}. This
latter improvement must also be taken into account
for precision LHC predictions.}
\footnote{We follow the development of Field in Ref.~\cite{rfield} in removing the mass singularities such that our evolution variable for the DGLAP-CS equation is $t=\ln(Q^2/\Lambda^2)$ as it is given in eq.(4.6.11) in Ref.~\cite{rfield}.}. These new kernels provide us with a new IR-improved resummed scheme for the parton density functions (PDF's) and the reduced cross section with the same value of $\sigma$ in (\ref{bscfrla}):
\begin{equation}
\begin{split}
F_j,\; \hat\sigma &\rightarrow F'_j,\; \hat\sigma'\; \text{for}\\
P_{Gq}(z)&\rightarrow P^{\exp}_{Gq}(z), \text{etc.}
\end{split}
\label{newscheme1}
\end{equation}
As discussed in Refs.~\cite{herwiri}, this new scheme has improved MC stability
-- in the attendant parton shower MC HERWIRI1.031 based on the new kernels there is no need
for an IR cut-off `$k_0$'
parameter. While the degrees of freedom
below the IR cut-offs in the usual showers are dropped in those showers,
in the showers in HERWIRI1.031, these degrees of freedom are included in the calculation and are integrated over in the process of generating the Gribov-Lipatov exponents $\gamma_A$ in (\ref{dglap19}). We note also that, {\em as the differences between them start in ${\cal O}(\alpha_s^2)$, the new kernels
agree with the usual kernels at ${\cal O}(\alpha_s)$}. Thus,
for the realization of exact
NLO ME/shower matching the MC@NLO and POWHEG frameworks apply directly
to the new kernels.
\par
In Fig.~1, in the interest of pedagogy, the basic physical idea,
discussed by Bloch and Nordsieck in Ref.~\cite{bn1}, which underlies the new kernels is illustrated -- we show this because our approach is still not generally a familiar one:
\begin{figure}[h]
\begin{center}
\epsfig{file=figdglp1new.eps,width=70mm}
\end{center}
\label{fig-bn-1}
\caption{Bloch-Nordsieck soft quanta for an accelerated charge.}
\end{figure}
a coherent state of very soft massless quanta of the attendant gauge field is
generated by an accelerated charge so that one cannot know which of the
infinity of possible states
has been made in the splitting process $q(1)\rightarrow q(1-z)+G\otimes G_1\cdots\otimes G_\ell,\; \ell=0,\cdots,\infty$ shown in Fig.~1.
This effect is taken into account in the new kernels
by resumming the
terms ${\cal O}\left((\alpha_s \ln(\frac{q^2}{\Lambda^2})\ln(1-z))^n\right)$
in the $z\rightarrow 1$ IR limit.
We see from (\ref{newscheme1}) and (\ref{bscfrla})
that when the usual kernels are used these terms
are generated order-by-order in the solution for the cross section
$\sigma$ in (\ref{bscfrla}). Thus, for a given order of exactness in the
input perturbative components therein our
resumming them enhances the convergence of the
representation in (\ref{bscfrla}).
This last remark is illustrated in the next Section in the context of the comparison of recent LHCb data to
NLO parton shower/matrix element matched predictions.\par
\section{Interplay of IR-Improved DGLAP-CS Theory and NLO Shower/ME Precision: Comparison with LHCb Data}
In the new MC HERWIRI1.031~\cite{herwiri} we have the first realization of the new IR-improved kernels in the HERWIG6.5~\cite{hwg} environment. In Refs.~\cite{herwiri,1305-0023}, we have made comparisons with the data of ATLAS and CMS on the single $Z/\gamma*$ production and decay to lepton pairs and have been encouraged by the results of these comparisons. Here, we extend our comparisons to the more forward acceptance of the LHCb for the same processes.\par
We need to make the following important observation regarding the results
from Refs.~\cite{herwiri,1305-0023}. We have shown that the exact ${\cal O}(\alpha_s)$ correction makes a $11.5\%$ -- $12.8\%$ correction to the spectra
in $p_T$ and $Y$ analyzed therein. Following the methodology in Ref.~\cite{jad-prec}, we then estimate the attendant physical theoretical precision error
due to uncalculated higher orders as one-half of this effect, which we then double to be on the safe side. Thus, we will take the conservative estimate
$\Delta\sigma^{phys}_{th}/\sigma\cong 13\%$ for the physical theoretical precision in the plots which make herein as to be understood in interpreting the discussion which we give in what follows. This will reflect what we have learned from Refs.~\cite{herwiri,1305-0023} regarding the errors in our work on the theoretical side. This estimate should be considered preliminary and serves as a lower bound; we expect our final value to be somewhat larger. We would, however, like to emphasize that what we do here represents
a somewhat different approach to the error estimate than what is done in Ref.~\cite{fewz2}, for example, wherein the various scales of the calculation are varied by some arbitrarily chosen factors. The authors in Ref.~\cite{blm} have emphasized the probable limitations of such approaches. Our approach is based
on our experiences in the precision LEP EW studies~\cite{jad-prec} and
more generically on the analyses of Refs.~\cite{zinn-justin}.
As a practical matter,
we will omit our physical theoretical precision error estimate from the $\chi^2/\text{d.o.f.}$ which we use to
compare theory and data, so that our $\chi^2/\text{d.o.f.}$ are respective upper limits accordingly.\par
Specifically, with the recent LHCb data~\cite{lhcbdata} as our baseline,
we compare the predictions from HERWIRI1.031 with HERWIG6.510, with
the MC@NLO~\cite{mcatnlo} exact ${\cal O}(\alpha_s)$ correction included,
to illustrate the interplay between the attendant precision in NLO ME matched parton shower MC's
and the new IR-improvement for the kernels.
In Fig.~\ref{figlhcb1} in panel (a) we show for the single $Z/\gamma*$ production at the LHC
the comparison between the LHCb rapidity data for the $e^+e^-$ channel
and the MC theory predictions and in panel
(b) in the same figure we show the analogous comparison with the LHCb data
for the $\mu^+\mu^-$ channel.
\begin{figure}[h]
\begin{center}
\setlength{\unitlength}{0.1mm}
\begin{picture}(1600, 930)
\put( 400, 770){\makebox(0,0)[cb]{\bf (a)} }
\put(1240, 770){\makebox(0,0)[cb]{\bf (b)} }
\put( -50, 0){\makebox(0,0)[lb]{\includegraphics[width=90mm]{lhcb-rap1-2011.eps}}}
\put( 800, 0){\makebox(0,0)[lb]{\includegraphics[width=90mm]{lhcb-rap2-2011-mu.eps}}}
\end{picture}
\end{center}
\caption{\baselineskip=8pt Comparison with LHCb data: (a), LHCb rapidity data on
($Z/\gamma^*$) production to $e^+e^-$ pairs, the circular dots are the data, the green(blue) squares are MC@NLO/HERWIG6.510($\rm{PTRMS}=2.2$ GeV/c)(MC@NLO/HERWIRI1.031);
(b), LHCb rapidity data on ($Z/\gamma^*$) production to (bare) $\mu^+\mu^-$ pairs, with the same graphical notation as that in (a). In both (a) and (b), the green triangles are MC@NLO/HERWIG6.510($\rm{PTRMS}=$0). These are otherwise untuned theoretical results.
}
\label{figlhcb1}
\end{figure}
These results should be considered from the perspectives of our analysis in Refs.~\cite{herwiri} of the FNAL data on the single $Z/\gamma^*$ production in
$\text{p}\bar{\text{p}}$ collisions at $1.96$ TeV and our analysis in Ref.~\cite{herwiri,1305-0023} of the LHC ATLAS and CMS data single $Z/\gamma^*$ production in
$\text{p}\text{p}$ collisions. More precisely, what we found in Refs.~\cite{herwiri,1305-0023} was that the IR-improvement in HERWIRI1.031 allowed it to give a better $\chi^2/\text{d.o.f}$ to the FNAL and ATLAS and CMS data without the need
of a large intrinsic value $\text{PTRMS}$ on the scale of the expectations from successful models of the proton~\cite{pwvfn}, which would have $\text{PTRMS}\simeq 0.4$ GeV/c, as would the precociousness of Bjorken scaling~\cite{bj1,scaling}. This is in contrast to the unimproved results from HERWIG6.5, wherein such a large value as $2.2$ GeV/c for $\text{PTRMS}$ was needed to get similar values of
$\chi^2/\text{d.o.f}$ for the $p_T$ spectra but for the rapidity data
such a large value of $\text{PTRMS}$
was not necessary to get the corresponding similar
values. What we see in Fig.~\ref{figlhcb1} is that the situation is similar
in the LHCb rapidity data. The values of the $\chi^2/\text{d.o.f}$
are 0.746, 0.814, 0.836 for the respective predictions from
MC@NLO/HERWIRI1.031, MC@NLO/HERWIG6.5($\text{PTRMS}=0$) and
MC@NLO/HERWIG6.5($\text{PTRMS}=2.2$ GeV/c) for the $e^+e^-$ data and are
0.773, 0.555, 0.537 for the respective predictions for the $\mu^+\mu^-$ data.
All three calculations give acceptable values of $\chi^2/\text{d.o.f}$.\par
When turn to the transverse momentum degrees of freedom, the situation changes
relative to what we found in Refs.~\cite{herwiri,1305-0023} for D0, CDF, ATLAS and CMS. We start with the $\phi_\eta^*$ LHCb data in Ref.~\cite{lhcbdata}. Specifically, we have
\begin{equation}
\phi_\eta^*=\tan(\phi_{\text{acop}}/2)\sqrt{1-\tanh^2(\Delta\eta/2)}
\label{eqn-phi-eta*}
\end{equation}
where $\Delta\eta=\eta^--\eta^+$ when $\eta^-$ and $\eta^+$ are the respective
negatively and positively charged lepton pseudo-rapidities and $\phi_{\text{acop}}=\pi - \Delta\phi$ with $\Delta\phi$ as defined in Sect. 1. This variable is not exactly the same as the $p_T$ of the produced $Z/\gamma^*$ but is correlated with it, where as shown in Ref.~\cite{atlas-phi-star} there is a longer and longer tail in the correlation as $\phi_\eta^*$ decreases toward smaller values.
Moreover, there is significant bin-to-bin migration in the lower bins of
the $\phi_\eta^*$ data as it has been unfolded in Ref.~\cite{lhcbdata}. Thus,
unless we have the same MC unfolding effect for each MC which we use in our comparisons, we cannot address the uncertainty of the comparison where the migration is substantial. Accordingly, we restrict the comparison to the regime where the unfolding effect is minimal until such a time when each MC will have been used for the unfolding corrections. Again, the comparisons we show here are otherwise untuned comparisons. And, we only show the MC@NLO/A results,
for A = HERWIG6.5($\text{PTRMS}=0$), HERWIG6.5($\text{PTRMS}=2.2$ GeV/c) and HERWIRI1.031, where
as usual we always set $\text{PTRMS}=0$ in HERWIRI1.031 simulations.
The corresponding results are shown in Fig.~\ref{figlhcb2}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=100mm]{lhcb-pestr-2011-mu-lowcut-3-15.eps}
\end{center}
\caption{\baselineskip=8pt Comparison with LHCb data on $\phi_\eta^*$ for the $\mu^+\mu^-$ channel in single $Z/\gamma^*$ production at the LHC. The legend (notation) for the plots is the same as in Fig.~\ref{figlhcb1}.
}
\label{figlhcb2}
\end{figure}
The respective $\chi^2/\text{d.o.f.}$ are 1.2, 0.23, 0.35 for the MC@NLO/ERWIRI1.031, MC@NLO/HERWIG6.5($\text{PTRMS}=0$), MC@NLO/HERWIG6.5($\text{PTRMS}=2.2$ GeV/c) simulations. Thus, all three simulations give acceptable fits to the data, with the curious result that the MC@NLO/HERWIG6.5\\
($\text{PTRMS}=0$) gives a very mildly better fit than does
MC@NLO/HERWIG6.5($\text{PTRMS}=2.2$ GeV/c) -- we would caution here
already that we really do not have the errors under such control that
we could take $|\Delta\chi^2/\text{d.o.f.}|\simeq 0.1$ as significant.
These results make us recall the difference between the $\phi_\eta^*$
variable and the $p_T$ of the $Z/\gamma^*$ as well as, perhaps,
the difference between the forward
and more central observations, as
we have seen in Refs.~\cite{herwiri,1305-0023} that a good $p_T$-fit for the central region with HERWIG6.5 is not possible with $\text{PTRMS}=0$. We thus look next at the more forward LHCb data on $p_T$.
\par
In analyzing the LHCb $p_T$ data in Ref.~\cite{lhcbdata}, we note that there is significant bin-to-bin migration at low $p_T$ in the first bin. Since the different MC's have significantly different predictions in the low $p_T$ regime below $4$ GeV/c, until these effects are determined with each MC, it is not appropriate in our approach to try to use the data to assess the accuracy of the MC's in this regime. What we will do again here is to use the data where these type of effects are minimal and wait until
we can get data wherein each MC has been used to assess these migration effects before we try to assess the regime below $p_T \cong 4$ GeV/c. We call this our ``conservative'' approach to the data.\par
With this understanding, we show in Fig.~\ref{figlhcb3} the comparisons between the three\\
MC@NLO/A predictions and the LHCb $p_T$ data on single $Z/\gamma^*$ production and decay to $\mu^+\mu^-$ pairs, where, again, A = HERWIG6.5($\text{PTRMS}=0$), HERWIG6.5($\text{PTRMS}=2.2$ GeV/c) and HERWIRI1.031, and
as usual we always set $\text{PTRMS}=0$ in HERWIRI1.031 simulations.
\begin{figure}[h]
\begin{center}
\includegraphics[width=100mm]{lhcb-pt-2011-mu-intm-3-15.eps}
\end{center}
\caption{\baselineskip=8pt Comparison with LHCb data on $p_T$ for the $\mu^+\mu^-$ channel in single $Z/\gamma^*$ production at the LHC. The legend (notation) for the plots is the same as in Fig.~\ref{figlhcb1}.
}
\label{figlhcb3}
\end{figure}
What we find is that the respective $\chi^2/\text{d.o.f.}$ are 0.183, 0.103, 0.789 respectively. We see that for the more forward LHCb data all three calculations give acceptable fits to the data with only a very mild indication that
the $\text{\rm PTRMS}=2.2$ GeV/c HERWIG6.5 results give a better fit than do the
$\text{\rm PTRMS}=0$ GeV/c HERWIG6.5 results, in contrast with what was found
in Refs.~\cite{herwiri,1305-0023} for the D0 and ATLAS $p_T$ data and in general
agreement with our results in Fig.~\ref{figlhcb2} for the $\phi_\eta^*$ variable when we recall again that we really do not have the errors under such control that we can consider $|\Delta\chi^2/\text{d.o.f.}|\simeq 0.1$ as significant. We conclude that, when we look over the data on single $Z/\gamma^*$ production at FNAL(CDF and D0) and at LHC(ATLAS, CMS and LHCb), as we have shown in Refs.~\cite{herwiri,1305-0023} and with the results presented here, HERWIRI1.031 gives a good fit to all the data analyzed without the need of ad hocly hard intrinsic $\text{\rm PTRMS}$ whereas HERWIG6.5 needs such a value at $\sim 2$ GeV/c in order to give a good fit to all of these data for both the rapidity and the transverse momentum based observables. This gives us a well-defined starting point from which to set a baseline~\cite{1407-7290} for a rigorous treatment
of the theoretical precision tag on such processes at the LHC and at FCC~\cite{fcc}.\par
\section{Conclusions}
What we have shown is the following. The realization of IR-improved DGLAP-CS theory
in HERWIRI1.031, when used in the MC@NLO/HERWIRI1.031 exact ${\cal O}(\alpha_s)$ ME matched parton shower framework,
affords one the opportunity to explain, on an event-by-event basis, both the rapidity and the $p_T$ dependent spectra of the $Z/\gamma^*$ in pp collisions
in the recent LHC data from the LHCb, respectively, without the need of an
unexpectedly hard intrinsic Gaussian $p_T$ distribution with rms value of $\rm{PTRMS}\cong 2$ GeV/c in the proton's wave function. This extends a similar conclusion to the LHCb that we had established in Refs.~\cite{herwiri,1305-0023} for the ATLAS and CMS data. Our view is that this can be interpreted as providing
further support for a rigorous basis for the phenomenological correctness
of such unexpectedly hard distributions insofar as describing these data using the usual unimproved DGLAP-CS showers is concerned. Accordingly, we
continue to propose
that comparison of other distributions such as the invariant mass distribution
with the appropriate cuts and the more detailed $Z/\gamma^*$ $p_T$
spectra in the regime below $10.0$GeV/c be used to
differentiate between these phenomenological
representations of parton shower physics
in MC@NLO/HERWIG6.510 and the fundamental description of the parton shower physics in MC@NLO/HERWIRI1.031. We recall that elsewhere~\cite{1305-0023} we have further emphasized that the precociousness of Bjorken scaling~\cite{scaling,bj1} argues against the fundamental correctness
of the {\em hard} scale intrinsic $p_T$ ansatz with the unexpectedly large value of $\rm{PTRMS}\cong 2$ GeV/c, as do the successful models~\cite{pwvfn} of the proton's wave function,
which would predict this value to be $\lesssim 0.4$ GeV/c. As we have emphasized as well elsewhere~\cite{1305-0023},
we point out that the fundamental description in MC@NLO/HERWIRI1.031 can be systematically improved to the NNLO parton shower/ME matched level~\cite{znder-hoeche} -- a level which we anticipate is a key ingredient in achieving the (sub-)1\% precision tag for such processes as single heavy gauge boson production at the LHC.
Our comparisons with the LHCb data are not inconsistent the proposition that our methods should work in its region of the acceptance phase space as well as they do in the
acceptance phase spaces for ATLAS and CMS. \par
In closing, one of us (B.F.L.W.)
thanks Prof. Ignatios Antoniadis for the support and kind
hospitality of the CERN TH Unit while part of this work was completed.\par
|
\section{Introduction}
\label{sec:intro}
The Cosmic Web (CW) is the most salient manifestation of the anisotropic nature of gravitational collapse, the motor behind
the formation of structure in the cosmos. Structure in the Universe has risen out of tiny primordial (Gaussian) density
and velocity perturbations by means of gravitational instability. N-body computer simulations have profusely illustrated
how the primordial density field transforms into a pronounced and intricate filigree of filamentary features, dented by
dense compact clumps at the nodes of the network \citep{Bond1996}. Moreover, because of the hierarchical nature
of the cosmic matter distribution, also filaments, sheets and voids emerge
by the gradual merging of smaller scale specimen. In the process, small filaments align themselves along the direction of
the emerging larger scale filament. The emerging picture is therefore one of a primordially and hierarchically defined skeleton
whose weblike topology is imprinted over a wide spectrum of scales. Weblike patterns on ever larger scales get to dominate
the density field as cosmic evolution proceeds, and as small scale structures merge into larger ones. On the other hand, within
the gradually emptying void regions the topological outline of the early weblike patterns remains largely visible as a faint
remnant of past glory (see {\it e.g.} Fig.~\ref{fig:nexus_web}).
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.6\textwidth]{NEXUS_web.eps}
\end{center}
\caption{Rendering of the Cosmic Web elements from Millenium Simulation II as identified by the NEXUS algorithm.
Different colours mark different elements of the Cosmic Web: dense nodes (clusters) are marked by red,
elongated filaments are yellow, the pervading network of walls we depict by green and the remaining empty
spaces correspond to cosmic voids.}
\label{fig:nexus_web}
\end{figure}
So far most of the authors have focused on the clustering and density-based statistics analysis of the CW, as it is
seen in computer N-body simulations and galaxy catalogues
\citep[{\frenchspacing\it e.g.}][]{Springel2006,AragonCalvo2010,Cautun_cweb_evo_2014,Metuki2014,Falck_cweb2014,Nuza_galaxy_cweb2014}. In this contribution
we want to adopt a different, yet complementary to previous studies, approach. Namely we will study the dynamics of the Cosmic
Web elements as reflected by the statistics of pairwise motions observed inside different environments. The dynamical studies
of the CW elements are very important, both for providing complementary understanding their internal dynamics and kinematics,
and for established whether the observed large-scale structure environments are just reflection of the geometrical and spatial
galaxy distribution correlations or consists of a more dynamically coherent objects.
\section{The pairwise motions}
\label{sec:pairwise}
In this work we will consider lower-moments of the statistics of pairwise Dark Matter (DM) velocities as tracers
of the underlying dynamics driven by gravity in different CW environments.
The statistic of the mean relative pairwise velocity of galaxies $v_{12}$ --
the \textit{streaming velocity} -- reflects the ``mean tendency of well-separated galaxies to approach each other''
\citep{1980Peebles}. This statistic was
introduced by Davis\&Peebels \citep{DavisPeebles_BBGKY} in the context of the kinetic BBGKY theory which describes the dynamical
evolution of a system of particles interacting via gravity.
In the fluid limit, its equivalent is the pair-density weighted relative velocity
\begin{equation}
\label{eqn:v12-wieghted}
\mathbf{v}_{12}(r) = \langle\mathbf{v}_1-\mathbf{v}_2\rangle_{\rho} =
{\langle(\mathbf{v}_1-\mathbf{v}_2)(1+\delta_1)(1+\delta_2)\rangle\over 1+\xi(r)}\,\,,
\end{equation}
where $\mathbf{v}_1$ and $\delta_1=\rho_1/\langle\rho\rangle-1$ stand for the peculiar velocity and fractional matter density
contrast taken at
point $\mathbf{r}_1$, $r=|\mathbf{r}_1-\mathbf{r}_2|$, and $\xi(r)=\langle\delta_1\delta_2\rangle$ is the 2-point density
correlation function.
The $\langle\cdots\rangle_{\rho}$ is the pair-weighted average, which differs from normal spatial averaging by the weighting factor
$\mathcal{W}=\rho_1\rho_2/\langle\rho_1\rho_2\rangle$. Note $\mathcal{W}$ is proportional to the number density of pairs. The
gravitational instability theory predicts that the $v_{12}(r)$ magnitude is governed by the 2-point correlation function $\xi(r)$ and
the growth rate of matter density perturbations $f\equiv d\ln D_{+}/d\ln a$ (where $D_{+}(a)$ is the linear growing mode solution,
and $a$ is the cosmological scale factor) through the pair conservation equation \citep{1980Peebles}.
\cite{Juszkiewicz1999} provided a closed-form expression that is a good approximation to the solution of the pair conservation
equation for universes with Gaussian initial conditions:
\begin{equation}
v_{12}=-{2\over3}H_0rf\bar{\bar{\xi}}(r)[1+\alpha\bar{\bar{\xi}}(r)]\,\,,
\label{eqn:v12_ansatz}
\end{equation}
where
\begin{equation}
\bar{\xi}(r)=(3/r^3)\int_0^r\xi(x)x^2dx\equiv\bar{\bar{\xi}}(r)[1+\xi(r)]\,\,.
\label{eqn:xi_spec}
\end{equation}
Here, $\alpha$ is a parameter that depends on the logarithmic slope of $\xi(r)$ and $H_0=100\,h\,$km s$^{-1}$ Mpc$^{-1}$ is
the present day value of
the Hubble constant. It is clear that $v_{12}(r)$ is a strong function of $\xi(r)$ and $f$, which both in general will
take different local averages in different CW environments. This motivates us to consider low-order moments of
the pairwise velocity distribution
as tracers of CW specific local dynamics that should be reflected in motions of galaxies and DM.
Specifically, in our analysis we will
consider the following quantities:
\begin{itemize}
\item the mean radial pairwise velocity, $v_{12}$;
\item the dispersion in the (radial) pairwise velocities (not centred), $\sigma_{\parallel}=\langle v_{12}^2\rangle^{0.5}$;
\item the mean transverse velocity of pairs, $v_{\perp}$;
\item the dispersion of the transverse velocity of pairs, $\sigma_{\perp}=\langle v_{\perp}^2\rangle^{0.5}$.
\end{itemize}
It is well known that the cosmic velocity field has large coherence length, much
larger than the density field \citep[{\frenchspacing\it e.g.}][]{Chod2002,Ciec2003}.
Thus the contribution of the large-scale velocity modes is always significant, when one concerns the peculiar velocities.
However this disadvantage (from the point of view of probing the velocity properties driven by local environment)
is not present for the pairwise velocity statistics we have just described above.
This is because at a given separation $r$ the velocity difference between a galaxy pair does not receive any net contribution from
modes with wavelengths larger than the considered separation, since those give the same contribution for both particles/galaxies.
Hence, at larger scales, for which the galaxies in a pair inhabit different haloes, the distribution of $v_{12}$ factorises into
two individual peculiar velocity distributions for each galaxy/particle and those are always sensitive to non-linearities driven
by virial motions within a given host halo alone (see \citet{Scoccimarro2004} for more details). Thanks to this,
the pairwise velocity statistics is automatically free of large-scale velocity modes contributions, and hence
is very well suited for the kind of analysis we want to perform.
\section{Dynamics of the Cosmic Web}
\label{sec:dynamics}
To study the dynamics of pairwise motions in different large-scale structure environments we use
the DM density and velocity fields from the Millennium-II simulation \citep{MS2}. To segment
the simulation volume into voids, walls, filaments and cluster voxels we use the \verb#NEXUS+# algorithm \citep{Nexus}
applied on the DM density field computed on a $256^3$ grid (with a grid cell width of $0.39h^{-1}\,{\rm Mpc}$). For such defined
setup we obtain a catalogue of voxels fully segmenting the 3D simulation cube into four possible different CW
environments (see Fig.~\ref{fig:nexus_web}). The construction of
the \verb#NEXUS+# algorithm is such that
the determination of the local environment is based on spatial variations along three Cartesian directions in the density field.
As a consequence
our voids correspond to spatial regions that are relatively empty with spherical-like symmetry, wall constitute thin planar
structures, finally the filaments are elongated fibres. Our node environment corresponds to regions centred around massive
clusters and their local infall regions.
Having assessed the segmentation of our simulation volume into the CW elements we have computed all related pairwise
statistics ($v_{12}$, $\sigma_{\parallel}$, $v_{\perp}$ and $\sigma_{\perp}$) for randomly Poisson sub-sampled DM particles.
We use $1/100$th of the original number of DM simulation particles. This sub-sampling leave us with a sufficient
number of DM particles to study pairwise motions statistics for separations $\geq0.5h^{-1}\,{\rm Mpc}$, since at this separation we
have an averaged number of 10 particle pairs per voxel.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.48\textwidth]{pdf_v12_dm_r05mpc.eps}
\includegraphics[width=0.48\textwidth]{pdf_v12_dm_r2mpc.eps}
\end{center}
\caption{{\scriptsize The probability density functions for the mean pairwise streaming velocity $v_{12}$ computed at
pair separation of $0.5h^{-1}\,{\rm Mpc}$ (left panel) and $2h^{-1}\,{\rm Mpc}$ (right panel). The line marking is the following:
black with open circles - averaged for all cosmic environments ({\it i.e.} whole volume), red line - node/cluster
environments, blue - filaments, orange - walls and we use green for cosmic voids.}}
\label{fig:pdf_v12}
\end{figure}
We begin by plotting the probability density functions for $v_{12}$ values computed at two different averaged pair
separations: $0.5$ and $2h^{-1}\,{\rm Mpc}$. The results are shown in the two panels of the Fig.~\ref{fig:pdf_v12}. The reference
point is marked always by a solid black line with open circle data points and corresponds to a result obtained from
the whole uniform volume ({\it i.e.} without segmenting into different morphology elements). The data corresponding
to the following environments we mark by: red (node/cluster), blue (filaments), orange (walls) and green for voids.
The both panels illustrate the striking difference of the $v_{12}$ distribution functions among different environments.
We denote that the width of the distribution (roughly corresponding here to dispersion) increase significantly
as we move from voids and walls towards denser environments of filaments and clusters. This is clearly visible at
both considered separations. This indicate that pairwise motions are much hotter in nodes and filaments than
in voids and walls. As we have described above, the mean streaming pairwise velocity is very sensitive
to perturbation modes whose length is smaller than a considered pair separation. This property
of the pairwise statistics is now clearly consistent with our measured $v_{12}$ PDFs. It is evident that in environments
like cosmic clusters and filaments the degree of non-linearity (in both the density and velocity field) is much
higher compared to relatively quieter cosmic walls and filaments. This difference is mostly driven by violent
relaxation processes and thermalised virial motions that tend to dominate the peculiar velocity field inside
clusters and dense filaments regions \citep[{\frenchspacing\it e.g.}][]{Cautun_cweb_evo_2014,Nexus}. Secondly we can denote that the degree
of (a)symmetry of the PDFs are subject to a significant environment-driven variation. This can be especially seen
on the left panel of Fig.~\ref{fig:pdf_v12}, where we plot the probability density functions for pairs at $R_p=2h^{-1}\,{\rm Mpc}$
separation. The {\it stable clustering} model \citep{1980Peebles} predicts that at weakly non-linear scales
there should be slight excess of negative pairwise velocities reflecting the gravitational clustering mechanism that
still operates at those scales. Namely between two regimes of fully virialised motions at small scales and Hubble
expansion dominated at large scales there should be a region where some significant part of the peculiar motions
still have potential character. We can observed this by noting that the cluster/node PDF function on the right panel
has much higher symmetry then the PDFs of the rest of the environments. Such symmetric $v_{12}$ PDF is a characteristic
of a fully randomised motions supported by velocity dispersion. This is the case inside virialised haloes, which clearly
dominate the pairwise flows signal in our node/cluster environments.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.48\textwidth]{v12_rp_envs_DM.eps}
\includegraphics[width=0.48\textwidth]{sig_p_sig_pa_envs_dm.eps}
\end{center}
\caption{{\scriptsize The minus mean streaming velocities - $-v_{12}$ (left panel) and parallel
and perpendicular pair velocities dispersions
(right panel) - $\sigma_{\parallel}$, $\sigma_{\perp}$ - as a function of a pair separation $R_p$.
The lines colouring is consistent with the previous figure.}}
\label{fig:v12_sigmas_Rp}
\end{figure}
As a complementary probe of the pairwise motions we also study the minus mean infall velocities ($-v_{12}$)
and parallel and perpendicular pair velocities dispersion ($\sigma_{\parallel}$, $\sigma_{\perp}$) as
a function of a pair separation $R_p$. We plot the corresponding results in the left and right panel
of the Fig.~\ref{fig:v12_sigmas_Rp}. We use the same colours as before to mark results for different
CW elements. On the left panel for comparison we also draw a purple line depicting the Hubble
expansion velocity. The results shown here are consistent with the previously discussed PDF properties.
We note both higher values of the mean infall velocities as well as much higher dispersions of both
the infall and perpendicular pair-velocity components for high density environments of nodes and filaments.
In contrast the pairwise motions in walls and especially in voids are characterised by very small
dispersions (are very cold). We can also denote another interesting observation regarding the
dynamics of pairwise motions in cosmic voids. Namely from the shape of the $-v_{12}(R_p)$ curve on the left
panel we can infer that, on average, the pairwise motions in voids have {\it super Hubble} character.
This means that inside the cosmic voids the expansion is faster then the background Hubble expansion rate.
Although this result was known for some time \citep[{\frenchspacing\it e.g.}][]{Seth2004}, as the inside regions of cosmic
voids can be treated locally as approximate Friedman Universes with lower effective $\Omega_m$ value
then the cosmic mean, it is very interesting to confirm this result in our data for pairwise motions.
The last thing we want the emphasise here is the lack of the offset between two different velocity directions
dispersions for the cosmic voids. This can be gauged by the offset between solid and dashed lines of the
same colour on the right panel of the Fig.~\ref{fig:v12_sigmas_Rp}. The offset between $\sigma_{\perp}$ and
$\sigma_{\parallel}$ is largest for the cluster environment and decrease as we move to filaments and walls.
However even for cosmic walls this offset is still significant (of the order of $\sim 33\%$), while
in the case of the cosmic voids we observe that values and shapes of the both dispersions lines are
fully consistent with each other. We speculate that this feature is driven by two facts: (i) the structure
formation has ceased inside cosmic voids, and (ii) there are no preferred directions for the voids segments
reflecting their average close-to spherical symmetry.
\section{Conclusions}
We have studied the dynamics and statistical properties of dark matter particles pairwise motions
in four different CW elements: nodes/clusters, filaments, walls and voids. We have shown that
both the probability density functions for the mean infall velocities as well as scale dependence
of $v_{12}$ and two connected velocity dispersions are taking significantly different values and shapes
in different large-scale environments. The flows in voids and wall are much colder and also have
smaller magnitude, when compare to pairwise motions in denser node and filament environments. This indicate
that the \verb#NEXUS+# algorithm is very robust in determining different CW elements. The algorithm
does not use any velocity data whatsoever, and yet the CW elements delineated by this method
are characterised by pairwise motions whose statistics are consistent with simple theoretical predictions.
Our finding also indicate that the CW and its elements, as first defined by \citet{Bond1996}, is not only real
in the sense of the DM/galaxy density field morphology, but is also present at the level of peculiar motions
of DM. This is clearly seen once the contribution of large-scale velocity modes was removed, which can be
obtained for example by considering pairwise statistics as we did in this study. Such a strong correlations between
the distribution and dispersion of mean streaming velocities and the CW morphology also indicate
that the cosmic velocity-density relation has different limits of validity in different environments.
This reflects the different degree of non-linearity in the structure formation in voids, walls, filaments and nodes.
\section*{Acknowledgements}
This work would not be possible without a significant contribution of Marius Cautun, Rien van de Weygaert
and Carlos S. Frenk. I am very grateful for their help and many stimulating discussions.
The author acknowledge the support received from Polish National Science Center in grant
no. DEC-2011/01/D/ST9/01960 and ERC Advanced Investigator grant of C.~S.~Frenk, COSMIWAY.
\bibliographystyle{hapj}
|
\section{Introduction}
Models based on the left--right symmetric gauge group $G_{3221} \equiv SU(3)_c \times SU(2)_L \times SU(2)_R \times U(1)_{B-L}$ \cite{lr}
are attractive extensions of the Standard Model (SM) with several interesting features. At the fundamental level Parity
is a good symmetry in these models. The observed Parity violation in weak interactions is explained by the spontaneous breaking of $SU(2)_R \times U(1)_{B-L}$ down to $U(1)_Y$ of the SM at a scale $v_R$ well above the masses of the $W$ and $Z$ bosons. The gauge structure requires the
existence of the right--handed neutrino, and thus leads naturally to small neutrino masses via the seesaw mechanism. In fact,
with the right--handed neutrino included, $G_{3221}$ is the maximal flavor--blind gauge symmetry that can be realized at a scale of order TeV, relevant to the ongoing LHC experiments.\footnote{There is a natural embedding of $G_{3221}$ into the Pati--Salam symmetry $G_{422} \equiv SU(4)_c \times SU(2)_L \times SU(2)_R$ \cite{ps}, however, the scale of $G_{422}$ symmetry breaking must be of order $10^5$ GeV or above, from
$K_L \rightarrow \mu e$ decay constraints. Embedding $G_{3221}$ (or $G_{422}$) into the unified symmetry group of $SO(10)$ is very natural, but that
symmetry breaking scale must be of order $10^{15}$ GeV, from constraints on nucleon decay and gauge coupling unification.} Because
of Parity invariance these models can potentially solve the strong CP problem \cite{strongcp} without introducing a global Peccei--Quinn symmetry and
the resulting axion.
Supersymmetric versions of left--right gauge models, denoted here as SUSYLR models, preserve the merits of $G_{3221}$ noted above, and in addition, can solve the gauge hierarchy problem. These models can have a natural dark matter candidate in the
lightest supersymmetric particle, with an unbroken $R$-parity emerging from the $U(1)_{B-L}$ gauge symmetry \cite{mohapR}.
It has been noted that the puzzle of small phases in the SUSY breaking sector
(arising from electric dipole moment constraints) has a natural explanation
in SUSY left--right models, by virtue of Parity symmetry \cite{susycp}. Several versions of the SUSYLR models have been proposed and
studied in the literature, with differing Higgs boson sectors used for symmetry breaking \cite{bhm,ma,km,susylr,bm0,amz}.
Here we undertake a systematic study of the Higgs potential in various realizations of these models, focussing on the
lightest neutral Higgs boson mass $M_h$. In many cases
we find that the tree--level constraint $M_h \leq m_Z$ of the MSSM is modified to the less stringent constraint $M_h \leq \sqrt{2}\, m_W$ \cite{bhm}. In some models this limit is relaxed even further.
This difference in the upper limit arises from the non-decoupling $D$--terms of $SU(2)_R \times U(1)_{B-L}$, which occurs
when the symmetry breaking scale $v_R$ and the SUSY breaking scale are of the same order. Thus, these models
would predict additional $W^\pm_R$ and $Z_R$ gauge bosons within reach of LHC experiments, in addition to SUSY particles in the parameter regime where the upper limit on $M_h$ is relaxed.
In the MSSM heavy stops ($m_{\tilde{t}} > 2$ TeV) with large mixing are typically needed in order to accommodate the Higgs boson of mass 126 GeV
discovered recently at LHC. Such a large mass of the stop puts the gauge hierarchy problem in a different perspective, since
some amount of tuning would be required. With the increased tree--level mass of $M_h$, SUSYLR models would allow for the stops
to be much lighter and less mixed, and thus would alleviate the tuning problem.
Our analysis focuses on two basic classes of SUSYLR models which have been developed in the literature.
In one class Higgs triplets are introduced for $SU(2)_R$ symmetry breaking along with $SU(2)_L \times
SU(2)_R$ bi-doublets which break the electroweak symmetry \cite{km,susylr,bm0,amz}. Fermion mass generation is via direct Yukawa couplings in this class of models,
including the Majorana mass of the right-handed neutrino. In a second class, Higgs doublets are used to break $SU(2)_R$ symmetry, with $SU(2)_L$
doublets and/or $SU(2)_L \times SU(2)_R$ bi-doublets breaking the electroweak symmetry. Additional fermions are necessary in this class for fermion mass
generation, at least in the neutrino sector. A specific example studied incorporates the
inverse seesaw mechanism for neutrino masses with the inclusion of gauge singlet fermions. Another
example, termed alternate left--right model \cite{ma,bhm}, has an $E_6$ inspired particle spectrum. A third example uses
a universal seesaw mechanism for quarks and leptons by introducing vector--like $SU(2)_L \times SU(2)_R$
gauge singlet quarks and leptons \cite{bm2,universal}. In each of the cases listed above,
we also allow for the presence of a Higgs singlet scalar, which would admit the possibility of $SU(2)_R \times U(1)_{B-L}$ symmetry breaking down to $U(1)_Y$ in the SUSY limit.
In the absence of such a gauge singlet Higgs field this symmetry breaking scale would be of the same order as the supersymmetry breaking scale, which is shown to be consistent with experimental limits.
In certain cases with Higgs singlets we present approximate analytic expressions for the limit on $M_h$ that interpolate between the limit $M_h \leq m_Z$ of the MSSM if the $SU(2)_R$ breaking scale is much higher than the
SUSY breaking scale, and the limit $M_h \leq \sqrt{2}\, m_W$ which arises if this scale is comparable to the SUSY breaking scale.
Non--decoupling $D$--term effects on the lightest Higgs boson mass in extensions of the MSSM have been studied by various
authors. In Ref. \cite{bhm} symmetry breaking in SUSYLR models with an $E_6$ inspired particle spectrum was studied and
a relation $M_h \leq \sqrt{2} \,m_W$ was derived. In Ref. \cite{amz} symmetry breaking of SUSYLR models with
Higgs triplets was studied and an enhancement of $M_h$ compared to the MSSM result was observed. Ref. \cite{bdkt} has studied extended gauge
sectors, including an extra $SU(2)$ added to the SM gauge symmetry. In some cases there is an unknown gauge coupling, which was chosen so that it remains perturbative all the way to a GUT scale, and significant increase in $M_h$ was observed. In Ref. \cite{malinsky}
non--decoupling effects of an additional $U(1)$ gauge symmetry was studied, which also showed a modest increase in $M_h$.
Our aim in this paper is to systematically study the Higgs boson sectors of various realizations of SUSYLR models, which
has some overlap with some of the earlier studies. In one case we reproduce and generalize the results of Ref. \cite{bhm}. In another case studied, where we provide an analytic formula for the upper limit on $M_h$ that interpolates between the decoupling and non--decoupling limits of left--right symmetry, our results agree roughly with the numerical results of Ref. \cite{amz}.
We provide complete listings of the Higgs spectra for each case studied. Many of the examples, such as the
inverse seesaw model and the universal seesaw model, are analyzed in the SUSYLR framework for the first time here.
When gauge singlets that couple to the MSSM Higgs fields are present in the theory, additional $F$--term contributions to
$M_h$ arise. In several cases this contribution is non--decoupling, a well-known case being the
next to minimal supersymmetric standard model (NMSSM) \cite{ellwanger}. Modest increase in
$M_h$ can arise from this contribution, although we find the non--decoupling $D$--term to be somewhat more significant in the SUSYLR models.
There are direct and indirect limits on the mass of the $W_R^\pm$ gauge boson of left--right symmetric models with or without supersymmetry. CMS collaboration has obtained a lower limit on $M_{W_R}$ that ranges from 1.8 TeV to 3 TeV, depending on the mass of the right--handed neutrino, if the gauge coupling of $W_R$ is the same as that of the Standard Model
$W$ boson \cite{cms}. This limit is obtained with 19.6 $fb^{-1}$ data collected at 8 TeV by looking for an excess in the $eejj$ and $\mu \mu jj$ channels in $pp$ collision \cite{keung}. For a discussion of limits on the $W_R$ mass in the minimal left-right symmetric model see Ref. \cite{miha}. Indirect limits on $W_R$ mass arise from box diagram contributions
to $K^0-\overline{K^0}$ mixing which is found in the minimal model to be $m_{W_R} > 1.8$ TeV \cite{soni}.
These limits are all compatible with the scale of left--right symmetry breaking comparable to that of SUSY breaking.
In minimal SUSYLR models with Higgs triplets used for $SU(2)_R$ symmetry breaking, a doubly charged scalar remains
light with its mass below a TeV. This state acquires its mass entirely through radiative corrections. The consistency
of such a framework was shown in Ref. \cite{bm0}. In this paper we carry out a complete one--loop calculation of the
doubly charged Higgs boson mass in these models arising from the Majorana Yukawa couplings. We demonstrate the finiteness
of the mass, and show that if the right--handed symmetry breaking scale is taken to be much above the SUSY breaking
scale, the squared mass of this field would be negative. Minimal models with Higgs triplets would be suggestive of low energy $SU(2)_R$ breaking.
The remainder of the paper is organized as follows. In Sec. II we introduce the various versions of SUSY left-right
model. Here we list the Higgs content and explain how realistic fermion masses, including neutrino masses, are generated. In Sec. III we analyze the Higgs potentials of SUSYLR models with triplet
scalars breaking the $SU(2)_R \times U(1)_{B-L}$ symmetry. Various scenarios are discussed here. For electroweak symmetry breaking we allow for one or two bi-doublets. We also allow for a gauge singlet that facilitates LR symmetry breaking in the SUSY limit. We focus on the lightest neutral Higgs boson mass and derive the tree--level constraint $M_h \leq \sqrt{2}\,m_W$
in one case. In Sec. IV we analyze inverse seesaw models which utilize Higgs doublets and bidoublets. Sec. V has our results on the universal seesaw models which contain
only Higgs doublet fields. In Sec. VI, an $E_6$ inspired SUSYLR model is studied in detail. Sec. VII is devoted to the calculation of one-loop radiative corrections to the doubly charged Higgs boson mass in SUSYLR models with triplet Higgs.
This particle is predicted to be light, below a TeV, regardless of the scale of SUSY breaking. Sec. VIII have our
conclusions and some discussions. We collect several results relevant for the symmetry breaking analyses in the Appendix.
\section {Variety of Supersymmetric Left--Right Models} \label{sec2}
In this section we introduce various realizations of the SUSY left--right model. All these realizations have an extended
gauge symmetry which is $SU(3)_c \times SU(2)_L \times SU(2)_R \times U(1)_{B-L}$. We develop and analyze different Higgs boson sectors that break this symmetry down to the SM symmetry and then to $U(1)_{\rm em}$.
The right-handed $SU(2)_R$ symmetry breaking can be achieved either by Higgs triplets or by Higgs doublets, while the electroweak symmetry may be broken by Higgs bidoublets or by $SU(2)_L$ doublets.
In each case we allow for the possibility of a gauge singlet scalar field as well, which would enable the
breaking of $SU(2)_R \times U(1)_{B-L}$ down to $U(1)_Y$ in the supersymmetric limit. If such a gauge singlet Higgs scalar is not present, which is also studied, then the scale of $SU(2)_R$ symmetry breaking should be comparable to the scale
of SUSY breaking. We also investigate the Higgs boson spectrum of an alternate left--right symmetric model motivated by
$E_6$ unification.
Each of these models has a common chiral fermion sector consisting of three families of quark and lepton superfields given as\footnote{The identification of normal quarks and leptons will be slightly different in the $E_6$ motivated SUSYLR model.}
\begin{eqnarray}
\!\!Q\!\!&=&\!\!\begin{pmatrix}
u\\ d \end{pmatrix} \sim \left (3,2, 1, \frac13 \right ),~~~~~~~~
Q^c\!=\!\begin{pmatrix}
d^c\\-u^c
\end{pmatrix} \sim \left ( 3^{\ast},1, 2, -\frac13
\right ),\nonumber \\
L&=&\begin{pmatrix}
\nu\\ e \end{pmatrix} \sim\left ( 1,2, 1, -1 \right ),~~~~~~~~
L^c=\begin{pmatrix}
e^c \\ -\nu^c \end{pmatrix} \sim \left ( 1,1, 2, 1 \right ),
\label{eq:matter}
\end{eqnarray}
where the quantum numbers under
$SU(3)_c \times SU(2)_L \times SU(2)_R \times U(1)_{B-L}$ gauge group are listed.
We require that each of the models studied must meet four basic criteria. First, a mechanism for $SU(2)_R$ breaking consistent with the experimental limits on the $W_R^\pm$ and $Z_R$ gauge boson masses must be present. Second, the model must be able to generate realistic quark and lepton masses. Third, there must be a mechanism to generate small neutrino masses.
This could be Type I or Type II seesaw, inverse seesaw or a universal seesaw mechanism.
Fourth, there should be an unbroken $R$-parity that provides a dark matter candidate. In models with Higgs triplets, this
last requirement turns out to be automatic, but with Higgs doublets fields, an additional $Z_2$ symmetry will be assumed
that distinguishes lepton doublets from the Higgs doublets, as is usually done within the MSSM framework.
We now describe briefly each of these models, with a more extended discussion on the Higgs boson spectrum delegated to subsequent sections.
\subsection{Models with Higgs triplets and bidoublets}\label{2.1}
This scenario satisfies our requirements for a consistent model in the most straightforward way. An $SU(2)_R$ triplet Higgs field $\Delta^c(1,1,3,-2)$ is introduced, which breaks the gauge symmetry once its neutral component acquires a nonzero vacuum expectation value (VEV). It also couples to the right-handed neutrinos and generates Majorana masses for them.
Two bidoublet fields $\Phi_a(1,2,2,0)$ are introduced which can have Yukawa couplings with the quarks and leptons, generating their (Dirac) masses and CKM mixing angles. (With only one bidoublet, the CKM mixing angles would
all vanish. We shall also study the simpler case of having only one bidoublet in the theory, which would ascribe the
CKM mixings to soft SUSY breaking \cite{bdm1}.)
Being a supersymmetric theory the right-handed $\Delta^c(1,1,3,-2)$ field must be accompanied by another $SU(2)_R$ triplet field $\overline \Delta^c(1,1,3,+2)$ for anomaly cancellation and for achieving symmetry breaking consistently.
In a left-right symmetric model, the right-handed triplets must also be accompanied by left-handed triplet partners $\Delta(1,3,1,2)$ and $\overline \Delta(1,3,1,-2)$ for parity conservation.
We allow for cases with and without an extra gauge singlet scalar field $S(1,1,1,0)$. In the absence of the field $S$, it is not possible to break the $SU(2)_R$ symmetry in the supersymmetric limit. This is a viable possibility, as we shall see.
The presence of the singlet field $S$ would enable decoupling the two symmetry breaking scales, allowing the $SU(2)_R$ symmetry to be broken at a much higher scale than SUSY breaking. Thus, the Higgs boson fields in this model are given as
\begin{eqnarray}
\Delta(1,3,1,2)&=&\begin{pmatrix}
\frac{\delta^{+}}{\sqrt{2}} & \delta^{++}\\ \delta^{0} & -\frac{\delta^{+}}{\sqrt{2}} \end{pmatrix},~~~~
\overline{\Delta}(1,3,1,-2)=\begin{pmatrix}
\frac{\overline{\delta}^{-}}{\sqrt{2}} & \overline{\delta}^{0}\\ \overline{\delta}^{--} & -\frac{\overline{\delta}^{-}}{\sqrt{2}} \end{pmatrix},\nonumber \\
\Delta^{c}(1,1,3,-2)&=&\begin{pmatrix}
\frac{\delta^{c^{-}}}{\sqrt{2}} & \delta^{c^{0}}\\ \delta^{c^{--}} & -\frac{\delta^{c^{-}}}{\sqrt{2}} \end{pmatrix},~~~~
\overline{\Delta}^{c}(1,1,3,2)=\begin{pmatrix}
\frac{\overline{\delta}^{c^{+}}}{\sqrt{2}} & \overline{\delta}^{c^{++}}\\ \overline{\delta}^{c^{0}} & -\frac{\overline{\delta}^{c^{+}}}{\sqrt{2}} \end{pmatrix},\nonumber \\
\Phi_i(1,2,2,0)&=&{\begin{pmatrix}
\phi^{+}_1 & \phi^{0}_{2} \\ \phi^{0}_{1} & \phi^{-}_{2} \end{pmatrix}_i}~~(i=1,2),~~~~ S(1,1,1,0).
\label{eq:triphig}
\end{eqnarray}
The nonzero VEVs of various component fields are denoted as
\begin{equation}
\left< \delta^{c^0} \right> = v_R,~~~~ \left< \overline \delta^{c^0} \right> = \overline v_R,~~~~ \left< \phi_{1_i}^0 \right> = v_{u_i},~~~~ \left< \phi^0_{2_i} \right> = v_{d_i}~.
\label{vev}
\end{equation}
None of the other fields acquire vacuum expectation value. In particular, $\langle \delta^0 \rangle = 0$ in this model,
so that there is no type II seesaw contribution to the neutrino masses. We shall take the limit where $v_R, \overline v_R >> v_u, v_d$.
The Yukawa couplings in the model are given by the superpotential
\begin{eqnarray}
W_Y &=& \sum_{j=1}^2 \left( Y_q^{(j)} Q^T \tau_2 \Phi_j \tau_2 Q^c + Y_l^{(j)} L^T \tau_2 \Phi_j \tau_2 L^c \right) + i \frac{f}{2} L^T \tau_2 \Delta L + i \frac{f^c}{2} {L^c}^T \tau_2 \Delta^c L^c ,
\end{eqnarray}
where $Y_q^j$ and $Y_l^j$ are the quark and lepton Yukawa coupling matrices and $f$ is the Majorana Yukawa coupling matrix which generates large Majorana masses for right-handed neutrinos. This superpotential is invariant under parity transformation under which $\Phi \rightarrow \Phi^{\dagger}, \Delta \rightarrow \Delta^{c^*}, \overline \Delta \rightarrow \overline \Delta^{c^*}, S \rightarrow S^*,Q \rightarrow Q^{c^*}, L \rightarrow L^{c^*}, \theta \rightarrow \overline{\theta}$, along with
$W_L^\pm \rightarrow W_R^{\pm *}$. Parity invariance requires the Yukawa coupling matrices $Y_q^j$ and $Y_l^j$ to be hermitian and $f^c=f$. Once the various fileds acquire VEVs as shown in Eq. (\ref{vev}), the following mass matrices for fermions will
be induced:
\begin{eqnarray}
M_u &=& Y_q^{(1)} v_{u_1} + Y_q^{(2)} v_{u_2}, ~~~ M_d = Y_q^{(1)} v_{d_1} + Y_q^{(2)} v_{d_2},\nonumber \\
M_\nu^D &=& Y_l^{(1)} v_{u_1} + Y_l^{(2)} v_{u_2},~~~ M_\ell = Y_l^{(1)} v_{d_1} + Y_l^{(2)} v_{d_2},\nonumber \\
M_R &=& f v_R
\end{eqnarray}
for the up quarks, down quarks, neutrino Dirac, charged lepton and right-handed Majorana neutrino sectors.
Note that with only a single bidoublet scalar, the up and down quark mass matrices would become proportional, resulting
in vanishing CKM angles.
Left--right symmetric models predict the existence of new gauge bosons, one charged $W_R^\pm$ and one neutral $Z_R$.
In the limit where the right-handed symmetry breaking VEVs are much bigger than the electroweak symmetry breaking VEVs, we can neglect the mixing between the left-handed and the right-handed gauge bosons and obtain relations for the masses of the heavier $W_R^\pm$ and $Z_R$ bosons. They are given by
\begin{equation}
M^2_{W_R^\pm} \simeq \frac{1}{2} g_R^2 (2 v_R^2+2\overline v_R^2+v_{u_i}^2+v_{d_i}^2),
\end{equation}
and
\begin{equation}
M^2_{Z_R} \simeq \frac{g_R^2}{2 \cos^2 \theta_W \cos 2\theta_W} \left[ 4 (v_R^2+\overline v_R^2)\cos^4 \theta_W+(v_{u_i}^2+v_{d_i}^2)\cos^2 2\theta_W\right],
\end{equation}
where $g_R$ is the $SU(2)_R$ gauge coupling, $\theta_W$ is the weak mixing angle, and
the index $i$ is summed over the number of bidoublets in the model.
Since the gauge boson masses must be consistent with the experimental limits, these expressions will be relevant in setting lower limits on the right-handed symmetry breaking scale $v_R,\,\overline v_R$.
\subsection{Inverse seesaw model}
$SU(2)_R$ gauge symmetry can be broken by Higgs doublet fields, instead of Higgs triplets of the previous subsection.
This would simplify the analysis of the Higgs boson sector considerably. Two bidoublet fields are assumed to be present in the Higgs spectrum as before, which can generate quark and charged lepton masses and CKM mixings. Unlike the Higgs triplet field of the previous subsection, the Higgs doublet fields do not directly couple to fermions, and so
right-handed neutrinos would not receive heavy Majorana masses needed for the seesaw mechanism.
This situation is remedied by slightly complicating the fermion sector with the introduction of gauge singlet neutral fermions $N_i$ ($i=1-3$), one for each generation, along with the chiral matter fields that are given in Eq.~(\ref{eq:matter}).
The $N$'s can have
gauge invariant Majorana masses which in turn will generate small masses for the ordinary neutrinos via an inverse seesaw mechanism \cite{inverse}. The Higgs sector of this model consists of the following fields:
\begin{eqnarray}
H_L(1,2,1,-1) &=& \begin{pmatrix}
H_L^0 \\ H_L^{-} \end{pmatrix},
\overline{H}_L(1,2,1,1) = \begin{pmatrix}
\overline{H}_L^+ \\ \overline{H}_L^0 \end{pmatrix},
H_R(1,1,2,1) = \begin{pmatrix}
H_R^+ \\ H_R^{0} \end{pmatrix},\nonumber \\
\overline{H}_R(1,1,2,-1) &=&\begin{pmatrix}
\overline H_R^0 \\ \overline H_R^{-} \end{pmatrix},
\Phi_a(1,2,2,0)= \begin{pmatrix}
\phi^{+}_1 & \phi^{0}_{2} \\ \phi^{0}_{1} & \phi^{-}_{2} \end{pmatrix}_a (a=1,2).
\label{eq:2dh}
\end{eqnarray}
Here the $SU(2)_R$ doublet $H_R(1,1,2,1)$ is accompanied by $\overline{H}_R(1,1,2,-1)$ for anomaly cancelation,
while $H_L(1,2,1,1)+\overline{H}_L(1,2,1,-1)$ are their parity partners.
The VEVs of the neutral components of these fields are parametrized as
\begin{equation}
\left< H_L^0 \right> = v_L,~~~ \left< \overline H_L^0 \right> = \overline v_L,~~~\left< H_R^0 \right> = v_R,~~~ \left< \overline H_R^0 \right> = \overline v_R,~~~ \left< \phi_{1_i}^0 \right> = v_{u_i},~~~ \left< \phi^0_{2_i} \right> = v_{d_i}.
\label{vev2}
\end{equation}
As noted earlier, in this model, a $Z_2$ symmetry is assumed that distinguishes the $H_L$ field and the lepton doublet
$L$ (and similarly $H_R$ and $L^c$ fields). Under this $Z_2$, the lepton fields are odd, while the Higgs doublet fields
are all even.
The superpotential relevant for quark and lepton mass generation in this case is given as
\begin{eqnarray}
W_Y = \sum_{j=1}^2 \left( Y_q^{(j)} Q^T \tau_2 \Phi_j \tau_2 Q^c + Y_l^{(j)} L^T \tau_2 \Phi_j \tau_2 L^c \right) + i f L^T \tau_2 \overline H_L N + i f^c {L^c}^T \tau_2 \overline H_R N +\frac 12 \mu_N N N
\end{eqnarray}
where $Y_q^{(j)}$ and $Y_l^{(j)}$ are the quark and lepton Yukawa coupling matrices, $f$ and $f^c$ are the
analogs of Dirac Yukawa couplings matrices with the singlet neutrino and $\mu_N$ is the lepton number violating Majorana mass term for $N$.
Under parity symmetry $N$ transforms as $N \rightarrow N^*$.
From parity invariance we see that $Y_q^{(j)}$ and $Y_l^{(j)}$ are hermitian, $f^c=f$ and $\mu_N$ is real. The $3 N_g \times 3 N_g$ neutrino mass matrix is given as (with $N_g$ being the number of generations)
\begin{equation}
M_\nu = \begin{pmatrix} 0& Y_l^{(i)} v_{1_i}&f \overline v_L \\ Y_l^{(i) T} v_{1_i} & 0 & f \overline v_R \\ f^T \overline v_L & f^T \overline v_R & \mu_N \end{pmatrix}.
\end{equation}
This is the left--right symmetric realization of the inverse seesaw mechanism \cite{inverse,barr}.
In the limit of small $\overline v_L$ and $\mu_N$, the light neutrino masses would vanish. Thus, the smallness of
these two parameters would provide an understanding of small neutrino masses. Note that the condition $\overline v_L$ being
small for explaining small neutrino masses is something specific to the left--right symmetric realization of the
inverse seesaw mechanism.
The heavy gauge boson masses in this model are given as:
\begin{equation}
M^2_{W_R^\pm} \simeq \frac{1}{2} g_R^2 (v_R^2+\overline v_R^2+v_{u_i}^2+v_{d_i}^2),
\label{eq:WR}
\end{equation}
and
\begin{equation}
M^2_{Z_R} \simeq \frac{g_R^2}{2 \cos^2 \theta_W \cos 2\theta_W} \left[(v_R^2+\overline v_R^2)\cos^4 \theta_W+(v_{u_i}^2+v_{d_i}^2)\cos^2 2\theta_W + (v_L^2+\overline v_L^2)\sin^4 \theta_W \right],
\label{eq:ZR}
\end{equation}
where $g_R$ is the $SU(2)_R$ gauge coupling, $\theta_W$ is the weak mixing angle, the index $i$ is summed over the number of bidoublets in the model and the VEVs are given
in Eq.~(\ref{vev2}). Comparing these expressions for the masses with the experimental limit on the heavy gauge bosons we will be able to set lower limits on $v_R$ and $\overline v_R$.
\subsection{Universal seesaw model}
One can choose an even simpler Higgs boson sector compared to the Higgs sector of the inverse seesaw model
in order to achieve the desired symmetry breaking. A doublet field $H_R(1,1,2,1)$ is sufficient to break the $SU(2)_R$ symmetry and analogously a doublet field $H_L(1,2,1,-1)$ suffices for $SU(2)_L$ symmetry breaking. Anomaly cancelation, needed in
Higgs sectors of supersymmetric models, would require the addition of a $\overline H_R(1,1,2,-1)$ and $\overline H_L(1,2,1,1)$ fields. This Higgs boson sector, without any bidoublet fields, will not be able to generate quark and lepton masses.
Additional vectorlike quarks and leptons which are $SU(2)_L \times SU(2)_R$ singlets are introduced for this purpose \cite{universal,bm2}. This scenario is termed ``universal seesaw", since all quarks and leptons
would acquire masses via a generalized seesaw involving mixing with the heavy fermions.
The chiral matter sector in this case would consist of the quarks and leptons listed in Eq.~(\ref{eq:matter}) along with a set of heavy singlet quarks and lepton fields, one per generation, denoted as $P(3,1,1,\frac{4}{3})$, $N(3,1,1,-\frac{2}{3})$ and $E(1,1,1,-2)$ and their conjugates $P^c(3,1,1,-\frac{4}{3}), N^c(3,1,1,\frac{2}{3})$ and $E^c(1,1,1,2)$.
There can also be a neutral singlet lepton denoted as $R(1,1,1,0)$ for generating small neutrino masses via the seesaw mechanism, however, this is not essential as tiny Dirac masses for the neutrinos can be generated at the two-loop level arising through the mixed exchange of from $W_L$ and $W_R$ gauge bosons without breaking lepton number \cite{bh}. The Higgs sector of the universal seesaw model is given by
\begin{eqnarray}
H_L(1,2,1,-1) &=& \begin{pmatrix}
{H_L}^0 \\ {H_L}^{-} \end{pmatrix},
\overline{H}_L(1,2,1,1) = \begin{pmatrix}
{\overline{H}_L}^+ \\ {\overline{H}_L}^{0} \end{pmatrix},\nonumber \\
H_R(1,1,2,1) &=& \begin{pmatrix}
{H_R}^+ \\ {H_R}^{0} \end{pmatrix},
\overline{H}_R(1,1,2,-1) =\begin{pmatrix}
{\overline H_R}^0 \\ {\overline H_R}^{-} \end{pmatrix}.
\label{eq:2ch}
\end{eqnarray}
The absence of bidoublet fields prevents any direct Yukawa coupling between the left-handed and the right-handed fermion fields
of Eq. (\ref{eq:matter}). The Higgs doublet fields can couple quarks with the vectorlike quarks and leptons with the vectorlike
leptons. The Yukawa superpotential of the model is given by
\begin{eqnarray}
W_Y &=& y_u Q \overline H_L P^c - y_d Q H_L N^c - y_l L H_L E^c + y_\nu L \overline H_L R \notag \\
&+& y_u^c Q^c \overline H_R P - y_d^c Q^c H_R N - y_l^c L^c H_R E + y_\nu^c L^c \overline H_R R\notag \\
&+& m_u P P^c + m_d N N^c + m_l E E^c + \frac12 m_R R R
\end{eqnarray}
where $y_i$ and $y^c_i$ stand for $3\times3$ Yukawa coupling matrices and $m_i$ are the heavy singlet fermions Majorana mass matrices. Under parity symmetry, $P \rightarrow P^{c*},\, N \rightarrow N^{c*}$ and
$E \rightarrow E^{c*}$. Parity invariance then requires $y_i^c = y_i^*$ and $m_i$ to be real. The fermion mass matrix
for the up--qurk, down--quark and charged lepton sectors are given by
\begin{eqnarray}
M_u = \left(\begin{matrix} 0 & y_u \overline v_L \cr y_u^\dagger \overline v_R & m_u \end{matrix}\right);~~
M_d = \left(\begin{matrix} 0 & y_d v_L \cr y_d^\dagger v_R & m_d \end{matrix}\right);~~
M_l = \left(\begin{matrix} 0 & y_l v_L \cr y_l^\dagger v_R & m_l \end{matrix}\right)~.
\end{eqnarray}
Here $m_u$ is multiplied by $(u,\,P)$ from the left and $(u^c,\,P^c)$ from the right, and so on.
The vacuum expectation values are defined as $\langle H_L^0 \rangle = v_L,\,\langle \overline H_L^0 \rangle =
\overline v_L,\, \langle H_R^0 \rangle = v_R,\,\langle \overline H_R^0 \rangle =
\overline v_R$. The light fermion mass can be obtained as $m_{\rm down} \sim |y_d|^2 v_L v_R/m_d$, etc.
Note that the determinant of the quark mass matrices are real, by virtue of
parity symmetry, provided that the VEVs are real. CP violation will occur via CKM mixings, since the Yukawa couplings themselves are not real.
This feature of the model has been used to explain the strong CP problem without the use of axions \cite{bm0}.
The heavy gauge boson masses can be obtained from Eq.~(\ref{eq:WR}) and Eq.~(\ref{eq:ZR}) by setting $v_{u_i}= v_{d_i} =0$.
We investigate two variations of this model with and without a singlet Higgs boson. The upper limit on the mass of the lightest CP-even Higgs boson will be quite different in the two cases as we show later.
\subsection{$E_6$ motivated left-right supersymmetric model}
This model is motivated by the low energy manifestation of heterotic superstring theory where the matter supermultiplets belong to the {\bf{27}} representation of $E_6$ group. The particle content of this representation under the subgroup $SU(3)_c\times SU(2)_L\times SU(2)_R\times U(1)$ is given as:
\begin{align}
&Q = (u,d):(3,2,1,\frac{1}{3}),~~ d^c:(\overline{3},1,1,\frac{2}{3}),~~Q^c = (h^c,u^c):(\overline{3},1,2,-\frac{1}{3}), \nonumber \\
&L^c = (e^c,n):(1,1,2,1),~~F = \begin{pmatrix}\nu_e&E^c\\epsilon&N^c_E \end{pmatrix} :(1,2,2,0),~~ h:(3,1,1-\frac{2}{3}),\notag \\
&\psi = (\nu_E,E):(1,2,1,-1),~~N:(1,1,1,0).
\end{align}
We can define an $R$-parity in this case under which the $\{u,d,\nu_e,e\}$ fields are even while the $\{h,E,\nu_E,N_E^c,n\}$ fields are odd. The $W_R^\pm$ gauge boson is also odd under this $R$-parity. The superpartners of these fields have opposite R-parity. The fermions and the gauge bosons have odd and even $R$-parity respectively, except for the $W_R^\pm$ gauge boson which is odd as it links particles of opposite $R$-parity.
The Higgs fields of this model can be identified as:
\begin{eqnarray}
H_L(1,2,1,-1) &=&\begin{pmatrix}
{H_L}^0 \\ {H_L}^{-} \end{pmatrix}=\begin{pmatrix}\tilde{\nu}_E\\ \tilde{E}\end{pmatrix},~~
H_R(1,1,2,1) = \begin{pmatrix}
{H_R}^+ \\ {H_R}^{0} \end{pmatrix}=\begin{pmatrix}\tilde{e}^c\\ \tilde{n} \end{pmatrix}, \notag \\
\Phi(1,2,2,0)&=&\begin{pmatrix}
\phi^{+}_1 & \phi^{0}_{2} \\ \phi^{0}_{1} & \phi^{-}_{2} \end{pmatrix}=\begin{pmatrix}\tilde{E}^c&\tilde{N}^c_E\\\tilde{\nu}_e&\tilde{e} \end{pmatrix}.~~~~
\label{eq:2ah}
\end{eqnarray}
The Yukawa interaction terms in the superpotential are given as
\begin{equation}
W_{Y}= \lambda_1 Q d^c \psi+\lambda_2 Q Q^c F+\lambda_3h Q^c L^c+\lambda_4 F L^c \psi+\lambda_5 F N F+\lambda_6h d^c N.
\end{equation}
This generates masses for quarks and leptons as well as CKM mixings. A small neutrino mass can be generated by the mixing of the $n, \nu_E$ and the $N_E^c$ fields with $\nu_e$ and $N$ fields. The neutrino mass matrix in the basis $(\nu_e, N,\nu_E,N^c_E,n)$ takes the form
\begin{equation}
m_{\nu} = \begin{pmatrix}
0 & \lambda_5 \left< \widetilde{N}^c_E \right> & 0 & 0 & \lambda_5 \left< \widetilde{N}\right>\\
\lambda_5 \left< \widetilde{N}^c_E\right> & 0 & 0 & 0 & \lambda_5 \left< \widetilde{\nu}_e \right> \\
0 & 0 & 0&\lambda_4 \left< \widetilde{n} \right>&\lambda_4 \left< \widetilde{N}^c_E \right>\\
0 & 0 & \lambda_4 \left< \widetilde{n} \right>&0&\lambda_4 \left< \widetilde{\nu}_E \right>\\
\lambda_5 \left< \widetilde{N}\right> & \lambda_5 \left< \widetilde{\nu}_e \right> & \lambda_4 \left< \widetilde{N}^c_E \right>&\lambda_4 \left< \widetilde{\nu}_E \right>&0
\end{pmatrix}~.
\end{equation}
In the limit of vanishing $\lambda_5$, $\nu_e$ form an almost Dirac neutrino with $N$. The VEV $\langle N \rangle$ can be small, since only the $\lambda_5$ coupling can induce its VEV, which should be small for neutrino masses. In our symmetry
breaking analysis we shall keep $\langle N \rangle=0$.
The heavy gauge boson masses can be obtained from Eq.~(\ref{eq:WR}) and Eq.~(\ref{eq:ZR}) by setting $\overline v_R$ and $\overline v_L$ to be zero.
\section {Symmetry Breaking with Higgs Triplet fields} \label{case1}
In this section we analyze the Higgs sectors of a class of models which use Higgs triplets for $Su(2)_R$ symmetry breaking.
Here we construct the relevant superpotential for symmetry breaking, compute the Higgs potential and from it the Higgs boson
spectrum. The mixing of Higgsinos with the gauginos is also analyzed. We concentrate on the lightest neutral CP even Higgs boson mass $M_h$ and study how it gets modified and its effect on the parameter space of each model.
Four models are studied under this class. The first case has a pair of Higgs triplets, one bidoublet and a gauge singlet
in the spectrum. In the second case, we integrate out the singlet field of case one, but keep its effective non-renormalizable interactions. The third case has no singlet field at all. Case four has two triplets and two bidoublets, which is fully realistic for fermion mass generation, including CKM mixing angles. The cases with only one bidoublet field should be thought
of as being special cases of this case, with or without an additional singlet.
\subsection{Case with a pair of triplets, a bidoublet and a gauge singlet} \label{case1d}
We first analyze the case with the triplet Higgs fields $\Delta,\overline \Delta,\Delta^c, \overline \Delta^c$, one bidoublet Higgs field $\Phi$ and a singlet Higgs field $S$. The quantum numbers and compositions of these fields are shown in Eq. (\ref{eq:triphig}).
For a fully realistic model we need two bidoublet fields to generate the quark mixings, but for simplicity we will only use a single bidoublet in our calculations in this section. This does not significantly affect the Higgs boson masses as will be shown in a later section. The most general superpotential terms involving only the Higgs boson fields in this case is given as:
\begin{eqnarray}
W&=& S\left[{\text{Tr}}(\lambda \Delta \overline{\Delta})+{\text{Tr}}(\lambda ^c \Delta^{c}\overline{\Delta}^{c}) +\dfrac{\lambda^{\prime}}{2} {\text{Tr}}(\Phi^{T}\tau_{2}\Phi\tau_{2})-M^2 \right] \notag \\
&+& \text{Tr} \left[ \mu_1 \Delta \overline \Delta+\mu_2 \Delta^c \overline \Delta^c+\frac{\mu}{2} \left( \Phi^T \tau_2 \Phi \tau_2 \right) \right] + \frac{\mu_S}{2} S^2+\frac{\kappa}{3} S^3,
\end{eqnarray}
where $\lambda^c = \lambda^*$, $\mu_1 = \mu_2^*$ and $\lambda^{\prime}, M^2,\mu$ and $\mu_S$ are real from parity invariance.
The Higgs potential consists of the $F$-terms, $D$-terms and soft supersymmetry-breaking terms,
\begin{equation}
V_{Higgs} = V_F+V_D+V_{Soft}.
\end{equation}
In this case, the relevant terms in the Higgs potential are given by:
\begin{eqnarray}
\label{eq:1df}
V_F&=& {\text{Tr}} \left| (\lambda \Delta \overline{\Delta})+(\lambda^* \Delta^{c}\overline{\Delta}^{c}) +\frac{\lambda^\prime}{2}(\Phi^{T}\tau_{2}\Phi\tau_{2})-M^2+\mu_S S+\kappa S^2 \right|^2+\text{Tr} \left| \mu \Phi + \lambda ^ \prime S \Phi \right|^2 \nonumber \\
&+& \text{Tr} \left[ \left| \mu_1 \Delta + \lambda S \Delta \right| ^2 + \left| \mu_1 \overline{\Delta} + \lambda S \overline{\Delta} \right| ^2 + \left| \mu_1^* \Delta^c + \lambda^* S \Delta^c \right| ^2 \right. \notag \\
&+& \left. \left| \mu_1^* \overline \Delta^c + \lambda^* S \overline \Delta^c \right| ^2 \right] , \\
\label{eq:two}
V_D&=&\frac{g_L^2}{8}\sum \limits_{a=1}^3 \left|{\text{Tr}}(2\Delta^\dagger \tau_a \Delta+2\overline\Delta^\dagger \tau_a \overline\Delta+\Phi^\dagger \tau_a \Phi)\right|^2\nonumber \\
&+&\frac{g_R^2}{8}\sum \limits_{a=1}^3 \left|{\text{Tr}}(2{\Delta^c}^\dagger \tau_a \Delta^c+2\overline{\Delta^c}^\dagger \tau_a \overline\Delta^c+\Phi^* \tau_a \Phi^T)\right|^2 \nonumber \\
&+& \frac{g_V^2}{2}\left|{\text{Tr}}(\Delta^\dagger \Delta- \overline\Delta^\dagger \overline\Delta-{\Delta^c}^\dagger \Delta^c+ \overline{\Delta^c}^\dagger \overline\Delta^c)\right|^2, \\
V_{Soft}&=&m_1^2{\text{Tr}}({\Delta^c}^\dagger\Delta^c)+m_2^2{\text{Tr}}({\overline{\Delta}^c}^\dagger\overline\Delta^c)+m_3^2{\text{Tr}}(\Delta^\dagger\Delta)+m_4^2{\text{Tr}}(\overline\Delta^\dagger\overline\Delta)\nonumber \\
&+&m_S^2 |S|^2+ m_5^2 {\text{Tr}}(\Phi^\dagger \Phi)
+\left[\lambda A_{\lambda} S {\text{Tr}}(\Delta \overline\Delta +\Delta^c \overline{\Delta}^c)+h.c.\right]\nonumber \\&+&[\lambda^\prime A_{\lambda^\prime} S {\text{Tr}}(\Phi^{T}\tau_{2}\Phi\tau_{2})+h.c.]+(\lambda C_{\lambda} M^2 S +h.c.) + \left( \mu_S B_S S^2 + h.c. \right) \notag \\
&+& \left[ \mu_1 B_1 \text{Tr} \left( \Delta \overline \Delta \right)+\mu_2 B_2 \text{Tr} \left( \Delta^c \overline \Delta^c \right)+\mu B \text{Tr} \left( \Phi^T \tau_2 \Phi \tau_2 \right) +\kappa A_\kappa S^3+ h.c. \right].
\label{eq:pot1d}
\end{eqnarray}
The soft mass terms $m_1^2$ and $m_3^2$ (and similarly $m_2^2$ and $m_4^2$) should be equal with exact parity symmetry, but
we shall allow for soft breaking of parity in these dimension two terms. With exact parity, consistent symmetry breaking cannot be achieved in this model, as there would be unwanted massless modes if parity is forced to be broken spontaneously. We use this potential to calculate the Higgs boson mass-squared matrices for the charged, neutral CP-even and neutral CP-odd Higgs bosons. For simplicity we will assume all the parameters in the potential to be real. The vacuum structure that we choose is given by:
\begin{equation} \label{eq:vev1a}
\left<\Delta^c\right> = \begin{pmatrix}
0 & v_R \\ 0 & 0 \end{pmatrix},~~
\left<\overline\Delta^c\right> = \begin{pmatrix}
0&0\\\overline{v}_R e^{i \phi_R} & 0 \end{pmatrix},~~
\left<\Phi\right> ={\begin{pmatrix}
0 & v_2 \\ v_1 e^{i \phi_1} & 0 \end{pmatrix}}, \left< S \right> = v_S e^{i \phi_S}.
\end{equation}
while the neutral components of $\Delta$ and $\overline{\Delta}$ fields do not get any vacuum expectation value. In the absence of the singlet field $S$, parity conservation dictated that all the soft supersymmetry breaking parameters must be real. The presence of the singlet does result in a nonzero phase but for simplicity we assume this to be zero. If all parameters in the Higgs potential are real, there should be a minimum that preserves CP invariance. We focus on this minimum.
So for the present case $\phi_R=0$, $\phi_1=0$ and $\phi_S = 0$. This choice of phases negates the mixing between the scalar and the pseudo-scalar Higgs bosons but does not significantly affect the mass of the lightest CP-even Higgs boson. The values of $v_R$ and $\overline v_R$ are of the order of the right-handed $W_R^\pm$ mass, while $v_1$ and $v_2$ are of electroweak scale and hence $v_R,\overline v_R >> v_1, v_2$.
We first look at the CP-even Higgs boson which is the main focus of this section. To easily identify the field corresponding to the lightest eigenvalue, we take a linear combination of the Higgs fields so that only two of the newly defined fields get non-zero vacuum expectation values -- one at the high right-handed symmetry breaking scale and the other at the lower electroweak symmetry breaking scale. The field redefinition that we use is given as:
\begin{align} \label{eq:basis1d}
\rho_1&=\frac{v_1 \phi_1^0 +v_2 \phi_2^0}{\sqrt{v_1^2+v_2^2}},~ \rho_2=\frac{v_2\phi_1^0-v_1 \phi_2^0}{\sqrt{v_1^2+v_2^2}},~ \rho_3=\frac{v_R {\delta^c}^0 +\overline v_R {\overline \delta^c}^0}{\sqrt{v_R^2+\overline v_R^2}},~ \rho_4=\frac{\overline v_R{\delta^c}^0-v_R {\overline \delta^c}^0}{\sqrt{v_R^2+\overline v_R^2}}.
\end{align}
In this rotated basis we calculate the mass matrix subject to the following minimization conditions:
\begin{align}
0= & v_1 [4 m_5^2 + g_L^2 (v_1^2 - v_2^2) +
g_R^2 (v_1^2- v_2^2 + 2 v_R^2 - 2 \overline v_R^2)]
- 8 \lambda^\prime A_{\lambda^\prime} v_2 v_S - 8 \mu B v_2 \notag \\
&+4 {\lambda^\prime} v_2 (M^2-\lambda v_R \overline v_R+ \lambda^\prime v_1 v_2 - \mu_S v_S-\kappa v_S^2) + 4 v_1(\mu+ \lambda^\prime v_S)^2, \notag \\
0= & v_2 [4 m_5^2 + g_L^2 (v_2^2 - v_1^2) +
g_R^2 (v_2^2 - v_1^2 - 2 v_R^2 + 2 \overline v_R^2)]
- 8 \lambda^\prime A_{\lambda^\prime} v_1 v_S - 8 \mu B v_1 \notag \\
&+4 {\lambda^\prime} v_1 (M^2-\lambda v_R \overline v_R+ \lambda^\prime v_1 v_2 - \mu_S v_S-\kappa v_S^2) + 4 v_2(\mu+ \lambda^\prime v_S)^2, \notag \\
0=& 2 m_1^2 v_R + g_R^2 v_R ( v_1^2 - v_2^2 + 2 v_R^2 - 2 \overline v_R^2) +
2 [g_V^2 v_R (v_R^2 -\overline v_R^2) + \lambda A_\lambda \overline v_R v_S +\mu_1 B_2 \overline v_R \notag \\
& + v_R \left(\lambda v_S + \mu_1 \right) ^2 +\lambda \overline v_R (- M^2 +\lambda v_R \overline v_R-\lambda^ \prime v_1 v_2 + \mu_S v_S+\kappa v_S^2 ), \nonumber \\
0=& 2 m_2^2 \overline v_R + g_R^2 \overline v_R ( -v_1^2 + v_2^2 - 2 v_R^2 + 2 \overline v_R^2) +
2 [g_V^2 \overline v_R (-v_R^2 +\overline v_R^2) + \lambda A_\lambda v_R v_S +\mu_1 B_2 v_R \notag \\
& + \overline v_R \left(\lambda v_S + \mu_1 \right) ^2 +\lambda v_R (- M^2 +\lambda v_R \overline v_R-\lambda^ \prime v_1 v_2 + \mu_S v_S+\kappa v_S^2 ) , \nonumber \\
0=& 2 \left[ m_S^2 v_S + C_\lambda M^2 \lambda -
2 \lambda^\prime A_{\lambda^\prime} v_1 v_2 + \lambda
A_\lambda v_R \overline v_R + {\lambda^\prime}^2 (v_1^2 + v_2^2) + \lambda^2 (v_R^2 +
\overline v_R^2)] v_S + \mu_S B_S v_S \right. \notag \\
& \mu \lambda^\prime (v_1^2+v_2^2)+ \lambda \mu_1 (v_R^2+\overline v_R^2) + 3 \kappa A_\kappa v_S^2 (\mu_S+2 \kappa v_S) \left( -M^2 + \lambda v_R \overline v_R -\lambda^\prime v_1 v_2 + \mu_S v_S + \kappa v_S^2 \right) {\Big]}.
\label{eq:min1d}
\end{align}
We first look at the CP even scalar Higgs boson masses.
We get a $5\times 5$ mass-squared matrix in the basis $({\text{Re}}\rho_1,{\text{Re}}\rho_2,{\text{Re}}\rho_3,{\text{Re}}\rho_4,{\text{Re}}{S})$ where one of the eigenvalues would remain light. The relevant terms in this $5\times 5$ mass-squared matrix are given as:
\begin{eqnarray} \label{eq:masmat1d}
M_{11}& =& \frac{g_L^2 (v_1^2 - v_2^2)^2 + g_R^2 (v_1^2 - v_2^2)^2 +
8 v_1^2 v_2^2 {\lambda^\prime}^2}{2 (v_1^2 + v_2^2)},\notag \\
M_{12}& =& \frac{ v_1 v_2 (v_1^2 - v_2^2) (g_L^2 + g_R^2 - 2 {\lambda^\prime}^2)}{ (v_1^2 + v_2^2)},\notag \\
M_{13}& =& \frac{g_R^2 (v_1^2 - v_2^2) (v_R^2 -\overline v_R^2) -
4 \lambda \lambda^\prime v_1 v_2 v_R \overline v_R}{\sqrt{(v_1^2 + v_2^2)(v_R^2 +\overline v_R^2)}}, \notag \\
M_{14}& =& \frac{2[ g_R^2 (v_1^2 - v_2^2)v_R \overline v_R -
\lambda \lambda^\prime v_1 v_2 (v_R^2 -\overline v_R^2)]}{\sqrt{(v_1^2 + v_2^2)(v_R^2 +\overline v_R^2)}}, \notag \\
M_{15} &=& \frac{2 \lambda^\prime [-2 A_{\lambda^\prime} v_1 v_2 +
(v_1^2 + v_2^2) (v_S \lambda^\prime+\mu)-(\mu_S+2 \kappa v_S) v_1 v_2]}{\sqrt{v_1^2 + v_2^2}}, \notag \\
M_{22}&=&{\Big{[}}(2 g_L^2 + 2 g_R^2) v_1^2 v_2^2 + 2 m_5^2 (v_1^2 + v_2^2) +
{\lambda^{\prime}}^2(v_1^2 - v_2^2)^2 +2 {\lambda^\prime}^2 v_S^2(v_1^2 +v_2^2) \notag \\
&+& 4 \lambda^\prime \mu v_S (v_1^2 + v_2^2)+
2 \mu^2(v_1^2 + v_2^2){\Big{]}}/(v_1^2 + v_2^2), \notag \\
M_{23}&=& \frac{2 \left[ g_R^2 v_1 v_2 (v_R^2 -\overline v_R^2) + \lambda \lambda^\prime (v_1^2 -
v_2^2) v_R \overline v_R\right]}{\sqrt{(v_1^2 + v_2^2) ( v_R^2 +\overline v_R^2)}}, \notag \\
M_{24}&=& \frac{4 g_R^2 v_1 v_2 v_R \overline v_R - \lambda \lambda^\prime (v_1^2 -
v_2^2) (v_R^2- \overline v_R^2)}{\sqrt{(v_1^2 + v_2^2)(
v_R^2 +\overline v_R^2)}}, \notag \\
M_{25}&=&\frac{\lambda^\prime (v_1^2 - v_2^2) (2 A_{\lambda^\prime} + \mu_S+2 \kappa v_S)}{\sqrt{v_1^2 + v_2^2}},\notag \\
M_{33} &=&\frac{2\left[ \left( g_R^2+g_V^2\right)(v_R^2 - \overline v_R^2)^2+2 \lambda^2 v_R^2 \overline v_R^2 \right]}{v_R^2+\overline v_R^2}, \notag \\
M_{34}&=& \frac{2 v_R \overline v_R (v_R^2 - \overline v_R^2)^2 \left(2 g_R^2+2g_V^2+\lambda^2 \right)}{v_R^2+\overline v_R^2}, \notag \\
M_{35} &=& \frac{2 \lambda \left[ A_\lambda v_R \overline v_R+ (\lambda v_S+\mu_1)(v_R^2+\overline v_R^2)
+v_R \overline v_R(\mu_S+2 \kappa v_S) \right]}{\sqrt{v_R^2+\overline v_R^2}}, \notag \\
M_{44}&=&\left[8 (g_R^2+g_V^2) v_R^2 \overline v_R^2 + (m_1^2+m_2^2 ) (v_R^2 + \overline v_R^2) + \lambda^2(v_R^2-\overline v_R^2)^2 \right. \notag \\
&+& \left. 2( \lambda v_S+\mu_1 )^2 (v_R^2 + \overline v_R^2) \right]/(v_R^2 + \overline v_R^2),\notag \\
M_{45} &=& -\frac{\lambda(v_R^2 - \overline v_R^2) (A_\lambda + \mu_S+2 \kappa v_S)}{\sqrt{v_R^2 + \overline v_R^2}},\notag \\
M_{55} &=& m_S^2 + {\lambda^\prime}^2(v_1^2 + v_2^2) + \lambda^2 (v_R^2 +\overline v_R^2)+ \mu_S^2 +2 \mu_S B_S \notag \\
&+& 2 \kappa[-M^2+\lambda v_R \overline v_R-\lambda^{\prime} v_1 v_2+3(A_\kappa+\mu_S+\kappa v_s) v_S.
\end{eqnarray}
From our choice of basis, we can guess that the $M_{11}$ element of the mass-matrix, along with the corrections from the off-diagonal elements, would approximately be the lightest eigenvalue for this matrix. We calculate the corrections to lightest eigenvalue coming from the off-diagonal $M_{12},M_{13},M_{14}$ and $M_{15}$ elements. It can be seen that the $M_{12}$ element is proportional to the square of the lighter VEVs that break electroweak symmetry while the diagonal $M_{22}$ element comes out to be proportional to the square of the heavy VEV $v_R$. Hence the $M_{12}$ term gives a negligible correction to the lightest eigenvalue, of order $v^4/v_R^2$, which is negligible. Noting that the off-diagonal corrections to the lightest eigenvalue are negative definite, in order to derive an upper limit on the lightest eigenvalue we choose parameters $\lambda$, $A_{\lambda^\prime}$ and $A_{\lambda}$ such that they make $M_{13}$, $M_{15}$ and $M_{35}$ zero respectively.
There is no longer any freedom left to make $M_{14}$ also vanish. However, parametrically we can choose the soft mass parameter
$m_1$ to be much bigger than $v_R, \overline v_R$, in which case the off-diagonal corrections from $M_{14}$ would be suppressed.
Thus we see that he upper limit on the lightest eigenvalue is simply $M_{11}$.\footnote{If we choose $\mu_S$ to be much greater than all the other mass scales in the model, we get back the familiar result where the upper limit of the neutral Higgs mass is bound by $M_Z$.} This gives us the largest allowed value of $M_h^2$ to be
\begin{equation}
M_{h_{tree}}^2 = 2 M_W^2 \cos^2 2\beta+ \lambda^2 v^2 \sin^2 2 \beta,
\label{limit1}
\end{equation}
where $\tan\beta = \frac{v_1}{v_2}$, $v^2 = v_1^2+v_2^2$ and we have assumed that the $SU(2)_R$ gauge coupling ($g_R)$ is equal to the $SU(2)_L$ gauge coupling $g_L$. Choosing a larger or smaller value of $g_R$ will lead to a larger or smaller value of the tree level Higgs boson mass respectively. For example, for $g_R^2=2g_L^2$ we get $M_{h_{tree}}^2 = 3 M_W^2 \cos^2 2\beta+ \lambda^2 v^2 \sin^2 2 \beta,$ while for $g_R^2=\frac{g_L^2}{2}$ we get
$M_{h_{tree}}^2 = \frac{3}{2} M_W^2 \cos^2 2\beta+ \lambda^2 v^2 \sin^2 2 \beta.$ Note that the second term in Eq. (\ref{limit1}) has the same origin as in NMSSM due to the presence of a gauge singlet scalar in the model.
Including the leading radiative corrections from the top and stop sector, the upper limit on the lightest CP even Higgs boson mass is:
\begin{eqnarray}
M_h^2 &=& (2 M_W^2 \cos^2 2 \beta+ \lambda^2 \sin^2 2 \beta) \Delta_1 + \Delta_2
\end{eqnarray}
where
\begin{eqnarray}
\Delta_1 &=& \left(1-\frac{3}{8 \pi^2} \frac{m_t^2}{v^2}t\right), \notag \\
\Delta_2 &=& \frac{3}{4 \pi^2} \frac{m_t^4}{v^2}\left[ \frac{1}{2} \tilde{X}_t+t+\frac{1}{16 \pi^2} \left(\frac{3}{2} \frac{m_t^2}{v^2}-32 \pi \alpha_3 \right) \left( \tilde{X}_t t+t^2 \right) \right],
\label{eq:rad}
\end{eqnarray}
and $m_t$ is the running top quark mass, $v=\sqrt{v_1^2+v_2^2}\approx 174$ GeV, $\alpha_3$ is the running QCD coupling, $\tilde{X}_t$ is the left--right stop squark mixing parameter, and $t = {\text{log}} \frac{M_{S}^2}{M_t^2}$ with $M_t$ being the top quark pole mass and $M_S$ being the geometric mean of the two stop squark masses.
\begin{figure}[h!]\centering
\includegraphics[width=3.1in]{1d_tb.jpg}
\includegraphics[width=3.05in]{1d_Ms.jpg}
\caption{{\sl(a) Variation of Higgs boson mass with $\tan \beta$, (b) Higgs boson mass as a function of $M_S$ for the case
with Higgs triplets, one bidoublet and a singlet.
The red region represents the band where 124 GeV $< M_h <$ 126 GeV. The light green region corresponds to
$X_t=0$ while the blue upper region is for $X_t=6$. The red shaded region is for all values of Higgs mass greater than 126 GeV and it is overlapped by the blue and the green regions. The black solid line represents the MSSM upper limit for the Higgs mass.
}}
\label{fig:1d}
\end{figure}
The upper limit of the Higgs boson mass in this case is plotted in Fig.~\ref{fig:1d}(a) as a function of $\tan \beta$. The red region in the figure represents the band where the mass is between 124 GeV and 126 GeV. Anything below this has not been included as that will be ruled out by experiments. Any point above this can always be lowered by choosing a different set of parameters, as one must remember that we have chosen our parameter space so as to maximize the lightest Higgs boson mass. The light green region represents the area where the stop squark mixing is minimum, i.e., $X_t=0$ while the blue upper region is for maximal mixing where $X_t=6$. The red shaded region is for all values of Higgs mass greater than 126 GeV and it is overlapped by the blue and the green regions. Fig.~\ref{fig:1d}(b) represents the upper limit of the Higgs mass and as a function $M_S$ in Fig.~\ref{fig:1d}(b). Again the red band is where the Higgs boson mass is between 124 GeV and 126 GeV, green region is for $X_t=0$, blue region represents $X_t=6$ and shaded red region is for all values of Higgs mass greater than 126 GeV which is overlapped by the green and the blue regions. The black solid line in each case represents the MSSM upper limit for the Higgs mass. We can see that a Higgs mass of 124 GeV can be very easily achieved in this case for a very small mass of stop squark and even for minimal mixing between them.
The $2\times 2$ mass-squared matrix corresponding to the neutral left-handed triplet scalar Higgs fields in the original basis is given as
\begin{equation}
\begin{bmatrix}m_3^2 + \frac{g_L^2}{2} (v_1^2 - v_2^2) +
2 g_V^2 (-v_R^2 +\overline v_R^2) + ( \lambda v_S + \mu_1)^2 & - \lambda (M^2 - \lambda v_R \overline v_R +
{\lambda^{\prime}} v_1 v_2 - \mu_S v_S-\kappa v_S^2)+\lambda A_\lambda v_S + \mu_1 B_1 \\ - \lambda (M^2 - \lambda v_R \overline v_R +
{\lambda^{\prime}} v_1 v_2 - \mu_S v_S-\kappa v_S^2)+\lambda A_\lambda v_S + \mu_1 B_1&
m_4^2 - \frac{g_L^2}{2} (v_1^2 - v_2^2) +
2 g_V^2 (v_R^2 -\overline v_R^2) + ( \lambda v_S + \mu_1)^2\end{bmatrix}
\label{eq:1dleft}
\end{equation}
We now look at the pseudo-scalar Higgs boson masses in this model. The structure of this sector is very similar to the scalar Higgs boson in the sense that the left-handed triplet fields decouple to form a $2\times 2$ matrix which is exactly the same as given in Eq.~(\ref{eq:1dleft}) while the imaginary components of the other neutral Higgs bosons form a $5\times 5$ matrix. We choose a basis given as
\begin{equation}
g_1 = \frac{v_1 \phi^0_{1} - v_2 \phi^0_{2}}{\sqrt{v_1^2+v_2^2}},~~~~
g_2 = \frac{v_R \delta^{c^0} -\overline v_R \overline \delta^{c^0}}{\sqrt{v_R^2+\overline v_R^2}},~~~~
h_1 = \frac{v_2 \phi^0_{1} + v_1 \phi^0_{2}}{\sqrt{v_1^2+v_2^2}},~~~~
h_2 = \frac{\overline v_R \delta^{c^0} + v_R \overline \delta^{c^0}}{\sqrt{v_R^2+\overline v_R^2}}.
\end{equation}
The Im($g_1$) and Im($g_2$) fields can be identified as the Goldstone bosons which are absorbed by the $Z_R$-boson and the $Z$-boson to make them massive. Integrating out these Goldstone states, the resulting $3\times 3$ matrix in the basis $({\text{Im}}(h_2),{\text{Im}}(h_1),{\text{Im}}(S))$ is given as
\begin{eqnarray}
M_{11} &=& m_1^2 + m_2^2 + \lambda^2(v_R^2+\overline v_R^2+2 v_S^2) +2 \mu_1 (2 \lambda v_s +\mu_1), \notag \\
M_{12} &=& -\lambda {\lambda^{\prime}} \sqrt{(v_1^2+v_2^2)(v_R^2+\overline v_R^2)} , \notag \\
M_{13} &=& \lambda (\mu _S+2 \kappa v_S-A_ \lambda) \sqrt{v_R^2+\overline v_R^2}, \notag \\
M_{22} &=& 2 m_5^2+{\lambda^{\prime}}^2(v_1^2+v_2^2+2 v_S^2)+2 \mu(2 {\lambda^{\prime}} v_S+\mu), \notag \\
M_{23} &=& {\lambda^{\prime}} (2A_{{\lambda^{\prime}}}-\mu_S-2 \kappa v_S) \sqrt{v_1^2+v_2^2}, \notag \\
M_{33} &=& m_S^2 +\lambda^2 (v_R^2+\overline v_R^2)+{\lambda^{\prime}}^2(v_1^2+v_2^2)-\mu_S(2 B_S-\mu_S) \notag \\
&+& 2\kappa(M^2-\lambda v_R \overline v_R +\lambda ^{\prime} v_1 v_2+\mu_S v_S+\kappa v_S^2-3 A_\kappa v_S).
\label{eq:1dpseudo}
\end{eqnarray}
The charged Higgs boson sector has six singly-charged fields in this model. Their mass-squared matrix can be split into two block diagonal matrices. There is a $2\times 2$ matrix corresponding to the $\delta^+$ and $\overline \delta^-$ fields which in its original basis is given as
\begin{equation}
\begin{pmatrix} g_V^2(\overline v_R^2-v_R^2)+ m_3^2 + (\mu_1+\lambda v_S)^2 & - \lambda (M^2 - \lambda v_R \overline v_R +
{\lambda^{\prime}} v_1 v_2 - \mu_S v_S-\kappa v_S^2-A_\lambda v_S) + \mu_1 B_1 \\ - \lambda (M^2 - \lambda v_R \overline v_R +
{\lambda^{\prime}} v_1 v_2 - \mu_S v_S-\kappa v_S^2-A_\lambda v_S) + \mu_1 B_1 & g_V^2(v_R^2-\overline v_R^2)+ m_4^2 + (\mu_1+\lambda v_S)^2 \end{pmatrix}.
\label{eq:1dchl}
\end{equation}
The other $4\times4$ block has two Goldstone bosons which are absorbed by $W_R$ and $W$ gauge bosons to get mass. We choose a basis given as:
\begin{equation}
\sigma_1^+ = \frac{v_1 \phi_1^+ - v_2 \phi_2^{-^*} }{\sqrt{v_1^2+v_2^2}},~\sigma_2^+ = \frac{v_2 \phi_1^+ + v_1 \phi_2^{-^*} }{\sqrt{v_1^2+v_2^2}},~\sigma_3^+ = \frac{v_R \overline \delta^{c^{+}} -\overline v_R \delta^{c^{-^*}}}{\sqrt{v_R^2+\overline v_R^2}},~\sigma_4^+ = \frac{\overline v_R \overline \delta^{c^{+}} + v_R \delta^{c^{-^*}}}{\sqrt{v_R^2+\overline v_R^2}},
\end{equation}
where the Goldstone eigenstates can be identified as
\begin{equation}
g_1^+ = \sigma_1^+,~~~~ g_2^+ = \frac{\sqrt{2(v_1^2+v_2^2)(v_R^2+\overline v_R^2)}\sigma_4^+ + (v_2^2-v_1^2)\sigma_2^+}{\sqrt{(v_2^2-v_1^2)^2+2(v_1^2+v_2^2)(v_R^2+\overline v_R^2)}}.
\end{equation}
This gives us the actual basis for the physical singly-charged Higgs bosons to be
\begin{equation}
h_1^+ = \sigma_3^+,~~~~
h_2^+ = \frac{(v_2^2-v_1^2)\sigma_4^+ - \sqrt{2(v_1^2+v_2^2)(v_R^2+\overline v_R^2)}\sigma_2^+}{\sqrt{(v_2^2-v_1^2)^2+2(v_1^2+v_2^2)(v_R^2+\overline v_R^2)}},
\end{equation}
and the $2\times 2$ singly-charged Higgs boson mass-squared matrix elements are given as
\begin{eqnarray}
M_{11}&=& \frac{g_R^2 \left\{(v_1^2-v_2^2)(v_R^2-\overline v_R^2)+2(v_R^2+\overline v_R^2)^2 \right\} }{(v_R^2 +\overline v_R^2)} \notag \\
&-& \frac{ 2(v_R^2+\overline v_R^2) \left\{ \lambda(-M^2+A_\lambda v_S +\lambda v_R \overline v_R-\lambda^{\prime}v_1 v_2+\mu_S v_S+\kappa v_S^2)+B_2 \mu_1\right\} }{v_R \overline v_R}, \notag \\
M_{12} &=& \frac{2 g_R^2 v_R \overline v_R \sqrt{
v_1^4 + 2 v_1^2 (-v_2^2 + v_R^2 +\overline v_R^2) +
v_2^2 [v_2^2 + 2 (v_R^2 +\overline v_R^2)]}}{v_R^2 + \overline v_R^2}, \notag \\
M_{22}&=&-\frac{g_R^2 (v_R^2 -\overline v_R^2) \left[v_1^4 + 2 v_1^2 (-v_2^2 + v_R^2 +\overline v_R^2) + v_2^4 +2 v_2^2 (v_R^2 +\overline v_R^2)\right]}{(v_1^2 - v_2^2) (v_R^2 +\overline v_R^2)}.
\label{eq:1dch}
\end{eqnarray}
Using the minimization conditions given in Eq.~(\ref{eq:min1d}), we eliminate $B_2$, $m_2$, $m_5$, $B$ and $C_\lambda$ in terms of the other parameters and numerically calculate the Higgs boson mass spectrum. We choose the parameters such that the lightest neutral scalar Higgs boson mass is 125 GeV after the radiative corrections and with a stop squark mass of 570 GeV and a stop squark mixing parameter $X_t$ = 4, as an example. The soft quadratic mass terms for the left-handed and right-handed triplets were chosen to be different since otherwise it leads to unphysical states with some negative eigenvalues for the Higgs boson mass-squared matrices. The numerical values of the chosen parameters and the masses obtained are given in Table~\ref{tab:one}. It is easy to identify the left-handed triplet Higgs boson eigenvalues since they decouple in each case as discussed earlier. In the table, $M_{H_1^\Delta}$ and $M_{H_2^\Delta}$ denote the mass-squared values for the left-handed triplet scalar Higgs bosons. Similarly $M_{A_1^\Delta}$, $M_{A_2^\Delta}$ and $M_{\Delta_1^+}$, $M_{\Delta_2^+}$ are the squared masses for the pseudo-scalar and the single-charged left-handed triplet Higgs bosons respectively. This numerical result shows the self-consistency of the model.
\vspace*{0.1in}
\noindent {\large{\bf{Chargino and Neutralino masses}}}
\vspace*{0.1in}
The particle spectrum of this model is much richer compared to the Minimal Supersymmetric Standard Model and hence the study of the chargino and neutralino masses is crucial for determining the lightest supersymmetric particle, a candidate for dark matter in the universe. The higgsinos and the gauginos mix to form charginos and neutralinos. The chargino mass matrix in this case is written as
\begin{equation}
{\mathcal{L}}_{ch} = -\frac{1}{2}\begin{pmatrix}\widetilde{\delta}^{c^{-}}&\widetilde{\overline \delta}^-&\widetilde{\phi}_2^-&\widetilde{W}_R^-&\widetilde{W}_L^-\end{pmatrix} \begin{pmatrix}\mu_1+\lambda^* v_S &0&0&-\sqrt{2}g_R v_R&0 \\ 0&\mu_1+\lambda v_S&0&0&0 \\ 0&0&\mu+\lambda^{\prime} v_S&g_R v_2&g_L v_2 \\ \sqrt{2} g_R \overline v_R&0&g_R v_1&M_R&0 \\ 0&0&g_L v_1&0&M_L \end{pmatrix}\begin{pmatrix}\widetilde{\overline \delta}^{c^+} \\ \widetilde{\delta}^+ \\ \widetilde{\phi}_1^+ \\ \widetilde{W}_R^+ \\ \widetilde{W}_L^+ \end{pmatrix}.
\label{eq:1dchargino}
\end{equation}
where the gaugino soft mass terms are given as
\begin{equation}
\mathcal{L}_G = -\frac{1}{2}\left(M_3 \widetilde{g}\widetilde{g}+ M_R \widetilde{W}_R \widetilde{W}_R + M_L \widetilde{W}_L \widetilde{W}_L + M_1 \widetilde{B} \widetilde{B} + h.c.\right)
\label{eq:gluino}
\end{equation}
The neutralino mass matrix splits into two matrices with the left-handed triplet Higgsino fields $\widetilde \delta^0$ and $\widetilde {\overline \delta^0}$ decoupling to form a $2\times 2$ matrix given as
\begin{equation}
\begin{pmatrix} 0&\mu_1+\lambda v_S \\ \mu_1+\lambda v_S&0 \end{pmatrix}~.
\end{equation}
The mass matrix for the other neutral fields in the basis $\begin{pmatrix}\widetilde{\delta}^{c^0}&\widetilde{\overline \delta}^{c^0}&\widetilde{\phi}_1^0&\widetilde{\phi}_2^0&\widetilde{B}&\widetilde{W}_{R_3}&\widetilde{W}_{L_3}&\widetilde{S}\end{pmatrix}$ is given as
\begin{equation}
\begin{pmatrix} 0&\mu_1+\lambda^* v_S&0&0&-\sqrt{2}g_V v_R&\sqrt{2}g_R v_R&0&\lambda^* \overline v_R \\ \mu_1+\lambda^* v_S&0&0&0&\sqrt{2} g_V \overline v_R&-\sqrt{2}g_R \overline v_R&0&\lambda^* v_R \\ 0&0&0&-\mu-\lambda^{\prime} v_S&0&-\frac{g_R v_1}{\sqrt{2}}&-\frac{g_L v_1}{\sqrt{2}}&-\lambda^{\prime} v_2 \\ 0&0&-\mu-\lambda^{\prime} v_S&0&0&\frac{g_R v_2}{\sqrt{2}}&\frac{g_L v_2}{\sqrt{2}}&-\lambda^{\prime} v_1 \\ -\sqrt{2} g_V v_R&\sqrt{2}g_V \overline v_R&0&0&M_1&0&0&0 \\ \sqrt{2} g_R v_R&-\sqrt{2}g_R \overline v_R&-\frac{g_R v_1}{\sqrt{2}}&\frac{g_R v_2}{\sqrt{2}}&0&M_R&0&0 \\ 0&0&-\frac{g_L v_1}{\sqrt{2}}&\frac{g_L v_2}{\sqrt{2}}&0&0&M_L&0 \\ \lambda ^* \overline v_R & \lambda^* v_R &-\lambda^{\prime} v_2&-\lambda^{\prime} v_1&0&0&0&\mu_S\end{pmatrix},
\end{equation}
where $M_R, M_L$ and $M_1$ are defined above. Parity invariance further demands $M_L=M_R^*$ and $M_1$ and $M_3$ are real.
The chargino and neutralino masses for this models are given in Table~\ref{tab:one}. Here again the states $\widetilde \Delta_1^+$ and $\widetilde \Delta_{1,2}^0$ refer to the chargino and neutralino states corresponding to the left-handed triplet Higgsinos.
\begin{center}
\begin{table}[h!]
\centering
\begin{tabular}{|C{2.8cm}|C{2.8cm}|C{2.8cm}|C{2.9cm}|C{3.1cm}|}
\hline
{\bf{Scalar Higgs boson masses}} & {\bf{Pseudo-scalar Higgs boson masses}} & {\bf{Single charged Higgs boson masses}}& \bf{Chargino masses}&{\bf{Neutralino masses}} \\ \hline
$M_{H_1}$=6.26 TeV, $M_{H_2}$=2.59 TeV, $M_{H_3}$=1.21 TeV, $M_{H_4}$=468 GeV, $M_{H_1^\Delta}$=4.51 TeV, $M_{H_2^\Delta}$=1.92 TeV & $M_{A_1}$=4.53 TeV, $M_{A_2}$=3.34 TeV, $M_{A_3}$=514 GeV, $M_{A_1^\Delta}$=4.51 TeV, $M_{A_2^\Delta}$=1.92 TeV & $M_{H_1^+}$=4.77 TeV, $M_{H_2^+}$=513 GeV, $M_{\Delta_1^+}$=4.51 TeV, $M_{\Delta_2^+}$=1.92 TeV & $M_{\widetilde \Delta_1^+}$= 2.65 TeV, $M_{\widetilde \chi_1^+}$= 4.23 TeV, $M_{\widetilde \chi_2^+}$= 2.38 TeV, $M_{\widetilde \chi_3^+}$= 809 GeV, $M_{\widetilde \chi_4^+}$= 348 GeV & $M_{\widetilde \Delta_{1,2}^0}$= 2.65 TeV, $M_{\widetilde \chi_1^0}$= 5.98 TeV, $M_{\widetilde \chi_2^0}$= 4.85 TeV, $M_{\widetilde \chi_3^0}$= 3.09 TeV, $M_{\widetilde \chi_4^0}$= 2.00 TeV, $M_{\widetilde \chi_5^0}$= 1.15 TeV, $M_{\widetilde \chi_6^0}$= 885 GeV, $M_{\widetilde \chi_7^0}$= 352 GeV, $M_{\widetilde \chi_8^0}$= 346 GeV\\ \hline
\end{tabular}
\caption{{Higgs boson, chargino and neutralino masses for a sample point for case with triplets using the parameters given as:
$\lambda^{\prime}$=0.7, $\lambda$=-0.3, $v_1$=173.14 GeV, $v_2$=17.3 GeV, $v_R$=3 TeV, $\overline v_R$=3.1 TeV, $\mu_1$=3.1 TeV, $\mu_2$=3.1 TeV, $\mu$=-1.4 TeV, $\mu_S$=-700 GeV, $m_S^2$=9 TeV$^2$, $v_S$=1.5 TeV, $B_S$=2 TeV, $m_1^2=m_3^2$=1 TeV$^2$, $\kappa$=0.1, $A_\kappa$=1 TeV, $m_4^2$=9 TeV$^2$, $A_{\lambda}$=-4 TeV, $A_{\lambda^{\prime}}$=-1 TeV, $M_R$=800 GeV, $M_L$=800 GeV, $M_1$=400 GeV and $B_1$ is chosen to be equal to $B_2$ which was fixed using the minimzation conditions.}}
\label{tab:one}
\end{table}
\end{center}
The doubly charged Higgs boson sector of the model is discussed in more detail in Sec. \ref{2chhiggs} where we also cary
out the one-loop radiative corrections to its mass and show the consistency of the framework.
\subsection{Symmetry breaking with a pair of Higgs triplets, a bidoublet and a heavy singlet}
\label{heavysinglet}
We now look at the case where the single Higgs $S$ is heavy and can be integrated out from the low energy sector of the model to give the following superpotential:
\begin{eqnarray}
W&=&\mu_1{\text{Tr}}(\Delta \overline{\Delta})+\mu_2{\text{Tr}}(\Delta^{c}\overline{\Delta}^{c}) +\epsilon \text{Tr}\left[\Delta^c \overline\Delta^c\right]^2 +\frac12 \mu {\text{Tr}}(\Phi^{T}\tau_{2}\Phi\tau_{2}).
\end{eqnarray}
Here $\epsilon$ is proportional to $1/{M_S}$ with $M_S$ being the scale at which the singlet is integrated out. Note that $\epsilon$ is a relevant operator which is kept in our analysis, although the field $S$ has been integrated out. Since $\epsilon$ is very small, we only kept the $\epsilon\text{Tr}(\Delta ^c \overline {\Delta}^c)^2$ term in the superpotential as other terms will have no significant effect to the lightest CP-even Higgs boson mass.
The $D$-term of the Higgs potential is exactly same as in Eq.~(\ref{eq:two}) but there will be different contributions to the $F$-term and the soft supersymmetry breaking terms. They are given by:
\begin{eqnarray}
V_F&=&|\mu_1|^2{\text{Tr}}(\Delta^{\dagger} \Delta+{\overline\Delta}^{\dagger} \overline\Delta)+{\text{Tr}}\left[ \left|\mu_2^2{\Delta^c}+2 \epsilon \Delta^c \overline\Delta^c \Delta^c \right|^2+\left|\mu_2^2 \overline\Delta^c+2 \epsilon \overline\Delta^c \Delta^c \overline\Delta^c \right|^2\right] \notag \\
&+& |\mu|^2 {\text{Tr}}(\Phi^\dagger \Phi), \\
V_{Soft}&=& m_1^2 {\text{Tr}}(\Phi^\dagger \Phi)+\left[ B \mu {\text{Tr}}(\Phi^T \tau_2 \Phi \tau_2)+h.c.\right]+m_3^2{\text{Tr}}(\Delta^\dagger\Delta)+m_4^2{\text{Tr}}(\overline\Delta^\dagger\overline\Delta) \nonumber \\
&+&m_5^2{\text{Tr}}({\Delta^c}^\dagger\Delta^c)+m_6^2{\text{Tr}}({\overline{\Delta}^c}^\dagger\overline\Delta^c)+{\text{Tr}}(B_1 \mu _1\Delta \overline\Delta + h.c.)\nonumber \\
&+&{\text{Tr}}(B_2 \mu_2\Delta^c \overline{\Delta}^c+h.c.)+\left[ \epsilon D_{\epsilon} {\text{Tr}} ( {\Delta^c} \overline\Delta^c )^2 + h.c. \right].
\label{eq:pot1c}
\end{eqnarray}
We use the same basis field redefinition as in Eq.~(\ref{eq:basis1d}). The minimization conditions are given as:
\begin{align}
0&= -4 B \mu {v_2}+ {v_1} \left( 4 {m_1}^2+ {g_L}^2 ( {v_1}^2- {v_2}^2)+ {g_R}^2 \left( {v_1}^2- {v_2}^2+2 {v_R}^2-2 \overline{v}_R^2\right)+4 \mu ^2\right), \nonumber \\
0& = -4 B \mu {v_1}+ {v_2} \left(4 {m_1}^2+ {g_L}^2 \left(- {v_1}^2+ {v_2}^2\right)+ {g_R}^2 \left(- {v_1}^2+ {v_2}^2-2 {v_R}^2+2 \overline{v}_R^2\right)+4 \mu ^2\right), \nonumber \\
0& = 2 B_2 \mu_2 {\overline v_R}+\left[2 {m_5}^2+2 {\mu_2}^2+ {g_R}^2 \left( {v_1}^2- {v_2}^2\right)\right] {v}_R+2 \left( {g_R}^2+ {g_V}^2\right) {v}_R \left(- {\overline v_R}^2+ {v}_R^2\right) \nonumber \\
&+ 4 \epsilon \overline{v}_R\left[D_{\epsilon} v_R \overline v_R +\mu_2 (3 v_R^2+\overline v_R^2)+2 \epsilon v_R \overline v_R(2 v_R^2+\overline v_R^2) \right], \nonumber \\
0& = 2 B_2 \mu_2 {v_R}+\left[2 {m_6}^2+2 {\mu_2}^2+ {g_R}^2 \left(- {v_1}^2+ {v_2}^2\right)\right] \overline{v}_R+2 \left( {g_R}^2 + {g_V}^2\right) \overline{v}_R \left(- {v_R}^2+ \overline{v}_R^2\right) \nonumber \\
&+ 4 \epsilon v_R\left[D_{\epsilon} v_R \overline v_R +\mu_2 (v_R^2+3\overline v_R^2)+2 \epsilon v_R \overline v_R( v_R^2+2\overline v_R^2) \right].
\label{eq:min1c}
\end{align}
Calculating the neutral CP-even Higgs boson mass-squared matrix subject to these minimization conditions, the matrix elements can be obtained from Eq.~(\ref{eq:masmat1d}) by putting all the triplet and bidoublet couplings to the singlet Higgs to be zero with some extra terms in the $M_{33},M_{34},M_{44}$ elements. The relevant terms in the mass-squared matrix are:
\begin{align}
&M_{11} = \frac{(g_L^2+g_R^2)(v_1^2-v_2^2)^2}{2(v_1^2+v_2^2)}, \notag \\
& M_{13} = \frac{g_R^2 (v_R^2-\overline v_R^2)(v_1^2-v_2^2)}{\sqrt{(v_1^2+v_2^2)(v_R^2+\overline v_R^2)}}, \notag \\
& M_{14} = \frac{2g_R^2 v_R \overline{v}_R(v_1^2-v_2^2)}{\sqrt{(v_1^2+v_2^2)(v_R^2+\overline v_R^2)}}, \\
& M_{33} = \frac{2(g_R^2+g_V^2)v_R^3-B_2 \mu_2\overline{v}_R-2 \epsilon \overline v_R \left[ \mu_2 (\overline{v}_R^2-3 v_R^2)-8 \epsilon v_R^3 \overline v_R \right]}{v_R}, \notag \\
& M_{34}= B_2 \mu_2-2(g_R^2+g_V^2)v_R\overline{v}_R+\epsilon \left[3 \mu_2(v_R^2+\overline v_R^2)+2 v_R \overline v_R(D_{\epsilon}+4 \epsilon(v_R^2+\overline v_R^2) \right], \notag \\
& M_{44} = \frac{2(g_R^2+g_V^2)\overline{v}_R^3-B_2 \mu_2v_R+2 \epsilon v_R \left[ \mu_2 (3 \overline{v}_R^2- v_R^2)+8 \epsilon \overline v_R^3 v_R \right]}{\overline v_R}.
\label{eq:masmat1c}
\end{align}
We calculate the contribution of the off-diagonal ($M_{13},M_{14}$) entries in the mass-squared matrix to the lightest eigenvalue using the seesaw formula. For simplicity we take the approximation $D_\epsilon =0$ and we get the following result:
\begin{equation}
M_{h_{tree}}^2 = 2 M_W^2 {\cos^2 2\beta}\left[1-\frac{x}{2 \left( \frac{g_R^4 x}{g_R^2-{g^\prime}^2}+y \right)} \right]
\label{inter}
\end{equation}
where
\begin{align}
x &= B_2 \mu_2 (v_R^2 - \overline v_R^2)^2 +
2 \epsilon (v_R^2 +
\overline v_R^2) \left[\mu_2 (v_R^4 - 10 v_R^2 \overline v_R^2 + \overline v_R^4) -
24 \epsilon v_R^3 \overline v_R^3\right], \nonumber \\
y &= 8 v_R \overline v_R \epsilon ( B_2 \mu_2^2 (v_R^2 + \overline v_R^2) +
\mu_2^2 [3 v_R^4 + 2 v_R^2 \overline v_R^2 + 3 \overline v_R^4) \epsilon +
2 \mu_2 \overline v_R (7 v_R^5 + 6 v_R^3 \overline v_R^2 + 7 v_R \overline v_R^4) \epsilon^2 \nonumber \\
& +
v_R \overline v_R \epsilon (3 m_8^2 (v_R^2 +\overline v_R^2) +
16 v_R \overline v_R (v_R^4 + v_R^2 \overline v_R^2 + \overline v_R^4) \epsilon^2], \notag
\end{align}
$\tan \beta = \frac{v_1}{v_2}$ and $g_R=g_L$.
This result shows that the lightest CP-even Higgs boson mass has an upper limit of $ \sqrt{2} M_W$ in this case which can be realized if $x=0$. If we consider $\overline v_R^2-v_R^2 \sim M_{SUSY}^2$ and $v_R, \overline v_R >> M_{SUSY}$, we get an upper limit of $M_Z$ for the lightest scalar Higgs boson mass. Eq. (\ref{inter}) interpolates between the two interesting cases of
$v_R \sim M_{SUSY}$ and $v_R \gg M_{SUSY}$. Taking the least constraining of the limits we have for the largest allowed
$M_h$ at tree level,
\begin{equation}
M^2_{h_{tree}} = 2 M_W^2 \cos^2 2 \beta.
\end{equation}
Including the one and two loop corrections from the top quark and stop squark, we get:
\begin{eqnarray}
M_{h_{max}}^2 &=& (2 M_W^2 \cos^2 2 \beta) \Delta_1 + \Delta_2,
\end{eqnarray}
where $\Delta_1$ and $\Delta_2$ are defined in Eq.~(\ref{eq:rad}).
\begin{figure}[h!]\centering
\includegraphics[width=3.1in]{1c_tb.jpg}
\includegraphics[width=3.1in]{1c_Ms.jpg}
\caption{{\sl(a) Variation of Higgs boson mass with $\tan \beta$, (b) Higgs boson mass as a function of $M_S$
for the case with Higgs triplets, one bidoublet and a heavy singlet.
Notation here is the same as in Fig. \ref{fig:1d}.}}
\label{fig:1c}
\end{figure}
The Higgs boson mass is plotted in Fig~\ref{fig:1c}(a) as a function of $\tan \beta$. The red region in the figure represents the band where the mass is between 124 GeV and 126 GeV. The light green region represents the area where the stop squark mixing is minimum i.e. $X_t=0$ while the blue upper region is for maximal mixing where $X_t=6$. The red shaded region is for all values of Higgs mass greater than 126 GeV and it is overlapped by the blue and the green regions. Fig.~\ref{fig:1c}(b) represents the upper limit of the Higgs mass and as a function $M_S$. Again the red band is where the Higgs boson mass is between 124 GeV and 126 GeV, green region is for $X_t=0$, blue region represents $X_t=6$ and shaded red region is for all values of Higgs mass greater than 126 GeV which is overlapped by the green and the blue regions. The black solid line in each case represents the MSSM upper limit for the Higgs mass.
The pseudo-scalar mass-squared matrix is again two $2\times2$ blocks which can be obtained by putting all the singlet couplings to zero in Eq.~(\ref{eq:1dleft}) and Eq.~(\ref{eq:1dpseudo}).
The charged Higgs boson mass-squared matrix is exactly the same as in Eq.~(\ref{eq:1dchl}) and Eq.~({\ref{eq:1dch}) with all the singlet couplings set to zero and in the limit where we take $D_{\epsilon}\rightarrow 0$.
\vspace*{0.1in}
\noindent {\large{\bf{Chargino and Neutralino masses}}}
\vspace*{0.1in}
We now look at the chargino and neutralino sector in this case. The chargino basis is exactly the same as in the case discussed in {{section 3.1}}. The chargino mass matrix in this case is written as
\begin{equation}
M_{ch} = \begin{pmatrix}\mu_2+\epsilon v_R \overline v_R &0&0&-\sqrt{2}g_R v_R&0 \\ 0&\mu_1&0&0&0 \\ 0&0&\mu&g_R v_2&g_L v_2 \\ \sqrt{2} g_R \overline v_R&0&g_R v_1&M_R&0 \\ 0&0&g_L v_1&0&M_L \end{pmatrix}.
\end{equation}
The neutralino mass matrix again splits into two matrices. The matrix in the basis $\begin{pmatrix}\widetilde{\delta}^0&{\widetilde{\overline \delta}^0} \end{pmatrix}$ is given as
\begin{equation}
\begin{pmatrix} 0&\mu_1 \\ \mu_1&0 \end{pmatrix},
\end{equation}
while the mass matrix in the basis $\begin{pmatrix}\widetilde{\delta}^{c^0}&\widetilde{\overline \delta}^{c^0}&\widetilde{\phi}_1^0&\widetilde{\phi}_2^0&\widetilde{B}&\widetilde{W}_{R_3}&\widetilde{W}_{L_3}\end{pmatrix}$ is given as
\begin{equation}
\begin{pmatrix} \epsilon \overline v_R^2&\mu_2+\epsilon v_R \overline v_R&0&0&-\sqrt{2}g_V v_R&\sqrt{2}g_R v_R&0 \\ \mu_2+\epsilon v_R \overline v_R&\epsilon v_R^2&0&0&\sqrt{2} g_V \overline v_R&-\sqrt{2}g_R \overline v_R&0 \\ 0&0&0&-\mu&0&-\frac{g_R v_1}{\sqrt{2}}&-\frac{g_L v_1}{\sqrt{2}} \\ 0&0&-\mu&0&0&\frac{g_R v_2}{\sqrt{2}}&\frac{g_L v_2}{\sqrt{2}} \\ -\sqrt{2} g_V v_R&\sqrt{2}g_V \overline v_R&0&0&M_1&0&0\\ \sqrt{2} g_R v_R&-\sqrt{2}g_R \overline v_R&-\frac{g_R v_1}{\sqrt{2}}&\frac{g_R v_2}{\sqrt{2}}&0&M_R&0 \\ 0&0&-\frac{g_L v_1}{\sqrt{2}}&\frac{g_L v_2}{\sqrt{2}}&0&0&M_L\end{pmatrix}.
\label{eq:1cneutalino}
\end{equation}
\subsection{Case with two pair of triplets and a bidoublet} \label{case1a}
This is a special case of the one discussed in {{Section~\ref{case1d}}}. We do not have the singlet Higgs and as a result it will be seen that the lightest Higgs boson mass upper limit becomes the same as MSSM. We also show explicitly the
self-consistency of this model which requires $v_R$ and $M_{SUSY}$ to be of the same order.
The most general superpotential relevant to our calculation is given by:
\begin{eqnarray}
W&=&\mu_1{\text{Tr}}(\Delta \overline{\Delta})+\mu_2{\text{Tr}}(\Delta^{c}\overline{\Delta}^{c}) +\frac12 \mu {\text{Tr}}(\Phi^{T}\tau_{2}\Phi\tau_{2}).
\end{eqnarray}
The $D$-term in the Higgs potential is exactly the same as given in Eq.~(\ref{eq:two}), the $F$-term can be obtained from Eq.~(\ref{eq:1df}) by putting all the singlet couplings to zero. The soft supersymmetry breaking terms are given by:
\begin{eqnarray}
V_{Soft}&=& m_1^2{\text{Tr}}({\Delta^c}^\dagger\Delta^c)+m_2^2{\text{Tr}}({\overline{\Delta}^c}^\dagger\overline\Delta^c)+m_3^2{\text{Tr}}(\Delta^\dagger\Delta)+m_4^2{\text{Tr}}(\overline\Delta^\dagger\overline\Delta) \nonumber \\
&+&m_5^2 {\text{Tr}}(\Phi^\dagger \Phi)+ \left[B \mu {\text{Tr}}(\Phi^T \tau_2 \Phi \tau_2)+h.c.\right]\nonumber \\
&+&\left[{B_1 \mu_1 \text{Tr}}(\Delta \overline\Delta) + h.c.\right] + \left[ {B_2 \mu_2\text{Tr}}(\Delta^c \overline{\Delta}^c)+h.c.\right].
\end{eqnarray}
We use this potential to calculate the Higgs boson mass-squared matrices for the charged, neutral CP-even and neutral CP-odd Higgs bosons. To easily identify the field corresponding to the lightest eigenvalue, we redefine the Higgs fields. This redefinition is the same as in Eq.~(\ref{eq:basis1d}).
The minimization conditions and the Higgs mass-squared in this case can again be obtained by putting all the singlet couplings to zero in the model of {{Section~\ref{case1d}}}.
Calculating the lightest eigenvalue for the CP-even Higgs boson mass-squared matrix we get:
\begin{equation}
M_{h_{tree}}^2 = \frac{g_L^4 ({g^\prime}^2 + g_R^2) (v_1^2 - v_2^2)^2}{2 [g_L^2 g_R^2 +
{g^\prime}^2 (g_L^2 - g_R^2)] (v_1^2 + v_2^2)}.
\end{equation}
If we assume that the $SU(2)_R$ gauge coupling ($g_R$) is equal to the $SU(2)_L$ gauge coupling ($g_L$), $\tan\beta = \frac{v_1}{v_2}$ and $v^2 = v_1^2+v_2^2$, then
\begin{equation}
M_{h_{tree}}^2 = \frac{(g_L^2+{g^ \prime}^2)}{2} v^2 \cos^2{2\beta}.
\end{equation}
The mass of the $Z$ boson in this model is $\sqrt{\frac{g_L^2+{g^\prime}^2}{2}} v$. So we see that the tree-level lightest CP-even Higgs mass has an upper limit of $M_Z$. This is same as the case of MSSM.
The charged mass-squared matrix is the same as in Eq.~(\ref{eq:1dch}) while the pseudo-scalar mass-squared matrix is composed of two $2\times2$ block which can be obtained from Eq.~(\ref{eq:1dleft}) and Eq.~(\ref{eq:1dpseudo}) by putting all the singlet couplings to zero.
The chargino mass matrix in this case is a special limit of {{Section~\ref{case1d}}} obtained by neglecting all the singlet couplings while the neutralino mass matrix is obtained from Eq.~(\ref{eq:1cneutalino}) by putting $\epsilon=0$.
\subsection{Case with two pair of triplets and two bidoublets} \label{case1b}
This case is a realistic model where, unlike previous cases, we can generate the CKM matrices for quarks directly. The calculation of the Higgs mass, though shows that the result for the upper limit on $M_h$ is exactly the same as the case with only one bidoublet. Due to the complexity of the calculations, we only discuss the neutral CP-even Higgs boson mass in this case and see that the largest $M_h$ is the same as with one bidoublet. The particle content of the Higgs sector will be exactly as in Eq.~(\ref{eq:triphig}) except in this case $a = 1,2$.
The superpotential of the model is given as:
\begin{eqnarray}
W&=&\mu_1{\text{Tr}}(\Delta \overline{\Delta})+\mu_2{\text{Tr}}(\Delta^{c}\overline{\Delta}^{c}) +\frac 12\mu_{ab} {\text{Tr}}(\Phi_a^{T}\tau_{2}\Phi_b\tau_{2}).
\end{eqnarray}
The relevant terms in the Higgs potential is given by:
\begin{eqnarray}
V_F&=&|\mu_1|^2{\text{Tr}}(\Delta^{\dagger} \Delta+{\overline\Delta}^{\dagger} \overline\Delta)+|\mu_2|^2{\text{Tr}}({\Delta^c}^{\dagger} \Delta^c+{\overline\Delta^c}^\dagger \overline\Delta^c) \notag \\
&+&\sum\limits_{a=1}^2 {\text{Tr}} |(\mu_{a1} \Phi_{1}+\mu_{a2} \Phi_2)|^2, \\
V_D&=&\frac{g_L^2}{8}\sum \limits_{a=1}^3 \left| {\text{Tr}}(2\Delta^\dagger \tau_a \Delta+2\overline\Delta^\dagger \tau_a \overline\Delta+(\Phi_1^\dagger \tau_a \Phi_1)+(\Phi_2^\dagger \tau_a \Phi_2)\right| ^2\nonumber \\
&+&\frac{g_R^2}{8}\sum \limits_{a=1}^3 \left| {\text{Tr}}(2{\Delta^c}^\dagger \tau_a \Delta^c+2\overline{\Delta^c}^\dagger \tau_a \overline\Delta^c+(\Phi_1^* \tau_a \Phi_1^T)+(\Phi_2^* \tau_a \Phi_2^T)\right|^2 \nonumber \\
&+& \frac{g_V^2}{2}\left| {\text{Tr}}(\Delta^\dagger \Delta- \overline\Delta^\dagger \overline\Delta-{\Delta^c}^\dagger \Delta^c+ \overline{\Delta^c}^\dagger \overline\Delta^c)\right| ^2, \\
V_{Soft}&=& m_{ab}^2 {\text{Tr}}(\Phi_a^\dagger \Phi_b)+\sum\limits_{a,b=1}^2{{B_{ab} \mu_{ab}}} \left[{\text{Tr}}(\Phi_a^T \tau_2 \Phi_b \tau_2)+h.c.\right]+m_3^2{\text{Tr}}(\Delta^\dagger\Delta)+m_4^2{\text{Tr}}(\overline\Delta^\dagger\overline\Delta) \nonumber \\
&+&m_5^2{\text{Tr}}({\Delta^c}^\dagger\Delta^c)+m_6^2{\text{Tr}}({\overline{\Delta}^c}^\dagger\overline\Delta^c)+ [B_1 \mu_1{\text{Tr}}(\Delta\overline\Delta)+h.c.] \notag \\
&+& [B_2 \mu_2{\text{Tr}}(\Delta^c\overline\Delta^c)+h.c.].
\label{eq:three}
\end{eqnarray}
We use this Higgs potential for this variation of the LRSUSY model and calculate the mass-squared matrix for the neutral CP-even Higgs boson. The vacuum structure for this model is given by:
\begin{equation}
\left<\Delta^c\right> = \begin{pmatrix}
0 & v_R \\ 0 & 0 \end{pmatrix},~
\left<\overline\Delta^c\right> = \begin{pmatrix}
0&0\\\overline{v}_R & 0 \end{pmatrix},~
\left<\Phi_1\right> ={\begin{pmatrix}
0 & v_{d_1} \\ v_{u_1} & 0 \end{pmatrix}},~
\left<\Phi\right> ={\begin{pmatrix}
0 & v_{{d_2}} \\ v_{{u_2}} & 0 \end{pmatrix}}.
\end{equation}
The left-handed triplet fields $\Delta$ and $\overline\Delta$ do not get any VEV. We do a field redefinition with the $\phi_{11}^0, \phi_{21}^0, \phi_{12}^0, \phi_{22}^0$ fields so that only one of the new fields get a non-zero vacuum expectation value. The transformation we use is given by:
\begin{eqnarray}
\rho_1&=&\frac{v_{{u_1}} \phi_{11}^0 +v_{{d_1}} \phi_{21}^0+v_{{u_2}}\phi_{12}^0+v_{{d_2}} \phi_{22}^0}{\sqrt{v_{{u_1}}^2+v_{{d_1}}^2+v_{{u_2}}^2+v_{{d_2}}^2}},~~\rho_2=\frac{v_{{d_1}}\phi_{11}^0-v_{{u_1}} \phi_{21}^0}{\sqrt{v_{{u_1}}^2+v_{{d_1}}^2}},~~\rho_3=\frac{v_{{d_2}}\phi_{21}^0-v_{{u_2}} \phi_{22}^0}{\sqrt{v_{{u_2}}^2+v_{{d_2}}^2}}, \nonumber \\
\rho_4&=&\frac{v_{{u_1}}(v_{{u_2}}^2+v_{{d_2}}^2)\phi_{11}^0+v_{{d_1}}(v_{{u_2}}^2+v_{{d_2}}^2)\phi_{21}^0 -v_{{u_2}}(v_{{u_1}}^2+v_{{d_1}}^2)\phi_{12}^0-v_{{d_2}}(v_{{u_1}}^2+v_{{d_1}}^2)\phi_{22}^0}{\sqrt{(v_{{u_1}}^2+v_{{d_1}}^2)(v_{{u_2}}^2+v_{{d_2}}^2)(v_{{u_1}}^2+v_{{d_1}}^2+v_{{u_2}}^2+v_{{d_2}}^2)}}.\notag
\end{eqnarray}
The $\rho_1$ field gets a VEV of $\sqrt{v_{{u_1}}^2+v_{{d_1}}^2+v_{{u_2}}^2+v_{{d_2}}^2}$, the other fields do not get any VEV. The $\Delta$ and $\overline\Delta$ fields decouple and we get a $6\times6$ mass-square matrix in the basis $({\text{Re}}\rho_1,{\text{Re}}\rho_2,{\text{Re}}\rho_3,{\text{Re}}\rho_4,{\text{Re}}{\delta^c}^0,{\text{Re}}{\overline\delta^c}^0)$. The minimization conditions for this case are given in the Appendix. The matrix elements for this case are not quoted here as they are lengthy and this case is not very interesting in terms of the final result which comes out to be exactly as {{section~\ref{case1a}}}.
Using the minimization conditions and the assumption that the right-handed symmetry breaking scale is much above the electroweak scale, we get the lightest eigenvalue to be:
\begin{equation}
M_{h_{tree}}^2 = \frac{\left(g_L^2+g^\prime\right) (v_{{u_2}}^2 - v_{{d_2}}^2 + v_{{u_1}}^2 - v_{{d_1}}^2)^2}{ 2(v_{{u_2}}^2 + v_{{d_2}}^2 +
v_{{u_1}}^2 + v_{{d_1}}^2)} = M_Z^2 \cos^2 2\beta
\end{equation}
where $\tan\beta = \frac{\sqrt{(v_{{u_1}}^2+v_{{u_2}}^2)}}{\sqrt{(v_{{d_1}}^2+v_{{d_2}}^2)}}$ and $v^2 = \sqrt{v_{{u_1}}^2+v_{{d_1}}^2+v_{{u_2}}^2+v_{{d_2}}^2}$. We have made the assumption that $g_R=g_L$.
This result is the same as the previous case with one bidoublet and gives an upper-limit for the tree-level mass of lightest CP-even neutral Higgs boson to be $M_Z$.
\section{Inverse seesaw model} \label{iseesaw}
The Higgs spectrum of this model is given in Eq.~(\ref{eq:2dh}). The most general superpotential terms needed for calculation of the Higgs boson mass are given as:
\begin{equation}
W = i \mu_1 H_L^T \tau_2 \overline H_L + i \mu_1 H_R^T \tau_2 \overline H_R + \lambda H_L^T \tau_2 \Phi \tau_2 H_R + \lambda\overline H_L^T \tau_2 \Phi \tau_2 \overline H_R + \mu {\text{Tr}}\left[\Phi \tau_2 \Phi^T \tau_2\right].
\end{equation}
The relevant Higgs potential in this case is given as:
\begin{eqnarray}
V_F &=& \text{Tr} \left[ \left| i \mu_1 \tau_2 \overline H_L + \lambda \tau_2 \Phi \tau_2 H_R \right| ^2 +\left| i \mu_1 \tau_2 \overline H_R + \lambda \tau_2 \Phi ^T \tau_2 H_L \right| ^2 \right. \nonumber \\
&+&~~~~~\left| -i \mu_1 \tau_2 H_L + \lambda \tau_2 \Phi \tau_2 \overline H_R \right|^2 + \left|- i \mu_1 \tau_2 H_R + \lambda \tau_2 \Phi ^T \tau_2 \overline H_L \right|^2 \nonumber \\
&+&~~~~~\left. \left| \lambda H_R H_L^T+ \lambda \overline H_R \overline H_L^T +2 \mu \phi ^T \right|^2 \right] , \\
V_D &=&\frac{g_L^2}{8}\sum \limits_{a=1}^3 \left| H_L^\dagger \tau_a H_L+\overline H_L^\dagger \tau_a \overline H_L+{\text{Tr}}(\Phi^\dagger \tau_a \Phi)\right| ^2 \nonumber \\ &+&\frac{g_R^2}{8}\sum \limits_{a=1}^3 \left| H_R^\dagger \tau_a H_R+\overline H_R^\dagger \tau_a \overline H_R+{\text{Tr}}(\Phi^* \tau_a \Phi^T)\right| ^2\nonumber \\
&+& \frac{g_V^2}{8}\left| H_R^\dagger H_R-\overline H_R^\dagger \overline H_R-H_L^\dagger H_L+\overline H_L^\dagger \overline H_L \right| ^2, \\
V_{Soft}&=& \text{Tr} \left[ m_1^2 H_L^\dagger H_L+m_2^2 H_R^ \dagger H_R+ m_3^2 \overline H_L^\dagger \overline H_L + m_4^2 \overline H_R^\dagger \overline H_R + m_5^2 \Phi ^ \dagger \Phi \right. \nonumber \\
&+&~~~ \left( \lambda A_{\lambda} H_L^T \tau_2 \Phi \tau_2 H_R + \lambda A_{\lambda} \overline H_L^T \tau_2 \Phi \tau_2 \overline H_R + h.c. \right)+ \left( B \mu \Phi^T \tau_2 \Phi \tau_2 + h.c. \right) \nonumber \\
&+&~~~ \left. \left( i B_1 \mu_1 H_L^T \tau_2 \overline H_L + i B_1 \mu_1 H_R^T \tau_2 \overline H_R + h.c \right) \right].
\end{eqnarray}
Parity conservation would require $m_1^2=m_2^2$ and $m_3^2=m_4^2$ but as in the previous case, we allow for soft breaking
of parity by the bilinear terms and choose these parameters to be different.
The vacuum expectation values of the Higgs fields are given as:
\begin{eqnarray}
\left<H_L\right> &=& \begin{pmatrix}
v_L \\ 0 \end{pmatrix},
\left<H_R\right> = \begin{pmatrix}
0 \\ v_R \end{pmatrix},
\left<\overline H_L\right> = \begin{pmatrix}
0 \\ \overline v_L \end{pmatrix},\nonumber \\
\left<\overline H_R\right> &=& \begin{pmatrix}
\overline v_R \\ 0 \end{pmatrix},
\left<\Phi\right> ={\begin{pmatrix}
0 & v_2 \\ v_1 & 0 \end{pmatrix}}.
\end{eqnarray}
We again choose a rotated basis given as
\begin{eqnarray}
\rho_1&=&\frac{v_{{u_1}} H_L^0 +v_{{d_1}} \overline H_L^0+v_{{u_2}}\phi_{1}^0+v_{{d_2}} \phi_{2}^0}{\sqrt{v_{{u_1}}^2+v_{{d_1}}^2+v_{{u_2}}^2+v_{{d_2}}^2}},~~\rho_2=\frac{v_{{d_1}}H_L^0-v_{{u_1}} \overline H_L^0}{\sqrt{v_{{u_1}}^2+v_{{d_1}}^2}},\notag \\
\rho_3&=&\frac{v_1 v_L H_L^0+v_1 \overline v_L \overline H_L^0-(v_L^2+\overline v_L^2) \phi_1^0}{\sqrt{(v_L^2+\overline v_L^2+v_1^2)(v_L^2+\overline v_L^2)}},\notag \\
\rho_4&=&\frac{v_2 v_L H_L^0+v_2 \overline v_L \overline H_L^0+v_1 v_2 \phi_1^0-(v_1^2+v_L^2+\overline v_L^2)\phi_2^0}{\sqrt{(v_L^2+\overline v_L^2+v_1^2+v_2^2)(v_L^2+\overline v_L^2+v_1^2)}},
\end{eqnarray}
such that only $\rho_1$ gets a non-zero vacuum expectation value at the electroweak symmetry breaking scale. The right-handed doublets get VEVs of order the right-handed symmetry breaking scale. The minimization conditions in this case are given as:
\begin{align}
0= & 2 m_5^2 v_1+\frac{v_1}{2}\left[ g_L^2 \left( v_1^2-v_2^2-v_L^2+\overline v_L^2 \right) +g_R^2
\left( v_1^2-v_2^2-v_R^2+\overline v_R^2 \right) + 4 \lambda^2 \left( v_L^2+v_R^2 \right) \right] \notag \\ & -2 \lambda A_{\lambda} v_L v_R+2 \lambda \mu_1 \left( v_L \overline v_R-v_R \overline v_L \right) + 4 \mu \left( \lambda \overline v_L \overline v_R-B v_2+2 \mu v_1 \right), \notag \\
0= & 2 m_5^2 v_2+\frac{v_2}{2}\left[ g_L^2 \left( -v_1^2+v_2^2+v_L^2-\overline v_L^2 \right) - g_R^2
\left( -v_1^2+v_2^2+v_R^2-\overline v_R^2 \right) + 4 \lambda^2 \left( \overline v_L^2+\overline v_R^2 \right) \right] \notag \\ & -2 \lambda A_{\lambda} \overline v_L \overline v_R-2 \lambda \mu_1 \left( v_L \overline v_R-v_R \overline v_L \right) + 4 \mu \left( \lambda v_L v_R-B v_1+2 \mu v_2 \right), \notag \\
0=& 2 m_1^2 v_L+\frac{v_L}{2}\left[ g_L^2 \left( -v_1^2+v_2^2+v_L^2-\overline v_L^2 \right) + g_V^2
\left( v_L^2-\overline v_L^2-v_R^2+\overline v_R^2 \right) + 4 \lambda^2 \left( v_1^2+ v_R^2 \right) \right] \notag \\ & -2 \lambda A_{\lambda} v_1 v_R+2 \lambda \mu_1 \overline v_R \left( v_1 -v_2 \right) + 2 \mu_1^2 v_L+2 B_1 \mu_1 \overline v_L + 4 \mu \lambda v_2 v_R, \notag \\
0=& 2 m_2^2 v_R+\frac{v_R}{2}\left[ g_R^2 \left( -v_1^2+v_2^2+v_R^2-\overline v_R^2 \right) + g_V^2
\left( -v_L^2+\overline v_L^2+v_R^2-\overline v_R^2 \right) + 4 \lambda^2 \left( v_1^2+ v_L^2 \right) \right] \notag \\ & -2 \lambda A_{\lambda} v_1 v_L-2 \lambda \mu_1 \overline v_L \left( v_1 -v_2 \right) + 2 \mu_1^2 v_R-2 B_1 \mu_1 \overline v_R + 4 \mu \lambda v_2 v_L, \notag \\
0=& 2 m_3^2 \overline v_L+\frac{\overline v_L}{2}\left[ g_L^2 \left( v_1^2-v_2^2-v_L^2+\overline v_L^2 \right) - g_V^2 \left( v_L^2-\overline v_L^2-v_R^2+\overline v_R^2 \right) + 4 \lambda^2 \left( v_2^2+\overline v_R^2 \right) \right] \notag \\ & -2 \lambda A_{\lambda} v_2 \overline v_R-2 \lambda \mu_1 v_R \left( v_1 -v_2 \right) + 2 \mu_1^2 \overline v_L+2 B_1 \mu_1 v_L + 4 \mu \lambda v_1 \overline v_R, \notag \\
0=& 2 m_4^2 \overline v_R+\frac{\overline v_R}{2}\left[ g_R^2 \left( v_1^2-v_2^2-v_R^2+\overline v_R^2 \right) + g_V^2 \left( v_L^2-\overline v_L^2-v_R^2+\overline v_R^2 \right) + 4 \lambda^2 \left( v_2^2+\overline v_L^2 \right) \right] \notag \\ & -2 \lambda A_{\lambda} v_2 \overline v_L+2 \lambda \mu_1 v_L \left( v_1 -v_2 \right) + 2 \mu_1^2 \overline v_R-2 B_1 \mu_1 v_R + 4 \mu \lambda v_1 \overline v_L.
\label{eq:iseesawmin}
\end{align}
The relevant mass-matrix elements in this case are given as:
\begin{eqnarray}
M_{11}&=& \frac{g_R^2 \left(v_1^2-v_2^2\right)^2+g_V^2 \left( v_L^2-\overline v_L ^2 \right)^2+g_L^2 \left(v_1^2-v_2^2-v_L^2+\overline{v}_L^2\right)^2+8 \lambda ^2 \left(v_1^2 v_L^2 +v_2^2
\overline{v}_L^2 \right) }{2 \left(v_1^2+v_2^2+v_L^2+\overline{v}_L^2\right)}, \nonumber \\
M_{12}&=& \frac{v_L \overline{v}_L \left(g_V^2 \left(v_L^2-\overline{v}_L^2\right)+g_L^2 \left(-v_1^2+v_2^2+v_L^2-\overline{v}_L^2\right)+2
\left(v_1^2-v_2^2\right) \lambda ^2\right)}{\sqrt{v_L^2+\overline{v}_L^2} \sqrt{v_1^2+v_2^2+v_L^2+\overline{v}_L^2}}, \nonumber \\
M_{13}&=&\left[v_1 \left\{g_V^2 (v_L^2-\overline{v}_L^2)^2+2 g_L^2 v_L^2
\left(-v_1^2+v_2^2+v_L^2-\overline{v}_L^2\right)\right. \right. \nonumber \\
&-& g_R^2 \left(v_1^2-v_2^2\right) \left(v_L^2+\overline{v}_L^2\right)
+\left. \left. 4 \lambda ^2 (v_1^2 v_L^2 - v_L^4 + v_2^2 \overline{v}_L^2 - v_L^2 \overline{v}_L^2)\right\}\right]\mbox {\Large$ /$}\notag \\
&~& ~\left(2 \sqrt{\left(v_L^2+\overline{v}_L^2\right) \left(v_1^2+v_L^2+\overline{v}_L^2\right)} \sqrt{v_1^2+v_2^2+v_L^2+\overline{v}_L^2}\right),\notag \\
M_{14}&=&\left[v_2 \left\{g_V^2 (v_L^2-\overline{v}_L^2)^2+2 g_L^2 (v_1^2+\overline v_L^2)
\left(v_1^2-v_2^2-v_L^2+\overline{v}_L^2\right)\right. \right. \nonumber \\
&+& \left. \left. g_R^2 \left(v_1^2-v_2^2\right) \left(2 v_1^2+v_L^2+\overline{v}_L^2\right)+4 \lambda ^2 \left( 2 v_1^2 v_L^2 -v_1^2 \overline{v}_L^2 + v_2^2 \overline{v}_L^2 -
v_L^2 \overline{v}_L^2 - \overline{v}_L^4 \right) \right\}\right]\mbox {\Large$ /$} \nonumber \\
&~&~ \left(2(v_1^2+v_2^2+v_L^2+\overline{v}_L^2)
\sqrt{\left(v_1^2+v_L^2+\overline{v}_L^2\right)}\right), \notag \\
M_{15} &=& \left[g_R^2 \left(-v_1^2+v_2^2\right) v_R+g_V^2 \left(-v_L^2+\overline{v}_L^2\right) v_R+4 \lambda \left\{ -A_\lambda v_1 v_L+\mu_1 (-v_1+v_2) \overline{v}_L \right. \right. \notag \\
&+& \left. \left. \lambda v_1^2 v_R + \lambda v_L^2 v_R +2 \mu v_2 v_L \right\}\right]/\left(2 \sqrt{v_1^2+v_2^2+v_L^2+\overline{v}_L^2}\right),\notag \\
M_{16} &=&\left[g_R^2 \left(v_1^2-v_2^2\right) \overline{v}_R+g_V^2 \left(v_L^2-\overline{v}_L^2\right) \overline{v}_R+4
\lambda \left\{\mu_1 (v_1-v_2) v_L-A_\lambda v_2 \overline{v}_L+\lambda v_2^2 \overline{v}_R \right. \right. \notag \\
&+&\left. \left. \lambda \overline{v}_L^2
\overline{v}_R +2 \mu v_1 \overline{v}_L \right\}\right]/\left(2 \sqrt{v_1^2+v_2^2+v_L^2+\overline{v}_L^2}\right),\notag \\
M_{55} &=& \mu_1^2+m_2^2-\frac14 \left[ g_R^2(v_1^2-v_2^2-3 v_R^2+\overline v_R^2)+g_V^2(v_L^2-\overline v_L^2-3 v_R^2+\overline v_R^2)\right] +\lambda^2(v_1^2+v_L^2),\notag \\
M_{56}&=& -\frac{(g_R^2+g_V^2) v_R \overline v_R+2B_1\mu_1}{2}, \notag \\
M_{66}&=&\mu_1^2+m_4^2+\frac14 \left[ g_R^2(v_1^2-v_2^2- v_R^2+3 \overline v_R^2)+g_V^2(v_L^2-\overline v_L^2- v_R^2+3\overline v_R^2)\right] +\lambda^2(v_2^2+\overline v_L^2).~~~~~~~~~~
\end{eqnarray}
All the other elements in the mass matrix are of order SUSY breaking scale squared or the right-handed symmetry breaking scale squared. The only matrix elements that can provide significant contributions to the lightest eigenvalue comes from $M_{15}$ and $M_{16}$. We focus on the $3\times3$ sector formed by $M_{11},M_{15},M_{16},M_{55},M_{56},M_{66}$. We choose some of the parameters such that the $M_{15}$ and $M_{16}$ terms become zero and check that we have enough freedom to consistently keep the other eigenvalues of the matrix to be positive. The smallest eigenvalue in this case for the lightest CP-even Higgs boson
is given by:
\footnotesize
\begin{equation}
M^2_{h_{tree}} =\frac{g_R^2 \left(v_1^2-v_2^2\right)^2+g_V^2 \left( v_L^2-\overline v_L ^2 \right)^2+g_L^2 \left(v_1^2-v_2^2-v_L^2+\overline{v}_L^2\right)^2+8 \lambda ^2 \left(v_1^2 v_L^2 +v_2^2
\overline{v}_L^2 \right) }{2 \left(v_1^2+v_2^2+v_L^2+\overline{v}_L^2\right)}.
\end{equation}
\normalsize
We define $ v_1 = v \sin \beta \cos \phi$, $ v_2= v \cos \beta \sin \psi$, $v_L = v \cos \beta \cos \psi$, $\overline v_L = v \sin \beta \sin \phi$ and $ g_R=g_L$. Maximizing this expression with respect to $\phi$ and $\psi$ gives the Higgs boson mass including the one and two loop corrections from the top and stop sector as:
\begin{eqnarray}
M_{h_{max}}^2 &=& \left(2 M_W^2 \sin^4 \beta +\frac{M_W^4}{2 M_W^2 - M_Z^2} \cos^4 \beta -\frac{M_W^2}{2} \sin^2 2\beta+ \lambda^2 v^2 \sin^2 2 \beta\right) \Delta_1 \notag \\
&+& \Delta_2,
\end{eqnarray}
where $\Delta_1$ and $\Delta_2$ are defined in Eq.~(\ref{eq:rad}). The coefficient of the $\Delta_1$ term is the tree-level lightest Higgs boson mass.
\begin{figure}[h!]\centering
\includegraphics[width=3.1in]{2d_tb.jpg}
\includegraphics[width=3.1in]{2d_ms.jpg}
\caption{{\sl(a) Variation of Higgs boson mass with $\tan \beta$, (b) Higgs boson mass as a function of $M_S$ in the inverse seesaw model. Notation same as in Fig. \ref{fig:1d}}}
\label{fig:2d}
\end{figure}
The Higgs boson mass is plotted in Fig~\ref{fig:2d}(a) as a function of $\tan \beta$. The red region in the figure represents the band where the mass is between 124 GeV and 126 GeV. The light green region represents the area where the stop squark mixing is minimum, i.e., $X_t=0$ while the blue upper region is for maximal mixing where $X_t=6$. The red shaded region is for all values of Higgs mass greater than 126 GeV and it is overlapped by the blue and the green regions. Fig.~\ref{fig:2d}(b) represents the upper limit of the Higgs mass and as a function $M_S$. Again the red band is where the Higgs boson mass is between 124 GeV and 126 GeV, green region is for $X_t=0$, blue region represents $X_t=6$ and shaded red region is for all values of Higgs mass greater than 126 GeV which is overlapped by the green and the blue regions. The black solid line in each case represents the MSSM upper limit for the Higgs mass.
The pseudo-scalar mass-squared matrix in this case is a $4\times4$ matrix after eliminating the two Goldstone states which are absorbed by the $Z_R$ and $Z$ bosons to get mass. We choose a transformation given as
\begin{align}
& \rho_1 = \frac{v_L \overline H_L^0 + \overline v_L H_L^0}{\sqrt{v_L^2+\overline v_L^2}},~~~~ \rho_2 = \frac{v_R \overline H_R^0 + \overline v_R H_R^0}{\sqrt{v_R^2+\overline v_R^2}},~~~~ \rho_3 = \frac{v_1 v_L H_R^0+v_1 v_R H_L^0+v_L v_R \phi_1^0}{\sqrt{v_1^2 v_L^2+v_1^2 v_R^2+v_L^2 v_R^2}}, \notag \\
&\rho_4 = \frac{v_2 \overline v_L \overline H_R^0+v_2 \overline v_R \overline H_L^0+\overline v_L \overline v_R \phi_2^0}{\sqrt{v_2^2 \overline v_L^2+v_2^2 \overline v_R^2+\overline v_L^2 \overline v_R^2}}.
\end{align}
The matrix elements in the basis $(\text{Im}\rho_1, \text{Im}\rho_2,\text{Im}\rho_3,\text{Im}\rho_4)$ are given as:
\begin{eqnarray}
M_{11} &=& -\frac{(v_L^2 + \overline v_L^2) \left[B_1 \mu_1 v_L \overline v_L -
\mu_1 (v_1 \overline v_L v_R + v_2 v_L \overline v_R) \lambda +
2 \mu (\lambda v_2 v_L v_R + \lambda v_1 \overline v_L \overline v_R -B v_1 v_2)\right]}{v_L^2 \overline v_L^2}, \notag \\
M_{12} &=& -\frac{2\mu \left(-B v_1 v_2 +\lambda v_2 v_L v_R+ \lambda v_1 \overline v_L \overline v_R\right)\sqrt{(v_L^2 + \overline v_L^2)(v_R^2 +\overline v_R^2)} }{v_L \overline v_L v_R\overline v_R}, \notag \\
M_{13} &=& -\frac{(\lambda \mu_1 \overline v_L v_R + 2 B \mu v_2 - 2 \lambda \mu \overline v_L \overline v_R) \sqrt{(v_L^2+\overline v_L^2)[v_L^2 v_R^2+v_1^2( v_L^2+ v_R^2)]}}{v_L^2 \overline v_L v_R}, \notag \\
M_{14} &=& -\frac{(\lambda \mu_1 \overline v_R v_L + 2 B \mu v_1 - 2 \lambda \mu v_L v_R) \sqrt{(v_L^2+\overline v_L^2)[\overline v_L^2 \overline v_R^2+v_2^2(\overline v_L^2+ \overline v_R^2)}}{v_L \overline v_L^2 \overline v_R}, \notag \\
M_{22} &=& \frac{(v_R^2 + \overline v_R^2) \left[B_1 \mu_1 v_R \overline v_R -
\mu_1 (v_2 \overline v_L v_R + v_1 v_L \overline v_R) \lambda -
2 \mu (\lambda v_2 v_L v_R + \lambda v_1 \overline v_L \overline v_R -B v_1 v_2)\right]}{v_R^2 \overline v_R^2}, \notag \\
M_{23} &=& \frac{(\lambda \mu_1 \overline v_R v_L - 2 B \mu v_2 + 2 \lambda \mu \overline v_L \overline v_R) \sqrt{(v_R^2+\overline v_R^2)[v_L^2 v_R^2+v_1^2( v_L^2+ v_R^2)]}}{v_L \overline v_R v_R^2}, \notag \\
M_{24} &=& \frac{(\lambda \mu_1 \overline v_L v_R - 2 B \mu v_1 + 2 \lambda \mu v_L v_R) \sqrt{(v_R^2+\overline v_R^2) [\overline v_L^2 \overline v_R^2+v_2^2(\overline v_L^2+ \overline v_R^2)]}}{v_R \overline v_L \overline v_R^2}, \notag \\
M_{33} &=& \frac{[v_L^2 v_R^2 + v_1^2 (v_L^2 + v_R^2)] \left[ \lambda A_{\lambda} v_L v_R + \lambda \mu_1(\overline v_L v_R - v_L \overline v_R) + 2 B \mu v_2 -2 \lambda \mu \overline v_L \overline v_R \right]}{v_1 v_L^2 v_R^2}, \notag \\
M_{34}&=& \frac{2 \mu B \sqrt{v_L^2 v_R^2+v_1^2(v_L^2+v_R^2)}\sqrt{\overline v_L^2 \overline v_R^2+v_2^2(\overline v_L^2+\overline v_R^2)}}{v_L v_R \overline v_L \overline v_R}, \notag \\
M_{44} &=& \frac{[\overline v_L^2 \overline v_R^2 + v_2^2 (\overline v_L^2 +\overline v_R^2)] \left[ \lambda A_{\lambda} \overline v_L \overline v_R - \lambda \mu_1(\overline v_L v_R - v_L \overline v_R) + 2 B \mu v_1 -2 \lambda \mu v_L v_R \right]}{v_2 \overline v_L^2 \overline v_R^2}.
\end{eqnarray}
The charged Higgs boson mass-squared matrix is a $6\times6$ matrix of which there are two zero mass eigenstates which are the Goldstone bosons required to give mass to the $W_R$ and the $W$ bosons. The elements of the $6\times6$ matrix in the original basis are given as:
\begin{eqnarray}
M_{11}&=&m_1^2 +\mu_1^2+\frac14 \left[ g_L^2(v_1^2-v_2^2+v_L^2+\overline v_L^2)+g_V^2(v_L^2-\overline v_L^2-v_R^2+\overline v_R^2)\right]+\lambda^2(v_2^2+v_R^2),\notag \\
M_{12}&=&-B_1 \mu_1+\frac{g_L^2 v_L \overline v_L}{2},\notag \\
M_{13}&=&\lambda(\lambda \overline v_L \overline v_R-A_\lambda v_2+2 \mu v_1), \notag \\
M_{14} &=&\lambda[\lambda v_R \overline v_L+\mu_1(v_1-v_2)],\notag \\
M_{15}&=&\frac12 g_L^2 v_1 v_L+\lambda( A_\lambda v_R-\mu_1 \overline v_R-\lambda v_1 v_L),\notag \\
M_{16}&=&\frac12 g_L^2 v_2 v_L-\lambda(\mu_1 \overline v_R-\lambda v_2 v_L-2\mu v_R),\notag \\
M_{22}&=&m_3^2 +\mu_1^2+\frac14 \left[ g_L^2(-v_1^2+v_2^2+v_L^2+\overline v_L^2)+g_V^2(-v_L^2+\overline v_L^2+v_R^2-\overline v_R^2)\right]+\lambda^2(v_1^2+\overline v_R^2),\notag \\
M_{23}&=&\lambda[\lambda v_L \overline v_R-\mu_1(v_1-v_2)],\notag \\
M_{24}&=&\lambda(\lambda v_L v_R-A_\lambda v_1+2 \mu v_2), \notag \\
M_{25}&=&\frac12 g_L^2 v_1 \overline v_L-\lambda(\mu_1 v_R+\lambda v_1 \overline v_L-2\mu \overline v_R),\notag \\
M_{26}&=&\frac12 g_L^2 v_2 \overline v_L+\lambda( A_\lambda \overline v_R-\mu_1 v_R-\lambda v_2 \overline v_L),\notag \\
M_{33}&=&m_2^2 +\mu_1^2+\frac14 \left[ g_R^2(v_1^2-v_2^2+v_R^2+\overline v_R^2)+g_V^2(-v_L^2+\overline v_L^2+v_R^2-\overline v_R^2)\right]+\lambda^2(v_2^2+v_L^2),\notag \\
M_{34}&=&B_1 \mu_1+\frac{g_R^2 v_R \overline v_R}{2},\notag \\
M_{35}&=&\frac12 g_R^2 v_2 v_R+\lambda(\mu_1 \overline v_L-\lambda v_2 v_R+2\mu v_L),\notag \\
M_{36}&=&\frac12 g_R^2 v_1 v_R+\lambda(\mu_1 \overline v_L+\lambda v_1 v_R+A_\lambda v_L),\notag \\
M_{44}&=&m_4^2 +\mu_1^2+\frac14 \left[ g_R^2(-v_1^2+v_2^2+v_R^2+\overline v_R^2)+g_V^2(v_L^2-\overline v_L^2-v_R^2+\overline v_R^2)\right]+\lambda^2(v_1^2+ \overline v_L^2),\notag \\
M_{45}&=&\frac12 g_R^2 v_2 \overline v_R+\lambda(\mu_1 v_L-\lambda v_2 \overline v_R+A_\lambda \overline v_L),\notag \\
M_{46}&=&\frac12 g_R^2 v_1 \overline v_R+\lambda(\mu_1 v_L-\lambda v_1 v_R+2\mu \overline v_L),\notag \\
M_{55}&=&m_5^2 +4 \mu^2+\frac14 \left[ g_L^2(v_1^2+v_2^2+v_L^2-\overline v_L^2)+g_R^2(v_1^2+v_2^2 -v_R^2+\overline v_R^2)\right]+\lambda^2(\overline v_L^2+v_R^2),\notag \\
M_{56}&=&\frac{(g_L^2+g_R^2)v_1 v_2+4 B \mu}{2}, \notag \\
M_{66}&=&m_5^2 +4 \mu^2+\frac14 \left[ g_L^2(v_1^2+v_2^2-v_L^2+\overline v_L^2)+g_R^2(v_1^2+v_2^2 +v_R^2-\overline v_R^2)\right]+\lambda^2( v_L^2+\overline v_R^2).
\end{eqnarray}
Using the minimization conditions given in Eq.~(\ref{eq:iseesawmin}), we eliminate $m_1$, $m_2$, $m_3$, $B$, $B_1$ and $m_5$. We then numerically calculate the pseudo-scalar and charged Higgs boson masses choosing the remaining parameters such the the lightest Higgs mass is 125 GeV after radiative corrections. We choose a stop squark mass of 500 GeV and the mixng parameter $X_t$=4. The numerical values of the Higgs masses are given in Table~\ref{tab:two} for this choice of parameters.
\vspace*{0.1in}
\noindent
{\large{\bf{Chargino and Neutralino masses}}}
\vspace*{0.1in}
The chargino mass terms in this case is written as
\begin{equation}
{\mathcal{L}}_{chargino} = -\frac{1}{2}\begin{pmatrix}\widetilde{H}_R^+&\widetilde{\overline H}_L^+&\widetilde{\phi}_1^+&\widetilde{W}_R^+ & \widetilde{W}_L^+\end{pmatrix} \begin{pmatrix}\mu_1&-\lambda v_2&\lambda v_L&g_R v_R&0 \\-\lambda v_1&-\mu_1&\lambda \overline v_R&0&g_L \overline v_L \\\lambda \overline v_L& \lambda v_R&2 \mu&g_R v_1&g_L v_1 \\ g_R \overline v_R&0&g_R v_2&M_R&0\\0&g_L v_L&g_L v_2&0&M_L \end{pmatrix} \begin{pmatrix}\widetilde{\overline H}_R^- \\ \widetilde{H}_L^-\\ \widetilde{\phi}_2^- \\ \widetilde{W}_R^- \\ \widetilde{W}_L^- \end{pmatrix},
\end{equation}
and the neutralino mass matrix in the basis $\begin{pmatrix}\widetilde{H}_R^0&\widetilde{H}_L^0&\widetilde{\overline H}_R^0&\widetilde{\overline H}_L^0&\widetilde{\phi}_1^0&\widetilde{\phi}_2^0&\widetilde{B}&\widetilde{W}_{R_3}&\widetilde{W}_{L_3}\end{pmatrix}$ is given as
\begin{equation}
M_{n} = \begin{pmatrix} 0&-\lambda v_1&-\mu_1&0&-\lambda v_L&0&\frac{g_V v_R}{\sqrt{2}}&-\frac{g_R v_R}{\sqrt{2}}&0 \\ -\lambda v_1&0&0&\mu_1&-\lambda v_R&0&-\frac{g_V v_L}{\sqrt{2}}&0&\frac{g_L v_L}{\sqrt{2}} \\ -\mu_1&0&0&-\lambda v_2&0&-\lambda \overline v_L&-\frac{g_V \overline v_R}{\sqrt{2}}&\frac{g_R \overline v_R}{\sqrt{2}}&0 \\0&\mu_1&-\lambda v_2&0&0&-\lambda \overline v_R &\frac{g_V \overline v_L}{\sqrt{2}}&0&-\frac{g_L \overline v_L}{\sqrt{2}} \\-\lambda v_L &-\lambda v_R&0&0&0&-2 \mu&0&-\frac{g_R v_1}{\sqrt{2}}&-\frac{g_L v_1}{\sqrt{2}} \\ 0&0&-\lambda \overline v_L& -\lambda \overline v_R&-2 \mu&0&0&\frac{g_R v_2}{\sqrt{2}}&\frac{g_L v_2}{\sqrt{2}} \\ \frac{g_V v_R}{\sqrt{2}}&-\frac{g_V v_L}{\sqrt{2}}&-\frac{g_V \overline v_R}{\sqrt{2}}&\frac{g_V \overline v_L}{\sqrt{2}}&0&0&M_1&0&0 \\ -\frac{g_R v_R}{\sqrt{2}}&0&\frac{g_R \overline v_R}{\sqrt{2}}&0&-\frac{g_R v_1}{\sqrt{2}}&\frac{g_R v_2}{\sqrt{2}}&0&M_R&0 \\ 0&\frac{g_L v_L}{\sqrt{2}}&0&-\frac{g_L \overline v_L}{\sqrt{2}}&-\frac{g_L v_1}{\sqrt{2}}&\frac{g_L v_2}{\sqrt{2}}&0&0&M_L\end{pmatrix},
\end{equation}
The chargino and neutralino masses are given in Table.~\ref{tab:two}.
\begin{center}
\begin{table}[h!]
\centering
\begin{tabular}{|C{2.7cm}|C{2.7cm}|C{2.8cm}|C{2.8cm}|C{2.9cm}|} \hline
{\bf{Scalar Higgs boson masses}} & {\bf{Pseudo-scalar Higgs boson masses}} & {\bf{Single charged Higgs boson masses}}& \bf{Chargino masses}&{\bf{Neutralino masses}} \\ \hline
$M_{H_1}$=5.80 TeV, $M_{H_2}$=5.43 TeV, $M_{H_3}$=3.08 TeV, $M_{H_4}$=694 GeV, $M_{H_5}$=436 GeV & $M_{A_1}$=29.6 TeV, $M_{A_2}$=4.67 TeV, $M_{A_3}$=2.80 TeV, $M_{A_4}$=478 GeV & $M_{H_1^+}$=5.80 TeV, $M_{H_2^+}$=5.21 TeV, $M_{H_3^+}$=3.08 TeV, $M_{H_4^+}$=454 GeV & $M_{\widetilde \chi_1^+}$=5.80 TeV, $M_{\widetilde \chi_2^+}$=3.87 TeV, $M_{\widetilde \chi_3^+}$=2.86 TeV, $M_{\widetilde \chi_4^+}$=1.88 TeV, $M_{\widetilde \chi_5^+}$=800 GeV & $M_{\widetilde \chi_{1,2}^0}$=5.80 TeV, $M_{\widetilde \chi_3^0}$=4.31 TeV, $M_{\widetilde \chi_4^0}$=2.90 TeV, $M_{\widetilde \chi_{5,6}^0}$=2.86 TeV, $M_{\widetilde \chi_7^0}$=2.09 TeV, $M_{\widetilde \chi_8^0}$=800 GeV, $M_{\widetilde \chi_9^0}$=526 GeV \\ \hline
\end{tabular}
\caption{{Higgs boson, chargino and neutralino masses for inverse seesaw model with a sample point given as:
$\lambda$=0.36, $v_1$=165.8 GeV, $v_2$=8 GeV, $v_L$=10 GeV, $\overline v_L$=51 GeV, $v_R$=3 TeV, $\overline v_R$=4 TeV, $\mu_1$=-2.68 TeV, $\mu$=-2.8 TeV, $m_4^2$=$-700^2~\text{GeV}^2$, $A_\lambda$=700 GeV, $M_R$=800 GeV, $M_L$=800 GeV, $M_1$=400 GeV.}}
\label{tab:two}
\end{table}
\end{center}
\section{Universal Seesaw model} \label{useesaw}
\subsection{Case with a Singlet} \label{3.1}
The particle spectrum for this case is given in Eq.~(\ref{eq:2ch}) with an additional singlet Higgs field $S$. The superpotential is given as:
\begin{eqnarray}
W&=& S(i \lambda H_L^T \tau_2 \overline{H}_L + i\lambda^c H_R^T \tau_2\overline{H}_R-M^2),
\end{eqnarray}
where $\lambda^c=\lambda^*$ and $M^2$ is real from parity invariance.
The $D$-terms, $F$-terms and the soft supersymmetry breaking terms are given as:
\begin{eqnarray}
V_F&=& \left| \lambda
{\text{Tr}} [i H_L^T \tau_2 \overline H_L + i H_R^T \tau_2 \overline H_R] -
M^2 \right|^2\nonumber \\
&+& |\lambda S|^2 {\text{Tr}}[H_L^{\dagger} H_L+{\overline H_L}^{\dagger} \overline H_L+H_R^{\dagger} H_R+{\overline H_R}^\dagger \overline H_R], \\
\label{eq:2cdterm}
V_D&=&\frac{g_L^2}{8}\sum \limits_{a=1}^3 |H_L^\dagger \tau_a H_L+\overline H_L^\dagger \tau_a \overline H_L|^2\nonumber \\
&+&\frac{g_R^2}{8}\sum \limits_{a=1}^3 |H_R^\dagger \tau_a H_R+\overline H_R^\dagger \tau_a \overline H_R|^2 \nonumber \\
&+& \frac{g_V^2}{8}|-H_L^\dagger H_L+\overline H_L^\dagger \overline H_L+H_R^\dagger H_R-\overline H_R^\dagger \overline H_R|^2 , \\
V_{Soft}&=& m_3^2(H_L^{\dagger} H_L)+m_4^2({H_R}^{\dagger} H_R)+m_5^2({\overline H_L}^{\dagger} \overline H_L)+m_6^2({\overline H_R}^\dagger \overline H_R)+m_S^2 |S|^2\nonumber \\
&+& \left[ \lambda A_\lambda S( H_L^T \tau_2 \overline{H}_L + H_R^T \tau_2\overline{H}) +h.c. \right] +(\lambda C_{\lambda} M^2 S +h.c.) .
\end{eqnarray}
We choose a rotated basis which is exactly the same as in Eq.~(\ref{eq:basis1d}) with $\phi_1 \rightarrow H_L, \phi_2 \rightarrow \overline H_L,{\delta^c}^0 \rightarrow H_R,{\overline \delta^c}^0 \rightarrow \overline H_R, v_1 \rightarrow v_L, v_2 \rightarrow \overline v_L$. The minimization conditions are slightly modified form of Eq.~(\ref{eq:min1d}) and are given by:
\begin{align}
0=& v_L [4 m_3^2 + g_L^2 (-\overline v_L^2 + v_L^2) +
g_V^2 (-\overline v_L^2 + v_L^2 - v_R^2 + \overline v_R^2)]
+ 4 \lambda A_\lambda \overline v_L v_S+ 4 \lambda^2 v_L v_S^2 \nonumber \\ &+
4 \lambda \overline v_L (-M^2 +\lambda v_L \overline v_L -\lambda v_R \overline v_R), \notag \\
0=& \overline v_L [4 m_5^2 + g_L^2 (-v_L^2 + \overline v_L^2) +
g_V^2 (-v_L^2 + \overline v_L^2 + v_R^2 - \overline v_R^2)]
+ 4 \lambda A_\lambda v_L v_S +4 \lambda^2\overline v_L v_S^2 \notag \\
&+ 4 \lambda v_L (-M^2 +\lambda v_L \overline v_L -\lambda v_R \overline v_R), \nonumber \\
0=& 4 m_4^2 v_R - g_V^2 v_R (-v_L^2 + \overline v_L^2 + v_R^2 - \overline v_R^2) + g_R^2 v_R (v_R^2 -\overline v_R^2)\nonumber \\
& - 4 \lambda A_\lambda \overline v_R v_S +4 \lambda \overline v_R (M^2 - \lambda v_L \overline v_L) +4 \lambda ^2 v_R (\overline v_R^2 + v_S^2), \nonumber \\
0=& 4 m_6^2 \overline v_R + g_V^2 \overline v_R (v_L^2 - \overline v_L^2 - v_R^2 + \overline v_R^2) +
g_R^2 \overline v_R (\overline v_R^2 - v_R^2)\nonumber \\
& - 4 \lambda A_\lambda v_R v_S +4 \lambda v_R (M^2 - \lambda v_L \overline v_L) +4 \lambda ^2 \overline v_R ( v_R^2 + v_S^2), \nonumber \\
0=& 2 m_S^2 v_S +2 C_\lambda M^2 \lambda +2 \lambda A_\lambda (v_L \overline v_L-v_R \overline v_R)+\lambda^2 (v_L^2 + \overline v_L^2 + v_R^2 +
\overline v_R^2) v_S.
\label{eq:min2c}
\end{align}
Using this minimization and the basis $({\text{Re}}\rho_1,{\text{Re}}\rho_2,{\text{Re}}{H_R}^0,{\text{Re}}{\overline H_R}^0)$, the relevant mass-squared matrix elements are given by:
\begin{eqnarray} \label{eq:masmat2c}
M_{11}& =& \frac{g_L^2 (v_L^2 - \overline v_L^2)^2 + g_V^2 (v_L^2 - \overline v_L^2)^2 +
8 v_L^2 \overline v_L^2 \lambda^2}{2 (v_L^2 + \overline v_L^2)},\notag \\
M_{12}& =& \frac{ v_L \overline v_L (v_L^2 - \overline v_L^2) (g_L^2 + g_V^2 - 2 \lambda^2)}{ (v_L^2 + \overline v_L^2)},\notag \\
M_{13}& =& \frac{-g_V^2 (v_L^2 - \overline v_L^2) (v_R^2 -\overline v_R^2) -
8 \lambda^2 v_L \overline v_L v_R \overline v_R}{\sqrt{(v_L^2 + \overline v_L^2)(v_R^2 +\overline v_R^2)}}, \notag \\
M_{14}& =& \frac{ -g_V^2 (v_L^2 - \overline v_L^2)v_R \overline v_R +
2 \lambda^2 v_L \overline v_L (v_R^2 -\overline v_R^2)}{\sqrt{(v_L^2 + \overline v_L^2)(v_R^2 +\overline v_R^2)}}, \notag \\
M_{15} &=& \frac{\lambda [2 A_\lambda v_L \overline v_L +
2 (v_L^2 + \overline v_L^2) v_S \lambda]}{\sqrt{v_L^2 + \overline v_L^2}}, \notag \\
M_{55} &=& m_S^2 + (v_L^2 + \overline v_L^2 + v_R^2 +\overline v_R^2) \lambda^2.
\end{eqnarray}
The other terms in the mass matrix are given in the appendix. We choose the ratio between $\overline v_R$ and $v_R$ such that the matrix element $M_{13}$ vanishes and we choose the value of $A_ \lambda$ such that $M_{15}$ becomes zero. Then we calculate the correction from the off-diagonal elements to the lightest eigenvalue of this mass-squared matrix. In the limit where the soft-supersymmetry breaking parameter $m_6$ is significantly larger $v_R$, we can show that this correction vanishes. We use the definitions of $\tan\beta = \frac{\overline v_L}{v_L}$ and $v^2 = v_L^2+\overline v_L^2$. Including the loop corrections from the top and stop sector, the Higgs boson mass is:
\begin{eqnarray}
M_h^2 &=& \left(\frac{M_W^4}{2M_W^2-M_Z^2}\cos^22\beta+\lambda^2 v^2 \sin^2 2 \beta\right) \Delta_1 + \Delta_2,
\end{eqnarray}
where $\Delta_1$ and $\Delta_2$ are defined in Eq.~(\ref{eq:rad}). As before the coefficient of the $\Delta_1$ term is the tree-level Higgs boson mass upper limit.
The Higgs boson mass is plotted in Fig~\ref{fig:2c}(a) as a function of $\tan \beta$. The red region in the figure represents the band where the mass is between 124 GeV and 126 GeV. The light green region represents the area where the stop squark mixing is minimum, i.e., $X_t=0$ while the blue upper region is for maximal mixing where $X_t=6$. The red shaded region is for all values of Higgs mass greater than 126 GeV and it is overlapped by the blue and the green regions. Fig.~\ref{fig:2c}(b) represents the upper limit of the Higgs mass and as a function $M_S$. Again the red band is where the Higgs boson mass is between 124 GeV and 126 GeV, green region is for $X_t=0$, blue region represents $X_t=6$ and shaded red region is for all values of Higgs mass greater than 126 GeV which is overlapped by the green and the blue regions. The black solid line in each case represents the MSSM upper limit for the Higgs mass.
\begin{figure}[h!]\centering
\includegraphics[width=3.0in]{2c_tb.jpg}
\includegraphics[width=3.1in]{2c_Ms.jpg}
\caption{{\sl(a) Variation of Higgs boson mass with $\tan \beta$, (b) Higgs boson mass as a function of $M_S$ in the universal seesaw model. Notation same as in Fig. \ref{fig:1d}.}}
\label{fig:2c}
\end{figure}
To obtain the pseudo scalar mass-squared matrix , we make the following transformation
\begin{align}
\rho_1 = \frac{v_R \overline H_R^0 + \overline v_R H_R^0}{\sqrt{v_R^2+\overline v_R^2}},~~\rho_2 = \frac{v_L \overline H_L^0 + \overline v_L H_L^0}{\sqrt{v_L^2+\overline v_L^2}}, ~~g_1 = \frac{v_R H_R^0 - \overline v_R \overline H_R^0}{\sqrt{v_R^2+\overline v_R^2}},~~ g_2 = \frac{v_L H_L^0 - \overline v_L \overline H_L^0}{\sqrt{v_L^2+\overline v_L^2}}.
\end{align}
The imaginary components of $g_1$ and $g_2$ are identified as the Goldstone states, and the $3\times3$ pseudo-scalar matrix elements in the basis $(\text{Im}\rho_1, \text{Im}\rho_2,\text{Im}S)$ are given as:
\begin{eqnarray}
M_{11} &=& m_4^2+ m_6^2+\lambda^2(v_R^2+\overline v_R^2+ v_S^2), \notag \\
M_{12} &=& -\lambda^2 \sqrt{(v_R^2+\overline v_R^2)(v_L^2+\overline v_L^2)} \notag \\
M_{13} &=& \lambda A_{\lambda}\sqrt{v_R^2+\overline v_R^2} \notag \\
M_{22} &=& m_3^2+m_5^2+\lambda^2 (v_L^2+\overline v_L^2+2 v_S^2) \notag \\
M_{23} &=& -\lambda A_{\lambda}\sqrt{v_L^2+\overline v_L^2} \notag \\
M_{33} &=& m_S^2 + \lambda^2(v_L^2+\overline v_L^2+v_R^2+\overline v_R^2)
\end{eqnarray}
The charged Higgs boson matrix is obtained by identifying the Goldstone boson states to be:
\begin{equation}
g_1^+ = \frac{v_R H_R^+-\overline v_R {\overline H_R^-}^*}{\sqrt{v_R^2+\overline v_R^2}},~~~~
g_2^+ = \frac{v_L {H_L^-}^*-\overline v_L\overline H_L^+}{\sqrt{v_L^2+\overline v_L^2}},
\label{eq:useechgold}
\end{equation}
and the physical charged Higgs boson mass eigenstates as:
\begin{equation}
h_1^+ = \frac{\overline v_R H_R^++v_R {\overline H_R^-}^*}{\sqrt{v_R^2+\overline v_R^2}},~~~~
h_2^+ = \frac{\overline v_L {H_L^-}^*+ v_L\overline H_L^+}{\sqrt{v_L^2+\overline v_L^2}}.
\end{equation}
The two eigenvalues of the charged Higgs boson mass-squared matrix in this case are given by:
\begin{eqnarray}
M^2_{h^+_1} &=& m_4^2+m_6^2+\frac{1}{2} g_R^2(v_R^2+\overline v_R^2)+2 \lambda^2 v_S^2, \notag \\
M^2_{h^+_2} &=& m_3^2+m_5^2+\frac{1}{2} g_L^2(v_L^2+\overline v_L^2)+2 \lambda^2 v_S^2.
\end{eqnarray}
We again use the minimization conditions given in Eq.~(\ref{eq:min2c}) to eliminate $m_3$, $m_4$, $m_5$, $m_6$ and $C_\lambda$. We take a stop squark mass of 600 GeV and $X_t$ = 6. Using all these constraints on the aforementioned parameters and making sure that the lightest CP-even Higgs boson mass is 125 GeV, we numerically calculate the masses of the charged and pseudo-scalar Higgs boson for a sample point. The results are given in Table~\ref{tab:three} for this choice of parameters.
\vspace*{0.1in}
\noindent
{\large{\bf{Chargino and Neutralino masses}}}
\vspace*{0.1in}
The chargino mass terms in this case is written as
\begin{equation}
{\mathcal{L}}_{chargino} = -\frac{1}{2}\begin{pmatrix}\widetilde{H}_R^+&\widetilde{\overline H}_L^+&\widetilde{W}_R^+ & \widetilde{W}_L^+\end{pmatrix}\begin{pmatrix}\lambda^* v_S&0&g_R v_R&0 \\ 0&- \lambda v_S&0&g_L \overline v_L \\ g_R \overline v_R&0&M_R&0\\0&g_L v_L&0&M_L \end{pmatrix} \begin{pmatrix}\widetilde{\overline H}_R^-\\ \widetilde{H}_L^- \\ \widetilde{W}_R^- \\ \widetilde{W}_L^- \end{pmatrix},
\label{eq:2cch}
\end{equation}
and the neutralino mass matrix in the basis $\begin{pmatrix}\widetilde{H}_R^0&\widetilde{H}_L^0&\widetilde{\overline H}_R^0&\widetilde{\overline H}_L^0&\widetilde{B}&\widetilde{W}_{R_3}&\widetilde{W}_{L_3}&\widetilde{S}\end{pmatrix}$ is given as
\begin{equation}
M_{n} = \begin{pmatrix} 0&0&-\lambda^* v_S&0&\frac{g_V v_R}{\sqrt{2}}&-\frac{g_R v_R}{\sqrt{2}}&0&-\lambda^* \overline v_R \\ 0&0&0&\lambda v_S&-\frac{g_V v_L}{\sqrt{2}}&0&\frac{g_L v_L}{\sqrt{2}}&\lambda \overline v_L \\ -\lambda^* v_S &0&0&0&-\frac{g_V \overline v_R}{\sqrt{2}}&\frac{g_R \overline v_R}{\sqrt{2}}&0&-\lambda^* v_R \\ 0&\lambda v_S&0&0&\frac{g_V \overline v_L}{\sqrt{2}}&0&-\frac{g_L \overline v_L}{\sqrt{2}}&\lambda v_L \\ \frac{g_V v_R}{\sqrt{2}}&-\frac{g_V v_L}{\sqrt{2}}&-\frac{g_V \overline v_R}{\sqrt{2}}&\frac{g_V \overline v_L}{\sqrt{2}}&M_1&0&0&0 \\ -\frac{g_R v_R}{\sqrt{2}}&0&\frac{g_R \overline v_R}{\sqrt{2}}&0&0&M_R&0&0 \\ 0&\frac{g_L v_L}{\sqrt{2}}&0&-\frac{g_L \overline v_L}{\sqrt{2}}&0&0&M_L&0 \\ -\lambda^* \overline v_R&\lambda \overline v_L&-\lambda^* v_R&\lambda v_L&0&0&0&0\end{pmatrix},
\end{equation}
where $\widetilde{W}_R,\widetilde{W}_L$ and $\widetilde{B}$ are the superpartners of the right-handed gauge bosons, left-handed gauge bosons and the $U(1)_{B-L}$ gauge boson and $M_R, M_L$ and $M_1$ are their soft masses respectively as given in Eq.~{\ref{eq:gluino}. The numerical values of the masses for the chosen sample point are given in Table~{\ref{tab:three}.
\begin{center}
\begin{table}[h!]
\centering
\begin{tabular}{|C{2.7cm}|C{2.7cm}|C{2.8cm}|C{2.9cm}|C{2.9cm}|} \hline
{\bf{Scalar Higgs boson masses}} & {\bf{Pseudo-scalar Higgs boson masses}} & {\bf{Single charged Higgs boson masses}}& \bf{Chargino masses}&{\bf{Neutralino masses}} \\ \hline
$M_{H_1}$=4.53 TeV, $M_{H_2}$=2.47 TeV, $M_{H_3}$=1.84 TeV, $M_{H_4}$=636 GeV & $M_{A_1}$=4.49 TeV, $M_{A_2}$=1.92 TeV, $M_{A_3}$=636 GeV & $M_{H_1^+}$=1.98 TeV, $M_{H_2^+}$=641 GeV & $M_{\widetilde \chi_1^+}$=2.55 TeV, $M_{\widetilde \chi_2^+}$=1.47 TeV, $M_{\widetilde \chi_3^+}$=809 GeV, $M_{\widetilde \chi_4^+}$=274 GeV & $M_{\widetilde \chi_1^0}$=2.96 TeV, $M_{\widetilde \chi_2^0}$=2.13 TeV, $M_{\widetilde \chi_{3}^0}$=2.02 TeV, $M_{\widetilde \chi_4^0}$=1.85 TeV, $M_{\widetilde \chi_4^0}$=809 GeV, $M_{\widetilde \chi_6^0}$=543 GeV, $M_{\widetilde \chi_7^0}$=281 GeV, $M_{\widetilde \chi_8^0}$=266 GeV \\ \hline
\end{tabular}
\caption{{Higgs boson, chargino and neutralino masses for Universal seesaw model with a singlet Higgs boson field using a sample point with parameters given as:
$\lambda$=0.46, $v_L$=7.4 GeV, $\overline v_L$=173.85 GeV, $v_R$=3 TeV, $\overline v_R$=3.1 TeV, $M^2$=-$2.2^2~ \text{TeV}^2$, $v_S$=600 GeV, $m_S^2$=16 TeV$^2$, $A_{\lambda}$=-1 TeV, $M_R$=800 GeV, $M_L$=800 GeV, $M_1$=400 GeV.}}
\label{tab:three}
\end{table}
\end{center}
\subsection{Case without singlet}\label{3.2}
The most general superpotential involving the Higgs fields in this case is given by:
\begin{eqnarray}
W&=&i \mu_1 H_L^T \tau_2 \overline{H}_L+i \mu_2 H_R^T \tau_2\overline{H}_R.
\end{eqnarray}
The $D$-terms in the superpotential is the same as in Eq.~(\ref{eq:2cdterm}). The $F$-terms and the soft supersymmetry breaking terms in the Higgs potential are given by:
\begin{eqnarray}
V_F&=&\mu_1^2(H_L^{\dagger} H_L+\overline H_L^{\dagger} \overline H_L)+\mu_2^2(H_R^{\dagger} H_R+\overline H_R^\dagger \overline H_R), \\
V_{Soft}&=& B_1 \mu_1(i H_L^T \tau_2 \overline{H}_L+h.c.)+B_2 \mu_2(i H_R \tau_2\overline{H}_R+h.c) \nonumber \\&+& m_3^2(H_L^{\dagger} H_L)+m_4^2({\overline H_L}^{\dagger} \overline H_L)+m_5^2({H_R}^{\dagger} H_R)+m_6^2({\overline H_R}^\dagger \overline H_R).
\label{eq:pot2b}
\end{eqnarray}
The vacuum structure in this case is given as:
\begin{equation}
\left<H_L\right> = \begin{pmatrix}
v_L \\ 0 \end{pmatrix},
\left<H_R\right> = \begin{pmatrix}
0 \\ v_R \end{pmatrix},
\left<\overline H_L\right> = \begin{pmatrix}
0 \\ \overline v_L \end{pmatrix},
\left<\overline H_R\right> = \begin{pmatrix}
\overline v_R \\ 0 \end{pmatrix}.
\end{equation}
We take a rotated basis given by:
\begin{equation} \label{eq:basis2b}
\rho_1=\frac{v_L {H_L}^0 +\overline v_L {\overline {H}_L}^0}{\sqrt{v_L^2+\overline v_L^2}},~~~~~~~~~~\rho_2=\frac{\overline v_L {H_L}^0-v_L \overline H_L^0}{\sqrt{v_L^2+\overline v_L^2}}.
\end{equation}
The minimization conditions are given by:
\begin{align}
0&=2 \mu_2^2 v_R + 2 m_4^2 v_R - 2 B_2 \mu_2 \overline v_R +
\frac12 v_R[ g_R^2 (v_R^2 - \overline v_R^2) -
g_V^2 (v_L^2 -\overline v_L^2 - v_R^2 +\overline v_R^2)], \notag \\
0&=2 \mu_2^2 \overline v_R + 2 m_6^2 \overline v_R-2 B_2 \mu_2 v_R +
\frac12 \overline v_R [g_R^2 (-v_R^2 +\overline v_R^2) +
g_V^2 (v_L^2 -\overline v_L^2 - v_R^2 +\overline v_R^2)], \notag \\
0&=2 \mu_1^2 v_L + 2 m_3^2 v_L + 2 B_1 \mu_1 \overline v_L +
\frac12 v_L [g_L^2 (v_L^2 -\overline v_L^2) +
g_V^2 (v_L^2 -\overline v_L^2 - v_R^2 +\overline v_R^2)], \notag \\
0&=2 \mu_1^2\overline v_L + 2 m_5^2\overline v_L + 2 B_1 \mu_1 v_L +
\frac12 \overline v_L [g_L^2 (- v_L^2 +\overline v_L^2) +
g_V^2 (\overline v_L^2 - v_L^2 + v_R^2 -\overline v_R^2)].
\label{eq:min2b}
\end{align}
Using the potential and minimization equations, we calculate the mass-squared matrix in the basis $({\text{Re}}\rho_1,{\text{Re}}\rho_2,{\text{Re}}{H_R}^0,{\text{Re}}{\overline H_R}^0)$. We get the following matrix:
\begin{equation}
\left[
\begin{array}{cccc}
\frac{(g_L^2+g_V^2)(v_L^2-\overline v_L^2)^2}{2(v_L^2+\overline v_L^2)}&\frac{(g_L^2+g_V^2)(v_L^2-\overline v_L^2)v_L\overline v_L}{(v_L^2+\overline v_L^2)}&-\frac{g_V^2 v_R(v_L^2-\overline v_L^2)}{2\sqrt{(v_L^2+\overline v_L^2)}}&\frac{g_V^2 \overline{v}_R(v_L^2-\overline v_L^2)}{2\sqrt{(v_L^2+\overline v_L^2)}}\\
\frac{(g_L^2+g_V^2)(v_L^2-\overline v_L^2)v_L\overline v_L}{(v_L^2+\overline v_L^2)}&\frac{ 2(g_L^2+g_V^2)(v_L^2 \overline v_L^2+(m_3^2+m_5^2+\mu_1^2)(v_L^2+\overline v_L^2)}{(v1_L^2+\overline v_L^2)}&-\frac{ g_V^2 v_L \overline v_L v_R}{\sqrt{v_L^2+\overline v_L^2}}&\frac{ g_V^2 v_L \overline v_L \overline{v}_R}{\sqrt{v_L^2+\overline v_L^2}} \\
-\frac{g_V^2 v_R(v_L^2-\overline v_L^2)}{2\sqrt{(v_L^2+\overline v_L^2)}}&-\frac{ g_V^2 v_L \overline v_L v_R}{\sqrt{v_L^2+\overline v_L^2}}&\frac{(g_R^2+g_V^2)v_R^3+2B_2 \mu_2\overline{v}_R}{2v_R}&\mbox {\footnotesize $ -B_2 \mu_2-\frac{1}{2}(g_R^2+g_V^2)v_R\overline{v}_R$ }\\
\frac{g_V^2 \overline{v}_R(v_L^2-\overline v_L^2)}{2\sqrt{(v_L^2+\overline v_L^2)}}&\frac{ g_V^2 v_L \overline v_L \overline{v}_R}{\sqrt{v_L^2+\overline v_L^2}}&\mbox {\footnotesize $-B_2 \mu_2-\frac{1}{2}(g_R^2+g_V^2)v_R\overline{v}_R $}&\frac{(g_R^2+g_V^2)\overline{v}_R^3+2B_2 \mu_2v_R}{2\overline{v}_R}
\end{array}
\right].
\label{eq:masmat2b}
\end{equation}
Here we have assumed $v_R,\overline{v}_R \neq 0$ in obtaining the mass matrix. We calculate the contribution of the off-diagonal elements to the lightest eigenvalue using the seesaw formula and this gives us the result
\begin{eqnarray}
M_{h_{tree}}^2 = M_Z^2 \cos^2 2 \beta,
\end{eqnarray}
where we have also assumed that the $SU(2)_R$ gauge coupling ($g_R$) is equal to the $SU(2)_L$ gauge coupling ($g_L$), $\tan\beta = \frac{\overline v_L}{v_L}$ and $v^2 = v_L^2+\overline v_L^2$.
The CP-odd Higgs boson mass-squared matrix is a $4\times4$ matrix which has two Goldstone states same as in Sec.~\ref{3.1}. The resulting matrix after eliminating the Goldstone states is a $2\times2$ matrix whose eigenvalues are given as:
\begin{eqnarray}
M^2_{A_1} = m_4^2+m_6^2+2 \mu_2^2, ~~~~~~M^2_{A_2} = m_3^2+m_5^2+2 \mu_1^2.
\end{eqnarray}
The charged Higgs boson matrix is again a $4\times4$ matrix with two Goldstome states which are the same as in Eq.~(\ref{eq:useechgold}). The eigenvalues of the remaining $2\times2$ charged Higgs boson mass-squared matrix in this case are given by:
\begin{eqnarray}
M^2_{h^+_1} &=& m_4^2+m_6^2+\frac{1}{2} g_R^2(v_R^2+\overline v_R^2)+2 \mu_2^2, \notag \\
M^2_{h^+_2} &=& m_3^2+m_5^2+\frac{1}{2} g_L^2(v_L^2+\overline v_L^2)+2 \mu_1^2.
\end{eqnarray}
Here we use the minimization conditions given in Eq.~(\ref{eq:min2b}) to eliminate $B_1$, $B_2$, $\mu_1$ and $\mu_2$. Since the light Higgs boson mass in this case is the same as in MSSM we use a stop squark mass of 1.5 TeV and maximal mixing between the stop squarks to get a Higgs boson mass of 125 GeV. The pseudo-scalar and charged Higgs boson masses are given in Table~\ref{tab:four}.
\vspace*{0.1in}
\noindent
{\large{\bf{Chargino and Neutralino masses}}}
\vspace*{0.1in}
The chargino mass terms in this case is written as
\begin{equation}
{\mathcal{L}}_{chargino} = -\frac{1}{2}\begin{pmatrix}\widetilde{H}_R^+&\widetilde{\overline H}_L^+&\widetilde{W}_R^+ & \widetilde{W}_L^+\end{pmatrix} \begin{pmatrix}\mu_2&0&g_R v_R&0 \\ 0&- \mu_1&0&g_L \overline v_L \\ g_R \overline v_R&0&M_R&0\\0&g_L v_L&0&M_L \end{pmatrix} \begin{pmatrix}\widetilde{\overline H}_R^-\\ \widetilde{H}_L^- \\ \widetilde{W}_R^- \\ \widetilde{W}_L^- \end{pmatrix},
\label{eq:2bch}
\end{equation}
and the neutralino mass matrix in the basis $\begin{pmatrix}\widetilde{H}_R^0&\widetilde{H}_L^0&\widetilde{\overline H}_R^0&\widetilde{\overline H}_L^0&\widetilde{B}&\widetilde{W}_{R_3}&\widetilde{W}_{L_3}\end{pmatrix}$ is given as
\begin{equation}
M_{n} = \begin{pmatrix} 0&0&-\mu_2&0&\frac{g_V v_R}{\sqrt{2}}&-\frac{g_R v_R}{\sqrt{2}}&0 \\ 0&0&0&\mu_1&-\frac{g_V v_L}{\sqrt{2}}&0&\frac{g_L v_L}{\sqrt{2}} \\ -\mu_2 &0&0&0&-\frac{g_V \overline v_R}{\sqrt{2}}&\frac{g_R \overline v_R}{\sqrt{2}}&0 \\ 0&\mu_1&0&0&\frac{g_V \overline v_L}{\sqrt{2}}&0&-\frac{g_L \overline v_L}{\sqrt{2}} \\ \frac{g_V v_R}{\sqrt{2}}&-\frac{g_V v_L}{\sqrt{2}}&-\frac{g_V \overline v_R}{\sqrt{2}}&\frac{g_V \overline v_L}{\sqrt{2}}&M_1&0&0 \\ -\frac{g_R v_R}{\sqrt{2}}&0&\frac{g_R \overline v_R}{\sqrt{2}}&0&0&M_R&0 \\ 0&\frac{g_L v_L}{\sqrt{2}}&0&-\frac{g_L \overline v_L}{\sqrt{2}}&0&0&M_L\end{pmatrix},
\end{equation}
where $M_R, M_L$ and $M_1$ are given in Eq.~{\ref{eq:gluino}. The numerical values of the masses for the chosen sample point are given in Table~{\ref{tab:four}.
\begin{center}
\begin{table}[h!]
\centering
\begin{tabular}{|C{2.7cm}|C{2.7cm}|C{2.8cm}|C{2.9cm}|C{2.9cm}|} \hline
{\bf{Scalar Higgs boson masses}} & {\bf{Pseudo-scalar Higgs boson masses}} & {\bf{Single charged Higgs boson masses}}& \bf{Chargino masses}&{\bf{Neutralino masses}} \\ \hline
$M_{H_1}$=5.71 TeV, $M_{H_2}$=2.32 TeV, $M_{H_3}$=360 GeV & $M_{A_1}$=5.07 TeV, $M_{A_2}$=2.32 TeV & $M_{H_1^+}$=5.50 TeV, $M_{H_2^+}$=2.32 TeV & $M_{\widetilde \chi_1^+}$=4.57 TeV, $M_{\widetilde \chi_2^+}$=1.11 TeV, $M_{\widetilde \chi_3^+}$=792 GeV, $M_{\widetilde \chi_4^+}$=389 GeV & $M_{\widetilde \chi_1^0}$=4.98 TeV, $M_{\widetilde \chi_2^0}$=3.39 TeV, $M_{\widetilde \chi_{3}^0}$=1.11 TeV, $M_{\widetilde \chi_4^0}$=1.10 TeV, $M_{\widetilde \chi_4^0}$=948 GeV, $M_{\widetilde \chi_6^0}$=793 GeV, $M_{\widetilde \chi_7^0}$=554 GeV \\ \hline
\end{tabular}
\caption{{Higgs boson, chargino and neutralino masses for Universal seesaw model using a sample point with parameters given as:
$v_L$=10 GeV, $\overline v_L$=173.71 GeV, $v_R$=3 TeV, $\overline v_R$=3.5 TeV, $m_3^2$=4 $\text{TeV}^2$, $m_4^2$=4 $\text{TeV}^2$, $m_5^2$=-1 $\text{TeV}^2$, $m_6^2$=-1 $\text{TeV}^2$, $M_R$=800 GeV, $M_L$=800 GeV, $M_1$=400 GeV.}}
\label{tab:four}
\end{table}
\end{center}
\section{$E_6$ Inspired Left-right Supersymmetric model} \label{e6}
The Higgs spectrum for this model is discussed in Eq.~(\ref{eq:2ah}). The relevant terms in the superpotential involving the $H_L$, $H_R$ and $\Phi$ fields are given as:
\begin{equation}
W = \lambda{H_L}^T \tau_2 \Phi \tau_2 H_R + \mu {\text{Tr}}\left[\Phi \tau_2 \Phi^T \tau_2\right],
\end{equation}
where the parameter $\lambda$ and $\mu$ must be real for the superpotential to be invariant under parity transformation.
The Higgs potential consisting of the $V_F$, $V_D$ and $V_{Soft}$ terms will be given as:
\begin{eqnarray}
V_F&=&{\text{Tr}}({|\lambda H_R^T \tau_2 \Phi \tau_2|}^2+{|\lambda H_L^T \tau_2 \Phi \tau_2|}^2)+{\text{Tr}}(|\lambda H_L H_R^T+2 \mu \Phi|^2), \\
V_D&=&\frac{g_L^2}{8}\sum \limits_{a=1}^3 |H_L^\dagger \tau_a H_L+{\text{Tr}}(\Phi^\dagger \tau_a \Phi)|^2+\frac{g_R^2}{8}\sum \limits_{a=1}^3 |H_R^\dagger \tau_a H_R+{\text{Tr}}(\Phi^* \tau_a \Phi^T)|^2\nonumber \\
&+& \frac{g_V^2}{8}|H_R^\dagger H_R-H_L^\dagger H_L|^2, \\
V_{Soft}&=& m_1^2 {\text{Tr}}(\Phi^\dagger \Phi)+ \left[{B \mu \text{Tr}}(\Phi^T \tau_2 \Phi \tau_2)+h.c.\right]+m_3^2 H_L^\dagger H_L + m_4^2 H_R^\dagger H_R \nonumber \\
&+& (A_{\lambda} \lambda H_L^T\tau_2 \Phi \tau_2 H_R+h.c.).
\label{eq:one}
\end{eqnarray}
Using this potential we calculate the Higgs boson mass-squared matrix. We choose the following vacuum structure for the Higgs fields:
\begin{equation}
\left<H_L\right> = \begin{pmatrix}
v_L \\ 0 \end{pmatrix},
\left<H_R\right> = \begin{pmatrix}
0 \\v_R \end{pmatrix},
\left<\Phi\right> ={\begin{pmatrix}
0 & v_2 \\ v_1 & 0 \end{pmatrix}}
\end{equation}
To easily identify the field corresponding to the lightest eigenvalue, we take a linear combination of the ${H_L}^0,\phi_1^0$ and $\phi_2^0$ fields. We make sure that only one of the newly defined fields get a non-zero vacuum expectation value(or VEV). The field redefinition that we used is:
\begin{eqnarray}
\rho_1&=&\frac{v_L {H_L}^0+v_1 \phi_1^0 +v_2 \phi_2^0}{\sqrt{v_L^2+v_1^2+v_2^2}},
\rho_2=\frac{v_L\phi_1^0-v_1 {H_L}^0}{\sqrt{v_1^2+v_L^2}},\notag \\
\rho_3&=&\frac{v_L v_2 {H_L}^0+v_1 v_2 \phi_1^0-(v_1^2+v_L^2)\phi_2^0}{\sqrt{(v_1^2+v_2^2)(v_1^2+v_2^2+v_L^2)}}.~~~~~~~~
\end{eqnarray}
With this choice, one can verify that only the $\rho_1$ field gets a non-zero vacuum expectation value of $\sqrt{v_1^2+v_2^2+v_L^2}$. We calculate the $4\times4$ mass-squared matrix for the neutral CP-even Higgs boson in the basis $({\text{Re}}\rho_1,{\text{Re}} {H_R}^0,{\text{Re}} \rho_2,{\text{Re}} \rho_3)$. It is easy to identify the lightest mass eigenvalue in this new basis. We use the minimization condition for the potential to express the soft SUSY breaking masses and the coefficient $\mu$ in terms of the other parameters in the model. The minimization conditions and mass-squared matrix is given in Appendix.
We assume that $v_R >> v_1,v_2,v_L$ and using this assumption we can get he lightest eigenvalue of the mass-squared matrix. It turns out that we can neglect the corrections from two of the off-diagonal matrix elements as they are of order of $\sim \frac{v_1^4}{v_R^2}$. So we effectively have a $2\times2$ matrix. Diagonalizing this matrix, we get the lightest neutral CP-even Higgs mass given by:
\begin{align}
M_{h_{tree}}^2 &= [g_R^2 (v_1^2 - v_2^2)^2 + g_V^2 v_L^4 + g_L^2 (-v_1^2 + v_2^2 + v_L^2)^2 +
8 v_1^2 v_L^2 \lambda^2 \notag \\
& - (g_V^2 v_L^2 + g_R^2 (-v_1^2 + v_2^2 + v_L^2) + 4 v_1^2 \lambda^2)^2/(g_R^2 + g_V^2)]/(2 (v_1^2 + v_2^2 + v_L^2)).
\end{align}
We then choose $v_1=v \sin\beta$, $v_2= v \cos\beta \cos\phi$ and $v_L = v \cos\beta \cos\phi$. Maximizing the resulting expression with respect to $\lambda$ and $\phi$ and choosing $g_R=g_L$, we get:
\begin{equation}
M_{h_{tree}}^2 =2 M_W^2 \cos^2 2\beta.
\end{equation}
This result is exactly the same as in {{Section~\ref{heavysinglet}}} and has been discussed in details in that section.
The pseudo-scalar Higgs boson mass-squared matrix is obtained by eliminating the Goldstone states and choosing the following basis:
\begin{equation}
\rho_1 = \frac{v_1 v_L H_R^0+v_1 v_R H_L^0+v_L v_R \phi_1^0}{\sqrt{v_1^2v_L^2+v_1^2v_R^2+v_L^2 v_R^2}},~~~~\rho_2 =\frac{v_2 \phi_1^0+v_1 \phi_2^0}{\sqrt{v_1^2+v_2^2}}.
\end{equation}
The $2\times2$ mass-squared matrix in the basis (Im$\rho_1$,Im$\rho_2$) can be written as
\begin{equation}
\begin{pmatrix}
\frac{\lambda(A_\lambda v_1-2\mu v_2)[v_L^2v_R^2+v_1^2(v_L^2+v_R^2)]}{v_1^2v_Lv_R}&\frac{2\lambda \mu \sqrt{(v_1^2+v_2^2)[v_L^2v_R^2+v_1^2(v_L^2+v_R^2)]}}{v_1^2}\\\frac{2\lambda \mu \sqrt{(v_1^2+v_2^2)[v_L^2v_R^2+v_1^2(v_L^2+v_R^2)]}}{v_1^2}&\frac{2(v_1^2+v_2^2)(\mu_2^2v_1-\lambda \mu v_L v_R)}{v_1^2 v_2}
\end{pmatrix}
\end{equation}
The charged Higgs boson mass-squared matrix is a $4\times4$ matrix of which there are two Goldstone states. In the original basis of (${H_L^-}^*,H_R^+,\phi_1^+,{\phi_2^-}^*$) the mass-squared matrix is given as:
\begin{eqnarray}
M_{11}&=&m_3^2 +\frac14 \left[ g_L^2(v_1^2-v_2^2+v_L^2)+g_V^2(v_L^2-v_R^2)\right]+\lambda^2(v_2^2+v_R^2),\notag \\
M_{12}&=&\lambda(-A_\lambda v_2+2 \mu v_1), \notag \\
M_{13}&=&\frac12 g_L^2 v_1 v_L+\lambda( A_\lambda v_R-\lambda v_1 v_L),\notag \\
M_{14}&=&\frac12 g_L^2 v_2 v_L-\lambda(\lambda v_2 v_L-2\mu v_R),\notag \\
M_{22}&=&m_3^2 +\frac14 \left[ g_R^2(v_1^2-v_2^2+v_R^2)+g_V^2(-v_L^2+v_R^2)\right]+\lambda^2(v_2^2+v_L^2),\notag \\
M_{23}&=&\frac12 g_R^2 v_2 v_R+\lambda(-\lambda v_2 v_R+2\mu v_L),\notag \\
M_{24}&=&\frac12 g_R^2 v_1 v_R-\lambda(\lambda v_1 v_R-A_\lambda v_L),\notag \\
M_{33}&=&m_1^2 +\frac14 \left[g_L^2(v_1^2+v_2^2+v_L^2)+g_R^2(v_1^2+v_2^2-v_R^2)\right]+\lambda^2 v_R^2 +4 \mu^2,\notag \\
M_{34}&=&\frac{(g_L^2+g_R^2)v_1 v_2+4 B \mu}{2}, \notag \\
M_{44}&=&m_1^2 +4 \mu^2+\frac14 \left[ g_L^2(v_1^2+v_2^2-v_L^2)+g_R^2(v_1^2+v_2^2 +v_R^2)\right]+\lambda^2 v_L^2.
\end{eqnarray}
We use the minimization conditions given in Eq.~(\ref{eq:e6min}) to eliminate $B, m_1, m_3$ and $m_4$. To get the correct CP-even lightest Higgs boson mass of 125 GeV, we choose a stop squark mass of 600 GeV and $X_t$ = 1. The numerical values of the masses of the Higgs boson physical states are given in Table~\ref{tab:five}.
\vspace*{0.1in}
\noindent
{\large{\bf{Chargino and Neutralino masses}}}
\vspace*{0.1in}
The higgsinos and the gauginos mix to form the charginos and the neutralinos. The chargino mass term in this case is written as
\begin{equation}
{\mathcal{L}}_{chargino} = -\frac{1}{2}\begin{pmatrix}\widetilde{H}_R^+&\widetilde{\phi}_1^+&\widetilde{W}_R^+ & \widetilde{W}_L^+\end{pmatrix} \begin{pmatrix}-\lambda v_2&\lambda v_L&g_R v_R&0 \\ \lambda v_R&2 \mu&g_R v_1&g_L v_1 \\ 0&g_R v_2&M_R&0\\gamma_L v_L&g_L v_2&0&M_L \end{pmatrix} \begin{pmatrix}\widetilde{H}_L^-\\ \widetilde{\phi}_2^- \\ \widetilde{W}_R^- \\ \widetilde{W}_L^- \end{pmatrix},
\end{equation}
and the neutralino mass matrix in the basis $\begin{pmatrix}\widetilde{H}_R^0&\widetilde{H}_L^0&\widetilde{\phi}_1^0&\widetilde{\phi}_2^0&\widetilde{B}&\widetilde{W}_{R_3}&\widetilde{W}_{L_3}\end{pmatrix}$ is given as
\begin{equation}
M_{n} = \begin{pmatrix} 0&-\lambda v_1&-\lambda v_L&0&\frac{g_V v_R}{\sqrt{2}}&-\frac{g_R v_R}{\sqrt{2}}&0 \\ -\lambda v_1&0&-\lambda v_R&0&-\frac{g_V v_L}{\sqrt{2}}&0&\frac{g_L v_L}{\sqrt{2}} \\ -\lambda v_L &-\lambda v_R&0&-2 \mu&0&-\frac{g_R v_1}{\sqrt{2}}&-\frac{g_L v_1}{\sqrt{2}} \\ 0&0&-2 \mu&0&0&\frac{g_R v_2}{\sqrt{2}}&\frac{g_L v_2}{\sqrt{2}} \\ \frac{g_V v_R}{\sqrt{2}}&-\frac{g_V v_L}{\sqrt{2}}&0&0&M_1&0&0 \\ -\frac{g_R v_R}{\sqrt{2}}&0&-\frac{g_R v_1}{\sqrt{2}}&\frac{g_R v_2}{\sqrt{2}}&0&M_R&0 \\ 0&\frac{g_L v_L}{\sqrt{2}}&-\frac{g_L v_1}{\sqrt{2}}&\frac{g_L v_2}{\sqrt{2}}&0&0&M_L\end{pmatrix}.
\end{equation}
where $M_R, M_L$ and $M_1$ are given in Eq.~{\ref{eq:gluino}. The numerical values of the masses for the chosen sample point are given in Table~{\ref{tab:five}.
\begin{center}
\begin{table}[h!]
\centering
\begin{tabular}{|C{2.7cm}|C{2.7cm}|C{2.8cm}|C{2.9cm}|C{2.9cm}|} \hline
{\bf{Scalar Higgs boson masses}} & {\bf{Pseudo-scalar Higgs boson masses}} & {\bf{Single charged Higgs boson masses}}& \bf{Chargino masses}&{\bf{Neutralino masses}} \\ \hline
$M_{H_1}$=3.04 TeV, $M_{H_2}$=1.72 TeV, $M_{H_3}$=890 GeV & $M_{A_1}$=3.05 TeV, $M_{A_2}$=888 GeV & $M_{H_1^+}$=3.04 TeV, $M_{H_2^+}$=898 GeV & $M_{\widetilde \chi_1^+}$=2.22 TeV, $M_{\widetilde \chi_2^+}$=2.09 TeV, $M_{\widetilde \chi_3^+}$=799 GeV, $M_{\widetilde \chi_4^+}$=2.53 GeV & $M_{\widetilde \chi_1^0}$=2.21 TeV, $M_{\widetilde \chi_2^0}$=2.20 TeV, $M_{\widetilde \chi_{3}^0}$=2.00 TeV, $M_{\widetilde \chi_4^0}$=1.63 TeV, $M_{\widetilde \chi_4^0}$=799 GeV, $M_{\widetilde \chi_6^0}$=24.3 GeV, $M_{\widetilde \chi_7^0}$=4.96 GeV \\ \hline
\end{tabular}
\caption{{Higgs boson, chargino and neutralino masses for Universal seesaw model using parameters given as:
$\lambda$=0.3, $v_L$=20 GeV, $v_1$=172.5 GeV, $v_2$=11 GeV, $v_R$=3 TeV, $\mu$=-1 TeV, $A_{\lambda}$=1 TeV, $M_R$=-800 GeV, $M_L$=-800 GeV, $M_1$=400 GeV.}}
\label{tab:five}
\end{table}
\end{center}
\section{Doubly-charged Higgs boson mass from loop corrections } \label{2chhiggs}
In the models discussed under section~\ref{case1}, the $SU(2)_R$ symmetry breaking is achieved by triplet Higgs bosons. Each triplet Higgs boson has a doubly-charged particle which should be relatively easy to detect experimentally if they can be produced at the colliders. These doubly-charged particles, if seen, can tell us a lot about the symmetry breaking pattern and their properties can help identify the underlying model. It turns out that in the minimal models, the doubly charged
Higgs boson remains light with a mass below a TeV regardless of the scale of $SU(2)_R$ breaking. This arises owing to an enhanced symmetry of the tree-level Higgs potential of the model. In this section we present a complete calculation of
the one-loop induced mass of this scalar proportional to its Majorana Yukawa coupling. This completes the calculation
initiated in Ref. \cite{bm0}.
We focus on a realistic left-right supersymmetric model where the $SU(2)_R \times U(1)_{B-L}$ symmetry is broken into $U(1)_Y$ by triplet Higgs boson field $\Delta^c$, and then the $SU(2)_L \times U(1)_Y$ symmetry breaking is achieved via bidoublet field $\Phi$. The chiral matter sector of this model is given in Eq.~(\ref{eq:matter}).
The Higgs boson sector is given in Eq.~(\ref{eq:triphig}). A singlet field $S$ is introduced so that the $SU(2)_R \times U(1)_{B_L}$ symmetry breaking can be achieved in the supersymmetric limit.
The superpotential of the model is given as:
\begin{eqnarray}
W&=& Y_{u} Q^{T} \tau_{2}\Phi_{1}\tau_{2}Q^{c} + Y_{d} Q^{T} \tau_{2}\Phi_{2}\tau_{2}Q^{c} +Y_{\nu} L^{T} \tau_{2}\Phi_{1}\tau_{2}L^{c} +Y_{l} L^{T} \tau_{2}\Phi_{2}\tau_{2}L^{c} \nonumber\\
&+&i(\frac{f}{2}^{*} L^{T}\tau_{2} \Delta L+\frac{f}{2} L^{c^{T}}\tau_{2}\Delta^{c}L^{c} ) \nonumber\\
&+&S[Tr(\lambda^{*} \Delta \overline{\Delta}+\lambda \Delta^{c}\overline{\Delta}^{c}) + \lambda_{ab}^{'} Tr(\Phi_{a}^{T}\tau_{2}\Phi_{b}\tau_{2}) -M_{R}^{2}] + W'
\label{LRSsup}
\end{eqnarray}
where
\begin{equation}
W'=\left[M_{\Delta} Tr(\Delta\overline{\Delta})+M_{\Delta}^* Tr(\Delta^c\overline{\Delta}^c)\right]+\mu_{ab} Tr\left(\Phi^T_a\tau_2 \Phi_b\tau_2\right)+M_{S}S^2+\lambda_{S}S^3.
\end{equation}
\noindent Here $Y_{u,d}$ and $Y_{\nu,l}$ are the Yukawa couplings for quarks and leptons respectively and $f$ is the Majorana neutrino Yukawa coupling matrix. This is the most general superpotential. $R$-parity is automatically preserved in this case. Putting $W'=0$ gives an enhanced $U(1)$ $R$-symmetry in the theory. Under this $R$-symmetry, $Q,Q^C,L,L^C$ fields have a charge of +1, $S$ has charge +2 and all other fields have charge zero with $W$ carrying a charge $+2$. Putting $W'=0$ also helps in understanding the $\mu$-problem. The doubly-charged left-handed and right-handed Higgsinos would be degenerate in mass in this case.
We will study the case where $W^{\prime}=0$. The left-handed triplets do not get any VEV and hence the masses of their doubly-charged particles are heavy. Thus we will concentrate on the right-handed Higgs boson triplet sector from here on. The Higgs potential consists of $F$ term, $D$ term and soft supersymmetry breaking terms which in this case are then given as
\begin{eqnarray}
V_F &=& \left|\lambda {\rm Tr}(\Delta^c \overline{\Delta}^c) +
\lambda'_{ab} {\rm Tr} \left( \Phi_a^T \tau_2 \Phi_b \tau_2
\right)-{\cal M}_R^2\right|^2 + |\lambda|^2 |S|^2 \left|{\rm
Tr}(\Delta^c \Delta^{c \dagger}) + {\rm
Tr}(\overline{\Delta}^c~\overline{\Delta}^{c \dagger}) \right|
\nonumber \\
V_{\rm soft} &=& M_1^2 {\rm Tr} (\Delta^{c \dagger} \Delta^c) +
M_2^2 {\rm Tr} (\overline{\Delta}^{c \dagger} \overline{\Delta}^c) +M_S^2 |S|^2\nonumber \\
&+& \{A_\lambda \lambda S {\rm Tr}(\Delta^c \Delta^{c \dagger}) -
C_\lambda {\cal M}_R^2 S + h.c.\} \nonumber \\
V_D &=& {g_R^2 \over
8}\sum_{a}\left|{\rm Tr}(2 \Delta^{c \dagger} \tau_a \Delta^c +
2 \overline{\Delta}^{c \dagger} \tau_a \overline{\Delta}^c + \Phi_a \tau_a^T \Phi_a^\dagger)\right|^2 \nonumber \\
&+& {g'^2 \over 8} \left|{\rm Tr}(2 \Delta^{c \dagger} \Delta^c +
2 \overline{\Delta}^{c \dagger} \overline{\Delta}^c )\right|^2~.
\end{eqnarray}
If we consider a charged breaking vacuum structure for the $\Delta^c$ and $\overline \Delta^c$ fields given as
\begin{equation}
\left<\Delta^c\right> = \begin{pmatrix}
0 & v_R \\ v_R & 0 \end{pmatrix},~~~~
\left<\overline\Delta^c\right> = \begin{pmatrix}
0&\overline v_R\\\overline{v}_R & 0 \end{pmatrix},
\end{equation}
it can be shown that the Higgs potential is lower compared to the charge conserving vacuum given in Eq.~(\ref{eq:vev1a}) \cite{km}. The $F$ term and the soft SUSY breaking terms will be the same for both vacuua whereas the $D$ term of the potential will vanish for the charged breaking vacuum while being positive definite for the charge conserving one. This would lead to a charge breaking vacuum to be the stable one which is not desirable. This is also the root cause for the doubly charged scalar of the model
receiving negative squared mass, which is unacceptable. The solution to these problems lies in the calculation of the loop correction to the Higgs potential which can make the mass of the doubly charged field positive and at the same time reverse
the roles of charge breaking and charge conserving vacuua.
The tree-level doubly-charged Higgs mass-squared matrix in the basis $(\delta^{c^{--^*}},\overline \delta^{c^{++}})$ is given as
\begin{eqnarray}
M^2_{\delta^{++}} = \begin{pmatrix} -2 g_R^2(|v_R|^2-|\overline{v}_R|^2)-{\overline{v}_R^*
\over v_R} Y^* & Y \\ Y^* & 2 g_R^2(|v_R|^2 - |\overline{v}_R|^2) - {v_R \over \overline{v}_R^*}Y \end{pmatrix}
\end{eqnarray}
where $Y = \lambda A_\lambda S +\left| \lambda \right| ^2 \left( v_R \overline v_R - \frac{M^2_R}{\lambda} \right)$ and the electroweak vev has been neglected. It can be easily seen that if the gauge couplings are neglected, then this matrix will have a massless mode. Thus in this limit, the loop corrections to this massless mode should remain finite \cite{Weinberg:1973ua}. We proceed to compute the one-loop corrections to the would-be Goldstone boson mass arising from its Majorana Yukawa couplings.
We first identify the eigenstate corresponding to the Goldstone state. It is given as
\begin{equation}
G^{++} = \frac{v_R^* \delta^{c^{--^*}}+\overline v_R \overline \delta^{c^{++}}}{\sqrt{v_R^2+\overline v_R^2}}.
\end{equation}
The couplings that we would need to consider include the direct coupling of the doubly-charged particles to the electron and selectron fields, doubly-charged Higgs coupling to the neutral Higgs triplet and singlet Higgs bosons and the coupling of these neutral fields to the neutrino and sneutrino fields. These are the fields that appear in one-loop diagrams that
induce a finite mass for $G^{++}$. We also need to calculate the masses of each of these particles.
We assume that the Majorana coupling $f$ of Eq. (\ref{LRSsup}) is significant for one generation of leptons. This coupling involves
an almost massless electron, a heavy right-handed neutrino, two degenerate selectrons and two sneutrinos. If we denote
\begin{equation}
\widetilde \nu^c = \frac{n_1+i n_2}{\sqrt{2}},~~~~\widetilde \nu^{c^*} = \frac{n_1-i n_2}{\sqrt{2}},
\end{equation}
then the masses of all the particles are then given as
\begin{align}
& M_{e^c} \approx 0,~~~~ M^2_{\widetilde{e}_{1,2}^c}= m_{L^c}^2,~~~~ M_{\nu ^c} = f v_R,\notag \\
&M^2_{n_{1,2}} = m_{L^c}^2 + \left[ f^2 v_R^2 \pm (f \lambda \overline v_R v_S+f A_f v_R) \right]
\end{align}
where $m^2_{L^c}$ is the soft mass for the sleptons and $A_f$ is the trilinear coupling associated with the Majorana
Yukawa coupling $f$.
The neutral Higgs sector relevant for our calculation would include the $\delta^{c^0}, \overline \delta^{c^0}$ and $S$ fields.
Let us write them as
\begin{equation}
\delta^{c^0} = \frac{X_1+i Y_1}{\sqrt{2}},~~~~ \overline \delta^{c^0} = \frac{X_2+i Y_2}{\sqrt{2}},~~~~ S = \frac{X_3+i Y_3}{\sqrt{2}}.
\end{equation}
If we choose all the couplings and the VEVs to be real, then we will get two $3\times 3$ mass-squared matrices for these fields-- one for the real part and another for the imaginary part. We only need to consider the real fields as the imaginary fields will have no relevant cubic couplings to the Goldstone field $G^{++}$. The relevant interaction terms in the Lagrangian which would be necessary for our calculation are given as
\begin{eqnarray}
-\mathcal{L}_{int} &=& G^{++} G^{--} \left[ (|\tilde{e}^c_1|^2+|\tilde{e}^c_2|^2) \frac{f^2 v_R^2}{v_R^2+\overline v_R^2} + \sqrt{2} \frac{\lambda ^2 v_R \overline v_R^2}{v_R^2+\overline v_R^2} X_1+\sqrt{2} \frac{\lambda ^2 v_R^2 \overline v_R}{v_R^2+\overline v_R^2} X_2 \right. \notag \\
&+& \left.\sqrt{2}\left(\lambda ^2 v_S + \frac{\lambda A_\lambda v_R \overline v_R}{v_R^2+\overline v_R^2} \right) X_3 \right]-\left[\frac{f A_f v_R+f \lambda \overline v_R v_S}{2 \sqrt{v_R^2+\overline v_R^2}}(\tilde{e}^c_1 \tilde{e}^c_1+\tilde{e}^c_2\tilde{e}^c_2)G^{--}\right]\notag \\
&+&\left[ \frac{f A_f}{2 \sqrt{2}}({n_1^2}-{ n_2^2})+\frac{f^2 v_R}{\sqrt{2}}({ n_1^2}+{ n_2^2})\right]X_1 \notag \\&+&\frac{f \lambda v_S}{2 \sqrt{2}}(n_1^2-n_2^2)X_2 + \frac{f \lambda \overline v_R}{2 \sqrt{2}}(n_1^2-n_2^2)X_3.
\end{eqnarray}
The mass-squared matrix for the neutral scalar Higgs bosons is given as
\begin{equation}
M_H^2 = \begin{pmatrix}
M_1^2+\lambda^2(v_S^2+\overline v_R^2)&\lambda^2 v_R \overline v_R+\lambda A_\lambda v_S - \lambda^2 M_R^2&2 \lambda^2 v_S v_R+\lambda A_\lambda\overline v_R \\
\lambda^2 v_R \overline v_R+\lambda A_\lambda v_S - \lambda^2 M_R^2&M_2^2 +\lambda^2 (v_S^2+v_R^2)&2 \lambda^2 v_S \overline v_R+\lambda A_\lambda v_R \\
2 \lambda^2 v_S v_R+\lambda A_\lambda\overline v_R & 2 \lambda^2 v_S \overline v_R+\lambda A_\lambda v_R &M_S^2+\lambda^2(v_R^2+\overline v_R^2)\\
\end{pmatrix}.
\end{equation}
Usually one would need to diagonalize this mass-squared matrix and identify the mass eigenstates. Fortunately that is not the case here. Let us choose a basis given as
\begin{equation}
\hat{X} = V^T X
\end{equation}
where $X = \begin{pmatrix}X_1 & X_2 & X_3 \end{pmatrix}^T$, $V$ is an orthogonal transformation matrix and $\hat{X}$ represents the mass eigenbasis. Then the diagonal mass-squared matrix is given as
\begin{equation}
D^2 = V^T M_H^2 V.
\end{equation}
All the couplings of the neutral Higgs bosons can now be written as
\begin{equation}
-\mathcal{L}_{\hat{X}} = P_i V_{ij}\hat{X}_j G^{++}G^{--} + Q_i V_{ij}\hat{X}_j n_1^2 + R_i V_{ij}\hat{X}_j n_2^2 + T_i V_{ij}\hat{X}_j \nu^c \nu^c
\end{equation}
where $P,Q,R$ and $T$ are vectors given as
\begin{eqnarray}
P&=&\begin{bmatrix} \sqrt{2} \frac{\lambda ^2 v_R \overline v_R^2}{v_R^2+\overline v_R^2},&\sqrt{2} \frac{\lambda ^2 v_R^2 \overline v_R}{v_R^2+\overline v_R^2},&\sqrt{2}\left(\lambda ^2 v_S + \frac{\lambda A_\lambda v_R \overline v_R}{v_R^2+\overline v_R^2}\right)\end{bmatrix},\notag \\
Q&=&\begin{bmatrix} \frac{f A_f}{2 \sqrt{2}}+\frac{f^2 v_R}{\sqrt{2}},&\frac{f \lambda v_S}{2 \sqrt{2}},&\frac{f \lambda \overline v_R}{2 \sqrt{2}}\end{bmatrix},\notag \\
R&=&\begin{bmatrix} \frac{-f A_f}{2 \sqrt{2}}+\frac{f^2 v_R}{\sqrt{2}},&-\frac{f \lambda v_S}{2 \sqrt{2}},&-\frac{f \lambda \overline v_R}{2 \sqrt{2}}\end{bmatrix},\notag \\
T&=&\begin{bmatrix}\frac{f}{\sqrt{2}},&0,&0\end{bmatrix}.
\end{eqnarray}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=1.6in]{dc15.pdf}
\includegraphics[width=1.6in]{dc11.pdf}
\end{center}
\caption{One-loop Feynman diagrams inducing finite mass for doubly charged Higgs boson with sneutrino and neutrino exchange.}
\label{fig:dc2}
\end{figure}
We can now calculate the one-loop corrections to the doubly-charged Higgs boson mass. The corrections coming from the right-handed neutrino and sneutrino sector are given by the Feynman diagrams in Fig.~{\ref{fig:dc2}}.
The corresponding amplitudes are given as
\begin{eqnarray}
M_1 &=& -\frac{i}{2}\left[ P^T M_h^{-2}Q \int \frac{d^4k}{(2\pi)^4} \frac{1}{k^2-m_{n_1}^2} + P^T M_h^{-2}R \int \frac{d^4k}{(2\pi)^4} \frac{1}{k^2-m_{n_2}^2} \right], \notag \\
M_2 &=&2 i M_{\nu^c} P^T M_h^{-2} T \int \frac{d^4k}{(2\pi)^4} \text{Tr}\left( \frac{\slashed{k}+M_{\nu^c}}{k^2 -M_{\nu^c}^2} \right).
\end{eqnarray}
\begin{figure}[h!]
\includegraphics[width=2.1in]{dc12.pdf}
\includegraphics[width=2.1in]{dc13.pdf}
\includegraphics[width=1.3in]{dc14.pdf}
\caption{Feynman diagrams for electron and selectron one-loop correction}
\label{fig:dc2a}
\end{figure}
The Feynman diagrams for the electron and selectron corrections are given in Fig.~\ref{fig:dc2a} and the corresponding amplitudes are given as
\begin{eqnarray}
M_3 &=& -\frac{i}{2}\frac{(f A_f v_R+f \lambda \overline v_R v_S)^2}{v_R^2+\overline v_R^2} \int \frac{d^4k}{(2\pi)^4} \frac{1}{k^2-m_{\tilde e^c}^2}, \notag \\
M_4 &=&-\frac{i f^2 v_R^2}{v_R^2+\overline v_R^2} \int \frac{d^4k}{(2\pi)^4} \frac{1}{k^2},\notag \\
M_5 &=& \frac{i f^2 v_R^2}{v_R^2+\overline v_R^2} \int \frac{d^4k}{(2\pi)^4} \frac{1}{k^2-m_{\tilde e^c}^2}
\end{eqnarray}
Summing over all the correction to the doubly-charged Higgs boson mass coming from these diagrams we get
\begin{eqnarray}
\Delta M_{G^{++}}^2&=& \frac{1}{16 \pi^2 \left( v_R^2 + \overline v_R^2 \right)} \left[ f^2 v_R^2 m^2_{\tilde{e}^c} \text{ln} \left( \frac{m^2_{\tilde{e}^c}}{M^2_{\nu^c}}\right) + \frac{f^2}{2} \left( \lambda \overline v_R v_S+A_f v_R \right)^2 \text{ln} \left(\frac{m^2_{\tilde{e}^c}}{M^2_{\nu^c}} +1 \right) \right. \notag \\
&-&\frac{f}{4} \left( A_f v_R+2 f v_R^2+\lambda \overline v_R v_S \right) m^2_{n_1} \text{ln} \left( \frac{m^2_{n_1}}{M^2_{\nu^c}} \right) \notag \\
&-&\left. \frac{f}{4} \left(- A_f v_R+2 f v_R^2-\lambda \overline v_R v_S \right) m^2_{n_2} \text{ln} \left( \frac{m^2_{n_2}}{M^2_{\nu^c}} \right) \right].
\label{eq:corr2}
\end{eqnarray}
A nontrivial check of the calculation is finiteness of the sum, although individual diagrams diverge.
It is interesting to see what happens to the mass as $v_R, \overline v_R \gg M_{SUSY}$ is taken. In this limit Eq. (\ref{eq:corr2})
reduces to
\begin{equation}
\Delta M_{G^{++}}^2 \simeq \frac{f^2}{16\pi^2}\frac{v_R^2}{v_R^2+\overline v_R^2}\left[m^2_{\tilde{e}^c}+\frac{1}{2}
\frac{(\lambda \overline v_R v_S + A_f v_R)^2} {v_R^2 + \overline v_R^2} \right]\left( {\rm ln} \frac{m^2_{\tilde{e}^c}}{M_{\nu^c}^2}-1\right)~.
\end{equation}
Since $m^2_{\tilde{e}^c}$ is of order SUSY breaking scale and $M_{\nu^c}^2$ is of order $v_R^2$, in this limit we
see that the loop correction to the doubly charged mass is negative! This suggests that the $SU(2)_R$ breaking
scale cannot be much above the SUSY breaking scale for consistency. When the two scales are comparable, the loop
correction can make the doubly charged Higgs boson squared mass to be positive for various choice of parameters.
One would expect the mass to be below a TeV, owing to the suppression factor $f^2/(16 \pi^2)$.
\vskip0.2in
\section{Discussion and Conclusion}
\label{sec:conclusion}
In this paper we have carried out a systematic investigation of the Higgs boson spectra in a variety of
supersymmetric left-right models. We have focussed on the lightest CP even Higgs boson mass and found its
theoretical upper limit at tree level deviates significantly from $M_h < M_Z$ of MSSM. Several variations
relax this limit to $M_h < \sqrt{2} m_W$, while other variations make it even weaker. Our analysis focussed on
two basic classes of models, one which uses Higgs triplets to break $SU(2)_R$ gauge symmetry, and the other
which uses doublet for this purpose. In the latter case additional fermion fields are needed in order to
generate realistic fermion masses. We studied models with inverse seesaw for neutrino masses, universal seesaw
model for fermion masses, and an $E_6$ inspired left-right model. The Higgs sectors of these models were analyzed
with or without a gauge singlet Higgs field present. The relaxed limit on $M_h$ suggests that large supersymmetric
contribution from the top--stop sector is not required, and fine tuning may be minimized compared to MSSM.
In the model with $SU(2)_R$ triplet Higgs fields, a doubly charged scalar remains light below a TeV, regardless
of the scale of $SU(2)_R$ breaking. We have computed one-loop corrections to its mass arising from Majorana Yukawa
couplings. For these corrections to be positive, the $SU(2)_R$ breaking scale should be not much above the
SUSY breaking scale.
SUSYLR models with Higgs triplet fields can only be extrapolated to energy scales of order $10^{12}$ GeV, at which
point some new dynamics should appear. On the other hand, models with Higgs doublets and bidoublets can be extrapolated all the way to the GUT scale.
\section*{Acknowledgements}
This research is supported in part by the National Science Foundation under Grant No. NSF PHY11-25915 (KSB)
and in part by the US Department of Energy Grant No. de-sc0010108.
AP was supported in part by the the NRF grant funded by Korea government of the MEST (No.2014R1A1A2057665).
\newpage
\begin{appendix}
\begin{center}
{\bf APPENDIX}
\end{center}
Here we summarize various expressions that were relevant for the Higgs boson mass spectra
in variations of the SUSYLR models.
{\bf{Minimization conditions and scalar Higgs boson mass-squared matrix}}
{\underline{$E_6$ Inspired LRSUSY model}}
The minimization conditions for the potential are given as:
\begin{eqnarray}
0&=& 2 m_1^2 v_1 +
\frac{1}{8}(-4 g_L^2 v_1 (-v_1^2 + v_2^2 + v_L^2) -
4 g_R^2 v_1 (-v_1^2 + v_2^2 + v_R^2)) - 2 \lambda A_\lambda v_L v_R +
2 v_1 (v_L^2 + v_R^2) \lambda^2 \notag \\
&-& 4 B v_2 \mu + 8 v_1 \mu^2, \notag \\
0&=& 2 m_3^2 v_L +
\frac12 v_L \left[ g_L^2 (-v_1^2 + v_2^2 + v_L^2) + g_V^2 (v_L^2 - v_R^2)\right] +
2 \lambda (\lambda v_1^2 v_L + \lambda v_L v_R^2 + 2\mu v_2 v_R - A_\lambda v_1 v_R), \notag \\
0&=&2 m_4^2 v_R +
\frac12 v_R \left[g_R^2 (-v_1^2 + v_2^2 + v_R^2) + g_V^2 (-v_L^2 + v_R^2)\right] +
2 \lambda( \lambda v_1^2 v_R + v_L^2 v_R \lambda + 2 \mu v_2 v_L-A_\lambda v_1 v_L), \notag \\
0&=&2 m_1^2 v_2 +
\frac12 v_2 (g_L^2 (-v_1^2 + v_2^2 + v_L^2) + g_R^2 (-v_1^2 + v_2^2 + v_R^2))+ 4 \mu (\lambda v_L v_R + 2 \mu v_2 -B v_1)
\label{eq:e6min}
\end{eqnarray}
The mass-squared matrix elements $M_{ij}(=M_{ji})$ in this case in the basis $({\text{Re}}\rho_1,{\text{Re}} {H_R}^0,{\text{Re}} \rho_2,{\text{Re}} \rho_3)$ is given by:
\begin{eqnarray}
M_{11}&=&\frac{g_R^2(v_1^2-v_2^2)^2+g_V^2v_L^4+g_L^2(v_1^2-v_2^2-v_L^2)^2+8\lambda^2 v_1^2v_L^2}{2(v_1^2+v_2^2+v_L^2)},\nonumber \\
M_{12}&=&\frac{g_R^2v_R(v_2^2-v_1^2)-g_V^2v_Rv_L^2+4\lambda\left\{-A_{\lambda}v_1v_L+\lambda v_R(v_1^2+v_L^2)+2\mu v_2 v_L\right\}}{2\sqrt{v_1^2+v_2^2+v_L^2}},\nonumber \\
M_{13}&=&\frac{v_1v_L\left[g_R^2(v_2^2-v_1^2)+g_V^2v_L^2+2g_L^2(v_2^2-v_1^2+v_L^2)+4\lambda^2(v_1^2-v_L^2)\right]}{2\sqrt{(v_1^2+v_2^2+v_L^2)(v_1^2+v_L^2)}},\nonumber \\
M_{14}&=&\frac{v_2\left[g_V^2v_L^4+2g_L^2v_1^2(v_1^2-v_2^2-v_L^2)+g_R^2(v_1^2-v_2^2)(2v_1^2+v_L^2)+8\lambda^2v_1^2v_L^2\right]}{2(v_1^2+v_2^2+v_L^2)\sqrt{(v_1^2+v_L^2)}},\nonumber \\
M_{22}&=&\frac{g_R^2v_R^3+g_V^2v_R^3+2\lambda A_{\lambda}v_1v_L-4\lambda \mu v_2 v_L}{2v_R},\nonumber \\
M_{23}&=&\frac{(g_R^2-g_V^2)v_1v_Lv_R+2\lambda\left[A_\lambda(v_L^2-v_1^2)+2\mu v_1v_2\right]}{2 \sqrt{v_1^2+v_L^2}},\nonumber \\
M_{24}&=&\frac{v_2v_R\left[-2g_R^2v_1^2-(g_R^2+g_V^2)v_L^2+4\lambda^2(v_1^2+v_L^2)\right]-4\lambda v_L\left[A_\lambda v_1v_2+\mu(v_1^2-v_2^2+v_L^2)\right]}{2\sqrt{(v_1^2+v_2^2+v_L^2)(v_1^2+v_L^2)}},\nonumber \\
M_{33}&=&\frac{(4g_L^2+g_R^2+g_V^2)v_1^3v_L^3+4B \mu v_2v_L^3+2\lambda A_\lambda v_R(v_1^2+v_L^2)^2-8\lambda^2v_1^3v_L^3-4\lambda \mu v_1^3v_2v_R}{2v_1v_L(v_1^2+v_L^2)},\nonumber \\
M_{34}&=&[-4m_2^2v_L(v_1^2+v_2^2+v_L^2)+v_1v_2v_L (4g_L^2v_1^2-g_V^2v_L^2+g_R^2(2v_1^2+v_L^2))+4\lambda^2 v_2v_L(v_L^2-v_1^2),\nonumber \\
&+& 4\lambda \mu v_R(v_1^2+v_2^2+v_L^2)]/ \left[2(v_1^2+v_L^2)\sqrt{v_1^2+v_2^2+v_L^2)}\right] \nonumber \\
M_{44}&=& \left[-g_R^2 \left( v_1^6 + 10 v_1^4 v_2^2 - v_1^2 v_2^4 - 2 v_1^4 v_L^2 + 10 v_1^2 v_2^2 v_L^2 - v_1^2 v_L^4 + 3 v_2^2 v_L^4+ v_1^4 v_R^2 - v_1^2 v_2^2 v_R^2 +
2 v_1^2 v_L^2 v_R^2+ v_L^4 v_R^2 \right)\right. \notag \\
&+& g_L^2 \left\{ -v_1^6 + v_1^4 (10 v_2^2 - v_L^2) + v_L^2 (v_2^2 + v_L^2)^2 +
v_1^2 (-v_2^4 + v_L^4)\right\} +g_V^2 (3 v_2^2 v_L^4 - v_2^2 v_L^2 v_R^2)+4 M_3^2 v_2^2 v_L^2 \notag \\
&+& 4 M_1^2 \left\{ v_1^4 + v_L^4 + v_1^2 (v_2^2 + 2 v_L^2)\right\} - 8 \lambda A_\lambda v_1 v_2^2 v_L v_R+ 4 \lambda^2( 6 v_1^2 v_2^2 v_L^2 + v_1^2 v_2^2 v_R^2 + v_2^2 v_L^2 v_R^2)
\notag \\
&+& \left. 16 \mu v_2 (v_1^2 + v_L^2) (B v_1 - \lambda v_L v_R ) +
16 \mu^2 \left\{ v_1^4 + v_L^4 + v_1^2 (v_2^2 + 2 v_L^2)\right\} \right]
\Big{/}\left[8 (v_1^2 + v_L^2) (v_1^2 + v_2^2 + v_L^2)\right].
\end{eqnarray}
{\underline{Case with a pair of triplets and two bidoublets}}
The minimization conditions for this case are given as:
\begin{eqnarray}
0&=&4 m_{12}^2 v_{{d_1}} + 4 {B_{12}\mu_{12}} v_{{d_2}} + 4 m_{11}^2 v_{{u_1}} + 8 {B_{11} \mu_{11}} v_{{u_2}} +
g_{L}^2 v_{{u_1}} (v_{{d_1}}^2 - v_{{d_2}}^2 + v_{{u_1}}^2 - v_{{u_2}}^2) \notag \\
&+& g_{R}^2 v_{{u_1}} (v_{{d_1}}^2 - v_{{d_2}}^2 + 2 {v_R}^2 - 2 \overline{v}_R^2 + v_{{u_1}}^2 - v_{{u_2}}^2) +
4 \left[ v_{{u_1}} (\mu_{11}^2 + \mu_{12}^2) + v_{{d_1}} \mu_{12} (\mu_{11} + \mu_{22}) \right],\notag \\
0&=&4 {B_{12}\mu_{12}} v_{{d_1}} + 4 m_{12}^2 v_{{d_2}} + 8 {B_{11}\mu_{11}} v_{{u_1}} + 4 m_{11}^2 v_{{u_2}} +
g_{L}^2 v_{{u_2}} (-v_{{d_1}}^2 + v_{{d_2}}^2 - v_{{u_1}}^2 + v_{{u_2}}^2) \notag \\
&+&
g_{R}^2 v_{{u_2}} (-v_{{d_1}}^2 + v_{{d_2}}^2 - 2 {v_R}^2 + 2 \overline{v}_R^2 - v_{{u_1}}^2 + v_{{u_2}}^2) +
4 \left[ v_{{u_2}} (\mu_{11}^2 + \mu_{12}^2) + v_{{d_2}} \mu_{12} (\mu_{11} + \mu_{22})\right],\notag \\
0&=& 4 m_{22}^2 v_{{d_1}} + 8 {B_{22} \mu_{22}} v_{{d_2}} + 4 m_{12}^2 v_{{u_1}} + 4 {B_{12} \mu_{12}} v_{{u_2}} +
g_{L}^2 v_{{d_1}} (v_{{d_1}}^2 - v_{{d_2}}^2 + v_{{u_1}}^2 - v_{{u_2}}^2)\notag \\
& +&
g_{R}^2 v_{{d_1}} (v_{{d_1}}^2 - v_{{d_2}}^2 + 2 {v_R}^2 - 2 \overline{v}_R^2 + v_{{u_1}}^2 - v_{{u_2}}^2 +
4 \left[v_{{u_1}} \mu_{12} (\mu_{11} + \mu_{22}) + v_{{d_1}} (\mu_{12}^2 + \mu_{22}^2)\right],\notag \\
0&=& 8 {B_{22}\mu_{22}} v_{{d_1}} + 4 m_{22}^2 v_{{d_2}} + 4 {B_{12} \mu_{12}} v_{{u_1}} + 4 m_{12}^2 v_{{u_2}} +
g_{L}^2 v_{{d_2}} (-v_{{d_1}}^2 + v_{{d_2}}^2 - v_{{u_1}}^2 + v_{{u_2}}^2)\notag \\
& + &
g_{R}^2 v_{{d_2}} (-v_{{d_1}}^2 + v_{{d_2}}^2 - 2 {v_R}^2 + 2 \overline{v}_R^2 - v_{{u_1}}^2 + v_{{u_2}}^2) +
4 \left[ v_{{u_2}} \mu_{12} (\mu_{11} + \mu_{22}) + v_{{d_2}} (\mu_{12}^2 + \mu_{22}^2))\right],\notag \\
0&=& 2 B_2 \mu_2 \overline{v}_R +
{v_R} (2 m_5^2 + 2 \mu_2^2 + 2 g_{V}^2 ({v_R}^2 - {\overline{v}_R}^2) +
g_{R}^2 (v_{{d_1}}^2 - v_{{d_2}}^2 + 2 {v_R}^2 - 2 \overline{v}_R^2 + v_{{u_1}}^2 - v_{{u_2}}^2),\notag \\
0&=& 2 B_2 \mu_2 {v_R} +
\overline{v}_R (2 m_6^2 + 2 \mu_2^2 + 2 g_{V}^2 (\overline{v}_R^2-{v_R}^2) +
g_R^2 (v_{{d_2}}^2 - v_{{d_1}}^2 - 2 {v_R}^2 + 2 \overline{v}_R^2 - v_{{u_1}}^2 + v_{{u_2}}^2).~~~~~~~~~~
\end{eqnarray}
{\underline{Universal seesaw model with a singlet}}
The mass-squared matrix elements are given by:
\begin{eqnarray}
M_{11} &=& \frac{g_L^2 \left( {v_L}^2- {\overline{v}_L}^2\right)^2+ { {g_V} }^2 \left( {v_L}^2- {\overline{v}_L}^2\right)^2+8 {v_L}^2 {\overline{v}_L}^2 \lambda ^2}{2 \left( {v_L}^2+ {\overline{v}_L}^2\right)}, \nonumber \\
M_{12} &=& \frac{ {v_L} {\overline{v}_L} \left( {v_L}^2- {\overline{v}_L}^2\right) \left( { g_L}^2+ { {g_V} }^2-2 \lambda ^2\right)}{ {v_L}^2+ {\overline{v}_L}^2}, \nonumber \\
{M_{13}} &=& \frac{- { {g_V} }^2 \left( {v_L}^2- {\overline{v}_L}^2\right) \left( {v_R}^2- {\overline{v}_R}^2\right)-8 {v_L} {\overline{v}_L} {v_R}
{\overline{v}_R} \lambda ^2}{2 \sqrt{ {v_L}^2+ {\overline{v}_L}^2} \sqrt{ {v_R}^2+ {\overline{v}_R}^2}} , \nonumber \\
{M_{14}} &=& \frac{ { {g_V} }^2 \left(- {v_L}^2+ {\overline{v}_L}^2\right) {v_R} {\overline{v}_R}+2 {v_L} {\overline{v}_L} \left( {v_R}^2- {\overline{v}_R}^2\right)
\lambda ^2}{\sqrt{ {v_L}^2+ {\overline{v}_L}^2} \sqrt{ {v_R}^2+ {\overline{v}_R}^2}}, \nonumber \\
{M_{15}} &=& \frac{2 \lambda \left( { A_{\lambda} } {v_L} {\overline{v}_L}+\left( {v_L}^2+ {\overline{v}_L}^2\right) {v_S} \lambda \right)}{\sqrt{ {v_L}^2+ {\overline{v}_L}^2}}, \nonumber \\
{M_{22}} &=& 2 [{ g_L}^2 {v_L}^2 {\overline{v}_L}^2+2 { {g_V} }^2 {v_L}^2 {\overline{v}_L}^2+ {m_3}^2 \left( {v_L}^2+ {\overline{v}_L}^2\right)+ {m_5}^2
\left( {v_L}^2+ {\overline{v}_L}^2\right)\nonumber \\
&+& {v_L}^4 \lambda ^2-2 {v_L}^2 {\overline{v}_L}^2 \lambda ^2+ {\overline{v}_L}^4 \lambda ^2+2 {v_L}^2 {v_S}^2
\lambda ^2+2 {\overline{v}_L}^2 {v_S}^2 \lambda ^2]/\left( {v_L}^2+ {\overline{v}_L}^2\right), \nonumber \\
{M_{23}} &=& \frac{ { {g_V} }^2 {v_L} {\overline{v}_L} \left(- {v_R}^2+ {\overline{v}_R}^2\right)+2 \left( {v_L}^2- {\overline{v}_L}^2\right) {v_R}
{\overline{v}_R} \lambda ^2}{\sqrt{ {v_L}^2+ {\overline{v}_L}^2} \sqrt{ {v_R}^2+ {\overline{v}_R}^2}} , \nonumber \\
{M_{24}} &=& \frac{-2 { {g_V} }^2 {v_L} {\overline{v}_L} {v_R} {\overline{v}_R}-\left( {v_L}^2- {\overline{v}_L}^2\right) \left( {v_R}^2- {\overline{v}_R}^2\right)
\lambda ^2}{\sqrt{ {v_L}^2+ {\overline{v}_L}^2} \sqrt{ {v_R}^2+ {\overline{v}_R}^2}}, \nonumber \\
{M_{25}} &=& -\frac{ { A_{\lambda} } \left( {v_L}^2- {\overline{v}_L}^2\right) \lambda }{\sqrt{ {v_L}^2+ {\overline{v}_L}^2}} , \nonumber \\
{M_{33}} &=& \frac{ {g_R}^2 \left( {v_R}^2- {\overline{v}_R}^2\right)^2+ { {g_V} }^2 \left( {v_R}^2- {\overline{v}_R}^2\right)^2+8 {v_R}^2
{\overline{v}_R}^2 \lambda ^2}{2 \left( {v_R}^2+ {\overline{v}_R}^2\right)}, \nonumber \\
{M_{34}} &=& \frac{ {v_R} {\overline{v}_R} \left( {v_R}^2- {\overline{v}_R}^2\right) \left( {g_R}^2+ { {g_V} }^2-2 \lambda ^2\right)}{ {v_R}^2+ {\overline{v}_R}^2}, \nonumber \\
{M_{35}} &=& \frac{\lambda \left(-2 { A_{\lambda} } {v_R} {\overline{v}_R}+2 \left( {v_R}^2+ {\overline{v}_R}^2\right) {v_S} \lambda \right)}{\sqrt{ {v_R}^2+ {\overline{v}_R}^2}} , \nonumber \\
{M_{44}} &=& [2 {g_R}^2 {v_R}^2 {\overline{v}_R}^2+2 { {g_V} }^2 {v_R}^2 {\overline{v}_R}^2+ {m_4}^2 \left( {v_R}^2+ {\overline{v}_R}^2\right)+ {m_6}^2
\left( {v_R}^2+ {\overline{v}_R}^2\right)\nonumber \\
&+& {v_R}^4 \lambda ^2-2 {v_R}^2 {\overline{v}_R}^2 \lambda ^2+ {\overline{v}_R}^4 \lambda ^2+2 {v_R}^2 {v_S}^2
\lambda ^2+2 {\overline{v}_R}^2 {v_S}^2 \lambda ^2]/\left( {v_R}^2+ {\overline{v}_R}^2\right), \nonumber \\
{M_{45}} &=& \frac{ { A_{\lambda} } \left( {v_R}^2- {\overline{v}_R}^2\right) \lambda }{\sqrt{ {v_R}^2+ {\overline{v}_R}^2}}, \nonumber \\
{M_{55}} &=& m_S^2 + (v_L^2 + \overline v_L^2 + v_R^2 +\overline v_R^2).
\end{eqnarray}
\end{appendix}
\newpage
|
\section{Introduction}
String theory contains various extended objects such as fundamental strings,
solitonic five-branes, and $\text{D} p$-branes.
These objects are known to couple to the standard background fields;
the $B$-field or the Ramond-Ramond fields.
If we consider a compactification on a seven-torus, $T^7_{3\cdots 9}$,
there arise additional objects, called \emph{exotic branes}
\cite{Obers:1998fb,Eyras:1999at,LozanoTellechea:2000mc,deBoer:2010ud,Bergshoeff:2011se,deBoer:2012ma}.
The exotic branes can exist only in the presence of compact isometry directions,
just like the Kaluza-Klein monopoles,
and have the tension proportional to $g_s^{\alpha}$ with $\alpha=-2,-3,-4$.
Among them, a $5^2_2$-brane, which has two isometry directions, has been well-studied recently
\cite{deBoer:2010ud,Bergshoeff:2011se,Kikuchi:2012za,deBoer:2012ma,Hassler:2013wsa,Geissbuhler:2013uka,Kimura:2013fda,Kimura:2013zva,Chatzistavrakidis:2013jqa,Kimura:2013khz,Andriot:2014uda,Kimura:2014upa,Okada:2014wma}.
Since the $5^2_2$ background has a non-vanishing (magnetic) \emph{$Q$-flux} \cite{Hassler:2013wsa,Geissbuhler:2013uka},
we can identify the $5^2_2$-brane as an object that magnetically couples to a bi-vector field $\beta^{ij}$
whose derivative gives the $Q$-flux.
This can be shown more explicitly by writing down
the worldvolume effective action of the $5^2_2$-brane \cite{Chatzistavrakidis:2013jqa,Kimura:2014upa}.
If we perform an $S$-duality transformation,
the $5^2_2$-brane is mapped to another exotic brane, called a $\text{D} 5_2$-brane,
which is a member of a family of exotic $p$-branes,
denoted by $\text{D} p_{7-p}$ \cite{Eyras:1999at,LozanoTellechea:2000mc}.
If we adopt a notation used in \cite{Obers:1998fb,deBoer:2012ma},
the exotic $p$-brane, which has $(7-p)$ special isometry directions,
is denoted by $p_3^{7-p}(n_1\cdots n_p,m_1\cdots m_{7-p})$
since the mass is written as
\begin{align}
M = \frac{1}{g_s^3\,l_s}\,\Bigl(\frac{R_{n_1}\cdots R_{n_p}}{l_s^p}\Bigr)\,\Bigl(\frac{R_{m_1}\cdots R_{m_{7-p}}}{l_s^{7-p}}\Bigr)^2 \quad
\bigl(\text{$R_i$: radius in the $x^i$-direction}\bigr)\,,
\end{align}
where $x^{n_i}$ are the extending directions while
$x^{m_i}$ are the special isometry directions.
They are also called the higher Kaluza-Klein branes \cite{LozanoTellechea:2000mc},
since the quadratic dependence on the radii in the isometry directions
is similar to the case of the Kaluza-Klein monopole, KK5$=5^1_2$\,.
For the special case of $p=7$, we frequently denote it by NS7 instead of $7_3$\,.
The duality relation between the standard branes and the exotic branes
is summarized in Figure \ref{fig:exotic}.
\begin{figure}[b]
\centering
\includegraphics[width=0.78\linewidth]{exotic.eps}
\caption{A family of exotic branes and the duality web.}
\label{fig:exotic}
\end{figure}
In spite of the presence of a symmetric structure between the exotic branes and the usual branes
(see Figure \ref{fig:exotic}),
little is known about the exotic branes;
e.g., the background fields which couple to the exotic branes
have not been studied in detail, other than the case of the $5^2_2$-brane.
The main interest in this paper is to identify the background fields
which couple to the exotic branes
and to write down the effective supergravity action for the background fields.
For the $5^2_2$-brane, the relevant background field is a bi-vector $\beta^{ij}$
which is a function of the standard NS-NS fields.
The effective theory for the $\beta$-field has been constructed in a series of works
\cite{Andriot:2011uh,Andriot:2012wx,Andriot:2012an,Andriot:2013xca,Andriot:2014qla}
and is called the \emph{$\beta$-supergravity}.
On the other hand, for the $\text{D} p_{7-p}$-brane,
the relevant background field is expected to be a $(7-p)$-vector $\gamma^{i_1\cdots i_{7-p}}$
whose derivative is called the non-geometric \emph{$P$-flux}
(see \cite{Aldazabal:2006up,Aldazabal:2008zza,Aldazabal:2010ef}
where the $\gamma$-fields are introduced
in the study of the exceptional generalized geometry,
\cite{Bergshoeff:2010xc,Bergshoeff:2011zk,Bergshoeff:2011se} where the relation between
mixed-symmetry tensors and exotic branes is discussed,
\cite{Chatzistavrakidis:2013jqa,Kimura:2014upa} where the effective $\text{D} 5_2$-brane action is written down and
the $\text{D} 5_2$-brane is found to couple to a bi-vector $\gamma^{ij}$ magnetically,
and \cite{Andriot:2013xca,Andriot:2014qla} where a possible relation between the polyvectors $\gamma$ and exotic branes is discussed).
However, the definition of the $\gamma$-fields and the effective action for the $\gamma$-fields are still not fully understood.
In this paper, assuming the existence of some isometry directions,
we construct effective actions for various mixed-symmetry tensors that couple to exotic branes.
We consider the cases of the exotic $5^2_2$-brane, the $1^6_4$-brane, and the $\text{D} p_{7-p}$-brane,
and argue that these exotic branes are the magnetic sources of the non-geometric fluxes
associated with polyvectors
$\beta^{ij}$, $\beta^{i_1\cdots i_6}$, and $\gamma^{i_1\cdots i_{7-p}}$, respectively.
As it is well-known, an exotic-brane background written in terms of the usual background fields
is not single-valued and has a $U$-duality monodromy.
However, with a suitable redefinition of the background fields,
the $U$-duality monodromy of the exotic-brane background
simply becomes a gauge transformation associated with a shift in a polyvector
(which corresponds to a natural extension of the \emph{$\beta$-transformation} known in the generalized geometry).
This kind of field redefinition and the rewriting of the action in terms of the new background fields are the main tasks of this paper.
We further find a new instanton solution that corresponds to the electric source
of the non-geometric flux, whose existence has been anticipated in section 7 of \cite{Bergshoeff:2011se}.
This paper is organized as follows.
In section \ref{sec:supergravity-description},
we review the supergravity description of various defect branes
with an emphasis on the \emph{exotic duality} \cite{Bergshoeff:2011se},
which relates the objects described in the upper half (exotic branes)
and the lower half (standard branes) of Figure \ref{fig:exotic}.
In section \ref{sec:beta-sugra}, we review the $\beta$-supergravity
and examine the $5^2_2$-brane as the magnetic source of the $Q$-flux.
We also find an instanton solution of the $\beta$-supergravity that has the electric charge associated with the $Q$-flux.
In section \ref{sec:S-dual-bega}, utilizing the techniques of the $\beta$-supergravity,
we derive the effective action for the bi-vector $\gamma^{ij}$ (whose derivative gives the non-geometric $P$-flux),
and find a (geometric) solution that corresponds to the $5^2_3$-brane.
In section \ref{sec:exotic-p-branes},
we derive the effective actions for various polyvectors
and show that the (reduced) action of the $\beta$-supergravity
and the action obtained in section \ref{sec:S-dual-bega} are reproduced as special cases.
We then find two kinds of solutions with either the magnetic or the electric charge associated with the non-geometric flux.
Section \ref{sec:summary} is devoted to summary and discussions.
\section{Supergravity description of defect branes}
\label{sec:supergravity-description}
In this section, we review the supergravity solutions corresponding to various defect branes,
and discuss the $\text{SL}(2,{\mathbb Z})$ duality \cite{Eyras:1999at,LozanoTellechea:2000mc}
which relates the standard branes and the exotic branes.
In the following, we basically follow the notations of \cite{Okada:2014wma}.
All defect-brane backgrounds considered in this paper can be obtained
from the following seven-brane background by performing the $T$- and $S$-dualities:
\begin{align}
\begin{split}
\underline{\text{7-brane:}}\quad & {\mathrm{d}} s^2= \rho_2^{-1/2}\, \bigl(\rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z}+{\mathrm{d}} x^2_{03\cdots 9}\bigr) \qquad
\bigl(z\equiv r \Exp{{\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}}\theta}\equiv x^1+{\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}} x^2\bigr)\,,
\\
&\Exp{2\phi} = \rho_2^{-2} \,,\quad
C^{(0)} = \rho_1 \,, \quad
C^{(8)} =-\rho_2^{-1}\,{\mathrm{d}} t\wedge {\mathrm{d}} x^3\wedge \cdots \wedge {\mathrm{d}} x^9\,.
\end{split}
\label{eq:7-branes}
\end{align}
This background satisfies the equations of motion for the type IIB supergravity
as long as the functions $\rho \equiv \rho_1+{\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}}\rho_2$
and $f$ are holomorphic functions of $z$.
In the following, we choose them as $\rho(z)={\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}} (\sigma/2\pi)\,\log(r_\text{c}/z)= (\sigma/2\pi)\,\bigl[\theta+{\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}} \log(r_\text{c}/r)\bigr]$
($r_\text{c}\,$: positive constant) and $f(z)=1$,
which makes the above background the well-known D7 background.
Here, the value of $\sigma$ depends on the duality frame (see Table \ref{tab:sigma})
and it is now given by the string coupling constant; $\sigma_{\DD7}=g_s$\,.
\begin{table}[b]
\centering
\begin{tabular}{|c||l||c||l|}\hline
$\text{D} p$ & $\sigma_{\text{D} p}=g_s\,\bigl(l_s^{7-p}/R_{m_1} \cdots R_{m_{7-p}}\bigr)$ &
$p^{7-p}_3$ & $\sigma_{p^{7-p}_3}=g_s^{-1}\,\bigl(R_{m_1} \cdots R_{m_{7-p}}/l_s^{7-p}\bigr)$
\\\hline
$\NS5$ & $\sigma_{\NS5}= l_s^2/R_{m_1}R_{m_2}$ &
$5^2_2$ & $\sigma_{5^2_2}= R_{m_1}R_{m_2}/l_s^2$
\\\hline
$\text{KK} 5$ & $\sigma_{\text{KK} 5}= R_m/R_\ell$ & \multicolumn{2}{c|}{---}
\\\hline
F1 & $\sigma_{\FF1}= g_s^2\,\bigl(l_s^6/R_{m_1}\cdots R_{m_6}\bigr)$ &
$1^6_4$ & $\sigma_{1^6_4}= g_s^{-2}\,\bigl(R_{m_1}\cdots R_{m_6}/l_s^6\bigr)$
\\\hline
P & $\sigma_{\text{P}}=g_s^2\,\bigl(l_s^8/R_{m_1}\cdots R_{m_6}R_n^2\bigr)$ &
$0^{(1,6)}_4$ & $\sigma_{0^{(1,6)}_4}=g_s^{-2}\,\bigl(R_{m_1}\cdots R_{m_6}R_n^2/l_s^8\bigr)$
\\\hline
\end{tabular}
\caption{Values of the dimensionless constant $\sigma$ in various duality frames.
The directions $x^n$, $x^{m_i}$, and $x^\ell$ are the same as those appearing in the
expressions for the background fields.}
\label{tab:sigma}
\end{table}
Note that in the solutions described below,
we can always redefine the holomorphic functions as (see \cite{Bergshoeff:2006jj})
\begin{align}
\rho(z) \to \rho'(z) = \frac{a\,\rho(z) + b}{c\,\rho(z) + d}\,,\quad
f(z)\to f'(z)=(c\,\rho + d)\,f(z)\,,\quad
\begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \text{SL}(2,{\mathbb Z})\,,
\label{eq:SL2Z}
\end{align}
which corresponds to the $\text{SL}(2,{\mathbb Z})$ symmetry in the type IIB theory.
Note also that all backgrounds considered in this paper, by construction,
have isometries in the $03\cdots 9$-directions.
The defect D$p(n_1\cdots n_p)$ background,
which corresponds to a D$p$-brane extending in the $x^{n_1},\dotsc,x^{n_p}$-directions
and smeared over the remaining directions, $x^{m_1},\dotsc,x^{m_{7-p}}$, of the seven-torus $T^7_{3\cdots 9}$ ($n_i,m_i=3,\dotsc,9$),
is given by
\begin{align}
\begin{split}
\underline{\text{ $\text{D} p$ :}}\quad &{\mathrm{d}} s^2
= \rho_2^{-1/2}\, \bigl(\rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z}+{\mathrm{d}} x^2_{0n_1\cdots n_p}\bigr) + \rho_2^{1/2}\, {\mathrm{d}} x^2_{m_1\cdots m_{7-p}} \,,
\\
&C^{(7-p)} = \rho_1\,{\mathrm{d}} x^{m_1}\wedge\cdots\wedge{\mathrm{d}} x^{m_{7-p}}\,,\quad
\Exp{2\phi} = \rho_2^{\frac{3-p}{2}} \,,
\\
&C^{(p+1)} = - \rho_2^{-1}\,\epsilon_{0n_1\cdots n_pm_{7-p}\cdots m_1}\,{\mathrm{d}} t \wedge {\mathrm{d}} x^{n_1}\wedge \cdots \wedge {\mathrm{d}} x^{n_p} \,,
\end{split}
\label{eq:Dp-brane}
\end{align}
where the totally antisymmetric symbol is given by $\epsilon_{03\cdots 9}\equiv 1$ and the indices are not summed.
On the other hand, the exotic $p$-brane background,
$\text{D} p_{7-p}(n_1\cdots n_p,m_1\cdots m_{7-p})$, is given by
\begin{align}
\begin{split}
\underline{\text{ $\text{D} p_{7-p}$ :}}\quad
&{\mathrm{d}} s^2 = \frac{\abs{\rho}}{\rho_2^{1/2}}\, \bigl(\rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z}+{\mathrm{d}} x^2_{0n_1\cdots n_p}\bigr) + \frac{\rho_2^{1/2}}{\abs{\rho}}\, {\mathrm{d}} x^2_{m_1\cdots m_{7-p}} \,,
\\
&C^{(7-p)} = -\frac{\rho_1}{\abs{\rho}^2}\,{\mathrm{d}} x^{m_1}\wedge \cdots \wedge {\mathrm{d}} x^{m_{7-p}}\,, \quad
\Exp{2\phi} = \biggl(\frac{\rho_2}{\abs{\rho}^2}\biggr)^{\frac{3-p}{2}} \,,
\\
&C^{(p+1)} = -\frac{\abs{\rho}^2}{\rho_2}\,\epsilon_{0n_1\cdots n_pm_{7-p}\cdots m_1}\,{\mathrm{d}} t \wedge {\mathrm{d}} x^{n_1}\wedge \cdots \wedge {\mathrm{d}} x^{n_p} \,.
\label{eq:KKp-brane}
\end{split}
\end{align}
As it has been noticed in \cite{Eyras:1999at,LozanoTellechea:2000mc}, this background is obtained from the D$p$-brane background
through the replacement
\begin{align}
\rho(z) \to -\rho^{-1}(z) \,,\quad \rho_2\,\abs{f}^2\to \rho_2\,\abs{f}^2\,,
\label{eq:exotic-dual}
\end{align}
which is a special case of the $\text{SL}(2,{\mathbb Z})$ transformation given in \eqref{eq:SL2Z}.
This kind of duality between a usual brane and an exotic brane is called the exotic duality \cite{Bergshoeff:2011se}.
The defect (or smeared) $\NS5(n_1\cdots n_5)$ and the $5^2_2(n_1\cdots n_5,m_1m_2)$ backgrounds are also related
to each other through the exotic-duality transformation \eqref{eq:exotic-dual}:
\begin{align}
\begin{split}
\underline{\text{ $\NS5$ :}}\quad
&{\mathrm{d}} s^2 = \rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} + {\mathrm{d}} x^2_{0n_1\cdots n_5} + \rho_2\, {\mathrm{d}} x_{m_1m_2}^2 \,, \quad
\Exp{2\phi} = \rho_2 \,,
\\
&B^{(2)}= \rho_1\,{\mathrm{d}} x^{m_1}\wedge {\mathrm{d}} x^{m_2}\,, \quad
B^{(6)}= \rho_2^{-1}\,\epsilon_{0n_1\cdots n_5m_2m_1}\,{\mathrm{d}} t\wedge {\mathrm{d}} x^{n_1}\wedge\cdots\wedge {\mathrm{d}} x^{n_5}\,,
\label{eq:NS5}
\end{split}
\\
\begin{split}
\underline{\text{ $5^2_2$ :}}\quad
&{\mathrm{d}} s^2 = \rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} +{\mathrm{d}} x^2_{0n_1\cdots n_5} + \frac{\rho_2}{\abs{\rho}^2}\,\, {\mathrm{d}} x_{m_1m_2}^2 \,, \quad
\Exp{2\phi} = \frac{\rho_2}{\abs{\rho}^2}\,,
\\
&B^{(2)} = -\frac{\rho_1}{\abs{\rho}^2}\,{\mathrm{d}} x^{m_1}\wedge {\mathrm{d}} x^{m_2} \,, \quad
B^{(6)}= \frac{\abs{\rho}^2}{\rho_2}\,\epsilon_{0n_1\cdots n_5m_2m_1}\,{\mathrm{d}} t\wedge {\mathrm{d}} x^{n_1}\wedge\cdots\wedge {\mathrm{d}} x^{n_5}\,.
\label{eq:522}
\end{split}
\end{align}
Applying a general $\text{SL}(2,{\mathbb Z})$ transformation to these five-brane backgrounds,
we can obtain the background of a \emph{defect $(p,q)$-five brane}
\cite{Kimura:2014wga,deBoer:2012ma},
which is a bound state of $p$ defect NS5-branes and $q$ $5^2_2$-branes.
Note that we can perform the $\text{SL}(2,{\mathbb Z})$ transformation even in the type IIA theory,
since, in this duality frame, the $\text{SL}(2,{\mathbb Z})$ transformation
is realized as a subgroup of the $T$-duality group.
The background of a defect KK5$(n_1\cdots n_5,m)$-brane smeared in the $x^\ell$-direction
and its exotic-dual background are given by
\begin{align}
\begin{split}
\underline{\text{ $\text{KK} 5$ :}}\quad
&{\mathrm{d}} s^2 = \rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} + {\mathrm{d}} x^2_{0n_1\cdots n_5} +\rho_2\,{\mathrm{d}} x_\ell^2 + \rho_2^{-1}\,\bigl({\mathrm{d}} x^m -\rho_1\,{\mathrm{d}} x^\ell\bigr)^2 \,,
\\
&\Exp{2\phi} =1\,,\quad B^{(2)} = 0 \,,
\end{split}
\\
\begin{split}
\underline{\text{ anti-$\KK5$ :}}\quad
&{\mathrm{d}} s^2 = \rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} + {\mathrm{d}} x^2_{0n_1\cdots n_5} +\rho_2\,{\mathrm{d}} x_m^2 + \rho_2^{-1}\,\bigl({\mathrm{d}} x^\ell +\rho_1\,{\mathrm{d}} x^m\bigr)^2 \,,
\\
&\Exp{2\phi} =1\,,\quad B^{(2)} = 0 \,.
\end{split}
\end{align}
The latter corresponds to the background of an anti-KK5$(n_1\cdots n_5,\ell)$-brane smeared in the $x^m$-direction.
Namely, under the exotic-duality transformation, the Taub-NUT direction is interchanged with the smeared direction.
The background of a bound state of these KK5-branes is also considered in \cite{Kimura:2014wga}.
Further, there are the following pairs of strings and pp-waves:
\begin{align}
\begin{split}
\underline{\text{ \FF1 :}}\quad
& {\mathrm{d}} s^2 = \rho_2^{-1}\, \bigl(\rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} +{\mathrm{d}} x_{0n}^2 \bigr)+ {\mathrm{d}} x^2_{m_1\cdots m_6} \,, \quad
\Exp{2\phi} = \rho_2^{-1} \,,
\\
&B^{(6)} = \rho_1\,{\mathrm{d}} x^{m_1}\wedge \cdots \wedge {\mathrm{d}} x^{m_{6}}\,, \quad
B^{(2)} = -\rho_2^{-1}\,\epsilon_{0nm_1\cdots m_6}\,{\mathrm{d}} t \wedge {\mathrm{d}} x^n \,,
\end{split}
\label{eq:F1-BG}
\\
\begin{split}
\underline{\text{ $1^6_4$ :}}\quad &{\mathrm{d}} s^2= \frac{\abs{\rho}^2}{\rho_2} \, \bigl(\rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} +{\mathrm{d}} x_{0n}^2 \bigr)+ {\mathrm{d}} x^2_{m_1\cdots m_6} \,, \quad
\Exp{2\phi} = \frac{\abs{\rho}^2}{\rho_2} \,,
\\
&B^{(6)}= -\frac{\rho_1}{\abs{\rho}^2}\,{\mathrm{d}} x^{m_1}\wedge \cdots \wedge {\mathrm{d}} x^{m_{6}}\,, \quad
B^{(2)} = -\frac{\abs{\rho}^2}{\rho_2}\,\epsilon_{0nm_1\cdots m_6}\,{\mathrm{d}} t \wedge {\mathrm{d}} x^n \,,
\label{eq:1^6_4-BG}
\end{split}
\\
\begin{split}
\underline{\text{ P :}}\quad & {\mathrm{d}} s^2 = -2\,{\mathrm{d}} t\,{\mathrm{d}} x^n + \rho_2\, {\mathrm{d}} x_n^2 + \abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} + {\mathrm{d}} x^2_{m_1\cdots m_6} \,,
\\
&\Exp{2\phi} = \rho_2^{-1} \,,\quad B^{(2)} = 0\,,
\end{split}
\\
\begin{split}
\underline{\text{ $0^{(1,6)}_4$ :}}\quad
& {\mathrm{d}} s^2 = -2\,{\mathrm{d}} t\,{\mathrm{d}} x^n + \frac{\rho_2}{\abs{\rho}^2}\, {\mathrm{d}} x_n^2 + \abs{\rho}^2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} + {\mathrm{d}} x^2_{m_1\cdots m_6} \,,
\\
&\Exp{2\phi} = \frac{\abs{\rho}^2}{\rho_2} \,,\quad
B^{(2)} = 0\,.
\end{split}
\end{align}
If we perform a timelike $T$-duality in the $\text{D} 0$ or the $\text{D} 0_7$ background,
we obtain the following defect D-instanton (or more precisely the defect E0-brane \cite{Hull:1998vg}) background or another instanton background,
to be called $\text{D}(-1)_8$:
\begin{align}
\begin{split}
\underline{\text{ $\text{D}(-1)$ \ [IIB$^\star$]\,:}}\quad
& {\mathrm{d}} s^2 = \Exp{\phi/2}\, \bigl(\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z}+ {\mathrm{d}} x^2_{03\cdots 9}\bigr) \,, \quad
\Exp{2\phi} = \rho_2^2 \,,
\\
&C^{(8)} = \rho_1\,{\mathrm{d}} t\wedge{\mathrm{d}} x^3\wedge\cdots\wedge{\mathrm{d}} x^9\,,\quad C^{(0)} = \rho_2^{-1} \,,
\end{split}
\\
\begin{split}
\underline{\text{ $\text{D}(-1)_8$ \ [IIB$^\star$]\,:}}\quad
& {\mathrm{d}} s^2 = \Exp{\phi/2}\, \bigl(\abs{\rho}^2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z}+ {\mathrm{d}} x^2_{03\cdots 9}\bigr) \,, \quad
\Exp{2\phi} = \frac{\rho_2^2}{\abs{\rho}^4} \,,
\\
&C^{(8)} = -\frac{\rho_1}{\abs{\rho}^2}\,{\mathrm{d}} t\wedge{\mathrm{d}} x^3\wedge\cdots\wedge{\mathrm{d}} x^9\,,\quad C^{(0)} = \frac{\abs{\rho}^2}{\rho_2} \,.
\end{split}
\end{align}
These backgrounds satisfy the equations of motion for the type IIB$^\star$ theory of \cite{Hull:1998vg}
since we have performed a timelike $T$-duality.
The corresponding backgrounds written as solutions of the (\emph{Euclideanised}) type IIB theory
are given by (note that the Euclideanisation is given by the replacements
$t\to {\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}}\tau$ and $C^{(0)}\to {\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}} C^{(0)}$, and then $C^{(8)}$ is defined by ${\mathrm{d}} C^{(8)}=-*{\mathrm{d}} C^{(0)}$)
\begin{align}
\begin{split}
\underline{\text{ $\text{D}(-1)$ \ [IIB]\,:}}\quad
& {\mathrm{d}} s^2 = \Exp{\phi/2}\, \bigl(\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z}+{\mathrm{d}}\tau^2+{\mathrm{d}} x^2_{3\cdots 9}\bigr) \,, \quad
\Exp{2\phi} = \rho_2^2 \,,
\\
&C^{(8)} = -\rho_1\,{\mathrm{d}} \tau\wedge{\mathrm{d}} x^3\wedge\cdots\wedge{\mathrm{d}} x^9\,,\quad
C^{(0)} = \rho_2^{-1} \,,
\label{eq:D-inst-IIB}
\end{split}
\\
\begin{split}
\underline{\text{ $\text{D}(-1)_8$ \ [IIB]\,:}}\quad
& {\mathrm{d}} s^2 = \Exp{\phi/2}\, \bigl(\abs{\rho}^2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z}+{\mathrm{d}}\tau^2+ {\mathrm{d}} x^2_{3\cdots 9}\bigr) \,, \quad
\Exp{2\phi} = \frac{\rho_2^2}{\abs{\rho}^4} \,,
\\
&C^{(8)} = \frac{\rho_1}{\abs{\rho}^2}\,{\mathrm{d}} \tau\wedge{\mathrm{d}} x^3\wedge\cdots\wedge{\mathrm{d}} x^9\,,\quad
C^{(0)} = \frac{\abs{\rho}^2}{\rho_2} \,.
\end{split}
\label{eq:D-inst-8}
\end{align}
Note that these objects are not related to each other by an $S$-duality,
as opposed to the case of seven branes.
Further details about instanton backgrounds are discussed in section \ref{sec:p=7}.
The uplifts of the above defect backgrounds to eleven dimensions are given in
\cite{LozanoTellechea:2000mc,Bergshoeff:2011se,deBoer:2012ma}
although we do not consider them in this paper.
\section{$\beta$-supergravity}
\label{sec:beta-sugra}
Recently, it has been pointed out in \cite{deBoer:2010ud}
that the $5^2_2$ background \eqref{eq:522} has a $T$-duality monodromy around the center.
That is, globally, we have to glue the background fields in different coordinate patches
using a $T$-duality transformation.
This kind of \emph{non-geometric background} which requires to use a $T$-duality transformation
as a transition function is called a \emph{$T$-fold}.
More generally, a non-geometric background which requires to use
a larger duality transformation in string theory, i.e.~the $U$-duality symmetry,
is called a \emph{$U$-fold} \cite{Hull:2004in}.
In the special case of $T$-folds,
we can globally describe the backgrounds within a framework of the double field theory (DFT),
which is a $T$-duality covariant reformulation of the low-energy supergravity theory
\cite{Hull:2009mi,Hull:2009zb,Hohm:2010jy,Hohm:2010pp,Hohm:2011dv,Hohm:2010xe,Jeon:2010rw,Jeon:2011cn,Jeon:2011vx,Hohm:2011nu,Jeon:2011sq,Jeon:2012kd,Hohm:2012gk,Jeon:2012hp,Berman:2013uda}.
In DFT, in addition to the usual spacetime coordinates $x^i$ ($i =0,1,\dotsc,9$),
we also introduce the ``dual'' coordinates $\tilde{x}_i$,
and treat them on an equal footing; $(x^I)\equiv (\tilde{x}_i,\,x^i)$\,.
The fundamental fields of DFT are the generalized metric
\begin{align}
(\mathcal H_{IJ}) \equiv
\begin{pmatrix} G^{-1} & -G^{-1}\,B \\ B\, G^{-1} & G-B\, G^{-1}\,B \end{pmatrix} \quad
\bigl[G\equiv (G_{ij})\,,\quad B\equiv(B_{ij})\bigr]\,,
\end{align}
and the ($T$-duality invariant) dilaton $\Exp{-2d}\equiv \sqrt{\abs{G}}\,\Exp{-2\phi}$.
According to a constraint, called the strong constraint,
these fields can depend only on the half of the coordinates.
If we choose the background fields to depend only on the usual coordinates $x^i$,
the DFT action reduces to the standard ten-dimensional action for the NS-NS fields.
In DFT, a choice of coordinates corresponds to a choice of a $T$-duality frame,
and indeed, a \emph{generalized coordinate transformation},
which is a gauge symmetry of DFT, corresponds to an $\text{O}(10,10)$ transformation.
More explicitly, under a coordinate transformation, $x^I\to x^{\prime\,I}(x^I)$,
the generalized metric (which is a \emph{generalized tensor}) transforms as \cite{Hohm:2012gk}
\begin{align}
\mathcal H'_{IJ} = \mathcal{F}_I{}^K\,\mathcal{F}_J{}^L\,\mathcal H_{KL} \,,\quad
\mathcal F_I{}^J
&\equiv \frac{1}{2}\,\Bigl(\frac{\partial x^K}{\partial x^{\prime\,I}}\frac{\partial x'_K}{\partial x_J}
+\frac{\partial x'_I}{\partial x_K}\frac{\partial x^J}{\partial x^{\prime\,K}}\Bigr) \,,
\end{align}
where indices are raised or lowered by using the $\text{O}(10,10)$-invariant metric,
$\eta = \bigl(\begin{smallmatrix} 0& \mathbf{1} \cr \mathbf{1} &0\end{smallmatrix}\bigr)$\,.
Now, let us consider a generalized coordinate transformation
\begin{align}
x^{\prime\,i} = \tilde{x}_i\,,\quad \tilde{x}'_i =x^i \,,
\label{eq:dual-map}
\end{align}
in a standard background that does not depend on the dual coordinates $\tilde{x}_i$\,.
Since $\mathcal F_I{}^J=\delta_I^J$, the functional form of the generalized metric $\mathcal H_{IJ}$ is invariant,
although the background fields become functions only of the dual coordinates $\tilde{x}_i$\,.
In the ``dual'' spacetime, spanned by the dual coordinates $\tilde{x}_i$\,,
a natural set of background fields
\footnote{For example, the string sigma model action for the dual coordinates $\tilde{X}_i(\sigma)$ can be written as (see e.g.~\cite{Hull:2004in})
\begin{align*}
S=- \frac{1}{4\pi\alpha'}\,\int{\mathrm{d}}^2\sigma\sqrt{-\eta}\,\bigl(\eta^{\alpha\beta}\,\tilde{G}^{ij}
+ \epsilon^{\alpha\beta}\,\tilde{B}^{ij}\bigr) \,\partial_\alpha \tilde{X}_i\, \partial_\beta \tilde{X}_j\,.
\end{align*}}
which we denote by $(\tilde{G}^{ij},\,\tilde{\varphi},\,\tilde{B}^{ij})$, is given by
\begin{align}
\begin{split}
\tilde{G}^{ij}&\equiv \bigl(E^{-1}\bigr)^{ik}\,\bigl(E^{-1}\bigr)^{jl}\,G_{kl} \,, \quad
E\equiv \bigl(E_{ij}\bigr)\equiv \bigl(G_{ij} + B_{ij}\bigr) \,,
\\
\tilde{B}^{ij}&\equiv \bigl(E^{-1}\bigr)^{ik}\,\bigl(E^{-1}\bigr)^{jl}\,B_{kl}\,, \quad
\Exp{-2\tilde{\varphi}}\sqrt{\abs{\tilde{G}}}\equiv \Exp{-2\phi}\sqrt{\abs{G}} \,.
\end{split}
\label{eq:tilde-T-dual}
\end{align}
These ``dual'' background fields satisfy the standard ten-dimensional equations of motion for the NS-NS fields
with the following replacements:
\begin{align}
x^i \to \tilde{x}_i\,,\quad \partial_i\to \tilde{\partial}^i\,,\quad
G_{ij}\to \tilde{G}^{ij} \,,\quad B_{ij}\to \tilde{B}^{ij}\,, \quad
\sqrt{\abs{G}}\Exp{-2\phi} \to \sqrt{\abs{\tilde{G}}}\Exp{-2\tilde{\varphi}} \,.
\end{align}
Under the generalized coordinate transformation \eqref{eq:dual-map}, the $5^2_2(34567,89)$ background
is mapped to the dual background;
\begin{align}
\begin{split}
\underline{\text{ $5^2_2$ :}}\quad
{\mathrm{d}} \tilde{s}^2 &\equiv \tilde{G}^{ij}\,{\mathrm{d}} \tilde{x}_i\,{\mathrm{d}} \tilde{x}_j
= \rho_2^{-1}\,\abs{f}^{-2}\,{\mathrm{d}} z\,{\mathrm{d}}\bar{z} + {\mathrm{d}} \tilde{x}^2_{034567} + \rho_2\, {\mathrm{d}} \tilde{x}_{89}^2 \,,
\\
\tilde{B}^{(2)} &\equiv \frac{1}{2}\,\tilde{B}^{ij}\,{\mathrm{d}} \tilde{x}_i\wedge {\mathrm{d}} \tilde{x}_j
= -\rho_1\,{\mathrm{d}} \tilde{x}_8\wedge {\mathrm{d}} \tilde{x}_9 \,,\quad
\Exp{2\tilde{\varphi}} = \abs{f}^{-2}\,\rho_2^{-1} \,.
\label{eq:522-dual}
\end{split}
\end{align}
This background has the same form with the (geometric) $\NS5(34567)$ background
in the usual ten-dimensional spacetime
(apart from the $1$-$2$ components of the metric and the dilaton).
That is, the $5^2_2$ background written in the dual fields \eqref{eq:522-dual}
is a geometric background in a sense that
the monodromy, $\rho_1\to \rho_1+\sigma$, is just a gauge transformation associated with the $\tilde{B}$-field
\footnote
See section 5.1 of \cite{Okada:2014wma} for another explanation that
the $5^2_2$ background (described as a doubled geometry) is geometric;
the monodromy is realized as a generalized diffeomorphism there.}
The redefinition of the background fields \eqref{eq:tilde-T-dual} is essential
in the formulation of the $\beta$-supergravity,
although we do not introduce the dual coordinates $\tilde{x}_i$ in the $\beta$-supergravity.
\subsection{The action and its reduction}
In the $\beta$-supergravity, the fundamental fields
$(\tilde{g}_{ij},\,\tilde{\phi},\,\tilde{\beta}^{ij})$ (which depend only on the usual ten-dimensional coordinates $x^i$)
are defined by
\begin{align}
\tilde{g}_{ij}&\equiv E_{ik}\,E_{jl}\,G^{kl} \,, \quad
\tilde{\beta}^{ij} \equiv \bigl(E^{-1}\bigr)^{ik}\,\bigl(E^{-1}\bigr)^{jl}\,B_{kl}\,,\quad
\Exp{-2\tilde{\phi}}\sqrt{\abs{\tilde{g}}}\equiv \Exp{-2\phi}\sqrt{\abs{G}}\,.
\label{eq:beta-fields}
\end{align}
Note that the metric $\tilde{g}_{ij}$ is the inverse of
the dual metric $\tilde{G}^{ij}$ introduced in \eqref{eq:tilde-T-dual},
and accordingly the dilaton $\tilde{\phi}$ is different from
$\tilde{\varphi}$ given in \eqref{eq:tilde-T-dual}.
Further, note that $\tilde{\beta}^{ij}$ behaves as a bi-vector under diffeomorphisms
(in the standard ten-dimensional spacetime)
although its functional form is the same with the ``2-form'' $\tilde{B}^{ij}$ in the dual spacetime.
In a series of papers \cite{Andriot:2011uh,Andriot:2012wx,Andriot:2012an,Andriot:2013xca},
an effective action for $(\tilde{g}_{ij},\,\tilde{\phi},\,\tilde{\beta}^{ij})$ was proposed
and it was shown that the action is equal to the usual action for the NS-NS fields (up to boundary terms)
if we rewrite these fields using the original background fields through the relation \eqref{eq:beta-fields}.
The effective action, written in a manifestly invariant form under diffeomorphisms, is given by \cite{Andriot:2013xca}
\begin{align}
\begin{split}
S\bigl[\,\tilde{g}_{ij},\,\tilde{\phi},\,\tilde{\beta}^{ij}\,\bigr]
&= \frac{1}{2\kappa_{10}^2}\,\int{\mathrm{d}}^{10}x\sqrt{\abs{\tilde{g}}}\,\Exp{-2\tilde{\phi}}\,
\Bigl(\tilde{R}+4\,\tilde{g}^{ij}\,\partial_i \tilde{\phi}\,\partial_j \tilde{\phi} + \widecheck{\mathcal R}(\tilde{g}_{ij},\,\tilde{\beta}^{ij})
\\
&\qquad\qquad\qquad\qquad\qquad\qquad - \frac{1}{2}\,\abs{R^{ijk}}^2
+ 4\,\abs{\tilde{\beta}^{ij}\,\partial_j \tilde{\phi} -\mathcal T^i(\tilde{g}_{ij},\,\tilde{\beta}^{ij})}^2\Bigr)\,,
\label{eq:beta-action}
\end{split}
\end{align}
where we defined
\begin{align}
\begin{split}
&\widecheck{\mathcal R}^{ij}(\tilde{g}_{ij},\,\tilde{\beta}^{ij})\equiv
-\tilde{\beta}^{kl}\,\partial_l\widecheck{\Gamma}^{ij}_k+\tilde{\beta}^{il}\,\partial_l\widecheck{\Gamma}^{kj}_k
+\widecheck{\Gamma}^{ij}_k\,\widecheck{\Gamma}^{lk}_l-\widecheck{\Gamma}^{li}_k\,\widecheck{\Gamma}^{kj}_l\,,
\\
&\widecheck{\mathcal R}(\tilde{g}_{ij},\,\tilde{\beta}^{ij})\equiv \tilde{g}_{ij}\,\widecheck{\mathcal R}^{ij}(\tilde{g}_{ij},\,\tilde{\beta}^{ij})\,,
\\
&\widecheck{\Gamma}^{ij}_k \equiv \frac{1}{2}\,\tilde{g}_{kl}\,\bigl(-\tilde{\beta}^{im}\,\partial_m\tilde{g}^{jl}-\tilde{\beta}^{jm}\,\partial_m\tilde{g}^{il}+\tilde{\beta}^{lm}\,\partial_m\tilde{g}^{ij}\bigr) +\tilde{g}_{kl}\,\tilde{g}^{m(i}\,\partial_m\tilde{\beta}^{j)l} -\frac{1}{2}\,\partial_k\tilde{\beta}^{ij} \,,
\\
&\mathcal T^i(\tilde{g}_{ij},\,\tilde{\beta}^{ij}) \equiv \widecheck{\Gamma}^{ki}_k = \partial_k\tilde{\beta}^{ik}-\frac{1}{2}\,\tilde{\beta}^{ij}\,\tilde{g}_{kl}\,\partial_j\tilde{g}^{kl}
= \nabla_k\tilde{\beta}^{ik}\,,
\\
&R^{ijk}\equiv 3\,\tilde{\beta}^{l[i}\partial_l\tilde{\beta}^{jk]}\,, \quad
\abs{V^{i_1\cdots i_p}}^2\equiv \frac{1}{p!}\,\tilde{g}_{i_1j_1}\cdots \tilde{g}_{i_pj_p}\, V^{i_1\cdots i_p}\,V^{j_1\cdots j_p}\,.
\end{split}
\label{eq:beta-definitions}
\end{align}
Note that we can assume that any derivative $\partial_i$ contracted with $\tilde{\beta}^{ij}$ vanishes,
as long as we consider the defect backgrounds which satisfy $\tilde{\beta}^{\mathbf{a} i}=0$ ($\mathbf{a}=1,2$) and
have isometries in the $03\cdots 9$-directions.
Under the assumption, the above action reduces to the following simple form \cite{Andriot:2011uh}:
\begin{align}
S\bigl[\,\tilde{g}_{ij},\,\tilde{\phi},\,\tilde{\beta}^{ij}\,\bigr]
= \frac{1}{2\kappa_{10}^2}\, \int \Exp{-2\tilde{\phi}}\,\Bigl(\tilde{*}\,\tilde{R} + 4\,{\mathrm{d}}\tilde{\phi}\wedge \tilde{*}\,{\mathrm{d}}\tilde{\phi}
-\frac{1}{4}\,\tilde{g}_{ik}\,\tilde{g}_{jl}\,\bm{Q}^{(1)\,ij}\wedge \tilde{*}\,\bm{Q}^{(1)\,kl} \Bigr) \,,
\label{eq:beta-action-reduced}
\end{align}
where the $Q$-flux, $\bm{Q}^{(1)\,ij}\equiv Q_k{}^{ij}\,{\mathrm{d}} x^k \equiv {\mathrm{d}} \tilde{\beta}^{ij}$,
is a mixed-symmetry tensor (i.e.~a bi-vector-valued 1-form)
\footnote{In fact, this is not a tensor as we can see from the definition.
However, if we assume $\tilde{\beta}^{\mathbf{a} i}=0$ ($\mathbf{a}=1,2$) and the existence
of isometries in the $03\cdots 9$-directions, $Q_k{}^{ij}$ transforms as a tensor
under diffeomorphisms, $x'^\mathbf{a}=x'^\mathbf{a}(x^1,x^2)$ and $x'^\mathbf{p}=\Lambda^\mathbf{p}{}_\mathbf{q}\,x^\mathbf{q}$ ($\mathbf{p},\mathbf{q}=0,3,\dotsc,9$,\ $\Lambda^\mathbf{p}{}_\mathbf{q}\in \text{GL}(8,{\mathbb R})$)\,,
which respect the assumptions.
\label{foot:tensor}}
In this case, the Bianchi identity and the equation of motion for $\tilde{\beta}^{ij}$ become
\begin{align}
{\mathrm{d}} \bm{Q}^{(1)\,ij} = 0\,,\quad {\mathrm{d}} \bm{Q}^{(9)}_{ij} =0 \,,\quad
\bm{Q}^{(9)}_{ij} \equiv \Exp{-2\tilde{\phi}}\,\tilde{g}_{ik}\,\tilde{g}_{jl}\,\tilde{*}\,\bm{Q}^{(1)\,kl} \,.
\label{eq:beta-Bianchi}
\end{align}
Assuming the isometries in the $03\cdots9$-directions
and that the background fields satisfy
$\tilde{g}_{\mathbf{a}\mathbf{p}}=0$ and $\tilde{\beta}^{\mathbf{a} i}=0$ ($\mathbf{a}=1,2$, $\mathbf{p}=0,3,\dotsc,9$),
we can show that $\bm{Q}^{(9)}_{ij}$ has the following form:
\begin{align}
\bm{Q}^{(9)}_{ij} = \bigl[\bm{q}^{1}_{ij}(x^1,x^2)\,{\mathrm{d}} x^1+\bm{q}^2_{ij}(x^1,x^2)\,{\mathrm{d}} x^2\bigr]\wedge {\mathrm{d}} t\wedge {\mathrm{d}} x^3\wedge\cdots\wedge{\mathrm{d}} x^9\,.
\end{align}
Then, using ${\mathrm{d}} \bm{Q}^{(9)}_{ij} =0$, we can always choose a gauge
in which the dual potential $\tilde{\beta}^{(8)}_{ij}$,
defined by $\bm{Q}^{(9)}_{ij}\equiv {\mathrm{d}} \tilde{\beta}^{(8)}_{ij}$, is proportional to ${\mathrm{d}} t\wedge {\mathrm{d}} x^3\wedge\cdots\wedge{\mathrm{d}} x^9$\,;
\begin{align}
\begin{split}
\tilde{\beta}^{(8)}_{ij} &= \bm{q}^0_{ij}(x^1,x^2)\,{\mathrm{d}} t\wedge {\mathrm{d}} x^3\wedge\cdots\wedge{\mathrm{d}} x^9
\\
\bigl[{\mathrm{d}} \bm{q}^0_{ij}&\equiv \bm{q}^1_{ij}(x^1,x^2)\,{\mathrm{d}} x^1+\bm{q}^2_{ij}(x^1,x^2)\,{\mathrm{d}} x^2\bigr]\,.
\label{eq:beta-8}
\end{split}
\end{align}
If we define a contraction of a mixed-symmetry tensor $\tau_{i_1\cdots i_n}^{(8)}$ by
\begin{align}
\begin{split}
&\bm{\iota} \tau_{i_1\cdots i_n}^{(8)}
\equiv \frac{1}{(8-n)!}\,\tau_{k_1\cdots k_{8-n}i_1\cdots i_n,\,i_1\cdots i_n}^{(8)}\,
{\mathrm{d}} x^{k_1}\wedge \cdots \wedge {\mathrm{d}} x^{k_{8-n}}\,,
\\
&\text{or }\quad \bigl(\bm{\iota} \tau_{i_1\cdots i_n}^{(8)}\bigr)_{k_1\cdots k_{8-n}}\equiv \tau_{k_1\cdots k_{8-n}i_1\cdots i_n,\,i_1\cdots i_n}^{(8)}\,,
\end{split}
\end{align}
equation \eqref{eq:beta-8} ensures that $\bm{\iota} \tilde{\beta}^{(8)}_{ij}$ is a (2-tensor-valued) 6-form.
Further, the relation
\begin{align}
\tilde{\beta}^{(8)}_{\mathbf{p}\mathbf{q}} = \frac{1}{2}\,\bm{\iota} \tilde{\beta}^{(8)}_{\mathbf{p}\mathbf{q}}\wedge {\mathrm{d}} x^\mathbf{p}\wedge {\mathrm{d}} x^\mathbf{q}\,,
\label{eq:contracted-beta8}
\end{align}
is shown to be satisfied.
\subsection{$5^2_2$-brane as a source of the $Q$-flux}
By using the fundamental fields of the $\beta$-supergravity,
the $5^2_2$ background can be written as
\begin{align}
\begin{split}
\underline{\text{ $5^2_2$ :}}\quad
&{\mathrm{d}} \tilde{s}^2 = \rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} + {\mathrm{d}} x^2_{034567} + \rho_2^{-1}\, {\mathrm{d}} x_{89}^2 \,, \quad
\Exp{2\tilde{\phi}} = \rho_2^{-1} \,,
\\
&\tilde{\beta}^{89}= -\rho_1 \,, \quad
\tilde{\beta}_{89}^{(8)} = \rho_2^{-1}\,{\mathrm{d}} t\wedge {\mathrm{d}} x^3\wedge \cdots \wedge {\mathrm{d}} x^9\,,
\end{split}
\label{eq:522-beta}
\end{align}
which indeed satisfies the equations of motion for the $\beta$-supergravity \cite{Hassler:2013wsa}.
This background is quite similar to the NS5-brane background \eqref{eq:NS5},
although the $8$-$9$ components of the metric and the dilaton are inverted.
Since the monodromy around the center simply becomes a gauge transformation ($\tilde{\beta}^{89}\to \tilde{\beta}^{89}-\sigma$),
called the $\beta$-transformation,
we can conclude that the $\beta$-supergravity can describe the $5^2_2$ background globally
\footnote{In fact, the background fields include a cutoff radius $r_\text{c}$ and the geometry gives a good description only for $r\ll r_\text{c}$.
However, we can smoothly extend the geometry beyond the cutoff radius by introducing additional defect branes,
which makes the total energy finite \cite{Greene:1989ya,Bergshoeff:2006jj},
and then the cutoff radius $r_\text{c}$ can be interpreted as the distance between the $5^2_2$-brane and a neighboring defect brane \cite{Kikuchi:2012za}.}
On the other hand, the NS5 background,
obtained from \eqref{eq:522-beta} with the replacements \eqref{eq:exotic-dual},
is not single-valued and non-geometric.
Namely, the $\beta$-supergravity can be considered as a reformulation of the usual supergravity
which is suitable for a global description of the $5^2_2$ background, instead of the NS5 background.
Similarly, all effective actions proposed in this paper
are suitable for a global description of an exotic-brane background, instead of a standard brane background.
Now, in order to discuss the coupling of the mixed-symmetry tensor
$\tilde{\beta}^{(8)}_{ij}$ to the $5^2_2$-brane,
we comment on the relation between the definition of $\tilde{\beta}^{ij}$ or $\tilde{\beta}^{(8)}_{ij}$
and that of the corresponding one, $\tilde{B}^{mn}$ or $\tilde{B}_{8}^{mn}$,
introduced in the study of the effective worldvolume theory of
the $5^2_2$-brane \cite{Chatzistavrakidis:2013jqa}.
In \cite{Chatzistavrakidis:2013jqa}, the $2\times 2$-matrix $\tilde{B}^{mn}$ is defined by
\begin{align}
\tilde{B}^{mn}\equiv \frac{\det B_{mn}}{\det E_{mn}}\,(B^{-1})^{mn} \quad (m,n=8,9)\,,
\end{align}
which is obtained by applying a double $T$-duality $T_{89}$ to $B_{mn}$.
On the other hand, $\tilde{\beta}^{ij}$, defined in \eqref{eq:beta-fields}, is obtained by applying the $T$-dualities in all spacetime dimensions to $B_{ij}$.
In order to compare these quantities,
we assume that the background satisfies $G_{\mu m}=B_{\mu i}=0$ ($\mu,\nu=0,\dotsc, 7$, $m,n=8,9$).
In this case, we can easily show that the $8$-$9$ components of $\tilde{\beta}^{ij}$
coincide with $\tilde{B}^{mn}$\,.
Moreover, with the same assumptions, the definition of $\tilde{B}_8^{mn}$
(see (5.14) of \cite{Chatzistavrakidis:2013jqa}) become
\footnote{Under the assumption $G_{\mu m}=B_{\mu i}=0$, $\tilde{H}^m$ and $\theta_m$ appearing in (5.14) of \cite{Chatzistavrakidis:2013jqa} vanish.
In addition, $\det(G_{np} + B_{np})$ in the same equation should be corrected as $(\det E_{mn})^2/(\det G_{mn})$\,.}
\begin{align}
{\mathrm{d}} \iota_9\iota_8\tilde{B}_{8}^{89}
= \Exp{-2\phi} \sqrt{\det G_{mn}}\, \frac{(\det E_{mn})^2}{\det G_{mn}}\, \hat{\star}\,{\mathrm{d}}\tilde{B}^{89} \,,
\end{align}
where $\hat{\star}$ is the Hodge star operator associated with the metric $G_{\mu\nu}$\,.
Using that $\tilde{g}_{ij}$ and $\Exp{-2\tilde{\phi}}$ are now given by
\begin{align}
(\tilde{g}_{ij})=\begin{pmatrix} G_{\mu\nu} & 0 \\ 0 & \frac{\det E_{mn}}{\det G_{mn}}\,G_{mn} \end{pmatrix} \,,\quad
\Exp{-2\tilde{\phi}}\sqrt{\det \tilde{g}_{mn}} = \Exp{-2\phi}\sqrt{\det G_{mn}}\,,
\end{align}
and using the isometries in the $89$-directions, we obtain
\begin{align}
{\mathrm{d}} \tilde{B}_{8}^{89}
= \Exp{-2\tilde{\phi}} \tilde{g}_{8k}\, \tilde{g}_{9l}\, \tilde{*}\,{\mathrm{d}} \tilde{B}^{kl} \,,
\end{align}
which coincides with the definition of $\tilde{\beta}^{(8)}_{89}$ given in \eqref{eq:beta-Bianchi}
with the identification $\tilde{B}^{89}=\tilde{\beta}^{89}$\,.
For a general case where $G_{\mu m}$ and $B_{\mu i}$ are no longer assumed to vanish,
definitions of $\tilde{\beta}^{ij}$ and $\tilde{B}^{mn}$ are different.
However, this is not a problem since the relation between the field $\beta^{ij}$
which magnetically couples to the exotic $5^2_2$-brane,
and the original background fields $(G_{ij},\,B_{ij})$ depends on a choice of the duality frame.
Namely, under a duality transformation
\begin{align}
\{G_{ij},\,B_{ij},\,\cdots\}
\to\bigl\{G'_{ij}(G_{ij},\,B_{ij},\,\cdots),\,B'_{ij}(G_{ij},\,B_{ij},\,\cdots),\,\cdots\bigr\}\,,
\end{align}
the relation between $\beta^{ij}$ and the original fields is changed;
$\beta^{ij}=\tilde{\beta}^{ij}(G_{ij},\,B_{ij})=\widehat{\beta}^{ij}(G'_{ij},\,B'_{ij})$\,.
In the following, we will basically use the relation
$\beta^{ij}=\tilde{\beta}^{ij}(G_{ij},\,B_{ij})$ and omit the tilde,
since with this definition, we can write down a covariant action \eqref{eq:beta-action}
for $\beta^{ij}$ with all $i$-$j$ components.
As shown in \cite{Chatzistavrakidis:2013jqa,Kimura:2014upa},
the Wess-Zumino term of the $5^2_2(34567,89)$-brane action (smeared in the isometry directions, $x^8$ and $x^9$) can be written as
\begin{align}
\begin{split}
S_{\text{WZ}}^{5^2_2}
&= -\mu_{5^2_2}\,n^{89}\,\int_{\mathcal M_6\times T^2_{89}}\bm{\iota}\beta_{89}^{(8)}\wedge \frac{{\mathrm{d}} x^8\wedge{\mathrm{d}} x^9}{(2\pi R_8)(2\pi R_9)}
= -\frac{\mu_{5^2_2}\,n^{89}}{(2\pi R_8)(2\pi R_9)}\,\int_{\mathcal M_6\times T^2_{89}} \beta_{89}^{(8)}
\\
&= -\mu_{5^2_2}\,\int \beta_{89}^{(8)}\wedge \bm{\delta}^{89}(x-X(\xi)) \quad \bigl(n^{89}\text{: number of the $5^2_2(34567,89)$-branes}\bigr)
\\
&\Bigl(\bm{\delta}^{\mathbf{p}_1\cdots\mathbf{p}_n}(x-X(\xi))
\equiv \frac{n^{\mathbf{p}_1\cdots\mathbf{p}_n}\,\delta^2(x-X(\xi))}{(2\pi R_{\mathbf{p}_1})\cdots(2\pi R_{\mathbf{p}_n})}\,{\mathrm{d}} x^1\wedge{\mathrm{d}} x^2\,,\quad n^{\mathbf{p}_1\cdots\mathbf{p}_n}\in{\mathbb Z} \Bigr)\,,
\end{split}
\label{eq:SWZ522}
\end{align}
where we used \eqref{eq:contracted-beta8} and $\mathcal M_6$ is the worldvolume of the $5^2_2$-brane,
and the Ramond-Ramond fields and the worldvolume gauge fields are turned off for simplicity.
Now, let us consider the dual action which is equivalent to \eqref{eq:beta-action-reduced},
and additionally includes the Wess-Zumino term:
\begin{align}
\begin{split}
S\bigl[\,\tilde{g}_{ij},\,\tilde{\phi},\,\beta^{(8)}_{ij}\,\bigr]
&= \frac{1}{2\kappa_{10}^2}\, \int \Bigl[ \Exp{-2\tilde{\phi}}\,\bigl(\tilde{*}\,\tilde{R} + 4\,{\mathrm{d}}\tilde{\phi}\wedge \tilde{*}\,{\mathrm{d}}\tilde{\phi} \bigr)
-\frac{1}{4}\,\Exp{2\tilde{\phi}}\,\tilde{g}^{ik}\,\tilde{g}^{jl}\,\bm{Q}^{(9)}_{ij}\wedge \tilde{*}\,\bm{Q}^{(9)}_{kl}\Bigr]
\\
&\quad - \mu_{5^2_2}\,\int \frac{1}{2}\,\beta_{\mathbf{p}\mathbf{q}}^{(8)}\wedge\bm{\delta}^{\mathbf{p}\mathbf{q}}(x-X(\xi)) \,.
\end{split}
\end{align}
Taking a variation with respect to $\beta_{\mathbf{p}\mathbf{q}}^{(8)}$,
we obtain the following equation of motion:
\begin{align}
\frac{1}{2\kappa_{10}^2}\,{\mathrm{d}} \bm{Q}^{(1)\,\mathbf{p}\mathbf{q}}
= \frac{\mu_{5^2_2}}{(2\pi R_{\mathbf{p}})(2\pi R_{\mathbf{q}})}\,n^{\mathbf{p}\mathbf{q}} \,\delta^2(x-X(\xi))\,{\mathrm{d}} x^1\wedge{\mathrm{d}} x^2 \,.
\label{eq:BI-522}
\end{align}
See \cite{Chatzistavrakidis:2013jqa,Andriot:2014uda} and Appendix D of \cite{Okada:2014wma}
for the above Bianchi identity for the $Q$-flux in the presence of the $5^2_2$-branes (or the $Q$-branes).
From \eqref{eq:BI-522}, we conclude that the current for the $5^2_2(n_1\cdots n_5,m_1m_2)$-brane
(in the absence of the Ramond-Ramond fields) is given by
\begin{align}
\tilde{*}\,j_{5^2_2(n_1\cdots n_5,\,m_1m_2)}
= \frac{(2\pi R_{m_1})(2\pi R_{m_2})}{2\kappa_{10}^2\,\mu_{5^2_2}}\,{\mathrm{d}} \bm{Q}^{(1)\,m_1m_2} \,.
\end{align}
Having identified the $5^2_2$-brane as a magnetic source of the $Q$-flux,
it is then natural to investigate an object which electrically couples to the $\beta$-field.
In the $\beta$-supergravity, such an object will be described by the following (Euclidean) solution:
\begin{align}
\begin{split}
{\mathrm{d}} \tilde{s}^2 &= \abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} + {\mathrm{d}} \tau^2 + {\mathrm{d}} x^2_{34567} + \rho_2\,{\mathrm{d}} x_{89}^2 \,,\quad
\Exp{2\tilde{\phi}} = \rho_2 \,,
\\
\beta^{(8)}_{89}&= -\rho_1\,{\mathrm{d}} \tau\wedge{\mathrm{d}} x^3\wedge\cdots\wedge{\mathrm{d}} x^9 \,,\quad
\beta^{89} = \rho_2^{-1}\,.
\end{split}
\label{eq:beta-2-instanton}
\end{align}
Indeed, this background has a monodromy given by the shift in the mixed-symmetry tensor,
$\beta^{(8)}_{03\cdots 9,89}\to \beta^{(8)}_{03\cdots 9,89} - \sigma$,
which gives a non-zero electric charge associated with the $Q$-flux;
\begin{align}
\int {\mathrm{d}} \beta^{(8)}_{03\cdots 9,89} = -\sigma\,(2\pi R_0)\,(2\pi R_3)\cdots (2\pi R_9) \,.
\end{align}
Since $\beta^{89}$ is a (bi-vector-valued) 0-form, the object which electrically couples to $\beta^{89}$ will be an instanton
that has two special isometry directions, $x^8$ and $x^9$.
See section \ref{sec:p=7} for further details about instanton backgrounds.
Now, let us consider how the $\FF1$ background (or the $1^6_4$ background) can be described in the $\beta$-supergravity.
If we use the relation \eqref{eq:beta-fields},
we will notice that, apparently, we cannot express the background \eqref{eq:F1-BG} (or \eqref{eq:1^6_4-BG})
in terms of the fundamental fields of the $\beta$-supergravity
since the matrix $E_{ij}$ is not invertible.
However, if we consider a gauge transformation $B_{03}\to B_{03}-a$ ($a$: constant)
in the defect $\FF1(3)$ background \eqref{eq:F1-BG}, we can calculate $(\tilde{g}_{ij},\,\tilde{\phi},\,\beta^{ij})$\,:
\begin{align}
\begin{split}
\underline{\text{ $\FF1$ :}}\quad
&{\mathrm{d}}\tilde{s}^2= \abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}}\bar{z} + \Exp{2\tilde{\phi}}\,\bigl({\mathrm{d}} t^2 - {\mathrm{d}} x_3^2\bigr) + {\mathrm{d}} x_{4\cdots 9}^2 \,, \quad
\Exp{2\tilde{\phi}}= a^2\,(2\,a^{-1}+ \rho_2)\,,
\\
&\beta^{(8)}_{03}= - a^2\,\rho_1\,{\mathrm{d}} t\wedge {\mathrm{d}} x^3\wedge\cdots\wedge {\mathrm{d}} x^9\,,\quad
\beta^{03} = \frac{a^{-2}}{2\,a^{-1}+ \rho_2} - a^{-1}\,.
\end{split}
\end{align}
If we choose $a=1$ and interchange $t$ with $x^3$, and perform a shift in the cutoff radius $r_\text{c}$ which makes $2+ \rho_2\to \rho_2$,
the $\FF1(3)$ background becomes
\begin{align}
\begin{split}
\underline{\text{ $\FF1$ :}}\quad
&{\mathrm{d}}\tilde{s}^2= \abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}}\bar{z} + \Exp{2\tilde{\phi}}\,\bigl(-{\mathrm{d}} t^2 + {\mathrm{d}} x_3^2\bigr) + {\mathrm{d}} x_{4\cdots 9}^2 \,, \quad
\Exp{2\tilde{\phi}}= \rho_2 \,,
\\
&\beta^{(8)}_{03}= \rho_1\,{\mathrm{d}} t\wedge {\mathrm{d}} x^3\wedge\cdots\wedge {\mathrm{d}} x^9 \,,\quad
\beta^{03} = -\rho_2^{-1} + 1\,,
\end{split}
\end{align}
which is globally well-defined.
Similarly, the $1^6_4$ background becomes
\begin{align}
\begin{split}
\underline{\text{ $1^6_4$ :}}\quad
&{\mathrm{d}}\tilde{s}^2= \abs{\rho}^2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}}\bar{z} + \Exp{2\tilde{\phi}}\,\bigl(-{\mathrm{d}} t^2 + {\mathrm{d}} x_3^2\bigr) + {\mathrm{d}} x_{4\cdots 9}^2 \,, \quad
\Exp{2\tilde{\phi}}= \frac{\rho_2}{\abs{\rho}^2} \,,
\\
&\beta^{(8)}_{03}= -\frac{\rho_1}{\abs{\rho}^2}\,{\mathrm{d}} t\wedge {\mathrm{d}} x^3\wedge\cdots\wedge {\mathrm{d}} x^9 \,,\quad
\beta^{03} = -\frac{\abs{\rho}^2}{\rho_2} + 1\,,
\end{split}
\end{align}
which is not single-valued as is the case with the original background \eqref{eq:1^6_4-BG}.
A possible way to globally describe the $1^6_4$ background is discussed in section \ref{sec:NS-NS}.
\section{Ramond-Ramond counterpart of the $\beta$-supergravity}
\label{sec:S-dual-bega}
In the previous section, we reviewed the $\beta$-supergravity
and explained that it is suitable for describing the non-geometric $5^2_2$ background.
If we perform an $S$-duality, the $5^2_2$-brane is mapped to the D5$_2$-brane (or the $5^2_3$-brane).
At the same time, the non-geometric $Q$-flux, sourced by the $5^2_2$-brane,
will be mapped to another non-geometric flux, called the $P$-flux,
which is related to the Ramond-Ramond 2-form $C^{(2)}$ instead of the $B$-field.
In this section, we write down an effective action for the $P$-flux
utilizing the techniques of the $\beta$-supergravity
and examine the $\DD5_2$-brane as a magnetic source of the $P$-flux.
\subsection{The effective action for the $P$-flux}
Let us begin with the type IIB action in the ten-dimensional Einstein frame:
\begin{align}
\begin{split}
S&= \frac{1}{2\kappa_{10}^2}\,
\int \Bigl[ *_{\text{E}} R_{\text{E}} - \frac{{\mathrm{d}}\tau\wedge *_{\text{E}}{\mathrm{d}}\overline{\tau}}{2\,(\operatorname{Im}\tau)^2}
\\
&\quad\qquad\qquad - \frac{1}{2\operatorname{Im}\tau}\,\bigl({\mathrm{d}} C^{(2)}-\tau\,{\mathrm{d}} B^{(2)}\bigr)\wedge *_{\text{E}}\bigl({\mathrm{d}} C^{(2)}-\overline{\tau}\,{\mathrm{d}} B^{(2)}\bigr)
-\frac{1}{4}\,F^{(5)}\wedge *_{\text{E}} F^{(5)}\Bigr] \,,
\end{split}
\label{eq:SIIB}
\end{align}
where the Chern-Simons term is dropped since it is irrelevant for the following discussions,
and we have defined the axio-dilaton $\tau$ by
\begin{align}
\tau \equiv C^{(0)} + {\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}}\Exp{-\phi}\,.
\end{align}
If we redefine the axio-dilaton by
\begin{align}
\widetilde{\tau} \equiv \widetilde{C}^{(0)} +{\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}} \Exp{- \widecheck{\phi}}
\equiv -1/\tau = \frac{-C^{(0)}+ {\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}} \Exp{-\phi}}{\abs{\tau}^2} \,,
\end{align}
the action becomes
\begin{align}
\begin{split}
S &= \frac{1}{2\kappa_{10}^2}\,
\int \Bigl[ *_{\text{E}} R_{\text{E}} - \frac{{\mathrm{d}} \widetilde{\tau}\wedge *_{\text{E}}{\mathrm{d}}\overline{\widetilde{\tau}}}{2\,(\operatorname{Im}\widetilde{\tau})^2}
\\
&\quad\qquad\qquad - \frac{1}{2\operatorname{Im}\widetilde{\tau}}\,\bigl({\mathrm{d}} B^{(2)}-\widetilde{\tau}\,{\mathrm{d}} C^{(2)}\bigr)
\wedge *_{\text{E}}\bigl({\mathrm{d}} B^{(2)}-\overline{\widetilde{\tau}}\,{\mathrm{d}} C^{(2)}\bigr)
-\frac{1}{4}\,F^{(5)}\wedge *_{\text{E}} F^{(5)}\Bigr] \,.
\end{split}
\end{align}
Then, in the ``string frame'' given by $\widecheck{g}_{ij} \equiv \Exp{\widecheck{\phi}/2}\,G^{\text{E}}_{ij}
= \Exp{\phi/2}\,\abs{\tau}\,G^{\text{E}}_{ij}= \abs{\tau}\,G_{ij}$\,,
the action can be rewritten as
\begin{align}
\begin{split}
S &= \frac{1}{2\kappa_{10}^2}\,
\int \Exp{-2\widecheck{\phi}}\,\bigl(\widecheck{*}\,\widecheck{R} + 4\,{\mathrm{d}}\widecheck{\phi}\wedge \widecheck{*}\,{\mathrm{d}}\widecheck{\phi}
- \frac{1}{2}\,{\mathrm{d}} C^{(2)}\wedge \widecheck{*}\,{\mathrm{d}} C^{(2)}\bigr)
\\
&\quad -\frac{1}{4\kappa_{10}^2}\,
\int \bigl[{\mathrm{d}} \widetilde{C}^{(0)}\wedge \widecheck{*}\,{\mathrm{d}} \widetilde{C}^{(0)}
+ \bigl({\mathrm{d}} B^{(2)}-\widetilde{C}^{(0)}\,{\mathrm{d}} C^{(2)}\bigr)\wedge \widecheck{*}\,\bigl({\mathrm{d}} B^{(2)}-\widetilde{C}^{(0)}\,{\mathrm{d}} C^{(2)}\bigr)
\\
&\quad\qquad\qquad\quad + \frac{1}{2}\,F^{(5)}\wedge \widecheck{*}\, F^{(5)} \bigr]\,.
\end{split}
\label{eq:SIIB-NS7}
\end{align}
In a simple case of $B^{(2)}= C^{(0)}= C^{(4)}=0$\,, the above action reduces to
\begin{align}
S\bigl[\,\widecheck{g}_{ij},\,\widecheck{\phi},\,C^{(2)}\,\bigr]
= \frac{1}{2\kappa_{10}^2}\,\int \Exp{-2\widecheck{\phi}}\,\Bigl(\widecheck{*}\,\widecheck{R}
+ 4\,{\mathrm{d}}\widecheck{\phi}\wedge \widecheck{*}\,{\mathrm{d}}\widecheck{\phi}
- \frac{1}{2}\,{\mathrm{d}} C^{(2)}\wedge \widecheck{*}\,{\mathrm{d}} C^{(2)}\Bigr) \,,
\end{align}
which has the same structure with the NS-NS action.
We can thus use the techniques of the $\beta$-supergravity
to rewrite the action into the following form:
\begin{align}
\begin{split}
S\bigl[\,g'_{ij},\,\phi',\,\gamma^{ij}\,\bigr]
&= \frac{1}{2\kappa_{10}^2}\,\int{\mathrm{d}}^{10}x \sqrt{\abs{g'}}\, \Exp{-2\phi'}\,
\Bigl(R'+4g^{\prime\, ij}\,\partial_i \phi\,\partial_j \phi + \widecheck{\mathcal R}(g'_{ij},\,\gamma^{ij})
\\
&\qquad\qquad\qquad\qquad\qquad\qquad - \frac{1}{2}\,\abs{S^{ijk}}^2
+ 4\,\abs{\gamma^{ij}\,\partial_j \phi -\mathcal T^i(g'_{ij},\,\gamma^{ij})}^2\Bigr)\,,
\end{split}
\end{align}
where $g'_{ij}$, $\gamma^{ij}$, $\phi$, and $S^{ijk}$ are defined by
\begin{align}
\begin{split}
g'_{ij}&\equiv F_{ik}\,F_{jl}\,\widecheck{g}^{kl} \,, \quad
\gamma^{ij} \equiv \bigl(F^{-1}\bigr)^{ik}\,\bigl(F^{-1}\bigr)^{jl}\,C^{(2)}_{kl}\,,\quad
\Exp{-2\phi'}\sqrt{\abs{g'}}\equiv \Exp{-2\widecheck{\phi}}\sqrt{\abs{\widecheck{g}}}\,,
\\
F&\equiv \bigl(F_{ij}\bigr)\equiv \bigl(\widecheck{g}_{ij} - C^{(2)}_{ij}\bigr)
= \bigl(\Exp{-\phi}G_{ij} - C^{(2)}_{ij}\bigr)\,,\quad
S^{ijk}\equiv 3\,\gamma^{l[i}\,\partial_l\gamma^{jk]} \,,
\end{split}
\end{align}
and $R'$ is the Ricci scalar associated with $g'_{ij}$\,,
and $\widecheck{\mathcal R}$ and $\mathcal T^i$ are the same as those defined in \eqref{eq:beta-definitions}.
We then make a further redefinition
\begin{align}
\tilde{g}_{ij} \equiv \Exp{-\phi'} g'_{ij}\,,\quad
\Exp{\tilde{\phi}}\equiv \Exp{-\phi'} \,,
\end{align}
which brings the action into the following form:
\begin{align}
\begin{split}
&\tilde{S}\bigl[\,\tilde{g}_{ij},\,\tilde{\phi},\,\gamma^{ij}\,\bigr]
\\
&= \frac{1}{2\kappa_{10}^2}\int{\mathrm{d}}^{10}x \sqrt{\abs{\tilde{g}}}\,
\Bigl[ \Exp{-2\tilde{\phi}} \,\bigl(\tilde{R}+4\,\tilde{g}^{ij}\,\partial_i \phi\,\partial_j \phi \bigr)
-\frac{1}{2}\,\Exp{-6\tilde{\phi}} \,\abs{S^{ijk}}^2
\\
&\qquad\qquad\qquad +\Exp{-4\tilde{\phi}}\,\Bigl(\tilde{g}_{ij}\,\widecheck{\mathcal R}^{ij}(\Exp{-\tilde{\phi}}\tilde{g}_{ij},\,\gamma^{ij})
+ 4\,\absb{\gamma^{ij}\,\partial_j \phi -\mathcal T^i(\Exp{-\tilde{\phi}}\tilde{g}_{ij},\,\gamma^{ij})}^2\Bigr)\Bigr] \,.
\end{split}
\label{eq:gamma-action}
\end{align}
To summarize, we performed the redefinition of fields
\begin{align}
\begin{split}
&\tilde{g}_{ij}= \frac{\abs{\det (\Exp{-\phi}G_{ij})}^{1/2}}{\abs{\det F_{ij}}^{1/2}}\,\Exp{2\phi} F_{ik}\,F_{jl}\,G^{kl} \,, \quad
\gamma^{ij} = \bigl(F^{-1}\bigr)^{ik}\,\bigl(F^{-1}\bigr)^{jl}\,C^{(2)}_{kl}\,,
\\
&\Exp{2\tilde{\phi}} = \frac{\abs{\det (\Exp{-\phi}G_{ij})}}{\abs{\det F_{ij}}}\,\Exp{2\phi}
\,,\quad
F_{ij} = \Exp{-\phi} G_{ij} - C^{(2)}_{ij} \,,
\end{split}
\label{eq:gamma-field-redefinition}
\end{align}
and the resulting action \eqref{eq:gamma-action} is equal to the type IIB action with $B^{(2)}= C^{(0)}= C^{(4)}=0$, up to total derivative terms.
As we expect naturally, the above action can also be obtained by making the following replacements (like the $S$-duality)
in the action of the $\beta$-supergravity:
\begin{align}
\begin{alignedat}{3}
G_{ij} &\to \Exp{-\phi}\, G_{ij}\,,\quad&
\phi &\to - \phi\,,\quad& B_{ij}&\to -C^{(2)}_{ij} \,,
\\
\tilde{g}_{ij} &\to \Exp{-\tilde{\phi}}\, \tilde{g}_{ij}\,, \quad&
\tilde{\phi}&\to -\tilde{\phi}\,,\quad& \beta^{ij}&\to -\gamma^{ij}\,.
\label{eq:gen-S-dual}
\end{alignedat}
\end{align}
As in the case of the $\beta$-supergravity,
if we assume that any derivative $\partial_i$ contracted with $\gamma^{ij}$ vanishes,
the above action reduces to the following simple action:
\begin{align}
\begin{split}
\tilde{S}\bigl[\,\tilde{g}_{ij},\,\tilde{\phi},\,\gamma^{ij}\,\bigr]
= \frac{1}{2\kappa_{10}^2} \int\Bigl[\,\Exp{-2\tilde{\phi}}\,
\bigl(\tilde{*}\,\tilde{R} + 4\,{\mathrm{d}}\phi\wedge \tilde{*}\,{\mathrm{d}}\phi\bigr)
-\frac{1}{4} \Exp{-4\tilde{\phi}}\,\tilde{g}_{ik}\,\tilde{g}_{jl}\,
\bm{P}^{(1)\,ij}\wedge \tilde{*}\bm{P}^{(1)\,kl}\,\Bigr] \,,
\end{split}
\label{eq:gamma-action-reduced}
\end{align}
where we defined the $P$-flux by
$\bm{P}^{(1)\,ij}\equiv P_k{}^{ij}\,{\mathrm{d}} x^k \equiv {\mathrm{d}} \gamma^{ij}$.
\subsection{$5^2_3$-brane as a source of the $P$-flux}
The $\text{D} 5_2(34567,89)$ background written in the new background fields $(\tilde{g}_{ij},\,\tilde{\phi},\,\gamma^{ij})$ is given by
\begin{align}
\begin{split}
\underline{\text{ $\text{D} 5_2$ :}}\quad
&{\mathrm{d}} \tilde{s}^2 = \rho_2^{1/2}\, \bigl(\rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z}+{\mathrm{d}} x^2_{034567}\bigr) + \rho_2^{-1/2}\, {\mathrm{d}} x^2_{89} \,, \quad
\Exp{2\tilde{\phi}} = \rho_2 \,,
\\
&\gamma^{89} = \rho_1 \,, \quad
\gamma^{(8)}_{89} = -\rho_2^{-1}\,{\mathrm{d}} t \wedge {\mathrm{d}} x^3\wedge \cdots \wedge {\mathrm{d}} x^9 \,.
\end{split}
\label{eq:523-gamma}
\end{align}
We can confirm that this background indeed satisfies the equations of motion
derived from the action \eqref{eq:gamma-action-reduced}
(see \eqref{eq:EOM1}--\eqref{eq:EOM3} for the explicit form).
Note that this is a geometric background since the monodromy is given by a gauge transformation
that corresponds to the shift in the $\gamma$-field, $\gamma^{89}\to \gamma^{89}+\sigma$,
to be called the \emph{$\gamma$-transformation}.
According to \cite{Chatzistavrakidis:2013jqa,Kimura:2014upa},
the Wess-Zumino term of the $5^2_3(34567,89)$-brane action
(smeared in the isometry directions, $x^8$ and $x^9$)
is written as
\begin{align}
S^{5^2_3}_{\text{WZ}} = - \mu_{5^2_3} \,\int \gamma_{89}^{(8)}\wedge\bm{\delta}^{89}(x-X(\xi)) \,,
\end{align}
where the $B$-field, the Ramond-Ramond 0- and 4-forms, and the worldvolume gauge fields
are turned off for simplicity, and $\bm{\delta}^{89}(x-X(\xi))$ is defined in \eqref{eq:SWZ522}.
As in the case of the $5^2_2$-brane, if we consider the action
\begin{align}
\begin{split}
S\bigl[\,\tilde{g}_{ij},\,\tilde{\phi},\,\gamma^{(8)}_{ij}\,\bigr]
&= \frac{1}{2\kappa_{10}^2}\, \int \Bigl[\Exp{-2\tilde{\phi}}\,\bigl(\tilde{*} \tilde{R} + 4\,{\mathrm{d}}\tilde{\phi}\wedge \tilde{*}\,{\mathrm{d}}\tilde{\phi} \bigr)
-\frac{1}{4}\,\Exp{4\tilde{\phi}}\,\tilde{g}^{ik}\,\tilde{g}^{jl}\,\bm{P}^{(9)}_{ij}\wedge \tilde{*}\,\bm{P}^{(9)}_{kl} \Bigr]
\\
&\quad - \mu_{5^2_3}\,\int \frac{1}{2}\, \gamma_{\mathbf{p}\mathbf{q}}^{(8)}\wedge\bm{\delta}^{\mathbf{p}\mathbf{q}}(x-X(\xi)) \,,
\end{split}
\end{align}
and take a variation with respect to $\gamma_{\mathbf{p}\mathbf{q}}^{(8)}$, we obtain the Bianchi identity for the $P$-flux with a source term:
\begin{align}
\frac{1}{2\kappa_{10}^2}\,{\mathrm{d}} \bm{P}^{(1)\,\mathbf{p}\mathbf{q}}
= \frac{\mu_{5^2_3}}{(2\pi R_\mathbf{p})(2\pi R_\mathbf{q})}\,n^{\mathbf{p}\mathbf{q}}\, \delta^2(x-X(\xi))\,{\mathrm{d}} x^1\wedge{\mathrm{d}} x^2 \,.
\end{align}
As in the case of the $\beta$-supergravity,
we can further find a solution corresponding to the (Euclidean) background of an instanton that couples to $\gamma^{ij}$ electrically.
The explicit form of the background fields is given later in the next section (see \eqref{eq:gamma-instantons}).
\section{Effective actions for non-geometric fluxes}
\label{sec:exotic-p-branes}
In this section, we derive the effective actions for various polyvectors $\gamma^{i_1\cdots i_{7-p}}$,
which are generalizations of the (simple) action \eqref{eq:gamma-action-reduced}
for $\gamma^{ij}$ to arbitrary $(7-p)$-vectors.
From these actions, we can also obtain the effective actions for
the non-geometric $Q$-fluxes; $Q_k{}^{ij}$ and $Q_k{}^{i_1\cdots i_6}$.
Since the derivation presented here does not rely on the results of the $\beta$-supergravity,
as for the action for the $Q$-flux,
the derivation can be regarded as another derivation of the (reduced) $\beta$-supergravity action.
Further, we find two kinds of solutions which correspond to
the exotic-brane backgrounds and their electric duals.
\subsection{Effective actions for the $P$-fluxes}
\label{sec:action-P-flux}
In this subsection, we consider the following ansatz,
which is crucial in deriving the effective actions for the non-geometric $P$-fluxes:
\begin{align}
\begin{split}
&{\mathrm{d}} s^2=G_{\mathbf{a}\mathbf{b}}\,{\mathrm{d}} x^\mathbf{a}\,{\mathrm{d}} x^\mathbf{b}+G_{\mathbf{p}\mathbf{q}}\,{\mathrm{d}} x^\mathbf{p}\,{\mathrm{d}} x^\mathbf{q} \qquad (\,\text{i.e.~}G_{\mathbf{a}\mathbf{p}}=0 \,)\,,
\\
&B^{(2)}=0\,,\quad
C^{(p+1)}_{\mathbf{a} i_1\cdots i_p}=0\,,\quad \text{isometries in the $x^\mathbf{p}$-directions}\,.
\label{eq:assumptions}
\end{split}
\end{align}
It is important to note that, throughout this paper,
the bold indices $\mathbf{a},\mathbf{b},\mathbf{c}$ run only over $1,2$,
while $\mathbf{p},\dotsc,\mathbf{t}$ run over $0,3,\dotsc,9$.
Now, due to a technical reason explained below,
we start from the type II$^\star$ theory \cite{Hull:1998vg}.
Since the $B$-field is assumed to vanish, the action is given by
\begin{align}
S = \frac{1}{2\kappa_{10}^2}\,
\int \Bigl[\Exp{-2\phi}\,\bigl(*\,R + 4\,{\mathrm{d}} \phi\wedge *\,{\mathrm{d}} \phi\bigr)
+ \sum_p \frac{a_p}{2}\, {\mathrm{d}} C^{(p+1)}\wedge *\,{\mathrm{d}} C^{(p+1)}\Bigr] \,,
\label{eq:typeII*}
\end{align}
where $p$ is summed over $0,2$ for the type IIA$^\star$ theory
while $-1,1,3$ for the type IIB$^\star$ theory,
and $a_p$ are constants given by $a_p=1$ ($p=-1,0,1,2$) and $a_3=1/2$.
Let us introduce $(7-p)$-vector fields $\gamma^{i_1\cdots i_{7-p}}$ by
\begin{align}
\begin{split}
&\gamma^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}
\equiv \frac{1}{(p+1)!}\,\epsilon^{\mathbf{q}_{p+1}\cdots \mathbf{q}_1\mathbf{s}_1\cdots \mathbf{s}_{7-p}}\,C^{(p+1)}_{\mathbf{q}_1\cdots \mathbf{q}_{p+1}} \,,
\\
&\gamma^{\mathbf{a} i_1\cdots i_{6-p}}\equiv 0 \qquad
\bigl(\epsilon^{\mathbf{p}_1\cdots \mathbf{p}_8}=\epsilon^{[\mathbf{p}_1\cdots \mathbf{p}_8]}\,,\ \epsilon^{03\cdots 9}=1\bigr) \,.
\end{split}
\label{eq:gamma-def}
\end{align}
Note that, if the metric is diagonal,
the map $C^{(p+1)}_{\mathbf{q}_1\cdots \mathbf{q}_{p+1}}\to \gamma^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}$
corresponds to performing eight $T$-dualities, $T_{03\cdots 9}$,
which includes a timelike $T$-duality
\footnote
If we instead define $\gamma$ as $C^{(p+1)}_{j_1\cdots j_{p+1}}= \epsilon_{j_{p+1}\cdots j_1i_1\cdots i_{9-p}}\,\gamma^{i_1\cdots i_{9-p}}/(9-p)!$\,,
which corresponds to taking $T$-dualities in all spacetime directions, $\gamma^{i_1\cdots i_{9-p}}$ coincides with $\tilde{C}^{i_1\cdots i_{9-p}}$
defined in (6.15) of \cite{Hohm:2011dv}.}
Thus, in order to make $\gamma^{i_1\cdots i_{7-p}}$ into a background field in the type II theory,
$C^{(p+1)}$ should be a field in the type II$^\star$ theory,
and this is the reason why we start from the type II$^\star$ theory.
In the following, we rewrite the action \eqref{eq:typeII*} regarding $\gamma^{i_1\cdots i_{7-p}}$
as a fundamental variable, instead of the form field $C^{(p+1)}$\,.
Under the assumptions \eqref{eq:assumptions},
the Ramond-Ramond part of the action \eqref{eq:typeII*} becomes
\begin{align}
\frac{1}{2\kappa_{10}^2}\int{\mathrm{d}}^{10}x \sqrt{\abs{G}}\,\sum_p\frac{a_p}{2\,(p+1)!}\,G^{\mathbf{a}\mathbf{b}}\,G^{\mathbf{p}_1\mathbf{q}_1}\cdots G^{\mathbf{p}_{p+1}\mathbf{q}_{p+1}}\,
\partial_\mathbf{a} C^{(p+1)}_{\mathbf{p}_1\cdots \mathbf{p}_{p+1}}\,\partial_\mathbf{b} C^{(p+1)}_{\mathbf{q}_1\cdots \mathbf{q}_{p+1}} \,.
\end{align}
By substituting the relation
\begin{align}
C^{(p+1)}_{\mathbf{q}_1\cdots \mathbf{q}_{p+1}}
= \frac{1}{(7-p)!}\,\epsilon_{\mathbf{q}_{p+1}\cdots \mathbf{q}_1\mathbf{s}_1\cdots \mathbf{s}_{7-p}}\,\gamma^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}} \quad
\bigl(\epsilon_{\mathbf{p}_1\cdots \mathbf{p}_8}=\epsilon_{[\mathbf{p}_1\cdots \mathbf{p}_8]}\,,\ \epsilon_{03\cdots 9}=1\bigr)\,,
\end{align}
the action can be written as
\begin{align}
-\frac{1}{2\kappa_{10}^2}\,\int{\mathrm{d}}^{10}x\sqrt{\abs{G}}\,\sum_p\frac{a_p\,\Delta^{-1}}{2\,(7-p)!}\,G^{\mathbf{a}\mathbf{b}}\,G_{\mathbf{r}_1\mathbf{s}_1}\cdots G_{\mathbf{r}_{7-p}\mathbf{s}_{7-p}}\,\partial_\mathbf{a} \gamma^{\mathbf{r}_1\cdots \mathbf{r}_{7-p}}\,\partial_\mathbf{b} \gamma^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}} \,,
\end{align}
where we defined $\Delta\equiv \abs{\det G_{\mathbf{p}\mathbf{q}}}$ and used the identity
\begin{align}
\begin{split}
&\frac{1}{(p+1)!\,[\,(7-p)!\,]^2}\,G^{\mathbf{p}_1\mathbf{q}_1}\cdots G^{\mathbf{p}_{p+1}\mathbf{q}_{p+1}}\,\epsilon_{\mathbf{p}_1\cdots \mathbf{p}_{p+1}\mathbf{r}_1\cdots \mathbf{r}_{7-p}}\,\epsilon_{\mathbf{q}_1\cdots \mathbf{q}_{p+1}\mathbf{s}_1\cdots \mathbf{s}_{7-p}}
\\
&= -\frac{\Delta^{-1}}{(7-p)!}\, G_{\mathbf{t}_1\mathbf{s}_1}\cdots G_{\mathbf{t}_{7-p}\mathbf{s}_{7-p}}\,
\delta^{[\mathbf{t}_1}_{[\mathbf{r}_1}\cdots \delta^{\mathbf{t}_{7-p}]}_{\mathbf{r}_{7-p}]}\,.
\end{split}
\end{align}
Thus, the action becomes
\begin{align}
S = \frac{1}{2\kappa_{10}^2}\,
\int \biggl[\Exp{-2\phi}\,\bigl(*\,R + 4\,{\mathrm{d}} \phi\wedge *\,{\mathrm{d}} \phi\bigr)
-\sum_p \frac{a_p\,\Delta^{-1}}{2\,(7-p)!}\,\bm{P}^{(1)\,i_1\cdots i_{7-p}}\wedge *\,\bm{P}^{(1)}_{i_1\cdots i_{7-p}}\biggr] \,,
\label{eq:RR-action}
\end{align}
where we defined the $P$-flux by
\begin{align}
\bm{P}^{(1)\, i_1\cdots i_{7-p}} \equiv P_j{}^{i_1\cdots i_{7-p}}\,{\mathrm{d}} x^j \equiv {\mathrm{d}} \gamma^{i_1\cdots i_{7-p}}\,.
\end{align}
Due to the presence of $\Delta^{-1}$,
if we treat $\bm{P}^{(1)\,i_1\cdots i_{7-p}}$ as tensors (see footnote \ref{foot:tensor})
this action is not invariant under diffeomorphisms in the eight-dimensional spacetime, spanned by $x^\mathbf{p}$.
This issue is resolved by making a redefinition of the metric and the dilaton.
Since the redefinition depends on the degree $p$,
in the following, we consider the case where only a Ramond-Ramond $p$-form (with $p\neq 3$) is non-vanishing:
\begin{align}
S = \frac{1}{2\kappa_{10}^2}\,
\int \biggl[\Exp{-2\phi}\,\bigl(*\,R + 4\,{\mathrm{d}} \phi\wedge *\,{\mathrm{d}} \phi\bigr)
- \frac{\Delta^{-1}}{2\,(7-p)!}\,\bm{P}^{(1)\,i_1\cdots i_{7-p}}\wedge *\,\bm{P}^{(1)}_{i_1\cdots i_{7-p}}\biggr] \,.
\label{eq:p-vector-action}
\end{align}
In this case, with the redefinition of the metric and the dilaton
\begin{align}
\begin{split}
{\mathrm{d}}\tilde{s}^2
&\equiv \Exp{\frac{4}{p-3}\,\phi} \Exp{\frac{p+1}{p-3}\,\eta} G_{\mathbf{a}\mathbf{b}}\,{\mathrm{d}} x^\mathbf{a}\,{\mathrm{d}} x^\mathbf{b} + \Exp{\eta} G_{\mathbf{p}\mathbf{q}}\,{\mathrm{d}} x^\mathbf{p}\,{\mathrm{d}} x^\mathbf{q}\,,
\\
\Exp{2\tilde{\phi}}&\equiv \Exp{2\phi+4\eta}\,,\quad
\Exp{\eta}\equiv \bigl(\Exp{-2\phi}\Delta^{1/2}\bigr)^{\frac{2}{p-3}} \,,
\label{eq:gamma-p-redefinitions}
\end{split}
\end{align}
we can show that the action \eqref{eq:p-vector-action}
is equal to the following action up to total derivative terms (see appendix \ref{app:derivation} for the detail):
\begin{align}
S = \frac{1}{2\kappa_{10}^2}\,
\int \biggl[\Exp{-2\tilde{\phi}}\, \bigl(\tilde{*}\,\tilde{R} + 4\,{\mathrm{d}} \tilde{\phi}\wedge \tilde{*}\,{\mathrm{d}} \tilde{\phi}\bigr)
- \frac{1}{2\,(7-p)!}\,\Exp{-4\tilde{\phi}}\bm{P}^{(1)\,i_1\cdots i_{7-p}}\wedge \tilde{*}\,\bm{P}^{(1)}_{i_1\cdots i_{7-p}}\biggr] \,,
\label{eq:gamma-action-p}
\end{align}
which reduces to the action \eqref{eq:gamma-action-reduced} in a special case of $p=5$\,.
The equations of motion are obtained as follows:
\begin{align}
&\tilde{R}+4\bigl(\tilde{\nabla}^i\partial_i\tilde{\phi}-\tilde{g}^{ij}\,\partial_i\tilde{\phi}\,\partial_j\tilde{\phi}\bigr)
-\frac{\Exp{-2\tilde{\phi}}}{(7-p)!}\,P_i{}^{j_1\cdots j_{7-p}}\,P^i{}_{j_1\cdots j_{7-p}} =0 \,,
\label{eq:EOM1}
\\
&\tilde{R}_{ij} +2\tilde{\nabla}_i\partial_j\tilde{\phi}-\frac{\Exp{-2\tilde{\phi}}}{2\,(7-p)!}\,
\Bigl(P_i{}^{k_1\cdots k_{7-p}}\,P_{jk_1\cdots k_{7-p}} -(7-p)\,P_{k_1i}{}^{k_2\cdots k_{7-p}}\,P^{k_1}{}_{jk_2\cdots k_{7-p}}
\nonumber\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad +\frac{1}{2}\,P_k{}^{l_1\cdots l_{7-p}}\,P^k{}_{l_1\cdots l_{7-p}} \,\tilde{g}_{ij} \Bigr) =0 \,,
\label{eq:EOM2}
\\
&{\mathrm{d}} \bm{P}^{(9)}_{i_1\cdots i_{7-p}}=0\,,\quad
\bm{P}^{(9)}_{i_1\cdots i_{7-p}} \equiv
\Exp{-4\tilde{\phi}}\, \tilde{g}_{i_1j_1}\cdots\tilde{g}_{i_{7-p}j_{7-p}}\, \tilde{*}\, \bm{P}^{(1)\,j_1\cdots j_{7-p}}\,.
\label{eq:EOM3}
\end{align}
In this theory, we can find the following two solutions.
The first one
\begin{align}
\begin{split}
\underline{\text{ $\text{D} p_{7-p}$ :}}\quad
&{\mathrm{d}} \tilde{s}^2 = \rho_2^{1/2}\,\bigl(\rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} +{\mathrm{d}} x^2_{0n_1\cdots n_p}\bigr) + \rho_2^{-1/2}\, {\mathrm{d}} x^2_{m_1\cdots m_{7-p}} \,,
\quad
\Exp{2\tilde{\phi}} = \rho_2^{\frac{p-3}{2}} \,,
\\
&\gamma^{m_1\cdots m_{7-p}} = \rho_1 \,,
\quad
\gamma^{(8)}_{m_1\cdots m_{7-p}} = -\rho_2^{-1}\,{\mathrm{d}} t \wedge {\mathrm{d}} x^3\wedge \cdots \wedge {\mathrm{d}} x^9 \,,
\end{split}
\label{eq:gamma-Dp_7-p}
\end{align}
is a generalization of \eqref{eq:523-gamma} and has a monodromy given by
a $\gamma$-transformation, $\gamma^{m_1\cdots m_{7-p}}\to \gamma^{m_1\cdots m_{7-p}} +\sigma$.
That is, it has a magnetic charge associated with the $P$-flux and should correspond to the background of the exotic $\text{D} p_{7-p}$-brane.
On the other hand, the second (Euclidean) solution
\begin{align}
\begin{split}
&{\mathrm{d}} \tilde{s}^2 = \rho_2^{-1/2}\, \bigl(\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z}+{\mathrm{d}}\tau^2+{\mathrm{d}} x^2_{n_1\cdots n_p}\bigr)+ \rho_2^{1/2}\, {\mathrm{d}} x^2_{m_1\cdots m_{7-p}} \,,
\quad
\Exp{2\tilde{\phi}} = \rho_2^{\frac{3-p}{2}} \,,
\\
&\gamma^{(8)}_{m_1\cdots m_{7-p}} = -\rho_1\,{\mathrm{d}} \tau \wedge {\mathrm{d}} x^3\wedge \cdots \wedge {\mathrm{d}} x^9 \,, \quad
\gamma^{m_1\cdots m_{7-p}} = \rho_2^{-1} \,,
\end{split}
\label{eq:gamma-instantons}
\end{align}
has an electric charge associated with the $P$-flux
and will correspond to an instanton with special isometry directions, $x^{m_1},\dotsc,x^{m_{7-p}}$,
which is similar to the solution obtained in \eqref{eq:beta-2-instanton}.
In the case of $p=3$, we cannot perform the redefinition \eqref{eq:gamma-p-redefinitions},
and do not know how to derive the action \eqref{eq:gamma-action-p}.
However, if the action is derived from a suitable field redefinition,
we can confirm that the above backgrounds with $p=3$ indeed satisfy the equations of motion.
\subsection{Relating the new and the original background fields}
In this subsection, we discuss the exotic duality further
and investigate a relation between the fundamental fields of the theory \eqref{eq:gamma-action-p}
and the standard background fields.
\paragraph*{\underline{Exotic duality}\\}
In order to examine the exotic duality, let us consider a simple configuration
\begin{align}
\begin{split}
&{\mathrm{d}} s^2 = G_{\mu\nu}\,{\mathrm{d}} x^\mu {\mathrm{d}} x^\nu+G_{\mathsf{m}\mathsf{n}}\,{\mathrm{d}} x^\mathsf{m} {\mathrm{d}} x^\mathsf{n}\quad (G_{\mathsf{m}\mathsf{n}}:\text{ diagonal}) \,,
\\
&C^{(7-p)} = C^{(7-p)}_{(p+3)\cdots 9}\,{\mathrm{d}} x^{p+3}\wedge\cdots\wedge{\mathrm{d}} x^9
\\
&(\mu,\nu=0,\dotsc,p+2\,,\quad \mathsf{m},\mathsf{n}=p+3,\dotsc,9)\,,
\end{split}
\label{eq:simple-config}
\end{align}
which includes the $\text{D} p(3\cdots (p+2))$ and the $\text{D} p_{7-p}(3\cdots (p+2),(p+3)\cdots 9)$ background in the type II theory.
For this kind of simple background, it is convenient to take a $T_{9\cdots (p+3)}$-duality and describe it in the type IIB theory.
The resulting type IIB background (in the Einstein frame) after performing the $T_{9\cdots (p+3)}$-duality is given by
\begin{align}
\begin{split}
{\mathrm{d}} \tilde{s}_{\text{E}}^2
&= \Exp{-\tilde{\phi}/2}\,\bigl(\tilde{G}_{\mu\nu}\,{\mathrm{d}} x^\mu {\mathrm{d}} x^\nu + \tilde{G}_{\mathsf{m}\mathsf{n}}\,{\mathrm{d}} x^\mathsf{m} {\mathrm{d}} x^\mathsf{n}\bigr) \quad
\bigl(\tilde{G}_{\mu\nu}\equiv G_{\mu\nu}\,,\ \tilde{G}_{\mathsf{m}\mathsf{n}}\equiv G^{\mathsf{m}\mathsf{n}}\bigr)\,,
\\
\tilde{\tau} &\equiv \tilde{C}^{(0)} +{\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}}\Exp{-\tilde{\phi}}
\equiv C^{(7-p)}_{(p+3)\cdots 9} + {\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}}\sqrt{\det G_{\mathsf{m}\mathsf{n}}}\, \Exp{-\phi} \equiv \tau_{[(p+3)\cdots 9]}\,.
\end{split}
\end{align}
Since the type IIB theory has the well-known $\text{SL}(2,{\mathbb Z})$-duality symmetry,
\begin{align}
G^{\text{E}}_{ij}\to G^{\text{E}}_{ij}\,,\quad
\tau_{[(p+3)\cdots 9]} \to \tau'_{[(p+3)\cdots 9]}
=\frac{a\,\tau_{[(p+3)\cdots 9]}+b}{c\,\tau_{[(p+3)\cdots 9]}+d}\,,\quad
\begin{pmatrix} a & b \\ c & d \end{pmatrix} \in \text{SL}(2,{\mathbb Z}) \,,
\end{align}
performing the $\text{SL}(2,{\mathbb Z})$ transformation followed by a $T_{(p+3)\cdots 9}$-duality transformation,
we obtain a new background in the original type II theory (see (2.3) of \cite{LozanoTellechea:2000mc} for a similar argument).
In particular, performing $T_{9\cdots (p+3)}$-$S$-$T_{(p+3)\cdots 9}$-duality transformations,
the background \eqref{eq:simple-config} is mapped to the following background (in the string frame):
\begin{align}
\begin{split}
&{\mathrm{d}} s'^2= \abs{\tau_{[(p+3)\cdots 9]}}\,G_{\mu\nu}\,{\mathrm{d}} x^\mu\,{\mathrm{d}} x^\nu
+\abs{\tau_{[(p+3)\cdots 9]}}^{-1}\,G_{\mathsf{m}\mathsf{n}}\,{\mathrm{d}} x^\mathsf{m}\,{\mathrm{d}} x^\mathsf{n} \,,
\\
&C'^{(7-p)} = -\frac{C^{(7-p)}_{(p+3)\cdots 9}}{\abs{\tau_{[(p+3)\cdots 9]}}^2}\,
{\mathrm{d}} x^{p+3}\wedge\cdots\wedge{\mathrm{d}} x^9\,,\quad
\Exp{2\phi'}=\abs{\tau_{[(p+3)\cdots 9]}}^{p-3}\,\Exp{2\phi} \,.
\end{split}
\label{eq:exotic-special}
\end{align}
For the D$p(3\cdots (p+2))$-brane background, we have $\tau_{[(p+3)\cdots 9]}=\rho_1+{\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}}\rho_2=\rho(z)$,
and we can confirm that the above transformation rule coincides with that of the exotic duality transformation
given in \eqref{eq:exotic-dual}.
The transformation rule \eqref{eq:exotic-special} under the exotic duality
can be applied only for $p$-branes extending in the $3\cdots (p+2)$-directions.
However, for a special case of $p=5$,
without assuming $C^{(2)} = C^{(2)}_{89}\,{\mathrm{d}} x^{8}\wedge{\mathrm{d}} x^9$,
we can write down the following transformation rule
which exchanges a $\text{D} 5$-brane and an exotic $\text{D} 5_2$-brane extending in arbitrary $x^\mathbf{p}$-directions
\footnote{Note that the transformation rule is deformed in the presence of other background fields, such as the $B$-field.
\begin{align}
\begin{split}
\widetilde{G}_{\mathbf{a}\mathbf{b}}&= \abs{\det F_{\mathbf{p}\mathbf{q}}}^{1/2}\,G_{\mathbf{a}\mathbf{b}}\,,\quad
\widetilde{G}_{\mathbf{p}\mathbf{q}}= \abs{\det F_{\mathbf{p}\mathbf{q}}}^{1/2}\,(F^{-1})^{\mathbf{p}\mathbf{r}}\,(F^{-1})^{\mathbf{q}\mathbf{s}}\,G_{\mathbf{r}\mathbf{s}}\,,
\\
\Exp{2\widetilde{\bm{\phi}}}&= \abs{\det F_{\mathbf{p}\mathbf{q}}}\,\Exp{2\phi}\,, \quad
\widetilde{C}^{(2)}_{\mathbf{p}\mathbf{q}} = (F^{-1})^{\mathbf{p}\mathbf{r}}\,(F^{-1})^{\mathbf{q}\mathbf{s}}\,C^{(2)}_{\mathbf{r}\mathbf{s}}\quad
\bigl(F_{\mathbf{p}\mathbf{q}}\equiv \Exp{-\phi}G_{\mathbf{p}\mathbf{q}}-C^{(2)}_{\mathbf{p}\mathbf{q}}\bigr)\,.
\end{split}
\label{eq:exotic-p=5}
\end{align}
This corresponds to an $S$-duality transformation followed by $T_{03\cdots 9}$-dualities and an $S$-duality transformations.
For general $p$, we do not know a covariant expression like \eqref{eq:exotic-p=5},
but we here assume the existence of a transformation rule which interchanges $\text{D} p$-branes with $\text{D} p_{7-p}$-branes:
\begin{align}
\bigl(G_{ij},\,\phi,\,C^{(7-p)}\bigr)
\overset{\text{exotic duality}}\longrightarrow
\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{C}^{(7-p)}\bigr) \,.
\end{align}
Performing an $S$-duality,
we can also obtain the transformation rule for the fields in the NS-NS sector.
That is, if we consider a configuration
\begin{align}
\begin{split}
&{\mathrm{d}} s^2 = G_{\mu\nu}\,{\mathrm{d}} x^\mu {\mathrm{d}} x^\nu+G_{\mathsf{m}\mathsf{n}}\,{\mathrm{d}} x^\mathsf{m} {\mathrm{d}} x^\mathsf{n}\quad (G_{\mathsf{m}\mathsf{n}}:\text{ diagonal}) \,,
\\
&B^{(7-p)} = B^{(7-p)}_{(p+3)\cdots 9}\,{\mathrm{d}} x^{p+3}\wedge\cdots\wedge{\mathrm{d}} x^9
\end{split}
\end{align}
with $p=1$ or $p=5$, and define a complex field
\begin{align}
\tau^{\text{NS}}_{[(p+3)\cdots 9]} \equiv B^{(7-p)}_{(p+3)\cdots 9} + {\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}}\sqrt{\det G_{\mathsf{m}\mathsf{n}}}\, \Exp{-\frac{5-p}{2}\,\phi} \,,
\end{align}
the configuration after the action of the exotic duality (i.e.~the $S$-dual of \eqref{eq:exotic-special}) is given by
\begin{align}
\begin{split}
&{\mathrm{d}} s'^2= \abs{\tau^{\text{NS}}_{[(p+3)\cdots 9]}}^{\frac{p-1}{2}}\,G_{\mu\nu}\,{\mathrm{d}} x^\mu\,{\mathrm{d}} x^\nu
+\abs{\tau^{\text{NS}}_{[(p+3)\cdots 9]}}^{\frac{p-5}{2}}\,G_{\mathsf{m}\mathsf{n}}\,{\mathrm{d}} x^\mathsf{m}\,{\mathrm{d}} x^\mathsf{n} \,,
\\
&B'^{(7-p)} = -\frac{B^{(7-p)}_{(p+3)\cdots 9}}{\abs{\tau^{\text{NS}}_{[(p+3)\cdots 9]}}^2}\,
{\mathrm{d}} x^{p+3}\wedge\cdots\wedge{\mathrm{d}} x^9\,,\quad
\Exp{2\phi'}=\abs{\tau^{\text{NS}}_{[(p+3)\cdots 9]}}^{-(p-3)}\,\Exp{2\phi} \,.
\end{split}
\end{align}
We can easily verify that this transformation rule with $p=5$ exchanges the $\NS5(34567)$ background \eqref{eq:NS5}
with the $5^2_2(34567,89)$ background \eqref{eq:522},
and that with $p=1$ exchanges the $\FF1(3)$ background \eqref{eq:F1-BG}
with the $1^6_4(3,456789)$ background \eqref{eq:1^6_4-BG} with each other.
Note that the combination $\tau^\text{NS}_{[89]}=B_{89}+{\hspace{0.7pt}\mathrm{i}\hspace{0.7pt}}\sqrt{\det G_{\mathsf{m}\mathsf{n}}}$ frequently appears in the discussion of the monodromy
of NS5(34567)- and $5^2_2(34567,89)$-branes (see e.g., \cite{deBoer:2012ma,Kimura:2014wga}).
\paragraph*{\underline{On the relation between $\bigl(\tilde{g}_{ij},\,\tilde{\phi},\,\gamma^{i_1\cdots i_{7-p}}\bigr)$ and $\bigl(G_{ij},\,\phi,\,C^{(7-p)}\bigr)$}\\}
Now, we discuss a possible relation between the $\gamma$-field, $\gamma^{i_1\cdots i_{7-p}}$,
and the standard background fields; $\bigl(G_{ij},\,\phi,\,C^{(7-p)}\bigr)$\,.
We here assume the ansatz \eqref{eq:assumptions}, and further, $G_{pq}$ is diagonal.
Then, we consider the following sequence of dualities:
\begin{align}
\bigl(G_{ij},\,\phi,\,C^{(7-p)}\bigr)
\overset{\text{exotic duality}}\longrightarrow
\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{C}^{(7-p)}\bigr)
\overset{T_{03456789}}\longrightarrow
\bigl(G^\star_{ij},\,\phi^\star,\,C^{\star(p+1)}\bigr)_{\text{II$^\star$}} \,.
\label{eq:dual-fields}
\end{align}
If we consider a non-geometric $\text{D} p_{7-p}$-brane background on the leftmost side,
the background in the middle has the same form with the geometric $\text{D} p$-brane background,
and the background on the rightmost side has the same form with the E$(7-p)$-brane background \cite{Hull:1998vg} in the type II$^\star$ theory.
Due to the assumption that $G_{\mathbf{p}\mathbf{q}}$ is diagonal, the relation between
$\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{C}^{(7-p)}\bigr)$
and
$\bigl(G^\star_{ij},\,\phi^\star,\,C^{\star(p+1)}\bigr)$
is given by
\begin{align}
\begin{split}
&\widetilde{G}_{\mathbf{a}\mathbf{b}}=G^\star_{\mathbf{a}\mathbf{b}}\,,\quad
\widetilde{G}^{\mathbf{p}\mathbf{q}}=G^\star_{\mathbf{p}\mathbf{q}}\,,\quad
\Exp{-2\widetilde{\bm{\phi}}}\sqrt{\abs{\det \widetilde{G}_{\mathbf{p}\mathbf{q}}}}=\Exp{-2\phi^\star}\sqrt{\abs{\det G^\star_{\mathbf{p}\mathbf{q}}}} \,,
\\
&\widetilde{C}^{(7-p)}_{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}
= \frac{1}{(p+1)!}\,\epsilon^{\mathbf{q}_{p+1}\cdots \mathbf{q}_1\mathbf{s}_1\cdots \mathbf{s}_{7-p}}\,C^{\star (p+1)}_{\mathbf{q}_1\cdots \mathbf{q}_{p+1}} \,.
\end{split}
\end{align}
Then, with the identification of $\bigl(G^\star_{ij},\,\phi^\star,\,C^{\star(p+1)}\bigr)_{\text{II$^\star$}}$
with the fields $\bigl(G_{ij},\,\phi,\,C^{(7-p)}\bigr)$ appearing in \eqref{eq:gamma-p-redefinitions},
we obtain the following expression for $\bigl(\tilde{g}_{ij},\,\tilde{\phi},\,\gamma^{i_1\cdots i_{7-p}}\bigr)$:
\begin{align}
\begin{split}
&(\tilde{g}_{ij})
= \begin{pmatrix} \Exp{\frac{4}{p-3}\,\phi^\star} \Exp{\frac{p+1}{p-3}\,\eta} G^\star_{\mathbf{a}\mathbf{b}}&0\\
0&\Exp{\eta} G^\star_{\mathbf{p}\mathbf{q}}
\end{pmatrix}
= \begin{pmatrix} \Exp{-\frac{4}{p-3}\,\widetilde{\bm{\phi}}}\, \Exp{\frac{7-p}{p-3}\,\eta} \, \widetilde{G}_{\mathbf{a}\mathbf{b}}&0\\
0&\Exp{\eta} \widetilde{G}^{\mathbf{p}\mathbf{q}}
\end{pmatrix} \,,
\\
&\Exp{2\tilde{\phi}} = \Exp{2\phi^\star+4\eta}=\Exp{-2\widetilde{\bm{\phi}}+(7-p)\,\eta} \,,\quad
\gamma^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}=\widetilde{C}^{(7-p)}_{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}\,,
\\
&\Exp{\eta} = \Bigl(\Exp{-2\phi^\star}\sqrt{\abs{\det G^\star_{\mathbf{p}\mathbf{q}}}}\Bigr)^{\frac{2}{p-3}}
= \Bigl(\Exp{-2\widetilde{\bm{\phi}}}\sqrt{\abs{\det \widetilde{G}_{\mathbf{p}\mathbf{q}}}}\Bigr)^{\frac{2}{p-3}} \,.
\end{split}
\label{eq:relations}
\end{align}
The inverse relation is given by
\begin{align}
\begin{split}
&(\widetilde{G}_{ij})
= \begin{pmatrix} \Exp{-\frac{4}{p-3}\,\tilde{\phi}}\, \Exp{\frac{7-p}{p-3}\,\eta}\, \tilde{g}_{\mathbf{a}\mathbf{b}}&0\\
0&\Exp{\eta} \tilde{g}^{\mathbf{p}\mathbf{q}}
\end{pmatrix} \,,\quad
\Exp{2\widetilde{\bm{\phi}}} = \Exp{-2\tilde{\phi}+(7-p)\,\eta} \,,\quad
\\
&\widetilde{C}^{(7-p)}_{\mathbf{s}_1\cdots \mathbf{s}_{7-p}} =\gamma^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}\,,
\quad
\Exp{\eta} = \Bigl(\Exp{-2\tilde{\phi}}\sqrt{\abs{\det \widetilde{g}_{\mathbf{p}\mathbf{q}}}}\Bigr)^{\frac{2}{p-3}} \,.
\end{split}
\label{eq:inv-relations}
\end{align}
In a case where $\bigl(G_{ij},\,\phi,\,C^{(7-p)}\bigr)$ in \eqref{eq:dual-fields}
is given by the $\text{D} p_{7-p}$ background,
$\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{C}^{(7-p)}\bigr)$
in \eqref{eq:dual-fields}
has the same form with the $\text{D} p$ background.
Since the $\text{D} p$ background \eqref{eq:Dp-brane} satisfies
$\eta=0$ and $\Exp{-\frac{4}{p-3}\,\widetilde{\bm{\phi}}}=\rho_2$
\footnote{Since $\eta$ is invariant under an $S$-duality or $T$-dualities in the $x^\mathbf{p}$-directions, all defect backgrounds
considered in section \ref{sec:supergravity-description} satisfy $\eta=0$.}
we obtain the following expression for $\bigl(\tilde{g}_{ij},\,\tilde{\phi},\,\gamma^{i_1\cdots i_{7-p}}\bigr)$:
\begin{align}
\begin{split}
\underline{\text{ $\text{D} p_{7-p}$ :}}\quad
&{\mathrm{d}}\tilde{s}^2 = \rho_2^{1/2}\,\bigl(\rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} +{\mathrm{d}} x^2_{0n_1\cdots n_p}\bigr) + \rho_2^{-1/2}\, {\mathrm{d}} x^2_{m_1\cdots m_{7-p}}\,,
\\
&\Exp{2\tilde{\phi}} = \rho_2^{\frac{p-3}{2}} \,, \quad \gamma^{m_1\cdots m_{7-p}} = \rho_1 \,,
\end{split}
\end{align}
which exactly reproduces the solution found in \eqref{eq:gamma-Dp_7-p}.
To summarize, the derivation of the theory presented in this section consists of two steps:
\begin{align}
\text{1st step:}\quad&\bigl(G_{ij},\,\phi,\,C^{(7-p)}\bigr)
\overset{\text{exotic duality}}\longrightarrow
\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{C}^{(7-p)}\bigr) \,,
\\
\text{2nd step:}\quad&\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{C}^{(7-p)}\bigr)
\overset{\text{field redefinition}}{\longrightarrow}
\bigl(\tilde{g}_{ij},\,\tilde{\phi},\,\gamma^{i_1\cdots i_{7-p}}\bigr) \,.
\end{align}
The first step is the exotic duality which is just a $U$-duality transformation that maps
a non-geometric background to a standard geometric background.
The second step is given by the field redefinition \eqref{eq:relations}
which converts a $(7-p)$-form field into a $(7-p)$-vector,
and after the redefinition, the action has the form \eqref{eq:gamma-action-p}.
Although we do not know the general expression for the transformation rule under the exotic duality,
for example, in the case of $p=5$ where the transformation is given by \eqref{eq:exotic-p=5},
we can obtain the following relation between
$\bigl(\tilde{g}_{ij},\,\tilde{\phi},\,\gamma^{i_1\cdots i_{7-p}}\bigr)$
and $\bigl(G_{ij},\,\phi,\,C^{(7-p)}\bigr)$:
\begin{align}
\begin{split}
(\tilde{g}_{ij})&= \frac{\abs{\det G_{\mathbf{p}\mathbf{q}}}^{1/2}}{\abs{\det F_{\mathbf{p}\mathbf{q}}}^{1/2}}\,\Exp{-2\phi}
\begin{pmatrix} \Exp{-2\phi}\,G_{\mathbf{a}\mathbf{b}} & 0 \\ 0 & F_{\mathbf{p}\mathbf{r}}\,F_{\mathbf{q}\mathbf{s}}\,G^{\mathbf{r}\mathbf{s}}\end{pmatrix}
\\
&= \frac{\abs{\det (\Exp{-\phi}G_{ij})}^{1/2}}{\abs{\det F_{ij}}^{1/2}}\,\Exp{2\phi}
\begin{pmatrix} F_{\mathbf{a} k}\,F_{\mathbf{b} l}\,G^{kl} & 0 \\ 0 & F_{\mathbf{p} k}\,F_{\mathbf{q} l}\,G^{kl}\end{pmatrix} \,,
\\
\Exp{2\tilde{\phi}}&= \frac{\abs{\det G_{\mathbf{p}\mathbf{q}}}}{\abs{\det F_{\mathbf{p}\mathbf{q}}}}\,\Exp{-6\phi}
= \frac{\abs{\det (\Exp{-\phi}G_{ij})}}{\abs{\det F_{ij}}}\,\Exp{2\phi} \,,
\\
\gamma^{\mathbf{p}\mathbf{q}} &= (F^{-1})^{\mathbf{p}\mathbf{r}}\,(F^{-1})^{\mathbf{q}\mathbf{s}}\,C^{(2)}_{\mathbf{r}\mathbf{s}}
= (F^{-1})^{\mathbf{p} i}\,(F^{-1})^{\mathbf{q} j}\,C^{(2)}_{ij}\,,
\end{split}
\end{align}
where we used the assumptions $G_{\mathbf{a}\mathbf{p}}=C^{(2)}_{\mathbf{a} i}=0$\,.
This is precisely equal to the relation \eqref{eq:gamma-field-redefinition},
and the presentation given in this section serves as an alternative derivation of
the theory \eqref{eq:gamma-action-reduced} with its generalization for general values of $p$ (with $p\neq 3$).
\paragraph*{\underline{Duality rules for the new background fields}\\}
Here, we comment on the duality rules for the new background fields $\bigl(\tilde{g}_{ij},\,\tilde{\phi},\,\gamma^{i_1\cdots i_{7-p}}\bigr)$
under the $U$-duality transformations.
By the construction of the theory presented in this section,
it will be natural to define the transformation rule in the following manner.
Let us begin with a configuration $\bigl(\tilde{g}_{ij},\,\tilde{\phi},\,\gamma^{i_1\cdots i_{7-p}}\bigr)$
which satisfies the equations of motion derived from the action \eqref{eq:gamma-action-p}.
In order to perform a $U$-duality, we first transform these fields into $\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{C}^{(7-p)}\bigr)$
using the relation \eqref{eq:inv-relations}.
Secondly, we use the standard transformation rules under the $U$-duality to obtain a new background
$\bigl(\widetilde{G}'_{ij},\,\widetilde{\bm{\phi}}',\,\widetilde{C}'^{(7-p')},\cdots\bigr)$,
where the ellipsis represents possible additional fields such as the $B$-field.
Finally, using the relation \eqref{eq:relations}
\footnote{Note that the relation \eqref{eq:relations} should be generalized if additional fields, such as the $B$-field, have non-vanishing values.}
we can obtain the $U$-dual background
$\bigl(\tilde{g}'_{ij},\,\tilde{\phi}',\,\gamma'^{i_1\cdots i_{7-p'}},\cdots\bigr)$.
At the same time, the original background fields $\bigl(G_{ij},\,\phi,\,C^{(7-p')}\bigr)$ should be transformed into $\bigl(G'_{ij},\,\phi',\,C'^{(7-p')}\bigr)$,
which is the exotic dual of the background
$\bigl(\widetilde{G}'_{ij},\,\widetilde{\bm{\phi}}',\,\widetilde{C}'^{(7-p')},\cdots\bigr)$.
As an example, let us consider a simple configuration given by
\begin{align}
{\mathrm{d}} \tilde{s}^2 = \tilde{g}_{\mathbf{a}\mathbf{b}}\,{\mathrm{d}} x^\mathbf{a} {\mathrm{d}} x^\mathbf{b} + \tilde{g}_{pq}\,{\mathrm{d}} x^p {\mathrm{d}} x^q
+ \tilde{g}_{99}\,{\mathrm{d}} x_9^2 \,, \quad
\gamma^{(p+2)\cdots 8} = a\qquad (p,\,q=0,3\dotsc,8)\,.
\end{align}
In this case, if we take a $T$-duality in the $x^9$-direction, from the above procedure,
we can straightforwardly obtain the following $T$-dual background:
\begin{align}
\begin{split}
{\mathrm{d}} \tilde{s}'^2 &= \Bigl(\Exp{\frac{4\tilde{\phi}}{(p-3)(p-4)}+\frac{4\eta}{p-3}}\,\tilde{g}_{99}^{-\frac{2}{p-4}}\Bigr)\,\tilde{g}_{\mathbf{a}\mathbf{b}}\,{\mathrm{d}} x^\mathbf{a} {\mathrm{d}} x^\mathbf{b} + \tilde{g}_{pq}\,{\mathrm{d}} x^p {\mathrm{d}} x^q
+ \Exp{-2\eta}\,\tilde{g}_{99}^{-1}\,{\mathrm{d}} x_9^2 \,,
\\
\Exp{2\tilde{\phi}'}&= \Exp{2\tilde{\phi}+2\eta}\,\tilde{g}_{99}^{-1} \,,\quad \gamma'^{(p+2)\cdots 9} = a\,.
\end{split}
\end{align}
We can easily confirm that this relation indeed maps the $\text{D} p_{7-p}(3\cdots(p+1)9,(p+2)\cdots 8)$ background \eqref{eq:gamma-Dp_7-p}
to the $\text{D} (p-1)_{8-p}(3\cdots(p+1),(p+2)\cdots 9)$ background.
In this way, we can show that the $\gamma$-fields follow the rule,
$\gamma^{i_1\cdots i_p}\to \gamma^{i_1\cdots i_py}$ or $\gamma^{i_1\cdots i_py}\to\gamma^{i_1\cdots i_p}$,
under the $T_y$-duality transformation,
which is similar to the rule for the Ramond-Ramond fields.
In the case of general $U$-duality transformation,
since multiple Ramond-Ramond fields and the $B$-field have non-vanishing values,
we can no more use the relation \eqref{eq:relations}.
However, as we discuss in the next subsection, in the case of an $S$-duality transformation,
we can obtain a relation similar to \eqref{eq:relations} and derive the $\beta$-supergravity.
\subsection{Effective actions for $Q$-fluxes}
\label{sec:NS-NS}
We here present another derivation of the (reduced) action \eqref{eq:beta-action-reduced}
for the $\beta$-supergravity from the action \eqref{eq:gamma-action-p}.
We further obtain an action for the non-geometric flux which has an exotic string
(i.e.~$1^6_4$-brane) as the magnetic source.
\paragraph*{\underline{Alternative derivation of the $\beta$-supergravity}\\}
Let us consider an $S$-duality transformation given by
\begin{align}
\begin{alignedat}{3}
\widetilde{G}_{ij} &\to \Exp{-\widetilde{\bm{\phi}}}\, \widetilde{G}_{ij}\,,\quad&
\widetilde{\bm{\phi}} &\to - \widetilde{\bm{\phi}}\,,\quad& \widetilde{C}^{(2)}_{ij}&\to -\widetilde{B}_{ij} \,,
\\
\tilde{g}_{ij} &\to \Exp{-\tilde{\phi}}\, \tilde{g}_{ij}\,, \quad&
\tilde{\phi}&\to -\tilde{\phi}\,,\quad& \gamma^{ij}&\to -\beta^{ij}\,,\quad \eta\to -\eta\,.
\end{alignedat}
\end{align}
Under the redefinition of fields, the action \eqref{eq:gamma-action-p} and the relation \eqref{eq:relations} with $p=5$ become
\begin{align}
S &= \frac{1}{2\kappa_{10}^2}\,
\int \biggl[\Exp{-2\tilde{\phi}}\, \bigl(\tilde{*}\,\tilde{R} + 4\,{\mathrm{d}} \tilde{\phi}\wedge \tilde{*}\,{\mathrm{d}} \tilde{\phi}\bigr)
- \frac{1}{4}\,\Exp{-2\tilde{\phi}}\,\bm{Q}^{(1)\,ij}\wedge \tilde{*}\,\bm{Q}^{(1)}_{ij}\biggr]\,,
\\
\begin{split}
(\tilde{g}_{ij})
&= \begin{pmatrix} \widetilde{G}_{\mathbf{a}\mathbf{b}}&0\\
0& \widetilde{G}^{\mathbf{p}\mathbf{q}}
\end{pmatrix} \,,
\quad
\Exp{2\tilde{\phi}} = \Exp{-2\widetilde{\bm{\phi}}+2\eta} \,,
\quad
\beta^{ij}=\widetilde{B}_{ij}\,,
\\
\Exp{\eta}&= \Bigl(\Exp{-2\widetilde{\bm{\phi}}}\sqrt{\abs{\det \widetilde{G}_{\mathbf{p}\mathbf{q}}}}\Bigr)^{-1} \,.
\label{eq:beta-tilde-widetilde}
\end{split}
\end{align}
Further, under the $S$-duality, the transformation rule for the exotic duality given in \eqref{eq:exotic-p=5} becomes
\begin{align}
\bigl(G_{ij},\,&\phi,\,B^{(2)}\bigr)
\overset{\text{exotic duality}}\longrightarrow
\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{B}^{(2)}\bigr) \,,
\\
\begin{split}
(\tilde{G}_{ij})
&= \begin{pmatrix}
G_{\mathbf{a}\mathbf{b}}&0 \\ 0& \bigl(E^{-1}\bigr)^{\mathbf{p} i}\,\bigl(E^{-1}\bigr)^{\mathbf{q} j}\,G_{ij}
\end{pmatrix}\,,
\\
\tilde{B}_{\mathbf{p}\mathbf{q}} &\equiv \bigl(E^{-1}\bigr)^{\mathbf{p} i}\,\bigl(E^{-1}\bigr)^{\mathbf{q} j}\,B_{ij}\,,\quad
\Exp{-2\tilde{\bm{\phi}}}\sqrt{\abs{\det\tilde{G}_{\mathbf{p}\mathbf{q}}}}
\equiv \Exp{-2\phi}\sqrt{\abs{\det G_{\mathbf{p}\mathbf{q}}}}\,.
\end{split}
\end{align}
Substituting this relation into \eqref{eq:beta-tilde-widetilde},
we can correctly reproduce the transformation rule \eqref{eq:beta-fields}
between $\bigl(G_{ij},\,\phi,\,B^{(2)}\bigr)$ and $\bigl(\tilde{g}_{ij},\,\tilde{\phi},\,\beta^{ij}\bigr)$.
This gives another derivation of the action \eqref{eq:beta-action-reduced} in the case $G_{\mathbf{a}\mathbf{p}}=B_{\mathbf{a} i}=0$ is satisfied.
\paragraph*{\underline{Supergravity action for the exotic string $1^6_4$}\\}
Now, let us consider applying an $S$-duality transformation to the action \eqref{eq:gamma-action-p} with $p=1$.
By using the $S$-duality transformation rule,
\begin{align}
\begin{alignedat}{3}
G_{ij} &\to \Exp{-\phi}\, G_{ij}\,,\quad&
\phi &\to - \phi\,,\quad& C^{(6)}_{i_1\cdots i_6}&\to B_{i_1\cdots i_6} \,,
\\
\tilde{g}_{ij} &\to \Exp{-\tilde{\phi}}\, \tilde{g}_{ij}\,, \quad&
\tilde{\phi}&\to -\tilde{\phi}\,,\quad& \gamma^{i_1\cdots i_6}&\to \beta^{i_1\cdots i_6}\,,\quad \eta\to -\eta\,,
\end{alignedat}
\end{align}
the action \eqref{eq:gamma-action-p} and the relation \eqref{eq:relations} with $p=1$ become
\begin{align}
&S= \frac{1}{2\kappa_{10}^2}\,
\int \biggl[\Exp{-2\tilde{\phi}}\, \bigl(\tilde{*} \tilde{R} + 4\,{\mathrm{d}} \tilde{\phi}\wedge \tilde{*}\,{\mathrm{d}} \tilde{\phi}\bigr)
- \frac{1}{2\cdot 6!}\,\Exp{-6\tilde{\phi}}\bm{Q}^{(1)\,i_1\cdots i_{6}}\wedge \tilde{*}\,\bm{Q}^{(1)}_{i_1\cdots i_{6}}\biggr] \,,
\label{eq:exotic-string-action}
\\
\begin{split}
&(\tilde{g}_{ij})
= \begin{pmatrix} \Exp{-4\widetilde{\bm{\phi}}+6\eta} \widetilde{G}_{\mathbf{a}\mathbf{b}}&0\\
0&\Exp{2\eta} \widetilde{G}^{\mathbf{p}\mathbf{q}}
\end{pmatrix} \,, \quad
\Exp{2\tilde{\phi}} = \Exp{-2\widetilde{\bm{\phi}}+6\eta} \,,
\\
&\beta^{\mathbf{s}_1\cdots \mathbf{s}_6} =\widetilde{B}^{(6)}_{\mathbf{s}_1\cdots \mathbf{s}_6}\,, \quad
\Exp{\eta} = \Exp{-2\widetilde{\bm{\phi}}}\sqrt{\abs{\det \widetilde{G}_{\mathbf{p}\mathbf{q}}}} \,.
\label{eq:exotic-string-relation}
\end{split}
\end{align}
In this case, it is difficult to find a transformation rule for the exotic duality
\begin{align}
\bigl(G_{ij},\,\phi,\,B^{(6)}\bigr)
\overset{\text{exotic duality}}\longrightarrow
\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{B}^{(6)}\bigr) \,.
\end{align}
However, as a special case, if we consider the $1^6_4(3,456789)$ background,
the fields with tilde $\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{B}^{(6)}\bigr)$
should have the same form with those in the $\FF1(3)$ background.
Then, from the relation \eqref{eq:exotic-string-relation},
we can obtain the following solution of the theory \eqref{eq:exotic-string-action},
which corresponds to the background of the $1^6_4(3,456789)$-brane:
\begin{align}
\begin{split}
\underline{\text{ $1^6_4$ :}}\quad
&{\mathrm{d}} \tilde{s}^2 = \rho_2\,\bigl(\rho_2\,\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z} +{\mathrm{d}} x^2_{03}\bigr) + {\mathrm{d}} x_{4\cdots 9}^2 \,, \quad
\Exp{2\tilde{\phi}} = \rho_2 \,,
\\
&\beta^{456789}= \rho_1 \,, \quad \beta_{456789}^{(8)} = - \rho_2^{-1}\,{\mathrm{d}} t\wedge {\mathrm{d}} x^3\wedge \cdots \wedge {\mathrm{d}} x^9\,.
\end{split}
\end{align}
We can also find a Euclidean solution which corresponds to the background of an instanton that electrically couples to $\beta^{i_1\cdots i_6}$:
\begin{align}
\begin{split}
&{\mathrm{d}} \tilde{s}^2 = \rho_2^{-1}\, \bigl(\abs{f}^2\,{\mathrm{d}} z\,{\mathrm{d}} \bar{z}+{\mathrm{d}}\tau^2+{\mathrm{d}} x^2_{n_1n_2}\bigr) + {\mathrm{d}} x^2_{m_1\cdots m_6} \,,
\quad
\Exp{2\tilde{\phi}} = \rho_2^{-1} \,,
\\
&\beta^{(8)}_{m_1\cdots m_6} = -\rho_1\,{\mathrm{d}} \tau \wedge {\mathrm{d}} x^3\wedge \cdots \wedge {\mathrm{d}} x^9 \,, \quad
\beta^{m_1\cdots m_6} = \rho_2^{-1} \,.
\end{split}
\label{eq:beta6-instantons}
\end{align}
\subsection{Seven branes and instantons}
\label{sec:p=7}
In the case of $p=7$, the exotic duality is the same as the $S$-duality:
\begin{align}
\widetilde{G}_{ij}=\abs{\tau}\,G_{ij}\,,\quad \Exp{2\widetilde{\bm{\phi}}}= \abs{\tau}^4\,\Exp{2\phi} \,,\quad \widetilde{C}^{(0)}=-\frac{C^{(0)}}{\abs{\tau}^2} \,.
\end{align}
In this case, from the relation \eqref{eq:relations} with $p=7$,
we can write down an explicit relation between $\bigl(G_{ij},\,\phi,\,C^{(0)}\bigr)$ and
$\bigl(\tilde{g}_{ij},\,\tilde{\phi},\,\gamma\bigr)$ as in the case of $p=5$\,:
\begin{align}
\begin{split}
(\tilde{g}_{ij})
&= \begin{pmatrix} \abs{\tau}^{-2}\,\Exp{-\phi}\, \tilde{G}_{\mathbf{a}\mathbf{b}}&0\\ 0&\Exp{\eta} \tilde{G}^{\mathbf{p}\mathbf{q}}
\end{pmatrix} \,,
\quad
\Exp{2\tilde{\phi}} = \frac{\Exp{-2\phi}}{\abs{\tau}^4} \,,\quad
\gamma = \widetilde{C}^{(0)} = -\frac{C^{(0)}}{\abs{\tau}^2} \,,
\\
\Exp{2\eta}&= \Exp{-2\bm{\phi}}\sqrt{\abs{\det G_{\mathbf{p}\mathbf{q}}}} \,.
\end{split}
\end{align}
From this relation, we can obtain the expression for $\bm{P}^{(9)}$ in terms of the original background fields $\bigl(G_{ij},\,\phi,\,C^{(0)}\bigr)$:
\begin{align}
\bm{P}^{(9)}
&= \Exp{-4\tilde{\phi}} \,\tilde{*}\,{\mathrm{d}} \gamma
= \Exp{-4\tilde{\phi}}\,\frac{\sqrt{\abs{\tilde{g}}}}{9!}\,\tilde{g}^{\mathbf{a}\mathbf{b}}\,\partial_\mathbf{a} \gamma\,\epsilon_{\mathbf{b} i_1\cdots i_9}\,{\mathrm{d}} x^{i_1}\wedge \cdots\wedge {\mathrm{d}} x^{i_9}
\nonumber\\
&= \abs{\tau}^4\,\frac{\sqrt{\abs{\tilde{G}}}}{9!}\,\tilde{G}^{\mathbf{a}\mathbf{b}}\,\partial_\mathbf{a} \gamma \,\epsilon_{\mathbf{b} i_1\cdots i_9}\,{\mathrm{d}} x^{i_1}\wedge \cdots\wedge {\mathrm{d}} x^{i_9}
= \abs{\tau}^4 *\,{\mathrm{d}} \biggl(-\frac{C^{(0)}}{\abs{\tau}^2}\biggr)
\nonumber\\
&= -C^{(0)}\,{\mathrm{d}} B^{(8)} + \abs{\tau}^2\, {\mathrm{d}} C^{(8)}\,,
\label{eq:P9-expression}
\end{align}
where we defined \cite{Eyras:1999at}
\begin{align}
{\mathrm{d}} B^{(8)}\equiv *\,{\mathrm{d}} \abs{\tau}^2 = \Exp{2\phi}*_{\text{E}} {\mathrm{d}} \abs{\tau}^2\,,\quad
{\mathrm{d}} C^{(8)}\equiv *\,{\mathrm{d}} C^{(0)} =\Exp{2\phi}*_{\text{E}}{\mathrm{d}} C^{(0)}\,.
\end{align}
Then, the equation of motion \eqref{eq:EOM3} can be written as
\begin{align}
{\mathrm{d}} \bm{P}^{(9)} = {\mathrm{d}} \bigl(-C^{(0)}\,{\mathrm{d}} B^{(8)} + \abs{\tau}^2\, {\mathrm{d}} C^{(8)}\bigr)=0\,.
\end{align}
In fact, this equation of motion can be derived, in the original theory,
as a conservation law for the Noether current associated with the $\text{SL}(2)$ symmetry \cite{Meessen:1998qm}.
Conversely, in the theory \eqref{eq:gamma-action-p}, the equation of motion, ${\mathrm{d}} F^{(9)}={\mathrm{d}}^2 C^{(8)}=0$,
of the original theory
will appear as a conservation law for the Noether current.
As it was shown in \cite{Eyras:1999at}, the dual potential $\gamma^{(8)}$ (or $\widetilde{C}^{(8)}$ given in \cite{Eyras:1999at})
is the field that electrically couples to an NS7-brane.
This can also be understood from the fact that a D7-brane couples to $C^{(8)}$
and the 8-forms transform under an $S$-duality (or the exotic duality) as follows \cite{Eyras:1999at}:
\begin{align}
C^{(8)} \to -\gamma^{(8)}\,,\quad \gamma^{(8)}\to -C^{(8)}\,,\quad B^{(8)}\to -B^{(8)} \,.
\end{align}
Further, as it is well-known, a D-instanton, or a $\text{D}(-1)$-brane, electrically couples to the Ramond-Ramond 0-form, $C^{(0)}$.
The ``mass,'' i.e., the on-shell value of the Euclidean action is calculated in \cite{Gibbons:1995vg}
and is proportional to $g_s^{-1}$, like the tension of the D$p$-branes.
If we perform an $S$-duality, the D-instanton is mapped to another instanton
which couples to the field $\gamma=\widetilde{C}^{(0)}$
and will have the ``mass'' proportional to $g_s$, since the $S$-duality maps $g_s\to 1/g_s$.
Indeed, in the Einstein frame, the action for $\gamma$ becomes
\begin{align}
S = \frac{1}{2\kappa_{10}^2}\,\int \sqrt{\abs{\tilde{g}_{\text{E}}}}\, \Bigl(\tilde{*}_\text{E} \tilde{R}_{\text{E}} -\frac{1}{2}\,{\mathrm{d}}\tilde{\phi}\wedge *_{\text{E}}{\mathrm{d}}\tilde{\phi}
- \frac{1}{2}\,\Exp{-2\tilde{\phi}}\, {\mathrm{d}} \gamma\wedge *_{\text{E}}{\mathrm{d}}\gamma \Bigr) \,,
\end{align}
and using the result (6.6) of \cite{Bergshoeff:2004fq}, we can find that the value of the on-shell action is proportional to $g_s$.
In the following, we will denote the instanton by $\mathbf{I}_{1}$\,.
Note that this instanton is a special case of the $(p,q)$-instanton,
which is a member of the \emph{$Q$-instantons} discovered in \cite{Bergshoeff:2008qq}.
By performing $T$-dualities in the $x^\mathbf{p}$-directions,
$\gamma$ will be mapped to a $(7-p)$-vector $\gamma^{i_1\cdots i_{7-p}}$,
and we will obtain an instanton, to be called $\mathbf{I}^{7-p}_{1}$, which electrically couples to $\gamma^{i_1\cdots i_{7-p}}$.
Recalling the transformation rule for the fundamental constants under the action of $T$-duality, $g_s\to g_s\, (l_s/R_{i_n})$ and $l_s\to l_s$,
we expect that the ``mass'' of $\mathbf{I}^{7-p}_{1}$ will be proportional to $g_s\,(l_s^{7-p}/R_{\mathbf{p}_1}\cdots R_{\mathbf{p}_{7-p}})$
(see section 7 of \cite{Bergshoeff:2011se} for a discussion on the existence of such objects).
The family of instantons, $\mathbf{I}^{7-p}_{1}$, can be thought of as generalizations of the pp-wave (whose mass is proportional to $(l_s/R_i)$),
much like the exotic branes can be thought of as generalizations of the Kaluza-Klein monopole.
The instanton $\mathbf{I}^{7-p}_{1}$ can exist only when there are $(7-p)$ compact isometry directions.
Since it is the electric source of $\bm{P}^{(1)\,i_1\cdots i_{7-p}}$,
the corresponding background should be given by \eqref{eq:gamma-instantons}.
Further, in the type IIB theory, by performing an $S$-duality,
the background fields $\gamma^{ij}$ and $\gamma^{i_1\cdots i_6}$ are transformed into $\beta^{ij}$ and $\beta^{i_1\cdots i_6}$\,.
At the same time, the instantons $\mathbf{I}^2_{1}$ and $\mathbf{I}^6_{1}$ will be mapped to other instantons, to be called $\mathbf{I}^2_0$ and $\mathbf{I}^6_{2}$,
whose ``mass'' will be proportional to $(l_s^2/R_{\mathbf{p}_1} R_{\mathbf{p}_2})$ and $g_s^2\,(l_s^{6}/R_{\mathbf{p}_1}\cdots R_{\mathbf{p}_6})$, respectively.
These instantons are also predicted in \cite{Bergshoeff:2011se}.
The corresponding background solutions will be given by \eqref{eq:beta-2-instanton} and \eqref{eq:beta6-instantons}.
\subsection{The mixed-symmetry tensors}
So far, we argued that the mixed-symmetry tensors
$\beta^{(8)}_{ij}$, $\beta^{(8)}_{i_1\cdots i_6}$, and $\gamma^{(8)}_{i_1\cdots i_{7-p}}$
electrically couple to exotic branes.
Here, we provide expressions for these mixed-symmetry tensors in terms of the original background fields,
like the relation \eqref{eq:P9-expression}.
For convenience, we denote the form fields $B^{(2)}$, $B^{(6)}$, and $C^{(7-p)}$ collectively as $\mathsf{A}^{(7-p)}$,
and the mixed-symmetry tensors as $\mathcal A^{i_1\cdots i_{7-p}}$ and $\mathcal A^{(8)}_{i_1\cdots i_{7-p}}$\,.
We again use the following tilde notation for the background fields which are related by the exotic duality:
\begin{align}
\bigl(G_{ij},\,\phi,\,\mathsf{A}^{(7-p)}\bigr)
\overset{\text{exotic duality}}\longrightarrow
\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{\mathsf{A}}^{(7-p)}\bigr) \,.
\end{align}
In the following, we derive an expression for the mixed tensor $\mathcal A^{(8)}_{i_1\cdots i_{7-p}}$
in terms of the background fields on the right-hand sides;
$\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{\mathsf{A}}^{(7-p)}\bigr)$.
We define the dual potential $\widetilde{\mathsf{A}}^{(p+1)}$ by
\begin{align}
{\mathrm{d}}\widetilde{\mathsf{A}}^{(p+1)}\equiv s_p\,\Exp{2\,(\alpha_s+1)\,\widetilde{\bm{\phi}}}\,\tilde{*}\,{\mathrm{d}} \widetilde{\mathsf{A}}^{(7-p)}\,,
\end{align}
where $s_p\equiv (-1)^{\frac{(p+2)(p+1)}{2}+\alpha_s+1}$ and
$\alpha_s=\{-1,\,-2,\,0\}$ for $\widetilde{\mathsf{A}}^{(p+1)}=\{\widetilde{C}^{(p+1)},\,\widetilde{B}^{(2)},\,\widetilde{B}^{(6)}\}$,
and $\tilde{*}$ is the Hodge star operator associated with $\widetilde{G}_{ij}$\,.
Then, using the identification
$\mathcal A^{i_1\cdots i_{7-p}}\equiv\{\gamma^{i_1\cdots i_{7-p}},\,\beta^{i_1i_2},\,\beta^{i_1\cdots i_6}\}
= \widetilde{\mathsf{A}}^{(7-p)}_{i_1\cdots i_{7-p}}$ and the definition
\begin{align}
{\mathrm{d}} \mathcal A^{(8)}_{i_1\cdots i_{7-p}}\equiv
\Exp{2\,(\alpha+1)\,\tilde{\phi}}\,\tilde{g}_{i_1j_1}\cdots\tilde{g}_{i_{7-p}j_{7-p}}\,\tilde{*}\,{\mathrm{d}} \mathcal A^{j_1\cdots j_{7-p}}\quad
(\alpha\equiv -\alpha_s-4)\,,
\end{align}
we obtain
\begin{align}
\partial_\mathbf{a} \widetilde{\mathsf{A}}^{(p+1)}_{\mathbf{s}_1\cdots \mathbf{s}_{p+1}}
&= s_p\,\Exp{2\,(\alpha_s+1)\,\widetilde{\bm{\phi}}}\,\frac{\sqrt{\abs{\widetilde{G}}}}{(7-p)!}\,
\widetilde{G}^{\mathbf{c}\mathbf{b}}\,\widetilde{G}^{\mathbf{q}_1\mathbf{r}_1}\cdots\widetilde{G}^{\mathbf{q}_{7-p}\mathbf{r}_{7-p}}\,
\epsilon_{\mathbf{c}\mathbf{q}_1\cdots \mathbf{q}_{7-p}\mathbf{a}\mathbf{s}_1\cdots \mathbf{s}_{p+1}}\,
\partial_\mathbf{b} \widetilde{\mathsf{A}}^{(7-p)}_{\mathbf{r}_1\cdots \mathbf{r}_{7-p}}
\nonumber\\
&= s_p\,\Exp{2\,(\alpha+1)\,\tilde{\phi}}\,\frac{\sqrt{\abs{\tilde{g}}}}{(7-p)!}\,\tilde{g}^{\mathbf{c}\mathbf{b}}\,
\tilde{g}_{\mathbf{q}_1\mathbf{r}_1}\cdots\tilde{g}_{\mathbf{q}_{7-p}\mathbf{r}_{7-p}}\,\epsilon_{\mathbf{c}\mathbf{q}_1\cdots \mathbf{q}_{7-p}\mathbf{a}\mathbf{s}_1\cdots \mathbf{s}_{p+1}}\,
\partial_\mathbf{b} \mathcal A^{\mathbf{r}_1\cdots \mathbf{r}_{7-p}}
\nonumber\\
&= s_p\,(-1)^{(p+2)(7-p)}\,\partial_\mathbf{a} \mathcal A^{(8)}_{\mathbf{s}_1\cdots \mathbf{s}_{p+1}\mathbf{q}_1\cdots \mathbf{q}_{7-p},\mathbf{q}_1\cdots \mathbf{q}_{7-p}} \,,
\end{align}
where the summation over the indices $\mathbf{q}_n$ is assumed in the second line but not in the third line,
and we used
\begin{align}
\widetilde{G}_{\mathbf{a}\mathbf{b}}\propto \tilde{g}_{\mathbf{a}\mathbf{b}}\,,\quad
\widetilde{G}^{\mathbf{p}\mathbf{q}}=\Exp{-(\alpha+2)\,\eta}\,\tilde{g}_{\mathbf{p}\mathbf{q}}\,,\quad
\Exp{\eta}=\Bigl(\Exp{-2\widetilde{\bm{\phi}}}\sqrt{\abs{\det \widetilde{G}_{\mathbf{p}\mathbf{q}}}}\Bigr)^{\frac{2}{p-3-4(\alpha+3)}} \,,
\end{align}
in the second equality.
Namely, we obtain
\begin{align}
\begin{split}
&\partial_\mathbf{a} \mathcal A^{(8)}_{\mathbf{s}_1\cdots \mathbf{s}_{p+1}\mathbf{q}_1\cdots \mathbf{q}_{7-p},\mathbf{q}_1\cdots \mathbf{q}_{7-p}}
= s_p\,(-1)^{(p+2)(7-p)}\,\partial_\mathbf{a} \mathsf{A}^{(p+1)}_{\mathbf{s}_1\cdots \mathbf{s}_{p+1}}
\\
&=(-1)^{(p+2)(7-p)}\,
\frac{\Exp{-2(\alpha+3)\tilde{\bm{\phi}}}}{(7-p)!}\,\varepsilon^{\mathbf{b}\mathbf{r}_1\cdots \mathbf{r}_{7-p}}{}_{\mathbf{a}\mathbf{s}_1\cdots \mathbf{s}_{p+1}}\,
\partial_\mathbf{b} \mathsf{A}^{(7-p)}_{\mathbf{r}_1\cdots \mathbf{r}_{7-p}}\quad \bigl(\varepsilon_{i_0\cdots i_9}\equiv \sqrt{\abs{\tilde{g}}}\,\epsilon_{i_0\cdots i_9}\bigr)\,,
\end{split}
\end{align}
which reproduces the proposed relations (30)--(32) of \cite{Bergshoeff:2011se}.
If we further use the relation between $\bigl(G_{ij},\,\phi,\,\mathsf{A}^{(7-p)}\bigr)$ and $\bigl(\widetilde{G}_{ij},\,\widetilde{\bm{\phi}},\,\widetilde{\mathsf{A}}^{(7-p)}\bigr)$,
we can also obtain the relation between $\bm{\iota}\mathcal A^{(8)}_{\mathbf{q}_1\cdots \mathbf{q}_{7-p}}$
and $\bigl(G_{ij},\,\phi,\,\mathsf{A}^{(7-p)}\bigr)$.
\subsection{Summary of the results}
In this section, we have presented various actions with the following form:
\begin{align}
S\bigl[\tilde{g}_{ij},\,\tilde{\phi},\,\mathcal A^{i_1\cdots i_{7-p}}\bigr]
&= \frac{1}{2\kappa_{10}^2}\,
\int \biggl[\Exp{-2\phi}\,\bigl(\tilde{*}\,\tilde{R} + 4\,{\mathrm{d}} \phi\wedge \tilde{*}\,{\mathrm{d}} \phi\bigr)
\nonumber\\
&\quad\qquad\qquad
- \frac{\Exp{2\,(\alpha+1)\,\tilde{\phi}}}{2\,(7-p)!}\,\tilde{g}_{i_1j_1}\cdots \tilde{g}_{i_{7-p}j_{7-p}}\,
\mathcal Q^{(1)\,i_1\cdots i_{7-p}}\wedge \tilde{*}\, \mathcal Q^{(1)\,j_1\cdots j_{7-p}}\biggr]\,,
\label{eq:action-general}
\end{align}
where $\mathcal Q^{(1)\,i_1\cdots i_{7-p}} \equiv {\mathrm{d}}\mathcal A^{i_1\cdots i_{7-p}}$ is a non-geometric flux
of which an exotic brane acts as the magnetic source,
and $\alpha$ is an integer given in Table \ref{tab:exotic}.
\begin{table}[t]
\centering
\begin{tabular}{|c||c|c|c|}\hline
$\mathcal A^{i_1\cdots i_{7-p}}$ & $\beta^{ij}$ & $\gamma^{i_1\cdots i_{7-p}}$ & $\beta^{i_1\cdots i_6}$ \\\hline
$\mathcal Q^{(1)\,i_1\cdots i_{7-p}}$ & \quad$\bm{Q}^{(1)\,ij}$\quad & \quad$\bm{P}^{(1)\,i_1\cdots i_{7-p}}$\quad & \quad$\bm{Q}^{(1)\,i_1\cdots i_6}$\quad \\\hline
$\alpha$ & $-2$ & $-3$ & $-4$ \\\hline
magnetic source $\bigl(p^{7-p}_{-\alpha}\bigr)$ & $5^2_2$ & $p^{7-p}_3$ & $1^6_4$ \\\hline
$\tilde{\alpha}\equiv -\alpha-2$ & $0$ & $1$ & $2$ \\\hline
electric source $\bigl(\mathbf{I}^{7-p}_{\tilde{\alpha}}\bigr)$ & $\mathbf{I}^2_0$ & $\mathbf{I}^{7-p}_{1}$ & $\mathbf{I}^6_{2}$ \\\hline
\end{tabular}
\caption{A list of non-geometric fluxes and their magnetic/electric sources.}
\label{tab:exotic}
\end{table}
The equations of motion are given by
\begin{align}
&\tilde{R}+4\bigl(\tilde{\nabla}^i\partial_i\tilde{\phi}-\tilde{g}^{ij}\,\partial_i\tilde{\phi}\,\partial_j\tilde{\phi}\bigr)
+\frac{(\alpha+1)\,\Exp{2\,(\alpha+2)\,\tilde{\phi}}}{2\,(7-p)!}\,\mathcal Q_i{}^{j_1\cdots j_{7-p}}\,\mathcal Q^i{}_{j_1\cdots j_{7-p}} =0 \,,
\\
&\tilde{R}_{ij} +2\tilde{\nabla}_i\partial_j\tilde{\phi}-\frac{\Exp{2\,(\alpha+2)\,\tilde{\phi}}}{2\,(7-p)!}\,
\Bigl(\mathcal Q_i{}^{k_1\cdots k_{7-p}}\,\mathcal Q_j{}_{k_1\cdots k_{7-p}} -(7-p)\,\mathcal Q_{k_1}{}^{k_2\cdots k_{7-p}}\,\mathcal Q^{k_1}{}_{jk_2\cdots k_{7-p}}
\nonumber\\
&\qquad\qquad\qquad\qquad\qquad\qquad -\frac{\alpha+2}{2}\,\mathcal Q_k{}^{l_1\cdots l_{7-p}}\,\mathcal Q^k{}_{l_1\cdots l_{7-p}} \,\tilde{g}_{ij} \Bigr) =0 \,,
\\
&{\mathrm{d}} \mathcal Q^{(9)}_{i_1\cdots i_{7-p}}=0\,,\quad
\mathcal Q^{(9)}_{i_1\cdots i_{7-p}} \equiv
\Exp{2\,(\alpha+1)\,\tilde{\phi}}\, \tilde{g}_{i_1j_1}\cdots\tilde{g}_{i_{7-p}j_{7-p}}\, \tilde{*}\, \mathcal Q^{(1)\,j_1\cdots j_{7-p}}
\equiv {\mathrm{d}}\mathcal A^{(8)}_{i_1\cdots i_{7-p}}\,.
\end{align}
If we regard the dual potential $\mathcal A^{(8)}_{i_1\cdots i_{7-p}}$ as a fundamental field,
the dual action is given by
\begin{align}
\begin{split}
S\bigl[\tilde{g}_{ij},\,\tilde{\phi},\,\mathcal A^{(8)}_{i_1\cdots i_{7-p}}\bigr]
&= \frac{1}{2\kappa_{10}^2}\,
\int \biggl[\Exp{-2\phi}\,\bigl(\tilde{*}\,\tilde{R} + 4\,{\mathrm{d}} \phi\wedge \tilde{*}\,{\mathrm{d}} \phi\bigr)
\\
&\quad\qquad\qquad
- \frac{\Exp{2\,(\tilde{\alpha}+1)\,\tilde{\phi}}}{2\,(7-p)!}\,\tilde{g}^{i_1j_1}\cdots \tilde{g}^{i_{7-p}j_{7-p}}\,
{\mathrm{d}} \mathcal A^{(8)}_{i_1\cdots i_{7-p}}\wedge \tilde{*}\, {\mathrm{d}} \mathcal A^{(8)}_{j_1\cdots j_{7-p}}\biggr] \,,
\end{split}
\end{align}
where we defined $\tilde{\alpha}\equiv -\alpha - 2$\,.
We can add the Wess-Zumino term of the exotic $p^{7-p}_{-\alpha}$-brane
extending in the $x^{\mathbf{r}_1},\cdots,x^{\mathbf{r}_p}$-directions and
smeared over the $x^{\mathbf{s}_1},\cdots,x^{\mathbf{s}_{7-p}}$-directions:
\begin{align}
S_{\text{WZ}} &= -\mu_{p^{7-p}_{-\alpha}}\,\sum_{\mathbf{s}_1,\cdots,\mathbf{s}_{7-p}}\int_{\mathcal M_{p+1}\times T^{7-p}_{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}}\frac{n^{\mathbf{s}_1\cdots\mathbf{s}_{7-p}}}{(7-p)!}\,\bm{\iota}\mathcal A_{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}^{(8)}\wedge \frac{{\mathrm{d}} x^{\mathbf{s}_1}\wedge\cdots\wedge{\mathrm{d}} x^{\mathbf{s}_{7-p}}}{(2\pi R_{\mathbf{s}_1})\cdots (2\pi R_{\mathbf{s}_{7-p}})}
\nonumber\\
&= -\mu_{p^{7-p}_{-\alpha}}\,\int\frac{1}{(7-p)!}\, \mathcal A_{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}^{(8)}\wedge \bm{\delta}^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}(x-X(\xi)) \,.
\end{align}
Then, taking variation, we obtain the following Bianchi identity as the equation of motion:
\begin{align}
{\mathrm{d}}^2 \mathcal A^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}
= 2\kappa_{10}^2\,\mu_{p^{7-p}_{-\alpha}}\,\frac{n^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}}{(2\pi R_{\mathbf{s}_1})\cdots (2\pi R_{\mathbf{s}_{7-p}})}\,
\delta^2(x-X(\xi))\,{\mathrm{d}} x^1\wedge{\mathrm{d}} x^2 \,.
\end{align}
If we choose $n^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}=1$ and integrate the equation, we obtain
\begin{align}
\sigma = \int{\mathrm{d}}^2 \mathcal A^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}
= \frac{2\kappa_{10}^2\,\mu_{p^{7-p}_{-\alpha}}}{(2\pi R_{\mathbf{s}_1})\cdots (2\pi R_{\mathbf{s}_{7-p}})} \,,
\end{align}
where we used $\mathcal A^{\mathbf{s}_1\cdots \mathbf{s}_{7-p}}=\rho_1$\,.
From this relation and the value of $\sigma$ given in Table \ref{tab:sigma},
we can confirm that $\mu_{p^{7-p}_{-\alpha}}$ is indeed equal to the tension of the exotic brane:
\begin{align}
\mu_{p^{7-p}_{-\alpha}} = \frac{\sigma\,(2\pi R_{\mathbf{s}_1})\cdots (2\pi R_{\mathbf{s}_{7-p}})}{(2\pil_s)^7\,l_s\,g_s^2}
= \frac{M_{p^{7-p}_{-\alpha}}}{(2\pi R_{\mathbf{r}_1})\cdots (2\pi R_{\mathbf{r}_{p+1}})}\,,
\end{align}
where we used $2\kappa_{10}^2=(2\pil_s)^7\,l_s\,g_s^2$\,.
\section{Summary and discussions}
\label{sec:summary}
In this paper, we have presented (truncated) effective actions for various polyvectors
whose magnetic sources can be identified with the exotic branes.
Requiring the existence of compact isometry directions,
which are required for the existence of exotic branes,
we showed that the effective actions can be derived
from the standard (truncated) supergravity actions.
In each theory, we found two solutions with either the magnetic or the electric charge associated with the non-geometric flux.
The former solution corresponds to an exotic-brane background
while the latter corresponds to a new instanton background.
By taking account of the $U$-duality symmetry of the string theory,
all defect branes including the standard branes and exotic branes should be treated as equals.
However, in the standard formulation of the supergravity,
only the backgrounds of the standard branes are well described globally.
Contrarily, in the reformulation of supergravity presented in this paper,
the background fields of the exotic branes are globally defined
while those of the standard branes are not single-valued.
In this sense, the effective theory considered in this paper is complementary to the standard supergravity.
Our reformulation is still not complete and we should investigate a further generalization
so as to allow for general backgrounds with multiple non-geometric fluxes.
Such generalization is necessary if we consider, for example,
the background representing a bound state of $p$ $5^2_2$-branes and $q$ $5^2_3$-branes,
which is the exotic dual of the $(p,q)$ five-brane (i.e., a bound state of $p$ NS5-branes and $q$ D5-branes).
In addition, there is another direction of generalization.
The action \eqref{eq:action-general} presented in this paper can be applied only for defect backgrounds
with isometries in the $03\cdots 9$-directions.
In the absence of these isometries, the action should be modified, like the action of the $\beta$-supergravity \eqref{eq:beta-action}
or its Ramond-Ramond counterpart \eqref{eq:gamma-action}.
Since the equations of motion derived from the simple action \eqref{eq:action-general}
do not coincide with those derived from the complete action
(i.e., the action \eqref{eq:beta-action} or \eqref{eq:gamma-action})
even though we impose the assumptions \eqref{eq:assumptions}
\footnote
We would like to thank David Andriot for pointing out this issue.
In the case of the $\beta$-supergravity,
if we assume that any derivative $\partial_i$ contracted with
$\beta^{ij}$ vanishes, the dilaton equation of motion and
the Einstein equations (see (1.28) and (1.29) of \cite{Andriot:2013xca})
coincide with those derived from the simple action \eqref{eq:beta-action}
(note that $\widecheck{\mathcal R}^{ij}= (1/4)\,\bigl(Q^{ikl}\,Q^{j}{}_{kl}-2\,Q^{kil}\,Q_k{}^j{}_l\bigr)$ and $\widecheck{\mathcal R}=-(1/4)\,Q^{ijk}\,Q_{ijk}$
can be shown by using the simplifying assumption).
However, the equation of motion for the $\beta$-field (see (1.30) of \cite{Andriot:2013xca})
does not coincide with \eqref{eq:beta-Bianchi},
although the difference disappears in the $5^2_2$ background.
}
it will be important to find the complete actions for any value of $p$
and check whether the exotic-brane backgrounds indeed satisfy
the equations of motion derived from the complete actions
\footnote{In the case of the $5^2_2$-brane (or the $Q$-brane), the background is shown to satisfy the equations of motion derived from the complete action
(see Appendix D.1 of \cite{Andriot:2014uda}).}
It will be also interesting to describe various non-geometric backgrounds with less isometries,
such as the background of the ``NS5-brane localized in winding space'' constructed in \cite{Berman:2014jsa} (see also the references therein),
as solutions of the complete theory without simplifying assumptions.
Further, it will be important to establish a formulation in which the background of an arbitrary defect brane
can be equally described globally;
that is, a manifestly $U$-duality covariant formulation of the supergravity theory.
A promising approach in this direction is taken in DFT,
which can globally describe both the usual brane (i.e.~NS5-brane) and the exotic brane (i.e.~$5^2_2$-brane),
and can reproduce the standard supergravity action or the $\beta$-supergravity action as a special limit.
Although DFT has already been generalized to incorporate the Ramond-Ramond fields \cite{Hohm:2011dv,Jeon:2012kd,Jeon:2012hp},
we cannot globally describe $U$-folds (such as the $\text{D} p_{7-p}$ background) in the framework of DFT,
since the generalized coordinate transformations in DFT do not include general $U$-dual transformations.
Recently, several generalizations of DFT have been studied in various papers
(see e.g.~\cite{Berman:2010is,Berman:2011kg,Berman:2011cg,Berman:2011jh,Berman:2012uy,Berman:2012vc,Park:2013gaj,Hohm:2013jma,Park:2014una},
and \cite{Hohm:2013pua,Hohm:2013vpa,Hohm:2013uia,Godazgar:2014nqa,Hohm:2014fxa,Blair:2014zba,Berman:2014hna,Musaev:2014lna}
where the \emph{exceptional field theory} has been proposed and studied),
which will be possible to describe all exotic-brane backgrounds globally.
It will be interesting to derive the effective theories proposed in this paper
as some special limits of these theories.
It will be also important to investigate a reformulation of the effective worldvolume theory of exotic branes
by using the newly introduced background fields $\bigl(\tilde{g}_{ij},\,\tilde{\phi},\,\mathcal A^{i_1\cdots i_{7-p}}\bigr)$.
More generally, it will be important to find a manifestly $U$-duality covariant formulation for
the effective worldvolume theory of exotic branes.
In our reformulation of the supergravity where a $p$-vector is regarded as a fundamental field,
there naturally appears the background of an instanton that electrically couples to a $p$-vector.
Depending on the type of the $p$-vector (i.e.~$\beta^{ij}$, $\gamma^{i_1\cdots i_{7-p}}$, or $\beta^{i_1\cdots i_6}$),
the value of the on-shell action for the instanton background
is expected to be proportional to $g_s^{\tilde{\alpha}}$ with $\tilde{\alpha}=0,1,2$.
Since these instantons have not been studied well,
it will be important to analyze their properties further.
Since the $\mathbf{I}_1^{7-p}$ backgrounds \eqref{eq:gamma-instantons} are similar
to the backgrounds of the D$p$-instantons (see \cite{Hull:1998vg} and references therein),
it will be natural to expect that the instantons $\mathbf{I}_1^{7-p}$ are the exotic dual of the D$p$-instantons,
and $\mathbf{I}^2_0$ and $\mathbf{I}^6_{2}$ are their $S$-dual objects.
\section*{Acknowledgments}
We acknowledge support by the National Research Foundation of Korea (NRF-MSIP)
grants 2005-0093843, 2010-220-C00003 and 2012-K2A1A9055280.
We would like to thank Yolanda Lozano and Soo-Jong Rey for helpful discussions.
|
\section{Introduction}
\input{tex/introduction.tex}
\section{Preliminaries}
\label{prelims sec}
\input{tex/preliminaries.tex}
\section{One point estimates for the inverse maps}
\label{inverse sec}
\input{tex/one-point-inverse.tex}
\section{One point estimates for the forward maps}
\label{time infty sec}
\input{tex/time-infinity.tex}
\section{Upper bounds for multifractal and integral means spectra}
\label{haus upper sec}
\input{tex/ubd.tex}
\section{Two point estimate}
\label{2pt sec}
\input{tex/two-point.tex}
\section{Lower bounds for multifractal and integral means spectra}
\label{haus lower sec}
\input{tex/dim-lower.tex}
\subsection{Lower bound for the integral means spectrum}
\label{ims lower sec}
\input{tex/ims-lower-fixed.tex}
\subsection{Basic notations}
\label{basic notations}
Given two variables $a$ and $b$, we say $b = o_a(1)$ if $b \rightarrow 0$ as $a \rightarrow 0$ (or as $a\rightarrow \infty$, depending on the context) and we say $b = O_a(1)$ if $b$ is bounded above by an $a$-independent constant for sufficiently small (or sufficiently small, depending on thecontext) values of $a$. We usually allow $o_a(1)$ and $O_a(1)$ terms to depend on certain parameters other than $a$, but not on others. We will describe this dependence as needed.
We say that $a \preceq b$ (resp.\ $a \succeq b$) if there is a constant $c$ which does not depend on the main parameters of interest such that $a \leq c b$ (resp.\ $a \geq c b$). We say $a \asymp b$ if $a \preceq b$ and $a \succeq b$. As in the case of $o_a(1)$ and $O_a(1)$ above, we usually allow the implicit constants in $\preceq, \succeq$, and $\asymp$ to depend on certain parameters, but not on others, and we describe this dependence as needed.
For a point $z\in \mathbf C$ and $r> 0$, we write $B_r(z)$ for the ball of radius $r$ centered at $z$. More generally, for a set $A\subset \mathbf C$, we write $B_r(A) = \bigcup_{z\in A} B_r(z)$.
For a curve $\eta : [0,T] \rightarrow \mathbf C$, we will often use the abbreviation
\begin{equation} \label{eta^t}
\eta^t = \eta([0,t]) .
\end{equation}
Furthermore, when there is no risk of ambiguity we will simply write $\eta$ for the entire image of $\eta$.
For a domain $D$ and $z\in D$, we write $\operatorname{hm}^z( \cdot ; D)$ for the harmonic measure from $z$ in $D$. That is, for $A\subset\partial D$, $\operatorname{hm}^z(A ; D)$ is the probability that a Brownian motion started from $z$ exits $D$ in $A$.
If $D' = D\setminus \eta$ for some non-self-crossing curve $\eta \in \overline D$ and $z$ is a point on $\eta$ which is visited only once, we will write $z^-$ (resp.\ $z^+$) for the prime end of $D$ corresponding to the left (resp.\ right) side of $z$. When we use this notation, our curve $\eta$ will have an obvious orientation and ``left" and ``right" are as viewed by someone walking along $\eta$ in the forward direction.
We will also use the following notation.
\begin{notation}
\label{arc notation}
Given a Jordan domain $D$ and $x,y\in\partial D$, we write $[x,y]_{\partial D} $ for the closed counterclockwise arc from $x$ to $y$ in $\partial D$. We similarly define the open arc $(x,y)_{\partial D}$ and the half-open arcs $(x,y]_{\partial D}$ and $[x,y)_{\partial D}$.
\end{notation}
\subsection{Reverse continuity conditions}
\label{G prelim}
\subsubsection{In the upper half plane}
Here we introduce a regularity condition which will arise frequently in the remainder of the paper.
\begin{defn} \label{mathcal M def}
We denote by $\mathcal M$ the set of increasing functions $\mu : (0,\infty) \rightarrow (0,\infty)$ with $\lim_{\delta\rta0}\mu(\delta) = 0$.
\end{defn}
\begin{defn}
\label{G def}
Let $f$ be a map from a subdomain $D$ of $\mathbf H$ into $\mathbf H$. For $\mu\in\mathcal M$, let $G(f,\mu)$ be the event that the following occurs. For any $\delta>0$ and any $x,y\in \mathbf R \cap \partial D$ with $|x| ,|y| \leq \delta^{-1}$ and $|x-y| \geq \delta $, we have $|f (x) | , |f (y)| \leq \mu(\delta)^{-1}$ and $|f (x) - f (y)| \geq \mu(\delta)$.
\end{defn}
The statement that $\mathcal G(f,\mu)$ holds is the same as the statement that $f^{-1}$ has a certain $\mu$-dependent modulus of continuity on $f(\mathbf R \cup \infty)$, with $\mathbf R \cup \infty$ given the one-point compactification topology.
We note that
\begin{equation} \label{G compose}
G(f,\mu_1) \cap G(g,\mu_2) \quad \Rightarrow \quad G(g\circ f , \mu_2 \circ \mu_1) .
\end{equation}
We are interested in the condition $G(f,\mu)$ (and the analogous conditions in the next subsection) for two reasons. The first is that these conditions imply bounds on the distance from certain subsets of $\partial D$ to certain subsets of $\mathbf R$ (or $\partial\mathbf D$ in the setting of the next subsection) and on its diameter (see Lemmas~\ref{G implies U} and~\ref{G dist} below).
The second reason for our interest in the condition of Definition~\ref{G def} is as follows. We will often want to study conformal maps which are normalized by specifying the images of certain marked boundary points. When composing various maps, our marked points might be mapped to somewhere other than where we want them to go. So, we will frequently need to apply a conformal automorphism (of $\mathbf D$ or $\mathbf H$) at the end of our arguments to move the marked points to their desired positions. The condition $G(\cdot,\mu)$ ensures that the images of the marked points are not too close together, and so allows us to control the derivative of this conformal automorphism.
\begin{lem}\label{G implies U}
Let $\eta$ be a simple curve in $\mathbf H$ parametrized by capacity which does not hit $\mathbf R$. Write $\eta^t := \eta([0,t])$. Let $f_t : \mathbf H\setminus \eta^t \rightarrow \mathbf H$ be the centered Loewner maps for $\eta$. Fix $T \in (0,\infty)$ and suppose that for some $\mu \in \mathcal M$, we have
\begin{equation}\label{f_T big}
f_T(-\delta ) - f_T(0^-) \leq -\mu(\delta) \leq \mu(\delta) \leq f_T( \delta ) - f_T(0^+) ,\qquad \forall \delta> 0 .
\end{equation}
Then there is a $\mu'\in\mathcal M$ and a $d > 0$ depending only on $\mu$ and $T$ such that
\begin{equation}\label{diam small}
\operatorname{diam} \eta^T \leq d \quad \operatorname{and} \quad \text{$\forall \delta>0$, $\forall z\in \eta^T$ with $|\operatorname{Re} z| \geq \delta $, we have $\operatorname{Im} z \geq \mu'(\delta)$}.
\end{equation}
Conversely, if \eqref{diam small} holds for some $d > 0$ and some $\mu'\in\mathcal M$, we can find $\mu\in\mathcal M$ depending only on $d$ and $\mu'$ such that $ G(f_T ,\mu)$ holds.
\end{lem}
Note that it is clear that $G(f_T, \mu)$ implies \eqref{f_T big}, so Lemma~\ref{G implies U} implies in particular that \eqref{diam small} holds for some $d$ and $\mu'$ depending only on $\mu$ whenever $G(f_T,\mu)$ occurs.
\begin{proof}[Proof of Lemma~\ref{G implies U}]
Let $\operatorname{hm}^\infty_T = \operatorname{hm}^\infty(\cdot ; \mathbf H\setminus \eta^T)$ denote harmonic measure from $\infty$ in $\mathbf H \setminus\eta^T$, so for a set $I\subset \partial (\mathbf H\setminus \eta^T )$ (viewed as a collection of prime ends),
\begin{equation*}
\operatorname{hm}^{\infty}_T I := \lim_{y\rightarrow \infty} y\mathbf P^{iy} (B_\tau \in I)
\end{equation*}
for $B$ a Brownian motion and $\tau $ its exit time from $\mathbf H\setminus\eta^T$. It follows from conformal invariance of Brownian motion that for any $I\subset \partial (\mathbf H\setminus \eta^T)$,
\begin{equation} \label{hm length}
\operatorname{hm}^\infty_T(I ) = \frac1\pi \operatorname{length} f_T(I) ,
\end{equation}
where by $\operatorname{length}$ we mean Lebesgue measure.
Now, assume \eqref{f_T big} holds.
For any $r>0$ and $x\in\mathbf R$, the harmonic measure from $\infty$ in $\mathbf H$ of the line segment $[x,x+ir]$ from $x$ to $x+ir$ is a constant depending only on $r$. For $\delta > 0$, we can find $r = r(\delta) > 0$ such that this constant is $< \pi \mu(\delta)$. If $\eta^T$ contains a point $x+i y$ with $x \geq \delta$ and $y\leq r$, then $\operatorname{hm}^\infty_T([0,\delta] ) \leq \operatorname{hm}^\infty_T([x , x+ir] ) < \pi \mu(\delta)$. This contradicts our hypothesis on \eqref{f_T big} and the relation \eqref{hm length}. A similar statement holds if we instead consider $x\leq -\delta$. Hence each point of $\eta^T$ with real part $\geq \delta$ in absolute value has imaginary part $\geq r$. This proves the second part of \eqref{diam small} with $\mu'(\delta) = r$.
For the first part of \eqref{diam small}, fix $\delta> 0$. Denote by $S_{\delta}$ the set of points in $\mathbf H$ with real part at least $\delta$. Then we have
\begin{equation} \label{hm S_delta}
\operatorname{hm}^\infty_T(\eta^T \cap S_{\delta}) \leq \frac{1}{\mu'(\delta)} \lim_{y\rightarrow \infty} y\mathbf E^{iy}\!\left( \operatorname{Im} B_\tau \mathbf 1_{B_\tau \in \eta^T \cap S_{\delta}} \right) .
\end{equation}
By \cite[Proposition~3.38]{lawler-book} we have
\begin{equation} \label{hcap gamma formula}
T = \operatorname{hcap} \eta^T = \lim_{y\rightarrow \infty} y\mathbf E^{iy} (\operatorname{Im} B_\tau)
\end{equation}
so \eqref{hm S_delta} is at most $ T/\mu'(\delta)$. On the other hand, \eqref{hcap gamma formula} and the Beurling estimate imply that $\sup_{z\in \eta^T} \operatorname{Im} z$ is bounded above by a constant $C_0$ depending only on $T$. The harmonic measure from $\infty$ in $\mathbf H$ of $[-\delta , \delta] \times [0,C_0]$ is at most a constant $C_1$ depending only on $\delta$ and $T$. Therefore
\[
\operatorname{hm}^\infty_T ( \eta^T) \leq T/\mu'(\delta) + C_1 .
\]
By \cite[equation 3.13]{lawler-book}, this implies $\operatorname{diam} \eta^T $ is bounded above by a constant depending only on $\mu$ and $T$.
Conversely, suppose \eqref{diam small} holds. For $\delta >0$, let $U_\delta$ be the set of points in $z\in\mathbf H$ with $|z| \leq d$ and either $|\operatorname{Re} z | \geq \delta/2$ or $\operatorname{Im} z \leq \mu'(\delta/2)$. The harmonic measure from $\infty$ of each sub-interval of $[\delta/2 , \delta^{-1}] \cup [-\delta^{-1} , -\delta/2]$ in $\mathbf H\setminus U_\delta$ of length $\delta/2$ is at least some constant $\mu_0(\delta)$ depending only on $\delta/2$ and $\mu'(\delta/2)$. By \eqref{hm length}, this implies that the length of the image of such an interval under $f_T$ is at least a $\pi \mu_0(\delta)$. On the other hand, \cite[Proposition~3.46]{lawler-book} implies that we can find $\mu_1(\delta) > 0$ depending only on $\delta$ and $d$ such that $|f_T(x)| \leq \mu_1(\delta)^{-1}$ for each $x\in [-\delta^{-1} , \delta^{-1}]$. This proves that $\mathcal G(f_T ,\mu)$ holds with $\mu = (\pi \mu_0)\vee\mu_1$.
\end{proof}
\subsubsection{In the disk} \label{G disk sec}
The following is the analogue of Definition~\ref{G def} for the unit disk $\mathbf D$.
\begin{defn}\label{G infty def}
Let $A\subset \mathbf D$ be a subdomain and let $I\subset \partial \mathbf D \cap \partial D$. Let $f : D\rightarrow \mathbf D$ be a conformal map. Let $\mu\in\mathcal M$ (Definition~\ref{mathcal M def}). We say that $\mathcal G_I(f, \mu)$ occurs if the following is true. For each $\delta >0$ and each $x,y\in I$ with $|x-y| \geq \delta$, we have $|f(x) - f(y) |\geq \mu(\delta)$. We abbreviate
\[
\mathcal G (f,\mu) = \mathcal G_{\partial\mathbf D\cap \partial D} (f,\mu).
\]
\end{defn}
We also make the following definition.
\begin{defn}\label{G' def}
Let $A\subset \overline{\mathbf D}$ be a closed set and $I\subset \overline{\partial \mathbf D \setminus A}$. (Oftentimes we will take $I$ to be a closed arc with endpoints in $A$, or a finite union of such arcs.) We say that $\mathcal G_I'(A , \mu)$ occurs if the following is true. For each $\delta > 0$, $A$ lies at distance at least $\mu(\delta)$ from $I\setminus B_\delta(I\cap A)$. We write
\[
\mathcal G'(A , \mu) = \mathcal G_{\overline{\partial \mathbf D \setminus A}}(A , \mu).
\]
\end{defn}
\begin{remark} \label{G pos}
We will frequently find ourselves in the following situation. Suppose we are given a deterministic arc $I\subset \partial \mathbf D$, a random closed subset $A\subset \overline{\mathbf D}$ with $I \subset\overline{\partial \mathbf D \setminus A}$ a.s., and a deterministic $\epsilon > 0$. In this case we can find (using monotonicity) a deterministic $\mu \in \mathcal M$ for which $\mathbf P\!\left(\mathcal G_I(A , \mu) \right) \geq 1-\epsilon$ where $\mathbf P$ is typically the law of SLE.
\end{remark}
The conditions of Definitions~\ref{G infty def} and~\ref{G' def} will serve as the main ``global regularity" conditions in our estimates starting from Section~\ref{time infty sec}. The relationship between the conditions $\mathcal G(\cdot)$ and $\mathcal G'(\cdot)$ is contained in the following lemma.
\begin{lem}
\label{G dist}
Let $A\subset \overline{\mathbf D}$ be a closed set and $I = [x,y]_{\partial\mathbf D}$ be an arc contained in $\overline{\partial \mathbf D \setminus A}$. Let $m\in (x,y)_{\partial\mathbf D}$ and suppose that $|x-m|$ and $|y-m|$ are each at least $\Delta>0$. Let $D $ be the connected component of $\mathbf D\setminus A$ containing $I$ on its boundary. Let $\Phi : D \rightarrow \mathbf D$ be the unique conformal map taking $x$ to $-i$, $y$ to $i$, and $m$ to~$1$.
\begin{enumerate}
\item For each $\mu\in\mathcal M$, there exists $\mu'\in\mathcal M$ depending only on $\mu$ and $\Delta$ such that if $\mathcal G_I (\Phi , \mu)$ occurs, then $\mathcal G_I'(A , \mu')$ occurs. \label{G to dist}
\item Conversely, suppose $I'\subset I$ and $\mathcal G_{I'}'(A , \mu)$ holds for some $\mu\in\mathcal M$. There is a $\mu'\in\mathcal M$ depending only on $\mu$ and $\Delta$ such that $\mathcal G_{I'}(\Phi , \mu')$ holds. In fact, the following superficially stronger statement is true. For each $\delta > 0$, $\Phi$ is Lipschitz continuous on $I'\setminus (B_\delta(x) \cup B_\delta(y))$ and $\Phi ^{-1}$ is Lipschitz continuous on $\Phi( I'\setminus (B_\delta(x) \cup B_\delta(y)) ) $ with Lipschitz constants depending only on $\mu(\delta)$, $\delta$, and $\Delta$. \label{dist to G}
\end{enumerate}
\end{lem}
\begin{proof}
The basic idea of the proof is similar to that of Lemma~\ref{G implies U}, but we consider harmonic measure from $m$ rather than harmonic measure from $\infty$.
Fix $\delta>0$. Let $x_{\delta }$ and $y_{\delta }$ be the unique points of $I$ lying at distance $\delta $ from $x$ and $y$, respectively. Let $\widehat D$ be the radial reflection of $\widehat D$ across $I$, viewed as a subset of the Riemann sphere. Extend $\Phi $ to $\widehat D $ by Schwarz reflection. Then $\Phi$ maps $\widehat D $ into $\mathbf C\setminus [i,-i]_{\partial\mathbf D}$, and maps $I$ to $[-i,i]_{\partial\mathbf D}$. Suppose $\delta>0$. Let $\widehat D_\delta = \widehat D\setminus [y_{\delta } ,y]_{\partial\mathbf D}$. Let $\widetilde y_\delta := \Phi(y_\delta)$. Then $\widetilde y_{\delta} $ is determined by the condition that the harmonic measure of $[y_{\delta},i]_{\partial\mathbf D}$ from $m$ in $\widehat D_\delta$ equals the harmonic measure of the side of $[\widetilde y_{\delta} , i]_{\partial\mathbf D}$ closer to 0 from 1 in $(\mathbf C \cup \infty)\setminus [y_{\delta}',-i]_{\partial\mathbf D}$.
If $\mathcal G^*_I(\Phi ,\mu)$ occurs, then $\widetilde y_{\delta} $ lies at distance at least $\mu(\delta)$ from $i$, which means that the harmonic measure of $[y_{\delta },y]_{\partial\mathbf D}$ from 1 in $\widehat D_\delta$ is at least some constant $\epsilon > 0$ depending only on $\mu(\delta)$. By symmetry, the same holds for $[x ,x_{\delta }]_{\partial\mathbf D}$.
By the Beurling estimate, we can find some $\zeta_0 > 0$ depending only on $\epsilon$ such that $\operatorname{dist}(m ,A) \geq \zeta_0$. We can find a $\zeta_1 >0$ such that if $z\in [x_{\delta } , y_{\delta }]_{\partial\mathbf D}$ lies at distance at least $ \zeta_0$ from $m$, then the probability that a Brownian motion started from $m$ hits $B_{\zeta_1}( z)$ before hitting $[i,-i]_{\partial\mathbf D}$ is at most $\epsilon$. If $\operatorname{dist}(z,A) < \zeta_1$ for such a $z$, then a Brownian motion started from 1 must hit $B_{\zeta_1}(z)$ before hitting either $[y_{\delta } , y]_{\partial\mathbf D}$ or $[x , x_{\delta } ]_{\partial\mathbf D}$. Hence we must have $\operatorname{dist}(z , A) \geq \zeta_1 \wedge \zeta_0$ for each $z\in [x_{\delta } , y_{\delta }]_{\partial\mathbf D}$. This proves assertion~\ref{G to dist} with $\mu'(\delta) = \zeta_1 \wedge\zeta_0$.
Conversely, suppose $I'\subset I$ and $\mathcal G_{I'}'(A , \mu)$ holds for some $\mu\in\mathcal M$. For $\delta >0$ let $x_\delta'$ be either $x_\delta$ (as defined just above) or the endpoint of $I'$ closest to $x$, whichever is closest to $x$. Define $y_\delta'$ similarly. A Brownian motion started from any point of $ [x_{\delta}' , y_{\delta }' ]_{\partial\mathbf D}$ as a positive probability depending only on $\delta $, $\mu(\delta)$, and $\Delta$ to stay within distance $\mu(\delta)$ of $I$ until it hits $ [y_{\delta }' , y]_{\partial\mathbf D}$ (resp.\ $[x,x_{\delta }']_{\partial\mathbf D}$). By the Beurling estimate there is a $\mu_0(\delta)$ depending only on $\mu(\delta)$, $\delta$, and $\Delta$ such that $\Phi( [x_{\delta}' , y_{\delta }' ]_{\partial\mathbf D} )$ lies at distance at least $\mu_0(\delta)$ from $[i,-i]_{\partial\mathbf D}$.
It remains to establish the Lipschitz continuity statement. For this, we observe that for any $z\in [x_{\delta}' , y_{\delta }' ]_{\partial\mathbf D}$, the Koebe quarter theorem implies
\[
\frac{ \operatorname{dist}(\Phi (z) , [i,-i]_{\partial\mathbf D}) }{4 \operatorname{dist}(z,A) \wedge \delta } \leq |\Phi '(z)| \leq \frac{ 4 \operatorname{dist}(\Phi (z) , [i,-i]_{\partial\mathbf D}) }{ \operatorname{dist}(z,A) \wedge \delta } .
\]
Hence
\[
\frac{ \mu_0(\delta) }{8 } \leq |\Phi '(z)| \leq \frac{ 8}{ \mu(\delta)\wedge\delta } .
\]
So, $|\Phi '|$ is bounded above and below by positive constants on $[x_\delta' ,y_\delta']_{\partial\mathbf D}$ depending only on $\mu(\delta)$, $\delta$, and $\Delta$ which establishes the desired Lipschitz continuity.
\end{proof}
\subsection{Schramm-Loewner evolution}
\label{sle prelim}
Let $t\mapsto W_t$ be a continuous function on $[0,\infty)$. The \emph{chordal Loewner equation} is the ordinary differential equation
\begin{equation}\label{loewner eqn}
\partial_t g_t(z) = \frac{2}{g_t(z) - W_t} ,\qquad g_0(z) = z .
\end{equation}
A solution to \eqref{loewner eqn} is a family of conformal maps $\{g_t : t\geq 0\}$ from subdomains of $\mathbf H$ to $\mathbf H$, satisfying the hydrodynamic normalization $\lim_{z\rightarrow \infty} (g_t(z) - z) = 0$. The complements $(K_t)$ of the domains of $(g_t)$ in $\mathbf H$ are an increasing family of closed subsets of $\mathbf H$ called the \emph{hulls} of the process. The \emph{centered Loewner maps} corresponding to $(g_t)$ are defined by
\[
f_t := g_t -W_t .
\]
A chordal \emph{Schramm-Loewner evolution} with parameter $\kappa > 0$ ($\operatorname{SLE}_\kappa$) is the random evolution obtained by solving \eqref{loewner eqn} where the driving process $W$ is $\sqrt\kappa$ times a Brownian motion. It can be shown \cite{schramm-sle} that this Loewner evolution is generated by a curve which we typically denote by $\eta$.
Chordal $\operatorname{SLE}_\kappa$ on other domains is defined by conformal mapping. We refer the reader to \cite{lawler-book} or \cite{werner-notes} for a more detailed introduction to SLE.
More generally, suppose we are given a vector of real weights $\overline \rho = (\rho^1 , \ldots , \rho^n )$ and a collection of points $z^1 , \ldots , z^n \in \mathbf H$.
Chordal $\operatorname{SLE}_\kappa(\underline \rho)$ is the random evolution obtained by solving \eqref{loewner eqn} with the driving function $W$ part of the solution to the system of SDE's
\begin{equation}\label{sle kappa rho}
dW_t = \sqrt\kappa dB_t + \sum_{i=1}^{n} \operatorname{Re} \frac{\rho^{i}}{W_t - V_t^i} dt, \qquad dV_t^{i } = \frac{2}{V_t^i - W_t} dt , \qquad W_0 = y \qquad V_0^{i } = z^{i } .
\end{equation}
The points $z^i$ are called the \emph{force points}. See \cite{lsw-restriction, sw-coord, ig1} for more on $\operatorname{SLE}_\kappa(\underline\rho)$.
We will also need to consider the \emph{reverse Loewner equation}. This is the ODE
\begin{equation}\label{reverse loewner eqn}
\partial_t g_t(z) = -\frac{2}{g_t(z) - W_t} ,\qquad g_0(z) = z ,
\end{equation}
whose solution is a family of conformal maps from $\mathbf H$ to sub-domains of $\mathbf H$. Reverse $\operatorname{SLE}_\kappa$ is obtained by taking $W_t$ to be $\sqrt\kappa$ times a Brownian motion. For each time $t$, the time $t$ centered Loewner map of a reverse $\operatorname{SLE}_\kappa$ has the same law as the inverse of the time $t$ centered Loewner map of a forward $\operatorname{SLE}_\kappa$ \cite[Lemma~3.1]{schramm-sle}.
We will also need to consider reverse $\operatorname{SLE}_\kappa(\underline \rho)$ with force points $z^1,\ldots,z^n$, which is obtained by solving \eqref{reverse loewner eqn} with the driving function $W$ part of the solution to the system of SDE's
\begin{equation*}
dW_t = \sqrt\kappa dB_t + \sum_{i=1}^{n} \operatorname{Re} \frac{\rho^{i}}{W_t - V_t^i} dt, \qquad dV_t^{i } = -\frac{2}{V_t^i - W_t} dt , \qquad W_0 = y \qquad V_0^{i } = z^{i } .
\end{equation*}
For a general $\underline\rho$ we do not have as simple a relation between forward and reverse $\operatorname{SLE}_\kappa(\underline\rho)$ as we do for ordinary $\operatorname{SLE}_\kappa$. However, there are various forward and reverse symmetries, some of which are discussed in \cite{wedges, shef-zipper}.
Throughout most of the rest of this paper we will fix $\kappa \in (0,4]$ and we will not always make dependence on $\kappa$ explicit.
\subsection{Gaussian free fields}
\label{gff prelim}
For some of our results, we will make use of couplings of $\operatorname{SLE}_\kappa$ with Gaussian free fields. In this section we give some basic background about the latter object.
Let $D$ be a domain in $\mathbf C$ with harmonically non-trivial boundary (i.e.\ a Brownian motion started in $D$ a.s.\ exits $D$ in finite time). We denote by $H(D)$ the Hilbert space completion of the subspace of $C^\infty(\overline D)$ consisting of those smooth, real-valued functions $f$ such that
\[
\int_{D} |\nabla f(z)|^2 \, dz <\infty ,\qquad \int_D f(z) \, dz = 0
\]
with respect to the Dirichlet inner product
\begin{equation}\label{dirichlet prod}
(f,g)_\nabla = \frac{1}{2\pi} \int_D \nabla f(z) \cdot \nabla g(z) \, dz .
\end{equation}
A \textit{free-boundary Gaussian free field} (GFF) on $D$ is a random distribution (in the sense of Schwartz) on $D$ given by the formal sum
\begin{equation}\label{h}
h = \sum_{j=1}^\infty X_j f_j
\end{equation}
where $\{f_j\}$ is an orthonormal basis for $H(D)$ and $(X_j)$ is a sequence of i.i.d.\ standard Gaussian random variables. It is defined as a pointwise function, but for each $g\in H(D)$, the formal inner product
\[
(h, g)_\nabla = \sum_{j=1}^\infty (f,g)_\nabla
\]
converges almost surely. Moreover, $(h,g)$ is a.s.\ defined for each fixed $g\in L^2(D)$ by the formula
\begin{equation}\label{by parts}
(h , g) = (h , -\Delta^{-1} g)_\nabla
\end{equation}
where $\Delta^{-1}$ denotes the inverse Laplacian with Neumann boundary conditions. More generally, this formula makes sense if $g$ is any distribution whose inverse Laplacian is in $H(D)$.
Similarly, one can define a \emph{zero-boundary GFF} on $D$ by replacing $H(D)$ with $H_0(D)$, defined as the Hilbert space completion of the space of smooth compactly supported functions on $D$ in the inner product \eqref{dirichlet prod}. A zero boundary GFF is defined without the need to make a choice of additive constant. A Gaussian free field with a given choice of boundary data on $\partial D$ is defined to be a zero boundary GFF plus the harmonic extension of the given boundary data to $D$.
If $V, V^\perp \subset H(D)$ are complementary orthogonal subspaces, then the formula \eqref{h} implies that $h$ decomposes as the sum of its projections onto $V$ and $V^\perp$. In particular, we can take $V$ to be the closure $H_0(D)$ of $C_c^\infty(D)$ in the inner product \eqref{dirichlet prod} and $V^\perp$ the set $\operatorname{Harm}_D$ of functions in $H(D)$ which are harmonic in $D$. This allows us to decompose a free boundary GFF as the sum of a zero boundary Gaussian free field and a random harmonic function $\mathfrak h$ on $D$, the latter defined modulo additive constant. We call these distributions the \emph{zero-boundary part} and \emph{harmonic part} of $h$, respectively.
We refer to \cite{shef-gff} and the introductory sections of \cite{ss-contour} and \cite{qle} for more details on GFF's.
\subsubsection{Reverse SLE/GFF coupling}
The following relation between free boundary GFFs and reverse $\operatorname{SLE}_\kappa(\underline\rho)$ is established in \cite[Section~4.2]{shef-zipper}. Let $(g_t)$ be the \emph{centered} Loewner maps of a reverse $\operatorname{SLE}_\kappa(\underline\rho)$ with force points $z^1,\ldots ,z^n$ as in Section~\ref{sle prelim}. Let $ h$ be a free boundary GFF on $\mathbf H$, independent from $(g_t)$.
For $t\geq 0$ let
\begin{equation*}
h_t = h \circ g_t + \frac{2}{\sqrt\kappa} \log |g_t(\cdot)| + \frac{1}{2\sqrt\kappa} \sum_{i=1}^n \rho^i G(g_t(z^i), g_t(\cdot) ) ,
\end{equation*}
where
\[
G(x,y) := -\log|x-y| - \log |\overline x- y|
\]
is the Green's function on $\mathbf H$ with Neumann boundary conditions. Let
\begin{equation}\label{Q def}
Q = \frac{2}{\sqrt\kappa} +\frac{\sqrt\kappa}{2} .
\end{equation}
Let $\tau$ be a stopping time for $\eta$ which is a.s.\ less than the first time $t$ that $f_t(z^i) =0$ for some $i$. Then \cite[Theorem~4.5]{shef-zipper} implies that $h_\tau + Q\log |g_\tau'| \overset{d}{=} h_0$, modulo additive constant.
\begin{figure}
\begin{center}
\includegraphics[scale=.8]{figures/drawing-reverse-coupling.pdf}
\caption{An illustration of the reverse SLE/GFF coupling in the case of a single force point (marked in blue). This is the case we will use in Section~\ref{inverse sec}.}
\end{center}
\end{figure}
There is also an analogue of the above coupling for a zero boundary GFF paired with a forward $\operatorname{SLE}_\kappa(\underline\rho)$, which we discuss in Section~\ref{ig prelim}.
\subsubsection{Estimates for the harmonic part}\label{gff lem sec}
In the course of proving our one-point estimate we will need some basic analytic lemmas about the harmonic part of a free boundary GFF which we will prove here.
\begin{lem}\label{harmonic cov}
Let $\mathfrak h$ be the harmonic part of a free boundary GFF on $\mathbf D$, normalized so that $\mathfrak h(0) = 0$. Then for any $z,w \in \mathbf D$, $\mathfrak h(z)$ and $\mathfrak h(w)$ are jointly Gaussian with means zero and covariance
\[
\mathbf E(\mathfrak h(z) \mathfrak h(w)) = -2 \log |1-z\overline w| .
\]
\end{lem}
\begin{proof}
For $n\geq 1$, let
\begin{equation}\label{phi psi}
\phi_n(z) = (2 / n)^{1/2} \operatorname{Re} z^n , \qquad \psi_n(z) = (2 / n)^{1/2} \operatorname{Im} z^n .
\end{equation}
Then $\{\phi_n , \psi_n : n\geq 1\} $ is an orthonormal basis for the set of harmonic functions on $\mathbf D$ in the Dirichlet inner product. So, by definition of the free boundary GFF, we can write
\begin{equation}\label{series formula}
\sum_{n=1}^\infty X_n \phi_n + \sum_{n=1}^\infty Y_n \psi_n ,
\end{equation}
where the $X_n$'s and $Y_n$'s are i.i.d.\ $N(0,1)$. From this expression, it follows that $(\mathfrak h(z) , \mathfrak h(w))$ is centered Gaussian for each $z,w\in \mathbf D$, and
\alb
\mathbf E(\mathfrak h(z) \mathfrak h(w)) &= \sum_{n=1}^\infty \phi_n(z)\phi_n(w) + \sum_{n=1}^\infty \psi_n(z)\psi_n(w) \\
&= 2\sum_{n=1}^\infty \frac{(\operatorname{Re} z^n)(\operatorname{Re} w^n) + (\operatorname{Im} z^n) (\operatorname{Im} w^n)}{n} \\
&= \sum_{n=1}^\infty \frac{(z\overline w)^n + (w\overline z)^n}{n} \\
&= -\log(1 - z\overline w) - \log (1-w\overline z)\\
&= -2\log|1 - z\overline w| .
\ale
\end{proof}
\begin{lem}\label{gff prob}
Let $h$ be a free boundary GFF on $\mathbf H$ with additive constant chosen so that its harmonic part vanishes at $a$ for some $a\in \mathbf H$. Let $A\subset \mathbf H$ be a deterministic hull lying at positive distance from $a$ and let $g : \mathbf H\rightarrow \mathbf H\setminus A$ be the inverse centered hydrodynamic map. Let $\widetilde h = h\circ g$ and let $(\widetilde h_\epsilon)$ be the circle average process for $\widetilde h$ (see \cite[Section 3.1]{shef-kpz} for more on the circle average process of a GFF). Fix $x\in \mathbf R$ and $\xi >1/2$. For any $\delta \geq \epsilon > 0$ we have
\begin{equation} \label{wt h_ep prob}
\mathbf P\!\left( |\widetilde h_\epsilon(x + i\delta )| > (\log \epsilon^{-1})^\xi \right) = o_\epsilon(1)
\end{equation}
with the $o_\epsilon(1)$ depending only on $x$, $a$, $\operatorname{diam} A$, $\xi$, and $\delta$, but uniform for $x$ in compact subsets of $\mathbf R$, $a$ in compact subsets of $\mathbf H$, and $\delta$ in compact subsets of $[\epsilon , \infty)$.
\end{lem}
\begin{proof}
Write $h = h^0 + \mathfrak h$, for $h^0$ a zero boundary GFF and $\mathfrak h$ an independent harmonic function. Let $\mathfrak h_A$ be projection of $h^0$ onto the set of functions which are harmonic on $\mathbf H\setminus A$ and let $h^0_A = h^0|_A - \mathfrak h_A$ be the zero-boundary part of $h^0|_A$. Then we can write
\begin{equation} \label{h|H minus A}
h|_{\mathbf H\setminus A} = h^0_A + \mathfrak h_A + \mathfrak h|_{\mathbf H\setminus A} ,
\end{equation}
with the three summands independent.
The function $g$ increases imaginary parts, so it follows from Lemma~\ref{harmonic cov} and a coordinate change to $\mathbf D$ that $\mathfrak h(g(x+i\delta))$ is centered Gaussian with variance $\leq 2\log \delta^{-1} + O_\epsilon(1)$.
By the Koebe distortion theorem, $|g'(x + i\delta)|$ is at least a constant depending only on $y$ times $\delta |g'(x + iy)|$ for any $y > \delta$. By \cite[Proposition~3.46]{lawler-book} and the Koebe quarter theorem, for large enough $y$ (depending only on $\operatorname{diam} A$), $|g'(x + iy)|$ is bounded above by a constant depending only on $\operatorname{diam} A$. By another application of the Koebe quarter theorem, we therefore have
\begin{equation} \label{dist succeq delta}
\operatorname{dist}(g(x+i \delta) , A) \succeq \delta^2 .
\end{equation}
It follows from \cite[Lemma~6.4]{ig1} that $\mathfrak h_A(g(x+i \delta ))$ is centered Gaussian with variance at most $2 \log \delta^{-1} + O_\epsilon(1)$.
By conformal invariance, $h_A^0 \circ g$ has the law of a zero boundary GFF on $\mathbf H$. By \eqref{dist succeq delta} and \cite[Proposition~3.1]{shef-kpz}, the circle average $(h_A^0 \circ g)_\epsilon(x+i\delta)$ is Gaussian with mean 0 and variance at most $2\log \epsilon^{-1} + O_\epsilon(1)$. By \eqref{h|H minus A},
\[
\widetilde h_\epsilon(x+i \delta ) = (h_A^0 \circ g)_\epsilon(x+i \delta) + \mathfrak h_A(g(x+i \delta)) + \mathfrak h(g(x+i\delta))
\]
is Gaussian with mean 0 and variance at most $6 \log \epsilon^{-1} + O_\epsilon(1)$.
We obtain \eqref{wt h_ep prob} from the Gaussian tail bound.
\end{proof}
\subsection{Imaginary geometry}
\label{ig prelim}
The proof of the lower bounds in our main theorems will make heavy use of the so-called forward coupling of $\operatorname{SLE}_\kappa$ or $\operatorname{SLE}_\kappa(\underline\rho)$ with the GFF with Dirichlet boundary conditions. In this coupling, $\operatorname{SLE}_\kappa(\underline\rho)$ for $\kappa \in (0,4)$ can be interpreted as the flow line of the formal vector field $e^{i h / \chi}$ where $h$ is a GFF and
\begin{equation}
\label{chi lambda}
\chi = \frac{2}{\sqrt{\kappa}} - \frac{\sqrt{\kappa}}{2}.
\end{equation}
For $\kappa > 4$, $\operatorname{SLE}_\kappa(\underline\rho)$ can be interpreted as a ``tree'' or ``light-cone'' of $\operatorname{SLE}_{16/\kappa}$ flow lines \cite{ig1}. The case $\kappa=4$ is somewhat degenerate (though simpler to analyze) since $\chi \to 0$ as $\kappa \to 4$. $\operatorname{SLE}_4(\underline\rho)$ has the interpretation of being a level line (rather than a flow line or light cone) of the GFF. See \cite{wang-wu-level-lines} for a detailed study of this case.
The coupling of $\operatorname{SLE}_4$ with the GFF was actually the first coupling in this family to be discovered \cite{ss-contour} (see also \cite{ss-dgff} which gives the convergence of the contours of the discrete GFF to $\operatorname{SLE}_4$). The existence of the forward coupling in the general setting is established in \cite{dubedat-coupling,ss-contour,shef-slides,ig1}; see \cite[Theorem~1.1]{ig1} for a precise statement. The theory of how different flow lines and light cones of the same GFF interact is developed in \cite{ig1,ig2,ig3,ig4}; these works are also where the term ``imaginary geometry'' is coined. At this point in time, there are several places which contain short ``crash courses'' on imaginary geometry which are sufficient to understand its usage in this work. We refer the reader to one of \cite[Section~2.2]{ig2}, \cite[Section~2.3]{ig4}, or \cite[Section~2.2]{miller-wu-dim}; \cite[Section~1]{ig1} and \cite[Section~4]{ig4} contain many of the main theorem statements in addition to more detailed overviews of the related literature.
\subsection{Properties of the multifractal spectrum sets}
\label{multifractal sets}
In this subsection we will prove some elementary properties of the sets of Section~\ref{multifractal def}, as well as a lemma which is relevant to the integral means spectrum.
Our first lemma tells us that the sets of Section~\ref{multifractal def} are only non-empty in the case $s\in [-1,1]$.
\begin{lem}\label{s>1 empty}
Let $D\subset \mathbf C$ be a simply connected domain and let $\phi : \mathbf D\rightarrow D$ be a conformal map. For each $x\in\partial\mathbf D$, there is a constant $C$ depending only on $\phi$ and $ \phi(x) $ but uniform for $\phi(x)$ in compact subsets of $\overline D$ s.t. for each sufficiently small $\epsilon>0$,
\begin{equation*}
C^{-1} \epsilon \leq |\phi'((1-\epsilon)x)| \leq C \epsilon^{-1} .
\end{equation*}
\end{lem}
\begin{proof}
By the Cauchy estimate,
\begin{equation*}
|\phi'((1-\epsilon)x)| \leq \epsilon^{-1} \sup_{z\in B_\epsilon((1-\epsilon) x)} |\phi(z)|
\end{equation*}
which gives the upper bound. For the lower bound, we apply the Koebe distortion theorem.
\end{proof}
Next we prove some lemmas which give that the multifractal spectrum sets are invariant under reasonable modifications of the definitions.
\begin{lem}\label{two curves}
Let $D\subset \mathbf C$ be a simply connected domain, $\phi :\mathbf D\rightarrow D$ a conformal map, and fix $x\in \partial \mathbf D$. Let $\gamma : [0,1] \rightarrow \overline{\mathbf D}$ be a simple smooth curve such that $\gamma(0) = x$, $\gamma((0,1]) \subset \mathbf D$, and $ \gamma'(0)$ is not tangent to $\partial\mathbf D$ at $x$. Then
\begin{equation} \label{limsup on curve}
\limsup_{\epsilon \rightarrow 0} \frac{\log |\phi'( (1-\epsilon)x)|}{-\log \epsilon } = \limsup_{\epsilon \rightarrow 0} \frac{\log |\phi'( \gamma(\epsilon) )|}{-\log \epsilon } .
\end{equation}
If one of the limsups is in fact a true limit, then the other is as well.
\end{lem}
\begin{proof}
By Taylor's formula, we can write
\begin{equation} \label{taylor for gamma}
\gamma(\epsilon) = x + \epsilon \gamma'(0) + O_\epsilon(\epsilon^2) .
\end{equation}
The function $w\mapsto \phi (\epsilon w + (1-\epsilon) x) / (\phi'((1-\epsilon) x) \epsilon)$ is a conformal map from $\mathbf D$ into $\mathbf C$ with unit derivative at the origin. By the Koebe distortion theorem applied to this map evaluated at $w = \epsilon^{-1}( \gamma(\epsilon) - (1-\epsilon) x)$ it follows that
\begin{equation} \label{koebe angle bound}
1- | \gamma'(0) + x + O_\epsilon(\epsilon ) | \preceq \frac{ |\phi'( \gamma(\epsilon) ) |}{ |\phi'( (1-\epsilon) x )| } \preceq \frac{1 }{(1- | \gamma'(0) + x + O_\epsilon(\epsilon ) |)^3} .
\end{equation}
Since $\gamma'(0)$ is not tangent to $\partial\mathbf D$ at $x$, there is some $c>0$ s.t. $|c \gamma'(0) +x| < 1$. It follows that we can perform a linear re-parametrization of $\gamma$ in such a way that $| \gamma'(0) + x| < 1$. Then \eqref{koebe angle bound} implies the statement of the lemma.
\end{proof}
\begin{lem}\label{D D'}
Let $D$ and $D'$ be two simply connected domains in $\mathbf C$, bounded by curves, which share a common boundary arc $I$. Let $z$ be a prime end lying in the interior of $I$. Then for each $s \in \mathbf R$, we have $z\in \Theta^s(D)$ iff $z \in \Theta^s(D')$. The same holds with $\Theta^{s , \geq}(\cdot)$ or $\Theta^{s;\leq}(\cdot)$ in place of $\Theta^s(\cdot)$.
\end{lem}
\begin{proof}
By comparing $D$ and $D'$ to the connected component of $D\cap D'$ with $I$ on its boundary, it suffices to consider the case where $D'\subset D$. Let $\phi : \mathbf D\rightarrow D$ and $\psi:\mathbf D\rightarrow D'$ be the corresponding conformal maps. We can factor $\phi = \psi \circ \xi$, where $\xi = \psi^{-1} \circ \phi$. Then
\begin{equation}\label{phi and xi}
\phi'((1-\epsilon) \phi^{-1}(z ) ) = \psi'(\xi( (1-\epsilon) \phi^{-1}(z ) ) ) \xi'((1-\epsilon) \phi^{-1}(z ))
\end{equation}
By Schwarz reflection, $\xi$ extends to be analytic in a neighborhood of $\phi^{-1}(z)$, so $|\xi'((1-\epsilon) \phi^{-1}(z ))|$ is bounded above and below by positive constants for small $\epsilon$. Let $\gamma(\epsilon) = \xi((1-\epsilon) \phi^{-1}(z )) $. Note that $\gamma$ is a simple curve in $\mathbf D$ with $\gamma(0) = \psi^{-1}(z)$. We have
\[
\gamma'(0) = -\xi'(\phi^{-1}(z)) \phi^{-1}(z) .
\]
Since $\xi$ maps a neighborhood of $\phi^{-1}(z)$ in $\partial\mathbf D$ into $\partial\mathbf D$, we have that $\xi'(\phi^{-1}(z))$ is a real multiple of $ \frac{ \xi(\phi^{-1}(z))}{\phi^{-1}(z)} = \frac{ \psi^{-1}(z) }{\phi^{-1}(z)}$. Hence $\gamma'(0)$ is a real multiple of $ \psi^{-1}(z) $.
In particular $\gamma$ is not tangent to $\partial\mathbf D$ at $\psi^{-1}(z)$ so the stated result follows from Lemma~\ref{two curves}.
\end{proof}
\begin{lem}\label{H to D}
Let $D\subset \mathbf C$ be a simply connected domain. Let $\phi : \mathbf D\rightarrow D$ and $\psi : \mathbf H\rightarrow D$ be conformal maps. For each prime end $z\in\partial D$ with $\psi^{-1}(z) \not=\infty$, one has
\begin{equation}
\limsup_{\epsilon \rightarrow 0} \frac{\log |\phi'( (1-\epsilon) \phi^{-1}(z) )|}{-\log \epsilon } = \limsup_{\epsilon \rightarrow 0} \frac{\log |\psi'( \psi^{-1}(z) + i \epsilon )|}{-\log \epsilon } .
\end{equation}
If one of the limsups is in fact a true limit, then the other is as well.
\end{lem}
\begin{proof}
We can write $\psi = \phi \circ \xi$ where $\xi = \phi^{-1} \circ \psi : \mathbf H \rightarrow \mathbf D$ is a conformal map. Then
\begin{equation*}
\psi'( \psi^{-1}(z) + i \epsilon ) = \phi'( \xi( \psi^{-1}(z) + i \epsilon ) ) \xi'(\psi^{-1}(z) + i \epsilon) .
\end{equation*}
The map $\xi'$ extends smoothly to $\partial\mathbf H$, so $|\xi'(\psi^{-1}(z) + i \epsilon)|$ is bounded above and below by positive constants for small $\epsilon$. Let $\gamma(\epsilon) = \xi( \psi^{-1}(z) + i \epsilon )$. Then $\gamma'(0)= i(\phi^{-1})'(z) /(\psi^{-1})'(z)$, which is not tangent to $\partial\mathbf D$ at $\phi^{-1}(z)$. Therefore the desired result follows from Lemma~\ref{two curves}.
\end{proof}
Lemma~\ref{H to D} implies in particular that if $\psi : \mathbf H \rightarrow D$ is a conformal map, then $\dim_{\mathcal H} \Theta^s(D)$ and $\dim_{\mathcal H} \widetilde\Theta^s(D)$ are unaffected if we replace their definitions from Section~\ref{multifractal def} with
\begin{equation} \label{theta with psi}
\widetilde\Theta^s(D) =
\left\{x\in \mathbf R: \lim_{\epsilon \rightarrow 0} \frac{\log |\psi' (x + i\epsilon )|}{-\log \epsilon } =s \right\} \quad \operatorname{and}\quad \Theta^s(D) = \psi ( \widetilde\Theta^s(D))
\end{equation}
The analogous statement also holds for the sets $\Theta^{s ; \geq}(D), \widetilde\Theta^{s; \geq}(D) , \Theta^{s;\leq}(D)$, and $\widetilde\Theta^{s;\leq}(D)$.
What follows is the analogue of Lemma~\ref{D D'} for the integral means spectrum.
\begin{lem}
\label{ims compare}
Let $D$ and $D'$ be two bounded Jordan domains in $\mathbf C$ and suppose there exists a connected boundary arc $I$ shared by $D$ and $D'$. Let $\phi \colon \mathbf D \to D$ and $\psi \colon \mathbf D \to D'$ be conformal maps. Let $J'$ be a closed subset of the interior of $I$ and let $J$ be a closed subset of the interior of $J'$. For $\epsilon > 0$, let $A_\epsilon$ be the set of $z\in \partial B_{1-\epsilon}(0)$ with $z/|z| \in \phi^{-1}(J)$ and let $A_\epsilon'$ be the set of $z\in \partial B_{1-\epsilon}(0)$ with $z/|z| \in \psi^{-1}(J')$. Then we have
\begin{equation} \label{ims phi psi}
\limsup_{\epsilon\rightarrow 0} \frac{\log \int_{A_\epsilon} |\phi'(z)|^a \, dz}{-\log \epsilon} \leq \limsup_{\epsilon\rightarrow 0} \frac{\log \int_{A_\epsilon'} |\psi'(z)|^a \, dz}{-\log \epsilon} .
\end{equation}
\end{lem}
\begin{proof}
Let $\xi$ be the conformal map from a subdomain of $\mathbf D$ to a subdomain of $D'\cap D$ which equals $\psi^{-1} \circ \phi$ wherever the latter is defined. By Schwarz reflection $\xi$ extends to a conformal map from a neighborhood of $\phi^{-1}(J')$ to a neighborhood of $\psi^{-1}(J')$. In particular $|\xi'| \asymp 1$ on a neighborhood of $\phi^{-1}(J')$, with implicit constants independent of $\epsilon$. By a change of variables, for sufficiently small $\epsilon > 0$ we have
\begin{equation} \label{ims change vars1}
\int_{A_\epsilon} |\phi'(z)|^a \,dz \asymp \int_{A_\epsilon} |\psi'(\xi(z))|^a \,dz \asymp \int_{\xi(A_\epsilon)} |\psi'(w)| \,dw .
\end{equation}
Let $p_\epsilon$ be the radial projection from $\mathbf D$ onto $\partial B_{1-\epsilon}(0)$. By the above application of Schwarz reflection (and the fact that $J$ is contained in the interior of $J'$), for sufficiently small $\epsilon > 0$, we have that $p_\epsilon$ restricts to a diffeomorphism from $\xi(A_\epsilon)$ to a subset $\widetilde A_\epsilon'$ of $A_\epsilon'$. Furthermore, since $|\xi'| \asymp 1$ on a neighborhood of $\psi^{-1}(J')$, we have $|p_\epsilon'| \asymp 1$ on $\xi(A_\epsilon)$ for sufficiently small $\epsilon$, and by the Koebe distortion theorem we have $|\psi'(p_\epsilon(w))| \asymp |\psi'(w)|$ for $w \in \xi(A_\epsilon)$ and sufficiently small $\epsilon$.
Therefore, a second change of variables yields
\begin{equation} \label{ims change vars2}
\int_{\xi(A_\epsilon)} |\psi'(w)| \,dw \asymp \int_{\widetilde A_\epsilon'} |\psi'(z)| \, dz \leq \int_{ A_\epsilon'} |\psi'(z)| \, dz .
\end{equation}
We obtain~\eqref{ims phi psi} by combining~\eqref{ims change vars1} and~\eqref{ims change vars2}.
\end{proof}
\subsection{Zero-one laws}
\label{zero one sec}
In this section we will prove that the multifractal spectrum and integral means spectrum of an $\operatorname{SLE}_\kappa(\underline\rho)$ curve are a.s.\ deterministic and do not depend on $\underline\rho$ or on which complementary component of the curve we consider.
\begin{prop} \label{theta zero one}
Let $D \subset \mathbf C$ be a smoothly bounded domain. Let $\kappa >0$ and let $\underline\rho$ be a vector of real weights. Let $\eta$ be a chordal $\operatorname{SLE}_\kappa(\underline\rho)$ process in $D$, with any choice of initial and target points and force points located anywhere in $\overline{ D}$, run up until the first time it either hits an interior force point or hits the continuation threshold after which it is no longer defined (c.f.\ \cite[Section~2.1]{ig1}). Fix $s\in (-1,1)$. Almost surely, the following is true.
Let $V$ be a connected component of $ D\setminus \eta$ or a connected component of $\ D\setminus \eta([0,t])$ for any $t > 0$ and let $\phi : \mathbf D\rightarrow V$ be a conformal map.
The Hausdorff dimension of each of the sets $\Theta^{s}(V) \setminus \partial D , \widetilde \Theta^{s}(V) \setminus \phi^{-1}(\partial D), \Theta^{s;\geq}(V)\setminus \partial D ,\widetilde \Theta^{s;\geq}(V) \setminus \phi^{-1}(\partial D) , \Theta^{s;\leq}(V)\setminus \partial D $, and $\widetilde\Theta^{s;\leq}(V)\setminus \phi^{-1}(\partial D)$ is equal to a deterministic constant which depends only on $\kappa$ and $s$. Furthermore, the a.s.\ Hausdorff dimensions of the corresponding sets for $\kappa$ and $16/\kappa$ are equal.
\end{prop}
\begin{proof}
We begin with some reductions.
By Lemma~\ref{H to D} we can consider $\operatorname{SLE}_\kappa$ from 0 to $\infty$ in $\mathbf H$ instead of $\operatorname{SLE}_\kappa$ between two arbitrary points in $ D$, and we can use the alternative definition \eqref{theta with psi} for the multifractal spectrum maps (with $(f_t)$ the centered Loewner maps for $\eta$ and $\psi = f_t^{-1}$).
Next, we note that $D\setminus \eta^t$ a.s.\ has only countably many connected components for any given $t > 0$. By continuity, countable stability of Hausdorff dimension, and Lemma~\ref{D D'}, it therefore suffices to prove that the statement of the proposition holds a.s.\ for some fixed but arbitrary choice of domain $V$ as in the statement of the proposition (chosen in some deterministic manner), rather than for all such $V$ simultaneously (it will be clear from the proof that the a.s.\ dimension obtained does not depend on how we choose $V$).
We will prove the result for $\Theta^s(V) \setminus \mathbf R$ and $\widetilde\Theta^s(V)\setminus f_t^{-1}(\mathbf R)$; the case for the other sets is similar.
First consider the case where $\kappa\leq 4$ and $\underline\rho = 0$, so $\eta$ is an ordinary $\operatorname{SLE}_\kappa$ process. In this case, the statement of the proposition for a complementary connected component $V$ of $\mathbf H\setminus \eta$ follows from the statement for $V = \mathbf H\setminus \eta^t$ by Lemma~\ref{D D'} and countable stability of Hausdorff dimension, so it suffices to prove the statement with $V = \mathbf H\setminus \eta^t$ for a general choice of $t > 0$.
By scale invariance the law of each $\Theta^s(\mathbf H\setminus \eta^t)$ is independent of $t$. Since the derivative of the conformal map $f_{t/2}$ is bounded above and below by positive (random) constants in a neighborhood of each point of $\eta^t \setminus \eta^{t/2}$, we infer that $\Theta^s(\mathbf H\setminus \eta^t) \setminus \eta^{t/2} = \Theta^s(\mathbf H\setminus f_{t/2}(\eta^t\setminus \eta^{t/2}))$.
Since conformal maps preserve Hausdorff dimension and by Lemma~\ref{D D'}, we thus have that the Hausdorff dimension of each $\Theta^s(\mathbf D\setminus \eta^t)$ is equal to the maximum of $\dim_{\mathcal H} \Theta^s(\mathbf H\setminus \eta^{t/2})$ and $\dim_{\mathcal H} \Theta^s(\mathbf H\setminus f_{t/2}(\eta^t\setminus \eta^{t/2}))$. These latter two sets are independent and identically distributed (by the Markov property of SLE) and their Hausdorff dimensions agree in law with that of $\Theta^s(\mathbf H\setminus \eta^t)$ (by the scale invariance property noted above). A random variable can be equal to the maximum of two independent random variables with the same law as itself only if it is a.s.\ constant.
To prove the analogous statement for $\widetilde \Theta^s(\mathbf H\setminus \eta^t)$, we observe that $\dim_{\mathcal H} \widetilde\Theta^s(\mathbf H\setminus \eta^t)$ is the maximum of $\dim_{\mathcal H} f_t^{-1} ( \widetilde \Theta^s(\mathbf H\setminus \eta^t) \cap \eta^{t/2})$ and $ \dim_{\mathcal H} f_t^{-1} ( \Theta^s(\mathbf H\setminus \eta^t) \setminus \eta^{t/2})$. By smoothness of the map $f_{t/2} \circ f_t^{-1}$ on $ f_{t/2}( \mathbf H\setminus \eta^{t/2})$ and of $f_t^{-1} $ on $\eta^t\setminus \eta^{t/2}$, respectively, these dimensions equal $\dim_{\mathcal H} f_{t/2}^{-1}\left( \widetilde \Theta^s(\mathbf H\setminus \eta^{t/2})\right) $ and $\dim_{\mathcal H} (f_t \circ f_{t/2}^{-1} )^{-1}\left( \widetilde \Theta^s( \mathbf H\setminus f_{t/2}( \eta^t \setminus \eta^{t/2}))\right)$, respectively. By the Markov property these latter two quantities are i.i.d., and we conclude as above.
Next suppose that $\kappa$ is still $\leq 4$, but that $\underline\rho$ is arbitrary. For $\delta > 0$, inductively define stopping times $\tau_j^\delta$ and $\sigma_j^\delta$ for $j\in\mathbf N$ as follow. Let $\tau_1^\delta$ be the first time $t > 0$ that either $\eta$ hits an interior force point or the continuation threshold; or $\operatorname{Im} \eta(t) \geq 2\delta$. Let $\sigma_1^\delta$ be the first time $s > \tau_1^\delta$ that $\eta$ hits an interior force point or the continuation threshold; or $\operatorname{Im} \eta(s) \leq \delta$. Inductively, if $j \geq 2$ and $(\tau_{j-1}^\delta , \sigma_{j-1}^\delta)$ has been defined, let $\tau_j^\delta$ be the first time $t > \sigma_{j-1}^\delta$ that $\eta$ either hits an interior force point or the continuation threshold; or $\operatorname{Im} \eta(t) \geq 2\delta$ and let $\sigma_j^\delta$ be the first time $s > \tau_j^\delta$ that either $\eta$ hits the continuation threshold or $\operatorname{Im} \eta(s) \leq \delta$. For each $\delta$ and each $j$, the law of $f_{\tau_j^\delta}(\eta|_{[\tau_j^\delta , \sigma_j^\delta]})$ is absolutely continuous with respect to the law of an ordinary $\operatorname{SLE}_\kappa$ stopped at some a.s.\ positive time. Therefore, Lemma~\ref{D D'} implies that if $V$ is as in the statement of the lemma and $\psi : \mathbf H\rightarrow V$ is a conformal map, then $\dim_{\mathcal H} \Theta^s(V) \cap \eta([\tau_j^\delta , \sigma_j^\delta]) $ and $\dim_{\mathcal H} \widetilde \Theta^s(V) \cap \psi^{-1}\left( \eta([\tau_j^\delta , \sigma_j^\delta]) \right)$, respectively, a.s.\ agree with the a.s.\ values of the corresponding sets for an ordinary $\operatorname{SLE}_\kappa$. We have
\[
\bigcup_{\delta > 0}\bigcup_{j\in \mathbf N} \eta( [\tau_j^\delta , \sigma_j^\delta] ) = \eta \setminus \mathbf R
\]
so by countable stability of Hausdorff dimension (restrict $\delta$ to a sequence tending to 0) and we obtain the statement of the proposition for a general $\operatorname{SLE}_\kappa(\underline\rho)$ process with $\kappa \leq 4$.
The statement for $\kappa > 4$ follows from the statement for $16/\kappa <4$ together with Lemma~\ref{D D'} and SLE duality (see, e.g.\ \cite{zhan-duality1,zhan-duality2,dubedat-duality,ig1,ig4}).
\end{proof}
For the proof of Corollary~\ref{ims cor}, we will also need the analogue of Proposition~\ref{theta zero one} for the integral means spectrum.
\begin{prop} \label{ims zero one}
Suppose we are in the setting of Proposition~\ref{theta zero one}. Fix $a \in\mathbf R$. Almost surely, the following is true. Let $V$ be a complementary connected component of either $D\setminus \eta$ or of $D\setminus \eta^t$ for any $t > 0$. Then $\operatorname{IMS}^{\operatorname{bulk}}_{V}(a) $ is equal to a deterministic constant which depends only on $\kappa$ and $a$. This deterministic constant is the same if we replace $\kappa$ with $16/\kappa$.
\end{prop}
\begin{proof}
The proof proceeds in a similar manner to that of Proposition~\ref{theta zero one}. By continuity of $\eta$ together with Lemma~\ref{ims compare}, it is enough to prove that the statement of the proposition holds a.s.\ for some fixed but arbitrary choice of domain $V$ as in the statement of the proposition (chosen in some deterministic manner), rather than for all such $V$ simultaneously (it will be clear from the proof that the a.s.\ dimension obtained does not depend on how we choose $V$).
Henceforth we fix a conformal map $\psi : D\rightarrow \mathbf H$ taking the initial and target points for $\eta$ to 0 and $\infty$, respectively, and assume that $\eta$ is parametrized so that the half-plane capacity of $\psi(\eta^t)$ is $t$ for each $t\geq 0$. We also let $(\widetilde f_t)$ be the centered Loewner maps for $\psi(\eta)$ and set $f_t := \psi^{-1} \circ \widetilde f_t \circ \psi$. Define the sets $I^\zeta(\cdot)$ and $A_\epsilon^\zeta(\cdot)$ as in Section~\ref{ims sec}.
First consider the case where $\kappa \leq 4$ and $\underline\rho = 0$, so that $\eta$ is an ordinary $\operatorname{SLE}_\kappa$. Fix $t > 0$. By scale invariance (applied to $\psi(\eta)$) the law of $\operatorname{IMS}^{\operatorname{bulk}}_{D\setminus \eta^t}(a)$ does not depend on $t$. Let $\widetilde\eta := f_{t/2}(\eta|_{[t/2,\infty)})$ and $\widetilde f_{t/2} := f_t \circ f_{t/2}^{-1}$, so that $\widetilde\eta$ is independent from $\eta^{t/2}$, $\widetilde\eta \overset{d}{=} \eta$, and $\widetilde f_{t/2} \overset{d}{=} f_{t/2}$. Since the derivative of the conformal map $f_{t/2}$ is bounded above and below by positive constants on each compact subset of $\eta^t \setminus \eta^{t/2}$, we obtain that a.s.\
\begin{equation*} \label{bulk IMS def}
\operatorname{IMS}_{ D\setminus \widetilde \eta^{t/2}}^{\operatorname{bulk}}(a) := \sup_{\zeta > 0} \limsup_{\epsilon\rightarrow 0} \frac{\log \int_{\widetilde A_\epsilon^\zeta(f_t^{-1}) } |\phi'(z)|^a \, dz }{-\log \epsilon} ,
\end{equation*}
where $\widetilde A_\epsilon^\zeta(f_t^{-1})$ is the set of $z\in \partial B_{1-\epsilon}(0)$ such that $f_t^{-1}(z/|z|) \in \eta^t\setminus \eta^{t/2}$, $|f_t^{-1}(z/|z|) - \eta(t)| \geq \zeta$, and $|f_t^{-1}(z/|z|) - \eta(t/2)| \geq \zeta$.
Note that $\widetilde A_\epsilon^\zeta(f_t^{-1}) \subset A_\epsilon^\zeta(f_t^{-1})$. Therefore, a.s.\
\begin{equation*}
\operatorname{IMS}_{D\setminus \eta^t}^{\operatorname{bulk}}(a) \geq \max\left\{ \operatorname{IMS}_{ D\setminus \widetilde \eta^{t/2}}^{\operatorname{bulk}}(a) , \operatorname{IMS}_{ D\setminus \eta^{t/2}}^{\operatorname{bulk}}(a) \right\} .
\end{equation*}
All three random variables in this inequality have the same law (by the considerations above), and the two on the right are independent. We infer that all three random variables are a.s.\ equal to a deterministic constant depending only on $a$ and $\kappa$.
It remains to treat the two complementary connected components of the whole curve. By symmetry, the a.s.\ value of $\operatorname{IMS}_{D\setminus \eta^t}^{\operatorname{bulk}}(a) $ is unaffected if, in the definition of the bulk integral means spectrum, we consider only the pre-images of prime ends on the right (or left) side of $\eta^t$. By transience of SLE, for any given $\zeta >0$, we can a.s.\ find $t > 0$ such that $\eta$ and $\eta^t$ agree outside $B_\zeta(\partial D)$. It therefore follows from Lemma~\ref{ims compare} that if $V$ is one of the two connected components of $D\setminus \eta$, then $\operatorname{IMS}_{V}^{\operatorname{bulk}}(a)$ is a.s.\ equal to the same a.s.\ value as $\operatorname{IMS}_{D\setminus \eta^t}^{\operatorname{bulk}}(a)$ for a finite time $t$. This completes the proof of the proposition in the case $\kappa \leq 4$ and $\underline\rho=0$.
Next suppose that $\kappa$ is still $\leq 4$, but that $\underline\rho$ is arbitrary. Note that in this case $\eta$ does not intersect itself, so the definition of the sets $I^\zeta(\cdot)$ in Section~\ref{ims sec} is unaffected if we ignore the points $x_V$ and $y_V$. For $\delta > 0$ and $j\in \mathbf N$, define the stopping times $\tau_j^\delta$ and $\sigma_j^\delta$ as in the proof of Proposition~\ref{theta zero one} but with $\operatorname{Im} \eta(t)$ replaced by $\operatorname{dist}(\eta(t) , \partial D)$. For each $\zeta >0$, only finitely many of the excursions $\eta([\tau_j^\delta , \sigma_j^\delta])$ intersect $D\setminus B_\zeta(\partial D)$. By applying Lemma~\ref{ims compare} to each of these excursions (with the arc $I$ equal to the excursion and the sets $J$ and $J'$ equal to its intersections with $D\setminus B_\zeta(\partial D)$ and $D\setminus B_{\zeta'}(\partial D)$ for appropriate $\zeta$ and $\zeta'$) and using absolute continuity as in the proof of Proposition~\ref{theta zero one}, we obtain the statement of the proposition for $\kappa\leq 4$ and a general choice of $\underline\rho$ in the case where $V$ is a complementary connected component of $\eta^t$ for some finite $t > 0$. For a complementary connected component of the whole curve $\eta$, we apply the same argument as in the case $\underline\rho =0$ (note that transience of $\operatorname{SLE}_\kappa(\underline\rho)$ for general $\underline\rho$ processes which do not hit the continuation threshold is proven in \cite[Theorem~1.3]{ig1}).
Finally, the case $\kappa > 4$ follows from SLE duality as in the proof of Proposition~\ref{theta zero one}, but with Lemma~\ref{ims compare} used in place of Lemma~\ref{D D'}.
\end{proof}
\subsection{Multifractal spectrum definition}
\label{multifractal def}
We will now introduce the sets whose Hausdorff dimension we will compute, in the setting of general domains in the complex plane. Our definitions are similar to those in \cite[Section~2]{lawler-viklund-tip}, but we deal with the boundary of a domain rather than the tip of a given curve.
Let $D\subset \mathbf C$ be a simply connected domain and let $\phi : \mathbf D\rightarrow D$ be a conformal map. For $s\in\mathbf R$, define
\begin{align}
\label{tilde theta}
\widetilde\Theta^{s }(D) &:= \left\{x\in \partial \mathbf D: \lim_{\epsilon \rightarrow 0} \frac{\log |\phi'((1-\epsilon)x)|}{-\log \epsilon } = s \right\}
\intertext{and}
\label{theta}
\Theta^{s }(D) &:=\phi( \widetilde\Theta^{s }(D) ).
\intertext{Also define}
\widetilde\Theta^{s ; \leq }(D) &:= \left\{x\in \partial \mathbf D: \limsup_{\epsilon \rightarrow 0} \frac{\log |\phi'((1-\epsilon)x)|}{-\log \epsilon } \leq s \right\} \notag\\
\Theta^{s ; \leq }(D) &:=\phi( \widetilde\Theta^{s ; \leq }(D) ) \notag \\
\widetilde\Theta^{s ; \geq }(D) &:= \left\{x\in \partial \mathbf D: \limsup_{\epsilon \rightarrow 0} \frac{\log |\phi'((1-\epsilon)x)|}{-\log \epsilon } \geq s \right\} \notag \\
\Theta^{s ; \geq }(D) &:=\phi( \widetilde\Theta^{s ; \geq }(D) ). \notag
\end{align}
The \emph{multifractal spectrum} of $D$ can be defined as one of the two functions $s \mapsto \dim_{\mathcal H}\Theta^{s}(D)$ or $s \mapsto \dim_{\mathcal H} \widetilde\Theta^{s }(D)$. It is easy to check that these definitions do not depend on the choice of conformal map $\phi$. We note that although the sets $\Theta^s(D)$ and $\widetilde\Theta^s(D)$ are defined for all $s\in \mathbf R$, these sets are empty for $s\notin [-1,1]$ (see Lemma~\ref{s>1 empty} below).
\subsection{Main results}
Our main result is the following theorem.
\begin{thm}
\label{main thm}
Let $\kappa\leq 4$. Let $\eta$ be a chordal $\op{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$. Let $D_\eta$ be the connected component of $\mathbf D\setminus \eta([0,\infty))$ containing 1 on its boundary. Let
\begin{align}
\widetilde\xi(s) &: =1 - \frac{(4 + \kappa)^2 s^2}{8 \kappa (1 + s)}\label{tilde xi(s)} \\
\xi(s) &:= \frac{ 8 \kappa ( 1 + s ) -(4+\kappa)^2 s^2}{8 \kappa ( 1 - s^2) } \label{xi(s)} \\
s_- &:= \frac{4 \kappa - 2 \sqrt2 \sqrt{\kappa (2 + \kappa) (8 + \kappa)}}{ (4 + \kappa)^2} \label{s-} \\
s_+ &:= \frac{4 \kappa + 2 \sqrt2 \sqrt{\kappa (2 + \kappa) (8 + \kappa)}}{ (4 + \kappa)^2}. \label{s+}
\end{align}
For $s\in (-1,1)$, a.s.
\begin{equation*}
\begin{aligned}[c]
\dim_{\mathcal H} \widetilde\Theta^{s }(D_\eta) &= \dim_{\mathcal H} \widetilde\Theta^{s; \geq }(D_\eta) = \widetilde\xi(s) ,\quad &0\leq s \leq s_+ \notag\\
\dim_{\mathcal H} \widetilde\Theta^{s }(D_\eta) &= \dim_{\mathcal H} \widetilde\Theta^{s; \leq }(D_\eta) = \widetilde\xi(s) , \quad & s_- \leq s \leq 0 \notag\\
\dim_{\mathcal H} \Theta^{s }(D_\eta) &= \dim_{\mathcal H} \Theta^{s; \geq }(D_\eta) = \xi(s) , \quad & \frac{\kappa}{4} \leq s \leq s_+ \notag\\
\dim_{\mathcal H} \Theta^{s }(D_\eta) &= \dim_{\mathcal H} \Theta^{s; \leq }(D_\eta) = \xi(s) , \quad & s_- \leq s \leq \frac{\kappa}{4} .\notag
\end{aligned}
\end{equation*}
Moreover, we a.s.\ have $\widetilde\Theta^{s }(D_\eta) = \Theta^{s }(D_\eta) = \emptyset$ for each $s\notin [s_- , s_+]$.
\end{thm}
\begin{remark}
\label{critical s remark}
The significance of $s_-$ and $s_+$ is that $\widetilde\xi(s) \geq 0$ for $s \in [s_- , s_+]$, and the significance of $s=\kappa/4$ is that it is the value which maximizes $\xi$. Note $s_- \in (-1,0)$ and $s_+ \in (0,1]$ for any $\kappa > 0$ and $s_+ = 1$ if and only if $\kappa=4$. We refer the reader to Remark~\ref{s=1 kappa=4} below for more detail regarding the case $\kappa=4$, $s=1$.
\end{remark}
\begin{figure}
\begin{center}
{\includegraphics[width=0.45\textwidth]{figures/xi_xit_plot_new4.pdf}}\hspace{0.08\textwidth}{\includegraphics[width=0.45\textwidth]{figures/ims-graph-new4.pdf}}
\end{center}
\caption{\label{xi xit graph} Left: A graph of the Hausdorff dimensions $\widetilde\xi(s)$ of $\widetilde\Theta^s(D_\eta)$ and $\xi(s)$ of $\Theta^s(D_\eta)$ from Theorem~\ref{main thm} as $s$ ranges from~$-1$ to~$1$ for $\kappa = 2$. The value of $s$ which maximizes $\widetilde{\xi}$ is $0$ and the value of $s$ which maximizes $\xi$ is $\kappa/4=1/2$. Note that $\xi(\kappa/4) = 1+\kappa/8$ which is the almost sure Hausdorff dimension of $\op{SLE}_\kappa$ \cite{beffara-dim}. Right: a graph of the bulk integral means spectrum $\operatorname{IMS}_{D_\eta}(a)$ of $D_\eta$ from Corollary~\ref{ims cor} as $a$ ranges from $-7$ to $7$ for $\kappa = 3$.}
\end{figure}
The $\operatorname{SLE}_\kappa(\underline{\rho})$ processes are an important variant of $\operatorname{SLE}$ in which one keeps track of extra marked points --- so-called force points. The force points can be either on the domain boundary or in its interior and are respectively referred to as boundary and interior force points. These processes were first introduced by Lawler, Schramm, and Werner in \cite[Section~8.3]{lsw-restriction} and, just like ordinary $\operatorname{SLE}_\kappa$, the $\operatorname{SLE}_\kappa(\rho)$ processes naturally arise in many different contexts. Since $\operatorname{SLE}_\kappa(\underline\rho)$ for different vectors of weights $\underline\rho$ has the same behavior when it is not interacting with its force points, one expects an analogue of Theorem~\ref{main thm} to be true for such processes provided we exclude points near the boundary of the domain and stop the path before interacting with an interior force point. Furthermore, by SLE duality, one expects an analogue of Theorem~\ref{main thm} for $\kappa >4$. Such results do indeed hold true, as described in the following corollary.
\begin{cor}
\label{dim for K_t}
Let $D \subset \mathbf C$ be a smoothly bounded domain. Let $\kappa >0$ and let $\underline\rho$ be a vector of real weights. Let $\eta$ be a chordal $\operatorname{SLE}_\kappa(\underline\rho)$ process in $D$, with any choice of initial and target points and force points located anywhere in $\overline{ D}$, run up until the first time it either hits an interior force point or hits the continuation threshold (c.f.\ \cite[Section 2.1]{ig1}). Fix $s \in (-1,1)$. Almost surely, the following is true.
Let $V$ be a connected component of $ D\setminus \eta$ or a connected component of $ D\setminus \eta([0,t])$ for any $t > 0$ before $\eta$ hits an interior force point or the continuation threshold and let $\phi : \mathbf D\rightarrow V$ be a conformal map. Then
\begin{equation*}
\begin{aligned}[c]
\dim_{\mathcal H} \left( \widetilde\Theta^{s }(V) \setminus \phi^{-1}(\partial D)\right) &= \dim_{\mathcal H} \left( \widetilde\Theta^{s; \geq }(V) \setminus \phi^{-1}(\partial D) \right) = \widetilde\xi(s) ,\quad & 0\leq s \leq s_+ \notag\\
\dim_{\mathcal H}\left( \widetilde\Theta^{s }(V) \setminus \phi^{-1}(\partial D)\right) &= \dim_{\mathcal H}\left( \widetilde\Theta^{s; \leq }(V) \setminus \phi^{-1}(\partial D)\right) = \widetilde\xi(s) , \quad & s_- \leq s \leq 0 \notag\\
\dim_{\mathcal H} \left(\Theta^{s } (V) \setminus \partial D\right) &= \dim_{\mathcal H} \left(\Theta^{s; \geq }(V) \setminus \partial D\right) = \xi(s) ,\quad & \frac{\kappa}{4} \leq s \leq s_+ \notag\\
\dim_{\mathcal H}\left( \Theta^{s } (V) \setminus \partial D\right) &= \dim_{\mathcal H}\left( \Theta^{s; \leq }(V) \setminus \partial D\right) = \xi(s) , \quad & s_- \leq s \leq \frac{\kappa}{4} \notag
\end{aligned}
\end{equation*}
That is, the conclusion of Theorem~\ref{main thm} holds a.s.\ away from the domain boundary at all times simultaneously for an $\operatorname{SLE}_\kappa(\underline\rho)$ with a general $\kappa > 0$ and vector of weights $\underline\rho$ up until the process either hits an interior force point or the continuation threshold.
\end{cor}
\begin{proof}
This follows from Theorem~\ref{main thm} combined with Proposition~\ref{theta zero one} below. Note that the functions $\widetilde \xi(s)$ and $\xi(s)$ are unaffected if we replace $\kappa$ by $16/\kappa$, as one would expect from SLE duality \cite{zhan-duality1,zhan-duality2,dubedat-duality,ig1,ig4}.
\end{proof}
\begin{remark}
We believe that the techniques developed in this paper could also be employed to describe the multifractal behavior of the $\operatorname{SLE}_\kappa(\underline{\rho})$ processes even near their intersection points with the domain boundary and near their tip, though we will not carry this out here.
\end{remark}
Roughly speaking, the \emph{harmonic measure spectrum} of a hull $A\subset \mathbf H$ gives, for each $\alpha \in (1/2,\infty)$, the Hausdorff dimension of the set $\Theta^\alpha_{\operatorname{hm}}(A)$ of points $x \in \partial A$ for which the harmonic measure from $\infty$ of $B_\epsilon(x)$ in $\mathbf H\setminus A$ decays like $\epsilon^\alpha$ as $\epsilon\rightarrow 0$ (or in the pre-image $\widetilde\Theta^\alpha_{\operatorname{hm}}(A)$ of $\Theta^\alpha_{\operatorname{hm}}(A)$ under a conformal map $\mathbf D\rightarrow \mathbf H\setminus A$). In \cite[Section~2.3]{lawler-viklund-tip}, the authors give a a rigorous treatment of the harmonic measure spectrum at the tip of a curve. A nearly identical construction works for the harmonic measure spectrum of a whole hull in $\mathbf H$. Similar constructions also work for hulls in $\mathbf D$ or $\mathbf C$. In particular, one has (see \cite[Lemma~2.3]{lawler-viklund-tip})
\begin{equation} \label{mf hm spec}
\Theta^s ( A) = \Theta^{\frac{1}{1-s}}_{\operatorname{hm}}(\mathbf H\setminus A) \quad \forall s \in (-1,1) .
\end{equation}
\begin{remark}
\label{bm spectrum}
In light of the relationship between $\op{SLE}_6$ and Brownian motion \cite{lsw-frontier}, we see that Corollary~\ref{dim for K_t} with $\kappa = 6$ yields the harmonic measure spectrum for the Brownian frontier computed in \cite{lawler-frontier, lsw-frontier, lsw-bm-exponents1, lsw-bm-exponents2, lsw-bm-exponents3}.
\end{remark}
\begin{remark}
In \cite{dup-mf-spec-bulk} (see in particular \cite[Equation~6]{dup-mf-spec-bulk}), Duplantier predicts that the harmonic measure spectrum for the bulk of the $\op{SLE}_\kappa$ curve is given by
\begin{equation} \label{dup f(a)}
f(\alpha) = \alpha + \frac{25-c}{24} \left( 1 - \frac12 \left(2\alpha - 1 + \frac{1}{2\alpha-1} \right) \right) ,
\end{equation}
where
\begin{equation*}
c = \frac{(6-\kappa) (6-16/\kappa)}{4}
\end{equation*}
is the central charge. The exponent~\eqref{xi(s)} is related to the exponent~\eqref{dup f(a)} by
\begin{equation*}
\xi(s ) = f\left(\frac{1}{1-s} \right).
\end{equation*}
This is what we would expect in light of~\eqref{mf hm spec}.
\end{remark}
\begin{comment}
c[kappa_] := (6 - kappa) (3 kappa - 8)/(2 kappa);
f[kappa_, a_] :=
a + (25 - c[kappa])/24 (1 - (1/2) (2 a - 1 + 1/(2 a - 1)))
alpha[kappa_, s_] := (4 + kappa)^2 s^2/(8 kappa (1 + s))
xi1[kapppa_,
s_] := (-8 kappa -
8 kappa s + (4 + kappa)^2 s^2)/(8 kappa (-1 + s^2))
xi[kappa_,
s_] := (8 kappa (1 + s) - (4 + kappa)^2 s^2)/(8 kappa (1 - s^2))
FullSimplify[xi[kappa, s] - xi1[kappa, s]]
FullSimplify[f[kappa, 1/(1 - s)] - xi[kappa, s]]
Solve[xi[kappa, s] == 0, s];
FullSimplify
\end{comment}
The dimension $\xi(s)$ attains a unique maximum value of $1+\kappa/8$ on $[-1,1]$ at $s = \kappa/4$. This maximum value coincides with the Hausdorff dimension of the $\op{SLE}_\kappa$ curve \cite{beffara-dim}, which suggests that near a ``typical point'' of $\eta$, the modulus of the derivative of a conformal map from $D_\eta$ to $\mathbf D$ grows like $\operatorname{dist}(z , \eta)^{ \frac{\kappa}{4-\kappa}}$. Hence Theorem~\ref{main thm} gives an alternative proof of the following.
\begin{cor}
Let $\kappa\leq 4$. The Hausdorff dimension of an $\op{SLE}_\kappa$ curve $\eta$ is a.s.\ equal to $1+\kappa/8$.
\end{cor}
We remark that we believe that the methods that we use to establish the lower bound in Theorem~\ref{main thm} could be employed to give an independent derivation of the lower bound of the dimension of $\op{SLE}_\kappa$ for all $\kappa > 0$, however we will not carry this out here.
\subsection{Integral means spectrum}
\label{ims sec}
The \emph{integral means spectrum} of a simply connected domain $D\subset \mathbf D$ is the function $\operatorname{IMS}_D : \mathbf R\rightarrow \mathbf R$ defined by
\begin{equation}
\label{IMS def}
\operatorname{IMS}_D(a) := \limsup_{\epsilon\rightarrow 0} \frac{\log \int_{\partial B_{1-\epsilon}(0)} |\phi'(z)|^a \, dz }{-\log \epsilon} ,
\end{equation}
where $\phi : \mathbf D\rightarrow D$ is a conformal map. (There is a three parameter family of such conformal maps, but $\operatorname{IMS}_D(a)$ does not depend on the specific choice of $\phi$.) The integral means spectrum is of substantial interest in complex analysis, primarily in the form of the \emph{universal integral means spectrum}, which is defined by
\[
\operatorname{IMS}^U(a) := \sup_{D} \operatorname{IMS}_D(a)
\]
where the supremum is over all simply connected domains $D\subset \mathbf C$. It has been conjectured by Kraetzer \cite{kraetzer-ims} that $\operatorname{IMS}^U(a) = t^2/4$ for $|t|\leq 2$ and $\operatorname{IMS}^U(a) = |t|-1$ for $|t|\geq 2$. This conjecture has several important consequences in complex analysis. See, e.g., \cite{pom-im-survey, bel-smirnov-survey, hs-ims-survey, pom-book} for more details. The integral means spectrum is often very difficult to compute in practice for deterministic domains. However, domains bounded by random fractals (e.g.\ the complement of an $\op{SLE}_\kappa$ curve) are sometimes more tractable. For example, in \cite{bel-smirnov-hm-sle} Beliaev and Smirnov give an explicit calculation of the average integral means spectrum of the complement of a whole plane $\op{SLE}_\kappa$ curve (which is defined as in~\eqref{IMS def} but with $|\phi'(z)|^a$ replaced by $\mathbf E(|\phi'(z)|^a)$).
In this paper we shall be interested in a slight refinement of the definition of the integral means spectrum for the complement of a curve which negates possible pathologies arising from unusual behavior at its endpoints or when it intersects itself or the boundary of the domain. Namely, let $D\subset\mathbf C$ be a bounded simply connected domain with smooth boundary and let $\eta : [0,T]\rightarrow \overline{ D}$ be a non-self-crossing curve (we allow $T = \infty$). Let $V$ be a connected component of $D\setminus \eta$. Let $x_V$ be the first (equivalently last) point of $\partial V$ hit by $\eta$ and let $\phi : \mathbf D\rightarrow V$ be a conformal map.
For $\zeta > 0$, let
\begin{equation} \label{I_zeta def}
I^\zeta(\phi) := \phi^{-1}\left( \partial V \setminus (B_\zeta(\eta(T)) \cup B_\zeta(x_V) \cup B_\zeta(\partial D) \right) .
\end{equation}
Let $A_\epsilon^\zeta(\phi)$ be the set of $z\in \partial B_{1-\epsilon}(0)$ with $z/|z| \in I^\zeta(\phi)$.
The \emph{bulk integral means spectrum} of $V$ is the function $\operatorname{IMS}_V : \mathbf R\rightarrow \mathbf R$ defined by
\begin{equation} \label{bulk IMS def}
\operatorname{IMS}_V^{\operatorname{bulk}}(a) := \sup_{\zeta > 0} \limsup_{\epsilon\rightarrow 0} \frac{\log \int_{A_\epsilon^\zeta(\phi) } |\phi'(z)|^a \, dz }{-\log \epsilon} .
\end{equation}
One can check that the definition~\eqref{bulk IMS def} does not depend on the choice of $\phi$.
We extract the following from the proof of Theorem~\ref{main thm}.
\begin{cor}
\label{ims cor}
For $a\in \mathbf R$ with $a < \frac{ (4+\kappa)^2}{8\kappa}$, let
\begin{equation} \label{s_*(a)}
s_*(a) := -1 + \frac{ 4 + \kappa }{\sqrt{ (4 + \kappa)^2 - 8 a \kappa}} .
\end{equation}
Also let $s_-$ and $s_+$ be as in~\eqref{s-} and~\eqref{s+} and let $a_-$ (resp.\ $a_+$) be the value of $a$ for which $s_*(a) = s_-$ (resp.\ $s_*(a) = s_+$). Set
\begin{equation} \label{IMS* def}
\operatorname{IMS}^*(a):=
\begin{dcases}
-1+ s_- a ,\qquad &a < a_- \\
-a + \frac{(4+\kappa)(4+\kappa - \sqrt{(4+\kappa)^2 -8a \kappa})}{4\kappa} ,\qquad &a\in [a_- , a_+] \\
-1+ s_+ a ,\qquad &a >a_+ .
\end{dcases}
\end{equation}
Suppose we are in the setting of Corollary~\ref{dim for K_t}. Almost surely, the following is true. Let $a \in \mathbf R$ and let $V$ be a complementary connected component of either $D\setminus \eta$ or of $D\setminus \eta^t$ for any $t > 0$ (before $\eta$ hits an interior force point or the continuation threshold if it is an $\operatorname{SLE}_\kappa(\underline{\rho})$ process). Then
\begin{equation} \label{ims eqn}
\operatorname{IMS}_{V}^{\operatorname{bulk}}(a) = \operatorname{IMS}^*(a) .
\end{equation}
\end{cor}
The result of Corollary~\ref{ims cor} is in agreement with the (rigorously proven) formula\footnote{The formula appearing in \cite[Theorem~1]{bel-smirnov-hm-sle} for the bulk integral means spectrum is actually equal to 5 plus the formula~\eqref{IMS* def}; the $5$ in their formula is a misprint.}
for the average bulk integral means spectrum of whole-plane $\op{SLE}$ in \cite[Theorem~1]{bel-smirnov-hm-sle} for $a \in [a_-, a_+]$,
and with \cite[Conjecture~1]{bel-smirnov-hm-sle} for the a.s.\ bulk integral means spectrum for all values of $a\in\mathbf R$.
\begin{remark}
As conjectured in~\cite{bel-smirnov-hm-sle}, the a.s.\ bulk integral means spectrum of Corollary~\ref{ims cor} differs from the average integral means spectrum computed in~\cite{bel-smirnov-hm-sle} for values of $a\notin [a_-,a_+]$. We explain why this is the case. First, as noted in~\cite{bel-smirnov-hm-sle}, we expect the average and a.s.\ bulk integral means spectra to differ because the function which gives the average bulk integral means spectrum does not satisfy Makarov's~\cite{makarov-fine} characterization of possible integral means spectra. At a more heuristic level, the average integral means spectrum for $a\notin[a_-,a_+]$ is distorted by the occurrence of the small (but still positive) probability event that a conformal map $\phi : \mathbf D\rightarrow V$ satisfies $|\phi'(z)| \approx (1-|z|)^{-s}$ for some $z$ close to $\partial\mathbf D$ and some $s\notin [s_- , s_+]$. However, this event a.s.\ does not occur in the limit (c.f.\ Theorem~\ref{main thm}) so does not effect the a.s.\ bulk integral means spectrum.
\end{remark}
\begin{comment}
xi[s_] := 1 - (4 + kappa)^2 s^2/(8 kappa (1 + s))
f[s_, a_] := xi[s] + a s - 1
Solve[D[f[s, a], s] == 0, s];
FullSimplify
FullSimplify[
f[s, a] /.
scrit[a_] := -1 + (4 + kappa)^2/
Sqrt[(4 + kappa)^2 (16 + kappa (8 - 8 a + kappa))]
sstar[a_] := -1 + (4 + kappa)/Sqrt[ (4 + kappa)^2 - 8 a kappa]
1
FullSimplify[
scrit[a] - sstar[a], {0 < kappa < 4, a < 3 (4 + kappa)^2/(32 kappa)}]
s1 = (4 kappa -
2 Sqrt[2] Sqrt[kappa (2 + kappa) (8 + kappa)])/(4 + kappa)^2;
s2 = (4 kappa +
2 Sqrt[2] Sqrt[kappa (2 + kappa) (8 + kappa)])/(4 + kappa)^2;
Solve[sstar[a] == s1, a];
FullSimplify
Solve[sstar[a] == s2, a];
FullSimplify
FullSimplify[f[s1, a], {0 < kappa < 4}]
FullSimplify[f[s2, a], {0 < kappa < 4}]
a1[kappa_] = -(Sqrt[
2] (16 Sqrt[kappa (2 + kappa) (8 + kappa)] + Sqrt[
kappa^5 (2 + kappa) (8 + kappa)]) +
4 kappa (16 + 3 Sqrt[2] Sqrt[kappa (2 + kappa) (8 + kappa)] +
kappa (10 + kappa)))/(2 kappa (4 + kappa)^2);
a2[kappa_] =
1/(2 kappa (4 + kappa)^2) (Sqrt[
2] (16 Sqrt[kappa (2 + kappa) (8 + kappa)] + Sqrt[
kappa^5 (2 + kappa) (8 + kappa)]) -
4 kappa (16 - 3 Sqrt[2] Sqrt[kappa (2 + kappa) (8 + kappa)] +
kappa (10 + kappa)));
mu[kappa_] := (4 + kappa)^2/(4 kappa);
amin[kappa_] := (-1 -
2 mu[kappa] - (1 + mu[kappa]) Sqrt[1 + 2 mu[kappa]])/mu[kappa];
amax[kappa_] := (-1 -
2 mu[kappa] + (1 + mu[kappa]) Sqrt[1 + 2 mu[kappa]])/mu[kappa];
FullSimplify[a1[kappa] - amin[kappa], {0 < kappa < 4}]
FullSimplify[a2[kappa] - amax[kappa], {0 < kappa < 4}]
(* agrees with Beliaev-Smirnov *)
Plot[a1[kappa] - (-1 - 3 kappa/8), {kappa, 0, 4}]
Plot[3 (4 + kappa)^2/(32 kappa) - a2[kappa], {kappa, 0, 4}]
\end{comment}
\subsection{Outline}
There is a systematic approach to computing Hausdorff dimensions of random fractal sets of the sort we consider here. One first gets a sharp estimate for the probability that a single point is contained in the set (the ``one-point estimate") and uses this to get an upper bound on the Hausdorff dimension. One then defines a subset of the set of interest (the ``perfect points") and obtains an estimate for the probability that any two given points are perfect (the ``two-point estimate"). This enables one to define a Frostman measure on the set of perfect points and thereby obtain a lower bound on the Hausdorff dimension of the set of interest (see \cite[Section~4]{peres-bm} for more on Frostman measures and their connection to Hausdorff dimension). We will follow this outline here. See, e.g., \cite{miller-wu-dim,mww-extremes,lawler-viklund-tip,msw-gasket} for more examples of this technique.
We will now give a moderately detailed outline of the remainder of this paper. The reader should note that this section does not constitute a precise description of all of the proofs in our paper, but rather is only a heuristic guide. For the sake of brevity, many technical details have been omitted, especially in regards to proof of the two-point estimate.
In Section~\ref{prelims sec}, we will give some background on the objects which appear in our proofs, including SLE, the GFF, and the various couplings between them. We will also establish some notations and prove some elementary lemmas which we will need in the sequel.
Next we will prove our one-point estimate. This is done in two stages. In Section~\ref{inverse sec}, we will establish pointwise derivative estimates for the inverse centered Loewner maps $(f_t^{-1})$ for an $\operatorname{SLE}_\kappa$. Roughly, our estimates will take the form
\begin{equation}
\label{rough inverse estimate}
\mathbf P(|(f_t^{-1})'(z)| \approx \epsilon^{-s} , \quad \text{regularity conditions}) \approx \epsilon^{\alpha(s)} ,\qquad \forall s \in (-1,1) ,\qquad \forall z\in \mathbf H \:\operatorname{with} \: \operatorname{Im} z = \epsilon .
\end{equation}
Here $\alpha(s) =\frac{(4 + \kappa)^2 s^2}{8 \kappa (1 + s)} $. The precise meaning of $\approx$ as well as precise regularity conditions in our estimate are given in Section~\ref{inverse setup sec}. The proof of these estimates is based on a family of non-negative martingales for the reverse Loewner flow $(g_t)$, analogous to the martingales for the forward $\operatorname{SLE}_\kappa$ flow in \cite[Section~5]{sw-coord}. The reverse Loewner flow is of interest because we have $g_t \overset{d}{=} f_t^{-1}$ for each fixed $t$ (see, e.g., \cite[Lemma~3.1]{schramm-sle}). For a given $z\in \mathbf H$ with $\operatorname{Im} z = \epsilon$, one can find a martingale $M_t^z$ with the property that $M_t \mathbf 1_{\underline E(z) } \approx \epsilon^{-\alpha(s)}$, where $\underline E(z)$ denotes the event in the probability in~\eqref{rough inverse estimate} with $g_t$ in place of $f_t^{-1}$. We then arrive at
\[
\mathbf P(\underline E(z))\approx \epsilon^{\alpha(s)} \mathbf P^z_*(\underline E(z)),
\]
where $\mathbf P^z_*$ denotes the measure obtained by re-weighting the law of the original $\operatorname{SLE}_\kappa$ process by $M$ (which will be the law of a reverse chordal $\operatorname{SLE}_\kappa(\rho)$ for an appropriate $\rho$). Hence we just need to show $\mathbf P_*^z(\underline E(z)) $ is uniformly positive, independent of $\epsilon$. This is done in two steps. First, to obtain $\mathbf P_*^z(|g_t'(z)| \approx \epsilon^{-s}) \rightarrow 1$ as $\epsilon\rightarrow 0$, we use a coupling of $g_t$ with a GFF together with a coordinate change argument similar in spirit to the proof of \cite[Theorem~8.1]{qle}. To obtain that the auxiliary regularity conditions hold with uniformly positive probability under $\mathbf P_*^z$, we use a combination of stochastic calculus, forward/reverse (in the sense of Loewner flows) SLE symmetry, and GFF coupling arguments.
In Section~\ref{time infty sec} we use the estimate of Section~\ref{inverse sec} to establish pointwise derivative estimates for the ``time infinity" conformal map $\Psi_\eta$ associated with an $\operatorname{SLE}_\kappa$ process $\eta$ from $-i$ to $i$ in the unit disk $\mathbf D$, defined as follows. Let $D_\eta$ be the right connected component of $\mathbf D\setminus \eta$, as in Theorem~\ref{main thm}. Let $\Psi_\eta : D_\eta \rightarrow \mathbf D$ be the unique conformal map fixing $-i$, $i$, and 1. Our estimates for $\Psi_\eta$ take the form
\begin{equation} \label{rough forward estimate}
\mathbf P\left(\operatorname{dist}(z , \eta) \approx \epsilon^{1-s} , \: |\Psi_\eta'(z)| \approx \epsilon^{ s} , \: \text{regularity conditions}\right) \approx \epsilon^{\gamma(s)} ,\qquad \forall s \in (-1,1) ,\qquad \forall z\in \mathbf D
\end{equation}
where $\gamma(s) = \alpha(s) -2s + 1$ and $\alpha(s)$ as above.
The idea of the proof of~\eqref{rough forward estimate} is as follows. First we observe using the Koebe quarter theorem that for each $\epsilon > 0$ and each $t>0$, the set of points $\underline A_\epsilon(t)$ in $\mathbf D$ for which the analogue of the event of~\eqref{rough inverse estimate} with $\mathbf D$ in place of $\mathbf H$ occurs is (approximately) the image under $f_t$ of the set $A_\epsilon(t)$ of points in $\mathbf D$ for which the event of~\eqref{rough forward estimate} holds with $\Psi_\eta$ replaced by $f_t$ and $\eta$ replaced by $\eta([0,t])$. Hence the estimate~\eqref{rough inverse estimate} together with an elementary change of variables yields $\mathbf E(\operatorname{Area} A_\epsilon(t)) \approx \epsilon^{\gamma(s)}$. We are then left to (a) transfer this area estimate from finite time to infinite time and (b) argue that the probability of the event~\eqref{rough forward estimate} does not depend too strongly on $z$. Both tasks will be accomplished by means of various conditioning arguments which rely crucially on the regularity conditions involved in the estimate~\eqref{rough inverse estimate}.
In Section~\ref{haus upper sec}, we will use the estimates~\eqref{rough inverse estimate} and~\eqref{rough forward estimate} to prove upper bounds for the Hausdorff dimensions of the sets $\widetilde\Theta^{s;*}(D_\eta)$ and $\Theta^{s;*}(D_\eta)$, where $*$ stands for $\geq$ or $\leq$ as well as an upper bound for the bulk integral means spectrum of $D_\eta$, as claimed in Corollary~\ref{ims cor}.
In Section~\ref{2pt sec} we prove our two-point estimate.
The first step of the proof is a slight modification of the estimate~\eqref{rough forward estimate}. Namely, let $\overline\eta$ denote the time reversal of $\eta$, which has the law of a chordal $\operatorname{SLE}_\kappa$ from $i$ to $-i$ \cite{zhan-reversibility}. Let $\tau_\beta$ (resp.\ $\overline\eta_\beta$) be the first time $\eta$ (resp.\ $\overline\eta$) hits the ball of radius $e^{-\beta}$ centered at the origin. Let $\eta^{\tau_\beta} = \eta([0,\tau_\beta])$, $\overline\eta^{\overline\tau_\beta} = \overline\eta([0,\overline\tau_\beta])$, and let $\phi_\beta$ be the conformal map from $\mathbf D\setminus (\eta^{\tau_\beta} \cup \overline\eta^{\overline\tau_\beta})$ to $\mathbf D$ which fixes $-i$, $i$, and 1. Then we will use the one-point estimate~\eqref{rough forward estimate} to show
\begin{equation} \label{rough hitting estimate}
\mathbf P\left(|\phi_\beta'(z)| \approx e^{-\beta q}, \: \text{regularity conditions}\right) \approx e^{-\beta \gamma^*(q)} ,\qquad \forall q \in (-1/2,\infty) .
\end{equation}
Here $q = s/(1-s)$ and $\gamma^*(q) = \gamma(s)/(1-s) = (q+1)\gamma(q)$, with $\gamma$ as in~\eqref{rough forward estimate}.
The estimate~\eqref{rough hitting estimate} allows us to break the event that $|\Psi_\eta'(0)| \approx e^{-n \beta}$ down into several stages and estimate each individually. Indeed, if we apply a conformal map from $\mathbf D\setminus (\eta^{\tau_\beta} \cup \overline\eta^{\overline\tau_\beta})$ to $\mathbf D$ which fixes 0, then the rest of the curve will be mapped to another curve whose law is the same as that of $\eta$ (modulo perturbations of its endpoints, which can be dealt with by growing out a little bit more of the curve). In this manner we can construct two approximately independent events $E_{0,1}$ and $E_{0,2}$ whose intersection is contained in the event $\{|\Psi_\eta'(0)| \approx e^{-2\beta} \}$. By iterating this procedure we construct a sequence of approximately independent events $E_{0,j}$ such that $|\Psi_\eta'(0)| \approx e^{-n\beta}$ on $E_n(0) := \bigcap_{j=1}^n E_{0,j}$ and $\mathbf P(E_{z,j}) \approx e^{-\beta \gamma^*(q)}$.\footnote{Actually, we will need to increase $\beta$ by a little bit at each stage for technical reasons, but the basic idea of the argument is the same if we consider a fixed but large $\beta$.}
We can similarly construct events $E_{z,j}$ and $E_n(z)$ for any $z\in\mathbf D$ by first mapping $z$ to 0.
For the lower bound on $\dim_{\mathcal H} \Theta^s(D_\eta)$, the perfect points will be, roughly speaking, the set of $z\in\mathbf D$ for which $E_n(z)$ occurs for every $n\in\mathbf N$. In order to obtain a lower bound on the Hausdorff dimension of the set of perfect points, we need to estimate the probability that $E_n(z)$ and $E_n(w)$ both occur for $z,w\in\mathbf D$, depending on $|z-w|$. To this end, suppose $|z-w| \approx e^{-\beta k}$. We condition on the event $E_k(z)$, corresponding to what happens before we get near $z$ and $w$. After we map out the part of the curve which is grown before the $k$th stage, $z$ and $w$ will be at constant order distance from each other. See Figure~\ref{2pt zoom in fig}.
\begin{figure}[ht!]
\begin{center}
\includegraphics[scale=.8]{figures/drawing-2pt-zoom-in.pdf}
\caption{\label{2pt zoom in fig} If $|z-w| \approx e^{-\beta k}$, then after applying a conformal map which takes the complement of the parts of $\eta$ and $\overline\eta$ involved in the event $E_0^k(z)$ to $\mathbf D$ and takes $z$ to 0, the images of $z$ and $w$ will be at constant order distance from each other. Note, however, that in this setting the derivatives of the stage $k+1$-map near $z$ and $w$ are not approximately independent, since they each depend on the whole curve in the picture on the right. }
\end{center}
\end{figure}
We would like to say that the behaviors of the curve near $z$ and near $w$ are approximately conditionally independent given $E_k(z) $. However, the derivatives of the maps we are interested in depend on the whole curve. Hence we need to localize our events. This is accomplished using a different coupling with a GFF, namely the forward SLE/GFF coupling, or ``imaginary geometry" coupling studied in \cite{dubedat-coupling,shef-zipper,shef-slides,ig1,ig2,ig3,ig4}.
At each stage in the construction of the events $E_n(z)$, we can add auxiliary curves, which are all flow lines (in the sense of \cite{ig1}; c.f.\ Section~\ref{ig prelim}) of the same GFF. These auxiliary curves will form pockets surrounding $z$ with the property that the parts of $\eta$ inside different pockets are independent once we condition on the pockets, and the derivative of $\Psi_\eta$ at a point inside a pocket can be estimated by the derivative of a map which depends only on the behavior of $\eta$ inside this pocket. We then define the event $E_{z,j}$ so that it depends only on the behavior of the curve inside the $j$th pocket. See Figure~\ref{2pt zoom in flow lines fig} for an illustration.
\begin{figure}[ht!]
\begin{center}
\includegraphics[scale=.8]{figures/drawing-2pt-zoom-in-with-flow-lines.pdf}
\caption{\label{2pt zoom in flow lines fig} A modified version of Figure~\ref{2pt zoom in fig} where we add auxiliary curves (shown in green) at each stage to form a pocket. Here we define the events at each stage in terms of only the part of the curve inside the previous pocket. This gives us the needed local independence of the events $E_{z,j}$ and $E_{w,j}$. }
\end{center}
\end{figure}
The independence of the parts of $\eta$ inside different pockets will eventually enable us to establish the two-point estimate needed for the proof of the lower bounds in Theorem~\ref{main thm}.
In Section~\ref{haus lower sec}, we use our two-point estimate to prove lower bounds for the Hausdorff dimensions of the sets $\widetilde\Theta^s(D_\eta)$ and $\Theta^s(D_\eta)$ as well as for the bulk integral means spectrum of $D_\eta$.
Appendices~\ref{local prelim} and~\ref{smac sec} contain several technical lemmas which are needed for the proofs in Section~\ref{2pt sec}.
\subsection{Statement of the estimates} \label{inverse setup sec}
Let $\kappa \in(0,4]$. Let $\eta$ be a chordal $\operatorname{SLE}_\kappa$ process from 0 to $\infty$ in $\mathbf H$. Let $(f_t)$ be its centered Loewner maps. For $z\in\mathbf H$ with $\operatorname{Im} z = \epsilon$, $u > 0$, $ s\in (-1,1]$, $c>0$, and $r>0$, let $\underline E^{s;u}(z;t) = \underline E^{s;u}(z; t,c , r )$ be the event that the following is true.
\begin{enumerate}
\item $c^{-1} \epsilon^{-s + u} \leq |(f_t^{-1})'(z)| \leq c \epsilon^{-s-u}$. \label{deriv reverse}
\item $\operatorname{Im} f_t^{-1}(z) \geq r$.\label{im z r}
\end{enumerate}
\begin{thm} \label{1pt chordal}
Let $\kappa \in (0,4]$. Let $(f_t)$ be the centered Loewner maps of a chordal $\operatorname{SLE}_\kappa$ process from 0 to $\infty$ in $\mathbf H$. Let $z\in \mathbf H$ with $\operatorname{Im} z = \epsilon $ and $R^{-1} \leq |\operatorname{Re} z| \leq R$ for some $R > 1$. Define the event $\underline E^{s;u}(z;t) = \underline E^{s;u}(z; t,c , r )$ as above. Let $G( f_t , \mu)$ be the event of Definition~\ref{G def}.
Let
\begin{equation} \label{alpha def}
\alpha(s) = \frac{(4 + \kappa)^2 s^2}{8 \kappa (1 + s)} ,\qquad \alpha_0(s) = \frac{ (4 + \kappa)^2 s (2 + s)}{8 \kappa (1 + s)^2} .
\end{equation}
For each $t , c,r >0$, each $\mu\in\mathcal M$, each $s\in (-1,1]$, and each $R > 1$, we have
\begin{equation} \label{alpha(s) asymp}
\mathbf P\left(\underline E^{s;u}(z ;t ) \cap G( f_t , \mu) \right) \preceq \epsilon^{\alpha(s ) - \alpha_0(s) u} .
\end{equation}
Furthermore, for each $ r >0$, there exists $ t_* >0$, s.t. for each $t\geq t_*$, we can find $\mu \in\mathcal M$ s.t. for each $c >0$, each $s \in (-1,1)$, and each $u > 0$,
\begin{equation} \label{alpha(s) asymp'}
\mathbf P\left( \underline E^{s;u}(z ;t ) \cap G( f_t , \mu) \right) \succeq \epsilon^{\alpha(s ) +\alpha_0(s) u } .
\end{equation}
In both~\eqref{alpha(s) asymp} and~\eqref{alpha(s) asymp'}, the implicit constants in $\preceq$ and $\succeq$ depend on the other parameters but not on $\epsilon$, and are uniform for $R^{-1} \leq |\operatorname{Re} z| \leq R$.
\end{thm}
\begin{remark}
There are two reasons why we include the condition $G(f_t ,\mu)$ in the estimates of Theorem~\ref{1pt chordal}. The first is that it implies an upper bound on the diameter of $\eta^t$ (see Lemma~\ref{G implies U}), which is needed to estimate some the auxiliary terms which arise in our proof. The second is that, in the sequel, we will often normalize our conformal maps by specifying the images of three marked points on the boundary. The condition $G(f_t , \mu)$ is needed so that the derivative of a conformal automorphism which takes the images of these marked points to where we want them to be is not too large or too small.
\end{remark}
\begin{remark}
Estimates similar to Theorem~\ref{1pt chordal} can be deduced in a somewhat more efficient manner from the results in \cite[Section~3]{schramm-sle}. In particular, \cite[Lemma~3.3]{schramm-sle} implies the upper bound~\eqref{alpha(s) asymp} for a restricted range of parameter values and an estimate similar to~\eqref{alpha(s) asymp'} can be deduced from \cite[ Corollary 3.5]{schramm-sle}. However, these results give optimal bounds only for a certain subset of $(\kappa,s)$ pairs, and do not include the additional regularity conditions on the events in the lower bound which we include here.
\end{remark}
\subsection{Reverse SLE martingales and upper bounds} \label{reverse upper sec}
Let $(g_t)$ be the centered Loewner maps of a reverse $\operatorname{SLE}_\kappa$ flow, so
\begin{equation} \label{reverse Loewner SDE}
d g_t(z) = -\frac{2}{g_t(z)} \, dt - dW_t ,\qquad g_0(z) = z
\end{equation}
for $W_t = \sqrt\kappa B_t$ and $(B_t)$ a standard linear Brownian motion. Our interest in $(g_t)$ stems from the fact that if $(f_t)$ is as in Theorem~\ref{1pt chordal}, then $g_t \overset{d}{=} f_t^{-1}$ for each $t$ (see, e.g. \cite[ Lemma~3.1]{schramm-sle}).
Let $K_t = \mathbf H\setminus g_t(\mathbf H)$ be the hulls corresponding to $(g_t)$. Since $f_t^{-1} \overset{d}{=} g_t$ for each $t$, it is only a minor abuse of notation to replace $f_t^{-1}$ with $g_t$ in the definition of the events of Theorem~\ref{1pt chordal}, and we do so in the remainder of this section.
\subsubsection{Reverse SLE martingales} \label{reverse sle sec}
We state here a result originally due to Lawler~\cite[Proposition~2.1]{lawler-reverse-multifractal}, but in a form which is more convenient for our purposes.
\begin{lem}\label{reverse mart}
Let $\kappa > 0$. Let $(g_t)$ be as above, $\rho\in \mathbf R$, $z\in \mathbf H$, and
\begin{equation} \label{M def}
M_t^z = |g_t'(z )|^{\tfrac{ (8 +2 \kappa - \rho )\rho}{8\kappa }} (\operatorname{Im} g_t(z ))^{-\tfrac{\rho^2}{8\kappa}} |g_t(z) |^{\rho/\kappa} .
\end{equation}
Then $M_t^z$ is a martingale. Let $\mathbf P_*^z$ be the law of $(g_t)$ weighted by $M^z$. The law of $(g_t)$ under $\mathbf P_*^z$ is that of the centered Loewner maps of a reverse $\operatorname{SLE}_\kappa(\rho)$ with a force point at $z$. That is, under the reweighted law,
\begin{equation} \label{dW_t}
dW_t = - \operatorname{Re} \frac{\rho}{g_t(z)} \, dt + \sqrt\kappa dB_t^z
\end{equation}
for $B_t^z$ a $\mathbf P_*^z$-Brownian motion.
\end{lem}
\begin{remark}
The martingale~\eqref{M def} is the reverse SLE analogue of the local martingale of \cite[Section~5]{sw-coord} in the case of a single force point.
\end{remark}
\subsubsection{Proof of the upper bound}
In this subsection we will prove~\eqref{alpha(s) asymp} of Theorem~\ref{1pt chordal}. We will actually prove something a little stronger, namely the following.
\begin{prop} \label{upper bound infty}
Let $\kappa > 0$. Let $\alpha(s)$ be as in~\eqref{alpha def} and let $(g_t)$ be the centered Loewner maps of a reverse $\operatorname{SLE}_\kappa$ as above. Let $c,t,d> 0$. For $s\in [0,1]$, let $\underline E^{s;\infty}(z;t) = \underline E^{s;\infty}(z;t,c, d )$ be the event that $|g_t'(z)| \geq c^{-1} \epsilon^{-s } $ and $|g_t(z)| \geq d^{-1}$. For $s \in (-1,0)$, let $\underline E^{s;\infty}(z;t) = \underline E^{s;\infty}(z ; t , c , d)$ be the event that $|g_t'(z)| \leq c \epsilon^{-s }$ and $|g_t(z)| \leq d$.
For any bounded stopping time $\tau$ for $(g_t)$ we have
\begin{equation} \label{upper bound infty eqn}
\mathbf P\left( \underline E^{s;\infty}(z;\tau ) \right) \preceq \epsilon^{\alpha(s ) } .
\end{equation}
For any $R >1$, the implicit constant in~\eqref{upper bound infty eqn} is uniform for $z\in \mathbf H$ with $R^{-1} \leq |\operatorname{Re} z| \leq R$.
\end{prop}
The estimate~\eqref{alpha(s) asymp} is immediate from Proposition~\ref{upper bound infty} in the case $s\in [0,1]$. To extract~\eqref{alpha(s) asymp} from Proposition~\ref{upper bound infty} in the case $s \in (-1,0)$, we observe that Lemma~\ref{G implies U} implies that $\operatorname{diam} K_t$ is bounded by a constant depending only on $t $ and $\mu$ on the event $G(g_t^{-1} , \mu)$ (c.f. the discussion following Definition~\ref{G def}). For $R^{-1} \leq |\operatorname{Re} z|\leq R$, \cite[eqn. 3.14]{lawler-book} then implies that $|g_t(z)|$ is bounded by a constant depending only on $t, \mu $, and $ R$ on $\underline E^{s;u}(z;t) \cap G(g_t^{-1} , \mu)$. Thus we have $\underline E^{s;u}(z;t) \cap G(g_t^{-1} , \mu) \subset \underline E^{s + u ; \infty }(z;t,c, d)$ for a suitable choice of $d$.
\begin{proof}[Proof of Proposition~\ref{upper bound infty}]
Throughout, we fix $R >1$ and require all implicit constants to be uniform for $z\in \mathbf H$ with $R^{-1} \leq |\operatorname{Re} z| \leq R$.
Let
\begin{equation} \label{optimal rho}
\rho = \rho(s) := \frac{(4 + \kappa) s}{1 + s }.
\end{equation}
and denote by $\mathbf P_*^z$ the law of $(g_t)$ re-weighted by the martingale of Lemma~\ref{reverse mart} with this choice of $\rho$.
By the Loewner equation, $\operatorname{Im} g_\tau(z)$ is bounded above by a constant depending only on the essential supremum of $\tau$. Therefore,
\begin{equation} \label{M_tau succeq}
M_{\tau }^z \mathbf 1_{\underline E^{s;\infty}(z;\tau) } \succeq \epsilon^{\tfrac{-s(8 +2 \kappa - \rho )\rho}{8\kappa} } \mathbf 1_{\underline E^{s;\infty}(z;\tau)}
\end{equation}
(we can replace the $\succeq$ with an $\asymp$ if we assume that $\operatorname{Im} g_t(z)$ is bounded below and $|g_t(z)|$ is bounded above).
Furthermore,
\begin{equation} \label{M_0^z}
M_0^z \asymp \epsilon^{-\tfrac{\rho^2}{8\kappa}} .
\end{equation}
Thus the optional stopping theorem implies
\begin{equation*}
\epsilon^{\tfrac{-s(8 +2 \kappa - \rho )\rho}{8\kappa} } \mathbf P(\underline E^{s;\infty}(z;\tau) )\asymp \mathbf E\left(M_{\tau }^z \mathbf 1_{\underline E^{s;\infty}(z;\tau)} \right)\preceq \epsilon^{-\rho^2/8\kappa} \mathbf P_*^z(\underline E^{s;\infty}(z;\tau) ) .
\end{equation*}
Therefore
\begin{equation} \label{M proportionality}
\mathbf P (\underline E^{s;\infty}(z;\tau) ) \preceq \epsilon^{ \tfrac{ s (8 +2 \kappa - \rho )\rho}{8\kappa } - \tfrac{\rho^2}{8\kappa}} \mathbf P_*^z(\underline E^{s;\infty}(z;\tau) ) .
\end{equation}
The value of the exponent on the right is maximized by taking $\rho = \rho(s)$, as in~\eqref{optimal rho}.
Choosing this value of $\rho$ yields the upper bound~\eqref{upper bound infty eqn}.
\end{proof}
\begin{comment}
Clear[f,kappa, rho];
f[kappa_, rho_, s_] :=
s (8 + 2 kappa - rho) rho/(8 kappa) - rho^2/(8 kappa)
g[kappa_,rho_,s_]:= s (8 + 2 kappa - rho) rho/(8 kappa)
Solve[D[f[kappa, rho, s], rho] == 0, rho]
FullSimplify[rho /.
FullSimplify[f[kappa, rho, s] /. {rho -> ((4 + kappa) s)/(1 + s)}]
FullSimplify[g[kappa, rho, s] /. {rho -> ((4 + kappa) s)/(1 + s)}]
\end{comment}
\subsection{Reduction of the lower bound to a result for a stopping time} \label{reverse lower sec}
Now we turn our attention to the lower bound~\eqref{alpha(s) asymp'} in Theorem~\ref{1pt chordal}. We continue to assume that we have replaced $f_t^{-1}$ with $g_t$ in the definition of the events of Theorem~\ref{1pt chordal}, as in Section~\ref{reverse upper sec}.
Let $T$ be the first time $t$ that $ \operatorname{Im} g_t(z) \geq r$ and fix a time $ \overline t > 0$. Put
\begin{equation}\label{tau def}
\tau = T \wedge \overline t .
\end{equation}
We claim that to prove that~\eqref{alpha(s) asymp'} holds at time $ \overline t$, and hence to finish the proof of Theorem~\ref{1pt chordal}, it is enough to prove the following statement.
\begin{prop} \label{P* pos}
Let $\rho = \rho(s)$ be as in~\eqref{optimal rho}. Let $\mathbf P_*^z$ be the law of a reverse $\operatorname{SLE}_\kappa(\rho)$ process $(g_t)$ with hulls $(K_t)$, with an interior force point located at $z \in \mathbf H$ with $\operatorname{Im} z = \epsilon$. Let $\tau$ be as in~\eqref{tau def}. Define the events $\underline E^{s;u}(z;\tau)$ as in Section~\ref{inverse setup sec}, but with $(g_t)$ in place of $(f_t)$ and the time $\tau$ hull $K_\tau$ for $(g_t)$ in place of $\eta^\tau$. There exists $r_* > 0$ such that for each $r \geq r_*$, we can find $t_* >0$ such that for each $t\geq t_*$, there exists $\mu\in\mathcal M$ such that for any $u>0$ and any $ \overline t \geq t_*$, it holds for each $z\in \mathbf H$ with $\operatorname{Im} z = \epsilon$ and $R^{-1} \leq |\operatorname{Re} z| \leq R$ that
\begin{equation}
\mathbf P_*^z \left( \underline E^{s;u}(z;\tau) \cap G(g_\tau^{-1} , \mu) \right)\succeq 1 .
\end{equation}
Here the implicit constant and all of the above parameters are independent of $\epsilon$ and uniform for $z$ with $R^{-1} \leq |\operatorname{Re} z| \leq R$.
\end{prop}
We will prove Proposition~\ref{P* pos} in the subsequent subsections. In the remainder of this subsection we deduce Theorem~\ref{1pt chordal} from Proposition~\ref{P* pos}.
First we note that the requirement that $r \geq r_*$ in Lemma~\ref{P* pos} is no obstacle; indeed, it is clear that the probability of the event of Theorem~\ref{1pt chordal} is decreasing in $r$.
Observe that $|g_\tau(z)|$ is a.s. bounded above by a positive constant on the event $\underline E^{s;u}(z;\tau) \cap \mathcal G(g_\tau^{-1} , \mu) $ (c.f. Section~\ref{reverse upper sec}). By combining this with the definition of $ \underline E^{s;u}(z;\tau) $ we see that
\begin{equation*}
M_{\tau }^z \mathbf 1_{\underline E^{s;u}(z;\tau) \cap \mathcal G(g_\tau^{-1} , \mu) } \preceq \epsilon^{\tfrac{-( s+u) (8 +2 \kappa - \rho )\rho}{8\kappa} } \mathbf 1_{\underline E^{s;u}(z;\tau) \cap G(g_\tau^{-1} , \mu)} .
\end{equation*}
By~\eqref{M_0^z} and our choice~\eqref{optimal rho} of $\rho$ we then have
\begin{equation} \label{P and P*}
\epsilon^{\alpha(s) + \alpha_0(s) u} \mathbf P_*^z\left(\underline E^{s;u}(z;\tau) \cap G(g_\tau^{-1} , \mu)\right) \preceq \mathbf P\left(\underline E^{s;u}(z;\tau) \cap G(g_\tau^{-1} , \mu) \right) .
\end{equation}
Assuming that Proposition~\ref{P* pos} holds,~\eqref{P and P*} implies~\eqref{alpha(s) asymp'} with $\tau$ in place of $t$. To get the desired bound at the deterministic time $ \overline t$, let $\widetilde g_{t } = g_{t+\tau} \circ g_\tau^{-1}$, so that (by the Markov property) the families of conformal maps $(g_t)$ and $(\widetilde g_t)$ are i.i.d. Let $(\widetilde K_t)$ be the hulls for $(\widetilde g_t)$. For $w\in \mathbf C$, $\mu' \in\mathcal M$ and $C>1$, let $\widetilde E(w) = \widetilde E (w; \overline t, C, \mu' )$ be the event that the following is true.
\begin{enumerate}
\item $C^{-1} \leq \operatorname{dist}(w , \widetilde K_t) \leq C$.
\item $C^{-1} \leq |\widetilde g_{t }'(w)| \leq C$ for each $t \leq \overline t$.
\item $G(\widetilde g_t^{-1} , \mu' )$ occurs for each $t\leq \overline t$.
\end{enumerate}
If $C$ is chosen sufficiently large, depending on $ \overline t $ but uniform for $w$ in compact subsets of $\mathbf H$ and $\mu' \in\mathcal M$ is chosen sufficiently small, then we have that $\mathbf P(\widetilde E (w ) )$ is at least a positive constant depending uniformly on $w$ in compact subsets of $\mathbf H$. Furthermore, since we have a bound on $\operatorname{diam} K_\tau$ on the event $\underline E^{s;u}(z;\tau) \cap G(g_\tau^{-1} , \mu) $, it follows from the Markov property that
\begin{equation*}
\mathbf P\left(\widetilde E (g_\tau(z) ) \cap \underline E^{s;u}(z;\tau) \cap G(g_\tau^{-1} , \mu) \right) \succeq \mathbf P\left(\underline E^{s;u}(z;\tau) \cap G(g_\tau^{-1} , \mu) \right) .
\end{equation*}
On the other hand, the definition of $\widetilde E(g_\tau(z) )$ implies that
\begin{equation*}
\widetilde E (g_\tau(z) ) \cap \underline E^{s;u}(z;\tau) \cap G(g_\tau^{-1} , \mu) \subset \underline E^{s;u}(z; \overline t , c' , r )\cap G(g_{\overline t}^{-1} , \mu \circ \mu')
\end{equation*}
for some $ c' >0 $ depending on the other parameters (here we use that $\operatorname{Im} g_t(z)$ is increasing in $t$ for the condition involving $r$). By making $c$ sufficiently small, we can make $c'$ as small as we like. We conclude that~\eqref{alpha(s) asymp'} at time $\tau$ implies~\eqref{alpha(s) asymp'} at time $ \overline t$.
Thus to prove Theorem~\ref{1pt chordal} it remains to prove Proposition~\ref{P* pos}. The proof is separated into two major steps: first we prove that the derivative condition~\ref{deriv reverse} in the definition of $\underline E^{s;u}(z)$ holds at time $\tau$ with $\mathbf P_*^z$-probability tending to 1 as $\epsilon = \operatorname{Im} z \rightarrow 0$. This is done in Section~\ref{gff deriv sec} via a coupling with a Gaussian free field. Then we prove that $\mathbf P_*^z\left( \{\tau < \overline t\} \cap G(g_\tau^{-1} , \mu) \right)$ is uniformly positive for sufficiently small $\mu$ and sufficiently large $\overline t$. This is done in Section~\ref{auxiliary sec} via a stochastic calculus argument.
\subsection{Coupling with a GFF}
\label{gff deriv sec}
Assume we are in the setting of Proposition~\ref{P* pos}. Let $ h$ be a free boundary GFF on $\mathbf H$, independent from $(g_t)$, normalized so that its harmonic part $\mathfrak h$ vanishes at $i y$ for some $y > 0$ (which we will specify below in such a way that it depends on $ \overline t$, but not $\epsilon$). Let $\mathbf P_h$ be the law of $h$.
For $t\geq 0$ let
\begin{equation} \label{h_t def}
h_t = h \circ g_t + \frac{2}{\sqrt\kappa} \log |g_t(\cdot)| + \frac{\rho}{2\sqrt\kappa} G(g_t(z), g_t(\cdot) ) ,
\end{equation}
where
\[
G(x,y) := -\log|x-y| - \log |\overline x- y|
\]
is the Green's function on $\mathbf H$ with Neumann boundary conditions.
Let $\tau$ be as in~\eqref{tau def}. By \cite[ Proposition~4.1]{shef-zipper}, we have $h_\tau + Q\log |g_\tau'| \overset{d}{=} h_0$, modulo additive constant, where $Q = \frac{2}{\sqrt\kappa} +\frac{\sqrt\kappa}{2} $
is as in~\eqref{Q def}. Let $b_\tau$ be this additive constant, so
\begin{equation}\label{eqD with constant}
h_\tau + Q\log |g_\tau'| - b_\tau \overset{d}{=} h_0 .
\end{equation}
The idea of the proof of~\eqref{P* pos} is to estimate the terms other than $\log |g_\tau'|$ in~\eqref{eqD with constant}, and thereby obtain an estimate on $|g_\tau'|$.
Let
\begin{equation} \label{h_0' def}
h_0 ' = h_\tau + Q\log |g_\tau'| - b_\tau
\end{equation}
so that by~\eqref{eqD with constant} we have $h_0' \overset{d}{=} h_0$. Rearranging the definition of $h_0'$ gives
\begin{align} \label{gff diff}
&Q\log |g_\tau'(w)| = h_0' - h_\tau + b_\tau \nonumber \\
&\quad = h' - h \circ g_\tau + \frac{2}{\sqrt\kappa} \log \frac{|w|}{|g_\tau(w)|} +\frac{\rho}{2\sqrt\kappa} \left( \log \frac{ |g_\tau(w) - g_\tau(z)|}{ |w- z| } + \log \frac{ |g_\tau(w) - \overline{g_\tau(z)}|}{ |w-\overline z| } \right) + b_\tau ,
\end{align}
where here $h'$ is a field with the same law as $h$ and we use $w$ instead of $\cdot$ as a dummy variable.
Since all of the non-GFF terms in~\eqref{gff diff} are harmonic away from $z$, the equation still holds for $w\not=z$ if we replace $ h'$ and $ h \circ g_\tau$ with the circle average processes $ h'_\epsilon$ and $( h\circ g_\tau )_\epsilon$ for these two fields. We will use~\eqref{gff diff} to estimate $b_\tau$ and then to estimate $|g_\tau'(z)|$.
\emph{Throughout this subsection, we require all implicit constants to be independent of $\epsilon$ and uniform for $R^{-1} \leq |\operatorname{Re} z | \leq R$ and all $o_\epsilon(1)$ terms to be uniform for $R^{-1} \leq |\operatorname{Re} z | \leq R$. }
\begin{lem} \label{b control}
Let $\xi >1/2$. If $y$ is chosen sufficiently large (independently of $\epsilon$ and uniform for $R^{-1} \leq |\operatorname{Re} z| \leq R$) then
\begin{equation} \label{b_tau to 0}
(\mathbf P_*^z \otimes \mathbf P_h)\left( \{ |b_\tau| > (\log \epsilon^{-1} )^\xi \} \cap G(g_\tau^{-1} , \mu) \cap \{ \tau < \overline t \} \right) = o_\epsilon(1) .
\end{equation}
\end{lem}
\begin{proof}
If we replace the GFF terms with circle averages in~\eqref{gff diff} and evaluate at $w = i y$, we get
\begin{align} \label{gff diff at iy}
Q\log |g_\tau'(i y)| &= h'_\epsilon(i y) - (h \circ g_\tau)_\epsilon( i y) + \frac{2}{\sqrt\kappa} \log \frac{y}{|g_\tau(i y)|}\nonumber\\
& +\frac{\rho}{2\sqrt\kappa} \left( \log \frac{ |g_\tau(i y) - g_\tau(z )|}{ |i y- z| } + \log \frac{ |g_\tau(i y) - \overline{g_\tau(z)}|}{ |i y -\overline z| } \right) + b_\tau .
\end{align}
By Lemma~\ref{G implies U} $\operatorname{diam} K_\tau \preceq 1$ on $G(g_\tau^{-1} , \mu)$. By \cite[Proposition~3.46]{lawler-book} we have $\operatorname{Im} g_\tau(i y) \asymp |g_\tau(i y)| \asymp 1$ on $ G(g_\tau^{-1} , \mu)$. By the Koebe quarter theorem we also have $|g_\tau'(i y)| \asymp 1$ on $G(g_\tau^{-1} , \mu)$. Hence each of the terms in~\eqref{gff diff at iy} except for $b_\tau$ and the two circle averages is $\asymp 1$ on $ G(g_\tau^{-1} , \mu) \cap \{\tau < \overline t\}$ if $y$ is chosen sufficiently large, depending only on $\mu$. By Lemma~\ref{gff prob}, for $\xi > 1/2$ we have
\[
(\mathbf P_*^z \otimes \mathbf P_h)\left(|h'_\epsilon(i y) - (h \circ g_\tau)_\epsilon( i y)| > (\log \epsilon)^\xi \right) = o_\epsilon(1).
\]
Note that we took $A = \emptyset$ in that lemma to estimate $h_\epsilon'(iy)$ and we took $A = K_\tau$ and used that $ K_\tau$ is independent of $h$ to estimate $(h\circ g_\tau)_\epsilon(iy )$.
By re-arranging~\eqref{gff diff at iy} we conclude.
\end{proof}
\begin{prop}
\label{gff deriv control}
Suppose we define $\rho = \rho(s)$ as in~\eqref{optimal rho}. For any $\mu \in\mathcal M$ and any $c>0$, we have
\begin{align} \label{P* deriv}
\mathbf P_*^z\left( \{ |g_\tau'(z)| \notin [c^{-1 } \epsilon^{-s+u} , c \epsilon^{-s -u}] \} \cap G(g_\tau^{-1} ,\mu) \cap \{ \tau < \overline t \} \right) = o_\epsilon(1) .
\end{align}
\end{prop}
\begin{proof}
Since the circle average process is continuous \cite[Proposition~3.1]{shef-kpz}, we can take the limit as $w\rightarrow z$ in~\eqref{gff diff} to get
\begin{align} \label{gff limit}
Q\log |g_\tau'(z)| &= h_\epsilon'(z) - ( h\circ g_\tau)_\epsilon(z) + \frac{\rho}{2\sqrt\kappa} \log|g_\tau'(z)| -\frac{\rho}{2\sqrt\kappa} \log \epsilon \nonumber \\
&+ \frac{2}{\sqrt\kappa} \log \frac{|z|}{|g_\tau(z)|} + \frac{\rho}{2\sqrt\kappa} \log | \operatorname{Im} g_\tau(z) | + b_\tau .
\end{align}
Since we have a uniform upper bound on $\operatorname{diam} K_\tau$ on the event $ G(g_\tau^{-1} , \mu) $ and $\operatorname{Im} g_\tau(z) = r$ on the event $\{\tau < \overline t\}$, the absolute value of the sum of the fifth and sixth terms in the right in~\eqref{gff limit} is $\preceq 1$ on $G(g_\tau^{-1} , \mu) \cap \{\tau < \overline t\}$.
By Lemma~\ref{gff prob} (applied as in the proof of Lemma~\ref{b control}), for any $\xi > 1/2$,
\begin{equation*}
(\mathbf P_*^z \otimes \mathbf P_h )\left( | h_\epsilon'(z) - ( h\circ g_\tau)_\epsilon(z)| \geq (\log \epsilon^{-1})^\xi \right) = o_\epsilon(1).
\end{equation*}
By Lemma~\ref{b control}, the probability that the last term in~\eqref{gff limit} is $\geq (\log \epsilon)^{1/2}$ and $G(g_\tau^{-1} , \mu) \cap \{\tau < \overline t\}$ occurs is of order $o_\epsilon(1)$. Hence except on an event of $\mathbf P_*^z \otimes \mathbf P_h $-probability of order $o_\epsilon(1)$, on the event $G(g_\tau^{-1} , \mu) \cap \{\tau< \overline t\}$ it holds that
\begin{equation*}
Q\log |g_\tau'(z)| = \frac{\rho}{2\sqrt\kappa} \log|g_\tau'(z)| +\frac{\rho}{2\sqrt\kappa} \log \epsilon^{-1} + o_\epsilon( \log \epsilon^{-1}) .
\end{equation*}
Rearranging, we get that except on an event of $\mathbf P_*^z \otimes \mathbf P_h $-probability of order $o_\epsilon(1)$, on the event $G(g_\tau^{-1} , \mu) \cap \{\tau < \overline t\}$ we have
\begin{align} \label{log g_tau}
\log |g_\tau'(z)| = \frac{\rho }{\kappa + 4 - \rho } \log \epsilon^{-1} + o_\epsilon( \log \epsilon^{-1}) .
\end{align}
With $\rho$ as in~\eqref{optimal rho} we have
\[
\frac{\rho }{\kappa + 4 - \rho } = s,
\]
so integrating out $\mathbf P_h$ yields~\eqref{P* deriv}.
\end{proof}
\subsection{Auxiliary conditions hold with positive probability}
\label{auxiliary sec}
\subsubsection{Setup}
In light of Proposition~\ref{gff deriv control}, to prove Proposition~\ref{P* pos}, and hence Theorem~\ref{1pt chordal}, it remains to prove the following.
\begin{prop} \label{d control}
Let $(g_t)$ be as in~\eqref{reverse Loewner SDE}. and let $(K_t)$ be the associated hulls. Let $z\in \mathbf H$. For $r > \operatorname{Im} z$ let $ T_r$ be the first time $t$ that $\operatorname{Im} g_t(z) = r$.
Let $\rho \in (-\infty , \kappa/2+2)$ and let $\mathbf P_*^z$ be the law of $(g_t)$ weighted by $M^z$, as in Lemma~\ref{reverse mart}. For any given $R > 1$, there exists $r_* > 0$ such that for each $r \geq r_*$, we can find $\mu\in\mathcal M$, $\overline t >0$, and $ p >0$ such that for $z\in \mathbf H$ with $ |\operatorname{Re} z |\leq R$ and $\operatorname{Im} z $ sufficiently small, we have
\[
\mathbf P_*^z\left( \{ T_r \leq \overline t \} \cap G(g_{T_r}^{-1} , \mu) \right) \geq p .
\]
\end{prop}
The proof of Proposition~\ref{d control} will be completed in two stages. First, we will show that we can move the real part of the force point from $\operatorname{Re} z$ to 0 without any pathological behavior (Lemma~\ref{Y to 0}). Then, we will use a forward/reverse SLE symmetry argument to rule out pathological behavior after the real part of the force point has first reached $0$.
We adopt the following notation. Fix $z\in \mathbf H$ with $|\operatorname{Re} z|\leq R$ and $\operatorname{Im} z = \epsilon$. Let
\begin{equation} \label{XYZ def}
Z_t=g_t(z) =X_t+iY_t .
\end{equation}
By~\eqref{dW_t}, we have that under $ \mathbf P_*^z $,
\begin{align} \label{X Y ODE}
dX_t= ( \rho-2 ) \frac{ X_t }{|Z_t|^2} \, dt - \sqrt\kappa dB_t^z ,\qquad dY_t= \frac{2Y_t }{ |Z_t|^2} \, dt , \qquad X_0= \operatorname{Re} z , \qquad Y_0 = \epsilon
\end{align}
for $B_t^z$ a $\mathbf P_*^z$-Brownian motion.
\subsubsection{Pushing the force point to the imaginary axis}
We use the notation~\eqref{XYZ def} and put
\begin{equation} \label{X tau}
S_0 := \inf\{t\geq 0 : X_t = 0\} .
\end{equation}
\begin{lem}\label{Y to 0}
Suppose we are in the setting of Proposition~\ref{d control}. Let $Z_t = X_t + i Y_t$ be as in~\eqref{XYZ def} and let $S_0$ be as in~\eqref{X tau}. Let $\zeta \in (0,1)$. There exists $d > 0$ and $p_0 > 0$, independent of $\epsilon$ and of $X_0 \in [-R,R]$, such that whenever $\epsilon \leq \zeta$ the probability of the event $E_0 = E_0(\zeta , d )$ that
\begin{enumerate}
\item $S_0 \leq \zeta$;\label{t0 cond}
\item $Y_{S_0} \leq 5\zeta^{1/2}$; \label{Y cond}
\item $\operatorname{diam} K_{S_0} \leq d $; \label{diam cond}
\end{enumerate}
is at least $ p_0 $.
\end{lem}
\begin{proof}
By symmetry we can assume without loss of generality that $X_0 >0$. We will treat the conditions in the definition of $E_0$ in order. \\
\noindent\textit{Condition~\ref{t0 cond}:}
Let
\begin{equation} \label{min nu}
\nu > 1\wedge \left( \frac{2(\rho-2)}{\kappa} + 1 \right) .
\end{equation}
\begin{comment}
Clear[kappa]
rho[s_] := ((4 + kappa) s)/(1 + s);
Solve[2 (rho[s] - 2)/kappa + 1 == 2, s]
\end{comment}
Let $\widetilde X $ be $\sqrt\kappa$ times a Bessel process driven by $-B_t^z$, started from $X_0$, of dimension $\nu$.
We have
\begin{equation*}
d(\widetilde X_t - X_t) = \frac{ X_t ( a X_t- (\rho-2)\widetilde X_t) + a Y_t^2 }{(X_t^2 + Y_t^2) \widetilde X_t} \, dt ,
\end{equation*}
where $a = \kappa (\nu-1)/2 > 0\wedge (\rho-2)$. This is strictly positive whenever $ X_t > \widetilde X_t$ (since $\widetilde X_t \geq 0$). This implies that a.s.\
\begin{equation}\label{X tilde X}
\widetilde X_t \geq X_t ,\qquad \forall t\leq S_0 .
\end{equation}
Our choice~\eqref{optimal rho} for $\rho$ implies that~\eqref{min nu} holds for some Bessel dimension $\nu \in (0,2)$, in which case~$\widetilde X$ hits~$0$ before time $\zeta$ with uniformly positive probability \cite[Proposition~1.21]{lawler-book}. From this and~\eqref{X tilde X} we conclude that we can find $p_0 > 0$ independent of $\epsilon$ and of $X_0 \in [-R,R]$ such that
\begin{equation} \label{zeta prob}
\mathbf P\left(S_0 \leq \zeta \right) \geq 2p_0 .
\end{equation}
\noindent\textit{Condition~\ref{Y cond}:}
By~\eqref{X Y ODE} we have that $Y$ is increasing and $\partial_t Y_t^2 \leq 4$. Hence $Y_t \leq 4t^{1/2} + \epsilon$, so on the event $\{S_0 \leq \zeta \}$ we have $Y_{S_0} \leq 5\zeta^{1/2}$.
\noindent\textit{Condition~\ref{diam cond}:}
Let $\widetilde X$ be a Bessel process of dimension $\nu$ started from $X_0$ as in the proof of Condition~\ref{t0 cond}. Since $\widetilde X$ and $B^z$ are a.s.\ bounded up to time $\zeta$ and their laws do not depend on $\epsilon$, it follows from~\eqref{X tilde X} and~\eqref{zeta prob} that we can find $C_0 > 0$, independent of $\epsilon$ and uniform for $X_0 \in [-R,R]$ such that the probability of the event $E_0^*$ that
\begin{enumerate}
\item $S_0 \leq \zeta$;
\item $Y_{S_0} \leq 5\zeta^{1/2}$;
\item $\sup_{t\leq \zeta} |\sqrt \kappa B_t^z| \leq C_0$;
\item $\sup_{t\leq \zeta} |X_t| \leq C_0$;
\end{enumerate}
is at least $p_0$. Note that for the last condition we use~\eqref{X tilde X}.
By~\eqref{X Y ODE}, we have for $t\leq S_0$ that
\begin{equation} \label{int of Re}
| \rho-2 | \int_0^t \frac{X_s}{X_s^2 +Y_s^2} \, ds \leq |X_0| + |X_t| + |\sqrt\kappa B_t| .
\end{equation}
In the case $\rho \not=2$, it follows from~\eqref{int of Re} that on the event $E_0^*$,
\begin{equation} \label{int of Re sup}
\int_0^t \frac{X_s}{X_s^2 +Y_s^2} \, ds \leq C_1 := \frac{R + 2C_0}{|\rho-2|} .
\end{equation}
In the case $\rho =2$, it follows from~\eqref{X Y ODE} that $X$ is a constant times a Brownian motion, so in this case we can (using condition~\ref{t0 cond}) find a possibly larger constant $C_1$, still independent of $\epsilon$, such that~\eqref{int of Re sup} holds with probability at least $1-p_0/2$. In this case we add this latter condition to the event $E_0^*$.
Now consider some $b \in \mathbf R$, $|b| > 1$. Let $\delta> 0$. Let $\tau_b$ be the first time $t$ that $|g_t(b)| \leq \delta$. By~\eqref{dW_t} and the Loewner equation, we have
\begin{equation*}
g_t(b) = -\int_0^t \frac{2}{g_s(b)} \, ds + \rho \int_0^t \frac{X_s}{X_s^2 +Y_s^2} \, ds -\sqrt\kappa B_t^z + b .
\end{equation*}
So, it follows from~\eqref{int of Re sup} that on $E_0^*$ we have
\begin{equation*}
\inf_{t \leq S_0 \wedge \tau_b} |g_t (b)| \geq |b| - C_2,
\end{equation*}
where
\[
C_2 = 2 \zeta \delta^{-1} + |\rho| C_1 + C_0 .
\]
Hence if we take $|b| > 2C_2$, then we have $\inf_{t \leq S_0 \wedge \tau_b} |g_t (b)| \geq C_2$, which implies $\tau_b > S_0$ (provided we choose $\delta < C_2/2$).
In particular, if $b>1$ is chosen sufficiently large (independent of $\epsilon$ and $X_0\in [-R,R]$), then $g_{S_0}(-b)$ and $g_{S_0}(b)$ lie in $\mathbf R$. Therefore the map $g_\tau^{-1}$ takes $\partial K_\tau$ into $[-b,b]$. This implies that the harmonic measure from $\infty$ of $K_\tau$ in $\mathbf H\setminus K_\tau$ is at most $2\pi b$, so by \cite[Equation 3.14]{lawler-book}, it follows that $\operatorname{diam} K_{S_0}$ is bounded by a constant $d >0$ independent of $\epsilon$ and $X_0\in [-R,R]$ on $E_0^*$. Since $\mathbf P(E_0^*) \geq p_0$, the lemma follows.
\end{proof}
\subsubsection{Pushing the force point starting from 0}
In light of the strong Markov property and Proposition~\ref{d control}, we now need to consider the behavior of the process~\eqref{X Y ODE} if we start $X_0$ from $0$ and $Y_0$ from $y \in [\epsilon ,5\zeta^{1/2}]$ for $\zeta$ small but fixed. For this, we first need to review some calculations from \cite[Section~3]{wedges}. Throughout this subsection, we assume $X_0 = 0$, $Y_0 = y \in [\epsilon , 5\zeta^{1/2}]$. Let
\begin{equation} \label{s_y def}
\theta_t = \operatorname{arg} Z_t \quad \operatorname{and} \quad s_y = \frac12 \log y .
\end{equation}
For $s\geq s_y$ define $\sigma(s)$ by
\begin{equation} \label{sigma(s) def}
s= \int_{0}^{\sigma(s)} \frac{ 1 }{|Z_u|^2} \, du + s_y ,
\end{equation}
so $d\sigma(s) = |Z_{\sigma(s)}|^2 \, ds$ and $\sigma(s_y) = 0$.
Denote processes under the time change $t\mapsto \sigma(s)$ by a star, so $\theta_s^* = \theta_{\sigma(s)}$, etc.
By some elementary calculations using It\^o's formula (see the proof of \cite[Proposition~3.8]{wedges}), we have $d\log Y_s^* = 2\, ds$ and
\begin{equation} \label{d theta*}
d \theta_s^* = \sqrt{\kappa} \sin \theta_s^* d\widehat B_s + \left( 2+\frac{\kappa}{2} - \frac{\rho }{2 } \right) \sin(2\theta_s^*) \, ds
\end{equation}
for $\widehat B_s$ a Brownian motion. Since $Y_{s_y}^* = Y_0 = y$, it follows that $Y_s^* = e^{2s}$. Furthermore, as explained in the proof of \cite[Proposition~3.8]{wedges}, there is a unique stationary distribution for the SDE~\eqref{d theta*} which takes the form
\begin{equation} \label{stationary distribution}
C \sin^\beta(\theta) \, d\theta ,\qquad \beta = \frac{8-2\rho }{\kappa} ,
\end{equation}
where $C$ is a normalizing constant.
Let $\widetilde\theta_s^*$ be a stationary solution to~\eqref{d theta*}. Let $\widetilde Z_s^* = \frac{ e^{2s} e^{ i \widetilde\theta_s^*} }{\sin \widetilde\theta_s^*}$, so that $\operatorname{Im} \widetilde Z_s^* = e^{2s}$ and $\operatorname{arg} \widetilde Z_s^* = \widetilde\theta_s^*$. Let $\widetilde W_s^*$ be determined by $\widetilde Z_s^*$ in the same manner that $W_s^*$ is determined by $Z_s^*$. Let
\begin{equation*}
\widetilde \sigma (t) := \int_0^t |\widetilde Z_u^* |^2 \, du .
\end{equation*}
Denote processes under the time change $s\mapsto \widetilde\sigma^{-1}(t)$ by removing the star. Then we have that $(\widetilde \theta_t , \widetilde Z_t , \widetilde W_t)$ are related in the same manner as $(\theta_t , Z_t , W_t)$. Moreover,
\[
\widetilde\sigma(s) = \inf\{t \in \mathbf R : Y_t = e^{2s}\} .
\]
Following \cite[Section~3]{wedges}, we define \emph{a reverse $\operatorname{SLE}_\kappa(\rho)$ process with a force point infinitesimally above 0} to be the Loewner evolution driven by $\widetilde W$.
\begin{lem} \label{stationary control}
Let $(\widetilde g_t)$ be the reverse Loewner maps of a reverse $\operatorname{SLE}_\kappa(\rho)$ process with a force point infinitesimally above 0, with hulls $(\widetilde K_t)$. We adopt the notation given just above, so in particular a star denotes processes under the time change $t\mapsto \widetilde\sigma(s)$. For $ \overline s \in\mathbf R$ and $s>0$, let
\begin{equation} \label{stationary composed maps}
\widetilde g_{\overline s , s}^* := \widetilde g_{s-\overline s }^* \circ (\widetilde g_{\overline s}^*)^{-1} \quad \operatorname{and } \quad \widetilde K_{\overline s,s}^*:= \widetilde K_{s -\overline s}^* \setminus \widetilde g_{ \overline s , s}^*(\widetilde K_{\overline s}^*) .
\end{equation}
For $t,d > 0$ and $\mu\in\mathcal M$, let $F = F (s, \overline s, d, t, \mu)$ be the event that the following is true.
\begin{enumerate}
\item $\widetilde \sigma(s ) \leq t $. \label{F time}
\item For each $\delta > 0$, the harmonic measure from $\infty$ of each of $[-\delta ,0]$ and of $[0 ,\delta ]$ in $\mathbf H\setminus \left( \widetilde K_s^* \cup \widetilde g_{\overline s , s}( B_{ d}(0) \cap \mathbf H) \right)$ is at least $\mu(\delta)$. \label{F3 G}
\end{enumerate}
For any given $d > 0$, $\overline s \in \mathbf R$, and $p \in (0,1)$, we can find $s_* >0$ (depending only on $d$, $\overline s$, and $p$) such that whenever $s \geq s_*$, there exists $ t > 0$ and $\mu \in \mathcal M$ (depending only $s$, $\overline s$, and $p$) such that
\begin{equation*}
\mathbf P\left(F \right) \geq 1-p .
\end{equation*}
\end{lem}
\begin{proof}
By \cite[Proposition~3.10]{wedges}, for each $s >0$, the law of $\widetilde K_{s }^*$ given $\widetilde Z_{s }^*$ is that of a forward chordal $\operatorname{SLE}_\kappa(\rho-8)$ hull with an interior force point at $\widetilde Z_{s}^*$ stopped at the first time it hits its force point. By \cite[Theorem 3]{sw-coord} this law is that same as that of a radial $\operatorname{SLE}_\kappa( \kappa + 2-\rho)$ from 0 to $\widetilde Z_{s }^*$ with a force point at $\infty$, run until the first time it hits $\widetilde Z_{s }^*$. Since $ \kappa +2 -\rho > \kappa/2 - 2$ (by our choice of $\rho$) \cite[Theorem~1.12]{ig4} implies that such a process is transient (i.e., almost surely tends to its target point) and \cite[Lemma~2.4]{ig4} implies that it a.s.\ does not intersect itself or hit $\mathbf R \cup \{\infty\}$. In particular, $\widetilde K_{s }^*$ is a.s.\ a simple curve which does not intersect $\mathbf R$ except at its starting point and has finite half-plane capacity. Therefore the same is a.s.\ true of $\widetilde K_{\overline s,s}^*$ for each $s , \overline s \in\mathbf R$.
By scale invariance (which follows from uniqueness of the stationary solution to~\eqref{d theta*}), for each $c\in\mathbf R$ we have
\[
\left\{ e^{2c} \widetilde g_{s - c}^*(e^{-2c} \cdot) \,: \, s\in\mathbf R \right\} \overset{d}{=} \{ \widetilde g_s^* \,:\, s\in \mathbf R\} .
\]
Hence for each $s > \overline s\in \mathbf R$ we have
\begin{equation} \label{comp scaling}
e^{ 2c} \widetilde g_{\overline s , s }^*(e^{-2c} \cdot) \overset{d}{=} \widetilde g_{\overline s , s + 2 c }^* .
\end{equation}
\begin{comment}
Inverse of $e^{2c} \widetilde g_{\overline s - c}^*(e^{-2c} \cdot)$ is $e^{2c} (\widetilde g_{s - c}^*)^{-1} (e^{-2c} \cdot)$. Compose this with function at time $s-\overline s$ to get
\alb
e^{2c} \widetilde g_{s -\overline s - c}^*( (\widetilde g_{\overline s-c}^*)^{-1}(e^{-2c} \cdot) ) &= e^{2c} \widetilde g_{\overline s -c , s-2c}^*( e^{-2c} \cdot ) .
\ale
This agrees in law with $\widetilde g_{\overline s , s }^*$. Replace $\overline s$ with $\overline s + c$ and $s$ with $s + 2c$ to get~\eqref{comp scaling}.
\end{comment}
Since each $K_{\overline s , s}^*$ a.s.\ does not intersect $\mathbf R$ except at its starting point, for any given $\overline s\in\mathbf R$ and $s>0$, there a.s.\ exists some $\delta > 0$ and $\lambda > 0$ (random) such that $\operatorname{Im} \widetilde g_{s, \overline s}^*(z) \geq \lambda $ for each $z\in B_{\delta}(0)$. By~\eqref{comp scaling}, if we are given $d > 0$, $\overline s \in \mathbf R$, and $p > 0$ we can find $s_* > 0$ such that whenever $s \geq s_*$, it holds with probability at least $1-p/2$ that $\operatorname{Im} \widetilde g_{s, \overline s}^*(z) \geq 1$ for each $z\in B_d(0)$. Since $K_{\overline s , s}^*$ a.s.\ does not intersect $\mathbf R$ and a.s.\ has finite half plane capacity, for each such $s$ we can find $t$ and $\mu$ as in the statement of the lemma such $\mathbf P(F) \geq 1-p$.
\end{proof}
By convergence of solutions of SDE's to their stationary distribution, the above lemma will allow us to control the behavior of $\theta_s^*$ after a sufficiently large amount of time has passed (see the proof of Lemma~\ref{Y from 0} below). We also need to rule out pathological behavior at short time scales, which is the purpose of the next lemma.
\begin{lem} \label{small times}
Let $s_y $ be as in~\eqref{s_y def}. For any $p \in (0,1)$ and $v >0$, there is a $b>0$ depending on $v$, $p$, and $\zeta$ but not $\epsilon$, $\zeta$, or the particular choice of $y\in [\epsilon , 5\zeta^{1/2}]$ such that
\begin{equation*}
\mathbf P_*^z\left( K_{s_y + v}^* \subset B_{b }(0) \right) \geq 1-p.
\end{equation*}
Here $K_s^* = K_{\sigma(s)}$, for $(K_t)$ the hulls of the reverse Loewner evolution driven by $(W_t)$.
\end{lem}
\begin{proof}
First note that $\theta_s^*$ cannot hit $0$ or $\pi$. To see this, one observes that $\theta_s^*$ is a time change of a constant multiple of the process of \cite[Section~1.11]{lawler-book} with $a = (4 +\kappa-\rho)/\kappa > 1/2$, so the claim follows from \cite[Lemma~1.27]{lawler-book}.
Therefore there exists $\delta > 0$ depending only on $v$ such that if $\theta_s^*$ is started at time $s_y$ with initial condition $\theta_{s_y}^* = \pi/2$ then with probability at least $1-p/2$ we have $\theta_s^* \in (\delta, 2\pi-\delta)$ for each $s\in [s_y , s_y + v]$. Let $G$ be the event that this occurs.
On the event $G$, we can find a constant $c >0$ depending only on $\delta$ such that $X_s^*/Y_s^* \leq c$ for $s \in [s_y , s_y+v]$. It then follows from~\eqref{X Y ODE} that on this event we have
\begin{equation*}
\partial_t Y_t \geq \frac{1}{c Y_t} ,\qquad \forall t\leq \sigma (s_y+v)
\end{equation*}
for a possibly larger $c $.
This implies
\begin{equation} \label{Y big}
Y_t^2 \geq c^{-1} t + y^2
\end{equation}
for a possibly larger constant $c$. Therefore, $t\leq \sigma ( s_y+v)$. In particular, $( e^{4v} -1) y^2 = Y_{\sigma(s_y + v)}^2 - y^2 \geq c^{-1} \sigma(s_y+v)$, so for some possibly larger constant $c $ we have
\begin{equation} \label{sigma small}
\sigma(s_y + v) \leq c y^2 .
\end{equation}
Let $B_t^z$ be the Brownian motion of~\eqref{dW_t}. We can find a $C >0$ depending only on $\zeta$ such that with probability at least $1-p/2$, we have $|\sqrt\kappa B_t^z| \leq Cy$ for each $t \in [0, c y^2]$. Let $G'$ be the event that this occurs and that $G$ occurs. It follows from~\eqref{dW_t} and~\eqref{sigma small} that on $G'$, we have
\begin{equation*}
\sup_{0 \leq t\leq \sigma ( s_y+v)} |W_t| \preceq 1 .
\end{equation*}
By \cite[Lemma~4.13]{lawler-book} we then have $\operatorname{diam} K_{\sigma(s_y + v)} \preceq 1$.
\end{proof}
\begin{lem}\label{Y from 0}
Suppose we are in the setting of this subsection (so that in particular $X_0 = 0$ and $Y_0 = y$). Let $T_r' = \inf\{t \geq 0 : Y_t = r\} = \sigma(\frac12 \log r)$. Also let $d> 0$ and $p \in (0,1)$. There is an $r_* > 0$ (depending on $\zeta , d$, and $p$) such that for $r \geq r_*$, there exists $t_1 > 0$ and $\mu\in\mathcal M$, independent of $\epsilon$ and the particular choice of $y\in [\epsilon , 5\zeta^{1/2}]$ such that the probability of the event $E_1 = E_1(r , d , t_1 , \mu )$ that
\begin{enumerate}
\item $T_r' \leq t_1$; \label{E1 time}
\item For each $\delta > 0$, the harmonic measure from $\infty$ of each of $[-\delta ,0]$ and of $[0 ,\delta ]$ in $\mathbf H\setminus \left( K_{T_r'} \cup g_{T_r'}( B_{ d}(0) \cap \mathbf H ) \right) $ is at least $\mu(\delta)$; \label{E1 hm condition}
\end{enumerate}
is at least $1-p$.
\end{lem}
\begin{remark}
The purpose of condition~\ref{E1 hm condition} is as follows. When we compose with $g_{S_0}$ on the event $E_0$ of Lemma~\ref{Y to 0}, the part of the hull grown before time $S_0$ is ``pushed" into $g_{T_r'}(B_d(0))$. The harmonic measure condition~\ref{E1 hm condition} together with Lemma~\ref{G implies U} will then imply the occurrence of $G(g_{T_r}^{-1} , \mu)$ on the event $E_0\cap E_1$.
\end{remark}
\begin{proof}[Proof of Lemma~\ref{Y from 0}]
Define the processes $X_s^* , Y_s^* ,Z_s^* , \sigma(s) , \theta_s^*$ as above.
Let $(\widetilde g_t)$ be the reverse Loewner maps of a reverse $\operatorname{SLE}_\kappa(\rho)$ process with a force point immediately above 0. We adopt the notation given just above Lemma~\ref{stationary control}, so that $\widetilde Z_t$ is the evolution of the force point and $\widetilde \theta_s^* = \operatorname{arg} \widetilde Z_s^*$ is the corresponding stationary solution to~\eqref{d theta*}.
By convergence of the law of the solution of~\eqref{d theta*} to its stationary distribution, there exists $v > 0$, independent of $\epsilon$, $\zeta$, and $y\in [\epsilon , 5\zeta^{1/2}]$, such that the following is true. The total variation distance between the law of $\theta_{s_y +v}^*$, started from $\pi/2$ at time $s_y$, and the stationary distribution~\eqref{stationary distribution} is at most $ p/4$. Let $\overline s = s_y +v$, and note that if $r$ is chosen sufficiently large depending only on $v$ then we have $\overline s \leq s_{r/2}$. We can couple $\theta^*$ with $ \widetilde \theta^*$ in such a way that with probability at least $1-p/3$, these two processes agree at time $\overline s$ and (by the Markov property) at every time thereafter. Let $F_1$ be the event that $\theta_{s}^* = \widetilde \theta_{s}^*$ for each $s \geq \overline s$.
Define the maps $\widetilde g_{ \overline s ,s}^*$ and the hulls $\widetilde K_{ \overline s , s}^*$ as in~\eqref{stationary composed maps}. Define $g_{s,\overline s}^*$ and $K_{\overline s,s}^*$ analogously but with $g_s^*$ and $K_s^*$ in place of $\widetilde g_s^*$ and $\widetilde K_s^*$. We have that $(\theta_s^* , e^{2s})$ determines $W_s^*$ and hence also $(g_s^*)$. Similarly for the corresponding processes under the stationary distribution. Therefore on $F_1$, we have
\begin{equation} \label{F1 agree}
g_{\overline s , s}^* = \widetilde g_{\overline s , s}^* ,\qquad K_{\overline s,s}^* = \widetilde K_{\overline s,s}^* ,\qquad \forall s \geq 0.
\end{equation}
By Lemma~\ref{small times} we can find a $b >0$ depending only on $v$ such that the probability of the event
\begin{equation*}
F_2:=\{ K_{\overline s}^* \subset B_{b }(0) \}
\end{equation*}
is at least $1-p/3$. By \cite[Proposition 3.46]{lawler-book} there is a deterministic constant $d' > 0$ depending only on $b$ and $\overline s$ such that on the event $F_2$ we have
\begin{equation} \label{K_ol s small}
K_{\overline s}^* \cup g_{\overline s}^*\left(B_d(0)\cap \mathbf H \right) \subset B_{d'}(0)\cap \mathbf H.
\end{equation}
Let $s_*$ be chosen so that the conclusion of Lemma~\ref{stationary control} holds with this choice of $d'$ in place of $d$, $p /3$ in place of $p$, and $\overline s$ as above. Let $s \geq s_*$ and let $t$ and $\mu$ be chosen so that with $F_3 = F (s , \overline s , t, d' , \mu)$ the event of Lemma~\ref{stationary control} we have $\mathbf P(F_3) \geq 1-p/3$.
Then we have
\[
\mathbf P\left(F_1 \cap F_2 \cap F_3 \right) \geq 1-p .
\]
We will now conclude the proof by showing that $F_1\cap F_2 \cap F_3 \subset E_1$ for an appropriate choice of parameters.
If we set $r_* = e^{2(s_* + \overline s)}$ and $r = e^{2(s + \overline s)}$, then $r$ ranges over $[r_* , \infty)$ as $s$ ranges over $[0,\infty)$. On the event $F_1 \cap F_2 \cap F_3$, we have
\alb
T_r '= \operatorname{hcap} K_{s + \overline s} &\leq \operatorname{hcap} K_{s,\overline s}^* + \operatorname{hcap} K_{\overline s}^* .
\ale
The first term is at most $t$ by condition~\ref{F time} in the definition of $F_3$ and~\eqref{F1 agree}. The second term is at most a finite constant depending only on $b$. Hence we can find a $t_1 > 0$ as in the statement of the lemma such that on $F_1\cap F_2 \cap F_3$ we have $T_r' \leq t_1$. Furthermore, on $F_1 \cap F_2 \cap F_3$,
\alb
K_{T_r'} \cup g_{T_r'}\left(B_d(0)\cap \mathbf H \right) &= K_{\overline s , s}^* \cup g_{\overline s , s }\left( K_{\overline s}^* \cup g_{\overline s}^*\left(B_d(0)\cap \mathbf H \right) \right) \\
&= \widetilde K_{\overline s , s}^* \cup \widetilde g_{\overline s , s_*}\left( K_{\overline s}^* \cup g_{\overline s}^*\left(B_d(0)\cap \mathbf H \right) \right) \quad \text{(by~\eqref{F1 agree})} \\
&\subset \widetilde K_{\overline s , s}^* \cup \widetilde g_{\overline s , s_*}\left( B_{d'}(0) \cap \mathbf H \right) \quad \text{(by~\eqref{K_ol s small})} .
\ale
Hence it follows from condition~\ref{F3 G} in the definition of $F_3$ that we can find $\mu\in\mathcal M$ (satisfying the conditions of the lemma) such that condition~\ref{E1 hm condition} in the definition of $E_1$ holds on $F_1\cap F_2 \cap F_3$.
\end{proof}
\subsubsection{Proof of the main proposition}
Now we can combine the results of the previous two subsections to complete the proof of Proposition~\ref{d control}.
\begin{proof}[Proof of Proposition~\ref{d control}]
Let $\zeta >0$, $d>0$, and $p_0 > 0$ be as in Lemma~\ref{Y to 0}, and let $E_0 = E_0(\zeta , d )$ be the event of that lemma, so that $\mathbf P(E_0) \geq p_0$.
Conditional on $\{g_t : t\leq S_0\}$, the law of $\{\widehat g_t : t\geq 0\} := \{g_{t+S_0} \circ g_{S_0}^{-1} : t\geq 0\}$ is the same as that of the process started from $X_0 = 0$ and $Y_0 = Y_{S_0}$. Note that $Y_{S_0} \in [\epsilon , 5\zeta^{1/2}]$ on $E_0$. Define the times $T_r'$ and the events $E_1 = E_1(r , t_1 , d, \mu)$ as in Lemma~\ref{Y from 0} but with $(\widehat g_t)$ in place of $(g_t)$. Let $r_*$, $\mu$, and $t_1$ satisfy the conclusion of this lemma for $d$ as above and $p=1/2$, so that if $r \geq r_*$ then $\mathbf P(E_1 | E_0) \geq 1/2$, whence $\mathbf P(E_0 \cap E_1) \geq p_0/2$.
By condition~\ref{t0 cond} in the definition of $E_0$ and condition~\ref{E1 time} in the definition of $E_1$ we have $T\leq \zeta + t_1$ on $E_0\cap E_1$. By condition~\ref{E1 hm condition} in the definition of $E_1$, on the event $E_0 \cap E_1$, the harmonic measure from $\infty$ of each of $[-\delta ,0]$ and $[0, \delta ]$ in $\mathbf H\setminus K_{T_r}$ is at least $\mu(\delta)$. By Lemma~\ref{G implies U} we can find $\mu'\in\mathcal M$ and $\overline t >0$ such that
\[
E_0\cap E_1 \subset G(g_{T_r}^{-1} , \mu') \cap \{T_r \leq \overline t\} .
\]
This proves the statement of the proposition.
\end{proof}
\subsection{Results for the disk}
\label{disk sec}
In the sequel we will work mostly in the unit disk $\mathbf D$ rather than in the upper half plane $\mathbf H$. In this brief subsection we make some trivial remarks about how Theorem~\ref{1pt chordal} generalizes to this setting.
Suppose $\eta$ is a chordal $\operatorname{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$. Let $\psi : \mathbf D\rightarrow \mathbf H$ be the conformal map taking $-i$ to $0$, $i$ to $\infty$, and having positive real derivative at 0. Suppose $\eta$ is parametrized in such a way that $\psi(\eta)$ is parametrized by half-plane capacity. For each time $t \geq 0$, let
\begin{equation*}
f_t : \mathbf D\setminus \eta^t \rightarrow \mathbf D
\end{equation*}
be defined so that $\psi \circ f_t \circ \psi^{-1}$ is the time $t$ centered Loewner map for $\psi(\eta)$.
For $s\in(-1,1)$, $u>0$, $z\in \mathbf D$ with $1-|z|=\epsilon$ and $t,c,d > 0$, let $\overline{ E}_\epsilon^{s;u}(z;t) = \overline{E}^{s;u}(z ; t , c , d ) $ be the event that the following is true.
\begin{enumerate}
\item $\epsilon^{-s+u} \leq |(f_t^{-1})'(z)| \leq \epsilon^{-s-u}$.
\item $f_t^{-1}(z) \in B_d(0)$.
\end{enumerate}
Then in this context Theorem~\ref{1pt chordal} reads as follows.
\begin{cor}[Theorem~\ref{1pt chordal} for the disk] \label{1pt chordal disk}
Suppose we are in the setting described just above. Let $\delta > 0$. Let $z\in \mathbf D$ with $|z-i| , |z + i| \geq \delta$ and $1-|z| = \epsilon$. Define the events $\mathcal G(\cdot)$ as in Definition~\ref{G infty def}.
For each $t , c, d , \delta >0$, each $s\in (-1,1]$, and each $\mu\in\mathcal M$,
\begin{equation} \label{disk asymp}
\mathbf P\left(\overline{ E}^{s;u}(z;t) \cap \mathcal G( f_t , \mu) \right) \preceq \epsilon^{\alpha(s ) - \alpha_0(s) u} .
\end{equation}
Furthermore, there exists $ t_* >0$ such that for each $t\geq t_*$, we can find $\mu \in\mathcal M$ and $d>0$ such that for each $c >0$, each $s \in (-1,1)$, and each $u > 0$,
\begin{equation} \label{disk asymp'}
\mathbf P\left( \overline E^{s;u}(z ;t ) \cap \mathcal G( f_t , \mu) \right) \succeq \epsilon^{\alpha(s ) +\alpha_0(s) u } .
\end{equation}
In both~\eqref{disk asymp} and~\eqref{disk asymp'}, the implicit constants in $\preceq$ and $\succeq$ depend on the other parameters but not on $\epsilon$, and are uniform for $z\in \mathbf D$ with $|z-i| , |z + i| \geq \delta$.
\end{cor}
\begin{proof}
This is immediate from Theorem~\ref{1pt chordal} and a coordinate change. Note that we use Lemma~\ref{G implies U} to obtain a $d > 0$, depending on $\mu$, such that~\eqref{disk asymp'} holds.
\end{proof}
\subsection{Statement of the estimates}
\label{time infty setup sec}
In this section we transfer the estimates of Theorem~\ref{1pt chordal} to estimates for certain ``time infinity" forward Loewner maps, which we will define shortly. We work in the setting of $\mathbf D$, rather than $\mathbf H$, as this setting will be more convenient for our two-point estimates. We start by defining our events.
Let $x,y\in\partial\mathbf D$ be distinct and let $m$ be the midpoint of the counterclockwise arc connecting $x$ and $y$ in $\partial\mathbf D$. Suppose we are given a simple curve $\eta$ in $\mathbf D$ connecting $x$ and $y$. Let $D_\eta$ be the connected component of $\mathbf D\setminus \eta$ containing $m$ on its boundary. Let $\Psi_\eta : D_\eta \rightarrow \mathbf D$ be the unique conformal map taking $x$ to $-i$, $y$ to $i$, and $m$ to 1. For $s\in\mathbf R$, $u >0$, $\epsilon > 0$, $c>1$, and $z \in \mathbf D$, let $\mathcal E_\epsilon^{s;u}(\eta , z ; c )$ be the event that
\begin{enumerate}
\item $z\in D_\eta$;
\item $ c^{-1} \epsilon^{1-s +u } \leq \operatorname{dist}(z, \partial D_\eta ) \leq c \epsilon^{1-s-u }$;
\item $ c^{-1} \epsilon^{ s + u } \leq |\Psi_\eta'(z)| \leq c \epsilon^{ s -u }$.
\end{enumerate}
For technical reasons it will also be convenient to consider the counterclockwise arc of $\partial\mathbf D$ from $y$ to $x$. We denote by $m^-$ the midpoint of this arc. Let $D_\eta^-$ be the connected component of $\mathbf D\setminus \eta$ containing $m^-$ on its boundary and we let $\Psi_\eta^- : D_\eta^- \rightarrow \mathbf D$ be the unique conformal map taking $x$ to $i$, taking $y$ to $-i$, and taking $m^-$ to $-1$. See Figure~\ref{domains fig} for an illustration.
Let $\mathcal A_\epsilon^{s;u}(\eta, c)$ be the set of $z \in \mathbf D$ for which $\mathcal E_\epsilon^{s;u}(\eta , z ; c )$ occurs.
\begin{figure}\label{domains fig}
\begin{center}
\includegraphics[scale=.7]{figures/drawing-D-domains.pdf}
\caption{An illustration of the domains and maps used in Theorem~\ref{1pt forward}.}
\end{center}
\end{figure}
\begin{thm} \label{1pt forward}
Let $\alpha(s)$ and $\alpha_0(s)$ be as in~\eqref{alpha def}. Let
\begin{equation}\label{gamma def}
\gamma(s) := \alpha(s) - 2s + 1 = \frac{(4 + \kappa)^2 s^2}{8 \kappa (1 + s)} -2s + 1 ,\qquad \gamma_0(s) := 2\alpha_0(s) + 2 = \frac{ 2(4 + \kappa)^2 s (2 + s)}{8 \kappa (1 + s)^2} + 2 .
\end{equation}
Let $\kappa \leq 4$. Suppose $\eta$ is a chordal $\operatorname{SLE}_\kappa$ from $x$ to $y$ in $\mathbf D$ and define the maps and events as above. Also define the events $\mathcal G(\cdot , \mu)$ as in Definition~\ref{G infty def}. For each $d\in (0,1)$, each $\mu\in\mathcal M$, and each $z\in B_d(0)$, we have
\begin{align} \label{E probs upper}
\mathbf P\left( \mathcal E_\epsilon^{s;u}(\eta , z ; c ) \cap \mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu) \right) \preceq \epsilon^{\gamma(s) - \gamma_0(s) u } .
\end{align}
Furthermore, there exists $\mu\in \mathcal M$ depending only on $d$ such that s.t.\ for $z\in B_d(0)$,
\begin{align}\label{E probs lower}
\mathbf P\left( \mathcal E_\epsilon^{s;u}(\eta , z ; c ) \cap \mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu) \right)\succeq \epsilon^{\gamma(s) + \gamma_0(s) u } .
\end{align}
In~\eqref{E probs upper} and~\eqref{E probs lower} the implicit constants are independent of $\epsilon$ and uniform for $z\in B_d(0)$ and for $|x-y|$ bounded below by a positive constant.
\end{thm}
The proof of Theorem~\ref{1pt forward} proceeds as follows. First we use Theorem~\ref{1pt chordal} to prove estimates for the area of certain finite-time analogues of the sets of Theorem~\ref{1pt forward}. This is done in Section~\ref{area sec}. This subsection also contains a result which allows us to extend the estimate for deterministic times to estimates for certain stopping times, which will be needed in the sequel. Then, in Section~\ref{compare sec}, we prove several lemmas comparing finite time and infinite time maps and use these lemmas to obtain estimates for the area of the set $\mathcal A_\epsilon^{s;u}(\eta ;c)$ of points where the events of Theorem~\ref{1pt forward} occur. Finally, we complete the proof of Theorem~\ref{1pt forward} in Section~\ref{time infty proof sec} by proving a lemma which gives that the probabilities of the events of Theorem~\ref{1pt forward} do not depend too strongly on $z$. In Section~\ref{finite time sec} we deduce an analogue of Theorem~\ref{1pt forward} for the curve stopped at a finite time.
\subsection{Area estimates and stopping estimates for finite time maps}
\label{area sec}
In this section we will prove estimates for the expected area of the set of points where finite-time analogues of the events of Theorem~\ref{1pt forward} occur. We will also prove a result which allows us to compare probabilities for events at stopping times whose difference is bounded. Suppose we are in the setting of Theorem~\ref{1pt forward}.
\begin{defn} \label{E def D}
Let $\eta$ be a chordal $\operatorname{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$. Define its centered Loewner maps $(f_t)$ as in Section~\ref{disk sec}. For $t,\epsilon,u,\delta,c>0$, and $s \in (-1,1)$, and $z\in \mathbf D$, let $ E_\epsilon^{s;u}(\eta , z;t,\delta, c)$ be the event that the following is true.
\begin{enumerate}
\item $c^{-1} \epsilon^{ s + u } \leq |f_t'(z)| \leq c \epsilon^{ s -u }$.\label{A cond dist}
\item $ c^{-1} \epsilon^{1-s + u} \leq\operatorname{dist}(z, \eta^t ) \leq c\epsilon^{1-s -u}$.\label{A cond deriv}
\item $|f_t(z) - i|$ and $ |f_t(z) +i| $ are both at least $ \delta$. \label{A cond delta}
\end{enumerate}
Let $A_\epsilon^{s;u}(\eta ; t,\delta,c)$ be the set of $z\in \mathbf D$ for which $E_\epsilon^{s;u}(\eta , z;t,\delta,c)$ occurs.
\end{defn}
\begin{lem} \label{coord change area H}
Suppose we are in the setting of Theorem~\ref{1pt forward} with $x=-i$ and $y=i$. Define the sets $A_\epsilon^{s;u}(\eta ; t,\delta , c)$ as in Definition~\ref{E def D} and the events $\mathcal G(f_t , \mu)$ as in Definition~\ref{G infty def}. For any choice of parameters $t,\delta, c , d , \mu$ and any $d\in (0,1)$,
\begin{align} \label{A areas}
\mathbf E \left[ \operatorname{Area} ( A_\epsilon^{s,u}(\eta; t ,\delta c) \cap B_d(0)) \mathbf 1_{\mathcal G(f_t , \mu) } \right] \preceq \epsilon^{\gamma(s) - \gamma_0(s) u}
\end{align}
where the implicit constants are independent of $\epsilon$. Moreover, if $t$ and $\mu$ are chosen sufficiently large then we can find $d\in (0,1)$, depending on $t$ and $\mu$ such that
\begin{align} \label{A areas'}
\mathbf E \left[ \operatorname{Area} ( A_\epsilon^{s;u}(\eta ; t,\delta , c ) \cap B_d(0)) \mathbf 1_{\mathcal G(f_t , \mu) } \right] \succeq \epsilon^{\gamma(s) + \gamma_0(s) u} .
\end{align}
\end{lem}
\begin{proof}
Let $\overline{ A}_\epsilon^{s;u} = \overline{ A}_\epsilon^{s;u}(\eta ; t ,\delta , c , d ) $ be the set of $z\in \mathbf D$ such that
\begin{enumerate}
\item $c^{-1} \epsilon^{1+u} \leq 1-|z| \leq c \epsilon^{1-u}$;
\item $|z - i| $ and $ |z +i|$ are each at least $\delta$;
\item The event $\overline{E}^{s;u}(z ; t , c , d )$ of Section~\ref{disk sec} occurs.
\end{enumerate}
By~\eqref{disk asymp} in Theorem~\ref{1pt chordal disk}, if the first two conditions in the definition of $\overline{ A}_\epsilon^{s;u} $ hold for some $z\in \mathbf D$, then
\begin{equation*}
\mathbf P\left( \overline{E}^{s;u}(z ; t , c , d ) \cap \mathcal G(f_t ,\mu) \right) \preceq \epsilon^{ \alpha(s) - \alpha_0(s) u } .
\end{equation*}
By integrating this over all such $z$, we get
\begin{equation} \label{ul A area}
\mathbf E\left[ \operatorname{Area}(\overline{ A}_\epsilon^{s;u} ) \mathbf 1_{G(f_t ,\mu)} \right] \preceq \epsilon^{ \alpha(s) + 1 - (\alpha_0(s)+1) u } .
\end{equation}
Similarly, suppose $t$, $d$, and $\mu$ are chosen so that~\eqref{disk asymp'} in Theorem~\ref{1pt chordal} holds. Then we have
\begin{equation} \label{ul A area'}
\mathbf E\left[ \operatorname{Area}( \overline{ A}_\epsilon^{s;u} ) \mathbf 1_{G(f_t ,\mu)} \right] \succeq \epsilon^{ \alpha(s) + 1 + (\alpha_0(s)+1) u } .
\end{equation}
By the change of variables formula we have
\begin{equation}\label{area integral}
\operatorname{Area}( A_\epsilon^{s;u}(\eta ; t ,\delta , c ) \cap B_d(0) ) = \int_{ f_t (A_\epsilon^{s;u}(\eta ; t ,\delta , c ) \cap B_d(0) )} |(f_t ^{-1})'(z)|^2 \, dz .
\end{equation}
The Koebe quarter theorem implies
\[
\overline{ A}_\epsilon^{s;u/2}(\eta ; t ,\delta , c' ,d ) \subset f_t \left(A_\epsilon^{s;u}(\eta ; t ,\delta , c ) \cap B_d(0)\right) \subset \overline{ A}_\epsilon^{s;2u}(\eta ; t ,\delta , c '' , d )
\]
for appropriate $c', c'' > 0$, depending only on $c$.
Thus~\eqref{ul A area} implies~\eqref{A areas}. Similarly~\eqref{ul A area'} implies~\eqref{A areas'}.
\end{proof}
In the remainder of this subsection we prove a result which allows us to transfer estimates between stopping times and deterministic times. We first need the following lemma.
\begin{lem} \label{chordal event uniform}
Let $ \eta$ be a chordal $\operatorname{SLE}_\kappa$ process from $-i$ to $i$ in $\mathbf D$. Let $( f_t)$ be its centered Loewner maps. For $w\in \mathbf D$, $\delta' >0$, $C>1$, and $\mu\in\mathcal M$ let $H(w; t) = H(w;t,\delta' , C, \mu)$ be the event that the following is true.
\begin{enumerate}
\item $ \operatorname{dist}(w , \eta^t ) \geq C^{-1} $. \label{H dist}
\item $ C^{-1} \leq | f_t'(w )| \leq C $. \label{H deriv}
\item $|f_t(w) - i | $ and $ |f_t(w) + i| $ are each at least $ \delta'$. \label{arg event}
\item The event $\mathcal G(f_t , \mu) $ occurs. \label{G event}
\end{enumerate}
For $ \zeta >0$, let $S_{\delta}^\zeta$ be the set of $w\in \mathbf D$ with $|w - i | , |w + i| \geq \delta$ and $1-|w| \leq \zeta$.
If $T>0$ and $\delta > 0$, then there exists $p, \zeta ,\delta'>0$, $\mu\in\mathcal M$, and $ C > 1$ depending on $T$ and $\delta $ such that
\begin{equation} \label{hat E all t}
\mathbf P\left( H(w; t) \quad \forall t\leq T, \quad \forall w\in S_{\delta}^\zeta \right) \geq p .
\end{equation}
\end{lem}
\begin{proof}
Fix $\delta' > 0$, to be determined later, and let
\[
U := \{z \in \mathbf D : |\operatorname{Re} z| \leq \delta'/2 \} .
\]
Let $E_0 = E_0(T)$ be the event that $\eta^T\subset U $.
By \cite[Lemma~2.3]{miller-wu-dim}, $\mathbf P(E_0) >0$. Thus, if $\zeta$ is chosen sufficiently small and $C$ is chosen sufficiently large, depending on $\delta$ and $\delta'$, then condition~\ref{H dist} holds a.s.\ on $E_0 $ for each $w\in S_{\delta'}^\zeta \supset S_\delta^\zeta$.
On $E_0 $, we have by Schwarz reflection that $f_t$ a.s.\ extends to be conformal on a neighborhood of $S_\delta^\zeta$ for each sufficiently small $\zeta$ and each $t\leq T$. It follows that for any such $\zeta$ we can find a (possibly larger) constant $C$ depending only on $\delta$, $T$, and $\zeta$ such that the probability of the event $E_1 $ that $E_0$ occurs and condition~\ref{H deriv} holds for each $w\in S_\delta^\zeta$ and each $t \in [0,T]$ is at least $\mathbf P(E_0)/2$.
By continuity, if $\delta'$ is chosen sufficiently small, depending only on $\delta$ and $T$, then the conditional probability given $E_1$ of the event
\begin{equation} \label{arg uniform event}
\left\{ f_t(w)\in S_{\delta'}^\zeta , \quad \forall t\leq T ,\qquad \forall w\in S_\delta^\zeta \right\}
\end{equation}
tends to 1 as $\zeta \rightarrow 0$. In particular, we can find $\zeta > 0$ sufficiently small that the probability of the event $E_2$ that~\eqref{arg uniform event} occurs and that $E_1$ occurs is at least $\mathbf P(E_1)/2$.
Since $\eta$ a.s.\ does not hit $\partial\mathbf D$ and $f_t^{-1}$ is a.s.\ continuous, we can find $\mu\in\mathcal M$ such that $\mathbf P(\mathcal G(f_t, \mu) | E_2) \geq 1/2$. If $E_2 \cap \mathcal G(f_t, \mu)$ occurs, then the event in~\eqref{hat E all t} occurs.
\end{proof}
\begin{lem}\label{stopping estimate}
Let $\eta$ be a chordal $\operatorname{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$ with centered Loewner maps $(f_t)$. Let $ \tau , \tau'$ be stopping times for $\eta$ and suppose there is a deterministic time $T > 0$ such that a.s.\ $ \tau \leq \tau' \leq T$. For any $c > 1$, $\mu\in\mathcal M$, and $\delta > 0$, we can find $c' > 0$, $\delta' > 0$, and $\mu'\in\mathcal M$ depending only on $T$ and $\delta$ such that for each $u>0$, $s\in (-1,1)$, and $z\in \mathbf D$,
\begin{equation} \label{stopping compare eqn}
\mathbf P\left( E_\epsilon^{s;u}(\eta , z;\tau , \delta , c) \cap \mathcal G(f_\tau ,\mu) \right) \preceq \mathbf P\left( E_\epsilon^{s;u}(\eta , z;\tau' , \delta' , c' ) \cap G(f_{\tau'} ,\mu') \right) ,
\end{equation}
with the implicit constant uniform for $z$ in compact subsets of $\mathbf D$ and independent of $\epsilon$.
\end{lem}
\begin{proof}
Let $(\mathcal F_t)$ be the filtration generated by $\eta$. Let $\widehat f_t = f_{t + \tau} \circ f_\tau^{-1}$ and $\widehat\eta = f_\tau(\eta|_{[\tau,\infty)})$. Fix $T$ and $\delta$ and let $p, \zeta , \delta' , C ,$ and $\mu' $ satisfy the conclusion of Lemma~\ref{chordal event uniform} (with $\mu'$ in place of $\mu$). Define the events $H(f_\tau(z) ; t, \delta' , C )$ as in Lemma~\ref{chordal event uniform} with $w = f_\tau(z)$ and $f_t$ replaced by $\widehat f_t$. That lemma together with the Markov property of SLE imply that if $\epsilon$ is chosen sufficiently small, then the conditional probability of the event
\begin{equation*}
\widehat H := \left\{ H(f_\tau(z) ; t ,\delta' , C , \mu' ) \quad \forall t\leq T \right\}
\end{equation*}
given $ E_\epsilon^{s;u}(\eta , z;\tau , \delta , c) \cap G(f_\tau ,\mu) $ is at least $p $. By inspection, for sufficiently small $\epsilon$ and any $t\leq T$, we have
\[
E_\epsilon^{s;u}(\eta , z;\tau , \delta , c) \cap \mathcal G(f_\tau ,\mu) \cap H(f_\tau(z) ; t ,\delta' , C , \mu' ) \subset E_\epsilon^{s;u}(\eta , z;\tau + t , \delta' , c') \cap G(f_{t+\tau},\mu' \circ\mu)
\]
for some $c' >0$ depending only on $C$ and $c$. Since $\tau'-\tau \leq T$ a.s., it follows that
\begin{equation*}
E_\epsilon^{s;u}(\eta , z;\tau , \delta , c) \cap \mathcal G(f_\tau ,\mu) \cap \widehat H \subset E_\epsilon^{s;u}(\eta , z;\tau' , \delta' , c') \cap G(f_{\tau'},\mu' \circ\mu)
\end{equation*}
for some $c' > 0$. Taking probabilities proves~\eqref{stopping compare eqn} (with $\mu'\circ\mu$ in place of $\mu'$).
\end{proof}
\subsection{Comparison lemmas}
\label{compare sec}
In this subsection we prove several lemmas comparing probabilities of sets associated with the finite time Loewner maps to probabilities of sets associated with the infinite time Loewner maps of Theorem~\ref{1pt forward}, and use these results to estimate the areas of the sets $\mathcal A_\epsilon^{s;u}(\eta ; c)$ of Theorem~\ref{1pt forward}.
Our first lemma is needed for the proof of the upper bound in Theorem~\ref{1pt forward}.
\begin{lem} \label{infty compare upper'}
Suppose we are in the setting of Theorem~\ref{1pt forward} with $x=-i$ and $y= i$. For each $d\in (0,1)$ and each $\mu\in\mathcal M$, there exists $\mu'\in \mathcal M$ and a bounded stopping time $\tau$ for $\eta$, both independent of $u$ and of $z \in B_d(0)$, such that for each $z\in B_d(0)$,
\begin{align} \label{infty compare upper' eqn}
\mathbf P\left( \mathcal E^{s;u}(\eta,z;c) \cap \mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu) \right) \preceq \mathbf P\left( E_\epsilon^{s;u}(z; \tau , \delta , c ' ) \cap\mathcal G(f_\tau , \mu ') \right)
\end{align}
with the implicit constants independent of $\epsilon$ and uniform for $z \in B_d(0)$.
\end{lem}
\begin{proof}
Suppose $ \mathcal E^{s;u}(\eta,z;c) \cap \mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu)$ occurs.
We will prove the lemma by growing some more of the curve out from $-i$ and $i$ to get a new curve $\widetilde\eta \overset{d}{=} \eta$ with the property that $E_\epsilon^{s;u}(\widetilde\eta , z; \tau , \delta , c ' ) \cap\mathcal G(f_\tau , \mu ')$ occurs for an appropriate bounded stopping time $\tau$ and the derivatives of the conformal maps associated with $\widetilde\eta^\tau$ and with $\eta$ at $z$ are comparable.
To this end, let $ \eta_0$ be a chordal $\operatorname{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$, independent of $\eta$. Let $\overline\eta_0$ be its time reversal. Then $\overline\eta_0$ has the law of a chordal $\operatorname{SLE}_\kappa$ from $i$ to $-i$~\cite{zhan-reversibility}. Let $ \delta_0, C , \beta , \zeta , r , a >0$, and $\mu_0 \in\mathcal M$.
Let $P $ be the event that the following is true.
\begin{enumerate}
\item Let $\overline T$ be the first time $\overline\eta_0$ gets within distance $e^{-\beta}$ of $z$. Then $\overline T < \infty$. \label{P hit}
\item For each $t\geq 0$, let $\phi_t : \mathbf D\setminus (\eta_0^t \cup \overline\eta_0^{\overline T})$ be the unique conformal map fixing $z$ and taking $\overline\eta_0(\overline T)$ to $i$. Let $T$ be the first time $t$ that $\phi_t(\eta_0(t)) = -i$ and $|\eta_0(t) - z| \leq 2e^{-\beta}$. Then $T < \infty$ and $\eta^T$ is disjoint from $B_{1/2}(i)$. \label{P forward}
\item Henceforth put $\phi = \phi_T$. We have $C^{-1} \leq |(\phi^{-1})'(w)| \leq C$ for each $w\in B_{(1+d)/2}(0)$.\label{P deriv}
\item We have $\phi^{-1}\left(B_{\delta_0}(-i) \cup B_{\delta_0}(i) \cup B_{1-r}(0)\right)\subset B_{(1-d)/2}(z)$. \label{P ball}
\item Let $\overline\sigma$ be the last exit time of $\overline\eta_0$ from $B_\zeta(i)$ before time $\overline T$. Then $\overline\eta_0^{\overline\sigma} \subset B_{2\zeta}(i)$. \label{P stay}
\item Let
\begin{equation} \label{K set def}
K:= \eta_0^T \cup \overline\eta_0([\overline\sigma,\overline T]) \cup B_{(1-d)/2}(z) .
\end{equation}
The harmonic measure from $i$ of each side of $K \cap B_{(1-d)/2}(i)$ and each side of $K \cap B_{(1-d)/2}(-i)$ in the Schwarz reflection of $\mathbf D\setminus K$ across $[-1,1]_{\partial\mathbf D}$ is at least $a$. \label{P hm}
\item $\mathcal G'(K , \mu_0 )$ occurs (Definition~\ref{G' def}). \label{P G}
\end{enumerate}
See Figure~\ref{infty pushing fig} for an illustration of the event $P$. In what follows, all implicit constants are required to depend only on $\mu$, $d$, and the parameters for $P$.
First we will argue that $\mathbf P(P) \succeq 1$ for a suitable choice of parameters. It follows from \cite[Lemma~2.3]{miller-wu-dim} and reversibility of SLE that for any given choice of $r$ and $\delta_0$, conditions~\ref{P hit},~\ref{P forward}, and~\ref{P stay} hold with positive probability depending only on $\beta$, $\zeta$, and $d$. By the Koebe growth theorem, if $\beta$ is chosen sufficiently large (depending on $r$ and $d$) and $\delta$ is chosen sufficiently small (depending only on $d$) then condition~\ref{P ball} also holds simultaneously with positive probability depending only on $\beta ,\zeta , d$, $\delta$, and $r$. By choosing $a$ and $C$ sufficiently large and $\mu_0$ sufficiently small (see Lemma~\ref{G pos}), depending only on $d$ and the other parameters for $P$, we can arrange that the remaining conditions in the definition of $P$ hold with probability arbitrarily close to 1. Thus we have $\mathbf P(P)\succeq 1$.
\begin{figure}
\begin{center}
\includegraphics[scale=.7]{figures/drawing-infty-pushing2.pdf}
\caption{An illustration of the event $P$ and the curve $\widetilde\eta$ used in the proof of Lemma~\ref{infty compare upper'}. The red points are $-i$, $i$, and 1 and their images under $\phi$. } \label{infty pushing fig}
\end{center}
\end{figure}
Let $\widetilde\eta = \eta_0$ on the event that $P$ does not occur. On $P$, let $\widetilde \eta = \phi^{-1}(\eta) \cup \eta_0^T \cup \overline\eta^{\overline T}$, parametrized in such a way that its image under the conformal map from $\mathbf D$ to $\mathbf H$ taking $-i$ to $0$, $i$ to $\infty$, and $0$ to $i$ is parametrized by capacity. By the Markov property and reversibility of SLE, $\widetilde\eta$ has the same law as $\eta$. Let $(\widetilde f_t)$ be the centered Loewner maps for $\widetilde\eta$. Let
\[
\widetilde E = \mathcal E^{s;u}_\epsilon(\eta,z;c) \cap \mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu) \cap P .
\]
Let $\tau$ be the hitting time of $B_{ \zeta}(i)$ by $\widetilde\eta$. Then $\tau$ is a bounded stopping time for $\widetilde\eta$.
We claim that if the parameters for $P$ are chosen appropriately, then we have
\begin{equation} \label{E G P contain}
\widetilde E \subset E_\epsilon^{s;u}(\widetilde\eta , z; \tau , \delta , \widetilde c ) \cap\mathcal G(\widetilde f_\tau , \widetilde \mu)
\end{equation}
for some $\widetilde c > 0$ and $\widetilde \mu \in \mathcal M$, depending only on $\mu$, $d$, and the parameters for $P$. Given the claim~\eqref{E G P contain}, our desired result~\eqref{infty compare upper' eqn} follows by taking probabilities and noting that $P$ is independent from $\eta$.
By condition~\ref{P ball} in the definition of $P$, on the event $\widetilde E$ we have $\widetilde\eta^\tau \subset K$, as in~\eqref{K set def}, provided $r$ is chosen sufficiently small, depending only on $\mu$ and $\delta_0$. By condition~\ref{P G} in the definition of $P$ and Lemma~\ref{G dist}, we can find $\widetilde\mu\in\mathcal M$ depending only on $\mu$, $d$ and the parameters for $P$ such that $\widetilde E\subset \mathcal G(\widetilde f_\tau , \widetilde \mu)$. By condition~\ref{P hm} in the definition of $P$, we can find $\delta > 0$ depending only on $a$ such that $\widetilde f_\tau(z) / |\widetilde f_\tau(z)| $ lies at distance at least $\delta $ from $\pm i$ on $\widetilde E$. That is, condition~\ref{A cond delta} in the definition of $E_\epsilon^{s;u}(\widetilde\eta , z; \tau , \delta ,\widetilde c )$ holds on $\widetilde E$.
By condition~\ref{P deriv} in the definition of $P$, we have $\operatorname{dist}(z , \widetilde \eta) \asymp \operatorname{dist}(z , \eta)$ on $P$. It therefore follows that condition~\ref{A cond dist} in the definition of $E_\epsilon^{s;u}(\widetilde\eta , z; \tau , \delta , \widetilde c )$ holds on $\widetilde E$ for some $\widetilde c \asymp 1$.
It remains to show that condition~\ref{A cond dist} in the definition of $E_\epsilon^{s;u}(\widetilde\eta , z; \tau , \delta , \widetilde c )$ holds on $\widetilde E$ provided $\widetilde c \asymp 1$ is chosen sufficiently large. It is enough to show $|\widetilde f_\tau'(z)| \asymp |\Psi_{ \eta}'(z)|$ on $\widetilde E$. We will do this in two stages. Let $\Psi_{\widetilde\eta }$ be as in Section~\ref{time infty setup sec} with $\widetilde\eta$ in place of $\eta$. First we will show that $|\Psi_\eta'(z)|\asymp |\Psi_{\widetilde\eta}'(z)|$, and then we will show that $|\Psi_{\widetilde\eta}'(z)| \asymp |\widetilde f_\tau'(z)|$.
For the first stage, let $g $ be the conformal automorphism of $\mathbf D$ taking $\Psi_\eta(\phi(-i^+))$ to $-i$, $\Psi_\eta(\phi( i^-))$ to $i$, and $\Psi_\eta(\phi(1))$ to 1. Then we have
\begin{equation} \label{Psi_hat eta decomp}
\Psi_{\widetilde\eta} = g \circ \Psi_{ \eta} \circ \phi .
\end{equation}
By condition~\ref{P G} in the definition of $P$, together with the definition of $\widetilde E$, we have
$|g '|\asymp 1$ uniformly on $\mathbf D$ on $\widetilde E$, so by condition~\eqref{P deriv} in the definition of $P$, we have $|\Psi_{\widetilde \eta}'(z)| \asymp |\Psi_\eta'(z)|$ on $\widetilde E$.
For the second stage, let $ \Psi_{\widetilde\eta^\tau}$ be the conformal map from $\mathbf D\setminus \widetilde \eta^\tau$ to $\mathbf D$ taking $-i^+$ to $-i$ and fixing $i$ and 1. Then $\Psi_{\widetilde\eta^\tau}$ differs from $\widetilde f_\tau$ by a conformal automorphism of $\mathbf D$ taking $\widetilde f_\tau(-i^+)$ to $-i$ and $\widetilde f_\tau(1)$ to 1. Since $\mathcal G(\widetilde f_\tau , \widetilde \mu)$ holds on $\widetilde E$, we have
\begin{equation} \label{f asymp Psi}
|\Psi_{\widetilde\eta^\tau} '(z)| \asymp |\widetilde f_\tau'(z)| .
\end{equation}
Let $I$ be the arc of $\partial\mathbf D$ of length $\zeta$ centered at 1. By condition~\ref{P G} in the definition of $P$ (c.f. Remark~\ref{phi hm hypotheses}), the lengths of $\Psi_{\widetilde\eta}(I)$ and $\Psi_{\widetilde\eta^\tau}(I)$ are $\succeq 1$ on $\widetilde E$. By conditions~\ref{P ball} and~\ref{P stay} in the definition of $P$ and a study of the harmonic measure from $1$ in the Schwarz reflection of $D_{\widetilde\eta}$, the distances from $\Psi_{\widetilde\eta}(z)$ to $\Psi_{\widetilde\eta}(I)$ and from $\Psi_{\widetilde\eta^\tau}(z)$ to $\Psi_{\widetilde\eta^\tau}(I)$ are $\succeq 1$ on $\widetilde E$ provided $\zeta$ is chosen sufficiently small relative to $d$. By Lemma~\ref{phi hm}, we therefore have
\begin{equation} \label{Psi_wt eta hm}
|\Psi_{\widetilde\eta}'(z)| \asymp \frac{\operatorname{hm}^z(I ; D_{\widetilde\eta} )}{ \operatorname{dist}(z , \widetilde\eta) } ,\qquad |\Psi_{\widetilde\eta^\tau}'(z)| \asymp \frac{\operatorname{hm}^z(I ;\mathbf D\setminus \widetilde\eta^\tau )}{ \operatorname{dist}(z , \widetilde\eta^\tau) }
\end{equation}
on $\widetilde E$.
By conformal invariance $\operatorname{hm}^z(I ; D_{\widetilde\eta} )$ is the same as the probability that a Brownian motion started from $\Psi_{\widetilde\eta^\tau}(z)$ exits $\mathbf D$ in $\Psi_{\widetilde\eta^\tau}(I)$ before hitting $\Psi_{\widetilde\eta^\tau}(\widetilde\eta([\tau,\infty))$. By conditions~\ref{P stay} and~\ref{P G} in the definition of $P$, if $\zeta$ is chosen sufficiently small, independently of $\epsilon$, then the diameter of $\Psi_{\widetilde\eta^\tau}(\widetilde\eta([\tau,\infty))$ is at most a constant less than 1 times its distance from $\Psi_{\widetilde\eta^\tau}(I)$ on $\widetilde E$ (here we again use harmonic measure from 1). Therefore, the probability that a Brownian motion started from $\Psi_{\widetilde\eta^\tau}(z)$ exits $\mathbf D$ in $\Psi_{\widetilde\eta^\tau}(I)$ before hitting $\Psi_{\widetilde\eta^\tau}(\widetilde\eta([\tau,\infty))$ is proportional to the probability that a Brownian motion started from $\Psi_{\widetilde\eta^\tau}(z)$ exits $\mathbf D$ in $\Psi_{\widetilde\eta^\tau}(I)$. That is, $\operatorname{hm}^z(I ; D_{\widetilde\eta} ) \asymp \operatorname{hm}^z(I ;\mathbf D\setminus \widetilde\eta^\tau )$ on $\widetilde E$. By combining this with~\eqref{f asymp Psi} and~\eqref{Psi_wt eta hm}, we conclude.
\end{proof}
The next lemma is needed for the proof of the lower bound in Theorem~\ref{1pt forward}.
\begin{lem}\label{infty compare lower}
Suppose we are in the setting of Theorem~\ref{1pt forward} with $x=-i$ and $y= i$. Let $d\in (0,1)$. For any $ \delta > 0 $ and $\mu\in\mathcal M$, there exists $\mu'\in\mathcal M$ such that for any $c > 0$, there exists $c' >0$, such that for $z\in B_d(0)$ and sufficiently small $\epsilon > 0$,
\begin{align} \label{infty compare lower eqn}
\mathbf P\left( \mathcal E_\epsilon^{s;u}(\eta , z; c' ) \cap \mathcal G(\Psi_\eta , \mu') \cap \mathcal G(\Psi_\eta^-,\mu') \right) \succeq \mathbf P\left( E_\epsilon^{s;u}(\eta , z; t, \delta , c ) \cap \{ \operatorname{Re} f_t(z) \geq 0 \} \cap \mathcal G(f_t,\mu) \right) ,
\end{align}
with implicit constants independent of $\epsilon$ and uniform on $B_d(0)$. By re-choosing $c$, we can make $c'$ as small as we like.
\end{lem}
\begin{proof}
Let $(f_t)$ be the centered Loewner maps for $\eta$ as in Section~\ref{area sec}. For $t\geq 0$, let $\eta_t = f_t(\eta|_{[t,\infty)})$. Let $ D_t$ be the connected component of $\mathbf D\setminus \eta_t$ containing 1 on its boundary and let $D_t^-$ be the other connected component of $\mathbf D\setminus \eta_t$. Let $\Psi_t : D_t \rightarrow \mathbf D$ (resp.\ $\Psi_t^- : D_t\rightarrow\mathbf D$) be the unique conformal maps fixing $-i,i,1$ (resp.\ $-i,i,-1$). Let $b_t$ (resp.\ $b_t^-$) be the image of the right (resp.\ left) side of $-i$ under $f_t$. Finally, let $\psi_t $ (resp.\ $\psi_t^- $) be the conformal automorphism of $\mathbf D$ fixing $i$, taking $\Psi_t(b_t)$ to $-i$, and taking $\Psi_t(f_t(1))$ to $1$ (resp.\ fixing $i$, taking $\Psi_t^-(b_t^-)$ to $-i$, and taking $\Psi_t^-(f_t(-1))$ to -1). Then for each $t$,
\begin{equation} \label{Phi decomp'}
\Psi_\eta = \psi_t \circ \Psi_t\circ f_t ,\qquad \Psi_\eta^- = \psi_t^- \circ \Psi_t^- \circ f_t .
\end{equation}
Moreover, $(\Psi_t , \Psi_t^-)$ and $f_t$ are independent and we have $\Psi_t\overset{d}{=} \Psi_\eta$, $\Psi_t^- \overset{d}{=} \Psi_\eta^-$. See Figure~\ref{f-phi-psi} for an illustration of some of these maps.
\begin{figure}
\begin{center}
\includegraphics[scale=.7]{figures/drawing-f-phi-psi.pdf}
\caption{An illustration of the maps used in the proof of Lemma~\ref{infty compare lower} for the right side of $\mathbf D$. The red points are the images of $-i,i$, and $1$ under the various maps. The last map $\psi_t$ takes these points back to their original positions so that by composing all three maps we recover the original map $\Psi_\eta$.} \label{f-phi-psi}
\end{center}
\end{figure}
For $C>1$, $\mu'\in\mathcal M$, and $w\in \mathbf D$, let $ F(w) = F(w;t ,C , \mu' )$ be the event that the following is true.
\begin{enumerate}
\item $w\in D_t$.
\item $C^{-1} \leq |\Psi_t'(w)| \leq C$. \label{C deriv}
\item $ \operatorname{dist}(w , \eta_t) = \operatorname{dist}(w , \partial\mathbf D)$. \label{C dist}
\item $\mathcal G(\Psi_t ,\mu' ) \cap \mathcal G(\Psi_t^- , \mu')$ occurs. \label{Phi_t(r)}
\end{enumerate}
By \cite[Lemma~2.3]{miller-wu-dim}, for each $\delta > 0$, we can find $C>1 $ and $\mu'\in\mathcal M$ such that for each $w\in \mathbf D$ lying at distance at least $\delta $ from $\pm i$ with $\operatorname{Re} w \geq 0$, we have that $\mathbf P(F(w) ) \succeq 1$, with the implicit constant independent of $\epsilon$ and uniform for $w$ satisfying the conditions above.
If we let
\begin{equation*}
F^*(z )
:= E_\epsilon^{s;u}(\eta , z; t,\delta , c) \cap \{\operatorname{Re} f_t(z) \geq 0\} \cap \mathcal G(f_t, \mu ) \cap F(f_t(z) ) ,
\end{equation*}
then by independence of $f_t$ and $\eta_t$ and our choice of parameters for $F(\cdot)$ we have
\begin{equation} \label{P(W)}
\mathbf P\left( F^*(z) \right) \asymp \mathbf P\left( E_\epsilon^{s;u}(\eta , z; t,\delta , c ) \cap \{ \operatorname{Re} f_t(z) \geq 0 \} \cap \mathcal G(f_t,\mu) \right) .
\end{equation}
By condition~\ref{Phi_t(r)} in the definition of $F(f_t(z))$, $|\psi_t'|$ and $|(\psi_t^-)'|$ are bounded above and below by positive $\epsilon$-independent constants on the event $F^*(z )$. Hence it follows from~\eqref{Phi decomp'} that $F^*(z )\subset \mathcal E_\epsilon^{s;u}(\eta , z; c' ) \cap \mathcal G(\Psi_\eta , \mu'') \cap \mathcal G(\Psi_\eta^-,\mu'') $ for some $c'>0$ and some $\mu''\in\mathcal M$ which do not depend on $\epsilon$ and are uniform for $z\in B_d(0)$. By combining this with~\eqref{P(W)} we get~\eqref{infty compare lower eqn} (with $\mu''$ in place of $\mu'$).
\end{proof}
Now we can transfer our area estimates for the finite time sets to area estimates for the time infinity sets.
\begin{lem} \label{infty area}
Suppose we are in the setting of Theorem~\ref{1pt forward} with $x = -i$ and $y=i$. For each $d\in(0,1)$, each $\mu\in\mathcal M$, and each $c>0$,
\begin{align} \label{A areas upper}
\mathbf E\left( \operatorname{Area} (\mathcal A_\epsilon^{s;u}(\eta ; c ) \cap B_d(0) ) \mathbf 1_{\mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu)} \right) \preceq \epsilon^{\gamma(s) - \gamma_0(s) u } .
\end{align}
Furthermore, there exists $d\in (0,1)$ and $\mu\in \mathcal M$ depending only on $d$ such that for $c>0$,
\begin{align} \label{A areas lower}
\mathbf E\left( \operatorname{Area}( \mathcal A_\epsilon^{s;u}(\eta ; c ) \cap B_d(0) ) \mathbf 1_{\mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu)} \right) \succeq \epsilon^{\gamma(s) + \gamma_0(s) u }.
\end{align}
In both~\eqref{A areas upper} and~\eqref{A areas lower} the implicit constants depend on the other parameters but not on $\epsilon$.
\end{lem}
\begin{proof}
The relation~\eqref{A areas upper} follows by integrating the estimate from Lemma~\ref{infty compare upper'} over $B_d(0)$, applying Lemma~\ref{stopping estimate} to replace the stopping time $\tau$ with a deterministic time, then applying~\eqref{A areas} from Lemma~\ref{coord change area H}.
For~\eqref{A areas lower}, choose parameters in such a way that~\eqref{A areas'} from Lemma~\ref{coord change area H}) holds. Given $c>0$, choose $\mu' \in\mathcal M$ and $c'>0$ such that the conclusion of Lemma~\ref{infty compare lower} holds for our choice of $t,\mu,$ and $d$. Integrate the estimate of Lemma~\ref{infty compare lower} over $B_d(0)$ to get
\begin{align} \label{integral compare f_t}
&\mathbf E\left( \operatorname{Area} ( \mathcal A_\epsilon^{s;u}(\eta ; c' ) \cap B_d(0)) \mathbf 1_{\mathcal G(\Psi_\eta , \mu') \cap \mathcal G(\Psi_\eta^- , \mu')} \right) \nonumber\\
&\qquad \succeq \mathbf E\left( \operatorname{Area} \left( \{z \in A_\epsilon^{s;u}(t,\delta, c ) \cap B_d(0) : \operatorname{Re} f_t(z) \geq 0\}\right) \mathbf 1_{\mathcal G(f_t , \mu) } \right) .
\end{align}
Since the law of $(f_t)$ is symmetric about the imaginary axis the right side of~\eqref{integral compare f_t} is at least
\begin{equation*}
\frac12 \mathbf E\left( \operatorname{Area} ( \widehat A_\epsilon^{s;u}(t,\delta, c ) \cap B_d(0) ) \mathbf 1_{\mathcal G(f_t , \mu) } \right) .
\end{equation*}
By combining this with~\eqref{A areas'} from Lemma~\ref{coord change area H} and re-labelling $c$ and $\mu$, we deduce~\eqref{A areas lower}.
\end{proof}
\subsection{Proof of the estimates}
\label{time infty proof sec}
In this subsection we will complete the proof of Theorem~\ref{1pt forward}.
To get Theorem~\ref{1pt forward} from the area estimates of Lemmas~\ref{infty compare upper'} and~\ref{infty compare lower}, we need to argue that the probabilities of the events of Theorem~\ref{1pt forward} do not depend too strongly on $z$. This is accomplished in the next lemma.
\begin{lem}\label{infty log continuity}
Suppose we are in the setting of Theorem~\ref{1pt forward} with $x=-i$, $y=i$. Fix $d\in (0,1)$. For any $\mu\in \mathcal M$ and $c>0$, we can find $\mu'\in\mathcal M$ and $c'>0$ such that for $z,w\in B_d(0)$,
\begin{equation} \label{infty log continuity eqn}
\mathbf P\left( \mathcal E_\epsilon^{s;u}(\eta , w ; c ) \cap \mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu) \right) \preceq \mathbf P\left( \mathcal E_\epsilon^{s;u}(\eta , z ; c' ) \cap \mathcal G(\Psi_\eta , \mu') \cap \mathcal G(\Psi_\eta^- , \mu') \right)
\end{equation}
with implicit constants independent of $\epsilon$ and uniform in $B_d(0)$. By re-choosing $c$, independently of $\epsilon$, we can make $c'$ as small as we like.
\end{lem}
\begin{proof}
The basic idea of the proof is as follows. First we apply a conformal map taking $z$ to $w$ and fixing $-i$. The image of $\eta$ under such a map will be an $\operatorname{SLE}_\kappa$ with a new target point $b$. To compare such a curve to our original curve, we grow a carefully chosen segment of the new curve backward from $b$ in such a way that when we map back to $\mathbf D$, we get a chordal $\operatorname{SLE}_\kappa$ from $-i$ to $i$. We now commence with the details.
For $z,w\in B_d(0)$, let $\phi = \phi_{z,w} : \mathbf D \rightarrow \mathbf D$ be the unique conformal map fixing $-i$ and taking $z$ to $w$. Let $b := \phi(i)$ and $ \eta^b = \phi(\eta)$. The law of $\eta^b $ is that of a chordal $\operatorname{SLE}_\kappa$ process from $-i$ to $b$ in $\mathbf D$.
The map $\phi$ depends continuously on $z$ in the topology of uniform convergence on compacts.
It follows that for any $\mu\in\mathcal M$ we can find a deterministic constant $c' > 0$ depending only on $ c$, $\mu$, and $d$, linearly on $c$, and a deterministic $\mu' \in \mathcal M$ depending only on $\mu$ and $d$ such that for $z,w \in B_d(0)$,
\begin{equation} \label{z w compare}
\mathcal E_\epsilon^{s;u}(\eta^b , w ; c ) \cap \mathcal G(\Psi_{\eta^b} , \mu ) \cap \mathcal G(\Psi_{ \eta^b}^- , \mu) \subset \mathcal E_\epsilon^{s;u}( \eta , z ; c' ) \cap \mathcal G(\Psi_{\eta} , \mu' ) \cap \mathcal G(\Psi_{ \eta^b}^- , \mu') .
\end{equation}
Let $\overline\eta^b $ be the time reversal of $\eta^b $. Then $\overline\eta^b $ is a chordal $\operatorname{SLE}_\kappa$ from $b$ to $-i$ in $\mathbf D$ \cite{zhan-reversibility}. We give $\overline \eta^b$ the usual chordal parametrization, so that it is the conformal image of a chordal $\operatorname{SLE}_\kappa$ parametrized by capacity from 0 to $\infty$ in $\mathbf H$. For each $t\geq 0$, let $\overline g_t : \mathbf D\setminus \overline\eta^b([0,t]) \rightarrow \mathbf D$ be the unique conformal map fixing $-i$ and $w$. Let $\tau $ be the first time $t$ that $\overline g_t (\overline\eta^b(t)) = i$.
Fix $\mu^b \in\mathcal M$ and let $\overline E^b $ be the event that the following occurs.
\begin{enumerate}
\item $\tau $ is less than or equal to the first time $t$ that $\overline\eta^b$ hits $B_{d^*}(0)$, where $d^* = 1 - \frac14 \inf_{z,w\in B_d(0)} \operatorname{dist}( \phi_{z,w}(B_d(0) ), \partial\mathbf D )$. \label{E bar tau}
\item $\mathcal G(\overline g_\tau , \mu^b )$ occurs. \label{E bar dist}
\end{enumerate}
By \cite[Lemma~2.3]{miller-wu-dim}, if $\mu^b$ is chosen sufficiently small then $\mathbf P(\overline E^b)$ is a positive constant depending only on $\mu^b$ and $B_d(0)$.
By the Markov property, conditional on $\overline E^b$, the law of $\overline g_\tau (\overline\eta^b|_{[\tau,\infty)})$ is that of a chordal $\operatorname{SLE}_\kappa$ process from $i$ to $-i$ in $\mathbf D$. Therefore its time reversal $\widehat \eta := \overline g_\tau^b(\eta|_{[0,\tau^b]})$, where $\tau^b$ is the time corresponding to $\tau$ under the time reversal, has the law of a chordal $\operatorname{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$. In particular, $\widehat\eta \overset{d}{=} \eta$.
Define the open sets $D_{\eta^b} , D_{\widehat\eta}$ and the maps $\Psi_{ \eta^b} , \Psi_{\widehat \eta}$ as in Section~\ref{time infty setup sec} with $ \eta^b , \widehat \eta$, resp., in place of $\eta$.
Let $\psi$ (resp.\ $\psi^-$) be the conformal automorphism of $\mathbf D$ which fixes $-i$, takes $(\Psi_{\widehat\eta} \circ \overline g_\tau )(i)$ to $i$, and takes $(\Psi_{\widehat\eta} \circ \overline g_\tau )(1)$ to $1$ (resp.\ fixes $-i$, takes $(\Psi_{\widehat\eta}^- \circ \overline g_\tau )(b^+)$ to $i$, and takes $(\Psi_{\widehat\eta}^- \circ \overline g_\tau )(-1)$ to $-1$). Then we have
\begin{equation*}
\Psi_{\eta^b} = \psi \circ\Phi_{\widehat\eta} \circ \overline g_\tau ,\qquad \Psi_{\eta^b}^- = \psi^- \circ \Psi_{\widehat\eta}^-\circ \overline g_\tau .
\end{equation*}
See Figure~\ref{cont maps fig} for an illustration of some of these maps.
\begin{figure}
\begin{center}
\includegraphics[scale=.7]{figures/drawing-continuity-maps-infty.pdf}
\caption{An illustration of the maps used in the proof of Lemma~\ref{infty log continuity} on the event $\overline E^b$.} \label{cont maps fig}
\end{center}
\end{figure}
By condition~\ref{E bar dist} in the definition of $\overline E^b$, on the event $\overline E^b \cap \mathcal E_\epsilon^{s;u}(\widehat\eta , w ; c ) \cap \mathcal G(\Psi_{\widehat \eta}, \mu ) \cap \mathcal G(\Psi_{\widehat\eta}^- , \mu)$, it holds that $|\psi'|$ and $|(\psi^-)'|$ are bounded above and below by deterministic positive constants depending only on $\mu^b$ and $\mu$. Furthermore, we have that $\mathcal G(\psi , \mu_2) \cap \mathcal G(\psi^- , \mu_2)$ holds for some $\mu_2\in\mathcal M$ depending on $\mu^b , \mu$. The Koebe distortion theorem and condition~\ref{E bar tau} in the definition of $\overline E^b$ imply that $|g_\tau'(w)|$ is bounded above and below by positive constants depending only on $d$ on the event $\overline E^b$.
Hence for some $c' > 0$, independent of $\epsilon$ and uniform for $z,w\in B_d(0)$, we have
\begin{equation} \label{E^b contained'}
\overline E^b \cap \mathcal E_\epsilon^{s;u}(\widehat\eta , w ; c ) \cap \mathcal G (\Psi_{\widehat \eta} , \mu) \cap \mathcal G (\Psi_{\widehat \eta}^- , \mu) \subset \mathcal E_\epsilon^{s;u}( \eta^b , w ; c') \cap \mathcal G(\Psi_{\eta^b} , \mu_2 \circ \mu \circ \mu^b ) \cap \mathcal G(\Psi_{\eta^b}^- , \mu_2 \circ \mu \circ \mu^b ) .
\end{equation}
By the Markov property and the fact that $\mathbf P(\overline E^b)$ is uniformly positive, we have
\begin{equation} \label{E^b Markov'}
\mathbf P\left( \overline E^b \cap \mathcal E_\epsilon^{s;u}(\widehat\eta , w ; c) \cap \mathcal G (\Psi_{\widehat \eta} , \mu) \cap \mathcal G(\Psi_{\widehat\eta}^- , \mu) \right) \asymp \mathbf P\left( \mathcal E_\epsilon^{s;u}(\widehat\eta , w ; c , \lambda,\ell ) \cap \mathcal G (\Psi_{\widehat \eta} , \mu)\cap \mathcal G(\Psi_{\widehat\eta}^- , \mu) \right) .
\end{equation}
Since $\widehat \eta \overset{d}{=} \eta$,~\eqref{infty log continuity eqn} now follows from~\eqref{E^b contained'},~\eqref{E^b Markov'}, and~\eqref{z w compare}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{1pt forward}]
By applying a coordinate change it is enough to consider the case $x=-i,$ $y=i$. By Lemma~\ref{infty log continuity}, for any $z\in B_d(0)$, we have, in the notation of that lemma,
\alb
\mathbf P\left( \mathcal E_\epsilon^{s;u}(\eta , z ; c ) \cap \mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu) \right) \preceq \mathbf E\left( \operatorname{Area} (\mathcal A_\epsilon^{s;u}(\eta , z ; c' ) \cap B_d(0)) \mathbf 1_{\mathcal G(\Psi_\eta , \mu') \cap \mathcal G(\Psi_\eta^- , \mu')} \right) \\
\mathbf P\left( \mathcal E_\epsilon^{s;u}(\eta , z ; c ' ) \cap \mathcal G(\Psi_\eta , \mu') \cap \mathcal G(\Psi_\eta^- , \mu') \right) \succeq \mathbf E\left( \operatorname{Area} (\mathcal A_\epsilon^{s;u}(\eta , z ; c ) \cap B_d(0)) \mathbf 1_{\mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu)} \right) .
\ale
We conclude by combining this with Lemma~\ref{infty area} (and re-labelling $c'$ in the lower bound).
\end{proof}
\subsection{Finite time estimates}
\label{finite time sec}
In this subsection we use Theorem~\ref{1pt forward} and the comparison lemmas of Section~\ref{compare sec} to prove estimates for the finite time Loewner maps. The results in this subsection are not needed for the proof of our main result, and are stated only for the sake of completeness.
\begin{thm} \label{1pt forward finite}
Let $\kappa \in (0,4]$. Let $( f_t)$ be the centered Loewner maps of a chordal $\operatorname{SLE}_\kappa$ process $ \eta$ from $-i$ to $i$ in $\mathbf D$. Define the events $E_\epsilon^{s;u}(z;t,\delta,c)$ as in Definition~\ref{E def D} and the sets $G(f_t ,\mu)$ as in Definition~\ref{G infty def}. For any $\mu\in\mathcal M$ and any $z\in\mathbf D$,
\begin{equation} \label{finite upper}
\mathbf P\left( E_\epsilon^{s;u}(\eta , z;t,\delta ,c) \cap \mathcal G(f_t ,\mu ) \cap \{\operatorname{Re} f_t(z) \geq 0\} \right) \preceq \epsilon^{\gamma(s) -\gamma_0(s) u } .
\end{equation}
Moreover, there exists $t_0 > 0$ depending on $z$ and uniform for $z$ in compacts such that for each $t\geq t_0$ and sufficiently small $\delta >0$, we can find $\mu \in\mathcal M$ such that for each $c>0$,
\begin{equation} \label{finite lower}
\mathbf P\left( E_\epsilon^{s;u}(\eta , z;t,\delta ,c) \cap \mathcal G(f_t ,\mu ) \right) \succeq \epsilon^{\gamma(s) +\gamma_0(s) u } .
\end{equation}
In~\eqref{finite upper} and~\eqref{finite lower} the implicit constants are independent of $\epsilon$ and uniform for $z$ in compacts. The estimate~\eqref{finite upper} holds with $t$ replaced by a bounded stopping time. The estimate~\eqref{finite lower} holds with $t$ replaced by a bounded stopping time which is a.s.\ $\geq t_0$.
\end{thm}
\begin{proof}
The statement for deterministic times follows by combining Theorem~\ref{1pt forward} with Lemmas~\ref{stopping estimate},~\ref{infty compare upper'} and~\ref{infty compare lower}. The statement for stopping times follows from this and Lemma~\ref{stopping estimate}.
\end{proof}
Theorem~\ref{1pt forward finite} implies the following corollary.
\begin{cor}
The statement of Theorem~\ref{1pt forward finite} also holds with $\operatorname{SLE}_\kappa$ replaced by $\operatorname{SLE}_\kappa(\underline\rho)$ with finitely many boundary force points located at positive distance from $-i$ provided $\underline\rho$ is chosen so that an $\operatorname{SLE}_\kappa(\underline\rho)$ process does not hit $\partial\mathbf D$.
\end{cor}
\begin{proof}
Lemma~\ref{G implies U} shows that for any $\delta > 0$, we can find a deterministic bounded simply connected open set $U\subset \mathbf D$ with $\partial U \cap \mathbf D \subset \partial\mathbf D\setminus B_\delta(-i)$ such that $\eta^t\subset U$ on the event $G(f_t,\mu)$. If $\delta $ is chosen to be smaller than half of the distance from any of the force points to the $-i$, then the form of the Radon-Nikodym derivatives in \cite[Section 5]{sw-coord} implies that the laws of a chordal $\operatorname{SLE}_\kappa$ and a chordal $\operatorname{SLE}_\kappa(\underline\rho)$ are stopped at their first exit time from $U$ are \hyperref[smac]{s.m.a.c.}\ (Definition~\ref{smac}) with constants depending only on $\underline\rho,\kappa,U$, and the distance of the force points to $U$. Hence the statement of the corollary follows from Theorem~\ref{1pt forward finite}.
\end{proof}
\subsection{Upper bound for the Hausdorff dimension of the subset of the circle}
In this subsection we use Proposition~\ref{1pt chordal} to obtain upper bounds on the Hausdorff dimension of the sets $\widetilde \Theta^s(\mathbf D \setminus K_t)$ of Section~\ref{multifractal def} for the hulls $(K_t)$ of a chordal $\operatorname{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$. In light of Lemma~\ref{theta zero one}, Proposition~\ref{bdy haus upper} implies the upper bounds for $\dim_{\mathcal H} \widetilde \Theta^{s;\geq}(D_\eta)$ and $\dim_{\mathcal H} \widetilde \Theta^{s;\leq}(D_\eta)$ in Theorem~\ref{main thm}.
\begin{prop} \label{bdy haus upper}
Let $\eta$ be a chordal $\operatorname{SLE}_\kappa$ process from $-i$ to $i$ in $\mathbf D$ with centered Loewner maps $(f_t)$ (defined as in Section~\ref{disk sec}) and hulls $(K_t)$.
Let $\widetilde \xi(s)$, $s_-$, and $s_+$ be as in \eqref{tilde xi(s)}. For each $t > 0$ and $s\in [-1,1]$, a.s.\
\begin{align}
\label{bdy haus upper eqn}
& \dim_{\mathcal H} \widetilde\Theta^{s; \geq }(\mathbf D\setminus K_t) \leq \widetilde\xi(s) ,\qquad 0\leq s \leq s_+ \nonumber \\
& \dim_{\mathcal H} \widetilde\Theta^{s; \leq }(\mathbf D\setminus K_t) \leq \widetilde\xi(s) ,\qquad s_-\leq s \leq 0 .
\end{align}
Almost surely, for each $s\notin [s_- , s_+]$ we have $\widetilde\Theta^{s }(\mathbf D\setminus K_t) = \emptyset$.
\end{prop}
\begin{remark}
If $\alpha(s)$ is as in \eqref{alpha def} in the statement of Theorem~\ref{1pt chordal}, we have $\widetilde{\xi}(s) = 1-\alpha(s)$.
\end{remark}
\begin{proof}[Proof of Proposition~\ref{bdy haus upper}]
For $\delta>0$ and $s\in (-1,1)$, let
\begin{equation*}
\widetilde\Theta^{ s;*}_\delta(\mathbf D\setminus K_t) := \widetilde\Theta^{ s;*}(\mathbf D\setminus K_t) \cap \left\{x\in \partial \mathbf D : |x-i| , |x+i| \geq \delta ,\quad 1- |f_t^{-1}(x)| \geq \delta \right\} ,
\end{equation*}
where $*$ stands for $\geq$ in the case $s \geq 0$ or $\leq$ in the case $s<0$.
The reason for this definition is that it will allow us to apply the estimates of Proposition~\ref{upper bound infty} after a change of coordinates from $\mathbf D$ to $\mathbf H$.
By countable stability of Hausdorff dimension, to prove \eqref{bdy haus upper eqn}, it is enough to show that a.s.\
\begin{equation*} \label{dim lower show}
\mathcal H^\beta(\widetilde\Theta^{ s;*}_\delta(\mathbf D\setminus K_t) ) = 0 \quad \forall \:\delta >0, \quad \forall \: \beta > \widetilde\xi(s) .
\end{equation*}
Henceforth fix $\delta$, $\beta$, and $s$ as above. Also let $s' \in [0,s)$ (if $s \geq 0$) or $s' \in (s , 0)$ (if $s < 0$) be chosen in such a way that $\widetilde\xi(s') < \beta$.
For $n\in \mathbf N$ and $k \in \{1,\dots,n\}$, let
\begin{equation} \label{B_n^k}
B_n^k : = \left\{w \in \mathbf D \,:\, \frac{ \pi (k-1) }{2^{n-1} } \leq \operatorname{arg} w \leq \frac{ \pi k}{2^{n-1}} , \quad 2^{-n } \leq 1-|w| \leq 2^{-n + 1} \right\} .
\end{equation}
Let $E_n^k$ be the event there is a $w \in B_n^k$ with $1-|f_t^{-1}(w)| \geq \delta/2$ and
\begin{equation} \label{deriv at w}
\begin{dcases}
&|(f_t^{-1})'(w)| \geq 2^{n s'} ,\quad s\geq 0\\
&|(f_t^{-1})'(w)| \leq 2^{n s'} ,\quad s <0 .
\end{dcases}
\end{equation}
Each $B_n^k$ can be covered by at most an $(n,k)$-independent constant number of balls of radius $<2^{-n-1}$, and each point of $B_n^k$ lies at distance at least $2^{-n}$ from $\partial\mathbf D$. So the Koebe distortion and growth theorems imply that for sufficiently large $n$, on the event $E_n^k$, we have that for the center $z$ of one of these balls, $|(f_t^{-1})'(z)|$ is at least (if $s \geq 0$) or at most (if $s < 0$) an $(n,k)$-independent constant times $2^{n s'}$; $|f_t^{-1}(z)| \leq \delta$; and $1-| f_t^{-1}(z)| \geq \delta$.
For $n\in\mathbf N$, let $\mathcal K_n$ be the set of those $k\in \{1,\dots,n\}$ such that $\exp(i \pi k/2^{n-1}) $ lies at distance at least $\delta/2$ from $-i$ and $i$. By Proposition~\ref{upper bound infty} and a change of coordinates to $\mathbf H$, whenever $k\in\mathcal K_n$, we have
\begin{equation} \label{P(E_k^n)}
\mathbf P(E_n^k ) \preceq 2^{-n( 1-\widetilde\xi(s') )}
\end{equation}
where the implicit constant is independent of $n$ and uniform for $k\in\mathcal K_n$.
For $n\in \mathbf N$ and $k \in \{1,\dots,n\}$, let
\begin{equation*}
I_n^k := \left\{x\in\partial \mathbf D \,:\, \frac{\pi (k-1) }{2^{n-1} } \leq \operatorname{arg} x \leq \frac{ \pi k}{2^{n-1}} \right\} .
\end{equation*}
For $m\in \mathbf N$, let $\mathcal I_m$ be the collection of those intervals $I_n^k$ for pairs $(n,k)$ such that $n\geq m$, $k\in\mathcal K_n$, and $E_n^k$ occurs. We claim that for each $m\in\mathbf N$, $\mathcal I_m$ is a cover of $\widetilde\Theta^{s;*}_\delta(\mathbf D\setminus K_t)$. Indeed, if $x \in \widetilde\Theta^{s;*}_\delta(\mathbf D\setminus K_t)$, then for any $m\in \mathbf N$ we can find $n\geq m$ and $w\in \mathbf D$ with $1-|w| \leq 2^{-n}$, $\operatorname{arg} w = \operatorname{arg} x $, $|(f_t^{-1})'(w)| \geq (1-|w|)^{- s'}$ (resp.\ $|(f_t^{-1})'(w)| \leq (1-|w|)^{- s'}$ if $s < 0$), and $1-|f_t^{-1}(w)| \geq \delta/2$. The point $w$ lies in $B_n^k$ for some pair $(n,k)$ with $I_{n,k} \in \mathcal I_m$. Since $\operatorname{arg} w = \operatorname{arg} x $, we have $x \in I_{n,k}$ for this choice of $(n,k)$.
Now, observe that \eqref{P(E_k^n)} implies
\begin{align}\label{haus bdy sum}
\mathbf E\left( \sum_{I\in\mathcal I_m} (\operatorname{diam} I)^\beta \right) &\asymp \sum_{n=m}^\infty \sum_{k\in\mathcal K_n} 2^{-n \beta } \mathbf P(E_n^k) \nonumber \\
&\preceq \sum_{n=m}^\infty 2^{-n( \beta - \widetilde\xi(s') )} .
\end{align}
This tends to 0 as $m\rightarrow \infty$ since $\beta > \widetilde\xi(s')$ (by our choice of parameters above). Since $\mathcal I_m$ is a covering of $\widetilde\Theta^{s;*}_\delta(\mathbf D\setminus K_t)$ by intervals of diameter tending to zero as $m\rightarrow \infty$, this proves $\mathcal H^\beta(\widetilde\Theta_\delta^{ s;*}(\mathbf D\setminus K_t) ) = 0$.
If $s \in [-1,1] \setminus [s_- , s_+]$, then $\widetilde\xi(s) < 0$, so the right side of \eqref{haus bdy sum} for $\beta = 0$ decays exponentially fast in $m$. Thus the expected number of sets in $\mathcal I_m$ tends to zero exponentially fast, and it follows from the Borel Cantelli lemma that a.s.\ $\mathcal I_m = \emptyset$ for sufficiently large $m$. Hence a.s.\ $\widetilde\Theta^{ s;*}_\delta (\mathbf D\setminus K_t)= \emptyset$ for each $\delta> 0$. Finally, it follows from Lemma~\ref{s>1 empty} that $\widetilde\Theta^{ s;*}(\mathbf D\setminus K_t) = \emptyset$ for $s\notin [-1,1]$.
\end{proof}
\subsection{Upper bound for the Hausdorff dimension of the subset of the curve}
In this subsection we will use Theorem~\ref{1pt forward} to give an upper bound for the Hausdorff dimension of the sets $\Theta^{s;\geq}(D )$ and $\Theta^{s;\leq}(D )$ of Section~\ref{multifractal def}, with $D = D_\eta$ as in Theorem~\ref{main thm}. For this purpose it will be convenient to introduce a slight variant of the sets of Section~\ref{multifractal def}. For a domain $D\subset \mathbf C$, a conformal map $\phi : \mathbf D\rightarrow D$, $s\in \mathbf R$, and $u > 0$, let
\begin{equation} \label{theta^su def}
\Theta^{s;u}(D) := \left\{x\in\partial D : s-u\leq \limsup_{\epsilon\rightarrow 0} \frac{\log |\phi'((1-\epsilon )\phi^{-1}(x) )|}{-\log \epsilon } \leq s+u \right\}.
\end{equation}
\begin{lem} \label{haus curve upper u}
Let $\eta$ be a chordal $\operatorname{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$ and let $D_\eta$, $\xi(s)$, $s_-$, and $s_+$ be as in Theorem~\ref{main thm}. Then a.s.\
\begin{equation} \label{haus curve eqn}
\dim_{\mathcal H} \Theta^{s;u}(D_\eta) \leq \xi(s ) + o_u(1),
\end{equation}
whenever $s \in [s_- , s_+]$, and a.s.\ $ \Theta^{s;u}(D_\eta) =\emptyset$ for sufficiently small $u$ otherwise. The $o_u(1)$ in \eqref{haus curve eqn} tends to 0 as $u\rightarrow 0$ and can be taken to be uniform for $s$ in compact subsets of $(-1,1)$.
\end{lem}
\begin{remark}
If $\alpha(s)$ is as in \eqref{alpha def}, $\gamma(s)$ is as in \eqref{gamma def}, and $ \xi(s)$ is as in \eqref{xi(s)}, we have
\begin{equation}\label{xi choice}
\xi(s ) = 2 - \frac{\gamma(s)}{1-s} = \frac{1-\alpha(s) }{1-s } .
\end{equation}
\end{remark}
To prove Lemma~\ref{haus curve upper u} we first need the following lemma.
\begin{lem}\label{ab and su}
Let $D\subset\mathbf C$ be a simply connected domain and let $\phi : \mathbf D\rightarrow D $ be a conformal map. Suppose $x\in \Theta^{s;u}(D)$ for some $s\in\mathbf R$ and $u >0$.
There is a sequence of points $(w_k)$ in $ D$ converging to $x$ such that
\begin{equation*}
\frac{-s-u}{1-s+u} \leq \liminf_{k\rightarrow\infty} \frac{\log |(\phi^{-1})'(w_k)|}{-\log\operatorname{dist}(w_k,\partial D)} \leq \limsup_{k\rightarrow\infty} \frac{\log |(\phi^{-1})'(w_k)|}{-\log\operatorname{dist}(w_k,\partial D)} \leq \frac{-s+u}{1-s-u } .
\end{equation*}
and
\begin{equation*}
\limsup_{k\rightarrow \infty} \frac{\log |w_k-x|}{ -\log \operatorname{dist}(w_k , \partial D)} \leq -\frac{1-s-u }{1-s+u } .
\end{equation*}
\end{lem}
\begin{proof}
Let $x\in \Theta^{s;u}(D)$.
For $\epsilon > 0$, put $z_\epsilon = \phi((1-\epsilon ) \phi^{-1}(x))$. By the Koebe quarter theorem, we have
\begin{equation} \label{z_ep dist}
\operatorname{dist}(z_\epsilon , \partial D) \asymp \epsilon |\phi'((1-\epsilon ) \phi^{-1}(x))| ,
\end{equation}
with proportionality constants $1/4$ and $4$ in each $\asymp$.
Clearly,
\begin{equation} \label{z_ep deriv}
(\phi^{-1} )'(z_\epsilon) = \frac{1}{\phi'((1-\epsilon ) \phi^{-1} (x)) }.
\end{equation}
Now let $\delta >0$.
By \cite[Proposition~2.7]{lawler-viklund-tip} we have
\begin{equation*}
\limsup_{\epsilon \rightarrow 0} \frac{ \log \epsilon^{1-s-u} }{\log v (x ;\epsilon) } \leq 1 ,
\end{equation*}
where $v(x ;\epsilon)$ is the length of the image of the curve $t\mapsto z_t$ for $t\in [0,\epsilon]$. Consequently, for sufficiently small $\epsilon > 0$ we have
\begin{equation} \label{z_ep dist >}
|z_\epsilon - x| \leq \epsilon^{1-s-u - \delta} .
\end{equation}
By assumption, for sufficiently small $\epsilon > 0$ we have $|\phi'((1-\epsilon ) \phi^{-1}(x))| \leq \epsilon^{- s - u -\delta}$, so by \eqref{z_ep dist}, $\operatorname{dist}(z_\epsilon , \partial D) \leq \epsilon^{1-s- u-\delta}$ and by \eqref{z_ep deriv}, $|(\phi^{-1})'(z_\epsilon)| \geq \epsilon^{s+ u + \delta}$. Furthermore, for any $k\in \mathbf N$, we can find $\epsilon_k $ such that $|\phi'((1+\epsilon_k) \phi^{-1}(x))| \geq \epsilon_k^{-s+ u+\delta}$. For each such $k$, we have $\operatorname{dist}(z_{\epsilon_k} , \partial D) \geq \epsilon_k^{1-s+ u + \delta}$ and $|(\phi^{-1})'(z_{\epsilon_k})| \leq \epsilon_k^{s - u-\delta}$.
Hence
\begin{equation} \label{liminf >}
\liminf_{k\rightarrow\infty} \frac{\log |(\phi^{-1})'(z_{\epsilon_k})|}{ -\log\operatorname{dist}(z_{\epsilon_k} ,\partial D)} \geq \frac{-s -u -\delta }{1-s+u + \delta } .
\end{equation}
and
\begin{equation} \label{limsup <}
\limsup_{k\rightarrow \infty} \frac{\log |(\phi^{-1})'(z_{\epsilon_k} )|}{-\log\operatorname{dist}(z_{\epsilon_k} ,\partial D)} \leq \frac{-s +u+\delta}{1-s -u-\delta} .
\end{equation}
By \eqref{z_ep dist >}, we also have
\begin{equation} \label{lim dist <}
\limsup_{k\rightarrow \infty} \frac{\log |z_{\epsilon_k} -x|}{-\log \operatorname{dist}(z_{\epsilon_k} , \partial D)} \leq -\frac{1-s-u-\delta}{1-s+u+\delta} .
\end{equation}
Since $\delta$ is arbitrary, by combining \eqref{liminf >}, \eqref{lim dist <}, and \eqref{limsup <} we obtain the statement of the lemma with $w_k = z_{\epsilon_k}$.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{haus curve upper u}]
The statement for $s \notin [s_- , s_+]$ follows from the analogous statement in Proposition~\ref{bdy haus upper}, so we henceforth assume $s \in [s_- , s_+]$.
By countable stability of Hausdorff dimension $\xi(s)$, to prove \eqref{haus curve eqn}, it is enough to show that a.s.\ $ \mathcal H^\beta( \Theta^{s ,u }(D_\eta) \cap B_d(0) ) = 0$ for each $\beta > \xi(s ) + o_u(1)$, and each $d\in (0,1)$. It is moreover enough to prove the result restricted to the event $\mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu)$ (in the notation of Theorem~\ref{1pt forward}) for an arbitrary choice of $\mu\in\mathcal M$.
Let
\[
r > \frac{1-s-u}{1-s+u} .
\]
Note that we can take $r = 1+o_u(1)$. For $n\in \mathbf N$ let $\mathcal D^n = 2^{-n(1-s) -4} \mathbf Z^2$ be the dyadic lattice of mesh size $2^{-n(1-s) -4}$.
For $z\in \mathcal D^n$, let $B_0^n(z) , B_1^n(z)$, $B_2^n(z)$, and $B_3^n(z)$ be the disks centered at $z$ of radii $2^{-n(1-s) - 4 }$, $2^{-n(1-s )-2}$, $2^{-n(1-s )+2}$, and $ 2^{-n(1-s ) r +1 }$, respectively.
Define $\Psi_\eta$ as in Section~\ref{time infty setup sec}. For $z\in \mathbf D$ let $E^n(z)$ be the event that the following occurs.
\begin{enumerate}
\item $\eta \cap B^n_2(z) \not=\emptyset$. \label{3disks hit}
\item $\eta \cap B^n_1(z) = \emptyset$.\label{3disks no hit}
\item There is a $w\in B_0^n(z)$ with $2^{-n(s + 2u )} \leq |\Psi_\eta '(w)| \leq 2^{-n(s - 2u )}$.\label{3disks deriv}
\end{enumerate}
On $E^n(z)$, we have
\begin{equation*}
2^{-n(1-s )} \preceq \operatorname{dist}(z , \partial D_\eta ) \preceq 2^{-n(1-s )} ,\qquad 2^{-n(s + 2u )} \preceq |\Psi_\eta '(z)| \preceq 2^{-n(s -2u )} ,
\end{equation*}
with constants uniform in $B_d(0)$ (the inequality for $|\Psi_\eta'|$ follows from the Koebe distortion theorem).
So, by Proposition~\ref{1pt forward},
\begin{equation}\label{E^n(z) prob}
\mathbf P\left( E^n(z) \cap \mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu) \right) \preceq 2^{-n(\gamma(s ) - 2\gamma_0(s) u )}
\end{equation}
with constants uniform in $B_d(0)$.
Let $\mathcal U^n$ be the set of disks $B^{ n}_3(z)$ for $z\in \mathcal D^n$ such that $z\in B_d(0)$ and $E^n(z)$ occurs. Note that the cardinality of the set of disks which can belong to $\mathcal U^n$ is of order $2^{2n(1-s )}$.
We claim that $ \Theta^{s;u}(D_\eta) \cap B_d(0) \subset \bigcup_{n \geq N} \bigcup_{B^n_3(z) \in\mathcal U^n} B^n_3(z)$ for each $N\in \mathbf N$.
Indeed, suppose $x\in \Theta^{s;u} (D_\eta ) \cap B_d(0)$.
By Lemma~\ref{ab and su}, for any $\delta >0$, we can find a sequence $n_k \rightarrow \infty$ and a sequence of points $w_k \in D_\eta$ converging to $x$ such that for each $k$, $2^{-n_k (1-s ) -2} \leq \operatorname{dist}(w_k , \partial D_\eta ) \leq 2^{-n_k(1-s ) }$, $|w_k - x| \leq 2^{-n_k(1-s )r}$, and $2^{-n_k(s +2u )} \leq |\Psi_\eta '(w_k)| \leq 2^{-n_k (s -2u)}$.
Each $w_k$ belongs to $B^{n_k}_0(z)$ for some $z\in \mathcal D^{n_k}$. Our hypothesis on the distance from $w_k$ to $\partial D_\eta$ implies that conditions~\ref{3disks hit} and~\ref{3disks no hit} in the definition of $E^{n_k}(z)$ hold for this $z$. Clearly, condition~\ref{3disks deriv} also holds for this $z$.
Thus for such a $z$, $E^n(z)$ holds and we have $x\in B_3^n(z)$ (here we use the condition on $|w_k - x|$). This proves our claim.
Thus, for any $n\in \mathbf N$, $\bigcup_{n\geq m} \mathcal U^n$ is a cover of $ \Theta^{s;u}(\partial D_\eta ) \cap B_d(0)$. Each set in this cover has diameter $\preceq 2^{-n(1-s ) r}$ and we have by \eqref{E^n(z) prob} that
\begin{align} \label{haus curve sum}
\mathbf E\left( \mathbf 1_{\mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu)} \sum_{n=m}^\infty \sum_{U\in \mathcal U_m} (\operatorname{diam} B^n(z))^\beta \right) &\preceq \sum_{n=m}^\infty \sum_{z\in\mathcal D_n \cap B_d(0)} 2^{-n \beta(1-s )r } \mathbf P\left( E^n(z)\cap\mathcal G(\Psi_\eta , \mu) \cap \mathcal G(\Psi_\eta^- , \mu) \right) \nonumber \\
&\preceq \sum_{n=m}^\infty 2^{2n(1-s )} 2^{-n\beta (1-s ) r} 2^{-n(\gamma(s)- 2\gamma_0(s) u )} .
\end{align}
This tends to 0 as $m\rightarrow \infty$ provided
\[
\beta > 2 - \frac{ \gamma(s) + 2\gamma_0(s) u}{(1-s) r} = \xi(s) + o_u(1) ,
\]
where the $o_u(1)$ can be taken to be uniform for $s$ in compact subsets of $(-1,1)$. Since $\mu$ is arbitrary we conclude that $ \mathcal H^\beta( \Theta^{s;u } (\partial D_\eta ) \cap B_d(0)) = 0$ for any such $\beta$.
\end{proof}
From Lemma~\ref{haus curve upper u}, we can deduce the upper bounds on $\dim_{\mathcal H} \Theta^{s;\geq}(D_\eta)$ and $\dim_{\mathcal H}(\Theta^{s;\leq}(D_\eta))$ in Theorem~\ref{main thm}.
\begin{prop} \label{haus curve upper}
Suppose we are in the setting of Theorem~\ref{main thm}. Then a.s.\
\alb
& \dim_{\mathcal H} \Theta^{s; \geq }(D_\eta) \leq \xi(s) ,\qquad \frac{\kappa}{4} \leq s \leq s_+ \\
& \dim_{\mathcal H} \Theta^{s; \leq }(D_\eta) \leq \xi(s) ,\qquad s_- \leq s \leq \frac{\kappa}{4} .
\ale
\end{prop}
\begin{proof}
For $s\leq \kappa/4$ and any $n\in\mathbf N$, we have
\begin{equation} \label{theta dim union}
\Theta^{s;\leq} (D_\eta )\subset \bigcup_{j=m_0}^{m_1} \Theta^{-j/n ; 1/n}(D_\eta) ,
\end{equation}
where $m_0$ is the greatest integer such that $m_0/n $ is smaller than $s_-$ and $m_1$ is the least integer such that $m_1/n\geq s$. We have that $\xi(s)$ is increasing on $[0,\kappa/4]$ and $\gamma_0(s)$ is uniformly bounded for $s \in [s_- , s_+]$ (and for $s \leq s_1$ for some $s_1 <1$ in the case $\kappa=4$). Our upper bound in the case $s \leq \kappa/4$ thus follows from Lemma~\ref{haus curve upper u} and \eqref{theta dim union} together with stability of Hausdorff dimension under unions. Similarly for the case $s\geq \kappa/4$.
\end{proof}
\subsection{Upper bound for the integral means spectrum}
\label{ims upper sec}
In this subsection we will prove the upper bound for the bulk integral means spectrum of the SLE curve in Corollary~\ref{ims cor}. In light of Lemma~\ref{ims zero one}, it will be enough to prove an upper bound for the bulk integral means spectrum of $\mathbf D\setminus \eta^t$ for given $t\geq 0$ in the case of an ordinary $\operatorname{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$ for $\kappa \leq 4$.
\begin{prop} \label{ims upper}
Let $\kappa \in (0,4]$ and let $ \operatorname{IMS}^*(a)$ be defined as in Corollary~\ref{ims cor}. Let $\eta$ be a chordal $\operatorname{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$. For each $t > 0$ and each $a \in \mathbf R$, a.s.\ $\operatorname{IMS}_{\mathbf D\setminus \eta^t}^{\operatorname{bulk}}(a) \leq \operatorname{IMS}^*(a) $.
\end{prop}
\begin{proof}
Let $(f_t)$ be the centered Loewner maps for $\eta$, as defined in Section~\ref{disk sec}.
For $\delta>0$, let $U_t(\delta)$ be the set of $z\in \mathbf D\setminus \eta^t$ with $1-|f_t^{-1}(z)| \geq \delta$ and $|z-i| , |z+i | \geq \delta$. Also define the sets $A_\epsilon^\zeta(f_t^{-1} )$ be as in Section~\ref{ims sec} (immediately following~\eqref{I_zeta def}). For any given $\zeta > 0$ there a.s.\ exists (random) $\delta>0$ such that $A_\epsilon^\zeta(f_t^{-1}) \subset \partial B_{1-\epsilon}(0) \cap U_t( \delta)$ for sufficiently small $\epsilon$. Therefore, it is enough to show that for each $\delta > 0$ and each $\beta > \operatorname{IMS}^*(a) $, we a.s.\ have
\begin{equation} \label{U(delta) sup}
\limsup_{\epsilon\rightarrow 0} \frac{\log \int_{\partial B_{1-\epsilon}(0) \cap U_t( \delta)} |(f_t^{-1})'(z)|^a \, dz }{-\log \epsilon} \leq \beta .
\end{equation}
Fix $\delta>0$ and $\beta > \operatorname{IMS}^*(a)$ as above. Also fix $t> 0$ and $c>2$ and define the events $\overline E^{s;u}(z) = \overline{E}^{s;u}(z ; t , c , 1-\delta ) $ as in Section~\ref{disk sec} with $d = 1-\delta$. Let $s_-$ and $s_+$ be as in the statement of Theorem~\ref{main thm}. For $n\in \mathbf N$ and $k\in \{0,\dots,n\}$, let
\[
u_n = \frac{s_+-s_-}{n} , \qquad s_k^n = s_0 + k u_n .
\]
For $n\in\mathbf N$ and $\epsilon > 0$, let $A_\epsilon^n(-)$ (resp. $A_\epsilon^n(+)$) be the set of $z\in \partial B_{1-\epsilon}(0) \cap U_t(\delta) $ such that $|(f_t^{-1})'(z)| \leq \epsilon^{-s_- + 1/n}$ (resp. $|(f_t^{-1})'(z)| \geq \epsilon^{-s_+ - 1/n}$). For $k \in\{0,\dots,n\} $, let $A_\epsilon^n(k)$ be the set of $z\in \partial B_{1-\epsilon}(0) \cap U_t(\delta)$ such that the $\overline{E}^{s_k^n ; u_n}(z)$ occurs. Let $\ell_\epsilon^n(k)$ be the Lebesgue measure of $A_\epsilon^n(k)$ and let $\ell_\epsilon^n(\pm)$ be the Lebesgue measure of $A_\epsilon^n(\pm)$.
In what follows, we require implicit constants to be independent of $\epsilon$, but not of $n$ or $k$, and we denote by $o_n(1)$ a term which tends to 0 as $n\rightarrow\infty$ and does not depend on $k$ or $\epsilon$.
By construction, we have $\partial B_{1-\epsilon}(0) \cap U_t(\delta) = A_\epsilon^n(-)\cup A_\epsilon^n(+) \cup \bigcup_{k=0}^n A^n_\epsilon $, whence
\alb
\int_{\partial B_{1-\epsilon}(0) \cap U_t( \delta) } |(f_t^{-1})'(z)|^a \, dz &\preceq \sum_{k=0}^n \epsilon^{-a s_k^n + o_n(1) } \ell_\epsilon^n(k) + \epsilon^{-a s_-} \ell_\epsilon^n(-) + \epsilon^{-a s_+} \ell_\epsilon^n(+) .
\ale
The proof of Lemma~\ref{bdy haus upper} shows that for each $n\in\mathbf N$, there a.s.\ exists a random $\epsilon_0^n > 0$ such that for $\epsilon < \epsilon_0^n$, the sets $A_\epsilon^n(-)$ and $A_\epsilon^n(+)$ are empty. Hence for $\epsilon < \epsilon_0^n$, we have
\begin{equation} \label{int < sup}
\int_{\partial B_{1-\epsilon}(0) \cap U_t( \delta) } |(f_t^{-1})'(z)|^{a} \, dz \preceq \sup_{k\in\{0,\dots,n\} } \epsilon^{-a s_k^n } \ell_\epsilon^n(k) .
\end{equation}
By Corollary~\ref{1pt chordal disk}, for $k\in \{0,\dots,n\}$, we have
\begin{equation*}
\mathbf E(\ell_\epsilon^n(k) ) \preceq \epsilon^{\alpha(s_k^n) + o_n(1) } .
\end{equation*}
By Chebyshev's inequality,
\begin{equation} \label{chebyshev k}
\mathbf P\left( \epsilon^{-a s_k^n} \ell_\epsilon^n(k) > \epsilon^{-\beta } \right) \preceq \epsilon^{\alpha(s_k^n) - a s_k^n +\beta + o_n(1)}.
\end{equation}
We have
\begin{equation} \label{ims inf}
\inf_{s \in [s_- , s_+]} \left(\alpha(s_k^n) - a s_k^n \right) = -\operatorname{IMS}^*(a).
\end{equation}
Note that the range $[a_- , a_+]$ in Corollary~\ref{ims cor} is precisely the set of $a\in \mathbf R$ for which the minimizer in~\eqref{ims inf} is not equal to $s_-$ or $s_+$.
It follows that for sufficiently large $n\in\mathbf N$, depending only on $\beta$, we have
\[
\mathbf P\left( \sup_{k\in\{0,\dots,n\} } \epsilon^{-a s_k^n } \ell_\epsilon^n(k) > \epsilon^{-\beta} \right) \preceq \epsilon^{\beta - \operatorname{IMS}^*(a) + o_n(1)} .
\]
Since $\beta > \operatorname{IMS}^*(a)$, if $n\in\mathbf N$ is chosen sufficiently large (depending only on $\beta$ and $a$), then the Borel-Cantelli lemma together with~\eqref{int < sup} implies that a.s.
\[
\int_{\partial B_{1-2^{-j}}(0) \cap U_t(\delta) } |(f_t^{-1})'(z)|^a \, dz \leq 2^{-j\beta}
\]
for sufficiently large $j\in\mathbf N$. By the Koebe distortion theorem, it follows that a.s.
\[
\limsup_{\epsilon\rightarrow 0} \frac{\log \int_{\partial B_{1-\epsilon}(0) \cap U_t(\delta) } |(f_t^{-1})'(z)|^a \, dz }{-\log \epsilon} \leq \beta .
\]
This proves (\ref{U(delta) sup}), and hence the statement of the proposition.
\end{proof}
\begin{comment}
Clear[alpha, s, a, f]
alpha[s_] := (4 + kappa)^2 s^2/(8 kappa (1 + s));
f[s_, a_] := alpha[s] - a s
Solve[D[f[s, a], s] == 0, s];
FullSimplify
FullSimplify[f[s, a] /.
\end{comment}
\subsection{Event at the hitting time} \label{2pt setup sec}
In this subsection we introduce an event which will serve as the basic building block for the ``perfect points" which we will use to prove our lower bounds on the Hausdorff dimensions of $\Theta^s(D_\eta)$ and $\widetilde\Theta^s(D_\eta)$.
\subsubsection{Definition of the event}
Suppose $\eta : [0,\infty ] \rightarrow \overline{\mathbf D}$ is a simple curve in $\mathbf D$ which connects distinct points $x,y \in \partial\mathbf D$ and does not otherwise hit $\partial\mathbf D$.
We recall the notation
\begin{equation*}
\eta^t = \eta([0,t]),\qquad \eta = \eta([0,\infty])
\end{equation*}
from Section~\ref{basic notations}.
Throughout this section,
for $\beta > 0$, we will write
\begin{equation*}
B_\beta := B_{e^{-\beta}}(0) .
\end{equation*}
Let $\overline \eta$ be the time reversal of $\eta$.
Fix parameters $\beta > 0$, $u, a, c > 0$, $v\geq 0$, $q\in (-1/2,\infty)$, and $\mu\in\mathcal M$. The parameter $q$ corresponds to $s/(1-s)$, for $s$ the parameter of Theorem~\ref{main thm}.
Let $E = E_\beta^{q;u}(\eta ; a, c , \mu )$ be the event that the following holds.
\begin{enumerate}
\item Let $\tau_\beta$ (resp.\ $\overline\tau_\beta$) be the first time that $\eta$ (resp.\ $\overline\eta$) hits $\partial B_\beta$. Then $\tau_\beta , \overline\tau_\beta < \infty$. \label{E hit}
\item The harmonic measure from 0 in $\mathbf D \setminus (\eta^{\tau_{\beta }} \cup \overline\eta^{\overline\tau_{\beta }})$ of each of the two sides of $\eta^{\tau_{\beta }} $ and each of the two sides of $ \overline\eta^{\overline\tau_{\beta }}$ is at least $a$. \label{E top}
\item Let $\phi_\beta :\mathbf D \setminus (\eta^{\tau_\beta} \cup \overline\eta^{\overline\tau_\beta}) \rightarrow\mathbf D$ be the unique conformal transformation which takes $x$ to $-i$, $y$ to $i$, and the midpoint $m$ of $[x,y]_{\partial\mathbf D}$ to 1. Then we have that $c^{-1} e^{-\beta (q+u)} \leq |\phi_{\beta }'(0)| \leq c e^{-\beta (q-u)}$. \label{E phi}
\item $\mathcal G'(\eta^{\tau_{\beta }} \cup \overline\eta^{\tau_{\beta }}, \mu)$ occurs (Definition~\ref{G' def}). \label{E G}
\end{enumerate}
We will be primarily interested in the case where $\eta$ is a chordal $\operatorname{SLE}_\kappa$ process from $x$ to $y$. In this case, we write $\mathcal F_\beta$ for the $\sigma$-algebra generated by $\eta|_{[0,\tau_\beta]}$ and $\overline\eta|_{[0,\overline\tau_\beta]}$.
\subsubsection{Setup and upper bound}
We need to estimate the probability of the event $E$ of the preceding subsection. To this end we will prove the following.
\begin{prop} \label{1pt at hitting}
Let $x,y\in\partial \mathbf D$. Let $\eta$ be a chordal $\operatorname{SLE}_\kappa$ process from $x$ to $y$ in $\mathbf D$. Let $E =E_\beta^{q;u}(\eta; a ,c , \mu )$ be as in Section~\ref{2pt setup sec}. Let $\gamma$ and $\gamma_0$ be as in~\eqref{gamma def} and
\begin{align} \label{gamma* def}
\gamma^*(q) &:= (q+1) \gamma\left(\frac{q}{1+q} \right) = \frac{8 \kappa + 8 \kappa q + (4 - \kappa)^2 q^2}{8 (\kappa + 2\kappa q)} .
\end{align}
There exists a continuous function $\gamma_0^* : (-1/2 , \infty) \rightarrow (0,\infty)$ (with $\gamma_0^*(q)$ depending only on $q$) such that the following is true for sufficiently small $u > 0$ (depending only on $q$). We have
\begin{equation}\label{1pt hitting upper}
\mathbf P(E) \preceq e^{-\beta (\gamma^*(q) + \gamma^*_0(q) u )} .
\end{equation}
Moreover, there exists $\mu\in\mathcal M$ depending only on $z$ but uniform for $z$ in compacts such that for each $q \in (-1/2,\infty)$, each $\beta > 0$, and each $a , u , c > 0$ with $\beta u \geq 100$,
\begin{equation} \label{1pt hitting lower}
\mathbf P(E ) \succeq e^{-\beta(\gamma^*(q) + \gamma^*_0(q) u ) }.
\end{equation}
The implicit constants are independent of $\beta$, uniform for $z$ in compacts, and uniform in $x,y$ provided $|x-y|$ is bounded below by a positive constant.
\end{prop}
First we will prove the upper bound~\eqref{1pt hitting upper}, which is a straightforward consequence of Theorem~\ref{1pt forward}.
\begin{proof}[Proof of Proposition~\ref{1pt at hitting}, upper bound]
Let $\widehat \eta$ be the image under $\phi_{\beta }$ of the part of $\eta$ lying between $\eta(\tau_{\beta })$ and $\overline\eta(\overline\tau_{\beta })$. Let $\widehat x = \phi_{\beta}(\eta(\tau_{\beta}))$ and $\widehat y = \phi_{\beta}(\overline\eta(\overline\tau_{\beta}))$, so that the conditional law of $\widehat\eta$ given $\mathcal F_{\beta}$ is that of an $\operatorname{SLE}_\kappa$ from $\widehat x$ to $\widehat y$ in $\mathbf D$. Let $C> 1$. Let $\widehat E = \widehat E(C)$ be the event that the following occurs.
\begin{enumerate}
\item $\widehat \eta$ does not exit $\phi_{\beta}(B_{1})$. \label{hat E stay}
\item Define the domain $D_{\widehat \eta}$ as in Section~\ref{time infty setup sec}. Then $\phi_{\beta}(0) \in D_{\widehat\eta}$ and $C^{-1} (1-|\phi_{\beta}(0)|) \leq \operatorname{dist}(\phi_{\beta}(0) ,\partial D_{\widehat\eta}) \leq C (1-|\phi_{\beta}(0)|)$. \label{hat E dist}
\item Let $\Phi_{\widehat\eta} : D_{\widehat\eta} \rightarrow \mathbf D$ be the conformal map fixing $-i$, $i$, and 1. Then $C^{-1} \leq |\Phi_{\widehat\eta}'(\phi_{\beta}(0))| \leq C$.
\end{enumerate}
It follows from condition~\ref{E top} in the definition of $E$ and \cite[Lemma~2.3]{miller-wu-dim} that we can find a $C > 0$ depending only on $a$ such that for sufficiently large $\beta$, $\mathbf P(\widehat E | E) \succeq 1$. Thus
\begin{equation} \label{E' and hat E}
\mathbf P(E) \asymp \mathbf P(E \cap \widehat E) .
\end{equation}
So, it will suffice to prove an upper bound for $\mathbf P\left(E \cap \widehat E \right)$.
Let $s \in (-1,1)$ and $\epsilon > 0$ be chosen so that
\begin{equation} \label{s ep choice}
\frac{s}{1-s} = q ,\qquad \epsilon^{1-s } = e^{-\beta }.
\end{equation}
Let $D_\eta$, $\Psi_\eta$, $\Psi_\eta^-$, and $\mathcal E_\epsilon^{s; u}(\eta ,0 ; c)$ be as in Section~\ref{time infty setup sec}. It follows from Lemma~\ref{G dist} and condition~\ref{E G} in the definition of $E$ that
\begin{equation} \label{E implies G}
E\subset \mathcal G(\phi_{\beta} , \mu')
\end{equation}
for some $\mu' \in \mathcal M$ depending only on $\mu$. By combining this with condition~\ref{hat E stay} in the definition $\widehat E$ we see that
that $ E\cap \widehat E \subset \mathcal G(\Psi_\eta , \mu') \cap \mathcal G(\Psi_{ \eta}^- , \mu')$ for some (possibly smaller) $\mu'\in\mathcal M$ depending only on $\mu$. We furthermore have
\[
\Psi_\eta = \Psi_{\widehat\eta} \circ \phi_{\beta} .
\]
Hence we have
\[
E\cap \widehat E \subset \mathcal E_\epsilon^{s; u}(\eta ,0 ; c) \cap \mathcal G(\Psi_\eta , \mu') \cap \mathcal G(\Psi_\eta^- , \mu')
\]
for suitable choice of $\mu'$ and $c$. Thus~\eqref{1pt hitting upper} follows from~\eqref{E' and hat E} and the upper bound in Theorem~\ref{1pt forward}. Note that we can take the dependence on $u$ to be linear (with slope depending on $q$) since the exponent in the upper bound in Theorem~\ref{1pt forward} depends smoothly on $s\in (-1,1)$ and $u > 0$ sufficiently small.
\end{proof}
\subsubsection{Lower bound}
The proof of the lower bound in Proposition~\ref{1pt at hitting} will take substantially more work than the proof of the upper bound. The basic idea is to stop $\eta$ and $\overline\eta$ at times $t$ and $\overline t$ for which the the following is true. On the event $\mathcal E_\beta^{s;u}(\cdot)$ of Theorem~\ref{1pt forward}, the conformal map from $\mathbf D\setminus (\eta^{t_0} \cup \overline\eta^{\overline t_0})$ to $\mathbf D$ which takes $x$ to $-i$, $y$ to $i$, and the midpoint of $[x,y]_{\partial\mathbf D}$ to 1 has the same derivative behavior as the conformal map $\Psi_\eta :\mathbf D\setminus \eta \rightarrow \mathbf D$ with the same normalization; the points $\eta(t_0)$ and $\overline\eta(\overline t_0)$ are at distance slightly less than $e^{-\beta}$ from 0; and the law of the remainder of the curve given $\eta^{t_0} \cup \overline\eta^{\overline t_0}$ has the law of a chordal $\operatorname{SLE}_\kappa$. We also need to require that $\eta(t_0)$ and $\overline\eta(\overline t_0)$ are sufficiently far apart in a conformal sense, so that they do not immediately link up after times $t_0$ and $\overline t_0$. We then condition on $\eta^{t_0} \cup \overline\eta^{\overline t_0}$ and use standard arguments to get that the curves reach $B_\beta$ without any pathological behavior. The main difficulty in the proof is constructing the times $t_0$ and $\overline t_0$.
We start by inductively defining a means of growing $\eta$ and $\overline\eta$ simultaneously to get an increasing family of hulls $K_t\subset \mathbf D$. Assume $\eta$ (resp.\ $\overline\eta$) is parametrized in such a way that its image under the conformal map $\mathbf D\rightarrow \mathbf H$ taking $-i$ to 0, $i$ to $\infty$, and 0 to $i$ (resp.\ the reciprocal of this conformal map) is parametrized by half plane capacity. Let $\sigma_1$ be the first time $t$ that $\operatorname{hm}^0(\eta^t ; \mathbf D\setminus \eta^t) = 1/2$. This time is a.s.\ finite since a Brownian motion started from 0 has probability at least $1/2$ to hit $\eta$ before $\partial\mathbf D$. For $t\leq \sigma_1$, let $K_t = \eta^t$. Let $\overline\sigma_1$ be the first $\overline t$ that either $\operatorname{hm}^0(\eta^{\overline t} ; \mathbf D\setminus (\eta^{\sigma_1} \cup \overline\eta^{\overline t}) ) =1/2$ or $\overline\eta(\overline t) = \eta(\sigma_1)$. For $t \in [\sigma_1 ,\sigma_1 + \overline\sigma_1 ]$ let $K_t = \eta^{\sigma_1} \cup \overline\eta^{t -\sigma_1}$.
Inductively, suppose $n \geq 2$ and $\sigma_{n-1}$, $\overline\sigma_{n-1}$, and $K_t$ for $t\leq \sigma_{n-1 } + \overline\sigma_{n-1}$ have been defined. If $K_{\sigma_{n-1} + \overline\sigma_{n-1} } =\eta$ we let $\sigma_n = \sigma_{n-1}$ and $\overline\sigma_n = \overline\sigma_{n-1}$. Otherwise, let $\sigma_n$ be the least $t \geq \sigma_{n-1}$ such that either $\operatorname{hm}^0( \eta^t ; \mathbf D\setminus ( \eta^t \cup \overline\eta^{\overline\sigma_{n-1}})) =1/2$ or $\eta(t) =\overline\eta(\overline\sigma_{n-1})$. Let $K_t = \eta^{t-\overline\sigma_{n-1}} \cup \overline\eta^{\overline\sigma_{n-1}}$ for $t\in [\sigma_{n-1} + \overline\sigma_{n-1} , \sigma_n + \overline \sigma_{n-1}]$. Let $\overline\sigma_n$ be the first time $\overline t \geq \overline \sigma_{n-1}$ such that either $\operatorname{hm}^0( \overline\eta^{\overline t} ; \mathbf D\setminus ( \eta^{\sigma_n} \cup \overline\eta^{\overline t})) =1/2$ or $\overline\eta(\overline t ) = \eta(\sigma_n)$. Let $K_t = \eta^{\sigma_n} \cup \overline\eta^{ t - \sigma_n}$ for $t\in [\sigma_n +\overline\sigma_{n-1}, \sigma_n + \overline \sigma_n]$.
For each $t \geq 0$, let $T_t$ (resp.\ $\overline T_t$) be the time such that $\eta(T_t)$ (resp.\ $\overline\eta(\overline T_t)$) is the tip of the part of $\eta$ (resp.\ $\overline\eta$) included in $K_t$. Observe that the Markov property and reversibility of SLE imply that for each $t$, the conditional law of $\eta \setminus K_t$ given $K_t$ is that of a chordal $\operatorname{SLE}_\kappa$ from $\eta(T_t)$ to $\overline\eta(\overline T_t)$ in $\mathbf D\setminus K_t$.
\begin{lem} \label{sigma hm converge}
Let $\sigma_\infty = \lim_{n\rightarrow\infty} \sigma_n$ and $\overline \sigma_\infty = \lim_{n\rightarrow\infty} \overline \sigma_n$ (the limits necessarily exist by monotonicity). Let $K_\infty = \eta^{\sigma_\infty} \cup \overline\eta^{\overline\sigma_\infty}$. Then we a.s.\ have
\[
\lim_{n\rightarrow\infty} \operatorname{hm}^0(\eta^{\sigma_n} ; \mathbf D\setminus K_{\sigma_n +\overline\sigma_n}) = \lim_{n\rightarrow\infty} \operatorname{hm}^0(\eta^{\sigma_n} ; \mathbf D\setminus K_{\sigma_n +\overline\sigma_{n-1}}) = \operatorname{hm}^0(\eta^{\sigma_\infty} ; \mathbf D\setminus K_\infty)
\]
and
\[
\lim_{n\rightarrow\infty} \operatorname{hm}^0(\overline\eta^{\overline\sigma_n} ; \mathbf D\setminus K_{\sigma_n +\overline\sigma_n}) = \lim_{n\rightarrow\infty} \operatorname{hm}^0(\overline\eta^{\overline\sigma_{n-1}} ; \mathbf D\setminus K_{\sigma_n +\overline\sigma_{n-1}}) = \operatorname{hm}^0(\overline\eta^{\overline\sigma_\infty} ; \mathbf D\setminus K_\infty) .
\]
\end{lem}
\begin{proof}
We a.s.\ have $0\notin\eta$ so it is a.s.\ the case that for each $\epsilon > 0$, we can find a random $\delta > 0$ such that for any $z\in \eta$, the probability that a Brownian motion started from 0 hits $B_\delta(z)$ before leaving $\mathbf D$ is at most $\epsilon$.
By a.s.\ continuity of $\eta$, we can a.s.\ find a (random) $N\in\mathbf N$ such that for $n\geq N$, $\eta([\sigma_n , \sigma_\infty])\subset B_\delta(\eta(\sigma_\infty))$ and $\overline\eta([\overline\sigma_n , \overline\sigma_\infty])\subset B_\delta(\overline\eta(\overline\sigma_\infty))$. Hence with probability at least $1-\epsilon$, a Brownian motion started from 0 exists $\mathbf D\setminus K_{\sigma_n + \overline\sigma_n}$ in the same place it exits $\mathbf D\setminus K_\infty$. This proves the limits involving $K_{\sigma_n + \overline\sigma_n}$. The limits involving $K_{\sigma_n + \overline\sigma_{n-1}}$ are proven similarly.
\end{proof}
\begin{lem} \label{sigma limit}
We a.s.\ have $K_\infty = \eta$. Let $z_\infty = \eta(\sigma_\infty) = \overline\eta(\overline \sigma_\infty)$ be the meeting point. On the event that 0 lies to the right of $\eta$ and $\operatorname{dist}(0 , \eta) \leq e^{-\beta}$, it holds a.s.\ that $\operatorname{hm}^0(\eta^{\sigma_\infty} ; D_\eta)$ and $\operatorname{hm}^0(\overline\eta^{\overline \sigma_\infty} ; D_\eta)$ are each at least $1/2 - o_\beta(1)$, where the $o_\beta(1)$ is a deterministic quantity which tends to 0 as $\beta\rightarrow 0$.
\end{lem}
\begin{proof}
First we argue that $K_\infty = \eta$. Suppose not. Almost surely, either $\operatorname{hm}^0(\eta^{\sigma_\infty} ; \mathbf D\setminus K_\infty)$ or $\operatorname{hm}^0(\overline\eta^{\overline\sigma_\infty} ; \mathbf D\setminus K_\infty)$ is $< 1/2$. Suppose $\operatorname{hm}^0(\eta^{\sigma_\infty} ; \mathbf D\setminus K_\infty) < 1/2$. The other case is treated similarly. By Lemma~\ref{sigma hm converge} we a.s.\ have $\operatorname{hm}^0(\eta^{\sigma_n} ; \mathbf D\setminus K_{\sigma_n +\overline\sigma_{n-1}}) < 1/2$ for sufficiently large $n$. By definition of $\sigma_n$ this can be the case only if $\eta(\sigma_n) = \overline\eta(\overline\sigma_{n-1})$ which implies $K_\infty = \eta$.
It is immediate from Lemma~\ref{sigma hm converge} that $\operatorname{hm}^0(\eta^{\sigma_\infty} ; D_\eta)$ and $\operatorname{hm}^0(\overline\eta^{\overline \sigma_\infty} ; D_\eta)$ are each at most $1/2$. Furthermore, the Beurling estimate implies $\operatorname{hm}^0(\partial\mathbf D ; D_\eta) = o_\beta(1)$. Hence
\[
\operatorname{hm}^0(\eta^{\sigma_\infty} ; D_\eta) = 1 - \operatorname{hm}^0(\overline\eta^{\overline \sigma_\infty} ; D_\eta) - \operatorname{hm}^0(\partial\mathbf D ; D_\eta) \geq 1/2 - o_\beta(1)
\]
and similarly for $\eta^{\overline\sigma_\infty}$.
\end{proof}
\begin{lem} \label{sigma good time}
For $t\geq 0$, let $\Phi_t $ be the conformal map from the connected component of $ \mathbf D\setminus K_t$ with $[x,y]_{\partial\mathbf D}$ on its boundary to $ \mathbf D$ taking $x^+$ to $-i$, $y^-$ to $i$, and the midpoint $m$ of $[x,y]_{\partial\mathbf D}$ to 1. Let $\widetilde\Phi_t$ be the conformal map from this same connected component to $\mathbf D$ which fixes 0 and takes $m$ to 1. Let $\Psi_\eta$ be as in Section~\ref{time infty setup sec}. Let $\mu\in\mathcal M$. There is a $C>1$ depending only on $\mu$ such that the following is true. If $\beta$ is chosen sufficiently large then on the event $\mathcal G(\Psi_\eta , \mu) \cap \{\operatorname{dist}(0 , \eta) \leq e^{-\beta}\} \cap \{0\in D_\eta\}$, there a.s.\ exists a time $\tau > 0$ such that the following is true.
\begin{enumerate}
\item $\operatorname{dist}(0,K_\tau ) \leq C \operatorname{dist}(0,\eta)$. \label{sigma good dist}
\item $C^{-1} |\Psi_\eta'(0)| \leq |\Phi_\tau '(0)| \leq C |\Psi_\eta'(0)|$. \label{sigma good deriv}
\item $\widetilde\Phi_\tau(\eta(T_\tau))$ and $\widetilde\Phi_\tau(\overline\eta(\overline T_\tau))$ lie in $[i,-i]_{\partial\mathbf D}$.
\item $\operatorname{hm}^0(\eta\setminus K_\tau ; D_\eta) \geq 1/4 - o_\beta(1)$. Here the $o_\beta(1)$ is deterministic and depends only on $\beta$. \label{sigma good mid}
\end{enumerate}
\end{lem}
\begin{proof}
Throughout, we assume we are working on the event $\mathcal G(\Psi_\eta , \mu) \cap \{\operatorname{dist}(0 , \eta) \leq e^{-\beta}\} \cap \{0\in D_\eta\}$ and we require all implicit constants to be deterministic and depend only on $\mu$.
Let $\widetilde\Psi_\eta : D_\eta \rightarrow \mathbf D$ be the conformal map which fixes 0 and takes $m$ to 1. If $z_\infty$ is as in Lemma~\ref{sigma limit} then by conformal invariance of harmonic measure we have
\begin{equation}\label{z_infty to 1}
|\widetilde\Psi_\eta(z_\infty) + 1| = o_\beta(1) .
\end{equation}
Let $\tau$ be the first time $t$ that $\widetilde\Psi_\eta(\eta(T_t))$ and $\widetilde\Psi_\eta(\overline \eta(\overline T_t))$ are both in $[i,-i]_{\partial\mathbf D}$. By Lemma~\ref{sigma limit} such a $t$ necessarily exists provided $\beta$ is large enough. Let $\widetilde A_\tau = [\widetilde\Psi_\eta(\overline \eta(\overline T_\tau)) , \widetilde\Psi_\eta(\eta(T_\tau)) ]_{\partial\mathbf D}$. By continuity one of the two endpoints of $\widetilde A_\tau$ is $-i$ or $i$ so by~\eqref{z_infty to 1} we have $\operatorname{hm}^0(\widetilde A_\tau ; \mathbf D )\geq 1/4-o_\beta(1)$. Furthermore, the harmonic measure from 0 in $\mathbf D$ of each of the two arcs connecting $\widetilde A_\tau$ and 1 is at least $1/4 - o_\beta(1)$.
Let $A_\tau = \widetilde\Psi_\eta^{-1 } (\widetilde A_\tau) = \eta \setminus K_\tau$.
By conformal invariance of harmonic measure we have that $\operatorname{hm}^0(\eta^{T_t} ; D_\eta)$, $\operatorname{hm}^0(\overline\eta^{\overline T_t} ; D_\eta)$, and $\operatorname{hm}^0(A_\tau; D_\eta)$ are each at least $1/4 - o_\beta(1)$.
By Lemma~\ref{phi hm arc} we have $\operatorname{dist}(0 , K_\tau) \asymp \operatorname{dist}(0 , \eta)$ and $|\Phi_\tau'(0)| \asymp |\Psi_\eta'(0)|$. Since $\widetilde\Psi_\eta(\eta(T_\tau))$ and $\widetilde\Psi_\eta(\overline\eta(\overline T_\tau))$ lie in $[i,-i]_{\partial\mathbf D}$ and removing $A_\tau$ can only increase the harmonic measure from 0 of parts of $\partial D_\eta$ outside of $A_\tau$, $\widetilde\Phi_\tau(\eta(T_\tau))$ and $\widetilde\Phi_\tau(\overline\eta(\overline T_\tau))$ must lie in $[i,-i]_{\partial\mathbf D}$.
Thus, the conditions of the lemma hold for this choice of $\tau$.
\end{proof}
\begin{figure}\label{E0 sigma fig}
\begin{center}
\includegraphics[scale=.8]{figures/drawing-sigma-lemma.pdf}
\end{center}
\caption{An illustration of the argument of Lemma~\ref{E0 event} in the case $\{T' < \infty\}$. The hull $K_{\tau_0}$ is shown in black. the curve $\eta'$ and its pre-image under $\widetilde\Phi_{\tau_0}$ are shown in blue. The extra part of the curve which we grow after growing $K_{\tau_0}$ is shown in orange.}
\end{figure}
\begin{lem} \label{E0 event}
Let $v > 0$, $\zeta > 0$, and $\mu_0\in\mathcal M$. For two times $t , \overline t> 0$, let $E_0(t, \overline t) = E_0(t, \overline t , v , \zeta , \mu_0)$ be the event that the following occurs.
\begin{enumerate}
\item $32 e^{- \beta } \leq \operatorname{dist}(z , \eta^t \cup \overline\eta^{\overline t}) \leq e^{-\beta(1-v)}$. \label{E0 dist}
\item Let $ \phi_{t,\overline t} : \mathbf D\setminus (\eta^t \cup \overline\eta^{\overline t}) \rightarrow \mathbf D$ be the conformal map which takes $x^+$ to $-i$, $y^-$ to $i$, and the midpoint $m$ of $[x,y]_{\partial\mathbf D}$ to 1. Then $ e^{-\beta(q+v)} \leq |\phi_{t,\overline t}'(0)| \leq e^{- \beta(q-v)}$.\label{E0 deriv}
\item Let $ \psi_{t,\overline t} : \mathbf D\setminus (\eta^t \cup \overline\eta^{\overline t}) \rightarrow \mathbf D$ be the conformal map which fixes 0 and takes $m$ to 1. Then $|\psi_{t,\overline t}(\eta(t )) - \psi_{t,\overline t}(\overline\eta(\overline t))| \geq \zeta$. \label{E0 separated}
\item $\mathcal G'(\eta^t \cup \overline\eta^{\overline t} , \mu_0)$ occurs. \label{E0 delta}
\end{enumerate}
There is a $\zeta>0$, a $\mu_0\in\mathcal M$, and random times $t_0 , \overline t_0$, all independent of $v$, such that
\begin{equation}\label{P(E_0)}
\mathbf P(E_0(t_0, \overline t_0) ) \succeq e^{-\beta(\gamma^*(q) + \gamma_0^*(q) v) } ,
\end{equation}
with the implicit constant independent of $\beta$ and uniform in $x,y$ with $|x-y|$ bounded below and $\gamma^*(q) , \gamma_0^*(q)$ as in Proposition~\ref{1pt at hitting}. Furthermore, we can choose $t_0$ and $\overline t_0$ in such a way that the conditional law given $\eta^{t_0} \cup \overline\eta^{\overline t_0}$ of the part of $\eta$ between $\eta(t_0)$ and $\overline\eta(\overline t_0)$ on the event $E_0(t_0 , \overline t_0)$ is that of a chordal $\operatorname{SLE}_\kappa$ from $\eta(t_0)$ to $\overline\eta(\overline t_0)$ in $\mathbf D\setminus ( \eta^{t_0} \cup \overline\eta^{\overline t_0})$.
\end{lem}
\begin{proof}
Let $s\in (-1,1)$, $v'\in (0,v)$, and $\epsilon > 0$ be chosen so that
\[
s =\frac{q}{q+1} + o_{v'}(1) , \qquad \epsilon^{1-s-2v'} \leq e^{-\beta(1-v)} ,\qquad \epsilon^{1-s+2v'} \geq 32 e^{-\beta} .
\]
Let $c >0$ and let $ \mathcal E^{s;v'}_{\epsilon}(\eta, 0 ; c)$ be the event of Section~\ref{time infty setup sec} (with $v'$ in place of $u$). Let $\Psi_{\eta} : D_{\eta} \rightarrow \mathbf D$ and $\Psi_{\eta}^- : D_{\eta}^- \rightarrow\mathbf D$ be as in that subsection. Let $\mu'\in\mathcal M$ and let
\begin{equation*}
\mathcal E := \mathcal E^{s;v'}_{\epsilon}(\eta, 0 ; c ) \cap \mathcal G(\Psi_{\eta} , \mu') \cap \mathcal G(\Psi_{\eta}^- , \mu') .
\end{equation*}
By Theorem~\ref{1pt forward}, if the parameter $\mu'$ is chosen appropriately then we have
\[
\mathbf P(\mathcal E ) \succeq e^{-\beta ( \gamma^*(q) + \gamma_0^*(q) v ) } ,
\]
for an appropriate choice of $\gamma_0^*(q)$. Lemma~\ref{G dist} implies that we can find $\mu_0\in\mathcal M$ depending only on $\mu'$ such that
\begin{equation} \label{t ol t G union}
\bigcup_{t , \overline t \geq 0} \mathcal G'(\eta^t \cup \overline\eta^{\overline t} , \mu_0) \subset \mathcal G(\Psi_{\eta} , \mu') \cap \mathcal G(\Psi_{\eta}^- , \mu') .
\end{equation}
Let $\tau_0$ be the first time $\tau$ that the first two conditions in the definition of $E_0(T_\tau , \overline T_\tau)$ are satisfied and that $\widetilde\Phi_\tau(\eta(T_\tau))$ and $\widetilde\Phi_\tau(\overline\eta(\overline T_\tau))$ (as defined just above Lemma~\ref{sigma hm converge}) both lie in $[i,-i]_{\partial\mathbf D}$. By Lemma~\ref{sigma good time} and the definition of $\mathcal E$, if $c$ is chosen sufficiently large then $\tau_0 < \infty$ a.s.\ on $\mathcal E$. Moreover, decreasing $\tau$ only increases $\operatorname{hm}^0(\eta\setminus K_{\tau } ; D_\eta)$, so on $\mathcal E$ we a.s.\ have
\begin{equation}\label{diam lower bound}
\operatorname{hm}^0(\eta\setminus K_{\tau_0} ; D_\eta) \geq 1/4 - o_\beta(1).
\end{equation}
Let $\eta' = \widetilde\Phi_{\tau_0}(\eta\setminus K_{\tau_0})$, with the parametrization it inherits from $\eta$. Observe that the conditional law of $\eta'$ given $K_{\tau_0}$ is that of a chordal $\operatorname{SLE}_\kappa$ from $x':= \widetilde\Phi_{\tau_0}(\eta(T_{\tau_0} ))$ to $y' := \widetilde\Phi_{\tau_0}(\overline\eta(\overline T_{\tau_0}))$ in $\mathbf D$ (here we used that we made $\tau_0$ the \emph{smallest} time for which our desired conditions are satisfied).
The definition of $E_0$ almost holds with $t_0 = T_{\tau_0}$ and $\overline t_0 = \overline T_{\tau_0}$, but $\widetilde\Phi_{\tau_0}(\eta(T_{\tau_0}))$ and $\widetilde\Phi_{\tau_0}(\overline\eta(\overline T_{\tau_0}))$ may be too close together. To this end, we will choose slightly larger times at which the images of the tips of $\eta$ and $\overline\eta$ are separated. Note that~\eqref{diam lower bound} implies $\operatorname{diam} \eta' \geq\zeta_0$ on $\mathcal E$ for some universal constant $\zeta_0\in (0,1/4)$.
Let $\overline\eta'$ be the time reversal of $\eta'$, with the parametrization it inherits from $\overline\eta$.
Let $T'$ (resp.\ $\overline T'$) be the first time $\eta'$ (resp.\ $\overline\eta'$) enters $B_{1-\zeta_0/4}(0)$. Let $T''$ be the first time $t$ that $\operatorname{arg} \eta'(t) \geq \operatorname{arg} x' + \zeta_0/8$. Let $\overline T''$ be the first time $\overline t$ that $\operatorname{arg} \overline \eta'(\overline t) \leq \operatorname{arg} y' - \zeta_0/8$ for some $\overline t'\leq \overline t$. Since $\operatorname{diam} \eta' \geq\zeta_0$ a.s.\ on $\mathcal E$, either $|x' - y'| \geq \zeta_0/8$ or one of $T', T'' $ or $\overline T''$ is finite on this event (if not, then $\eta'$ is contained in the wedge $\{z\in\mathbf D: \operatorname{arg} y' - \zeta_0/8 \leq \operatorname{arg} z \leq \operatorname{arg} x' + \zeta_0/8 ,\: |z| \geq 1-\zeta_0/8\}$ and this wedge has diameter $<\zeta_0$).
Hence the intersection with $\mathcal E$ of at least one of the events $\{|x' - y'| \geq \zeta_0/8\}$, $\{T' < \infty\}$, $\{T''< T'\}$, or $\{\overline T'' < \overline T'\}$ has probability at least $\frac14 \mathbf P(\mathcal E ) \succeq e^{-\beta ( \gamma^*(q) + \gamma_0^*(q) v ) }$.
It is therefore enough to show that the conclusion of the lemma is true in each of the four possible cases. We will do this by choosing $t_0$ to be one of $T_{\tau_0} , T', $ or $T''$ and $\overline t_0$ to be one of $\overline T_{\tau_0} , \overline T'$, or $\overline T''$. It is clear that the last statement of the lemma holds for any such choice. Furthermore, for any such choice, $(\eta')^{t_0} \cup (\overline\eta')^{\overline t_0}$ lies at uniformly positive distance from the segment $[0,1]$ on the event $\mathcal E$. Hence Lemma~\ref{phi hm} and conformal invariance of harmonic measure imply that condition~\ref{E0 deriv} in the definition of $E_0$ holds for any such choice. Clearly, condition~\ref{E0 dist} in the definition of $E_0$ holds a.s.\ on $\mathcal E$ for any such choice. By~\eqref{t ol t G union}, condition~\ref{E0 delta} holds for any such choice.
Therefore we just need to verify that condition~\ref{E0 separated} holds in each of the five cases.
\begin{enumerate}
\item If $\mathbf P( |x' - y'| \geq \zeta_0/8 ,\:\mathcal E) \succeq e^{-\beta ( \gamma^*(q) + \gamma_0^*(q) v ) }$ then we can just set $t_0 = T_{\tau_0}$, $\overline t_0 = \overline T_{\tau_0}$, and $\zeta = \zeta_0/8$.
\item If $\mathbf P(T' < \infty ,\:\mathcal E) \succeq e^{-\beta ( \gamma^*(q) + \gamma_0^*(q) v ) }$ then we set $t_0 = T'$ and $\overline t_0 = \overline T_{\tau_0}$. A Brownian motion has probability at least a constant $\zeta >0$ depending only on $\zeta_0$ to exit $B_{1-\zeta_0/16}(0)$ within distance $\zeta_0/4$ of 1 and then make a counterclockwise loop around the origin before leaving $\mathbf D\setminus B_{1-\zeta_0/16}(0)$. In this case it necessarily exits $\mathbf D\setminus (\eta')^{T'}$ on the left side of $(\eta')^{T'}$. See Figure~\ref{E0 sigma fig} for an illustration in this case.
\item If $\mathbf P(T'' < T' ,\:\mathcal E) \succeq e^{-\beta ( \gamma^*(q) + \gamma_0^*(q) v ) }$ then we set $t_0 = T'\wedge T''$ and $\overline t_0 = \overline T_{\tau_0}$. A Brownian motion has probability at least a constant $\zeta >0$ depending only on $\zeta_0$ to exit $\mathbf D$ before hitting any point outside of $\mathbf D\setminus B_{1-\zeta_0/8}(0)$ whose argument is not between $\operatorname{arg} x'$ and $\operatorname{arg} x' + \zeta_0/8$. If this is the case and $T' \leq T''$, then a Brownian motion necessarily exits $\mathbf D\setminus (\eta')^{T' \wedge T''}$ on the left side of $(\eta')^{T'\wedge T''}$.
\item The case for $\{\overline T'' < \overline T'\}$ is treated in the same manner as the case for $\{T'' < T'\}$.
\end{enumerate}
Thus, we have exhausted all possible cases and we conclude that the statement of the lemma holds.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{1pt at hitting}, lower bound]
Suppose $\zeta$, $\mu_0$, $t_0$, and $\overline t_0$ are chosen so that the conclusion of Lemma~\ref{E0 event} holds. Let $E_0 = E_0(t_0 , \overline t_0 , \zeta , \mu_0)$ be as in that lemma.
Let $\beta_0 = -\log \operatorname{dist}(0 , \eta^{t_0} \cup \overline\eta^{\overline t_0} )$. Note
\[
\beta(1-v) \leq \beta_0 \leq \beta - \log 32 .
\]
Fix $r \in ( \log16 , \log 32)$.
Let $\eta_1$ be the image under $\psi_{t_0 , \overline t_0}$ (defined as in Lemma~\ref{E0 event}) of the part of $\eta$ between $\eta(t_0)$ and $\overline\eta(\overline t_0)$. Let $x_1$ and $y_1$ be its endpoints. Let $\tau_1'$ (resp.\ $\overline\tau_1'$) be the first time $\eta_1$ (resp.\ $\overline\eta_1$) hits $\psi_{t_0 , \overline t_0}(B_{\beta_0 + r})$. Let $G_1$ be the event that
\begin{enumerate}
\item $|\eta_1(\tau_1') - \overline\eta_1(\overline\tau_1')| \geq (1/32) e^{-r}$.
\item $\eta_1^{\tau_1'} \cup \overline\eta_1^{\overline\tau_1'} \subset \psi_{t_0 , \overline t_0}(B_1)$.
\item $\eta_1^{\tau_1'} \cup \overline\eta_1^{\overline\tau_1'}$ is disjoint from the $\zeta/2$-neighborhood of the segment connecting 0 and the midpoint of the shorter arc between $x_1$ and $y_1$.
\end{enumerate}
By the Koebe quarter theorem we have
\[
B_{r+\log 16} \subset \psi_{t_0 , \overline t_0}(B_{ \beta_0 + r}) \subset B_{r - \log 16} .
\]
Hence by \cite[Lemma~2.3]{miller-wu-dim}, condition~\ref{E0 separated} in the definition of $E_0$, and the last statement of Lemma~\ref{E0 event} we have that $\mathbf P(G_1 | E_0)$ is at least a $\beta$-independent positive constant.
For $k = 1,2,3, \ldots$, let $\widetilde\psi_k$ be the map from $\mathbf D\setminus ( \eta^{\tau_{\beta_0 + k r}} \cup \overline\eta^{\overline\tau_{\beta_0 + k r}} )$ to $\mathbf D$ with $\widetilde\psi_k(0) = 0$ and $\widetilde{\psi}_k'(0)>0$.
For $k\geq 2$, let $ \eta_k$ be the image under $\widetilde\psi_{k-1}$ of the part of $\eta$ which lies between $\eta(\tau_{\beta_0 +(k-1)r})$ and $\overline\eta(\overline\tau_{\beta_0 + (k-1) r})$. The law of $\eta_k$ given $\mathcal F_{\beta_0 + (k-1) r }$ is that of a chordal $\operatorname{SLE}_\kappa$ from $x_k := \widetilde\psi_{k-1}( \eta(\tau_{\beta_0 + (k-1) r}) )$ to $y_k:=\widetilde\psi_{k-1}(\overline\eta(\overline\tau_{\beta_0 + (k-1) r}))$.
Let $\overline\eta_k$ be the time reversal of $\eta_k$.
Let $\tau_k'$ and $\overline\tau_k'$ be the hitting times of $\widetilde\psi_{k-1}( B_{\beta_0 +k r} )$ by $\eta_k$ and $\overline\eta_k$, respectively. Fix $\delta>0$ and for $k\geq 1$ let $G_k$ be the event that $\eta^{\tau_k}$ (resp.\ $\overline\eta^{\overline\tau_k}$) is contained in the $\delta$-neighborhood of the segment $[x_k , 0]$ (resp.\ $[y_k , 0]$).
By the Koebe quarter theorem, whenever $\widetilde\psi_{k-1}$ is defined we have
\[
B_{r+\log 16} \subset \widetilde\psi_{k-1}(B_{ \beta_0 +kr}) \subset B_{r - \log 16} .
\]
By conformal invariance of harmonic measure, on $G_{k-1} $ for $k\geq 2$, $| x_k - y_k|$ is at least a universal constant provided $\delta$ is taken sufficiently small.
It now follows from from \cite[Lemma~2.3]{miller-wu-dim} that for each $k\geq 2$, $\mathbf P(G_k | G_{k-1}) \geq p$ for some $p>0$ which depends only on $\delta$.
Let $k_*$ be the least integer $k$ such that $k r + \beta_0 \geq \beta$. Note that $k_* \leq \beta v/r$. Let
\[
G^* := \bigcap_{k= 1}^{k_*} G_k .
\]
It is clear that on the event $E_0 \cap G^*$, conditions~\ref{E hit},~\ref{E top}, and~\ref{E G} in the definition of $E $ hold provided we take $\delta$ sufficiently small, depending on $a$.
It remains to deal with condition~\ref{E phi}. For $k\geq 1$, let $\widehat\eta_k$ be the curve obtained by connecting $\eta(\tau^*_{\beta_0 +k r})$ and $\overline\eta(\overline\tau^*_{\beta_0 + k r})$ via the arc of $ B_{\beta_0 +k r}^* $ which does not disconnect 0 from $[x_* , y_*]_{\partial\mathbf D}$. Let $\Psi_{\widehat \eta_k}$ be the conformal map from the connected component of $\mathbf D\setminus \widehat\eta_k$ containing $[x_* , y_*]_{\partial\mathbf D}$ on its boundary to $\mathbf D$ which takes $x_*$ to $-i$, $y_*$ to $i$, and the midpoint of $[x_* , y_*]_{\partial\mathbf D}$ to 1. By Lemma~\ref{phi hm arc}, we have
\[
C^{-1} |\Psi_{\widehat \eta_k}'(0)| \leq |\phi_{\beta'}'(0)| \leq C |\Psi_{\widehat \eta_k}'(0)| ,\qquad\forall \beta' \in [\beta_0 + (k-1) r , \beta_0 + kr] ,\qquad \forall k\geq 2 .
\]
on $G^*$, for some deterministic $C>0$ depending only on $a$, $r$, and $\mu$. A similar statement holds for $k=1$ provided we replace $C$ with a constant $C_1>0$ which is allowed to depend on $\zeta$ (and hence on $v$), but not $\beta$.
In particular, on $G^*$ we have
\[
C_1^{-1} C^{- \beta v/r } e^{-\beta (q+ v )} \leq |\phi_{\beta }'(0)| \leq C_1^{-1} C^{ \beta v/r }e^{-\beta (q- v )} .
\]
If we choose $v$ such that $v \leq u/2$ and $C^{ v/r} \leq e^{u/2}$ and $c_1$ sufficiently small, then condition~\ref{E phi} in the definition of $E $ holds on $E_0\cap G^*$.
By Lemma~\ref{E0 event} and our choice of parameters above we then have
\[
\mathbf P(E ) \geq \mathbf P(G_1 |E_0) p^{k_*-1} e^{-\beta(\gamma^*(q) + v) } \succeq e^{-\beta (\gamma^*(q) + \gamma_0^*(q) u)} .
\]
\end{proof}
\subsection{Events for the perfect points}
\label{perfect setup sec}
In this subsection we will define events $E_{z,j}$ for each $z\in \mathbf D$ and $j\in \mathbf N$ which we will eventually use to construct subsets of $\Theta^s(D_\eta)$ and $\widetilde\Theta^s(D_\eta)$ (called the ``perfect points") whose Hausdorff dimension can be bounded below. The definition of the events $E_{z,j}$ involves a number of auxiliary objects which we list below.
Let $\chi = 2/\sqrt\kappa - \sqrt\kappa/2$ and let $\lambda = \pi/\sqrt\kappa$ be as in~\eqref{chi lambda}. Let $h$ be a zero boundary GFF on $\mathbf D$ plus a harmonic function chosen in such a way that if $\psi : \mathbf H\rightarrow \mathbf D$ is the conformal map taking $0$ to $-i$, $\infty$ to $i$, and $0$ to $i$, then $h\circ\psi - \chi\operatorname{arg}\psi'$ is a GFF on $\mathbf H$ with boundary data $-\lambda$ on $(-\infty ,0 ]$ and $\lambda$ on $[0,\infty)$. By \cite[Theorem~1.1]{ig1} the zero-angle flow line $\eta$ of $h$ started from $-i$ is a chordal $\operatorname{SLE}_\kappa$ from $-i$ to~$i$ in~$\mathbf D$.\footnote{In the case $\kappa = 4$, we replace flow lines of $h$ with a given angle by level lines of $h$ at a given level (see \cite{ss-dgff,ss-contour,wang-wu-level-lines}). Everything that follows works identically with this replacement. In fact, since (in contrast to the situation for flow lines) the time reversal of a level line is also a level line \cite[Theorem~1.1.5]{wang-wu-level-lines}, some of the proofs are easier for $\kappa=4$. }
Fix $ \Delta > \widetilde\Delta > 0$, $q \in (-1/2,\infty)$, $ \dot\Delta, c , a , \theta , \delta_0 , r , \dot r , p_L >0$, and $\mu , \mu_L , \mu_F \in\mathcal M$. Assume that the parameters $a$ and $\mu$ are chosen in such a way that the conclusion of Proposition~\ref{1pt at hitting} holds. Also fix sequences $\beta_j \rightarrow \infty$ and $u_j \rightarrow 0$, which we will choose later.
For each $z\in\mathbf D$ and $j\in\mathbf N$, we will inductively define the following objects.
\begin{itemize}
\item Events $L_{z,j}$, $\widetilde E_{z,j}$, $F_{z,j}$, $ E_{z,j}$.
\item Points $x_{z,j}, y_{z,j} , x_{z,j}^* , y_{z,j}^*, x_{z,1}^F , y_{z,1}^F, b_{z,j} , \overline b_{z,j}$.
\item Conformal maps $\phi_{z,j}, p_{z,j} , \widehat p_{z,j} , \psi_{z,j} , \psi_{z,j}^F$.
\item Random times $\sigma_{z,j} , \overline\sigma_{z,j}$, $\tau_{z,j}$, $ \overline \tau_{j,z}$, $\tau_{z,1}^*$, $\overline\tau_{z,j}^*$, $T_{z,j}$, $\overline T_{z,j}$, $T_{z,j}^*$, $\overline T_{z,j}^*$, $t_{z,j}^\pm$, and $\widetilde t_{z,j}^\pm$.
\item Curves $\eta_{z,j}$, $\widetilde \eta_{z,j}$, $\eta_{z,j}^\pm$, and $\widehat\eta_{z,j}^\pm$.
\item Domains $D_{z,j} $ and $\widehat D_{z,j}$.
\item $\sigma$-algebras $\mathcal F_{z,j}^0$ and $\mathcal F_{z,j}$.
\end{itemize}
First consider the case $j=1$. Let $f_{z,1}$ be the conformal automorphism of $\mathbf D$ satisfying $f_{z,1}(z) = 0$ and $f_{z,1}' (0) >0$. Let $x_{z,1} = x_{z,1}^* = f_{z,1}(-i)$ and $y_{z,1} = y_{z,1}^* = f_{z,1}(i)$. Let $\eta_{z,1}^0 = f_{z,1}(\eta)$.
Let $\sigma_{z,1}$ be the first time $\eta_{z,1}^0$ hits $B_{\Delta }$. For $\alpha \geq 0$ let $\overline t_\alpha$ be the first time $\overline\eta_{z,1}$ hits $B_\alpha$. Let $\psi_{z,1}^\alpha$ be the conformal map from the connected component of $ \mathbf D\setminus( (\eta_{z,1}^0)^{\sigma_{z,1}} \cup (\overline\eta_{z,1}^0)^{\overline t_\alpha})$ containing the origin to $\mathbf D$ which fixes 0 and takes $ \eta_{z,1}^0(\sigma_{z,1})$ to $-i$.
\footnote{This connected component is all of $ \mathbf D\setminus( (\eta_{z,1}^0)^{\sigma_{z,1}} \cup (\overline\eta_{z,1}^0)^{\overline t_\alpha}) $ in the case $j=1$, but may not be in the case $j \geq 2$ since then $\eta_{z,j}^0$ may hit $\partial\mathbf D$.}
Let $\overline \sigma_{z,1} = \overline t_\alpha$ for the least $\alpha \geq \Delta$ such that $\psi_{z,1}^\alpha(\overline\eta_{z,1}^0(\overline t_\alpha)) = i$, provided such an $\alpha$ exists, and $\overline\sigma_{z,1} = \infty$ otherwise. Let $\psi_{z,1} = \psi_{z,1}^\alpha$ for this $\alpha$ if such an $\alpha$ exists and let $\psi_{z,1} = \operatorname{Id}$ otherwise.
Let $L_{z,1}$ be the event that the following occurs.
\begin{enumerate}
\item $(\eta_{z,1}^0)^{\sigma_{z,1}}$ (resp.\ the part of $ (\overline\eta_{z,1}^0)^{\overline\sigma_{z,1}}$ traced before it hits $B_{\Delta}$) is contained in the $e^{-2\Delta}$-neighborhood of the segment $[x_{z,1} , 0]$ (resp.\ $[y_{z,1} ,0]$). \label{L hit}
\item $\eta_{z,1}^0$ and $\overline\eta_{z,1}^0$ do not leave $B_{\Delta_0}$ after hitting $B_{\Delta /2}$. \label{L stay}
\item $\overline\sigma_{z,1} \leq \overline t_{\Delta + \log 2}$ and $\overline\eta_{z,1}^0([\overline t_\Delta , \overline\sigma_{z,1}] ) \subset B_{\Delta/2}$. \label{L leq}
\item $\psi_{z,1}^{-1}$ takes $B_{1-\mu(\delta_0)}(0) \cup B_{\delta_0}(-i) \cup B_{\delta_0}(i)$ into $B_{ \widetilde\Delta } $. \label{L ball}
\item $e^{-\dot\Delta + \dot r/4} \leq |\psi_{z,1}'(0)| \leq e^{-\dot\Delta + \dot r/2}$. \label{L deriv}
\item $\mathcal G_{[x_{z,1}^* , y_{z,1}^*]}( \psi_{z,1} , \mu_L )$ occurs (Definition~\ref{G infty def}). \label{L G}
\item The conditional probability given $(\eta_{z,j}^0)^{\sigma_{z,j}} \cup (\overline\eta_{z,j}^0)^{\overline \sigma_{z,j}} $ that the part of $\eta_{z,1}^0$ lying between $\eta_{z,1}^0(\sigma_{z,1} )$ and $\overline\eta_{z,1}^0(\overline\sigma_{z, 1} )$ never exits $B_{\widetilde\Delta }$ is at least $p_L$. \label{L cond}
\end{enumerate}
Let $\eta_{z,1}$ be the image under $\psi_{z,1}$ of the part of $\eta_{z,1}^0$ lying between $\eta_{z,1}^0(\sigma_{z,1} )$ and $\overline\eta_{z,1}^0(\overline\sigma_{z, 1} )$. Let $\mathcal F_{z,1}^0$ be the $\sigma$-algebra generated by $\eta_{z,1}^0|_{[0, \sigma_{z,1} ]}$ and $\overline\eta_{z,1}^0|_{[0, \overline\sigma_{z,1} ]}$.
\begin{figure}
\begin{center}
\includegraphics[scale=.7]{figures/drawing-L-event.pdf}
\caption{An illustration of the the parts of the curve $\eta_{z,1}^0$ (left) and $\eta_{z,1}$ (right) associated with the events $L_{z,1}$ and $\widetilde E_{z,1} $. For clarity, the disks here are shown larger than they actually are in practice. The same is true in the other figures in this section.} \label{L fig}
\end{center}
\end{figure}
\begin{remark} \label{L motivation}
The reason for introducing $\eta_{z,1}$ instead of working directly with $\eta_{z,1}^0$ is that we need the laws of the curves at each stage in our construction to be \hyperref[smac]{s.m.a.c.}\ (Definition~\ref{smac}). Otherwise we will end up with additional proportionality constants in our probability estimates. When we iterate these estimates several times, the proportionality constants will produce an exponential factor which will have a significant impact on our final estimates. The reason for condition~\ref{L ball} in the definition of $L_{z,1}$ is that it implies $\psi_{z,1}^{-1}(\eta_{z,1}) \subset B_{\widetilde\Delta }$ provided condition~\ref{E G} in the definition of $E_{\beta_1}^{q;u_1}(\eta_{z,1} , c,a,\mu)$ holds.
\end{remark}
\begin{remark} \label{L motivation cond}
By \cite[Lemma~2.3]{miller-wu-dim}, reversibility, and the Markov property, condition~\ref{L cond} in the definition of $L_{z,1}$ (for sufficiently small $p_L$) follows from the other conditions in the case of an ordinary $\operatorname{SLE}_\kappa$. However, when we define the events $L_{z,j}$ for $j\geq 2$, the curve $\eta_{z,j}^0$ will be an $\operatorname{SLE}_\kappa(\rho^L ; \rho^R)$ for certain $\rho^L , \rho^R \in (-2,0)$, in which case condition~\ref{L cond} (with $j$ in place of 1) is non-trivial. The reason for introducing condition~\ref{L cond} is as follows. By Lemma~\ref{rho abs cont}, we can estimate the law of the remainder of the middle part of $\eta_{z,j}^0$ given $(\eta_{z,1}^0)^{\sigma_{z,1}} \cup (\overline\eta_{z,1}^0)^{\overline \sigma_{z,1}}$ \emph{conditioned} on the event that this curve does not leave $B_\Delta$. However, we want to estimate the law \emph{restricted} to the event that the curve does not leave $B_\Delta$. For this, we need the probability that it leaves $B_\Delta$ to be $\asymp 1$.
\end{remark}
\begin{lem}\label{L prob}
For any $\widetilde\Delta > 0$, $\delta_0 > 0$, and $\mu \in \mathcal M$,
and $\dot r > 0$,
we can find $\Delta >0$, and
and $\dot\Delta > 0$
depending only on $\Delta$, $\mu_L \in\mathcal M$ depending only on $\mu$, and $p_L > 0$ depending only on the other parameters in the definition of $L_{z,1}$, all depending uniformly on $z$ in compacts, such that $\mathbf P(L_{z,1} ) \succeq 1$, with the implicit constant uniform for $z$ in compacts.
\end{lem}
\begin{proof}
This follows from \cite[Lemma~2.3]{miller-wu-dim}. Note that we apply the Koebe growth theorem to $\psi_{z,1}^{-1}$ in order to make $\widetilde\Delta $ as large as we like.
\end{proof}
Let
\[
\widetilde E_{z,1} = L_{z,1} \cap E_{\beta_1}^{q;u_1}(\eta_{z,1} , c, a , \mu ) ,
\]
with the latter event as in Section~\ref{2pt setup sec}. Let $\tau_{z,1} $, and $\overline\tau_{z,1}$ be the stopping times $\tau_{\beta_1}$ and $\overline\tau_{\beta_1}$ from Section~\ref{2pt setup sec}. Let $\phi_{z,1} : \mathbf D\setminus (\eta_{z,1}^{\tau_{z,1} } \cup \overline\eta_{z,1}^{\overline\tau_{z,1} }) \to {\BB D}$ be the conformal map $\phi_{\beta_1}$ from Section~\ref{2pt setup sec}. Let $T_{z,1}$ be the time for $\eta$ corresponding to $\tau_{z,1} $.
Let $\widehat\eta_{z,1}^-$ and $\widehat\eta_{z,1}^+$ be the flow lines of $h $ started from $\eta(T_{z,1})$ with angles $\theta$ and $-\theta$, respectively.\footnote{Since $T_{z,1}$ is not a stopping time for $\eta$, we define the flow lines $\widehat\eta_{z,1}^\pm$ by taking the limit of the flow lines of angles $\mp\theta$ of $h$ started from $\eta(t)$ as $t$ increases to $T_{z,1}$ along rational times. The limiting object can equivalently be defined as the left or right outer boundary of a certain counterflow line of $h$ started from $-i$ and stopped at the first time it hits $\eta(T_{z,1})$ (c.f.\ \cite[Section~5]{ig1}), so in particular is a simple curve.}
Note that the flow line with a negative sign has positive angle and vice versa. This is because a flow line with a negative angle a.s.\ stays to the right of $\eta$, and a flow line with a positive angle a.s.\ stays to the left of $\eta$. See \cite[Theorem~1.5]{ig1}. Let $\eta_{z,1}^\pm = (\psi_{z,1} \circ f_{z,1})(\widehat\eta_{z,1}^\pm)$.
By the results of \cite[Section~7]{ig1}, the conditional law of $\widehat \eta_{z,1}^+$ (resp.\ $\widehat \eta_{z,1}^-$) given $\eta$ is that of a $\operatorname{SLE}_\kappa(\rho^0 ; \rho^1)$ processes from $\eta(T_{z,1})$ to $i$ in the right (resp.\ left) connected component of $\mathbf D\setminus \eta$, where
\begin{equation} \label{rho0 rho1}
\rho^0 = -\frac{\theta\chi}{\lambda} ,\qquad \rho^1 = \frac{\theta \chi}{\lambda}-2
\end{equation}
and the force points are located immediately to the left and right of $\eta(T_{z,1})$.
By \cite[Remark 5.3]{ig1}, $\widehat\eta_{z,1}^\pm$ a.s.\ fail to hit $\eta^{T_{z,1}} \cup \partial\mathbf D$ provided $ -\theta \chi/\lambda \geq \kappa/2-2$.
Furthermore, $ \eta_{z,1}^\pm$ a.s.\ intersect (but do not cross) each other provided $2\theta\chi/\lambda - 2 < \kappa/2-2$.
Since $\kappa \leq 4$ we can choose a small $\theta > 0$ depending only on $\kappa$ in such a way that $ \eta_{z,1}^\pm$ a.s.\ do not intersect $\partial\mathbf D $ but do a.s.\ intersect each other. We henceforth assume that $\theta$ has been chosen in this manner.
Let $\widehat D_{z,1}$ be the connected component of $\mathbf D\setminus (\widehat\eta_{z,1}^+ \cup \widehat\eta_{z,1}^-)$ which contains $z$. Let $D_{z,1}$ be the connected component of $\mathbf D\setminus ( \eta_{z,1}^+ \cup \eta_{z,1}^-)$ which contains the origin. Let $p_{z,1} : D_{z,1} \rightarrow \mathbf D$ be the conformal map with $p_{z,1}(0) = 0$ and $p_{z,1}'(0) > 0$. Let $\widehat p_{z,1} = p_{z,1} \circ \psi_{z,1} \circ f_{z,1}$, so that $\widehat p_{z,1} : \widehat D_{z,1} \rightarrow \mathbf D$ and takes $z$ to 0. See Figure~\ref{flow lines fig1} for an illustration of the event $\widetilde E_{z,1}$ and the flow lines $\widehat\eta_{z,1}^\pm$.
\begin{figure}
\begin{center}
\includegraphics[scale=.7]{figures/drawing-flow-lines1.pdf}
\caption{An illustration of the curve $\eta$ grown up to its hitting time of a quasi-disk centered at $z$ on the event $ E_{z,1} $, together with the flow lines $\widehat\eta_{z,1}^\pm$ (shown in green) used in the definition of the event $F_{z,1}$. The part of the curve $\eta$ associated with the event $L_{z,1}$ is shown in purple. The domain $\widehat D_{z,1}$ is the complementary connected component of the green flow lines which contains $z$. Also shown are the segments of $\eta$ involved in the last part of the event $E_{z,1}$ (in orange) and the stopping times $T_{z,1}^*$ and $\overline T_{z,1}^*$ defined just after the definition of $E_{z,1}$. } \label{flow lines fig1}
\end{center}
\end{figure}
Let $t_{z,1}^+$ be the first time that $\eta_{z,1}^+$ hits $\eta_{z,1}^-$ after the first time it exits the disk of radius $ e^{-\beta_1 - \dot r}$ centered at $\eta_{z,1}(\tau_{z,1} )$. Let $t_{z,1}^-$ be the time $t$ such that $\eta_{z,1}^-(t) = \eta_{z,1}^+(t_{z,1}^+)$. Let $\overline b_{z,1} = \eta_{z,1}^\pm (t_{z,1}^\pm)$ and let $b_{z,1}$ be the last intersection point of $\eta_{z,1}^\pm$ before $\overline b_{z,1}$. Also let $\widetilde t_{z,1}^\pm$ be the first exit times of $\eta_{z,1}^\pm$ from the annulus $B_{\beta_1 - \log 2} \setminus B_{\beta_1 + \log 2}$.
Let $F_{z,1}$ be the event that the following occurs.
\begin{enumerate}
\item $t_{z,1}^+ \leq \widetilde t_{z,1}^+$ and $t_{z,1}^- \leq \widetilde t_{z,1}^-$. \label{F contained}
\item $|b_{z,1} | \leq e^{-\beta_1 - \dot r}$ and $\overline b_{z,1} \notin \overline\eta_{z,1}^{\overline\tau_{z,1}}$. \label{F b}
\item Let $\psi_{z,1}^F : \mathbf D\setminus (\eta_{z,1}^{\tau_{z,1}} \cup \overline\eta_{z,1}^{\overline\tau_{z,1}})$ be the conformal map with $\psi_{z,1}^F(0) = 0$ and $(\psi_{z,1}^F)'(0) > 0$. Let $x_{z,1}^F = \psi_{z,1}^F( \eta_{z,1}( \tau_{z,1}))$ and $y_{z,1}^F = \psi_{z,1}^F(\overline\eta_{z,1}(\overline\tau_{z,1}))$. Then $|\psi_{z,1}^F( b_{z,1}) - x_{z,1}^F|$ and $|\psi_{z,1}^F(\overline b_{z,1}) - y_{z,1}^F|$ are each at most $r$. \label{F b bar}
\item Each point of $\psi_{z,1}^F( (\eta_{z,1}^+)^{t_{z,1}^+})$ (resp.\ $\psi_{z,1}^F((\overline\eta_{z,1}^+)^{\overline t_{z,1}^+})$) lies within distance $r$ of $[x_{z,1}^F , y_{z,1}^F]_{\partial\mathbf D}$ (resp.\ $[y_{z,1}^F , x_{z,1}^F]_{\partial\mathbf D}$). \label{F hm}
\item $\mathcal G'( \psi_{z,1}^F( (\eta_{z,1}^+)^{t_{z,1}^+} \cup (\eta_{z,1}^-)^{t_{z,1}^-}) , \mu_F)$ occurs. \label{F G}
\end{enumerate}
See Figure~\ref{F fig} for an illustration of the event $F_{z,1}$.
It follows from condition~\ref{F b bar} that $D_{z,1}$ is the ``pocket" formed by $\eta_{z,1}^\pm$ between their hitting times of $b_{z,1}$ and $\overline b_{z,1}$ on the event $F_{z,1}$.
\begin{figure}
\begin{center}
\includegraphics[scale=.7]{figures/drawing-flow-lines2.pdf}
\caption{An illustration of the event $F_{z,1}$ that $\widetilde E_{z,1} $ occurs and the flow lines $\eta_{z,1}^\pm$ started at $\eta_{z,1}(\tau_{z,1})$ behave in the manner we would like. The domain $D_{z,1}$ is that which lies between the parts of $\eta_{z,1}^\pm$ traced between their hitting times of $b_{z,1}$ and $\overline b_{z,1}$. } \label{F fig}
\end{center}
\end{figure}
\begin{remark}
Our reason for introducing the auxiliary flow lines $\eta_{z,1}^\pm$ is as follows. The conditional law of the part of $\eta$ which lies inside $\widehat D_{z,1}$ is conditionally independent of the part of $\eta$ which lies outside $\widehat D_{z,1} $ given the flow lines $\eta_{z,1}^\pm$ (see Lemma~\ref{F cond} below). When applied at various scales, this fact this will eventually allow us to get the needed ``near independence" of the events $E_{z,j}$ and $E_{w,j}$ for $z\not=w$.
\end{remark}
\begin{lem} \label{F prob}
Given $r>0$ we can choose $\mu_F$ and $\dot r$ independently of $\beta_1$ and $u_1$ and uniform for $z$ in compacts in such a way that $\mathbf P(F_{z,1} \,|\, \widetilde E_{z,1} ) \asymp 1$ with the implicit constants depending on the other parameters but not on $\beta_1$, $u_1$ or $z$.
\end{lem}
\begin{proof}
Let $ \eta_{z,1}^F$ be the image under $\psi^F_{z,1}$ of the part of $\eta_{z,1}$ between $\eta_{z,1}(\tau_{z,1} )$ and $\overline\eta_{z,1}(\overline\tau_{z,1} )$. Note that the distance between the endpoints $ x_{z,1}^F$ and $y_{z,1}^F $ of $\eta_{z,1}^F$ is uniformly positive on $\widetilde E_{z,1} $ by condition~\ref{E top} in the definition of $E_{\beta_1}^{q;u_1}(\cdot)$.
Let $\widetilde r \in (0,r)$ and let $U$ be the $\widetilde r$-neighborhood of the line segment from $ x^F_{z,1}$ to $ y^F_{z,1}$. Let $\mu_F' \in \mathcal M$. Let $S_{z,1}$ be the event that $ \eta_{z,1}^F \subset U$ and $\mathcal G'( \eta_{z,1}^F , \mu_F')$ occurs.
By the Markov property and reversibility of SLE, the conditional law of $ \eta_{z,1}^F$ given a realization of $\mathcal F_{z,1}^0$ for which $L_{z,1}$ and a realization of $\eta_{z,1}^{\tau_{z,1}} \cup \overline\eta_{z,1}^{\overline\tau_{z,1}}$ is that of a chordal $\operatorname{SLE}_\kappa$ from $ x_{z,1}^F$ to $ y_{z,1}^F$ in $\mathbf D$. By \cite[Lemma~2.3]{miller-wu-dim}, we thus have $\mathbf P(S_{z,1} \,|\, \widetilde E_{z,1}) \succeq 1$ provided $ \mu_F'$ is chosen sufficiently small, independently of $\beta_1$, $u_1$, and $z$.
The conditional law of $\psi^F_{z,1}( \eta_{z,1}^+ )$ given $\eta_{z,1}^0$ is that of a $\operatorname{SLE}_\kappa(\rho^0 ; \rho^1)$ process in the right connected component of $\mathbf D\setminus \eta_{z,1}^F$ from $ x_{z,1}^F$ to $\psi^F_{z,1}(i^-)$ with force points located on either side of $ x^F_{z,1}$, where $\rho^0$ and $\rho^1$ are as in~\eqref{rho0 rho1}.
Our choice of $\theta$ implies that such a process a.s.\ does not hit $[x_{z,1}^F , \psi^F_{z,1}(i^-)]_{\partial\mathbf D}$. Condition~\ref{E top} in the definition of $E_{\beta_1}^{q;u_1}(\cdot)$ implies that $\psi^F_{z,1}(i^-)$ lies at uniformly positive distance from $ x_{z,1}^F$ and $y^F_{z,1}$ on $\widetilde E_{z,1}$.
Similar statements hold with $-$ in place of $+$ and ``left" in place of ``right". By \cite[Lemma~2.5]{miller-wu-dim} (and a straightforward complex analysis argument to make a suitable choice of $\dot r$), we have $\mathbf P(F_{z,1} \,| \, \widetilde E_{z,1} \cap S_{z,1} ) \succeq 1$ provided $\widetilde r$, $\mu_F$, and $\dot r$ are chosen sufficiently small, independently of $\beta_1$, $u_1$, and $z$.
\footnote{To get that the flow lines $ \eta_{z,1}^\pm$ intersect one another where we want them to with uniformly positive probability, we can further condition on a second pair of flow lines $\widetilde\eta_{z,1}^\pm$ with the same angles as $\eta_{z,1}^\pm$, started at a point near where we want the intersection to occur. We then apply \cite[Lemma~2.5]{miller-wu-dim} to the conditional law of $\eta_{z,1}^\pm$ given $\widetilde\eta_{z,1}^\pm$ and $\eta$, and observe that $\eta_{z,1}^\pm$ merge with $\widetilde\eta_{z,1}^\pm$ upon intersecting; and that $\widetilde\eta_{z,1}^\pm$ a.s.\ intersect one another at points arbitrarily close to their starting points. See \cite{miller-wu-dim} for several examples of similar arguments.}
We conclude by observing that
\[
\mathbf P(F_{z,1}\,|\, \widetilde E_{z,1}) \geq \mathbf P(F_{z,1} \cap S_{z,1} \,|\, \widetilde E_{z,1}) = \mathbf P(F_{z,1}\,|\, \widetilde E_{z,1} \cap S_{z,1}) \mathbf P(S_{z,1} \,|\, \widetilde E_{z,1}) .
\]
\end{proof}
Let $ \tau_{z,1}^*$ (resp.\ $\overline\tau_{z,1}^*$) be the time $ \eta_{z,1}$ (resp.\ $\overline\eta_{z,1}$) hits $ b_{z,1}$ (resp.\ $\overline b_{z,1}$). Note that these times are a.s.\ finite since $\eta_{z,1}$ a.s.\ lies between $\eta_{z,1}^\pm$). Let $E_{z,1} $ be the event that the following occurs.
\begin{enumerate}
\item $\widetilde E_{z,1} \cap F_{z,1}$ occurs. \label{E_z F}
\item $ \psi_{z,1}^F ( \eta_{z,1}([\tau_{z,1} , \tau_{z,1}^*]) )$ (resp.\ $ \psi_{z,1}^F(\overline\eta_{z,1}( [\overline \tau_{z,1} , \overline\tau_{z,1}^*])$) is contained in the disk of radius $2r$ centered at $x_{z,1}^F$ (resp.\ $y_{z,1}^F$) (notation as in condition~\ref{F b bar} in the definition of $F_{z,1}$). \label{E_z stay}
\end{enumerate}
\begin{remark}\label{E stay remark}
By \cite[Theorem~1.5]{ig1} $\eta_{z,1}$ cannot cross $\eta_{z,1}^\pm$. By combining this with condition~\ref{L ball} in the definition of $L_{z,1}$, condition~\ref{E G} in the definition of $E_{\beta_1}^{q;u_1}(\cdot)$, and condition~\ref{E_z stay} in the definition of $E_{z,1}$, it follows that the whole of $\psi_{z,1}^{-1}(\eta_{z,1})$ (i.e. the part of $\eta_{z,1}^0$ between $\eta_{z,1}^0(\sigma_{z,1})$ and $\overline\eta_{z,1}^0(\overline\sigma_{z,1})$) is contained in $B_{\widetilde\Delta }$ on the event $E_{z,1}$.
\end{remark}
Let $T_{z,1}^*$ and $\overline T_{z,1}^*$, resp., be the times for $\eta$ and $\overline \eta$, resp., corresponding to $\tau_{z,1}^*$ and $\overline\tau_{z,1}^*$.
Let $\mathcal F_{z,1}$ be the $\sigma$-algebra generated by $\eta |_{[0, T_{z,1}^*]}$, $\overline\eta |_{[0, \overline T_{z,1}^*]}$, and $\widehat \eta_{z,1}^\pm|_{[0, t_{z,1}^\pm]}$.
\begin{lem} \label{E prob}
We have $\mathbf P(E_{z,1} \, |\, \widetilde E_{z,1} \cap F_{z,1}) \asymp 1 $ with the implicit constant depending on the other parameters but not on $\beta_1$ or on $z$.
\end{lem}
\begin{proof}
Define the open set $U$ and the event $S_{z,1}$ as in the proof of Lemma~\ref{F prob}. By Bayes' rule we have
\[
\mathbf P\left(S_{z,1} \, | \, \widetilde E_{z,1} \cap F_{z,1} \right) = \frac{\mathbf P(F_{z,1}\cap S_{z,1} \,|\, \widetilde E_{z,1})}{\mathbf P(F_{z,1} \, |\, \widetilde E_{z,1} ) } .
\]
By the argument of Lemma~\ref{F prob}, this is at least a positive $(\beta_1, z)$-independent constant. On the other hand we have
\[
\mathbf P\left(E_{z,1} \,|\, \widetilde E_{z,1} \cap S_{z,1} \cap F_{z,1} \right) \succeq 1.
\]
\end{proof}
By combining Proposition~\ref{1pt at hitting}, Lemma~\ref{L prob}, Lemma~\ref{F prob}, and Lemma~\ref{E prob} we infer
\begin{lem} \label{E_1 prob}
Provided $\beta_1$ is chosen sufficiently large and the other parameters are chosen appropriately, independently of $\beta_1$ and $u_1$ and uniformly for $z$ in compacts, we have $e^{-\beta_1 (\gamma^*(q) + \gamma_0^*(q) u_1 )} \preceq \mathbf P(E_{z,1}) \preceq e^{-\beta_1 (\gamma^*(q) - \gamma_0^*(q) u_1 )}$ with the implicit constant depending on the other parameters (including $u_1$) but not on $\beta_1$, and uniform for $z$ in compacts.
\end{lem}
Now suppose $j\geq 2$ and the objects have been defined for all positive integers $l \leq j-1$.
Let $\eta_{z,j}^0$ be the image under $p_{z,j-1}$ of the part of $\eta_{z,j-1}$ which lies in $D_{z,j-1}$. Let $x_{z,j} $ and $y_{z,j} $ be its initial and terminal points. Define the map $\psi_{z,j}$, the times $\sigma_{z,j}$ and $\overline\sigma_{z,j}$, and the curve $\eta_{z,j}$ in the same manner as in the case $j=1$, but with all of the 1's replaced by $j$'s (the auxiliary parameters remain unchanged).
Let $[x_{z,j}^* , y_{z,j}^*]_{\partial\mathbf D}$ be the largest sub-arc of $[x_{z,j} , y_{z,j}]_{\partial\mathbf D}$ not disconnected from the origin by $(\eta_{z,j}^0)^{\sigma_{z,j}} \cup (\overline\eta_{z,j}^0)^{\overline\sigma_{z,j}}$ (this arc need not be equal to $[x_{z,j} , y_{z,j}]_{\partial\mathbf D}$ since $\eta_{z,j}^0$ can intersect $\partial\mathbf D$).
Fix a constant $p_L >0$. Define the event $L_{z,j}$ in exactly the same manner as in the case $j=1$ but with all 1's replaced by $j$'s.
Let $\mathcal F_{z,j}^0$ be the $\sigma$-algebra generated by $\eta_{z,j}^0|_{[0,\sigma_{z,j}]}$ and $\overline\eta_{z,j}^0|_{[0,\overline\sigma_{z,j}]}$.
Define the event $\widetilde E_{z,j}$, the times $\tau_{z,j} $ and $\overline\tau_{z,j}$, and the map $\phi_{z,j}$ in exactly the same manner as the corresponding objects for $j=1$, but with all the 1's replaced by $j$'s.
Let $T_{z,j}$ and $\overline T_{z,j} $ be the times for $\eta$ and $\overline\eta$ corresponding to $\tau_{z,j} $ and $\overline \tau_{z,j} $.
Let $\widehat\eta_{z,j}^\pm$ be the flow lines of $\eta$ started from $\eta(T_{z,j} )$ with angles $\mp \theta$. Let $\widehat D_{z,j}$ be the connected component of $\mathbf D\setminus (\widehat\eta_{z,j}^+ \cup \widehat\eta_{z,j}^-)$ containing $z$.
Let $ \eta_{z,j}^\pm $ be the images of $\widehat\eta_{z,j}^\pm$ under the map $\psi_{z,j} \circ \widehat p_{z,j-1 }$.
Define the domain $D_{z,j}$, the event $F_{z,j}$, the map $p_{z,j}$, and the times $t_{z,j}^\pm$ and $\widetilde t_{z,j}^\pm$ in exactly the same manner as in the $j=1$ case except with all of the 1's replaced by $j$'s. Let $\widehat p_{z,j} = p_{z,j} \circ \psi_{z,j} \circ \widehat p_{z,j-1}$ so that $\widehat p_{z,j} : \widehat D_{z,j} \rightarrow \mathbf D$ takes $z$ to 0.
Also define the event $E_{z,j}$ and the times $\tau_{z,j}^*$ and $\overline\tau_{z,j}$ in exactly the same manner as in the $j=1$ case except with all of the 1's replaced by $j$'s.
Let $T_{z,j}^*$ and $\overline T_{z,j}^*$ be the times for $\eta$ and $\overline\eta$ corresponding to $\tau_{z,j}^*$ and $\overline \tau_{z,j}^*$.
Let $\mathcal F_{z,j}$ be the $\sigma$-algebra generated by $\eta|_{[0,T_{z,j}^*]}$, $\overline\eta|_{[0,\overline T_{z,j}^*]}$, and $\widehat \eta_{z,l}^\pm|_{[0, t_{z,l}^\pm]}$ for $l \leq j$.
\begin{figure}[ht!]
\begin{center}
\includegraphics[scale=.7]{figures/drawing-flow-lines-E2.pdf}
\caption{An illustration of the curve $\eta$ and the auxiliary flow lines $\widehat \eta_{z,1}^\pm$ and $\widehat \eta_{z,2}^\pm$ on the event $E_{z,1} \cap E_{z,2}$. Note that only a neighborhood of $z$ is shown, not the whole disk. The parts of $\eta$ involved with the last parts of the definitions of $E_{z,1}$ and $E_{z,2}$ are shown in orange. The parts of $\eta$ involved with the event $L_{z,2}$ are shown in purple. The domain $\widehat D_{z,1}$ (resp.\ $\widehat D_{z,2}$) is the region enclosed by the outer (resp.\ inner) green curves.} \label{flow lines fig E2}
\end{center}
\end{figure}
\subsection{k-perfect points}
\label{perfect setup sec'}
Fix sequences $\beta_j \rightarrow \infty$ and $u_j \rightarrow 0$ as above. Define the events $E_{z,j}^m$ in the same manner as the events $E_{z,j}$ in the preceding subsection but with the sequences $(\beta_j)$ and $(u_j)$ replaced by $(\beta_{j+m })$ and $(u_{j+m })$. Also fix $d\in (0,1)$. The estimate of Lemma~\ref{E_1 prob} becomes
\begin{equation} \label{C_u}
C_{u_{m+1}}^{-1} e^{-\beta_{m+1} (\gamma^*(q) + \gamma_0^*(q) u_{m+1} )} \leq \mathbf P(E_{z,1}^m ) \leq C_{u_{m+1}} e^{-\beta_{m+1} (\gamma^*(q) - \gamma_0^*(q) u_{m+1} )} ,
\end{equation}
where for $u > 0$, $C_u$ is a constant which is allowed to depend on $u $ and on the other parameters in the preceding subsection but not on any of the $\beta_j$'s, and is uniform for $z \in B_d(0)$. We can take $C_u$ to be decreasing in $u$. We also let
\begin{equation}\label{bar beta}
\overline \beta_{m_1 , m_2} = \sum_{j=m_1+1 }^{m_2 } \beta_j ,\qquad \overline u_{m_1 , m_2} = \sum_{j=m_1+1}^{m_2 } \beta_j u_j ,\qquad \overline\beta_m = \overline\beta_{0,m}, \qquad \overline u_m = \overline u_{0,m} .
\end{equation}
\begin{remark}
The reason we allow $\beta$ and $u$ to vary here is that we eventually want to get a lower bound for the Hausdorff dimension of the sets $\Theta^s(D_\eta)$ and $\widetilde\Theta^s(D_\eta)$. If we fixed $u$, we would instead get the Hausdorff dimension of the sets where the limits in the definitions of $\Theta^s(D_\eta)$ and $\widetilde\Theta^s(D_\eta)$ are between $s-u$ and $s+u$. In order to allow $u$ to vary, we also need to allow $\beta$ to vary, for otherwise the constants $C_u$ in~\eqref{C_u} would be larger than $e^\beta$ when $u$ is very small. The idea in Lemma~\ref{beta u choice} below is to let $u_j \rightarrow 0$ and $\beta_j \rightarrow \infty$ slowly enough that our estimates are not much different than they would be with fixed $\beta$ and $u$.
\end{remark}
\begin{lem} \label{beta u choice}
Given $d$ and the parameters from Section~\ref{perfect setup sec}, one can choose sequences $(\beta_j)$ and $(u_j)$ such that the following is true.
\begin{enumerate}
\item $\beta_j$ increases to $\infty$ and $u_j$ decreases to 0 as $j\rightarrow\infty$. \label{beta u limit}
\item For each $j \in \mathbf N$, $\beta_{j+1} \leq \beta_j + o_j(1)$. \label{beta u diff}
\item For each fixed $k \in \mathbf N$, $\beta_{jk} \leq \overline \beta_j o_j(1) $. \label{beta u k}
\item For each $j\in \mathbf N$, $C_{u_j} \leq e^{\beta_j u_j \gamma_0^*(q)}$. \label{beta u C}
\item $\beta_j u_j \rightarrow \infty$ as $j\rightarrow \infty$. \label{beta u infty}
\end{enumerate}
\end{lem}
\begin{proof}
Choose $\beta_0$ much larger than $\Delta \vee \gamma_0^*(q)^{-1} \log C_1$ and set $\beta_j = \log j + \beta_0$. It is clear that $\beta_j\rightarrow \infty$ as $j\rightarrow \infty$ and assertions~\ref{beta u diff} and~\ref{beta u k} hold. We now inductively choose $(u_j)$.
Start with a sequence $\{u_m^*\}_{m\in\mathbf N} \subset (0,1)$ which decreases to 0. Let $j_1$ be the least positive integer $j$ such that $C_{ u_2^* } \leq e^{ \beta_j u_2^* \gamma^*_0(q) }$. Such a $j$ exists since $\beta_j \rightarrow \infty$ as $j\rightarrow\infty$. Set $u_j = u_1^*$ for $j\leq j_1$. Inductively, suppose $m \geq 2$ and $j_1 , \ldots , j_{m-1}$ and $u_j$ for $j\leq j_{m-1}$ have been defined. Let $j_m$ be the least integer $j \geq j_{m-1}+ 1$ such that $C_{u_{m+1}^*} \leq e^{\beta_{j } u_{m+1}^* \gamma_0^*(q)}$. Let $u_j = u_m^*$ for $j = j_{m-1} + 1 , \ldots , j_m$. It is clear that condition~\ref{beta u C} holds for this choice of $(u_j)$. Observe that condition~\ref{beta u C} still holds if we increase $u_j$. Hence by increasing $u_j$ we can arrange that~\ref{beta u infty} holds, and we still have $u_j\rightarrow 0$.
\end{proof}
We henceforth assume the sequences $(\beta_j)$ and $(u_j)$ are chosen as in Lemma~\ref{beta u choice}.
For $n\geq k \geq 0$, define
\begin{equation} \label{E_z^n def}
E_{k,n}^m(z) = \bigcap_{j = k+1}^n E_{z,j}^m .
\end{equation}
Also let
\[
E_{k,n}(z) := E_{k,n}^0(z) ,\qquad E_n(z) := E_{0,n} (z) .
\]
The set of \emph{$n$-perfect points} is
\begin{equation} \label{mathcal P_n def}
\mathcal P_n := \left\{z\in \mathbf D : E_n(z) \: \operatorname{occurs} \right\}
\end{equation}
\subsection{Analytic properties}
\label{flow line analytic sec}
In this subsection we study some analytic properties of the events of Sections~\ref{perfect setup sec} and~\ref{perfect setup sec'}. The results of this subsection are needed to analyze the correlation structure of our events in the next subsection and to show that the perfect points are in fact contained in the sets whose Hausdorff dimension we want to compute in Section~\ref{haus lower sec}. The main result of this subsection is the following.
\begin{lem}\label{E_z basics}
Assume we are in the setting of Sections~\ref{perfect setup sec} and~\ref{perfect setup sec'}. For $n\in\mathbf N$ let $\Phi_{z,n}$ be the conformal map from $\mathbf D\setminus (\eta^{T_{z,n}^*} \cup \overline\eta^{\overline T_{z,n}^*})$ to $\mathbf D$ which takes $-i^+$ to $-i$, $i^-$ to $i$, and 1 to 1. The following holds a.s.\ on $E_n(z)$, with all implicit constants deterministic and independent of $n$ and of $z\in B_d(0)$.
\begin{enumerate}
\item We have \label{Phi' asymp}
\begin{equation*}
e^{-\overline\beta_{ n} q - 2 \overline u_{ n} } \preceq |\Phi_{z,n }'(z)| \preceq e^{ - \overline\beta_{ n} q + 2 \overline u_{ n} } .
\end{equation*}
\item There is a constant $\lambda > 0$, independent of $n$ and $z\in B_d(0)$, such that \label{eta dist}
\begin{equation*}
e^{-\overline\beta_{ n} - \lambda n } \leq \operatorname{dist}\left(z , \eta^{T_{z,n}^*} \cup \overline\eta^{\overline T_{z,n}^*}\right) \leq e^{ - \overline\beta_{ n} + \lambda n } .
\end{equation*}
\item We have \label{eta tips}
\[
|\eta(T_{z,n}^*) - z| \asymp |\overline\eta(\overline T_{z,n}^*) - z| \asymp \operatorname{dist}(z , \eta^{T_{z,n}^*} \cup \overline\eta^{\overline T_{z,n}^*})
\]
\item We have \label{wh D diam}
\[
e^{ -\overline\beta_n - \lambda n } \leq \operatorname{dist}(z , \partial \widehat D_{z,n}) \leq \operatorname{diam} \widehat D_{z,n } \leq e^{-\overline\beta_n + \lambda n} .
\]
\end{enumerate}
\end{lem}
To prove Lemma~\ref{E_z basics} we will need to compare the derivatives of several different maps. To this end, we will define the following objects.
\begin{enumerate}
\item Conformal maps $\phi_{z,j}^0$, $\widetilde \phi_{z,j} $, $\widehat \phi_{z,j}$, $f_{z,j}$, and $g_{z,j}$.
\item Random times $\sigma_{z,j}^*$, $\overline\sigma_{z,j}^*$, $\widetilde\tau_{z,j}^*$, and $\overline{\widetilde \tau}_{z,j}^*$.
\item Points $\widetilde x_{z,j}$, $\widetilde y_{z,j}$, and $\widetilde m_{z,j}$.
\item Curves $\widetilde \eta_{z,j} $.
\end{enumerate}
For $j\in\mathbf N$, let $\widehat\psi_{z,j}$ be the conformal map from $\mathbf D\setminus (\eta^{T_{z,j}^*} \cup \overline\eta^{\overline T_{z,j}^*})$ to $\mathbf D$ which fixes 0 and whose derivative at 0 has the same argument as $ \Phi_{z,j}'(z)$
Let $\sigma_{z,j}^*$ and $\overline\sigma_{z,j}^*$ be the stopping times for $\eta_{z,j}^0$ corresponding to $\tau_{z,j}^*$ and $\overline\tau_{z,j}^*$ (equivalently to $T_{z,j}^*$ and $\overline T_{z,j}^*$). Let $\phi_{z,j}^0$ be the conformal map from the connected component of $\mathbf D\setminus ((\eta_{z,j}^0 )^{\sigma_{z,j}^*} \cup (\overline\eta_{z,j}^0 )^{ \overline\sigma_{z,j}^*})$ containing 0 to $\mathbf D$ which takes $x_{z,j}^*$ to $-i$, $y_{z,j}^*$ to $i$, and the midpoint $m_{z,j}^*$ of $[x_{z,j}^* , y_{z,j}^*]_{\partial\mathbf D}$ to 1.
For $j=1$, the map $f_{z,1}$ has already been defined in Section~\ref{perfect setup sec}. For $j\geq 2$, we let $f_{z,j}$ be the conformal map which takes $\Phi_{z,j-1}(z)$ to $0$ with $\Phi_{z,j-1}'(z) > 0$. Observe that $\widehat\psi_{z,j-1} = f_{z,j} \circ \Phi_{z,j-1}$ (where here we take $\Phi_{z,0}$ to be the identity map and $\widehat \psi_{z,j} = f_{z,j}$ in the case $j=1$). For $j\geq 1$, let $\widetilde \eta_{z,j} $ be the image under $\widehat\psi_{z,j-1}$ of the part of $\eta$ between $\eta(T_{z,j-1}^*)$ and $\overline\eta(\overline T_{z,j-1}^*)$. Note that $\widetilde\eta_{z,j}$ involves the same part of the curve as $\eta_{z,j}^0$, which is larger than the part of the curve involved in the definition of $\eta_{z,j}$. Let $\widetilde\tau_{z,j}^* $ and $\overline{\widetilde \tau}_{z,j}^* $ be the times for $\widetilde\eta_{z,j} $ and $\overline{\widetilde \eta}_{z,j} $ corresponding to the times $T_{z,j}^*$ and $\overline T_{z,j}^*$ for $\eta$ and $\overline\eta$.
Let $\widetilde x_{z,j}$ and $\widetilde y_{z,j}$ be the start and end points for $\widetilde\eta_{z,j} $. Let $\widetilde\phi_{z,j} $ be the conformal map from $ \mathbf D \setminus (\widetilde\eta_{z,j}^{\widetilde\tau_{z,j}^* } \cup \overline{\widetilde\eta}_{z,j}^{\overline{\widetilde\tau}_{z,j}^*}) $ to $\mathbf D$ which takes $\widetilde x_{z,j}^+$ to $-i$, $\widetilde y_{z,j}^-$ to $i$ and the midpoint $\widetilde m_{z,j}$ of $[\widetilde x_{z,j} , \widetilde y_{z,j}]_{\partial\mathbf D}$ to 1. Let $g_{z,j} : \mathbf D\rightarrow \mathbf D$ be the conformal map taking $(\widetilde\phi_{z,j} \circ f_{z,j})(b)$ to $b$ for $b = -i^+ , i^- ,1$. Let
\begin{equation*}
\widehat\phi_{z,j} := g_{z,j} \circ \widetilde \phi_{z,j} \circ f_{z,j} .
\end{equation*}
Observe that (with $\Phi_{z,j}$ as in Lemma~\ref{E_z basics})
\begin{equation} \label{Phi wh phi}
\Phi_{z,j} = \widehat\phi_{z,j} \circ \cdots \circ \widehat \phi_{z,1}.
\end{equation}
See Figure~\ref{flow lines fig maps} for an illustration of these maps in the case $j=2$ (which has all of the features of the general case).
\begin{figure}[ht!]
\begin{center}
\includegraphics[scale=.7]{figures/drawing-flow-lines-maps.pdf}
\caption{An illustration of some of the maps associated with the events $E_{z,1}$ and $E_{z,2}$. The images of $-i$, $i$, and 1 are shown in red. The images of $z$ are shown in blue. The black curves are associated with the event $E_{z,1}$ and the blue curves are associated with the event $E_{z,2}$. The map $\widehat\phi_{z,2}$ is the composition of the last four maps in the figure. The map $\Phi_{z,2}$ is the composition of all four maps.} \label{flow lines fig maps}
\end{center}
\end{figure}
\begin{lem} \label{p and psi}
If $\beta_1$ is chosen sufficiently large, then on the event $E_n(z)$, we have
\begin{equation} \label{p* asymp}
|\widehat p_{z,n}'(z)| \asymp |\widehat \psi_{z,n}'(z)| ,
\end{equation}
with the implicit constants independent of $n$ and uniform for $z\in B_d(0)$.
\end{lem}
\begin{proof}
Assume we are working on the event $ E_n(z)$. Let $p_{z,n-1}^*$ be the conformal map from $\widehat\psi_{z,n}(\widehat D_{z,n-1})$ to $\mathbf D$ with $p_{z,n-1}^*(0) = 0$ and $(p_{z,n-1}^*)'(0) > 0$ (in the case $n = 1$, we take $p_{z,n-1}^* $ to be the identity). Let $p_{z,n}^{**}$ be the conformal map from $(p_{z,n -1}^* \circ \widehat \psi_{z,n})(\widehat D_{z,n})$ to $\mathbf D$ with $p_{z,n}^{**}(0)=0$ and $(p_{z,n}^{**})'(0) > 0$. We then have
\begin{equation} \label{wh psi decomp}
\widehat p_{z,n} = p_{z,n}^{**} \circ p_{z,n-1}^* \circ \widehat \psi_{z,n} .
\end{equation}
By the Beurling estimate and \cite[Exercise 2.7]{lawler-book} the diameter of $\mathbf D\setminus \widehat\psi_{z,n}(\widehat D_{z,n-1})$ tends to 0 as $\beta_n \rightarrow \infty$ (and hence also as $\beta_1 \rightarrow \infty$). Therefore, if $\beta_1$ is chosen sufficiently large, then $|(p_{z,n-1}^*)'(0)|\asymp 1$.
By condition~\ref{E_z stay} in the definition of $E_{z,n}$, the distance from 0 to $(p_{z,n-1}^* \circ \psi_{z,n})(\partial \widehat D_{z,n})$ is proportional to the distance from 0 to $\psi_{z,n}^F(\partial D_{z, n})$, with $\psi_{z,n}^F$ as in the definition of $F_{z,n}$. By condition~\ref{F contained} in the definition of $F_{z,n}$, this distance is $\asymp 1$. Consequently, $|(p_{z,n}^{**})'(0)|\asymp 1$ so our desired result follows from~\eqref{wh psi decomp}.
\end{proof}
\begin{lem} \label{wt phi G}
Let $\zeta \in (0 , a/100)$. If we make the parameter $r$ in the definition of $E_{z,1}$ sufficiently small (depending on $\zeta$ and the other parameters but not on $\beta_1$ or $u_1$ and uniform for $z \in B_d(0)$) then for any sub-arc $I$ of $[\widetilde x_{z,n+1} , \widetilde y_{z,n+1}]_{\partial\mathbf D}$ lying at distance at least $\zeta $ from $\widetilde x_{z,n+1}$ and $\widetilde y_{z,n+1}$, we have that $\widetilde\phi_{z,n+1}$ is Lipschitz on $I $ and $\widetilde\phi_{z,n+1}^{-1}$ is Lipschitz on $\widetilde\phi_{z,n+1}(I)$ on the event $E_n(z)$ with Lipschitz constants independent of $\beta_1$ and $u_1$ and uniform for $z \in B_d(0)$.
\end{lem}
\begin{proof}
See Figure~\ref{phi G maps} for an illustration of the argument. Throughout, we work on the event $E_n(z)$.
Let $A := \widehat\psi_{z,n }((\widehat \eta_{z,n}^+)^{t_{z,n}^+})$. Then $A$ disconnects $\widetilde\eta_{z,n+1} $ from $I $ in $\partial\mathbf D$. We claim that there is a constant $\delta > 0$, independent of $\beta_1$ and $u_1$ and uniform for $z\in B_d(0)$, such that
\begin{equation}\label{A zeta' dist}
E_{z,n-1} \cap E_{z,n} \subset \{\operatorname{dist}(A , I) \geq \delta \} .
\end{equation}
Given the claim, the statement of the lemma follows from Lemma~\ref{G dist}.
Let $\psi_{z,n}^*$ be a conformal map from the connected component of $\mathbf D\setminus (\eta_{z,n}^{\tau_{z,n}^*} \cup \overline\eta_{z,n}^{\overline\tau_{z,n}^*})$ containing 1 on it boundary to $\mathbf D$ which fixes 0. This map is defined only up to a rotation, which we will specify shortly. Let $\eta_{z,n+1}^*$ be the image under $\psi_{z,n}^*$ of the part of $\eta_{z,n}$ between $\eta_{z,n}(\tau_{z,n}^*)$ and $\overline\eta_{z,n}(\overline\tau_{z,n}^*)$. We can normalize $\psi_{z,n}^*$ in such a way that we have
\[
\eta_{z,n+1}^* = p_{z,n-1}^*( \widetilde\eta_{z,n+1} ) ,
\]
with $p_{z,n-1}^*$ as in the proof of Lemma~\ref{p and psi}.
By condition~\ref{E top} in the definition of $\widetilde E_{z,n}$ and condition~\ref{E_z stay} in the definition of $E_{z,n}$, we have that $\mathbf D\setminus \widehat\psi_{z,n }(\widehat D_{z,n-1})$ lies at distance at least a positive constant depending only on $a$ from $\widetilde x_{z,n} $ and $\widetilde y_{z,n} $ on $E_{z,n }$. Since the diameter of $\mathbf D\setminus \widehat\psi_{z,n }(\widehat D_{z,n-1})$ tends to 0 as $\beta_n \rightarrow \infty$ (by the argument of Lemma~\ref{p and psi}), we have that $(p_{z,n-1}^*)^{-1}$ is nearly constant near $\widetilde x_{z,n} $ and $\widetilde y_{z,n} $ if $\beta_1$ is large.
By the Schwarz lemma $(p_{z,n-1}^*)^{-1} $ decreases distances to $\partial\mathbf D$. Hence the distance from $A$ to $I$ is at least a $\beta_1$-independent constant times the distance from $A^*$ to $I$ if $\beta_1$ is large, where
\[
A^* = p_{z,n-1}^*(A) = \psi_{z,n}^*((\eta_{z,n}^+)^{t_{z,n}^+}) .
\]
Hence it is enough to prove~\eqref{A zeta' dist} with $A^*$ in place of $A$.
Let $I' \supset I$ be a slightly larger arc. By condition~\ref{E_z stay} in the definition of $E_{z,n}$ the distance from $A^*$ to $I$ is $\succeq$ the distance from $\psi_{z,n}^F( (\eta_{z,n}^+)^{t_{z,n}^+})$ to $I'$ if $r$ is chosen sufficiently small, where $\psi_{z,n}^F $ is as in the definition of $F_{z,n}$.
We conclude by applying condition~\ref{F G} in the definition of $F_{z,n}$.
\end{proof}
\begin{figure}
\begin{center}
\includegraphics[scale=.8]{figures/drawing-phi-G-maps.pdf}
\caption{An illustration of the maps used in the proof of Lemma~\ref{wt phi G}. In order to control the distance from $\widetilde\eta_{z,n+1} $ to an arc on the right boundary of the disk, we compare $\widetilde\eta_{z,n+1} $ to the curve $\eta_{z,n+1}^*$ and then to the curve $\eta_{z,n }^F$. The distance from the last curve to the right boundary is bounded below by condition~\ref{F G} in the definition of $F_{z,n}$.} \label{phi G maps}
\end{center}
\end{figure}
\begin{lem} \label{phi' asymp}
We can choose the parameters in Section~\ref{perfect setup sec} independently of $n$ in such a way that on $E_n(z)$ we have
\begin{equation} \label{phi' asymp eqn}
e^{- \beta_n (q + u_n) } \preceq |\phi'(w)| \preceq e^{-\beta_n (q-u_n)}
\end{equation}
where the pair $(\phi , w)$ is any one of $(\phi_{z,n} , 0) , (\phi_{z,n}^0, 0) , ( \widetilde \phi_{z,n} , 0)$, or $ (\widehat \phi_{z,n} , \Phi_{z,n-1}(z))$. The implicit constants are independent of $n$ and uniform for $z$ in compacts.
\end{lem}
\begin{proof}
By definition of $E_{z,n}$ (in particular, of $\widetilde E_{z,n}$), the statement of the lemma is true for $(\phi , w) = (\phi_{z,n} , 0)$.
We will now transfer the estimate~\eqref{phi' asymp eqn} from $\phi_{z,n}$ to $\phi_{z,n}^0$ to $\widetilde\phi_{z,n}$ to $\widehat\phi_{z,n}$. This latter map is our primary interest, mostly because of~\eqref{Phi wh phi}. Throughout, we assume that $E_{z,n-1} \cap E_{z,n}$ occurs and require all implicit constants to be independent of $m$ and $n$ and uniform for $z$ in compacts.
Let $\phi_{z,n}^{**}$ be the conformal map from the connected component of $\mathbf D \setminus (\eta_{z,n}^{\tau_{z,n}^*} \cup \overline\eta_{z,n}^{\overline\tau_{z,n}^*})$ containing 0 to $\mathbf D$ which takes $-i^+$ to $-i$, $i^-$ to $i$, and 1 to 1.
Let $g_{z,n}^*$ be the conformal automorphism of $\mathbf D$ which fixes $-i$ and $i$
and takes $(\phi_{z,n}^{**} \circ \psi_{z,n})(m_{z,n}^*)$ to $1$. Then we have
\begin{equation} \label{phi^* decomp}
\phi_{z,n}^0 = g_{z,n}^* \circ \phi_{z,n}^{**}\circ \psi_{z,n}
\end{equation}
By condition~\ref{E_z stay} in the definition of $E_{z,n}$ and Lemma~\ref{phi hm arc}, we have $|(\phi_{z,n}^{**})'(0)| \asymp |\phi_{z,n}'(0)|$ provided $r$ is chosen sufficiently small independently of $n$ and uniformly for $z$ in compacts.
Furthermore, by Lemma~\ref{G dist} and condition~\ref{E G} in the definition of $E_{\beta_n}^{q;u_n}(\cdot)$, we have that $\mathcal G(\phi_{z,n}^{**} , \mu')$ holds on $E_{z,n}$ for some $\mu'\in\mathcal M$ depending only on $\mu$. By condition~\ref{L hit} in the definition of $L_{z,n}$ and condition~\ref{F hm} in the definition of $F_{z,n-1}$, $|x_{z,n}^* - y_{z,n}^*| \succeq 1$ on $E_{z,n-1} \cap E_{z,n}$. By combining this with condition~\ref{L G} in the definition of $L_{z,n}$ we have $|(g_{z,n}^*)'(z)| \asymp 1$ on all of $\mathbf D$ on the event $E_{z,n}$. Hence~\eqref{phi^* decomp} implies~\eqref{phi' asymp eqn} for $\phi_{z,n}^0$.
By Lemma~\ref{phi hm D} applied with $U = \mathbf D\setminus ((\widetilde \eta_{z,n} )^{\widetilde\tau_{z,n}^*} \cup (\overline{\widetilde\eta}_{z,n} )^{\overline{\widetilde \tau}_{z,n}^*})$, $D = \widehat\psi_{z,n-1}(\widehat D_{z,n-1})$, $\phi = \widetilde\phi_{z,n}$, $\widehat\phi = \phi_{z,n}^0$, and $z = \widehat z = 0$, the estimate~\eqref{phi' asymp eqn} for $\phi_{z,n}^0$ implies the estimate~\eqref{phi' asymp eqn} for $\widetilde\phi_{z,n}$. Note that the conditions of Lemma~\ref{phi hm D} (with parameters $\zeta , \Delta , \delta \succeq 1$) follow easily from the definition of $E_n(z)$ together with Lemmas~\ref{p and psi} and~\ref{wt phi G}
To get the estimate for $\widehat\phi_{z,n}$, write
\begin{equation} \label{wh phi_2 decomp}
|\widehat \phi_{z,n}'(\Phi_{z,n-1}(z))| = |g_{z,n}'( \widetilde \phi_{z,n} (0) )| |\widetilde \phi_{z,n}'(0)| |f_{z,n}'(\Phi_{z,n-1}(z)) | .
\end{equation}
By the Koebe quarter theorem, $ |f_{z,n}'(\Phi_{z,n-1}(z)) | \asymp (1-|\Phi_{z,n-1}(z)|)^{-1}$. In fact, we have
$|f_{z,n}'(w)| \asymp 1-|\Phi_{z,n-1}(z)|$ on subsets of $\mathbf D$ at positive distance from $\Phi_{z,n-1}(z)$.
By condition~\ref{E top} in the definition of $E_{\beta_{n-1}}^{q;u_{n-1}}(\cdot)$, we can find $\zeta > 0$ depending only on $a$ such that $ f_{z,n} ([-i,i]_{\partial\mathbf D})$ lies at distance at least $\zeta $ from $\widetilde x_{z,n}$ and $\widetilde y_{z,n}$ on $E_{z,n-1}$. By Lemma~\ref{wt phi G}, on $E_{z,n-1}$, it holds that $\widetilde\phi_{z,n}$ distorts the distances between points in $f_{z,n} ([-i,i]_{\partial\mathbf D})$ by at most a constant factor.
Since conformal automorphisms of $\mathbf D$ depend smoothly on their parameters, it follows that
\begin{equation} \label{g' asymp n}
|g_{z,n}'| \asymp |(f_{z,n}^{-1})'|
\end{equation}
on the whole left half of $\mathbf D$. By combining this with~\eqref{wh phi_2 decomp} and the first statement of the lemma we conclude.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{E_z basics}]
Assume $E_n(z)$ holds. Assertion~\ref{Phi' asymp} is immediate from Lemma~\ref{phi' asymp} and the relation~\eqref{Phi wh phi}. Note that we can absorb the implicit constants in~\eqref{phi' asymp eqn} into an additional factor of $e^{\overline u_n}$ due to condition~\ref{beta u infty} of Lemma~\ref{beta u choice}.
To prove item~\ref{eta dist}, we induct on $n$. The case $n=1$ is immediate from the definitions of the events.
Now suppose $n\geq 2$ and the result has been proven with $n$ replaced by $n-1$.
Since $\widehat \psi_{z,n-1}^{-1}$ maps $\mathbf D\setminus (\widetilde \eta_{z,n }^{\widetilde\tau_{z,n }^*} \cup \overline{\widetilde \eta}_{z,n }^{\overline{\widetilde\tau}_{z,n }^*})$ to $\mathbf D\setminus (\eta^{T_{z,n }^*} \cup \overline{ \eta}^{\overline{T}_{z,n }^*})$ and fixes 0, the Koebe quarter theorem implies
\begin{equation} \label{dist compare}
\operatorname{dist}\left(z , \eta^{T_{z,n }^*} \cup \overline{ \eta}^{\overline{T}_{z,n }^*}\right) \asymp |(\widehat \psi_{z,n-1}^{-1})'(0)| \operatorname{dist}\left(0 , \widetilde \eta_{z,n }^{\widetilde\tau_{z,n }^*} \cup \overline{\widetilde \eta}_{z,n }^{\overline{\widetilde\tau}_{z,n }^*} \right) .
\end{equation}
By a second application of the Koebe quarter theorem,
\begin{equation} \label{wh psi koebe}
|(\widehat \psi_{z,n-1}^{-1})'(0)| \asymp \operatorname{dist}\left(z , \eta^{T_{z,n-1}^*} \cup \overline\eta^{\overline T_{z,n-1}^*}\right) .
\end{equation}
By the inductive hypothesis,
\begin{equation} \label{wh psi inductive}
e^{- \overline\beta_{n-1} - \lambda (n-1) } \preceq \operatorname{dist}\left(z , \eta^{T_{z,n-1}^*} \cup \overline{ \eta}^{\overline{T}_{z,n-1}^*}\right) \preceq e^{- \overline\beta_{n-1} + \lambda (n-1) } .
\end{equation}
By Lemma~\ref{p and psi} (applied with $n$ replaced by $n-1$), we have
\begin{equation} \label{dist compare 0}
\operatorname{dist}\left(0 , \widetilde \eta_{z,n }^{\widetilde\tau_{z,n }^*} \cup \overline{\widetilde \eta}_{z,n }^{\overline{\widetilde\tau}_{z,n }^*} \right) \asymp \operatorname{dist}\left(0 , (\eta_{z,n}^0)^{\sigma_{z,n}^*} \cup (\overline \eta_{z,n}^0)^{\overline\sigma_{z,n}^*} \right).
\end{equation}
By definition of $E_{z,n}$, this last distance is $\asymp e^{-\beta_n}$ on $E_{z,n}$. Provided $\lambda$ is chosen sufficiently large, independently of $n$ and $z\in B_d(0)$, we can now complete the induction by combining~\eqref{dist compare}, ~\eqref{wh psi koebe},~\eqref{wh psi inductive}, and~\eqref{dist compare 0}.
By condition~\ref{E top} in the definition of the event from Section~\ref{2pt setup sec}, condition~\ref{E_z stay} in the definition of $E_{z,n}$, and the Beurling estimate, the harmonic measure from $z$ of each of the two sides of $\eta^{T_{z,n}^*}$ (resp.\ each of the two sides of $\overline\eta^{\overline T_{z,n}^*}$) in $\mathbf D\setminus ( \eta^{T_{z,n}} \cup \overline\eta^{\overline T_{z,n}})$ is at least some $n $-independent constant. This implies assertion~\ref{eta tips}.
For item~\ref{wh D diam}, we use assertion~\ref{eta dist} and the Koebe quarter theorem to see that there exists radii $\rho ' > \rho > 0 $ such that $\rho \succeq e^{-\overline\beta_{n-1} - \lambda n }$, $\rho' \preceq e^{-\overline\beta_{n-1} +\lambda n }$, $ \widehat \psi_{z,n-1}^{-1}(B_{\beta_n +\log 2}) \supset B_\rho(z)$, and $\widehat \psi_{z,n-1}^{-1}(B_{\beta_n -\log 2}) \subset B_{\rho'}(z)$. By combining this with condition~\ref{F contained} in the definition of $F_{z,j}$ we see that assertion~\ref{wh D diam} holds (after possibly increasing $\lambda$).
\end{proof}
\subsection{Probabilistic properties}
In this subsection we study the events defined above from a probabilistic perspective, and eventually prove our two-point estimate.
\begin{prop} \label{2pt estimate}
Let $z,w\in B_d(0)$. Let $\lambda$ be as in assertion~\ref{eta dist} of Lemma~\ref{E_z basics}. Choose $ k \in \mathbf N$ such that $e^{- \overline \beta_{k +1} -\lambda (k+1) } \leq |z-w| \leq e^{-\overline \beta_{k } - \lambda k }$. Then for any $n \in \mathbf N$ with $\overline\beta_n -\lambda n \geq \overline \beta_{k +1} + \lambda (k+2) $,
\begin{equation} \label{2pt estimate eqn}
\mathbf P( E_n(z) \cap E_n(w)) \preceq e^{\overline\beta_k o_k(1)} \frac{\mathbf P( E_n(z)) \mathbf P( E_n(w))}{\mathbf P( E_k(w))}
\end{equation}
with the implicit constants independent of $n$ and $k$, the $o_k(1)$ independent of $n$, and both deterministic and uniform for $z,w \in B_d(0)$.
\end{prop}
\begin{remark}
In the setting of Proposition~\ref{2pt estimate} we have
\[
e^{-\overline\beta_k } = |z-w|^{1 + o_{|z-w|}(1)}
\]
so by Lemma~\ref{P(E_z)} below we can rewrite the estimate~\eqref{2pt estimate eqn} as
\[
\mathbf P( E_n(z) \cap E_n(w)) \preceq |z-w|^{-\gamma^*(q) + o_{|z-w|}(1)} \mathbf P( E_n(z)) \mathbf P( E_n(w)) .
\]
This is the form of the estimate we will use when we prove lower bounds for the Hausdorff dimensions of our sets.
\end{remark}
In order to prove Proposition~\ref{2pt estimate}, we first need to introduce an additional family of auxiliary flow lines which form the same pocket as $\eta_{z,n}^\pm$ on the event $E_n(z)$, but whose law is easier to analyze.
To this end, let $\dot \eta_{z,1}^0 = \eta_{z,1}^0$. Let $\dot \tau_{z,1} $ be the first time $\dot \eta_{z,1}^0 $ hits $B_{\beta_1 + \dot\Delta}$ (here $\dot\Delta$ is as in the definition of $L_{z,1}$). Let $\dot h_{z,1} = h \circ f_{z,1}^{-1} - \chi \operatorname{arg} (f_{z,1}^{-1})'$. Let $\dot \eta_{z,1}^\pm$ be the flow lines of $\dot h_{z,1}$ started from $\dot\tau_{z,1}$ with angles $\mp \theta$. Let $\dot D_{z,1}$ be the connected component of $\mathbf D\setminus (\dot\eta_{z,1}^+ \cup \dot\eta_{z,1}^-)$ which contains the origin. Let $\dot p_{z,1}$ be the conformal map from $\dot D_{z,1}$ to $\mathbf D$ with $\dot p_{z,1}(0) = 0$ and $\dot p_{z,1}'(0) > 0$.
Inductively, suppose $j\geq 2$ and that $\dot \eta_{z,j-1}$, $\dot \tau_{z,j-1}$, $\dot h_{z,j-1}$, $\dot\eta_{z,j-1}^\pm$, $\dot D_{z,j-1}$, and $\dot p_{z,j-1}$ have been defined. Let $\dot\eta_{z,j}^0 := \dot p_{z,j-1}(\dot\eta_{z,j-1}^0 \cap \dot D_{z,j-1})$. Let $\dot\tau_{z,j}$ be the first time $\dot\eta_{z,j}^0$ hits $B_{\beta_1 + \dot\Delta}$. Let $\dot h_{z,j} = \dot h_{z,j-1} \circ p_{z,j-1}^{-1} - \chi \operatorname{arg} (p_{z,j-1}^{-1})'$. Let $\dot \eta_{z,j}^\pm$ be the flow lines of $\dot h_{z,j}$ started from $\dot\tau_{z,j}$ with angles $\mp \theta$. Let $\dot D_{z,j}$ be the connected component of $\mathbf D\setminus (\eta_{z,j}^+ \cup \eta_{z,j}^-)$ which contains the origin. Let $\dot p_{z,j}$ be the conformal map from $\dot D_{z,j}$ to $\mathbf D$ with $\dot p_{z,j}(0)=0$ and $\dot p_{z,j}'(0) > 0$. See Figure~\ref{dot fig} for an illustration of this construction.
\begin{figure}
\begin{center}
\includegraphics[scale=.8]{figures/drawing-flow-lines-dot-construction.pdf}
\end{center}
\caption{An illustration of two stages of the construction of the pockets $\dot D_{z,j}$. The pocket $\dot D_{z,1}$ is the region containing the origin (shown in blue) surrounded by the outermost orange curves in the first figure. The pocket $\dot D_{z,2}$ is the region containing the origin surrounded by the orange curves in the second figure. Since the orange flow lines are started at stopping times for $\eta_{z,j}^0$, we can easily compute the conditional law of the objects involved in this construction using the result of \cite{ig1}. }
\end{figure} \label{dot fig}
Let
\begin{equation*}
\widehat{\dot \eta}_{z,j}^\pm := \left( f_{z,1}^{-1} \circ \dot p_{z,1}^{-1} \circ \cdots \circ \dot p_{z,j-1}^{-1} \right)(\dot\eta_{z,j}^\pm) \quad \operatorname{and} \quad \widehat{\dot D}_{z,j} := \left( f_{z,1}^{-1} \circ \dot p_{z,1}^{-1} \circ \cdots \circ \dot p_{z,j-1}^{-1} \right)(\dot D_{z,j})
\end{equation*}
be, respectively, the flow lines for $h$ corresponding to $\dot\eta_{z,j}^\pm$ and the pocket they form surrounding $z$.
Our interest in the above objects stems from the following lemma.
\begin{lem} \label{F cond}
Define the objects $\dot\eta_{z,j}^0$, $\dot D_{z,j}$, etc. as above, and retain the notation of Section~\ref{perfect setup sec}. If $\beta_1$ (and hence each $\beta_j$ for $j\geq 1$) is chosen sufficiently large, depending only on the parameter $\dot r$, then the following is true.
\begin{enumerate}
\item The field $\dot h_{z,j}$ is conditionally independent from $h|_{\mathbf D\setminus \widehat{\dot D}_{z,j-1}}$ given $\widehat{\dot D}_{z,j}$ and $h|_{\partial \widehat{\dot D}_{z,j-1}}$. \label{dot field}
\item Let $\dot x_{z,j}$ and $\dot y_{z,j}$ be the start and end points of $\dot \eta_{z,j}^0$. Conditional on $\widehat{\dot D}_{z,j-1}$ and $h|_{\mathbf D\setminus \widehat{\dot D}_{z,j-1}}$, we have that $\dot \eta_{z,j}^0$ is the zero-angle flow line of $\dot h_{z,j}$ from $\dot x_{z,j}$ to $\dot y_{z,j}$. In particular, the law of $\dot\eta_{z,j}^0$ under this conditioning is that of a chordal $\operatorname{SLE}_\kappa(\rho^1 ; \rho^1)$ from $\dot x_{z,j}$ to $\dot y_{z,j}$ in $\mathbf D$, with $\rho^1$ as in~\eqref{rho0 rho1}. \label{dot curve}
\item On the event $E_n(z)$, we a.s.\ have $\eta_{z,j}^0 = \dot\eta_{z,j}^0$ for each $j\leq n+1$ and $\widehat{\dot D}_{z,j} = \widehat D_{z,j}$ for each $j\leq n$. \label{dot equal on E}
\end{enumerate}
\end{lem}
\begin{proof}
To obtain assertion~\ref{dot field}, we start by observing that by \cite[Theorem~1.1]{ig1} and induction, for each $j\geq 2$, the set $\dot A_{z,j-1} := \eta^{\dot T_{z,j-1}} \cup \widehat{\dot \eta}_{z,j-1}^+ \cup \widehat{\dot\eta}_{z,j-1}^-$ is a local set for $h$ in the sense of \cite[Section~3.3]{ss-contour}. Hence the conditional law of $h $ given $\dot A_{z,j-1}$ and $h|_{\dot A_{z,j-1}}$ in each complementary connected component of $\mathbf D\setminus \dot A_{z,j-1}$ is independently that of a zero-boundary GFF plus a certain $(\dot A_{z,j-1} , h|_{\dot A_{z,j-1}})$-measurable harmonic function. In particular, the conditional law of $\dot h_{z,j-1}$ given $\dot A_{z,j-1}$ is that of a GFF on $\mathbf D$ with boundary data $\lambda - \theta \chi- \chi\cdot \operatorname{winding}$ on $[\dot y_{z,j} , \dot x_{z,j}]_{\partial\mathbf D}$ and $-\lambda +\theta\chi$ on $\lambda - \theta \chi - \chi\cdot \operatorname{winding}$ on $[\dot x_{z,j} , \dot y_{z,j}]_{\partial\mathbf D}$, where $\lambda$ and $\chi$ are as in Section~\ref{ig prelim}.
By \cite[Theorem~1.2]{ig1} and locality, $\dot A_{z,j-1}$ is a.s.\ determined by $\widehat{\dot D}_{z,j-1}$ and $h|_{\mathbf D\setminus \widehat{\dot D}_{z,j-1}}$.
Hence we get the same conditional law for $\dot h_{z,j}$ if we instead condition on $\widehat{\dot D}_{z,j-1}$ and $h|_{\mathbf D\setminus \widehat{\dot D}_{z,j-1}}$.
Since this law depends only on $\widehat{\dot D}_{z,j-1}$ and $h|_{\partial \widehat{\dot D}_{z,j-1}}$, we obtain assertion~\ref{dot field}.
Assertion~\ref{dot curve} follows immediately from our description of the conditional law of $\dot h_{z,j}$ and \cite[Theorems~1.1 and~2.4]{ig1}.
It remains to prove assertion~\ref{dot equal on E}. See Figure~\ref{merge fig} for an illustration. By the Koebe distortion theorem, we have that as $\beta_j \rightarrow \infty$, $e^{\beta_j + \dot\Delta} \operatorname{diam} \psi_{z,j}(B_{\beta_j + \dot\Delta})$ and $e^{\beta_j + \dot\Delta} \operatorname{dist}(0 , \partial \psi_{z,j}(B_{\beta_j + \dot\Delta}))$ tend uniformly to $|\psi_{z,j}'(0)|^{-1} $. Thus if $\beta_1$ (and hence $\beta_j$) is chosen sufficiently large, then by condition~\ref{L deriv} in the definition of $L_{z,j}$ we have
\begin{equation} \label{psi balls}
B_{\beta_j + \dot r} \subset \psi_{z,j}(B_{\beta_j + \dot\Delta}) \subset B_{\beta_j} .
\end{equation}
By~\eqref{psi balls} and condition~\ref{F b} in the definition of $F_{z,1} $, it follows that on $E_1(z)$, the starting point for the curves $\psi_{z,1}(\dot \eta_{z,1}^\pm)$ is disconnected from $\partial\mathbf D$ by the parts of $\eta_{z,1}^\pm$ traced before they hit $b_{z,1}$. Therefore, $ \eta_{z,1}^\pm $ a.s.\ hit and (by \cite[Theorem~1.5]{ig1}) subsequently merge with $\psi_{z,1}(\dot \eta_{z,1}^\pm)$ before reaching $b_{z,1}$. This proves assertion~\ref{dot equal on E} in the case $n=1$. The general case follows from~\eqref{psi balls} and induction.
\end{proof}
\begin{figure}
\begin{center}
\includegraphics[scale=.8]{figures/drawing-flow-lines-merge.pdf}
\end{center}
\caption{An illustration of the proof of assertion~\ref{dot equal on E} of Lemma~\ref{F cond} in the case $j=1$. The inner dotted quasi-circle is the image of $B_{\beta_1 + \dot\Delta }$ under $\psi_{z,1}$. The orange flow lines $\dot\psi_{z,1}(\dot\eta_{z,1}^{\pm })$ quickly merge with $\eta_{z,1}^\pm$ and form the same pocket around zero as the flow lines $\eta_{z,1}^\pm$ on the event $E_{z,1} $.}
\end{figure} \label{merge fig}
Note that Lemma~\ref{F cond} implies in particular that the conditional law of $\eta_{z,n}^0$ given $\widehat D_{z,n-1}$ and $h|_{\mathbf D\setminus \widehat D_{z,n-1}}$ on the event $E_{n-1}(z)$ is that of a chordal $\operatorname{SLE}_\kappa(\rho^1 ; \rho^1)$ process from $x_{z,n}$ to $y_{z,n}$ in $\mathbf D$. Furthermore, $\eta_{z,n}^0$ is equal to a flow line of $h|_{\widehat D_{z,n-1}}$ on this event. Since the parts of $\eta$ and the auxiliary flow lines which generated $\mathcal F_{z,n-1}$ all lie outside $\widehat D_{z,n-1}$, \cite[Theorem~1.2]{ig1} implies that $\mathcal F_{z,n-1}$ is a.s.\ determined by $\widehat D_{z,n-1}$ and $h|_{\mathbf D\setminus \widehat D_{z,n-1}}$ on the event $E_{n-1}(z)$.
For $n\geq 1$, define the event $\dot L_{z,n}$ and the curve $\dot \eta_{z,n}$ in the same manner as the event $L_{z,n}$ and the curve $\dot \eta_{z,n}$ but with $\dot\eta_{z,n}^0$ in place of $\eta_{z,n}^0$. Also fix $\dot a \in (0,2)$ and let
\begin{equation} \label{dot K}
\dot K_{z,n} = \dot K_{z,n}(\dot a) := \left\{ |\dot x_{z,n+1} - \dot y_{z,n+1}| \geq \dot a \right\} .
\end{equation}
We can now prove the analogue of Lemma~\ref{L prob} for $n\geq 2$.
\begin{lem} \label{L prob j}
Suppose we are in the setting of Lemma~\ref{F cond}. If the parameters involved in the definition of $L_{z,1}$ are chosen appropriately, independently of $n$, $\beta_n$, $u_n$, and $z \in B_d(0)$ then we can find a deterministic $p > 0$ which does not depend on $n$, $\beta_n$, $u_n$, and $z \in B_d(0)$ such that for each $n\geq 2$, we have
\begin{equation*}
\mathbf P\left(L_{z,n} \,|\, \mathcal F_{z,n-1} \right) \mathbf 1_{E_{n-1}(z)} \geq p \mathbf 1_{E_{n-1}(z)} ,\qquad \mathbf P\left(\dot L_{z,n} \,|\, \widehat{\dot D}_{z,n-1} , \, h|_{\mathbf D\setminus \widehat{\dot D}_{z,n-1}} \right) \mathbf 1_{\dot K_{z,n-1 }} \geq p \mathbf 1_{\dot K_{z,n-1 }} .
\end{equation*}
\end{lem}
\begin{proof}
We prove only the statement for $ L_{z,n}$. The statement for $\dot L_{z,n}$ is proven in exactly the same manner. Let $A$ be the event that the part of $ \eta_{z,n}^0$ lying between $ \eta_{z,n}^0 (\sigma_{z,1} )$ and $\overline\eta_{z,1}^0(\overline\sigma_{z, 1} )$ never exits $B_{\widetilde\Delta}$. Let $B$ be the event that all of the conditions in the definition of $L_{z,n}$ except possibly condition~\ref{L cond} occur.
In light of Lemma~\ref{F cond}, we can apply~\cite[Lemma~2.5]{miller-wu-dim} to get that if the parameters are chosen appropriately, independently of $n$,
then $\mathbf P\left(A\cap B \,|\, \mathcal F_{z,n-1} , \right) \mathbf 1_{E_{n-1}(z)} \succeq \mathbf 1_{E_{n-1}(z)}$ (here we use condition~\ref{F hm} in the definition of $F_{z,n-1}$ to get that $x_{z,n}$ and $y_{z,n}$ lie at uniformly positive distance from one another on $E_{n-1}(z)$). Hence $\mathbf P\left(A \,|\, B,\, \mathcal F_{z,n-1} \right) \mathbf 1_{E_{n-1}(z)} \succeq 1$. It follows that there exists a sub-event of $B$ with uniformly positive probability and a $p_L > 0$ on which $\mathbf P\left(A\,|\, (\eta_{z,n}^0)^{\sigma_{z,n}} \cup (\overline\eta_{z,n}^0)^{\overline \sigma_{z,n}} ,\, \mathcal F_{z,n-1} \right) \mathbf 1_{E_{n-1}(z)} \geq p_L \mathbf 1_{E_{n-1}(z)}$. On this sub-event, $L_{z,n}$ occurs.
\end{proof}
For $n \in \mathbf N$ and $j\in\mathbf N$, define flow lines $\dot \eta_{z,j}^{n , \pm}$, times $\widetilde{\dot t}_{z,j}^{n,\pm}$ and $\dot t_{z,j}^{n,\pm}$, and events $\dot E_{z,j}^n$ in the same manner as $\eta_{0,j}^\pm$, $\widetilde t_{0,j}^\pm$, $t_{z,j}^\pm$, and $E_{0,j}$ but with $h$ replaced by $\dot h_{z,n}$, $\eta_{0,1}^0 = \eta$ replaced by $\dot\eta_{z,n}^0$, and the sequence $(\beta_j , u_j)_{j\in\mathbf N}$ replaced by $(\beta_{n+j} , u_{n+j})_{j\in\mathbf N}$ (we retain the above definition of $\dot L_{z,n}$). Also let
\begin{equation} \label{dot E cap}
\dot E_{k }^n(z) := \bigcap_{j=1}^k \dot E_{z,j}^n .
\end{equation}
\begin{lem}\label{eta abs cont}
Suppose we are in the setting of Section~\ref{perfect setup sec}. If the auxiliary parameters are chosen appropriately, independently of $n$, $(\beta_j , u_j)_{j\geq 1}$, and $z \in B_d(0)$, and if $\beta_1$ is chosen sufficiently large, then the following three laws are a.s.\ \hyperref[smac]{s.m.a.c.}\ (Definition~\ref{smac}) with deterministic (i.e., independent of whatever realization we are conditioning on) constants uniform in $n$, $(\beta_j , u_j)_{j\geq 1}$, and $z \in B_d(0)$.
\footnote{The flow lines $\eta_{z,j}^\pm$ are only defined on the event $\eta_{z,j}$ hits a certain ball centered at the origin. We take these flow lines to be equal to a ``graveyard point" in our probability space on the event that $\eta_{z,j}$ does not hit such a ball.}
\begin{enumerate}
\item The conditional joint law of $ \eta_{z,n}$ and $\left\{ (\eta_{z,j}^+)^{ \widetilde t_{z,j}^+} , (\eta_{z,j}^-)^{\widetilde t_{z,j}^-} \right\}_{j\geq n}$ given any realization of $\mathcal F_{z,n-1} \vee \mathcal F_{z,n}^0$ for which $E_{n-1}(z) \cap L_{z,n} $ occurs, restricted to the event $\mathcal G'(\eta_{z,n} , \mu) $. \label{eta n law}
\item The conditional joint law of $\dot \eta_{z,n}$ and $\left\{ (\dot \eta_{z,j}^{n , +})^{ \widetilde{\dot t}_{z,j}^{n,+}} , (\dot \eta_{z,j}^{n , -})^{ \widetilde{\dot t}_{z,j}^{n,-}} \right\}_{j\geq 1}$ given any realization of $\widehat{\dot D}_{z,n-1}$ and $h|_{\mathbf D\setminus \widehat{\dot D}_{z,n-1}}$ for which $\dot K_{z,n-1} \cap \dot L_{z,n}$ occurs, restricted to the event $\mathcal G'(\dot\eta_{z,n} , \mu) $. \label{dot n law}
\item The conditional joint law of $\eta_{z,1}$ and $\left\{ (\eta_{z,1}^+)^{\widetilde t_{z,1}^+} , (\eta_{z,1}^-)^{\widetilde t_{z,1}^-}\right\}_{j \geq 1}$ given any realization of $\mathcal F_{z,1}^0$ for which $L_{z,1}$ occurs, restricted to the event $\mathcal G'(\eta_{z,1} , \mu)$, with the sequence $(\beta_j , u_j)_{j\in\mathbf N}$ replaced by $(\beta_{n+j} , u_{n+j})_{j\in\mathbf N}$. \label{eta 1 law}
\end{enumerate}
\end{lem}
\begin{proof}
Throughout, we assume that all implicit constants for \hyperref[smac]{s.m.a.c.}\ are deterministic, independent of $n$ and of $(\beta_j , u_j)_{j\geq 1}$, and uniform for $z \in B_d(0)$. We will prove only the strict mutual absolute continuity of the laws~\ref{eta n law} and~\ref{eta 1 law}. Strict mutual absolute continuity of of the laws~\ref{eta n law} and~\ref{dot n law} is immediate from assertions~\ref{dot field} and~\ref{dot equal on E} of Lemma~\ref{F cond}.
Let $\omega$ be a realization of $\mathcal F_{z,n-1} \vee \mathcal F_{z,n}^0$ for which $E_{n-1}(z) \cap L_{z,n}$ occurs. Let $A$ be the event that the part of $\eta_{z,n}^0$ between $\eta_{z,n}^0(\sigma_{z,n})$ and $\overline\eta_{z,n}^0(\overline\sigma_{z,n})$ never exits $B_{\widetilde\Delta}$.
By Lemma~\ref{F cond} the conditional law of $\eta_{z,n}^0$ given any realization of $\mathcal F_{z,n-1}$ for which $E_{n-1}(z)$ occurs is that of an $\operatorname{SLE}_\kappa(\rho^1 ; \rho^1)$ process from $x_{z,n}$ to $y_{z,n }$ in $\mathbf D$ with force points located on either side of $x_{z,n}$.
Since the start and end points of $\eta_{z,n}$ are equal to $-i$ and $i$, Lemma~\ref{rho abs cont} implies that (provided $\widetilde\Delta$ is chosen sufficiently large) the conditional law of $\eta_{z,n}$ given $\omega$ and $A$ is \hyperref[smac]{s.m.a.c.}\ with respect to the law of a chordal $\operatorname{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$ given (equivalently, restricted to) the event that it never leaves $\psi_{z,n}(B_{\widetilde\Delta})$. By condition~\ref{L cond} in the definition of $L_{z,n}$, the same is true of the conditional law of $\eta_{z,n}$ given $\omega$, \emph{restricted} to the event $A$. By condition~\ref{L ball} in the definition of $L_{z,n}$ we have $\mathcal G'(\eta_{z,n} , \mu) \cap L_{z,n} \subset A$, so the same is also true of the conditional law of $\eta_{z,n}$ given $\omega$, restricted to the event $\mathcal G'(\eta_{z,n} , \mu)$.
By the results of \cite[ Section~7]{ig1}, the conditional law of $\eta_{z,n}^+ $ given $\omega$ and $\eta $ is that of a chordal $\operatorname{SLE}_\kappa(\rho^0 ; \rho^1 , \rho^2)$ from $\eta_{z,n}(\tau_{z,n})$ to $\psi_{z,n}(y_{z,n}^-)$ in the connected component of $\mathbf D \setminus \eta_{z,n} $ containing the right side of $-i$ on its boundary, with $\rho^0 , \rho^1$ as in~\eqref{rho0 rho1} and $\rho^2$ depending only on $\theta$ and $\kappa$. The force points corresponding to the weights $\rho^0$ and $\rho^1$ are located on either side of $\eta_{z,n}(\tau_{z,n})$.
The extra force point corresponding to $\rho^2$ is located at $\psi_{z,n}(x_{z,n}^+)$.
By \cite[Lemma~2.8]{miller-wu-dim} the conditional law of $(\eta_{z,n}^+)^{\widetilde t_{z,n}^+} $ given $\omega$ and any realization of $\eta_{z,n}$ for which $\mathcal G'(\eta_{z,n},\mu)$ occurs is a.s.\ \hyperref[smac]{s.m.a.c.}\ with respect to the law of a chordal $\operatorname{SLE}_\kappa(\rho^0; \rho^1)$ from $ \eta_{z,n}(\tau_{z,n}) $ to any given point $v$ on the right boundary of the aforementioned connected component of $\mathbf D \setminus \eta_{z,n} $ stopped at an appropriate time.
A similar statement holds for $(\eta_{z,n}^-)^{\widetilde t_{z,n}^-}$.
Since $\eta_{z,n}^\pm$ are conditionally independent given $\eta_{z,n}^0$ and $\mathcal F_{z,n-1}$, it follows that the joint law of $((\eta_{z,n}^+)^{\widetilde t_{z,n}^+} , (\eta_{z,n}^-)^{\widetilde t_{z,n}^-})$ given $\omega$ and $\omega_\eta$ is a.s.\ \hyperref[smac]{s.m.a.c.}\ with respect to the law of a pair of curves with the same description as the joint law of $((\eta_{z,1}^+)^{\widetilde t_{z,1}^+} , (\eta_{z,1}^-)^{\widetilde t_{z,1}^-})$ given $\eta_{z,1}$ and $\mathcal F_{z,1}^0$, restricted to the event $L_{z,1} \cap \mathcal G'(\eta_{z,1} , \mu)$, but with $\beta_1$ replaced by $\beta_n$.
It remains to deal with the conditional laws of the remaining flow lines. Let $h_{z,n+1} = h\circ \widehat p_{z,n}^{-1} - \chi \operatorname{arg} \widehat p_{z,n}^{-1}$. By the results of \cite[Section~7]{ig1}, the conditional law of $h_{z,n+1}$ given $\{ \eta , (\eta_{z,n}^+)^{ t_{z,n}^+} , (\eta_{z,n}^-)^{ t_{z,n}^-}\}$ and $\mathcal F_{z,n-1} $ is that of an independent GFF in each of the complementary connected components of $\mathbf D\setminus \eta_{z,n+1}^0$ with boundary data depending only on $\theta$ and $\kappa$.
By \cite[Theorem~1.2]{ig1}, under this conditioning the collection of curves
\[
\left\{( \widehat p_{z,n}( (\widehat \eta_{z,j}^+)^{\widetilde t_{z,j}^+} ) , \widehat p_{z,n}( (\widehat \eta_{z,j}^-)^{\widetilde t_{z,j}^-} ) ) : j \geq n+1 \right\}
\]
is a.s.\ determined by $h_{z,n+1}$, $\eta_{z,n}$, $(\eta_{z,n}^+)^{ t_{z,n}^+}$, and $ (\eta_{z,n}^-)^{ t_{z,n}^-}$, in the same manner that the collection of curves
\[
\left\{( \widehat p_{z,1}( (\widehat \eta_{z,j}^+)^{\widetilde t_{z,j}^+} ) , \widehat p_{z,1}( (\widehat \eta_{z,j}^-)^{\widetilde t_{z,j}^-} ) ) : j \geq 2 \right\}
\]
is determined by $h_{z,2}$, $\eta_{z,1}$, $(\eta_{z,1}^+)^{ t_{z,1}^+}$, and $ (\eta_{z,1}^-)^{ t_{z,1}^-}$, except with $(\beta_j , u_j)_{j\in\mathbf N}$ replaced by $(\beta_{n+j} , u_{n+j})_{j\in\mathbf N}$. By combining everything, we get the statement of the lemma.
\end{proof}
\begin{lem} \label{E_0^n split}
Let $z \in \mathbf D$, $m\in\mathbf N$, and $k\leq n \in \mathbf N$. Recall the definitions of the events from Section~\ref{perfect setup sec'}. We have
\begin{equation} \label{E_0^n split eqn}
\mathbf P\left(E_{0,n}^m(z)\right) \asymp \mathbf E\left( E_{0,k}^m(z)\right) \mathbf P\left(E_{0,n-k}^{m+k}(z)\right)
\end{equation}
with proportionality constants independent of $n$, $m$, and $k$ and uniform for $z \in B_d(0)$.
\end{lem}
\begin{proof}
We have
\[
\mathbf P\left(E_{0,n}^m(z)\right) = \mathbf P\left(E_{0,k}^m(z)\right) \mathbf P\left(E_{k,n}^{m } (z) \, |\, E_{0,k}^m(z) \right) .
\]
Furthermore
\begin{equation} \label{E to L}
\mathbf P\left(E_{k,n}^{m } (z) \, |\, E_{0,k}^m(z) \right) = \mathbf P\left(L_{z,k+1} \, |\, E_{0,k}^m(z) \right) \mathbf P\left(E^{m }_{k,n}(z) \, |\, L_{z,k+1} \cap E_{0,k}^m(z) \right) .
\end{equation}
The first factor on the right in~\eqref{E to L} is bounded below by a positive constant by Lemma~\ref{L prob j}. Conditional on $\mathcal F_{z,k+1}^0 \vee \mathcal F_{z,k}$, the event $E_{k,n}^{m } (z)$ is determined by $\eta_{z,k+1}$ and the auxiliary flow lines $(\eta_{z,j}^\pm)^{\widetilde t_{z,j}^\pm}$ for $j\geq k+1$. Hence that Lemma~\ref{eta abs cont} implies $ \mathbf P\left(E_{k,n}^{m } (z) \, |\, L_{z,k+1} \cap E_m^k(z) \right) \asymp \mathbf P\left(E_{0,n-k}^{m+k}(z)\right)$, as required.
\end{proof}
\begin{lem} \label{P(E_z)}
For each $z\in B_d(0)$ and each $n,m\in \mathbf N$, we have
\[
e^{ - \overline \beta_{m,m+n} \gamma^*(q) - 3 \overline u_{m,m+n} \gamma_0^*(q) } \preceq \mathbf P\left(E_{0,n}^m(z) \right) \preceq e^{ - \overline \beta_{m,m+n} \gamma^*(q) + 3 \overline u_{m,m+n} \gamma_0^*(q) }
\]
with the implicit constants independent of $n$ and uniform for $z\in B_d(0)$.
\end{lem}
\begin{proof}
Let $C_*^{-1}$ and $C_*$ be the proportionality constants in~\eqref{E_0^n split eqn}. By Lemma~\ref{E_0^n split} we have
\[
C_*^{-n} \prod_{j=0}^{n-1} \mathbf P\left(E_{z,1}^{m+j} \right) \leq \mathbf P\left(E_{0,n}^m(z)\right) \preceq C_*^n \prod_{j=0}^{n-1} \mathbf P\left(E_{z,1}^{m+j} \right) ,
\]
with the implicit constant independent of $\beta$, $n$, and $m$ and uniform for $z$ in compacts. By~\eqref{C_u} we then have
\begin{equation} \label{E_z eqn 1}
e^{-\overline\beta_{m,m+n} \gamma^*(q) - \overline u_{m,m+n} \gamma_0^*(q)} C_*^{-n} \prod_{j=m }^{m+n-1} C_{u_j}^{-1} \leq \mathbf P(E_{0,n}^m(z)) \leq e^{-\overline\beta_{m,m+n} \gamma^*(q) - \overline u_{m,m+n} \gamma_0^*(q)} C_*^{ n} \prod_{j=m }^{m+n-1} C_{u_j} .
\end{equation}
By condition~\ref{beta u infty} in Lemma~\ref{beta u choice} we have $C_*^n \preceq e^{\overline u_{m,m+n} \gamma_0^*(q) }$. By condition~\ref{beta u C} in Lemma~\ref{beta u choice} we also have $\prod_{j=m }^{m+n-1} C_{u_j} \leq e^{\overline u_{m,m+n} \gamma_0^*(q) }$. By plugging these estimates into~\eqref{E_z eqn 1} we get the desired conclusion.
\end{proof}
\begin{lem} \label{near independence}
Let $z,w\in B_d(0)$. Let $\lambda$ be as in assertion~\ref{eta dist} of Lemma~\ref{E_z basics}. Choose $ k \in \mathbf N$ such that $e^{- \overline \beta_{k +1} -\lambda (k+1) } \leq |z-w| \leq e^{-\overline \beta_{k } - \lambda k }$. Then for any $n \in \mathbf N$ with $\overline\beta_n -\lambda n \geq \overline \beta_{k +1} + \lambda (k+2) $, we have
\begin{equation*}
\mathbf P\left( E_{k ,n} (z) \cap E_{k ,n} (w) \,|\, E_k(z) \cap E_k(w)\right) \preceq e^{\overline\beta_k o_k(1)} \mathbf P( E^{k }_{0,n-k }(z) ) \mathbf P( E^{k }_{0,n-k }(w))
\end{equation*}
with the implicit constants independent of $n$ and $k$, the $o_k(1)$ independent of $n$, and both deterministic and uniform for $z,w \in B_d(0)$.
\end{lem}
\begin{proof}
Throughout, we require implicit constants and $o_k(1)$ terms to satisfy the conditions of the statement of the lemma.
Let $k'$ be the least integer such that $\overline\beta_{k'} -\lambda k' \geq \overline \beta_{k +1} + \lambda (k+2) $. Note $k'\leq n$.
Let $\dot P_{z,k'}$ be the event that $\operatorname{diam}( \widehat{\dot D}_{z,k'} ) \leq \tfrac12 e^{- \overline\beta_{k'} + \lambda k' }$ and the event $\dot K_{z,k'}$ defined in~\eqref{dot K} occurs. Define $\dot P_{w,k'}$ similarly. By Lemma~\ref{F cond}, we have $\widehat{\dot D}_{z,k'} = \widehat D_{z,k'}$ on $E_{k'}(z)$, so the objects involved in the definition~\eqref{dot E cap} of $\dot E_{0,n-k'}^{k'}(z)$ agree with the objects involved in the definition of $E_{k',n}(z)$ on this event. Similar statements hold with $w$ in place of $z$. By combining this observation with Condition~\ref{E top} in Subsection~\ref{2pt setup sec} and assertion~\ref{wh D diam} of Lemma~\ref{E_z basics}, we get that if $\dot a$ (as defined just above~\eqref{dot K}) is chosen sufficiently small, depending only on $a$, then
\begin{equation*}
E_n(z) \subset \dot E_{n-k'}^{k'} (z) \cap \dot P_{z,k'} ,\quad \operatorname{and} \quad E_n(w) \subset \dot E_{n-k'}^{k'} (w) \cap \dot P_{w,k'} .
\end{equation*}
Therefore,
\begin{align} \label{E to dot E}
&\mathbf P\left( E_{k ,n}^{k }(z) \cap E_{k ,n}^{k }(w) \,|\, E_k(z) \cap E_k(w)\right) = \mathbf P\left(E_n(z) \cap E_n(w) \,| \, E_k(z) \cap E_k(w) \right) \nonumber\\
&\qquad \leq \mathbf P\left(\dot E_{n-k'}^{k'} (z) \cap \dot P_{z,k'} \cap \dot E_{n-k'}^{k'} (w) \cap \dot P_{w,k'} \,| \, E_k(z) \cap E_k(w) \right) \nonumber\\
&\qquad= \mathbf P\left(\dot E_{n-k'}^{k'} (z) \cap \dot E_{n-k'}^{k'} (w) \,| \, E_k(z) \cap E_k(w) \cap \dot P_{z,k'} \cap \dot P_{w,k'} \right) \mathbf P\left( \dot P_{z,k'} \cap \dot P_{w,k'} \right) .
\end{align}
So, we need only estimate the last line of~\eqref{E to dot E}.
On the event $\dot P_{z,k'} \cap \dot P_{w,k'}$, the domains $\widehat{\dot D}_{z,k'}$ and $\widehat{\dot D}_{w,k'}$ are disjoint.
By assertion~\ref{wh D diam} of Lemma~\ref{E_z basics}, on the event $E_k(z) \cap E_k(w) \cap \dot P_{z,k'} \cap \dot P_{w,k'}$ we have $\widehat{\dot D}_{z,k'} \cup \widehat{\dot D}_{w , k'} \subset \widehat D_{z,k} \cap \widehat D_{w,k}$. The event $E_k(z)$ is determined by $\widehat D_{z, k}$ and $h|_{\mathbf D\setminus \widehat D_{z,k}}$.
Consequently, the event $E_k(z) \cap \dot P_{z,k'}$ is determined by $\widehat{\dot D}_{z, k'}$ and $h|_{\mathbf D\setminus \widehat{\dot D}_{z, k'}}$.
Similar statements hold with $w$ in place of $z$.
Let $\mathcal H$ be the $\sigma$-algebra generated by $\widehat{\dot D}_{z,k'}$, $\widehat{\dot D}_{w,k'}$, $h|_{\partial \widehat{\dot D}_{z,k'}}$, and $h|_{\partial \widehat{\dot D}_{w,k'}}$.
The above considerations together with Lemma~\ref{F cond} imply that the events $\dot E_{n-k'}^{k'} (z)$ and $\dot E_{n-k' }^{k'}(w)$ are conditionally independent given $E_k(z) \cap E_k(w) $ and $\mathcal H$ on the event $\dot P_{z,k'} \cap\dot P_{w,k'}$.
We thus have
\begin{align} \label{dot E cond}
&\mathbf P\left(\dot E_{n-k'}^{k'} (z) \cap \dot E_{n-k' }^{k'}(w) \,|\, \mathcal H ,\, E_k(z) \cap E_k(w) \right) \mathbf 1_{ \dot P_{z,k'} \cap \dot P_{w,k'}} \nonumber \\
&= \mathbf P\left(\dot E_{n-k'}^{k'} (z) \,|\, \mathcal H ,\, E_k(z) \cap E_k(w) \right) \mathbf P\left(\dot E_{n-k'}^{k'} (w) \,|\, \mathcal H ,\, E_k(z) \cap E_k(w) \right) \mathbf 1_{ \dot P_{z,k'} \cap \dot P_{w,k'}} .
\end{align}
By Lemmas~\ref{L prob j} and~\ref{eta abs cont} we have
\begin{equation} \label{dot E prob}
\mathbf P\left(\dot E_{n-k'}^{k'} (z) \,|\, \mathcal H ,\, E_k(z) \cap E_k(w) \right) \mathbf 1_{ \dot P_{z,k'} \cap \dot P_{w,k'}} \asymp \mathbf P\left(E_{0,n-k'}^{k'}(z)\right) \mathbf 1_{ \dot P_{z,k'} \cap \dot P_{w,k'}}
\end{equation}
and similarly with $z$ and $w$ interchanged. By~\eqref{E to dot E},~\eqref{dot E cond}, and~\eqref{dot E prob},
\begin{equation} \label{k,n to n-k'}
\mathbf P\left( E_{k ,n}^{k }(z) \cap E_{k ,n}^{k }(w) \,|\, E_k(z) \cap E_k(w)\right) \preceq \mathbf P\left(E_{0,n-k'}^{k'}(z)\right) \mathbf P\left(E_{0,n-k'}^{k'}(w)\right) .
\end{equation}
By Lemma~\ref{E_0^n split} we have
\begin{equation} \label{3 E's}
\mathbf P\left(E_{0,n-k'}^{k'}(z)\right) \asymp \frac{\mathbf P\left(E_{0,n-k}^{k}(z) \right)}{\mathbf P\left(E_{0,k'-k}^k(z) \right) } .
\end{equation}
By Lemma~\ref{P(E_z)},
\[
\mathbf P\left(E_{0,k'-k}^k(z) \right) \succeq e^{ - \overline \beta_{k , k'} \gamma^*(q) - 3 \overline u_{k , k'} \gamma_0^*(q) } .
\]
By our choice of $k'$ we have $\overline\beta_{k , k'} \leq \lambda (k + k'+1) + \beta_{k' }$. Since $\beta_j$ is increasing in $j$ we have $k' - k = o_k(1) \overline \beta_k$ and $\lambda (k + k'+1) = o_k(1) \overline\beta_k$. By condition~\ref{beta u k} in Lemma~\ref{beta u choice} we have $\beta_{k' } \leq \beta_{k+o_k(1)} \leq \overline\beta_k o_k(1)$. Therefore $\mathbf P\left(E_{0,k'-k}^k(z) \right) = e^{k o_k(1)} \overline \beta_k$.
Hence~\eqref{3 E's} implies
\[
\mathbf P\left(E_{0,n-k'}^{k'}(z)\right) \preceq e^{\overline\beta_k o_k(1)} \mathbf P\left(E_{0,n-k}^{k}(z) \right) .
\]
A similar assertion holds with $w$ in place of $z$. We conclude by combining this with~\eqref{k,n to n-k'}.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{2pt estimate}]
We have
\alb
\mathbf P\left( E_n(z) \cap E_n(w)\right) &= \mathbf P\left( E^k_{k,n}(z) \cap E^k_{k,n}(w) \cap E_k(z) \cap E_k(w) \right) \quad \text{(by definition}) \\
&\leq \mathbf P\left( E^k_{k,n}(z) \cap E^k_{k,n}(w) \,|\, E_k(z) \cap E_k(w) \right) \mathbf P\left( E_k(z) \right) \\
&\preceq e^{\overline\beta_k o_k(1)} \mathbf P\left( E^k_{0,n-k}(z) \right) \mathbf P\left( E^k_{0,n-k}(w)\right) \mathbf P\left( E_k(z) \right) \quad \text{(by Lemma~\ref{near independence})} .
\ale
By Lemma~\ref{E_0^n split} we have
\[
\mathbf P\left( E^k_{0,n-k}(w)\right) \asymp \frac{\mathbf P\left( E_n(w)\right)}{\mathbf P\left( E_k(w)\right)} , \qquad \mathbf P\left( E^k_{0,n-k}(z) \right) \mathbf P\left( E_k(z) \right) \asymp \mathbf P\left( E_n(z)\right)
\]
By combining the above relations we get~\eqref{2pt estimate eqn}.
\end{proof}
\subsection{Setup}\label{lower setup sec}
Let $\eta$ be a chordal $\operatorname{SLE}_\kappa$ from $-i$ to $i$ in $\mathbf D$. Let $D_\eta$ be as in Theorem~\ref{main thm} and define the sets $\widetilde\Theta^{s }(D_\eta)$ and $\Theta^s(D_\eta)$ as in Section~\ref{multifractal def}. The goal of this section is to obtain lower bounds on $\dim_{\mathcal H} \widetilde\Theta^s( D_\eta)$ and $\dim_{\mathcal H} \Theta^s(D_\eta)$, and thereby complete the proof of Theorem~\ref{main thm}. We accomplish this using the estimates of Section~\ref{2pt sec}.
Throughout this section we use the notation defined in Sections~\ref{perfect setup sec} and~\ref{perfect setup sec'}, with
\[
q = \frac{s}{1-s} .
\]
In particular, we recall the definition~\eqref{mathcal P_n def} of the $n$-perfect points and the definition of the exponents $\gamma^*(q)$ and $\gamma_0^*(q)$ from~\eqref{gamma* def}.
In the next two sections we will use the $n$-perfect points to define various notions of ``perfect points" which are contained in the sets we are interested in and which will allow us to obtain lower bounds on their Hausdorff dimensions.
In order to prove that the perfect points are contained in our sets of interest, we will need the following technical lemma.
\begin{lem} \label{I stuff}
Let $\Psi_\eta : D_\eta \rightarrow \mathbf D$ be as in Section~\ref{time infty setup sec}. Suppose $z\in \mathcal P_k \cap D_\eta$. For $n \leq k-1$ let $ I_{z,n}$ be the image under $\Psi_\eta$ of $\eta \cap \widehat D_{z,n}$. If the parameters for the events of Section~\ref{perfect setup sec} are chosen appropriately, independently of $n$ and of $z\in B_d(0)$, we have the following.
\begin{enumerate}
\item We have $ e^{-\overline\beta_n (q+1 ) - 3\overline u_n} \preceq \operatorname{length} I_{z,n} \preceq e^{-\overline\beta_n (q+1 ) + 3\overline u_n}$. \label{I length}
\item If $n\leq k-2$ the distance from $I_{z,n+1}$ to $\partial I_{z,n}$ is proportional to the length of $I_{z,n}$. \label{I nest}
\item If $x\in I_{z,n}$ then there exists $\delta_n > 0$ such that $|(\Psi_\eta^{-1})'((1-\delta_n)x)| \asymp \delta_n^{s+o_n(1)}$ and $\delta_n \asymp e^{-\overline\beta_n (q+1+o_n(1))}$.\label{I deriv}
\end{enumerate}
The implicit constants are independent of $n$ and both the $o_n(1)$ and the implicit constants are deterministic and independent of $k$, $x$, and $z\in B_d(0)$
\end{lem}
\begin{proof}
Fix such a $z$, $n$, and $k$ as in the statement of the lemma. Throughout the proof we assume $E_k(z)$ occurs and require all constants (either referred to as such or implicit in $\asymp$, etc.) to be independent of $n , k , x$, and $z\in B_d(0)$. See Figure~\ref{I stuff fig} for an illustration of the argument.
Recall the pocket $\widehat D_{z,n}$ formed by the auxiliary flow lines $\widehat\eta_{z,n}^\pm$ from Section~\ref{perfect setup sec}. The map $\widehat p_{z,n } : \widehat D_{z,n } \rightarrow \mathbf D$ defined in Section~\ref{perfect setup sec} takes $z$ to 0 and $\eta \cap \widehat D_{z,n}$ to the curve $\eta_{z,n+1}^0$, whose endpoints are $x_{z,n+1}$ and $y_{z,n+1}$. Note that condition~\ref{E top} in the definition of $E_{\beta_n}^{q;u_n}(\cdot)$ implies lower bound on $|x_{z,n+1} -y_{z,n+1}|$, depending only on the parameter $a$.
By conditions~\ref{L hit} and~\ref{L ball} in the definition of $L_{z,n+1}$, there is a unique arc $A^0$ of $\partial B_{\widetilde\Delta/2}$ which disconnects $\eta_{z,n+1}^0 \cap B_{\widetilde\Delta}$ from $[x_{z,n+1}^* , y_{z,n+1}^*]_{\partial\mathbf D}$ in $\mathbf D\setminus \eta_{z,n+1}^0$ (c.f.\ Remark~\ref{E stay remark}). Let $w^0$ be the point of $A^0$ closest to the midpoint of $[x_{z,n+1}^* , y_{z,n+1}^*]_{\partial\mathbf D}$. Let $D^0$ be the connected component of $\mathbf D\setminus \eta_{z,n+1}^0$ containing $[x_{z,n+1}^* , y_{z,n+1}^*]_{\partial\mathbf D}$ on its boundary.
Observe that the harmonic measure from $w^0$ in $D^0$ of any sub-arc of $[x_{z,n+1}^* , y_{z,n+1}^*]_{\partial\mathbf D}$ is proportional to the length of that sub-arc. Furthermore, $\operatorname{hm}^{w^0}(\eta_{z,n+1}^0 ; D^0) \asymp 1$. Define $\psi^F_{z,n }$ as in the definition of $F_{z,n}$.
By condition~\ref{E top} in the definition of $E_{\beta_n}^{q;u_n}$ and condition~\ref{E_z stay} in the definition of $E_{z,n}$, the arc of $\partial\mathbf D$ which is the image of the right side of $\eta_{z,n}^{\tau_{z,n}^*}$ (resp. the left side of $\overline\eta_{z,n}^{\overline\tau_{z,n}^*}$) under $\psi_{z,n}^F$ has length $\asymp 1$. By conformal invariance of Brownian motion and condition~\ref{F hm} the definition of $F_{z,n}$, provided $r$ is chosen sufficiently small relative to $\widetilde\Delta$, the harmonic measure from $(\psi_{z,n}^F \circ p_{z,n}^{-1} ) (w^0) $ in the right connected component of $\mathbf D\setminus \psi_{z,n}^F(\eta_{z,n}^0)$ of each of these two sub-arcs is $\asymp 1$.
Let $w = \widehat p_{z,n }^{-1} ( w^0)$. It follows from the above considerations and conformal invariance of Brownian motion that (notation as in Section~\ref{basic notations})
\begin{equation} \label{hm w}
\operatorname{hm}^w ( \eta^{T_{z,n}^*} ; D_\eta) \asymp \operatorname{hm}^w(\overline \eta^{\overline T_{z,n}^*} ;D_\eta) \asymp 1, \quad \operatorname{and} \quad \operatorname{hm}^w(\eta \cap \widehat D_{z,n} ; D_\eta) \asymp 1 .
\end{equation}
By Lemma~\ref{phi hm arc} and condition~\ref{L hit} in the definition of $L_{z,1}$, we thus have
\begin{equation} \label{dist and deriv}
|\Psi_\eta'(w)| \asymp |\Phi_{z,n}'(w)|,\quad \operatorname{and} \quad \operatorname{dist}(w ,\eta) \asymp \operatorname{dist}\left( w, \eta^{T_{z,n}^*} \cup \overline\eta^{\overline T_{z,n}^*}\right) .
\end{equation}
If $\widetilde\Delta$ is chosen sufficiently large, independently of $n$ and $z\in B_d(0)$, then by the Koebe growth theorem applied to $\widehat p_{z,n}^{-1}$, we have $|w - z| \leq \frac12 \operatorname{dist}(z , \eta^{T_{z,n}^*} \cup \overline\eta^{\overline T_{z,n}^*})$. By the Koebe distortion theorem we then have $|\Phi_{z,n}'(w) \asymp |\Phi_{z,n}'(z)|$, which by assertion~\ref{Phi' asymp} of Lemma~\ref{E_z basics} is bounded between constants times $e^{-\overline\beta_n (q+1 ) + 2\overline u_n}$ and $e^{-\overline\beta_n (q+1 ) - 2\overline u_n}$. Therefore,
\begin{equation} \label{Psi_eta at w}
e^{-\overline\beta_n q + 2\overline u_n} \preceq |\Psi_\eta'(w)| \preceq e^{-\overline\beta_n q - 2\overline u_n} .
\end{equation}
Moreover, by assertion~\ref{eta dist} of Lemma~\ref{E_z basics} and assertion~\ref{beta u infty} of Lemma~\ref{beta u choice}, $\operatorname{dist}(z , \eta^{T_{z,n}^*} \cup \overline\eta^{\overline T_{z,n}^*})$ is bounded between constants times $e^{-\overline\beta_n (q+1 ) + \overline u_n}$ and $e^{-\overline\beta_n (q+1 ) - \overline u_n}$, so
\begin{equation} \label{eta dist at w}
e^{-\overline\beta_n - \overline u_n} \preceq \operatorname{dist}( w, \eta^{T_{z,n}^*} \cup \overline\eta^{\overline T_{z,n}^*})\preceq e^{-\overline\beta_n + \overline u_n} .
\end{equation}
Let $\widetilde w = \Psi_\eta(w)$. By~\eqref{dist and deriv},~\eqref{Psi_eta at w}, and~\eqref{eta dist at w} we have $e^{-\overline\beta_n (q+1 ) - 3\overline u_n} \preceq 1-|\widetilde w| \preceq e^{-\overline\beta_n (q+1 ) + 3\overline u_n}$.
By~\eqref{hm w} and conformal invariance of harmonic measure,
\begin{equation} \label{length asymp}
\operatorname{dist}(\widetilde w , I_{z,n}) \asymp \operatorname{length}(I_{z,n}) \asymp 1-|\widetilde w| .
\end{equation}
This proves assertion~\ref{I length}.
To get assertion~\ref{I nest}, we observe that the harmonic measure from $w^0$, as defined above, of each of $(\eta_{z,n+1}^0)^{\sigma_{z,n+1}}$ and $(\overline\eta_{z,n+1}^0)^{\overline\sigma_{z,n+1}}$ is $\asymp 1$. It therefore follows from conformal invariance of harmonic measure the distance from the endpoints of $I_{z,n}$ to the endpoints of $I_{z,n+1}$ is proportional to $1-|\widetilde w|$. We conclude by means of~\eqref{length asymp}.
To complete the proof of assertion~\ref{I deriv}, suppose given $x\in I_{z,n}$. By~\eqref{length asymp} the angle between the tangent line to $\partial\mathbf D$ at $x$ and the segment $[x,\widetilde w]$ is $\asymp 1$. Hence we can find $\delta_n \asymp 1-|\widetilde w|$ and $r \in (0,1)$ with $\log r \asymp 1$ such that $\widetilde w \in B_{r \delta_n}((1-\delta_n) x)$. By the Koebe distortion theorem we have $|(\Psi_\eta^{-1})'((1-\delta_n)x)| \asymp |(\Psi_\eta^{-1})'(\widetilde w)|$. By combining this with \ref{eta dist at w} we conclude that assertion~\ref{I deriv} holds.
\end{proof}
\begin{figure}\label{I stuff fig}
\begin{center}
\includegraphics{figures/drawing-I-stuff.pdf}
\end{center}
\caption{An illustration of the proof of Lemma~\ref{I stuff}. We use harmonic measure from $w$ and its images under various maps to deduce the desired conditions on the orange arc $I_{z,n}$ from the definitions of the events of Section~\ref{perfect setup sec}. Also shown (in red) is the arc $I_{z,n+1}$ appearing in assertion~\ref{I nest} and its images under the various maps. }
\end{figure}
\subsection{Lower bound for the Hausdorff dimension of the subset of the curve}
In this subsection we will prove a lower bound on the Hausdorff dimension of the sets $ \Theta^{s }(D_\eta)$. Assume we are in the setting of Section~\ref{lower setup sec}.
We define the set $\mathcal P$ of \emph{perfect points} as follows. Let $\lambda$ be the constant from Proposition~\ref{2pt estimate}. For $n\in\mathbf N$, let $n'$ be the greatest integer such that $\overline\beta_n -\lambda n \geq \overline\beta_{n'+1} +\lambda(n'+2)$. Let
\begin{equation}\label{ep_n def}
\epsilon_n := e^{-\overline\beta_{n'+1} -\lambda(n'+2) } .
\end{equation}
Note that Lemma~\ref{beta u choice} implies $e^{-\overline\beta_n } = \epsilon_n^{1+o_n(1)}$. Our reason for choosing this value of $\epsilon_n$ is that the pockets $\widehat D_{z,n}$ and $\widehat D_{w,n}$ are disjoint on $E_n(z) \cap E_n(w)$ provided $|z-w| \geq \epsilon_n$ (see Lemma~\ref{E_z basics}).
Choose a collection $\mathcal C_n$ of $ \asymp \epsilon_n^{-2} $ points in $B_d(0)$, no two of which lie within distance $ \epsilon_n $ of each other. For $z\in \mathcal C_n$ let $B^n(z)$ be the ball of radius $\epsilon_n$ centered at $z$. Let $\mathcal C_n' = \mathcal C_n\cap \mathcal P_n$ be the set of $z\in\mathcal C_n$ for which $E_n(z)$ occurs. Let
\begin{equation} \label{mathcal P def}
\mathcal P := \bigcap_{n\geq 1} \overline{\bigcup_{k\geq n} \bigcup_{z\in \mathcal C_k'} B^k(z)} .
\end{equation}
\begin{lem} \label{perfect pt contained}
Define $\mathcal P$ as in~\eqref{mathcal P def}. We have $\mathcal P \subset \Theta^{s }(D_\eta)$ for $s = q/(q+1)$. In fact, if $w\in\mathcal P$, then for $\epsilon > 0$ we have
\begin{equation} \label{deriv at ep}
|(\Psi_\eta^{-1})'( (1-\epsilon ) \Psi_\eta(w))| \asymp \epsilon^{-s+o_\epsilon(1)} ,
\end{equation}
with the implicit constants and the $o_n(1)$ deterministic and uniform for $w\in B_d(0)$.
\end{lem}
\begin{proof}
Fix $w\in \mathcal P$. Since $\eta$ is closed, it is clear that $w \in \eta$. It remains to prove~\eqref{deriv at ep}. By definition of $\mathcal P$, if we are given $n\in\mathbf N$, then we can find $k\geq n+1$ and $z\in \mathcal C_k'$ such that $|z-w| \leq e^{- 2 \overline\beta_{n+1} } $.
By Lemma~\ref{E_z basics} we have $w \in \widehat D_{z, n}$, so $\Psi_\eta(w) \in I_{z,n}$, as defined in Lemma~\ref{I stuff}. Let $\delta_n$ be as in that lemma with $x = \Psi_\eta(w)$.
By the Koebe distortion theorem, for $\epsilon \in [\delta_{n+1} ,\delta_n]$ we have
\begin{equation} \label{Psi' at ep}
\frac{1- (\delta_n - \delta_{n+1})/\delta_n }{(1+ (\delta_n - \delta_{n+1})/\delta_n )^3} \leq \frac{ |(\Psi_\eta^{-1})'((1-\epsilon) \Psi_\eta(w) )| }{ |(\Psi_\eta^{-1})'((1-\delta_n) \Psi_\eta(w) )| } \leq \frac{1+ (\delta_n - \delta_{n+1})/\delta_n }{(1- (\delta_n - \delta_{n+1})/\delta_n )^3} .
\end{equation}
By Lemma~\ref{I stuff} we have
\[
1- (\delta_n - \delta_{n+1})/\delta_n \asymp e^{- \beta_n + \overline\beta_n o_{n}(1) } ,
\]
which is proportional to $e^{\overline\beta_n o_{n}(1)} \asymp \epsilon^{o_\epsilon(1)}$ by Lemma~\ref{beta u choice}. We furthermore have $\delta_n = \epsilon^{1+o_\epsilon(1)}$. Hence~\eqref{Psi' at ep} and Lemma~\ref{I stuff} imply $|(\Psi_\eta^{-1})'( (1-\epsilon ) \Psi_\eta(w))| \asymp \epsilon^{-s+o_\epsilon(1)}$, as required.
\end{proof}
\begin{prop} \label{haus curve lower}
Let $s_- , s_+$ be as in Theorem~\ref{main thm}. For each $s \in (s_- , s_+)$, we a.s.\ have
\begin{equation*}
\dim_{\mathcal H} \Theta^s(D_\eta) \geq \xi(s) ,
\end{equation*}
where $\xi(s)$ is as in~\eqref{xi(s)}.
\end{prop}
\begin{proof}
For a Borel measure $\nu$ on a metric space $X$ and $\alpha > 0$, write
\begin{equation} \label{I_alpha def}
I_\alpha(\nu) = \int_X \int_X \frac{d\nu(z) \, d\nu(w)}{|z-w|^\alpha}
\end{equation}
for the $\alpha$-energy of $\nu$. By standard results for Hausdorff dimension (see \cite[Theorem~4.27]{peres-bm}) a metric space which admits a positive finite measure with finite $\alpha$-energy has Hausdorff dimension at least $\alpha$. In view of Lemma~\ref{perfect pt contained}, we are lead to construct such a measure $\nu$ on $\mathcal P $ for each $\alpha < \xi(s)$. We do this via the usual argument (see, e.g. \cite{miller-wu-dim,hmp-thick-pts, beffara-dim}).
Define the events $E_n(z)$ as in Section~\ref{perfect setup sec'} and the sets of points $\mathcal C_n$ and $\mathcal C_n'$ and the balls $B^n(z)$ as in the definition of $\mathcal P$ (right above~\eqref{mathcal P def}). Let $\epsilon_n$ be as in~\eqref{ep_n def}.
For each $n\in\mathbf N$, define a measure $\nu_n$ on $\mathbf D$ by
\[
d\nu_n(x) = \sum_{z\in \mathcal C_n } \frac{\mathbf 1_{E_n(z) }}{\mathbf P(E_n(z) )} \mathbf 1_{(x\in B^n(z))} \, dx .
\]
Then $\mathbf E(\nu_n(\mathbf D)) \asymp 1$. Moreover,
\alb
\mathbf E(\nu_n(\mathbf D)^2) &\preceq \epsilon_n^4 \sum_{z\not= w\in \mathcal C_n } \frac{\mathbf P( E_n(z) \cap E_n(w) ) }{\mathbf P(E_n(z) ) \mathbf P( E_n(w) )} + \epsilon_n^4 \sum_{z\in\mathcal C_n } \frac{1}{\mathbf P(E_n(z) )} .
\ale
By Lemma~\ref{P(E_z)} and Proposition~\ref{2pt estimate}, this is bounded by an $n$-independent constant times
\alb
\epsilon_n^4 \sum_{z\not= w\in \mathcal C_n } |z-w|^{- \gamma^*(q) + o_{|z-w|}(1) + o_n(1) } + \epsilon_n^4 \sum_{z\in\mathcal C_n } \epsilon_n^{- \gamma^*(q) + o_n(1) } ,
\ale
with the $o_{|z-w|}(1)$ independent of $n$. For $s \in (s_- , s_+)$ we have $\gamma^*(q) = \gamma(s)/(1-s) < 2$. Therefore, for sufficiently large $n$, $\mathbf E(\nu_n(\mathbf D)^2)$ is bounded above by a finite, $n$-independent constant. By the Vitalli convergence theorem, we can a.s.\ find a subsequence of the measures $\nu_n$ which converges weakly to a measure $\nu$ whose total mass is bounded above by some deterministic constant, and whose expected mass is positive.
On the other hand, we have
\alb
\mathbf E(I_\alpha(\nu_n)) &= \sum_{z, w\in \mathcal C_n } \frac{\mathbf P(E_n(z) \cap E_n(w) ) }{\mathbf P(E_n(z) ) \mathbf P( E_n(w))} \iint_{B^n(z) \times B^n(w)} \frac{1}{|x-y|^\alpha} \, dx \,dy \\
&= \sum_{z\not= w\in \mathcal C_n } \frac{\mathbf P( E_n(z) \cap E_n(w) ) }{\mathbf P(E_n(z) E_n(z) ) \mathbf P(E_n(z) )} \iint_{B^n(z) \times B^n(w)} \frac{1}{|x-y|^\alpha} \, dx \,dy \\
&+ \sum_{z \in \mathcal C_n } \frac{1}{\mathbf P(E_n(z) ) } \iint_{B^n(z) \times B^n(z)} \frac{1}{|x-y|^\alpha} \, dx \,dy\\
&\preceq \sum_{z\not= w\in \mathcal C_n } \frac{\mathbf P(E_n(z) \cap E_n(w) ) }{\mathbf P(E_n(z) ) \mathbf P(E_n(z) )} \frac{\epsilon_n^4}{|z-w|^{ \alpha}} + \sum_{z \in \mathcal C_n } \frac{\epsilon_n^{4-\alpha} }{\mathbf P(E_n(z) ) } \\
&\preceq \sum_{z\not= w\in \mathcal C_n } |z-w|^{-\gamma^*(q) - \alpha + o_{|z-w|}(1)+ o_n(1)} \epsilon_n^4 + \epsilon_n^{2-\alpha - \gamma^*(q) + o_n(1)} .
\ale
We have $\gamma^*(q) + \alpha < 2$ for $s \in (s_- , s_+)$ and $\alpha < \xi(s)$, so the above expression is $\preceq 1$.
We conclude that any subsequential limit $\nu$ of the measures $(\nu_n)$ satisfies $I_\alpha(\nu) <\infty$. Our definition~\eqref{mathcal P def} of $\mathcal P$ is such that $\nu$ is necessarily supported on $\mathcal P$. Hence \cite[Theorem~4.27]{peres-bm} and Lemma~\ref{perfect pt contained} imply that with positive probability, we have $\dim_{\mathcal H} \Theta^s(D_\eta) \geq \xi(s) $. Proposition~\ref{theta zero one} implies that this in fact a.s.\ holds.
\end{proof}
\begin{comment}
Clear[qq]
gamma1[kappa_,
q_] := (8 kappa +
8 kappa q + (-4 + kappa)^2 q^2)/(8 (kappa + 2 kappa q))
alpha[kappa_, s_] := (4 + kappa)^2 s^2/(8 kappa (1 + s));
xi[kappa_,
s_] := (16 s^2 + kappa^2 s^2 +
8 kappa (-1 - s + s^2))/(8 kappa (-1 + s^2))
qq[s_] := s/(1 - s)
FullSimplify[gamma1[kappa, qq[s]] + xi[kappa, s] ]
\end{comment}
\subsection{Lower bound for the Hausdorff dimension of the subset of the circle}
In order to get a lower bound on the Hausdorff dimension of $\widetilde\Theta^s(D_\eta)$, we will need a different set of perfect points. Define $\epsilon_n$, the sets $\mathcal C_n$, $\mathcal C_n'$ as in the definition~\eqref{mathcal P def} of $\mathcal P$. For $z\in \mathcal C_n'$, let $I_{z,n-1}$ be as in the statement of Lemma~\ref{I stuff}. Let $x_{z,n}$ be the midpoint of $I_{z,n-1}$ and let $I_{z,n}'$ be the arc of length $\epsilon_n^{q+1}$ centered at $x_{z,n}$. By Lemma~\ref{I stuff} we have $\operatorname{length}( I_{z,n}') = \operatorname{length}(I_{z,n-1})^{1+o_n(1)}$. Let
\begin{equation} \label{wt mathcal P def}
\widetilde{\mathcal P} := \bigcap_{n\geq 1} \overline{\bigcup_{k\geq n} \bigcup_{z \in \mathcal C_k'} I_{z,k-1}' } .
\end{equation}
\begin{lem} \label{perfect pts contained tilde}
Define $\widetilde{\mathcal P}$ as in~\eqref{wt mathcal P def}. If $\widetilde\Delta$ and $\beta_1$ are chosen sufficiently large then $\widetilde{\mathcal P} \subset \widetilde\Theta^{s }(D_\eta)$ for $s = q/(q+1)$. In fact, if $x\in \widetilde{\mathcal P}$, then for $\epsilon > 0$ we have
\begin{equation*}
|(\Psi_\eta^{-1})'( (1-\epsilon ) x)| \asymp \epsilon^{-s+o_\epsilon(1)} ,
\end{equation*}
with the implicit constants and the $o_\epsilon(1)$ deterministic and uniform in $x$.
\end{lem}
\begin{proof}
If $x\in \widetilde{\mathcal P}$ then for any given $n \in \mathbf N$ we can find $k\geq n$ and $z\in\mathcal C_k'$ such that $x$ lies within distance $ \operatorname{length}(I_{z,n}')^2$ of $I_{z,k}'$. If $k$ is chosen sufficiently large, depending on $n$, then by assertions~\ref{I length} and~\ref{I nest} of Lemma~\ref{I stuff} we have $x\in I_{z,n}$. We then conclude as in the proof of Lemma~\ref{perfect pt contained}.
\end{proof}
\begin{lem} \label{x_z dist}
For each $n$ there is an integer $m_n \leq n$ such that the following is true. We have $ \overline\beta_{n } - \overline\beta_{m_n} = \overline\beta_{n } o_n(1)$ and if $z,w\in\mathcal C_n'$ with $|z-w| \geq e^{-\overline\beta_{m_n+1}}$ then we have $\operatorname{dist}(I_{z,n}' , I_{w,n}' ) \succeq |z-w|^{q+1+o_{|z-w|}(1) }$, with the $o_{|z-w|}(1)$ and implicit constants deterministic, independent of $n$, and uniform for $z,w\in \mathcal C_n'$.
\end{lem}
\begin{proof}
We argue as in the proof of Lemma~\ref{near independence}. Choose $k\in \mathbf N$ such that $ e^{-\beta_{k+1} - \lambda(k+1)} \leq |z-w| \leq e^{-\beta_k - \lambda k}$.
Let $k'$ be the least integer such that $\overline\beta_{k'} -\lambda k' \geq \overline \beta_{k +1} + \lambda (k+1) $. By our choice~\eqref{ep_n def} of $\epsilon_n$ we have $k' \leq n-1$. By Lemma~\ref{E_z basics}, $\widehat D_{z,k'} \cap \widehat D_{w,k'} = \emptyset$ and hence $I_{z,k'} \cap I_{w,k'} = \emptyset$. If $\operatorname{length}(I_{z,n}') \leq \operatorname{length}(I_{z,k'+1 })$ then by assertions~\ref{I length} and~\ref{I nest} of Lemma~\ref{I stuff} we have
\[
\operatorname{dist}(x_{z,n} , x_{w,n} ) \succeq e^{-\overline\beta_{k'+1 } (q+1) - 3 \overline u_{k'+1} } \succeq |z-w|^{q+1+o_{|z-w|}(1) } .
\]
On the other hand, by assertion~\ref{I length} of Lemma~\ref{I stuff} we have $\operatorname{length}(I_{z,n}') \leq \operatorname{length}(I_{z,k'+1 })$ provided $\overline\beta_{k' +1} (q+1) + 3 \overline u_{k'+1} \leq (\overline\beta_{n } - \lambda n + \overline\beta_{n } o_n(1))(q+1)$, or equivalently provided
\begin{equation*}
\overline\beta_{n } - \overline\beta_{k'+1} \geq \frac{3\overline u_{k'+1} + \lambda n + \overline\beta_{n } o_n(1) }{q+1} .
\end{equation*}
It follows from Lemma~\ref{beta u choice} that we can choose $m_n \leq n$ such that $ \overline\beta_{n } - \overline\beta_{m_n} = \overline\beta_{n } o_n(1) $ and $\operatorname{length}(I_{z,n}') \leq \operatorname{length}(I_{z,k' +1})$ whenever $k'\leq m_n$.
\end{proof}
\begin{prop} \label{haus circle lower}
Let $s_- , s_+$ be as in Theorem~\ref{main thm}. For each $s\in (s_- , s_+)$ we a.s.\ have
\begin{equation*}
\dim_{\mathcal H} \widetilde\Theta^s(D_\eta) \geq \widetilde\xi(s) ,
\end{equation*}
where $\widetilde\xi(s)$ is as in~\eqref{tilde xi(s)}.
\end{prop}
\begin{proof}
We argue as in the proof of Proposition~\ref{haus curve lower}. In particular, for any given $\alpha < \widetilde\xi(s)$, we will construct a positive finite measure $\widetilde \nu$ on $\widetilde{\mathcal P}$ (as defined in~\eqref{wt mathcal P def} with finite $\alpha$-energy, as defined in~\eqref{I_alpha def}).
Define $\epsilon_n$ as in~\eqref{ep_n def}. We require all implicit constants and $o_{|z-w|}(1)$ terms to be independent of $n$ and uniform for $z,w\in \mathcal C_n$. For $n\in\mathbf N$, define a measure $\widetilde \nu_n$ on $\partial \mathbf D$ by
\[
d\widetilde\nu_n(x) = \epsilon_n^{ 1-q} \sum_{z\in \mathcal C_n' } \frac{\mathbf 1_{E_n(z) }}{\mathbf P(E_n(z) )} \mathbf 1_{(x\in I_{z,k}')} \, dx .
\]
Then we have $\mathbf E(\widetilde \nu_n(\partial \mathbf D)) \asymp 1$.
As in the proof of Proposition~\ref{haus curve lower} we have
\alb
\mathbf E(\widetilde \nu_n(\partial\mathbf D)^2) &\preceq \epsilon_n^{4} \sum_{z\not=w\in\mathcal C_n } \frac{\mathbf P( E_n(z) \cap E_n(w) ) }{\mathbf P(E_n(z) ) \mathbf P( E_n(w) )} + \epsilon_n^4 \sum_{z\in\mathcal C_n} \epsilon_n^{-\gamma^*(q) + o_n(1) } \preceq 1 .
\ale
Let $m_n$ be as in Lemma~\ref{x_z dist} and let $\mathcal K_n$ be the set of pairs $(z,w)\in\mathcal C_n \times\mathcal C_n$ with $|z-w| \leq e^{-\overline\beta_{m_n }}$ and $z\not=w$. By Lemma~\ref{x_z dist} we have $\# \mathcal K_n \leq \epsilon_n^{2 - o_n(1)}$.
By Lemma~\ref{P(E_z)}, Proposition~\ref{2pt estimate}, and Lemma~\ref{x_z dist}, we have
\alb\label{circle energy}
\mathbf E(I_\alpha(\widetilde \nu_n ) )
&= \epsilon_n^{2- 2q} \sum_{(z,w)\in\mathcal C_n\times \mathcal C_n } \frac{\mathbf P(E_n(z) \cap E_n(w))}{\mathbf P(E_n(z) ) \mathbf P( E_n(w))} \iint_{I_{z,k} \times I_{w,k}} \frac{1}{|x-y|^\alpha} \, dx \,dy \nonumber\\
&\preceq \sum_{(z,w) \notin \mathcal K_n ,\: z\not=w} |z-w|^{-\gamma^*(q) + o_{|z-w|}(1) } |x_{z,n} - x_{w,n}|^{-\alpha} \epsilon_n^{2(1+q) + 2-2q } \nonumber \\
& + \sum_{(z,w)\in \mathcal K_n } |z-w|^{-\gamma^*(q) + o_{|z-w|}(1) } \epsilon_n^{(2-\alpha)(q+1) +2-2q + o_n(1) } \nonumber \\
& + \sum_{z\in \mathcal C_n } \epsilon_n^{(2-\alpha)(q+1) +2-2q - \gamma^*(q) + o_n(1) } \nonumber\\
&\preceq \epsilon_n^{4} \sum_{z \not=w \in \mathcal C_n' } |z-w|^{-\gamma^*(q) - \alpha (q+1) + o_{|z-w|}(1) } + \epsilon_n^{(2-\alpha)(q+1) - 2q- \gamma^*(q) + o_n(1)}+ \epsilon_n^{(2-\alpha)(q+1) - 2q- \gamma^*(q) + o_n(1)} .
\ale
Note that for the middle term we used $|z-w| \succeq \epsilon_n$ and $\# \mathcal K_n \leq \epsilon_n^{-2 - o_n(1)}$. If $s \in (s_- , s_+)$ and $q = s/(1-s)$ we have $\gamma^*(q) + \alpha(q+1) < 2$ and $(2-\alpha)(1+q) -2q - \gamma^*(q) >0$ for $\alpha < \widetilde\xi(s)$. It follows that we can a.s.\ find a subsequence of the measures $(\widetilde\nu_n)$ which converges weakly a.s.\ to a finite positive limiting measure supported on $\widetilde{\mathcal P}$ with finite $\alpha$-energy. We then conclude using~\cite[Theorem~4.27]{peres-bm}, Lemma~\ref{perfect pts contained tilde}, and Proposition~\ref{theta zero one}.
\end{proof}
\begin{comment}
Clear[qq]
gamma1[kappa_,
q_] := (8 kappa +
8 kappa q + (-4 + kappa)^2 q^2)/(8 (kappa + 2 kappa q))
alpha[kappa_, s_] := (4 + kappa)^2 s^2/(8 kappa (1 + s));
xi[kappa_, s_] := 1 - alpha[kappa, s]
qq[s_] := s/(1 - s)
r[kappa_, s_] := (2 - xi[kappa, s]) (1 + qq[s]) - 2 qq[s] -
gamma1[kappa, qq[s]]
FullSimplify[r[kappa, s]]
FullSimplify[gamma1[kappa, qq[s]] + xi[kappa, s] (1 + qq[s])]
\end{comment}
\subsection{Proof of Theorem~\ref{main thm}}
This follows by combining Propositions~\ref{bdy haus upper},~\ref{haus curve upper},~\ref{haus curve lower}, and~\ref{haus circle lower}.
\qed
\begin{remark} \label{s=1 kappa=4}
In the case $\kappa=4$, we have $s_+ = 1$, so the sets $\Theta^1(D_\eta)$ and $\widetilde\Theta^1(D_\eta)$ for $\kappa=4$ can be non-empty. We do not explicitly mention these sets in Theorem~\ref{main thm} because the statement of our results in this case is slightly different. We are now going to explain how to obtain analogs of the statements of Theorem~\ref{main thm} in the case $\kappa=4$, $s=1$. We prove in Proposition~\ref{bdy haus upper} that a.s.\ $\dim_{\mathcal H} \widetilde\Theta^1(D_\eta) = 0$. Since $\xi(1) = \dim_{\mathcal H}(\eta) = 3/2$ for $\kappa=4$, the upper bound for $\dim_{\mathcal H} \Theta^1(D_\eta)$ in the case $\kappa=4$ is trivial. We can obtain a lower bound for $\dim_{\mathcal H} \Theta^1(D_\eta)$ provided we replace the limit in~\eqref{tilde theta} with a $\limsup$, as follows. By Frostman's lemma~\cite[Theorem~4.30]{peres-bm} there exists a finite positive measure $\nu$ on $\eta$ and a constant $C>0$ such that $\nu(A) \leq C \operatorname{diam}(A)^{3/2}$ for each $A\subset \eta$. By the upper bound for $\dim_{\mathcal H} \Theta^{s;\leq}(D_\eta)$ in Theorem~\ref{main thm} for $s < 1$ and \cite[Theorem~4.19]{peres-bm}, we have $\nu\left(\Theta^{s;\leq}(D_\eta)\right) = 0$ for each $s<1$. Therefore $\nu$ assigns zero mass to the set of points $z \in \eta$ for which
\begin{equation*}
\limsup_{\epsilon \rightarrow 0} \frac{\log |(\Psi_\eta^{-1})'((1-\epsilon) \Psi_\eta( x) )|}{-\log \epsilon} < 1 .
\end{equation*}
By combining this with Lemma~\ref{s>1 empty} we obtain that $\nu$ a.s.\ assigns full mass to the set of $z\in\eta$ for which
\begin{equation*}
\limsup_{\epsilon \rightarrow 0} \frac{\log |(\Psi_\eta^{-1})'((1-\epsilon) \Psi_\eta( x) )|}{-\log \epsilon} = 1 .
\end{equation*}
Hence this set a.s.\ has Hausdorff dimension $\geq 3/2$.
\end{remark}
|
\section{Introduction}
Many physical phenomena can be described by mathematical models that involve partial
differential equations. Therefore, in recent years, researchers look for new numerical methods which are more cost effective and
simple in implementation to solve partial differential equations. Investigation of exact and approximate
solution helps us to understand meaning and relevance of these mathematical models. Several techniques
including scattering method \cite{vak}, sine-cosine method \cite{wazwaz}, homotopy analysis method \cite{liao}, \cite{jafari}, homotopy perturbation method [5-9], differential transform method \cite{solt}, variational iteration method \cite{he3}, \cite{biazar}, or decomposition methods [13-20] have been used for solving these problems, but mostly for two dimensional partial differential equations.
Inspired and motivated by ongoing research in this area, we apply new iterative scheme for solving heat- and
wave-type equations. Several examples are given to verify reliability and efficiency of proposed technique.
\section{Preliminaries}
Let $\Omega$ be a compact subset of $\RR^k$. Denote $J = [-\delta,\delta] \times \Omega$, where $\delta>0$ will be specified later, then $J$ is a compact subset of $\RR^{k+1}$. Let $u(t,x)=u(t,x_1,\dots,x_k)$
be a real function of $k+1$ variables defined on $J$. We introduce the following operators: $\nabla = (\frac{\partial}{\partial t},\frac{\partial}{\partial x_1}, \dots,
\frac{\partial}{\partial x_k})$ and $D = (\frac{\partial}{\partial x_1}, \dots, \frac{\partial}{\partial x_k})$. We deal with partial differential equations
\begin{equation}\label{eq1}
\frac{\partial^n}{\partial t^n} u(t,x) = F (t, x, u, \nabla u, \dots, \nabla^m u) \quad \text{for } m<n
\end{equation}
and
\begin{equation}\label{eq2}
\frac{\partial^n}{\partial t^n} u(t,x) = F (t, x, u, \nabla u, \dots, \nabla^{n-1} u, D \nabla^{n-1} u, D^2 \nabla^{n-1} u, \dots, D^{m-(n-1)} \nabla^{n-1} u)
\quad \text{for } m \geq n.
\end{equation}
In both cases, left-hand side of the equation contains only the highest derivative with respect to $t$. We do not consider equations where the order of partial derivatives
with respect to $t$ is $n$ or higher on the righthand side, including mixed derivatives.
When convenient, we will use multiindex notation as well:
$$
\frac{\partial^{\vert \alpha \vert}}{\partial x^{\alpha}} = \frac{\partial^{\vert \alpha \vert}}{\partial x_1^{\alpha_1} \partial x_2^{\alpha_2} \dots \partial x_k^{\alpha_k}},
$$
where $\vert \alpha \vert = \alpha_1 + \alpha_2 + \dots + \alpha_k$.
Denote $N=\max \{ m,n \}$.
We consider equation \eqref{eq1} or \eqref{eq2} with the set of initial conditions
\begin{align}\label{eq3}
u(0,x) &= c_1 (x),\notag \\
\frac{\partial}{\partial t} u(0,x) &= c_2 (x), \notag \\
&\vdots \notag \\
\frac{\partial^{n-1}}{\partial t^{n-1}} u(0,x) &= c_n (x),
\end{align}
where initial functions $c_i (x)$, $i=1,\dots,n$ are taken from space $C^N (\Omega, \RR)$. It means that we are looking for classical solutions.
For the purpose of clarity, we emphasize that our formulation covers for instance heat, wave, Burger, Boussinesq or Korteweg-de Vries (KdV) equations.
Obviously, $F \colon J \times \RR^k \times \RR \times \RR^{k+1} \times \RR^{(k+1)^2} \times \dots \times \RR^{(k+1)^m} \to \RR$ if $m<n$,
and $F \colon J \times \RR^k \times \RR \times \RR^{k+1}
\times \RR^{(k+1)^2} \times \dots \times \RR^{(k+1)^{n-1}} \times \RR^{k(k+1)^{n-1}} \times \RR^{k^2(k+1)^{n-1}} \times \dots \times \RR^{k^{m-n+1}(k+1)^{n-1}} \to \RR$ if $m \geq n$.
Denote
\begin{equation}
K_1 = \frac{(k+1)^{m+1} -1}{k} \quad \text{for } m<n
\end{equation}
and
\begin{equation}
K_2 = \frac{(k+1)^{n-1} -1}{k} + (k+1)^{n-1} \frac{k^{m-n+2} -1}{k-1} \quad \text{for } m \geq n.
\end{equation}
Then, if we consider $u$ as dependent variable, we see that $F$ is a function of $k+1+K_1$ variables in case $m<n$ or $k+1+K_2$ variables in case $m \geq n$.
Denote
\begin{equation}\label{eq4}
u_0 (t,x) = \sum_{i=1}^n c_i (x) \frac{t^{i-1}}{(i-1)!} = \sum_{i=1}^n \left( \frac{\partial^{i-1}}{\partial t^{i-1}} u(0,x) \right) \frac{t^{i-1}}{(i-1)!}.
\end{equation}
Then $u_0 \in C^N (J, \RR)$.
We suppose that $F$ is Lipschitz continuous in last $K_1$ ($m<n$) or $K_2$ ($m \geq n$) variables, i.e. $F$ satisfies condition
\begin{equation}\label{eq5}
\vert F (t,x,y_1, \dots, y_{K_l}) - F (t,x,z_1, \dots, z_{K_l}) \vert \leq L \left( \sum_{i=1}^{K_l} \vert y_i - z_i \vert \right), \quad l = 1 \text{ or } 2,
\end{equation}
on a compact set which is defined as follows: There is $R \in \RR, R>0$ such that \eqref{eq5} holds on
\begin{equation}\label{eq6}
J \times \prod_{\alpha_0 + \vert \alpha \vert \leq m}
[c_{\alpha_0, \alpha}, d_{\alpha_0, \alpha} ],
\end{equation}
where
\begin{equation}\label{eq7}
c_{\alpha_0, \alpha} = \min_{(t,x) \in J} \left[ \frac{\partial^{\alpha_0 + \vert \alpha \vert}}{\partial t^{\alpha_0} \partial x^\alpha} u_0 (t,x) \right] -R,
\quad
d_{\alpha_0, \alpha} = \max_{(t,x) \in J} \left[ \frac{\partial^{\alpha_0 + \vert \alpha \vert}}{\partial t^{\alpha_0} \partial x^\alpha} u_0 (t,x) \right] +R,
\end{equation}
and $\alpha_0 <n$ in all cases.
Since $F$ is continuous on compact set, $\vert F \vert$ attains its maximal value on this set, denote it $M$. Then we put
\begin{equation}\label{eq10}
\delta = \left( \frac{R \cdot (n-1)!}{M} \right)^{1/n}.
\end{equation}
\section{Main results}
\begin{theorem}\label{th1}
Let the condition \eqref{eq5} hold. Then problem consisting of equation \eqref{eq1} or \eqref{eq2} and initial conditions \eqref{eq3} has a unique local solution on $(-\delta, \delta) \times \Omega$, where $\delta$ is defined by \eqref{eq10}.
\end{theorem}
{\bf Proof.} First define the following operator
\begin{equation}\label{eq8}
T u(t,x) = u_0 (t,x) + \int_0^t \frac{(t-\xi)^{n-1}}{(n-1)!} F \bigl(\xi, x, u (\xi,x), \nabla u(\xi,x), \dots \bigr) d \xi,
\end{equation}
where function $F$ has either $k+1+K_1$ or $k+1+K_2$ arguments and the last $K_1$ or $K_2$ arguments involve dependent variable $u$.
Starting with equation \eqref{eq1}, respective \eqref{eq2}, and using repeated integration by parts, it can be easily proved that if $u$ is a solution of equation $u=Tu$, i.e., if $u$ is a fixed point of operator $T$, then it is a solution of problem \eqref{eq1}, \eqref{eq3}, respective \eqref{eq2}, \eqref{eq3}.
Denote $J_1 = [-\delta_1, \delta_1] \times \Omega$, where $0<\delta_1<\delta$. Then $J_1$ is compact. Let $C^N (J_1, \RR)$ be the space of functions from $J_1$ to $\RR$ with continuous partial derivatives up to order $N$. This space is a Banach space with respect to the norm
\begin{equation}\label{eq9}
\Vert u \Vert_{C^N} = \sum_{\alpha_0 + \vert \alpha \vert \leq N} \max_{(t,x) \in J_1} \left| \frac{\partial^{\alpha_0 + \vert \alpha \vert}}{\partial t^{\alpha_0} \partial x^\alpha} u(t,x) \right|.
\end{equation}
It is obvious that, considering space $C^N$, the order of partial derivatives with respect to $t$ is allowed to be greater than or equal to $n$ in case $m \geq n$ when calculating the norm. Further, we define closed ball $B_R (u_0) \subseteq C^N (J_1,\RR)$ as follows:
$$
B_R (u_0) = \left\{ y \in C^N (J_1,\RR) \colon \Vert y-u_0 \Vert_{C^N} \leq R \right\}.
$$
It is not difficult to verify that $F$ composed with any $y \in B_R (u_0)$ and its appropriate derivatives satisfies Lipschitz condition \eqref{eq5} and that upper bound $\vert F \vert \leq M$ remains valid as well. Indeed, $J_1 \subseteq J$ and if $\Vert y-u_0 \Vert_{C^N} \leq R$, then
\begin{equation*}
\left\vert \frac{\partial^{\alpha_0 + \vert \alpha \vert}}{\partial t^{\alpha_0} \partial x^\alpha} \Bigl( y(t,x)-u_0 (t,x) \Bigr) \right\vert \leq R
\end{equation*}
for all $(t,x) \in J_1$ and for all $\alpha_0 + \vert \alpha \vert \leq N$. Hence point $\bigl(t, x, y(t,x), \nabla y(t,x), \dots, \nabla^m y(t,x) \bigr)$ or\\
$\bigl(t, x, y(t,x), \nabla y(t,x), \dots, D^{m-(n-1)} \nabla^{n-1} y(t,x) \bigr)$ respectively lies in the set $J \times \prod\limits_{\alpha_0 + \vert \alpha \vert \leq m}
[c_{\alpha_0, \alpha}, d_{\alpha_0, \alpha} ]$ defined by \eqref{eq6} and \eqref{eq7}.
We need to show that $T$ is a contraction on $B_R (u_0)$ for sufficiently small $\delta_1$.
In the first step, we prove that $T$ maps $B_R (u_0)$ into itself. Take any $y \in B_R (u_0)$. Then
\begin{align*}
\Vert Ty - u_0 \Vert_{C^N} &= \Bigl\Vert u_0 + \int_0^t \frac{(t-\xi)^{n-1}}{(n-1)!} F \bigl(\xi, x, y (\xi,x), \nabla y(\xi,x), \dots \bigr) d \xi - u_0 \Bigr\Vert_{C^N} \\
&\leq \Bigl\Vert \int_0^t \frac{t^{n-1}}{(n-1)!} F \bigl(\xi, x, y (\xi,x), \nabla y(\xi,x), \dots \bigr) d \xi \Bigr\Vert_{C^N} \leq \frac{\delta_1^{n-1}}{(n-1)!} M \Bigl\Vert \int_0^t d \xi \Bigr\Vert_{C^N} \\
&\leq \frac{\delta_1^n}{(n-1)!} M < R \ \ \text{for } \ 0<\delta_1 < \left( \frac{R \cdot (n-1)!}{M} \right)^{1/n}.
\end{align*}
Thus $Ty \in B_R (u_0)$ for $0<\delta_1 < \left( \frac{R \cdot (n-1)!}{M} \right)^{1/n}$.
The second step is to show that $T$ is a contraction. Choose arbitrary $y, z \in B_R (u_0)$. Then we have
\begin{align*}
\Vert Ty - Tz \Vert_{C^N} &= \Bigl\Vert \int_0^t \frac{(t-\xi)^{n-1}}{(n-1)!} \Bigl[ F \bigl(\xi, x, y (\xi,x), \nabla y(\xi,x), \dots \bigr) - F \bigl(\xi, x, z (\xi,x), \nabla z(\xi,x), \dots \bigr) \Bigr] d \xi \Bigr\Vert_{C^N} \\
&\leq \Bigl\Vert \frac{t^{n-1}}{(n-1)!} \int_0^t \bigl\vert F \bigl(\xi, x, y (\xi,x), \nabla y(\xi,x), \dots \bigr) - F \bigl(\xi, x, z (\xi,x), \nabla z(\xi,x), \dots \bigr) \bigr\vert d \xi \Bigr\Vert_{C^N}\\
&\leq \frac{\delta_1^{n-1}}{(n-1)!} \Bigl\Vert \int_0^t L \cdot \Biggl( \sum_{\alpha_0 + \vert \alpha \vert \leq m} \Biggl\vert \frac{\partial^{\alpha_0 + \vert \alpha \vert}}{\partial t^{\alpha_0} \partial x^\alpha} \Bigl( y(t,x)-z(t,x) \Bigr) \Biggr\vert \Biggr) d \xi \Bigr\Vert_{C^N}\\
&\leq \frac{\delta_1^{n-1}}{(n-1)!} L \ \Biggl\Vert \Biggl( \sum_{\alpha_0 + \vert \alpha \vert \leq m} \max_{(t,x) \in J_1} \Biggl\vert \frac{\partial^{\alpha_0 + \vert \alpha \vert}}{\partial t^{\alpha_0} \partial x^\alpha} \Bigl( y(t,x)-z(t,x) \Bigr) \Biggr\vert \Biggr) \int_0^t d \xi \Biggr\Vert_{C^N} \leq \frac{\delta_1^n}{(n-1)!} L \ \Vert y-z \Vert_{C^N},
\end{align*}
where $\alpha_0 <n$ for equation \eqref{eq2} and $L$ is a Lipschitz constant for $F$ introduced in \eqref{eq5}. It follows that $T$ is a contraction for $0<\delta_1 < \left( \frac{(n-1)!}{2L} \right)^{1/n}$.
Combining all results, we obtain that if $0<\delta_1 \leq \min \left\{ \frac{\delta}{2}, \left( \frac{R \cdot (n-1)!}{2M} \right)^{1/n}, \left( \frac{(n-1)!}{2L} \right)^{1/n} \right\}$, then operator $T$ is a contraction on $B_R (u_0)$. Applying Banach contraction principle, we can conclude that $T$ has a unique fixed point in $B_R (u_0)$ which is a unique solution of problem \eqref{eq1}, \eqref{eq3}, respective \eqref{eq2}, \eqref{eq3}.
Since $\delta_1$ depends only on the Lipschitz constant $L$ and on the distance $R$ from initial data to the boundaries of the intervals $[c_{\alpha_0, \alpha}, d_{\alpha_0, \alpha} ]$ wherein the estimate $M$ holds, we can
apply our result repeatedly to get a unique local solution defined for $(t,x) \in (-\delta, \delta) \times \Omega$.
\qed
\begin{theorem}\label{th2}
Assume that condition \eqref{eq5} holds. Then iterative scheme $u_p = T u_{p-1}$, $p \geq 1$ with initial approximation $u_0$ defined by \eqref{eq4}, where $T$ is defined by \eqref{eq8}, converges to unique local solution $u(t,x)$ of problem \eqref{eq1}, \eqref{eq3}, respective \eqref{eq2}, \eqref{eq3}. Moreover, we have the following error estimate for this scheme:
\begin{equation}\label{eq13}
\Vert u(t,x) - u_p (t,x)\Vert_{C^N} \leq \frac{R \cdot \gamma^p}{1-\gamma}
\end{equation}
on $(-\delta_1, \delta_1) \times \Omega$, where $\delta_1$ is chosen such that operator $T$ is a contraction,
\begin{equation}\label{eq14}
\gamma = \frac{L \cdot \delta_1^n}{(n-1)!},
\end{equation}
and constants $L$ and $R$ are defined by \eqref{eq5} and \eqref{eq6}.
\end{theorem}
{\bf Proof.} First we need to show that sequence $\bigl( u_p \bigr)_{p=0}^\infty$ is convergent. We prove it by showing that it is a Cauchy sequence.
Take any $p,q \in \mathbb N$, $q \geq p$. Then
\begin{align*}
\Vert u_q - u_p \Vert_{C^N} &= \Bigl\Vert \int_0^t \frac{(t-\xi)^{n-1}}{(n-1)!} \Bigl[ F \bigl(\xi, x, u_{q-1} (\xi,x), \nabla u_{q-1}(\xi,x), \dots \bigr) - F \bigl(\xi, x, u_{p-1} (\xi,x), \nabla u_{p-1}(\xi,x), \dots \bigr) \Bigr] d \xi \Bigr\Vert_{C^N} \\
&\leq \frac{\delta_1^n}{(n-1)!} L \ \Vert u_{q-1}-u_{p-1} \Vert_{C^N} \leq \gamma \ \Vert u_{q-1}-u_{p-1} \Vert_{C^N},
\end{align*}
where $ 0 \leq \gamma \leq \frac{1}{2} <1$. Put $q=p+1$. We obtain
\begin{equation*}
\Vert u_{p+1} - u_p \Vert_{C^N} \leq \gamma \Vert u_{p}-u_{p-1} \Vert_{C^N} \leq \gamma^2 \Vert u_{p-1}-u_{p-2} \Vert_{C^N} \leq \ldots \leq \gamma^p \Vert u_1-u_0 \Vert_{C^N}.
\end{equation*}
Now, using the triangle inequality, we get
\begin{align*}
\Vert u_q - u_p \Vert_{C^N} &= \Vert u_q - u_{q-1} \Vert_{C^N} + \ldots + \Vert u_{p+2} - u_{p+1} \Vert_{C^N} + \Vert u_{p+1} - u_p \Vert_{C^N} \leq \bigl( \gamma^{q-1} + \ldots + \gamma^{p+1} + \gamma^{p} \bigr) \Vert u_1-u_0 \Vert_{C^N} \\
&\leq \gamma^p \bigl( 1 + \gamma + \gamma^2 + \ldots + \gamma^{q-p-1} \bigr) \Vert u_1-u_0 \Vert_{C^N} \leq \gamma^p \frac{1-\gamma^{q-p}}{1-\gamma} \Vert u_1-u_0 \Vert_{C^N}.
\end{align*}
Since $\gamma <1$, then $1-\gamma^{q-p} <1$ as well, and we estimate
\begin{align}
\Vert u_q - u_p \Vert_{C^N} &\leq \frac{\gamma^p}{1-\gamma} \Vert u_1-u_0 \Vert_{C^N} \leq \frac{\gamma^p}{1-\gamma} \Vert Tu_0-u_0 \Vert_{C^N} = \frac{\gamma^p}{1-\gamma} \Bigl\Vert \int_0^t \frac{(t-\xi)^{n-1}}{(n-1)!} F \bigl(\xi, x, u_0 (\xi,x), \nabla u_0(\xi,x), \dots \bigr) d \xi \Bigr\Vert_{C^N} \notag \\
&\leq \frac{\gamma^p}{1-\gamma} \cdot \frac{\delta_1^{n}}{(n-1)!} \cdot M < \frac{\gamma^p}{1-\gamma} \cdot R \leq \frac{\left( \frac{1}{2} \right)^p}{\frac{1}{2}} \cdot R. \label{eq15}
\end{align}
It follows that for arbitrary $\epsilon>0$ there is a $P \in \mathbb N$, $P > 1- \log_2 \epsilon + \log_2 R$ such that if $p,q \geq P$, then $\Vert u_q - u_p \Vert_{C^N} < \epsilon$. Thus sequence $\bigl( u_p \bigr)_{p=0}^\infty$ is a Cauchy sequence and consequently a convergent sequence.
Then there is a limit $u=\lim\limits_{p \to \infty} u_p$ such that $u=Tu$. Hence $u$ is a fixed point of operator $T$. Applying Theorem \ref{th1} we conclude that this fixed point is unique and it is a unique local solution of problem \eqref{eq1}, \eqref{eq3}, respective \eqref{eq2}, \eqref{eq3}.
Error estimate \eqref{eq13} follows immediately from \eqref{eq15} by taking a limit for $q \to \infty$.
\qed
\begin{corol}\label{c1}
Let condition \eqref{eq5} be valid and suppose that $F$ can be written as $G+g$:
\begin{equation}\label{eq12}
F \bigl(t, x, z (t,x), \nabla z(t,x), \dots \bigr) = G \bigl(t, x, z (t,x), \nabla z(t,x), \dots \bigr) + g(t,x).
\end{equation}
Then we may choose initial approximation
\begin{equation}\label{eq11}
\bar{u}_0 = u_0 + \int_0^t \frac{(t-\xi)^{n-1}}{(n-1)!} g(\xi, x) d \xi = \sum_{i=1}^n \left( \frac{\partial^{i-1}}{\partial t^{i-1}} u(0,x) \right) \frac{t^{i-1}}{(i-1)!} + \int_0^t \frac{(t-\xi)^{n-1}}{(n-1)!} g(\xi, x) d \xi.
\end{equation}
\end{corol}
{\bf Proof.} Denote $M_2 = \max \{ \vert G \vert + \vert g \vert \}$. According to the proof of Theorem \ref{th1}, we only need to show that $\bar{u}_0 \in B_R (u_0)$, i.e., $\Vert \bar{u}_0 - u_0 \Vert_{C^N} \leq R$. Indeed, we have
\begin{align*}
\Vert \bar{u}_0 - u_0 \Vert_{C^N} &= \Bigl\Vert u_0 + \int_0^t \frac{(t-\xi)^{n-1}}{(n-1)!} g(\xi, x) d \xi - u_0 \Bigr\Vert_{C^N} = \Bigl\Vert \int_0^t \frac{(t-\xi)^{n-1}}{(n-1)!} g(\xi, x) d \xi \Bigr\Vert_{C^N}\\
&\leq \Bigl\Vert \int_0^t \frac{(t-\xi)^{n-1}}{(n-1)!} \bigl\vert g(\xi, x) \bigr\vert d \xi \Bigr\Vert_{C^N} \leq \frac{t^{n-1}}{(n-1)!} \Bigl\Vert \int_0^t \bigl\vert g(\xi, x) \bigr\vert + \bigl\vert G \bigl(\xi, x, z (\xi,x), \nabla z(\xi,x), \dots \bigr) \bigr\vert d \xi \Bigr\Vert_{C^N}\\
&\leq \frac{\delta_2^{n-1}}{(n-1)!} M_2 \Bigl\Vert \int_0^t d \xi \Bigr\Vert_{C^N} \leq \frac{\delta_2^n}{(n-1)!} M_2 < R \ \ \text{for } \ 0<\delta_2 < \left( \frac{R \cdot (n-1)!}{M_2} \right)^{1/n}.
\end{align*}
Consequently, integral
\begin{equation*}
\int_0^t \frac{(t-\xi)^{n-1}}{(n-1)!} g(\xi, x) d \xi
\end{equation*}
is small enough for sufficiently small $\delta_2$ and thus $\bar{u}_0 \in B_R (u_0)$ for this $\delta_2$.
\qed
\section{Applications}
We demonstrate potentiality of our approach on several initial value problems (IVP's).
{\bf Example 1.} Consider the following two-dimensional heat-type equation
\begin{equation}\label{p}
\frac{\partial u(x,y,t)}{\partial t} - \frac{\partial^2 u(x,y,t)}{\partial x^2}+ \frac{\partial^2 u(x,y,t)}{\partial y^2} + u(x,y,t) = (1+t) \sinh{(x+y)}
\end{equation}
with initial condition
\begin{equation}\label{p1}
u(x,y,0)= \sinh{(x+y)}.
\end{equation}
Then
$$
u_0(x,y,t) =\sum\limits_{k=0}^{n-1} {\frac{\partial ^ku\left({x,y,0} \right)}{\partial t^k}} \frac{t^k}{k!} + \int_0^t g(x,y,\xi) d\xi
= \sinh{(x+y)} \left(1 + t + \frac{t^2}{2}\right)
$$
and
$$
u_p(x,y,t) = u_0(x,y,t)+\int_0^t \left(\frac{\partial^2u_{p-1}(x,y,\xi)}{\partial x^2}- \frac{\partial^2 u(x,y,\xi)}{\partial y^2}- u(x,y,\xi)\right) d\xi, \ \ p \geq 1.
$$
Hence
\begin{eqnarray}
u_1(x,y,t) &=& \sinh{(x+y)} \left(1 + t + \frac{t^2}{2}\right)-\int_0^t \sinh{(x+y)} \left(1 + \xi + \frac{\xi^2}{2}\right) d\xi \nonumber \\
&=& \sinh (x+y)\left( 1- \frac{t^3}{3!}\right), \nonumber \\
u_2(x,y,t) &=& \sinh{(x+y)} \left(1 + t + \frac{t^2}{2}\right) - \int_0^t \sinh{(x+y)}\left( 1- \frac{\xi^3}{3!}\right)d\xi\nonumber \\
&=& \sinh{(x+y)}\left(1+ \frac{t^2}{2}+ \frac{t^4}{4!}\right),\nonumber \\
u_3(x,y,t) &=&\sinh{(x+y)} \left(1 + t + \frac{t^2}{2}\right)- \int_0^t \sinh{(x+y)}\left(1+ \frac{\xi^2}{2}+ \frac{\xi^4}{4!}\right)d\xi \nonumber \\
&=& \sinh{(x+y)}\left( 1 + \frac{t^2}{2}-\frac{t^3}{3!}- \frac{t^5}{5!}\right),\nonumber \\
u_4(x,y,t) &=& \sinh{(x+y)} \left(1 + t + \frac{t^2}{2}\right)- \int_0^t \sinh{(x+y)}\left( 1+ \frac{\xi^2}{2}-\frac{\xi^3}{3!}- \frac{\xi^5}{5!}\right)d\xi \nonumber \\
&=& \sinh{(x+y)}\left(1+ \frac{t^2}{2}-\frac{t^3}{3!} +\frac{t^4}{4!}+ \frac{t^6}{6!}\right),\nonumber \\
u_5(x,y,t) &=&\sinh{(x+y)} \left(1 + t + \frac{t^2}{2}\right)- \int_0^t \sinh{(x+y)}\left( 1+ \frac{\xi^2}{2}-\frac{\xi^3}{3!}+\frac{\xi^4}{4}+ \frac{\xi^6}{6!} \right) d\xi \nonumber \\
&=& \sinh{(x+y)}\left( 1+\frac{t^2}{2}-\frac{t^3}{3!} +\frac{t^4}{4!}-\frac{t^5}{5!}- \frac{t^7}{7!}\right),\nonumber \\
\vdots \nonumber
\end{eqnarray}
We can see that so-called self-canceling terms appear between various components (see, for example, $u_1,u_2,u_3$). Keeping the remaining non-canceled terms, we have
$$
u(x,y,t) = \sinh (x+y)\left( 1 + \frac{t^2}{2!} - \frac{t^3}{3!} + \frac{t^4}{4!} - \frac{t^5}{5!} + \dots \right) = ( t+ e^{-t} )\sinh (x+y),
$$
which is unique exact solution of IVP \eqref{p}, \eqref{p1}.\\[3mm]
{\bf Example 2.} Consider the following initial value problem for two-dimensional heat-type equation with variable coefficients
\begin{equation}\label{3}
\frac{\partial u(x,y,t)}{\partial t} = \frac{y^2}{2} \frac{\partial^2 u(x,y,t)}{\partial x^2}+ \frac{x^2}{2} \frac{\partial^2 u(x,y,t)}{\partial y^2}
\end{equation}
with initial condition
\begin{equation}\label{4}
u(x,y,0)= y^2.
\end{equation}
Then
$$
u_0(x,y,t) = y^2
$$
and
$$
u_p(x,y,t) = y^2+\int_0^t \left(\frac{y^2}{2} \frac{\partial^2 u_{p-1}(x,y,\xi)}{\partial x^2}+ \frac{x^2}{2} \frac{\partial^2 u_{p-1}(x,y,\xi)}{\partial y^2}\right)d\xi,\ \ p \geq 1.
$$
Thus
\begin{eqnarray}
u_1(x,y,t) &=& y^2+\int_0^t x^2 d\xi = y^2+x^2t, \nonumber \\
u_2(x,y,t) &=& y^2+\int_0^t (y^2\xi + x^2 ) d\xi = y^2\left( 1+\frac{t^2}{2!}\right) +x^2t, \nonumber \\
u_3(x,y,t) &=& y^2+ \int_0^t \left( y^2\xi+ x^2 +x^2\frac{\xi^2}{2}\right) d\xi = y^2\left( 1+\frac{t^2}{2!}\right)+x^2 \left( t+ \frac{t^3}{3!} \right), \nonumber \\
u_4(x,y,t) &=& y^2\left(1+\frac{t^2}{2!} +\frac{t^4}{4!}\right ) + x^2 \left( t+ \frac{t^3}{3!} \right), \nonumber \\
u_5(x,y,t) &=& y^2\left(1+\frac{t^2}{2!} +\frac{t^4}{4!}\right ) + x^2 \left( t+ \frac{t^3}{3!}+ \frac{t^5}{5!} \right), \nonumber \\
\vdots \nonumber \\
u_{2l-1}(x,y,t) &=& y^2\left(1+\frac{t^2}{2!} +\frac{t^4}{4!} + \dots + \frac{t^{2l-2}}{(2l-2)!}\right )+ x^2 \left( t+ \frac{t^3}{3!}+ \frac{t^5}{5!}+\dots +\frac{t^{2l-1}}{(2l-1)!}\right), \nonumber \\
u_{2l}(x,y,t) &=& y^2\left(1+\frac{t^2}{2!} +\frac{t^4}{4!} + \dots + \frac{t^{2l}}{(2l)!}\right )+ x^2 \left( t+ \frac{t^3}{3!}+ \frac{t^5}{5!}+\dots +\frac{t^{2l-1}}{(2l-1)!}\right), \nonumber \\
\vdots \nonumber
\end{eqnarray}
Hence unique solution of IVP \eqref{3}, \eqref{4} has the form
\begin{eqnarray}
u(x,y,t) &=& x^2\left( t+ \frac{t^3}{3!} + \frac{t^5}{5!}+ \dots + \frac{t^{2l-1}}{(2l-1)!}+ \dots \right) + y^2\left( 1+ \frac{t^2}{2!} +
\frac{t^4}{4!}+ \dots + \frac{t^{2l}}{(2l)!}+ \dots \right) \nonumber \\
&=& x^2\cosh{t} +y^2\sinh{t}. \nonumber
\end{eqnarray}
{\bf Example 3.} Consider nonlinear wave-type equation
\begin{equation}\label{pa}
\frac{\partial^2 u(x,y,t)}{\partial t^2}= 2x^2+2y^2 +\frac{15}{2} x \left( \frac{\partial^2 u(x,y,t)}{\partial x^2}\right )^2 + \frac{15}{2}y \left( \frac{\partial^2 u(x,y,t)}{\partial y^2}\right)^2
\end{equation}
with initial conditions
\begin{equation}\label{pa1}
u(x,y,0) = \frac{\partial u(x,y,0)}{\partial t} = 0.
\end{equation}
Then
$$
u_0(x,y,t) = \int_0^t (t-\xi)(2x^2+2y^2)d\xi = t^2(x^2+y^2)
$$
and
$$
u_p(x,y,t) = t^2(x^2+y^2)+ \frac{15}{2}\int_0^t (t-\xi)\left(x \left( \frac{\partial^2 u_{p-1}(x,y,\xi)}{\partial x^2}\right )^2 + y \left( \frac{\partial^2 u_{p-1}(x,y,\xi)}{\partial y^2}\right)^2 \right) d\xi,
$$
$p\geq 1.$ From here we obtain the following iterations:
\begin{eqnarray}
u_1(x,y,t)&=& t^2(x^2+y^2)+ 30 \int_0^t (t-\xi)(x\xi^4 + y\xi^4) d\xi = t^2(x^2+y^2)+ t^6 (x+y), \nonumber \\
u_2(x,y,t)&=& t^2(x^2+y^2)+ 30 \int_0^t (t-\xi)(x\xi^4 + y\xi^4) d\xi = t^2(x^2+y^2)+ t^6 (x+y), \nonumber \\
\vdots \nonumber \\
u_l(x,y,t) &=& t^2(x^2+y^2)+ t^6 (x+y), \ l\geq 2.
\end{eqnarray}
Then we obtain required unique solution of \eqref{pa}, \eqref{pa1} as
$$
u(x,y,t)= t^2(x^2+y^2) + t^6(x+y).
$$
{\bf Example 4.} Consider three-dimensional wave-type equation with variable coefficients
\begin{equation}\label{h}
\frac{\partial^2 u}{\partial t^2}- \frac{1}{2} \left( x^2 \frac{\partial^2 u}{\partial x^2}+ y^2\frac{\partial^2 u}{\partial y^2}+z^2\frac{\partial^2 u}{\partial z^2}\right) = x^2+y^2+z^2
\end{equation}
with initial conditions
\begin{equation}\label{h1}
u(x,y,z,t) =0, \quad \frac{\partial u(x,y,z,t)}{\partial t} = x^2+y^2-z^2.
\end{equation}
Then
$$
u_0(x,y,z,t) = t(x^2+y^2-z^2) + \int_0^t (t-\xi)(x^2+y^2+z^2)d\xi = t(x^2+y^2-z^2)+\frac{t^2}{2!}(x^2+y^2+z^2)
$$
and
\begin{eqnarray*}
u_p(x,y,z,t) &=& t(x^2+y^2-z^2)+\frac{t^2}{2!}(x^2+y^2+z^2) \\
&+& \frac{1}{2}\int_0^t(t-\xi)\left( x^2 \frac{\partial^2 u_{p-1}}{\partial x^2}+ y^2\frac{\partial^2 u_{p-1}}{\partial y^2}+z^2\frac{\partial^2 u{p-1}}{\partial z^2}\right) d\xi, \ p\geq 1.
\end{eqnarray*}
From here we get
\begin{eqnarray*}
u_1(x,y,z,t)&=& t(x^2+y^2-z^2)+\frac{t^2}{2!}(x^2+y^2+z^2) \\
&+& \frac{1}{2}\int_0^t (t-\xi)\left[ (x^2+y^2)(2\xi+\xi^2) +z^2(-2\xi + \xi^2 )\right ] d\xi \\
&=& t(x^2+y^2-z^2)+\frac{t^2}{2!}(x^2+y^2+z^2)+\frac{t^3}{3!}(x^2+y^2-z^2) +\frac{t^4}{4!}(x^2+y^2+z^2) \\
&=& (x^2+y^2)\left(t+\frac{t^2}{2!} + \frac{t^3}{3!}+\frac{t^4}{4!} \right) + z^2 \left(-t+\frac{t^2}{2!}-\frac{t^3}{3!}+\frac{t^4}{4!}\right),\\
u_2(x,y,z,t)&=& (x^2+y^2)\left(t+\frac{t^2}{2!} + \frac{t^3}{3!}+\frac{t^4}{4!}+\frac{t^5}{5!} +\frac{t^6}{6!}\right) + z^2 \left(-t+\frac{t^2}{2!}-\frac{t^3}{3!}+\frac{t^4}{4!}
-\frac{t^5}{5!} +\frac{t^6}{6!}\right),\\
\vdots \\
u_k(x,y,z,t) &=& (x^2+y^2)\left(t+\frac{t^2}{2!} + \frac{t^3}{3!}+ \dots +\frac{t^{2k+2}}{(2k+2)!}\right) + z^2 \left(-t+\frac{t^2}{2!}-\frac{t^3}{3!}+ (-1)^k \frac{t^{2k+2}}{(2k+2)!}\right),\\
\vdots
\end{eqnarray*}
Thus
\begin{eqnarray}
u(x,y,z,t) &=& \lim_{k\rightarrow \infty} \bigg [(x^2+y^2) \left( t+ \frac{t^2}{2!}+ \frac{t^3}{3!}+\dots + \frac{t^{2(k+1)}}{(2(k+1)!}\right) \nonumber \\
&+& z^2 \left (-t+\frac{t^2}{2!} - \frac{t^3}{3!}+ \dots -\frac{t^{2k+1}}{(2k+1)!} \right)\bigg] \nonumber \\
&=& (x^2+y^2)e^t + z^2e^{-t} -(x^2+y^2+z^2) \nonumber
\end{eqnarray}
which is unique exact solution of IVP \eqref{h},\eqref{h1}.
\begin{remark} Some above mentioned initial value problems have been solved using homotopy analysis method, homotopy perturbation method or Adomian decomposition method respectively
(see \cite{jafari},\cite{jin},\cite{wazwaz1},\cite{adomian2}, \cite{din}). However, in contrast to iterative technique proposed in this paper, these methods require complicated calculations of multidimensional integrals, high order derivatives or Adomian's or He's polynomials as it can be seen in cited papers.
\end{remark}
\section{Conclusion}
\begin{itemize}
\item
We conclude that iterative algorithm presented in this paper is a powerful and efficient analytical technique suitable for numerical approximation of a solution of initial problem for wide class of partial differential equations of arbitrary order.
\item
There is no need for calculating multiple integrals or derivatives, only one integration in each step is performed. Less computational work is demanded compared to other methods (Adomian decomposition method, variational iteration method, homotopy perturbation method, homotopy analysis method).
\item
Expected solution is a limit of a sequence of functions, in contrast to other frequently used methods where a sum of a functional series is considered. Consequently, the form of a solution can be immediately controlled in each step.
\item
All notations are carefully described and proofs are treated rigorously, compared to many recently presented algorithms and methods.
\item
Region and rate of convergence depend on Lipschitz constant for righthand side $F$.
\item
Using presented approach, we are able not only to obtain approximate solution, but even there is a possibility to identify unique solution of initial problem in closed form.
\item
A specific advantage of this technique over any purely numerical method is that it offers a smooth, functional form of the solution in each step.
\item
Another advantage is that using our approach we avoided discretization, linearization or perturbation of the problem.
\item
There is a possibility to reduce computational effort by combining presented algorithm with Laplace transform since there is a convolution integral inside the iterative formula.
\item
Finally, a subject of further investigation is to develop the presented technique for systems of PDE's, to find modifications for solving equations with deviating arguments and for other types of problems (e.g. BVP's).
\end{itemize}
\section{Acknowledgements}
\label{sec6}
The first author was supported by the project CZ.1.07/2.3.00/30.0039 of Brno
University of Technology. The work of the second author was realised in CEITEC - Central European Institute of Technology with research infrastructure supported by the project CZ.1.05/1.1.00/02.0068 financed from European Regional Development Fund and by the project FEKT-S-11-2-921 of Faculty of Electrical Engineering and Communication, Brno University of Technology.
This support is gratefully acknowledged.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
A malicious software or \emph{malware} may refer to any kind of software that its functionality is to cause harm to a user, computer, or network \cite{SiHo}. Thus, any software with malicious purposes can be considered as malware.
The most hard-to-detect malware mutation is the metamorphic malware. According to the definitions given in
\cite{YoKa}, {\it metamorphism} is the process of transforming a piece of code, utilizing a mutation module called {\it metamorphic engine}, responsible for the replication of malware into copies that are structurally different. However, these copies tend to exhibit the same behavior. Specifically, a very important clue upon which is based our detection approach, is the fact that every new copy has modified structure, code sequence size and syntactic properties \cite{RaMaSu}, while its behavior remains the same.
\vspace*{0.15in}
\noindent {\bf Malware Detection.} The term \textit{malware detection} is referred to the process of determining whether a given program $\pi$ is malicious or benign according to an a priori knowledge \cite{ChJhSeSoBr,MaHi,IdMa,AlLaVeWa}. For this purpose there have been proposed several techniques that leverage various characteristics for distinguishing malicious from benign programs. However, an efficient malware detection is based on an important process, called {\it malware analysis}, which collects the required information.
More precisely, malware analysis \cite{BaMoKrKi} is the process of determining the purpose and the functionality or, in general, the behavior of a given malicious code. Such a process is a necessary prerequisite in order to develop efficient and effective detection and also classification methods; malware analysis is divided into two main categories, namely {\it Static} and {\it Dynamic} analysis \cite{SiHo}.
\begin{itemize}
\renewcommand{\labelitemi}{\scriptsize$\circ$}
\item{\bf Static analysis:} In static analysis the specimen (i.e., test sample) is examined without its execution, performing the analysis on its source code.
\item{\bf Dynamic analysis:} In dynamic analysis an execution of the malware has to be performed in order to collect the required data, concerning the behavior of a program. However this approach needs more expertise while is extremely dangerous for the host environment. As a result, in most of times dynamic analysis is performed in a virtual environment.
\end{itemize}
It is well known that the behavior of a program can be modeled based upon system-call dependencies as they capture its interaction with its hosting environment, the operating system. As easily one can understand, a representation that captures a sequences of system-calls would be liable since any reorder or addition of one or more system-calls could change the sequence. Thus, a more flexible representation that would capture their in between relations, as a graph in example, could address that problem \cite{KoCoKrKiZhWa}.
As mentioned in \cite{FRJhChSaYa}, most malware relies on system-calls in order to deliver their payload. Additionally, since the behavior of a malware program could be reflected by the effect on its host operating system's state, then its behavior can be modeled by a directed acyclic graph, generated from system-call traces collected during its execution \cite{PaReMuSu}, the so called behavior graph.
\vspace*{0.15in}
\noindent {\bf Our Contribution.} In this paper we present an elaborated graph-based algorithmic technique that effectively addresses the problem of malware detection. Our approach for malware detection is based on the, so far unexploited, information that system-calls of a program $\pi$ of similar functionality can be classified into the same group and also on a set of various similarity metrics concerning the dependencies between these groups.
More precisely, having an instance of a ScD graph $D[\pi]$, constructed by the system-calls invoked by a program $\pi$, we decided to empower our model against strong mutations by applying our detection technique on a weighted directed graph $D^*[\pi]$, which we call group dependency graph (or, for short GrD graph), resulting from $D[\pi]$ after grouping disjoint subsets of its vertices. Additionally, we propose the similarity metric NP-similarity that combines similarity metrics on qualitative, quantitative, and relational characteristics that are spread among the members of known malware families to achieve a distinction between a malware and a benign program
Finally, we evaluate our detection model and compare our results against the results achieved by a variety of techniques proving the potentials of our model.
\vspace*{0.15in}
\noindent {\bf Related Work.} Our model design is inspired by the use of system-call dependency graphs as described in \cite{BaReSo,ChJhKr,FRJhChSaYa}.
In~\cite{FRJhChSaYa}, Fredrikson {\it et al.} proposed an automatic technique for extracting optimally discriminative behavioral specifications, based on graph mining and concept analysis, that have a low false positive rate and at the same time are general enough, when used by a behavior based malware detector, to efficiently distinguish malicious from benign programs.
Christodorescu {\it et al.} \cite{ChJhKr} propose an algorithm that automatically constructs {\it specifications} of malicious behavior needed by AV's in order to detect malware. The proposed algorithm constructs such specifications by comparing the execution behavior of a known malware against the corresponding behaviors produced by benign programs.
Finally, Babic {\it et al.} \cite{BaReSo} propose an approach to learn and generalize from the observed malware behaviors based on tree automate interference where the proposed algorithm infers $k$-testable tree automata from system-call data flow dependency graphs in order to be utilized in malware detection.
\vspace*{0.15in}
\noindent {\bf Road Map.} The remainder of this work is organized as follows. In Section~2 we present and analyze our proposed model for malware detection based on group dependency graphs and describe the corresponding graph construction procedure. In Section~3 we present a similarity metric that combines qualitative, quantitative, and relational characteristics. In Section~4 we analyze our data set, describe our experimental design, evaluate our proposed model's implementation against real malware samples, and compare our results against the ones achieved by other models. Finally, in Section~5 we conclude our paper and discuss possible future extensions.
\section{Model Design}
In this section, we leverage the so far unexploited grouping of system-calls, invoked by a program $\pi$, into groups of similar functionality and construct a graph that its vertex set consists of super-nodes containing the system-calls belonging to the same group, while its edge set contains the interconnection between the system-calls of these groups.
\subsection{The System-call Dependency Graph $G$}
It is well known that the actions performed by a program, depicting its behavior, rely on system-calls. Tracing the system-calls performed during the execution of a malware program $\pi$, we can represent its behavior interpreting this information with a graph, so called {\it System-call Dependency Graph} (or, ScD for short); throughout the paper, we shall denote a ScD graph by $D[\pi]$ and the system-calls invoked by $\pi$ by $S_i$, $1 \leq i \leq n$.
The vertex set of a ScD graph $D[\pi]$ is consisted by all the system-calls that take place during the execution of a program, i.e., $S_1$, $S_2$, $\ldots$, $S_n$, while its edge set contains the pairs of system-calls that exchanged arguments during the execution. Thus, an edge of ScD graph $D[\pi]$ is a tuple of type $(S_{i}$:$k$, $S_{j}$:$\ell)$ indicating that the system-call $S_{i}$ invokes $S_{j}$ and the $k^{th}$ output argument of $S_{i}$ is passed as the $\ell^{th}$ input argument of $S_{j}$.
\begin{table}[!h]
\begin{minipage}[c]{2.7in}
\centering
\begin{tabular}{|c|l|c|c|}
\hline
\textbf{ID} & \textbf{\ System-call Name} &\textbf{In} & \textbf{Out} \\ \hline \hline
0 & \ NtOpenSection \ & 2 & 1 \\ \hline
1 & \ ACCESS\_MASK \ & 0 & 1 \\ \hline
2 & \ POBJECT\_ATTRIBUTES \ & 0 & 1 \\ \hline
3 & \ NtQueryAttributesFile \ & 1 & 1 \\ \hline
4 & \ NtRaiseHardError \ & 5 & 0 \\ \hline
5 & \ NTSTATUS \ & 0 & 1 \\ \hline
6 & \ ULONG \ & 0 & 1 \\ \hline
7 & \ PULONG\_PTR \ & 0 & 1 \\ \hline
8 & \ HARDERROR\_RESPONSE\_OPTION \ & 0 & 1 \\ \hline
\end{tabular}
\vspace{0.1in}
\end{minipage}
\begin{minipage}[c]{2.7in}
\centering
\vspace{-0.07 in}
\begin{tabular}{|c|c|}
\hline
\textbf{Trace} & \textbf{Edge}\\ \hline \hline
\ 1:1,0:1 \ & $S_{1} \longrightarrow S_{0} $\\ \hline
\ 2:1,0:1 \ & $S_{2} \longrightarrow S_{0}$ \\ \hline
\ 2:1,3:1 \ & $S_{2} \longrightarrow S_{3} $\\ \hline
\ 2:1,3:1 \ & $S_{2} \longrightarrow S_{3} $\\ \hline
\ 2:1,3:1 \ & $S_{2} \longrightarrow S_{3} $\\ \hline
\ 2:1,3:1 \ & $S_{2} \longrightarrow S_{3} $\\ \hline
\ 2:1,3:1 \ & $S_{2} \longrightarrow S_{3} $\\ \hline
\ 2:1,3:1 \ & $S_{2} \longrightarrow S_{3} $\\ \hline
\ 2:1,3:1 \ & $S_{2} \longrightarrow S_{3} $\\ \hline
\ 2:1,3:1 \ & $S_{2} \longrightarrow S_{3} $\\ \hline
\ 2:1,3:1 \ & $S_{2} \longrightarrow S_{3} $\\ \hline
\ 2:1,3:1 \ & $S_{2} \longrightarrow S_{3} $\\ \hline
\ 5:1,4:1 \ & \ $S_{5} \longrightarrow S_{4} $ \ \\ \hline
\ 6:1,4:2 \ & \ $S_{6} \longrightarrow S_{4} $ \ \\ \hline
\ 6:1,4:3 \ & \ $S_{6} \longrightarrow S_{4} $ \ \\ \hline
\ 7:1,4:4 \ & \ $S_{7} \longrightarrow S_{4} $ \ \\ \hline
\ 8:1,4:5 \ & \ $S_{8} \longrightarrow S_{4}$ \ \\ \hline
\end{tabular}
\vspace{0.19in}
\end{minipage}
\centering \hspace{0.7 in} {\bf (a)} \hspace{2.3 in} {\bf (b)} \vspace{0.19in}
\vspace*{-0.1 in}
\caption{(a) System-calls appeared during the execution of a program $\pi$ from a malware family Hupigon, (b) System-call dependencies \cite{Data-Set}.}
\label{Tab1}
\end{table}
\begin{figure}[t!]
\hrule\medskip\smallskip
\centering
\includegraphics[scale=0.50]{ScDG-1.eps}
\centering
\smallskip\medskip\hrule\medskip
\caption{\small{A system-call dependency graph $D[\pi]$ of a program $\pi$.}}
\label{fig:fig1}
\end{figure}
Next, we illustrate a simple example that includes the system-call traces obtained through dynamic taint analysis \cite{BaReSo} during the execution of a sample from malware family Hupigon, downloaded from Domagoj Babic's personal webpage~\cite{Data-Set}, and we explain how the ScD graph is constructed after the whole process. Observing the data from the Table~\ref{Tab1}, we can see the construction of the ScD graph $D[\pi]$ that is a directed acyclic graph (dag); see, Figure~\ref{fig:fig1}. It is easy to see that the vertex set of this graph is consisted from the system-calls appeared during the execution of the sample and its edge set is consisted by their in between data-flow dependencies; see, Tables~\ref{Tab1}(a) and \ref{Tab1}(b).
Finally, we recall a well known fact that is the suspicious sample needs to be executed in a contained environment (i.e., a virtual machine), where during its execution time, taint analysis is performed in order to capture system-call traces.
\subsection{The Group Dependency Graph $D^*[\pi]$}
The key idea of our detection model is based on the, so far unexploited, information that system-calls of program $\pi$ of similar functionality can be classified into the same group, as we firstly presented it in \cite{ChNiPo}. For a proper system-call grouping we utilized the grouping provided by NtTrace \cite{NtTrace}, a system-call monitoring tool for MS Windows, complying with Micorsoft's documentation, where each system-call has a detailed description indicating the group it belongs to; we denote by $\mathcal{C}^*$ the set of system-call groups for a given operating system and by $\mathcal{C}_1$, $\mathcal{C}_2$, $\ldots$, $\mathcal{C}_{n^*}$ the groups of $\mathcal{C}^*$.
Thus, if a system-call dependency graph $D[\pi]$ of a given program $\pi$ is composed by $n$ system-calls $S_1$, $S_2$, $\ldots$, $S_n$, then each system-call $S_i$, $1 \leq i \leq n$, belongs to exactly one group $\mathcal{C}_j$, $1 \leq j \leq n^*$.
Having the grouping $\mathcal{C}^*$ and a system-call dependency graph $D[\pi]$, we next construct the key component of our model that is the {\it Group Dependency Graph} (or, GrD for short). The GrD graph, which we denote by $G^*[\pi]$, is a directed weighted graph on $n^*$ nodes $u_1$, $u_2$, $\ldots$, $u_{n^*}$; it is constructed as follows:
\begin{itemize}
\item[(i)] we first define a bijective function $f: V(G^*[\pi]) \longrightarrow \mathcal{C}_i$ from the node set $V(G^*[\pi])$ $=$ $\{u_1, u_2, \ldots, u_{n^*}\}$ to the set of groups $\mathcal{C}_i$ $=$ $\{\mathcal{C}_1, \mathcal{C}_2, \ldots, \mathcal{C}_{n^*}\}$;
\item[(ii)] for every pair of nodes $\{u_i, u_j\} \in V(G^*[\pi])$, we add the directed edge~$(u_i, u_j)$ in $E(G^*[\pi])$ if $(S_p,S_q)$ is an edge in $E(G[\pi])$ and, $S_p \in \mathcal{C}_i$ and $S_q \in \mathcal{C}_j$, $1 \leq i, j \leq n^*$;
\item[(iii)] for each directed edge $(u_i, u_j) \in E(G^*[\pi])$, we assign the weight $w$ if there are $w$ invocations from a system-call in group $f(u_i)= \mathcal{C}_i$ to a system-call in group $f(u_j)= \mathcal{C}_j$, $1 \leq i, j \leq n^*$.
\end{itemize}
\noindent Figure~\ref{fig:fig2} depicts the GrD graph~$D^*[\pi]$ of the ScD graph~$D[\pi]$ of Figure~\ref{fig:fig1}; the set $Iset$ contains all the isolated nodes of $D^*[\pi]$.
\begin{figure}[t!]
\hrule\medskip\smallskip
\centering
\includegraphics[scale=0.50]{GrDG-1.eps}
\centering
\smallskip\medskip\hrule\medskip
\caption{\small{The group dependency graph $D^*[\pi]$ of the graph of Figure~\ref{fig:fig1}.}}
\label{fig:fig2}
\end{figure}
In Table~\ref{Tab2}, we present the groups of system-calls $\mathcal{C}_1$, $\mathcal{C}_2$, $\ldots$, $\mathcal{C}_{n^*}$ and the number of system-calls inside each group. We point out that the number of non-isolated nodes of graph $D^*[\pi]$ equals the number of groups formed by the system-call of graph $D[\pi]$; note that, the total number of nodes of $D^*[\pi]$ is always $n^*$. For example, the $9$ system-calls of ScD graph belong to $7$ groups (see, Table~\ref{Tab3}), the ScD graph $D[\pi]$ of Figure~\ref{fig:fig1} contains $9$ nodes, while its corresponding GrD graph $D^*[\pi]$ contains $7$ non-isolated nodes and thus $23$ isolated nodes in $Iset$; see, Figure~\ref{fig:fig2}.
It is extremely important to point out that while the ScD graph $D[\pi]$ is by definition an acyclic directed graph, the produced GrD graph $D^*[\pi]$ is not, in general, acyclic. As easily one can see that by grouping nodes in $D[\pi]$ it is very likely to create directed circles and/or self-loops; an indicative example appears in graph $D^*[\pi]$ of Figure~\ref{fig:fig2}.
\begin{table}[t!]
\centering
\begin{tabular}{|l|c|l|c|}
\hline
\textbf{\ Group Name} & \textbf{Size} & \textbf{\ Group Name} & \textbf{Size} \\ \hline \hline
\ ACCESS\_MASK \phantom{XX} & 1 & \ PHANDLE & 1 \\ \hline
\ Atom & 5 & \ PLARGE\_INTEGER & 1 \\ \hline
\ BOOLEAN & 1 & \ Process & 49 \\ \hline
\ Debug & 17 & \ PULARGE\_INTEGER \phantom{XX} & 1 \\ \hline
\ Device & 31 & \ PULONG & 1 \\ \hline
\ Environment & 12 & \ PUNICODE\_STRING & 1 \\ \hline
\ File & 44 & \ PVOID\_SIZEAFTER & 1 \\ \hline
\ HANDLE & 1 & \ PWSTR & 1 \\ \hline
\ Job & 9 & \ Registry & 40 \\ \hline
\ LONG & 1 & \ Security & 36 \\ \hline
\ LPC & 47 & \ Synchronization & 38 \\ \hline
\ Memory & 25 & \ Time & 5 \\ \hline
\ NTSTATUS & 1 & \ Transaction & 49 \\ \hline
\ Object & 19 & \ ULONG & 1 \\ \hline
\ Other & 36 & \ WOW64 & 19 \\ \hline
\end{tabular}
\medskip\medskip
\caption{The 30 system-call groups.}
\label{Tab2}
\vspace*{-0.2 in}
\end{table}
\begin{table}[b!]
\centering
\begin{tabular}{|c|l|c|}
\hline
\textbf{ID} & \textbf{\ System-call} & \textbf{Group} \\ \hline \hline
0 & \ NtOpenSection \ & Memory \\ \hline
1 & \ ACCESS\_MASK \ & \ ACCESS\_MASK \ \\ \hline
2 & \ POBJECT\_ATTRIBUTES \ & Object \\ \hline
3 &\ NtQueryAttributesFile \ & File\\ \hline
4 &\ NtRaiseHardError \ & Process\\ \hline
5 & \ NTSTATUS \ & NTSTATUS \\ \hline
6 & \ ULONG \ & ULONG \\ \hline
7 & \ PULONG\_PTR \ & Process\\ \hline
8 & \ HARDERROR\_RESPONSE\_OPTION \ & Process \\ \hline
\end{tabular}
\medskip\medskip
\caption{The 9 system-calls of Figure~\ref{fig:fig1} and their corresponding groups.}
\label{Tab3}
\end{table}
\subsection{Family Identity Matrix}
\label{sec:53}
In this section we will describe the construction of an informative adjacency matrix that will act as a unique identity for each malware family. Our approach is based on the intuition that malware samples belonging to an individual malware family tend to share common characteristics. This is a quite valuable information, that we leveraged in order to develop a technique that will utilize these characteristics in order to decide if an unknown sample is malware or not.
Defining the term {\it characteristic} when working on GrD graph $D^*[\pi]$, we could claim that a characteristic is an edge between two system-call groups, since in order for an individual task to be performed, system-calls of specific functionality need to be utilized and of course in different malware variants they can be substituted by equivalent ones. Thus, we decided to focus on edges that exist in most of the members' GrD graphs $D^*[\pi]$, constituting hence a qualitative characteristic of their family.
So, easily one can understand that, if in a malware family, a specific edge, appears in the majority of the members, then this edge exposes a greater significance, in contrast with another one that exists in the minority of the members of this family. Hence, in order to represent the significance of an edge we take into account the percentage of the members in a family in which this edge has a non-zero value. To this point we ought to underline that, since the values in the cells of adjacency matrix refer to the weight $w$ of the corresponding edge, in order to claim about the significance of an edge as a qualitative characteristic of a family, we are interested only on the non-zero weights.
Thus, having collected this valuable information we proceed by filtering it as to decide the significant edges that will indicate the characteristics of each family. Hence, having computed the percentage of appearance of each edge we can assign weights to each cell (i.e., edge) on this matrix constructing hence the {\it ID-matrix} of the family as shown in Figure~\ref{FamilyID}.
Finally, in order to assign weights we partition the values (ranging from 0 to 100) to three categories. However, before we assign the significance tags we ought to define the value ranges. So, we first define a threshold about 95$\%$ and the tags are arranged based on this threshold. Thus, we mark each cell with a tag either {\tt Red}, {\tt Gray}, or {\tt White}, with value {\tt 4}, {\tt 3}, or {\tt 2}, respectively.
In our model, the {\tt Red} tags cover cells containing values in the range $[0.95-1]$, the {\tt Gray} tags cover cells containing values in the range $(0.05-0.95)$, while the {\tt White} tags cover the ones containing values in the range $[0-0.05]$.
\begin{figure}[t!]
\hrule\vspace*{0.3cm}
\centering
\includegraphics[scale=0.50]{ID-Matrix-1.eps}
\centering
\medskip\hrule\medskip
\caption{\small{The structure of the ID-matrix.}}
\medskip
\label{FamilyID}
\end{figure}
\section{Model Metrics}
In this section, we present a similarity metric that combines qualitative, quantitative, and relational characteristics that are spread among the members of known malware families to achieve a distinction between a malware and a benign software.
\subsection{GrD-Similarity Metrics}
\label{sec:GrDSM}
\noindent We next present the proposed methodology for malware detection. Specifically, we present the computation of the similarity between a test sample $T$ and every malware family $F_{k}$ of the set $\mathcal{F}$, $1 \leq k\leq |\mathcal{F}|$. Before we proceed with the presentation of the similarity metrics let us give some definitions and notations used throughout the paper.
\begin{definition}
Let $A$, $B$ two $n \times n$ matrices with elements $A(\cdot)$, $B(\cdot) \in \mathrm{R}$, and let $p, q \in \mathrm{Z}$. Then, we define
\end{definition}
\begin{equation}
|A(i,j) \cap_{p\rightarrow q} B(i,j)|=
\begin{cases}
1, & \mbox{if } A(i,j)=p \mbox{ and } B(i,j)=q, \ 1 \leq i,j \leq n \\
0, & \mbox{otherwise}
\end{cases}
\end{equation}
\noindent and
\begin{equation}
|A \cap_{p\rightarrow q} B\vert = \sum\limits_{i=1}^{n}\sum\limits_{j=1}^{n}\vert A(i,j) \cap_{p\rightarrow q} B(i,j)|.
\end{equation}
\vspace*{0.1in}
\noindent By definition, $|A\cap_{p\rightarrow p} A|$ is the number of the elements of matrix $A$ with value $p$; hereafter, this number is referred to as $|A=p|$.
Next, we provide a set of GrD graph similarity metrics along with a description of their qualitative, quantitative, and relational characteristics. The following adjacency matrices represent the GrD graph $D^*[\pi]$
\vspace*{0.2in}
\noindent {\bf A. Family-Test Similarity}
\vspace*{0.05in}
\noindent The Family-Test similarity metric focuses on the computation of the similarity between the test sample and any malware family using the family's ID-matrix. We distinguish two similarity approaches, namely
\begin{itemize}
\item[$\circ$] Family.to.Test cover, and
\item[$\circ$] Test.to.Family cover.
\end{itemize}
\noindent The main purpose of the Family.to.Test cover similarity metric is to compute the rate of satisfiability on the qualitative characteristics of any family of the set $\mathcal{F}$ by a test sample, while the main purpose of the Test.to.Family cover similarity metric is to compute the rate of satisfiability, in terms of edge existence, of a test sample's adjacency matrix by the qualitative characteristics of any family represented by its ID-matrix.
In order to compare the test sample's adjacency matrix $T$ with the ID-matrix $F_k$ of the $k$th family of the set $\mathcal{F}$, we first need to make a cast on test sample's adjacency matrix $T$. Recall that, the cells of the test sample's adjacency matrix have either zero or non-zero values. Thus, we cast any non-zero values existed in test sample's adjacency matrix $T$ into 1s resulting the casted matrix $T^{c}$.
\vspace*{0.1in}
\noindent {\bf Family.to.Test cover:} The main process of this similarity metric is to cover in some fashion some qualitative characteristics of family $F_k$ with the edge existence of the test sample. We achieve such a coverage by first computing the number of cells with a Red tag or, equivalently, a value 4 in family's ID-matrix $F_k$ that their corresponding cells in the test sample's casted matrix $T^{c}$ have value 1, i.e., $F_{k}(i,j)=4$ and $T^{c}(i,j)=1$, and then dividing this number by the total number of cells in family's ID-matrix that have a Red tag, i.e., $F_{k}(i,j)=4$.
We call the above proposed cover {\it Four.to.One cover} and denote it, in a form of function, as $Four.to.One(F_{k},T^{c})$. Thus, the formula that gives the Family.to.Test cover similarity metric is the following:
\begin{equation}
Four.to.One(F_{k},T^{c})=\dfrac{\vert F_{k} \cap_{4\rightarrow 1} T^{c} \vert}{ \vert F_{k}=4 \vert}
\end{equation}
\noindent where, $F_{k}$ is the ID-matrix of the $k^{th}$ family of a set $\mathcal{F}$ of families and $T^{c}$ is the test sample's casted adjacency matrix.
\vspace*{0.1in}
\noindent {\bf Test.to.Family cover:} This metric computes the satisfiability of the edge existence in the test sample, represented by the topology of 1s in the casted matrix $T^{c}$, by the qualitative characteristics of a family $F_k$, represented by the topology of Red tags in its ID-matrix. In a similar way, we achieve a Test.to.Family coverage by first computing the number of cells in the test sample's casted matrix $T^{c}$ having value 1 that their corresponding cells in family's ID-matrix $F_k$ have a Red tag or, equivalently, a value 4, i.e., $T^{c}(i,j)=1$ and $F_{k}(i,j)=4$, , and then dividing this number by the total number of cells in the test sample's casted matrix $T^{c}$ that having value 1, i.e., $T^{c}(i,j)=1$.
As above, we also call the proposed cover {\it One.to.Four cover} and denote it, in a form of function, as $One.to.Four(T^{c},F_{k})$. Thus, the formula that gives the Test.to.Family cover similarity metric is the following:
\begin{equation}
One.to.Four(T^{c},F_{k})=\dfrac{\vert T^{c} \cap_{1\rightarrow 4} F_{k} \vert}{\vert T^{c}=1 \vert}
\end{equation}
\noindent where, again $F_{k}$ is the ID-matrix of the $k^{th}$ family of a set $\mathcal{F}$ of families and $T^{c}$ is the test sample's casted adjacency matrix.
\vspace*{0.2in}
\noindent {\bf B. Jaccard Similarity}
\vspace*{0.05in}
\noindent One more similarity metric we utilize to empower our formula for malware detection is the Jaccard index \cite{jaccard_tanimoto}. The reason we choose to utilize the Jaccard similarity is the fact that it is mostly applied on binary vectors and thus it seems to efficiently work for the comparison between two graph-objects in terms of edge existence. More precisely, this metric measures the similarity of relational characteristics, in terms of edge existence, between the test sample's casted matrix $T^{c}$ and a member's casted matrix $M^{c}$ of a malware family $F_{k}$.
We first utilize the Jaccard index to compute the maximum value produced by the most similar member of the family to the test sample, and then we compute the mean similarity between the test sample and all the members of a malware family .
The computation of the Jaccard similarity metric is achieved by first computing the number of cells that have value 1 in the test sample's casted matrix $T^{c}$ and their corresponding cells in the member's casted matrix $M^{c}$ have also value 1, i.e., $T^{c}(i,j)=1$ and $M^{c}(i,j)=1$, and then dividing this number by the number of the cells that either in $T^{c}$ or in $M^{c}$ have value 1, i.e., $T^{c}(i,j)=1$ or $M^{c}(i,j)=1$. Hence the computation of Jaccard similarity can be computed as follows:
\begin{equation}
J(T^{c},M^{c})=\dfrac{\vert T^{c} \cap_{1\rightarrow 1} M^{c} \vert }{ \vert T^{c} \cap_{1\rightarrow 1}M^{c}\vert + \vert T^{c} \cap_{1\rightarrow 0}M^{c} \vert + \vert T^{c} \cap_{0\rightarrow 1}M^{c} \vert} ,
\end{equation}
\vspace*{0.06in}
\noindent where $M^{c}$ is the member's casted adjacency matrix and $T^{c}$ is the test sample's casted adjacency matrix.
Let $F_{k}$ be a malware family containing $m_k$ members and let $M^{c}_{1}, M^{c}_{2}, \ldots, M^{c}_{m_k}$ be the casted matrices of the members of $F_{k}$. The maximum Jaccard similarity, produced by the most similar member of $F_{k}$ to the test sample, is defined as follows:
\begin{equation}
J_{\text{max}}(T^{c},F_{k})=\max\limits_{1\leq \ell \leq m_{k}}[J(T^{c},M^{c}_{\ell})],
\end{equation}
\vspace*{0.06in}
\noindent while the mean Jaccard similarity, produced by the mean of all the Jaccard similarity values between the test sample and the members of $F_{k}$, is defined by the following formula:
\begin{equation}
J_{\text{mean}}(T^{c},F_{k})=\dfrac{\sum\limits_{\ell=1}^{m_{k}}J(T^{c},M^{c}_{\ell})}{m_{k}},
\end{equation}
\noindent where $F_{k}$ is the $k^{th}$ malware family of a set $\mathcal{F}$ and $T^{c}$ is the test sample's casted adjacency matrix.
\vspace*{0.2in}
\noindent {\bf C. Bray-Curtis Similarity}
\vspace*{0.05in}
\noindent The last similarity metric we utilize to empower our formula for malware detection is the Bray-Curtis dissimilarity \cite{Bray_Curtis}. The reason we select the Bray-Curtis dissimilarity is the fact that it is mostly applied for the computation of diversity between two object represented by vectors of continuous values. More precisely, this metric measures the similarity of quantitative characteristic, in terms of edge weights, between the test samples' adjacency matrix $T$ and a member's adjacency matrix $M$. However, since the return value of the Bray-Curtis metric refers to the distance between any two objects, it is in the range $[0,1]$ with maximum value the 0. Thus, in order to reverse this property we perform a subtraction from 1.
As with the Jaccard similarity, we also utilize the Bray-Curtis similarity to compute first the maximum value produced by the most similar member of the family to the test sample, and then we compute the mean similarity between the test sample and all the members of a malware family.
The computation of Bray-Curtis similarity is achieved by computing the sum of subtractions of the corresponding values of cells $T(i,j)$ and $M(i,j)$ and dividing this number by the sum of their additions. Hence, the Bray-Curtis similarity is given by the following formula:
\begin{equation}
BC(T,M) =1-\dfrac{\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{n} (T(i,j) - M(i,j))}
{\sum\limits_{i=1}^{n}\sum\limits_{j=1}^{n} (T(i,j) + M(i,j))} ,
\end{equation}
\noindent where $T$ is the $n \times n$ adjacency matrix of test sample and $M$ is the $n \times n$ adjacency matrix of the member of family $F_{k}$ under consideration.
Next, we first define the maximum Bray-Curtis similarity, produced by the most similar member of $F_{k}$ to the test sample, as follows:
\begin{equation}
BC_{\text{max}}(T,F_{k})=\max\limits_{1\leq \ell \leq m_{k}}[BC(T,M_{\ell})] ,
\end{equation}
\noindent and then we define the mean Bray-Curtis similarity, produced by the mean of all the Bray-Curtis similarity values between the test sample and the members of $F_{k}$, as follows:
\begin{equation}
BC_{\text{mean}}(T,F_{k})=\dfrac{\sum\limits_{\ell=1}^{m_{k}}BC(T,M_{\ell})}{m_{k}} ,
\end{equation}
\noindent where $F_{k}$ is the $k^{th}$ malware family of the set $\mathcal{F}$, $m_{k}$ is the number of members of the family $F_{k}$, $M_{\ell}$ is the $\ell^{th}$ member's adjacency matrix and $T$ is the test sample's adjacency matrix.
\vspace*{0.2in}
\noindent {\bf D. Tanimoto Similarity}
\vspace*{0.05in}
\noindent Finally we use the Tanimoto similarity \cite{jaccard_tanimoto} for filtering purposes in our proposed malware detection model. The Tanimoto similarity is a mechanism for computing the Jaccard coefficient when the set under comparison are represented as bit vectors.
As with the previously described similarity metrics, we also utilize the Tanimoto similarity to compute the maximum value produced by the most similar member of the family to the test sample.
The computation of Tanimoto similarity is achieved by computing the sum of subtractions of the corresponding values of cells $T(i,j)$ and $M(i,j)$ and dividing this number by the sum of their additions. Hence, the Tanimoto similarity is given by the following formula:
\begin{equation}
TN(A,B)=\dfrac{\sum\limits_{i=1}^{n} (T(i,j) \times M(i,j))}{\sum\limits_{i=1}^{n} (T(i,j))^2 + \sum\limits_{i=1}^{n} (T(i,j))^2 - \sum\limits_{i=1}^{n} (T(i,j) \times M(i,j))} ,
\end{equation}
\noindent where $T$ is the $n \times n$ adjacency matrix of test sample and $M$ is the $n \times n$ adjacency matrix of the member of family $F_{k}$ under consideration.
Next, we define the max Tanimoto similarity, produced by the most similar member of $F_{k}$ to the test sample, as follows:
\begin{equation}
TN_{\text{max}}(T,F_{k})=\max\limits_{1\leq \ell \leq m_{k}}[TN(T,M_{\ell})] ,
\end{equation}
\noindent where $F_{k}$ is the $k^{th}$ malware family of the set $\mathcal{F}$, $m_{k}$ is the number of members of the family $F_{k}$, $M_{\ell}$ is the $\ell^{th}$ member's adjacency matrix and $T$ is the test sample's adjacency matrix.
\subsection{NP-similarity Metric}
\label{sec:NP}
Having presented several variants of Family-Test, Jaccard, and Bray-Curtis similarity metrics, let us now describe the NP-similarity metric which we have developed in order to detect whether an unknown test sample is a malicious or a benign one. More precisely, this metric globally measures the similarity between a test sample and a malware family combining, in a specific manner, the aforementioned similarity metrics taking into account the qualitative, the quantitative, and the relational characteristics of the objects under consideration.
The NP-similarity incorporates a combination of the similarity metrics referenced previously, where their contribution to its final result is affected by assigning different weights to each one of these similarity metrics. For our purpose, we choose four factors $a$, $b$, $c_{1}$, and $c_{2}$, and define three similarity-components namely $F_1$, $F_2$, and $F_3$.
The first similarity-components $F_1$ of our NP-similarity metric concerns the qualitative characteristics. Thus, we utilize the
Four.to.One cover similarity along with the One.to.Four similarity assigning greater weight factors in the Four.to.One. Indeed, we choose the factor $a=4$ for the Four.to.One similarity and the factor $b=2$ for the One.to.Four, while we choose greater weight factor for the case where both $Four.to.One()$ and $One.to.Four()$ take the maximum value 1; we express our choice by the following function:
\begin{equation*}
\phi= a \cdot Four.to.One(F_{k},T^{c}) + b \cdot One.to.Four(T^{c},F_{k})
\end{equation*}
\noindent Our choice is based on the intuition that, if the test sample is malicious, then it should be an expansion of a malware family inheriting and hence satisfying its qualitative characteristics. Moreover, the reason that we multiply the $Four.to.One()$ similarity by a greater factor is the fact that when this similarity metric is maximized it is indicating that the sample is a direct extension of the malware family.
Additionally, in the case where $Four.to.One()=1$ and $One.to.Four()=1$, the topology of 4s in Family ID-Matrix is identical to the topology of 1s in test samples casted matrix and thus we multiply the function $\phi$ by the factor $c_{1}=1.5$, otherwise we multiply it by the factor $c_{2}=1.2$. Hence, we define the first similarity-components $F_1$ of our NP-similarity metric as follows:
\begin{equation}
F_{1}=
\begin{cases}
c_1 \cdot (a + b), & \mbox{if $Four.to.One(F_{k},T^{c})=One.to.Four(T^{c},F_{k})=1$} \\
c_2 \cdot \phi, & \mbox{otherwise}
\end{cases}
\end{equation}
\noindent where, $a=4$, $b=2$, $c_{1}=1.5$ and $c_{2}=1.2$, while $F_{k}$ is the $k^{th}$ malware family of a set $\mathcal{F}$ and $T^{c}$ is the test sample's casted adjacency matrix.
The second similarity-components $F_2$ of our NP-similarity metric measures the similarity of relational
characteristics between the test sample and a malware family as described by the Jaccard index.
We assign appropriate weights on the max and mean Jaccard similarities, i.e., $J_{\text{max}}()$ and $J_{\text{mean}}()$, by using the factors $a$ and $b$, as follows:
\begin{equation}
F_{2}=a \cdot J_{max}(T^{c},F_{k}) + b \cdot J_{mean}(T^{c},F_{k})
\end{equation}
\noindent where, $a=4$ and $b=2$.
We next proceed by defining the third similarity-components $F_3$ that measures the similarity of qualitative
characteristics assigning, as before, appropriate weights on the max and mean Bray-Curtis similarities as follows:
\begin{equation}
F_{3}=a \cdot BC_{max}(T,F_{k}) + b \cdot BC_{mean}(T,F_{k})
\end{equation}
\noindent where, $a=4$ and $b=2$.
We point out that in similarity-components $F_{2}$ and $F_{3}$ we assign a greater weight to the max Jaccard and max Bray-Curtis since, as we describe above, it is more likely for the test sample to be a direct mutation of a member of a malware family in the case where it is a malware.
We finally define our NP-similarity metric by combining the three similarity-components $F_1$, $F_2$, and $F_3$, as follows:
\begin{equation}
NP(F_{k},T)=\dfrac {F_{1} \cdot F_{2} \cdot F_{3}}{Q},
\end{equation}
\noindent where, $Q$ is a normalization factor equals the maximum value of the product $F_1 \cdot F_2 \cdot F_3$ so that $NP() \in [0,1]$.
\bigskip
\noindent \textbf{Intuition.} The whole process of the NP-similarity construction, by the aspect of weights assignment, is based on the intuition that during the polymorphism procedure, it is more probable for a new strain to be a direct mutation from a member of a malware family. Hence, so for first component, in the case of the Four.to.One similarity, as for the other two, in the cases of max Jaccard and max Bray-Curtis similarities respectively, we assign a greater weight on to them as to emphasize that probability.
\subsection{Malware Detection using NP Similarity}
\label{sec:MDGrD}
Next, we show how can utilize the NP-similarity for malware detection based on GrD graphs; recall that, for a malware program $\pi$ the GrD graph is denoted by $D^*[\pi]$. The methodology we follow is simple: given a test sample $T$, we compute the NP-similarity metric between $T$ and all the malware families $F_{1}, F_{2}, \ldots, F_{n}$ of a set $\mathcal{F}$, and then we accordingly compute the max Tanimoto similarity exhibited by a member of each family. Finally, for the families with the maximum Tanimoto similarity, we check the one with the corresponding maximum NP-similarity and if this maximum value, indicating the most similar family $F_{k}$ to $T$ according to NP-similarity, is above the specified threshold $\lambda$, i.e., $NP(F_{k},T)~\geq~\lambda$, we claim that the test sample belongs to $\mathcal{F}$ and thus it is a malware.
It is worth noting to mention that, as we will discuss later, it is experientially proven that the application of the NP-similarity archives a satisfying distinction between the GrD graphs $D^*[\pi]$ representing malware and those ones representing benign software.
\section{Evaluation}
In this section we first present our experimental design and discuss the reasons that we adopt the proposed evaluation setup. Then, we discuss how we divide our data set into train-set and test-set and how we tune our threshold parameters according to feedback produced by a series experiments. Finally, we present our detection results after the application of NP-similarity and compare our results with those of other models.
\subsection{Experimental Design}\label{sec:61}
In order to evaluate our proposed malware detection technique we use a dataset of $2631$ malware samples from a set $\mathcal{F}$ of $48$ malware families $F_1$, $F_2$, $\ldots$, $F_{48}$, each $F_k$ containing from $3$ to $317$ malware members, and also a set of $33$ benign samples.
Additionally, it is of major importance to mention that we do not perform any taint malware analysis on the samples due to the risk posed to the systems connected to the same network. Thus, we downloaded the initial System-call Dependency Graphs produced by taint analysis from the web-page of Domagoj~Babic~\cite{Data-Set} and transformed each sample's ScD graph $D[\pi]$ into GrD graph $D^*[\pi]$, based on the grouping of system-calls presented in Table~\ref{Tab2}. The set $\mathcal{F}$ of the $48$ malware families along with their sizes (i.e., number of members) are listed in Table~\ref{Tab6}.
\begin{table}[h!]
\begin{center}
\begin{tabular}{|l|c|l|c|}
\hline
\textbf{\ Family Name} & \textbf{Size} & \textbf{\ Family Name} & \textbf{Size} \\ \hline \hline
\ ABU,Banload & 16 & \ Hupigon,AWQ & 219 \\ \hline
\ Agent,Agent & 42 & \ IRCBot,Sdbot & 66 \\ \hline
\ Agent,Small & 15 & \ LdPinch,LdPinch & 16 \\ \hline
\ Allaple,RAHack & 201 & \ Lmir,LegMir & 23 \\ \hline
\ Ardamax,Ardamax & 25 & \ Mydoom,Mydoom & 15 \\ \hline
\ Bactera,VB & 28 & \ Nilage,Lineage & 24 \\ \hline
\ Banbra,Banker & 52 & \ OnLineGames,Delf & 11 \\ \hline
\ Bancos,Banker & 46 & \ OnLineGames,LegMir & 76 \\ \hline
\ Banker,Banker & 317 & \ OnLineGames,Mmorpg & 19 \\ \hline
\ Banker,Delf & 20 & \ OnLineGames,OnLineGames\phantom{xx} & 23 \\ \hline
\ Banload,Banker & 138 & \ Parite,Pate & 71 \\ \hline
\ BDH,Small & 5 & \ Plemood,Pupil & 32 \\ \hline
\ BGM,Delf & 17 & \ PolyCrypt,Swizzor & 43 \\ \hline
\ Bifrose,CEP & 35 & \ Prorat,AVW & 40 \\ \hline
\ Bobax,Bobic & 15 & \ Rbot,Sdbot & 302 \\ \hline
\ DKI,PoisonIvy & 15 & \ SdBot,SdBot & 75 \\ \hline
\ DNSChanger,DNSChanger \phantom{xx} & 22 & \ Small,Downloader & 29 \\ \hline
\ Downloader,Agent & 13 & \ Stration,Warezov & 19 \\ \hline
\ Downloader,Delf & 22 & \ Swizzor,Obfuscated & 27 \\ \hline
\ Downloader,VB & 17 & \ Viking,HLLP & 32 \\ \hline
\ Gaobot,Agobot & 20 & \ Virut,Virut & 115 \\ \hline
\ Gobot,Gbot & 58 & \ VS,INService & 17 \\ \hline
\ Horst,CMQ & 48 & \ Zhelatin,ASH & 53 \\ \hline
\ Hupigon,ARR & 33 & \ Zlob,Puper & 64 \\ \hline
\end{tabular}
\medskip\medskip
\caption{The set $\mathcal{F}$ of the $48$ malware families $F_1, F_2, \ldots, F_{48}$, along with their sizes, i.e., number of members, downloaded from~\cite{Data-Set}.}
\label{Tab6}
\end{center}
\vspace*{-0.2 in}
\end{table}
For evaluation purposes of our model, we perform $5$-fold cross validation utilizing the dataset we described above. Additionally, we set the detection threshold $\lambda=0.56$ (see, Section~\ref{sec:MDGrD}), after performing a number of experiments focusing on maximizing the ratio of true-positives by the false-positives.
\subsection{Detection Results}
Next, we present our results after performing a set of 5-fold cross validation experiments partitioning the data set described above into 5 buckets using in each experiment one bucket as test-set and the other four as train-set. In Table~\ref{Tab7} we cite our results concerning the detection rates and the corresponding false positives for various values of threshold $\lambda$ as we described it previously. To this point we ought to notice that due to the 5-fold cross validation process the percentage values below are averaged over the five buckets.
\begin{table}[h!]
\begin{center}
\begin{tabular}{|c||c||c|}
\hline
\textbf{\ \ Threshold $\lambda$ \ \ }& \textbf{\ \ Detection Rate \ \ } & \textbf{\ \ False Positives \ \ }\\ \hline
$\lambda$ = 0.35 & 98.14 \% & 68.57 \% \\ \hline
$\lambda$ = 0.42 & 96.70 \% & 56.00 \% \\ \hline
$\lambda$ = 0.51 & 94.06 \% & 29.00 \% \\ \hline \hline
$\lambda$ = 0.56 & 91.32 \% & 13.70 \% \\ \hline \hline
$\lambda$ = 0.61 & 85.28 \% & 6.85 \% \\ \hline
$\lambda$ = 0.67 & 74.42 \% & 4.00 \% \\ \hline
$\lambda$ = 0.74 & 63.03 \% & 2.28 \% \\ \hline
$\lambda$ = 0.81 & 39.64 \% & 0.00 \% \\ \hline
\end{tabular}
\medskip\medskip
\caption{Malware detection results.}\label{Tab7}
\end{center}
\vspace*{-0.2 in}
\end{table}
In Figure \ref{np_chart} we provide an observation of the detection rates (True Positives) and False Positives of our proposed model, while changing the value of threshold $\lambda$. As mentioned previously, due to the $k$-fold cross validation process all the percentage values shown in the figures are averaged over the $k$-folds, (i.e., 5 folds for our proposed model).
In Table~\ref{Tab8}, we illustrate a comparison of our detection rates (i.e., true-positives) and the fail detections (i.e., false-positives) against those presented in other research works. We compare our results to graph-based and other techniques both using different data-sets. More precisely, the first column refers to the result's host, the second one refers to the utilized technique, while the third and fourth columns refer to the detection and false-positive rates, respectively.
\begin{table}[b!]
\begin{center}
\begin{tabular}{|r||l|c|c|}
\hline
\textbf{In:}\phantom{x} & \textbf{\ Technique} & \textbf{True Positives} & \textbf{False Positives} \\ \hline\hline
\cite{AlLaVeWa}\phantom{x} & \ SVM classifier (API-sequences) & 89.74 \% & 09.74 \% \\ \hline
\cite{YeDiTaDo}\phantom{x} & \ OOA rules (API-sequences) & 97.19 \% & 00.12 \% \\ \hline
\cite{ChJhSeSoBr}\phantom{x} & \ Templates (CFG) & 97.50 \% & 00.00 \% \\ \hline\hline
\cite{KoCoKrKiZhWa}\phantom{x} & \ Sequence Matching (ScD) & 64.00 \% & 00.00 \% \\ \hline
\cite{LuTa}\phantom{x} & \ Graph-Grading(ScD) & 80.09 \% & 11.00 \% \\ \hline
\cite{FRJhChSaYa}\phantom{x} & \ Graph Mining (ScD) & 92.40 \% & 6.1 \% \\ \hline
\cite{BaReSo}\phantom{x} & \ Tree Automata Inference (ScD) \phantom{x} & 80.00 \% & 05.00 \% \\ \hline
{this paper}\phantom{x} & \ NP-Similarity (GrD) & 91.32 \% & 13.70 \% \\ \hline
\end{tabular}
\medskip\medskip
\caption{Malware detection results comparison. Note that this paper uses the same dataset as \cite{BaReSo} and \cite{FRJhChSaYa}.}\label{Tab8}
\end{center}
\vspace*{-0.2 in}
\end{table}
Alazab {\it et~al.}~\cite{AlLaVeWa} developed a fully automated system that disassembles and extracts API-call features from executables and then, using $n$-gram statistical analysis, is able to distinguish malicious from benign executables. The mean detection rate exhibited was 89.74\% with 9.72\% false-positives when used a Support Vector Machine (SVM) classifier by applying $n$-grams.
Ye {\it et al.}~\cite{YeDiTaDo} described an integrated system for malware detection based on API-sequences. This is also a different model from ours since the detection process is based on matching the API-sequences on OOA rules (i.e., Objective-Oriented Association) in order to decide the maliciousness or not of a test program.
\begin{figure}[t!]
\hrule\vspace*{0.3cm}
\centering
\includegraphics[scale=0.55]{NP_Chart.eps}
\centering
\medskip\hrule\medskip
\caption{\small{Detection results for multiple values of $\lambda$ depicting true and false positives variations.}}
\label{np_chart}
\end{figure}
An important work of Christodorescu {\it et al.}, presented in~\cite{ChJhSeSoBr}, proposes a malware detection algorithm, called $A_{\text{MD}}$, based on instruction semantics. More precisely, templates of control flow graphs are built in order to demand their satisfiability when a program is malicious. Although their detection model exhibits better results than the ones produced by our model, since it exhibits 0 false-positives, it is a model based on static analysis and hence it would not be fair to compare two methods that operate on different objects.
Kolbitch {\it et al.}~\cite{KoCoKrKiZhWa} proposed an effective and efficient approach for malware detection, based on behavioral graph matching by detecting string matches in system-call sequences, that is able to substitute the traditional anti-virus system at the end hosts. The main drawback of this approach is the fact that although no false-positives where exhibited, their detection rates are too low compared with other approaches.
Luh and Tavolato~\cite{LuTa} present one more detection algorithm based on behavioral graphs that distinguishes malicious from benign programs by grading the sample based on reports generated from monitoring tools. While the produced false-positives are very close to ours, the corresponding detection ratio is even lower.
Fredrikson {\it et al.}~\cite{FRJhChSaYa} proposed an automatic technique for extracting optimally discriminative specifications based on graph mining and concept analysis that, when used by a behavior based malware detector, it can efficiently distinguish malicious from benign programs. The proposed technique can yield an 86.5\% detection rate with 0 false-positives. Since we compare only the maximum detection rates exhibitied by each technique, in Table \ref{Tab8} we show the maximum detection rate 99.4\% which however exhibits higher false-positives (57.14\%). However, a more fair comparison would be the one depicted in Figures \ref{np_chart} and \ref{other_chart} where for specific values of $t$ and $\lambda$ (i.e. $t=0.96$ and $\lambda=0.56$) someone can observe that our model reaches the detection rates of the proposed model presented in \cite{FRJhChSaYa} with barely $0.03\%$ more false positives, proving the potentials of our model in a further improvement.
Finally, Babic {\it et al.}~\cite{BaReSo} achieved the malware detection by $k$-testable tree automata inference from system-call data flow dependence graphs. To this point we ought to underline that in this work the authors use the same data-set that we borrow from Domagoj~Babic's web-page~\cite{Data-Set}. Thus, this work provides a fair instance to compare our model's results. However, while Babic {\it et al.} perform 2-fold cross validation using the first half of data-set as train-set and the second one as test-set, we perform 5-fold cross validation. Comparing the results exhibited in \cite{BaReSo} with ours, easily we can claim that our proposed model is quite competitive to Babic's especially for specific values of $\lambda$ ($0.61$ and $0.67$ respectively).
\section{Concluding Remarks}
We have presented an elaborated graph-based algorithmic technique for efficient malware detection by exploiting main properties of system-call dependency graphs. We leveraged the partitioning of system-calls in order to construct the GrD graph $D^*[\pi]$ that depicts the interconnection of specific groups of system-calls. Then, we developed the NP-similarity metric that, operating on GrD graphs, combines a set of similarity metrics in order to distinguish whether an unknown test sample is malicious or not based on a predefined threshold.
We evaluated our model's detection ability and compared its potentials against other results from several models either graph-based or not. The evaluation was performed on a set of $2630$ malware samples from $48$ malware families and $33$ benign commodity programs. The detection process exhibited a 91.3\% rate with 13.7\% false positives making it competing against other detection models.
Finally, an interesting perspective is the extension of our model for malware indexing, i.e., to classify a test sample in a malware family, if it has been detected as malware; we leave such an extension as a problem for further research.
\frenchspacing
|
\section{Introduction}\label{s12}
Let $Z^{(\alpha)}$ be a symmetric $\alpha$-stable process in $\Re^d$ with $\alpha\in (0,2)$, that is, a L\'evy process with
$$
\E e^{i(\xi, Z_t)}=e^{-t|\xi|^\alpha}, \quad \xi\in \Re^d.
$$
It is well known that for $Z^{(\alpha)}$, considered as a Markov process, its generator is defined on $C^2_\infty(\Re^d)$ by
$$
L^{(\alpha)}f(x)= \hbox{P.V.} \int_{\Re^d}\Big(f(x+u)-f(x)\Big){c^{(\alpha)}\over |u|^{d+\alpha}}du.
$$
Operator $L^{(\alpha)}$ is called also the \emph{fractional Laplacian}, and is denoted by $-(-\Delta)^{\alpha/2}$.
Given a scalar function $a:\Re^d\to (0, \infty)$ and a vector field $b:\Re^d\to \Re^d$,
consider the formal pseudo-differential operator
\begin{equation}\label{symbol}
L(x, D)f(x)=a(x) L^{(\alpha)}f(x)+\Big(b(x), \nabla f(x)\Big), \quad f\in C_\infty^2(\Re^d).
\end{equation}
This paper is addressed to the following problem: does an operator of type \eqref{symbol} correspond to a Markov process, and if so, what properties of the corresponding process one can deduce? Let us briefly summarize the results presented below.
We construct a strongly continuous semigroup $\{P_t, \, t\geq 0\}$ on $C_\infty(\Re^d)$, which corresponds to a Markov process $X$ on $\Re^d$, whose generator is a closure of the operator $(L(x,D), C_\infty^2(\Re^d))$. This semigroup will be obtained in the form
\be\label{semi}
P_tf(x)=\int_{\Re^d}p_t(x,y)f(y)\, dy, \quad f\in C_\infty(\Re^d),
\ee
which means that the Markov process $X$ admits the transition probability density $p_t(x,y)$. In addition, we prove that $p_t(x,y)$ is continuous w.r.t. $x$, that is, the Markov process $X$ is a strong Feller one. We also give two-sided bounds on the transition probability density $p_t(x,y)$, and clarify the probabilistic structure of the Markov process $X$. Namely, it will be shown that $X$ is the unique week solution to the following SDE driven by the $\alpha$-stable process $Z^{(\alpha)}$:
\be\label{SDE}
dX_t=b(X_t)\, dt+\sigma(X_{t-})\, dZ^{(\alpha)}_t.
\ee
Here and below we denote $\sigma(x):=a^{1/\alpha}(x)$.
The method we use is based on the \emph{parametrix} construction, which is a classical tool for constructing and estimating the fundamental solutions in the context of the Cauchy problem for a parabolic 2nd order PDE's, see \cite{Fr64}. Various extensions of the classical parametrix construction are available in the literature, which allow one to treat the Cauchy problem for pseudo-differential operators as well. In particular, there exists a variety of publications devoted to the analysis of properties of Markov processes with formal generators of the form (\ref{symbol}); we postpone their discussion to Section \ref{over} below. Here we just mention that in the results, available so far, it is required that either $b\equiv 0$ or $\alpha>1$, which corresponds heuristically to an assumption that the ``gradient part'' $(b(x), \nabla)$ should be dominated, in a sense, by the ``jump part'' $a(x)L^{(\alpha)}$, and therefore the ``jump part'' should present the ``main term'' in $L(x, D)$. We emphasize that our approach does not involve such a ``domination'' assumption, and we are able to treat formal generators of the form (\ref{symbol}) with $\alpha\leq 1$ and non-trivial ``gradient term'' $(b(x), \nabla)$.
The paper is organized as follows. In Section~\ref{s20} we formulate our main results; an outline of the method, an overview of the available results, and a relative discussion of our main results are also given therein. Section~\ref{s2} is devoted to the construction and estimation of the parametrix series, and to the continuity properties of $p_t(x,y)$. Similar estimates for the time-wise derivative of $p_t(x,y)$ are given in Section ~\ref{s3}. Section~\ref{s4} is devoted to the justification of the method, and contains in particular the proofs of the well-posedness of the martingale problem related to \eqref{symbol}, and of the uniqueness of the weak solution to \eqref{SDE}. The basic notation is collected in Appendix A. Appendices B and C are contain some auxiliary results, used in the proofs.
\section{The main results: outline, formulation, and discussion}\label{s20}
\subsection{Outline of the method}~\label{model}
In order to simplify the further exposition, let us briefly outline the parametrix construction our approach is based on.
We are looking for a fundamental solution to the Cauchy problem for a pseudo-differential operator
\be\label{L_ful}
\prt_t-L(x,D),
\ee
i.e. for such a function $p_t(x,y)$ that
\be\label{L_fund}
\Big(\prt_t-L(x, D)\Big)p_t(x,y)=0, \quad t>0, \quad x,y\in \Re^d,
\ee
\be\label{L_delta}p_t(x, \cdot)\to \delta_x, \quad t\to 0+, \quad x\in \Re^d.\ee
Consider \emph{some} approximation $p_t^0(x,y)$ to this function, and denote by $r_t(x,y)$ the residue w.r.t. this approximation; that is, write
\be\label{sol}
p_t(x,y)=p_t^0(x,y)+r_t(x,y).
\ee
Put
\be\label{Phi}
\Phi_t(x,y)=-\Big(\prt_t-L(x, D)\Big)p_t^0(x,y),\quad t>0, \quad x,y\in\Re^d.
\ee
Observe that since $p_t(x,y)$ is aimed to be the fundamental solution for the operator (\ref{L_ful}), one should have
$$
\Big(\prt_t-L(x, D)\Big)r_t(x,y)=\Phi_t(x,y).
$$
Resolving formally this equation in terms of the unknown fundamental solution $p_t(x,y)$, and then using (\ref{sol}), we get the following equation for $r_t(x,y)$:
$$
r_t(x,y)=(p\star \Phi)_t(x,y)=(p^0\star \Phi)_t(x,y)+(r\star \Phi)_t(x,y).
$$
The formal solution to this equation is given by the convolution
\be\label{r}
r=p^0\star \Psi,
\ee
where $\Psi$ is represented by the sum of convolution powers of $\Phi$:
\be\label{Psi}
\Psi_t(x,y)=\sum_{k\geq 1}\Phi^{\star k}_t(x,y).
\ee
On a formal level, this represents the required fundamental solution $p_t(x,y)$ in the form (\ref{sol}) with the residue given by (\ref{r}). In what follows, we make this formal representation meaningful by giving analytical bounds, which in particular provide that the series (\ref{Psi}) converge and the convolution (\ref{r}) is well defined.
The second principal part of our research is devoted to justification of the parametrix construction; that is, to the analysis of the semigroup properties of the kernel $p_t(x,y)$ defined by relations (\ref{sol}) -- (\ref{Psi}). We show that relation (\ref{semi}) indeed defines a strongly continuous, conservative, and non-negative semigroup on $C_\infty(\Re^d)$. Then we identify the corresponding Markov process $X$ in terms of the initial operator (\ref{symbol}). Finally, we show that the Markov process $X$ is the unique week solution to
equation \eqref{SDE}.
In fact, we will see below that the parametrix construction exposed above is very flexible: since we can vary the choice of the ``main term'' $p_t^0(x,y)$, we have a variety of possible methods rather than a fixed one.
\subsection{The main results}\label{main}
Our standing assumption on the intensity coefficient $a(x)$ is that it is \emph{bounded}, \emph{uniformly elliptic}, and \emph{H\"older continuous} with some index $\gamma>0$; that is, there exist $0<c_1<c_2$ and $C$ such that
\be\label{a_bdd_Hol}
c_1\leq |a(x)|\leq c_2, \quad |a(x)-a(y)|\leq C|x-y|^\gamma, \quad x,y\in \Re^d.
\ee
The drift coefficient $b(x)$ is assumed to be \emph{continuous and bounded}. In addition, we consider three following groups of assumptions.
{\it Assumption} \textbf{A}. $\alpha\in (1,2)$.
{\it Assumption} \textbf{B}. Function $b$ is H\"older continuous with the index $\gamma$, and $\alpha\in
((1+\gamma)^{-1}, 2)$.
{\it Assumption} \textbf{C}. Function $b$ is Lipschitz continuous; that is,
$$
|b(x)-b(y)|\leq C|x-y|, \quad x, y\in \Re^d.
$$
\begin{rem} Apparently, in the case of Assumption \textbf{B} it is required that both $a(x)$ and $b(x)$ are H\"older continuous, and then $\gamma$ denotes the least of their H\"older indices.
\end{rem}
\begin{rem}\label{r22} One can easily give examples showing that none of assumptions \textbf{A} -- \textbf{C} is contained in any other of them. These assumptions, in a sense, relate the regularity of $b$ to the stability index $\alpha$: when $b$ is assumed to be just bounded, one requires that $\alpha>1$ (\textbf{A}), while in order to be able to tackle arbitrary $\alpha\in (0,2)$ one requires $b$ to be Lipschitz continuous (\textbf{C}). Note that in the ``intermediate case'' \textbf{B}, when $\gamma\to 1$, the lower bound $1/(1+\gamma)$ for $\alpha$ tends to $1/2$, which differs from the lower bound $\alpha>0$ in the case \textbf{C}.
\end{rem}
When $b$ is Lipschitz continuous, the Cauchy problem for the ODE
\begin{equation}\label{flow1}
dv_t=b(v_t)\, dt,
\end{equation}
provides the \emph{flow of solutions} $\{\chi_t, t\in \Re\}$. Denote by $\{\theta_t=\chi_t^{-1}, t\in \Re\}$ the \emph{inverse flow}, which in fact solves the Cauchy problem for the ODE
\begin{equation}\label{flow2}
dv_t=-b(v_t)\, dt.
\end{equation}
\begin{thm}\label{t1} (On convergence of the parametrix series). Let one of three following cases hold true:
\begin{itemize}
\item[(a)] Assumption \textbf{A} holds and
\be\label{p_a}
p_t^0(x,y)={1\over t^{d/\alpha}a^{d/\alpha}(y)}g^{(\alpha)}\left({y-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right).
\ee
\item[(b)] Assumption \textbf{B} holds and
\be\label{p_b}
p_t^0(x,y)={1\over t^{d/\alpha}a^{d/\alpha}(y)}g^{(\alpha)}\left({y-tb(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right).
\ee
\item[(c)] Assumption \textbf{C} holds and
\be\label{p_c}
p_t^0(x,y)={1\over t^{d/\alpha}a^{d/\alpha}(y)}g^{(\alpha)}\left({\theta_t(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right).
\ee
\end{itemize}
Then the function $p_t(x,y)$ is well defined by relations (\ref{sol}) -- (\ref{Psi}), i.e.,
\begin{itemize}
\item the function (\ref{Phi}) and its convolution powers are well defined;
\item the series (\ref{Psi}) converges for every $t>0, x,y\in \Re^d$;
\item the convolution (\ref{r}) is well defined.
\end{itemize}
\end{thm}
In what follows, we denote $L=L(x, D)$, where $L(x,D)$ is the initial operator given by (\ref{symbol}).
\begin{thm}\label{t2} (On the properties of $p_t(x,y)$). In each of the cases (a) -- (c) of Theorem \ref{t1}, the function $p_t(x,y)$ defined by (\ref{sol}) -- (\ref{Psi}) has the following properties.
\begin{itemize}
\item[I.] Identity (\ref{semi}) defines a strongly continuous conservative semigroup on $C_\infty(\Re^d)$, which corresponds to a (strong) Feller Markov process $X$.
\item[II.] Process $X$ is a solution to
the martingale problem
\be\label{mart}
(L, C^2_\infty(\Re^d)).
\ee
\item[III.] The $C_\infty$-generator of the semigroup $\{P_t\}$ equals the $C_\infty$-closure of the operator $L$, defined on $\mathcal{D}(L)=C^2_\infty(\Re^d)$. Consequently, the martingale problem (\ref{mart}) is well posed, and the process $X$ is uniquely determined as its unique solution.
\end{itemize}
\end{thm}
Theorem \ref{t1} shows that the parametrix construction is feasible, and Theorem \ref{t2} justifies this construction in the sense of the martingale problem (\ref{mart}). The following theorem provides an additional information about the probabilistic structure of the corresponding Markov process $X$. Using the bounds for $p_t(x,y)$ from Theorem \ref{t3} below and general criteria (see, for example, \cite[Chapter~4, Theorem~2.7]{EK86}) we easily deduce that $X$ has a c\'adl\'ag modification. Denote by $\P_x$ the law of the Markov process $X$ with $X_0=x$ in the Skorokhod space $\DD([0, \infty),\Re^d)$ of c\'adl\'ag functions $[0, \infty)\to \Re^d.$
\begin{thm}\label{tweak} (On a weak solution to the SDE with a stable noise). In each of the cases (a) -- (c) of Theorem \ref{t1}, for any $x\in \Re^d$ the SDE (\ref{SDE}) with the initial condition
$X_0=x$ has a unique weak solution, and the law of this solution in $\DD([0, \infty),\Re^d)$ equals $\P_x$.
\end{thm}
The last two theorems in this section contain explicit estimates for the target density $p_t(x,y)$ and its derivative w.r.t. the time variable. These results serve as auxiliary ones for us in the proofs of Theorems \ref{t1} -- \ref{tweak}, but they also are of independent interest. Since in all three cases (a) -- (c) of Theorem \ref{t1} the ``main term'' $p_t^0(x,y)$ in (\ref{sol}) is given explicitly, it is practical to estimate the target density $p_t(x,y)$ in the terms of the bounds for the residue in (\ref{sol}). In the theorem below we will see that such bounds are given in the following form:
\be\label{r_bound}
|r_t(x,y)|\leq Ct^{\delta}H_t(x,y), \quad t\in (0, t_0], \quad x,y\in \Re^d.
\ee
The kernels $H_t(x,y)$ will be chosen in such a way that for every $T>0$
\be\label{H_bound}
C_{1,T}\leq \int_{\Re^d} H_t(x,y)\, dy\leq C_{2,T}, \quad t\in (0, T].
\ee
We call them ``hull kernels'' by the reason which will become clear from the proof below. Due to (\ref{H_bound}), the term $t^\delta$ in (\ref{r_bound}) is naturally interpreted as an accuracy rate (in the integral sense) of $p_t(x,y)$ with respect to the ``main term'' $p_t^0(x,y)$, i.e. it gives the ``size'' of the residue. Respectively, the ``hull kernel'' $H_t(x,y)$ controls the ``shape'' of the residue.
To shorten the formulae below, denote
$$
\alpha'=\alpha\vee 1.
$$
Clearly, in the case (a) of Theorem \ref{t1} we have $\alpha'=\alpha$.
\begin{thm}\label{t3} (On the bounds for the residue). In cases (a) -- (c) of Theorem \ref{t1}, the following bounds for $r_t(x,y)$ are available, respectively.
\begin{itemize}
\item[(a)] For any $\kappa\in(0,\gamma]$ one has (\ref{r_bound}) with
$$\delta=\left({\kappa\over \alpha}\right)\wedge \left(1-{1\over \alpha}\right),\quad
H_t(x,y)={1\over t^{d/\alpha}}G^{(\alpha-\kappa)}\left({y-x\over t^{1/\alpha}}\right).
$$
\item[(b)] For any $\kappa\in(0,\gamma]\cap (0, \alpha)$ one has (\ref{r_bound}) with $$\delta=\left({\kappa\over \alpha'}\right)\wedge \left(1-{1\over \alpha}+{\gamma\over \alpha'}\right),\quad
H_t(x,y)={1\over t^{d/\alpha}}G^{(\alpha-\kappa)}\left({y-tb(y)-x\over t^{1/\alpha}}\right).
$$
\item[(c)] For any $\kappa\in(0,\gamma]\cap (0, \alpha)$ one has (\ref{r_bound}) with
$$
\delta={\kappa\over \alpha'},\quad H_t(x,y)={1\over t^{d/\alpha}}G^{(\alpha-\kappa)}\left({\theta_t(y)-x\over t^{1/\alpha}}\right).
$$
\end{itemize}
\end{thm}
\begin{thm}\label{t4} (On the derivative $\prt_tp_t(x,y) $). In each of the cases (a) -- (c) of Theorem \ref{t1}, the following statements fold true.
\begin{enumerate}
\item There exists a set $\Upsilon\subset (0, \infty)\times \Re^d$ of zero Lebesgue measure such that the function $p_t(x,y)$ defined by (\ref{sol}) -- (\ref{Psi}) has a derivative
$$
\prt_tp_t(x,y), \quad x\in \Re^d,\quad (t,y)\in \Upsilon,
$$ which for every fixed $(t,y)\in \Upsilon$ is continuous w.r.t. $x$.
\item The derivative $\prt_tp_t(x,y)$ possesses the bound
$$
|\prt_tp_t(x,y)|\leq C\Big(t^{-1}\vee t^{-1/\alpha}\Big)H_t(x,y),\quad x\in \Re^d,\quad (t,y)\in \Upsilon,
$$
where $H_t(x,y)$ for cases (a) -- (c) are given, respectively, in Theorem \ref{t3}.
\item For any $f\in C_\infty(\rd)$ the function
$$
(0,\infty)\ni t\mapsto P_tf\in C_\infty(\rd)
$$
is continuously differentiable, and its derivative is given by
$$
(\prt_tP_tf)(x)=\int_\rd \prt_tp_t(x,y) f(y)dy.
$$
\end{enumerate}
\end{thm}
\subsection{Overview and discussion}\label{over}
For the description and the background of the parametrix construction of the fundamental solution to a Cauchy problem for a parabolic 2nd order PDE's, we refer to the monograph of Friedman \cite{Fr64}; see also the original paper by E.Levi \cite{Le07}. This construction was extended to equations with pseudo-differential operators in \cite{Dr77}, \cite{ED81}, \cite{Ko89}, and \cite{Ko00}; see also the reference list and an extensive overview in the monograph \cite{EIK04}. In \cite{Dr77}, \cite{ED81}, and \cite{Ko89} the ``main term'' in the pseudo-differential operator is assumed to have the form $a(x)L^{(\alpha)}$ (in our notation) with $\alpha>1$; in \cite{Ko00} the stability index $\alpha$ is allowed to be $\leq 1$, but in this case the gradient term should not be involved into the equation. The list of subsequent and related publications is large, and we can not discuss it in details here. Let us only mention two recent pre-prints \cite{CZ13}, where two-sided estimates, more precise than those from \cite{Ko89}, are obtained, and \cite{BK14}, where a probabilistic interpretation of the parametrix construction and its application to the Monte-Carlo simulation is developed.
In all the references listed above it is required that either the stability index $\alpha$ is $>1$, or the gradient term is not involved in the equation. This is the common assumption in all the references available for us in this direction, with the one important exception given by the recent paper \cite{DF13} (see also \cite{FP10}). In \cite{DF13}, for a L\'evy driven SDE with $\alpha$-stable like noise, a question of \emph{existence} of a distribution density is studied my means of a different method, based on an approximation of the initial SDE and a discrete integration by part formula. Such an approach is applicable to SDE's with the stability index of the noise $\alpha<1$ and non-trivial drift, but it does not give proper tools neither to obtain explicit estimates for this density, nor even to prove the existence and uniqueness of the solution to an initial SDE. Hence the scopes of our approach, based on the parametrix construction, differ from those of \cite{DF13} substantially.
Our version of the parametrix construction contains a substantial novelty, which makes it eligible in the case $\alpha\leq 1$ with non-trivial gradient term. To explain this modified construction in the most transparent form, we took the ``jump component'' in a comparatively simple form $a(x)Z^{(\alpha)}$. Clearly, such a choice is not unique and one can think about considering e.g. $\alpha$-stable symbols with state dependent spectral measures; cf. \cite{Ko00}. This, however, would lead to additional cumbersome but inessential technicalities, and here we avoid to do in such a way.
Observe that there is an interplay between the value of $\alpha$ and the regularity properties required for $b$; see Remark \ref{r22}. Heuristically, this means that increasing of $\alpha$ should relax the assumptions on $b$; this well corresponds to the effect, observed first in \cite{Po94}, \cite{PP95}, that the parametrix construction is still feasible for (possibly unbounded) $b\in L_{p}(\Re^d)$, $p>d/(\alpha-1)$. In \cite{BJ07} this effect was rediscovered in a stronger form: it is required therein that $b$ belongs to the Kato class $\mathbb{K}_{d,\alpha-1}$. In \cite{KS14}, this result is extended even further, with $b$ being allowed to be a generalized function equal to the derivative of a measure from the Kato class $\mathbb{K}_{d,\alpha-1}$.
In general, there is a substantial distance between just constructing a ``candidate for the fundamental solution'' (i.e. proving that relations (\ref{sol}) -- (\ref{Psi}) are feasible) on one hand, and justifying this construction (i.e. proving that $p_t(x,y)$ generates a strong Feller semigroup and relating this semigroup with the initial symbol) on the other hand. The first way of such a justification, proposed in \cite{Ko89}, is an extension of the approach used in \cite{Fr64}. In \cite{Fr64}, by a proper analysis of the parametrix series in the diffusion setting, it is proved that $p_t(x,y)$ is twice continuously differentiable w.r.t. $x$ and satisfies (\ref{L_fund}) in the classical sense. In the $\alpha$-stable case the natural upper bound $\prt_{xx}^2p_t^0(x,y)\leq Ct^{-2/\alpha}$ is strongly singular for small $t$, and therefore it is difficult to prove using the parametrix construction that $\prt_{xx}^2p_t(x,y)$ is well defined. Instead of that, in \cite{Ko89} a naturally extended domain for $L^{(\alpha)}$ is introduced in the terms of ``hyper-singular integrals'', and it is proved that $p_t(x,y)$ satisfies (\ref{L_fund}) in the corresponding ``extended'' sense. Once (\ref{L_fund}) is proved, the required properties of the Markov process associated with $p_t(x,y)$ follow from the positive maximum principle in a rather standard way. Another way to justify the parametrix construction, proposed in \cite{Ko00}, is to guarantee the required smoothness of $p_t(x,y)$ using the integration-by-parts procedure, but this approach seems to be only partially relevant; see Remark~\ref{r_Kol} below.
Partially, one can resolve the justification problem by using an approximative procedure (e.g. \cite{Po94}, \cite{PP95}) or by analysing the perturbation of the resolvent kernels (cf. \cite{BJ07}). However, the most difficult point here is to relate \emph{uniquely} the initial symbol and the Markov process associated with $p_t(x,y)$. This problem was solved recently in \cite{KS14}, in the framework of a singular gradient perturbation of an $\alpha$-stable generator, in terms of the weak solutions to a corresponding SDE (see also \cite{CW13}, where the martingale problem approach was used instead). The technique therein is closely related to the one introduced (in the diffusive setting) in \cite{BC03}, and apparently strongly relies on the structural assumption that the resolvent which corresponds to $p_t(x,y)$ is a perturbation of the resolvent for an $\alpha$-stable process.
We propose a new method of justification, based on the notion of the \emph{approximative fundamental solution} to the Cauchy problem for (\ref{L_ful}); see Section \ref{s4} and especially the discussion at the beginning of Section \ref{s52}. This method strongly exploits the properties of the parametrix series, which gives a possibility to make this method free from any additional structural assumptions. Therefore we expect that this method will be well applicable for other systems, where the parametrix construction is feasible; this is the subject of our further research.
Let us discuss briefly another large group of results, focused on the construction of a \emph{semigroup} for a Markov process with a given symbol rather than of a transition probability density $p_t(x,y)$ for it. An approach based on properties of the symbol of the operator and on the Hilbert space methods, is developed in the works of Jacob \cite{Ja94}, see also the monograph \cite{Ja96} for the more details. It allows to show the existence in $C_\infty(\rd)$ of the closed extension of a given pseudo-differential operator, and that this extension is a generator of a Feller semigroup. This approach was further developed by \cite{Ho98a}, \cite{Ho98b}, in \cite{Bo05}, \cite{Bo08}, and relies on the symbolic calculus approach for the construction of the parametrix (cf. \cite{Ku81}, also the original papers \cite{Ts74}, \cite{Iw77}); see also \cite{Ja01}--\cite{Ja05} for the detailed treatment.
Finally, we mention the group of results which are devoted to the well-posedness of the martingale problem for an integro-differential operator of certain type. For different types of perturbations of an $\alpha$-stable generator this problem was treated in the works by \cite{Ko84a}, \cite{Ko84b}, \cite{Ts70}, \cite{Ts73}, \cite{MP92a}--\cite{MP12a}, \cite{Ba88},
see also \cite{Ho94}, \cite{Ho95} for yet another approach for rather wide class of operators.
\section{Constructing and estimating the parametrix series. Continuity properties of $p_t(x,y)$. Proofs of Theorem~\ref{t1} and Theorem~\ref{t3}} \label{s2}
This section is mainly aimed at the construction of analytical bounds on the function $\Phi$ defined by (\ref{Phi}), the convolution powers of $\Phi$, and their convolutions with the ``main part'' $p_t^0(x,y)$. When these bounds are obtained, we easily deduce the proofs of Theorem~\ref{t1} and Theorem~\ref{t3}. In addition, the basic continuity properties of $p_t(x,y)$ and related integrals follow then as an easy consequence of these bounds.
\subsection{Generic calculations}\label{s21}
In all three cases (a) -- (c) of Theorem \ref{t1} we will use the same approach, which allows us to control unanimously the whole sequence of convolution powers $\Phi^{\star k}, k\geq 1$. For the sake of reader's convenience, let us explain this approach separately.
\emph{Assume} that $p_t^0(x,y)$ is chosen in such a way that the function $\Phi$ is well defined by (\ref{Phi}) and satisfies the following analogue of (\ref{r_bound}):
\be\label{Phi_bound}
|\Phi_t(x,y)|\leq C_{\Phi, T} t^{-1+\delta}H_t(x,y), \quad t\in (0,T], \quad x,y\in \Re^d,
\ee
with some positive $\delta$, and some kernel $H_t(x,y)$. In general, the terminal time point $T$ may vary, and the positive constant $C_{\Phi, T}$ may depend on $T$. \emph{Assume also} that the kernel $H_t(x,y)$ enjoys the following \emph{sub-convolution property}.
\begin{dfn} A kernel $\{H_{t}(x,y), t>0, x,y\in \Re^d\}$ has a sub-convolution property, if for every $T>0$ there exists a constant $C_{H, T}$ such that
$$
(H_{t-s}* H_s)(x,y)\leq C_{H, T} H_{t}(x,y), \quad t\in (0, T], \quad s\in (0, t), \quad x,y\in \Re^d.
$$
\end{dfn}
\begin{lem}\label{l1} Let $\Phi$ satisfy (\ref{Phi_bound}) with a kernel $H_t(x,y)$ which admits the sub-convolution property. Then for every $k\geq 1$
\be\label{Phi_k_bound}
|\Phi^{\star k}_t(x,y)|\leq {(C_{\Phi, T} \Gamma(\delta))^k(C_{H, T})^{k-1} \over \Gamma(k\delta)} t^{-1+k\delta}H_t(x,y), \quad t\in (0, T], \quad x,y\in \Re^d,
\ee
where $\Gamma(\cdot)$ denotes the Gamma-function and $C_{H, T}$ denotes the constant from the sub-convolution property for $H_t(x,y)$.
\end{lem}
\begin{proof} When $k=1$, the bound (\ref{Phi_k_bound}) coincides with the assumption (\ref{Phi_bound}). Let us use the induction: assuming that (\ref{Phi_k_bound}) holds true for $k$, consider the same bound for $k+1$. Using first (\ref{Phi_k_bound}) and (\ref{Phi_bound}), and then the sub-convolution property for $H_t(x,y)$, we get
\be\label{21}\ba
|\Phi_t^{\star (k+1)}(x,y)|&=\left|\int_0^t(\Phi^{\star k}_{t-s}*\Phi_s)(x,y)\, ds\right|
\\&\leq {(C_{\Phi, T} \Gamma(\delta))^k(C_{H, T})^{k-1} \over \Gamma(k\delta)} C_{\Phi, T}\int_0^t(t-s)^{-1+k\delta}s^{-1+\delta}(H_{t-s}*H_s)(x,y)\, ds
\\& \leq {(C_{\Phi, T} \Gamma(\delta))^k(C_{H, T})^{k-1} \over \Gamma(k\delta)} C_{\Phi, T} C_{H, T} H_{t}(x,y)\int_0^t(t-s)^{-1+k\delta}s^{-1+\delta}\, ds
\\&={(C_{\Phi, T} \Gamma(\delta))^k(C_{H, T})^{k-1} \over \Gamma(k\delta)} C_{\Phi, T} C_{H, T} {\Gamma(k\delta)\Gamma(\delta)\over \Gamma((k+1)\delta)} H_{t}(x,y)\\&={(C_{\Phi, T} \Gamma(\delta))^{k+1}(C_{H, T})^{k} \over\Gamma((k+1)\delta)} H_{t}(x,y),
\ea
\ee
which is just (\ref{Phi_k_bound}) with $k+1$ instead of $k$.
\end{proof}
\begin{cor}\label{c1} The function $\Psi_t(x,y)$ in \eqref{Psi} is well defined for $t\in (0, T]$, $x,y\in \Re^d$, and satisfies the inequality, similar to (\ref{Phi_bound}):
\be\label{Psi_bound}
|\Psi_t(x,y)|\leq C_{\Psi, T} t^{-1+\delta}H_t(x,y), \quad t\in (0,T], \quad x,y\in \Re^d,
\ee
where the constant $C_{\Psi, T}$ can be expressed explicitly in terms of $C_{\Phi,T}, C_{H,T}$ and $T$.
\end{cor}
Finally, \emph{assume} that the $*$-convolution of $p^0$ and $H$ is dominated by $H$; that is, there exists $\tilde C_{H,T}$ such that
\be\label{domin}
(p^0_{t-s}* H_s)(x,y)\leq \tilde C_{H, T} H_{t}(x,y), \quad t\in (0, T], \quad s\in (0, t), \quad x,y\in \Re^d.
\ee
Then using the calculation similar to those from the proof of Lemma~\ref{l1}, we get from Corollary~\ref{c1} that $r=p^0\star \Phi$ is well defined and satisfies
\be\label{r_bound_T}
|r_t(x,y)|\leq C_Tt^{\delta}H_t(x,y), \quad t\in (0, T], \quad x,y\in \Re^d,
\ee
where the constant $C_{T}$ can be expressed explicitly through $C_{\Phi,T}, C_{H,T}, \tilde C_{H,T}$ and $T$.
To summarize, in order to prove Theorem \ref{t1} and Theorem \ref{t3}, it is sufficient to prove that in each of the cases (a) -- (c),
\begin{itemize}
\item[(i)] the corresponding function $\Phi$ (which of course depends on the choice of $p^0$) satisfies (\ref{Phi_bound});
\item[(ii)] the corresponding kernel $H$ has the sub-convolution property;
\item[(iii)] the bound (\ref{domin}) is available.
\end{itemize}
Note that (\ref{domin}) would easily follow from the sub-convolution property for $H$ if we had
\be\label{domin_p}
p^0_t(x,y)\leq \hat C_{H, T} H_{t}(x,y), \quad t\in (0, T], \quad x,y\in \Re^d.
\ee
\subsection{Case (a)}\label{s22}
Fix $\xi\in \Re^d$, and denote
$$
L^\xi=a(\xi)L^{(\alpha)}.
$$
The fundamental solution $q^\xi_t(x,y)$ for the operator $(\prt_t-L^\xi)$ is just the transition probability density of the process $Z^{(\alpha)}$ with the time, re-scaled by $a(\xi)$:
$$
q_t^\xi(x,y)={1\over t^{d/\alpha}a^{d/\alpha}(\xi)}g^{(\alpha)}\left({y-x\over t^{1/\alpha}a^{1/\alpha}(\xi)}\right).
$$
Hence, in the case (a) our choice of the ``main part'' in the representation (\ref{sol}) is the following:
$$
p_t^0(x,y)=q_t^\xi(x,y)|_{\xi=y}.
$$
This choice follows the main line of the classical parametrix construction for paradolic PDE's, see \cite{Fr64}: one should take the ``principal part'' of $L=L(x, D)$, then ``freeze'' the coefficients in it at the ``arrival point'' $y$, and use the fundamental solution of this ``frozen'' equation as the ``main part'' in representation (\ref{sol}) for the unknown fundamental solution. This strategy has been applied successfully for pseudo-differential operators in various settings, see also e.g. \cite{Ko89} and \cite{Ko00}. Therefore, calculations in this subsection are not genuinely new; nevertheless, in order to make the presentation self-contained, we give them explicitly. Now, when we already have the generic calculations from the previous subsection, this can be done in a short and transparent form.
Let $p_t^0(x,y)$ be given by (\ref{p_a}). Then for any fixed $x,y\in \Re^d$ it is differentiable w.r.t. $t$ on $(0, \infty)$, and for any fixed $t\in (0, \infty)$, $y\in \Re^d$ it belongs to $C^2_\infty(\Re^d)$ as a function of $x$.
Since both $\nabla$ and $L^{(\alpha)}$ are well defined on the class $C^\infty_b(\Re^d)$, the function $\Phi_t(x,y)$ is well defined by (\ref{Phi}).
Let us show that $\Phi_t(x,y)$ admits the upper bound of the form (\ref{Phi_bound}). Since $q^\xi$ is the fundamental solution for $\prt_t-L^\xi$, one has
\be\label{Phi_a}\ba
\Phi_t(x,y)&=-\Big(\prt_t-L^y_x\Big)p_t^0(x,y)+(L_x-L^y_x)p_t^0(x,y)\\
&= \Big(a(x)-a(y)\Big) L^{(\alpha)}_xp_t^0(x,y)+\Big(b(x), \nabla_x p_t^0(x,y)\Big)\\
&= \Big(a(x)-a(y)\Big) {1\over t^{d/\alpha+1}a^{d/\alpha+1}(y)}(L^{(\alpha)} g^{(\alpha)})\left({y-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\\
&\hspace*{0.5cm}-
{1\over t^{(d+1)/\alpha}a^{(d+1)/\alpha}(y)}\left(b(x), (\nabla g^{(\alpha)})\left({y-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\right)=:\Phi_t^1(x,y)+\Phi_t^2(x,y).
\ea
\ee
Here and below the subscript $x$ means that an operator is applied w.r.t. the variable $x$.
Recall that $a, b$ are bounded, and $a$ is bounded away from zero. Hence by (\ref{g_a_der}) and (\ref{G1}) we have
$$
|\Phi_t^2(x,y)|\leq Ct^{-(d+1)/\alpha}G^{(\alpha+1)}\left({y-x\over t^{1/\alpha}}\right), \quad t>0, \quad x,y\in \Re^d.
$$
Fix $\kappa\in(0,\gamma]$ and note that $\kappa<\alpha$, because $\gamma\leq 1$ and in case (a) we have $\alpha>1$. Since $a$ is bounded and $\gamma$-H\"older continuous, we have
\be\label{a_hol_in}
|a(x)-a(y)|\leq C|x-y|^\kappa, \quad x,y\in \rd.
\ee
Then using (\ref{g_a_frac}) and (\ref{G3}) we get
$$
|\Phi_t^1(x,y)|\leq Ct^{-d/\alpha-1+\kappa/\alpha}G^{(\alpha-\kappa)}\left({y-x\over t^{1/\alpha}}\right).
$$
Combining this estimate with the bound for $\Phi^2_t(x,y)$ found above and taking into account (\ref{G2}), we obtain (\ref{Phi_bound})
with $$\delta=\left({\kappa\over \alpha}\right)\wedge \left(1-{1\over \alpha}\right),\quad
H_t(x,y)={1\over t^{d/\alpha}}G^{(\alpha-\kappa)}\left({y-x\over t^{1/\alpha}}\right).
$$
The above kernel $H_t(x,y)$ has the sub-convolution property, see Proposition~\ref{h-conv} below. In addition, property (\ref{domin_p}) now clearly holds true, because $G^{(\alpha-\kappa)}\geq G^{(\alpha)}$ and
$$
p_t^0(x,y)\asymp {1\over t^{d/\alpha}}G^{(\alpha)}\left({y-x\over t^{1/\alpha}}\right).
$$
Hence, all three items (i)-- (iii) from Section \ref{s21} are verified, which completes the proof of Theorem \ref{t1} and Theorem \ref{t3} in the case (a).\qed
\subsection{Case (b)}\label{s23} Again, we fix $\xi\in \Re^d$, and define
$$
L^\xi=a(\xi)L^{(\alpha)}+\Big(b(\xi), \nabla \Big).
$$
Now the fundamental solution $q^\xi_t(x,y)$ to the operator $(\prt_t- L^\xi)$ equals
$$
q_t^\xi(x,y)={1\over t^{d/\alpha}a^{d/\alpha}(\xi)}g^{(\alpha)}\left({y-x-tb(\xi)\over t^{1/\alpha}a^{1/\alpha}(\xi)}\right);
$$
that is, the transition probability density of the process $Z^{(\alpha)}$ with the time, re-scaled by $a(\xi)$ and with the constant drift $tb(\xi)$ added.
Hence, in the case (b), we again have the representation
$$
p_t^0(x,y)= q_t^\xi(x,y)|_{\xi=y},
$$
but with another fundamental solution $q^\xi$: we take the whole operator $L$ as its own ``main part'', and ``freeze'' its coefficients to get the required fundamental solution.
Since $q^\xi$ is the fundamental solution for $\prt_t-L^\xi$, we have in the same way as in (\ref{Phi_a})
\be\label{Phi_b}\ba
\Phi_t(x,y)&=-\Big(\prt_t- L^y_x\Big)p_t^0(x,y)+(L_x-L^y_x)p_t^0(x,y)\\
&= \Big(a(x)-a(y)\Big) L^{(\alpha)}_xp_t^0(x,y)+\Big(b(x)-b(y), \nabla_xp_t^0(x,y)\Big)\\
&= \Big(a(x)-a(y)\Big) {1\over t^{d/\alpha+1}a^{d/\alpha+1}(y)}(L^{(\alpha)} g^{(\alpha)})\left({y-tb(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\\&\hspace*{0.5cm}+
{1\over t^{(d+1)/\alpha}a^{(d+1)/\alpha}(y)}\left(b(y)-b(x), (\nabla g^{(\alpha)})\left({y-tb(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\right)\\
&=:\Phi_t^1(x,y)+\Phi_t^2(x,y).
\ea
\ee
Let us proceed with estimates for $\Phi_t(x,y)$ in the way similar to those from the previous subsection. Take some $\kappa\in(0,\gamma]$. Again, we have (\ref{a_hol_in}), but we need to modify this bound because in the argument of $L^{(\alpha)}g^{(\alpha)}$ we have the term $y-tb(y)-x$ instead of $y-x$. Recall that $\kappa\leq \gamma\leq 1$. Then using the elementary inequality $(u+v)^\kappa\leq u^\kappa+v^\kappa$, $u,v>0$, we get
\be\label{a_hol_in_b}
|a(x)-a(y)|\leq C(|y-tb(y)-x|+t|b(y)|)^\kappa\leq C|y-tb(y)-x|^\kappa+Ct^\kappa.
\ee
Since $b(\cdot)$ is bounded, using (\ref{g_a_frac}) and (\ref{G3}) we arrive at
\be\label{Phi_1_bound-b}
|\Phi_t^1(x,y)|\leq Ct^{-d/\alpha-1+\kappa/\alpha}G^{(\alpha-\kappa)}\left({y-tb(y)-x\over t^{1/\alpha}}\right)+Ct^{-d/\alpha-1+\kappa}G^{(\alpha)}\left({y-tb(y)-x\over t^{1/\alpha}}\right).
\ee
Similar argument can be applied to $\Phi^2_t(x,y)$, with possibly different $\kappa'\in(0,\gamma]$. Namely, inequality (\ref{a_hol_in_b}) still holds true if we take $b(x)$ instead of $a(x)$, and $\nabla g^{(\alpha)}$ satisfies (\ref{g_a_der}). Thus,
\be\label{Phi_2_bound-b}
|\Phi_t^2(x,y)|\leq Ct^{-d/\alpha-1/\alpha+\kappa'/\alpha}G^{(\alpha-\kappa'+1)}\left({y-tb(y)-x\over t^{1/\alpha}}\right)+Ct^{-d/\alpha-1/\alpha+\kappa'}G^{(\alpha+1)}\left({y-tb(y)-x\over t^{1/\alpha}}\right).
\ee
Now we come to a slightly cumbersome procedure of choosing the parameters $\kappa, \kappa'$ in order to balance the final estimate for $\Phi$. The general feature here is that making, say, the parameter $\kappa$ larger, we improve the respective power of $t$, but make the ``tails'' of the function $G^{(\alpha-\kappa)}$ ``heavier''. Note that for any choice of $\kappa'\leq \gamma$ functions $G^{(\alpha+1)}$ and $G^{(\alpha-\kappa'+1)}$ are dominated by
$G^{(\alpha)}$, which in turn is dominated by $G^{(\alpha-\kappa)}$. Therefore, it is reasonable to choose $\kappa'$ to be the maximal possible, i.e. $\kappa'=\gamma$. This leads to the estimate
$$
|\Phi_t(x,y)|\leq Ct^{-d/\alpha}G^{(\alpha-\kappa)}\left({y-tb(y)-x\over t^{1/\alpha}}\right)\Big[t^{-1+\kappa/\alpha}+t^{-1+\kappa}+t^{-1/\alpha+\gamma/\alpha}+t^{-1/\alpha+\gamma}\Big].
$$
Note that one should take $\kappa<\alpha$ in order $G^{(\alpha-\kappa)}$ to be integrable, and thus the whole estimate to make sense.
Hence, we have the bound (\ref{Phi_bound}) with
$$\delta=\kappa\wedge \left({\kappa\over \alpha}\right)\wedge \left(1-{1\over \alpha}+\gamma\right)\wedge \left(1-{1\over \alpha}+{\gamma\over \alpha}\right),\quad
H_t(x,y)={1\over t^{d/\alpha}}G^{(\alpha-\kappa)}\left({y-tb(y)-x\over t^{1/\alpha}}\right).
$$
Note that $\delta>0$ because $\alpha>(1+\gamma)^{-1}$.
Similarly to what we obtained in the previous subsection, the kernel $H_t(x,y)$ given above has the sub-convolution property, see Proposition~\ref{h-conv} below. The property (\ref{domin_p}) also holds true because
$$
p_t^0(x,y)\asymp {1\over t^{d/\alpha}}G^{(\alpha)}\left({y-tb(y)-x\over t^{1/\alpha}}\right).
$$
Thus, items (i)-- (iii) from Section \ref{s21} are verified, which completes the proof of Theorem \ref{t1} and Theorem \ref{t3} in case (b).\qed
\subsection{Case (c)}\label{s24} In case (c) we can not treat $p_t^0(x,y)$ as a fundamental solution for some approximation to $\prt_t-L$, hence we just calculate $\Phi$ explicitly. Put
$$
g_t^{(\alpha)} (x,y)=t^{-d/\alpha}g^{(\alpha)}\left(t^{-1/\alpha}(y-x)\right).
$$
Since $g_t^{(\alpha)}(x,y)$ is the transition probability density of the $\alpha$-stable process $Z^{(\alpha)}$, we have
\be\label{41}
(\prt_t-L^{(\alpha)}_x)g_t^{(\alpha)}(x,y)=0, \quad t>0, \quad x,y\in \Re^d.
\ee
Then the function $p_t^0(x,y)$ can be written as
$$
p_t^0(x,y)=g_{a(y)t}^{(\alpha)}(x,\theta_t(y)),
$$
and thus by (\ref{41}) we have
$$
\prt_tp_t^0(x,y)=a(y)L^{(\alpha)}_x g_{a(y)t}^{(\alpha)}(x,\theta_t(y))+\Big(\prt_t\theta_t(y), \nabla_x g_{a(y)t}^{(\alpha)}(x,\theta_t(y))\Big).
$$
On the other hand,
$$
L_xp_t^0(x,y)=a(x)L^{(\alpha)}_x g_{a(y)t}^{(\alpha)}(x,\theta_t(y))+\Big(b(x), \nabla_x g_{a(y)t}^{(\alpha)}(x,\theta_t(y))\Big).
$$
Recall that by definition of the flow $\{\theta_t(y)\}$ one has $\prt_t\theta_t(y)=-b(\theta_t(y))$. Using that $\nabla_x$ and $L^{(\alpha)}$ are homogeneous operators of the orders $1$ and $\alpha$, respectively, we can finally write
\be\label{Phi_c}\ba
\Phi_t(x,y)&=\Big(a(x)-a(y)\Big) {1\over t^{d/\alpha+1}a^{d/\alpha+1}(y)}(L^{(\alpha)} g^{(\alpha)})\left({\theta_t(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\\&+
{1\over t^{(d+1)/\alpha}a^{(d+1)/\alpha}(y)}\left(b(\theta_t(y))-b(x), (\nabla g^{(\alpha)})\left({\theta_t(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\right)=:\Phi_t^1(x,y)+\Phi_t^2(x,y).
\ea
\ee
Similarly to (\ref{a_hol_in_b}), we have
\be\label{a_hol_in_c}
|a(x)-a(y)|\leq C|\theta_t(y)-x|^\kappa+Ct^\kappa,
\ee
where we used that $|\theta_t(y)-y|\leq ct$ because $b$ is bounded. Then for $\kappa\in (0, \gamma]\cap(0,\alpha)$
$$
|\Phi_t^1(x,y)|\leq Ct^{-d/\alpha-1+\kappa/\alpha}G^{(\alpha-\kappa)}\left({\theta_t(y)-x\over t^{1/\alpha}}\right)+Ct^{-d/\alpha-1+\kappa}G^{(\alpha)}\left({\theta_t(y)-x\over t^{1/\alpha}}\right).
$$
Since $b(x)$ satisfies the Lipschitz condition, we have
$$
|\Phi_t^2(x,y)|\leq C t^{-d/\alpha}\left|{\theta_t(y)-x\over t^{1/\alpha}}\right|
\left|(\nabla g^{(\alpha)})\left({\theta_t(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\right|.
$$
Note that now the same expression $\theta_t(y)-x$ stands both in the argument of $(\nabla g^{(\alpha)})$ and in the multiplier term, and hence an additional term like $Ct^\kappa$ in (\ref{a_hol_in_c}) would not appear. This is the advantage which comes from the particular choice of the argument in $p_t^0(x,y)$ ``in the precise form'' $\theta_t(y)-x$. This advantage appears to be substantial, since now the upper bound for $\Phi_t^2(x,y)$ simplifies a lot, when compared with (\ref{Phi_2_bound-b}). Namely, using (\ref{g_a_der}) and (\ref{G3}), we get
$$
|\Phi_t^2(x,y)|\leq C t^{-d/\alpha}G^{(\alpha)}\left({\theta_t(y)-x\over t^{1/\alpha}}\right).
$$
Hence, we have the bound (\ref{Phi_bound}) with
$$\delta=\kappa\wedge \left({\kappa\over \alpha}\right),\quad
H_t(x,y)={1\over t^{d/\alpha}}G^{(\alpha-\kappa)}\left({\theta_t(y)-x\over t^{1/\alpha}}\right).
$$
As in previous subsections, the kernel $H_t(x,y)$ given above has the sub-convolution property, see Proposition~\ref{h-conv} below. Estimate (\ref{domin_p}) also holds true, since
$$
p_t^0(x,y)\asymp {1\over t^{d/\alpha}}G^{(\alpha)}\left({\theta_t(y)-x\over t^{1/\alpha}}\right).
$$
Thus, we verified statements (i)-- (iii) from Section \ref{s21}, which completes the proof of Theorem \ref{t1} and Theorem \ref{t3}. \qed
\subsection{Continuity properties of $p_t(x,y)$}\label{s25}
From the above construction and estimates, we easily deduce the following lemma.
\begin{lem} In each of the cases (a) -- (c) of Theorem \ref{t1}, the following properties hold true.
\begin{enumerate}\label{cont}
\item The functions $p_t(x,y)$ and $\Psi_t(x,y)$ are continuous w.r.t. $(t,x,y)\in (0,\infty)\times\Re^d\times\Re^d$.
\item For any $f\in C_\infty(\rd)$ the function $P_tf(x)= \int_\rd p_t(x,y) f(y)dy$ belongs to $C_\infty(\rd)$.
\item For any $f\in C_\infty(\rd)$
$$
P_tf\to f, \quad t\to 0+
$$
in $C_\infty$.
\end{enumerate}
\end{lem}
\begin{proof} Since the proof is easy, we just sketch the argument. It can be verified explicitly that in each of the cases (a) -- (c) of Theorem \ref{t1} the function $p_t^0(x,y)$ and the corresponding $\Phi_t(x,y)$, calculated explicitly in Sections \ref{s22} - \ref{s24}, are continuous w.r.t. $(t,x,y)\in (0,\infty)\times\Re^d\times\Re^d$. The integrals which define convolution powers of $\Phi_t(x,y)$ converge uniformly on every compact set in $(0,\infty)\times\Re^d\times\Re^d$, hence each of these convolution powers is continuous by the dominated convergence theorem. Similarly, by (\ref{Phi_k_bound}) the series which defines $\Psi$ converges uniformly on every compact set in $(0,\infty)\times\Re^d\times\Re^d$, and since we just have proved that each summand is continuous, the whole sum is also continuous. Finally, by the same arguments the convolution (\ref{r}) is again continuous. This gives property 1.
The proof of continuity of $P_tf$ can be made within the same lines. To prove that $P_tf(x)$ vanishes as $|x|\to \infty$, it is sufficient to note that for each of the kernels $H_t(x,y)$ constructed in Sections~\ref{s22}--\ref{s24} one has for every $T>0$
\begin{equation}\label{H1}
C_{1,T}\leq \int_{\Re^d} H_t(x,y)\, dy\leq C_{2,T}, \quad t\in (0, T],
\end{equation}
(see Proposition~\ref{H-bound} below), and for every $t>0$
$$
\sup_x\int_{y:|x-y|>R}H_t(x,y)\, dy\to 0, \quad R\to\infty.
$$
This gives property 2.
By Theorem~\ref{t3} and Proposition~\ref{H-bound} one has
$$
\sup_x \int_{\Re^d}|r_t(x,y)|\, dy\to 0, \quad t\to 0+.
$$
Since one can verify explicitly that
\be\label{delta}
\sup_x\left|\int_{\Re^d}p_t^0(x,y) f(y)\, dy-f(x)\right|\to 0, \quad t\to 0+,
\ee
this gives property 3.
\end{proof}
\section{Time-wise derivative of $p_t(x,y)$. Proof of Theorem \ref{t4}}\label{s3}
\subsection{Outline}
The main goal of this section is to prove Theorem~\ref{t4}; that is, we establish existence of the time-wise derivative $\prt_tp_t(x,y)$ and give estimates for it. Let us outline the argument, and indicate the main difficulty which arises therein.
It is an easy calculation to verify that, under the conditions of Theorem~\ref{t1} for the ``main part'' $p_t^0(x,y)$ of $p_t(x,y)$, the statements similar (and simpler) to those claimed for the whole $p_t(x,y)$ in parts 1 and 2 of Theorem~\ref{t4}, are valid. Namely, the lemma below holds true.
\begin{lem}In each of the cases (a) -- (c) of Theorem~\ref{t1}, the following statements take place.
\begin{enumerate}
\item The function $p_t^0(x,y)$ has a derivative
$$
\prt_tp_t^0(x,y),\quad t>0, \quad x,y\in \Re^d,
$$
which is continuous on $(0,\infty)\times\Re^d\times\Re^d$.
\item The derivative $\prt_tp_t^0(x,y)$ possesses the bound
$$
|\prt_tp_t^0(x,y)|\leq C\Big(t^{-1}\vee t^{-1/\alpha}\Big)H_t(x,y),\quad t>0, \quad x,y\in \Re^d,
$$
where functions $H_t(x,y)$ for the cases (a) -- (c), respectively, are given in Theorem \ref{t3}.
\end{enumerate}
\end{lem}
Using this lemma, one can try to expand these properties of $p_t^0(x,y)$ to the whole $p_t(x,y)$ using e.g. the integral representation
$$
p_t(x,y)=p_t^0(x,y)+\int_0^t\int_{\Re^d}p_{t-s}^0(x,z)\Psi_s(z,y)\,dzds.
$$
However, this can not be done straightforwardly, because $\prt_tp_{t-s}^0(x,z)$ has a non-integrable singularity $(t-s)^{-1}\vee (t-s)^{-1/\alpha}$ at the point $s=t$. Therefore we rewrite the integral representation for $p_t(x,y)$ in the following way (see also \cite[p.747]{Ko00}):
\be\label{46}\ba
p_t(x,y)&
=p_t^0(x,y)+\int_0^{t/2}\int_{\Re^d}p_{t-s}^0(x,z)\Psi_s(z,y)\,dzds+\int_{0}^{t/2}\int_{\Re^d}p_{s}^0(x,z)\Psi_{t-s}(z,y)\,dz\, ds.
\ea
\ee
Now we avoid the singularities, but instead we have to establish the differential properties of $\Psi$ w.r.t. the time variable. We will do this similarly to what was done in Section \ref{s21} before: first we prove such properties for $\Phi$, then for its convolution powers, and finally for $\Psi$. On this way, we meet a minor difficulty that, in the case (c) of Theorem \ref{t1}, the function $b(x)$ is not supposed to be from the class $C^1$, and therefore $\Phi_t(x,y)$ is not continuously differentiable w.r.t. $t$. This difficulty is of a completely technical nature, and will be resolved by choosing a "proper form" for differentiability of $\Phi_t(x,y)$ and its convolution powers.
\subsection{Time-wise derivatives of $\Phi$, $\Phi^{\star k}$ and $\Psi$.}
Consider first the following ``smooth'' case.
\begin{lem}\label{deriv_Phi} Assume that either one of cases (a) or (b) of Theorem \ref{t1} holds true, or case (c) of Theorem \ref{t1} holds true with an additional assumption that $b\in C^1_b(\Re^d)$. Then the statements below take place.
\begin{enumerate}
\item Function $\Phi_t(x,y)$ defined by (\ref{Phi}) has a derivative
$$
\prt_t\Phi_t(x,y),\quad t>0, \quad x,y\in \Re^d,
$$ continuous on $(0,\infty)\times\Re^d\times\Re^d$.
\item For any $\kappa\in (0, \gamma]\cap(0,\alpha)$ and $T>0$, the derivative $\prt_t\Phi_t(x,y)$ possesses the bound
$$
|\prt_t\Phi_t(x,y)|\leq C\Big(t^{-1}\vee t^{-1/\alpha}\Big)t^{-1+\delta} H_t(x,y),\quad t\in(0, T], \quad x,y\in \Re^d,
$$
where the respective indices $\delta$ and the hull kernels $H_t(x,y)$ for the cases (a) -- (c) are given in Theorem \ref{t3}.
\end{enumerate}
\end{lem}
\begin{proof} Let us give the calculations for the case (c), only; the other cases are similar and simpler. One can easily see that $\Phi$ given by (\ref{Phi_c}) is in this case continuously differentiable w.r.t. $t$; let us estimate the derivative. For the part, which in (\ref{Phi_c}) is denoted by $\Phi_t^1(x,y)$, one has
$$\ba
|\prt_t\Phi_t^1(x,y)|&\leq C\Big|a(x)-a(y)\Big|\bigg\{{1\over t^{d/\alpha+2}}\Big|(L^{(\alpha)} g^{(\alpha)})\left({\theta_t(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\Big|\\&\hspace*{1cm}+{1\over t^{d/\alpha+1+1/\alpha}}\Big|(\nabla L^{(\alpha)} g^{(\alpha)})\left({\theta_t(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\Big|\\&\hspace*{1cm}+{1\over t^{d/\alpha+2}}\Big|(\nabla L^{(\alpha)} g^{(\alpha)})\left({\theta_t(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\Big|\left|{\theta_t(y)-x\over t^{1/\alpha}}\right|\bigg\},
\ea
$$
because $\prt_t\theta_t(y)=-b(\theta_t(y))$, which is bounded. Apply (\ref{a_hol_in_c}) with $\kappa\in (0, \gamma]\cap(0,\alpha)$ for the 1st and 3rd term, and with $\kappa=\gamma$ for the 2nd term. Then by (\ref{g_a_frac}) and (\ref{g_a_frac_der}) we derive
$$
\ba
|\prt_t\Phi_t^1(x,y)|&\leq Ct^{-d/\alpha-2+\kappa/\alpha}G^{(\alpha-\kappa)}\left({\theta_t(y)-x\over t^{1/\alpha}}\right)+Ct^{-d/\alpha-2+\kappa}G^{(\alpha)}\left({\theta_t(y)-x\over t^{1/\alpha}}\right)\\&+ Ct^{-d/\alpha-1-1/\alpha+\gamma/\alpha}G^{(\alpha-\gamma+1)}\left({\theta_t(y)-x\over t^{1/\alpha}}\right)+Ct^{-d/\alpha-1-1/\alpha+\gamma}G^{(\alpha+1)}\left({\theta_t(y)-x\over t^{1/\alpha}}\right).
\ea
$$
Since $\kappa\leq \gamma\leq 1$, we can write the above bound in a simpler (but more rough) form:
$$
|\prt_t\Phi_t^1(x,y)|\leq C(t^{-1}\vee t^{-1/\alpha})t^{-1+\delta} H_t(x,y),
$$
$$
\delta=\kappa\wedge \left({\kappa\over \alpha}\right),\quad H_t(x,y)=t^{-d/\alpha}G^{(\alpha-\kappa)}\left({\theta_t(y)-x\over t^{1/\alpha}}\right).
$$
Similarly, for the part which in (\ref{Phi_c}) is denoted by $\Phi_t^2(x,y)$, we obtain
$$\ba
|\prt_t\Phi_t^2(x,y)|&\leq Ct^{-d/\alpha-1/\alpha}\Big|(\nabla g^{(\alpha)})\left({\theta_t(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\Big|\\
&+Ct^{-d/\alpha-1/\alpha-1}\Big|b(\theta_t(y))-b(x)\Big|\Big|(\nabla g^{(\alpha)})\left({\theta_t(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\Big|\\ &+Ct^{-d/\alpha-1/\alpha-1}\Big|b(\theta_t(y))-b(x)\Big|\Big|(\nabla^2 g^{(\alpha)})\left({\theta_t(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right)\Big|\left|{\theta_t(y)-x\over t^{1/\alpha}}\right|\\
&+Ct^{-d/\alpha-2/\alpha}\Big|b(\theta_t(y))-b(x)\Big|\Big|(\nabla^2 g^{(\alpha)})\left({\theta_t(y)-x\over t^{1/\alpha}a^{1/\alpha}(y)}\right);
\ea
$$
here we have used that $\nabla b$ and $\partial_t \theta_t (y)$ are bounded. Therefore, using the Lipschitz condition for $b$ and (\ref{g_a_der}), (\ref{g_a_der2}), we can write a shorter (and less precise) estimate
$$
|\prt_t\Phi_t^2(x,y)|\leq C(t^{-1}\vee t^{-1/\alpha}) H_t(x,y),
$$
which combined with the estimate for $\Phi^1_t(x,y)$ completes the proof.\end{proof}
\begin{lem} Under the conditions and notation of Lemma \ref{deriv_Phi}, the following statements hold true.
\begin{enumerate}
\item The functions $\Phi_t^{\star k}(x,y)$ and $\Psi_t(x,y)$, defined by (\ref{Phi}) have derivatives
$$
\prt_t\Phi_t^{\star k}(x,y),\quad \prt_t\Psi_t(x,y),\quad t>0, \quad x,y\in \Re^d,
$$ continuous on $(0,\infty)\times\Re^d\times\Re^d$.
\item For any $\kappa\in (0, \gamma]\cap(0,\alpha)$ and $T>0$ there exist $C_1, C_2, C>0$ such that
\be\label{42}
|\prt_t\Phi^{\star k}_t(x,y)|\leq {C_1(C_2)^k \over \Gamma(k\delta)}\Big(t^{-1}\vee t^{-1/\alpha}\Big) t^{-1+k\delta}H_t(x,y), \quad t\in (0, T], \quad x,y\in \Re^d,
\ee
\be\label{43}
|\prt_t\Psi_t(x,y)|\leq Ct^{-1+\delta}\Big(t^{-1}\vee t^{-1/\alpha}\Big) H_t(x,y),\quad t>0, \quad x,y\in \Re^d.
\ee
\end{enumerate}
\end{lem}
\begin{proof} Since the proof is similar to that of Lemma \ref{l1}, we just sketch the argument. Write
\be\label{44}
\Phi^{\star (k+1)}_t(x,y)=\int_0^{t/2}\int_{\Re^d}\Phi_{t-s}^{\star k}(x,z)\Phi_s(z,y)\,dz ds+\int_0^{t/2}\int_{\Re^d}\Phi_{s}^{\star k}(x,z)\Phi_{t-s}(z,y)\,dz ds.
\ee
By induction it can be easily shown that each $\Phi^{\star k}_t(x,y)$ has a continuous derivative w.r.t. $t$, and
\be\label{45}\ba
\prt_t\Phi^{\star (k+1)}_t(x,y)&=\int_0^{t/2}\int_{\Re^d}(\prt_t\Phi^{\star k})_{t-s}(x,z)\Phi_s(z,y)\,dz ds+
\int_0^{t/2}\int_{\Re^d}\Phi_{s}^{\star k}(x,z)(\prt_t\Phi)_{t-s}(z,y)\,dz ds\\&+\int_{\Re^d}\Phi_{t/2}^{\star k}(x,z)\Phi_{t/2}(z,y)\, dz.
\ea
\ee
Observe that
$$
\Big((t-s)^{-1}\vee (t-s)^{-1/\alpha}\Big)\leq \Big(2\vee 2^{1/\alpha}\Big)\Big(t^{-1}\vee t^{-1/\alpha}\Big), \quad s\in (0, t/2).
$$
Hence, the bound (\ref{42}) can be obtained by the induction in the same way as (\ref{21}). In addition, for fixed $y\in \Re^d$ each term in the sum has a derivative w.r.t. $t$, continuous w.r.t. $(t,x)\in (0, \infty)\times \Re^d$, and by (\ref{42}) the series for the derivative is also uniformly convergent. Thus, $\Psi$ has a derivative w.r.t. $t$, which is continuous w.r.t. $(t,x)\in (0, \infty)\times \Re^d$ and satisfies (\ref{43}).
\end{proof}
The technical difficulty, which is ``hidden'' in the above argument, is that (in the case (c) only) we need to differentiate w.r.t. $t$ the term
$$
b(\theta_t(y))
$$
in the expression for $\Phi_t^2(x,y)$. If $b$ does not belong to $C^1$, this term may not be continuously differentiable. Nevertheless, in the case (c) the function $b$ is assumed to be Lipschitz continuous, and then by the Rademacher theorem it has a derivative a.e. w.r.t. the Lebesgue measure on $\Re^d$. This observation leads to the following version of the above results in a ``non-smooth'' case.
\begin{lem}\label{deriv_irr} In the case (c) of Theorem \ref{t1}, the following statements hold true.
\begin{enumerate}
\item There exists a set $\Upsilon\subset (0, \infty)\times \Re^d$ of zero Lebesgue measure such that the functions $\Phi_t^{\star k}(x,y)$, $k\geq 1$, and $\Psi_t(x,y)$ are differentiable w.r.t. $t$ at any points $x\in \Re^d$ and $(t,y)\not \in \Upsilon$.
\item For any $(t,y)\not\in \Upsilon$, time-wise derivatives $\prt_t\Phi_t^{\star k}(x,y), k\geq 1$, and $\prt_t\Psi_t(x,y)$ are continuous w.r.t. $x$ and satisfy the bounds (\ref{42}), (\ref{43}).
\end{enumerate}
\end{lem}
\begin{proof} Denote by $\Upsilon_b$ the exceptional set of zero Lebesgue measure, such that $b$ is differentiable at every point outside $\Upsilon_b$. Since $\theta_t$ is a diffeomorphism of $\Re^d$ (see Theorem~I.2.3 and a comment in Chapter I $\S$5 from \cite{CL55}), the set $\Upsilon_{t,b}=\{y:\theta_t(y)\in \Upsilon_b\}$ is again of zero Lebesgue measure, and since $\prt_t\theta_t(y)=-b(\theta_t(y))$, the derivative $\prt_tb(\theta_t(y))$ is well defined for any $y\in \Upsilon_{t,b}$. This derivative is given by
$$
\prt_tb(\theta_t(y))=-\sum_{j=1}^d\prt_{j}b(\theta_t(y))b_j(\theta_t(y)),
$$
where the partial derivatives $\prt_jb$ are now well defined on $\Upsilon_b$ and bounded, because $b$ is Lipschitz continuous. The term $b(\theta_t(y))$ comes into the formula for $\Phi$ in a multiplicative way, with all other terms having derivatives w.r.t. $t$, continuous w.r.t. $(t,x,y)$. Hence, repeating the calculations from the proof of Lemma \ref{deriv_Phi}, we get the (part of) required statements for $\Phi$, with the exceptional set
$$
\Upsilon^1=\{(t,y):y\in \Upsilon_{t,b}\}.
$$
Further, it is easy to get by induction the same statements for $\Phi^{\star k}, k\geq 2$, with the exceptional set
$$
\Upsilon=\Upsilon^1\bigcup\left\{(0,\infty)\times \{y:\int_0^\infty1_{y\in \Upsilon_{s,b}}\,ds>0\}\right\}.
$$
Indeed, express the ratio
$$
{\Phi^{\star (k+1)}_{t+\triangle t}(x,y)-\Phi^{\star (k+1)}_t(x,y)\over \triangle t}
$$
using (\ref{44}), and observe that if $(t,y)\in \Upsilon$ then the respective ratios under the integrals converge $ds$-a.e. to the derivatives
$(\prt_t\Phi^{\star k})_{t-s}(x,z)$ and $\prt_t\Phi_{t-s}(z,y)$. Convergence of the integrals is then provided by the bounds (\ref{42}) and (\ref{Phi_k_bound}) via the dominated convergence theorem, which proves that the derivative $\prt_t\Phi^{\star (k+1)}_t(x,y)$ exists and admits representation (\ref{45}). The bound (\ref{42}) for it follows by induction. Its continuity w.r.t. $x$ also follows by induction and the dominated convergence theorem.
Similarly, the required statement for $\Psi$ can be obtained. Recall that $\Psi_t(x,y)$ is given by (uniformly convergent) series, and for each summand both its differentiability w.r.t. $t$ and the bound (\ref{42}) are proved for $(t, y)\not \in \Upsilon$. Again, by the dominated convergence theorem this leads to the same properties for the whole sum. To get the continuity w.r.t. $x$, we again use the dominated convergence theorem.
\end{proof}
\subsection{Proof of Theorem \ref{t4}}
Now we can easily finalize the proof of Theorem \ref{t4}. Again, we consider only a most cumbersome case (c) with $b$ being just Lipschitz continuous.
By representation (\ref{46}), the first two statements of the theorem follow from the given above statements about time-wise derivatives of $p_t^0(x,y)$ and $\Psi_t(x,y)$; the proofs here are completely analogous to those of Lemma \ref{deriv_irr}, and therefore are omitted. To prove statement 3, note that the set $\Upsilon$ constructed in Lemma \ref{deriv_irr} is such that for every fixed $t>0$ the set $\{y: (t,y)\in \Upsilon\}$ has zero Lebesgue measure. Together with the bounds for $\prt_tp_t(x,y)$ from statement 2, this makes it possible to use the dominated convergence theorem and prove that, for given $t>0$ and $f\in C_\infty(\Re^d)$,
$$
{P_{t+\triangle t}-P_tf(x)\over \triangle t}\to \int_{\Re^d}\prt_tp_t(x,y)f(y)\, dy, \quad \triangle t\to 0,
$$
uniformly in $x\in \Re^d$, which gives statement 3.\qed
\begin{rem}\label{r_Kol} In the above proof of Theorem \ref{t4}, based on (\ref{46}) and consequent parametrix-type iteration of convolutions, we were strongly motivated by the idea used in the proof of Theorem 3.1 in \cite{Ko00}. According to this idea, a derivative in each of two convolution integrals in (\ref{46}) should be moved to a ``least singular'' term, and we just have demonstrated the way this can be made with the derivative $\prt_t$. Unfortunately, we can not proceed in the same manner with the derivative $\prt_x$ unless $p_t^0(x,y)$ depends on $t$ and $x-y$, only. This clearly makes a severe structural limitation (this limitation seems to be hidden in the notation used in the proof of Theorem 3.1 in \cite{Ko00}, but it is still present in calculation therein). Therefore we do not use this argument for the derivative $\prt_x$, and develop another way to justify the whole method.
\end{rem}
\section{Justification of the method. Proofs of Theorem~\ref{t2} and Theorem~\ref{tweak}}\label{s4}
In Section \ref{s2} we proved that the parametrix construction described in Section \ref{s12} is feasible. The function
$$
p_t(x,y), \quad t>0, \quad x,y\in \Re^d,
$$
obtained within this construction is naturally presumed to be a fundamental solution to the Cauchy problem for the pseudo-differential operator given by (\ref{L_ful}), since the whole construction in Section \ref{s12} was based (of course, just formally) on this operator. In this section we justify the whole procedure. We begin with the outline of this procedure.
If we already knew that $p_t(x,y)$ is the fundamental solution, the proofs of Theorem~\ref{t2} and Theorem~\ref{tweak} could be managed in the classical way based on the positive maximum principle for $L(x, D)$; see, for example, \cite[p.165]{EK86}, also \cite[Corollary~4.5.14]{Ja01}. To prove that $p_t(x,y)$ is the fundamental solution, we should verify relations (\ref{L_fund}) and (\ref{L_delta}). Nevertheless, we can not do this directly: the main difficulty here is that we can not prove $p_t(x,y)$ to be smooth enough w.r.t. $x$. To overcome this difficulty, we construct the family $p_{t,\eps}(x,y), \eps>0$ (see (\ref{pe}) below), which we call the \emph{approximative fundamental solution}, because by the construction we would have the following:
\begin{itemize}
\item[(a)] $p_{t,\eps}(x,y), \eps>0$, are smooth enough and approximate $p_t(x,y)$ as $\eps\to 0$ in a proper sense (see Lemma \ref{aux-e} below);
\item[(b)] for $p_{t,\eps}(x,y), \eps>0$, identity (\ref{L_fund}) turns into an approximative identity and instead of (\ref{L_delta}) a similar convergence as $t, \eps\to 0$ holds true (see Lemma \ref{l5} below).
\end{itemize}
These properties allows us to modify properly the standard argument based on the positive maximum principle, which we now apply to the approximative fundamental solution $p_{t, \eps}(x,y)$ instead of $p_t(x,y)$ itself. On this way we prove three basic properties of $p_t(x,y)$, which mainly contain statements I and II of Theorem \ref{t2}, see Section \ref{s52} below. The proof of the ``uniqueness'' statement III of Theorem \ref{t2} (see Section \ref{s53} below) is the most delicate within the whole procedure, and is based mainly on the fact that, in addition to the above properties (a), (b) one has
\begin{itemize}
\item[(c)] $\prt_t p_t(x,y)$ is well defined (see Section \ref{s3} above), and $\prt_tp_{t,\eps}(x,y), \eps>0$, approximates $\prt_t p_t(x,y)$ as $\eps\to 0$ in a proper sense (see Lemma \ref{aux-e} below).
\end{itemize}
We note that exactly this point was our main reason to study in detail the time-wise derivatives of $p_t(x,y)$ in Theorem \ref{t4} and Section \ref{s3}.
\subsection{Approximative fundamental solution: construction and basic properties}
For $\eps>0$ denote
\begin{equation}\label{pe}
p_{t,\epsilon}(x,y):=p_{t+\epsilon}^0(x,y) + \int_0^t \int_{\rd} p_{t-s+\eps}^0(x,z) \Psi_s(z,y) dzds.
\end{equation}
\begin{lem}\label{aux-e}
\begin{enumerate}
\item For any $\epsilon>0$ the function $p_{t,\epsilon}(x,y)$ is continuously differentiable in $t$ and belongs to the class $C^2_\infty(\Re^d)$ in $x$.
\item For any $\epsilon>0$ and $f\in C_\infty(\rd)$
the function
$$
\int_{\Re^d}p_{t,\epsilon}(x,y)f(y)\, dy
$$
is continuously differentiable in $t$ and belongs to the class $C^2_\infty(\Re^d)$ in $x$.
\item $p_{t,\epsilon}(x,y)\to p_t(x,y)$ as $\epsilon\to 0$, uniformly on compact subsets of $(0,\infty)\times \rd\times \rd$.
\item For any $f\in C_\infty(\rd)$, one has
$$
\int_{\Re^d}p_{t,\epsilon}(x,y)f(y)\, dy\to P_tf(x)=\int_\rd p_{t}(x,y)f(y)dy, \quad \epsilon\to 0,
$$
uniformly w.r.t. $(t,x)\in [\tau,T]\times \rd$ for any $\tau>0$, $T>\tau$.
\item For any $f\in C_\infty(\rd)$ \, one has
$$
\prt_t\int_\rd p_{t,\eps}(x,y)f(y)dy\to \prt_tP_tf(x)=\int_\rd \prt_t p_{t}(x,y)f(y)dy, \quad \epsilon\to 0,
$$
uniformly w.r.t. $(t,x)\in [\tau,T]\times \rd$ for any $\tau>0$, $T>\tau$.
\end{enumerate}
\end{lem}
\begin{proof}
\emph{1,2}. The required smoothness of $p_{t,\epsilon}(x,y)$ follows from the smoothness of $p_{t}^0(x,y)$, because we introduce in (\ref{pe}) an additional time shift by positive $\eps$, which removes the singularity at the point $s=t$. In the notation of Theorem \ref{t3}, we have bounds on $p_{t}^0(x,y)$ and its derivatives of the form
$$
p_{t}^0(x,y)\leq CH_t(x,y), \quad |\nabla_x p_{t}^0(x,y)|\leq Ct^{-1/\alpha}H_t(x,y),\quad |\nabla^2_{xx} p_{t}^0(x,y)|\leq Ct^{-2/\alpha}H_t(x,y).
$$
Combined with the bound (\ref{Psi_bound}) on $\Psi_t(x,y)$, this easily yields that $p_{t,\epsilon}(x,y)$ belongs to the class $C^2_\infty(\Re^d)$ w.r.t. the variable $x$. Similar argument applies for the statement 2.
\emph{3,4}. It is clear from the explicit formulae for $p^0_t(x,y)$ that in each of the cases (a) -- (c) of Theorem \ref{t1} one has $p_{t+\epsilon}^0(x,y)\to p_t^0(x,y)$ as $\epsilon\to 0$ uniformly on compact sets in $(0,\infty)\times \rd\times \rd$. For $\Psi$ we have upper bounds of the form (\ref{Psi_bound}). By the dominated convergence theorem, this yields statements 3 and 4; the argument here is the same as those used in the proof of Lemma~\ref{cont}.
\emph{5.} Similarly to (\ref{46}) we have
$$\ba
\prt_t\int_\rd p_{t,\eps}(x,y)f(y)dy&=\int_\rd \prt_tp_{t,\eps}(x,y)f(y)dy=\int_\rd \prt_t p_{t+\eps}^0(x,y))f(y)\, dy\\&+\int_0^{t/2}\int_{\Re^d}\int_{\Re^d}(\prt_t p^0)_{t-s+\eps}(x,z)\Psi_s(z,y)f(y)\,dzdyds
\\&+\int_0^{t/2}\int_{\Re^d}\int_{\Re^d} p^0_{s+\eps}(x,z)(\prt_t\Psi)_{t-s}(z,y)f(y)\,dzdyds\\&+\int_{\Re^d}\int_{\Re^d} p^0_{t/2+\eps}(x,z)\Psi_{t/2}(z,y)f(y)\,dzdy.
\ea
$$
Using the bounds for $p^0, \Psi$ and their $\prt_t$-derivatives obtained above, it is easy to get the required convergence.
\end{proof}
Denote
$$
q_{t,\epsilon}(x,y): = \Big( L_x-\prt_t \Big) p_{t,\epsilon} (x,y).
$$
Observe that $L_xp_{t,\epsilon} (x,y)$ is well defined due to statement 1 in Lemma \ref{aux-e}.
\begin{lem}\label{l5} For any $f\in C_\infty(\rd)$ we have
\begin{itemize}
\item[(i)] \be\label{conv_loc}
\int_\rd q_{t,\eps}(x,y)f(y)dy\to 0, \quad \epsilon\to 0,
\ee
uniformly w.r.t. $(t,x)\in [\tau,T]\times \rd$ for any $\tau>0$, $T>\tau$, and
\be\label{conv_int}
\int_0^t \int_\rd q_{t,\eps}(x,y) f(y)dyds\to 0, \quad \epsilon\to 0,
\ee
uniformly w.r.t. $(t,x)\in [0,T]\times \rd$ for any $T>0$;
\item[(ii)]
$$
\int_{\Re^d}p_{t,\epsilon}(x,y)f(y)\, dy\to f(x), \quad t,\epsilon\to 0,
$$
uniformly w.r.t. $x\in \Re^d$.
\end{itemize}
\end{lem}
\begin{proof} The property (ii) is just a slight variation of statement 3 in Lemma \ref{cont}, and the proof is completely analogous:
from (\ref{domin_p}), (\ref{Psi_bound}), and (\ref{domin}) we deduce for $r_{t,\eps}(x,y)=p_{t,\eps}(x,y)-p_{t+\eps}^0(x,y)$ the bound
$$
|r_{t,\eps}(x,y)|\leq C(T,\eps_0)(t+\eps)^\delta H_{t+\eps}(x,y), \quad t\in (0, T], \quad \eps\in(0, \eps_0].
$$
Then by (\ref{H1})
$$
\sup_x \int_{\Re^d}|r_{t, \eps}(x,y)|\, dy\to 0, \quad t, \eps\to 0+.
$$
Combining this with (\ref{delta}), we complete the proof of (ii).
Let us proceed with the proof of (i). We have already mentioned that, because of an additional shift by $\eps$ of the time variable for $p^0_t(x,y)$ in the formula (\ref{pe}), the singularity at the point $s=t$ therein does not appear, and $p_{t,\eps}(x,y)$ inherits the smoothness properties from $p^0_t(x,y)$. The same reasoning leads to the formulae
$$
\prt_t p_{t, \epsilon}(x,y)=\prt_tp_{t+\epsilon}^0(x,y) + \int_0^t \int_\rd \prt_t p_{t-s+\epsilon}^0(x,x)\Psi_s(z,y)dzds+ \int_\rd p_{\epsilon}^0(x,z)\Psi_t(z,y)dz,
$$
$$
L_x p_{t+\epsilon}(x,y)=L_x p_{t+\epsilon}^0(x,y) + \int_0^t \int_\rd L_x p_{t-s+\epsilon}^0(x,z)\Psi_s(z,y)dzds.
$$
The proofs can be performed in a standard way using e.g. a combination of the integral formula for $L_x$ and the dominated convergence theorem (we omit the details). Recall that
$$
(L_x-\prt_t)p_t^0(x,y)=\Phi_t(x,y).
$$
Hence,
$$
q_{t,\epsilon}(x,y) = \Phi_{t+\epsilon} (x,y)-\int_\rd p_{\epsilon}^0(x,z)\Psi_t (z,y)dz+\int_0^t \int_\rd \Phi_{t-s+\epsilon}(x,z) \Psi_s(z,y)dzds.
$$
Since $\Psi$ comes from the formal representation (\ref{Psi}), which is justified by the bounds obtained in Section \ref{s2}, we have
$$
\Phi_t(x,y)=\Psi_t(x,y)-\int_0^t\int_{\Re^d}\Phi_{t-s}(x,z)\Psi_s(z,y)\,dzds.
$$
Then
$$\ba
q_{t,\epsilon}(x,y)& = \left(\Psi_{t+\epsilon} (x,y)-\int_\rd p_{\epsilon}^0(x,z)\Psi_t (z,y)dz\right)-\int_t^{t+\epsilon} \int_\rd \Phi_{t-s+\epsilon} (x,z)\Psi_s(z,y)dzds\\&=:q_{t,\epsilon}^1(x,y) +q_{t,\epsilon}^2(x,y).\ea
$$
The limit behavior of the integrals w.r.t. $q_{t,\epsilon}^{1,2}(x,y)$ is now well understood. Indeed, by statement 1 of Lemma \ref{cont}, one has that $\Psi$ is uniformly continuous on every compact set in $(0,\infty)\times\Re^d\times\Re^d$. Combined with the bound (\ref{Psi_bound}), this yields that for $f\in C_\infty(\Re^d)$
$$
\sup_{t\in [\tau, 1], x\in \Re^d}\left|\int_{\Re^d}\Psi_{t+\epsilon} (x,y)f(y)\, dy-\int_{\Re^d}\Psi_{t} (x,y)f(y)\, dy\right|\to 0, \quad \eps\to 0.
$$
Uniform continuity of $\Psi$ on compacts also provides, combined with (\ref{delta}) and (\ref{Psi_bound}), that
$$
\sup_{t\in [\tau, 1], x\in \Re^d}\left|\int_{\Re^d}\int_{\Re^d}p_\eps^0(x,z)\Psi_{t} (z,y)f(y)\, dzdy-\int_{\Re^d}\Psi_{t} (x,y)f(y)\, dy\right|\to 0, \quad \eps\to 0.
$$
This proves (\ref{conv_loc}) with $q_{t,\epsilon}^{1}(x,y)$ instead of $q_{t,\epsilon}(x,y)$. Applying (\ref{Psi_bound}) once more, we deduce for $q_{t,\epsilon}^{1}(x,y)$ (\ref{conv_int}) from (\ref{conv_loc}).
Observe that by \eqref{Phi_bound} and \eqref{Psi_bound} we have
\begin{equation}\label{int-sing}
\int_t^{t+\epsilon} \int_\rd |\Phi_{t-s+\epsilon} (x,z)\Psi_s(z,y)|dzds \leq C \delta^{-1} t^{-1+\delta}\epsilon^{\delta} H_{t+\epsilon}(x,y), \quad t\in (0,1], \quad x,y\in \Re^d.
\end{equation}
This gives immediately both (\ref{conv_loc}) and (\ref{conv_int}) with $q_{t,\epsilon}^{2}(x,y)$ instead of $q_{t,\epsilon}(x,y)$: to get the first bound we just multiply (\ref{int-sing}) by $|f(y)|$ and integrate it either with respect to $dy$ (in the first case), or with respect to $dydt$ (in the second case).
While doing that, one should also take into account the explicit form of the hull kernels $H_t(x,y)$, and that $f(y)$ vanishes at $\infty$ (this is simple, and hence we omit the details).
\end{proof}
\begin{comment}
\begin{rem} Below, we will also require the following point-wise versions of (\ref{conv_loc}), (\ref{conv_int}): for every $y\in \rd$
\be\label{conv_loc_p}
q_{t,\eps}(x,y)\to 0, \quad \epsilon\to 0,
\ee
uniformly w.r.t. $(t,x)\in [\tau,T]\times \rd$ for any $\tau>0, T>\tau$, and
\be\label{conv_int_p}
\int_0^t q_{t,\eps}(x,y) ds\to 0, \quad \epsilon\to 0,
\ee
uniformly w.r.t. $(t,x)\in [0,T]\times \rd$ for any $T>0$. The proof of (\ref{conv_loc_p}), (\ref{conv_int_p}) repeats almost literally those of (\ref{conv_loc}), (\ref{conv_int}), and therefore is omitted.
\end{rem}
\end{comment}
\subsection{Positive maximum principle, applied to the approximative fundamental solution}\label{s52}
\begin{lem}\label{posit} We have
$$
p_{t} (x,y) \geq 0, \quad t>0, \quad x,y\in \rd.
$$
\end{lem}
\begin{proof} Since $p_t(x,y)$ is continuous in $(t,x,y)$, it is enough to show that
\begin{equation}\label{int1}
\int_\rd p_t(x,y)f(y)dy \geq 0
\end{equation}
holds true for any $f\geq 0$, $f\in C_\infty(\rd)$. Suppose that \eqref{int1} fails for some $f$; without loss of generality we may assume that \be\label{f_int}\int_{\Re^d}f(y)dy=1.\ee Then there exist $t_0>0$ and $x_0\in \rd$ such that for some $\theta>0$ we have
\begin{equation}\label{int2}
\int_\rd p_{t_0}(x_0,y)f(y)dy <-\theta.
\end{equation}
Since $f\geq 0$, then by property (ii) from Lemma \ref{l5} there exist $\tau_0>0$, $\eps_0>0$, such that
\begin{equation}\label{tau5}
\inf_{x\in \Re^d, \tau\in(0, \tau_0], \eps\in (0, \eps_0]}\int_\rd p_{\tau, \eps}(x,y)f(y)dy>-\theta/3.
\end{equation}
Fix $\tau\in (0,\tau_0\wedge t_0)$ and $T\in (t_0, \infty)$.
By (\ref{int2}) and statement 4 in Lemma~\ref{aux-e} there exists $\epsilon_{\tau,T}>0$ such that
\begin{equation}\label{inf}
\inf_{t\in [\tau,T],x\in \rd} \int_\rd p_{t,\epsilon} (x,y)f(y)dy \leq -\theta,\quad \eps\in (0, \eps_{\tau, T}).
\end{equation}
Define the function
\begin{equation}\label{ptilde}
\tilde{p}_{t,\epsilon} (x,y)= p_{t,\epsilon} (x,y)+ t \theta/(2T).
\end{equation}
Then by (\ref{inf}) and our convention (\ref{f_int}) we have
$$
\inf_{t\in [\tau,T],x\in \rd} \int_\rd \tilde p_{t,\epsilon} (x,y)f(y)dy \leq -\theta/2<0,\quad \eps\in (0, \eps_{\tau, T}).
$$
On the other hand, by statement 2 of Lemma~\ref{aux-e},
\be\label{lalala}
\lim_{|x|\to\infty} \int_\rd \tilde{p}_{t,\epsilon} (x,y)f(y)dy \to t \theta/(2T)>0
\ee
uniformly w.r.t. $t\in [\tau, T]$. Thus, for every $\eps\in (0, \eps_{\tau, T})$ there exist $x_\eps\in \Re^d, t_\eps\in [\tau, T]$, such that
\be\label{glob_min}
\int_\rd \tilde p_{t_\eps,\epsilon} (x_\eps,y)f(y)dy=\min_{t\in [\tau,T],x\in \Re^d} \int_\rd \tilde p_{t,\epsilon} (x,y)f(y)dy\leq -\theta/2<0.
\ee
Observe that, by the construction, we have the following:
\begin{itemize}
\item since the convergence in (\ref{lalala}) is uniform, all points $x_\eps$, $\eps\in (0,\eps_{\tau, T})$, belong to some compact set $K(\tau, T,f)$;
\item by (\ref{tau5}), for $\eps\in (0,\eps_{\tau, T}\wedge \eps_0)$ we have $t_\eps>\tau$.
\end{itemize}
Now all the preparations are done, and we can finalize the proof in a standard way based on the positive maximum principle. Take $\eps\in (0,\eps_{\tau, T}\wedge \eps_0)$; since the minimum in (\ref{glob_min}) w.r.t. $(t,x)\in [\tau, T]\times \Re^d$ is attained at some point $(t_\eps, x_\eps)\in (\tau, T]\in \Re^d$, we conclude that
$$
\int_\rd \prt_t \tilde{p}_{t,\epsilon}(x,y)f(y)dy|_{(t_\eps,x_\eps)} \leq 0
$$
(the inequality may appear if $t_\eps=T$). It is straightforward to check that $(L, C^2_\infty(\Re^d))$ possesses the positive maximum principle, hence we have
$$
L_x\int_\rd \tilde{p}_{t,\epsilon}(x,y)f(y)dy|_{(t_\eps,x_\eps)}= \int_\rd L_x\tilde{p}_{t,\epsilon}(x,y)f(y)dy|_{(t_\eps,x_\eps)} \geq 0.
$$
Thus,
\begin{equation}\label{max1}
\Big( L_x -\prt_t \Big) \int_\rd \tilde{p}_{t,\epsilon}(x,y)f(y)dy\big|_{(t_\eps,x_\eps)}\geq 0.
\end{equation}
On the other hand, by (\ref{conv_loc}) we have
\begin{equation}\label{max2}
\Big(L_x-\prt_t \Big) \int_\rd \tilde{p}_{t,\epsilon}(x,y)f(y)dy=\int_\rd q_{t,\epsilon}(x,y)f(y)dy -\theta/(2T)\to -\theta/(2T), \quad \eps\to 0,
\end{equation}
uniformly w.r.t. $t\in [\tau, T]$ and $x$ in any compact set $K$. Taking $K$ equal to $K(\tau, T,f)$ which contains all $x_\eps$, $\eps\in (0,\eps_{\tau, T})$, we get a contradiction with (\ref{max1}).
\end{proof}
\begin{lem}\label{uniq}
The family $\{p_t(x,y),t>0\}$ possesses the semigroup property
\begin{equation}\label{semi_1}
p_t(x,y)= \int_\rd p_{t-s}(x,z)p_s(z,y)dz, \quad 0<s<t.
\end{equation}
\end{lem}
\begin{proof}
We use the maximum principle in the way similar to those in the previous proof. Namely, let us first prove that for any $f\in C_0(\rd)$, $f\geq 0$, $\int_\rd f(y)dy=1$, we have
\begin{equation}\label{semi_geq}
\int_\rd\left(p_t(x,y)- \int_\rd p_{t-s}(x,z)p_s(z,y)dz\right) f(y)dy\geq 0, \quad 0<s<t.
\end{equation}
Assuming the contrary, we will have for some fixed $s_0>0$ and $\theta>0$, for any (small) $\tau$ and (large) $T$, that
$$
\inf_{t\in [\tau, T], x\in \Re^d}\int_\rd \left(p_t(x,y)-\int_\rd p_{t-s_0}(x,z)p_{s_0}(z,y)dz\right)f(y)dy<-\theta,
$$
and consequently for (small enough) $\eps>0$
$$
\inf_{t\in [\tau, T], x\in \Re^d}\int_\rd \left(p_{t,\eps}(x,y)-\int_\rd p_{t-s_0, \eps}(x,z)p_{s_0}(z,y)dz\right)f(y)dy<-\theta.
$$
This yields that for every (small) $\eps>0$ there exist $t_\eps\in [\tau, T]$ and $x_\eps$ such that
\begin{align*}\int_\rd \Big(\tilde p_{t_\eps,\eps}(x_\eps,y)&-\int_\rd p_{t_\eps-s_0,\eps}(x_\eps,z)p_{s_0}(z,y)dz \Big)f(y)dy\\
&=\min_{t\in [\tau, T], x\in \Re^d}\int_\rd \left(\tilde p_{t,\eps}(x,y)-\int_\rd p_{t-s_0, \eps}(x,z)p_{s_0}(z,y)dz\right)f(y)dy<-\theta/2,
\end{align*}
see (\ref{ptilde}) for the definition of $\tilde p_{t,\eps}(x,y)$. If $\tau$ is chosen properly small, we have $t_\eps\in (\tau, T]$. In addition all $x_\eps$ belong to some compact set
$K(\tau, T, s_0, f)\subset\Re^d$. This yields
$$
\Big( L_x -\prt_t\Big)\int_\rd \left(\tilde p_{t,\eps}(t,y)-\int_\rd p_{t-s_0, \eps}(x,z)p_{s_0}(z,y)dz\right)f(y)dy\bigg|_{(t,x)=(t_\eps, x_\eps)}\geq 0,
$$
which contradicts to the fact that by (\ref{conv_loc})
$$
\Big(L_x-\prt_t \Big) \int_\rd \left(\tilde p_{t,\eps}(t,y)-\int_\rd p_{t-s_0, \eps}(x,z)p_{s_0}(z,y)dz\right)f(y)dy\bigg|_{(t,x)=(t_\eps, x_\eps)}\to -\theta/(2T), \quad \eps\to 0,
$$
uniformly w.r.t. $t\in [\tau, T]$ and $x$ in any compact set $K$. This contradiction proves (\ref{semi_geq}).
The same argument applied to the function
$$
\int_\rd \Big( \int_\rd p_{t-s_0}(x,z)p_{s_0}(z,y)\,dz-p_t(x,y)\Big)f(y)dy
$$
proves the reverse inequality.
\end{proof}
Recall that for any $f\in {C}_\infty^2 (\rd)$ the function $h_f(x):=Lf(x)$ is well defined and belongs to $C_\infty(\rd).$
\begin{lem}\label{mart1}
For any $f\in {C}_\infty^2 (\rd)$ one has
\begin{equation}\label{eq-norm}
\int_\rd p_t(x,y)f(y)dy=f(x)+ \int_0^t \int_\rd p_s (x,y)h_f(y)dyds, \quad t>0.
\end{equation}
\end{lem}
\begin{proof}
We use the same approach as in Lemma~\ref{uniq}. Namely, we first prove that
\begin{equation}\label{mart_geq}
\int_\rd p_t(x,y)f(y)dy\geq f(x)+ \int_0^t \int_\rd p_s (x,y)h_f(y)dyds.
\end{equation}
Assuming the contrary, we will have that for some fixed $s>0$ and $\theta>0$, for any (small) $\tau$ and (large) $T$
$$
\inf_{t\in [\tau, T], x\in \Re^d}\left(\int_\rd p_t(x,y)f(y)dy-\int_0^t\int_\rd p_{s} (x,y)h_f(y)\,dyds- f(x)\right)<-\theta,
$$
and consequently for $\eps>0$ small enough
$$
\inf_{t\in [\tau, T], x\in \Re^d}\left(\int_\rd p_{t, \eps}(x,y)f(y)dy-\int_0^t \int_\rd p_{s,\eps} (x,y)h_f(y)dyds- f(x)\right)<-\theta.
$$
This yields that for every (small) $\eps>0$ there exist $t_\eps\in [\tau, T]$ and $x_\eps$ such that the function
$$
\tilde u_{\eps}(t,x):=\int_\rd p_{t, \eps}(x,y)f(y)dy-\int_0^t\int_{\Re^d} p_{s,\eps} (x,y)h_f(y)dyds- f(x)+t\theta/(2T)
$$
attains its minimal value on the set $\{t\in [\tau, T], x\in \Re^d\}$ at the point $(t_\eps, x_\eps)$, and this minimal value is less than $-\theta/2$. Again, taking at the very beginning $\tau$ small enough (this choice should depend only on $\theta$ and bounds for $p_t^0(x,y)$, $\Psi_t(x,y)$), we can guarantee that $t_\eps>\tau$. This leads to the inequality
\be\label{ineq}
\Big(L_x- \prt_t\Big) \tilde{u}_\epsilon(t,x)\big|_{(t,x)=(t_\eps, x_\eps)} \geq 0.
\ee
In addition, all $x_\eps$ for small $\eps>0$ belong to some compact set
$K(\tau, T, f)\subset\Re^d$.
On the other hand, since $Lf(x)=h_f(x)$, we have
\begin{align*}
\Big( L_x-\prt_t \Big) \tilde u_{\epsilon} (t,x)&= \int_\rd (L_x-\prt_t) p_{t,\epsilon}(x,y) f(y)dy - \int_0^t \int_\rd L_x p_{s,\eps} (x,y)h_f(y)\,dyds\\
& + \int_{\Re^d} p_{t,\eps} (x,y)h_f(y)dy -h_f(x)-\theta/(2T)\\
&= \int_\rd (L_x-\prt_t) p_{t,\epsilon}(x,y) f(y)dy - \int_0^t \int_\rd (L_x -\prt_s) p_{s,\eps} (x,y)h_f(y)\,dyds -\theta/(2T)\\
& = \int_\rd q_{t,\epsilon}(x,y) f(y)dy - \int_0^t \int_\rd q_{s,\eps} (x,y)h_f(y)\,dyds -\theta/(2T),
\end{align*}
where $q_{t,\eps}(x,y)$ comes from Lemma \ref{l5}. Applying this lemma, we get that
$$
\Big( L_x-\prt_t \Big) \tilde u_{\epsilon} (t,x)\to -\theta/(2T), \quad \eps\to 0,
$$
uniformly in $t\in [\tau, T]$ and $x$ in any compact set $K$. Taking $K=K(\tau, T, f)$ containing all $x_\eps$, we get a contradiction to (\ref{ineq}). This proves (\ref{mart_geq}). The same argument applied to the function
$$
\hat u_{\eps}(t,x)=-\int_\rd p_{t, \eps}(x,y)f(y)dy+\int_0^t\int_{\Re^d} p_{s,\eps} (x,y)h_f(y)dyds+ f(x)+t\theta/(2T)
$$
proves the inverse inequality.
\end{proof}
\emph{Proof of statements I and II of Theorem \ref{t2}.} Note that it follows from Lemma \ref{mart1} that
\be\label{id}
\int_\rd p_t(x,y)dy=1, \quad t>0, \quad x\in \Re^d.
\ee
Indeed, take $f\in C^2_\infty(\Re^d)$ such that $f\equiv 1$ on the unit ball in $\Re^d$, and put $f_k(x)=f(k^{-1}x)$. Then
$$
f_k(x)\to 1, \quad Lf_k(x)\to 0, \quad k\to \infty,
$$
and both $|f_k|$ and $|Lf_k|$ are bounded by some constant independent on $k$. Note that for every $x\in \Re^d$, $t>0$, the functions
$$
p_t(x,\cdot), \quad \int_0^t p_s (x,\cdot)ds
$$
are integrable on $\Re^d$: this follows e.g. from (\ref{r_bound}), (\ref{domin_p}), and Proposition~\ref{H-bound}. Hence we can use the dominated convergence theorem in order to pass to the limit as $k\to \infty$ in (\ref{eq-norm}) with $f_k$ instead $f$; this gives (\ref{id}).
Let us summarize. Every $P_t, t>0$, defined by (\ref{semi}), is a positive contraction operator in $C_\infty(\Re^d)$ (statement 2 of Lemma \ref{cont}, Lemma \ref{posit}), and (\ref{id})). With the natural convention that $P_0$ is the identity operator, we have that the family $\{P_t, t\geq 0\}$ is a semigroup (Lemma \ref{uniq}), which is conservative and strongly continuous ((\ref{id}) and statement 3 of Lemma \ref{cont}). Since the semigroup $\{P_t, t\geq0\}$ possesses a continuous transition probability density $p_t(x,y)$, the respective Markov process $X$ is strong Feller. This completes the proof of statement I in Theorem \ref{t2}.
Finally, denote by $\E_x$ the expectation w.r.t. $\P_x$; that is, w.r.t. the law of $X$ with $X_0=x$. Using the Markov property of $X$, it is easy to deduce from (\ref{eq-norm}) and the semigroup property for $p_t(x,y)$ the following: for given $f\in C_\infty^2(\Re^d)$, $t_2>t_1$, and $x\in\Re^d$, for any $m\geq 1$, $r_1, \dots r_m\in [0, t_1]$, and bounded measurable $G:(\Re^d)^m\to \Re$ the identity
$$
\E_x\left[f(X_{t_2})-f(X_{t_2})-\int_{t_1}^{t_2}h_f(X_s)\, ds\right]G(X_{r_1}, \dots, X_{r_m})=0
$$
holds true. This means that for every $f\in C_\infty^2(\Re^d)$ the process
\be\label{Mf}
M^f_t=f(X_t)-\int_{0}^{t}h_f(X_s)\, ds, \quad t\geq 0
\ee
is a $\P_x$-martingale for every $x\in \Re^d$; that is, $X$ is a solution to the martingale problem (\ref{mart}). This proves statement II of Theorem \ref{t2}. \qed
\subsection{The generator of the semigroup $\{P_t, t\geq 0\}$}\label{s53}
Now we can describe the generator $A$ of the semigroup $\{P_t, t\geq 0\}$. The first step is to show that $A$ is well defined on ${C}_\infty^2 (\rd)$, and its restriction to this space coincides with $L$. The argument here is quite standard: first note that since for $f\in {C}_\infty^2 (\rd)$ the process (\ref{Mf}) is a $\P_x$-martingale for every $x\in \Re^d$, and $h_f$ is continuous, by Doob's optional sampling theorem the Dynkin operator $U$ (cf. \cite[Chapter II.5]{GS74}) is well defined on $f$ and $Uf=Lf$. Since $Uf$ is continuous, by \cite[Theorem II.5.1]{GS74} we get that $f$ belongs to the domain of the generator $A$, and $Af=Uf=Lf$.
Hence $(L, C_\infty^2 (\rd))$ is the restriction of $(A, \mathcal{D}(A))$. Since $A$ is a closed operator, this yields that $(L, C_\infty^2 (\rd))$ is closable. Let us show that its closure coincides with whole $(A, \mathcal{D}(A))$; this would prove statement III of Theorem \ref{t2}.
Take $f\in C_\infty(\Re^d)$ which belongs to the domain $\mathcal{D}(A)$ of the generator $A$.
Fix $t>0$, and consider the element $f_{t}=P_{t}f$. Write
$$
f_{t,\eps}(x)=\int_{\Re^d}p_{t,\eps}(x,y)f(y)\, dy,
$$
and observe that we have the following properties.
\begin{itemize}
\item By statement 2 in Lemma \ref{aux-e}, $f_{t,\eps}\in C^2_\infty(\Re^d)$.
\item Since $A$ is an extension of $L$, we have that $f_{t,\eps}$ belongs to the domain of $A$, and $Af_{t, \eps}=L f_{t,\eps}$.
\item By statement 4 in Lemma \ref{aux-e}, one has $f_{t,\eps}\to f_t$ in $C_\infty(\Re^d)$ as $\eps\to 0$.
\item By statement 5 in Lemma \ref{aux-e}, one has $\prt_tf_{t,\eps}\to \prt_tf_t$ in $C_\infty(\Re^d)$ as $\eps\to 0$.
\item By Lemma \ref{l5}, one has $(\prt_t-L)f_{t,\eps}\to 0$ in $C_\infty(\Re^d)$ as $\eps\to 0$.
\end{itemize}
Recall that $f\in \mathcal{D}(A)$, and therefore $\prt_t f_t=Af_t.$ Hence summarizing all stated above, we get that $f_{t,\eps}\in C^2_\infty(\Re^d)$ approximates $f_t$, and $Lf_{t,\eps}$ approximates $Af_t$ in $C_\infty(\Re^d)$ as $\eps\to 0$. This gives that the domain of the $C_\infty(\Re^d)$-closure of $(L, {C}_\infty^2 (\rd))$ contains every element of the form
$$
f_t=P_tf, \quad t>0,\quad f\in \mathcal{D}(A).
$$
This clearly yields that this closure coincides with whole $(A, \mathcal{D}(A))$.\qed
\subsection{Process $X$ as the unique weak solution to (\ref{SDE}): proof of Theorem \ref{tweak}}
It follows straightforwardly from the It\^o formula that any weak solution to (\ref{SDE}) is a solution to the martingale problem (\ref{mart}). Since we have already proved that this martingale problem is well posed, this immediately proves \emph{uniqueness} of the weak solution to (\ref{SDE}). The proof of the \emph{existence} of a weak solution is easier, and can be conducted e.g. in the following standard way.
\begin{itemize}
\item Consider a family of equations with smooth coefficients approximating the coefficients of (\ref{SDE}); this can be made in such a way that corresponding transition probability densities satisfy uniformly the estimates from Theorem \ref{t3}.
\item Consider the sequence of pairs $(X^n, Z^{(\alpha)})$, where $X^n$ denotes the strong solution to (\ref{SDE}) with given $\alpha$-stable process $Z^{(\alpha)}$. Using the uniform bounds for the transition probability densities for $X^n$ it is easy to apply standard criteria (e.g. \cite[Chapter 3, Theorem 8.6]{EK86}) in order to prove that this sequence is weakly compact in $\DD(\ax, (\Re^{d})^2)$.
\item Then there exists a weak limit point $(\tilde X, \tilde Z^{(\alpha)})$ of a sequence $(X^n, Z^{(\alpha)}), n\geq 1$. By \cite[Theorem 2.2]{KP91}, this weak limit point gives a weak solution to (\ref{SDE}); observe that the principal condition C2.2(i) of this theorem holds true because the second components in the above sequence are performed by the same semimartingale $Z^{(\alpha)}$ (see \cite[Remark 2.5]{KP91}).
\end{itemize}
|
\section{Introduction and Results}
\subsection{Introduction and Background}
In gauge theories, the Wilson loop $W_{\bf R} [C]$ constitutes an important class of defect operator localized on a contour $C$. Carrying representation ${\bf R}$ under gauge group $G$, it has played important roles in probing ground-state phases of gauge theories. It has also been used for studying excitation dynamics: Wilson loops on a light-like cusp or a polygon $C_n$ with $n$-cusps located at $x_1, \cdots, x_n$ probed deeply inelastic scattering of hadrons or high-energy parton-parton scattering with associated evolution equation for regge trajectory.
\cite{Korchemskaya:1992je}
In this paper, we undertake perturbtive study of light-like Wilson loop in the ABJM theory\cite{Maldacena:2008a} for the contour $C$ a lightlike polygon with $n$-cusps ($n \ge 4)$ located at $x_1, \cdots, x_n$. The ABJM theory is $(2+1)$-dimensional, parity-invariant Chern-Simons matter theory having ${\cal N}=6$ superconformal symmetry and $SU (N) \times SU (N)$ gauge symmetry with Chern-Simons levels $+k, -k $, respectively. We shall take planar limit $N \rightarrow \infty$ while holding `t Hooft coupling $\lambda = (N/k)$ fixed and study the Wilson loop up to second order in $\lambda$.
Our motivation comes from the equivalence of the lightlike polygon Wilson loops with parton scattering amplitudes on one hand and correlation functions on another in $(3+1)$-dimensional, planalr $\mathcal{N}=4$ Super Yang-Mills (SYM) theory, so we shall briefly review the results. A stepping stone was provided by the remarkable ansatz of Bern, Dixon and Smirnov (BDS) for the all-loop resummation formula for the four-point scattering amplitudes \cite{Bern:2005a}, built upon recursion relations at one-, two- and three-loops and upon universality of the splitting function for extending the ansatz to higher-point scattering amplitudes. The BDS ansatz prompted to investigate the scattering amplitudes in strongly coupled ${\cal N}=4$ SYM theory. Utilizing the AdS/CFT correspondence, Alday and Maldacena mapped the problem to a problem in string theory on weakly curved, $(4+1)$-dimensional anti-de Sitter (AdS) spacetime background. By exploiting conformal invariance, they calculated the four-point scattering amplitude on a D-brane displaced to the Coulomb branch by fermionic T-dualizing it to a holographic Wilson loop on a tetragon enclosed by light-like edges. The infrared (IR) finite part of the Wilson loop matched with the ultraviolet (UV) finite part of the all-loop BDS conjecture for four-point scattering amplitudes. Replacing cusp anomalous dimension to weakly coupled counterpart, the IR divergent part also agreed with the all-loop BDS conjecture. At weak `t Hooft coupling regime, the scattering amplitude - Wilson loop duality relates the $n$-particle MHV scattering amplitudes $A_n^{\rm MHV} = A_n^{\rm tree} K_n$ and the lightlike $n$-gon Wilson loops $\langle W_n \rangle$ by \cite{Drummond:2007a}
\bea
\log K_n (p_1, \cdots, p_n) = \log \langle W_n \rangle (x_1, \cdots, x_n).
\label{duality-relation}
\eea
The equivalence was confirmed by explicit computation for $n=4, 5$ and up to two loops.
The situation changed for higher-point scattering amplitudes. A discrepancy was found between the BDS conjecture of six-point scattering amplitude and the hexagon Wilson loop \cite{Drummond:2007b}. This result suggested either breakdown of the fermionic T-duality between scattering amplitude and Wilson loop or breakdown of the BDS conjecture. The first possibility was excluded by direct computation of parity even part of six gauge boson amplitude \cite{Bern:2008a}. There also appeared non-trivial differences to the BDS conjecture. This quantity, the discrepancy from the BDS conjecture, constitutes the remainder function. Detailed investigation of the lightlike hexagon Wilson loop at two-loop order \cite{Heslop:2009a} concluded that the BDS conjecture indeed breaks down. Numerically, they found that this remainder function depends only on conformal cross-ratios. Moreover, absence of the remainder function at tetragon and pentagon Wilson loops was understood as a consequence of the dual conformal symmetry and conformal Ward identity \cite{Drummond:2007a, Drummond:2008b}. This conformal Ward identity fixed functional structure of the Wilson loop expectation value and limited the number of free variables in the remainder function.
The analytic structure of this remainder function of hexagon Wilson loop was obtained in \cite{Duhr:2010a}. There, central tool was to utilize the special limit of the quasi-multi Regge kinematics(QMRK). It was found that the remainder function could be written with the uniform transcendentality in terms of the Goncharov polylogarithms. The structure was very involved but was reduced to a combination of classical polylogarithms of uniform transcendentality of degree four \cite{Spradlin:2010a}. These techniques were also utilized in subsequent works investigating analytic structure of the remainder function up to three loop order \cite{Heslop:2011a,Drummond:2011a}. It was further extended to analytic structure of the non-maximal helicity violating (nMHV) amplitudes at two loop order with help of the symbols \cite{Drummond:2011b}.
Moreover, there is a growing evidence that, in the collinear limit of adjacent edges of the polygon Wilson loop, the remainder function can be systematically expanded in powers of collinearity and that each term can be computed to all order in the `t Hooft coupling by utilizing the integrability of the four-dimensional planar ${\cal N}=4$ SYM theory \cite{Basso:2013vsa}, \cite{Basso:2013aha}, \cite{Basso:2014koa}, \cite{Basso:2014nra}. For the hexagon Wilson loop, the expansion takes schematically the form
\bea
\mbox{Rem}_6 = \sum_{\ell=1}^\infty \lambda^\ell \mbox{Rem}_6^{(\ell)}
\eea
where, in terms of suitable parametrization $(\tau, \sigma, \phi)$ of kinematical variables,
\bea
\mbox{Rem}_6^{(\ell)} = \sum_{m=1}^\infty e^{- m \tau} \sum_{p=0}^{[m/2]} \cos (m- 2 p) \phi \sum_{n=0}^{\ell - 1} \tau^n F^{(\ell)}_{m,p,n}(\sigma).
\eea
The collinear limit corresponds $\tau \rightarrow \infty$ and $m$ is the number of particles excited on a color-electric flux tube stretching between two edges that become collinear. Most interestingly, the dynamics of these excitations on the flux tube is described by an integrable spin chain.
We now turn to the ABJM theory and compare known results about scattering amplitudes and lightlike Wilson loops with those of the ${\cal N}=4 $ SYM theory. Given that both theories are superconformal, one might anticipate from the AdS/CFT correspondence that the Wilson loops / scattering amplitudes duality would hold equally and result in identical functional structure, at least at strong coupling regime. This is not the case. In the ABJM theory, the $n$-parton scattering amplitudes is definable only for even integers $n = 4, 6, 8, \cdots$ (since the dynamical matter fields are in bifundamental representations), while the $n$-gon Wilson loops exist for all integers $n = 4, 5, 6, \cdots$. So, a possibility is that the duality Eq.(\ref{duality-relation}) holds only for even integers $n$, while leaving Wilson loops of odd-sided polygons a class of its own. Secondly, the open superstring in AdS$_4 \times \mathbb{CP}^3$ does not possess the fermionic T-duality. Absence of fermionic T-duality symmetry implies no relation between scattering amplitudes and light-like Wilson loops. This then rules out the duality relation Eq.(\ref{duality-relation}) in the ABJM theory.
The four-point scattering amplitude was computed in two different approaches : the unitarity-cut method \cite{Huang:2011a} and direct superspace diagram calculation \cite{Bianchi:2011a}. Parallel to the $\mathcal{N}=4$ SYM theory, the four-point scattering amplitudes turned out related to the tetragon Wilson loop expectation value. It was observed that the scattering amplitudes in ABJM and $\mathcal{N}=4$ SYM theories are similar: two-loop four-point amplitude in the ABJM theory is equivalent to that of $\mathcal{N}=4$ SYM theory to all orders in $\epsilon$ \cite{Bianchi:2011b}. The sixe-point scattering amplitudes were computed at one-loop \cite{Beisert:2012a,Leoni:2012a,Brandhuber:2012a} and at two-loop \cite{Huang:2012a}, respectively. The one-loop six-point scattering amplitudes turned out non-vanishing, being proportional to shifted tree-level six-point amplitude.
The Wilson loops come in two types, one for $SU(N)_k$ and another for $SU(N)_{-k}$. Firstly, from the strong coupling side, one expects that the $AdS_4$ structure dictates that the minimal surface associated with the ABJM Wilson loop ought to behave the same as that for the ${\cal N}=4$ SYM Wilson loop.
there are two Wilson loop operators the expectation value at weak `t Hooft coupling regime starts perturbative corrections start at two loops. The Wilson loop operator of each gauge group preserves $1/6$ of supersymmetry and its expectation value is a function of $\lambda$, starting at the second order. On the other hand, average of the two Wilson loop operators preserves $1/2$ of the supersymmetry and its expectation value is a function of $\lambda^2$, starting at the second order.
Secondly, the perturbative corrections depend crucially on the radiatively-induced, parity-even gauge boson self-energy. Feynman rules for this gauge boson self-energy propagation in ABJM theory resembles propagators of gauge boson in $\mathcal{N}=4$ SYM\cite{Rey:2008a, Plefka:2008a}. This suggests that matter contributed part of two-loop Wilson loop in ABJM theory could be considered as one-loop Wilson loop in $\mathcal{N}=4$ SYM. Moreover, it was found that pure Chern-Simons part could be absorbed to matter part by replacing regularization energy scale. Indeed, anomalous conformal Ward identity again works for the ABJM theory. Indeed, it was conjectured that Wilson loop in the ABJM theory almost have same structure with Wilson loop in $\mathcal{N}=4$ SYM. Up to 4-point, this was numerically confirmed \cite{Plefka:2010a}. This suggests that bosonic Wilson loop expectation value is not equivalent object to scattering amplitude since odd loop Wilson loop expectation values are 0 in ABJM theory. 2-loop 6 point result seems more complex, appeared by linear combination of tree level amplitude and shifted tree level amplitude.
Despite of all these, explicit computation poses interesting puzzle. The light-like tetragon Wilson loop at two loops displays striking similarity to the Wilson loop of SYM theory at one loop. Moreover, the Wilson loop result was dual to the four-particle scattering amplitudes, thus satisfying the duality relation Eq.(\ref{duality-relation}). This might be a coincidence because, in ABJM theory, scattering amplitudes and tetragon Wilson loops are MHV-like observables for $n=4$ but not for higher $n$.
Another our motivation concerns comparative study between the lightlike polygon Wilson loop in $(2+1)$-dimensional ABJM theory and that in $(3+1)$-dimensional $\mathcal{N}=4$ Super Yang-Mills theory. In \cite{Plefka:2010a}, they computed Wilson loop expectation value for two loop tetragon case. The result was quiet remarkable, its structure is almost equivalent to one-loop Wilson loop in $\mathcal{N}=4$ Super Yang-Mills theory. There is no reason that existence of such similarity for arbitrary n($n>4$)-gon case. Our goal is constructing n-gon Wilson loop expectation value and comparing with $\mathcal{N}=4$ Super Yang-Mills theory's one.
Yet another motivation is to investigate whether and how the color electric flux-tube picture advanced in \cite{Basso:2013vsa}, \cite{Basso:2013aha}, \cite{Basso:2014koa}, \cite{Basso:2014nra} can be extended to the ABJM theory. While such an extension is much anticipated from the geometric similarity in the strong coupling regime, it is far from being clear in the weak coupling regime. The flux tube configuration in the ABJM theory is not transparent and also particle excitations on the flux tube worldsheet appear quite different. Nevertheless, results in this work indicates that a variety of geometric structures are quite similar for both theories but for rather different reasons. We expect that the soft-collinear operator product expansion we develop in this paper and the universal antenna function we derived would provide a starting point bridging conceptual and caculational gaps between the two theories.
\subsection{Main Results}
Firstly, we obtained analytic result of the light-like hexagon Wilson loop at second-order in `t Hooft coupling constant $\lambda = (N/k)$. Regularizing the UV divergence by supersymmetric dimensional reduction scheme with $d = 3 -2 \epsilon$, the result reads
\bea
\big< W_\Box [C_6] \big>^{(2)}_{\rm ABJM}= \lambda^2 \left[-{1 \over 2} \sum_{i=1}^6 \frac{(x_{i,i+2}^2 \tilde{\mu}^2)^{2\epsilon}}{(2\epsilon)^2}+\mbox{BDS}^{(2)}_{6}(x) + \left( \frac{9}{2} \mbox{Log}^2(2) + \frac{\pi^2}{3} \right) \right].
\label{hexagon}
\eea
Here, $x_i \ (i=1, \cdots, 6)$ are hexagon vertex positions, and $\tilde{\mu}^2 := 8 \pi e^{\gamma_E} \mu^2$. The second term is the UV-finite "BDS" function \footnote{Hereafter, we shall adopt the terminology of corresponding quantities in the ${\cal N}=4$ SYM theory} :
\bea
\mbox{BDS}^{(2)}_{6}(x) \! &=& \! {1 \over 2} \! \sum_{i=1}^6 \left[\frac{1}{4} \Log^2\left(\frac{x_{i,i+3}^2}{x_{i+1,i+4}^2}\right)- \! \Log\left(\frac{x_{i,i+2}^2}{x_{i,i+3}^2}\right) \! \! \Log\left(\frac{x_{i+1,i+3}^2}{x_{i,i+3}^2}\right) -\frac{1}{2} \! \mbox{Li}_2 \left(1-\frac{x_{i,i+2}^2 x_{i+3,i+5}^2}{x_{i,i+3}^2 x_{i+2,i+5}^2} \right)\right] \nonumber \\
&+& {\pi^2 \over 2}. \label{BDS6}
\eea
This function depends on the vertex positions the same way as the UV-finite, leading order BDS function of the $\mathcal{N}=4$ SYM theory. All the finite parts in Eqs.(\ref{hexagon}, \ref{BDS6}) exhibit the property of the uniform transcendentality.
Secondly, we obtained the ABJM antenna function relevant for lightlike polygon Wilson loops. Recall that, in the ${\cal N}=4$ SYM theory, the splitting function in the scattering amplitudes is defined by the IR factorization associated with multiple collinear limit of massless particles. Here, in ABJM theory, we dwell on special kinematic configuration and focus on the limit of the Wilson loop contour. For reasons that will become clear later, we focus on a sort of operator product expansion involving two collinear edges of fraction $h_1, h_3$ and one soft edge of fraction $h_2$ in between. This limit defines the triple antenna function: lightlike $n$-gon Wilson loop is decomposable into a product of $(n-2)$-gon light-like Wilson loop and this antenna function. At second order in `t Hooft coupling $\lambda$, the triple antenna function consists of two parts: the pure Chern-Simons part and matter-dependent part. Our result is
\bea
\left[ {\mbox{Ant}^{(2)}[C_n]
\over \mbox{Ant}^{(0)}[C_n]}
\right]_{\rm CS} &=& \frac{\Log(2)}{2\epsilon} \nonumber \\
&+&\frac{1}{2}\log(2) \mbox{Log} (h_1) +\frac{1}{2}\Log(2) \Log(h_3) +\frac{1}{2}\Log(2) \Log (x_{24}^2) +\frac{1}{2}\Log(2) \Log (x_{35}^2)
\nonumber \\
&-&\frac{7\pi^2}{24}+\Log^2(2)
\eea
for the pure Chern-Simons part and
\bea
\left[ {\mbox{Ant}^{(2)}[C_n] \over \mbox{Ant}^{(0)}[C_n] } \right]_{\rm matter} &=& \frac{1}{4\epsilon^2} + \frac{1}{4\epsilon}(\Log (h_1) + \Log(h_3) + \Log(x_{24}^2) + \Log(x_{35}^2)) \nonumber \\
&+&\frac{1}{2} \Log( h_1) \Log(x_{24}^2) +\frac{1}{2} \Log (h_3) \Log(x_{35}^2) +\frac{1}{2} \Log (x_{35}^2) \Log(x_{24}^2) \nonumber \\
&-&\frac{1}{2} \Log (h_1) \Log (h_3) - \frac{\pi^2}{6}
\eea
for the matter-dependent part. The result demonstrates that splitting function also displays the property of maximal transcendentality. Moreover, the result is independent of $n$, suggesting that the triple antenna function holds universally for all $n$. The total triple antenna function is strikingly similar to triple splitting function in the limit middle parton becomes soft. We found, however, they are still subtly different.
Thirdly, combining the two results above, we obtained the simplest functional form of lightlike polygon ABJM Wilson loop to the second-order in `t Hooft coupling. The requisite shape of lightlike polygon must obey the positivity condition and satisfy vanishing Gram determinant condition.
It turns out such kinematics requirements limit the lightlike polygon only to the one with even number of cusps, marking a stark difference from the ${\cal N}=4$ SYM theory. Demanding IR factorization of the Wilson loop with the universal antenna function, we obtained a version of operator product expansion, leading to a linear recursion relation among the light-like Wilson loops:
\bea
\Big< W_\Box [C_n ] \Big> \quad \longrightarrow \quad \mbox{Ant}[C_n] \cdot
\Big< W_\Box[C_{n-2}] \Big> .
\eea
Solving this recursion with the hexagon Wilson loop (\ref{hexagon}) as an input, we finally find that
\bea
\boxed{
\Big< W_\Box [C_n] \Big>^{(2)}_{\rm ABJM}=\lambda^2 \left[ -{1 \over 2} \sum_{i=1}^n \frac{(x_{i,i+2}^2 \tilde{\mu}^2)^{2\epsilon}}{(2\epsilon)^2} +\mbox{BDS}^{(2)}_n(x) + \mbox{Rem}^{(2)}_n(u) \right],
}
\eea
where $\mbox{Rem}_n(u)$ is the remainder function that depends on the Mandelstam invariants only through conformal cross-ratios $u$\rq{}s. At two loops, the remainder function is independent of $u$\rq{}s and reads
\bea
\boxed{
\mbox{Rem}^{(2)}_n(u) = \left[ n\left(\frac{\pi^2}{12} +\frac{3}{4} \log^2(2) \right) - \frac{\pi^2}{6} \right].
}
\eea
Here, we extracted this analytic result by utilizing the PSLQ algorithm to the high precision numerical integrations. As a nontrivial check, we derived the spacelike circular Wilson loop expectation value from the $n \rightarrow \infty$ continuum limit and found perfect agreement with the previous results.
This paper is organized as follows. In section 2, we set our notations for ABJM Wilson loops and summarize known results for lightlike tetragon Wilson loop. In section 3, we study lightlike hexagon Wilson loop at two loop order. Planar Feynman diagrams contributing to this order include ladder diagrams, triple-vertex diagrams and matter-dependent diagrams.
Details of integrand are relegated to the Appendix. In section 4, we explain computational details. We evaluate the matter-dependent part analytically. The ladder and the triple-vertex diagrams are more involved. We compute them numerically. To achieve high precision, we study so-called two-parameter configurations and utilize Mellin-Barnes transformation. In section 5, we construct the antenna function for the ABJM theory. Requiring the OPE-like factorization conditions, we show that the antenna function relates $n$-gon Wilson loop to $(n-2)$-gon Wilson loop. Solving this relation recursively general expression of the lightlike $n$-gon Wilson loop can be obtained. In section 6, we carry this out and obtain analytic expression for arbitrary $n$-gon. As a consistency check, we examine the $n \rightarrow \infty$ limit and reproduce the known exact result of circular Wilson loop. In section 7, we compare the structure of the lightlike polygon Wilson loop for the ABJM theory with that for the ${\cal N}=4$ SYM theory, and discuss physical implications.
\section{Light-like Polygon Wilson loop in ABJM Theory}
\subsection{ABJM Theory}
The ABJM theory describes $(2+1)$-dimensional supersymmetric matter interacting with Chern-Simons gauge system. It has ${\cal N}=6$ superconformal symmetry (having 24 conserved supercharges) and $U(N) \times \overline{U(N)}$ gauge group with Chern-Simons levels $+k, -k$, respectively. The gauge fields are denoted as $A_m(x) \in u(N)$ and $\overline{A}_m(x) \in \overline{u(N)}$. For our notations and conventions of the field contents, Lagrangian and Feynman rules, see the Appendix A. For foregoing considerations, it suffices to note that the action includes the pure Chern-Simons density \cite{Maldacena:2008a}
\bea
S_{\rm CS}&=& + \frac{k}{4\pi} \int \rmd^dx \epsilon^{mnp} \mbox{Tr} \Big( A_m \partial_{n} A_p + \frac{2i}{3} A_{m} A_{n} A_{p} \Big) \label{CS_action_a1}\\
\overline{S}_{\rm CS}&=& -\frac{k}{4\pi} \int \rmd^dx \epsilon^{mnp} \overline{\mbox{Tr}} \Big( \overline{A}_m \partial_{n} \overline{A}_p + \frac{2i}{3} \overline{A}_{m} \overline{A}_{n} \overline{A}_{p} \Big). \label{CS_action_b1}
\eea
Here, the Chern-Simons density has levels $+k$ and $-k$, respectively. Invariance of the action under large gauge transformation puts $k$ integer-valued. The action is invariant under the generalized parity that simultaneously reverts one spatial coordinates and exchanges the two gauge fields. In this theory, the ABJM Wilson loop operator in the fundamental representation $(\Box, 1) \oplus (1, \Box)$ of the gauge group $U(N) \times \overline{U(N)}$ is defined by \cite{Rey:2008a}:
\bea
\mathcal{W}_\Box [C]:=\frac{1}{2}\Big(W_\Box [C]+\overline{W}_\Box [C] \Big), \label{def:Wilson}
\eea
where $W_\Box [C]$ and $\overline{W}_\Box [C]$ refer to the Wilson loop of the fundamental representation of $U(N)$ and $\overline{U(N)}$ gauge groups, respectively.
The close contour $C$ is a geometric datum of the Wilson loop operator. Hereafter, we shall exclusively deal with Lorentzian contour $C_n$ connecting $n$ vertices $x_1, x_2, \cdots, x_n$ whose adjacent points are lightlike-separated. The total set $C_n$ with $n = 4, 5, 6, \cdots$ form lightlike $n$-gons. Denote the distance vectors between a pair of vertices by
\bea
x_{i,j} \equiv [x_i - x_j] \qquad \quad i,j = 1, \cdots, n.
\eea
Among them are the lightlike-separated edges $x_{i+1, i}$. Denote a point on $i$-th edge by $z_i$. In parametrized form, it is
\bea
z_i (\tau) = x_i + y_i \tau \qquad \mbox{where} \qquad y_i \equiv x_{i+1, i}, \qquad 0 \le \tau \le 1.
\eea
We relegate notations for various Lorentz invariants of $x_i$'s to Appendix A.
The lightlike $n$-gon Wilson loop operators for $SU(N)$ and $\overline{SU(N)}$ gauge groups take the form
\bea
W_\Box [C_n]&=&\frac{1}{N} \mbox{Tr}\mathcal{P} \mbox{exp} \left[i \oint_{C_n} \rmd\tau A_m(x(\tau)) \dot{x}^m(\tau) \right] \label{def:Wilson_a}\\
\overline{W}_\Box [C_n]&=& \frac{1}{N} \overline{\mbox{Tr}}\mathcal{P} \mbox{exp} \left[i \oint_{C_n} \rmd\tau \overline{A}_m(x(\tau)) \dot{x}^m(\tau) \right]. \label{def:Wilson_b}
\eea
Both are $1/6$-BPS operators preserving 4 supercharges. Under the generalized parity, the two Wilson operators are interchanged each other. On the other hand, the ABJM Wilson loop $\mathcal{W}_\Box [C]$ is $1/2$-BPS operator preserving 12 supercharges. By construction, it is invariant under the generalized parity. The $n$ vertices of $C_n$ break all supersymmetries. This implies that the expectation values of these Wilson loops receive quantum corrections. Analyzing these corrections in the regime of infinite number of color $N \rightarrow \infty$ and weak `t Hooft coupling $\lambda = (N/k) \ll 1$ is the main focus of this paper.
\subsection{Previous Results}
Our goal is to compute the vacuum expectation value of the lightlike polygon Wilson loop. In the planar limit, we evaluate it in perturbation theory of the `t Hooft coupling $\lambda$:
\bea
\big< {\cal W}_\Box[C] \big> = \sum_{\ell = 0}^\infty \lambda^\ell \ \big< {\cal W}_\Box[C] \big>^{(\ell)}
\eea
and similarly for $\big< W_\Box[C] \big>$ and $\big< \overline{W}_\Box [C] \big>$. The Wilson loops $\big< W_\Box[C] \big>$ and $\big< \overline{W}_\Box [C] \big>$ are 1/6-BPS configurations and in general receive perturbative corrections to all orders in $\lambda$. On the other hand, the ABJM Wilson loop $\big< {\cal W}_\Box[C] \big>$ is 1/2-BPS configuration and receive perturbative corrections only at even order of $\lambda$. This is an elementary consequence of the fact that the ABJM Wilson loop is invariant under the generalized parity. Since the net effect of the generalized parity is to flip $k$ to $-k$, equivalently, $\lambda$ to $-\lambda$, it follows immediately that
\begin{equation}
\big< W_\Box [C]\big>^{\ell=\mbox{odd}}=-\big< \overline{W}_\Box [C]\big>^{\ell = \mbox{odd}}
\end{equation}
Actually, the result is stronger at linear order in $\lambda$. At this order, kinematical considerations indicate that $\big< W_\Box [C_n] \big>^{(1)}$ and $\big< \overline{W}_\Box [C_n]\big>^{(1)}$ vanish separately. By the generalized parity transformation, it also follows that
\begin{equation}
\big< W_\Box [C]\big>^{\ell = \mbox{even}}=\big< \overline{W}_N[C]\big>^{\ell = \mbox{even}}.
\end{equation}
We conclude that
\bea
\big< {\cal W}_\Box[C]\big> = \sum_{\ell = 0}^\infty \lambda^{2 \ell} \big< W_\Box[C]\big>^{(2 \ell)} = \sum_{\ell = 0}^\infty \lambda^{2 \ell} \big< \overline{W}_\Box[C]\big>^{(2\ell)}.
\eea
The leading-order correction arises at two-loop order $O(\lambda^2)$. The diagrams contributing to this order are categorized to three groups \cite{Plefka:2010a}: matter-dependent diagrams, gauge boson ladder diagrams, and gauge boson triple-vertex diagrams. The contribution of the matter diagrams is equivalent to one-loop contribution in the $\mathcal{N}=4$ SYM theory. This is because, in the ABJM theory, the finite one-loop correction to the gauge boson propagator is precisely the same as the tree-level gauge boson propagator in the ${\cal N} = 4$ SYM theory \cite{Rey:2008a}. This means that differences between the ABJM theory and the $\mathcal{N}=4$ SYM originate from ladder diagrams and triple-vertex diagrams. Both diagrams originate from gauge boson interactions through the Chern-Simons parts. Computationally, these two contributions are the most complicated.
The general structure of the two-loop corrections to the light-like Wilson loop expectation value can be obtained by requiring the anomalous conformal Ward identities. For this consideration, we can split the contributions to two parts: the matter contribution and the Chern-Simons contribution.
As explained above, the matter contribution is structurally the same as the one-loop contribution to the lightlike Wilson loops in $\mathcal{N}=4$ SYM theory. Therefore, it is useful to recall how the anomalous conformal Ward identities determined the Wilson loop expectation value in the $(3+1)$-dimensional SYM theory. There, the dilatation generator $\mathbb{D}$ and the special conformal generator $\mathbb{K}$ were perturbatively modified by quantum corrections. The dilatation symmetry is broken by the UV regularization and its Ward identity gets anomalous. To ${\cal O}(\lambda_{\rm SYM})$,
the $(3+1)$-dimensional SYM theory exhibits
\begin{equation}
\mathbb{D}\big< W [C_n]\big> \Big\vert_{\rm SYM} = -\lambda_{\rm SYM} \left[ \sum \frac{(x_{i-1,i+1}^2 \mu^2)^\epsilon}{\epsilon} + \mathcal{O}(\epsilon^0) \right].
\end{equation}
The $\mathcal{O}(\epsilon^0)$ term refers that this Ward identity is verified up to $\epsilon^0$-order.
Using the elementary relation
\begin{equation}
\mathbb{D} \Big( (x_{i,j}^2)^\epsilon \Big) = 2\epsilon (x_{i,j}^2)^{\epsilon},
\end{equation}
we can find particular solution to the dilatational Ward identity as
\begin{equation}
\big< W [C_n]\big> \Big\vert_{\rm SYM} = \lambda_{\rm SYM} \left[ -\frac{1}{2} \sum \frac{(x_{i-1,i+1}^2 \mu^2)^\epsilon}{\epsilon^2} + \mathcal{O}(\epsilon^0) \right].
\end{equation}
Consideration of the special conformal generator $\mathbb{K}$ confirmes the result and further provides information for the ${\cal O}(\epsilon^0)$ part, so-called the BDS function, $\mbox{BDS}_n$. Homogeneous solution to the conformal Ward identities is referred as the remainder function $\mbox{Rem}_n$. It depends only on the conformal cross-ratios $u$ of the $n$-sided polygon. Putting together and replacing $\lambda_{\rm SYM}$ by $\lambda^2$, we deduce that the matter contribution in the ABJM theory takes the form
\bea
\big< W_\Box [C_n] \big>^{(2)}\Big|_{\rm matter} = \left[ - {1 \over 2} \sum_{i=1}^n {( - x^2_{i, i +2} \mu^2)^{2 \epsilon} \over (2 \epsilon)^2} + \mbox{BDS}^{(2)}_n (x) + \mbox{Rem}^{(2)}_{n, \rm SYM} (u) + {\cal O}(\epsilon) \right].
\eea
The subscript in the remainder function refers to the fact that it was deduced from the one-loop counterpart in the ${\cal N}=4$ SYM theory.
The pure Chern-Simons contribution is subject to the UV divergence. To regulate the divergence while preserving the supersymmetry, we use the dimensional reduction scheme, $d = (3 - 2 \epsilon)$. The scheme also contributes anomalies to the conformal and special conformal Ward identities. The resulting anomalous Ward identities are \cite{Plefka:2010a}
\bea
\mathbb{D}\big< W [C_n]\big> \Big\vert_{\rm CS} &=&
\Log(2) \Big(\sum_{i=1}^n 1 \Big) + {\cal O}(\epsilon) \nonumber \\
\mathbb{K}^m \big< W [C_n]\big> \Big\vert_{\rm CS} &=& -
\Log(2) \Big(\sum_{i=1}^n x_i^m \Big) + {\cal O}(\epsilon)
\eea
The full solution to these equations takes the form
\begin{equation}
\big< W_\Box [C_n]\big>^{(2)} \Big\vert_{\rm CS} = - {\Log(2) \over 2} \sum_{i=1}^n \frac{(-x_{i,i+2}^2 \mu^2)^{2\epsilon}}{2\epsilon}+\mbox{Rem}^{(2)}_{n, \rm CS}(u) +\mathcal{O}(\epsilon). \label{Wilson-anomalous}
\end{equation}
\color{black}
For the tetragon Wilson loop, $n=4$, the two-loop result was computed in \cite{Plefka:2010a}. The Chern-Simons contribution in (\ref{Wilson-anomalous}) is absorbable to the matter-dependent part by redefining the UV regularization scale $\mu$. Remarkably, the final result coincides with the one loop result in $\mathcal{N}=4$ SYM theory. Explicitly, the matter-dependent contribution and the ladder plus triple-vertex contribution take the form \cite{Bianchi:2013a}
\bea
\big< W_\Box [C_4]\big>^{(2)}\Big\vert_{\rm matter}&=&
- \frac{(-x_{13}^2 4\pi e^{\gamma_E}\mu^2)^{2\epsilon}}{(2\epsilon)^2} - \frac{(-x_{24}^2 4\pi e^{\gamma_E}\mu^2)^{2\epsilon}}{(2\epsilon)^2}
+ {1 \over 2} \Log^2 \Big(\frac{x_{13}^2}{x_{24}^2} \Big) + \mbox{Rem}_{4}^{(2)} (u) \Big\vert_{\rm matter}
\label{4pt_matter}\\
\big< W_\Box [C_4]\big>^{(2)}\Big\vert_{\rm CS}\ \ \ &=&
- {\Log(2) \over 2} \sum_{i=1}^4 \frac{(-x_{i,i+2} \pi e^{\gamma_E} \mu^2)^{2\epsilon}}{2\epsilon}
+ \mbox{Rem}_{4}^{(2)} (u) \Big\vert_{\rm CS}
\label{4pt_CS}
\eea
Hereafter, we denote $\mbox{Rem}_{n,\rm{matter}}^{(2)} (u)$ for the IR finite part of $\big< W_\Box [C_n]\big>^{(2)}_{\rm matter}$ modulo the BDS finite part. Also, $\mbox{Rem}_{n,\rm{CS}}^{(2)} (u)$ is the IR finite part of $\big< W_\Box [C_n]\big>^{(2)}_{\rm CS}$. For the tetragon Wilson loop, $n=4$, these numerical constants are given by
\bea
&& \mbox{Rem}_{4}^{(2)} (u) \Big\vert_{\rm matter} = {\pi^2 \over 4} \nonumber \\
&& \mbox{Rem}_{4}^{(2)} (u) \Big\vert_{\rm CS} = \frac{5 \pi^2}{12} -2 \mbox{Log}^2 (2).
\label{constants}
\eea
Finally, the two contributions, (\ref{4pt_matter}) and (\ref{4pt_CS}), can be combined to the following compact form
for the ABJM theory
\bea
\big< W_\Box [C_4]\big>^{(2)}_{\rm ABJM}=
-\frac{(-x_{13}^2 \tilde\mu^{2})^{2\epsilon}}{(2\epsilon)^2}
-\frac{(-x_{24}^2 \tilde\mu^{2})^{2\epsilon}}{(2\epsilon)^2}
+{1 \over 2}\mbox{Log}^2 \left(\frac{x_{13}^2}{x_{24}^2} \right)
+\mbox{Rem}^{(2)}_4 (u)
+\mathcal{O}(\epsilon). \label{4pt_ABJM}
\eea
Here, $\tilde{\mu}$ is the uniformizing UV regulator scale related to $\mu$ by
\bea
\widetilde\mu^{2}=8\pi e^{\gamma_E} \mu^2.
\label{regulator}
\eea
The remainder function $\mbox{Rem}^{(2)}_4(u)$ is
\bea
\mbox{Rem}_4^{(2)} (u) &=& \mbox{Rem}_4(u) \Big\vert_{\rm matter} + \mbox{Rem}_4 (u) \Big\vert_{\rm CS} + 5 \Log^2(2) \nonumber \\
&=& + 3 \Log^2(2) +\frac{2\pi^2}{3}.
\label{res_tet}
\eea
The last term in the first line is from the uniformization (\ref{regulator}) of the regulator scale.
The remainder function is independent of the conformal cross-ratios $u$'s, much the same way as the one-loop result in the ${\cal N}=4$ SYM theory. Moreover, it displays the uniform transcendentality property. \color{black}
\section{Hexagon Wilson Loops at Two Loops}
Our goal in this paper is to obtain the remainder function $\mbox{Rem}^{(2)}_n(u)$ in (\ref{Wilson-anomalous}) for general $n \ge 6$. For later convenience, we decompose the second-order corrections to the Wilson loop expectation value as
\bea
\big< W_\Box [C_n]\big>^{(2)}_{\rm ABJM} &=&
\left[ \big< W_\Box [C_n]\big>^{(2)}_{\rm matter}+\big< W_\Box [C_n]\big>^{(2)}_{\rm ladder}+\big< W_\Box [C_n]\big>^{(2)}_{\rm vertex}\right]_{\widetilde{\mu}}
\nonumber \\
&=& \left[ \big< W_\Box [C_n]\big>^{(1)}_{\mathcal{N} = 4 \ \rm SYM} \right]_{\rm BDS} +\mbox{Rem}^{(2)}_n \ .
\label{wilson-dec}
\eea
In the second line, we related the functional form of the ABJM Wilson loop expectation value to that of the ${\cal N}=4$ SYM Wilson loop expectation value. The BDS part is abelian, so it must be that both are the same. The remainder function is theory specific. In ABJM theory,
$\mbox{Rem}_n$ is related b
\begin{equation}
\mbox{Rem}^{(2)}_n := \mbox{Rem}^{(2)}_{n}\Big\vert_{\rm{matter}}+\mbox{Rem}^{(2)}_{n}\Big\vert_{\rm CS} + \frac{5}{4}n \Log^2(2).
\end{equation}
The last term constant originated from uniformizing the UV regulator scale as in (\ref{regulator}). The contribution ${\rm Rem}^{(2)}_{n, \rm CS}$ is computationally most complicated.
Our first task is to compute ${\rm Rem}^{(2)}_{n}(u)$ for $n=6$ analytically. For $n > 6$, we will determine ${\rm Rem}^{(2)}_n(u)$ using recursion relations that we will derive in section 7 from soft-collinear factorization of the light-like Wilson loop and analytic result for $n=6$ as an input.
\vskip0.5cm
\begin{center}
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\PhotonArc[clock](82.875,-45.688)(50.196,81.84,-11.856){4.5}{6.5}
\PhotonArc(69,7)(55.154,-67.62,-22.38){4.5}{3.5}
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\PhotonArc(208.5,142.75)(182.466,-130.499,-106.722){4.5}{6.5}
\PhotonArc(45,43)(97.949,-62.65,-40.03){4.5}{2.5}
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\PhotonArc(208.5,142.75)(182.466,-130.499,-106.722){4.5}{6.5}
\PhotonArc[clock](132,-44)(26.833,-153.435,-243.435){4.5}{3.5}
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\PhotonArc(120,10)(30.594,-168.69,-11.31){4.5}{6.5}
\Photon(120,-26)(120,-56){4.5}{3}
\end{picture}
}
\\
(a) \quad \quad \quad \quad \quad \quad \quad \quad \quad (b) \quad \quad \quad \quad \quad \quad \quad \quad \quad (c) \quad \quad \quad \quad \quad \quad \quad \quad \quad (d)
\\
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\Photon(84,-8)(150,-44){4.5}{6}
\Photon(90,4)(156,-32){4.5}{6}
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\Photon(84,-8)(150,-44){4.5}{6}
\Photon(90,4)(150,4){4.5}{5}
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\Photon(84,-8)(126,-56){4.5}{5}
\Photon(90,4)(156,-32){4.5}{6}
\end{picture}
}
\\
(e) \quad \quad \quad \quad \quad \quad \quad \quad \quad (f) \quad \quad \quad \quad \quad \quad \quad \quad \quad (g)
\\
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\Photon(84,-8)(126,-56){4.5}{5}
\Photon(90,4)(150,4){4.5}{5}
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\Photon(84,-8)(120,-56){4.5}{5}
\Photon(90,4)(138,-56){4.5}{6}
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\Photon(90,-44)(150,-44){4.5}{5}
\Photon(90,4)(150,4){4.5}{5}
\end{picture}
}
\\
(h) \quad \quad \quad \quad \quad \quad \quad \quad \quad (i) \quad \quad \quad \quad \quad \quad \quad \quad \quad (j)
\\
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\PhotonArc[clock](85,-21)(23.537,77.735,-77.735){4.5}{6.5}
\Vertex(108,-20){6}
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\PhotonArc[clock](-400.333,-331)(593.845,34.341,27.586){4.5}{7.5}
\Vertex(108,-26){6}
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(88,76) (76,-58)
\SetWidth{2.0}
\SetColor{Black}
\Line(96,16)(144,16)
\Line(96,16)(78,-20)
\Line(78,-20)(96,-56)
\Line(96,-56)(144,-56)
\Line(144,-56)(162,-20)
\Line(162,-20)(144,16)
\SetWidth{1.0}
\PhotonArc(171.857,75.571)(108.734,-138.835,-98.386){4.5}{7.5}
\Vertex(120,-20){6}
\end{picture}
}
\\
(k) \quad \quad \quad \quad \quad \quad \quad \quad \quad (l) \quad \quad \quad \quad \quad \quad \quad \quad \quad (m)
\\
\end{center}
\noindent
Figure 1 :
{\sl
The Feynman diagrams contributing to the lightlike hexagon Wilson loop consist of (a) $\sim$ (m) and cyclic permutations of the six edges. We classify them by (a) $\sim$ (d) as triple-vertex contributions, (e) $\sim$ (j) as ladder contributions, and (k) $\sim$ (m) as matter contributions.}
\rm
\vskip0.5cm
It turns out the anomalous conformal Ward identities demand that the Wilson loop expectation value must take the form Eq.(\ref{wilson-dec}). Here, we want to determine the remainder function $\mbox{Rem}_n$ in Eq.(\ref{wilson-dec}). To this end, we evaluate all contributing Feynman diagrams to two loop orders. We shall regularize the UV divergences in the dimensional regularization $d = (3 - 2 \epsilon)$ and adopt the dimensional reduction scheme DRED that treats the Levi-Civita symbol $\epsilon_{mnp}$ as 3-dimensional tensor while all others as $d$-dimensional tensors.
In Figure 1, we display the relevant diagrams. The complete list of the contributing diagrams include them and their cyclic permutations with respect to the hexagon edges. For foregoing discussions, we classify the diagrams in Figure 1 into three groups: triple-vertex diagrams for (a)-(d), ladder diagrams for (e)-(j), and matter-dependent diagrams (k)-(m). Computationally, we found that the triple-vertex diagrams the most complex. All of them involve the gauge field propagator ${(\Delta_{mn})}(x,y)$. We take the Landau gauge. In this gauge, the tree-level gauge field propagator is parity-odd and is given in position space by
\bea
{(\Delta_{mn})}^{(0)} (x,y)= {\lambda \over N} \ \mathbb{I} \otimes \mathbb{I} \ Z_{\rm o} \frac{\epsilon_{mnp} (x-y)^p}{[(x-y)^2]^{\frac{d}{2}}} \qquad
\mbox{where} \qquad Z_{\rm o} =
\pi^{(2-d)/2} \Gamma(d/2). \label{gluon_CS}
\eea
For derivation, see Appendix B. In the rest of this section, we present integral expressions of each group.
\subsection{Matter Contribution}
For the diagrams (k)-(m) in Figure 1, it suffices to first consider the self-energy of the gauge fields. At one-loop, the gauge field propagators receive corrections from vacuum polarization of matter fields. The one-loop corrected self energy is equal to the tree-level gauge field propagator in $\mathcal{N}=4$ SYM\cite{Rey:2008a, Plefka:2008a} In position space, the one-loop corrected gauge field propagator ${\Delta}^{(1)}_{mn}(x,y)$ is parity-even and takes the form
\bea
{\Delta}^{(1)}_{mn}(x,y)=- \frac{\lambda^2}{N} \ \mathbb{I} \otimes \mathbb{I} \ Z_{\rm e} \frac{g_{mn}}{((x-y)^2)^{d-2}} \qquad
\mbox{where} \qquad
Z_{\rm e} = \pi^{2-d} \Gamma^2 \Big(d/2 -1 \Big).
\label{gluonSYM}
\eea
See Appendix C for derivation and physical interpretation.
The matter contribution is computable parallel to the leading-order in the $\mathcal{N}=4$ SYM theory, except replacing the propagator with ${\Delta}^{(1)}_{mn}(x,y)$ in (\ref{gluonSYM}):
\bea
\big< W_\Box [C_n]\big>^{(2)}_{\rm matter}&=& {1 \over \lambda^2} {1 \over N} \mbox{Tr} {\cal P} \oint \rmd x_i^m \oint \rmd x_j^n \big( i^2 {\Delta}^{(1)}_{mn}(z_i,z_j)\big) \nonumber \\
&=& \Big((4\pi e^{\gamma_E})^{2\epsilon}+\frac{\pi^2}{2}\epsilon^2+\mathcal{O}(\epsilon^3) \Big) \sum_{i>j = 1}^n I_{ij} \label{matter_expression}
\eea
Here, $I_{ij}$ is the integral of one gauge boson exchange between edges $i, j$ along the contour $C_n$:
\bea
I_{ij} (x) =\int_0^1 \rmd \tau_i \int_0^1 \rmd \tau_j \frac{y_i \cdot y_j}{[(z_i-z_j)^2]^{d-2}} \qquad (i,j=1,2,\cdots,n). \label{one_loop}
\eea
It is straightforward to evaluate these integrals (\ref{one_loop}), as was done in \cite{Heslop:2007a}. Singular loci of the denominator are where the UV divergences arise and they occur precisely at the cusps, viz. when the gauge propagator connects two points on adjacent edges and approach toward the cusp in between.
For the adjacent diagrams, the integration is straightforward. The leading UV divergence is readily obtained as
\bea
I_{i+1,i} (x) &=& \int \rmd\tau_{i} \int \rmd\tau_{i+1} \ {y_i \cdot y_{i+1} \over [(z_{i+1}-z_{i})^2]^{d-2}}
= -\frac{1}{2} {(x_{i,i+2}^2)^{2\epsilon} \over (2 \epsilon)^2}.
\eea
Non-adjacent diagrams are UV finite. Summing them over all possible distinct permutations, we obtain the so-called the BDS function $\mbox{BDS}_n^{(2)}$:
\bea
\boxed{
\mbox{BDS}_n^{(2)}(x) \equiv \sum_{i>j+1}^n I_{ij} (x).
}
\eea
These integrals can be evaluated analytically, as was done in \cite{Heslop:2007a}:
\bea
I_{ij}(x) = {1 \over 2} \left[-\mbox{Li}_2(1-as)-\mbox{Li}_2(1-at)+\mbox{Li}_2(1-aP^2)+\mbox{Li}_2(1-aQ^2) \right]_{ij}.
\label{mat_a}
\eea
Here, the parameter $a$ is given by \cite{Heslop:2007a}
\bea
a=\frac{s+t-P^2-Q^2}{st-P^2 Q^2}, \quad \mbox{where} \quad P^2=x_{i,j+1}^2, \quad Q^2=x_{i+1,j}^2, \quad s=x_{i,j}^2, \quad t=x_{i+1,j+1}^2.
\label{mat_b}
\eea
Combining this with Eq.(\ref{matter_expression}), it follows that the matter contribution to the Wilson loop expectation value is given by
\bea
\boxed{
\big< W_\Box[C_n]\big>^{(2)}_{\rm matter}
= - \frac{1}{2}\sum_{i=1}^n
{(x_{i,i+2}^2 4\pi e^{\gamma_E} \mu^2)^{2\epsilon} \over (2\epsilon)^2}
+\mbox{BDS}^{(2)}_n(x)
+ \mbox{Rem}^{(2)}_ n(u)\Big\vert_{\rm matter}
+{\cal O}(\epsilon).}
\label{matterWvev}
\eea
Here, the matter contribution to the remainder function is given by
\bea
\boxed{
\mbox{Rem}_n^{(2)}(u) \Big\vert_{\rm matter} = -\frac{1}{16}n \pi^2 .
}
\label{remainder-matter}
\eea
For the special case of $n=4$, this result reproduces (\ref{4pt_matter}) and the remainder function (\ref{constants}).
\subsection{Gauge Boson Ladder diagram}
The pure Chern-Simons term generates ladder diagrams and triple-vetex diagrams. The ladder diagram contributes to the Wilson loop expectation value as
\begin{equation}
\big< W_\Box[C_n]\big>^{(2)}_{\rm ladder}= \left(\frac{\Gamma\big(\frac{d}{2}\big)}{\pi^{\frac{d-2}{2}}} \right)^2 \sum_{{\cal P}(i,j,k,l)}I_{\rm ladder}(x)
\end{equation}
Here, ${\cal P}(i,j,k,l)$ refers to sum over path-ordered, pairwise connections among the four segments $(i,j,k, l)$ and the $I_{\rm ladder}$ integral is given by
\begin{equation}
I_{\rm ladder}^{\{i,j,k,l\}}=\int \rmd \tau_i \cdots \int \rmd\tau_{l} \frac{\epsilon(y_i,y_l,z_i-z_l)}{[(z_i-z_l)^2]^{\frac{d}{2}}} \frac{\epsilon(y_j,y_k,z_j-z_k)}{[(z_j-z_k)^2]^{\frac{d}{2}}}, \label{Ladder}
\end{equation}
where the superscript $\{i,j,k,l\}$ labels the edges that the gauge field is attached. For instance, for the hexagon, the six configurations
\begin{equation}
\{i,j,k,l\} = \{4,4,1,1\} ,\{5,4,1,1\} ,\{4,3,1,1\} ,\{5,3,1,1\} ,\{3,3,1,1\} , \{5,4,2,1\}
\end{equation}
and their cyclic permutations should be summed over . Importantly, these ladder diagrams are all UV finite. Explicit form of the integrals are tabulated in Appendix E.
\subsection{Triple-Vertex Diagram}
The triple-vertex diagrams are reduced to tensor integrals involving the Levi-Civita tensor $\varepsilon_{mnp}$. We deal with such tensor integrals by reducing them to scalar integrals via the relations
\begin{equation}
I^{mnp}(x,y,z)={\partial \over \partial y^n} {\partial \over \partial z^p} I^m(y-x,z-x), \label{trick}
\end{equation}
where $I^m(y-x,z-x)$ is given by
\begin{equation}
I^m(a,b)=\int \rmd^dw \frac{w^m}{|w|^d |w-(y-x)|^d |w-(z-x)|^d} \ .
\end{equation}
Contracting the Levi-Civita tensors with the segment vectors of the polygon, one obtains integrals in readily evaluatable forms.
Triple-vertex diagram contributes to the Wilson loop expectation value as
\bea
\big< W_\Box[C_n]\big>^{(2)}_{\rm vertex}= \frac{i}{2\pi} \Big(\frac{\Gamma\big(\frac{d}{2} \big)}{\pi^{\frac{d-2}{2}}} \Big)^3 \sum_{\tiny{\mbox{path-ordered}}} I^{\{i,j,k\}}_{\rm vertex}
\eea
Here again, the superscript $\{i,j,k\}$ labels the edges where the gauge field is attached. The path ordering restricts $i>j>k$ case only. In self-explaining notation, the integral takes the form
\bea
I_{\rm vertex}^{\{i,j,k\}} (x) =\int_{\mathbb{R}^{2,1}} \rmd^d w \left[ \int \cdots \int \rmd \tau_i \rmd \tau_j \rmd \tau_k \epsilon^{abc} \frac{\epsilon(y_i, a, w-z_i) \epsilon(y_j, b, w-z_j) \epsilon(y_k, b, w-z_k)}{[(w-z_i)^2]^{\frac{d}{2}} [(w-z_j)^2]^{\frac{d}{2}}[(w-z_k)^2]^{\frac{d}{2}}} \right]. \label{Vertex}
\eea
In the case of hexagon, the four configurations
\bea
\{i,j,k\} = \{3,2,1\}, \{4,2,1\}, \{4,3,1\}, \{5,3,1\}
\eea
and their cyclic permutation generate all possible diagrams. Among them, divergence appears only through $\{3,2,1\}$-type configuration. For integral expression of the triple-vertex diagrams, see Appendix G.
Note that the triple-vertex diagrams are UV-divergent. These divergences arise from configurations whose three attached points of the gauge bosons approach a single segment. The $\{3,2,1\}$ diagram is an example of such configuration. After the Mellin-Barnes transformation, the integral $I_{\rm vertex}^{\{3,2,1\}}$ can be brought to a form that can be evaluated in part analytically and in part numerically with high precision.
The result reads
\bea
I_{\rm vertex}^{\{3,2,1\}}(x)
\frac{i\pi^{\frac{d}{2}}\Gamma(d-1)}{8\Gamma(\frac{d}{2})^3}
\Big(4\pi \Log(2) \frac{(x_{13}^2 \mu^2)^{2\epsilon}}{\epsilon} +4\pi \Log(2) \frac{(x_{24}^2 \mu^2)^{2\epsilon}}{\epsilon} + I^{\{3,2,1 \}}_{\rm finite}\Big).
\label{I_321_a}
\eea
Summing over all possible path-ordered triples $(i,j,k)$, we find that
\bea
\sum_{{\cal P}(i,j,k)} I^{\{i,j,k\}}_{\rm vertex}(x) =
\frac{i\pi^{\frac{d}{2}}\Gamma(d-1)}{8\Gamma(\frac{d}{2})^3}
\left(8\pi \Log(2) \sum_{i=1}^n \frac{(x_{i,i+2}^2 \mu^2)^{2\epsilon}}{2\epsilon} + I_{\rm finite}\right),
\label{vertex_result_a}
\eea
The leading UV-divergence is $\frac{1}{\epsilon}$, in contrast to $\frac{1}{\epsilon^2}$ leading UV-divergence in matter contribution.
\subsection{Wilson Loop of the Pure Chern-Simons Theory}
In pure Chern-Simons theory, the contribution $\big< W_\Box[C_n]\big>^{(2)}_{\rm CS}$ is obtained by combining $\big< W_\Box[C_n]\big>^{(2)}_{\rm ladder}$ and $\big< W_\Box[C_n]\big>^{(2)}_{\rm vertex}$. To evaluate these expectation values, we carry out tensor integral $\sum_{i>j>k>l} I_{i,j,k,l}^{\rm ladder}$ and $\sum_{i>j>k} I_{i,j,k}^{\rm vertex}$.
The result is
\bea
\big< W_\Box[C_n]\big>^{(2)}_{\rm CS} &=&
\big< W_\Box[C_n]\big>^{(2)}_{\rm ladder} + \big< W_\Box[C_n]\big>^{(2)}_{\rm vertex} \nonumber \\
&=& \Big(\frac{\Gamma\big(\frac{d}{2}\big)}{\pi^{\frac{d-2}{2}}} \Big)^2 \sum_{i>j=1}^n I^{i,j}_{\rm ladder} + \frac{i}{2\pi} \Big(\frac{\Gamma\big(\frac{d}{2} \big)}{\pi^{\frac{d-2}{2}}} \Big)^3 \sum_{i>j>k=1}^n I^{i,j,k}_{\rm vertex}.
\eea
Inserting (\ref{vertex_result_a}), we finally obtain
\bea
\big< W_\Box[C_n]\big>^{(2)}_{\rm CS}&= &
- \Big(\frac{\Gamma(d-1)}{2} \frac{\pi^{2-d}}{8}\Big) \Big(4\pi \log(2) \sum_{i=1}^n \frac{(x_{i,i+2}^2 \mu^2)^{2\epsilon}}{\epsilon} + I_{\rm vertex}^{\rm finite} \Big) + \Big(\frac{\Gamma\big(\frac{d}{2}\big)}{\pi^{\frac{d-2}{2}}} \Big)^2 \sum_{i>j>k>l} I^{i,j,k,l}_{\rm ladder} \nonumber \\
&=&
- \Big(\frac{\Gamma(d-1)}{2} \frac{\pi^{2-d}}{8}\Big) \Big(4\pi \log(2) \sum_{i=1}^n \frac{(x_{i,i+2}^2 \mu^2)^{2\epsilon}}{\epsilon} \Big) + I_{\rm CS} + \mathcal{O}(\epsilon) \nonumber \\
&=&
- {\log(2) \over 2} \sum_{i=1}^n \frac{(x_{i,i+2}^2 \pi e^{\gamma_E} \mu^2)^{2\epsilon}}{2\epsilon} +{\rm Rem}^{(2)}_{n, \rm CS}(u) + \mathcal{O}(\epsilon) \label{CSWvev}
\eea
In second line, we used the fact that $I_{\rm vertex}^{\rm finite}$ and $\sum_{i>j>k>l} I^{i,j,k,l}_{\rm ladder}$ are finite quantity. For convenience, we defined here
\bea
I_{\rm CS} = - \frac{1}{16\pi} I_{\rm vertex}^{\rm finite} + \frac{1}{4} \sum_{i>j>k>l} I^{i,j,k,l}_{\rm ladder}.
\eea
Explicit expansion of the last line in (\ref{CSWvev}) yields relation between $I_{\rm CS}$ and ${\rm Rem}_{n, \rm CS}(u)$:
\bea
{\rm Rem}^{(2)}_{n, \rm CS}(u) = I_{\rm CS} + \frac{n}{2} \Log(2).
\eea
We will evaluate $I_{\rm CS}$ numerically. Before proceeding, we will need to digress to general consideration of free kinematic variables in light-like polygon, viz. the moduli space of light-like polygon. For the result of the remainder function, the reader may skip to the end of section 7.
\section{Euclid, Mandelstam and Gram}
The first step in evaluating the remainder function is to specify the geometry of lightlike polygon. We shall call it the kinematics.
In this section, we present general considerations of the moduli space of a lightlike $n$-gon $C_n$.
\subsection{Moduli Space of Lightlike Polygon}
The contour $C_n$ is specified by the set of points $x_1, \cdots, x_n$. They are lightlike separated with adjacent neighbors, and can always be brought to
\bea
x_1 + \cdots + x_n = 0.
\label{translation}
\eea
by translation invariance \footnote{In other words, the center of mass of the polygon $C_n$ can always be put at the origin.}.
Equivalently, $C_n$ can be specified by the segment vectors $y_1, \cdots, y_n$. They are all light-like ($y_i^2=0$), and trivially satisfy the closedness condition
\bea
y_1 + \cdots + y_n = 0.
\label{conservation}
\eea
The two are discrete, polygon counterpart of the statement that a smooth curve can be described either by specifying position vectors of the curve or by specifying tangent vectors of the curve. Either way, one finds that the moduli space ${\cal M}[C_n]$ of $n$-sided polygon $C_n$ in $d$-dimensional embedding space is given by
\bea
\mbox{dim}{\cal M}[C_n] = (dn - n) - d - {1 \over 2} d(d-1).
\label{modulispacedim}
\eea
The dimension of the moduli space (\ref{modulispacedim}) grows linearly with $n$, the number of $x$'s or $y$'s.
For instance, consider the $n=6$ hexagon. We can specify 6 position vectors, $x_1, \cdots, x_6$ subject to (\ref{translation}). Out of $6 \times 3 =18$ components, light-like conditions $x_{i, i+1}^2 =0$ eliminates 6, (\ref{translation}) eliminates 3 and $so(2,1)$ Lorentz transformation eliminates 3. The remaining 6 independent variables are the moduli of $C_6$.
Alternatively, we can also specify 6 segment vectors $y_1, \cdots, y_6$ subject to (\ref{conservation}). Out of $6 \times 3 = 18$ components, light-like conditions $y_i^2 = 0$ eliminate 6, (\ref{conservation}) eliminates 3 and $so(2,1)$ Lorentz transformation eliminates 3. The remaining 6 independent variables are the moduli of $C_6$.
On the other hand, by the Poincar\'e invariance, the lightlike Wilson loops are not functions of $x_i$'s or $y_i$'s themselves, but are functions of the Mandelstam invariants $x_{ij}^2$, $i, j = 1, \cdots, n$. They vanish for $j = i, i \pm 1$, so the net number of nontrivial invariants is given by
\bea
\mbox{dim} M(C_n) = {1 \over 2} n (n - 3).
\label{mandelstamdimension}
\eea
Alternative choice of the Mandelstam invariants are $y_{ij}^2$. They range over $i,j = 1, \cdots, (n-1)$ because of the closedness condition (\ref{conservation}). They also vanish for $j=i$. Altogether, the net number of nontrivial invariants is given again by (\ref{mandelstamdimension}).
Their number grows quadratically with $n$, so would outgrow the dimension of $n$-gon moduli space (\ref{modulispacedim}). It must be that many of the Mandelstam invariants are redundant.
The projection of the space of Mandelstam invariants to the space of polygon moduli is achieved by the geometric condition that $n$ vectors in $d$ dimensional spacetime are necessarily linearly dependent for $n > d$. To this end, consider the Gram matrix $G$, whose $(i,j)$ entry is given by $y_i \cdot y_j$:
\bea
G \equiv M^{\rm T} \cdot M =
\begin{pmatrix}
y_1 \cdot y_1 && y_1 \cdot y_2 && y_1 \cdot y_3 && \cdots && y_1 \cdot y_n \\
y_2 \cdot y_1 && y_2 \cdot y_2 && y_2 \cdot y_3 && \cdots && y_2 \cdot y_n \\
y_3 \cdot y_1 && y_3 \cdot y_2 && y_3 \cdot y_3 && \cdots && y_3 \cdot y_n \\
\vdots && \vdots && \vdots && \ddots && \vdots \\
y_n \cdot y_1 && y_n \cdot y_2 && y_n \cdot y_3 && \cdots && y_n \cdot y_n
\end{pmatrix}
\eea
Here, $M$ is $(d \times n)$ matrix whose entries are the segment vectors
$ M = (y^m_1, y^m_2, \cdots, {y}^m_n)$.
Determinant of $G$, called Gram determinant, is nothing but the square of the hypercube volume spanned by the segment vectors:
\begin{eqnarray}
\mbox{Det} G(i, j)
= \vert\vert y_1 \wedge y_2 \wedge y_3 \wedge \cdots \wedge y_n \vert\vert^2.
\end{eqnarray}
Because of the closedness condition (\ref{conservation}), the Gram determinant vanishes identically. Moreover, $d$-dimensional spacetime accommodates at most $d$ many linearly independent vectors. Hence, in Gram matrix, determinant of any $(d+1) \times (d+1)$ sub-matrices ought to vanish identically. There are $(n-d-1)(n-d)/2$ many such choices, so these Gram sub-determinant conditions project the space of Mandelstam variables down to the space of independent scalar invariants of dimension
\bea
\mbox{dim} \Pi_G M (C_n) =
{1 \over 2} n (n - 3) - {1 \over 2} (n - d - 1) (n - d) = (d-1) n - {1 \over 2} d (d + 1).
\label{poincaredof}
\eea
This matches precisely with the dimension of the moduli space of $n$-sided lightlike polygon (\ref{modulispacedim}).
\subsection{Positivity Condition}
In evaluating the lightlike polygon Wilson loop operator expectation value , the input data of $C_n$ are the vectors $x_i$'s or $y_i$'s of the polygon. On the other hand, the expectation value is Poincar\'e invariant, so it must depend on these vectors only through scalar products:
\bea
y_i \cdot y_j = {1 \over 2} \left[ x_{i,j+1}^2+x_{i+1,j}^2-x_{i,j}^2-x_{i+1,j+1}^2 \right].
\eea
This suggests it natural to take the Mandelstam variables as input parameters. This is what we shall do for numerical computations. On the other hand, as we saw above, the Mandelstam variables are not mutually independent and need to be further supplemented by the Gram sub-determinant conditions.
A complication is that, typically, the Gram sub-determinant conditions are too involved to solve explicitly.
In evaluating the Feynman loop integrals, we shall employ the Mellin-Barnes transformations. During the evaluation, we shall provisionally assume that the Mandelstam variables are linearly independent until we perform the Mellin-Barnes transformations. We then evaluate the transformed expressions numerically, and at this stage we shall impose the Gram sub-determnant conditions by taking special kinematics of $C_n$ such that it becomes consistent with these conditions.
We found numerically that $\mbox{Rem}(u)$ yields physically meaningful values when Mandelstam variables are restricted by the Gram sub-determinant conditions and that, in solving the anomalous conformal Ward identity, the remainder function $\mbox{Rem}(u)$ is expressed in terms of cross ratios only after the Gram sub-determinant conditions are imposed to the Mandelstam variables.
Often, the Mellin-Barnes transformed integrals involve spurious poles. To avoid them, it is necessary
to impose all the Mandelstam variables to have the same sign. We shall call this condition as "positivity
condition". It turns out that, for the edge vectors $y_i$'s, the condition is satisfied by making timelike
components of adjacent edge vectors to have alternating signs. As the edge vectors are subject to the
closedness condition, this condition then implies that only even numbers of edges $n = 2\mathbb{N}$ are permissible. This purely geometric consideration imposes the polygons relevant for the lightlike ABJM Wilson
loops \footnote{The same restriction applies to all three dimensional conformal field theories.} restricted to those with even numbers of the edge. Though we do not have a fully general argument, we think that this is a general geometric condition.
To illustrate this, consider the case of hexagon. A choice of the edge vectors $y_1, \cdots, y_6$ satisfying the positivity condition and the closedness condition $y_1 + \cdots + y_6 = 0$ are
\begin{align}
\begin{array}{ccc}
y_1=(-\sqrt{a^2+b^2},a,b) & y_2=(+\sqrt{c^2+d^2},c,d) & y_3=(-\sqrt{e^2+f^2},e,f) \\
y_4=(+\sqrt{g^2+h^2},g,h) & y_5=(-\sqrt{p^2+q^2},p,q) & y_6=(+\sqrt{r^2+s^2},r,s)
\label{coordsetup}
\end{array}
\end{align}
First, we set the time-component of the edge vectors of alternating sign so that Mandelstam variables are positive. Take for example the hexagon. Among the nine Mandelstam variables, six variables($x_{13}^2, x_{24}^2, x_{35}^2$, $x_{46}^2, x_{15}^2, x_{26}^2$) are inner product of consecutive segment vector, viz. $2 y_i \cdot y_{i+1}$. Then, for example,
\begin{equation}
x_{1,3}^2 = 2 y_1 \cdot y_2 = \ 2 \sqrt{a^2+b^2} \sqrt{c^2+d^2} + 2 a c + 2 b d.
\end{equation}
By triangle inequality, sign of this Mandelstam variable is determined by the first term, regardless of signature of each parameters. To make this Mandelstam variable positive, we see that the edge vectors must be chosen to have consecutively alternating signs of their time components.
Such kinematical restrictions bear the following geometric implications to the `triple-collinear factorization' we will study in the next section. Recall that, by construction, a lightlike polygon is made of oriented edges which are all lightlike. When we take a polygon and let two non-adjacent vertices $x_i, x_j (j \ne i \pm 1)$ become lightlike, we see we can decompose the lightlike contour of the parent polygon as a sum of two lightlike contours of daughter polygons.
The absence of polygons with odd numbers of the edge also puts the constraint that the factorization must involve even number of consecutive vertices. This condition is also compatible with the requirement that the time component of edge vectors must be sign alternating. We see that such factorization gives rise to a nonlinear recursion relations among the lightlike Wilson loops.
\subsection{Moduli Space of Conformal Lightlike Polygon}
Up until now, in counting the moduli space of lightlike polygons, we only took into account the Poincar\'e symmetry of embedding spacetime. We now further endow the polygons with conformal symmetry. Replacing the Poincar\'e symmetry $so(d-1, 1)$ by the conformal symmetry $so(d, 2)$, we see that the dimension of moduli space of conformal lightlike polygons is modified to
\bea
\mbox{dim} {\cal M}_{\rm c} [C_n] = (d -1) n - {1 \over 2} (d+1) (d+2).
\label{dim-conformal-polygon}
\eea
On the other hand, we elaborated in the previous section that the geometry of lightlike polygons is more conveniently described in terms of Mandelstam variables but these variables are not mutually independent. The requisite projection of the Mandelstam variables is the Gram condition. Below, we explain how this can be achieved.
The dimension of parameter space of conformally invariant Mandelstam variables, viz. the conformal cross-ratios
\bea
u_{i,j} = \frac{x_{i,j+1}^2 x_{i+1,j}^2}{x_{i,j}^2 x_{i+1,j+1}^2},
\eea
is given by
\bea
\mbox{dim} M_c [C_n] = {1 \over 2} n (n-1) - n - n = {1 \over 2} n (n - 5),
\label{dim-crossratio}
\eea
\begin{center}
\fcolorbox{white}{white}{
\begin{picture}(140,130) (69,-27)
\SetWidth{2.0}
\SetColor{Black}
\Line(90,64)(162,76)
\SetWidth{1.0}
\Line[dash,dashsize=10](90,64)(90,-8)
\Line[dash,dashsize=10](90,-8)(162,76)
\Line[dash,dashsize=10](162,76)(162,-20)
\Line[dash,dashsize=10](162,-20)(90,64)
\SetWidth{2.0}
\Line(90,-8)(162,-20)
\Text(78,-20)[lb]{\Large{\Black{$x_j$}}}
\Text(168,-32)[lb]{\Large{\Black{$x_{j+1}$}}}
\Text(78,70)[lb]{\Large{\Black{$x_i$}}}
\Text(66,28)[lb]{\Large{\Black{$x_{i,j}^2$}}}
\Text(168,82)[lb]{\Large{\Black{$x_{i+1}$}}}
\Text(174,28)[lb]{\Large{\Black{$x_{i+1,j+1}^2$}}}
\Text(138,4)[lb]{\Large{\Black{$x_{i,j+1}^2$}}}
\Text(138,46)[lb]{\Large{\Black{$x_{i+1,j}^2$}}}
\end{picture}
}
\quad \quad
\fcolorbox{white}{white}{
\begin{picture}(140,130) (69,-27)
\SetWidth{2.0}
\SetColor{Black}
\Line(90,64)(162,76)
\SetWidth{1.0}
\Line[dash,dashsize=10](90,-8)(162,76)
\Line[dash,dashsize=10](162,76)(162,-20)
\Line[dash,dashsize=10](162,-20)(90,64)
\SetWidth{2.0}
\Line(90,-8)(162,-20)
\Text(78,-20)[lb]{\Large{\Black{$x_j$}}}
\Text(168,-32)[lb]{\Large{\Black{$x_{j+1}$}}}
\Text(78,70)[lb]{\Large{\Black{$x_i$}}}
\Text(66,28)[lb]{\Large{\Black{$x_{i,j}^2$}}}
\Text(168,82)[lb]{\Large{\Black{$x_{i+1}$}}}
\Text(174,28)[lb]{\Large{\Black{$x_{i+1,j+1}^2$}}}
\Text(138,4)[lb]{\Large{\Black{$x_{i,j+1}^2$}}}
\Text(138,46)[lb]{\Large{\Black{$x_{i+1,j}^2$}}}
\Line(90,64)(90,-8)
\end{picture}
\label{figure_cross}
}
\\
(a) \hskip7cm (b)
\end{center}
Figure 2: {\sl Cross-ratios or anharmonic ratios are conformally invariant Mandelstam variables of lightlike polygon.}
\vskip0.5cm
The counting is simple. To construct a cross-ratio, we need two distinct edges as in Figure 2. There are $n(n-1)/2$ possible pairs of edges. However, the resulting cross-ratio vanishes if the two edges chosen are nearest neighbors or next-nearest neighbors.
We are again in a situation that, in a given spacetime dimension, the dimension of the moduli space of conformal cross-ratios (\ref{dim-crossratio}) outgrows dimension of moduli space of conformal lightlike polygon (\ref{dim-conformal-polygon}) we want to describe. The requisite projection to the conformal cross-ratios is achieved by the Gram condition modulo conformal equivalence relations.
Let's be more explicit. A conformal covariant vector $x^m$ in $\mathbb{R}^{d-1, 1}$ can be equivalently described by projection of a vector $X^A = (X_1, X_0, X_1, \cdots, X_d)$ in embedding Minkowski space $\mathbb{R}^{d, 2}$ onto the lightlike hyperboloid:
\bea
\eta_{AB} X^A X^B = - (X_{-1})^2 - (X_0)^2 + (X_1)^2 + \cdots + (X_d)^2 = \eta_{mn} X^m X^n - 2 X^+ X^- = 0 .
\label{hyperboloid}
\eea
Choosing the lightcone coordinates $X^\pm = (X_{-1} \pm X_d)/\sqrt{2}$ are the lightcone coordinates, the vector $x^m$ is projectively obtained by
\bea
x^m = {X^m \over X^+}.
\eea
The action of the conformal group $SO(d, 2)$ to the vector $x^m$ is equivalent to the action of linear transformations acting on $X^A$ lying on the lightlike hyperboloid (\ref{hyperboloid}). It is known that the space of $x$-vectors is $\mathbb{R}^{d-1, 1}$ provided the $SO(d,2)$ is gauge-fixed to $X^+ = 1$. In this gauge,
\bea
X_{ij}^2 = - 2 X_i \cdot X_j = x^2_{ij}.
\eea
From this, it also follows that
\bea
Y_i := (X_{i+1} = X_i) = (y_i, 0, Y^-_i) \qquad (i=1, \cdots, n)
\eea
are lightlike in $\mathbb{R}^{d,2}$. We thus associated the conformal edge vectors $y_i$'s of a conformal lightlike polygon in the physical spacetime $\mathbb{R}^{d-1,1}$ with the edge vectors $Y_i$'s of a lightlike polygon in the embedding space $\mathbb{R}^{d,2}$. This then implies that the space of conformal cross-ratios in $\mathbb{R}^{d-1, 1}$ is the same as the space of Mandelstam variables in $\mathbb{R}^{d,2}$. Therefore, the dimension of the moduli space of conformal cross-ratios is given by
\bea
\mbox{dim} M_{\rm c} (C_n) = {1 \over 2} n (n - 1) - n - n
= {1 \over 2} n (n - 5).
\eea
We subtracted $n$ for choosing adjacent edge pairs, and $n$ for choosing next-adjacent edge pairs.
How do we match this moduli space to the moduli space of conformal lightlike polygons? The idea is that the Gram sub-determinants of the vectors project the cross-ratios down to the space of independent ones. The Gram determinant in the embedding space is now given by
\bea
G_c = M_c^T \cdot M_c =
\begin{pmatrix}
Y_1 \cdot Y_1 && Y_1 \cdot Y_2 && \cdots && Y_1 \cdot Y_n \\
Y_2 \cdot Y_1 && Y_2 \cdot Y_2 && \cdots && Y_2 \cdot Y_n \\
\vdots && \vdots&& \ddots && \vdots \\
Y_n \cdot Y_1 && Y_n \cdot Y_2 && \cdots && Y_n \cdot Y_n
\end{pmatrix}
\eea
Because of the closedness condition, the Gram determinant itself vanishes identically. Since the embedding space $\mathbb{R}^{d,2}$ accommodates at most $(d+2)$ many linearly independent vectors, Therefore, there are $(n - (d + 1)) (n - (d +1) -1)/2$ many Gram sub-determinant conditions. Therefore, the dimension of conformal cross-ratios is
\bea
\mbox{dim} \Pi_G M_c [C_n] = {1 \over 2} n (n - 5) - {1 \over 2} (n - d - 1)(n - d - 2) = (d-1) n - {1 \over 2} (d+1) (d+2).
\eea
This matches precisely the dimension of moduli space of conformal lightlike polygons (\ref{dim-conformal-polygon}).
\section{The Hexagon Remainder Function}
In this section, we compute the remainder function $\mbox{Rem}^{(2)}_{6, \rm CS}$, relevant for the hexagon ABJM Wilson loop expectation value. We explained in section 3 that this computation involves multi-dimensional scalar integrals. In this section, we compute them.
We expect from the anomalous conformal Ward identity that the remainder function $\mbox{Rem}_{n, \rm CS}$ depends only on the conformal cross-ratios. In setting up the computation, we can readily verify this property of the remainder function by varying shapes, equivalently, Mandelstam variables of the lightlike polygon. Not all the Mandelstam variables are independent and, as we explained in section 4, it is necessary to impose the Gram sub-determinant conditions. This condition turns out a stark difference from what were known for extracting the remainder function in the four-dimensional ${\cal N}=4$ SYM. In section 5.1, we recall this situation in detail.
For the computation of multi-dimensional integrals, we utilize public packages. The scalar integrals we need to compute span up to 8-dimensional complex integrations. The traditional {\ttfamily{MB}} package \cite{Czakon:2005a} turns out not powerful enough to render the result with requisite numerical precisions. Instead, we utilize the package {\ttfamily{FIESTA2}}, \cite{Smirnov:2009a}. In the following subsections, we present details of the computation. In section 5.2, we present numerical computations performed using the {\ttfamily{FIESTA2}} package. In section 5.3, for the special shapes of the hexagon discussed in the previous section, we reduce our multi-dimensional integrals to lower-dimensional integrals. The reduction facilitates to achieve high precision to the numerical computations. In section 5.4, we utilize the PSLQ algorithm and infer analytic expressions of the $\mbox{Rem}_{6, \rm CS}$ from the numerical results.
\subsection{Remainder Function in ${\cal N}=4$ Super Yang-Mills Theory}
In section 4, we explained that the Mandelstam variables are the Lorentz scalars convenient for specifying the geometry of lightlike
polygon, they need to be further projected down to the space of conformal cross-ratios since they are not mutually independent. We alluded that such projection is achieved by the Gram sub-determinant conditions. Therefore, in the numerical computation in this section of the hexagon remainder function, we shall cover the moduli space of the lightlike hexagon by varying the Mandelstam invariants over the subspace that the Gram sub-determinant conditions are satisfied.
While our prescription is the most natural steps to take, this was not what was practiced when the hexagon remainder function was computed in the (3+1)-dimensional $\mathcal{N}=4$ SYM theory. There, the anomalous conformal Ward identities also put the remainder function to be a function of conformal cross-ratios. The lightlike Wilson loops were again specified by Mandelstam invariants. Remarkably, it was observed that two sets of Mandelstam invariants, one obeying the Gram sub-determinant conditions and another not, yielded an identical result for the remainder function. In so far as the cross-ratios are the same, any choice of the Mandelstam variable set is allowed regardless of solving the Gram sub-determinant conditions. Indeed, this explains why lightlike Wilson loops with odd numbers of edges are admissible configurations in the $\mathcal{N}=4$ SYM theory.
As the Mandelstam variables can be chosen freely thus they can be taken `unphysical' values outside the moduli space of the hexagon, wide variety of kinematic limits become available in so far as evaluation of the remainder function is concerned. In the $(3+1)$-dimensional ${\cal N}=4$ SYM theory, this freedom was maximally taken into advantage. A particularly useful limit was the quasi multi-Regge kinematics (QMRK), since this kinematics enabled determination of the hexagon remainder function and understanding its analytic structure. For $(2+1)$-dimensional ABJM theory, we concluded in section 4 that such kinematic limits are not available and we should impose the Gram sub-determinant conditions throughout.
The Gram sub-determinant conditions essential for the ABJM theory bears further impact. In the the $(3+1)$-dimensional ${\cal N}=4$ SYM theory, another useful kinematic limit was to take the lightlike polygon to $(1+1)$-dimensional subspace. This limit brought in enormous simplification and facilitated computation of the remainder function analytically tractable. Unfortunately, for kinematical reasons again, this limit is also not available for $(2+1)$-dimensional ABJM theory. This is because the $(1+1)$-dimensional kinematics cruially relies on the positivity condition and the closedness of edge vectors. Take for instance the lightlike hexagon and restrict it to the $(1+1)$-dimensional lightlike basis, $(1,-1)$ and $(1,1)$. The 6 edge vectors obeying the positivity condition are parametrized as
\bea
y_1 = (a,-a), \quad y_2 = (b,b), \quad y_3 = (c,-c), \quad y_4 = (d,d), \quad y_5 = (e,-e), \quad y_6 = (f, f),
\eea
where $a,b,c,d,e,f$ are restricted to be positive. To obey the closedness, both $a+b+c+d+e+f = 0$ and $a-b+c-d+e-f = 0$ should be satisfied. We see that these conditions cannot be met, since the positivity of $a,b,c,d,e,f$ violates first equation. Therefore, $(1+1)$-dimensional lightlike condition, positivity condition and closedness are not mutually compatible.
\subsection{Scalar Invariants and Gram Sub-Determinant Conditions}
Here, we first study how the hexagon remainder function depends on the Mandelstam variables and the Gram sub-determinant conditions. We shall find that the dependence in the $(2+1)$-dimensional ABJM theory is very different from the dependence in the $(3+1)$-dimensional ${\cal N}=4$ SYM theory.
We computed numerically both the triple-vertex diagrams and the ladder diagrams listed in Figure 2. Adding them, we obtained the hexagon remainder function at two loops, $\mbox{Rem}^{(2)}_{6, \rm CS}$ as a function of 9 Mandelstam variables of the hexagon.
\setlength{\tabcolsep}{3pt}
\setlength{\extrarowheight}{1.5pt}
{
\begin{table}[ht]
\centering
\begin{tabular}{|c||ccccccccc|c|}
\hline
& $x_{13}^2$ & $x_{24}^2$ & $x_{35}^2$ & $x_{46}^2$ & $x_{15}^2$ & $x_{26}^2$ & $x_{14}^2$ & $x_{25}^2$ & $x_{36}^2$ & Rem$^{(2)}_{6, \rm CS}$ \\ \hline \hline
A &$ -1.0000$ & $-1.0000$ & $-1.0000$ & $-1.0000$ & $-1.0000$ & $-1.0000$ & $-1.0000$ & $-1.0000$ & $-1.0000$ & $-3.47537352$ \\ \hline
B & $-6.8764$ & $-18.194$ & $-21.887$ & $-77.498$ & $-48.781$ & $-14.780$ & $-24.467$ & $-30.720$ & $-3.3327$ & $-3.47342610$ \\ \hline
C & $-4.8757$ & $-11.282$ & $-6.1981$ & $-42.828$ & $-19.339$ & $-8.1903$ & $-15.616$ & $-10.007$ & $-2.5719$ & $-3.47622947$ \\ \hline
D & $-3.5979$ & $-7.3282$ & $-1.4275$ & $-24.543$ & $-7.9792$ & $-4.5361$ & $-10.424$ & $-2.6875$ & $-1.9989$ & $-3.47688979$ \\ \hline
E & $-116.29$ & $-4.0000$ & $-116.29$ & $-2.0350$ & $-4.0000$ & $-2.0350$ & $-4.0000$ & $-4.0000$ & $-59.160$ & $-3.48197748$ \\ \hline
F & $-4.0000$ & $-2.3528$ & $-9.0000$ & $-1.0000$ & $-1.3057$ & $-1.0000$ & $-1.0000$ & $-2.2500$ & $-3.6892$ & $-3.47579959$ \\ \hline
G & $-4.0000$ & $-1.0000$ & $-8.8965$ & $-4.4482$ & $-1.0000$ & $-2.0000$ & $-1.0000$ & $-1.0000$ & $-5.5504$ & $-3.47576202$ \\ \hline
H & $-1.2027$ & $-2.5332$ & $-2.0000$ & $-3.0000$ & $-6.2344$ & $-13.512$ & $-2.1782$ & $-3.6253$ & $-0.82827$ & $-3.47561202$ \\ \hline
\end{tabular}
\label{table1}
\caption{\sl Results of $R_{CS,6}$ for eight configurations of hexagon's Mandelstam variables. It suggests that R$^{(2)}_{6, \rm CS}$ takes a constant value over wide ranges of the conformal cross-ratios.}
\end{table}
}
\vskip0.5cm
In Table 1, we generated eight configurations (A)$\sim$(H) of the 9 Mandelstam variables $x_{13}^2, \cdots, x_{36}^2$, subject to the Gram sub-determinant conditions. Equivalently, these configurations are generated by lightlike segment vectors $y_1, \cdots, y_6$ subject to the $SO(3,2)$ conformal invariance. The results indicates that the hexagon remainder function Rem$^{(2)}_{6, \rm CS}$ is a constant number, independent of the Mandelstam variables and hence the conformal cross-ratios.
To test neessity of the Gram sub-determinant conditions, we chose a configuration, say (D), and permuted subset of the nine Mandelstam variables while keeping their conformal cross-ratios fixed. Obviously, permuting the Mandelstam variables so violates the Gram sub-determinant conditions. We computed the hexagon remainder function $\mbox{Rem}^{(2)}_{6, \rm CS}$ and the results are tabulated in Table 2.
\setlength{\tabcolsep}{3pt}
\setlength{\extrarowheight}{1.5pt}
{
\begin{table}[ht]
\centering
\begin{tabular}{|c||ccccccccc|c|}
\hline
& $x_{13}^2$ & $x_{24}^2 $ & $x_{35}^2$ & $x_{46}^2$ & $x_{15}^2$ & $x_{26}^2$ & $x_{14}^2$ & $x_{25}^2$ & $x_{36}^2$ & $\mbox{Rem}_{6, \rm CS}^{(2)}$ \\ \hline \hline
D1 & $-3.5979$ & $-7.9792$ & $-1.4274$ & $-24.543$ & $-7.3282$ &
$-4.5361$ & $-10.424$ & $-2.6875$ & $-1.9989$ & $-3.70845563$ \\ \hline
D2 & $-3.5979$ & $-7.9792$ & $-4.5361$ & $-24.543$ & $-7.3282$ & $-14.780$ & $-10.424$ & $-2.6875$ & $-1.9989$ & $-3.99210938$ \\ \hline
\end{tabular}
\caption{\sl We examined whether $\mbox{Rem}^{(2)}_{6, \rm CS}$ maintain the same values for the above Mandelstam variables. They all have same conformal cross ratios. One remarkable observation on the remainder function in $\mathcal{N}=4$ SYM was that it has the same value for all Mandelstam variables so long as their conformal cross ratios are the same, even if the Gram sub-determinant conditions were not satisfied. This result suggests that such feature no long holds in the ABJM theory.}
\end{table}
}
\vskip0.5cm
In Table 2, we generated configurations (D1) and (D2) that have the same conformal ratios as (D) but violates the Gram sub-determinant condition \footnote{This is equivalent to saying that there is no suitable choice of $x_i$'s or $y_i$'s vectors which generate (D1) and (D2) configurations.}. We observe that the remainder function at (D),(D1),(D2) do not agree one another even though all three sets have one and the same conformal cross-ratio. We thus conclude that, in stark contrast to the $(3+1)$-dimensional ${\cal N}=4$ SYM theory, the lightlike hexagon Wilson loop expectation value in the ABJM theory is consistent with the anomalous conformal Ward identity only if the Mandelstam variables were to satisfy the Gram sub-determinant conditions. Therefore, if two sets of the hexagon Mandelstam variables satisfy the Gram sub-determinant conditions and yield the same conformal cross-ratio, then their values of the remainder function should be the same. In our numerical computations, we have confirmed this.
\setlength{\tabcolsep}{3pt}
\setlength{\extrarowheight}{1.5pt}
{
\begin{table}[ht]
\centering
\begin{tabular}{|c||ccccccccc|c|}
\hline
& $x_{13}^2$ & $x_{24}^2$ & $x_{35}^2$ & $x_{46}^2$ & $x_{15}^2$ & $ x_{26}^2$ & $ x_{14}^2$ & $ x_{25}^2 $ & $ x_{36}^2$ & $\mbox{Rem}^{(2)}_{6, \rm CS}$ \\ \hline \hline
X & $-6.0000$ & $-2.0000$ & $-3.0000$ & $-9.0000$ & $-5.0000$ & $-7.0000$ & $-1.0000$ & $-4.0000$ & $-8.0000$ & $-3.99713002$ \\ \hline
Y & $-1.0000$ & $-5.0000$ & $-\frac{3}{16}$ & $-27.000$ & $-1.0000$ & $-7.0000$ & $-2.0000$ & $-1.0000$ & $-2.0000$ & $-3.84236164$ \\ \hline
Z & $-1.0000$ & $-\frac{1}{3}$ & $-1.0000$ & $-1.0000$ & $-2.0000$ & $-\frac{2}{3}$ & $-1.0000$ & $-\frac{2}{3}$ & $-1.0000$ & $-3.41789832$
\\ \hline
\end{tabular}
\caption{\sl Three random choices of the Mandelstam variables that do not obey the Gram sub-determinant conditions. The values of the remainder function $R_{CS,6}$ do not agrees with the values in Table 1 for `physical' Mandelstam variables.}
\end{table}
}
As another check, in Table 3, we considered randomly chosen configuration (X) and another configuration (Y) having the same conformal cross-ratios as (X). The two configurations yield different values for the remainder function. This affirms that configurations violating the Gram sub-determinant condition do not obey the anomalous conformal Ward identitie since these identities put the remainder function to a function only of conformal cross-ratios. As such, we call them 'unphysical' configurations. We also considered the configuration (Z) whose cross-ratios all have value 1 and hence relevant for the $(1+1)$-dimensional configuration of the hexagon. Result (Z), however, shows that it does not yield the physical result, because the closedness, $(1+1)$-dimensional lightlikeness, and the positivity conditions cannot be satisfied simultaneously.
Summarizing, we learned that although the polygon kinematics is most conveniently described in terms of the Mandelstam invariants, they are subject to various restrictions to correspond to physical configurations. Some of these restrictions are universal, independent of spacetime dimensions, while some other restrictions are specific to $(2+1)$-dimensional spacetime. Unfortunately, the latter restrictions were stringent enough not to allow the QMRK that played powerful role in understanding the analytic structure of the Wilson loop expectation values in the $(3+1)$-dimensional $\mathcal{N}=4$ SYM theory.
To avoid such difficulty, we identified alternative special kinematics that satisfy the Gram sub-determinant conditions and also permit continuous deformation within the moduli space of the lightlike polygon. The idea is to take the deformation parameters to asymptotic limit and reduce Mellin-Barnes integrals as simple as possible. We shall study these kinematic limits in the next sections. For now, we present numerical result for several configurations that turn out representative of 1- and 2-parameter subspaces.
\newline
\setlength{\tabcolsep}{3pt}
\setlength{\extrarowheight}{1.5pt}
{
\begin{table}[ht]
\centering
\begin{tabular}{|c||ccccccccc|c|}
\hline
& $x_{13}^2$ & $x_{24}^2$ & $x_{35}^2$ & $x_{46}^2$ & $x_{15}^2$ & $x_{26}^2$ & $x_{14}^2$ & $x_{25}^2$ & $x_{36}^2$ & $\mbox{Rem}^{(2)}_{6, \rm CS}$ \\ \hline \hline
J1 & $-100.00$ & $-1.0000$ & $-1.0000$ & $-\frac{1}{100}$ & $-1.0000$ & $-1.0000$ & $-1.0000$ & $-1.0000$ & $-1.0000$ & $-3.4857518$ \\ \hline
J2 & $-100.00$ & $-2000.0$ & $-100.00$ & $-100.00$ & $-5.0000$ &
$-100.00$ & $-100.00$ & $-100.00$ & $-100.00$ & $-3.4778556$ \\ \hline
\end{tabular}
\caption{\sl Mandelstam variable choices satisfying the Gram sub-determinant conditions. We checked (J1) and (J2) could be generated from suitably chosen edge vectors $y_i$'s. Since we want to take asymptotic limit while maintaining the Gram sub-determinant conditions, suitable special kinematics were considered. These results provide numerical evidence that both one-parameter family and two parameter family indeed yield satisfactory results for the remainder function. See Table 1 for comparison.}
\end{table}
}
\noindent
The configuration (J1) belongs to 1-parameter group, while (J2) belongs to 2-parameter group. We examined numerically the effect of changing these free parameters. As seen in Table 4, the remainder function $\mbox{Rem}^{(2)}_{6, \rm CS}$ takes a constant value over the ranges we changed these parameters. The result hints that we can take certain asymptotic limits of these moduli parameters and simplify the Mellin-Barnes transformation integrals.
\subsection{Special Shapes and Asymptotic Limits}
\subsubsection{The strategy}
Our goal is to compute the hexagon remainder function with high precision and infer from it analytic result. In the previous subsection, we presented the remainder function computed using the package {\ttfamily{FIESTA2}}. The numerical error is rather large, ${\cal O}(10^{-2})$. Here, we propose an alternative strategy for computing the remainder function with better numerical precision than {\ttfamily{FIESTA2}}. We begin with the Mellin-Barns transformation to our 2-loop integrals, for which we used the Mathematica package {\ttfamily{MB}}. The problem of this transformation is that it results in multi-dimensional scalar integrals, for which numerical precision is difficult to attain. The idea is to lower the dimension of numerical integral maximally so that higher numerical precision can be achieved. The way we achieve this is as follows. Recall that the dimension of numerical integral is closely related to the number of independent terms inside the denominator $\Delta_y$ in the two-loop integral involving the gauge boson triple-vertex diagrams (See Appendix G). By choosing judiciously a set of the Mandelstam variables that satisfy the Gram sub-determinant conditions and that reduce the number of terms in $\Delta_y$, we can bring down the dimension of numerical integrals and obtain the result with high numerical precisions. Below, we explain how we performed high precision numerical computation for the gauge boson triple-vertex diagrams. The integrals are relegated in the Appendix G. Attentive readers may skip technical details and go directly to the final results (\ref{res_hex}) and (\ref{result_hex}).
\subsubsection{Computational Details}
Our strategy for the numerical computation is as follows. We apply the Mellin-Barnes transformation to every loop integrals resulting from the gauge boson triple-vertex and gauge boson ladder diagrams derived in section 3. We then take special limits of the polygon shape deformed by one- or two-moduli parameters. The integrals are defined in the complex domain. Utilizing the packages {\ttfamily{MB}} and {\ttfamily{MBresolve}}\cite{Smirnov:2009b}, we resolve singularity structure of each complex integrals. We then apply the Barnes lemma to reduce the integrals to lower-dimensional integrals. We made this procedure automatic using the package {\ttfamily{barnesroutines}}\cite{Kosower}. Next, we apply the package {\ttfamily{MBasymptotics}}\cite{Czakon} to the chosen moduli parameters and obtain simpler expressions for the integrals. We find that these expressions are reducible to at most three-dimensional complex integrals. Finally, we evaluate them using the {\ttfamily{MB}} package and obtain numerical result with high precision.
What special limits can we choose for the Mandelstam variables of the lightlike hexagon? Subject to the Gram sub-determinant conditions, let's consider the following two special limits: the first one has 1-moduli parameter, while second one has 2-moduli parameters.
\begin{itemize}
\item
{\bf one-parameter hexagon}
\setlength{\tabcolsep}{14.5pt}
\setlength{\extrarowheight}{1.5pt}
{
\begin{table}[ht]
\centering
\begin{tabular}{|ccccccccc|}
\hline
$x_{13}^2$ & $x_{24}^2$ & $x_{35}^2$ & $x_{46}^2$ & $x_{15}^2$ & $x_{26}^2$ & $x_{14}^2$ & $x_{25}^2$ & $x_{36}^2$ \\ \hline \hline
$- e^a$ & $-1$ & $- 1$ & $- e^{-a}$ & $- 1$ & $- 1$ & $- 1$ & $-1$ & $- 1$ \\ \hline
\end{tabular}
\end{table}
}
\newline
The moduli parameter $a$ ranges over $- \infty < a < + \infty$. We take the configuration that obeys the positivity condition, and this puts all the Mandelstam variables to positive definite values.
\item {\bf two-parameter hexagon}
\setlength{\tabcolsep}{12pt}
\setlength{\extrarowheight}{1.5pt}
{
\begin{table}[ht]
\centering
\begin{tabular}{|ccccccccc|}
\hline
$x_{13}^2$ & $x_{24}^2$ & $x_{35}^2$ & $x_{46}^2$ & $x_{15}^2$ & $x_{26}^2$ & $x_{14}^2$ & $x_{25}^2$ & $x_{36}^2$ \\ \hline \hline
$-e^{\alpha}$ & $-e^{\alpha + \beta}$ & $-e^\alpha$ & $-e^\alpha$ & $-e^{\alpha - \beta}$ & $-e^\alpha$ & $-e^\alpha$ & $-e^\alpha$ & $-e^\alpha$
\\ \hline
\end{tabular}
\end{table}
}
\newline
The moduli parameters $\alpha, \beta$ range over $-\infty < \alpha, \beta < + \infty$. Again, taking the configuration obeying positivity condition, all the Mandelstam variables are positive definite.
\end{itemize}
We found that the hexagon remainder function $\mbox{Rem}^{(2)}_{6, \rm CS}$ remains constant-valued for a wide range of the moduli parameters $a, \alpha, \beta$. In the previous subsection, we already presented one such choice in the result for the configurations (J1) and (J2) in the previous subsection. The result suggested that the hexagon remainder function $\mbox{Rem}^{(2)}_{6, \rm CS}$ is indeed a constant up to two loops in the ABJM theory. We performed numerical computation for both configurations and found that the two-parameter configuration yields the result with better numerical precision. Hereafter, we will exclusively discuss the two-parameter configuration results. The simplest integral is $I_{321}$ in (\ref{321_delta}). Inserting the two-moduli parameter contour to (\ref{321_delta}), we observe that the four-fold integration is reduced to three-fold integration. For instance, the denominator is reduced to
\begin{equation}
\Delta_y \Big\vert_{\mbox{2-parameter}}\quad \longrightarrow \quad e^\alpha \cdot x\bar{x}y\bar{s_1}s_2+e^{\alpha - \beta} \cdot \bar{x}\bar{y}\bar{s_2}s_3+e^{\alpha - \beta} \cdot x\bar{y}s_1s_3+ e^\alpha \cdot x\bar{y}\bar{s_1}.
\end{equation}
By itself, five terms in the denominator $\Delta_y$ are reduced to three terms, so the two-parameter configuration does not appear to simplify the multi-dimensional integrals considerably. It turned out the two-parameter configuration is more effective for other triple-vertex diagrams involving higher-dimensional integrals. The most complicated integrals resulted from the contribution $I_{531}$. The Mellin-Barns transformation of this contribution yielded 8-dimensional complex integrals. With the two-parameter special kinematics, we were able to reduce these integrals to five-dimensional integrals. We could do even better. By taking the asymptotic limits for $\alpha, \beta$ sequentially,
\bea
\alpha \rightarrow - \infty \quad \mbox{then} \quad \beta \longrightarrow +\infty.
\eea
we were able to reduce the five-dimensional integrals down to at most three-dimensional integrals.
\subsubsection{Result}
The high precision computation yielded
\begin{align}
\mbox{Rem}^{(2)}_{6, \rm CS} = - 3.470168804. \quad \Big(0.000489814\Big).
\end{align}
Utilized the PSLQ algorithm, we converted thi to an analytic expression. The result is
\begin{equation}
\mbox{Rem}^{(2)}_{6, \rm CS}= -\frac{17}{4} \zeta(2)+ 3\mbox{Log}(2)+ 3\mbox{Log}^2(2). \label{res_hex}
\end{equation}
Numerical value of the right-hand side is $-3.470169200670522$, and this agrees to our numerical result $-3.47016880435048$ within the ${\cal O}(10^{-6})$ precision.
The final result for the two-loop, lightlike hexagon ABJM Wilson loop expecation value is obtained by combining the purely abelian, matter-dependent contribution (\ref{matterWvev}) and the pure Chern-Simons contribution (\ref{CSWvev}) for $n=6$. It takes the form:
\bea
\big< W_\Box [C_6]\big>^{(2)}_{\rm ABJM}&=& -
\Big[\frac{1}{2} \mbox{Log}(2) \sum_{i=1}^6 \frac{(x_{i,i+2}^2 \pi e^{\gamma_E} \mu^2)^{2\epsilon}}{2\epsilon} +\mbox{Rem}^{(2)}_{6, \rm CS} -3 \mbox{Log}(2) \nonumber \\
&& \hskip0.1cm + {1 \over 2}\sum_{i=1}^n \frac{(x_{i,i+2}^2 4\pi e^{\gamma_E} \mu^2)^{2\epsilon}}{(2\epsilon)^2}-\mbox{BDS}^{(2)}_6 +\frac{3}{8} \pi^2 \Big] \nonumber \\
&=& - \Big[ {1 \over 2} \sum_{i=1}^6 \frac{(x_{i,i+2}^2 8 \pi e^{\gamma_E} \mu^2)^{2\epsilon}}{(2\epsilon)^2}-\mbox{BDS}^{(2)}_6 + \mbox{Rem}^{(2)}_{6, \rm CS}-3 \mbox{Log}(2)+\frac{3 \pi^2}{8} -\frac{15}{2} \mbox{Log}^2(2) \Big] \nonumber \\
&=& - {1 \over 2} \sum_{i=1}^6 \frac{(x_{i,i+2}^2 8 \pi e^{\gamma_E} \tilde\mu^2)^{2\epsilon}}{(2\epsilon)^2} +\mbox{BDS}^{(2)}_6 + \Big(\frac{9}{2} \mbox{Log}^2(2) + \frac{ \pi^2}{3}\Big), \label{result_hex}
\eea
where the BDS contribution $\mbox{BDS}^{(2)}_6$ is already known. This is one of the main results of this paper. Like the lightlike tetragon Wilson loop expectation value, the UV finite part in (\ref{result_hex}) exhibits the uniform transcendentality.
While we have succeeded in obtaining two-loop analytic result for the hexagon Wilson loop expectation value, we have yet no clue for the structure of the remainder function $\mbox{Rem}^{(2)}_{n, \rm CS}$ for polygons of $n\ge 8$. To crack down its structure, we will need to understand further configurational structures of the lightlike polygon Wilson loop expectation value. This is what we will undertake in the next section.
\section{Lightlike Factorization and Antenna Function}
Conformal field theories are subject to infrared divergences due to collinear and soft bremsstrahlung partons. These divergences then allow universal factorization and scaling behavior of physical processes. A class of such processes is the parton scattering amplitudes in gauge theories. The universal factorization and scaling behavior allowed accurate prediction at fixed order perturbation theory and resummation of dominant logarithms.
Our goal in this section is to demonstrate that universal factorization and scaling behavior are also present in the lightlike polygon Wilson loops. We then introduce `universal antenna function' for a certain limit of the polygon shape, which we will utilize it in the next section to solve for the ABJM Wilson loop expectation value for arbitrary $n$.
\subsection{Infrared Factorization in Gauge Theories}
Let us recall the IR factorization in gauge theories and draw intuitions for what we may expect for the lightlike Wilson loops.
The color-ordered scattering amplitude in gauge theories has the factorization property with respect to the IR divergence. First, consider the collinear limit that lightlike momenta $k_i$ and $k_{i+1}$ of two partons $i$ and $i+1$ become parallel and coalesce to a new lightlike momentum $k_P$. Kinematically, this situation described by
\bea
k_{i} \longrightarrow h k_P \quad \mbox{and} \quad k_{i+1} \longrightarrow (1-h) k_P, \hskip2cm ( 0 \le h \le 1)
\eea
so the two collinear partons carry the fraction $h, (1 - h)$ of the momentum $k_P$.
For the $L$-loop $n$-point scattering amplitude $A_n^{(L)} (k_1, \cdots, k_n)$, the collinear limit exhibits factorization \cite{Bern:1999a}
\begin{equation}
A_{n}^{(L)}(k_1, \cdots, k_n) \longrightarrow \sum_{\lambda=\pm} \sum_{\ell =0}^L \mbox{Split}_{-\lambda}^{(L-\ell)}(h;k_{i},\lambda_{i};k_{i+1},\lambda_{i+1}) A_{n-1}^{(\ell)}(k_P,\lambda, \cdots k_n))
\label{split}
\end{equation}
Here $\lambda$ labels the polarization state of the factorizing parton. In the summation, $L, \ell=0$ denote the tree-level amplitude.
Helicity structure is fixed by the Poincar\'e invariance, so both the scattering amplitudes and the splitting functions can be decomposed to their tree-level counterparts times scalar functions summarizing loop corrections.
We define reduced scattering amplitudes $M_{n}^{(L)}$ for the ratio of the $L$-loop scattering amplitude to the tree-level scattering amplitude:
\begin{align}
A_{n}^{(L)} (k_1, \lambda_1, \cdots, k_n, \lambda_n) = A_{n}^{(0)} (k_1, \lambda_1, \cdots, k_n \lambda_n) \cdot M_{n}^{(L)} (k_1, \cdots, k_n).
\end{align}
Similarly, we define the reduced splitting functions $R_s^{(L)}(\epsilon,z,k_P)$ for the ratio of the $L$-loop splitting function to the tree-level splitting function
\begin{align}
\mbox{Split}_{-\lambda}^{(L)}(h;k_{i},\lambda_{i};k_{i+1},\lambda_{i+1})= \mbox{Split}_{-\lambda}^{(0)}(h;k_{i},\lambda_{i};k_{i+1},\lambda_{i+1}) \cdot R_s^{(L)}(\epsilon,h;k_P),
\label{rs}
\end{align}
where we use the dimensional regularization for the IR divergences.
In the collinear limit, the tree-level scattering amplitudes are expected to factorize as follows:
\begin{equation}
A_{n}^{(0)} \longrightarrow \sum_{\lambda=\pm} \mbox{Split}_{-\lambda}^{(0)}(h;k_{i},\lambda_{i};k_{i+1},\lambda_{i+1}) A_{n-1}^{(0)}(k_P,\lambda). \label{tree}
\end{equation}
This is illustrated in next Figure.
\begin{center}
\fcolorbox{white}{white}{
\begin{picture}(104,80) (71,-41)
\SetWidth{1.0}
\SetColor{Black}
\Arc(120,2)(26.077,148,508)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](96,14)(72,26)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](96,-10)(72,-22)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](138,20)(156,38)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](144,14)(168,26)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](144,2)(174,2)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](144,-10)(168,-22)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](132,-22)(144,-40)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](138,-16)(162,-34)
\end{picture}
}
\\
\fcolorbox{white}{white}{
\begin{picture}(220,88) (63,-9)
\SetWidth{1.0}
\SetColor{Black}
\Arc(132,28)(18.974,162,522)
\Arc(228,28)(18.974,162,522)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](114,34)(78,58)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](114,22)(78,4)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](150,28)(174,28)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](186,28)(210,28)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](240,46)(264,64)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](246,40)(276,58)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](252,34)(282,46)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](252,28)(282,28)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](252,22)(282,10)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](252,10)(276,-8)
\Text(60,58)[lb]{\Large{\Black{$k_i$}}}
\Text(60,4)[lb]{\Large{\Black{$k_{i+1}$}}}
\Text(162,40)[lb]{\Large{\Black{$-$}}}
\Text(192,40)[lb]{\Large{\Black{$-$}}}
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(220,88) (63,-9)
\SetWidth{1.0}
\SetColor{Black}
\Arc(132,28)(18.974,162,522)
\Arc(228,28)(18.974,162,522)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](114,34)(78,58)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](114,22)(78,4)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](150,28)(174,28)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](186,28)(210,28)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](240,46)(264,64)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](246,40)(276,58)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](252,34)(282,46)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](252,28)(282,28)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](252,22)(282,10)
\Line[arrow,arrowpos=0.5,arrowlength=5,arrowwidth=2,arrowinset=0.2](252,10)(276,-8)
\Text(60,58)[lb]{\Large{\Black{$k_i$}}}
\Text(60,4)[lb]{\Large{\Black{$k_{i+1}$}}}
\Text(162,40)[lb]{\Large{\Black{$+$}}}
\Text(192,40)[lb]{\Large{\Black{$+$}}}
\end{picture}
}
\end{center}
Figure 3. {\sl Factorization of scattering amplitudes in gauge theory. The $\lambda = \pm$ refers to the polarization of the intermediate, factorized particle state.}
\vskip0.5cm
Inserting the relation (\ref{split}) to (\ref{tree}), we get
\begin{equation}
M_n^{(L)}\longrightarrow \sum_{\ell =0}^L R_s^{(\ell)} M_{n-1}^{(L-\ell)} \label{split_M}
\end{equation}
By definition, $R_s^{(0)}=1$ and $M_{n}^{(0)}=1$. The reduced amplitudes $M_n^{(L)}$ at one- and two-loops factorize to
\begin{align}
M_n^{(1)} &\longrightarrow M_{n-1}^{(1)}+R_s^{(1)} \label{ratio1} \\
M_n^{(2)} &\longrightarrow M_{n-1}^{(2)}+R_s^{(1)}M_{n-1}^{(1)}+R_s^{(2)}. \label{ratio2}
\end{align}
In the $(3+1)$-dimensional SYM theory, it is known that (\ref{ratio1}) and (\ref{ratio2}) are related each other by the collinear relation \cite{Anastasiou:2003kj}:
\bea
M_n^{(2)}(\epsilon)=\frac{1}{2}\big( M_n^{(1)}(\epsilon)\big)^2+f^{(2)}(\epsilon) M_n^{(1)}(2\epsilon) + C^{(2)}.
\label{m}
\eea
Here, $C^{(2)}$ is a finite constant, equal to $-{1 \over 2} \zeta_2^2$. Also,
$f^{(2)}(\epsilon) = - (\zeta_2 + \zeta_3 \epsilon + \zeta_4 \epsilon^2 + \cdots )$.
Substituting (\ref{ratio1}) to (\ref{m}),
\bea
M_n^{(2)}(\epsilon) &=& {1 \over 2} (M^{(1)}_{n-1}(\epsilon) + R^{(1)}_s(\epsilon))^2 + f^{(2)}(\epsilon) (M^{(1)}_{n-1}(2 \epsilon) + R_s^{(1)}(2 \epsilon)) + C^{(2)} \nonumber \\
&=& M^{(2)}_{n-1} (\epsilon) + R^{(1)}_s (\epsilon) M^{(1)}_{n-1} + {1 \over 2} (R_s^{(1)}(\epsilon))^2 + f^{(2)}(\epsilon) R^{(1)}_s (2 \epsilon).
\eea
In the second line, we utlized the above collinear relation for $M^{(2)}_{n-1}$. Comparing this with (\ref{ratio2}), we obtain recursive relation for the splitting function:
\bea
R^{(2)}_s(\epsilon) = {1 \over 2} (R_s^{(1)})^2 + f^{(2)}(\epsilon) R^{(1)}_s(2 \epsilon) + {\cal O}(\epsilon).
\eea
More generally, the scalar splitting function $R_s^{(\ell)}$ also follows from the BDS-like relation for all higher $\ell > 1$ loops. Indeed, for QCD and $(3+1)$-dimensional ${\cal N}=4$ SYM theory, the scalar splitting function $R_s^{(1)}$ was calculated explicitly and its universality was established \cite{Kosower:1999rx}, \cite{Anastasiou:2003kj}.
Another source of the IR divergences in gauge theories is emission of the soft partons. These divergences also provide another kind of factorization. More explicitly, in the limit of one parton becomes soft, the scattering amplitudes exhibit an abelian factorization that it becomes a product of an eikonal factor with a lower-point scattering amplitude. At tree-level, when $b$-parton becomes soft, $k_b \simeq 0$, the soft factorization is given by
\begin{equation}
A_{n}^{(0)}(k_1, \cdots, k_a, k_b, k_c, \cdots k_n) \quad \longrightarrow \quad S^{(0)} (k_a, k_b, k_c) A_{n-1}^{(0)}(k_1, \cdots, k_a, k_c, \cdots, k_n) \qquad \mbox{for} \qquad k_b \rightarrow 0,
\end{equation}
where $S^{(0)}(k_a, k_b, k_c)$ denotes the tree-level eikonal factor,
\bea
S^{(0)}(k_a, k_b, k_c) = −{ 1 \over \sqrt{2}} \left[ {\epsilon_b^\pm \cdot k_a \over k_a \cdot k_b} - {\epsilon_b^\pm \cdot k_c \over k_b \cdot k_c} \right].
\eea
The soft bremsstrahlung factorization has the feature that this eikonal factor does not depend on the helicity of external particles.
The soft factorization also holds at higher loops. For example, at one loop, the scattering amplitude factorizes in the soft limit as
\bea
&& A^{(1)}_n (k_1, \cdots, k_a, k_b, k_c, \cdots, k_n)
\nonumber \\
&& \hskip0.5cm \longrightarrow S^{(0)}(k_a, k_b, k_c) A_{n-1}^{(1)}(k_1, \cdots, k_a, k_c, \cdots, k_n) + S^{(1)}(k_a,k_b, k_c) A^{(0)}_{n-1} (k_1, \cdots, k_a, k_c, \cdots, k_n). \qquad \qquad
\eea
Here, $S^{(1)}$ is the one-loop eikonal function. In dimensional regularization, it reads \cite{Bern:1998sc}:
\bea
S^{(1)}(k_a, k_b, k_c) = - S^{(0)}(k_a, k_b, k_c) {1 \over (4 \pi)^{2 - \epsilon}} {\Gamma(1 + \epsilon) \Gamma^2 (1 - \epsilon) \over \Gamma (1 - 2 \epsilon)} {1 \over \epsilon^2} \left( {(-s_{ac}) \mu^2 \over (- s_{ab}) (-s_{bc})} \right)^{\epsilon} {\pi \epsilon \over \sin (\pi \epsilon)},
\eea
where $s_{ab} = 2 k_a \cdot k_b$, etc. So, the soft factorization behavior is analogous to that of the collinear limit, just replacing the splitting function of the latter to the eikonal function.
The antenna function is a universal function introduced to describe in a unified manner all leading infrared singularities of tree-level scattering amplitudes as the color-connected set of momenta becomes collinear or soft. Consider, in color-order scattering amplitude, two hard momenta $k_a, k_b$ and one momentum $k_c$ in between. The unified factorization then takes the form
\begin{equation}
A_{n}^{(0)}(k_1, \cdots, k_a, k_c, k_b, \cdots, k_n) \quad \rightarrow \quad \sum_{\lambda} \mbox{Ant}(\hat{a},\hat{b} \leftarrow a,c,b)A_{n-1}(k_1, \cdots, -k_{\hat{a}},-k_{\hat{b}}, \cdots, k_n),
\end{equation}
where the antenna function Ant contains information of the parton $c$:
\begin{itemize}
\item collinear splitting function for $k_c \cdot k_a \rightarrow 0$ and $k_c \cdot k_b = $ finite ($k_{\hat{a}}=-(k_a+k_c), k_{\hat{b}}=-k_b$)
\item collinear splitting function for $k_c \cdot k_b \rightarrow 0$ and $k_c \cdot k_a = $ finite ($k_{\hat{a}} = - k_a, k_{\hat{b}} = - (k_c + k_b)$)
\item soft eikonal function for both $k_c \cdot k_a \rightarrow 0$ and $k_c \cdot k_b \rightarrow 0$
($k_{\hat{a}}=-k_a, k_{\hat{b}}=-k_b$).
\end{itemize}
The momentua $k_{\hat{a}}, k_{\hat{b}}$ are reconstructed from the original momenta via the reconstruction function \cite{Kosower:2002su}. The antenna function can also be extended to higher loops in terms of parton currents $J$ that was used in the Berends-Giele recursion relations \cite{Berends:1987me}. At $L$-loops,
\bea
\mbox{Ant}^{(L)}(\hat{a}, \hat{b} \leftarrow a, c, \cdots, m, b)
= \sum_{\ell = 0}^L \sum_{i=1}^m J^{(\ell)}(a, c, \cdots, i; \hat{a}) J^{(L-\ell)}(i+1, \cdots, m, b; \hat{b})
\eea
Then, the factorization of the leading-color contribution to higher-loop scattering amplitudes can be derived by matching to known purely collinear limit or purely soft bremsstrahlung limit. This leads to
\begin{equation}
A_{n}^{(L)}(k_1, \cdots, k_n) \longrightarrow \sum_{\ell=0}^L \sum_{\lambda} \mbox{Ant}^{(\ell)} (\hat{a},\hat{b} \leftarrow a,1,b) \cdot A_{n-1}^{(L-\ell)} (k_1, \cdots, -k_{\hat{a}},-k_{\hat{b}}, \cdots, k_n).
\label{antenna-loop}
\end{equation}
This can be generalized to multiple collinear singularities that involve simultaneous vanishing of Mandelstam invariants in these collinear momenta and one of the two hard momenta $a$ or $b$. We can also generalize this to multiple collinear-soft or purely multiple soft singularities that arise from vanishing of additional Mandelstam invariants involving other hard momenta as well.
Note that the leading singularities in the additional Mandelstam invariants are already incorporated to the antenna function. Therefore, these singularities also capture the leading behavior in the multiple collinear-soft or multiple soft singularities. Indeed, the $h \rightarrow 0$ limit of the collinear splitting function must also describe the soft bremsstrahlung eikonal. As such, (\ref{antenna-loop}) describes the leading singularity behavior of $L$-loop leading-color scattering amplitudes in all singular limits involving the color-connected singular set of momenta $k_1, \cdots, k_m$.
One can generalize the factorization to multi-parton kinematics. The next level of factorization involves two unresolved parton kinematics. The factorization in doubly unresolved limit is given at $L$ loops by
\begin{equation}
A_n^{(L)}(k_1, \cdots, k_n) \longrightarrow \sum_{\ell=0}^L \sum_{\lambda} \mbox{\rm Ant}^{(\ell)}(\hat{a},\hat{b} \leftarrow a,1,2,b) \cdot A_{n-2}^{(\ell)}(k_1, \cdots,-k_{\hat{a}},-k_{\hat{b}},\cdots, k_n). \label{antftn}
\end{equation}
This antenna function have various channels, for instance, triple collinear, double collinear, collinear soft and double soft. Among them, we will focus on the first case that $s_{a1},s_{12},t_{a12}$ goes to 0.
Much the way the splitting function or the eikonal function are universal, we expect the antenna function also have universal structures.
\subsection{Lightlike Factorization of Wilson Loop}
One expects that the lightlike polygon Wilson loops provides another class of processes that exhibit IR divergences and factorizations thereof. Indeed shape or geometry of the lightlike contour $C_n$ exhibits two types of move that can be viewed as the soft and the collinear limits. The soft bremsstrahlung limit takes place when two adjacent vertex points coalesce. The collinear limit takes place when two adjacent edges coalesce, equivalently, when three consecutive vertices become lightlike arrayed.
\vskip0.5cm
\begin{center}
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(148,157) (142,-53)
\SetWidth{2.0}
\SetColor{Black}
\Line(208,93)(160,69)
\Line(160,69)(144,29)
\Line(144,29)(152,-19)
\Line(152,-19)(200,-51)
\Line(200,-51)(280,-19)
\Line(288,29)(280,-19)
\SetWidth{0.3}
\Line(208,93)(288,29)
\SetWidth{1.0}
\Line[dash,dashsize=10](208,93)(280,85)
\Line[dash,dashsize=10](280,85)(288,29)
\Vertex(224,93){2.828}
\Vertex(280,85){2.828}
\SetWidth{2.0}
\Line(224,93)(288,29)
\Line(224,93)(208,93)
\SetWidth{1.0}
\SetColor{Red}
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](280,101)(224,101)
\end{picture}
}
\qquad
\qquad
\qquad
\fcolorbox{white}{white}{
\begin{picture}(148,148) (142,-62)
\SetWidth{2.0}
\SetColor{Black}
\Line(208,84)(160,60)
\Line(160,60)(144,20)
\Line(144,20)(152,-28)
\Line(152,-28)(200,-60)
\Line(200,-60)(280,-28)
\Line(288,20)(280,-28)
\SetWidth{0.3}
\Line(208,84)(288,20)
\SetWidth{1.0}
\Line[dash,dashsize=10](208,84)(280,76)
\Line[dash,dashsize=10](280,76)(288,20)
\Vertex(256,52){2.828}
\Vertex(280,76){2.828}
\SetWidth{2.0}
\Line(208,84)(256,52)
\Line(256,52)(288,20)
\SetWidth{1.0}
\SetColor{Red}
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](288,68)(272,52)
\end{picture}
}
\\
\medskip
\end{center}
Figure 4. {\sl Infrared singularities of lightlike polygon. There are two limits a contour $C_{n}$ can be reduced to $C_{n-1}$. The left figure describes the soft bremsstrahlung limit. The vertex $x_{i+1}$ coalesces to the adjacent vertex $x_i$, equivalently, the edge vector $y_i$ approaches to $0$. The right figure describes the collinear limit. The vertex $x_{i}$ approaches the lightlike edge connecting the vertices, $x_{i-1}$ and $x_{i+1}$, equivalently, two adjacent edge vectors $y_{i-1}$ and $y_i$ coalesce to a new lightlike vector. Although resulting topologies of are the same, the two limits should be distinguished. By analyticity, this is possible only if the limits are singular.}
\vskip0.5cm
The significance of these two processes is evident from geometric considerations among the vertex points $x_1, \cdots, x_n$. Generically, two non-adjacent vertex points are not lightlike separated. From either configurations, if we take succession of the above two processes for either vertex vectors or edge vectors within a lightlike polygon, we see that two non-adjacent cusp points of the polygon can be made lightlike separated. The limiting configuration is a lightlike polygon split to two lightlike polygons. Hereafter, this kinematic limit will be referred as {\sl lightlike factorization}. The classification is purely geomeric, so it must hold for observables defined for general quantum field theories of arbitrary spacetime dimensions.
In applying the above infrared factorizations of lightlike contour to ABJM Wilson loops, there is one further issue to be considered. We have shown in the last section that the lightlike ABJM Wilson loop cannot be defined on a polygon of odd numbers of edges since it does not permit configuration obeying the positivity condition \footnote{Recall also that this parallels to the fact that the ABJM scattering amplitudes involve even number of partons, though reasons are entirely different.}.
Whereas infrared factorization of the $(3+1)$-dimensional ${\cal N}=4$ SYM theory requires a single parton to fuse to other hard partons, infrared factorization of the $(2+1)$-dimensional ABJM theory requires two partons to fuse to other hard partons.
Thus, in the ABJM theory, we need to define an antenna function for double parton emissions. \footnote{Antenna function was studied for scattering amplitudes in QCD and other gauge theories. We are adopting the same terminology to lightlike Wilson loop expectation values. In $(3+1)$-dimensional ${\cal N}=4$ SYM theory, the scattering amplitude - Wilson loop duality relates the universal splitting function for collinear limit in scattering amplitudes to the universal factorization for lightlike limit of Wilson loop expectation values.}.
Intuitively, the above discussion makes it clear that the lightlike factorization is {\sl universal} --- the factorization should be independent of geometric details of spectator vertices or edges in the rest of the polygon. In the ABJM theory, we explained in the previous section that the positivity condition of the Mandelstam variables and the closedness condition of the edge vectors restricted the contour to even number of vertices, equivalently, even number of edges. Consistency with these conditions require that the infrared singularity must involve odd numbers of consecutive edges fusing to a single edge and consecutive vertices pairing up to dimerized configuration. Therefore, the basic building block of the lightlike factorization of a polygon Wilson loop is the collinear-soft-collinear limit among 4 consecutive vertices, equivalently, 3 consecutive edges. We shall introduce the ABJM {\sl antenna function} that describes in a unified way all leading singularities of such processes.
Incidentally, we do not consider the limit where three consecutive edges are purely collinear. This is because the corresponding edge vectors in general violate the positivity condition. We will further discuss this restriction below. We also do not consider the limit where two consecutive edges are purely soft. Although kinematically permitted, this limit requires to take several Mandelstam invariants to zero simultaneously. Numerically, such a limit is technically involved and difficult to handle. In this paper, we will not study this corner of the moduli space and simply contend that the universal antenna function we derive below be reduced to the correct double eikonal function of such processes once relevant factorization is taken judicially.
Our next goal is to explicitly check the universality of the antenna function for the lightlike Wilson loop. By definition, the Wilson loop operator is color-ordered. Therefore, we can describe its collinear factorization in a manner similar to the color-ordered scattering amplitudes in QCD exhibits the factorization with respect to the collinear divergences \cite{Bern:1999a}. So, take the collinear limit that three adjacent edge vectors $y_{i}, y_{i+1}$ and $y_{i+2}$ become lightlike parallel and coalesce to a new lightlike edge vector $y_P$. This situation is described by what we call `doubly unresolved limit' of $C_n \rightarrow C_{n-2}$:
\bea
y_{i} \rightarrow (1 - h_1 - h_2) y_P, \qquad y_{i+1} \rightarrow h_1 y_P, \quad y_{i+2} \rightarrow h_2 y_P \qquad
\mbox{where} \qquad y_P^2 = 0.
\label{kinematics}
\eea
In the doubly unresolved limit, we expect the lightlike polygon Wilson loop expectation value at $L$-loop factorizes universally as
\bea
\big< W_\Box [C_n] \big>^{(L)} \longrightarrow
\sum_{\ell=0}^L \mbox{Ant}^{(\ell)}(h_1, h_2; y_i, y_{i+1}, y_{i+2}) \big< W_\Box [C_{n-2}]
\big>^{(L-\ell)}.
\eea
We will abbreviate the antenna function that arises from factorization of the polygon $C_n$ Wilson loop as $\mbox{Ant}[C_n]$.
In the kinematics Eq.(\ref{kinematics}), the antenna function is closely related to the Wilson loop expectation value for the collapsing tetragon made of the edges $y_i, y_{i+1}, y_{i+2}, - y_P$. Our goal is to show that this antenna function is actually independent of the number of edges $n$ of the contour $C_n$ and hence universal. Note that the tree-level factorization for $L, \ell =0$:
\bea
\big< W_\Box[C_n] \big>^{(0)} \quad \longrightarrow \quad
\mbox{Ant}^{(0)}[C_n] \cdot \big< W_\Box [C_{n-2}] \big>^{(0)}
\eea
is actually a trivial statement since, in our normalization, all the quantities involved are 1.
At two-loop order, the doubly-unresolved configuration leads to the factorization:
\bea
\big< W_\Box [C_n] \big>^{(2)} \quad \longrightarrow \quad && \mbox{Ant}^{(2)}[C_4] \cdot \big< W_\Box [C_{n-2}] \rangle^{(0)} + \mbox{Ant}^{(0)} [C_4] \cdot \big< W_\Box [C_n] \big>^{(2)}
\nonumber \\
=&& \big< W_\Box [C_{n-2}]\big>^{(2)} + \mbox{Ant}^{(2)}[C_4].
\eea
The antenna function $\mbox{Ant}[C_4]$ is local in color-ordered contour geometry, so it is independent of $n$ and universal:
\bea
\boxed{
\mbox{Ant}^{(2)}[C_4] = \big< W_\Box[C_n] \big>^{(2)} - \big< W_\Box[C_{n-2}] \big>^{(2)} \quad \mbox{for \ all} \ \ n.
}
\label{antftn}
\eea
\section{Antenna Function for the ABJM Wilson Loops}
Built upon the idea of the previous section, we now construct the antenna function for the lightlike polygon Wilson loops in the ABJM theory. From (\ref{antftn}), the antenna function is obtained by subtracting $\big< W_\Box [C_{n-2}] \big>^{(2)}$ from $\big< W_\Box[C_n] \big>^{(2)}$. At first sight, it appears imperative to calculate $\big< W_\Box [C_{n-2}] \big>^{(2)}$ and $\big< W_\Box[C_n] \big>^{(2)}$. This turns out not the case, as most of the Feynman diagrams cancel each other. Eventually, only a small subset of Feynman diagrams contributes to the antenna function. In fact, the number of these diagrams are fixed regardless of $n$, which again is an indicative of the universality of the antenna function.
\subsection{Moduli Space of Lightlike Polygon Factorization}
In section 2, we already learned that anomalous conformal Ward identity offers hints on the analytic structure of the Wilson loop expectation value, separately for matter contribution and gauge boson contribution. As such, we shall consider the factorization limit for each contribution. In this subsection, we focus on lightlike factorization of matter contribution.
\subsubsection{triple collinear and soft-collinear kinematics}
As alluded above, we will need to deal with doubly-unresolved configuration involving four consecutive vertices, say, $x_2, x_3, x_4, x_5$ on a polygon $C_n$. The lightlike factorization takes place when $x_2$ and $x_5$ are lightlike separated. To reach this configuration, take two step. First, take $x_2$ and $x_5$ lightlike seperated. This does not yet put the edge vectors $y_2,y_3,y_4$
parallel, nor the contour $C_n$ factorized into two parts. Next, take $x_{24}^2$ and $x_{35}^2$ to 0. This gives the triple collinear / soft collinear limit of the edge vectors $y_2, y_3, y_4$. Upon taking these limits, the upper tetragon flattened, reducing $C_n$ to $C_{n-2}$. Note that this limit still leaves the vertices $x_3$ and $x_4$ as unrestricted moduli parameters. A corner of this moduli space where the 3 Mandelstam variables $x_{24}^2,x_{35}^2,x_{25}^2$ go to zero. There are two ways to approach this corner:
\begin{itemize}
\item {\bf triple-collinear limit}
\bea
y_2 \parallel y_3 \parallel y_4
\eea
\item {\bf soft-collinear limit}
\bea
y_3=0, \quad y_2 \parallel y_4
\eea
\end{itemize}
\vskip0.5cm
\begin{center}
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(124,84) (64,-38)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,0)(72,-36)
\SetWidth{0.4}
\Line(66,0)(186,-6)
\SetWidth{2.0}
\Line(186,-6)(180,-36)
\SetWidth{1.0}
\Line[dash,dashsize=10](66,0)(90,42)
\Line[dash,dashsize=10](90,42)(162,42)
\Line[dash,dashsize=10](162,42)(186,-6)
\SetWidth{2.0}
\Line(66,0)(84,6)
\Line(84,6)(168,6)
\Line(168,6)(186,-6)
\SetWidth{1.0}
\SetColor{Red}
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](96,36)(90,12)
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](156,36)(162,12)
\SetColor{Black}
\Vertex(168,6){2.828}
\Vertex(90,42){2.828}
\Vertex(84,6){2.828}
\Vertex(162,42){2.828}
\end{picture}
}
\quad \quad \qquad \qquad
\fcolorbox{white}{white}{
\begin{picture}(124,84) (64,-38)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,0)(72,-36)
\SetWidth{0.4}
\Line(66,0)(186,-6)
\SetWidth{2.0}
\Line(186,-6)(180,-36)
\SetWidth{1.0}
\Line[dash,dashsize=10](66,0)(90,42)
\Line[dash,dashsize=10](90,42)(162,42)
\Line[dash,dashsize=10](162,42)(186,-6)
\Vertex(162,42){2.828}
\Vertex(90,42){2.828}
\Vertex(96,6){2.828}
\Vertex(84,6){2.828}
\SetWidth{2.0}
\Line(66,0)(84,6)
\Line(84,6)(96,6)
\Line(96,6)(186,-6)
\SetWidth{1.0}
\SetColor{Red}
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](96,36)(90,12)
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](156,36)(108,12)
\end{picture}
}
\end{center}
Figure 5. {\sl There are two different special limits in doubly-unresolved geometries. Left figure describes the triple-collinear limit. Two vertices $x_{i+1}$ and $x_{i+2}$ approach a point on the edge connecting $x_i$ and $x_{i+3}$. Right figure describes the soft-collinear limit. The vertices $x_{i+1}$, $x_{i+2}$ come close each other and the edge vectors $y_{i-1}$ and $y_{i+1}$ become parallel.}
\vskip0.5cm
The triple-collinear limit is described by the contour geometry
\bea
y_2 \equiv h_1 y_C, \quad y_3 \equiv h_2 y_C, \quad y_4=h_3 y_C \quad \mbox{where} \quad h_1, h_2, h_3 \ge 0, \quad h_1+h_2+h_3=1 \quad \mbox{and} \quad y_C^2=0,
\eea
where the 'parton fraction' $h_1, h_2, h_2$ spans the local chart of the moduli space. Naively, the dimension of this moduli space is $[0, 1] \times [0, 1]$. The actual moduli space turns out $[0, 1]$, as we now explain. The kinematics describes the limit that three consecutive segment vectors are parallel one another. The corresponding Mandelstam invariants are
\bea
x_{13}^2= x_{14}^2=h_1 x_{15}^2, \quad x_{46}^2= x_{36}^2=h_3 x_{26}^2 \quad
x_{24}^2=(1-h_3)^2 y_C^2, \quad x_{35}^2=(1-h_1)^2 y_C^2, \quad x_{25}^2=y_C^2 \rightarrow 0. \label{hextc_profile}
\eea
So we see that, as the three edge vectors $y_2, y_3, y_4$ become parallel one another, the three Mandelstam invariants $x_{25}^2, x_{24}^2, x_{35}^2$ goes to 0. Their ratios are fixed with respect to the parton fractions $h_1, h_2, h_3$. We then recall that the Mandelstam invariants of physical configuration must satisfy the Gram sub-determinant conditions. For the above triple-collinear configuration, the Gram sub-determinant condition requires $h_2^2 x_{15}^2 x_{26}^2 =0 $ and is solved by $h_2 \rightarrow 0$. Therefore, for all $n$, we must set $h_2 = 0$. This then leads to the moduli space of the triple-collinear limit to be the domain of $h_1 = - h_3$, viz. $[0, 1]$.
The soft-collinear limit is described by the contour geometry
\bea
y_2 \equiv h_1 y_C, \quad y_3 \equiv y_{S}, \quad y_4=h_3 y_C \quad \mbox{where} \quad h_1, h_3 \ge 0 , \quad h_1+h_3=1, \quad y_C^2=0, \quad y_{S} \simeq 0.
\eea
The moduli space of this configuration is given by the domain of $h_1 = - h_3$, viz. $[0, 1]$. This can be checked straightforwardly.
The contour geometry describes the limit that a diminishing edge vector is squeezed between two collinear edge vectors. The corresponding Mandelstam invariants are
\bea
x_{13}^2 =
x_{14}^2=h_1x_{15}^2, \quad
x_{46}^2=
x_{36}^2=h_3 x_{26}^2, \quad
x_{24}^2=h_1 x_{25}^2, \quad x_{35}^2=h_3 x_{25}^2, \quad x_{25}^2=2y_C \cdot y_{S} \label{hexsc_profile_a}
\eea
It is straightforward to check that this kinematics automatically satisfy the Gram sub-determinant conditions provided $y_C^2=0$ and $y_{S} \rightarrow 0$. The four Mandelstam invariants $x_{13}^2$, $x_{14}^2$, $x_{46}^2$ and $x_{36}^2$ coincides with the triple-collinear limit invariants if $h_2$ is taken to 0. However, the ratios among $x_{25}^2, x_{24}^2, x_{35}^2$ are different from the triple-collinear limit, so should be considered separately.
\subsubsection{Factorization and Positivity Condition}
When computing the antenna function in perturbation theory, we need to impose two conditions to the moduli parameters of the polygon $C_n$: the Gram sub-determinant condition and the positivity condition. We identified that the Gram sub-determinant conditions is satisfied by both the triple collinear geometry on the subspace $h_2=0$ and the soft-collinear geometry. What about the positivity condition? Here, we show that the soft-collinear geometry of the polygon is uniquely singled out as the configuration that satisfy the positivity condition.
The triple collinear geometry is inconsistent with the positivity condition. To see this, start from
\bea
y_2 = h_1 y_C, \quad y_3 = h_2 y_C, \quad y_4 = h_3 y_C \quad \mbox{with} \quad h_1, h_2, h_3 \ge 0, \quad h_1+h_2+h_3=1, \quad y_C^2=0.
\nonumber
\eea
For a given lightlike vector $y_C$, the time-component of the $y_2,y_3,y_4$ vectors have the same sign since $h_1, h_2, h_3$ are all positive. On the other hand, the positivity condition requires alternating sign flip of the time-component. As such, the triple-collinear geometry contradicts this condition. We discard the triple collinear limit hereafter.
On the other hand, the soft-collinear geometry turns out to satisfy the positivity condition. To illustrate this, take the hexagon and consider the following parametrization of the edge vectors
\begin{align}
y_2&=h_1(\sqrt{2},1,1), \quad \quad y_3=a(-\sqrt{2},1,1), \quad \quad y_4=h_3(\sqrt{2},1,1) \label{p2p3p4}
\end{align}
We introduced a small parameter $a \simeq 0$ to render the vector $y_3$ soft. Fusion of these three edge vectors result in a new lightlike edge vector $y_C = (\sqrt{2}, 1, 1)$. The other edge vectors $y_5, y_6$ and $y_1$ are set to
\begin{align}
y_5&=(-\sqrt{f^2+g^2},f,g), \quad \quad y_6=(\sqrt{r^2+s^2},r,s), \quad \quad y_1=(-\sqrt{b^2+d^2},b,d),
\end{align}
where $g,r,s$ are free parameters while other three parameters $b,d,f$ will be fixed by the closedness condition. The Mandelstam variables $x^2_{ij}$ can then be read from these edge vectors from the identities $- 2 y_i \cdot y_j = x^2_{i+1, j} + x^2_{i, j+1} - x^2_{i, j} - x^2_{i+1, j+1}$. This configuration satisfies the positivity condition.
\subsection{Matter Contribution to Antenna Function}
Let's begin with the ABJM matter contribution to the antenna function. By (\ref{antftn}), this contribution to the antenna function is extracted from
\bea
\mbox{Ant}^{(2)}_{\rm matter} [C_4] = \big< \widetilde{W}_\Box[C_n] \big>^{(2)}_{\rm matter} - \big< \widetilde{W}_\Box [C_{n-2}] \big>^{(2)}_{\rm matter}.
\eea
Here, $\big< \widetilde{W}_\Box[C_n] \big>^{(2)}_{\rm matter}$ refers to Wilson loop expectation value for the polygon $C_n$ of soft-collinear geometry.
After the soft-collinear limit is taken, the reduced polygon $C_{n-2}$ consists of $(n-2)$ edge vectors $y_1,y_C,y_4, \cdots y_{n}$. Again, the new set of Mandelstam invariants $x_{ij}^2$ are obtained by inner product of these edge vectors.
The Wilson loop expectation value $\big< \widetilde{W}_\Box [C_{n-2}] \big>^{(2)}_{\rm matter}$ is obtained by inserting this new Mandelstam invariants to $\big< W_\Box [C_{n-2}] \big>^{(2)}_{\rm matter}$.
In case the three edge vectors $y_2, \ y_3, \ y_4$ coalesce in the soft-collinear limit, we shall call the vectors $y_1,y_2,y_3,y_4,y_5$ as `relevant edges'. We can then classify the matter-dependent 2-loop diagrams according to the locations the one-loop gauge propagator is attached:
\begin{align}
\mbox{Group A} \quad &: \quad \mbox{Neither end is attached to the relevant edges} \nonumber \\
\mbox{Group B} \quad &: \quad \mbox{One end is attached to the relevant edges while the other end is attached elsewhere} \nonumber \\
\mbox{Group C} \quad &: \quad \mbox{Both ends are attached to the relevant edges} \nonumber
\end{align}
We claim that Feynman diagrams belonging to Group A and Group B do not contribute to the antenna function. In other words,
\bea
&& \big< \widetilde{W}_\Box[C_n] \big>^{(2)}_{\rm matter}\Big|_{\mbox{\tiny{Group A}}} = \big< \widetilde{W}_\Box [C_{n-2}] \big>^{(2)}_{\rm matter}\Big|_{\mbox{\tiny{Group A}}}
\nonumber \\
&& \big< \widetilde{W}_\Box[C_n] \big>^{(2)}_{\rm matter} \Big|_{\mbox{\tiny{Group B}}} = \big< \widetilde{W}_\Box [C_{n-2}] \big>^{(2)}_{\rm matter} \Big|_{\mbox{\tiny{Group B}}}
\eea
Nontrivial contributions to the antenna function stem entirely from 9 Feynman diagrams belonging to the Group C.
\vskip0.5cm
\begin{center}
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(94,94) (76,-34)
\SetWidth{2.0}
\SetColor{Black}
\Line(78,28)(96,52)
\Line(96,52)(138,58)
\Line(138,58)(168,40)
\Line(168,40)(162,-8)
\Line(162,-8)(138,-32)
\Line(138,-32)(108,-32)
\Line(108,-32)(78,-8)
\Line(78,-8)(78,28)
\SetWidth{1.0}
\PhotonArc(133.637,60.726)(26.776,-163.23,-33.365){4}{5.5}
\Vertex(132,34){6}
\Line[dash,dashsize=10](78,-8)(162,-8)
\end{picture}
}
\quad \quad
\fcolorbox{white}{white}{
\begin{picture}(94,93) (76,-34)
\SetWidth{2.0}
\SetColor{Black}
\Line(78,27)(96,51)
\Line(96,51)(138,57)
\Line(138,57)(168,39)
\Line(168,39)(162,-9)
\Line(78,-9)(78,27)
\SetWidth{1.0}
\PhotonArc(133.637,59.726)(26.776,-163.23,-33.365){4}{5.5}
\Vertex(132,33){6}
\SetWidth{2.0}
\Line(78,-9)(162,-9)
\SetWidth{0.1}
\Line[dash,dashsize=10](78,-9)(108,-33)
\Line[dash,dashsize=10](108,-33)(144,-33)
\Line[dash,dashsize=10](144,-33)(162,-9)
\end{picture}
}
\\
\bigskip
(a) \qquad \qquad \qquad \qquad \qquad \qquad (b)
\\
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(94,94) (76,-34)
\SetWidth{2.0}
\SetColor{Black}
\Line(78,28)(96,52)
\Line(96,52)(138,58)
\Line(138,58)(168,40)
\Line(168,40)(162,-8)
\Line(78,-8)(78,28)
\SetWidth{1.0}
\PhotonArc[clock](177,-26)(75,-180,-253.74){4}{8.5}
\Vertex(120,22){6}
\SetWidth{2.0}
\Line(78,-8)(108,-32)
\Line(108,-32)(144,-32)
\Line(144,-32)(162,-8)
\SetWidth{1.0}
\Line[dash,dashsize=10](78,-8)(162,-8)
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(94,94) (76,-34)
\SetWidth{2.0}
\SetColor{Black}
\Line(78,28)(96,52)
\Line(96,52)(138,58)
\Line(138,58)(168,40)
\Line(168,40)(162,-8)
\Line(78,-8)(78,28)
\SetWidth{1.0}
\PhotonArc[clock](165.143,-2.286)(49.144,-142.797,-259.278){4}{8.5}
\Vertex(120,16){6}
\SetWidth{2.0}
\Line(78,-8)(108,-32)
\Line(108,-32)(144,-32)
\Line(144,-32)(162,-8)
\SetWidth{1.0}
\Line[dash,dashsize=10](78,-8)(162,-8)
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(94,94) (76,-34)
\SetWidth{2.0}
\SetColor{Black}
\Line(78,28)(96,52)
\Line(96,52)(138,58)
\Line(138,58)(168,40)
\Line(168,40)(162,-8)
\Line(78,-8)(78,28)
\SetWidth{1.0}
\PhotonArc[clock](156,16)(30,-90,-270){4}{8.5}
\Vertex(126,16){6}
\SetWidth{2.0}
\Line(78,-8)(108,-32)
\Line(108,-32)(144,-32)
\Line(144,-32)(162,-8)
\SetWidth{1.0}
\Line[dash,dashsize=10](78,-8)(162,-8)
\end{picture}
}
\fcolorbox{white}{white}{
\begin{picture}(94,93) (76,-34)
\SetWidth{2.0}
\SetColor{Black}
\Line(78,27)(96,51)
\Line(96,51)(138,57)
\Line(138,57)(168,39)
\Line(168,39)(162,-9)
\Line(78,-9)(78,27)
\SetWidth{1.0}
\PhotonArc[clock](162.75,1.5)(44.021,-166.201,-261.18){4}{6.5}
\Vertex(126,27){6}
\SetWidth{2.0}
\Line(78,-9)(162,-9)
\SetWidth{0.1}
\Line[dash,dashsize=10](78,-9)(108,-33)
\Line[dash,dashsize=10](108,-33)(144,-33)
\Line[dash,dashsize=10](144,-33)(162,-9)
\end{picture}
}
\\
\bigskip
(c) \qquad \qquad \qquad \qquad \qquad (d) \qquad \qquad \qquad \qquad \qquad (e) \qquad \qquad \qquad \qquad \qquad (f)
\\
\end{center}
Figure 6. {\sl The Feynman diagrams belonging to Group A and Group B. Upper diagrams belong to the Group A. (a) is equivalent to (b) when collinear limit taken. The diagrams in the second line belong to the Group B. (c)+(d)+(e) is equivalent to (f) after collinear limit. They are cancelled, therefore do not contribute to Antenna function.}
\vskip0.5cm
We found that, for any $n$, there are always 9 types of diagram that contribute to the antenna function:
\begin{align}
\mbox{Group C} = \Big\{ I_{21}, \ \ I_{32}, \ \ I_{43}, \ \ I_{54}, \ \ I_{31}, \ \ I_{41}, \ \ I_{52}, \ \ I_{53}, \ \ I_{42} \Big\}
\end{align}
In other words, all nontrivial contributions to the antenna function are from `local moves' around the three edge vectors fusing one another. This features a heuristic and intuitive explanation for the universality.
We calculated these 9 diagrams and computed the antenna function. From the known analytic results of these diagrams, we took the soft-collinear limit and subtracted the relevant $C_{n-2}$ diagrams. Up to ${\cal O}(\epsilon^0)$, we found the result as
\bea
\mbox{Ant}^{(2)}_{\rm matter} &=& \frac{1}{4\epsilon^2} + \frac{1}{4\epsilon}\big(\mbox{Log} h_1 + \mbox{Log} h_3 + \mbox{Log}(x_{24}^2 + \mbox{Log}(x_{35}^2) \big) \nonumber \\
&+& \frac{1}{2} \mbox{Log} h_1 \mbox{Log}(x_{24}^2) +\frac{1}{2} \mbox{Log} h_3 \mbox{Log}(x_{35}^2) +\frac{1}{2} \mbox{Log} (x_{35}^2) \mbox{Log}(x_{24}^2) -\frac{1}{2} \mbox{Log}h_1 \mbox{Log}h_3 \nonumber \\
&-& \frac{\pi^2}{6}.
\eea
In obtaining this result, we used the Abel's identity for the dilogarithms :
\begin{equation}
\mbox{Li}_2(u)+\mbox{Li}_2(v)-\mbox{Li}_2(u v)=\mbox{Li}_2\Big(\frac{u-uv}{1-uv}\Big)+\mbox{Li}_2\Big(\frac{v-uv}{1-uv}\Big)-\mbox{log}\Big(\frac{1-u}{1-u v}\Big) \mbox{log}\Big(\frac{1-v}{1-u v}\Big)
\end{equation}
and the Landen's identity:
\begin{equation}
\mbox{Li}_2(x)+\mbox{Li}_2(\frac{1}{x})=\frac{\pi^2}{3}-\frac{1}{2}\mbox{Log}^2(x)-i\pi \mbox{Log}(x).
\end{equation}
\subsection{Chern-Simons Contribution to Antenna function}
We now turn to the contribution of the pure Chern-Simons sector to the antenna function.
Again, we expect that
\bea
\mbox{Ant}^{(2)}_{\rm CS}[C_4] = \big< \widetilde{W}_\Box[C_n] \big>^{(2)}_{\rm CS} - \big< \widetilde{W}_\Box [C_{n-2}] \big>^{(2)}_{\rm CS}
\eea
Here, the Wilson loops $ \big< \widetilde{W}_\Box[C_n] \big>^{(2)}_{CS}$ and $\big< \widetilde{W}_\Box [C_{n-2}] \big>^{(2)}_{CS}$ are defined the same way as we defined for the matter contributions.
As explained in the previous section, the pure Chern-Simons contribution consists of the ladder diagrams and the triple-vertex diagrams. We found that the ladder diagrams does not give rise to infrared divergences, so they do not contribute to the antenna function.
We thus focus on the triple-vertex diagrams. We can again classify the relevant Feynman diagrams according to the combinatorics the triple gauge bosons are attached to the polygon $C_n$. As for the matter contributions, we showed in the last subsection that only `local moves' to the relevant edges contribute to the leading IR singularities. This turns out also the case for the Chern-Simons part: the contribution is completely determined by the triple-vertex diagrams whose gauge bosons are all attached to the relevant edges, $y_1, y_2, y_3, y_4, y_5$.
There are also IR divergences arising from `semi-local moves'. For instance, $I_{654}$ in $C_6$ is divergent. However, this divergence is cancelled by the diagram $I_{65P}$ in $C_{4}$ and $y_2 \parallel y_3 \parallel y_4$ when we compute the antenna function as the difference between the Wilson loop of $C_n$ and the Wilson loop of $C_{n-2}$. One readily notes that nontrivial contributions to the antenna function come from (1) the process that is divergent in $C_n$ but finite in $C_{n-2}$ and (2) the process that is finite in $C_n$ but divergent in $C_{n-2}$. These two processes are completely captured by the local moves to the relevant edges.
Recall the pure Chern-Simons result of the lightlike Wilson loop expectation value for hexagon and tetragon:
\begin{align}
\big< W_{\Box}[C_6]\big>^{(2)}_{\rm CS}&=- \Big[ \frac{\Log(2)}{2} \frac{\sum_{i=1}^6(x_{i,i+2}^2 \pi e^{\gamma_E} \mu^2)^{2\epsilon}}{2\epsilon} - \frac{17}{16}\zeta_2 + {3 \over 4} \mbox{Log}^2(2) \Big] \nonumber \\
\big< W_{\Box}[C_4]\big>^{(2)}_{\rm CS}&=- \Big[\frac{\Log(2)}{2} \frac{\sum_{i=1}^4(x_{i,i+2}^2 \pi e^{\gamma_E} \mu^2)^{2\epsilon}}{2\epsilon} - \ \frac{5}{8}\ \zeta_2 + {1 \over 2}\mbox{Log}^2(2) \Big]. \label{4and6}
\end{align}
Then, $\big< \widetilde{W}_\Box[C_6] \big>^{(2)}_{\rm CS}$ is obtained by taking the soft-collinear geometry (\ref{hexsc_profile_a}) to $\big< W_\Box [C_6]\big>^{(2)}_{\rm CS}$. For $\big< \widetilde{W}_\Box [C_{4}] \big>^{(2)}_{\rm CS}$, we replace $x_{13}^2$ and $x_{24}^2$ in (\ref{4and6}) by $x_{15}^2$ and $x_{26}^2$. The contribution to the antenna function is then obtained from the difference
\begin{align}
\mbox{Ant}^{(2)}_{\rm CS}[C_6]&= \big< \widetilde{W}_\Box[C_6] \big>^{(2)}_{\rm CS} - \big< \widetilde{W}_\Box [C_{4}] \big>^{(2)}_{\rm CS} \nonumber \\
&=\Big[\frac{\mbox{Log}(2)}{2\epsilon} +\frac{1}{2}\mbox{Log}(2) \ \mbox{Log}( h_1) +\frac{1}{2}\mbox{Log}(2) \ \mbox{Log}( h_3) +\frac{1}{2}\mbox{Log}(2) \ \mbox{Log} (x_{24}^2) +\frac{1}{2}\mbox{Log}(2) \ \mbox{Log} (x_{35}^2) \nonumber \\
&+ \mbox{Ant}^{(2)}[C_6]\Big|_{\rm finite} \Big]. \label{ant64exp}
\end{align}
Here,
\bea
\mbox{Ant}^{(2)}[C_6] \Big|_{\rm finite} = -\frac{7}{4}\zeta_2 + \mbox{Log}^2(2).
\eea
It turned out we need to numerically evaluate the Chern-Simons contribution to the antenna function. Hereafter,
we shall explicitly compute the contribution from the soft-collinear factorization of hexagon $C_6$ and octagon $C_8$ contours.
\subsubsection{hexagon $\longrightarrow$ tetragon}
For computational simplicity, let's first consider the soft-collinear factorization of hexagon $C_6$ to tetragon $C_4$. We computed contribution of the Chern-Simions contribution to the antenna function $\mbox{Ant}^{(2)}_{\rm CS}[C_6]$. Earlier, we alluded that
the gauge boson ladder diagrams do not contribute to the antenna function, though they do exhibit leading IR
singularities. We can classify the ladder diagrams into three groups:
\begin{align}
\mbox{Group A} &: \quad \{I_{3366},\ \ I_{6634},\ \ I_{6623}, \ \ I_{6624},\ \ I_{4466},\ \ I_{6622} \} \nonumber \\
\mbox{Group B} &: \quad \{ I_{5511} \}\nonumber \\
\mbox{Group C} &: \quad \mbox{all \ other \ diagrams} \nonumber
\end{align}
We found numerically that diagrams belonging to Group C vanishes in the soft-collinear limit.
We also checked numerically that the following identity holds:
\bea
\big< \widetilde{W}_\Box[C_n] \big>^{(2)}_{\rm ladder}\Big\vert_{A} - \big< \widetilde{W}_\Box [C_{n-2}] \big>^{(2)}_{\rm ladder} \Big\vert_{A} &=0 \nonumber \\
\big< \widetilde{W}_\Box[C_n] \big>^{(2)}_{\rm ladder} \Big\vert_{B} - \big< \widetilde{W}_\Box [C_{n-2}] \big>^{(2)}_{\rm ladder} \Big\vert_{B} &=0. \nonumber
\eea
We conclude that the ladder diagrams do not contribute to the antenna function.
\vskip0.5cm
\begin{center}
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(292,100) (76,-28)
\SetWidth{2.0}
\SetColor{Black}
\Line(84,46)(114,70)
\Line(114,70)(150,70)
\Line(150,70)(174,40)
\Line(174,16)(174,40)
\Line(78,4)(84,46)
\Line(78,4)(108,-26)
\Line(108,-26)(144,-26)
\Line(144,-26)(174,16)
\SetWidth{1.0}
\SetColor{Red}
\Line[dash,dashsize=10](84,4)(174,16)
\SetColor{Black}
\Photon(84,34)(126,40){4}{4}
\Photon(126,40)(90,-8){4}{6}
\Photon(126,40)(162,-2){4}{5}
\SetWidth{2.0}
\Line(276,46)(306,70)
\Line(306,70)(342,70)
\Line(342,70)(366,40)
\Line(366,16)(366,40)
\SetWidth{1.0}
\Photon(318,40)(318,10){4}{3}
\SetWidth{2.0}
\Line(270,4)(276,46)
\SetWidth{1.0}
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](204,22)(246,22)
\SetWidth{2.0}
\SetColor{Red}
\Line(270,4)(366,16)
\SetWidth{1.0}
\SetColor{Black}
\Photon(276,34)(318,40){4}{4}
\Photon(318,40)(354,16){4}{4}
\end{picture}
}
\\
\bigskip
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(292,100) (76,-28)
\SetWidth{2.0}
\SetColor{Black}
\Line(84,46)(114,70)
\Line(114,70)(150,70)
\Line(150,70)(174,40)
\Line(174,16)(174,40)
\Line(78,4)(84,46)
\Line(78,4)(108,-26)
\Line(108,-26)(144,-26)
\Line(144,-26)(174,16)
\SetWidth{1.0}
\SetColor{Red}
\Line[dash,dashsize=10](84,4)(174,16)
\SetWidth{2.0}
\SetColor{Black}
\Line(276,46)(306,70)
\Line(306,70)(342,70)
\Line(342,70)(366,40)
\Line(366,16)(366,40)
\Line(270,4)(276,46)
\SetWidth{1.0}
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](204,22)(246,22)
\SetWidth{2.0}
\SetColor{Red}
\Line(270,4)(366,16)
\SetWidth{1.0}
\SetColor{Black}
\PhotonArc[clock](133.764,-65.182)(69.204,127.12,65.92){4}{5.5}
\Photon(126,-2)(132,-26){4}{3}
\PhotonArc[clock](318.62,10.145)(35.861,-173.363,-350.603){4}{8.5}
\Photon(324,46)(318,10){4}{3}
\end{picture}
}
\\
\end{center}
Figure 7. {\sl Examples of two-loop Feynman diagrams that belong to the set $TV$. These diagrams are `local' and hence yields nontrivial contribution to the antenna function. The red edge denotes the segment vector $y_C$ resulting from taking the soft-collinear limit.}
\vskip0.5cm
This brings us to the contribution of the triple-vertex diagrams.
We found that only the following 10 diagrams give rise to leading IR singularities
and hence can contribute to the antenna function:
\begin{equation}
TV =\{I_{321}, \ \ I_{432}, \ \ I_{543}, \ \ I_{421}, \ \ I_{532}, \ \ I_{431}, \ \ I_{542}, \ \ I_{521}, \ \ I_{531}, \ \ I_{541} \}
\end{equation}
We computed these diagrams numerically using the Mathematica package {\tt FIESTA}.
Table 6 summarizes inputs and numerical results of the IR finite part of the antenna function
derived from the hexagon Wilson loop.
\setlength{\tabcolsep}{1.8pt}
\setlength{\extrarowheight}{1.5pt}
{
\begin{table}[ht]
\centering
\begin{tabular}{|c||ccccccccc|c|c|}
\hline
&$x_{13}^2 $ & $x_{24}^2$ & $x_{35}^2$ & $x_{46}^2$ & $x_{15}^2$ & $x_{26}^2$ & $x_{14}^2$ & $x_{25}^2$ & $x_{36}^2$ & $\big< W \big>_{\rm finite}$ & $\mbox{Ant}_{\rm finite}$ \\ \hline \hline
(1) & -1.95797 & -0.0005 & -0.0005 & -2.13973 & -3.91357 & -4.28024 &
-1.95511 & -0.00100 & -2.14001 &-3.47673 & -2.38762 \\ \hline
(2) &-5.02424 & -0.00100 & -0.00100 & -5.09307 & -10.048 & -10.188
& -5.02275 & -0.00200 & -5.09389 & -3.47701 & -2.38886 \\ \hline
(3) &-8.83791 &-0.00100 & -0.00100 & -13.6207 & -17.6764 & -27.2428
& -8.83749 & -0.00200 & -13.6211 & -3.47126 & -2.39603 \\ \hline
(4) &-11.4515 & -0.00199 &-0.00100 & -5.23415 & -22.9042 & -10.4691
& -11.4517 & -0.00200 & -5.23392 & -3.4768 & -2.39128 \\ \hline
\end{tabular}
\caption{Table 6. \sl Numerical result for the IR finite part of the lightlike hexagon Wilson loop expectation value and the antenna function. Notice that, for different configurations of the Mandelstam invariants, the results suggest that the IR finite part of the antenna function maintains a constant value.}
\end{table}
}
\vskip0.5cm
We also have an alternative method for calculating the antenna function. As in the two-parameter configuration of the hexagon, we can reduce the number of terms in the denominator of the Mellin-Barnes integrals by taking the soft-collinear configuration.
Moreover, we take a hint from the previous numerical results that the IR finite part of
the antenna function is independent of the polygon geometry.
This allows us to take the asymptotic limits for 3 of the Mandelstam invariants
$x_{25}^2$, $x_{15}^2$, $x_{26}^2$ and also for the the parton fractions
$h_1 \rightarrow 0, h_3 \rightarrow 1$. Taking these limits,
we succeeded in reducing to maximally 2-dimensional complex integrals.
Evaluating these integrals numerically, we find that
\bea
\mbox{Ant}^{(2)}_{\rm CS} &=& \frac{0.346574}{\epsilon} \label{ant64} \\
&+&\frac{1}{2}\Log(2) \ \Log (z_1) +\frac{1}{2} \Log (2) \ \Log (z_3) +\frac{1}{2} \Log(2) \ \Log(x_{24}^2) +\frac{1}{2}\Log (2) \
\Log (x_{35}^2) \nonumber \\
&-& 2.398181603. \nonumber
\eea
This result fits to what we expect from (\ref{ant64exp}). The numerical constant in (\ref{ant64}) can be identified with
\bea
-2.398181603 := -\frac{7}{4}\zeta_2 + \log^2(2) = -2.398181603066195
\nonumber
\eea
within the precision of $\mathcal{O}(10^{-7})$. This result reassures our intuitive picture that only those Feynman diagrams that are local move to the soft-collinear fusion contribute to the antenna function.
\subsubsection{octagon $\longrightarrow$ hexagon}
To convince that the antenna function we derived is universal for all $n$, we also computed the Chern-Simons part of the antenna function $\mbox{Ant}^{(2)}_{\rm CS}[C_8]$ for the factorization of octagon $C_8$ to hexagon $C_6$. Again, the set of Feynman diagrams that contribute to the antenna function comes only from the triple-vertex diagrams and consists of the 10 diagrams $TV$. This is because there are 5 relevant edge vectors for the soft-collinear kinematics. In fact, upon careful diagrammatic considerations, we confirmed that this argument holds for arbitrary $n$.
\vskip0.5cm
\begin{center}
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(106,58) (64,-52)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,4)(72,-26)
\Line(72,-26)(96,-50)
\Line(96,-50)(138,-50)
\Line(138,-50)(162,-26)
\Line(162,-26)(168,4)
\SetWidth{1.0}
\Line[dash,dashsize=10](72,-26)(162,-26)
\PhotonArc[clock](83.871,-43.198)(42.675,113.287,-9.171){4}{7.5}
\Photon(108,-14)(84,-38){4}{3}
\end{picture}
}
\quad
\fcolorbox{white}{white}{
\begin{picture}(106,58) (64,-52)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,4)(72,-26)
\Line(72,-26)(96,-50)
\Line(96,-50)(138,-50)
\Line(138,-50)(162,-26)
\Line(162,-26)(168,4)
\SetWidth{1.0}
\Line[dash,dashsize=10](72,-26)(162,-26)
\PhotonArc[clock](117,-39.8)(33.049,176.878,3.122){4}{8.5}
\Photon(114,-8)(114,-50){4}{4}
\end{picture}
}
\quad
\fcolorbox{white}{white}{
\begin{picture}(106,58) (64,-52)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,4)(72,-26)
\Line(72,-26)(96,-50)
\Line(96,-50)(138,-50)
\Line(138,-50)(162,-26)
\Line(162,-26)(168,4)
\SetWidth{1.0}
\Line[dash,dashsize=10](72,-26)(162,-26)
\Photon(126,-20)(150,-38){4}{3}
\PhotonArc(145.364,-47.545)(37.444,63.622,183.759){4}{7.5}
\end{picture}
}
\\
\bigskip
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(106,58) (64,-52)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,4)(72,-26)
\Line(72,-26)(96,-50)
\Line(96,-50)(138,-50)
\Line(138,-50)(162,-26)
\Line(162,-26)(168,4)
\SetWidth{1.0}
\Line[dash,dashsize=10](72,-26)(162,-26)
\Photon(114,-8)(114,-50){4}{4}
\PhotonArc(104.872,-55.553)(48.421,21.254,132.756){4}{8.5}
\end{picture}
}
\quad
\fcolorbox{white}{white}{
\begin{picture}(106,58) (64,-52)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,4)(72,-26)
\Line(72,-26)(96,-50)
\Line(96,-50)(138,-50)
\Line(138,-50)(162,-26)
\Line(162,-26)(168,4)
\SetWidth{1.0}
\Line[dash,dashsize=10](72,-26)(162,-26)
\PhotonArc[clock](176.4,-137.6)(135.86,132.852,93.545){4}{7.5}
\Photon(126,-14)(126,-50){4}{4}
\end{picture}
}
\quad
\fcolorbox{white}{white}{
\begin{picture}(106,58) (64,-52)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,4)(72,-26)
\Line(72,-26)(96,-50)
\Line(96,-50)(138,-50)
\Line(138,-50)(162,-26)
\Line(162,-26)(168,4)
\SetWidth{1.0}
\Line[dash,dashsize=10](72,-26)(162,-26)
\PhotonArc[clock](79.543,-84.457)(84.394,98.548,33.399){4}{6.5}
\Photon(120,-14)(114,-50){4}{3}
\end{picture}
}
\\
\bigskip
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(106,58) (64,-52)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,4)(72,-26)
\Line(72,-26)(96,-50)
\Line(96,-50)(138,-50)
\Line(138,-50)(162,-26)
\Line(162,-26)(168,4)
\SetWidth{1.0}
\Line[dash,dashsize=10](72,-26)(162,-26)
\PhotonArc[clock](196.616,-181.657)(181.922,127.846,99.05){4.5}{6.5}
\Photon(120,-14)(150,-38){5}{3}
\end{picture}
}
\quad
\fcolorbox{white}{white}{
\begin{picture}(106,58) (64,-52)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,4)(72,-26)
\Line(72,-26)(96,-50)
\Line(96,-50)(138,-50)
\Line(138,-50)(162,-26)
\Line(162,-26)(168,4)
\SetWidth{1.0}
\Line[dash,dashsize=10](72,-26)(162,-26)
\PhotonArc[clock](117.87,-88.957)(86.983,124.183,59.513){4.5}{6.5}
\Photon(108,-8)(84,-38){4.5}{3}
\end{picture}
}
\quad
\fcolorbox{white}{white}{
\begin{picture}(106,58) (64,-52)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,4)(72,-26)
\Line(72,-26)(96,-50)
\Line(96,-50)(138,-50)
\Line(138,-50)(162,-26)
\Line(162,-26)(168,4)
\SetWidth{1.0}
\Line[dash,dashsize=10](72,-26)(162,-26)
\PhotonArc[clock](117.87,-88.957)(86.983,124.183,59.513){4.5}{6.5}
\Photon(108,-8)(120,-50){4.5}{4}
\end{picture}
}
\\
\bigskip
\bigskip
\fcolorbox{white}{white}{
\begin{picture}(106,58) (64,-52)
\SetWidth{2.0}
\SetColor{Black}
\Line(66,4)(72,-26)
\Line(72,-26)(96,-50)
\Line(96,-50)(138,-50)
\Line(138,-50)(162,-26)
\Line(162,-26)(168,4)
\SetWidth{1.0}
\Line[dash,dashsize=10](72,-26)(162,-26)
\PhotonArc[clock](117.87,-88.957)(86.983,124.183,59.513){4.5}{6.5}
\Photon(108,-8)(150,-38){4.5}{3}
\end{picture}
}
\\
\end{center}
Figure 8. {\sl Elements of $TV$. Regardless of $n$, there are always 5 relevant segment vectors for the soft-collinear kinematics. These 10 diagrams are expected to contribute to the leading IR singularities and hence to the antenna function.}
\vskip0.5cm
Fortuitously, the seven Feynman diagrams $\{I_{321}, \ I_{432}, \ I_{543}, \ I_{421}, \ I_{532}, \ I_{431}, \ I_{542}\}$ for the octagon $C_8$ are exactly the same as those for the hexagon $C_6$. The remaining three diagrams $\{ I_{521}, \ I_{541}, \ I_{531} \}$ depend on the invariant $y_1 \cdot y_5$. We also need to modify this invariant according to the substitution
\bea
- 2y_1 \cdot y_5 = x_{25}^2 - x_{15}^2 - x_{26}^2 \quad \mbox{(Hexagon)} \quad \longrightarrow \quad - 2y_1 \cdot y_5 = x_{16}^2 + x_{25}^2 - x_{15}^2 - x_{26}^2 \quad \mbox{(Octagon)},
\eea
from which we see that it generates an additional Mandelstam invariant $x_{16}^2$.
For details, see appendix H.
Such change of the Mandelstam invariant is an exception for the hexagon to octagon and are not needed for the polygon with $n \ge 8$. That is, the sum of diagrams in the set $TV$ yields the same result for all $n \ge 8$.
Therefore, we expect the Chern-Simons contribution to the antenna function is the same for all $n \ge 8$. Of course, this is the feature we expect from the universality of the antenna function.
To evaluate the ten Feynman diagrams belonging to the set $TV$, we start with the soft-collinear geometry of the octagon $C_8$:
\bea
y_2 \equiv h_1 y_C, \quad y_3 \equiv y_{S}, \quad y_4=h_3 y_C&, \quad h_1, h_3 \ge 0, \quad h_1+h_3=1, \quad y_C^2=0 \quad y_{S} \simeq 0 .
\nonumber
\eea
In this limit, the Mandelstam invariants scale as
\bea
x_{13}^2=
x_{14}^2=h_1 x_{15}^2, \quad x_{46}^2=
x_{36}^2=h_3 x_{26}^2, \quad
x_{24}^2=h_1 x_{25}^2, \quad x_{35}^2=h_3 x_{25}^2, \quad x_{25}^2=2 y_C\cdot y_{S} \nonumber
\eea
\bea
x_{47}^2=h_3 x_{27}^2+ h_1 x_{57}^2, \quad x_{38}^2 = h_1 x_{58}^2 + h_3 x_{28}^2, \quad
x_{37}^2=h_1 x_{57}^2+ h_3 x_{27}^2, \quad x_{48}^2= h_3 x_{28}^2+ h_1 x_{58}^2.
\label{octtc_profile}
\eea
Algorithmically, we can generate this configuration starting from the hexagon by adding two edge vectors $y_7$ and $y_8$ and then imposing the positivity condition. The results of our numerical computation are summarized in Table 6. The results suggest that the finite part of the antenna function is independent of the choice of input Mandelstam invariants and that its numerical value is consistent with the numerical value extracted from the hexagon counterpart $\mbox{Ant}^{(2)}_{\rm CS, \ finite}$.
\setlength{\tabcolsep}{10pt}
\setlength{\extrarowheight}{1.5pt}
{
\begin{table}[!h]
\centering
\begin{tabular}{|c||c|c|}
\hline
& \{$x_{13}^2, x_{24}^2, x_{35}^2, x_{46}^2, x_{57}^2, x_{68}^2, x_{17}^2, x_{28}^2, x_{14}^2, x_{25}^2, x_{36}^2, x_{47}^2, x_{58}^2,$ & $\mbox{Ant}_{\rm CS}[C_8]_{\rm finite}$ \\
& $x_{16}^2, x_{27}^2, x_{38}^2, x_{15}^2, x_{26}^2, x_{37}^2, x_{48}^2$\} & \\ \hline \hline
& \{-0.53645, -0.00010, -0.00010, -0.47705, -0.25192, -11.3322 & \\
(1) & -2.27493, -7.89600, -0.53633, -0.00020, -0.47688, -0.53960, -0.87527, & -2.39734 \\
& -2.00140, -0.82752, -4.38564, -1.07288, -0.95403, -0.53974, -4.38553 \} & \\ \hline
& \{-2.35947, -0.00020, -0.00020, -3.04945, -12.433, -4.83899 & \\
(2) & -15.7136, -14.7954, -2.35936, -0.00040, -3.04955, -15.0670, -8.89897, & -2.39495 \\
& -0.450812, -17.7019, -11.8471, -4.71903, -6.09920, -15.0677, -11.8471 \} & \\ \hline
& \{-1.25316, -0.00020, -0.00020, -2.83305, -25.2163, -5.38963 & \\
(3) & -40.8014, -13.3704, -1.25260, -0.00040, -2.83315, -27.9871, -5.72270, & -2.39414 \\
& -1.94647, -30.7599, -9.54667, -2.50596, -5.66640, -27.9889, -9.5462 \} & \\ \hline
& \{-2.04042, -0.00020, -0.00020, -2.85248, -25.2976, -0.636597, & \\
(4) & -47.2655, -8.46627, -2.04047, -0.00040, -2.85243, -28.1002, -2.39373, & -2.39518 \\
& -0.349137, -30.9049, -5.42984, -4.08109,-5.70511, -28.1021, -5.42996 \} & \\ \hline
& \{-2.28437, -0.00010, -0.00010, -6.81235, -63.2938, -9.61232, & \\
(5) & -91.7177, -14.8614, -2.28411, -0.00020, -6.81221, -69.9258, -0.960108, & -2.39859 \\
& -11.3101, -76.5589, -7.91075, -4.56858, -13.6247, -69.9268, -7.91066 \} & \\ \hline
& \{-0.654001, -0.00006, -0.00034, -5.93175, -30.7878, -3.90942, & \\
(6) & -60.0102, -8.71534, -0.653408, -0.00040, -5.93188, -36.6137, -0.131069, & -2.38945 \\
& -7.86413, -37.6436, -7.42770, -4.35975, -6.97872, -36.6154, -7.42736 \} & \\ \hline
\end{tabular}
\caption{\sl Numerical result for the Chern-Simons contribution to the two-loop antenna function $\mbox{Ant}^{(2)}_{\rm CS}[C_8]_{\rm finite}$.
In numerical computation, we took $x_{25}^2 \simeq 0$ as a small quantity but not exactly zero. As we decrease $x_{25}^2$, we observed that $\mbox{Ant}^{(2)}_{\rm CS}[C_8]_{\rm finite}$ approaches to -2.39818.}
\end{table}
}
\vskip0.5cm
From the numerical results based on the hexagon and octagon Wilson loops, we find that the Chern-Simons contribution to the two-loop antenna function is given by
\bea
\mbox{Ant}^{(2)}_{\rm CS}[C_n] &=& \frac{\Log(2)}{2\epsilon} \nonumber \\
&+& \frac{1}{2}\Log(2) \ \Log (z_1) +\frac{1}{2}\Log (2) \ \Log (h_3) +\frac{1}{2}\Log(2) \ \Log(x_{24}^2) +\frac{1}{2}\Log(2) \ \Log (x_{35}^2)
\nonumber \\
&-&\frac{7}{4}\zeta(2) + \Log^2(2).
\eea
\subsection{ABJM Antenna Function}
The two-loop antenna function of the ABJM theory is then obtained by adding the matter-dependent contribution $\mbox{Ant}^{(2)}_{\rm matter}[C_n]$ and the pure Chern-Simons contribution $\mbox{Ant}^{(2)}_{\rm CS}[C_n]$ and suitably rescaling the regulator energy scale. The result reads
\begin{align}
& \mbox{Ant}^{(2)}_{\rm ABJM}\nonumber \\
&=\frac{1}{4\epsilon^2} + \frac{1}{4\epsilon}\big(\Log (z_1) + \Log (z_3) + \Log(x_{24}^2 \hat{\mu}^2) + \Log(x_{35}^2 \hat{\mu}^2) \big) \nonumber \\
&+\frac{1}{2} \Log(z_1) \Log(x_{24}^2 \hat{\mu}^2) +\frac{1}{2} \Log( h_3) \Log(x_{35}^2 \hat{\mu}^2) +\frac{1}{2} \Log (x_{35}^2 \hat{\mu}^2) \Log(x_{24}^2 \hat{\mu}^2) -\frac{1}{2} \Log (z_1) \Log (z_3) \nonumber \\
&+ \frac{1}{2}\Log^2(2) -\frac{11}{4}\zeta(2) .
\end{align}
Here, $\hat{\mu}^2 = 2 \mu^2$ is the rescaled regularization scale. The result is independent of $n$, confirming our intuition that the IR factorization is a local move and hence the antenna function should be a universal quantity.
\section{Recursion Relations and ABJM Wilson Loop Expecation Value}
Having obtained the universal antenna function, in this section, we shall obtain the Wilson loop expectation value for arbitrary polygon with $n \ge 8$. The strategy is to utilize the lightlike factorization and derive recursion relations between Wilson loops for polygon contours $C_n$ of different $n$.
Let's start with the Chern-Simons contribution. The two-loop antenna function takes the form:
\begin{align}
\mbox{Ant}^{(2)}_{\rm CS}[C_n] &= \mbox{Ant}^{(2)}_{CS}[C_n]\Big|_{\rm div} + \mbox{Ant}^{(2)}_{\rm CS}[C_n] \Big|_{\rm finite}
\end{align}
Here, $\mbox{Ant}^{(2)}_{\rm CS}[C_n] \Big|_{\rm div}$ and $\mbox{Ant}^{(2)}_{\rm CS}[C_n] \Big|_{\rm finite}$ are IR divergent, respectively, IR finite parts:
\begin{align}
&\mbox{Ant}^{(2)}_{\rm CS}[C_n]\Big|_{\rm div} = \frac{\Log(2)}{2\epsilon} + \frac{1}{2}\Log(2) \left[ \Log (h_1) + \Log (h_3) + \Log (x_{24}^2) + \Log (x_{35}^2) \right] \nonumber \\
&\mbox{Ant}^{(2)}_{\rm CS}[C_n] \Big|_{\rm finite}=-\frac{7}{4}\zeta(2) + \Log^2(2).
\end{align}
In deriving the Wilson loop expectation value, we are primarily interested in the analytic structure of the remainder function Rem$_{n,\rm CS}^{(2)}$. Therefore, it suffices to concentrate on the finite part, $\mbox{Ant}^{(2)}_{\rm CS}[C_n] \Big|_{\rm finite}$.
Intuitively, we can guess for the IR finite part of the remainder function, Rem$_{n,\rm CS}^{(2)}-\frac{n}{2} \Log(2)$. As a first step, consider $n=8$ octagon. Before imposing the Gram sub-determinant conditions, there are twelve conformal cross-ratios for the octagon. The remainder function for the octagon is a function of these cross-ratios:
\bea
\mbox{Rem}_{8,\rm CS}^{(2)} = \mbox{Rem}_{8, \rm CS}^{(2)}\big(u_{14},u_{25},u_{36},u_{47},u_{58},u_{16},u_{27},u_{38},u_{15},u_{26},u_{37},u_{48}
\big)
\eea
In the soft-collinear limit, $y_2 \parallel y_4$ and $y_3 \sim 0$, these cross-ratios are restricted accordingly:
\bea
u_{14} = u_{25} = u_{36} = u_{38} = u_{37} = 1, \quad u_{15} = 0 \nonumber
\eea
\bea
u_{47} = \frac{1-u_{26}}{u_{48}}, \quad u_{27} = \frac{1-u_{48}}{u_{26}}.
\eea
On the other hands, in the soft-collinear limit, the octagon $C_8$ is reduced to the hexagon $C_6$, for which the following nine Mandelstam invariants are relevant:
\bea
x_{16}^2, \ \ x_{15}^2, \ \ x_{26}^2, \ \ x_{17}^2, \ \ x_{27}^2, \ \ x_{28}^2, \ \ x_{57}^2, \ \ x_{58}^2, \ \ x_{68}^2.
\eea
From these invariants, we can form the following three conformal cross-ratios:
\bea
u_{58} = \frac{x_{15}^2 x_{68}^2}{x_{58}^2 x_{16}^2}, \qquad u_{16} = \frac{x_{17}^2 x_{26}^2}{x_{16}^2 x_{27}^2}, \qquad u_{27} u_{37} u_{47} = \frac{x_{28}^2 x_{57}^2}{x_{27}^2 x_{58}^2}
\eea
Therefore, by the lightlike factorization, the finite part of the antenna function $\mbox{Ant}^{(2)}_{\rm CS}[C_8] \Big|_{\rm finite}$ must be reduced to
\bea
&&\mbox{Rem}_{8, \rm CS}^{(2)} \big(1,1,1, (1-u_{26})/u_{48},u_{58},u_{16}, (1-u_{48})/u_{26},1,0,u_{26},1,u_{48}
\big) - 4 \Log(2) \nonumber \\
&=& \mbox{Rem}_{6,\rm CS}^{(2)} \big(u_{58},u_{16},u_{27} u_{37} u_{47} \big) -3 \Log(2) + \mbox{Ant}^{(2)}_{\rm CS}[C_8] \Big|_{\rm finite} \nonumber \\
&=& \mbox{Rem}_{6,\rm CS}^{(2)} \big(u_{58},u_{16},u_{27} u_{37} u_{47} \big) -3 \Log(2) -\frac{7}{4}\zeta(2) + \Log^2(2) \nonumber \\
&=& -6\zeta(2) + 4\Log^2(2).
\eea
In the last expression, we used the numerical result that $\mbox{Rem}_{6, \rm CS}^{(2)} \big(u_{58},u_{16},u_{27} u_{37} u_{47} \big)$ is a constant, independent of the input values of the conformal cross-ratios.
With such restricted information, it is impossible to determine general structure of the remainder function $\mbox{Rem}_{8, \rm CS}^{(2)}(u_{14},u_{25},u_{36},\cdots,u_{48})$ . However, for a given analytic structure of the remainder function, its soft-collinear limit should be controlled by the universal antenna function. This enables us to draw a conjecture that is consistent with the soft-collinear geometry to be
\bea
\mbox{Rem}_{8, \rm CS}^{(2)} = -6\zeta(2) + 4\mbox{Log}^2(2) + 4 \Log(2).
\eea
In other words, our conjecture is that the remainder function is independent of the twelve conformal cross-ratios. Moreover, utilizing numerical evidence that $\mbox{Rem}_{n, \rm CS}^{(2)}$ is constant-valued for arbitrary $n$, we find that the remainder function obeys the recursion relation
\bea
\boxed{
\mbox{Rem}_{n,\rm CS}^{(2)} - \frac{n}{2} \Log(2) = \mbox{Rem}_{n-2, \rm CS}^{(2)} -\frac{n-2}{2} \Log(2) + \mbox{Ant}^{(2)}_{\rm CS}[C_n]\Big|_{\rm finite}.
}
\eea
We can now iteratively solve this recursion relation along with our conjecture as the input:
\bea
\mbox{Rem}_{n,\rm CS}^{(2)} - \frac{n}{2} \Log(2) &=& \mbox{Rem}_{n-2, \rm CS}^{(2)} -\frac{n-2}{2} \Log(2) + \mbox{Ant}^{(2)}_{\rm CS}[C_n]\Big|_{\rm finite} \nonumber \\
&=& \mbox{Rem}_{n-2,\rm CS}^{(2)} -\frac{n-2}{2} \Log(2) -\frac{7}{4}\zeta(2) + \Log^2(2) \nonumber \\
&=& \cdots \nonumber \\
&=& \mbox{Rem}_{6, \rm CS}^{(2)} -3\Log(2) + \frac{n-6}{2} \big(-\frac{7}{4}\zeta(2) + \Log^2(2) \big) \nonumber \\
&=& \big[ \frac{1}{2} \Log^2(2) - \frac{7 \pi^2}{48}\big] n + \frac{\pi^2}{6}.
\eea
Our conjecture is further supported by numerical estimation of $\mbox{Rem}_{n, \rm CS}^{(2)}$ for $n=8,10, 12, \cdots, 20$ \cite{Wiegandt:2011a}.
Putting together, we now have the analytic result for the Chern-Simons contribution to the lightlike polygon Wilson loop expectation value as
\begin{equation}
\big< W_\Box [C_n]\big>^{(2)}_{\rm CS}= - \frac{\mbox{Log}(2)}{2} \sum_{i=1}^n \frac{(x_{i,i+2}^2 \pi e^{\gamma_E} \mu^2)^{2\epsilon}}{2\epsilon}+n \left( \frac{7 \pi^2}{48} - \frac{1}{2} \mbox{Log}^2(2)\right) - \frac{\pi^2}{6}. \label{w_result}
\end{equation}
Combining with the matter contribution (\ref{remainder-matter}), we finally arrives at the central result of this paper:
\begin{align}
\boxed{
\big< W_\Box [C_n]\big>^{(2)}_{\rm ABJM}= - {1 \over 2} \sum_{i=1}^n \frac{(x_{i,i+2}^2 8 \pi e^{\gamma_E} \mu^2)^{2\epsilon}}{(2\epsilon)^2} + \mbox{BDS}_n^{(2)} + n \left( \frac{\pi^2}{12} +\frac{3}{4} \Log^2(2) \right) - \frac{\pi^2}{6}.
}
\end{align}
\section{Test: Spacelike Circular Wilson Loop}
Having obtained the ABJM Wilson loop expectation value for lightlike polygon contour, we now would like to put the result to a test.
In this section, we shall consider a specific thermodynamic limit of $n\rightarrow \infty$ so that $C_n$ approaches a spacelike circle or ellipse.
Exact result of the Wilson loop expectation value for the circular contour is already known for pure Chern-Simons theory and for the ABJM theory by localization methods. We compare our result in this specific $n \rightarrow \infty$ limit with these exact results.
In Euclidean space, a circle (more generally an ellipse) is obtainable from a lightlike polygon $C_n$ by inscribing its vertices to touch the circle and taking the continuum limit $n \rightarrow \infty$. In Lorentzian spacetime, a spacelike circle (more generally an ellipse) is obtainable from a lightlike polygon $C_n$ by inscribing its edges to cross the circle and taking the continuum limit $n \rightarrow \infty$. It should therefore be possible to obtain a spacelike circular Wilson loop expectation value from the specified continuum limit of the lightlike polygon Wilson loop expectation value. Such test was first studied in \cite{Maldacena:2009a}.
Geometrically, the biggest difference of the lightlike polygon from the spacelike circular loop is the existence of vertices, where the polygon contour forms sharp cusps. Associated with these cusps are the UV divergences of the Wilson loop. However, these divergent parts are essentially abelian and they can be exponentiated and factored out. The remaining finite part should then be relatable to the spacelike circular Wilson loop expectation value in the thermodynamic limit. Roughly speaking, each vertices can be viewed as an elementary excitation along the loop and the $n \rightarrow \infty$ limit will populate the excitations such that the Wilson loop can be treated as a statistical system.
To obtain the circular Wilson loop, we shall arrange the geometry of the $n$ vertices as \cite{Heslop:2010a}
\bea
x_{2k} = \left(2 \ \mbox{sin} \frac{\pi}{2n}, \quad \mbox{cos} \frac{(2k+1)\pi}{n}, \quad \mbox{sin} \frac{(2k+1)\pi}{n} \right), \qquad x_{2k+1} = \left(0, \quad \mbox{cos} \frac{2k\pi}{n}, \quad \mbox{sin} \frac{2k\pi}{n} \right).
\label{heslop}
\eea
This configuration yields a polygon whose contour is sandwiched between a stack of two spacelike circles, all lying within $(1+1)$-dimensional subspace. Mandelstam invariants of the polygon are given by
\begin{align}
x_{2k,2j}^2&=-4\mbox{sin}^2\frac{(k-j)\pi}{n} \label{man_a} \\
x_{2k,2j+1}^2&= - 4\Big(\mbox{sin}^2\frac{(k-j-\frac{1}{2})\pi}{n} - \mbox{sin}^2\frac{\pi}{2n} \Big) \label{man_c} \\
x_{2k+1,2j+1}^2&=x_{2k,2j}^2 = -4\mbox{sin}^2\frac{(k-j)\pi}{n}. \label{man_b}
\end{align}
Evidently, the polygon satisfies the positivity condition and the closedness condition. Also, $(x_{2k+1}-x_{2k})^2 = 0$ holds.\\
\begin{center}
\fcolorbox{white}{white}{
\begin{picture}(440,156) (55,-15)
\SetWidth{1.0}
\SetColor{Black}
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](176,36)(176,124)
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](176,36)(120,-12)
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](176,36)(264,36)
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](276,76)(312,76)
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](372,36)(372,124)
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](372,36)(460,36)
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](372,36)(316,-12)
\SetWidth{2.0}
\Oval(176,36)(16,56)(0)
\Oval(176,76)(16,56)(0)
\Oval(372,36)(16,56)(0)
\SetWidth{1.0}
\SetColor{Red}
\Line(120,36)(136,68)
\Line(136,68)(152,20)
\Line(152,20)(176,60)
\Line(176,60)(192,20)
\Line(192,20)(216,68)
\Line(216,68)(232,36)
\Line(232,36)(216,84)
\Line(120,36)(136,84)
\Line[dash,dashsize=10](136,84)(152,52)
\Line[dash,dashsize=10](152,52)(168,92)
\Line[dash,dashsize=10](168,92)(192,52)
\Line[dash,dashsize=10](192,52)(216,84)
\Line(324,44)(316,36)
\Line(316,36)(328,28)
\Line(328,28)(352,20)
\Line(352,20)(368,20)
\Line(368,20)(380,20)
\Line(380,20)(396,24)
\Line(396,24)(416,28)
\Line(416,28)(428,36)
\Line(428,36)(424,44)
\Line(424,44)(412,48)
\Line(408,48)(396,52)
\Line(396,52)(384,52)
\Line(384,52)(368,52)
\Line(368,52)(356,52)
\Line(360,52)(344,48)
\Line(344,48)(340,48)
\Line(340,48)(324,44)
\SetColor{Black}
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](100,56)(100,36)
\Line[arrow,arrowpos=1,arrowlength=5,arrowwidth=2,arrowinset=0.2](100,56)(100,80)
\Text(52,56)[lb]{\Large{\Black{$2 \mbox{sin} \frac{\pi}{2n}$}}}
\Text(276,88)[lb]{\Large{\Black{$n \rightarrow \infty$}}}
\Text(184,120)[lb]{\Large{\Black{$t$}}}
\Text(108,-4)[lb]{\Large{\Black{$x$}}}
\Text(460,48)[lb]{\Large{\Black{$y$}}}
\Text(380,120)[lb]{\Large{\Black{$t$}}}
\Text(304,-4)[lb]{\Large{\Black{$x$}}}
\Text(264,48)[lb]{\Large{\Black{$y$}}}
\end{picture}
}
\label{largen}
\end{center}
Figure 9. {\sl The geometry of thermodynamic limit of lightlike polygon inscribed between two $(1+1)$-dimensional spacelike circular loops. }
\vskip0.5cm
There is a slight complication. As mentioned in Section 3, this $(1+1)$-dimensional kinematics would not satisfy the Gram sub-determinant conditions. While this is true for finite $n$, we now argue that the conditions are fully satisfied in the $n \rightarrow \infty$ thermodynamic limit.
First, focus on the pure Chern-Simons contribution to the Wilson loop expectation value. From the results in the section 8, we expect the result to be
\begin{equation}
\big< W_\Box [C_n]\big>^{(2)}_{\rm CS}= - \frac{\mbox{Log}(2)}{2} \sum_{i=1}^n \frac{(x_{i,i+2}^2 \pi e^{\gamma_E} \mu^2)^{2\epsilon}}{2\epsilon}+n \left( \frac{7 \pi^2}{48} - \frac{1}{2} \mbox{Log}^2(2)\right) - \frac{\pi^2}{6}. \label{CSstructure}
\end{equation}
were if the configuration to satisfy the Gram sub-determinant conditions. As said, the configuration (\ref{heslop}) does not satisfy them. However, we note that ($k \times k$) Gram sub-determinant conditions consist of $k$-th power of $\mbox{sin}^2 \frac{\pi}{n}$, because of all the Mandelstam invariants (\ref{man_c}) are accompanied by the factor of $\mbox{sin}^2 \frac{\pi}{n}$. Therefore, all the Gram sub-determinant conditions would vanish in the thermodynamic limit $n \rightarrow \infty$ and (\ref{CSstructure}) would approach the correct result. Geometrically, distance between the two enveloping circles $2 \mbox{sin} \frac{\pi}{2n}$ goes to zero, and the polygon collapses to a spacelike circle. This is illustrated in Figure \ref{largen}.
Dropping off the UV divergent parts, the Chern-Simons contribution reads
\begin{equation}
\big< W_\Box [C_n]\big>^{(2)}_{\rm CS}\Big|_{\rm{finite}}= n \left( \frac{7 \pi^2}{48} - \frac{1}{2} \mbox{Log}^2(2)\right) - \frac{\pi^2}{6}.
\end{equation}
The result is independent of the shape or the geometry of $C_n$. This fits well with the expectation that the pure Chern-Simons theory is topological. The thermodynamic limit $n \rightarrow \infty$ gives rise to a linear divergence, proportional to the perimeter $n := 2 \pi R/a$ for a circle of radius $R$ and short-distance defining scale $a$. It would be very interesting to understand this from the viewpoint of polygon regularization of the topological link invariants.
For the matter contribution part, the lightlike polygon Wilson loop expectation value was
\bea
\big< W_\Box[C_n]\big>^{(2)}_{\rm matter}
&=& - {1 \over 2}\sum_{i=1}^n \frac{1}{(2\epsilon)^2}(x_{i,i+2}^2 4\pi e^{\gamma_E} \mu^2)^{2\epsilon}+\mbox{BDS}^{(2)}_n-\frac{1}{16}n \pi^2 +{\cal O}(\epsilon).
\eea
Again, dropping off the UV divergent part of the result of Section 3.1, we have
\bea
\big< W_\Box[C_n]\big>^{(2)}_{\rm matter}\Big|_{\rm finite} &=& \sum_{i>j+1}^n I_{i,j}(x) ,
\eea
where
\bea
I_{i,j}(x) \equiv -\mbox{Li}_2(1-as)-\mbox{Li}_2(1-at)+\mbox{Li}_2(1-aP^2)+\mbox{Li}_2(1-aQ^2)-\frac{1}{16}n \pi^2.
\eea
\label{reg_pola}
We now recall from Section 3.1 that the parameters were defined by
\bea
a=\frac{s+t-P^2-Q^2}{st-P^2 Q^2}, \quad P^2=x_{i,j+2}^2, \quad Q^2=x_{i+1,j}^2, \quad s=x_{i,j}^2, \quad t=x_{i+1,j+1}^2.
\eea
Unlike the pure Chern-Simons part, the matter contribution part is sensitive to the shape or the geometry of the polygon $C_n$. Again,
this fits to our intuition that, once the matter is coupled, the Chern-Simons theory is no longer topological.
The next step is to consider a suitable geometrical limit so that the lightlike polygon $C_n$ asymptotes to the spacelike circle $O$. This regular polygon geometry is specified by (\ref{man_a}), (\ref{man_b}) and (\ref{man_c}) and the thermodynamic limit $n \rightarrow \infty$. We took this configuration to $\langle W_\Box[C_n] \rangle^{(2)}_{\rm matter}\Big\vert_{\rm finite}$ and evaluated its value numerically with respect to $n$.
Geometrically, the regular polygon has the property $x_{i,j}^2 = x_{i+1, j+1}^2 = \cdots = x_{i-1,j-1}^2$. This leads to the relation
\bea
I_{i,j}(x) = I_{i+1,j+1}(x) = \cdots = I_{i-1,j-1}(x) . \label{perm}
\eea
Define $F(n)$ and $f(n)$ by
\bea
F(n) &\equiv& \sum_{\substack{i>j \\ i \neq j+1}} I_{i,j} = \sum_{\substack{\rm{permutation} \\ \rm{of} \ j_*}} \sum_{\substack{i=j_*+2}}^{n+j_*-2} I_{i,j=j_*} \nonumber \\
f(n) &\equiv& \sum_{\substack{i=j_*+2}}^{n+j_*-2} I_{i,j=j_*},
\eea
where $j_*$ is some arbitrary reference number. By definition, $F(n)$ is equal to $\big< W_\Box[C_n]\big>^{(2)}_{\rm matter}|_{\rm finite}$. Also, (\ref{perm}) ensures that $f(n)$ is equal to ${1 \over n} F(n)$. We numerically computed the function $f(n)$ with high precision and fitted with 3-parameter $f(n)=\frac{a}{n^2}+\frac{b}{n}+c$. As we seen in Figure \ref{figure_regular}, this fitting works extremely well over two-orders of magnitudes, $6 \le n \le 500$. We found that the least chi-square fit of the coefficients $a,b,c$ to be
\bea
a = -21.1513 (\pm 0.005399), \qquad b = 9.86953(\pm 0.001135), \qquad c = -0.573233 (\pm 0.0002458)
\eea
Note that the value of $b$ very close to $\pi^2 := 9.869604401$.
With this result, we have
\bea
\big< W_\Box[C_n]\big>^{(2)}_{\rm matter}|_{\rm finite} = F(n) = \frac{a}{n} + \pi^2 + c n.
\eea
The first term dies out in the thermodynamic limit $n \rightarrow \infty$.
\begin{figure}[!h]
\centering
\includegraphics[scale=0.6]{plotresult.eps}
\label{figure_regular}
\end{figure}
Figure 10. {\sl The least chi-square fit of the function $f(n)$ with respect to $n$. Red crosses are numerical data obtained from (\ref{reg_pola}). The green line is fitting function $f(n)=\frac{a}{n^2}+\frac{b}{n}+c$.}
\vskip0.5cm
Ultimately, we expect the UV finite part of the ABJM Wilson loop for the regular polygon asymptotes at large $n$ limit to
\begin{align}
\big< W_{\Box}[C_n^{\rm{regular}}]\big>^{(2)}_{\rm{ABJM}}|_{\rm{finite}} &= \big< W_{\Box}C_n^{\rm{regular}}]\big>^{(2)}_{\rm{matter}}|_{\rm{finite}} + \big< W_{\Box}[C_n^{\rm{regular}}]\big>^{(2)}_{\rm{CS}}|_{\rm{finite}}\nonumber \\
&=\rho n + \big(\pi^2 -\frac{\pi^2}{6} \big) \label{result}
\end{align}
We replaced the $n$-independent constant by $\pi^2$, as evidenced by the above numerical fitting, and included $-\frac{\pi^2}{6}$ inherited from the $\big< W_\Box [C_n]\big>^{(2)}_{\rm CS}|_{\rm{finite}}$.
Let us compare this result with the previous results known for a spacelike circular ABJM Wilson loop.
In \cite{Rey:2008a}, the circular Wilson loop expectation value in ABJM theory was first computed and the result can be expanded in perturbative series of $\lambda$ was
\begin{equation}
\big< W_{\Box}[\bigcirc]\big>_{\rm{ABJM}}=1+\lambda^2 \big(\pi^2 -\frac{\pi^2}{6} \big)+ {\cal O}(\lambda^4).
\end{equation}
Also, using exact supersymmetry localization technique, the circular Wilson loop expectation value in ABJM theory was computed \cite{Marino:2009a}. The result, expanded in perturbative series of $\lambda$, is given by
\begin{equation}
\big< W_\Box[\bigcirc] \big>_{\rm ABJM} = 1+\frac{5\pi^2}{6} \lambda^2 + \mathcal{O}(\lambda^3)
\end{equation}
Our results in the $n \rightarrow \infty$ thermodynamic limit (\ref{result}) reproduces these previous results, thus passing a nontrivial consistency check.
\section{Comments on ABJ Theory}
In this paper, we studied quantum properties of the lightlike polygon Wilson loop expectation value in the ABJM theory. By invoking lightlike factorization and recursion relations thereof, we obtained the result for arbitrary polygon $C_n$ up to two loops. As a checkpoint, we took the continuum limit $n \rightarrow \infty$ and extracted the expectation value of spacelike circular Wilson loop. The result matched perfectly with previously known analytic results.
There are also variant conformal field theories related to the ABJM theory. The so-called ABJ theory has gauge group of different rank. As most of our considerations were based on the conformal invariance, it is of interest to explore the expectation value of the lightlike Wilson loop in this theory.
We already explained that the only difference of the ABJ theory from the ABJM theory is that the rank of gauge group is unequal: $G = U(N_1) \times U(N_2)$, but with the same Chern-Simons level. In the limit of infinite number of colors, $N_1, N_2 \rightarrow \infty$ , there are now two `t Hooft coupling constants, $\lambda_1 \equiv \frac{N_1}{k}$ and $\lambda_2 \equiv \frac{N_2}{k}$. The contribution of pure Chern-Simons system come from loops of gauge fields in the adjoint representations. The corresponding diagrams have weights $N_1^2$ or $N_2^2$. On the other hands, the contribution of matter system come from loops of bosons and fermions in the bi-fundamental representations. The corresponding diagrams have weight $N_1 N_2$. The one-loop corrections vanish for the same reasons as the ABJM theory. The two-loop corrections are nonzero and have weights $\frac{1}{2}(\lambda_1^2+\lambda_2^2)$ for the pure Chern-Simons contribution and $\lambda_1 \lambda_2$ for the matter contribution. Reflecting this fact, we obtain Wilson loop expectation value in ABJ theory.
\begin{align}
\big< W_{\Box}[C_n]\big>^{(2)}_{\mbox{\small{ABJ}}}&=\lambda_1 \lambda_2 \Big[ - {1 \over 2} \sum_{i=1}^n \frac{(x_{i,i+2}^2 4\pi e^{\gamma_E} \mu^2)^{2\epsilon}}{(2 \epsilon)^2}
+ \mbox{BDS}^{(2)}_n - \frac{1}{16}n \pi^2 \Big]
\nonumber \\
& \frac{\lambda_1^2+\lambda_2^2}{2}\Big[ - \frac{\log(2)}{2} \sum_{i=1}^n \frac{(x_{i,i+2}^2 \pi e^{\gamma_E} \mu^2)^{2\epsilon}}{2\epsilon} +\frac{7n\pi^2}{48} - \frac{n}{2}\log^2(2) - \frac{\pi^2}{6} \Big]
\end{align}
As for the ABJM theory, we might combine the pure Chern-Simons contribution and the matter contribution to a single expression. Doing so, one would get
\bea
\big< W_{\Box}[C_n]\big>^{(2)}_{\mbox{\small{ABJ}}}=\Delta \Big[ &-& \sum_{i=1}^n \frac{\big(x^2_{i, i+2} \mu'^{2}\big)^{2\epsilon}}{8\epsilon^2} + \mbox{BDS}_n^{(2)}
\nonumber \\
&+&\frac{1}{96} \Big( \pi^2 \big( 8(n-2)+(7n-8) \delta^2 \big) - 6n(2+\delta^2)(6+\delta^2) \log^2(2) \Big) \Big] \label{ABJ_res}
\eea
Here, $\mu'^2$ is
\begin{equation}
\mu'^2= \mu^2 \pi e^{\gamma_E} 2^{3+\frac{1}{2} \delta^2}
\end{equation}
and
\begin{equation}
\Delta^2= \lambda_1 \lambda_2, \quad \quad \quad \delta \equiv \frac{\lambda_1 - \lambda_2}{\sqrt{\lambda_1 \lambda_2}}
\end{equation}
Hence, ABJM result obtained at $\Delta \rightarrow \lambda$ and $\delta \rightarrow 0$ limit. Also, the result exhibits uniform transcendentality as for the ABJM theory.
The result (\ref{ABJ_res}), however, somewhat weird. Suppose there is exact result on Wilson loop expectation value of ABJ theory. Then, perturbative expansion with respect to $\lambda_1$ or $\lambda_2$ should reproduce (\ref{ABJ_res}). But this is impossible, shifted energy scale $\mu'^2$ contains $2^{\frac{(\lambda_1-\lambda_2)^2}{2\lambda_1 \lambda_2}}$ term. Therefore, combining matter dependent part and pure Chern-Simons part looks not consistent with perturbative expansion.
\section*{Acknowledgement}
We thank Benjamin Basso, James Drummond, Claude Duhr, Johannes Henn, Paul Heslop, Yu-Tin Huang, Hyunsoo Min and, especially, Jaesung Park for numerous helpful discussions. SJR acknowledges organizers and participants of IAS Focused Program "Scattering Amplitudes in Hong Kong" (17-21 November, 2014) for excellent scientific program, constructive inputs and stimulating discussions. This work was supported in part by the National Research Foundation of Korea grants 2005-0093843, 2010-220-C00003 and 2012K2A1A9055280.
\section*{Appendix}
|
\section{Introduction}
In recent years many physicists and some mathematician started to be interested in the possibility of giving a physical meaning to some non self-adjoint hamiltonian with real eigenvalues. The whole story started essentially with Bender and Boettcher in 1998, \cite{ben1}, with the famous $p^2+i\,x^3$ hamiltonian. Since then, hundreds of papers have been written, mainly from physical and numerical points of view. On the other hand, the mathematically oriented papers were much less. More recently, the number of such papers increased significantly also because several authors, mainly coming from functional analysis and operator algebras, joined the community starting to be interested into this topic. Some of these papers are listed in \cite{mathpap}, \cite{mathpap2} and \cite{mathpap3}, where more references can be found.
In several contributions different authors discuss some aspects of manifestly non self-adjoint hamiltonians which are quadratic in the position and in the momentum operators, or in some of their combinations. Quite often, in their analysis, they are able to deduce the explicit form of the eigenvectors, which are not mutually orthogonal, and of the related eigenvalues, which are real. In some previous papers by one of us (FB), it has been shown that some of the proposed models can be discussed in terms of the so-called $\D$-pseudo bosons ($\D$-PBs), that is of some {\em excitations} arising from properly deformed commutation rules. In this paper we set up a systematic analysis of several two-dimensional models which can be described completely in terms of $\D$-PBs, some of them already considered in the past and others, in our knowledge, new. Doing so, we propose a list of models which are simply {\em $\D$-deformed two-dimensional harmonic oscillators}, for which the approach discussed in, say, \cite{bagnewpb} and reviewed in \cite{bagbook} can be adopted (see also \cite{baggar} for models involving anti-commutation relations).
This article is organized as follows: to keep the paper self-contained, in the next section we review the definition and few results on $\D$ pseudo-bosons ($\D$-PBs). In Section III we discuss in details two models, deducing the eigenvectors of two different non self-adjoint hamiltonians. Among other things, we prove that the eigenstates do not form bases, but they are complete in $\Lc^2({\Bbb R}^2)$. In Section IV we give a list of other hamiltonians which can be discussed using the same techniques, and leading to similar conclusions. Section V contains our final remarks.
\section{$\D$ pseudo-bosons}\label{sectII}
We briefly review here few facts and definitions on $\D$-PBs. More details can be found in \cite{bagnewpb} and \cite{bagbook}. To simplify the notation, we consider here the one-dimensional case, since nothing essential changes going from one to more dimensions\footnote{This is true for commutative models, which are the ones considered all along this paper. When dealing with non-commutative quantum mechanics, differences may arise.}.
Let $\Hil$ be a given Hilbert space with scalar product $\left<.,.\right>$ and related norm $\|.\|$. Let further $a$ and $b$ be two operators
on $\Hil$, with domains $D(a)$ and $D(b)$ respectively, $a^\dagger$ and $b^\dagger$ their adjoint, and let $\D$ be a dense subspace of $\Hil$
such that $a^\sharp\D\subseteq\D$ and $b^\sharp\D\subseteq\D$, where $x^\sharp$ is $x$ or $x^\dagger$. Incidentally, it may be worth noticing
that we are not requiring here that $\D$ coincides with, e.g. $D(a)$ or $D(b)$. Nevertheless, for obvious reasons, $\D\subseteq D(a^\sharp)$
and $\D\subseteq D(b^\sharp)$.
\begin{defn}\label{def21}
The operators $(a,b)$ are $\D$-pseudo bosonic ($\D$-pb) if, for all $f\in\D$, we have
\be
a\,b\,f-b\,a\,f=f.
\label{21}\en
\end{defn}
Sometimes, to simplify the notation, instead of (\ref{21}) we will simply write $[a,b]=\1$, having in mind that both sides of this equation
have to act on some $f\in\D$.
\vspace{2mm}
Our working assumptions are the following:
\vspace{2mm}
{\bf Assumption $\D$-pb 1.--} there exists a non-zero $\varphi_{ 0}\in\D$ such that $a\,\varphi_{ 0}=0$.
\vspace{1mm}
{\bf Assumption $\D$-pb 2.--} there exists a non-zero $\Psi_{ 0}\in\D$ such that $b^\dagger\,\Psi_{ 0}=0$.
\vspace{2mm}
Then, if $(a,b)$ satisfy Definition \ref{def21}, it is obvious that $\varphi_0\in D^\infty(b):=\cap_{k\geq0}D(b^k)$ and that $\Psi_0\in D^\infty(a^\dagger)$, so
that the vectors \be \varphi_n:=\frac{1}{\sqrt{n!}}\,b^n\varphi_0,\qquad \Psi_n:=\frac{1}{\sqrt{n!}}\,{a^\dagger}^n\Psi_0, \label{22}\en
$n\geq0$, can be defined and they all belong to $\D$. We introduce the sets $\F_\Psi=\{\Psi_{ n}, \,n\geq0\}$ and
$\F_\varphi=\{\varphi_{ n}, \,n\geq0\}$. Since each
$\varphi_n$ and each $\Psi_n$ belong to $\D$, they also belong to the domains of $a^\sharp$, $b^\sharp$, $N=ba$ and $N^\dagger=a^\dagger b^\dagger$. We have
\be
\left\{
\begin{array}{ll}
b\,\varphi_n=\sqrt{n+1}\varphi_{n+1}, \qquad\qquad\quad\,\, n\geq 0,\\
a\,\varphi_0=0,\quad a\varphi_n=\sqrt{n}\,\varphi_{n-1}, \qquad\,\, n\geq 1,\\
a^\dagger\Psi_n=\sqrt{n+1}\Psi_{n+1}, \qquad\qquad\quad\, n\geq 0,\\
b^\dagger\Psi_0=0,\quad b^\dagger\Psi_n=\sqrt{n}\,\Psi_{n-1}, \qquad n\geq 1,\\
\end{array}
\right.
\label{23}\en as well as the following eigenvalue equations: $N\varphi_n=n\varphi_n$ and $N^\dagger\Psi_n=n\Psi_n$, $n\geq0$. Then, choosing the normalization of $\varphi_0$ and $\Psi_0$ in such a way $\left<\varphi_0,\Psi_0\right>=1$, we deduce that
\be \left<\varphi_n,\Psi_m\right>=\delta_{n,m}, \label{34}\en
for all $n, m\geq0$, so that $\F_\varphi$ and $\F_\Psi$ are biorthogonal sets. The third assumption introduced in \cite{bagnewpb} is the following:
\vspace{2mm}
{\bf Assumption $\D$-pb 3.--} $\F_\varphi$ is a basis for $\Hil$.
\vspace{1mm}
This is equivalent to the request that $\F_\Psi$ is a basis for $\Hil$ as well, \cite{bagnewpb}. In particular, if $\F_\varphi$ and $\F_\Psi$ are Riesz basis for $\Hil$, the $\D$-PBs are called {\em regular}\footnote{We recall that a set $\E=\{e_n\in\Hil,\,n\geq0\}$ is a (Schauder) basis for $\Hil$ if any vector $f\in\Hil$ can be written, uniquely, as an (in general)
infinite linear combination of the $e_n$'s: $f=\sum_{n=0}^\infty c_n(f)e_n$. Here $c_n(f)$ are complex numbers depending on the vector $f$ we
want to expand. If $\E$ is an orthonormal basis, that is when we also have $\left<e_n,e_m\right>=\delta_{n,m}$, then $c_n(f)=\left<e_n,f\right>$.
A Riesz basis $\F=\{f_n\in\Hil,\,n\geq0\}$ is a set of vectors for which a bounded
operator $T$ on $\Hil$ exists, with bounded inverse, and an orthonormal basis $\E=\{e_n\in\Hil,\,n\geq0\}$, such that $f_n=Te_n$, for all
$n\geq0$. Also, a set of vectors $\F$ is complete if the only vector $h\in\Hil$ which is orthogonal to all the vectors in $\F$ is the zero vector.}
\vspace{2mm}
In \cite{bagnewpb} a weaker version of Assumption $\D$-pb 3 has also been introduced, particularly useful for physical applications: for that, let $\G$ be a suitable dense subspace of $\Hil$. Two biorthogonal sets $\F_\eta=\{\eta_n\in\G,\,n\geq0\}$ and $\F_\Phi=\{\Phi_n\in\G,\,n\geq0\}$ have been called {\em $\G$-quasi bases} if, for all $f, g\in \G$, the following holds:
\be
\left<f,g\right>=\sum_{n\geq0}\left<f,\eta_n\right>\left<\Phi_n,g\right>=\sum_{n\geq0}\left<f,\Phi_n\right>\left<\eta_n,g\right>.
\label{25}
\en
Is is clear that, while Assumption $\D$-pb 3 implies (\ref{25}), the reverse is false. However, if $\F_\eta$ and $\F_\Phi$ satisfy (\ref{25}), we still have some (weak) form of resolution of the identity. Now Assumption $\D$-pb 3 may be replaced by the following:
\vspace{2mm}
{\bf Assumption $\D$-pbw 3.--} $\F_\varphi$ and $\F_\Psi$ are $\G$-quasi bases, for some subspace $\G$ dense in $\Hil$.
\vspace{2mm}
Let us now assume that Assumption $\D$-pb 1, $\D$-pb 2, and $\D$-pbw 3 are satisfied. Sometimes, even if it is not strictly necessary, it is convenient to consider $\G=\D$. Then, let us consider a self-adjoint, invertible, operator $\Theta$, which leaves, together with $\Theta^{-1}$, $\D$ invariant: $\Theta\D\subseteq\D$, $\Theta^{-1}\D\subseteq\D$. Hence, \cite{bagnewpb}, we say that $(a,b^\dagger)$ are $\Theta-$conjugate if $af=\Theta^{-1}b^\dagger\,\Theta\,f$, for all $f\in\D$. We recall that $(a,b^\dagger)$ are $\Theta-$conjugate if and only if $(b,a^\dagger)$ are $\Theta-$conjugate. Moreover,
$(a,b^\dagger)$ are $\Theta-$conjugate if and only if $\Psi_n=\Theta\varphi_n$, for all $n\geq0$. Furthermore,
if $(a,b^\dagger)$ are $\Theta-$conjugate, then $\left<f,\Theta f\right>>0$ for all non zero $f\in \D$. These results are all proved in the first paper in \cite{mathpap2}, where also other details of $\D$-PBs, including some interesting intertwining relations, are discussed. In view of the concrete applications considered here it might be useful to stress that, in some explicit models, \cite{bagbook}, $\D$ and $\G$ must be taken different. Of course, this has nothing to do with the Assumptions $\D$-pb 1, $\D$-pb 2, and $\D$-pb 3 (or $\D$-pbw 3).
\section{The two-dimensional deformed harmonic oscillators: two detailed examples}\label{SectIII}
The main ingredient of our construction is a standard two dimensional harmonic oscillator whose hamiltonian can be written as
\be
H_0=\omega_1A_1^\dagger A_1+\omega_2A_2^\dagger A_2+\omega_3\1,
\label{31}
\en
where $\omega_j\in{\Bbb R}$, and $[A_j,A_k^\dagger]=\delta_{j,k}\,\1$, $j,k=1,2$, all the other commutators being zero. It is clear that $H_0=H_0^\dagger$. The Hilbert space of the model is $\Hil=\Lc^2({\Bbb R}^2)$, with scalar product $\left<f,g\right>=\int_{\Bbb R}\int_{\Bbb R}\overline{f(x_1,x_2)}\,g(x_1,x_2)\,dx_1\,dx_2$.
What is interesting for us is to consider some {\em deformed} versions of $H_0$, which we write (formally, for the moment) as
$$
H=e^XH_0e^{-X},
$$
for some suitable operator $X$. Because of $(\ref{31})$, introducing (again formally at this stage) $a_j=e^XA_je^{-X}$ and $b_j=e^XA_j^\dagger e^{-X}$, we can write $H$ as follows
\be
H=\omega_1b_1 a_1+\omega_2b_2 a_2+\omega_3\1,
\label{32}
\en
and we have $[a_j,b_k]=\delta_{j,k}\,\1$, $j,k=1,2$. Hence these operators appear to satisfy the pseudo-bosonic commutation rules, since, in general, $b_j\neq a_j^\dagger$, due to the fact that $X^\dagger$ is not assumed to coincide with $-X$.
But, as we have already pointed out several times, this is formal. The reason is simple: most of the times the operators involved in this procedure are unbounded, so that one should pay attention to domain problems. In particular, since $H_0$ is unbounded, if $X$ is also unbounded, there is no reason, a priori, for $H$ to be well defined on a dense (or, even, simply non empty) subset of $\Hil$. This depends, of course, on the form of $X$. For the same reason, the commutator $[a_j,b_k]$ could not be densely defined. This has been discussed, for instance, in \cite{mathpap3}.
In what follows, we will make this transformation concrete and rigorous, by checking in details, for two different choices of $X$,
the validity of Assumptions $\D$-pb 1, $\D$-pb 2 and $\D$-pb 3, or its weaker form, $\D$-pbw 3.
In order to help the reader to identify the expression of $H$ we adopt in this paper, as dynamical variables, the position and momentum operators, rather than raising and lowering operators\footnote{The relation between them is the usual one: $A_j=\frac{x_j+ip_j}{\sqrt{2}}$, $j=1,2$.}. This, we believe, can be useful, although not necessary, since many models, in the existing physical literature on the subject, are written adopting this choice. Therefore we rewrite \eqref{31} as
\be
H_0=\tilde\omega_1(x_1^2+p_1^2)+\tilde\omega_2(x_2^2+p_2^2)+\tilde\omega_3\1,
\label{33}
\en
where, to simplify the notation, we have introduced $\tilde\omega_1=\omega_1/2$, $\tilde\omega_2=\omega_2/2$, $\tilde\omega_3=\omega_3-(\omega_1+\omega_2)/2 $. Here, as in ordinary (i.e. commutative) quantum mechanics, $[x_j,p_k]=i\delta_{j,k}\,\1$, $j,k=1,2$, while all the other commutators are assumed to be trivial.
\subsection{The first model}\label{secIII1}
In this section we consider the following quadratic choice for $X$:
\be
X=\gamma(\sqrt{2}(x_1+x_2)+2x_1x_2),
\label{310}
\en
where $\gamma\in {\Bbb R}$. We see that the momentum operators do not appear in $X$, and for this reason $U(\pm\gamma):=e^{\pm X}$ are just (unbounded, and therefore not everywhere defined) multiplication operators. In term of the bosonic operators $A_j$, $A_j^\dagger$, $X$ looks as
\be
X=\gamma(A_1^\dagger A_2+A_2^\dagger A_1+A_1 A_2+A_1^\dagger A_2^\dagger+A_1+A_1^\dagger +A_2+A_2^\dagger),
\label{310b}
\en
which appears rather more complicated.
The operators $a_j$ and $b_j$ introduced before can be rewritten as
\be
\left\{
\begin{array}{ll}
a_1=(x_1+ip_1)/\sqrt{2}-\gamma\left(x_2\sqrt{2}+\1\right), \quad a_2=(x_2+ip_2)/\sqrt{2}-\gamma\left(x_1\sqrt{2}+\1\right),\\
b_1=(x_1-ip_1)/\sqrt{2}+\gamma\left(x_2\sqrt{2}+\1\right), \quad b_2=(x_2-ip_2)/\sqrt{2}+\gamma\left(x_1\sqrt{2}+\1\right),\\
\end{array}
\right.
\label{311}\en
or as
\be
\left\{
\begin{array}{ll}
a_1=A_1-\gamma\left(A_2+A_2^\dagger+\1\right), \quad a_2=A_2-\gamma\left(A_1+A_1^\dagger+\1\right),\\
b_1=A_1^\dagger+\gamma\left(A_2+A_2^\dagger+\1\right), \quad b_2=A_2^\dagger+\gamma\left(A_1+A_1^\dagger+\1\right).\\
\end{array}
\right.
\label{311bis}\en
It is easy to check that they obey, formally, the pseudo-bosonic rule $[a_j,b_k]=\delta_{j,k}\1$, $j,k=1,2$,
the other commutators being zero. However, to go from formal to rigorous results, we now take a completely
different point of view, showing that the four operators \underline{defined} as in (\ref{311}), or as in (\ref{311bis}), satisfy the two-dimensional version of Assumptions $\D$-pb 1, $\D$-pb 2 and $\D$-pbw 3 of Section II. We will also show that assumption $\D$-pb 3 does not hold, so that the vectors we will construct extending (\ref{22}) are not bases for $\Hil$.
First, using (\ref{311}), we rewrite equations $a_1\varphi_{0,0}=a_2\varphi_{0,0}=0$ in a differential form as
$$
\frac{\partial\varphi_{0,0}(x_1,x_2)}{\partial x_1}+\left(x_1-\gamma(2x_2+\sqrt{2})\right)\varphi_{0,0}(x_1,x_2)=0
$$
and
$$
\frac{\partial\varphi_{0,0}(x_1,x_2)}{\partial x_2}+\left(x_2-\gamma(2x_1+\sqrt{2})\right)\varphi_{0,0}(x_1,x_2)=0.
$$
The solution is easily deduced:
\be
\varphi_{0,0}(x_1,x_2)=N\,e^{-\frac{1}{2}(x_1^2+x_2^2)+\sqrt{2}\,\gamma(x_1+x_2)+2\gamma x_1 x_2},
\label{312}\en
where $N$ is a normalization constant we will fix in the following. Now, since $b_j^\dagger$ coincides with $a_j$, but with $\gamma$ replaced by $-\gamma$, it is clear that the solution of $b_1^\dagger\Psi_{0,0}=b_2^\dagger\Psi_{0,0}=0$ can be deduced from (\ref{312}) with a similar replacement:
\be
\Psi_{0,0}(x_1,x_2)=N'\,e^{-\frac{1}{2}(x_1^2+x_2^2)-\sqrt{2}\,\gamma(x_1+x_2)-2\gamma x_1 x_2}.
\label{313}\en
Here $N'$ is another normalization constant, which needs not to coincide with $N$. A direct computation shows that, if $\gamma\notin\left]-\frac{1}{2},\frac{1}{2}\right[$, then both $\varphi_{0,0}(x_1,x_2)$ and $\Psi_{0,0}(x_1,x_2)$ are not square integrable. On the other hand, if we take $\gamma\in\left]-\frac{1}{2},\frac{1}{2}\right[ $ and we choose $N=N'=\frac{1}{\sqrt{\pi}}$, then $\left<\varphi_{0,0},\Psi_{0,0}\right>=1$. Among other things this means that, even if formally $(a_j,b_j)$ satisfy the pseudo-bosonic rules for all possible values of $\gamma$, they are surely not $\D$-pseudo bosonic operators if $\gamma\notin\left]-\frac{1}{2},\frac{1}{2}\right[$. For this reason, from now on we will assume that $\gamma \in \left]-\frac{1}{2},\frac{1}{2}\right[ $. In this case both $\varphi_{0,0}(x_1,x_2)$ and $\Psi_{0,0}(x_1,x_2)$ are square integrable and, more than this, they belong to $\Sc({\Bbb R}^2)$, the set of $C^\infty$-functions which decrease to zero, together with their derivatives, faster than any inverse power. For future convenience, it is also worth noticing that they also both belong to the set
$$
\D:=\left\{f(x_1,x_2)\in\Sc({\Bbb R}^2):\, e^{\beta_1x_1+\beta_2x_2}f(x_1,x_2)\in\Sc({\Bbb R}^2),\, \forall \beta_1,\beta_2\in{\Bbb C}\right\}.
$$
This set, which we are taking as our $\D$, is dense in $\Lc^2(\Bbb R^2)$, since it contains $D(\Bbb R^2)$, the set of the $C^\infty$ functions with compact support. Moreover, as required, $\D$ is stable under the action of both $a_j^\sharp$ and $b_j^\sharp$, and in fact we deduce that, by using a two-dimensional version of (\ref{22}),
\be
\varphi_{n_1,n_2}(x_1,x_2)=\frac{\tilde N}{\sqrt{2^{n_1+n_2}\,n_1!\,n_2!}}\,H_{n_1}(x_1)H_{n_2}(x_2)\,e^{-\frac{1}{2}(x_1^2+x_2^2)+\sqrt{2}\gamma(x_1+x_2)+2\gamma x_1x_2},
\label{313b}\en
for all $n_j\geq0$. Here $\tilde N$ is a suitable normalization, related to $N$, which is not particularly important in our analysis. Notice that, not surprisingly, each $\varphi_{n_1,n_2}$ belongs to $\D$ (so that, in particular, it belongs to $\Sc(\Bbb R^2)$). Incidentally, it is worth mentioning that $\D$ is invariant also under the action of $e^{\pm X}$.
The function $\Psi_{n_1,n_2}(x_1,x_2)$ can be deduced by $\varphi_{n_1,n_2}(x_1,x_2)$ simply replacing $\gamma$ with $-\gamma$. Therefore, also $\Psi_{n_1,n_2}(x_1,x_2)$ belong to $\D$. Our conclusion, so far, is that Assumptions $\D$-pb 1 and $\D$-pb 2 are indeed satisfied. To check whether $\D$-pb 3 is also satisfied, we first define the sets $\F_\Psi=\{\Psi_{n_1,n_2}(x_1,x_2),\,n_j\geq0\}$ and $\F_\varphi=\{\varphi_{n_1,n_2}(x_1,x_2),\,n_j\geq0\}$. Their vectors are eigenstates of $N_j$ and $N_j^\dagger$: $N_j\varphi_{n_1,n_2}=n_j\varphi_{n_1,n_2}$ and $N_j^\dagger\Psi_{n_1,n_2}=n_j\Psi_{n_1,n_2}$, $j=1,2$, and are mutually orthogonal. To show that neither $\F_\Psi$ nor $\F_\varphi$ are bases we use a standard argument, \cite{dav}: we prove that the norms of $\varphi_{n_1,n_2}(x_1,x_2)$ and $\Psi_{n_1,n_2}(x_1,x_2)$ are both divergent for diverging $n_1$ or $
n_2$.
As a matter of fact, due to the relation between these vectors, it is enough to check that $\|\varphi_{n_1,n_2}\|$ diverges with $n_j$. Moreover, it is sufficient to consider the case when $n_2=0$, and to prove that $\|\varphi_{n,0}\|\rightarrow\infty$ when $n\rightarrow\infty$. Indeed, since $H_0(x)=1$, we easily find that
$$
I_n:=\frac{\|\varphi_{n,0}\|^2}{\tilde N^2}=\frac{1}{2^n\,n!}\sqrt{\frac{\pi}{1-4\gamma^2}}\,e^{\frac{4\gamma^2}{1-2\gamma}}\int_{\Bbb R} H_n\left(\frac{t+t_0}{\sqrt{1-4\gamma^2}}\right)^2\,e^{-t^2}\,dt.
$$
The integral can be now estimated by first changing variable: let us put $s=\frac{t}{\sqrt{1-4\gamma^2}}$. Then, since
$$
\int_{\Bbb R} H_n\left(\frac{t+t_0}{\sqrt{1-4\gamma^2}}\right)^2\,e^{-t^2}\,dt \geq \sqrt{1-4\gamma^2} \int_{\Bbb R} H_n(s+s_0)^2\,e^{-s^2}\,ds,
$$
where $s_0=\frac{t_0}{\sqrt{1-4\gamma^2}}=\frac{\sqrt{2}\,\gamma}{1-2\gamma}$, we deduce that
$$
I_n\geq\pi \,e^{\frac{4\gamma^2}{1-2\gamma}}L_n\left(-\left(\frac{2\gamma}{1-2\gamma}\right)^2\right),
$$
which, for all non zero allowed $\gamma$, diverges when $n\rightarrow\infty$, see \cite{szego}. A similar estimate can be repeated for $\|\Psi_{n,0}\|^2$. Hence, $\F_\varphi$ and $\F_\Psi$ are not bases. Still, it is possible to check that they are both complete in $\Lc^2(\Bbb R^2)$. This follows, for instance, from their analytical expression, see (\ref{313b}), and by a simple extension of the completeness argument for functions of the form $x^n f(x)$, where $|f(x)|\leq Ce^{-\delta x}$, $\delta>0$, and $n=0,1,2,\ldots$, to a two-dimensional case, see \cite{kolm}, pg 426\footnote{The same conclusion can be deduced following \cite{sjo}, Lemma 3.12}.
Now, even if $\F_\varphi$ and $\F_\Psi$ are not bases, Assumption $\D$-pbw 3 could still be true. This is important since, as it is proved in \cite{mathpap2,bagnewpb}, this milder condition is enough to deduce several interesting consequences.
We first observe that, since $a_j\varphi_{0,0}=U(\gamma)A_jU^{-1}(\gamma)\varphi_{0,0}=0$, then $\varphi_{0,0}$ belongs to the domain of $U^{-1}(\gamma)=U(-\gamma)$. Moreover, because of the uniqueness of the vacuum of $A_j$, $\Phi_{0,0}(x_1,x_2)$, we deduce that $\varphi_{0,0}(x_1,x_2)$ must be proportional to $U(\gamma)\Phi_{0,0}(x_1,x_2)$. Similarly, $\Psi_{0,0}$ belongs to the domain of $U^\dagger(\gamma)=U(\gamma)$, and it must be proportional to $U(-\gamma)\Phi_{0,0}(x_1,x_2)$. Moreover, because of condition $\left<\varphi_{0,0},\Psi_{0,0}\right>=1$, we can fix these proportionality constants as follows:
\be
\left\{
\begin{array}{ll}
\varphi_{0,0}(x_1,x_2)=U(\gamma)\Phi_{0,0}(x_1,x_2),\\
\Psi_{0,0}(x_1,x_2)=U(-\gamma)\Phi_{0,0}(x_1,x_2).\\
\end{array}
\right.
\label{314}
\en
Incidentally, this is in agreement with our previous remark on the role of $\gamma$ in $\varphi_{0,0}(x_1,x_2)$ and $\Psi_{0,0}(x_1,x_2)$. Now we can check explicitly that: (i) $\varphi_{n_1,n_2}(x_1,x_2)\in D(U(-\gamma))$, for all $n_j\geq0$; (ii) $\Psi_{n_1,n_2}(x_1,x_2)\in D(U(\gamma))$, for all $n_j\geq0$; (iii ) $\varphi_{n_1,n_2}(x_1,x_2)=U(\gamma)\Phi_{n_1,n_2}(x_1,x_2)$ and $\Psi_{n_1,n_2}(x_1,x_2)=U(-\gamma)\Phi_{n_1,n_2}(x_1,x_2)$, for all $n_j\geq0$. Here $$\Phi_{n_1,n_2}(x_1,x_2)=\frac{1}{\sqrt{n_1!\,n_2!}}\,{A_1^\dagger}^{n_1}{A_2^\dagger}^{n_2}\Phi_{0,0}(x_1,x_2)$$ are the (well known) eigenstates of the hamiltonian $H_0$ in (\ref{31}), which form an o.n. basis for $\Hil$.
Now, to check that $\F_\varphi$ and $\F_\Psi$ are $\G$-quasi bases for a suitable $\G$, we start defining this set as the (finite) linear span of the vectors $\Phi_{n_1,n_2}(x_1,x_2)$, which is dense in $\Hil$. Now\footnote{This is a consequence of the fact that, as we have already seen, the vectors $\varphi_{n_1,n_2}=U(\gamma)\Phi_{n_1,n_2}$ and $\Psi_{n_1,n_2}=U(-\gamma)\Phi_{n_1,n_2}$, for all $n_j\geq0$, are all well defined in $\Hil$.}, since $\G\subseteq D(U(\gamma))\cap D(U(-\gamma))$, we can check that, taken $f,g\in \G$,
$$
\sum_{n_1,n_2}\left<f,\varphi_{n_1,n_2}\right>\left<\Psi_{n_1,n_2},g\right>=\sum_{n_1,n_2}\left<f,U(\gamma)\Phi_{n_1,n_2}\right>\left<U(-\gamma)\Phi_{n_1,n_2},g\right>=
$$
\be
=\sum_{n_1,n_2}\left<U(\gamma)f,\Phi_{n_1,n_2}\right>\left<\Phi_{n_1,n_2},U(-\gamma)g\right>=\left<U(\gamma)f,U(-\gamma)g\right>=\left<f,g\right>.
\label{add2}\en
Hence, $\F_\Psi$ and $\F_\varphi$ are $\G$-quasi bases.
Consequences of the validity of the three assumptions is that we can introduce a new self-adjoint operator,
$\Theta(\gamma)$, such that $(a_j,b_j^\dagger)$ are $\Theta(\gamma)$ conjugate. Moreover, $\Theta(\gamma)$
is positive, and some intertwining relations hold.
{In fact, by considering the relations between the vectors of $\mathcal F_\varphi$ and $\mathcal F_\Psi$ with the vectors $\Phi_{n_1,n_2}$, we can write:
$$\varphi_{n_1 ,n_2}=U^2(\gamma)\Psi_{n_1 ,n_2}=U(2\gamma)\Psi_{n_1 ,n_2},$$
which, see Section II, suggests to define $\Theta(\gamma)=U(-2\gamma)=U^{-1}(2\gamma)$. Now, using $\Theta(\gamma)$ defined in this way, we can further easily deduce the following
(weak) intertwining relations:
\begin{equation}\label{intertw}
N_j^\dagger \Theta(\gamma)\varphi_{n_1 ,n_2}=\Theta(\gamma)N_j\varphi_{n_1 ,n_2}, \quad \Theta^{-1}(\gamma)N_j^\dagger\Psi_{n_1 ,n_2}=N_j \Theta^{-1}(\gamma)\Psi_{n_1 ,n_2},
\end{equation}
which, as we can see, are defined respectively on $\F_\varphi$ and $\F_\Psi$, but not of course, on the whole $\Hil$.
}
Moreover, $\varphi_{n_1,n_2}$ are eigenstates of $H=\omega_1b_1a_1+\omega_2b_2a_2+\omega_3\1=\omega_1N_1+\omega_2N_2+\omega_3\1$
with eigenvalues $E_{n_1,n_2}=\omega_1n_1+\omega_2n_2+\omega_3$, while $\Psi_{n_1,n_2}$ are eigenstates of $H^\dagger$ with the same eigenvalues.
Hence $\Theta(\gamma)$ also intertwines (on a suitable domain) between $H$ and $H^\dagger$.
Finally, in terms of the bosonic operators $x_j,p_j$, the deformed hamiltonian $H$ turns out to be the following operator:
\bea
H=x_1^2(\tilde\omega_1-4\gamma^2\tilde\omega_2)+x_2^2(\tilde\omega_2-4\gamma^2\tilde\omega_1)+\tilde\omega_1 p_1^2+\tilde\omega_2 p_2^2-4\sqrt{2}\gamma^2\tilde\omega_2 x_1-4\sqrt{2}\gamma^2\tilde\omega_1 x_2+\nonumber\\
+2\sqrt{2}i\tilde\omega_1 \gamma p_1
+2\sqrt{2}i\tilde\omega_2 \gamma p_2+
4i\tilde\omega_2 \gamma x_1 p_2+4i\tilde\omega_1 \gamma x_2 p_1+\nonumber\\
+(\tilde\omega_3-2\gamma^2 \tilde\omega_1-2\gamma^2 \tilde\omega_2+2\gamma \tilde\omega_1+2\gamma \tilde\omega_2)\1,\label{add3}
\ena
which, apart an additive constant, looks like an asymmetric two-dimensional harmonic oscillator with a manifestly non self-adjoint perturbation, with a linear and a quadratic parts. Therefore we can conclude that the non self-adjoint hamiltonian (\ref{add3}) is just a very complicated way to write a much simpler hamiltonian, whose eigenvalues and eigenvectors can be easily deduced adopting the strategy described in Section II. Also, it is easy to find the eigenvectors of $H^\dagger$ with the same strategy. What it is also interesting is that neither of these two sets are bases for $\Hil$, but they are (both) $\G$-quasi bases.
\subsection{The second model}\label{secIII2}
We now consider the following choice for $X$:
\be
X=\gamma(x_1x_2 +p_1p_2),
\label{320}
\en
where $\gamma\in {\Bbb R}$. This is again quadratic, but compared with our previous choice,
involves also the momentum operators. In (\ref{33}) we assume that $\tilde\omega_1\neq\tilde\omega_2$ since, otherwise, the situation trivializes (see the expression \eqref{316h} for $H$ below),
meaning that the hamiltonian becomes, if $\tilde\omega_1=\tilde\omega_2$, a purely two dimensional, self-adjoint, harmonic oscillator.
As before we introduce $U(\gamma)=e^X$, and using the Baker-Campbell-Hausdorff formula, the operators $a_j$ and $b_j$ are found to be
\be
\left\{
\begin{array}{ll}
a_1=(C_{\gamma}x_1-S_{\gamma}x_2+i(C_{\gamma} p_1 -S_{\gamma}p_2))/\sqrt{2}=C_{\gamma} A_1-S_{\gamma}A_2,\\
a_2=(C_{\gamma}x_2-S_{\gamma}x_1+i(C_{\gamma} p_2 -S_{\gamma}p_1))/\sqrt{2}=C_{\gamma} A_2-S_{\gamma}A_1,\\
b_1=(C_{\gamma}x_1+S_{\gamma}x_2-i(C_{\gamma} p_1 +S_{\gamma}p_2))/\sqrt{2}=C_{\gamma} A_1^\dagger+S_{\gamma}A_2^\dagger,\\
b_2=(C_{\gamma}x_2+S_{\gamma}x_1-i(C_{\gamma} p_2 +S_{\gamma}p_1))/\sqrt{2}=C_{\gamma} A_2^\dagger+S_{\gamma}A_1^\dagger,\\
\end{array}
\right.
\label{321}\en
where $C_{\gamma}:=\cosh{\gamma}$ and $S_{\gamma}:=\sinh{\gamma}$, while the deformed hamiltonian $H$ in (\ref{32}) becomes
\begin{eqnarray}
\label{316h}
H=(\tilde\omega_1 C_{\gamma}^2-\tilde\omega_2 S_{\gamma}^2)x_1^2 +(\tilde\omega_2 C_{\gamma}^2-\tilde\omega_1 S_{\gamma}^2)x_2^2 + (\tilde\omega_1 C_{\gamma}^2-\tilde\omega_2 S_{\gamma}^2)p_1^2+(\tilde\omega_2 C_{\gamma}^2-\tilde\omega_1 S_{\gamma}^2)p_2^2+\\
\nonumber +2i(\tilde\omega_2 - \tilde\omega_1) C_{\gamma} S_{\gamma}(x_1p_2-x_2p_1)+\tilde\omega_3\1.
\end{eqnarray}
Because of the (\ref{321}) it is clear that the vacua of $a_j$, $\varphi_{0,0}(x_1,x_2)$, and of $b_j^\dagger$, $\Psi_{0,0}(x_1,x_2)$, both coincide with the vacuum of the standard bosonic operators $A_j$, $\Phi_{0,0}(x_1,x_2)$:
$$
\varphi_{0,0}(x_1,x_2)=\Psi_{0,0}(x_1,x_2)=\Phi_{0,0}(x_1,x_2).
$$
By introducing the shorthand notation
$$\xi (a,b,c,d)=\sqrt{\frac{(a+b-c-d)!(c+d)!}{a!b!}}, \ \ \ a,b,c,d \in \mathds N,$$
the expressions of $\varphi_{n_1,n_2}$ and $\psi_{m_1,m_2}$ can be written as follows:
$$
\varphi_{n_1,n_2}(x_1,x_2)=\frac{1}{\sqrt{n_1!n_2!}}b_1^{n_1}b_2^{n_2}\Phi_{0,0}(x_1,x_2)= \\$$$$
=\sum_{k=0}^{n_1}\sum_{j=0}^{n_2}\binom{n_1}{k}\binom{n_2}{j}C_{\gamma}^{n_1+j-k}S_{\gamma}^{n_2+k-j}\xi (n_1, n_2, j, k)\Phi_{n_1+n_2-j-k,j+k}(x_1,x_2)=$$
$$
=\binom{n_1+n_2}{n_1}^{\frac{1}{2}}\sum_{k=0}^{n_1}\sum_{j=0}^{n_2}\binom{n_1}{k}\binom{n_2}{j}\binom{n_1+n_2}{j+k}^{-\frac{1}{2}}C_{\gamma}^{n_1+j-k}
S_{\gamma}^{n_2+k-j}\Phi_{n_1+n_2-j-k,j+k}(x_1,x_2),
$$
and
$$
\Psi_{m_1,m_2}(x_1,x_2)=\frac{1}{\sqrt{m_1!m_2!}}(a_1^\dagger)^{m_1}(a_2^\dagger)^{m_2}\Phi_{0,0}(x_1,x_2)=\\$$
$$
=\sum_{i=0}^{m_1}\sum_{l=0}^{m_2}\binom{m_1}{i}\binom{m_2}{l}\bar C_{\gamma}^{m_1+l-i}(-\bar S_{\gamma})^{m_2+i-l}\xi (m_1,m_2,l,i)\Phi_{m_1+m_2-l-i,l+i}(x_1,x_2)=$$$$
=\binom{m_1+m_2}{m_1}^{\frac{1}{2}}\sum_{i=0}^{m_1}\sum_{l=0}^{m_2}\binom{m_1}{i}\binom{m_2}{l}\binom{m_1+m_2}{l+i}^{-\frac{1}{2}}\bar C_{\gamma}^{m_1+l-i}(-\bar S_{\gamma})^{m_2+i-l}\Phi_{m_1+m_2-i-l,i+l}(x_1,x_2).
$$
Hence, as these formulas show, the eigenstates of $H$ and $H^{\dagger}$ (and obviously of $N_j$ and $N_j^{\dagger}$)
are linear combinations of the eigenstates of $H_0$. Using the operator $U(\gamma)$, and repeating the same reasonings as before, they can be written as
$$\varphi_{n_1,n_2}=U(\gamma)\Phi_{n_1 , n_2} \ \ \ \ \ \Psi_{m_1,m_2}=(U^{-1}(\gamma))^{\dagger}\Phi_{m_1 , m_2}=U^{-1}(\gamma)\Phi_{m_1 , m_2}.$$
From the biorthogonality condition we deduce the following, non trivial, summation rule:
$$\left \langle\Psi_{m_1,m_2},\varphi_{n_1,n_2} \right \rangle=$$$$ =\sum_{i=0}^{m_1}\sum_{l=0}^{m_2}\sum_{k=0}^{n_1}\sum_{j=0}^{n_2}\binom{m_1}{i}\binom{m_2}{l}\binom{n_1}{k}\binom{n_2}{j} \xi (n_1, n_2, j, k) \xi (m_1,m_2,l,i) \times$$ $$ \times C_{\gamma}^{m_1+n_1+l-i+j-k}(- 1)^{m_2+i-l}S_{\gamma}^{m_2+n_2+i-l+k-j}\delta_{m_1+m_2,n_1+n_2}\delta_{l+i,j+k}=$$$$
=\delta_{m_1,n_1}\delta_{m_2,n_2}.
$$
To check if the two sets $\mathcal F_{\varphi}$ and $\mathcal F_{\Psi}$ are bases of $\mathcal L^2 (\mathds R^2)$,
we check, as in the previous section, whether $\|\varphi_{n_1,n_2}\|$ diverges with $n_j$. For simplicity, we set $n_2 =0$, as in Section \ref{secIII1}. Then
$$
\lVert\varphi_{n_1,0} \rVert^2=$$
$$=\left\langle\frac{1}{\sqrt{n_1!}}\sum_{k=0}^{n_1}\binom{n_1}{k}C_{\gamma}^{n_1-k}S_{\gamma}^k\sqrt{(n_1-k)!k!}\phi_{n_1-k,k},
\frac{1}{\sqrt{n_1!}}\sum_{i=0}^{n_1}\binom{n_1}{i}C_{\gamma}^{n_1-i}S_{\gamma}^i\sqrt{(n_1-i)!i!}\phi_{n_1-i,i}\right\rangle=
$$
$$
=\frac{1}{n_1!}\sum_{k=0}^{n_1}\sum_{i=0}^{n_1}\binom{n_1}{k}\binom{n_1}{i}C_{\gamma}^{2n_1-k-i}S_{\gamma}^{k+i}\sqrt{(n_1-k)!k!}\sqrt{(n_1-i)!i!}\delta_{k,i}=
$$
$$
=\sum_{k=0}^{n_1}\binom{n_1}{k}(|C_{\gamma}|^2)^{n_1-k}(|S_{\gamma}|^2)^k=(|C_{\gamma}|^2+|S_{\gamma}|^2)^{n_1}=\cosh^{n_1}(2\gamma)
$$
Since the hyperbolic cosine of a real number is always greater than one, for each $\gamma\neq0$, the norms of these vectors diverge as $n_1\rightarrow\infty$.
The same results is obviously obtained if we put $n_1=0$ and consider $n_2\rightarrow\infty$. Similar conclusions can be deduced working with $\Psi_{m_1,m_2}$.
Therefore
the sets of eigenstates of $N_j$ and $N_j^\dagger$ are not biorthonormal bases, although they are still complete in $\mathcal L^2(\mathds R^2)$, for the same reasons discussed in Section \ref{secIII1}.
However, we can check easily, repeating the same arguments as before, that
$\mathcal F_{\varphi}$ and $\mathcal F_{\Psi}$ are $\mathcal G$-quasi bases, where $\G$ is the linear span of the vectors $\Phi_{n_1,n_2}(x_1,x_2)$ as in the first model.
Due to the validity of the three assumptions $\D$-pb1, $\D$-pb2, $\D$-pbw3, we can consider the self-adjoint operator defined as in the previous section
as $\Theta(\gamma)=U(-2\gamma)$, such that $(a_j,b_j^\dagger)$ are $\Theta(\gamma)$-conjugate. Moreover, $\Theta(\gamma)$
is positive and the same intertwining relations as in \eqref{intertw} hold, and in the same sense.
We conclude that the hamiltonian {introduced in this section}
is not really a {\em new model}, but it is just a sufficiently, but not completely, regular, deformed two dimensional harmonic oscillator. More deformations will be listed in the next section.
\section{Other deformed hamiltonians}\label{secIII3}
This section is devoted to a list of other manifestly non self-adjoint hamiltonians which allow a $\D$-pseudo bosonic treatment,
since they eventually appear to be of the form $H=e^XH_0e^{-X}$, for some suitable $X$, which we don't assume here to be necessarily self-adjoint, and for $H_0$ as in (\ref{31}). Our list is rather concise. We just give the expression of the operator $X$ mapping $H_0$ into $H$, the expression of $H$ itself, and some bibliographic information. On the other hand, what we do not consider here, are all the mathematical details we have discussed in Section \ref{SectIII}, leaving open, for instance, the basis problem for the eigenstates of $H$ and $H^\dagger$, as well as the existence of the sets $\D$ and $\G$.
In other words, some results contained in this section are formal\footnote{This is not true for all the hamiltonians considered in this section, some of which have been treated rigorously in terms of $\D$-PBs already in recent papers.} but, we believe, still interesting in view of possible comparison with the literature: all the hamiltonians we are going to list, can in fact be rewritten, some of them at least formally, some others rigorously, in terms of $\D$-PBs.
Then their eigenvalues and eigenvectors, as well as those of their adjoint, can be deduced quite easily, in principle.
Another aspect which we are not going to consider in this section, and which is relevant for a deeper analysis, is whether the parameters of the transformation $e^X$, see below, should be constrained or not, as it happens Section \ref{secIII1}.
\begin{enumerate}
\item
\begin{flalign*} &\begin{cases}
X=-\alpha(p_1+p_2)+\beta(x_1+x_2), \quad \alpha,\beta \in {\Bbb R},\\
H=\tilde\omega_1(x_1^2+p_1^2)+\tilde\omega_2(x_2^2+p_2^2)+2i\alpha(\tilde\omega_1x_1+\tilde\omega_2x_2)+2i\beta(\tilde\omega_1p_1+\tilde\omega_2p_2)+\\
-\left(\tilde\omega _1+\tilde\omega _2-1\right) \left(\alpha^2+\beta^2\right)\1,\\
\tilde\omega_3=\alpha^2+\beta^2\text{ in \eqref{33}}.
\end{cases} & \end{flalign*}
This hamiltonian can be found in the literature in \cite{li}, with $\tilde\omega_1=\tilde\omega_2=1/2$.
\item
\begin{flalign*} &\begin{cases}
X=-\theta\frac{x_1p_1+p_1x_1}{2}, \quad \theta \in (-\pi/4,\pi/4),\\
H=\tilde\omega_1e^{2i\theta}x_1^2+\tilde\omega_1e^{-2i\theta}p_1^2+\tilde\omega_2x_2^2+\tilde\omega_2p_2^2+\tilde\omega_3\1.
\end{cases} & \end{flalign*}
This model is the two-dimensional version of the one-dimensional Swanson model as discussed in \cite{prov}, and first introduced in \cite{swa},
with $\tilde\omega_1=\sec(2\theta)/2$.
\item
\begin{flalign*}&\begin{cases}
X=\gamma(\sqrt{2}(x_1+x_2)+2x_1x_2),\quad \gamma\in {\Bbb R},\\
H=x_1^2(\tilde\omega_1-4\gamma^2\tilde\omega_2)+x_2^2(\tilde\omega_2-4\gamma^2\tilde\omega_1)+\tilde\omega_1 p_1^2+\tilde\omega_2 p_2^2-4\sqrt{2}\gamma^2\tilde\omega_2 x_1-4\sqrt{2}\gamma^2\tilde\omega_1 x_2+\\
+2\sqrt{2}i\tilde\omega_1 \gamma p_1+
+2\sqrt{2}i\tilde\omega_2 \gamma p_2+
4i\tilde\omega_2 \gamma x_1 p_2+4i\tilde\omega_1 \gamma x_2 p_1+\\
+(\tilde\omega_3-2\gamma^2 \tilde\omega_1-2\gamma^2 \tilde\omega_2+2\gamma \tilde\omega_1+2\gamma \tilde\omega_2)\1.
\end{cases}& \end{flalign*}
This is the hamiltonian introduced in Section \ref{secIII1}
\item
\begin{flalign*}&\begin{cases}
X=\gamma(x_1x_2 +p_1p_2),\quad \gamma\in {\Bbb R}\\
H=(\tilde\omega_1 C_{\gamma}^2-\tilde\omega_2 S_{\gamma}^2)x_1^2 +(\tilde\omega_2 C_{\gamma}^2-\tilde\omega_1 S_{\gamma}^2)x_2^2 + (\tilde\omega_1 C_{\gamma}^2-\tilde\omega_2 S_{\gamma}^2)p_1^2+(\tilde\omega_2 C_{\gamma}^2-\tilde\omega_1 S_{\gamma}^2)p_2^2+\\
\nonumber +2i(\tilde\omega_2 - \tilde\omega_1) C_{\gamma} S_{\gamma}(x_1p_2-x_2p_1)+\tilde\omega_3\1.\end{cases}& \end{flalign*}
This is the hamiltonian introduced in Section \ref{secIII2}
\item
\begin{flalign*}&\begin{cases}
X=\gamma_1(x_1x_2) +\gamma_2(p_1p_2),\quad \gamma_1,\gamma_2\in {\Bbb R},\\
H=x_1^2\left(\tilde\omega _1 C ^2_{\sqrt {\gamma _1 \gamma _2}}-\frac{\gamma _1 \tilde\omega _2 S^2_{\sqrt {\gamma _1 \gamma _2}}}{\gamma _2}\right)+
x_2^2\left(\tilde\omega _2 C ^2_{\sqrt {\gamma _1 \gamma _2}}-\frac{\gamma _1 \tilde\omega _1 S^2_{\sqrt {\gamma _1 \gamma _2}}}{\gamma _2}\right)+\\
+p_1^2\left(\tilde\omega _1 C ^2_{\sqrt {\gamma _1 \gamma _2}}-\frac{\gamma _2 \tilde\omega _2 S^2_{\sqrt {\gamma _1 \gamma _2}}}{\gamma _1}\right)+
p_2^2\left(\tilde\omega _2 C ^2_{\sqrt {\gamma _1 \gamma _2}}-\frac{\gamma _2 \tilde\omega _1 S^2_{\sqrt {\gamma _1 \gamma _2}}}{\gamma _1}\right)+\\
+2ix_1 p_2 S_{\sqrt {\gamma _1 \gamma _2}} C_{\sqrt {\gamma _1 \gamma _2}}\left( \sqrt{\frac{\gamma _1}{\gamma _2}} \tilde\omega _2 - \sqrt{\frac{\gamma _2}{\gamma _1}} \tilde\omega _1 \right)
+2ix_2 p_1 S_{\sqrt {\gamma _1 \gamma _2}} C_{\sqrt {\gamma _1 \gamma _2}}\left( \sqrt{\frac{\gamma _1}{\gamma _2}} \tilde\omega _1 - \sqrt{\frac{\gamma _2}{\gamma _1}} \tilde\omega _2 \right)+\omega_3\1,\\
C_{\sqrt {\gamma _1 \gamma _2}}:=\cosh(\sqrt{\gamma_1 \gamma_2}),\quad S_{\sqrt {\gamma _1 \gamma _2}}:=\sinh(\sqrt{\gamma_1 \gamma_2}).
\end{cases}& \end{flalign*}
\item
\begin{flalign*}&\begin{cases}
X=\gamma(x_1+p_1+x_1 p_1),\quad \gamma\in {\Bbb R},\\
H=\tilde\omega_1 x_1^2e^{-2i\gamma}+\tilde\omega_2 x_2^2+\tilde\omega_1 p_1^2e^{2i\gamma}+\tilde\omega_2 p_2^2+\\
+2x_1\tilde\omega_1 e^{i \gamma } \left(-1+e^{i \gamma }\right)+2 p_1\tilde\omega_1e^{i \gamma } \left(-1+e^{i \gamma }\right)+e^{-2 i \gamma } \left(\left(-1+e^{i \gamma }\right)^2 \left(1+e^{2 i \gamma }\right) \tilde\omega _1+e^{2 i \gamma } \tilde\omega _3\right)\1.
\end{cases}& \end{flalign*}
\item
\begin{flalign*}&\begin{cases}
X=\gamma_1(x_1 p_1)+\gamma_2x_1,\quad \gamma_1,\gamma_2\in {\Bbb R},\\
H=\tilde\omega_1 x_1^2e^{-2i\gamma}+\tilde\omega_2 x_2^2+\tilde\omega_1 p_1^2e^{2i\gamma}+\tilde\omega_2 p_2^2+p_1\frac{2 e^{i \gamma _1} \left(-1+e^{i \gamma _1}\right) \gamma _2 \tilde\omega _1}{\gamma _1}+ \left(\frac{\left(-1+e^{i \gamma _1}\right){}^2 \gamma _2^2 \tilde\omega _1}{\gamma _1^2}+\tilde\omega_3\right)\1.
\end{cases}& \end{flalign*}
\item
\begin{flalign*}&\begin{cases}
X=\gamma \left(x_1^2-p_1^2+\sqrt{2}x_1\right),\quad \gamma\in {\Bbb R},\\
H=\tilde\omega_1 x_1^2 (\cos^2{2 \gamma }-\sin^2{2\gamma})+
\tilde\omega_1 p_1^2 (\cos^2{2 \gamma }-\sin^2{2\gamma})+\tilde\omega_2x_2^2+\tilde\omega_2p_2^2+\\
+\sqrt{2}\tilde\omega_1 x_1 ( \cos^2{2 \gamma}-\cos{2 \gamma }-\sin^2{2 \gamma })+i \sqrt{2}\tilde\omega_1 p_1( \cos{2\gamma}\sin{2\gamma}-\sin{2 \gamma })+\\
+4 i \tilde\omega_1 x_1p_1\cos{2 \gamma } \sin{2 \gamma }+ \\
+\Bigg(\tilde\omega_1(\frac{1}{2}-\cos{2 \gamma }+\frac{1}{2} \cos^2{2 \gamma }-\frac{1}{2} \sin^2{2 \gamma}+2\cos{2 \gamma } \sin{2 \gamma })+\tilde\omega_3 \Bigg)\1.
\end{cases}& \end{flalign*}
\item
\begin{flalign*}&\begin{cases}
X=\gamma \left(x_1^2+p_1^2+x_1\right),\quad \gamma\in {\Bbb R},\\
H=\tilde\omega_1x_1^2+\tilde\omega_1p_1^2+\tilde\omega_2x_2^2+\tilde\omega_2p_2^2+\tilde\omega_1x_1+(\frac{\tilde\omega_1}{4}+\tilde\omega_3)\1.
\end{cases}& \end{flalign*}
\item
\begin{flalign*}&\begin{cases}
X=\gamma \left(x_1^2-p_1^2+2x_1 p_1\right),\quad \gamma\in {\Bbb R},\\
H= \tilde\omega _1 x_1^2 \left(\cos \left(4 \sqrt{2} \gamma \right)-\frac{i \sin \left(4 \sqrt{2} \gamma \right)}{\sqrt{2}}\right)+
\tilde\omega _1 p_1^2 \left(\cos \left(4 \sqrt{2} \gamma \right)+\frac{i \sin \left(4 \sqrt{2} \gamma \right)}{\sqrt{2}}\right)+\\
+\tilde\omega_2x_2^2+\tilde\omega_2p_2^2 +
x_1 p_1\left( 2i \sqrt{2} \tilde\omega _1 \sin \left(2 \sqrt{2} \gamma \right) \cos \left(2 \sqrt{2} \gamma \right)\right)+\\
+\left(\tilde\omega_3+ i\sqrt{2} \tilde\omega _1 \sin \left(2 \sqrt{2} \gamma \right) \cos \left(2 \sqrt{2} \gamma \right)\right)\1.
\end{cases}& \end{flalign*}
\item
\begin{flalign*}&\begin{cases}
X= \gamma_1(x_1+x_2+x_1 x_2)+\gamma_2(p_1+p_2+p_1p_2),\quad \gamma_1,\gamma_2 \in {\Bbb R},\\
H=x_1^2 \left(C^2_{\sqrt{\gamma _1 \gamma _2}} \tilde\omega _1-\frac{S^2_{\sqrt{\gamma _1 \gamma _2}} \gamma _2 \tilde\omega _2}{\gamma _1}\right)+p_1^2\left(C^2_{\sqrt{\gamma _1 \gamma _2}} \tilde\omega _1-\frac{S^2_{\sqrt{\gamma _1 \gamma _2}} \gamma _1 \tilde\omega _2}{\gamma _2}\right)+ \\
+x_2^2 \left(C^2_{\sqrt{\gamma _1 \gamma _2}} \tilde\omega _2-\frac{C^2_{\sqrt{\gamma _1 \gamma _2}} \gamma _1 \tilde\omega _1}{\gamma _2}\right)+p_2^2 \left(C^2_{\sqrt{\gamma _1 \gamma _2}} \tilde\omega _2-\frac{S^2_{\sqrt{\gamma _1 \gamma _2}} \gamma _2 \tilde\omega _1}{\gamma _1}\right)+\\
+ x_1\frac{2 \gamma _2 \tilde\omega _2 S^2_{\sqrt{\gamma _1 \gamma _2}}+\gamma _1 \left(\tilde\omega _1 \left(-i \sqrt{\frac{\gamma _2}{\gamma _1}} S_{2 \sqrt{\gamma _1 \gamma _2}}+2 C^2_{\sqrt{\gamma _1 \gamma _2}}-2 C_{\sqrt{\gamma _1 \gamma _2}}\right)-i \sqrt{\frac{\gamma _2}{\gamma _1}} \tilde\omega _2 \left(2 S_{\sqrt{\gamma _1 \gamma _2}}-S_{2 \sqrt{\gamma _1 \gamma _2}}\right)\right)}{\gamma _1}+\\
+p_1\frac{-2 \gamma _1 \tilde\omega _2 S^2_{\sqrt{\gamma _1 \gamma _2}}+\gamma _2 \left(\tilde\omega _1 \left(-i \sqrt{\frac{\gamma _1}{\gamma _2}} S_{2 \sqrt{\gamma _1 \gamma _2}}+2 C^2_{\sqrt{\gamma _1 \gamma _2}}-2 C_{\sqrt{\gamma _1 \gamma _2}}\right)+i \sqrt{\frac{\gamma _1}{\gamma _2}} \tilde\omega _2 \left(2 S_{\sqrt{\gamma _1 \gamma _2}}-S_{2 \sqrt{\gamma _1 \gamma _2}}\right)\right)}{\gamma _2}+ \\
+ x_2\frac{2 \gamma _1 \tilde\omega _1 S^2_{\sqrt{\gamma _1 \gamma _2}}+\gamma _2 \left(\tilde\omega _2 \left(-i \sqrt{\frac{\gamma _1}{\gamma _2}} S_{2 \sqrt{\gamma _1 \gamma _2}}+2 C^2_{\sqrt{\gamma _1 \gamma _2}}-2 C_{\sqrt{\gamma _1 \gamma _2}}\right)-i \sqrt{\frac{\gamma _1}{\gamma _2}} \tilde\omega _1 \left(2 S_{\sqrt{\gamma _1 \gamma _2}}-S_{2 \sqrt{\gamma _1 \gamma _2}}\right)\right)}{\gamma _2}+\\
+p_2\frac{-2 \gamma _2 \tilde\omega _1 S^2_{\sqrt{\gamma _1 \gamma _2}}+\gamma _1 \left(\tilde\omega _2 \left(-i \sqrt{\frac{\gamma _1}{\gamma _2}} S_{2 \sqrt{\gamma _1 \gamma _2}}+2 C^2_{\sqrt{\gamma _1 \gamma _2}}-2 C_{\sqrt{\gamma _1 \gamma _2}}\right)+i \sqrt{\frac{\gamma _2}{\gamma _1}} \tilde\omega _1 \left(2 S_{\sqrt{\gamma _1 \gamma _2}}-S_{2 \sqrt{\gamma _1 \gamma _2}}\right)\right)}{\gamma _1}+ \\
+2 x_2p_1 i C_{\sqrt{\gamma _1 \gamma _2}} S_{\sqrt{\gamma _1 \gamma _2}}\sqrt{\frac{\gamma _1}{\gamma _2}} (\tilde\omega _1-\tilde\omega_2)+2 x_1p_2 i C_{\sqrt{\gamma _1 \gamma _2}} S_{\sqrt{\gamma _1 \gamma _2}} \sqrt{\frac{\gamma _2}{\gamma _1}} (\tilde\omega _2-\tilde\omega_1)+\\
+\1 \Bigg(\tilde\omega_3+C_{\sqrt{\gamma _1 \gamma _2}} S_{\sqrt{\gamma _1 \gamma _2}} \sqrt{\frac{\gamma _1}{\gamma _2}} (\tilde\omega _1-\tilde\omega_2)+ C_{\sqrt{\gamma _1 \gamma _2}} S_{\sqrt{\gamma _1 \gamma _2}} \sqrt{\frac{\gamma _2}{\gamma _1}} (\tilde\omega _2-\tilde\omega_1)+\\
-\frac{2 \left(\tilde\omega _1+\tilde\omega _2\right) S^2_{\frac{1}{2} \sqrt{\gamma _1 \gamma _2}} \left((\gamma _1^2 +\gamma _2^2) \left(C_{\sqrt{\gamma _1 \gamma _2}}+1\right)+2 \gamma _2 \gamma _1 \left(i \left(\sqrt{\frac{\gamma _1}{\gamma _2}}+\sqrt{\frac{\gamma _2}{\gamma _1}}\right)S_{\sqrt{\gamma _1 \gamma _2}}-C_{\sqrt{\gamma _1 \gamma _2}}+1\right)\right)}{\gamma _1 \gamma _2} \Bigg).
\end{cases}& \end{flalign*}
\end{enumerate}
\vspace{2mm}
As one can see from this list, in some examples the operator $X$ in invariant under the exchange $(x_1,p_1)\,\leftrightarrow\,(x_2,p_2)$.
This is the case of examples 1, 3, 4, 5 and 11. In the other examples, this exchange produces a different operator $X$ and, recalling that $H_0$ is invariant with respect to this operation, a different $H$, which is still exactly diagonalizable by means of $\D$-PBs.
\vspace{3mm}
For completeness, we conclude this section by listing other hamiltonians which also can be described in pseudo-bosonic terms, and which has already been introduced in the literature along the years. In these cases the transformation is more involved, and will not be given here. These other hamiltonians are:
$$ \left\{
\begin{array}{ll}
H=H_1+H_2+H_3,\qquad \mbox{where}\\
H_1=\frac{1}{2}(p_1^2+ x_1^2)+\frac{1}{2}( p_2^2+ x_2^2),\\
H_2=\frac{\theta}{2\gamma^2}\,(p_1x_2-p_2x_1) ,\\
H_3=\frac{i}{\gamma}\left[A(x_1+x_2)+\frac{1}{\gamma^2}\left(p_1\left(B+\theta \frac{A}{2}\right)+p_2\left(B-\theta \frac{A}{2}\right)\right)\right],
\end{array}
\right.
$$
$$H=\frac{\nu}{2}\left(p_1^2+x_1^2+p_2^2+x_2^2\right)+i\sqrt{2}\,(p_1+p_2),$$
$$ H=(p_1^2+x_1^2)+(p_2^2+x_2^2+2ix_2)+2\epsilon
x_1x_2,$$
and
$$
H=\frac{1}{2}(p_1^2+x_1^2)+\frac{1}{2}(p_2^2+x_2^2)+i\left[A(x_1+x_2)+B(p_1+p_2)\right].$$
Here $\gamma$, $\nu$, $\theta$, $\epsilon$, $A$ and $B$ are real parameters, see \cite{bagbook}. Also, we recall that a general class of quadratic hamiltonian is discussed, from a different point of view, in \cite{sjo}.
\section{Conclusions}
In this paper we have considered several hamiltonians,
quadratic in two-dimensional position and momentum operators, which
can be analyzed in terms of $\D$-PBs.
This is particularly interesting when the hamiltonian
under analysis is manifestly non self-adjoint.
In this case, we have seen under which conditions on the values of the parameters of the hamiltonian our strategy works properly by checking
the validity of the $\D$-PBs assumptions introduced in Section \ref{sectII}. This was performed in details in Section \ref{secIII1} and \ref{secIII2}
We have also listed in \ref{secIII3}
several deformed hamiltonians, some of which already introduced elsewhere, which allow, at least formally, a pseudo-bosonic treatment. As for the physical content of the models, we should say that this is under debate, and we would say that these models make physical sense if PT-quantum mechanics, and its relatives, makes sense. Nevertheless, the mathematical aspects of these models appear surely worth of a deeper investigation.
\section*{Acknowledgements}
The authors wish to thank Prof. J. P. Gazeau for useful discussions during a preliminary stage of their work.
This work was partially supported by the University of Palermo and by G.N.F.M.
|
\section{\label{sec:intro}Introduction}
Lithium alanate (LiAlH$_{4}$) has been considered as a potential material for hydrogen storage due to its high hydrogen density and relatively low decomposition temperature.~\cite{orimo_chem_rev_2007} Yet the atomistic mechanisms behind the decomposition and dehydrogenation processes in this complex hydride are far from understood. For the purpose of optimizing its hydrogen storage and release capacity, it is desirable to understand the rate-limiting processes involved in the hydrogen desorption. First-principles calculations based on density functional theory have been demonstrated to be powerful for addressing defect-related processes in solids, and have provided valuable insights in the atomistic mechanisms involved in mass transport and hydrogen release.~\cite{peles_prb_2007,hoang_prb_2009,wilson-short_prb_2009,hoang_angew} Here we apply this approach to explore possible mechanisms for the decomposition and dehydrogenation of LiAlH$_{4}$.
It has been observed that LiAlH$_{4}$ desorbs hydrogen through a two-step process, similar to that reported in the widely studied NaAlH$_{4}$, i.e.,~\cite{dilts_1972,dymova_1994,blanchard_2005,andreasen_2005}
\begin{equation}
\label{eq:step1}
\rm{LiAlH_{4} \rightarrow \frac{1}{3} Li_{3}AlH_{6} + \frac{2}{3}Al + H_{2}},
\end{equation}
\begin{equation}
\label{eq:step2}
\rm{\frac{1}{3}Li_{3}AlH_{6} \rightarrow LiH + \frac{1}{3}Al + \frac{1}{2}H_{2}}.
\end{equation}
The first reaction occurs around 112$-$220$^\circ$C with a theoretical hydrogen release of 5.3 wt\%. It was observed to initiate by the melting of LiAlH$_{4}$ in the temperature range 150$-$170$^\circ$C, although isothermal decomposition without melting has also been reported.~\cite{andreasen_2005,wiench_2004,varin_2010} The second reaction takes place around 127$-$260$^\circ$C and releases 2.6 wt\% hydrogen. On the other hand, using nuclear magnetic resonance (NMR) studies of the decomposed sample during isothermal heating of LiAlH$_{4}$ at 150$^\circ$C over 2h, Wiench {\it et al}.~\cite{wiench_2004} observed an apparent deviation of the sample composition from that predicted by the above two-step mechanism. They suggested that thermal decomposition of LiAlH$_{4}$ may follow several different reaction paths, including
\begin{equation}
\label{eq:alternative}
\rm{LiAlH_{4} \rightarrow LiH + Al + \frac{3}{2}H_{2}}.
\end{equation}
Later studies carried out by Varin {\it et al}.,~\cite{varin_2010} however, did not support such direct decomposition of LiAlH$_{4}$ in the solid state into LiH and Al.
Significant efforts have been devoted to studying the decomposition kinetics of LiAlH$_{4}$ and determining the activation energy for hydrogen desorption.~\cite{andreasen_2005,andreasen_2006,blanchard_2005,varin_2010} From kinetic measurements carried out under isothermal conditions, Andreasen {\it et al}.~\cite{andreasen_2005} obtained an apparent activation energy of 0.85 eV for eqn (\ref{eq:step1}) in the solid state. For the same reaction but in the liquid state, Andreasen obtained an activation energy of 0.84 eV.~\cite{andreasen_2006} Blanchard {\it et al}.~\cite{blanchard_2005} reported an activation energy of 1.06 eV for the main desorption stage of LiAlD$_{4}$. More recently, Varin and Zbroniec estimated the activation energy for eqn (\ref{eq:step1}) to be 1.15 eV and 0.96 eV for as-received and ball-milled LiAlH$_{4}$, respectively.~\cite{varin_2010} Ball milling thus results in a slightly lower activation energy.
It has also been reported that metal additives such as Ti and Ni improve the dehydrogenation properties of LiAlH$_{4}$.~\cite{blanchard_2004,blanchard_2005,wang_adsorption_2005,chen_jpc_2001,kojima_2008,liu_jacs_2009,sun_ijhe_2008,andreasen_2006,naik_ijhe_2009,langmi_jpc_2010,rafiuddin_2010,varin2010928,varin20111167} For example, ball-milling LiAlH$_{4}$ with NiCl$_{2}$ was found to reduce the onset decomposition temperature by about 50$^\circ$C.~\cite{sun_ijhe_2008} Doping with TiCl$_{3}$ was also found to lower the onset temperature of eqn (\ref{eq:step1}) by 60$-$75$^\circ$C, bringing it well below the melting point of LiAlH$_{4}$.~\cite{langmi_jpc_2010} Recently, Liu {\it et al}.~\cite{liu_jacs_2009} demonstrated that Ti-doped LiAlH$_{4}$ can operate as a reversible hydrogen storage material that can release up to 7 wt\% hydrogen commencing at temperatures as low as 80$^\circ$C, and recharge can be achieved by employing liquid dimethyl ether as a solvent. In another study, nanometric TiC and Ni additives have been found to reduce the effective decomposition temperature and enhance the hydrogen desorption of LiAlH$_{4}$.\cite{rafiuddin_2010,varin2010928,varin20111167}
In spite of the consensus among different experimental reports that Ti and Ni lower the onset decomposition temperature, the conclusions regarding the effects of metal additives on the kinetics of isothermal decomposition of LiAlH$_{4}$ are not without conflict. On the one hand, Andreasen~\cite{andreasen_2006} reported apparent activation energies of 0.84 and 0.92 eV for undoped and Ti-doped LiAlH$_{4}$, respectively. Since the differences in the activation energies between undoped and Ti-doped samples were within the experimental uncertainty, the author suggested that the effect of Ti doping on the dehydrogenation kinetics of LiAlH$_{4}$ was mainly on the prefactor.~\cite{andreasen_2006} Blanchard {\it et al}.\cite{blanchard_2005} also reported small differences between the activation energies of doped and undoped samples; e.g., 1.06 and 0.99 eV for undoped and TiCl$_{3}$$\cdot$1/3AlCl$_{3}$-doped LiAlD$_{4}$, respectively. On the other hand, Chen {\it et al}.~\cite{chen_jpc_2001} found that the activation energy for hydrogen desorption of LiAlH$_{4}$ doped with 2 mol\% TiCl$_{3}$$\cdot$1/3AlCl$_{3}$ is 0.44 eV for eqn (\ref{eq:step1}) in the solid state, which is much smaller than the reported values (0.84$-$1.15 eV) for undoped LiAlH$_{4}$. Other research groups reported smaller, but still significant, reductions in the activation energy for decomposition when the compound was ball-milled with metal additives.~\cite{rafiuddin_2010,varin2010928} For example, Varin {\it et al}.\cite{varin2010928} reported an activation energy of 0.73 eV for LiAlH$_{4}$ ball-milled with nanometric Ni, which is 0.23 eV lower than that for undoped LiAlH$_{4}$.\cite{varin_2010}
To resolve this situation, clearly one needs to understand the fundamental mechanisms behind the decomposition and dehydrogenation processes and the interaction between the metal additives and the host material. Theoretical studies have so far focused mainly on bulk LiAlH$_{4}$ and its thermodynamic properties.~\cite{lovvik_prb_2004,kang_jpc_2004,vajeeston_2005,yoshino_2005,vansetten_prb_2007,jang_2006} Experimental data, on the other hand, suggested that the decomposition process involves mass transport by native point defects.~\cite{andreasen_2005,Ares20081263} This motivates us to perform first-principles calculations of native defects and transition-metal (Ti and Ni) impurities in LiAlH$_{4}$. As we will discuss in this paper, these calculations enable us to explore possible mechanisms for the decomposition and dehydrogenation involving mass transport mediated by native point defects and investigate the effects of Ti and Ni impurities in the material.
\section{\label{sec;metho}Methodology}
{\bf Computational details.} Our calculations were based on density functional theory within the generalized-gradient approximation (GGA)~\cite{GGA} and the projector augmented wave method,~\cite{PAW1,PAW2} as implemented in the VASP code.~\cite{VASP1,VASP2,VASP3} Calculations for bulk LiAlH$_{4}$ (24 atoms/unit cell) were performed using a 10$\times$6$\times$6 Monkhorst-Pack $\mathbf{k}$-point mesh.~\cite{monkhorst-pack} For calculations of native defects and transition-metal impurities, we used a (2$\times$2$\times$2) supercell containing 192 atoms/cell, and a 2$\times$2$\times$2 $\mathbf{k}$-point mesh. The plane-wave basis-set cutoff was set to 400 eV and convergence with respect to self-consistent iterations was assumed when the total energy difference between cycles was less than 10$^{-4}$ eV and the residual forces were less than 0.01 eV/{\AA}. The migration of selected native defects in LiAlH$_{4}$ was studied using the climbing image nudged elastic band method (NEB).~\cite{ci-neb}
{\bf Defect formation energies.} The likelihood of forming a defect is given by its formation energy ($E^{f}$). In thermal equilibrium, the concentration of defect X at temperature $T$ can be obtained via the relation~\cite{walle:3851,janotti2009}
\begin{equation}\label{eq:concen}
c(\mathrm{X})=N_{\mathrm{sites}}N_{\mathrm{config}}\mathrm{exp}[-E^{f}(\mathrm{X})/k_BT],
\end{equation}
where $N_{\mathrm{sites}}$ is the number of high-symmetry sites in the lattice per unit volume on which the defect can be incorporated, and $N_{\mathrm{config}}$ is the number of equivalent configurations per site. Obviously, defects with lower formation energies are more likely to form and occur in higher concentrations. Note that the energy in eqn (\ref{eq:concen}) is, in principle, a free energy; however, the entropy and volume terms are often neglected because they are negligible at relevant experimental conditions.\cite{janotti2009}
The formation energy of a defect X in charge state $q$ is defined as~\cite{walle:3851}
\begin{equation}\label{eq:eform}
E^f({\mathrm{X}}^q)=E_{\mathrm{tot}}({\mathrm{X}}^q)-E_{\mathrm{tot}}({\mathrm{bulk}})-\sum_{i}{n_i\mu_i} +q(E_{\mathrm{v}}+\Delta V+\mu_{e}),
\end{equation}
where $E_{\mathrm{tot}}(\mathrm{X}^{q})$ and $E_{\mathrm{tot}}(\mathrm{bulk})$ are, respectively, the total energies of a supercell containing the defect X and of a supercell of the perfect bulk material. $\mu_{i}$ is the chemical potential of species $i$; $\mu_{i}$=$\mu_{i}^{0}$+$\tilde{\mu}_{i}$, where $\mu_{i}^{0}$ equals the chemical potential of element $i$ in its standard state. $n_{i}$ denotes the number of atoms of species $i$ that have been added ($n_{i}$$>$0) or removed ($n_{i}$$<$0) to form the defect. $\mu_{e}$ is the electron chemical potential, i.e., the Fermi level, referenced to the valence-band maximum in the bulk ($E_{\mathrm{v}}$). $\Delta V$ is the ``potential alignment'' term, i.e., the shift in the band positions due to the presence of the charged defect and the neutralizing background, obtained by aligning the average electrostatic potential in regions far away from the defect to the bulk value.~\cite{walle:3851}
{\bf Chemical potentials.} The atomic chemical potentials $\mu_{i}$ are variables and can be chosen to represent experimental conditions. In the following discussions, we assume that LiAlH$_{4}$, Al, and Li$_{3}$AlH$_{6}$ are stable and in equilibrium. The chemical potentials of Li, Al, and H can then be obtained from the equations that express the stability of LiAlH$_{4}$, Al, and Li$_{3}$AlH$_{6}$.~\cite{walle:3851} This gives, approximately, $\tilde{\mu}_{\rm Li}$=$-$0.862 eV, $\tilde{\mu}_{\rm Al}$=0 eV, and $\tilde{\mu}_{\rm H}$=0 eV. Note that this set of chemical potentials also approximately corresponds to assuming that LiAlH$_{4}$, H$_{2}$, and Al (or Li$_{3}$AlH$_{6}$) are in equilibrium. For the impurities (Ti and Ni), the chemical potentials are fixed to the energy of the bulk metals, $\tilde{\mu}_{i}$=0 eV, which is the upper bound.
{\bf Transition levels.} We will refer to $\epsilon(q_{1}/q_{2})$ as the thermodynamic transition level which is defined as the Fermi-level position where the charge states $q_{1}$ and $q_{2}$ have equal formation energies.~\cite{walle:3851} It can be shown from eqn (\ref{eq:eform}) that the Fermi level at which the transition takes place is
\begin{equation}\label{eq:translevel}
\epsilon(q_{1}/q_{2})=\frac{E_{\mathrm{tot}}({\mathrm{X}}^{q_{1}})-E_{\mathrm{tot}}({\mathrm{X}}^{q_{2}})+(q_{1}-q_{2})E_{\mathrm{v}}}{q_{2}-q_{1}},
\end{equation}
where $q_{1}$ and $q_{2}$ are the initial and final charge states, respectively. Clearly, $\epsilon(q_{1}/q_{2})$ is independent of the choice of atomic chemical potentials.
\section{\label{sec;bulk}Bulk properties}
\begin{figure}
\begin{center}
\includegraphics[width=2.2in]{LiAlH4_struct}
\end{center}
\vspace{-0.15in}
\caption{Relaxed structure of monoclinic LiAlH$_{4}$. Large (gray) spheres are Li, medium (blue) spheres Al, and small (red) spheres H.}\label{fig;struct}
\end{figure}
We start by presenting the results for the basic structural and electronic properties of bulk LiAlH$_{4}$. The compound was reported to crystallize in the monoclinic structure, space group $P\mathrm{2}_{\mathrm{1}}/c$, with lattice parameters $a$=4.817 {\AA}, $b$=7.802 {\AA}, $c$=7.821 {\AA}, and $\beta$=112.228$^{\circ}$ at 8 K.~\cite{hauback_jac_2002} It can be regarded as an ordered arrangement of Li$^{+}$ and (AlH$_{4}$)$^{-}$ units. Figure \ref{fig;struct} shows the optimized structure of LiAlH$_{4}$. The calculated lattice parameters $a$=4.860 {\AA}, $b$=7.817 {\AA}, $c$=7.832 {\AA}, and $\beta$=111.808$^{\circ}$ are in agreement with the experimental values.
\begin{figure}
\begin{center}
\includegraphics[width=3.4in]{LiAlH4_DOS}
\end{center}
\vspace{-0.15in}
\caption{Total density of states (DOS) and projected density of states (PDOS) of LiAlH$_{4}$. The zero of energy is set to the highest occupied state.}\label{fig;DOS}
\end{figure}
Figure \ref{fig;DOS} shows the total density of states and projected density of states of LiAlH$_{4}$. The valence-band maximum (VBM) consists of the bonding state of Al $p$ and H $s$, whereas the conduction-band minimum (CBM) consists of the antibonding state of Al $p$ and H $s$ and contribution from Li $s$. The calculated band gap is 4.64 eV, very close to that reported previously (4.67 eV).~\cite{vansetten_prb_2007} As we will illustrate in the next sections, knowing the structural and electronic properties of LiAlH$_{4}$ is essential to understand the properties of native defects and the interaction between impurities and the host compound.
\section{\label{sec:defects}Formation of native defects}
In insulating, wide band-gap materials such as LiAlH$_{4}$, native point defects are expected to exist in charged states other than neutral, and charge neutrality requires that defects with opposite charge states coexist in equal concentrations.~\cite{peles_prb_2007,hoang_prb_2009,wilson-short_prb_2009,hoang_angew} We therefore investigated hydrogen-, lithium-, and aluminum-related point defects in all possible charge states. Defect complexes are also considered, with special attention to Frenkel pairs, i.e., interstitial-vacancy pairs of the same species. In the following, we present the results for the defects in each category. The role of these defects in ionic and mass transport in LiAlH$_{4}$ will be discussed in the next section.
\begin{figure}
\begin{center}
\includegraphics[width=3.0in]{LiAlH4_FE_H}
\end{center}
\vspace{-0.15in}
\caption{Calculated formation energies of hydrogen-related defects in LiAlH$_{4}$, plotted as a function of Fermi level with respect to the valence-band maximum.}\label{fig;FE;H}
\end{figure}
{\bf Hydrogen-related defects.} Figure \ref{fig;FE;H} shows the calculated formation energies of hydrogen vacancies ($V_{\mathrm{H}}$), interstitials (H$_{i}$), and interstitial molecules (H$_{2}$)$_{i}$ in LiAlH$_{4}$. Among these defects, the positively charged hydrogen vacancy ($V_{\mathrm{H}}^{+}$), positively charged hydrogen interstitial (H$_{i}^{+}$), and negatively charged hydrogen interstitial (H$_{i}^{-}$) have the lowest formation energies over a wide range of Fermi-level values. (H$_{2}$)$_{i}$ has the lowest formation energy in a relatively small range near $\mu_{e}$=2.84 eV where the formation energies of $V_{\mathrm{H}}^{+}$ and H$_{i}^{-}$ are equal. The neutral hydrogen vacancy ($V_{\mathrm{H}}^{0}$) and interstitial (H$_{i}^{0}$) are energetically less favorable than their respective charged defects over the entire range of the Fermi-level values, which is a characteristic of negative-$U$ centers.~\cite{negativeU}
\begin{figure}
\begin{center}
\includegraphics[width=3.2in]{LiAlH4_H}
\end{center}
\vspace{-0.15in}
\caption{Structures of (a) H$_{i}^{-}$, (b) $V_{\mathrm{H}}^{+}$, (c) $V_{\mathrm{H}}^{-}$, and (d) H$_{i}^{+}$. Only Li and Al atoms in [010] planes (and their coordinated H atoms) are shown.}\label{fig;H}
\end{figure}
The creation of H$_{i}^{-}$ involves adding one H atom and an electron (i.e., H$^{-}$) to the LiAlH$_{4}$ supercell (hereafter referred to as ``the system''). This H$^{-}$ combines with an (AlH$_{4}$)$^{-}$ unit to form (AlH$_{5}$)$^{2-}$, see Fig.~\ref{fig;H}(a). In the (AlH$_{5}$)$^{2-}$, the average Al$-$H distance is 1.71 {\AA}, compared to 1.63 {\AA} in the perfect bulk crystal. One may expect that due to Coulomb interaction, H$_{i}^{-}$ would prefer to stay near a Li$^{+}$ unit. In fact, we find that this configuration is not stable, except in some cases where H$_{i}^{-}$ is created simultaneously with other defect(s) (which will be illustrated later when we present our results for aluminum-related defects).
$V_{\mathrm{H}}^{+}$ is created by removing H$^{-}$ from the system. This leads to formation of an AlH$_{3}$-H-AlH$_{3}$ complex, or (Al$_{2}$H$_{7}$)$^{-}$, with the Al$-$Al distance being of 3.26 {\AA} (compared to 3.99 {\AA} in the bulk), see Fig.~\ref{fig;H}(b). $V_{\mathrm{H}}^{-}$, on the other hand, can be thought as the extraction of an H$^{+}$ ion from the system. This results in an AlH$_{3}$ unit, see Fig.~\ref{fig;H}(c). (H$_{2}$)$_{i}$ involves adding an H$_{2}$ molecule to the supercell. This interstitial molecule prefers to stay in an interstitial void, with the calculated bond length of 0.75 {\AA} being equal to that calculated for an isolated H$_{2}$ molecule. Finally, H$_{i}^{+}$ is created by adding an H$^{+}$ ion into the system. This results in an AlH$_{3}$-H-AlH$_{3}$ complex plus a H$_{2}$ interstitial molecule, see Fig.~\ref{fig;H}(d). H$_{i}^{+}$ can therefore be regarded as a complex of $V_{\mathrm{H}}^{+}$ and (H$_{2}$)$_{i}$. The formation energy of H$_{i}^{+}$ is, however, lower than the sum of the formation energies of $V_{\mathrm{H}}^{+}$ and (H$_{2}$)$_{i}$, giving H$_{i}^{+}$ a binding energy of 0.44 eV with respect to its constituents. From these analyses, it is evident that H$_{i}^{-}$, $V_{\rm{H}}^{+}$, $V_{\rm{H}}^{-}$, and (H$_{2}$)$_{i}$ are elementary native point defects, meaning the structure and energetics of the other defects can be interpreted in terms of these basic building blocks.
For the migration of H$_{i}^{-}$, $V_{\rm{H}}^{+}$, $V_{\rm{H}}^{-}$, and (H$_{2}$)$_{i}$, we find energy barriers of 0.15, 0.63, 0.90, and 0.23 eV, respectively. The energy barriers for $V_{\rm{H}}^{+}$ and $V_{\rm{H}}^{-}$ are relatively high because the diffusion of these defects involves breaking Al$-$H bonds from AlH$_{4}$ units. The diffusion of $V_{\rm{H}}^{-}$, for example, involves moving an H atom from an AlH$_{4}$ unit to the vacancy. The saddle-point configuration in this case consists of a H atom located midway between two AlH$_{3}$ units (i.e., AlH$_{3}$-H-AlH$_{3}$). Such a configuration is favorable in the case of $V_{\mathrm{H}}^{+}$, but high in energy for $V_{\rm{H}}^{-}$. For H$_{i}^{+}$, which is a complex of $V_{\mathrm{H}}^{+}$ and (H$_{2}$)$_{i}$, the migration barrier is larger than or equal to that of the least mobile constituent,\cite{wilson-short_prb_2009} i.e., 0.63 eV, the value for $V_{\mathrm{H}}^{+}$.
\begin{figure}
\begin{center}
\includegraphics[width=2.4in]{LiAlH4_Frenkel_H}
\end{center}
\vspace{-0.15in}
\caption{Structure of (H$_{i}^{-}$,$V_{\mathrm{H}}^{+}$) Frenkel pair in LiAlH$_{4}$. Only Li and Al atoms in [100] planes (and their coordinated H atoms) are shown. }\label{fig;Frenkel;H}
\end{figure}
Considering that hydrogen vacancies and interstitials can be stable as oppositely charged defects, charge and mass conservation conditions suggest that these native defects may form in the interior of the material in the form of Frenkel pairs. Since in LiAlH$_{4}$, H$_{i}^{+}$ is a complex defect, the only possible hydrogen Frenkel pair is (H$_{i}^{-}$,$V_{\mathrm{H}}^{+}$), whose structure is shown in Fig.~\ref{fig;Frenkel;H}. The configurations of the individual defects are preserved in this complex. (H$_{i}^{-}$,$V_{\mathrm{H}}^{+}$) has a formation energy of 1.23 eV, and a binding energy of 0.20 eV with respect to isolated H$_{i}^{-}$ and $V_{\mathrm{H}}^{+}$.
\begin{figure}
\begin{center}
\includegraphics[width=3.0in]{LiAlH4_FE_Li}
\end{center}
\vspace{-0.15in}
\caption{Calculated formation energies of lithium-related defects in LiAlH$_{4}$, plotted as a function of Fermi level with respect to the valence-band maximum.}\label{fig;FE;Li}
\end{figure}
{\bf Lithium-related defects.} Figure \ref{fig;FE;Li} shows the calculated formation energies of lithium vacancies ($V_{\mathrm{Li}}$), interstitials (Li$_{i}$), and $V_{\mathrm{LiH}}^{0}$ (removing one Li and one H atom) in LiAlH$_{4}$. Among these lithium-related defects, Li$_{i}^{+}$ and $V_{\mathrm{Li}}^{-}$ have the lowest formation energies for the entire range of Fermi-level values. These two defects have equal formation energies at $\mu_{e}$=2.83 eV.
The creation of $V_{\mathrm{Li}}^{-}$ involves removing a Li$^{+}$ ion from the system. This causes very small changes to the lattice geometry. On contrary, $V_{\mathrm{Li}}^{+}$, created by removing a Li atom and an extra electron, strongly disturbs the system. Besides the void formed by the removed Li, there are two AlH$_{3}$-H-AlH$_{3}$ complexes and a H$_{2}$ interstitial molecule that all can be identified as 2$V_{\mathrm{H}}^{+}$ and (H$_{2}$)$_{i}$. Therefore, $V_{\mathrm{Li}}^{+}$ can be regarded as a complex of $V_{\mathrm{Li}}^{-}$, 2$V_{\mathrm{H}}^{+}$, and (H$_{2}$)$_{i}$.
Li$_{i}^{+}$ is created by adding a Li$^{+}$ ion to the system. Like $V_{\mathrm{Li}}^{-}$, Li$_{i}^{+}$ does not cause much disturbance to the lattice geometry. On the other hand, Li$_{i}^{-}$, which is created by adding a Li atom and an extra electron to the system, strongly disturbs the system by breaking Al$-$H bonds and forming AlH$_{3}$ and AlH$_{5}$ units which can be identified as $V_{\mathrm{H}}^{-}$ and H$_{i}^{-}$, respectively. This defect, therefore, is considered as a complex of Li$_{i}^{+}$, $V_{\mathrm{H}}^{-}$, and H$_{i}^{-}$. Similarly, $V_{\mathrm{LiH}}^{0}$ can be regarded as a complex of $V_{\mathrm{Li}}^{-}$ and $V_{\mathrm{H}}^{+}$. Thus, Li$_{i}^{+}$ and $V_{\mathrm{Li}}^{-}$ can be considered as the elementary defects in the Li sublattice.
The migration of Li$_{i}^{+}$ involves an energy barrier of 0.28 eV. For $V_{\mathrm{Li}}^{-}$, the migration involves moving Li$^{+}$ from a nearby lattice site to the vacancy, and this gives an energy barrier as low as 0.14 eV. The migration barriers for Li$_{i}^{+}$ and $V_{\mathrm{Li}}^{-}$ are relatively small, suggesting that they are highly mobile. For Li$_{i}^{-}$, which can be considered as a complex of Li$_{i}^{+}$, $V_{\mathrm{H}}^{-}$, and H$_{i}^{-}$, the migration barrier is estimated to be at least 0.90 eV, the value for $V_{\mathrm{H}}^{-}$. Similarly, the estimated migration barrier of $V_{\mathrm{Li}}^{+}$ and $V_{\mathrm{LiH}}^{0}$ is at least 0.63 eV, the value for $V_{\mathrm{H}}^{+}$.
We also investigated possible formation of lithium Frenkel pairs. Since Li$_{i}^{-}$ and $V_{\mathrm{Li}}^{+}$ are not elementary defects, the only possibility is (Li$_{i}^{+}$,$V_{\mathrm{Li}}^{-}$). The distance between Li$_{i}^{+}$ and $V_{\mathrm{Li}}^{-}$ is 3.64 {\AA}. This pair has a formation energy of 0.75 eV and a binding energy of 0.40 eV. The formation energy is, therefore, much lower than that of the hydrogen Frenkel pair (H$_{i}^{-}$,$V_{\mathrm{H}}^{+}$). This result indicates that LiAlH$_{4}$ may be prone to Frenkel disorder on the Li sublattice.
\begin{figure}
\begin{center}
\includegraphics[width=3.0in]{LiAlH4_FE_Al}
\end{center}
\vspace{-0.15in}
\caption{Calculated formation energies of aluminum-related defects in LiAlH$_{4}$, plotted as a function of Fermi level with respect to the valence-band maximum.}\label{fig;FE;Al}
\end{figure}
{\bf Aluminum-related defects.} Figure \ref{fig;FE;Al} shows the calculated formation energies of Al vacancies ($V_{\mathrm{Al}}$), AlH vacancies ($V_{\mathrm{AlH}}$), AlH$_{2}$ vacancies ($V_{\mathrm{AlH_{2}}}$), AlH$_{3}$ vacancies ($V_{\mathrm{AlH_{3}}}$), and AlH$_{4}$ vacancies ($V_{\mathrm{AlH_{4}}}$). Only the lowest energy charge states for each defect are included. We find that $V_{\mathrm{AlH_{3}}}^{0}$, $V_{\mathrm{AlH_{4}}}^{+}$, $V_{\mathrm{Al}}^{3-}$, and $V_{\mathrm{Al}}^{3+}$ have the lowest formation energies for certain ranges of Fermi-level values. $V_{\mathrm{AlH_{4}}}^{+}$ corresponds to the removal of an entire (AlH$_{4}$)$^{-}$ unit from the system. The migration of $V_{\mathrm{AlH_{4}}}^{+}$ involves moving a nearby (AlH$_{4}$)$^{-}$ unit to the vacancy with an energy barrier of 0.43 eV.
\begin{figure}
\begin{center}
\includegraphics[width=3.0in]{LiAlH4_VAlH3}
\end{center}
\vspace{-0.15in}
\caption{Structure of $V_{\mathrm{AlH}_{3}}^{0}$ in LiAlH$_{4}$. The defect can be regarded as a complex of $V_{\mathrm{AlH}_{4}}^{+}$ (presented by empty spheres) and H$_{i}^{-}$. }\label{fig;VAlH3}
\end{figure}
The creation of $V_{\mathrm{AlH_{3}}}^{0}$ involves removing one Al and three H atoms from the system. This is equivalent to removing an (AlH$_{4}$)$^{-}$ unit and adding an H$^{-}$ simultaneously, as shown in Fig.~\ref{fig;VAlH3}. $V_{\mathrm{AlH_{3}}}^{0}$ therefore can be regarded as a complex of $V_{\mathrm{AlH_{4}}}^{+}$ and H$_{i}^{-}$ with a binding energy of 0.42 eV with respect to its constituents. The structure of H$_{i}^{-}$ in this complex is, however, significantly different from that presented earlier, {\it cf.}~Fig.~\ref{fig;H}(a). We find that the H$^{-}$ stays near three Li$^{+}$ ions with the Li$-$H distances being 1.80, 1.84, and 1.85 {\AA}; see Fig.~\ref{fig;VAlH3}. This configuration of the H$_{i}^{-}$ component can play an important role in the formation of LiH. In this case, a certain amount of LiAlH$_{4}$ may decompose directly into LiH, Al, and H$_{2}$ following eqn (\ref{eq:alternative}). The migration barrier of $V_{\mathrm{AlH_{3}}}^{0}$ is estimated to be at least 0.43 eV, given by that of $V_{\mathrm{AlH_{4}}}^{+}$.
We also find an alternative configuration of $V_{\mathrm{AlH_{3}}}^{0}$ where the H$^{-}$ unit combines with an (AlH$_{4}$)$^{-}$ unit to form (AlH$_{5}$)$^{2-}$, which is similar to the local structure of H$_{i}^{-}$ shown in Fig.~\ref{fig;H}(a). This configuration is 0.16 eV higher in energy than the lowest configuration of $V_{\mathrm{AlH_{3}}}^{0}$ mentioned above. Our results thus indicate that H$_{i}^{-}$ prefers staying close to Li$^{+}$ when it is created simultaneously with and in the vicinity of other defects such as $V_{\mathrm{AlH_{4}}}^{+}$.
Similar to $V_{\mathrm{AlH_{3}}}^{0}$, the structure and energetics of other aluminum-related defects can also be interpreted in terms of $V_{\mathrm{AlH_{4}}}^{+}$ and the elementary hydrogen-related defects. $V_{\mathrm{AlH}}^{0}$, for example, can be considered as a complex of $V_{\mathrm{AlH_{4}}}^{+}$, H$_{i}^{-}$, and (H$_{2}$)$_{i}$. The H$_{i}^{-}$ component in this defect is also composed of H$^{-}$ staying close to three Li$^{+}$ ions with the Li$-$H distances being 1.82, 1.85, and 1.85 {\AA}. $V_{\mathrm{AlH_{2}}}^{+}$, on the other hand, can be regarded as a complex of $V_{\mathrm{AlH_{4}}}^{+}$ and (H$_{2}$)$_{i}$. Likewise, $V_{\mathrm{Al}}^{+}$ can be considered as a complex of $V_{\mathrm{AlH_{4}}}^{+}$ and 2(H$_{2}$)$_{i}$; $V_{\mathrm{Al}}^{3-}$ as a complex of $V_{\mathrm{AlH_{4}}}^{+}$, (H$_{2}$)$_{i}$, and 2H$_{i}^{+}$; and $V_{\mathrm{Al}}^{3+}$ as a complex of $V_{\mathrm{AlH_{4}}}^{+}$ and 4H$_{i}^{-}$.
Overall, we find that H$_{i}^-$, $V_{\rm{H}}^-$, $V_{\rm{H}}^+$, (H$_{2})_{i}$, Li$_{i}^+$, $V_{\rm{Li}}^-$, and $V_{\rm{AlH_4}}^+$ are the elementary defects, and that the other defects can be regarded as complexes involving these basic constituents. Understanding the structure and energetics of these native point defects is, therefore, key to describing ionic and mass transport and the decomposition process in LiAlH$_{4}$. Since these point defects, except (H$_{2})_{i}$, are charged, their creation in the interior of the material necessarily requires both mass and charge conservation. As discussed earlier, hydrogen- and lithium-related defects can be formed inside LiAlH$_{4}$ via Frenkel pair mechanisms, i.e., moving H (or Li) from one lattice site to an interstitial site; this leads to formation of interstitial-vacancy complexes such as (H$_{i}^{-}$,$V_{\mathrm{H}}^{+}$) and (Li$_{i}^{+}$,$V_{\mathrm{Li}}^{-}$). Aluminum-related defects such as $V_{\mathrm{AlH_{4}}}^{+}$ and $V_{\mathrm{AlH_{3}}}^{0}$, on the other hand, can only be created at the surface or interface since the creation of such defects inside the material requires creation of the corresponding aluminum-related interstitials which are too high in energy.
\section{\label{sec:transport}Ionic and mass transport}
In the absence of electrically active impurities that can affect the Fermi-level position, or when such impurities occur in much lower concentrations than charged native defects, the Fermi-level position of LiAlH$_{4}$ is determined by oppositely charged defects with the lowest formation energies.~\cite{peles_prb_2007,hoang_prb_2009,wilson-short_prb_2009,hoang_angew} These defects are Li$_{i}^{+}$ and $V_{\rm{Li}}^{-}$ which pin the Fermi level at $\mu_{e}$=2.83 eV (hereafter referred to as $\mu_{e}^{\rm int}$, the Fermi-level position determined by intrinsic/native defects), where the formation energies and hence, approximately, concentrations of Li$_{i}^{+}$ and $V_{\rm{Li}}^{-}$ are equal. We list in Table \ref{tab:lialh} formation energies and migration barriers of the most relevant native point defects and defect complexes in LiAlH$_{4}$. The formation energies for charged defects are taken at $\mu_{e}^{\rm int}$.
\begin{table}[t]
\caption{Formation energies ($E^{f}$) and migration barriers ($E_{m}$) for native defects in LiAlH$_{4}$. Migration barriers denoted by an asterisk ($^{\ast}$) are the lower bounds, estimated by considering the defect as a complex and taking the highest of the migration barriers of the constituents.}\label{tab:lialh}
\begin{center}
\begin{tabular}{cccc}
Defect&$E^f$ (eV)& $E_m$ (eV)&Constituents \\
\hline
H$_{i}^{+}$ &0.73&0.63$^{\ast}$ &$V_{\rm{H}}^{+}$+(H$_{2}$)$_{i}$\\
H$_{i}^{-}$ &0.72&0.15 & \\
$V_{\rm{H}}^{+}$ &0.70&0.63 & \\
$V_{\rm{H}}^{-}$ &1.42&0.90 \\
(H$_{2}$)$_{i}$&0.47&0.23 \\
Li$_{i}^{+}$ &0.58&0.28 \\
$V_{\rm{Li}}^{-}$&0.58&0.14 \\
$V_{\rm{AlH_{4}}}^{+}$&0.68&0.43 \\
$V_{\rm{AlH_{3}}}^{0}$&0.99&0.43$^{\ast}$&$V_{\rm{AlH}_{4}}^{+}$+H$_{i}^{-}$ \\
$V_{\rm{AlH}_{2}}^{+}$&0.69&0.43$^{\ast}$&$V_{\rm{AlH}_{4}}^{+}$+(H$_{2}$)$_{i}$\\
$V_{\rm{AlH}}^{0}$&1.02&0.43$^{\ast}$&$V_{\rm{AlH}_{4}}^{+}$+(H$_{2}$)$_{i}$+H$_{i}^{-}$ \\
$V_{\rm{Al}}^{+}$&0.71&0.43$^{\ast}$&$V_{\rm{AlH}_{4}}^{+}$+2(H$_{2}$)$_{i}$\\
\end{tabular}
\end{center}
\begin{flushleft}
\end{flushleft}
\end{table}
{\bf Lithium-ion conduction.} We note from Table \ref{tab:lialh} that Li$_{i}^{+}$ and $V_{\rm{Li}}^{-}$ have low formation energies and are highly mobile. Both defects can contribute to the ionic conductivity. However, since the calculated migration barrier of $V_{\rm{Li}}^{-}$ is lower than that of Li$_{i}^{+}$, we expect that, in LiAlH$_{4}$, lithium diffusion via vacancy mechanism is dominant. The activation energy for ionic conduction is estimated to be 0.72 eV, the summation of the formation energy and migration barrier of $V_{\rm{Li}}^{-}$ ({\it cf.}~Table \ref{tab:lialh}). This value is in good agreement with the reported experimental value (0.76 eV).\cite{oguchi:096104}
{\bf Decomposition mechanism.} As suggested by experimental data,~\cite{andreasen_2005,Ares20081263} the decomposition of LiAlH$_{4}$ into Li$_{3}$AlH$_{6}$, Al, and H$_{2}$, i.e., eqn (\ref{eq:step1}), necessarily involves hydrogen and/or aluminum mass transport in the bulk. Besides, local and global charge neutrality must be maintained while charged defects are migrating. Keeping these considerations in mind, we identify the following native defects as essential to the decomposition process:
First, H$_{i}^{-}$, which is expected to help form Li$_{3}$AlH$_{6}$ and/or LiH as discussed in the previous section. The activation energy for self-diffusion of H$_{i}^{-}$ is 0.87 eV, the summation of its formation energy and migration barrier. For other hydrogen-related charged defects such as H$_{i}^{+}$, $V_{\rm{H}}^{+}$, and $V_{\rm{H}}^{-}$, the activation energies are 1.36, 1.33, and 2.32 eV, respectively. These values are much higher than that reported for undoped LiAlH$_{4}$, suggesting that H$_{i}^{+}$, $V_{\rm{H}}^{+}$, and $V_{\rm{H}}^{-}$ are not the native defects that drive the decomposition process.
Second, $V_{\mathrm{AlH_{4}}}^{+}$, which is needed for the diffusion of aluminum-related species and the formation of Al phase. The activation energy for self-diffusion of $V_{\mathrm{AlH_{4}}}^{+}$ is 1.11 eV. Note that some other aluminum-related defects such as $V_{\mathrm{AlH_{2}}}^{+}$ and $V_{\mathrm{Al}}^{+}$ can also play this role. However, $V_{\mathrm{AlH_{4}}}^{+}$ has a lower activation energy and is thus expected to be dominant. Some others such as $V_{\mathrm{AlH_{3}}}^{0}$ and $V_{\mathrm{AlH}}^{0}$ have formation energies that are higher than that of $V_{\mathrm{AlH_{4}}}^{+}$, except for a small range of Fermi-level values above $\mu_{e}$, {\it cf.}~Fig.~\ref{fig;FE;Al}. The activation energies for the diffusion of $V_{\mathrm{AlH_{3}}}^{0}$ and $V_{\mathrm{AlH}}^{0}$ are at least 1.42 and 1.45 eV, respectively, which are also much higher than the experimental values.
Third, Li$_{i}^{+}$ and $V_{\mathrm{Li}}^{-}$, which can be created in the interior of the material in the form of a (Li$_{i}^{+}$,$V_{\mathrm{Li}}^{-}$) Frenkel pair. These low-energy and mobile native point defects can act as accompanying defects in hydrogen/aluminum mass transport, providing local charge neutrality as positively and negatively charged hydrogen- and aluminum-related defects moving in the bulk. They can also participate in mass transport that assists, e.g., the formation of Li$_{3}$AlH$_{6}$ and LiH. The activation energy for the formation and migration of (Li$_{i}^{+}$,$V_{\mathrm{Li}}^{-}$) is estimated to be 0.89 eV, which is the formation energy of the Frenkel pair plus the migration barrier of $V_{\mathrm{Li}}^{-}$.
Given these native defects and their properties, the decomposition of LiAlH$_{4}$ can be described in terms of the following mechanism: $V_{\mathrm{AlH_{4}}}^{+}$ is created at the surface or interface. This is equivalent to removing one (AlH$_{4}$)$^{-}$ unit from the bulk LiAlH$_{4}$. The formation energy of $V_{\mathrm{AlH_{4}}}^{+}$ is relatively low (0.68 eV) in the bulk. This energy is expected to be even lower at the surface, given that the bonding environment at the surface is less constrained than in the bulk. Formation of $V_{\mathrm{AlH_{4}}}^{+}$ is therefore quite likely during decomposition. Since (AlH$_{4}$)$^{-}$ is not stable outside the material, it breaks down into AlH$_{3}$ and H$^{-}$; AlH$_{3}$ subsequently leaves the material and dissociates into Al and H$_{2}$, whereas H$^{-}$ stays at the surface or interface. H$^{-}$ can diffuse into the bulk in form of H$_{i}^{-}$ and combine with an (AlH$_{4}$)$^{-}$ unit to form (AlH$_{5}$)$^{2-}$, an intermediate toward forming (AlH$_{6}$)$^{3-}$, which is essential in forming Li$_{3}$AlH$_{6}$ from LiAlH$_{4}$ according eqn (\ref{eq:step1}). In going from (AlH$_{4}$)$^{-}$ to (AlH$_{6}$)$^{3-}$, the anion unit attracts more and more Li$_{i}^{+}$ due to Coulomb interaction. Here, the highly mobile Li$_{i}^{+}$ will also help maintain local charge neutrality in the region near H$_{i}^{-}$. Note that, instead of forming (AlH$_{5}$)$^{2-}$, some H$_{i}^{-}$ may also be stable in the configuration where the hydrogen interstitial stands near Li$^{+}$ units to form a Li-H complex, {\it cf.}~Fig.~\ref{fig;VAlH3}. This complex may act as a nucleation site for the formation of LiH from LiAlH$_{4}$ according to eqn (\ref{eq:alternative}). In order to maintain the reaction, (AlH$_{4}$)$^{-}$ has to be transported to the surface/interface, which is equivalent to $V_{\mathrm{AlH_{4}}}^{+}$ diffusing into the bulk. As $V_{\mathrm{AlH_{4}}}^{+}$ is migrating, local charge neutrality is maintained by having the highly mobile $V_{\mathrm{Li}}^{-}$ in the vacancy's vicinity.
In the mechanism we just proposed, possible rate-limiting processes are the formation and migration of H$_{i}^{-}$, $V_{\mathrm{AlH_{4}}}^{+}$, and (Li$_{i}^{+}$,$V_{\mathrm{Li}}^{-}$) in the bulk LiAlH$_{4}$. Since $V_{\mathrm{AlH_{4}}}^{+}$ gives the highest activation energy (1.11 eV) among the three defects, with respect to the chosen set of atomic chemical potentials, we believe that the decomposition of LiAlH$_{4}$ is rate-limited by the formation and migration of this defect. The calculated activation energy is in agreement with the reported experimental values (0.84$-$1.15 eV).\cite{andreasen_2005,andreasen_2006,blanchard_2005,varin_2010} Also, because the decomposition and dehydrogenation processes occur at the surface/interface, ball milling that enhances the specific surface area and/or shorten the diffusion paths is expected to slightly enhance the hydrogen desorption kinetics. This is consistent with experimental observations.\cite{andreasen_2005,Ares20081263,varin_2010}
\section{\label{sec:impurities}Transition-metal impurities}
\begin{figure*}
\begin{center}
\includegraphics[width=5.2in]{LiAlH4_TM}
\end{center}
\vspace{-0.15in}
\caption{Calculated formation energies of transition-metal impurities in different charge states ($q$=0, $\pm$1, $\pm$2, and $\pm$3) in LiAlH$_{4}$, plotted as a function of Fermi level with respect to the valence-band maximum: (a) Ni$_{\mathrm{Al}}$, (b) Ni$_{\mathrm{Li}}$, (c) Ni$_{i}$, (d) Ti$_{\mathrm{Al}}$, (e) Ti$_{\mathrm{Li}}$, and (f) Ti$_{i}$.}\label{fig;FE;TM}
\end{figure*}
We investigate the role played by transition-metal impurities in the hydrogen desorption kinetics of LiAlH$_{4}$ by carrying out calculations for Ti and Ni impurities nominally on the Al and Li sites (i.e., M$_{\mathrm{Al}}$ and M$_{\mathrm{Li}}$, where M= Ti, Ni) and at interstitial sites (M$_{i}$). For each impurity, we performed calculations for several configurations by slightly breaking the symmetry in order to avoid local minima and to obtain the lowest-energy configuration. Figure \ref{fig;FE;TM} shows the calculated formation energies of Ti and Ni impurities in various charge states ($q$=0, $\pm$1, $\pm$2). The results are also summarized in Table \ref{tab:impurities} where we give values for the thermodynamic transition level $\epsilon$(+$q$/$-q$) between charge states $+q$ and $-q$, the associated $U$ value which is defined as $U$=$\epsilon$(0/$-q$)$-$$\epsilon(+q/0)$, the corresponding formation energy $E^{f}$, and the shift in Fermi-level position away from $\mu_{e}^{\rm int}$=2.83 eV (determined by the native defects) which is defined as $\Delta$=$\epsilon$(+$q$/$-$$q$)$-$$\mu_{e}^{\rm int}$. We find that all the substitutional impurities are negative-$U$ centers. Of highest interest are M$_{\mathrm{Al}}^{\pm1}$ and M$_{\mathrm{Li}}^{\pm1}$, which have the lowest formation energies near $\mu_{e}^{\rm int}$=2.83 eV. Regarding the interstitials, Ni$_{i}$ is a positive-$U$ center, whereas Ti$_{i}$ is a negative-$U$ center corresponding to a transition between Ti$_{i}^{2+}$ and Ti$_{i}^{2-}$.
\begin{table}[t]
\caption{Characteristics of Ti and Ni impurities in LiAlH$_{4}$. For Ni$_{i}$ where $U>$0, $E^{f}$ is the formation energy in the neutral charge state (Ni$_{i}^{0}$); this impurity is not effective in shifting the Fermi level and is therefore marked with a $\times$ sign. All the quantities are given in electronvolt (eV).}\label{tab:impurities}
\begin{tabular}{ccrr}
&&Ni&Ti \\
\hline
Al site
&$\epsilon(+/-)$&2.71&2.88 \\
&$U$&$-$1.56&$-$1.90 \\
&$E^{f}$&0.93&0.72 \\
&$\Delta$&$-$0.12&+0.05\\
\hline
Li site
&$\epsilon(+/-)$&2.39&2.41 \\
&$U$&$-$0.85 &$-$2.10 \\
&$E^{f}$&0.92&0.41 \\
&$\Delta$&$-$0.44&$-$0.42 \\
\hline
Interstitial
&$\epsilon$(+2/$-$2)&2.63&3.03 \\
&$U$&+0.95&$-$0.12 \\
&$E^{f}$&0.41&1.19 \\
&$\Delta$&$\times$&+0.20 \\
\end{tabular}
\begin{flushleft}
\end{flushleft}
\end{table}
\begin{figure*}
\begin{center}
\includegraphics[width=5.0in]{LiAlH4_TiHn}
\end{center}
\vspace{-0.15in}
\caption{Local structures of several substitutional and interstitial impurities in LiAlH$_{4}$: (a) Ti$_{\mathrm{Al}}^{+}$ and Ti$_{\mathrm{Al}}^{-}$, (b) Ti$_{\mathrm{Li}}^{+}$ and Ti$_{\mathrm{Li}}^{-}$, and (c) Ti$_{i}^{2+}$ and Ti$_{i}^{2-}$. Large (dark yellow) spheres are Ti, medium (blue) spheres Al, and small (red) spheres H. Only Ti$-$H and Al$-$H bonds with bond lengths smaller than 2 {\AA} are shown.}\label{fig;TiHn}
\end{figure*}
There are drastic changes in the lattice geometry as Ti and Ni impurities are incorporated. The transition-metal impurities tend to pull (AlH$_{4}$)$^{-}$ units closer and/or, in some cases, break Al$-$H bonds to form M$\mathrm{H}_{n}$ complexes. Figure \ref{fig;TiHn} shows the local structures of several substitutional and interstitial Ti configurations in LiAlH$_{4}$. The creation of Ti$_{\mathrm{Al}}^{+}$, for example, leads to the formation of a complex in which the Ti atom is surrounded by $n$=9 hydrogen atoms with the Ti$-$H distances ranging from 1.81 to 1.93 {\AA}, {\it cf.}~Fig.~\ref{fig;TiHn}(a). These Ti$-$H bond lengths are comparable to that in bulk TiH$_{2}$, which is 1.92 {\AA}. Other impurities such as Ti$_{\mathrm{Al}}^{-}$, Ti$_{\mathrm{Li}}^{+}$, Ti$_{\mathrm{Li}}^{-}$, Ti$_{i}^{2+}$, and Ti$_{i}^{2-}$ also lead to formation of TiH$_{n}$ ($n$=6$-$8) complexes with the Ti$-$H distances ranging from 1.78 to 1.95 {\AA}. Similarly, Ni impurities also form NiH$_{n}$ complexes, however the $n$ value is smaller than those of Ti. For example, Ni$_{\mathrm{Al}}^{+}$ is surrounded by 6 hydrogen atoms, instead of 9 hydrogen atoms in the case of Ti$_{\mathrm{Al}}^{+}$. This difference is due to the fact that Ni has fewer empty 3$d$ states than Ti.
Following our argumentation presented previously for other complex hydrides such as NaAlH$_{4}$, LiBH$_{4}$, and Li$_{4}$BN$_{3}$H$_{10}$,~\cite{peles_prb_2007,hoang_prb_2009,wilson-short_prb_2009} the Fermi level of the system is determined by these electrically active impurities if their concentration is larger than that of the native defects. Chances are that this Fermi-level position will be different from the one that would occur in the absence of the impurities. In this case the formation energy of some of the native defects will be reduced. For those impurities that have negative-$U$ character, the Fermi level will simply coincide with the $\epsilon$(+$q$/$-q$) transition level of the impurities.\cite{hoang_prb_2009} For those impurities that exhibit positive $U$ and $\epsilon$(+$q$/0)$<$$\mu_{e}^{\rm int}$$<$$\epsilon$(0/$-q$) such as Ni$_{i}$, {\it cf.}~Fig.~\ref{tab:impurities}, the the impurity have no effect since it prefers the neutral charge state over a range of Fermi-level values that includes the Fermi-level position where the lithium-related defects switch charge state.\cite{hoang_prb_2009} Clearly, the effects of Ni and Ti impurities on shifting the Fermi level depend sensitively on how the impurities are incorporated into LiAlH$_{4}$.
For Ni$_{\rm{Al}}$, Ni$_{\rm{Li}}$, and Ti$_{\rm{Li}}$, the Fermi-level shift will result in lowering the formation energy of $V_{\mathrm{AlH_{4}}}^{+}$. Since the formation and diffusion of $V_{\mathrm{AlH_{4}}}^{+}$ corresponds to formation of H$_{2}$ (and Al) in the mechanism we proposed for the decomposition and dehydrogenation of LiAlH$_{4}$, this may lead to a reduction in the onset decomposition temperature (or the effective decomposition temperature) as seen in experiments. This shift, however, would also result in increasing the formation energy of H$_{i}^{-}$. In this way, Ni$_{\rm{Li}}$, Ni$_{\rm{Al}}$, and Ti$_{\rm{Li}}$ lower the activation energy associated with $V_{\mathrm{AlH_{4}}}^{+}$ but increase that associated with H$_{i}^{-}$. Upon introducing the impurities into LiAlH$_{4}$, $\Delta$ changes from 0 eV up to $-$0.12 eV (Ni$_{\rm{Al}}$), $-$0.44 eV (Ni$_{\rm{Li}}$), or $-$0.42 eV (Ti$_{\rm{Li}}$). Note that, as the Fermi level is shifted, some charged native defects may occur with higher concentrations to counteract the effects of the electrically active impurities, and reduce the magnitude of $\Delta$. When $\Delta$ reaches $-$0.12 eV, the activation energies associated with $V_{\mathrm{AlH_{4}}}^{+}$ and H$_{i}^{-}$ are both equal to 0.99 eV. For lower $\Delta$ values, the activation energy associated with H$_{i}^{-}$ is higher than that associated with $V_{\mathrm{AlH_{4}}}^{+}$. As a result, the process associated with H$_{i}^{-}$ becomes the rate-limiting step in hydrogen desorption. Our analyses, therefore, indicate that Ni$_{\rm{Al}}$, Ni$_{\rm{Li}}$, and Ti$_{\rm{Li}}$ can modify the activation energy. In fact, these impurities can reduce the activation energy by up to about 0.12 eV, which is in good agreement with some experimental observations.~\cite{blanchard_2005,rafiuddin_2010,varin2010928} Using similar arguments, we predict that Ti$_{i}$ would not lower the activation energy for hydrogen desorption according to the proposed mechanism.
\section{\label{sec:sum}Conclusion}
We have carried out a comprehensive first-principles study of native defects and transition-metal impurities in LiAlH$_{4}$. The compound is found to be prone to Frenkel disorder on the Li sublattice. Lithium interstitials and vacancies have low formation energies and are highly mobile, which can participate in lithium-ion conduction, and act as accompanying defects in hydrogen mass transport. We have proposed a specific mechanism for the decomposition of LiAlH$_{4}$ that involves the formation and migration of H$_{i}^{-}$, $V_{\mathrm{AlH_{4}}}^{+}$, and (Li$_{i}^{+}$,$V_{\mathrm{Li}}^{-}$) in the bulk LiAlH$_{4}$. Our calculated activation energy is in agreement with the experimental values. In light of this atomistic mechanism, we are able to explain the decomposition and dehydrogenation of LiAlH$_{4}$, the rate-limiting step in the hydrogen desorption kinetics, and the effects of transition-metal (Ti and Ni) impurities on the onset decomposition temperature and on the activation energy. Our results also suggest that it is the structure of H$_{i}^{-}$ that determines the hydride phase (Li$_{3}$AlH$_{6}$ or LiH) in the decomposition products. This relationship should be further explored in other complex hydrides.
\begin{acknowledgments}
K.~H.~was supported by General Motors Corporation, and A.~J.~by the U.S.~Department of Energy (Grant No. DE-FG02-07ER46434). This work made use of NERSC resources supported by the DOE Office of Science under Contract No.~DE-AC02-05CH11231 and of the CNSI Computing Facility under NSF Grant No.~CHE-0321368. The writing of this paper was partly supported by Naval Research Laboratory through Grant No.~NRL-N00173-08-G001.
\end{acknowledgments}
|
\section{Introduction}
Cells and organisms sense, compute, and make decisions: to proliferate, to find food, to protect themselves against predators and unfavorable environmental changes, to act in unison with their neighbors within a collective, and---broadly speaking---to invest their limited resources to their maximal benefit. These processes span a wide range of temporal and spatial scales, and reflect dynamics in phase spaces of widely varying intrinsic dimensionality. In bacteria, for example, the chemotactic signaling network comprises just a handful of chemical species, responds to changes in the local nutrient concentration on the one second timescale, and modulates bacterium's swimming by controlling the direction of flagellar rotation. At the other extreme of complexity, millions of neurons in the brain of a pianist are engaged in processing roughly one megabit per second of visual, auditory, and tactile information, combining these data with the score retrieved from memory, and computing the fine motor program controlling the pianist's fingers. Other processes take place over much longer timescales: DNA sequences in a population of organisms are subject to the evolutionary forces of mutation, selection, and random drift, and changes in the environment slowly shape the distribution of genomic sequences.
In all these cases, intuition tells us that information is ``flowing'' from the outside environment to a representation internal to the organism or a population, and that this information flow is essential for life. Can this intuition be formalized? If so, can it be connected, quantitatively, to experimental data? Most ambitiously, can a focus on information flow help us to unify our description of the complex and diverse phenomena that we see in the living world?
The purpose of this review is to explain how the intuition about ``biological information flows'' can be formalized in the language of information theory, and what the resulting research program teaches us about the physics of living systems. We introduce the basic concepts of information theory (Section~\ref{sec2}), show how this mathematical structure leads to new views of experimental data (Section~\ref{sec3}), and explore the possibility that some of life's mechanisms can be understood as solutions to the problem of optimizing information flow subject to physical constraints (Section~\ref{sec5}). Throughout, we follow the physics tradition of emphasizing the unity of theoretical ideas at the expense of some of the biological complexity.
Before we begin, it seems worth discussing how ideas about information flow relate to the more familiar theoretical structures in physics. Intuitions about information flow are not confined to biological systems. When we look a ferromagnet, for example, we know that all the details of the electronic configurations on one side of the sample are not relevant if we try to predict those on the other side of the sample; what does carry information across long distances is the order parameter, in this case the magnetization. But precisely because we can identify the order parameter, we don't need to speak abstractly about information. Rather than trying to turn ``carry information'' into a precise mathematical statement, we just compute correlation functions of the order parameter evaluated at distant points.
For most of the biological systems that we will be discussing, finding something like an order parameter remains a distant hope. Symmetries are absent, and even locality is not much of a guide to understanding, for example, a network of neurons in which each cell may be connected to many thousands of neighbors. Under these conditions, some more general approaches to formalizing our intuitions about information flow would be useful. Interestingly, as we write this, information theoretic ideas are becoming more important in the study of topological phases in correlated electron systems, where no local order parameter is possible.
\section{Building blocks of information theory}
\label{sec2}
Here we give a brief exposition of information theory, focusing on topics that are most relevant for our subsequent discussion; see also Ref \cite{bialek_12}. The foundational papers by Shannon are wonderfully readable \cite{shannon_48}, there is an excellent textbook account aimed at the core of the subject \cite{cover+thomas_91}, and a more recent text that emphasized connections between ideas information theory and statistical physics \cite{mezard+montanari_09}. Ideas which have been most widely used in thinking about biological systems are not necessarily those that have been most useful in the (enormously successful) application of information theory to communication technology, so our perspective here is somewhat different from that in Ref \cite{cover+thomas_91}.
When we ask a question, and hear the answer, we gain information. In 1948 Shannon asked whether we could attach a number to the amount of information we gain \cite{shannon_48}. He imagined that, upon asking the question, we could envision a set of possible answers (${\rm n} = 1,\, 2,\, \cdots ,\, N$) and that we could assign probabilities ${\mathbf p} \equiv \{p_1, \, p_2,\, \cdots ,\, p_N\}$ to each of these answers. Any acceptable measure of information must obey some constraints: if all $N$ possible answers are equally likely, then the information must grow with $N$; if the question can be decomposed into independent parts, then the information gained on hearing the answer to each part should add to the total; if we can decompose the full question into a tree of choices (as we do when playing twenty questions), then the total information should be the weighted sum along the paths through the tree. Remarkably, these postulates allow only one possible measure of information, the entropy of the distribution of answers,\footnote{Shannon's theorem makes precise the vague statement made in many statistical mechanics classes, that entropy is related to our lack of information about the microscopic state of a system. The more positive version is that the entropy is the information we would gain upon learning the full microscopic state.} $S[{\mathbf p}]= -k\sum_{\rm n} p_{\rm n}\log p_{\rm n}$.
A seemingly very different question concerns the amount of space required to write down, or represent the answer to our question. While each possible answer might require a different amount of space, if this is a question we can ask many times (what will the temperature be at noon tomorrow?) then it makes sense to ask about the minimum amount of space required, per answer, in the limit of a long sequence of question/answer pairs. The answer is again the entropy. There is some ambiguity about units---the number of characters we need to write the answer depends on our choice of alphabet, and the entropy itself has an arbitrary constant $k$---but if we choose, for example, a binary alphabet, we can set $k=1$ and $\log \equiv \log_2$, then all of the ambiguity is resolved. When entropy or information is measured in these units, they are called `bits.'
Getting the precise answer to a question is a rare thing. More common is that we are interested in $x$ but can only measure $y$. Before a measurement we know only that $x$ is drawn from the distribution $P_X(x)$; after the measurement we know that it is drawn from the conditional distribution $P(x|y)$. The same arguments that establish the uniqueness of the entropy now show that the only acceptable measure of the information about $x$ gained by observing $y$ is the reduction in entropy,
\begin{eqnarray}
I (y\rightarrow x) &=& -\sum_x P_X(x) \log_2 P_X(x) \nonumber\\
&&\,\,\,\,\, - \left[ -\sum_x P(x|y) \log_2 P(x|y) \right].
\label{IdiffS1}
\end{eqnarray}
If we ask not about the information gained from observing a particular $y$, but the average over all the $y$'s we can observe, this average information is actually symmetric in $x$ and $y$, and is usually called the ``mutual information.'' It can be written in a manifestly symmetric form, and unlike the entropy itself it is well defined even when $x$ and $y$ are continuous variables,
\begin{equation}
I(x;y) = \iint dx\;dy\; P(x,y) \log_2\left[\frac{P(x,y)}{P_X(x)P_Y(y)}\right] , \label{info1}
\end{equation}
where $P(x,y)$ is the joint distribution of $x$ and $y$. This can be extended to the case where $x$ and $y$ are multidimensional, or even functions, in which case the sums over states become functional integrals as in quantum mechanics or field theory.
One way to think about the mutual information is as a measure of dependency or correlation. Clearly we have $I(x;y) = 0$ if the variables $x$ and $y$ are independent, so that $P(x,y) = P_X(x)P_Y(y)$, and Shannon's arguments tell us that $I(x;y)$ is the only measure of correlation that satisfies a set of sensible requirements. If $x$ and $y$ are jointly Gaussian, then
\begin{equation}
I(x;y) = -\frac{1}{2}\log_2(1-c^2),
\end{equation}
where $c$ is the usual linear correlation coefficient. But it is easy to imagine relations between $x$ and that are highly informative but not linear. Even if the underlying variables are linearly correlated, the things we measure might be nonlinearly transformed version of these variables, and in this case computing the mutual information becomes an especially powerful way of analyzing the data; see Section \ref{parameters} below.
The more conventional way of thinking about the mutual information is in terms of a communication channel. There is a input message $x$, and it is sent along channel that produces the output $y$. The fact that $y$ is chosen from a probability distribution [$P(y|x) = P(x|y)P_Y(y)/P_X(x)$] means that the communication is noisy, and this must limit the amount of information that can be transmitted. If we think of the mutual information as being a functional of the distributions $P_X(x)$ and $P(y|x)$, then the convexity of the logarithm means that there is a maximum as functional of the distribution of inputs; this maximum is called the channel capacity,\footnote{It is a remarkable fact that one can transmit information {\em without error} along a channel that has a non--zero noise level, provided that the rate of transmission does not exceed this capacity. Strictly speaking ``capacity'' usually is used to describe this maximum rate of errorless transmission, rather than the maximum of the mutual information between input and output \cite{cover+thomas_91}.} and is the limit to information transmission set by noise in the channel.\footnote{In engineered systems, much emphasis is placed on devising coding schemes, i.e., algorithms that transform inputs $x$ into messages to be sent through the channel (and likewise recover $x$ from the channel output), so that information can be transmitted over noisy channels as close to capacity as possible using bounded processing resources. While interesting, coding questions have remained largely unexplored in biological systems.}
We note that actually finding the capacity of a channel, given a detailed model of the input/output relation and noise as represented by $P(y|x)$, can be challenging. Simple models, such as $y$ being a linearly filtered version of $x$ with added Gaussian noise, are tractable, and generalize in a straightforward fashion to cases where the signals are varying in time. For this reason, Gaussian channels have dominated the discussion of both biological and engineered systems, and the capacity-achieving strategies that depend on the system's input and output noise levels are well-understood \cite{bialek_12,cover+thomas_91,berger_71}. Beyond the Gaussian channel, progress has been made using various approximative schemes, such as by assuming that noise is small compared to the dynamic range of the signal (the small-noise approximation), but there is certainly room for further theoretical developments using biophysically-informed noise models.
Examples of information transmission through a noisy channel are everywhere in biological systems. Most obviously, our eyes and ears take inputs from the outside world and the neurons emerging from our sense organs generate streams of identical electrical pulses, termed action potentials or spikes. Even single celled organisms have sensors, as with the bacterial chemotaxis system, which responds to changing concentrations of molecules in the environment and generates at the output changes in the direction of swimming. There are also purely internal examples, as when a cell synthesizes ``transcription factor'' proteins that bind to DNA and influence the rate at which encoded information is read out to make other proteins; here the transcription factor concentration is the input and the resulting protein concentration is the output. In some of those cases variables are naturally continuous (sound pressure), in other cases they are naturally discrete (counting individual molecules), and in many cases it is useful to pass between descriptions that are either discrete or continuous (e.g, concentration instead of counting molecules); it is attractive that we can discuss information transmission in all these cases, in the same language and the same units.
Information is a difference in entropies, as in Eq (\ref{IdiffS1}), and because information is mutual we can write this in two ways,
\begin{eqnarray}
I(x;y) &=& S[P_Y(y)] - \langle S[P(y|x)]\rangle_{P_X(x)} \label{info2} \\
&=& S[P_X(x)] - \langle S[P(x|y)]\rangle_{P_Y(y)}, \label{info3}
\end{eqnarray}
where $\langle\cdot\rangle_{p(x)}$ denotes an average over $p(x)$. If either the input or the output is discrete, then both the entropy and the conditional entropy are positive semi--definite, and hence the entropy sets an upper bound on the information. Further, if we can assign a cost to every output state $y$---e.g., because generating more action potentials costs more energy, then there is a maximum entropy consistent with a given average cost (for more on maximum entropy see Section \ref{maxent}). Thus there is a rigorous path from limits on physical resources to limits on information transmission.
Another way of bounding information is through the data processing inequality. If the output $y$ receives information about the input $x$ only through some intermediate $z$, we can write
\begin{equation}
P(y|x) = \sum_z P_{\rm II} (y|z) P_{\rm I} (z|x) .
\end{equation}
Now there are two communication channels, first $x\rightarrow z$ (channel $\rm I$) and then $z\rightarrow y$ (channel $\rm II$). Each channel is associated with a mutual information, $I_{\rm I}(x; z)$ and $I_{\rm II} (z; y)$, and each of these is an upper bound on the information transmission through the combined channel, $I(x; y)$. Another way of saying this is that any processing from $z$ to $y$ can only result in a loss of information, so that $I(x;y) \leq I_{\rm I} (x;z)$. In this sense, information can be destroyed, but not created. This result is conceptually important, and we will see that it is of practical use in analyzing experiments (Section \ref{parameters}). As for the mutual information, so too for the channel capacities: to support mutual information $I(x;y)$, between $x$ and $y$, the capacities of both channels $\rm I$ and $\rm II$ must be larger than $I(x;y)$.
The universality of information-theoretic quantities is their strength, but it also leaves one with an uncomfortable feeling that there is more to the story of biological signal processing than bits. In particular, not all bits may be of equal value or utility to the organism. Information processing usually has a goal beyond communication, and this seems especially true in biological systems. Perhaps because Shannon called his work a mathematical theory of communication, it is not as widely appreciated that information theory has anything to say about such goal--defined tasks. The key idea is rate--distortion theory, which again traces back to Shannon \cite{shannon_59,berger_71}.
As an example, consider a single celled organism in an environment where various nutrients are present at concentrations ${\mathbf c} = \{c_1,\, c_2,\, \cdots ,\, c_N\}$, and over time this environmental state varies. To metabolize these molecules, the cell synthesizes $e_1$ molecules of enzyme $1$, $e_2$ copies of enzyme $2$, and so on up to enzyme $K$; to fully convert the inputs into useful molecules, there may need to be more enzymes than nutrients. The growth rate of the cell depends on both the environment and the state of the cell, $\lambda ({\mathbf c},{\mathbf e})$. Under any particular environmental condition $\mathbf c$, there is an optimal setting of the internal state $\mathbf e$ that maximizes the growth rate. If the cell finds this optimum exactly, then by looking at $\mathbf e$ we would know exactly the environmental variable $\mathbf c$, and this exact knowledge of real variables corresponds to an infinite amount of information. At the opposite extreme, the cell could choose the copy numbers of various enzymes at random, and hope that, averaged over the distribution of environments, the average growth rate is not so bad; in this case the internal state of the cell carries zero information about the environment. But neither infinite nor zero information seems plausible; what happens in between? Rate--distortion theory tells us that if the cell needs to achieve a given average growth rate, then there is a minimum required mutual information between the state of the environment and the internal state of the cell, $I ({\mathbf c},{\mathbf e}) = I_{\rm min} (\langle\lambda\rangle )$.
More generally, if we have inputs $x$ and outputs $y$, then given some measure of performance ${\cal U}(x,y)$, we can plot the average performance vs. the mutual information $I(x;y)$. The rate--distortion curve $I_{\rm min} (\langle {\cal U}\rangle )$ divides this plane into accessible and inaccessible regions. Given a certain number of bits, there is a maximal achievable performance, and conversely achieving any given level of performance requires a minimum number of bits. Thus, although collecting and transmitting bits cannot be the goal of an organism, any real goal requires a minimum number of bits. If bits are cheap, then this is a true but not very useful observation. If, on the other hand, the limited resources available to cells and organisms mean that more bits are hard to get, then selection for better performance will translate into selection for mechanisms that gather and process more bits with the same resources.
Rate distortion is a framework that formalizes lossy data compression: signals $x$ can be sent through the channel to yield outputs $y$ which are then used to find the best achievable reconstruction $\hat{x}(y)$ of the input signal. The metric for the reconstruction quality is the distortion function, for example, the RMS error between the true signal $x$ and its reconstruction $\hat{x}$. In this sense, rate distortion tells us the minimum number of bits to recover the input signal to within a desired level of quality. When thinking about biological systems, it is often convenient to turn this reasoning around and explicitly construct various decoding or readout schemes that reconstruct the input signals from the activity of the biological network. Classic examples involve decoding of motion trajectories from the spiking activity of neurons in the visual system \cite{bialek+al_91,marre+al_13}, or the decoding of hand trajectories from the activity of neurons in motor cortex during drawing \cite{schwartz_94}; for more on decoding of neural activity see Ref \cite{spikes}. Decoding schemes can be used to put a lower bound on information transmission (Section~\ref{sec3}) and might further suggest how much of the total information is encoded explicitly, in a way that can be extracted using biologically plausible mechanisms. Indeed, the relationship between the total information and the component which is accessible in simple form is an issue in neural coding that needs to be explored more fully.
\section{Information and data analysis}
\label{sec3}
In complex systems, we may be unsure which features of the system dynamics are the most relevant. Even if we can identify the correct features, we might not know the biologically meaningful metric that measures the similarity or difference along these relevant dimensions. Information theory gives us a way out of these difficulties, because it singles out measures that are completely general: while some particular correlation function of $X$ and $Y$ might vanish, if these variables are related in any way then there must be a nonzero mutual information $I(X;Y)$; further, there is an obvious sense in which the stronger the relationship, the greater the mutual information. Against this generality is the difficulty of actually estimating the entropy of a probability distribution from a limited set of samples.
\subsection{Quantifying real information flows}
\label{measure_info}
In thermodynamics, entropy changes are related to heat flows, and this can be used to measure the entropy in equilibrium systems. Away from equilibrium statistical mechanics, we have no such general method for measuring entropy. Indeed, the entropy depends on the entire probability distribution, and so ``measuring'' the entropy is quite unlike measuring the usual observables.
We can estimate the entropy of a probability distribution by counting how often each state occurs, using this frequency as an estimate of the probability, and plugging into the definition $S = -\sum_{\rm n}p_{\rm n} \log_2 p_{\rm n}$. While the errors in estimating the individual probabilities $p_{\rm n}$ are random, the convexity of the logarithm means that these random errors contribute to a systematic error than declines as $1/\sqrt{N_s}$, where $N_s$ is the number of samples. The difficulty is that the coefficient of this systematic error is proportional to the number of possible states of the system, and hence accurate estimates require many more samples than states. The problem becomes worse when we try to estimate the mutual information between two variables, since the number of possible states is the product of the number of states for each variable taken separately.
The problem of systematic errors in estimating entropy and information was appreciated very early, in the 1950s, as information theoretic ideas began to be applied to biological problems \cite{miller_55}. Motivated in large part by efforts to measure the information that neural responses carry about sensory inputs, there has been a substantial effort to calculate and correct for these systematic errors \cite{panzeri+treves_95,strong+al_98,paninski_03}. A different direction is to note that the single number, entropy, might be well determined even if the full distribution is not. This idea can be made concrete in a Bayesian approach, placing a prior on the space of probability distributions; priors that are uniform in entropy, as opposed to uniform in probability, seem especially effective \cite{nemenman+al_02,nemenman+al_04,archer+pillow_12,archer+pillow_13}. In many cases, the variables of interest are not naturally discrete, and so the problem of entropy estimation is tangled with that of density estimation \cite{victor_02}. Although much progress has been made, the difficulty of estimating entropy and information from finite samples leads to periodic inventions of other measures that might serve similar purposes (e.g., quantifying dependency beyond linear correlation); for a recent reminder of why mutual information is not just one measure among many, see Ref \cite{kinney+atwal_14}.
An important ingredient in estimating entropy and information from real data has been the use of bounds. It may be difficult to estimate the entire probability distribution, but it often is feasible to measure, with small error bars, several average quantities (e.g., moments), and construct the probability distribution that matches these averages exactly but has the maximum possible entropy (Section \ref{maxent}). The data processing inequality means, for example, that if we want to measure the mutual information between $x$ and $y$, and $y$ is of high dimensionality, any projection into a lower dimensional space, $y\rightarrow z$, can only result in a loss of information, $I(z; x) \leq I(y;x)$. Reducing dimensionality makes the problem of estimating distributions much easier, and thereby eases the sampling problems of entropy estimation; since the lower dimensional variable provides a lower bound to the true information, it makes sense to search for optimal projections (Section \ref{parameters}). Finally, as emphasized long ago by Ma \cite{ma_81}, the probability that two randomly chosen states are the same, $P_c = \sum_{\rm n} p_{\rm n}^2$, provides a lower bound to the entropy, $S \geq -\log_2 P_c$, that can be estimated reliably from relatively few samples. For more details, see Appendix A.8 of Ref \cite{bialek_12}.
A modest though widespread use of information theory in the analysis of biological data concerns the description of regulatory sites along DNA. The readout of information encoded in the genome---the synthesis of proteins---is regulated in part by the binding of ``transcription factor'' proteins (TFs) to these sites, and the strength with which one TF binds depends on the DNA sequence at the binding site. In this sense, the local sequences can be thought of as ``addresses'' that distinguish certain genes from the rest, but it is known that transcription factors do not do exact string matching; instead, they bind to an ensemble of sites that permit some variation. Early work used simple arguments to compute how much ``information'' is contained in these binding site ensembles, i.e., as the difference between the entropy of random sequences and the difference of the ensemble that the TF will bind \cite{berg+hippel}, and this has given rise to a popular graphical way of depicting TF binding site specificity in the form of ``sequence logos'' \cite{seqlogo}. Subsequent work has refined these arguments and asked whether the information in the binding sites is sufficient to address specific genes in genomes of pro- and eukaryotes \cite{wunderlich+mirny}.
Meaningful quantitative statements about information flow obviously depend on having high quality data. In the nervous system, the fact that the relevant output signals are electrical means that quantitative measurements have been the norm since the invention of the vacuum tube \cite{adrian_28}. Steady improvements in technology now make it possible to collect sufficient data for information theoretic analyses in almost any neural system, even in freely behaving animals; in practice, much of the motivation for these analyses has come from ideas about the optimization of the neural code, and so these measurements are reviewed below (Section IV.B). In biochemical and genetic networks, the fact that outputs are the concentrations of molecules means that there is no universal measurement tool.
Classical methods for measuring the concentrations of signaling and control molecules inside cells include raising antibodies against particular molecules of interest, and then tagging antibodies against these antibodies with fluorescent labels. Recent work in the fly embryo shows that, with care, this classical technique can pushed to give measurements of concentration that are accurate to $\sim 3\%$ of the relevant dynamic range \cite{dubuis+al_13a}. More modern methods involve genetically engineering the organism to produce fusion of the protein of interest with a fluorescent protein \cite{tsien_98}, and here too there has been progress in demonstrating that the fusion proteins can replace the function of the native protein, quantitatively \cite{gregor+al_07a,morrison+al_12}. Finally, rather than monitoring protein concentrations, one can count the messenger RNA molecules (mRNA) that are an intermediate step in protein synthesis; recent work has pushed to the point of single molecule sensitivity, so that counting is literal \cite{raj+al_08,little+al_13}. Similarly, progress has been made in quantifying the activity of intracellular signaling networks, and subsequently using information theoretic approaches to compare how signals propagate in different network architectures \cite{cheong+al_11}, or how they are encoded in the dynamical properties of the response \cite{selimkhanov+al_14}.
\subsection{Quantifying correlations}
In the last decade it has become possible to perform simultaneous recordings from tens or hundreds of nodes in interacting biological interacting networks. Gene expression arrays and sequencing technology have enabled such readouts for gene expression patterns; cheap sequencing can map the distributions of gene variants in populations; multi-electrode arrays and lately optical imaging have pushed the boundaries in neuroscience to the point where full temporal activity traces for every neuron in small brains---a zebrafish larva, the worm {\em C elegans}, or the fruit fly---don't seem out of reach; and in the field of collective behavior, simultaneous tracing of positions of hundreds or thousands of individuals can be performed. Because the barriers between subfields of biology are large, each of these experimental developments has led to an independent discovery of the need for new and more powerful analysis tools. Even the simple question of when the activity at two nodes in the network is correlated is complicated, because it is surely not true that linear correlations capture the full relationships among pairs of variables in these complex systems.
As microarray methods provided the first global pictures of gene expression in cells, several groups tried to analyze the covariation in expression levels simple by computing correlations. With enough data, it became possible to estimate the mutual information between expression levels \cite{slonim+al_05a}. These initial analyses revealed that, as expected, linear correlations did not exhaust the relationships between genes; while the most correlated pairs had the highest mutual information, at intermediate or even minimal levels of correlation one could find pairs of genes whose expression levels shared nearly half the mutual information of the most correlated pairs. In addition, it became clear that pairs of expression levels could share more than one bit of mutual information, which contributed to the realization that the control of gene expression involved more than just on/off switches.
Measurements of mutual information also provide a new basis for clustering, in which we try to maximize the average mutual information between the expression levels of pairs of genes in the same cluster \cite{slonim_05}. This approach is interesting mathematically because it is invariant to any invertible transformations of the expression levels (and hence to systematic distortions in the measurement), and because the clusters have no ``centers'' or prototypical members. These information theoretic methods were able to reveal biologically significant structures even inside tightly correlated ``modules'' where other methods failed. Further developments place the estimates of mutual information among many pairs of gene expression levels at the core of algorithms for reconstructing the underlying network of interactions, where the data processing inequality provides a guide to disentangling direct interactions from more widespread correlations \cite{aracne}.
One can look at interactions between the genes in a very different way, analyzing joint presence or absence of particular genes in the sequences of many organisms. Here, each ``sample,'' i.e., one organism's genome, is represented as a long binary vector, where 1 represents a presence of a particular gene and 0 its absence.\footnote{Some care is needed here. Two different organisms almost never have genes with exactly the same sequences, but if we are comparing organisms that are not too far separated in evolutionary distance, then finding the corresponding or ``homologous'' genes is relatively straightforward. It is more difficult to be certain of the historical path that generated these homologs: descent from a common ancestor (orthologs'), convergent evolution, gene duplication (paralogs), ... . For more about sequence similarity, see Section \ref{maxent}.} Estimating the mutual information between the presence/absence of pairs of genes is technically straightforward, and reveals the organization of genes into clusters \cite{bowers}; this approach can be extended beyond pairs to probe higher--order interactions among genes. Information theoretic measures can also be used to quantify the correlations between the presence/absence of particular genes and various phenotypes, and these analyses point toward functional modularity in the organization of genomes \cite{slonim+elemento}.
\subsection{Systems with many degrees of freedom}
\label{maxent}
One of the great triumphs of the twentieth century is the identification and characterization of the molecular building blocks of life. But life is more than the sum of its parts. From the spectacular aerial displays by flocks of birds down to the coordinated movements of cells in a gastrulating embryo, many of the most striking phenomena in the living world are the result of interactions among hundreds, thousands, or even millions of individual elements. The enormous success of statistical physics in describing emergent phenomena in equilibrium systems has led many people to hope that it could provide a useful language for describing emergence in biological systems as well. Perhaps the best developed set of ideas in this direction concerns the dynamics of neural networks \cite{hopfield_82,amit_89,hertz_91}, but there have also been important developments in the description of collective animal behavior \cite{toner+tu_95,toner+tu_98,ramaswamy_10} and in models for genetic networks. While each case has its own special features, it seems fair to say that, in all these cases, the search for simple, analyzable models often takes us away from contact with the details of measurements on the real biological systems. In extreme cases, it is not clear whether these models are just metaphors (the alignment of birds in a flock is like the alignment of spins in a magnet), or if they are to be taken seriously as theories that make quantitative predictions. Over the past decade, progress in experimental techniques, as noted in the previous section, have held out the promise of much more detailed confrontations between theory and experiment.
Although there are many interesting questions about emergent behavior in biological networks, we focus here on one question where information theoretic ideas have been crucial: if there a variable $y_{\rm i}$ that lives at each node of the network, can we make a model for the joint distribution of these variables, $P_y({\mathbf y}) \equiv P_y(y_1 ,\, y_2, \, \cdots ,\, y_N)$? Progress on this problem has been made using the idea of maximum entropy \cite{jaynes_57}.
Let us assume that experiments provide $T$ observations on the state of the system, $\mathbf{y}$. Often, the dimension of the system is so large that $T$ samples are completely insufficient to empirically sample $P_y(\mathbf{y})$. However, we might be able to estimate certain statistics on the data, which we denote by $\langle \hat{O}_\mu\rangle_{\rm expt}$, where $\mu=1,\dots,M$. These could be, for example, expectation values of individual network nodes, $\langle y_i\rangle$, the covariances, $\langle y_iy_j\rangle$, or any other function for which the samples provide a reliable estimate. We look for a model distribution $\hat{P}_y(\mathbf{y})$ that will exactly reproduce the measured expectation values, but will otherwise be as unstructured, or random, as possible---hence maximum entropy. This amounts to solving a variational problem for $\hat{P}$, for which the solution is:
\begin{equation}
\hat{P}^{(\hat{O}_1,\dots,\hat{O}_M)}(\mathbf{y})=\frac{1}{Z(\{g_\mu\})}\exp\left[\sum_{\mu=1}^M g_\mu \hat{O}_\mu(\mathbf{y})\right]. \label{maxent1}
\end{equation}
Thus, the maximum entropy distribution has an exponential form, where the coupling constants $g_\mu$ have to be chosen so that the constraints $\langle \hat{O}_\mu\rangle_{\rm expt}=\langle \hat{O}_\mu\rangle_{\rm \hat{P}}$ are exactly satisfied. Equation (\ref{maxent1}) has the form of a Boltzman distribution in which every state of the system is assigned an ``energy'' $E({\mathbf y}) = \sum_{\mu=1}^M g_\mu \hat{O}_\mu(\mathbf{y})$, and this equivalence to equilibrium statistical mechanics is an important source of intuition and calculational tools. But Eq (\ref{maxent1}) is {\em not} a description of a system in thermal equilibrium. If the only thing we measure about a system is its energy, then asking for the maximum entropy distribution consistent with the average energy does indeed correspond to a description of thermal equilibrium, but in most biological systems the quantities that we measure have no relation to the energy, and we are measuring many of them.
The maximum entropy method does not correspond to a single model for the system, but rather provides a systematic strategy for building a hierarchy of models that provide increasingly good approximations to the true distribution. At each step in the hierarchy, the models are parsimonious, having the minimal structure required to reproduce the expectation values that we are trying to match. For every new expectation value that we can measure, the new maximum entropy model provides a tighter upper bound on the entropy of the true distribution, and can uniquely decompose the total amount of correlation among the $N$ nodes in the network into contributions from pairwise, triplet-, $\dots$, $K$-fold interactions \cite{schneidman+al_03}; the hope is that we arrive at a good approximation while $K\ll N$.
Equation (\ref{maxent1}) tells us the form of the maximum entropy distribution, but to complete the construction of the model we need to find the coupling constants $\{g_\mu\}$. Usually in statistical mechanics we are given the coupling and try to compute the expectation values or correlation functions ($\{g_\mu\} \rightarrow \{\langle{\hat O}_\mu ({\mathbf y})\rangle\}$). Here we need to do inverse statistical mechanics, mapping the experimentally measured expectation values back to the coupling constants ($\{\langle{\hat O}_\mu ({\mathbf y})\rangle\} \rightarrow \{g_\mu\}$). Such inverse problems have a history dating back (at least) to the demonstration by Keller and Zumino \cite{keller+zumino_59} that the temperature dependence of the second virial coefficient in a classical gas determines the interaction potential between molecules uniquely, provided that this potential is monotonic; the search for rigorous statements about inverse statistical mechanics has attracted the attention of mathematical physicists \cite{chayes+al_84,caglioti+al_06}, although many questions are open. At a practical level, estimating the coupling constants from measured expectation values can be seen as part of the broader problem of learning probabilistic models from data, and hence part of machine learning in computer science. For a small sampling of the interplay between ideas in statistical physics and machine learning, see Refs \cite{mezard+montanari_09,tishby+al_90,yedidia+al_03,mezard_03,mehta+schwab_14}.
Interest in maximum entropy approaches to biological networks was stimulated by results from an analysis of activity patterns in small groups of neurons from the vertebrate retina \cite{schneidman+al_06}. As the retina responds to natural movies, pairwise correlations are weak, of mixed signs, and widespread. Models that ignore these weak correlations make predictions about the frequency of patterns in groups of ten neurons that can be wrong by many orders of magnitude, but these errors are largely cured in maximum entropy models that match the pairwise correlations. Even in these small networks, there are hints that what one sees in the network are genuinely collective states. By now, these ideas have been extended and applied to several different neural systems \cite{tkacik+al_06,shlens+al_06,tang+al_08,tkacik+al_09,shlens+al_09,ohiorhenuan+al_10,ganmor+al_11,simplest,sdme,tkacik+al_14a,mora+al_14}. Most recently it has been possible to build models for the joint of activity of more than 100 neurons in the retina, and these models are so accurate that they can predict distribution of the effective energy deep into the tail of patterns that occur only once in a two hour long experiment \cite{tkacik+al_14a}. The states that dominate the distribution are indeed collective, so one can reconstruct the time dependent response of single cells from the behavior of the population, without reference to the visual stimulus. Most provocatively, that distribution that we reconstruct from the measured expectation values exhibits many of the signature of an equilibrium statistical mechanics problem near its critical point \cite{tkacik+al_14b}.
An independent stream of work has used maximum entropy ideas to think about the ensemble of amino acid sequences that form a family of proteins, folding into essentially the same three--dimensional structures. This work was inspired by experiments that actively explored sequence space, trying to respect pairwise correlations among amino acids at different sites \cite{socolich+al_05,russ+al_05}, using an algorithm that is equivalent in some limits to maximum entropy \cite{bialek+ranganathan_07}. Although these correlations extend over long distances, the effective interactions among amino acids in these models were found to be spatially local; this raises the possibility that we can infer the spatial neighbor relations among amino acids from data on sequence variation, in effect folding the protein via sequence comparisons \cite{weigt_al_09,marks+al_11,hopf+al_12,sulkowska+al_12}. This new approach to the protein folding problem has generated considerable excitement, but there are many open questions about the sufficiency of the available data and the approximations that have been used to solve the inverse problem. Note that this effort to describe the distribution of sequences consistent with a given protein structure is the inverse of the usual protein folding problem, where we are given the sequence and asked to predict the structure, and there is a big conceptual question about the relation between these two problems.
In the immune system we can look at special cases of sequences ensembles, the distribution of antibody molecules that organisms use to combat infection and the distribution of viral sequences in patients that have long term infections such at HIV. In the problem of antibody diversity, maximum entropy models built from pairwise correlations provide a surprisingly accurate description of the distribution as a whole \cite{mora+al_10}, while for HIV the effective energy in maximum entropy models predicts viral fitness and vulnerability to immune attack \cite{ferguson+al_13}. This is a small window into a much larger literature on statistical physics approaches to the immune system.\footnote{See, for example, the program at the Kavli Institute for Theoretical Physics, http://immuno-m12.wikispaces.com.}
Birds in a flock are thought to follow only their near neighbors, yet the entire flock can decide to move in a single direction and at nearly uniform speed. Order spreads through the flock much as through spins in a magnet, and this is more than a metaphor: maximum entropy models consistent with measured local correlations in a flock provide successful quantitative, parameter--free predictions of correlations in flight direction and speed throughout the flock \cite{bialek+al_12,bialek+al_14}. It is also possible to do this analysis in more detail, building maximum entropy models that match the correlations between a bird and its nearest neighbor, its second neighbor, and so on; one finds that the resulting couplings which link each bird to its $k^{\rm th}$ neighbor decline exponentially with $k$, providing some of the most direct support yet for the hypothesis that interactions are local \cite{cavagna+al_14a}. Real starling flocks exhibit long--ranged correlations of the fluctuations in flight direction and speed \cite{cavagna+al_10}; analysis of the maximum entropy models shows that the directional correlations arise from Goldstone modes associated with the spontaneous polarization of the flock \cite{bialek+al_12}, while the speed correlations are long--ranged only because the parameters of the model are tuned close to a critical point \cite{bialek+al_14}; signatures of criticality have also been found in the behavior of insect swarms \cite{attanasi+al_14}.
\subsection{Information-theoretic parameter inference}
\label{parameters}
A standard tool for fitting parametric models to data is maximum likelihood inference. This procedure, however, depends on our ability to write down the likelihood function: the probability of making specific experimental observations given the parameters of the model. In studying biological systems we often are faced with situations where this is impossible, for two very distinct reasons. First, while we might have a good idea for how a part of the system works (e.g., that transcription factors recognize and bind DNA sequences in a way that is governed by equilibrium statistical mechanics), what we can measure is far ``downstream'' of these events (the activation level of a gene), and we may not have a quantitative model for the intervening steps. Second, and less fundamentally, many experiments still are not as calibrated as one might hope, so we don't have a model for the noise in the measurement itself.
Without the ability to write down the likelihood function, is there any way to do unbiased parametric inference of the process of interest? There is isn't much we can do about the general version of this problem, but in many biological systems one crucial step in the process that we are trying to describe involves a substantial reduction of dimensionality. Thus, in the visual system, we expect that neurons respond not to every possible detail of the image or movie falling onto the retina, but only to some limited set of features; in the simplest case these features might be described simply as linear projections onto a template or ``receptive field.'' If this picture is right, then the transformation from input movie $x$ to output spikes $z$ can be written as $x\rightarrow y \rightarrow z$, where $x$ is a vector in a space of very large dimensionality (the space of movies), while both $y$ and $z$ are low dimensional, in the simplest case just scalars. The mapping $y\rightarrow z$ then is easy to characterize; the problem is how to choose the projection that describes $x\rightarrow y$. The data processing inequality tells us that the best possible predictions of $z$ will be generated by a model in which the mapping $x\rightarrow y$ captures as much information as possible about $z$. Importantly, the difficulty of searching for such ``maximally informative projections'' does not vary much with the structure of the distribution of the inputs $x$, which means that we can use this approach to characterize the encoding of complex, naturalistic input signals \cite{sharpee+al_04}. Applications and extensions of this idea have been explored in several different systems \cite{sharpee+al_06,fitzgerald+al_11,rajan+bialek_12,eickenberg+al_12,rajan+al_13}.
The idea of searching for maximally informative projections has also been used to analyze experiments on the interactions of transcription factors with their target sites along DNA; here the projection is from the local DNA sequence to the binding energy of the TF \cite{kinney+al_07}. An important success of this initial effort was to show that two independent experiments which characterized the same transcription factor in fact led to consistent models of TF--DNA binding, despite earlier claims to the contrary. This work also demonstrated the formal equivalence between information-theoretic parameter inference and ``error-model-averaged'' maximum likelihood inference. Subsequent experiments exploited these idea to give a much more detailed view of how a transcription factor and the RNA polymerase bind and interact at the lac operon \cite{kinney+al_10}, and related methods were developed to allow analysis of a wider range of input data \cite{elemento+al_07}.
Because the issues of model inference for processes embedded into unknown larger systems or subject to unknown noise processes are prevalent in biology, we expect to see further applications for information theoretic inference in biology. We also note that these applications can be seen as specific examples of a more general framework that relates learning (including parameter inference) to information flow between the data and its compressed representations. There are many interesting links between the learning problem, statistical mechanics, and field theory, which are beyond the scope of this review.
\section{Optimizing information flow}
\label{sec5}
Information is essential for life, but bits are not free. Is it then possible that organisms are driven to extract and represent the maximum possible information given the physical constraints? This is an idea that arose almost immediately after Shannon's original work, and has been a productive source for thinking about many specific biological systems, across many scales.
\subsection{The genetic code}
Five years after Shannon's papers, Watson and Crick proposed the double helical structure for DNA, and after another five years Crick could articulate ``the sequence hypothesis'': the sequence of amino acids in proteins is determined by the sequence of bases along the DNA \cite{crick_58}. Once one has this idea, the mapping between two major classes of biopolymers becomes a problem of coding, and it seemed natural to bring the new ideas of information theory to bear on this central problem of life \cite{crick_63}. In particular, it was hoped that the code might be efficient in the sense defined by Shannon, using as few bases as possible to encode the amino acid sequence of a protein. As we now know, the genetic code maps triplets of bases (codons) into one of twenty amino acids (plus ``stop''), so that there is substantial redundancy.
The failure of the real genetic code to instantiate any of the early ideas about efficiency or optimality must have been a disappointment, and it is hard to find much serious interaction between information theory and molecular biology for some time after the code was worked out. This split may have been premature. In a triplet code, there are three possible ``reading frames,'' and the first organism to have its entire genome sequenced---the bacteriophage $\Phi$X174---makes use of the same DNA in multiple reading frames \cite{sanger+al_77}, an unexpected multiplexing that enhances efficiency in the original Shannon sense. Another observation is that the redundancy of the code is structured, so that random changes of a single base along the DNA are likely to map one amino acid either into itself (`silent' mutations) or into amino acids with very similar chemical properties. This would not be true in a random assignment of codons to amino acids, and it has been suggested that the real code may even be nearly optimal in this respect \cite{freeland+hurst_98,freeland+al_00}; if correct this means that the genetic code is not an optimal noiseless compression, as in the earliest (incorrect) proposals, but may be near the bounds set by rate--distortion theory \cite{tlusty_08}. A full development of this idea will require understanding the cost of amino acid substitutions, which relates back to the analysis of sequence ensembles in Section \ref{maxent}.
The genetic code really is a strategy for communicating from DNA bases to amino acids, and here too there are errors. Although one often pictures the specificity of life's mechanisms as deriving from the specific binding between complementary molecular components (as in the pairing of A with T and C with G in DNA), a beautiful chapter in the physics of life is the appreciation that these equilibrium interactions are not sufficiently selective, and hence that cells must build Maxwell demons to sort molecules with higher precision; the broad class of mechanisms that implement these demons is called kinetic proofreading \cite{hopfield_74,ninio_75} [see also Section 4.5 of Ref \cite{bialek_12}]. This creates an interesting tradeoff for the cell, which can operate at high fidelity but high energetic cost (paying for the demon), or trade the cost of errors against the reduced energetic cost of correcting them \cite{ehrenberg+kurland_84}. Although the costs and benefits of information flow in reading the genetic code have been much discussed, this still has not been given a fully information theoretic formulation.
\subsection{Efficient representations in sensory processing}
One of the earliest uses of information theory to think about biological systems was an attempt to estimate the amount of information carried by neuronal action potentials \cite{mackay+mcculloch_52,spikes}. The answer depends crucially on the time resolution with which such spikes can be generated and detected by neurons. If the average rate of action potentials is $\bar r$, and the system can work with a time resolution $\Delta\tau \ll 1/\bar r$, then the stream of spikes becomes a binary sequence of $1/\Delta\tau$ symbols per second, and the probability of a 1 (spike) is ${\bar r}\Delta\tau$. The maximum entropy of this sequence, which sets the capacity for the neuron to carry information, is then $S\sim \log_2(e/\bar r \Delta\tau)$ bits/spike. In contrast, if the system is limited to counting spikes in windows much larger than the typical interspike interval, then the entropy becomes $S\sim \log_2 (e\bar r\Delta\tau)/(\bar r\Delta\tau)$ bits/spike. Evidently higher time resolution allows for higher information rates, and dramatically so: if spike rates are a few per second (as often true, on average), then time resolutions of a few milliseconds could allow for up to ten bits per spike. In 1952, when MacKay and McCulloch first made such estimates, locating real neural codes on the continuum from timing to counting was a matter of speculation. But this work established an agenda: could we measure the information that spike trains carry, for example about sensory inputs, and does this information approach the limits set by the entropy of the spike sequences themselves?
MacKay and McCulloch were thinking about a single neuron. But signals often are conveyed by larger populations of neurons, for example in the retina where many neurons respond to overlapping regions of the visual world. As early as 1959, Barlow suggested that processing in the retina would have the function of removing redundancy among the outputs of neighboring ganglion cells, and argued that several major features of these cells' responses could be understood in this way: a reduction in the response to sustained stimuli (termed adaptation) serves to reduce redundancy in time, while the tendency of cells to respond to differences between light intensity in a small patch and the surrounding region (``center--surround'' organization) serves to reduce redundancy in space \cite{barlow_59,barlow_61}.
It would take many years until these ideas about efficient coding in sensory processing would be made quantitative. An important qualitative idea, however, is that the efficiency of a code depends on the distribution of inputs. Thus, we might expect that real organisms use codes that are matched to the distribution of signals that they encounter in their natural environment; put simply, the function of the code would make sense only in context, and the relevant context is quite complex. This realization stands in contrast to the long tradition of using highly simplified stimuli in laboratory studies of sensory function, and has led to the emergence of a subfield of neuroscience concerned specifically with the structure of natural sensory signals and the neural response to these signals \cite{simoncelli+olshausen_01,geisler_08}.
The first stages of retinal coding occur with neurons that produce analog voltages rather than discrete spikes. Laughlin focused on cells in the first stage of processing after the photoreceptors in the fly retina, and asked if the nonlinear input/output relation that transforms light intensity into voltage is matched to the environmental distribution of light intensities \cite{laughlin_81}. If the voltage noise is small and constant throughout the dynamic range, then optimal information transmission implies that a (normalized) input/output relation should be equal to the cumulative probability distribution of the light intensity. By sampling natural scenes with a custom-built photodetector to predict the optimal input/output relation and comparing it with the measured one, Laughlin found a remarkable agreement with the theory, especially considering that there are no free parameters. Although there are obvious open questions, this was a really beautiful result that inspired the community to take these ideas more seriously.
For spiking neurons, measurements of information transmission in response to complex, dynamic stimuli began to appear in the 1990s. Initial measurements were based on decoding the spike trains to recover estimates of the input sensory signals \cite{bialek+al_91,rieke+al_93,borst+theunissen_99}, and then analysis tools for more direct estimates were developed \cite{strong+al_98,ruyter+al_97}. In the primary sensory neurons of the frog inner ear, the mechanical sensors of the cockroach, ganglion cells in the vertebrate retina \cite{koch+al_06} and subsequent layers of visual processing in mammals \cite{reinagel+reid_00,liu+al_01}, and motion--sensitive neurons deeper in fly visual system, as well as other systems, it was found that spike trains could transmit more than one bit per spike, and typically reached 30--50\% of the limit set by the spike train entropy, even with time resolutions in the millisecond range. In accord with the idea of matching to natural signals, information rates and coding efficiencies are larger with distributions of stimuli that capture features of the natural ensemble \cite{rieke+al_95}, and in some cases truly natural stimuli reveal codes that maintain their efficiency down to sub--millisecond resolution \cite{nemenman+al_08}.\footnote{These results depend on getting control over the problems of estimating entropy and information from limited samples, as discussed in Section \ref{measure_info}.} Despite much prejudice to the contrary, there are demonstrations of efficient coding, with time resolutions of a few milliseconds, by single neurons in the cerebral cortex of our primate cousins \cite{buracas+al_98,kara+al_00}, suggesting strongly that these principles extend to brains much like our own.
If neurons implement coding strategies that are matched to the distribution of sensory inputs, what happens when this distribution changes? When we walk from a dark room out into the bright sunlight, the mean light intensity changes, and it takes a moment for our visual system to adapt to this new level. This adaptation is visible in the response of individual neurons in the retina, and this is one of the classical observations about neural coding, having roots in Adrian's first experiments in the 1920s \cite{adrian_28}. But the efficient coding hypothesis predicts that the code should adapt not just to the mean, but to the whole distribution. In particular, natural images have an intermittent structure, so that the variance in light intensity varies significantly from point to point in a scene, and hence the variance seen by neurons responding to a small patch of the world will vary in time, so adaptation to the variance would enhance efficiency \cite{ruderman+bialek_94}, and similar arguments could be made about other sensory modalities. Early experiments in the vertebrate retina demonstrated this adaptation to variance explicitly \cite{smirnakis+al_97}. Subsequent experiments have shown that retinal adaptation is an even richer phenomenon, so that properly chosen distributions of inputs can drastically change even very basic features of the code exhibited by single neurons, and this is part of a growing realization that the retina is capable of much more complex computations than originally believed \cite{gollisch+meister_10}.
Adaptation to the variance of input signals has now been seen in many systems \cite{wark+al_07}: motion--sensitive neurons of the fly visual system \cite{brenner+al_00a,fairhall+al_01}, auditory neurons at several stages of processing in mammals and birds \cite{kvale+schreiner_04,dean+al_05,nagel+doupe_06,wen+al_09,dahmen+al_10,rabinowitz+al_11}, whisker motion sensing neurons in rodents \cite{maravall+al_07}, and in the farthest reaches of primate visual cortex, where cells responsive to the shapes of objects adapt to the variance along different parameters describing these shapes \cite{debaene+al_07}. In many of these systems, changing the variance of the input distribution simply produces a rescaling of the input/output relation, as first observed in the fly motion--sensitive cells \cite{brenner+al_00a}. If the typical signals are large compared to internal noise levels but not so large as to saturate the system, then the dynamic range of the inputs themselves provides the only scale in the problem, and hence optimization must lead to the observed scaling behavior; in the fly it was even possible to check that the scaling factor chosen by the adaptation mechanism served to optimize the transmitted information \cite{brenner+al_00a}. When the input distribution shifts suddenly, the input/output relation adjusts with remarkable speed, perhaps so rapidly that there is no sharp distinction between adaptation and the nonlinear transient response. The time scales involved are essentially the shortest ones allowed by the need to gather statistics and verify that the distribution has, in fact, changed \cite{fairhall+al_01,wark+al_09}.
Barlow's original idea about redundancy reduction in the retina was implemented mathematically by Atick and Redlich \cite{atick+redlich_90}, who approximated retinal ganglion cells as analog devices that respond to a linearly filtered version of their visual inputs. They searched for optimal filters, and showed that in low background lights ganglion cells should integrate over small regions of space and time, while at high light levels the response should be differentiating; this corresponds to a crossover between noise reduction and redundancy reduction being the dominant factors in enhancing information transmission, and agrees with the data.
If we imagine that neurons filter their inputs to remove redundancy, then the outputs should become close to uncorrelated or ``white.'' But many natural input signals have power spectra in time that are approximately $1/\omega$, and whitening this spectrum requires a transfer function between input and output $\sim \sqrt \omega$, which is not so easy to realize. Nonetheless, the linear transfer function from the photoreceptor cells to the second order neurons studied by Laughlin has precisely this form \cite{laughlin+ruyter_96}. The transfer of the signal from one cell to the next is mediated, here as elsewhere in the brain, by the neurotransmitter molecules that are packaged into discrete vesicles. Careful measurements of the signal and noise characteristics in both cells leads to very high estimates of the information capacity of this connection (synapse), more than one thousand bits per second \cite{ruyter+laughlin_96}, and this is close to the physical limit set by the entropy of counting the vesicles \cite{spikes,bialek_12}.
In the actual retina, of course, not all cells are identical, and is interesting to ask if we can predict the distribution over cell types from optimization principles. There is an old discussion about the choice of three color sensitivities for the photoreceptors in the bee's eye, arguing that the particular combination that is found in a whole family of related species serves to provide maximal information about the identity of the flowers from which bees collect nectar \cite{chittka+menzel_92}. In the primate retina, more rigorous arguments have been used to predict the distribution of the three cone types, which is both highly non--uniform and varies with position on the retina \cite{garrigan+al_10}. Retinal processing involves a plethora of cell types, but all visual systems divide signals into `on' and `off' pathways, which respond with more action potentials when the light level rises above or below its local average value; Ratliff et al \cite{ratliff+al_10} have argued that the preponderance of off cells is matched to an asymmetry in the distribution of light intensities found in natural scenes, and that other differences between on an off pathways also serve to optimize information transmission.
If we focus on a single class of retinal ganglion cells, we find that they form a lattice across the surface of the retina [see, for example, Ref \cite{shlens+al_06}]. The geometrical properties of this lattice, specifically, the size over which each cell integrates the signal relative to the lattice spacing, can be predicted from information optimization \cite{borghuis+al_08}. These lattices are, however, not perfectly regular due to discreteness and imperfections in the spatial arrangement of the photoreceptors, and in principle these irregularities should lead to a loss of information. Remarkably, these variations turn to be correlated with the local variations in the connectivity from the photoreceptor array to the ganglion cell layer in such a way as to almost perfectly compensate for the information loss \cite{liu+al_09}.
Attempts to analyze the optimization problem for spiking neurons were quite scattered until the work of Smith and Lewicki on the coding of complex sounds \cite{smith+lewicki_05,smith+lewicki_06}. Rather than trying to build models of the process by which continuous inputs are converted into spikes, and then optimizing the parameters of these models, they chose the spike times themselves as the variables describing the encoding process and optimized the code assuming that the goal was to reconstruct the inputs using a simple linear algorithm, following \cite{bialek+al_91}. The result of the calculation is then a set of spike trains for a population of neurons responding to acoustic inputs, and these spike trains could be analyzed in the same way that experimentalists routinely analyze data from real neurons. The results were impressive: optimal spike encoders for speech--like sounds, but not for other sound ensembles, almost perfectly match the characteristics of real auditory neurons. Further, the rate--distortion functions for these neural codes reach or even surpass the performance of the best artificial systems.
The problem of optimal coding has also been studied from a theoretical angle. In a population of rate neurons that suffer from negligible intrinsic noise, the optimal code decorrelates the inputs: this should happen not only at pairwise order (by whitening), as suggested by early works of Barlow, van Hateren and others \cite{barlow_59,vanHateren_92}, but at all orders of correlation; in other words, the optimal code will minimize the mutual information between different neurons at the output. This led to the idea of `independent component analysis' (ICA) and sparse codes. Imagining that neurons in the visual pathway perform linear filtering on the incoming stimuli and optimizing the filters using ICA over natural scenes, two groups recovered the well-known Gabor-like filters found in the primary visual cortex \cite{olshausen+field_96,bell+sejnowski_97}.
In the presence of nonlinearities and noise, optimizing information is much more difficult; early work in this direction was done by Linsker \cite{linsker_88}. If neurons are noisy, perfect decorrelation is no longer optimal. In this case, some amount of redundancy (in engineered codes in the form of error-correcting bits) can be beneficial \cite{barlow_01}. It is interesting to note that real neuronal spike trains are usually weakly correlated at the pairwise level and redundant \cite{Puchalla}, and at the level of a large population this redundancy can suffice to predict well the behavior of one neuron from the state of the rest of the network \cite{tkacik+al_14}. If noise is small, information optimization can be carried out in a continuous nonlinear system, and the tradeoffs between decorrelation and redundancy reduction can be explored in detail; in this framework, Karklin and Simoncelli managed to derive many properties of the retina simultaneously from the statistical structure of natural scenes \cite{karklin+simoncelli_11}. This optimization, as well as previous work \cite{balasubramanian+al_01}, also incorporated an important constraint on metabolic efficiency (cost of spiking). By departing from analytically tractable cases, numerical exploration has been used to study optimal coding at arbitrary noise levels with high-dimensional stimuli \cite{tkacik+al_10}. Using a model for an interacting neural population that can be linked to maximum maximum entropy models with an Ising form (Section \ref{maxent}), small networks experienced a transition between redundant coding and decorrelation based on the statistical structure of the inputs and the level of noise in single units. Surprisingly, when noise in single units is high enough, the redundant strategy with information-maximizing connectivity yields a network with strong attractor states to which different stimuli map uniquely, recovering Hopfield-like associative memory \cite{hopfield_82} from an optimization principle. Mapping out a complete `phase diagram' of optimal coding strategies for high-dimensional noisy nonlinear input/output maps remains a challenge.
Finally, we note that many popular models for coding by large populations of neurons are inefficient in the extreme. It is widely assumed that neurons in a single region of the visual cortex, for example, are selective for a very limited number of features (e.g., the velocity of motion, or the orientation of an edge). If all the neurons in the population respond independently to these features, with each of the $N$ cells preferring a slightly different value, and we assume that what matters is the number of spikes each neuron generates in a relatively large window of time (``rate coding''), then the response of the population as a whole will allow us to determine the value of the feature with a precision $\sim 1/\sqrt{N}$. But this means that the information conveyed by the population is $\sim \log(N)$, and thus the information per neuron vanishes as $N$ becomes large. Are we missing something? One possibility is that the diversity of dynamical responses among neurons with redundant static feature selectivity can be sufficient to create a combinatorial code for stimulus timing, allowing the information to grow linearly with the number of neurons over a much larger dynamic range \cite{osborne+al_08}. A more radical alternative is suggested by the discovery of ``grid cells,'' neurons that encode an animal's position relative to a lattice \cite{hafting+al_05}. In this system, it is possible for a population of neurons with different lattice spacing to generate a truly efficient code, with more neurons allowing exponentially more accuracy in the representation of position \cite{fiete+al_08,sreenivasan+fiete_11}. These ideas point to the need for direct measurement of the information carried by large populations of cells, under reasonably natural conditions.
\subsection{Biochemical and genetic networks}
Even the simplest of bacteria have genes that code for many hundreds of proteins, and our own genome codes for more than ten thousand proteins. No single cell needs to make all these proteins at once, and there are many layers of regulation to control the readout or expression of individual genes. One important class of these mechanisms is the binding of protein molecules, called transcription factors, to specific sequences along the DNA, close to the start of the coding region of a gene. Transcription factor binding can both repress or activate protein synthesis; in bacteria there are simple geometrical pictures for how this can happen that are probably good approximations to reality, but the situation in higher organisms is less clear. In all cases, transcription factors act at concentrations in the nanoMolar range, and $1\,{\rm nM} \sim 1\,{\rm molecule/}\mu{\rm m}^3$. But a bacterium has a volume of only a few cubic microns, which means that the regulation of transcription is controlled by just handfuls of molecules.
The small numbers of molecules involved means that the regulation of gene expression is, inevitably, noisy, and this noise was measured directly in the early 2000s \cite{elowitz+al_02,ozbudak+al_02,blake+al_03}. How does this noise limit the transmission of information through the system, from the input concentration of transcription factors to the output level of the protein encoded by the target gene? Here information provides a measure of regulatory power, counting the number of reliably distinguishable output states that can be accessed by adjusting the inputs \cite{tkacik+al_08b,tkacik+walczak_11}. We recall that if the input/output relation and noise in the system are fixed, then we can optimize information transmission by adjusting the distribution of input signals. This matching condition, which is essentially the same as in Laughlin's discussion of the fly retina, allows the system to get as much regulatory power as possible out of a limited number of molecules. An important difference from Laughlin's discussion is that noise in the case of transcriptional regulation must be (and is measured to be) strongly dependent on the levels of input and output.
Early events in development of the fruit fly embryo provide a productive testing ground for ideas about the regulation of gene expression. During the construction of the egg, the mother places the mRNA for a transcription factor at the end that will become the head; the combination of synthesis, diffusion and degradation of the protein (Bicoid) translated from this mRNA results in a spatially varying concentration along the length of embryo---a gradient in the concentration of a primary morphogen, to use the conventional language. Bicoid is the input to an interacting network of genes (called ``gap genes'' and ``pair-rule genes'') that end up being expressed in distinct spatial patterns, which will define the final body plan of the larval fly. Importantly, these patterns are visible within a few hours after the egg is laid, before cells start to move and even before there are membranes marking the boundaries between the cells.
A first generation of quantitative experiments on this system mapped the relationship between the input transcription factor concentration and the output level of one of the gap genes, making measurements nucleus by nucleus in single embryos, and demonstrating that the results are consistent across embryos \cite{gregor+al_07a,gregor+al_07b}. If we think of $x$ as the input concentration and $y$ as the output concentration, then these experiments allow us to estimate the behavior of the communication channel, $P(y|x)$. We can solve numerically for the distribution of inputs that optimizes information transmission, and compare with experiment. Because the mean input/output relation has broad, flat segments, many different input distributions give essentially the same mutual information between $x$ and $y$, but all these near--optimal input distributions have the same output distribution, so we make the comparison of the predicted $P_{\rm out}(y)$ with experiment \cite{tkacik+al_08a}. The agreement, which involves no adjustable parameters, is excellent. Importantly, we can also estimate directly the mutual information $I(x;y)$ from the data, and find that it is $0.88\pm 0.09$ of the optimum, so that the system really is transmitting nearly the maximal information given the measured noise levels. This information is about a bit and a half, showing that the regulation of gene expression in this case is more than just an on/off switch.
Thinking of transcriptional regulation in the early embryo as having one input and one output is a drastic oversimplification, but dealing with multiple inputs and multiple outputs is very difficult. It is thus useful to note that the information of relevance for development is information about the position of the cell. If we can measure the expression levels of multiple genes simultaneously, we can ask how much information these level provide about position even if we can't also measure all the input signals. Indeed, the qualitative concept of positional information is very old \cite{wolpert_69}, and it is attractive to try and make this quantitative \cite{dubuis+al_13b,tkacik+al_14c}. Refinement of classical staining methods make it possible to measure, simultaneously and with high accuracy, the normalized concentrations of four gap genes that provide dominant signals along the central 80\% of the embryo's length \cite{dubuis+al_13a}. Analysis of these experiments shows that is is possible to ``decode'' the expression levels and estimate the position of a nucleus to an accuracy of better than 1\% of the length of egg, corresponding to a total of $4.14\pm 0.05$ bits of information about position \cite{dubuis+al_13b}. This precision is comparable to the spacing between neighboring cells, and to the precision of subsequent developmental events, suggesting that the embryo really does use this information.
Each of the four gap genes has a complicated profile of concentration vs. position along the embryo, and the noise levels have an even more complicated structure. But since cells are almost uniformly distributed, matching of the distribution on inputs to the noise characteristics of the system would require that the error in estimating position be almost uniform as well. This is, in fact, what we see in the data \cite{dubuis+al_13b}; indeed, if we imagine trying to redistribute the cells so as to transmit more information, we could gain less than 2\%. This seems a striking signature of optimization.
Optimization of information flow should involve more than matching: we should be able to find the architecture and parameters of genetic networks that allow for maximal information transmission. One ambitious idea is to define a dynamics in the space of networks that is a schematic of the evolutionary process, and then use information as a surrogate for fitness \cite{francois+siggia_10}. The networks that emerge from such simulations have many qualitative features in common with the networks functioning in real embryos, including the fruit fly. One can find examples of multiple genes driven by a common maternal morphogen, each one activating itself and repressing other genes in the network, much as for the gap genes. A conceptually striking feature of this work is that it explicitly constructs paths through the space of possible networks that lead from simpler to more complex body plans, through gene duplication followed by mutation. Importantly, almost all the steps along the path are either neutral or increase the fitness. As the authors emphasize, this ``... provides a quantitative demonstration that continuous positive selection can generate complex phenotypes from simple components by incremental evolution, as Darwin proposed.''
Rather than reconstructing entire evolutionary trajectories for networks, we can try to find locally optimal networks within some broad class of models \cite{genesI,genesII,genesIII}. Here again we can see features that remind us of real networks, but there are surprises. When the available concentration of transcription factors is severely limited, the optimal strategy for transmitting information is for a single transcription factor to broadcast to multiple, completely redundant targets, essentially using the redundancy to make multiple measurements of a weak signal. Observations of redundancy in genetic networks usually is taken as prima facie evidence against optimization and in favor of an interpretation in which network architecture is just an artifact of evolutionary history, but here we have a quantitative demonstration that such qualitative arguments can be wrong. More subtly, these calculations show that even qualitative features of the networks which maximize information transmission depend on some molecular details of the regulatory mechanisms. This is slightly worrisome, but also suggests a path for some of the intricacy of complex molecular interactions to influence macroscopic behavior. A related approach to these same problems is inspired not by embryos but by artificial genetic circuits that have more limited topologies and inputs, allowing for an exhaustive exploration \cite{ziv+al_07}. In these cases, the optimization of information transmission leads to very broad maxima in the space of network parameters, so that network topology is the dominant consideration. These efforts at full optimization have been limited to static inputs and outputs; information transmission in dynamic contexts raises new questions, which are just starting to be explored \cite{tostevin+wolde_09,ronde+al_10}. Finally, interesting questions that have been highlighted recently include the link between information transmission and cellular decision making \cite{bowsher+swain_14}, and using information theory to put strict bounds on the reliability of biochemical processes even when many of these processes are unknown or unobserved \cite{lestas+al_10}.
\subsection{More ambitious goals}
Success at gambling, and perhaps at life, plausibly depends on how much information we have. If we know more about the outcome of a coin flip (or a horse race, or the stock market, or ... ), then we should be able to win our bets more often, and hence make more money. But it is not obvious that the knowledge measured by Shannon's bits quantifies the useful knowledge in this context. Kelly analyzed a simple gambling game and showed that the maximum rate at which winnings can grow is precisely the information that bettors have about the outcome, and his proof is constructive, so we actually know how to achieve this maximum \cite{kelly_56}. We can think of the choices that organisms make---e.g., to express different levels of proteins---as being bets on the state of the environment, and there have been several efforts to construct scenarios that match the structure of Kelly's argument, so that the growth rate of the organisms becomes equal to the information that they have about environmental conditions \cite{bergstrom+lachmann_05,kussell+leibler_05,rivoire+leibler_11}. More generally, from rate distortion theory we know that achieving a given average growth rate across a wide range of conditions will require some minimum amount of information about the environment \cite{taylor+al_07}. These connections between growth rates and information provide a path for evolution to select for organisms that gather and represent more information about their environment.
Another interesting connection of information theoretic optimization to biological function is in the problem of search \cite{vergassola+al_07}. Organisms search for food, for shelter, for mates, and more; it may even be possible to think more generally about computations or decisions as the search for solutions. If we imagine that we are looking for something in one of $N$ possible places, we can order these by our current estimate of the probability $P_n$ that we will find what we are looking for in the $n^{\rm th}$ place, so that $P_1 > P_2 > \cdots > P_n$. Then if we look first in the most likely place, and then down the list, the mean number of tries before we find what we are looking for is $\bar n = \sum_{n=1}^N n P_n$. Given $\bar n$, there is a probability distribution ($P_n \propto e^{-\lambda n}$) that has maximal entropy. Conversely, if our uncertainty corresponds to a probability distribution with entropy $S$, then the mean search time $\bar n$ has a minimum, independent of all other details. The only way we can speed up our search is to reduce the entropy of the distribution, or to gain information in the Shannon sense; while there is no guarantee that our search will go as fast as allowed by this information, it cannot go faster.
The idea of ``infotaxis'' is to replace the task of finding something with the task of gaining information about its location \cite{vergassola+al_07}. In the context of a flying insect searching for the source of an odor, where new information arrives only intermittently when the flight path crosses odor plumes, optimizing the gathering of information by properly alternating upwind flight with cross--wind `casting' in fact leads to the source, and the flight paths that are generated by this algorithm are reminiscent of the paths taken by insects in natural environments.\footnote{As an aside, we note that insects seem to ``know'' about the intermittent structure of turbulent flows \cite{mafra-neto+carde_94}. In particular, moths exposed to smooth ribbon plumes, which provide a continuous signal and should in principle be easy to follow, fly along strongly zig--zag trajectories, spending most their time casting crosswind, while the same moths exposed to intermittent or turbulent plumes fly efficiently toward the source of the odor. This is one of the strongest, qualitative results demonstrating that neural processing is tuned to the statistical structure of physical signals that arise in the natural environment.} For the purposes of this review, perhaps the most important result from the infotaxis idea is the demonstration that a biologically grounded, goal--directed behavior can be replaced with an abstract, information theoretic optimization principle. This certainly encourages us to think that some of the nitty--gritty of biology should be derivable from more general principles.
In their earliest work, Attneave and Barlow expressed the hope that the search for efficient representations of sensory data would drive discovery of features and objects that form our perceptual vocabulary, and Attneave even tried simple experiments to show that human subjects use efficient strategies in abstracting simple figures into sketches \cite{barlow_59,barlow_61,attneave_54}. As an example, if our world consists of a finite set of rigid objects, each seen from many different perspectives, then the most efficient representation of the image data reaching our eyes would consist of a discrete label for object identity and a few parameters to describe the viewing geometry. In this case, optimization in an information theoretic sense would require our brains to recognize objects, and the labeling of these objects would constitute a `naming' of the meaningful features of the visual world, without any additional instructions. At its most ambitious, the claim would be that these features acquire meaning because they are maximally informative, as opposed to having meaning imposed upon them.
Shannon frequently used English as an example of information theoretic ideas, and even played a ``guessing game'' with human subjects to estimate the entropy of written texts \cite{shannon_51}. On the other hand, there is a persistent suspicion that there must be more than just statistics involved in the structure and meaning of language, and one of the foundational papers of modern linguistics is essentially an attack on Shannon's probabilistic models for language \cite{chomsky_56}. Some of these early arguments are related to the crude Markovian models that Shannon used as illustrations, and some are related to the confusion between learning a probabilistic model and simply counting events,\footnote{Thus, Chomsky famously argued that statistical approaches could not distinguish between sentences that have never occurred but are instantly recognizable as grammatical (Colorless green ideas sleep furiously) and sentences that never occurred because they are forbidden by grammatical rules (Furiously sleep ideas green colorless) \cite{chomsky_56}. As a concrete rejoinder, Pereira constructed a simple Markov model in which successive words belong to clusters, so that $P(W_{\rm n+1}, W_{\rm n}) = \sum_C Q(W_{\rm n+1}|C) Q(W_{\rm n}|C) P(C)$, and learned the model on a modest body of newspaper texts using just sixteen clusters. He found that Chomsky's two example sentences have probabilities that differ by a factor of more than $10^5$ \cite{pereira_00}. But between the dates of these references, a much more sophisticated view of the learning problem had emerged.} but questions remain, and Shannon himself eschewed any connection to the problem of meaning. In what proved to be a more direct attack on this question, Pereira et al \cite{pereira+al_93} developed an algorithm to cluster nouns based on the distribution of verbs that occur in the same sentence; their approach can be thought of as compressing the description of the noun while preserving information about the identity of the verb, and hence is yet another example of efficient representation \cite{tishby+al_99}. The resulting clusters exhibit a striking degree of semantic (rather than just syntactic) coherence, and it hard to resist saying that this purely information theoretic approach is extracting meaning. These references are just a small sampling from the roots of a huge literature. By now, we all interact with probabilistic models of language every day (through internet search engines, automatic translation, ...), so perhaps the larger conceptual question of whether the search for efficient representation captures meaning has become less urgent. Nonetheless, we suspect that there are many open questions here, including whether we really understand how to write down models that capture the very long--ranged correlations that are present in texts.
The entropy of a probability distribution is defined without reference to anything else, at least in the case of discrete variables, but information is always about something. Thus it makes no sense to ask about the information content of a signal; one has to ask how much information the signal provides about some quantity of interest.\footnote{We could be interested in the value of the signal itself, in which case the information is equal to the entropy. But this is seldom the case. If the signal is the text of a newspaper, for example, we probably are not interested in the text itself, but what it tells us about events in the world. In particular, the entropy of the text is increased if we introduce random spelling errors, but the information about events is decreased.} Arguably the most successful uses of information theory in thinking about biological systems are in those cases where the question ``information about what?'' has a clear answer. But this makes us worry that more of the predictive power of our theories resides in this qualitative knowledge of the biologically relevant information than in the more mathematical principle of optimizing information transmission. Can we do something more general than just making a list of which information is relevant for each particular system?
Information cannot be useful to an organism unless it carries predictive power: if we use information to guide our actions, it takes time to act, and hence any measure of the success of our actions will be in relation to a future state of the world \cite{bialek+al_07}. While not all predictive information is equally useful, non--predictive information is useless, and so separating predictive from non--predictive information is an important task. If we observe a system for time $T$, then the distribution from which trajectories are drawn has an entropy $S(T)$ that we expect to be extensive at large $T$. But the mutual information between, for example, observations on the past ($-T < t \leq 0$) and and equally long future ($0 < t \leq T$) is given by $I_{\rm pred}(T) = 2S(T) - S(2T)$. Importantly, any extensive component of the entropy cancels, and the information that the past provides about the future must be sub--extensive, even if we allow for an unlimited future, so most of the bits we observe must be discarded if we want to isolate the predictive bits. This sub--extensive entropy is related to our intuitive notions of the complexity of the underlying processes \cite{bialek+al_01a,bialek+al_01b}, and providing an efficient representation of predictive information is connected to problems ranging from signal processing to learning \cite{bialek+al_07,creutzig+al_09}. Recent work provides preliminary evidence that populations of neurons in the retina indeed provide such an efficient representation, separating predictive information from the non--predictive background at least in simplified visual worlds \cite{palmer+al_13}, but much more remains to be done in this direction.
\section{Outlook}
We can trace the use of information theoretic ideas to describe biological systems back roughly sixty years, almost to the origins of information theory itself. We have emphasized that these connections between information and the phenomena of life have two very different flavors, one grounded in data and the other aiming for a theory in which the behaviors of real biological systems are derivable from some sort of optimization principle. Both directions have been productive, and as the volume of quantitative experimental data on biological systems grows, it seems clear that the utility of the approaches to data analysis described in Section \ref{sec3} will grow in parallel. The search for real theories is more ambitious.
It is important that we have concrete examples of biological systems that are operating near an optimum of information transmission or efficiency of representation, as reviewed in Section \ref{sec5}. What we don't know is whether these are isolated instances, or examples of a general principle. It is relatively easy to work out what optimization means in simplified settings, but much harder to address the theoretical issues that arise in more realistic contexts, e.g. with many interacting degrees of freedom and the complexities of fully natural signals. At the same time, it took decades to get to the point where even the most basic quantities---the actual information carried by biological signals---could be measured reliably. Now that many of these foundational steps have been solidified, we believe that the coming years will see much more meaningful confrontations between theory and experiment, in a wide range of systems.
\begin{acknowledgments}
We thank our many collaborators and colleagues, whose names appear in the references, for teaching us so much and for making the exploration of these ideas so pleasurable. Our work was supported in part by the US National Science Foundation (PHY--1305525 and CCF--0939370), by the Austrian Science Foundation (FWF P25651), by the Human Frontiers Science Program, and by the Simons and Swartz Foundations.
\end{acknowledgments}
|
\section{Introduction}
In this paper we present a new approach to analyze and extract physical properties of spacetimes around rotating black holes using curvature scalar invariants. This builds on earlier work \cite{lake1,abdelqader1,Pelavas2001}. However, in this paper we go beyond visualization, and use the invariants, and only the invariants, to locate the horizon and ergosufrace, then calculate the mass and angular momentum of the Kerr black hole.
One of the main applications would be in the analysis of numerical relativity simulations. Currently, extracting information about the mass and angular momentum of black holes in numerical simulations requires finding the event horizon of the black hole, calculating the area and angular momentum of the horizon, then using the relationship between the area, mass and angular momentum in order to calculate the mass \cite{numericalbook}.
In section (\ref{invars}), we state the five curvature invariants on which we base the entire calculations that follow. Furthermore, out of the five curvature invariants, we construct and introduce three new dimensionless invariants. These dimensionless invariants serve as detectors for the horizon and ergosurface of the Kerr black hole.
In section (\ref{sec:global}), we present a global approach to analyze the Kerr spacetime and provide a method to extract its mass and angular momentum. First, we show how some dimensionless invariants can be used to locate the event horizon and ergosurface. Next, the area of these two surfaces can be calculated, and this in turn leads to the angular momentum and mass of the black hole. It is also possible to locate the inner event horizon using one of the dimensionless invariants. Therefore, this could provide an alternative technique to find a region to excise around the singularity to be used in the excision method in numerical relativity.
In section (\ref{sec:local}), we present a procedure to calculate the mass and angular momentum locally. In general, the goal and approach we take in this section are similar to the ones in \cite{Ferrando98, Ferrando2009, GomezLobo2007, Backdahl2010}, but the choice of invariants used to carry out the calculations differs. The invariants we use are of degree 2 (i.e. the contractions involve up to 2 factors of the curvature tensor), and order 3 (i.e. up to the 3rd derivative of the metric tensor). On the other hand, the objects used in the references above are of degree 3 and order 3. We present the procedure for the Kerr metric in general, then derive the simplified expressions for the Schwarzschild black hole (i.e. zero angular momentum), and in the weak field limit with angular momentum. Furthermore, as a byproduct of the calculation of the mass and angular momentum locally, it is possible to produce the Boyer-Lindquist coordinates at each point.
The intermediate steps of the procedures are not unique, and it is possible to take a different approach at each step. However, after tedious trial and error, the steps presented here are the ones we found to complete the calculation in the least amount of steps, with the simplest expressions algebraically. However, we show some alternative steps in the Appendix.
In section (\ref{kerrness}) we construct the ``Kerrness'' invariant that serves as an invariant local measure of the spacetime deviation from Kerr. It is a dimensionless invariant that ranges from 0 to 1, where a value of 1 indicates a perfect Kerr spacetime locally. A set of invariants were proposed to achieve the same goal but were based on a different choice invariants can be found in \cite{GomezLobo2012}.
Finally, in section (\ref{sec:discussion}) we discuss the possible applications of the results we present here in the analysis of exact and numerical spacetimes. It is worth noting that the three different procedures presented in sections~(\ref{sec:global}, \ref{sec:local}, and \ref{kerrness}) are complimentary but independent. In other words, each procedure can be performed and completed separately. The only common steps between them are the initial ingredients, which are the invariants presented in section (\ref{invars}).
\section{The curvature invariants of Kerr}\label{invars}
Constructing a minimal list of independent curvature invariants that characterize a spacetime is still an active research field \cite{CM91,ZM97}. In Kerr spacetime, all of the Ricci scalars vanish since it is a vacuum solution, and it has been shown that for the Kerr metric there are at most four independent invariants \cite{Coley2009}. Nonetheless, we consider the following seven invariants in this paper
\footnote{The first two invariants are often discussed in the literature as the real and imaginary parts of the complex Weyl invariant. In vacuum solutions such as the Kerr metric, ${C}_{\alpha \beta \gamma \delta} ={R}_{\alpha \beta \gamma \delta}$, where ${R}_{\alpha \beta \gamma \delta}$ is the Riemann tensor. Therefore, in this case $I_1$ equals the Kretschmann scalar. Furthermore, $I_1$ and $I_2$ can be expressed in terms of invariants in the Newman-Penrose formalism, and in the Kerr spacetime $I_1/48=\Re{(\Psi_2)}^2-\Im{(\Psi_2)}^2$, and $I_2/48=-2\Re{(\Psi_2)}\,\Im{(\Psi_2)}$. For a thorough review of the relationship between many curvature invariants in different notations in general see \cite{Cherubini2002}. The differential invariants $I_3$ and $I_4$ were first introduced and analyzed in \cite{karlhede82} and often refered to as the Karlhede invariants, and the differential invariants $I_5$, $I_6$, and $I_7$ were first introduced in \cite{Lake2004}. The gradient fields $k_{\mu}$ and $l_{\mu}$ for the Kerr metric were thoroughly analyzed in \cite{abdelqader1}.}:
\begin{equation}
I_1 \equiv C_{\alpha \beta \gamma \delta}\;C^{\alpha \beta \gamma \delta} \label{defi1} \;,
\end{equation}
\begin{equation}
I_2 \equiv {C^{*}}_{\alpha \beta \gamma \delta}\;C^{\alpha \beta \gamma \delta}\label{defi2} \;,
\end{equation}
\begin{equation}
I_3 \equiv \nabla_{\mu} C_{\alpha \beta \gamma \delta} \;\nabla^{\mu} C^{\alpha \beta \gamma \delta}\label{defi3}\;,
\end{equation}
\begin{equation}
I_4 \equiv \nabla_{\mu} C_{\alpha \beta \gamma \delta} \;\nabla^{\mu} {C^{*}}^{\; \alpha \beta \gamma \delta}\label{defi4}\;,
\end{equation}
\begin{equation}
I_5 \equiv k_{\mu} k^{\mu}\label{defi5}\;,
\end{equation}
\begin{equation}
I_6 \equiv {l}_{\mu} {l}^{\mu}\label{defi6}\;,
\end{equation}
and
\begin{equation}
I_7 \equiv {k}_{\mu} {l}^{\mu}\label{defi7}\;,
\end{equation}
where $ {C}_{\alpha \beta \gamma \delta} $ is the Weyl tensor, $ {C^{*}}_{\alpha \beta \gamma \delta} $ its dual, $ k_{\mu} \equiv - \nabla_{\mu}\, I_1\,$, and $ {l}_{\mu} \equiv - \nabla_{\mu}\, I_2 \;$. The explicit expression of these invariants for the Kerr spacetime is given in the appendix in a compact form.
Only four of the above seven invariants are actually independent. Most importantly, the calculations in the next two sections to locate the horizon, and calculate the mass and angular momentum are carried out with five of the invariants only, without the need for $I_3$ and $I_4$. However, we present them here for completeness, and in order to explore the three syzygies, or constraining equations, in Kerr spacetime between the full set of the seven non-vanishing invariants in Section~(\ref{kerrness}).
We introduce and define the following three dimensionless invariants constructed entirely out of the five curvature invariants $I_1$, $I_2$, $I_5$, $I_6$, and $I_7$ stated in Eq.~(\ref{defi1}), (\ref{defi2}), and (\ref{defi5}--\ref{defi7})
\footnote{An earlier preprint version of this manuscript required the use of $I_3$ and $I_4$ in the definition of $Q_1$ and $Q_2$. Those earlier definitions are presented here in the Appendix Eqs.~(\ref{q1c}) and (\ref{q2c}). However, we are grateful to Don Page for pointing out two additional syzygies between the seven invariants that we previously missed \cite{Page2015}, which completely eliminated the need for $I_3$ and $I_4$, and simplified the definitions to the ones we present here in Eqs.~(\ref{q1}) and (\ref{q2}).}
\begin{align}
Q_1 & \equiv \frac{1}{3\sqrt{3}} \frac{({I_1}^2-{I_2}^2)(I_5-I_6)+4\,I_1\,I_2\,I_7}{({I_1}^2+{I_2}^2)^{9/4}} \; , \label{q1}
\end{align}
\begin{align}
Q_2 & \equiv \frac{1}{27} \frac{I_5\,I_6-{I_7}^2}{({I_1}^2+{I_2}^2)^{5/2}} \; , \label{q2}
\end{align}
and
\begin{align}
Q_3 \equiv & \frac{1}{6\,\sqrt{3}}\;\frac{I_5+I_6}{({I_1}^2+{I_2}^2)^{5/4}} \; . \label{q3}
\end{align}
It is worth noting that the term $({I_1}^2+{I_2}^2)$ is positive-definite in Kerr spacetime, and it is used in the denominator simply to make the invariants $Q_1$, $Q_2$, and $Q_3$ dimensionless. The significance of $Q_1$ and $Q_2$ will become evident below, as they represent the long sought after invariant detectors for the Kerr black hole ergosurface and event horizon respectively \cite{Moffat,Gass}. $Q_3$ will be used in the calculation of the spin of the black hole locally, but it is redundant since the same calculation could be done with $Q_2$. Nonetheless, we introduce $Q_3$ for the simplicity of the resulting expressions.
\section{Global Approach for Locating the Horizon and Calculating the mass and spin}\label{sec:global}
In this section we present a global approach to calculate the mass and angular momentum in the Kerr spacetime completely based on curvature invariants. To start, we use the invariants to locate two uniquely defined 2D submanifolds. Afterwards, calculating their areas leads us to the mass and spin parameter of the black hole.
\subsection{Locating the Horizon and Ergosurface}
The two submanifold we consider here are the outer horizon and outer ergosurface. After evaluating and simplifying $Q_1$, which was defined in Eq.~(\ref{q1}), in Boyer-Lindquist (BL) coordinates we get
\begin{equation}\label{q1bl}
Q_1=\frac{ \left( r^2-a^2\,\cos^2 \theta \right) \left( r^2-2\,m\,r+a^2\,\cos^2 \theta \right)}{m\,{\left( r^2+a^2\,\cos^2 \theta \right)}^{3/2}} \;.
\end{equation}
Therefore, $Q_1$ vanishes when $r=\pm\, a\,\cos \theta$, and at the ergosurfaces where $r=m\pm\sqrt{m^2-a^2\cos^2 \theta}$. Most importantly, $Q_1$ is strictly positive outside the outer ergosurface, vanishes at the ergosurface, then becomes negative as soon as we cross it. Therefore, it is a very convenient invariant to use to detect the ergosurface in Kerr spacetime. Note that $Q_1$ also vanishes at the inner ergo surface, and at $r=\pm\, a\,\cos \theta$. However, these surfaces lie strictly within the outer ergosurface regardless of the values of $m$ and $a$. Therefore these additional roots of $Q_1$ do not affect its power to detect the outer ergosurface. We should also note that $I_3$ alone has been proposed as a detector for the outer ergosurface, since it does actually vanish at that surface \cite{karlhede82}. However, $I_3$ has many additional roots (nine roots in addition to the outer ergosurface), and some of these roots define surfaces that lie outside the ergosurface, some are inside it, and some actually cross it depending on the values of $m$ and $a$ \cite{Lake2004}. Therefore, it is very difficult to rely on $I_3$ alone as a detector of the outer ergosurface.
Similarly, after evaluating and simplifying $Q_2$, which was defined in Eq.~(\ref{q2}), in (BL) coordinates we get
\begin{equation}\label{q2bl}
Q_2=\frac{ a^2\,\sin^2 \theta \left( r^2-2\,m\,r+a^2 \right)}{m^2\,{\left( r^2+a^2\,\cos^2 \theta \right)}} \;.
\end{equation}
Therefore, $Q_2$ vanishes on the axis of rotation ($\theta=0$), and on the horizon where $r=m\pm\sqrt{m^2-a^2}$. Most importantly, $Q_2$ is strictly positive outside the outer horizon (except for on the axis of symmetry where it vanishes, but clearly never switches signs crossing the axis), vanishes at the outer horizon, then becomes negative as soon as we cross it. The invariant vanishes again at the inner horizon, and switches signs to positive inside the inner horizon. Therefore, $Q_2$ is a very convenient invariant to use for detecting the horizons in Kerr spacetime. The ability to locate the inner horizon efficiently could be exploited for the excision method in numerical relativity, providing an alternative approach to choose a region around the black hole singularity that is required to be located within the outer event horizon.
An earlier preprint version of this manuscript required the use of $I_3$ and $I_4$ in the definition $Q_2$, which we present in Eq.~(\ref{q2c}). It was recently noted by Page and Shoom \cite{Page2015} that the numerator of $Q_2$ in Eq.~(\ref{q2c}), which dictates its roots, can be written as $I_5\,I_6-{I_7}^2=(k\cdot k)(l\cdot l)-{(k\cdot l)}^2$. In other words, the invariant $Q_2$ vanishes when the two gradient fields $k_{\mu}$ and $l_{\mu}$ are parallel. This led them to propose a generalization of $Q_2$, and introduced an invariant that vanishes on Killing horizons in stationary spacetimes in general \cite{Page2015}.
In the case of zero angular momentum ($a=0$), the Kerr solution reduces to the Schwarzschild metric. Note that in this case, $Q_2$ vanishes everywhere, since $I_6=I_7=0$ in Eq.~(\ref{q2}), or $a=0$ in Eq.~(\ref{q2bl}). However, $Q_1$ does not vanish and it serves as the horizon detector since the ergosurface coincides with the horizon in Schwarzschild spacetime, where it reduces to $Q_1=I_5/{(I_1)}^{5/2}$. However, $I_1$ is positive-definite in Schwarzschild, so it is simply the invariant $I_5$ that vanishes and switches signs at the horizons, and only at the horizons. The same is true for the Karlhede invariant $I_3$ which was first observed in \cite{karlhede82}, and the connection between the two invariants is easy to see since $I_5=\tfrac{12}{5}I_1\,I_3$ in Schwarzschild, which is a syzygy that will be explored in detail in Section~(\ref{kerrness}).
\subsection{Calculating the Mass and Spin}
\begin{figure}[t]
\centering
\includegraphics[width=3.0in,angle=0]{fig1.eps}
\caption{\label{fig:area}The ratio between the area of the outer ergosurface to the area of the outer horizon (${\mathcal{A}_{Ergo}}/{\mathcal{A}_{H}}$) for the Kerr black hole as a function of the dimensionless spin parameter $A$.}
\end{figure}
In this subsection we present a method to extract the mass and angular momentum of the Kerr black hole based on the curvature invariants presented above. Once the outer horizon and outer ergosurface are found using the invariants $Q_1$ and $Q_2$, their surface areas can be measured directly, based on the geometry of the spacetime. However, the surface area of the outer horizon $\mathcal{A}_{H}$ is also determined by its mass and angular momentum, and it is given by
\begin{equation}\label{horizonA}
\frac{\mathcal{A}_{H}}{8\pi \, m^2}=1+\sqrt{1-A^2}\; ,
\end{equation}
where $A \equiv a/m$ is the dimensionless spin parameter. The surface area of the outer ergosurface $\mathcal{A}_{Ergo}$ is also determined by the mass and angular momentum, and can be evaluated using
\begin{align}\label{ergoA}
\frac{\mathcal{A}_{Ergo}}{8\pi \, m^2}= &\int\limits_0^{\pi} \Bigg\lbrace
\sqrt{\frac{\sin^2 \theta \left( 1+\sqrt{1-A^2\cos^2 \theta}+A^2\sin^2 \theta \right)}{4 \left(1-A^2\cos^2 \theta \right)}} \nonumber \\
&\times \sqrt{\left( 1+\sqrt{1-A^2\cos^2 \theta} \right) } \Bigg\rbrace \; \mathrm{d}\theta .
\end{align}
The above formula can be expressed in a closed form using elliptical functions, and a thorough analysis of the ergosurface can be found here \cite{Pelavas2001}.
The ratio between the two areas, ${\mathcal{A}_{Ergo}}/{\mathcal{A}_{H}}$, is a one-to-one, strictly increasing function of $A$, which is plotted in Fig.~\ref{fig:area}. Therefore, once $\mathcal{A}_{H}$ and $\mathcal{A}_{Ergo}$ have been measured from the geometry of the spacetime, we can find $A$ directly from the ratio ${\mathcal{A}_{Ergo}}/{\mathcal{A}_{H}}$. Once $A$ is found, we can substitute its value into Eq.~(\ref{horizonA}), and then solve for $m$ to find the value of the mass.
It is worth emphasizing the point that the details of this procedure to extract the mass and spin of the Kerr black hole are not unique, and could be implemented in many ways. In principle, it is possible to construct dimensionless invariants other than $Q_1$ and $Q_2$ such that their roots uniquely define two other surfaces (other than the outer horizon and ergosurface). However, any other two invariants would only be useful for this application if the ratio between the resulting two surface areas is a strictly increasing or decreasing function of the spin parameter $A$. After that, similar steps can be followed as we describe above, where the areas can be measured based on the geometry of the spacetime, then $A$ and $m$ can be extracted from the areas and their ratio.
\section{Local Approach for Calculating the mass and spin}\label{sec:local}
In this section we present another approach to calculate the mass and spin of the Kerr black hole, still within the context of using curvature invariants alone. However, in contrast to the method presented in the previous section, here we calculate $m$ and $A$ based solely on knowing the five invariants ($I_1$,$I_2$,$I_5$,$I_6$, and $I_7$) locally at any point in the spacetime. One of the advantages of this approach is that it does not require locating the black hole or its event horizon. Therefore it could be used to find the total mass and angular momentum of a 2-body system before the merger by applying it in the weak field limit relatively far away from the rotating masses.
We present here the minimal steps needed to carry out the calculations, and include the derivation in the Appendix. To start, we define the dimensionless parameter $ p_1 \equiv a\,\cos \theta / r $, and calculate it using
\begin{align}
p_1 &=-\tan\left[\frac{5}{2} \tan^{-1}\left(\frac{I_2}{I_1}\right) - \tan^{-1}\left(\frac{2\,I_7}{I_5-I_6}\right) \right] \label{eqp1} \; .
\end{align}
Next, we introduce a second dimensionless parameter $ p_2 \equiv r/m $, and calculate it using
\begin{equation}\label{eqp2}
p_2 = \frac{2}{1+p_1^2} + Q_1 \frac{\sqrt{1+p_1^2}}{1-p_1^2} \; .
\end{equation}
The mass $m$ can now be found using
\begin{equation}\label{eqmass}
m_{\text{Kerr}}=\frac{2\sqrt[4]{3}}{ {p_2}^{3/2} \; {(1+{p_1}^2)}^{3/4}\;\sqrt[8]{ {I_1}^2+{I_2}^2} } \; ,
\end{equation}
and the dimensionless spin parameter $A$ can now be found using
\begin{equation}\label{eqA}
A=\sqrt{p_2 \left[1+ Q_3 \sqrt{1+{p_1}^2} - {p_2} \left( 1- {p_1}^2 \right)/2\right] } \quad .
\end{equation}
As it was in the case of the global approach, again the details of this method are not unique. In principle, the steps in Eqs.~(\ref{eqp1}--\ref{eqA}) used to calculate $p_1$, $p_2$, $m_{\text{kerr}}$, and $A$ could be done in many alternative ways based on the same invariants, and we include some examples in the Appendix. However, the steps we present here were the simplest we could achieve from an aesthetic point of view.
The resulting equations for $ A $ and $ m$ above are coordinate independent by definition as they were constructed from scalar invariants. However, we can still extract, for example, the (BL) coordinates as a byproduct of the calculations. More specifically, using the definitions of $p_1$ and $ p_2$, we can calculate the BL radius and polar angle:
\begin{align}
\label{rblcoords}
r & = m\; p_2 \;,
\end{align}
and
\begin{align}
\label{tblcoords}
\cos(\theta)& = p_1\;r/a= p_1\;p_2/A \; .
\end{align}
\subsection{Special Case 1: The Schwarzschild Spacetime (a=0) }\label{sbsec:schw}
In the Schwarzshild black hole (i.e. $a=0$), four of the seven invariants we started the calculation with vanish; $I_2$, $I_4$, $I_6$, and $I_7$. This leads to a simple and exact expression for the mass
\begin{align}
{m}_{\text{Schw}} &\equiv \frac{2\sqrt[4]{3}}{
{\left(2+\tfrac{I_5}{3\sqrt{3}\,{I_1}^{5/2}}\right)}^{3/2} \;\sqrt[4]{{I_1}} } \; . \label{eqmass1}
\end{align}
Also the expression for the BL radius in Schwarzschild spacetime (i.e. the areal radius, which we refer to as $\bar{r}$ here) simplifies to
\begin{align}
\bar{r} &\equiv \frac{2\sqrt[4]{3}}{
{\left(2+\tfrac{I_5}{3\sqrt{3}\,{I_1}^{5/2}}\right)}^{1/2} \;\sqrt[4]{{I_1}}} \; . \label{rschw}
\end{align}
Alternatively, the two equations above can be expressed in terms of the Karlhede invariant $I_3$ instead of $I_5$, by substituting $I_5=\tfrac{12}{5}I_1\,I_3$, which is a syzygy that will be explored in Section~(\ref{kerrness}).
\subsection{Special Case 2: Weak Field Limit ($r/m \gg 1 $)}\label{sbsec:wfl}
In the weak field limit where $p_2=r/m \gg 1$, we have $(1/p_2)\ll 1$ and $|p_1| \ll 1$, since $p_1=A\cos \theta/(r/m)$. To leading order in $p_1$ we find that $p_1\cong I_2/6I_1$. Also, to leading order in $p_1$, Eq.~(\ref{eqmass}) simplifies to
\begin{align}
{m_{\text{Kerr}}} &\cong \frac{2\sqrt[4]{3}}{
{\left(2+\tfrac{I_5-I_6}{3\sqrt{3}\,{I_1}^{5/2}}\right)}^{3/2} \;\sqrt[4]{{I_1}} } \; . \label{mwfl}
\end{align}
The expression for the BL radius in Eq.~(\ref{rblcoords}) simplifies to
\begin{align}
r &\cong \frac{2\sqrt[4]{3}}{
{\left(2+\tfrac{I_5-I_6}{3\sqrt{3}\,{I_1}^{5/2}}\right)}^{1/2} \;\sqrt[4]{{I_1}}} \; . \label{rwfl}
\end{align}
Furthermore, the expression for the dimensionless spin parameter $A$ found in Eq.~(\ref{eqA}) simplifies to
\begin{align}\label{eqAwf}
A &\cong \sqrt{\left(2+\tfrac{I_5-I_6}{3\sqrt{3}\,{I_1}^{5/2}}\right) \Big( I_6\,I_1- \tfrac{8}{7}I_2\,I_7 \Big) \Big/ \sqrt{27\,{I_1}^7\,}\; }\; .
\end{align}
The equation above might be the most relevant to the field of numerical relativity, since it can provide a simple and direct way to approximate the total angular momentum of a binary black hole system when evaluated relatively far away from the system before they merge in the weak field limit. Alternatively, the three equations above can be expressed in terms of $I_3$ instead of $I_5-I_6$, and $I_4$ instead of $I_7$, since in the weak field limit $I_5-I_6 \cong \tfrac{12}{5}I_1\,I_3$, and $I_7\cong \tfrac{21}{10} I_1\,I_4$ as a result of the syzygies that will be explored in the next section.
In a recent paper, a procedure to calculate special relativistic linear and angular momentum based on curvature invariants was proposed \cite{Flanagan:2014}. The procedure requires defining two quantities, $M$ and $r$, based on similar invariants we use here. These quantities resemble the zeroth order approximation of ${m}_{\text{Schw}}$ and $\bar{r}$ found in Eqs.~(\ref{eqmass1}) and (\ref{rschw}). However, using Eqs.~(\ref{mwfl}) and (\ref{rwfl}) for $M$ and $r$ instead might improve the accuracy of the procedure proposed in \cite{Flanagan:2014}.
\section{Invariant syzygies and ``Kerrness'' Invariant}\label{kerrness}
\begin{figure}[t]
\centering
\includegraphics[width=3.3in,angle=0]{fig2.eps}
\caption{\label{fig:kernesscurzon}Contour plot of the Kerrness invariant $K$, which is defined in Eq.~(\ref{K}), in the Curzon-Chazy spacetime. In this plot $s=5$, and the contour levels are 0.01, 0.1, 0.5, 0.85, and 0.96. The lowest contour region (i.e. $0<K<0.01$) is red, and the highest contour region (i.e. $0.96<K<1$) is blue.}
\end{figure}
Only four of the seven invariants introduced in Section~(\ref{invars}) are independent. There are three syzygies, or constraining equations, between the invariants:
\begin{equation}\label{syzygy}
I_6-I_5+ \tfrac{12}{5} \left(I_1\,I_3-I_2\,I_4 \right)=0 \;,
\end{equation}
\begin{equation}\label{syzygy2}
I_7- \tfrac{6}{5} \left(I_1\,I_4+I_2\,I_3 \right)=0 \;,
\end{equation}
and
\begin{multline}\label{syzygy3}
4I_1\,I_2\,I_3({I_1}^2-{I_2}^2)({I_3}^2-3{I_4}^2) \\ =I_4(3{I_3}^2-{I_4}^2)({I_1}^4-6{I_1}^2\,{I_2}^2+{I_2}^4) \;.
\end{multline}
The first syzygy, Eq.~(\ref{syzygy}), was discovered by accident, and the other two, Eqs.~(\ref{syzygy2}) and (\ref{syzygy3}), along with a simple way to derive the syzygies were pointed out to us by Don Page \cite{Page2015}, and we include that derivation in the Appendix.
These syzygies can be exploited to construct a geometric invariant measure of the ``Kerrness'' of a spacetime locally. For example, consider the dimensionless invariant $\chi$ defined as
\begin{equation}\label{chi}
{\chi} \equiv \frac{I_6-I_5+ \tfrac{12}{5} \left(I_1\,I_3-I_2\,I_4 \right)}{{\left({I_1}^2+{I_2}^2 \right)}^{5/4}} \; .
\end{equation}
In Kerr spacetime, evidently $\chi=0$ everywhere as a result of Eq.~(\ref{syzygy}). Furthermore, if we find that $\chi \neq 0$ at some point, this indicates that the local geometry deviates from that of Kerr. However, it is difficult to get an intuitive feel of the scale of this deviation from Kerr directly from the value of $\chi$. Therefore, we construct another dimensionless invariant $K$ based on $\chi$, defined as
\begin{equation}\label{K}
K \equiv e^{-s{\chi}^2} \; ,
\end{equation}
where the constant $s$ is an arbitrary positive number and can be thought of as a sensitivity parameter. By construction, contour levels of $K$ are also contours levels of $\chi$, and both are dimensionless. However, $K$ ranges from 0 to 1. It provides an intuitive measure of how close the spacetime is to the Kerr metric locally, where a value of 1 indicates a perfect Kerr spacetime. The specific value of $K$ is of no real significance, as it is meant to be used in a relative sense, comparing two different points of the spacetime. For example, it can be used to produce what could resemble a heat map of the spacetime, where the regions with the highest values indicate that they are closest to Kerr. We call $K$ the ``Kerrness'' invariant, and propose using it as a spacetime analysis and visualization tool for both exact solutions and numerical relativity simulations.
As an example, we show in Fig.~\ref{fig:kernesscurzon} the contour plot of $K$ for the Curzon-Chazy spacetime \cite{Curzon,Chazy}. This spacetime is believed to be a vacuum spacetime containing a non-rotating singular ring. For a detailed analysis of the Curzon-Chazy solution and its curvature invariants see \cite{abdelqader2} and references within.
\section{Discussion and Conclusion}\label{sec:discussion}
We have presented an invariant characterization of the Kerr spacetime. The physical properties of rotating black hole such as its mass, angular momentum, event horizon, and ergosurface can be defined and described in a coordinate-independent and observer-independent formulation based only on curvature invariants.
The dimensionless invariants introduced in section (\ref{invars}) serve as convenient detectors of the black hole's event horizons and ergosurfaces. This can provide an alternative approach to locate black holes in numerical relativity, and choose an appropriate region to excise around the singularity in the excision method by locating the inner horizon.
Furthermore, the area of the outer ergosurface, along with the area of the outer horizon, provide an alternative method to calculate the mass and angular momentum of rotating black holes as explained in section (\ref{sec:global}). In contrast with current methods used in numerical relativity analysis, this method is completely path independent, and does not require finding the Killing field on the horizon.
Another approach was presented in section (\ref{sec:local}), where the mass and spin parameter can be calculated locally. This procedure could be useful in numerical relativity as well, but will only be reliable in the regions of the spacetime dominated a single Kerr black hole. In numerical simulations of black hole mergers, we expect this procedure to be most reliable near each black hole, where the gravitational field is dominated by one of them, and very far from the system, where the spacetime asymptotically approaches Kerr again but representing the combined mass and angular momentum of the system. However, in the region in between the black holes, the calculations could produce unphysical values of the spin parameter outside the range of 0 to 1, or imaginary numbers. Nonetheless, one of the main advantages of this method is that it does not require finding the horizon, or calculating any surface areas. Therefore, when it is applied in the weak field limit in numerical simulations of black hole mergers, we expect it to give reliable results for the combined mass and angular momentum of the system even before merging.
The Kerrness invariant was introduced in section (\ref{kerrness}), and it can be used in the analysis of exact and numerical spacetimes. For example, it can be used to produce what could resemble a heat map of the spacetime, where the regions with the highest values indicate that this region is closest to Kerr. This provides an invariant and intuitive method to compare and visualize spacetimes. Furthermore, it can be used in combination with local calculations of the mass and spin, as it can indicate the regions where the calculations can be trusted.
\newline
\begin{acknowledgments}
The authors would like to thank Don Page for his constructive review of an early version of this manuscript and pointing out two of the syzygies, which greatly improved and simplified the presentation of the results. MA would like to thank Alan Coley and the Department of Mathematics and Statistics at Dalhousie University for their hospitality, where part of this research was conducted while visiting. MA would like to thank Alan Coley, Robert Owen, and Nathan Deg for helpful discussions. KL would like to thank Eric Poisson and Matt Visser for discussions at a very early stage of this work. This work was supported in part by a grant (to KL) from the Natural Sciences and Engineering Research Council of Canada. Portions of this work were made possible by use of \textit{GRTensorII} \cite{grt}.
\end{acknowledgments}
|
\section{Introduction}
Branes probing toric Calabi-Yau (CY) cones
offer an infinite family of AdS/CFT models
with explicit AdS solutions and field theory Lagrangians.
In particular, D3-branes probing a toric CY$_3$ cone produce a
$D=4$, ${\cal N}=1$ quiver gauge theory which flows to a superconformal
field theory.
The brane tiling model \cite{Hanany:2005ve,Franco:2005rj}
encodes the gauge groups, matter fields, and super-potentials of
the gauge theory into a bipartite graph on a torus.
Algorithms to translate between a toric diagram and the corresponding brane tiling are known.
M2-branes probing a CY$_4$ cone similarly give rise to
a $D=3$, ${\cal N}=2$ superconformal field theory.
But, the problem of constructing the field theory
for an arbitrary toric diagram still has not been solved completely.
An M-theoretic analog of the brane tiling model,
dubbed `brane crystal model' \cite{Lee:2006hw,Lee:2007kv,Kim:2007ic},
helped finding some abelian gauge theories but the non-abelian generalization was obstructed by the lack of a Lagrangian description
for the M5-brane theory.
Progress was made by applying brane tiling methods to Chern-Simons-matter (CSm) theories \cite{Martelli:2008rt,Martelli:2008si,Hanany:2008cd,Ueda:2008hx,Imamura:2008qs,Hanany:2008fj,Franco:2008um,Aganagic:2009zk,Benini:2009qs,Benini:2011cma,Closset:2012ep}. The key idea is to reduce M-theory to IIA string theory along one of the $U(1)^4$ isometry orbits. The gauge theory can be constructed in the IIA setup as usual. The information on the M-theory circle is encoded in the CS levels.
In terms of toric diagrams, the brane tiling model for M2-branes
begins by projecting a three dimensional toric diagram down to two dimensions which gets uplifted back to three dimensions
by the CS levels.
This projection/uplifting procedure is known to work only
for a limited families among all possible toric diagrams.
One of the most detailed confirmation of the toric AdS$_5$/CFT$_4$ correspondence is the equivalence between $a$-maximization \cite{Intriligator:2003jj} and volume-minimization \cite{Martelli:2005tp,Martelli:2006yb},
which was first proved in \cite{Butti:2005vn,Butti:2005ps}.
The $a$-function is a cubic function of the trial R-charge
which is a linear combination of all global $U(1)$ symmetries.
The coefficients of the cubic polynomial
are areas of triangles in the toric diagram
\cite{Benvenuti:2006xg,Lee:2006ru}.
The global symmetries have two types: mesonic and baryonic.
Geometrically, mesonic symmetries are the $U(1)^3$ isometries
of the CY cone, whereas baryonic symmetries correspond to
homology 3-cycles. In the proof of the equivalence \cite{Butti:2005vn,Butti:2005ps},
the $a$-function is maximized with respect to baryonic components first.
After the baryonic components are eliminated,
the remaining $a$ as a function of mesonic components
is shown to be equal, up to an overall numerical factor, to the inverse of the volume \cite{Martelli:2005tp} as a function of the Reeb vector components. The Reeb vector is the geometric counterpart of the R-charge.
The $a$-function is defined in terms of 't Hooft anomaly and has no counterpart in odd dimensions. For $D=3$, ${\cal N}=2$ theories,
the supersymmetric free energy on three-sphere, $F = - \log|Z_{S^3}|$,
was argued to play the role of the $a$-function \cite{Kapustin:2009kz,Jafferis:2010un,Hama:2010av}.
Much like the $a$-function, $F$ decreases along an RG flow, and
the superconformal R-charge
can be determined by extremizing $F$; see \cite{Closset:2012vg} for a proof.
In the large $N$ limit, the free energy
is related to the volume of the Sasaki-Einstein seven-manifold as \cite{Martelli:2011qj,Cheon:2011vi,Jafferis:2011zi}
\begin{align}
F = N^{3/2} \sqrt{\frac{2\pi^6}{27{\rm Vol}(Y_7)}} \,.
\label{F-vol-1}
\end{align}
The current paper addresses the question of establishing
the $F$ vs volume relation \eqref{F-vol-1} for arbitrary toric CY$_4$ cone,
with both sides regarded as functions of mesonic charges.
Compared to the original $a$-max/vol-min problem,
this question poses several additional difficulties.
Originating from a 't Hooft anomaly, the $a$-function is
a cubic polynomial of the coefficients of the trial R-charge.
But, there is no a priori reason for the $F$-function
to take a simple polynomial form.
Even when the large $N$ limit of the $F$-function takes a simple form,
it is not visible until the last stage of localization computation.
Computing $F$ for many examples would be desirable. But, as mentioned earlier, there is no general method to construct the gauge theory
for arbitrary toric diagram.
Even when the gauge theory Lagrangian is known,
some $U(1)$ global symmetries are realized non-perturbatively
and make it difficult to include in the trial R-charge
with independent coefficients.
Despite these obstacles, Amariti and Franco \cite{Amariti:2012tj}
made some remarkable progress. (See \cite{Amariti:2011uw} for an earlier attempt.)
They constructed gauge theories dual to
a few infinite families of toric diagrams with up to eight vertices,
and computed $F$ in the large $N$ limit.
Trying to interpret the results in a geometric way,
they argued that $F$ should take the general form,
\begin{align}
\frac{F^2}{N^3} \propto \sum_{I,J,K,L} V_{IJKL} \Delta^I \Delta^J \Delta^K \Delta^L + (\mbox{corrections}).
\label{Q-intro}
\end{align}
The $\Delta^I$ are the coefficients of the trial R-charge,
each associated to a vertex of the toric diagram,
and $V_{IJKL}$ is proportional to the volume of the tetrahedron
formed by four vertices of the toric diagram.
So, the leading term is a natural generalization of the cubic form of $a$
\cite{Benvenuti:2006xg,Lee:2006ru}.
They also argued that
the correction terms should be assigned to internal edges of the toric diagram.
They determined the precise form of the correction term
for 5-vertex models, and gave some preliminary results for 6- and 8-vertex
models.
In the current paper, we propose a purely geometric method to determine the correction terms in the quartic polynomial \eqref{Q-intro}
without restrictions from gauge theory realizations.
We begin with the Amariti-Franco proposal with unknown coefficients for the correction terms. We decompose the trial R-charge into baryonic and mesonic components. Schematically, we have
\begin{align}
F^2 \sim t^4 + t^3 s + t^2 s^2 + t s^3 + s^4\,,
\label{F-st-intro}
\end{align}
where $t$ and $s$ represent baryonic and mesonic components.
Our main result consists of two statements. First, the correction terms are uniquely determined by demanding that the $t^4$ and $t^3$ terms cancel out. Second, once the baryonic components are eliminated by extremizing
$F^2$ in \eqref{F-st-intro}, the remaining function of mesonic components
coincide precisely with the inverse volume of the toric Sasaki-Einstein manifold \cite{Martelli:2005tp,Martelli:2006yb}.
We verify our claims explicitly for most general 5- and 6-vertex models
and some 8-vertex models, leaving the general case as a conjecture.
Our proposal for the geometric free energy was inspired by
an analogous decoupling of baryonic charges in the $a$-max/vol-min
problem in the AdS$_5$/CFT$_4$ setup together with
the concrete form of Amariti-Franco proposal for 5-vertex models.
In section \ref{sec:geo}, we will review the aspects of toric geometry relevant to our problem and spell out the precise statement of our proposal. In section \ref{sec:qft}, we reproduce the field theory computation
of \cite{Amariti:2012tj} and confirm that our proposal is
consistent with all infinite families of examples.
In section \ref{sec:grav}, we turn to the AdS side of AdS$_4$/CFT$_3$.
In particular, we explore the possibility of
using the same quartic polynomial as the prepotential
in the gauged supergravity.
We compute the gauge kinetic terms in Kaluza-Klein gravity and gauged supergravity descriptions. While the mesonic sector exhibits perfect agreement, the baryonic sector shows some mild discrepancy.
We conclude with a comment on how to resolve the discrepancy.
\section{Geometry \label{sec:geo}}
After two short reviews, we will present the geometric free energy proposal, which is the main result of the whole paper. We will give explicit form of the free energy for general 5-vertex and 6-vertex models,
and close the section with a discussion on generalization.
\subsection{Toric Sasaki-Einstein manifold \label{sec:toreview}}
An $n$-dimensional toric cone $X$ is constructed by a GLSM quotient of $\{ Z^I \} \in \mathbb{C}^{d}$ with respect to
integer-valued charges $Q_a^I$ $(a = 1, \cdots, d-n)$,
\begin{eqnarray}
\label{glsm}
X = \left\{ \sum_{I=1}^d Q_a^I |Z^I|^2 =0 \right\}
/
(Z^I \sim e^{\theta^a Q_a^I} Z^I) \,.
\end{eqnarray}
The cone is Calabi-Yau (CY) if and only if $\sum_I Q_a^I = 0$ for each $a$.
Let $v^i$ ($i=1,2,\ldots, n$) be the kernel of the map
$Q_a : \mathbb{Z}^d \rightarrow \mathbb{Z}^{d-n}$, {\it i.e.}, $Q_a^I v_I^i = 0$.
One may regard $v_I^i$ as $d$ lattice vectors in $\mathbb{Z}^n$
and use them to parametrize $|Z^I|^2 = v_I \cdot y \equiv v_I^i y_i$
$(y \in \mathbb{R}^n)$. The allowed values of $y$
form a polyhedral cone $\triangle$ defined by $\{ v_I \cdot y \ge 0\}$ in $\mathbb{R}^n$.
The cone $X$ is then a fibration of $n$ angles
$\{\phi^i\}$ over the base $\triangle$.
Using the CY condition $\sum_I Q_a^I = 0$,
one can choose $v_I^n=1$ for all $I$, as this assignment satisfies
$Q_a^I v_I^n = 0$ automatically. With $v_I^n=1$,
the collection of the remaining components of $v_I$'s drawn on $\mathbb{Z}^{n-1}\in \mathbb{R}^{n-1}$
will be called the toric diagram.
By construction, the toric $X$ has $n$ isometries
$K_i = \partial/\partial \phi^i$. The Reeb vector $K_R$ is in general
a linear combination, $K_R = b^i K_i$. In \cite{Martelli:2005tp},
it was shown that the Reeb vector characterizes all the essential
geometric properties of the cone $X$. In particular, the base $Y$ of the cone is defined as $Y = X \cap \{b\cdot y = 1/2\}$. Supersymmetric cycles of $Y$ are given by
$\Sigma^I = Y \cap \{v_I \cdot y=0 \}$.
By definition, $X$ being K\"ahler or Ricci-flat is equivalent to
$Y$ being Sasakian or Einstein, respectively.
The Reeb vector determines a unique Sasakian metric on $Y$.
The volume of $Y$ can be computed by summing over the volume
of the supersymmetric cycles $\Sigma^I$ associated to
the vertices $v_I$ of the toric diagram.
The CY condition on $X$ fixes $b^n =n$. The metric of $Y$ become Einstein at the
minimum of $\mbox{Vol}(Y)$ as $(b^1, b^2, \ldots, b^{n-1}; b^n=n)$ is varied inside the polyhedral
cone: $b \in \triangle$.
This is the volume-minimization to be compared with field theory results
via AdS/CFT.
Concretely, for $n=3$, the volume as a function of the Reeb vector is given by the Martelli-Sparks-Yau formula \cite{Martelli:2005tp},
\begin{align}
\frac{\mbox{Vol}_{\rm MSY}(Y_5)}{\mbox{Vol}(S^5)} = \frac{1}{b^3} \sum_I \frac{\langle
v_{I-1},v_I,v_{I+1} \rangle}{\langle b,v_{I-1},v_I \rangle \langle
b, v_I, v_{I+1} \rangle} \equiv \frac{1}{b^3} \sum_I L^I(b) \,.
\label{YS5}
\end{align}
Here, $\langle u, v, w \rangle$ denotes the determinant of the
$(3\times 3)$ matrix made out of vectors $u, v, w$.
For $n=4$, the volume is again expressed as a sum over the vertices of
the toric diagram,
\begin{align}
\frac{\mbox{Vol}_{\rm MSY}(Y_7)}{\mbox{Vol}(S^7)} = \frac{1}{b^4} \sum_I L^I(b) \,,
\label{YS7}
\end{align}
but the precise form of $L^I(b)$ depends on how many neighboring vertices the vertex $v_I$ has. In the simplest case of three nearest neighbors, say, $\{v_J, v_K, v_L\}$, it is given by
\begin{align}
L^I(b) = M^I_{JKL}(b) \equiv \frac{\langle v_I,v_J,v_K,v_L \rangle^2}{
\langle b,v_I,v_J,v_K \rangle
\langle b,v_I,v_K,v_L \rangle
\langle b,v_I,v_L,v_J \rangle} \,.
\end{align}
Our convention for the orientation of the vertices are explained in Figure~\ref{fig:MSY-CY4}.
When there are more then three neighboring vertices,
we can triangulate the ``polygon" composed of neighboring vertices
to compute $L^I$. For instance, with four neighboring vertices, we obtain
\begin{align}
L^I = M^I_{JKL} + M^I_{JLM} = M^I_{JKM} + M^I_{MKL} \,.
\end{align}
The generalization to more neighboring vertices is straightforward.
\begin{figure}[h!]
\centering
\includegraphics[width=7cm]{MSY-CY4.pdf}
\caption{The volume of the supersymmetric cycle associated to a vertex $v_I$. When viewed from the ``outside" of the toric diagram, the neighboring vertices $\{ v_J, v_K, \cdots\}$ are aligned along the ``polygon" in a clock-wise order. (a) Three neighboring vertices introduce a tetrahedron (b) Four neighboring vertices lead to a triangulation composed of two tetrahedra}
\label{fig:MSY-CY4}
\end{figure}
As explained in \cite{Franco:2005sm}, when $Y$ is simply-connected,
which we assume for the rest of this paper, the homology group of
$Y$ is given by $H_{2n-3}(Y,\mathbb{Z})=\mathbb{Z}^{d-n}$. If $C^a$ ($a=1,\cdots ,d-n$)
form a basis of $(2n-3)$-cycles of $Y$, it can be shown that
$\Sigma^I = Q_a^I C^a$
with $Q_a^I$ being precisely the GLSM data \eqref{glsm}.
As the harmonic $(2n-3)$-forms $\omega_a$
dual to $C^a$ measure the baryonic charges of $\Sigma^I$, we have
\begin{eqnarray} \label{baryoncharge}
B_a\left[\Sigma^I\right] = \int_{\Sigma^I} \omega_a =
Q_a^I . \end{eqnarray}
As one can see from the torus action in the GLSM description (\ref{glsm}),
for simply connected $Y$, the baryonic charges $Q_a^I$ and the mesonic charges $(K_i = \partial/\partial \phi^i)$ together span $\mathbb{Z}^d$.
This means that the toric relation $Q_a^I v_I^i=0$ can be extended to
\begin{align}
\label{extor}
\begin{pmatrix} Q_a{}^I \\ F_i{}^I \end{pmatrix}
\begin{pmatrix} u_I{}^{b} & v_I{}^j \end{pmatrix}
= \begin{pmatrix} \delta_a^b & 0 \\ 0 & \delta_i^j \end{pmatrix},
\end{align}
for some integer-valued matrices $F_i^I$ and $u_I^b$ \cite{Lee:2006ru}.
The volume of a supersymmetric cycle is mapped to the superconformal R-charge of the corresponding local operator via AdS/CFT \cite{Berenstein:2002ke}. For later convenience, we follow \cite{Martelli:2005tp} to define the geometric R-charge $\Delta^I_{\rm MSY}(b)$ as
\begin{align}
\Delta^I_{\rm MSY}(b) = \frac{2L^I(b)}{\sum_I L^I(b)} \,.
\end{align}
\subsection{A-maximization revisited \label{sec:a-max} }
For toric theories, the $a$-function takes a simple geometric form
\cite{Benvenuti:2006xg,Lee:2006ru}
\begin{align}
\bar{a}(\Delta) \equiv \frac{a(\Delta)}{N^2} = \frac{9}{32} C_{IJK} \Delta^I \Delta^J \Delta^K = \frac{9}{64} |\langle v_I, v_J, v_K \rangle | \Delta^I \Delta^J \Delta^K \,,
\end{align}
where each coefficient
\begin{align}
C_{IJK} = \frac{1}{2} |\langle v_I, v_J, v_K \rangle |
\label{ccc}
\end{align}
is the area of the triangle composed of three vertices $(I,J,K)$ on the toric diagram.
\begin{figure}[h!]
\centering
\includegraphics[width=11cm]{cy3-toric.pdf}
\caption{Toric diagram of CY$_3$.}
\label{fig:cy3-toric}
\end{figure}
\paragraph{Volume as the inverse of $a$: overview}
The equivalence between $a$-maximization and volume-minimization was originally proved in \cite{Butti:2005vn}. The proof was simplified in \cite{Lee:2006ru} using the triangle formula \eqref{ccc}. The proof
roughly consists of three steps.
First, we decompose the trial R-charges into
a linear combination of the baryonic and the mesonic charges,
\begin{align}
\Delta^I = t^a Q_a^I + s^i F_i^I \,.
\label{decomp-st1}
\end{align}
In terms of the $t$ and $s$ variables, the $a$-function decomposes into, schematically,
\begin{align}
a \sim t^3 + t^2 s + t s^2 + s^3\,.
\label{a-st1}
\end{align}
Second, by mathematical induction using \eqref{ccc} \cite{Benvenuti:2006xg,Lee:2006ru}, we can show that the $t^3$ terms vanish identically for any toric theory. The remaining terms can be reorganized as
\begin{align}
a = -m_{ab}(s) t^a t^b + 2 n_a(s) t^a + R(s) \,,
\label{a-st2}
\end{align}
where $m_{ab}$, $n_a$, $R$ are homogeneous polynomials of degree one, two and three in $s$, respectively.
Extremizing $a$ with respect to $t$, we obtain an intermediate result,
\begin{align}
\Delta^I(s) = Q_a^I m^{ab}(s) n_b(s) + s^i F_i^I \,,
\qquad
\bar{a}(s) = R(s) + m^{ab}(s) n_a(s) n_b(s) \,,
\end{align}
where $m^{ab}$ is the matrix inverse of $m_{ab}(s)$.
Finally, the equivalence between $a$-maximization and
volume-minimization is established by
proving that
\begin{align}
\Delta^I(s) = \Delta^I_{\rm MSY}(b) \left. \right|_{s=(2/3)b} \,,
\quad
\bar{a}(s) = \left. \frac{\pi^3}{4{\rm Vol}_{\rm MSY}(b)} \right|_{s=(2/3)b} \,.
\label{av0}
\end{align}
\paragraph{Some details}
We review parts of the proof of the assertions above
that will be relevant for generalizations to the AdS$_4$/CFT$_3$ setting. To begin with, for general CY$_n$, we define the normalized Reeb vector $x^i$ by
\begin{align}
x^i = \frac{s^i}{2} = \frac{b^i}{n} \qquad (i=1,\ldots,n) \,,
\label{bsbsx}
\end{align}
such that $x^n=1$ and the domain of $x^{i=1,\ldots,n-1}$ is precisely the interior of the toric diagram.
We rewrite the relation \eqref{YS5}, \eqref{YS7} between $\mbox{Vol}(Y)$ and $\mbox{Vol}(\Sigma^I)$ as
\begin{align}
S(x) = \frac{1}{x^n} \sum_I L^I(x) \,,
\end{align}
As shown in \cite{Martelli:2005tp}, it is a part of a more general relation,
\begin{align}
\sum L^I(x) v_I^i = \frac{x^i}{x^n} \sum_I L^I(x) = x^i S(x) \,,
\label{LS-id}
\end{align}
which can be proved by applying Stokes' theorem in the toric diagram.
Specializing to CY$_3$, with $(b^1, b^2, b^3) = 3(x^1 ,x^2, x^3=1)$, we introduce \cite{Lee:2006ru}
\begin{align}
&r_I = (v_I^1, v_I^2) - (x^1 ,x^2) \,,
\quad
w_I = (v_{I+1}^1,v_{I+1}^2) - (v_{I}^1, v_{I}^2) \,,
\nonumber \\
&A_I = \langle r_I, w_I \rangle \equiv
{\rm det}
\begin{pmatrix}
r_I^1 & r_I^2 \\
w_I^1 & w_I^2
\end{pmatrix} \,,
\quad
L^I(x) = \frac{\langle w_{I-1}, w_I\rangle}{A_{I-1}A_I}\,,
\quad
S(x) = \frac{1}{x^3} \sum_I L^I(x) \,.
\end{align}
See Figure~\ref{fig:cy3-toric}(b) for the geometric meaning of each quantity.
Now, the first half of the proof of \eqref{av0} asserts that the baryon charges decouple
from the maximization process:
\begin{align}
{\rm Tr} B R^2 |_{\Delta^I_{\rm MSY}} = 0
\quad \mbox{or equivalently} \quad
C_{IJK} B^I L^J L^K = 0 \,.
\label{av1}
\end{align}
The other half states that
\begin{align}
a_{\rm CFT}|_{\Delta^I_{\rm MSY}} = \frac{\pi^3}{4{\rm Vol}_{\rm MSY}}
\quad \mbox{or equivalently} \quad
C_{IJK} L^I L^J L^K = 3 S^2 \,.
\label{av2}
\end{align}
As proved in \cite{Benvenuti:2006xg,Lee:2006ru}, both \eqref{av1} and \eqref{av2} follow from a single lemma:
\begin{align}
c_I \equiv C_{IJK} L^J L^K = 3 S + \langle r_I , u \rangle ,
\label{lemma}
\end{align}
where $u$ is some vector independent of the label $I$.
Once the lemma is proved, (\ref{av1}) follows from $\sum_I Q_a^I = 0 = \sum_I Q_a^I
v_I$ and (\ref{av2}) from $\sum_I L^I r_I = 0$.
\subsection{Geometric free energy}
\label{sec:geom-free-energy}
Amariti and Franco \cite{Amariti:2012tj} computed the large $N$ free energy of
a large class of toric CFT$_3$'s. They found that, for all
examples they considered, the following relation holds:
\begin{align}
\bar{F}^2(\Delta) \equiv \frac{9F^2(\Delta)}{2\pi^2 N^3} = \frac{2}{3} C_{IJKL} \Delta^I \Delta^J \Delta^K \Delta^L \,,
\end{align}
where the coefficients take the general form,
\begin{align}
C_{IJKL} = |\langle v_I, v_J, v_K, v_L \rangle | + \mbox{(corrections)}\,.
\label{cijkl}
\end{align}
The normalization for $\bar{F}^2$ is chosen such that
$\bar{F}^2=1$ for CY$_4=\mathbb{C}^4$.
We warn the readers that the ``correction" terms are not meant to be smaller than the ``leading" terms. They are just less obvious than the leading terms. Amariti and Franco also noticed that all correction terms are somehow associated to internal lines of the toric diagram.
More specifically, there is a type 1 correction term for each internal line,
and a type 2 correction term for each pair of internal lines.
The goal of this section is to turn the observations of Amariti and Franco to a general conjecture for the form of correction terms and
to gain some geometric understanding. As an application of the conjecture, we will determine
the correction terms explicitly for generic toric diagrams with 5 or 6 vertices and some specific diagrams with 7 or 8 vertices.
The key idea behind the conjecture is that the correction terms
Amariti and Franco found for particular examples are such that
the quartic and cubic terms in baryonic components of the trial R-charge
(to be called $t^4$ and $t^3$ terms below)
vanish identically. We reverse the logic and base our conjecture on four central assumptions.
\begin{enumerate}
\item
The leading term always take the same form as in \eqref{cijkl}.
\item
The type 1 and type 2 terms explained below \eqref{cijkl} exhaust
all possible corrections.
\item
The coefficients of the correction terms are rational functions of
$|\langle v_I, v_J, v_K, v_L \rangle |$.
\item
The vanishing of $t^4$ and $t^3$ constrains the correction coefficients.
\end{enumerate}
The decoupling of baryonic charges goes in close parallel
with the AdS$_5$/CFT$_4$ story reviewed in the previous subsection.
We decompose the trial R-charges as
\begin{align}
\Delta^I = t^a Q_a^I + s^i F_i^I \,.
\label{decomp-st}
\end{align}
The charges are subject to $\sum_I \Delta^I = 2$,
which is equivalent to $s^4=2$ and $b^4=4$.
It is a special case of \eqref{bsbsx} at $n=4$.
In terms of the $t$ and $s$ variables, the function $F^2$ looks like
\begin{align}
F^2 \sim t^4 + t^3 s + t^2 s^2 + t s^3 + s^4\,.
\label{F-st}
\end{align}
Our conjecture propose that the correction terms should be chosen such that the $t^4$ and $t^3$ terms vanish. A priori, the existence and the uniqueness of such correction terms are not obvious at all. At the time of writing, we do not know how to prove or disprove the conjecture. We will simply explore the possibilities by starting from the simplest case and proceeding to more complicated ones.
Assuming the vanishing of $t^4$ and $t^3$ terms in \eqref{F-st}, we
can organize the remaining terms as follows,
\begin{align}
F^2 = -m_{ab}(s) t^a t^b + 2 n_a(s) t^a + R(s) \,.
\label{F-st2}
\end{align}
The functions $m_{ab}$, $n_a$ and $R$ are homogeneous polynomials
of $s$ of degree $2$, $3$ and $4$, respectively.
Maximizing $F^2$ with respect to $t$ gives $t^a = m^{ab} n_b$. Inserting it back to \eqref{decomp-st} and \eqref{F-st2}, we obtain
\begin{align}
\bar{\Delta}^I(s) = Q_a^I m^{ab}(s) n_b(s) + s^i F_i^I \,,
\nonumber \\
\bar{F}^2(s) = R(s) + m^{ab}(s) n_a(s) n_b(s) \,.
\end{align}
Further extremization of $\bar{F}^2$ determines the ``vacuum" value of $s$, which we call $s_*$.
A major sequel to our conjecture is that $\bar{\Delta}^I$ and $\bar{F}^2$ match their geometric counterparts even before extremization with respect to $s$, just as in the proof \cite{Butti:2005vn} of the $a$-maximization vs volume-minimization.
\begin{align}
\Delta^I(s) = \Delta^I_{\rm MSY}(b) \left. \right|_{s=b/2} \,,
\quad
\bar{F}^2(s) = \left. \frac{\pi^4}{3{\rm Vol}_{\rm MSY}(b)}\right|_{s=b/2} \,.
\label{vol-inv-fsq}
\end{align}
Again, it is not clear how this result follows from our conjecture.
In the following subsections, we will verify this claim for several families of concrete examples and sketch some ideas for the general proof.
\subsection{5-vertex models}
As shown in Figure~\ref{fig:5v-general},
a generic toric diagram with 5 vertices contains one internal line.
The non-generic configuration
with no internal line can be smoothly reached from the generic case.
For instance,
one can move the vertex $4$ in Figure~\ref{fig:5v-general} continuously, with all others fixed, until the internal line $\overline{45}$ intersects the external edge $\overline{12}$.
\begin{figure}[htbp]
\centering
\includegraphics[width=3cm]{5v-general.pdf}
\caption{A generic toric diagram with 5 vertices has one internal line.}
\label{fig:5v-general}
\end{figure}
\noindent
The Amariti-Franco proposal \cite{Amariti:2012tj} for the 5-vertex model is
\begin{align}
\bar{F}^2(\Delta) = \sum_{I<J<K<L} V_{IJKL} \Delta^I \Delta^J \Delta^K \Delta^L
- \frac{V_{1245}V_{2345}V_{3145}}{V_{1234}V_{1235}} (\Delta^4\Delta^5)^2 \,,
\end{align}
where we defined $V_{IJKL} = |\langle v_I, v_J, v_K, v_L \rangle|$.
This proposal is the simplest non-trivial case of our general conjecture.
The simplicity of the 5-vertex model allows us to prove the vanishing of $t^4$ and $t^3$ terms by straightforward computation.
\paragraph{Vanishing of $t^4$ and $t^3$ terms}
Taking account of the relative orientations of the vertices,
one can remove the absolute value sign from the definition of $V_{IJKL}$,
\begin{align}
&\qquad\qquad\qquad
V_{1234} = - \langle v_1, v_2, v_3, v_4 \rangle\,,
\quad
V_{1235} = + \langle v_1, v_2, v_3, v_5 \rangle\,,
\nonumber \\
&
V_{1245} = \langle v_1, v_2, v_4, v_5 \rangle\,,
\quad
V_{2345} = \langle v_2, v_3, v_4, v_5 \rangle\,,
\quad
V_{3145} = \langle v_3, v_1, v_4, v_5 \rangle\,,
\quad
\end{align}
The 5-vertex models have only one set of GLSM charges $Q^I_{a=1}\equiv Q^I$.
One may define
\begin{align}
(I,J,K,L) \equiv \langle v_I Q^I, v_J Q^J, v_K Q^K, v_L Q^L \rangle
\qquad
\mbox{(no sum over indices)} \,.
\end{align}
Using the fact that $\sum_I v_I Q^I =0$, one can replace all
$(I,J,K,L)$'s by, say, $(1,2,3,4)$:
\begin{align}
(1,2,3,5) = -(1,2,3,4)\,,
\quad (2,3,4,5) = -(2,3,4,1) = +(1,2,3,4)\,.
\end{align}
Now, the coefficient of the $t^4$ term, $C_{IJKL}Q^I Q^J Q^K Q^L$, is proportional to
\begin{align}
&-(1234) + (1235) + (1245)+(2345)+(3145) + \frac{(1245)(2345)(3145)}{(1234)(1235)}
\nonumber \\
&\qquad = \{-1-1+1+1+1\} (1,2,3,4) - \frac{(1,2,3,4)^3}{(1,2,3,4)^2} = 0 \,.
\end{align}
Next, the coefficients of $t^3$ terms are proportional to $T_I \equiv C_{IJKL}Q^J Q^K Q^L$. $T_1$ is proportional to
\begin{align}
&-(1,2,3,4)+ (1,2,3,5)+ (1,2,4,5)+(3,1,4,5)
\nonumber \\
&\qquad = \{-1-1+1+1\} (1,2,3,4) = 0 \,,
\end{align}
and similarly for $T_2$ and $T_3$.
On the other hand, $T_4$ is proportional to
\begin{align}
&-(1234)+ (1245)+(2345)+(3145) + 2 \frac{(1245)(2345)(3145)}{(1234)(1235)}
\nonumber \\
& \qquad = \{-1+1+1+1\} (1,2,3,4) - 2 \frac{(1,2,3,4)^3}{(1,2,3,4)^2} = 0 \,,
\end{align}
and similarly for $T_5$. This completes the proof of the vanishing of
all $t^4$ and $t^3$ terms for general 5-vertex models.
\paragraph{Volume as the inverse of $F^2$}
For general 5-vertex models, it is straightforward, albeit tedious,
to integrate out the $t$ variable
and prove the identity \eqref{vol-inv-fsq}
relating $\bar{F}^2(s)$ to the inverse of ${\rm Vol}_{\rm MSY}(b)$.
In practice, the algebraic manipulation is most easily done
with the aid of a computer program.
\subsection{6-vertex models}
The Amariti-Franco proposal \cite{Amariti:2012tj} does not cover all generic 6-vertex models. As explained earlier, we use the vanishing of $t^4$ and $t^3$ terms to find the form of the correction terms. Under the general assumptions of our conjecture, the correction terms are uniquely determined. Moreover, once the $t$ variables are integrated out,
the resulting $\bar{F}^2(s)$ is shown to be proportional to ${\rm Vol}_{\rm MSY}(b)$ as in \eqref{vol-inv-fsq}.
The computation involves quite a few variables. The position of the 6 vertices in $\mathbb{R}^3$ are specified by 18 parameters. Using the homogeneity of $F^2$ as well as the $SL(3,\mathbb{Z})$ and translation symmetries of the toric diagram,
we can fix 12, leaving 6 free parameters. The Reeb vector components add 3 variables. Proving identities among rational functions of 9 variables is often impractical even with a computer program.
We use the well-known fact that two rational functions are identical to each other if they yield the same value at sufficiently many different ``sampling" points. The number of points should be greater than the sum
of degrees of the numerator and the denominator of the rational function.
Throughout this subsection, it should be understood that
the vanishing of $t^4$, $t^3$ terms
and the equivalence between $\bar{F}^2$ and ${\rm Vol}_{\rm MSY}$
have been verified by the sampling method.
The 6-vertex models have a number of distinct configurations of internal lines.
One way to proceed is to begin with a toric diagram with no internal line and to add internal lines one at a time by deforming the position of some of the vertices.
\paragraph{Two internal lines meeting at a vertex}
One such example is depicted in Figure~\ref{fig:6v(a)}. We begin with a `triangular prism' which has no internal line. By pushing the vertex 4 toward the edge $\overline{56}$, we introduce two internal lines $\overline{24}$ and $\overline{34}$.
\begin{figure}[htbp]
\centering
\includegraphics[width=7cm]{6v-a.pdf}
\caption{A toric diagram with 6 vertices and two internal lines emanating from the same vertex.}
\label{fig:6v(a)}
\end{figure}
\noindent
After some trial and error in numerical experiment, we find the two
types of corrections terms:
\begin{align}
&\delta_1(F^2) =
- \frac{V_{2456}V_{2461}V_{2415}}{V_{2561}V_{4561}} (\Delta^2\Delta^4)^2
- \frac{V_{3465}V_{3451}V_{3416}}{V_{3561}V_{4561}} (\Delta^3\Delta^4)^2 \,,
\nonumber \\
&\delta_2(F^2) = - 2\frac{V_{2415}V_{3461}}{V_{4561}}
\left(\frac{V_{4abc}}{V_{1abc}}\right) (\Delta^2\Delta^4)(\Delta^3\Delta^4) \,.
\label{6v-2a}
\end{align}
Here, the indices $abc$ are three elements from $\{2,3,5,6\}$. The choice
of which three elements does not affect the result since the four vertices
$\{v_2,v_3,v_5,v_6\}$ lie on the same plane.
\paragraph{Two internal lines not meeting each other}
Another example with two internal lines is depicted in Figure~\ref{fig:6v(b)}. We begin again with the triangular prism and push the vertex 4 slightly parallel to the edge $\overline{56}$.
\begin{figure}[htbp]
\centering
\includegraphics[width=7cm]{6v-b.pdf}
\caption{A toric diagram with 6 vertices and two non-intersecting internal lines.}
\label{fig:6v(b)}
\end{figure}
\noindent
The correction terms in this case are
\begin{align}
&\delta_1(F^2) =
- \frac{V_{2453}V_{2431}V_{2415}}{V_{2531}V_{4531}} (\Delta^2\Delta^4)^2
- \frac{V_{1634}V_{1645}V_{1653}}{V_{1345}V_{6345}} (\Delta^1\Delta^6)^2 \,,
\nonumber \\
&\delta_2(F^2) = 2 \frac{V_{2415}V_{3461}}{V_{1345}}
(\Delta^2\Delta^4)(\Delta^1\Delta^6) \,.
\label{6v-2b}
\end{align}
This example meets the previous one when the vertices $\{v_1, v_3, v_4, v_6\}$ fall onto the same plane such that $V_{3461}=0$.
The coefficients of the $(\Delta^2 \Delta^4)^2$ term in \eqref{6v-2a} and \eqref{6v-2b} look different, but they can be shown to be equal when $\{v_1, v_3, v_4, v_6\}$ lie on the same plane.
\paragraph{Three connected internal lines}
We deform Figure~\ref{fig:6v(a)} further by turning on the third internal line $\overline{35}$. The result is depicted in Figure~\ref{fig:6v(d)}.
\begin{figure}[htbp]
\centering
\includegraphics[width=7cm]{6v-d.pdf}
\caption{A toric diagram with 6 vertices and three internal lines
connected at vertices.}
\label{fig:6v(d)}
\end{figure}
\noindent
The correction terms turn out to be
\begin{align}
&\delta_{1}(F^2) =
- \frac{V_{2456}V_{2461}V_{2415}}{V_{2561}V_{4561}} (\Delta^2\Delta^4)^2
- \frac{V_{5321}V_{5316}V_{2356}}{V_{2561}V_{2361 }} (\Delta^3\Delta^5)^2
\nonumber \\
&\qquad \qquad \quad - (1-R) \frac{V_{2346}V_{1345}V_{3461}}{V_{1236}V_{1456}} (\Delta^3\Delta^4)^2
\,,
\nonumber \\
&\delta_{2}(F^2) = - 2\frac{V_{2415}V_{3461}V_{4256}}{V_{4561}V_{1256}} (\Delta^2\Delta^4)(\Delta^3\Delta^4) - 2\frac{V_{2356}V_{3461}V_{3125}}{V_{1236}V_{6125}} (\Delta^3\Delta^4)(\Delta^3\Delta^5) \nonumber \\
& \qquad \qquad \quad +2 \frac{V_{1245}V_{2356}}{V_{1256}} (\Delta^2\Delta^4)(\Delta^3\Delta^5)
\,.
\label{6v-2d}
\end{align}
Here, $R$ denotes the ratio of products of volumes,
\begin{align}
R= \frac{V_{1245}V_{2356}V_{3164}}{V_{3145}V_{1256}V_{2364}} \,,
\label{R-def}
\end{align}
which is non-zero only when all three internal lines are turned on.
\paragraph{Three disconnected internal lines}
We deform Figure~\ref{fig:6v(b)} further by turning on the third internal line $\overline{35}$. The result is depicted in Figure~\ref{fig:6v(c)}.
\begin{figure}[htbp]
\centering
\includegraphics[width=7cm]{6v-c.pdf}
\caption{A toric diagram with 6 vertices and three non-intersecting internal lines.}
\label{fig:6v(c)}
\end{figure}
\noindent
The correction terms turn out to be
\begin{align}
&\delta_1(F^2) =
- \frac{1}{1+R} \left(\frac{V_{4256}}{V_{1256}}\frac{V_{2134}}{V_{5134}} V_{2415}(\Delta^2\Delta^4)^2
+ \mbox{(cyclic)} \right)\,,
\nonumber \\
&\delta_{2A}(F^2) = \frac{2}{1+R} \left(\frac{V_{1245}V_{3164}}{V_{3145}}
(\Delta^2\Delta^4)(\Delta^1\Delta^6)
+ \mbox{(cyclic)} \right) \,,
\nonumber \\
&\delta_{2B}(F^2) = -\frac{2R}{1+R}
\left( V_{2416} (\Delta^2\Delta^4)(\Delta^1\Delta^6)
+ \mbox{(cyclic)} \right) \,,
\label{6v-2c}
\end{align}
where $R$ is as defined in \eqref{R-def}
and ``$+{\rm(cyclic)}$" means a sum over the cyclic permutations,
\begin{align}
(123;456) \;\rightarrow\;
(231;564) \;\rightarrow\;
(312;645) \;\rightarrow\;
(123;456) \,.
\end{align}
In the limit where $\overline{35}$ disappears, $R$ vanishes
and $\delta_1$ and $\delta_{2A}$ reproduce \eqref{6v-2b}.
To make the comparison,
aside from reshuffling some indices, we need to use some identities
that hold when $\{v_2, v_3, v_5, v_6\}$ are coplanar.
The new term, $\delta_{2B}$, is visible only if all three internal lines are turned on.
As a further check for \eqref{6v-2c}, we can take the limit
where all three internal lines meet at a point,
as is the case for the example (A.4) of \cite{Amariti:2012tj}.
In that limit, $\delta_{2B}$ vanishes again, not because $R=0$
but because $V_{2416}$ and its cyclic permutations vanish.
For the particular example (A.4) of \cite{Amariti:2012tj},
it turns out that $R=1$
and \eqref{6v-2c} reproduces eq.~(A.5) of \cite{Amariti:2012tj}
including the precise normalization.
\footnote{Caution: there is an overall factor of 4
difference between our normalization and that of \cite{Amariti:2012tj}.}
Note that while \eqref{6v-2c} agrees with eq.~(A.5) of \cite{Amariti:2012tj}
numerically for arbitrary choices of the variables $\{X_{1,2}, Y_{1,2}, Z_{1,2}\}$, our geometric interpretation for the coefficients
of the correction terms differs from one suggested by \cite{Amariti:2012tj}. For instance, the coefficient of the $(\Delta^2 \Delta^4)^2$ term in \eqref{6v-2c} is
\begin{align}
- \frac{V_{4256}V_{2134}V_{2415}}{(1+R)V_{1256}V_{5134}} = -
\frac{V_{4256}V_{2134}V_{2415}V_{2364}}{V_{1245}V_{2356}V_{3164}+V_{3145}V_{1256}V_{2364}}\,.
\label{6v-t1c}
\end{align}
In contrast, eq.~(6.8) of \cite{Amariti:2012tj} suggests an interpretation
of the form
\begin{align}
- \frac{V_a V_b V_c + V_b V_c V_d + V_c V_d V_a+ V_d V_a V_b}{V_e V_f}\,,
\end{align}
which appears quite different from \eqref{6v-t1c}.
\subsection{Generalization}
\paragraph{Some 7-vertex and 8-vertex models}
Conceptually, our strategy to find the correction terms can be applied
to toric diagrams with arbitrary number of vertices.
However, the brute force computation becomes intractable as early as at 7-vertex, even with the aid of a computer.
To collect more evidence for our conjecture while keeping the computational complexity under control, we explored a few non-generic 7-vertex and 8-vertex models. Two such examples are depicted in
Figure~\ref{fig:7v8v}. In all examples we considered,
the correction terms were uniquely determined, in accordance with our conjecture.
\begin{figure}[htbp]
\centering
\includegraphics[width=7cm]{7v8v.pdf}
\caption{Some non-generic toric diagrams with 7 or 8 vertices.}
\label{fig:7v8v}
\end{figure}
We have not been able to derive a more systematic way
to determine the correction terms. In the rest of this subsection,
we sketch some ideas which may prove useful
in future attempts to find new systematic methods.
\paragraph{Flop transition}
Consider a generic toric diagram with $d$ vertices. By ``generic", we
mean that the boundary surface of the convex polytope can be decomposed into triangles such that no two triangles lie on the same plane. It is easy to show that
\begin{align}
\mbox{\#(external edges)} = 3d-6\,, \quad
\mbox{\#(internal edges)} = \frac{(d-3)(d-4)}{2} \,.
\end{align}
Recall that all the correction terms of the geometric free energy
formula were associated to internal lines.
As we deform the toric diagram continuously, the form of the
correction terms remain unchanged until a ``crossing" occurs.
By ``crossing", we mean the crossing of an internal line with an external edge. Whenever a crossing occurs,
a pair of neighboring triangles go through a ``flop" transition as depicted in Figure~\ref{fig:Flop}.
\begin{figure}[htbp]
\centering
\includegraphics[width=9cm]{Flop.pdf}
\caption{A ``flop" transition.}
\label{fig:Flop}
\end{figure}
It seems reasonable to assume that the terms in $F^2$ that are completely independent of the four vertices involved in the flop transition will remain unchanged. At least, this assumption is consistent with
all explicit results we have obtained up to 8-vertex models. The terms that will change can be organized as follows:
\begin{align}
&\mbox{Type 0.}\qquad V_{IJKL} \Delta^I \Delta^J \Delta^K \Delta^L \,,
\nonumber \\
&\mbox{Type 1(a).}\quad (\Delta^I \Delta^K)^2 \,, \quad (\Delta^J \Delta^L)^2 \,,
\nonumber \\
&\mbox{Type 1(b).}\quad (\Delta^I \Delta^A)^2 \,, \quad (\Delta^J \Delta^A)^2 \,, \quad (\Delta^K \Delta^A)^2 \,, \quad (\Delta^L \Delta^A)^2 \,,
\nonumber \\
&\mbox{Type 2(a).}\quad (\Delta^I \Delta^K)(\Delta^A \Delta^B) \,, \quad (\Delta^J \Delta^L)(\Delta^A \Delta^B) \,,
\nonumber \\
&\mbox{Type 2(b).}\quad (\Delta^I \Delta^A)(\Delta^J \Delta^B)\,,\;
(\Delta^J \Delta^A)(\Delta^K \Delta^B)\,,\;
(\Delta^K \Delta^A)(\Delta^L \Delta^B)\,,\;
\nonumber \\
&\qquad \qquad \qquad
(\Delta^L \Delta^A)(\Delta^I \Delta^B)\,,\;
(\Delta^I \Delta^A)(\Delta^K \Delta^B)\,,\;
(\Delta^J \Delta^A)(\Delta^L \Delta^B)\,,
\end{align}
where the the vertices $v_A, v_B$ does not belong to $\{ v_I,v_J,v_K,v_L\}$.
We may take the following approach to determine the coefficients of the correction terms. (1) Assume that we have some value of $C_{IJKL}$ such that $t^3$ and $t^4$ terms vanish. (2) When going through the ``flop", we know how the Type 0 term changes. (3) We could try to determine how other terms should change in order to maintain the vanishing of $t^3$ and $t^4$ terms. Some preliminary studies indicate that, although this approach gives rise to a set of constraints on the unknown coefficients, the constraints are not sufficient by themselves to determine all coefficients completely.
\paragraph{Recursive approach}
In a recursive approach, after finishing the study of toric diagrams
with $d$ vertices, we may add a new ``$(d+1)$-th" vertex and
see how things change:
\begin{align}
v_I{}^i
\quad \rightarrow \quad
\tilde{v}_I{}^i =
\begin{pmatrix}
v_I{}^i
\\ \hline
v_{d+1}^i
\end{pmatrix} \,.
\end{align}
To proceed, we need the GLSM charge matrix for
the new toric diagram whose rank should be $(d+1)-n$.
We will use the following recursive construction:
\begin{align}
Q_a{}^I
\quad \rightarrow \quad
\tilde{Q}_a{}^I =
\begin{pmatrix}
\begin{array}{c|c}
Q_a{}^I & 0
\\ \hline
v_{d+1}^i F_i{}^I & -1
\end{array}
\end{pmatrix} \,.
\end{align}
Generically, the new vertex produces $(d-3)$ extra internal lines.
Since the $t^4$, $t^3$ terms from all the pre-existing vertices
cancel out among themselves, the same cancellation should occur
among the additional leading and correction terms.
\paragraph{Some geometric identities}
We want to see how much information from section \ref{sec:a-max}
can be carried over to the current setup.
Recall from \eqref{LS-id} that
\begin{align}
\sum L^I(x) v_I^i = \frac{x^i}{x^4} \sum_I L^I(x) = x^i S(x) \,,
\end{align}
where $x^i$ is the normalized Reeb vector
and $L^I$ and $S$ are defined in \ref{sec:toreview}.
For $i\neq 4$, the identity can be understood as a consequence of the following relation,
\begin{align}
L^I r_I = \sum_{J \in N_I} c^{IJ} w_{IJ}
\qquad (c^{IJ} = c^{JI}, w_{IJ} = -w_{JI}) \,.
\end{align}
Here, $J\in N_I$ means that vertices $J$ and $I$ are neighbors sharing an external edge.
The explicit form of the coefficients is known
\begin{align}
w_{IJ} = v_I - v_J, \quad c^{IJ} = \frac{V_{IJKL}}{\langle x, v_I,v_J,v_K \rangle \langle x, v_I,v_J,v_L \rangle} \,,
\end{align}
with $J$, $K$ being the vertices of the two triangles meeting over the edge $\overline{IJ}$; see Figure~\ref{fig:C-IJ}.
\begin{figure}[htbp]
\centering
\includegraphics[width=4cm]{C-IJ.pdf}
\caption{The formula for $c^{IJ}$.}
\label{fig:C-IJ}
\end{figure}
In the CY$_3$ setup reviewed in section \ref{sec:a-max},
there was an interesting identity \eqref{lemma}:
\begin{align}
C_{IJK} L^J L^K = \frac{3S}{x^3} + \langle v_I, x , u \rangle
\quad \Longrightarrow \quad
C_{IJK} L^I L^J L^K = 3S^2 \left. \right|_{x^3=1} \,.
\label{LL3}
\end{align}
where the vector $u$ is independent of the vertex label $I$.
We propose that a CY$_4$ analog of \eqref{LL3} may hold, namely,
\begin{align}
c_I \equiv C_{IJKL} L^J L^K L^L = \frac{4S^2}{x^4} + \langle v_I, x , u \rangle
\quad \Longrightarrow \quad
C_{IJKL} L^I L^J L^K L^L = 4 S^3 \left. \right|_{x^4=1}\,.
\label{LL4}
\end{align}
for some ``two-form" $u$.
We content ourselves with verifying the proposal \eqref{LL4} for 5-vertex models, leaving a more general analysis for a future work.
We set $x^4=1$ and define $r_I^i = v_I^i - x^i$ such that
a $(4\times 4)$ determinant can be rewritten as a $(3\times 3)$ determinant
\begin{align}
\langle v_I, v_J, v_K, x \rangle
\equiv
\begin{pmatrix}
v_I^1 & v_I^2 & v_I^3 & 1 \\
v_J^1 & v_J^2 & v_J^3 & 1 \\
v_K^1 & v_K^2 & v_K^3 & 1 \\
x^1 & x^2 & x^3 & 1
\end{pmatrix}
=
\begin{pmatrix}
r_I^1 & r_I^2 & r_I^3 \\
r_J^1 & r_J^2 & r_J^3 \\
r_K^1 & r_K^2 & r_K^3 \\
\end{pmatrix}
\equiv
\langle r_I , r_J, r_K \rangle
\end{align}
Similarly, for a ``two-form" $u$ with vanishing components along the $x^4$ direction, we may write $\langle v_I, x , u \rangle = \langle r_I , u \rangle$. We further abbreviate $\langle r_I , r_J, r_K \rangle$ as $\langle I, J,K \rangle$ in what follows.
After some manipulations, it is possible to show that
\begin{align}
&c_1 = -S(\langle 1,2,4 \rangle L^2 L^4 + \langle 1,3,5 \rangle L^3 L^5 ) \,,
\nonumber \\
&c_2 = -S(\langle 2,3,4 \rangle L^3 L^4 + \langle 2,1,5 \rangle L^1 L^5 ) \,,
\nonumber \\
&c_3 = -S(\langle 3,1,4 \rangle L^1 L^4 + \langle 3,2,5 \rangle L^2 L^5 ) \,.
\end{align}
Combining this fact with a particular choice of basis for $u$,
\begin{align}
u = -\frac{S}{\langle 1,2,3 \rangle} (a_1 r_2 \wedge r_3 + a_2 r_3 \wedge r_1 +a_3 r_1 \wedge r_2) \,,
\end{align}
we obtain an exact expression for $u$ with
\begin{align}
&a_1 = \langle 1,2,4 \rangle L^2 L^4 + \langle 1,3,5 \rangle L^3 L^5 +4S \,,
\nonumber \\
&a_2 = \langle 2,3,4 \rangle L^3 L^4 + \langle 2,1,5 \rangle L^1 L^5 +4S \,,
\nonumber \\
&a_3 = \langle 3,1,4 \rangle L^1 L^4 + \langle 3,2,5 \rangle L^2 L^5 +4S \,.
\end{align}
Another lengthy but straightforward computation verifies the identity $c_I = 4S^2 + \langle r_I, u \rangle$ for the remaining $I=4,5$.
This expression for $u$ is fairly simple and exhibits
the symmetries ($1\rightarrow 2 \rightarrow 3 \rightarrow 1$, $4 \leftrightarrow 5$), but the generalization
to more vertices does not seem obvious.
\section{Field theory \label{sec:qft}}
In this section, we review the field theory computation performed in \cite{Amariti:2012tj}.
We first review the general method of constructing field theory models
and of computing the free energy in the large $N$ limit.
Then we examine a few infinite families of field theories
considered in \cite{Amariti:2012tj}.
By comparing the field theory result and their geometric counterpart,
we verify that all the results of \cite{Amariti:2012tj}
agree perfectly with our main conjecture.
\subsection{Construction of field theory models}
\subsubsection{Lifting algorithm}
We restrict our attention to 3d toric CS theories
that have some 4d ``parent" theory.
In particular, we will take the $L^{a,b,a}$ geometry
for the parent theory.
\begin{figure}[!htb]
\begin{center}
\begin{tikzpicture}
\draw[fill] (0,1) circle (2pt) coordinate (a0);
\draw[fill] (1,1) circle (2pt) coordinate (a1);
\draw[fill] (2,1) circle (2pt) coordinate (a2);
\draw[fill] (3,1) circle (2pt) coordinate (a3);
\draw[fill] (0,0) circle (2pt) coordinate (b0);
\draw[fill] (1,0) circle (2pt) coordinate (b1);
\draw[fill] (2,0) circle (2pt) coordinate (b2);
\draw[fill] (3,0) circle (2pt) coordinate (b3);
\draw[fill] (4,0) circle (2pt) coordinate (b4);
\draw[fill] (5,0) circle (2pt) coordinate (b5);
\draw (a0)--(a3)--(b5)--(b0)--(a0) ;
\node () at ($ (a0) + (0,0.4) $) {0} ;
\node () at ($ (a1) + (0,0.4) $) {1} ;
\node () at ($ (a2) + (0,0.4) $) {$\cdots$} ;
\node () at ($ (a3) + (0,0.4) $) {$a$} ;
\node () at ($ (b0) + (0,-0.4) $) {0} ;
\node () at ($ (b1) + (0,-0.4) $) {1} ;
\node () at ($ (b3) + (0,-0.4) $) {$\dots$} ;
\node () at ($ (b5) + (0,-0.4) $) {$b$} ;
\end{tikzpicture}
\end{center}
\vspace{-16pt}
\caption{Toric diagram for $L^{a,b,a}$. We assume $b\ge a$ without loss of generality.}
\label{fig:Laba}
\end{figure}
\begin{figure}[htb]
\begin{center}
\begin{tikzpicture}[scale=1, every node/.style={transform shape}, arrow head=1.4mm]
\node[circle,draw,fill=black] (1) at (2,0) {} ;
\node[circle,draw,fill=black] (2) at (3,0) {} ;
\node[circle,draw,fill=black] (3) at (4,0) {} ;
\node[circle,draw] (4) at (5,0) {} ;
\node[circle,draw,fill=black] (5) at (6,0) {} ;
\node[circle,draw] (6) at (7,0) {} ;
\node[circle,draw,fill=black] (7) at (8,0) {} ;
\node[circle,draw] (8) at (9,0) {} ;
\node[circle,draw,fill=black] (9) at (10,0) {} ;
\node[circle] (L) at (11,0) {} ;
\tikzset{
rightbi/.style={decoration={
markings,
mark=at position 0.6 with {\arrow{angle 90 new}},
},
postaction=decorate},
leftbi/.style={decoration={
markings,
mark=at position 0.5 with {\arrow{angle 90 new}},
},
postaction=decorate},
adj/.style={decoration={
markings,
mark=at position 0.3 with {\arrow{angle 90 new}},
},
postaction=decorate}
}
\draw[rightbi] (1) to[out=30,in=150] (2) ;
\draw[rightbi] (2) to[out=30,in=150] (3) ;
\draw[rightbi] (3) to[out=30,in=150] (4) ;
\draw[rightbi] (4) to[out=30,in=150] (5) ;
\draw[rightbi] (5) to[out=30,in=150] (6) ;
\draw[rightbi] (6) to[out=30,in=150] (7) ;
\draw[rightbi] (7) to[out=30,in=150] (8) ;
\draw[rightbi] (8) to[out=30,in=150] (9) ;
\draw[rightbi] (9) to[out=30,in=150] (L) ;
\draw[leftbi] (2) to[out=-150,in=-30] (1) ;
\draw[leftbi] (3) to[out=-150,in=-30] (2) ;
\draw[leftbi] (4) to[out=-150,in=-30] (3) ;
\draw[leftbi] (5) to[out=-150,in=-30] (4) ;
\draw[leftbi] (6) to[out=-150,in=-30] (5) ;
\draw[leftbi] (7) to[out=-150,in=-30] (6) ;
\draw[leftbi] (8) to[out=-150,in=-30] (7) ;
\draw[leftbi] (9) to[out=-150,in=-30] (8) ;
\draw[leftbi] (L) to[out=-150,in=-30] (9) ;
\draw[adj] (1) to[out=60,in=0] (2,1) to[out=180,in=120] (1) ;
\draw[adj] (2) to[out=60,in=0] (3,1) to[out=180,in=120] (2) ;
\draw[adj] (3) to[out=60,in=0] (4,1) to[out=180,in=120] (3) ;
\draw [decoration={ brace, mirror, raise=0.5cm }, decorate] (1.west) -- (3.east)
node [pos=0.5,anchor=north,yshift=-0.55cm] {$(b-a)$};
\draw [decoration={ brace, mirror, raise=0.5cm }, decorate] (4.west) -- (9.east)
node [pos=0.5,anchor=north,yshift=-0.55cm] {$2a$};
\end{tikzpicture}
\vspace{-16pt}
\end{center}
\caption{Quiver diagram for $L^{a,b,a}$ model. This diagram is for $a=3$, $b=6$.
This is originally a circular diagram. We cut it and place it on a line,
keeping in mind that
the right end and the left end should be identified.}
\label{quiver}
\end{figure}
We will use an algorithm for uplifting this toric diagram to three dimensions, which correspond to the 3d CS theory.
The uplifting algorithm to be used in this paper
is a special case of a more general method discussed in
\cite{Ueda:2008hx,Imamura:2008qs,Benini:2009qs,Benini:2011cma,Closset:2012ep}.
In the toric diagram, we assign an integer $Q_{\alpha}$ $(\alpha=1,\ldots,a)$ to each vertex on the upper row except the leftmost one. Similarly,
we assign an integer $P_{\beta}$ $(\beta=1,\ldots,b)$ to each vertex
on the lower row except the leftmost one.
We also assign a degeneracy to each vertex. The $\mu$-th vertex on the upper row has degeneracy $_aC_\mu$, and the $\nu$-th vertex on the lower row has degeneracy $_bC_\nu$.
The degenerate points on each vertex move in the ``vertical" direction as follows.
Let us focus on the upper row. The ``elevation" of each of the $_aC_\mu$ degenerate points is equal to the partial sum of $\mu$ elements taken from
the set $\{Q_\alpha\}$.
For example, consider
\begin{align}
Q_\alpha &= (0,1,0,2) \qquad (a=4) \,,
\end{align}
and $\mu =2$ as illustrated in Figure~\ref{fig:uplift}.
There are $_4C_2 = 6$ pairs of $Q_\alpha$. The partial sums are
\begin{align}
&Q_1 +Q_2 = 1 \,, \quad
Q_1 +Q_3 = 0 \,, \quad
Q_1 +Q_4 = 2 \,,
\nonumber \\
&Q_2 +Q_3 = 1 \,, \quad
Q_2 +Q_4 = 3 \,, \quad
Q_3 +Q_4 = 2 \,.
\end{align}
Thus, among the $_4C_2=6$ degenerate points, one stays at the bottom,
two move up one step, two move up two steps, and one moves up three steps.
The same manipulation should be done for all points in the upper row as well as those in the lower row, producing the 3d toric diagram.
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=11cm]{uplift.pdf}
\vspace{-24pt}
\end{center}
\caption{Uplifting degenerate points for $Q_\alpha=(0,1,0,2)$.}
\label{fig:uplift}
\end{figure}
The 3d gauge theory has the same gauge groups and matter fields
as its parent theory. What change the vacuum moduli space from CY$_3$
to CY$_4$ are the CS terms.
To determine the CS levels, we align $Q_\alpha$ and $P_\beta$ in a particular order to define $p_i$ $(i=1,\ldots,a+b)$
\begin{align}
p_i = (P_1, P_2, P_3, Q_1, P_4, Q_2, P_5, Q_3, P_6) \,. \qquad \mbox{($a=3,~b=6$ example)}
\label{eq:order}
\end{align}
and determine the level $k_i$ as the differences in $p_i$,
\begin{align}
k_i=p_i-p_{i-1}
\label{eq:cslevel}
\end{align}
We may reorder the integers $Q_\alpha$ and/or $P_\beta$
but it will not affect the large $N$ free energy \cite{Amariti:2012tj}.
This is consistent with the uplifting algorithm to construct the 3d toric diagram discussed above,
which is clearly independent of the reordering.
\paragraph{Flip symmetry}
By an $SL(3,\mathbb{Z})$ transformation,
the toric diagram of an $L^{a,b,a}$ model in
Figure~\ref{fig:Laba} can be transformed to a flipped form
in Figure~\ref{fig:Labaflip}. The flip reveals a slightly hidden
left-right (in the $x$-direction) symmetry of the toric diagram, which will
give a restriction on the critical value of the Reeb vector components.
The flip symmetry may or may not survive the uplifting procedure
depending on the assignment of $Q_\alpha$, $P_\beta$.
The 3d toric diagram may also have some additional symmetries.
\begin{figure}[!htb]
\begin{center}
\begin{tikzpicture}
\draw[fill] (2,1) circle (2pt) coordinate (a0);
\draw[fill] (3,1) circle (2pt) coordinate (a1);
\draw[fill] (4,1) circle (2pt) coordinate (a2);
\draw[fill] (5,1) circle (2pt) coordinate (a3);
\draw[fill] (0,0) circle (2pt) coordinate (b0);
\draw[fill] (1,0) circle (2pt) coordinate (b1);
\draw[fill] (2,0) circle (2pt) coordinate (b2);
\draw[fill] (3,0) circle (2pt) coordinate (b3);
\draw[fill] (4,0) circle (2pt) coordinate (b4);
\draw[fill] (5,0) circle (2pt) coordinate (b5);
\draw (a0)--(a3)--(b5)--(b0)--(a0) ;
\node () at ($ (a0) + (0,0.4) $) {0} ;
\node () at ($ (a1) + (0,0.4) $) {1} ;
\node () at ($ (a2) + (0,0.4) $) {$\cdots$} ;
\node () at ($ (a3) + (0,0.4) $) {$a$} ;
\node () at ($ (b0) + (0,-0.4) $) {0} ;
\node () at ($ (b1) + (0,-0.4) $) {1} ;
\node () at ($ (b3) + (0,-0.4) $) {$\dots$} ;
\node () at ($ (b5) + (0,-0.4) $) {$b$} ;
\end{tikzpicture}
\end{center}
\vspace{-16pt}
\caption{Flipped toric diagram for $L^{a,b,a}$.}
\label{fig:Labaflip}
\end{figure}
\subsubsection{Brane realization}
We explain how to determine the CS terms for the $L^{a,b,a}$ models
from a brane configuration of the ABJM type \cite{Aharony:2008ug}.
\begin{figure}[htb]
\begin{center}
\begin{tikzpicture}[scale=1, every node/.style={transform shape}, arrow head=1.4mm]
\draw (-2,0)--(2,0) node[right] {D3} ;
\draw (0,1)--(0,-1) node[below] {$k$ D5} ;
\draw[dashed] (-0.8,-0.8)--(0.8,0.8) node[above,xshift=1mm] {NS5};
\draw[thick, densely dotted] (-0.5,0) to[out=90,in=180] (0,0.5);
\draw[thick, densely dotted] (0,-0.5) to[out=0,in=-90] (0.5,0) node[xshift=5mm,yshift=-5mm] {\small string};
\begin{scope}[xshift=5cm,yshift=0.5cm]
\def\px{0.5};
\def\qx{-0.5};
\coordinate (p1) at (\px,0.6);
\coordinate (p1') at (\px,1);
\coordinate (p2) at (\qx,-0.6);
\coordinate (p2') at (\qx,-1);
\draw (-2,0)--(2,0)
\draw (p1')--(p1) (p2)--(p2')
\draw[dashed] (p1)-- ++(0.5,0)
\draw[dashed] (p2)-- ++(-0.5,0)
\draw[thick, densely dotted] (-0.5,0) to[out=90,in=180] (0.5,0.8);
\draw[thick, densely dotted] (-0.5,-0.8) to[out=0,in=-90] (0.5,0)
\draw[decorate, decoration={complete sines, number of sines=10, amplitude=0.5mm}]
(p2)--(p1)
\end{scope}
\begin{scope}[xshift=10cm]
\draw (-2,0)--(2,0) node[right] {D3};
\draw[decorate, decoration={complete sines, number of sines=10, amplitude=0.5mm}]
(-0.6,-1)--(0.6,1) node[above,xshift=1mm] {$(1,k)$};
\node at (0.8,-0.5) {$k$} ;
\node at (-1.2,-0.5) {$-k$} ;
\end{scope}
\end{tikzpicture}
\vspace{-0pt}
\end{center}
\caption{CS term from brane configuration.}
\label{CSfrombrane}
\end{figure}
As illustrated in Figure~\ref{CSfrombrane},
when a NS5-brane and $k$ D5-branes
merge to produce a $(1,k)$ brane,
the string connecting D3 and D5 branes become massive.
As the massive state is integrated out,
a fermion loop generates a CS term.
Due to the relative orientation,
the CS level for the left and right neighboring D3-brane
is $k$ and $-k$, respectively.
\begin{figure}[!htb]
\begin{center}
\begin{tikzpicture}[scale=1, every node/.style={transform shape}, arrow head=1.4mm]
\draw[thick, densely dotted] (0,0) ellipse[x radius=4cm, y radius=1cm];
\draw[fill] (-3.8,0.312) circle (0pt) coordinate (b1);
\draw[fill] (-2,0.866) circle (0pt) coordinate (b2);
\draw[fill] (0,1) circle (0pt) coordinate (b3);
\draw[fill] (2,0.866) circle (0pt) coordinate (a1);
\draw[fill] (3.8,0.312) circle (0pt) coordinate (b4);
\draw[fill] (2.8,-0.714) circle (0pt) coordinate (a2);
\draw[fill] (0.9,-0.974) circle (0pt) coordinate (b5);
\draw[fill] (-0.9,-0.974) circle (0pt) coordinate (a3);
\draw[fill] (-2.8,-0.714) circle (0pt) coordinate (b6);
\draw[thick] ($(b1) + (0,-0.5)$)--($(b1) + (0,0.5)$) node[above] {1};
\draw[thick] ($(b2) + (0,-0.5)$)--($(b2) + (0,0.5)$) node[above] {2};
\draw[thick] ($(b3) + (0,-0.5)$)--($(b3) + (0,0.5)$) node[above] {3};
\draw[thick] ($(b4) + (0,-0.5)$)--($(b4) + (0,0.5)$) node[above] {4};
\draw[thick] ($(b5) + (0,-0.5)$)--($(b5) + (0,0.5)$) node[above] {5};
\draw[thick] ($(b6) + (0,-0.5)$)--($(b6) + (0,0.5)$) node[above] {6};
\draw[thick,densely dashed] ($(a1) + (-0.4,-0.4)$)--($(a1) + (0.4,0.4)$) node[above] {1};
\draw[thick,densely dashed] ($(a2) + (-0.4,-0.4)$)--($(a2) + (0.4,0.4)$) node[above] {2};
\draw[thick,densely dashed] ($(a3) + (-0.4,-0.4)$)--($(a3) + (0.4,0.4)$) node[above] {3};
\begin{scope}[xshift=5cm]
\draw[thick, densely dotted] (0,0.8)--(1,0.8) node[right] {D3 brane} ;
\draw[thick, densely dashed] (0,0)--(1,0) node[right] {$(1,Q_\alpha)$ 5 brane} ;
\draw[thick] (0,-0.8)--(1,-0.8) node[right] {$(1,P_\beta)$ 5 brane} ;
\end{scope}
\end{tikzpicture}
\vspace{-16pt}
\end{center}
\caption{Brane configuration of $L^{a,b,a}$ model ($a=3$, $b=6$).}
\label{Lababrane}
\end{figure}
The brane figuration for the $L^{a,b,a}$ model
is depicted in Figure~\ref{Lababrane}.
Each stack of $N$ D3-branes between two neighboring 5-branes
gives rise to a $U(N)$ gauge group.
The strings connecting two sides of a 5-brane produce (anti-)bifundamental fields. When two consecutive 5-branes are
of the same type $(1,P_\beta)$, the gauge group in the middle hosts an adjoint field as well. Thus the gauge theory can be summarized by the
quiver diagram in Figure~\ref{quiver}. The brane realization
also explains why the CS levels for the gauge theory
are given by \eqref{eq:cslevel}.
\subsubsection{Perfect matching}
Perfect matching maps each vertex of the toric diagram,
including degenerate ones, to a global symmetry of the CS theory.
External perfect matchings, those associated to
non-degenerate external vertices, carry non-vanishing trial R-charges.
For $L^{a,b,a}$ models, the vertices on the upper row of the toric diagram in Figure~\ref{fig:Laba}
correspond to the bi-fundamental and adjoint fields
attached to the right side of white circles in Figure~\ref{quiver}.
The vertices on the lower row correspond to the bi-fundamentals
attached to the right side of black dots.
The detailed map between the degenerate vertices and the matter fields are as follows.
Again, let us focus on the upper row first.
There are $\sum_{\mu=0}^a {}_aC_\mu = 2^a$ vertices in the upper row.
Each vertex corresponds to a global charge.
On the other hand,
there are $a$ white circles in Figure~\ref{quiver},
and each white circle has a pair of bifundamental fields (left-pointing and right-pointing ones)
on the right.
Let us take one bifundamental field from each pair.
There are $2^a$ possible choices. Among those,
there are $_aC_\mu$ ways to choose $\mu$ left-pointing bifundamental fields and $(a-\mu)$ right-pointing bifundamental fields from the $a$ pairs.
The selected bifundamental fields and all the adjoint fields have
a unit charge for a global symmetry related to the $\mu$-th vertex,
and $_aC_\mu$ ways of the selection corresponds to the degeneracy.
Similarly,
on the lower row,
the $\nu$-th vertices corresponds to $_bC_\nu$ global symmetries
for which
$\nu$ left-pointing bifundamental fields
and $(b-\nu)$ right-pointing bifundamental fields
from the $b$ pairs
to the right of the black dots (but no adjoint field) have a unit charge.
There is a slightly different but equivalent explanation.
When we uplift a 2d toric diagram,
we considered the combinations of $\{Q_\alpha\}$.
For the bifundamental fields,
we can also consider the combinations of left-pointing
and right-pointing fields.
For the upper row, pairs of bifundamental fields
(left-pointing and right-pointing ones)
to the right of the white circle in Figure~\ref{quiver} are relevant.
For the degenerate vertices at the $\mu$-th point,
we picked $\mu$ out of $a$ $\{Q_\alpha\}$ charges.
Similarly,
we pick $\mu$ out of $a$ left-pointing bifundamental fields
and $(a-\mu)$ right-pointing bifundamental fields from the $a$ pairs.
Then, the selected bifundamental fields as well as all the adjoint fields
have a unit charge for a global symmetry.
The specified global symmetry in this procedure corresponds to
the shifted vertex by the choice of $\{Q_\alpha\}$ charges.
Even after the shift, some of the vertices are still degenerate.
The residual degeneracy will not affect later discussions,
since the trial R-charges are associated to external, non-degenerate vertices only.
For the lower row,
we do the same procedure for the bifundamentals to the right of the black dots. The only difference from the upper row is that the adjoint fields are not included.
\subsubsection{Computation of free energy}
\label{sec:comp-free-energy}
The method to calculate the large $N$ free energy for a vector-like theory is well explained in, {\it e.g.}, \cite{Herzog:2010hf}.
Here, we only give a minimal summary of the procedure, mainly to establish our notation.
The supersymmetric localization method reduces
a path integral to a finite dimensional integral over the
eigenvalues of some scalar fields.
In the large $N$ limit,
the eigenvalues are described approximately
by a continuous distribution.
In the end, the large $N$ free energy
can be expressed in terms of integrals over the eigenvalue distribution.
\begin{align}
F_\mathrm{CS}^i &=
\frac{N^{3/2}}{2\pi}\int x\rho(x) k_i y_i dx ,\\
F_\mathrm{adj}^i &= \frac{2N^{3/2}}{3} \pi^2 \Delta_i (1-\Delta_i)(2-\Delta_i)\int \rho^2 dx ,\\
F_\mathrm{bi}^{i,j} &= -N^{3/2}\frac{2-\Delta_{ij}^+}{2}b\int \rho^2 dx \left\{ \left(y_i-y_j +\pi \Delta_{ij}^- \right)^2
-\frac{\pi^2}{3}\Delta_{ij}^+(4-\Delta_{ij}^+) \right\} .
\end{align}
Here, $x$ is the real part of the normalized eigenvalue, $y$ is the imaginary part, and $\rho(x)$ is the eigenvalue density.
The first contribution comes from the CS terms of $U(N)_i$ gauge groups, the second from adjoint fields,
and the last from a pair of bifundamental fields.
$\Delta_i$ are the R-charges of adjoint fields, and
$\Delta_{ij}^+$ and $\Delta_{ij}^-$ are the sum and difference of R-charges of a pair of bifundamental fields
between gauge groups $U(N)_i$ and $U(N)_j$.
The free energy for the $L^{a,b,a}$ model is given by
\begin{align}
F_{aba} &= \sum_{i=1}^{a+b} F_\mathrm{CS}^i +\sum_{i=1}^{b-a} F_\mathrm{adj}^i +\sum_{i=1}^{a+b} F_\mathrm{bi}^{i,i+1} ,
\end{align}
where $a+b+1 = 1$ (mod $a+b$) is understood.
Note that this expression only depends on $\delta y_i = y_i -y_{i+1}$;
$\sum_i k_i y_i = \sum_i \delta y_i p_i$ where $p_i$ are ones defined in \eqref{eq:order}.
The final expression can be derived by
minimizing this expression in terms of $\rho$ and $\delta y_i$'s
subject to three constraints:
\begin{align}
\int \rho(x) dx = 1 \,, \quad
\sum_{i} \delta y_i = 0 \,, \quad
\mid \delta y_i +\pi \Delta_i^- \mid \leq \pi \Delta_i^+ \,.
\end{align}
\subsection{Infinite families}
In this subsection, we will reproduce a few infinite series of examples
from \cite{Amariti:2012tj} with slight changes of notations
to facilitate the comparison with other sections in the present paper.
In each example, we begin with the assignment of
$(Q_\alpha, P_\beta)$ and construct the toric diagram
using the uplifting algorithm.
We use the $SL(4,\mathbb{Z})$ freedom to put the toric diagram
in a frame where the symmetries of the diagram become manifest.
We will mostly focus on the $k=1$ case.
General value of $k$ can be reached by taking a $\mathbb{Z}_k$ orbifold of the $k=1$ case.
The goal of this subsection is to verify
that the field theory results from \cite{Amariti:2012tj}
agree with our geometric free energy.
Precisely how the comparison is made, however, requires
some explanation. In all but the simplest examples to be considered,
turning on all possible trial R-charge components
make the field theory computation unwieldily complex.
Fortunately, all the toric diagrams have enough symmetry
to reduce the number of free component of trial R-charge to one.
We will denote the free component by $\Delta$ without any indices.
The precise map between $\Delta$ and the Reeb vector components
can be deduced from MSY volume formulas.
Once the consistency between the field theory
result and the MSY formula is fully verified,
it remains to show that our geometric free energy
also agrees with the MSY formula.
The latter connection is stronger since
we can keep all three components of the Reeb vector
$(b^1, b^2, b^3;b^4=4)$ as free parameters.
\subsubsection{4 vertex models}
We consider the assignment, $Q_\alpha = 0$, $P_\beta = k$.
The CS level is determined by \eqref{eq:cslevel},
\begin{align}
\overrightarrow k = \left(0,\ldots,0 \mid -k,k,\ldots,-k,k \right) .
\end{align}
The 3d toric diagram obtained by the uplifting method is depicted in Figure~\ref{fig:4v}(a). In what follows, we will use the diagram in Figure~\ref{fig:4v}(b) related to the original one by an $SL(4,\mathbb{Z})$ transformation.
\begin{figure}[htb]
\begin{center}
\begin{tikzpicture}[scale=1, every node/.style={transform shape}, arrow head=3mm]
\draw[-stealth new,lightgray, thick] (0,0)--(4,0) node[right] {$x$};
\draw[-stealth new,lightgray, thick] (0,0)--(0,2) node[above] {$y$};
\draw[-stealth new,lightgray, thick] (0,0)--(-1.5,-2) node[below, xshift=-1.5mm] {$z$};
\draw[fill] (0,1) circle (2pt) coordinate (a0) node[xshift=-8mm, yshift=3.5mm] {{\bf 3}{\small $(0,1,0)$}};
\draw[fill] (2,1) circle (2pt) coordinate (a1) node[xshift=7mm, yshift=3.5mm] {{\bf 4}{\small $(a,1,0)$}};
\draw[fill] (0,0) circle (2pt) coordinate (b0) node[xshift=-8mm, yshift=3mm] {{\bf 1}{\small $(0,0,0)$}};
\draw[gray] (3,0) circle (2pt) coordinate (b1p);
\draw[fill] (1.7,-1.733) circle (2pt) coordinate (b1) node[xshift=9mm, yshift=0mm] {{\bf 2}{\small $(b,0,b)$}};
\draw node[yshift=-25mm] {(a)} ;
\node[circle] (b1pa) at (b1p) {} ;
\node[circle] (b1a) at (b1) {} ;
\draw[-latex new, gray, densely dashed] (b1pa)--(b1a) ;
\draw (a0)--(a1)--(b0)--(a0)--(b1)--(a1) (b0)--(b1) ;
\draw[-angle 60 new, line width=2pt] (5,0)--(6,0) ;
\node at (5.5,0.6) {$SL(4,\mathbb{Z})$} ;
\begin{scope}[xshift=8cm]
\draw[-stealth new,lightgray, thick] (0,0)--(4,0) node[right] {$y$};
\draw[-stealth new,lightgray, thick] (0,0)--(0,2) node[above] {$z$};
\draw[-stealth new,lightgray, thick] (0,0)--(-1.5,-2) node[below, xshift=-1.5mm] {$x$};
\draw[fill] (0,1) circle (2pt) coordinate (a0) node[xshift=7mm, yshift=3mm] {{\bf 3}{\small $(0,0,1)$}};
\draw[fill] (-1.3,-0.733) circle (2pt) coordinate (a1) node[xshift=6mm, yshift=-4mm] {{\bf 4}{\small $(a,0,1)$}};
\draw[fill] (0,0) circle (2pt) coordinate (b0) node[above, xshift=7mm] {{\bf 1}{\small $(0,0,0)$}};
\draw[fill] (3,0) circle (2pt) coordinate (b1) node[xshift=5mm,yshift=4mm] {{\bf 2}{\small $(0,b,0)$}};
\draw (a0)--(a1)--(b0)--(a0)--(b1)--(a1) (b0)--(b1) ;
\draw node[yshift=-25mm] {(b)} ;
\end{scope}
\end{tikzpicture}
\vspace{-16pt}
\end{center}
\caption{4 vertex model.}
\label{fig:4v}
\end{figure}
In the field theory computation of the free energy \cite{Amariti:2012tj},
it is possible to turn on all four components of the trial R-charge.
Each components are mapped to external perfect matchings
on the toric diagram. The result, taken from \cite{Amariti:2012tj}, is
\begin{align}
\bar{F}^2_\textrm{ft} = 16 ab \Delta^1 \Delta^2 \Delta^3 \Delta^4 \,.
\end{align}
The subscript ``ft" stands for field theory.
The agreement with our geometric formula is obvious:
$\bar{F}^2_\textrm{ft} = \bar{F}^2_\textrm{geo}$.
The comparison with the MSY formula is also straightforward.
The MSY volume formula gives
\begin{align}
Z_\mathrm{MSY} = \frac{\mathrm{Vol}(S^7)}{\mathrm{Vol}_\mathrm{MSY}(b)}= \frac{a b}{b^1 b^2 \left( b^1 -a b^3 \right) \left( b^2 +b (b^3 - b^4) \right)} \,.
\end{align}
The geometric values for the R-charge components are
\begin{align}
\Delta_\mathrm{MSY}^1 = -\frac{b^2 +b (b^3-b^4)}{2b} \,, \quad
\Delta_\mathrm{MSY}^2 = \frac{b^2}{2b} \,, \quad
\Delta_\mathrm{MSY}^3 = -\frac{b^1 -ab^3}{2a} \,, \quad
\Delta_\mathrm{MSY}^4 = \frac{b^1}{2a} \,.
\end{align}
In terms of the R-charge components, the MSY volume takes the orbifold form
\begin{align}
Z_\mathrm{MSY} &= \frac{1}{16 ab (\Delta^1 \Delta^2 \Delta^3 \Delta^4)_\mathrm{MSY}} \,.
\end{align}
Thus, we find $\bar{F}^2 = Z_\mathrm{MSY}^{-1}$ as expected.
For later convenience, let us illustrate how the flip symmetry
of the 3d toric diagram
reduces free components of the R-charge.
{}The geometric R-charges for those external vertices
exchanged by the flip symmetry should be equated: ($b^4=4$)
\begin{align}
\Delta^1 = \Delta^2 \quad \Longrightarrow \quad 2 b^2 = b (b^3 -4) \,,
\qquad
\Delta^3 = \Delta^4 \quad \Longrightarrow \quad 2 b^1 = a b^3 \, .
\end{align}
Note that the vertices $\bm 3$ and $\bm 4$ are flipped along the $x$-direction
and $\bm 1$ and $\bm 2$ are flipped along the $y$-direction.
Each flip gives information of a corresponding component of the Reeb vector.
Now we can parametrize the volume
in terms of one parameter, say, $b^3 = 4 \Delta$:
\begin{align}
&b^1 = 2 a \Delta \,, \quad
b^2 = 2 b (1-\Delta) \,, \quad
b^3 = 4 \Delta \,, \quad
b^4 = 4 \,, \\
&\Delta_\mathrm{MSY}^1 = \Delta_\mathrm{MSY}^2 = 1-\Delta \,,
\quad
\Delta_\mathrm{MSY}^3 = \Delta_\mathrm{MSY}^4 = \Delta \,, \\
&Z_\mathrm{MSY} = \frac{1}{16 ab \Delta^2 (1-\Delta)^2} \,.
\end{align}
\subsubsection{6 vertex models}
In all 6-vertex and 8-vertex models to be considered below,
we will use the symmetry of the toric diagrams
to reduce the number of free parameters in the Reeb vector to one
from the very beginning.
\paragraph{Family 1}
Consider the $(P,Q)$ charges
\begin{align}
Q_\alpha &= (\underbrace{k,\ldots,k}_a) \,,
\quad P_\beta = (\underbrace{0,\ldots,0}_{b-a},\underbrace{k,\ldots,k}_a)
\end{align}
The CS level is determined by \eqref{eq:cslevel},
\begin{align}
\overrightarrow k = \left(-k,0,\ldots,0 \mid k, 0,\ldots, 0 \right)
\end{align}
The 3d toric diagram, with labels and coordinates of the vertices, is depicted in Figure~\ref{fig:6v1}.
\begin{figure}[!htb]
\begin{center}
\begin{tikzpicture}[scale=1, every node/.style={transform shape}, arrow head=3mm]
\draw[-stealth new,lightgray, thick] (0,0)--(3,0) node[right] {$y$};
\draw[-stealth new,lightgray, thick] (0,0)--(0,4) node[above] {$z$};
\draw[-stealth new,lightgray, thick] (0,0)--(-2.6,-3.466) node[below, xshift=-1.5mm] {$x$};
\draw[fill] (0,0) circle (2pt) coordinate (1) node[xshift=6mm, yshift=-3mm] {{\bf 1}{\small $(0,0,0)$}};
\draw[fill] (0,3) circle (2pt) coordinate (2) node[xshift=10mm, yshift=3mm] {{\bf 2}{\small $(0,0,b-a)$}};
\draw[fill] (-2,-2.66) circle (2pt) coordinate (3) node[xshift=6mm, yshift=-3mm] {{\bf 3}{\small $(a,0,0)$}};
\draw[fill] (-2,0.33) circle (2pt) coordinate (4) node[xshift=-12mm] {{\bf 4}{\small $(a,0,b-a)$}};
\draw[fill] (2,0) circle (2pt) coordinate (5) node[xshift=6mm, yshift=-3mm] {{\bf 5}{\small $(0,1,0)$}};
\draw[fill] (0,-2.66) circle (2pt) coordinate (6) node[xshift=6mm, yshift=-3mm] {{\bf 6}{\small $(a,1,0)$}};
\draw (1)--(2)--(4)--(3)--(1)--(5)--(2) (3)--(6)--(4) (5)--(6) ;
\end{tikzpicture}
\vspace{-16pt}
\end{center}
\caption{6 vertex model, Family 1.}
\label{fig:6v1}
\end{figure}
The field theory computation in \cite{Amariti:2012tj},
with only one free parameter in the trial R-charge turned on,
gave
\begin{align}
\bar{F}^2_\mathrm{ft} = a(b-a)\Delta (1-\Delta)^2 \,.
\end{align}
On the geometry side, the MSY volume formula gives
\begin{align}
Z_\mathrm{MSY} &=
\frac{a (b-a) b^4}{b^1 b^2 b^3 (b^1 - a b^4) ( (b-a) b^2 +b^3 + (a -b)b^4 )} \,.
\end{align}
The geometric R-charges are
\footnote{To avoid clutter, we omit the subscript MSY when
the meaning is clear from the context.}
\begin{align}
\Delta^1 &= -\frac{(b^1 -a b^4) ( (b-a) b^2 +b^3 + (a -b)b^4 )}{a (a-b) b^4} ,\\
\Delta^2 &= \frac{b^3 (b^1 -a b^4)}{a (a-b) b^4} ,\quad
\Delta^3 = \frac{b^1 ( (b-a) b^2 +b^3 + (a -b)b^4 )}{a (a-b) b^4} ,\\
\Delta^4 &= -\frac{b^1 b^3}{a (a-b) b^4} ,\quad
\Delta^5 = -\frac{b^2 (b^1 -a b^4)}{a b^4} ,\quad
\Delta^6 = \frac{b^1 b^2}{a b^4} .
\end{align}
Again, we set $b^4 = 4$ and impose the flip symmetry.
The flip along $x$-direction exchanges
$\bm 1$ and $\bm 3$, $\bm 2$ and $\bm 4$, and $\bm 5$ and $\bm 6$.
The $x$-flip determines the value of $b^1$\,,
\begin{align}
\Delta^1 = \Delta^3 \,, \quad \Delta^2 = \Delta^4 \,,\quad \Delta^5 = \Delta^6 \quad \Longrightarrow \quad b^1 = 2 a .
\end{align}
Similarly, the $z$-flip symmetry solves $b^3$ for other parameters.
\begin{align}
\Delta^1 = \Delta^2 \,, \quad \Delta^3 = \Delta^4
\quad \Longrightarrow \quad \quad b^3 = \frac{1}{2} (b-a) (4-b^2) .
\end{align}
The field theory result and the geometric result can be identified if we relabel $b^2 = 4 \Delta$. Other variables
depend on $\Delta$ as
\begin{align}
& b^1 = 2a \,,\quad
b^2 = 4\Delta \,,\quad
b^3 = 2(b-a)(1-\Delta) \,,\quad
b^4 = 4 \,, \\
& \Delta^1 = \Delta^2 = \Delta^3 = \Delta^4 = \frac{1}{2}(1-\Delta) \,, \quad \Delta^5 = \Delta^6 = 2\Delta \,.
\end{align}
Inserting these into the MSY volume formula, we find
\begin{align}
Z_\mathrm{MSY} =
\frac{1}{16a(b-a)\Delta (1-\Delta)^2 } = (\bar{F}^2_\mathrm{ft})^{-1}\,.
\end{align}
It is straightforward to compare these results with
the main conjecture of section~\ref{sec:geo}. For this particular family,
the toric diagram contains no genuine internal line,
the free energy receives no correction term.
The geometric free energy is
\begin{align}
\bar{F}^2 = \; &
\Delta^1 \Delta^2\Delta^5(\Delta^3+ \Delta^4+ \Delta^6)
+ \Delta^3 \Delta^4\Delta^6(\Delta^1+ \Delta^2+ \Delta^5)
\nonumber \\
&+ \Delta^1 \Delta^2 \Delta^6 (\Delta^3 +\Delta^4) +
\Delta^2 \Delta^3 \Delta^5 (\Delta^4 +\Delta^6) +
\Delta^1 \Delta^4 \Delta^5 (\Delta^3 +\Delta^6)
\,.
\end{align}
Decomposing $\Delta^I$ into mesonic and baryonic variables as in \eqref{decomp-st1}
and integrating out the baryonic ones,
we get $\bar{F}^2 = Z_\mathrm{MSY}^{-1}$ with $b^i = 2s^i$.
\paragraph{Family 2}
We set $b=2a$ for simplicity. The $(P,Q)$ data are
\begin{align}
Q_\alpha = (\underbrace{k,\ldots,k}_a) \,,\quad
P_\beta = (\underbrace{0,\ldots,0}_{a},\underbrace{2k,\ldots,2k}_a) \,.
\end{align}
The CS level is determined by \eqref{eq:cslevel},
\begin{align}
\overrightarrow k = \left(-2k,0,\ldots,0 \mid k,k,-k,k, -k,\ldots,k,-k,k \right) \,.
\end{align}
The 3d toric diagram, with labels and coordinates of the vertices, is depicted in Figure~\ref{fig:6v2}.
\begin{figure}[!htb]
\begin{center}
\includegraphics[width=11cm]{v6f2.pdf}
\vspace{-24pt}
\end{center}
\caption{6 vertex model, Family 2.}
\label{fig:6v2}
\end{figure}
\noindent
The volume and geometric R-charges are given by
\begin{align}
& Z_\mathrm{MSY} = \frac{ 2 a A }{ b^1 B C D E F G} \,, \nonumber \\
& \Delta^1 =
\frac{ B C D E }{a A} \,, \quad
\Delta^2 =
-\frac{ B C b^1 (2 a b^1 +3 a b^4 -b^2 )}{ A } \,, \quad
\Delta^3 =
-\frac{ F G b^1 (a b^1 +a b^4 +b^2 ) }{ A } \,, \nonumber \\
& \Delta^4 =
\frac{C E G (a b^1 +2 a b^4 -2 b^3) }{2 a A } \,, \quad
\Delta^5 =
\frac{B D F (a b^1 +2 a b^4 +2 b^3 ) }{2 a A} \,, \quad
\Delta^6 =
\frac{D E F G}{a A} \,,
\end{align}
where we introduced some short-hand notations,
\begin{align}
& A = a \left(a^2 b_1^3 +5 a^2 b_1^2 b_4 +8 a^2 b_1 b_4^2 +4 a^2 b_4^3-a b_1^2 b_2
-2 a b_1 b_2 b_4 +b_1 b_2^2-3 b_1 b_3^2 -4 b_3^2 b_4 \right) \,, \nonumber \\
& B = (b^2-b^3) \,, \qquad\qquad
C = (b^2+b^3) \,, \qquad\qquad\qquad
D = (a (b^1 + b^4) -b^3) \,, \nonumber \\
& E = (a (b^1 + b^4) +b^3) \,, \quad
F = (a b^1 +2 a b^4 -b^2 -b^3) \,, \quad
G = (a b^1 +2 a b^4 -b^2 +b^3) \,.
\end{align}
The $z$-flip symmetry, which identifies vertices $\bm 4$ and $\bm 5$, demands that $b^3 = 0$. The $y$-flip symmetry,
which follows from the 2d toric diagram of the parent theory, implies
\begin{align}
\Delta^1 = \Delta^6 \quad &\Longrightarrow \qquad b^2 = \frac{a}{2}(b^1+8) ,\\
\Delta^2 = \Delta^3 \quad &\Longrightarrow \qquad b^2 = \frac{a}{2}(b^1+8)
\qquad \mbox{(less trivial)}.
\end{align}
Here and later, less trivial means that there are multiple solutions
to the equation.
However, the requirement that the Reeb vector should lie inside
the toric diagram rules out the extra unphysical solution.
Relabeling $b^1 = -4 \Delta$, we rewrite the Reeb vector and the trial R-charges as
\begin{align}
& b^1 = -4\Delta \,,\quad
b^2 = 2a(2-\Delta) \,,\quad
b^3 = 0 \,,\quad
b^4 = 4 \,, \\
& \Delta^1 = \Delta^6 = \frac{4(1-\Delta)^2}{4-3\Delta} \,, \quad
\Delta^2 = \Delta^3 = 2\Delta \,, \quad
\Delta^4 = \Delta^5 = \frac{2(1-\Delta)(2-\Delta)}{4-3\Delta}\,.
\end{align}
The MSY volume formula gives
\begin{align}
Z_\mathrm{MSY} &=
\frac{4-3\Delta}{32a^2 \Delta (1-\Delta)^2 (2-\Delta)^2 } \,.
\end{align}
This coincides with the inverse of $\bar{F}^2_\mathrm{ft}$ computed from
the field theory \cite{Amariti:2012tj}.
We can test our main conjecture on this example.
The internal lines of the toric diagram give rise to non-trivial corrections. Applying the methods of section~\ref{sec:geo}, we find
\begin{align}
\delta_1 = -\frac{a^2}{2} \left( (\Delta^1 \Delta^3)^2 +(\Delta^2 \Delta^6)^2 \right) \,, \quad
\delta_2 = +a^2 \Delta^1 \Delta^3 \Delta^2 \Delta^6 \,.
\end{align}
Including the correction terms and integrating out
the baryonic variables, we again confirm $\bar{F}^2 = Z_\mathrm{MSY}^{-1}$.
\subsubsection{8 vertex models}
\paragraph{Family 1}
The $(P,Q)$ charges are
\begin{align}
Q_\alpha = (\underbrace{k,\ldots,k}_a) \,,
\quad P_\beta = (0,\underbrace{k,\ldots,k}_{b-2},2k) \,.
\end{align}
The CS level is determined by \eqref{eq:cslevel}:
\begin{align}
\overrightarrow k = \left(-2k,k,0,\ldots,0 \mid 0,\ldots, 0,k \right)
\end{align}
The 3d toric diagram is depicted in Figure~\ref{fig:8v1}.
\begin{figure}[!htb]
\begin{center}
\begin{tikzpicture}[scale=1, every node/.style={transform shape}, arrow head=3mm]
\draw[-stealth new,lightgray, thick] (0,0)--(2,0) node[right] {$y$};
\draw[-stealth new,lightgray, thick] (0,0)--(0,2) node[above] {$z$};
\draw[-stealth new,lightgray, thick] (0,0)--(-3.5,-4.666) node[below, xshift=-1.5mm] {$x$};
\draw[fill] (0,0) circle (2pt) coordinate (1) node[xshift=10mm, yshift=3mm] {{\bf 1}{\small$(0,0,0)$}};
\draw[fill] (-1.5,-0.666) circle (2pt) coordinate (2) node[xshift=-9mm, yshift=3mm] {{\bf 2}{\small$(1,-1,0)$}};
\draw[fill] (0.5,-0.666) circle (2pt) coordinate (3) node[xshift=9mm, yshift=0mm] {{\bf 3}{\small$(1,1,0)$}};
\draw[fill] (-3.5,-3.333) circle (2pt) coordinate (4) node[xshift=-12mm, yshift=3mm] {{\bf 4}{\small$(b-1,-1,1)$}};
\draw[fill] (-1.5,-3.333) circle (2pt) coordinate (5) node[xshift=12mm, yshift=-3mm] {{\bf 5}{\small$(b-1,1,1)$}};
\draw[fill] (-3,-4) circle (2pt) coordinate (6) node[xshift=6mm, yshift=-3mm] {{\bf 6}{\small$(b,0,0)$}};
\draw[fill] (0,1) circle (2pt) coordinate (7) node[xshift=8mm, yshift=3mm] {{\bf 7}{\small$(0,0,1)$}};
\draw[fill] (-2,-1.666) circle (2pt) coordinate (8) node[xshift=9mm, yshift=0mm] {{\bf 8}{\small$(a,0,1)$}};
\draw (1)--(2)--(4)--(6)--(5)--(3)--(1) (1)--(7)--(8)--(6) (2)--(7)--(3) (4)--(8)--(5) ;
\draw[red, densely dotted] (1)--(8) (6)--(7) ;
\end{tikzpicture}
\vspace{-16pt}
\end{center}
\caption{8 vertex model, Family 1.}
\label{fig:8v1}
\end{figure}
We use the $y$-flip symmetry of the toric diagram to set $b^2 = 0$.
We also impose the $x$-flip symmetry:
\begin{align}
\Delta^1 = \Delta^6 \,,\quad \Delta^7 = \Delta^8 \;\; \mbox{(less trivial)}
\quad \Longrightarrow \qquad b^1 = \frac{1}{2} (4 b + (a - b) b^3) \,.\end{align}
Relabeling $b^3 = 4 \Delta$, we rewrite the Reeb vector and the trial R-charges as
\begin{align}
& b^1 = 2(b + (a - b) \Delta) \,,\quad
b^2 = 0 \,,\quad
b^3 = 4 \Delta \,,\quad
b^4 = 4 \,, \nonumber \\
& \Delta^1 = \Delta^6 = \frac{4(1-\Delta)^2}{(2+b)(1-\Delta) +a \Delta } \,,\quad
\Delta^7 = \Delta^8 = 2\Delta \,, \nonumber \\
& \Delta^2 = \Delta^3 = \Delta^4 = \Delta^5 =
-\frac{(1-\Delta)(b(1-\Delta)+a\Delta)}{(2+b)(1-\Delta) +a \Delta }
\,.
\end{align}
The MSY volume formula gives
\begin{align}
Z_\mathrm{MSY} &=
\frac{(2+b)(1-\Delta)+a\Delta}{32 \Delta (1-\Delta)^2 (b(1-\Delta) +a\Delta)^2 } \,.
\label{MSY-81}
\end{align}
It agrees with the field theory computation of \cite{Amariti:2012tj}
with the same parametrization in $\Delta$.
We do not have a general form of the correction terms
for arbitrary 8-vertex models. But, from the number of internal lines
and the symmetries of the toric diagram, we know
that the correction terms have only two independent coefficients.
Demanding that the $t^4$ and $t^3$ terms vanish as we did in section~\ref{sec:geo}, we can determine the correction terms uniquely:
\begin{align}
\delta_1 = -\frac{a}{2} \left( (\Delta^1 \Delta^8)^2 +(\Delta^6 \Delta^7)^2 \right) \,,
\quad
\delta_2 = +a \Delta^1 \Delta^6 \Delta^7 \Delta^8 \,.
\end{align}
Upon eliminating baryonic charges and
imposing the same symmetries for the mesonic charges,
we recover the same result for the volume \eqref{MSY-81}.
\paragraph{Family 2}
The $(P,Q)$ charges are
\begin{align}
Q_\alpha = (\underbrace{0,\ldots,0}_Y,\underbrace{k,\ldots,k}_{a-Y}) \,,
\quad P_\beta = (\underbrace{0,\ldots,0}_{X},\underbrace{k,\ldots,k}_{b-X}) \,.
\end{align}
Here, $X$ and $Y$ are integers satisfying $0<X<b$, $0<Y<a$.
The CS level can be determined by \eqref{eq:cslevel},
but its form depends on the values of $a$, $b$, $X$, $Y$.
\begin{itemize}
\item $b-X > a$
\begin{align}
\overrightarrow k = &\left(-k_1,0,\ldots,0,k_{X+1},0,\ldots,0,\right. \nonumber \\
&\qquad \left. -k_{b-a+1}, k_{b-a+2}, \ldots, -k_{b-a+2Y-1}, k_{b-a+2Y}, 0,\ldots, 0 \right)
\label{eq:8v2cs1}
\end{align}
\item $a \ge b-X > a-Y$
\begin{align}
\overrightarrow k = &\left(-k_1,0,\ldots,0,k_{a-b+2X+2},-k_{a-b+2X+3},
\right. \nonumber \\
&\qquad \left. k_{a-b+2X+4}, \ldots, -k_{b-a+2Y-1}, k_{b-a+2Y}, 0,\ldots, 0 \right)
\label{eq:8v2cs2}
\end{align}
\item $a-Y \ge b-X$
\begin{align}
\overrightarrow k = & \left(-k_1,0,\ldots,0,k_{b-a+2Y+1},-k_{b-a+2Y+2}, \right. \nonumber \\
&\qquad \left.
k_{b-a+2Y+3}, \ldots, -k_{a-b+2X}, k_{a-b+2X+1}, 0,\ldots, 0 \right)
\label{eq:8v2cs3}
\end{align}
\end{itemize}
where $k_i = k$ and the subscript $i$ refers to the positions of
the non-vanishing entries.
\begin{figure}[!htb]
\begin{center}
\begin{tikzpicture}[scale=1, every node/.style={transform shape}, arrow head=3mm]
\draw[-stealth new,lightgray, thick] (0,0)--(5,0) node[right] {$y$};
\draw[-stealth new,lightgray, thick] (0,0)--(0,2) node[above] {$z$};
\draw[-stealth new,lightgray, thick] (0,0)--(-2,-2.666) node[below, xshift=-1.5mm] {$x$};
\draw[fill] (0,0) circle (2pt) coordinate (1) node[xshift=8mm, yshift=3mm] {{\bf 1}{\small$(0,0,0)$}};
\draw[fill] (-1.5,-2) circle (2pt) coordinate (2) node[xshift=8mm, yshift=-3mm] {{\bf 2}{\small$(X,0,0)$}};
\draw[fill] (2.5,-2) circle (2pt) coordinate (3) node[xshift=12mm, yshift=-3mm] {{\bf 3}{\small$(X,b-X,0)$}};
\draw[fill] (4,0) circle (2pt) coordinate (4) node[xshift=10mm, yshift=3mm] {{\bf 4}{\small$(0,b-X,0)$}};
\draw[fill] (0,1) circle (2pt) coordinate (5) node[xshift=8mm, yshift=3mm] {5{\small$(0,0,1)$}};
\draw[fill] (-1,-0.333) circle (2pt) coordinate (6) node[xshift=-8mm, yshift=-3mm] {{\bf 5}{\small$(Y,0,1)$}};
\draw[fill] (2,-0.333) circle (2pt) coordinate (7) node[xshift=-10mm, yshift=-3mm] {{\bf 6}{\small$(Y,a-Y,1)$}};
\draw[fill] (3,1) circle (2pt) coordinate (8) node[xshift=12mm, yshift=3mm] {{\bf 7}{\small$(0,a-Y,1)$}};
\draw (1)--(2)--(3)--(4)--(1)--(5)--(6)--(7)--(8)--(5) (2)--(6) (3)--(7) (4)--(8) ;
\draw[red, densely dotted] (1)--(7) (4)--(6) ;
\draw[blue, densely dotted] (3)--(5) (2)--(8) ;
\end{tikzpicture}
\vspace{-16pt}
\end{center}
\caption{8 vertex model, Family 2.}
\label{fig:8v2}
\end{figure}
The 3d toric diagram is depicted in Figure~\ref{fig:8v2}.
We introduce the parametrization $b^3 = 4\Delta$ from the outset
and impose the $x$- and $y$-flip symmetries:
\begin{align}
\Delta^1 = \Delta^2 \quad &: \qquad b^1 = 2 (X(1-\Delta) +Y\Delta) ,\\
\Delta^1 = \Delta^4 \quad &: \qquad b^2 = 2 ((b-X)(1-\Delta) + (a - Y)\Delta)
\end{align}
The geometric R-charges take a simple form,
\begin{align}
\Delta^1 = \Delta^2 = \Delta^3 = \Delta^4 = 1-\Delta \,, \quad
\Delta^5 = \Delta^6 = \Delta^7 = \Delta^8 = \Delta \,.
\end{align}
The MSY volume formula gives
\begin{align}
Z_\mathrm{MSY} &=
\frac{1}{16 \Delta (1-\Delta) ((1 -\Delta) X + \Delta Y) ((b-X)(1-\Delta) + (a-Y) \Delta ) } \,.
\label{MSY-82}
\end{align}
It agrees with the gauge theory result \cite{Amariti:2012tj} .
The corrections terms are determined by the geometric method as usual.
\begin{align}
\delta_1 &= -\frac{(b - X) X (a - Y) Y}{aX +(b-2X)Y} \left( (\Delta^1 \Delta^7)^2 +(\Delta^4 \Delta^6)^2
+(\Delta^2 \Delta^8)^2 +(\Delta^3 \Delta^5)^2 \right) \,, \\
\delta_2 &=
\frac{2(b - X) X (a - Y) Y}{aX +(b-2X)Y} \left(
\Delta^1 \Delta^7 \Delta^2 \Delta^8
+\Delta^1 \Delta^7 \Delta^4 \Delta^6
+\Delta^2 \Delta^8 \Delta^3 \Delta^5
+\Delta^3 \Delta^5 \Delta^4 \Delta^6
\right)
\nonumber \\ &\qquad +
\begin{cases}
\frac{2 (b - X)^2 Y^2}{a X + (b - 2 X) Y} \left(
\Delta^1 \Delta^7 \Delta^3 \Delta^5
+\Delta^2 \Delta^8 \Delta^4 \Delta^6
\right) \qquad \mbox{for } aX \le bY \,.
\\
\frac{2 X^2 (a - Y)^2}{a X + (b - 2 X) Y} \left(
\Delta^1 \Delta^7 \Delta^3 \Delta^5
+\Delta^2 \Delta^8 \Delta^4 \Delta^6
\right) \qquad \mbox{for } aX > bY \,.
\end{cases}
\end{align}
While there are 10 correction terms altogether,
the symmetries of the toric diagram leaves
only three independent coefficients.
Under the general assumptions explained in section~\ref{sec:geo},
the coefficients are uniquely determined.
Upon eliminating baryonic charges and
imposing the same symmetries for the mesonic charges,
we recover the same volume as \eqref{MSY-82}.
\section{Gravity \label{sec:grav}}
We turn to the last topic of this paper, namely,
the gravity side of the AdS$_4$/CFT$_3$ correspondence.
The geometric free energy discussed earlier
is always a quartic polynomial.
In this section, we raise the possiblity of using
the same quartic polynomial as the prepotential
in the AdS gauged supergravity.
For general toric models, a consistent truncation of
the eleven dimensional supergravity is not available.
We circumvent the difficulty by focusing on
the gauge kinetic terms when the fluctuation of gauge fields are small.
Comparing Kaluza-Klein (KK) gravity and gauged supergravity
descriptions, we find perfect agreement in the mesonic sector
but small discrepancy in the baryonic sector.
\subsection{Kaluza-Klein supergravity}
\paragraph{M-theory}
Our convention for the bosonic part of the eleven dimensional supergravity is
\begin{align}
2\kappa_{11}^2 {\cal L} = * R - {\textstyle \frac{1}{2}} G\wedge *G -
\textstyle{\frac{1}{6}} C \wedge G \wedge G\,,
\label{Lag-11}
\end{align}
where $C$ is the 3-form field and $G=dC$.
The 11-dimensional Planck length is defined
by
\begin{align}
2\kappa_{11}^2 = (2\pi)^8 l_{11}^9 \,.
\end{align}
The Einstein equation is given by
\begin{align}
R_{MN} = \frac{1}{2\cdot 3!} G_{MPQR}G_N{}^{PQR}
- \frac{1}{6} g_{MN} \left( \frac{1}{4!} G_{PQRS} G^{PQRS} \right) \,.
\end{align}
It admits the vacuum AdS$_4\times Y_7$ solution in the form
\begin{align}
\overline{ds}^2 = (L/2)^2 ds_{AdS_4}^2 + L^2 ds_{Y_7}^2 ,
\quad
\bar{G} = 3(L/2)^3 \, \mbox{vol}_{AdS_4} ,
\label{ads4-vac}
\end{align}
where we use the unit normalization for the AdS$_4$ and the $Y_7$ factors,
\begin{align}
AdS_4 : R_{\mu\nu} = -3 g_{\mu\nu} , \quad Y_7 : R_{\alpha\beta} = 6 g_{\alpha\beta}\,,
\end{align}
and ${\rm vol}_{AdS_4}$ denotes the standard volume-form.
The flux quantization condition of M-theory determines the radius $L$ of $Y$:
\begin{align}
\frac{1}{(2\pi l_{11})^6} \int *G = N
\;\;\; \Longrightarrow \;\;\;
L^6 = \frac{(2\pi l_{11})^6 N}{6\mbox{Vol}(Y)} \,.
\end{align}
In what follows, we will abbreviate ${\rm Vol}(Y)$ to $V$
to simplify equations.
\paragraph{Baryonic gauge fields}
We follow \cite{Benvenuti:2006xg} to normalize the baryon charges by
\begin{align}
\label{bch}
B_a\left[\Sigma^I\right] = \frac{2\pi}{V} \int_{\Sigma^I} \omega_a = Q_a^I .
\end{align}
In other words, $\{ \frac{2\pi}{V}\omega_a\}$ form an integral basis of $H^5(Y,\mathbb{R})$. The Kaluza-Klein ansatz for the gauge fields
in the baryonic sector is given by
\begin{align}
G = \bar{G} + 6L^3 (*_4 F^a) \wedge (*_7 \omega_a) \,.
\label{kk-baryon}
\end{align}
The normalization of the fluctuation term is fixed by the requirement that the probe M5-branes wrapping the cycles $\Sigma^I$
are correctly normalized,
\begin{align}
T_{\rm M5} \int_{\Sigma^I\times \mathbb{R}_t} \tilde{C}_6
= Q_a{}^I \int_{\mathbb{R}_t} A^a \,.
\end{align}
Here, $\tilde{C}_6$ is the electromagnetic dual form field locally defined by $d\tilde{C}_6 = *dC_3$, and $A^a$ is the gauge field for the field strength in \eqref{kk-baryon}, $F^a = dA^a$. The tension
of an M5-brane is $T_{\rm M5} = 1/(2\pi)^5 l_{11}^6$.
At the linearized level, the gauge field satisfy the free field equation,
\begin{align}
d*F^a = 0 = dF^a \,,
\end{align}
and does not mix with metric fluctuations.
It is straightforward to compute the gauge kinetic term
in the 4-dimensional KK gravity. It is convenient to pull out overall factors of $L$ and ${\rm Vol}(Y)$, such that
the 4-dimensional Lagrangian is dimensionless.
\begin{align}
2\kappa_4^2 \, \mathcal{L}_{\rm KK} = *(R + 6) -N_{ab} F^a \wedge *F^b + \cdots \,.
\end{align}
The 4-dimensional metric is unit-normalized as before; it satisfies $R_{\mu\nu} = - 3g_{\mu\nu}$ at the vacuum. The 4-dimensional Newton constant is
\begin{align}
\frac{1}{2\kappa_4^2} = \frac{L^9 V}{4(2\pi)^8 l_{11}^9} = \frac{\pi}{2} V \left( \frac{N}{V}\right)^{3/2} \,.
\end{align}
In this convention, the gauge kinetic term, derived from
the 11-dimensional Lagrangian and the KK ansatz, is given by
\begin{align}
N_{ab} = \frac{9}{V} \int \omega_a \wedge *\omega_b \,.
\label{nab-kk}
\end{align}
\paragraph{Mesonic gauge fields}
The correct normalization for the flavor charges is
\begin{eqnarray}
\label{fch}
F_i^I = \frac{2\pi}{V} \int_{\Sigma^I} (*dK_i/12) \;\;\;\; (i=1,2,3,4).
\end{eqnarray}
As a consistency check, note that
\begin{eqnarray}
\label{qcheck}
\Delta^I = \frac{1}{2} b^i F_i^I = \frac{\pi}{12 V} \int_{\Sigma^I} *d K_R
= \frac{\pi}{6 V} \mbox{Vol}(\Sigma^I) \,.
\end{eqnarray}
We are abusing the notations a bit and use $K$ to
denote both a Killing vector $K^\alpha (\partial/\partial x^\alpha)$
and its dual one-form $g_{\alpha\beta}K^\alpha dx^\beta$. In the last step of
(\ref{qcheck}), we used the local $U(1)_R$ fibration description
of the SE manifold $Y$:
\begin{align}
\label{yb1}
ds_Y^2 = (e^0)^2 + ds_B^2,
& \quad
e^0 \equiv \frac{1}{4} d\psi + \sigma,
\quad
K_R = 4 \frac{\partial}{\partial \psi},
\\
\label{yb2}
R_{\mu\nu}^{(B)} = 8 g_{\mu\nu}^{(B)},
& \quad
d\sigma = 2 J_B,
\quad \quad
\mbox{vol}_\Sigma = e^0 \wedge {\textstyle \frac{1}{2}} J_B^2.
\end{align}
The KK ansatz for the mesonic gauge field is slightly involved
but well-known. The metric fluctuation takes the standard form;
the internal part of the metric is deformed by
\begin{align}
g_{\alpha\beta} dx^\alpha dx^\beta
\;\; \rightarrow \;\;
g_{\alpha\beta} (dx^\alpha + K_i^\alpha A^i_\mu dx^\mu )
(dx^\beta + K_j^\beta A^j_\nu dx^\nu ) \,.
\end{align}
This metric fluctuation must be accompanied by
a fluctuation of the 4-form flux \cite{Barnes:2005bm,Barnes:2005bw},
\begin{align}
G = \bar{G} + 2L^3 (*F^i) \wedge (dK_i/12) \,.
\end{align}
The mixing is needed to satisfy the linearized field equation,
\begin{align}
\nabla^{M} G_{M\mu\nu\alpha} &= \nabla^\beta (\delta G_{\beta\mu\nu\alpha}) -
g^{\lambda\sigma} (\delta \Gamma^{\rho}_{\lambda\alpha}) \bar{G}_{\sigma\mu\nu\rho}
= 3 \epsilon_{\mu\nu}{}^{\lambda\sigma} ( - F^i + F^i)_{\lambda\sigma} K_{i\alpha} = 0 \,.
\end{align}
Collecting both contributions, we obtain the kinetic term
for the mesonic gauge fields,
\begin{align}
2\kappa_4^2 \,\mathcal{L}_{\rm KK}|_{\rm mesonic} = -N_{ij} F^i \wedge * F^j \,,
\quad
N_{ij} = \frac{2}{V} \int K_i \wedge *K_j \,.
\label{nij-kk}
\end{align}
\subsection{Gauged supergravity}
We follow the conventions of \cite{de Wit:1984pk, Andrianopoli:1996vr, Craps:1997gp, Louis:2002ny}
for $D=4$, ${\cal N}=2$ gauged supergravity.
\paragraph{Special geometry}
The vector multiplet part of the $D=4$, ${\cal N}=2$ gauged supergravity
is governed by the prepotential ${\cal F}$.
It is a homogeneous function of degree two
in vector-multiplet scalars $X^I$.
\begin{align}
{\cal F}(\lambda X) = \lambda^2 {\cal F}(X).
\end{align}
The derivatives of $F$ are denoted by
\begin{eqnarray}
{\cal F}_I = \partial_I {\cal F}, \quad
{\cal F}_{IJ} \equiv \partial_I\partial_J {\cal F}~, \quad
{\cal F}_{IJK} \equiv \partial_I\partial_J\partial_K {\cal F}\,.
\end{eqnarray}
The K\"ahler potential,
the K\"ahler metric and the Yukawa couplings are given by
\begin{align}
e^{-K} &=
i(\bar{X}^I {\cal F}_I - X^I \bar{{\cal F}}_I) = -2 F_{IJ} X^I \bar{X}^J \,,
\\
e^{-K} g_{i\bar{j}} &= e^{-K} \partial_i \partial_{\bar{j}} K
= 2 D_i X^I D_{\bar{j}} \bar{X}^J F_{IJ}\,,
\\
C_{ijk} &=
D_i X^I D_j X^J D_k X^K {\cal F}_{IJK}\,.
\end{align}
where we defined $F_{IJ} \equiv \mathop{\rm Im}\nolimits {\cal F}_{IJ}$. The following relations hold:
\begin{align}
&F_{IJ} D_i X^J = D_i {\cal F}_J \,,
\qquad
F_{IJ} \bar{X}^I D_i X^J = 0 \,,
\\
&\bar{\partial}_{\bar{i}} D_j \Omega = g_{\bar{i} j} \Omega ~, \quad
\nabla_{i} D_{j} \Omega = i \,e^K C_{ij}{}^{\bar{k}} \bar{D}_{\bar{k}}
\bar{\Omega} \,,
\\
&R_{i\bar{j}k\bar{l}}
= g_{i\bar{j}}g_{k\bar{l}} + g_{i\bar{l}}g_{k\bar{j}}
- e^{2K} C_{ik}{}^{\bar{m}} \bar{C}_{\bar{j}\bar{l}\bar{m}}\,.
\end{align}
\paragraph{Supergravity Lagrangian}
To write down the vector multiplet part of the $D=4$, ${\cal N}=2$ supergravity Lagrangian (see \cite{de Wit:1984pk, Andrianopoli:1996vr, Craps:1997gp, Louis:2002ny} for details), we need to introduce
\begin{eqnarray}
\label{nn}
{\cal N}_{IJ} = \bar{{\cal F}}_{IJ} +
2i \frac{F_{IK} X^K F_{JL} X^L}{F_{MN} X^M X^N} ,
\;\;\;\;\;
N_{IJ} = -\mathop{\rm Im}\nolimits {\cal N}_{IJ},
\;\;\;\;\;
M_{IJ} = \mathop{\rm Re}\nolimits {\cal N}_{IJ}.
\end{eqnarray}
Some basic properties follow immediately.
\begin{align}
\label{nn1}
{\cal N}_{IJ} X^J = {\cal F}_I\,,
& \quad
\overline{{\cal N}}_{IJ} D_i X^J = D_i {\cal F}_I
\\
\label{nn2}
2\, N_{IJ} X^I \bar{X}^J = e^{-K}\,,
& \quad
2\, N_{IJ} D_i X^I D_{\bar{j}}\bar{X}^J = e^{-K} g_{i\bar{j}},
\\
\label{nn3}
N_{IJ} D_i X^I X^J = 0\,,
& \quad
N^{IJ} = 2 e^K (X^I \bar{X}^J + g^{\bar{i} j}D_{\bar{i}} \bar{X}^I D_j X^J).
\end{align}
The bosonic part of the Lagrangian is
\begin{eqnarray}
\label{lag}
{\cal L} = * (R - V) - 2 g_{i\bar{j}} dt^i \wedge * d\bar{t}^{\bar{j}}
- N_{IJ} F^I \wedge \star F^J - M_{IJ} F^I \wedge F^J.
\end{eqnarray}
The scalar potential is determined by some real coefficients $P_I$:
\begin{eqnarray}
\label{spo}
V = \left(N^{IJ} - 8 e^K X^I \bar{X}^J \right) P_I P_J
=
2 e^K \left( g^{\bar{i}j} D_{\bar{i}} \bar{W} D_j W -3 |W|^2 \right),
\;\;\;\;\;
W = P_I X^I.
\end{eqnarray}
The parameters $P_I$ originate from vacuum expectation values of some
hyper-multiplet scalars.
Each solution to $D_i W=0$ gives a supersymmetric AdS vacuum.
We normalize the potential such that $V|_* = -6$,
which amounts to setting the AdS radius to be unity: $R_{\mu\nu}=-3 g_{\mu\nu}$.
The second derivatives of the potential at the vacuum gives
the mass of the scalars. They can be computed using
the special geometry relations
\begin{eqnarray}
\bar{\partial}_{\bar{i}} D_j W = g_{\bar{i} j} W ,
\;
D_i D_j W = i e^K C_{ij}{}^{\bar{k}} \bar{D}_{\bar{k}} \bar{W}
\;\;
\Longrightarrow
\;\;
\bar{\partial}_{\bar{i}} \partial_j V|_* = - 2 g_{\bar{i} j},
\;
\partial_i \partial_j V|_* = 0.
\end{eqnarray}
The mass yields the expected value for the conformal weight of the lowest component of the current superfield:
\begin{eqnarray}
m^2 = \delta(\delta-3) = -2
\;\;\;\;\;
\mbox{or}
\;\;\;\;\;
\delta = 1\,.
\end{eqnarray}
\paragraph{Free energy vs prepotential - I.}
Our proposal for the prepotential is
\begin{align}
\mathcal{F} = i \sqrt{ \bar{F}^2(X)} \,.
\label{pre-pro}
\end{align}
with $\bar{F}^2$ taken from the geometric free energy formula.
We further assume that ${\rm Re}(X^I)$ (``axions") vanishes at the supergravity vacuum and ${\rm Im}(X^I)$ (``dilatons")
is proportional to $\Delta^I$ of the field theory:
\begin{align}
X^I = 0 + i \kappa \Delta^I \,.
\end{align}
For ${\cal N}=4$ or higher supersymmetry, this proposal was proposed earlier and verified to reproduce
the abelian truncation of the gauged supergravity \cite{Lee:2008zzi,sungjay,kll}. Let us review the simplest ${\cal N}=8$ case in which
the $SO(8)$ gauged supergravity is trucated to its $U(1)^4$ subsector.
The consistent truncation of this $U(1)^4$ supergravity from
the eleven dimensional supergravity was performed in \cite{Cvetic:1999xp}.
For simplicity, we focus on the axion-free sector.
The reduction ansatz for the metric is \begin{align}
ds^2_{11} = H^{2/3} ds_4^2 + 4 H^{-1/3} \sum_I X_i^{-1} (d\phi^i + \mu_i^2 A^i/2)^2
\,, \;\;
H = \sum_{i=1}^4 X_i \mu_i^2
\,,\;\;
\sum_i \mu_i^2 = 1 \,,
\end{align}
where we used a normalization equivalent to $L=2$ in \eqref{ads4-vac}.
The reduction ansatz for the 4-form field strength can be found in \cite{Cvetic:1999xp}. It is convenient to parametrize the scalars $X_i$,
which satisfy $X_1X_2X_3X_4=1$, with three scalars $\vec{\varphi} = (\varphi_1,\varphi_2,\varphi_3)$ as $X_i = e^{-\frac{1}{2} \vec{a}_i\cdot \vec{\varphi}}$, where
\begin{align}
\vec{a}_1 = (1,1,1), \quad
\vec{a}_2 = (1,-1,-1), \quad
\vec{a}_3 = (-1,1,-1), \quad
\vec{a}_4 = (-1,-1,1).
\end{align}
Then the resulting four dimensional supergravity Lagrangian reads
\begin{align}
&{\cal L} = *(R-V) -\frac{1}{2} (\partial \vec{\varphi})^2 -\frac{1}{2} \sum_{i=1}^4 e^{\vec{a}_i\cdot \vec{\varphi}} F_i \wedge * F_i \,,
\nonumber
\\
&V= -2 (\cos\varphi_1 +\cos\varphi_2 +\cos\varphi_3)\,.
\label{n8-lag}
\end{align}
Clearly, the vacuum of this potential is at $\vec{\varphi}=0$ or $X_i =1$.
In \cite{sungjay}, it was shown that
the prepotential ${\cal F} = i \sqrt{X_1X_2X_3X_4}$
with the gauge choice $X_1X_2X_3X_4=1$
and the recipe to derive the bosonic Lagrangian \eqref{lag}
exactly reproduces the Lagrangian \eqref{n8-lag}.
The agreement between the consistent truncation
and the gauged supergravity continues to hold
even if the axions are turned on.
The comparison was also extended to the abelian truncation of ${\cal N}=4$ orbifold theories and perfect agreement was found.
\paragraph{Free energy vs prepotential - II.}
Guided by the success for ${\cal N}\ge 4$ theories,
we test the proposal \eqref{pre-pro} for general ${\cal N}=2$ toric models.
If we focus on the computation of the gauge kinetic terms $N_{IJ}$
at the vacuum, we can use the following simplified formula,
\begin{align}
N_{IJ} = \frac{1}{2} \left( \frac{\partial_I \partial_J (F^2)}{(F^2)^{1/2}} - \frac{\partial_I(F^2) \partial_J(F^2)}{(F^2)^{3/2}} \right) \,.
\label{N-simple}
\end{align}
The derivation of this formula goes as follows.
We will take $X^I$ to be purely imaginary from the beginning,
but will leave ${\rm Im}X^I$ undetermined until the very end.
\begin{align}
&{\cal F} = i \sqrt{F^2(X)} \,, \qquad F^2(X) = \frac{1}{24} C_{IJKL} X^I X^J X^K X^L \,,
\nonumber \\
&{\cal F}_{IJ} = i \partial_I \left( \frac{\partial_J(F^2)}{2(F^2)^{1/2}} \right)
= i \left( \frac{1}{2} \frac{\partial_I\partial_J(F^2)}{(F^2)^{1/2}}-\frac{1}{4} \frac{\partial_I(F^2)\partial_J(F^2)}{(F^2)^{3/2}} \right) = i F_{IJ} \,,
\nonumber \\
&{\cal N}_{IJ} = \bar{{\cal F}}_{IJ} + 2i \frac{F_{IK}X^K F_{JL} X^L}{F_{MN} X^M X^N}
= -i \left( F_{IJ} -2 \frac{F_{IK}X^K F_{JL} X^L}{F_{MN} X^M X^N} \right) = -i N_{IJ} \,.
\label{N-step1}
\end{align}
So far, we have used reality conditions only. We can simplify the formula further using the homogeniety of $F^2$.
\begin{align}
&F_{IK} X^K = \frac{1}{2} \frac{X^K \partial_K(\partial_I(F^2))}{(F^2)^{1/2}} -\frac{1}{4} \frac{\partial_I(F^2) X^K \partial_K(F^2)}{(F^2)^{3/2}}
= \frac{3}{2} \frac{\partial_I(F^2)}{(F^2)^{1/2}} - \frac{\partial_I(F^2)}{(F^2)^{1/2}} = \frac{1}{2} \frac{\partial_I(F^2)}{(F^2)^{1/2}} \,,
\nonumber \\
&F_{MN} X^M X^N = \frac{1}{2} \frac{X^M\partial_M(F^2)}{(F^2)^{1/2}} = 2(F^2)^{1/2} \,.
\label{N-step2}
\end{align}
Inserting \eqref{N-step2} into \eqref{N-step1}, we arrive at \eqref{N-simple}.
Let us proceed to examine the value of $N_{IJ}$ at the vacuum.
To compare the result with those of KK supergravity,
we decompose the gauge kinetic coefficients into
the baryonic, mesonic, and the R-symmetry directions.
In the notations of section \ref{sec:geo},
\begin{align}
N_{ij} = F_i{}^I F_j{}^J N_{IJ} \,,
\quad
N_{ab} = Q_a{}^I Q_b{}^J N_{IJ} \,.
\end{align}
A straightforward computation shows that
\begin{align}
&N_{ab} = \left. \frac{m_{ab}}{2(F^2)^{1/2}} \right|_* \,,
\quad
N_{Ra} = 0\,,
\quad
N_{ia} = 0 \,,
\nonumber \\
& N_{ij} = \left. \frac{\partial_i \partial_j F^2}{2(F^2)^{1/2}} \right|_* \,,
\quad
N_{Ri} = 0 \,,
\quad
N_{RR} = \frac{1}{2} \,.
\end{align}
Here $m_{ab}$ is the quadratic function introduces in \eqref{F-st2}.
The decoupling of the R-symmetry component from
all others is as expected \cite{Barnes:2005bm}.
The mesonic coefficients $N_{ij}$ mathches precisely with
those obtained from the KK supergravity \eqref{nij-kk}
as can be proved by identities for toric geometry \cite{Martelli:2005tp}.
As for the baryonic ones, we do not have general formula to relate
$m_{ab}$ and the KK formula \eqref{F-st2}. However, in all examples
we have tested, the two results differ by an overall constant.
\begin{align}
N_{ab}(\mbox{KK}) = \frac{3}{4}N_{ab}(\mbox{prepotential}) \,.
\end{align}
This discrepancy does not lead to an immediate contradiction.
Our proposal for the prepotential was carried over from previous work
for ${\cal N} \ge 4$ theories, but there was no a priori reason
for its validity for general ${\cal N}=2$ theories.
It would be still desirable to gain further insight on the close resemblance between the free energy and the prepotential.
Since the free energy is obtained by a localization computation on the CFT$_3$ side, it might be a good to apply the localization technique in the AdS$_4$ supergravity. In a recent work \cite{Dabholkar:2014wpa}, a localization computation for supergravity was performed for ${\cal N} \ge 3$ AdS$_4$/CFT$_3$ models, which made use of a square-root prepotential
originally proposed in \cite{Gauntlett:2009zw}.
It would be interesting to apply the ideas of \cite{Dabholkar:2014wpa}
to the toric models considered in this paper.
\vskip 1cm
\acknowledgments
SL thanks Seok Kim and Sungjay Lee for collaborations on
a closely related unpublished work in 2007-2008, and
Kevin Goldstein, Yuji Tachikawa, and Sandip Trivedi for helpful discussions over the same period.
We thank Raju Roychowdhury for collaboration at an early stage of this work. This work was supported in part by the National Research Foundation of Korea (NRF) Grants 2012R1A1B3001085 and 2012R1A2A2A02046739.
The work of DY is supported in part by Perimeter Institute for Theoretical Physics. Research at Perimeter Institute is supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Economic Development and Innovation.
\newpage
|
\section{Introduction\label{intro}}
Concepts are envisaged as the structural units of human thought. Understanding the nature of these units can lead to a first principles basis for a foundational theory of cognition, especially for the study of decision making, problem solving, and communication~\cite{r1973,g2000}. Concepts can be used to represent very different forms of meaning, ranging from concrete objects (e.g., ``this cup of coffee in front of me'') to abstract forms of experience (e.g., ``absolute free will'').
Although concepts have been conceived to represent classes of similar items, it is generally accepted that
they do not have a fixed representational structure~\cite{f1998}.
Concretely, this
view
is based on three conceptual phenomena. The first is `vagueness', i.e. concepts have no sharp boundaries~\cite{h2007}. The second is `contextuality', i.e. concepts acquire meaning depending on the situation they are elicited~\cite{r1976}. These two phenomena have been accounted for, at least to some extent, by fuzzy set theory and classical probability theory \cite{z1989}. The third phenomenon corresponds to `concept combination', i.e. how
the meaning of combined concepts
relates to the meaning of their constituents. This phenomenon has been vastly investigated by psychologists and cognitive scientists~\cite{r1995,kp1995,h1997}. Experiments in concept combination
traditionally request participants to estimate
typicality or membership of certain items of concepts and their combinations. The goal is to obtain a model where
typicality or membership of the combined concept
is related to typicality or memberships of the constituent concepts. It has been shown that experimental estimates of
typicalities \cite{os1981} and of memberships
\cite{a2009a} cannot be modeled respectively within fuzzy logics and classical probability.
In particular, given two concepts $A$ and $B$, and an item $x$, the membership of $x$ with respect to a combined concept `$A$ and $B$' (`$A$ or $B$') is very often larger (smaller) than the membership of $x$ with respect to both $A$ and $B$. This contradicts fuzzy and classical probabilistic
rules~\cite{os1981,h1988a,h1988b}.
In this paper we support the view that a novel theoretical framework to cope with the problems of concept combination is needed. In particular, we have developed a modeling approach for concept combination that uses the mathematical formalism of quantum theory \cite{a2009a,ag2005a,ag2005b,as2011,ags2013,s2014b,asv2014b}. This approach enables faithful representation of a large amount of data collected on conjunctions and disjunctions of two concepts (Sect. \ref{conjdisj}) and, more recently, conjunctions and negations (Sect. \ref{CC-Neg}). Starting from the success of our quantum-theoretic approach, we formulate in Sect. \ref{nature} an explanatory hypothesis, according to which two types of reasoning simultaneously occur in a human decision making (conceptual membership estimation, human probability judgement, gamble preference under uncertainty, etc.) and, more generally, in a cognitive process. The first type is `emergent reasoning', the second is `logical reasoning'. This hypothesis is sustained and justified in Sects. \ref{conjdisj} and \ref{CC-Neg} by considering relevant highlights in concept theory. Finally, we investigate in Sect. \ref{aristotle} the implications of our explanatory hypothesis on the role played by the observed deviations from classicality in human decision making, and elaborate some important insights into the origin of logic and the nature of human reasoning.
\section{An explanatory hypothesis on the nature of human reasoning\label{nature}}
We elaborate in this section our explanation for the
appearance of genuine quantum structures in cognitive processes, which has matured during our research on conceptual combination and human decision making. This explanation reveals very stable patterns of human
reasoning, enlightening
at the same time
some fundamental traits of its deepest nature.
That the fundamental nature of human reasoning should include probabilistic aspects was already known to cognitive psychologists at the beginning of the seventies, when Rosch's experiments confirmed that the combination of natural concepts in human thought exhibits `graded typicality', or `vagueness'. However, one believed that the observed probabilities could be cast into the Kolmogorovian
structures of classical probability theory, thus revealing an underlying classical logical behaviour, possibly extended to include fuzzy set logic and its basic connectives. This was the `admirable illusion' of cognitive psychology in the last century.
This `human rational behaviour myth' was put at stake by different discoveries in concept theory. Osherson and Smith observed that human subjects estimate some items to be more typical examples of the concept conjunction `$A$ and $B$' than of $A$ and $B$, separately \cite{os1981}. Hampton observed that, for several items, the `membership weight', that is, the degree of memberhip, of the item with respect to the conjunction `$A$ and $B$' is generally higher than its membership weight with respect to $A$ or $B$, while the membership weight with respect to the disjunction `$A$ or $B$' is generally lower than the membership weight with respect to $A$ or $B$ \cite{h1988a,h1988b}. In the same period, the so-called `Tversky and Khaneman program' revealed the `disjunction effect' and the `conjunction fallacy' in human decision making \cite{tk1983,ts1992}, which has a counterpart in the `Allais', `Ellsberg' and `Machina paradoxes' of behavioural economics \cite{e1961,m2009}. The general attitude was then to consider these deviations from classicality as `fallacies', as `effects', as `paradoxes', or as `contradictions'. In other words, traditional approaches to cognition interpreted such deviations as `biases of classical logical reasoning'.
A different approach to cognitive psychology, initiated two decades ago, has meanwhile matured into a new domain of research, called `quantum cognition'. Its main feature is the use of the mathematical formalism of quantum theory as modeling tool for these cognitive situations where traditional classically based approaches fail. Important results have been obtained in the quantum modeling of the above decision making situations. Specifically, we have investigated the dynamics of concepts and how they combine as our contribution to quantum cognition, and were able to represent a huge amount of data collected in different experiments on conceptual combinations in a
quantum-mechanical framework.
Our quantum-theoretic approach on concepts and their combinations leads us to formulate a specific hypothesis on the mechanisms that underlie human
reasoning, not only when it combines concepts to form sentences and texts, or in a decision process, but, more generally, in any cognitive process.
According to our explanatory hypothesis, human reasoning is a specifically structured superposition of two processes, a `logical reasoning' and a `emergent reasoning'. The former
`logical reasoning' combines cognitive entities, such as concepts,
combinations of concepts, or propositions, by applying the rules of logic, though
generally in a probabilistic
way. The latter
`emergent reasoning' enables formation of combined cognitive entities as
newly emerging
entities, in the case of concepts, new concepts, in the case of propositions, new propositions,
carrying new meaning, linked to the meaning of the constituent cognitive entities, but with a linkage not defined by the algebra of logic.
The two mechanisms act simultaneously and in superposition in human thought
during a reasoning process, the first one is guided by an algebra of `logic', the second one
follows a mechanism of `emergence'. In this perspective, human reasoning can be mathematically formalized in a two-sector Fock space, where
the states of conceptual entities are represented
by unit vectors of this Fock space. More specifically, `sector 1 of Fock space', that is, an individual Hilbert space, models `conceptual emergence', hence the combination of two concepts is represented by a superposition vector of
the vectors representing the component concepts in this Hilbert space, allowing `quantum interference' between conceptual entities
to play a role in the process of emergence. `Sector 2 of Fock space', that is, a tensor product of two
versions of this Hilbert
space, models a conceptual combination from the combining concepts by requiring
the rules of logic
for the logical connective used for the combining, i.e. conjunction or disjunction,
to be satisfied in a probabilistic setting. This quantum-theoretic modeling suggested us to call `quantum conceptual thought' the process occurring in sector 1 of Fock space, `quantum logical thought' the process occurring in sector 2. \\
The relative importance of emergence
or logic
in a specific cognitive process is measured by the `degree of participation' of sectors 1 and 2, as it will be clear in the next sections.
The abundance of evidence of deviations from classical logical reasoning in concrete human decisions (paradoxes, fallacies, effects, contradictions),
together with our results, led us to draw the conclusion that emergence constitutes the dominant dynamics of human reasoning, while logic is only a secondary form of dynamics. Hence, with respect to the aforementioned deviations from classicality, we
put forward the view that what has been called a fallacy, an effect, a deviation, or a contradiction, is a consequence of the dominant dynamics and its nature is emergence, while what has been considered as a default to deviate from, namely classical logical reasoning, is a consequence of a secondary form
of dynamics, its nature being logic. This claim will be sustained and justified in the following sections, where we will apply our quantum-mechanical framework to model specific conceptual combinations and the respective experimental data.
We also identify aspects of the history and nature of human though explaining the value with respect to the evolution of the human species of the presence of the two types of reasoning, logical reasoning with maximum value for situations about macroscopical physical entities and their dynamics and interactions, and emergent reasoning more proper for situations made of conceptual entities and their interactions, e.g. like they arise from both human minds during a conversation.
\section{Modeling conceptual conjunction and disjunction in Fock space\label{conjdisj}}
We present here our quantum modeling approach in Fock space for the conjunction and the disjunction of two concepts.
In the case of two combining entities, a Fock space $\mathcal F$ consists of two sectors: `sector 1' is a Hilbert space $\cal H$, while `sector 2' is a tensor product $\cal H \otimes \cal H$
of two isomorphic versions of $\cal H$.
Let us now consider the membership weights of items of concepts and their conjunctions/disjunctions measured by Hampton \cite{h1988a,h1988b}. He identified systematic deviations from classical set (fuzzy set) conjunctions/disjunctions, an effect known as `overextension' or `underextension'.
Let us
firstly consider conjunctions. It can be shown that a large part of Hampton's data cannot be modeled in a classical probability space satisfying the axioms of Kolmogorov. For example, Hampton estimated the membership weight of the item {\it Mint} with respect to the concepts {\it Food}, {\it Plant} and their conjunction {\it Food And Plant}, finding $\mu_{Mint}(Food)=0.87$, $\mu_{Mint}(Plant)=0.81$, $\mu_{Mint}(Food \ And \ Plant)=0.9$, respectively
\cite{h1988a}. Thus, the item \emph{Mint} presents overextension with respect to the conjunction \emph{Food And Plant} of the concepts \emph{Food} and \emph{Plant}, and no classical probability representation exists for these data.
Let us now come to disjunctions. Also in this case, a large part of Hampton's data cannot be modeled in a classical Kolmogorovian probability space. For example, Hampton estimated the membership weight of the item
{\it Sunglasses} with respect to the concepts
{\it Sportswear}, {\it Sports Equipment} and their disjunction
{\it Sportswear Or Sports Equipment}, finding
$\mu_{Sunglasses}(Sportswear)=0.4$, $\mu_{Sunglasses}(Sports \ Equipment)=0.2$, $\mu_{Sunglasses}(Sportswear \ Or \ Sports \ Equipment)=0.1$. Thus, the item
\emph{Sunglasses} presents underextension with respect to the disjunction \emph{Sportswear Or Sports Equipment} of the concepts \emph{Sportswear} and \emph{Sports Equipment}, and no classical probability representation exists for these data
\cite{h1988b}.
It can be proved that a quantum probability model in Fock space exists for Hampton's data
on conjunction and disjunction \cite{a2009a,ags2013}. Let us again start with the conjunction of two concepts. Let $x$ be an item and let $\mu(A)$, $\mu(B)$ and $\mu(A \ {\rm and} \ B)$ be the membership weights of $x$ with respect to the concepts $A$, $B$ and `$A \ \textrm{and} \ B$' respectively. Let ${\cal F}={\cal H} \oplus ({\cal H} \otimes {\cal H})$ be the Fock space where we represent the conceptual entities. The states of the concepts $A$, $B$ and $`A \ \textrm{and} \ B'$ are represented by the unit vectors $|A\rangle, |B\rangle \in {\cal H}$ and $|A \ \textrm{and} \ B\rangle \in {\cal F}$, respectively, where
\begin{eqnarray}
|A \ \textrm{and} \ B\rangle=m e^{i\lambda}|A\rangle\otimes|B\rangle+ne^{i\nu}{1\over \sqrt{2}}(|A\rangle+|B\rangle)
\end{eqnarray}
The superposition vector ${1 \over \sqrt{2}}(|A\rangle+|B\rangle)$ describes `$A$ and $B$' as a new emergent concept, while the product vector $|A\rangle\otimes|B\rangle$ describes `$A$ and $B$' in terms of concepts $A$ and $B$. The positive numbers $m$ and $n$ are such that $m^{2}+n^{2}=1$. The decision measurement of a subject who estimates the membership of the item $x$ with respect to the concept `$A \ \textrm{and} \ B$' is represented by the orthogonal projection operator $M\oplus (M \otimes M)$ on ${\cal F}$, where $M$ is an orthogonal projection operator on ${\cal H}$. Hence, the membership weight of $x$ with respect to `$A \ \textrm{and} \ B$' is given by
\begin{eqnarray} \label{AND}
\mu(A \ \textrm{and} \ B)&=&\langle A \ \textrm{and} \ B|M \oplus (M \otimes M)|A \ \textrm{and} \ B \rangle \nonumber \\
&=&m^2\mu(A)\mu(B)+n^2 \left ( {\mu(A)+\mu(B) \over 2}+\Re\langle A|M|B\rangle \right )
\end{eqnarray}
where $\mu(A)=\langle A|M|A\rangle$ and $\mu(B)=\langle B|M|B\rangle$. The term $\Re\langle A|M|B\rangle$ is the `interference term' of quantum theory. A solution of Eq. (\ref{AND}) exists where this interference term given by
\begin{equation}
\Re\langle A|M|B\rangle=
\left\{
\begin{array}{ccc}
\sqrt{1-\mu(A)}\sqrt{1-\mu(B)}\cos\theta & & {\rm if} \ \mu(A)+\mu(B)>1 \\
\sqrt{\mu(A)}\sqrt{\mu(B)}\cos\theta & & {\rm if} \ \mu(A)+\mu(B)\le 1
\end{array}
\right.
\end{equation}
($\theta$ is the `interference angle'). Coming to the example above, namely, the item {\it Mint} with respect to {\it Food}, {\it Plant} and {\it Food And Plant}, we have that Eq. (\ref{AND}) is satisfied with $m^2=0.3$, $n^2=0.7$ and $\theta=50.21^{\circ}$ (cfr., Ref. \cite{ags2013}, Sect. 3).
Keeping in mind the explanation we have given in Sect. \ref{nature}, we interpret this result on conceptual conjunction as follows. Whenever a subject is asked to estimate whether a given item $x$ belongs to the vague concepts $A$, $B$, $`A \ {\rm and} \ B'$, two mechanisms act simultaneously and in superposition in the subject's thought. A `quantum logical thought', which is a probabilistic version of the classical logical reasoning, where the subject considers two copies of item $x$ and estimates whether the first copy belongs to $A$ and the second copy of $x$ belongs to $B$,
and further the probabilistic version of the conjunction is applied to both estimates. But also a `quantum conceptual thought' acts, where the subject estimates whether the item $x$ belongs to the newly emergent concept `$A \ {\rm and} \ B$'.
The place whether these superposed processes can be suitably structured is Fock space. Sector 1 hosts the latter process, while sector 2 hosts the former, while the weights $m^2$ and $n^2$ measure the `degree of participation' of sectors 2 and 1, respectively, in the case of conjunction.
In the case of {\it Mint}, subjects consider {\it Mint} to be more strongly a member of the concept {\it Food And Plant}, than they consider it to be a member of {\it Food} or of {\it Plant}. This is an effect due to a strong presence of quantum conceptual thought, the newly formed concept {\it Food And Plant} being found to be a better fitting category for {\it Mint} than the original concepts {\it Food} or {\it Plant}.
And indeed, in the case of {\it Mint}, considering the values of $n^2$ and $m^2$, the combination process mainly occurs in sector 1 of Fock space, which means that emergence prevails over logic.
Consider instead a situation where logical aspects are prevalent over emergent aspects in our quantum-theoretic modeling. This situation is an example of a `borderline contradiction'. Suppose that a large sample of human subjects is asked to estimate the truth values of the sentences ``John is tall'', ``John is not tall'' and ``John is tall and not tall'', for a given subject John showed to the eyes of the subjects. And suppose that the fractions of positive answers are $0.01$, $0.95$ and $ 0.15$, respectively \cite{ap2011}. This `borderline case' is clearly problematical from a classical logical perspective, and can be modeled in terms of overextension. Indeed,
let us denote by $\mu(A)$, $\mu(A')$ and $\mu (A \ {\rm and} \ A')$ the probabilities that the sentences ``John is tall'', ``John is not tall'' and ``John is tall and not tall'' are true, and interpret them as membership weights of the item {\it John} with respect to the concepts {\it Tall}, {\it Not Tall} and {\it Tall And Not Tall}, respectively. Then Eq. ({\ref{AND}}) is solved for $m^{2}=0.77$, $n^{2}=0.23$ and $\theta=0^{\circ}$ \cite{s2014a}. The explanation of this behaviour is that the reasoning process of the subject mainly occurs in sector 2 of Fock space, hence
logical reasoning is dominant, although emergent reasoning is also present, and it is its presence which evoked the name `contradiction' for this situation.
Let us now come to the disjunction of two concepts. Let $x$ be an item and let $\mu(A)$, $\mu(B)$ and $\mu(A \ {\rm or} \ B)$ be the membership weights of $x$ with respect to the concepts $A$, $B$ and `$A \ \textrm{or} \ B$', respectively. Let ${\cal F}={\cal H} \oplus ({\cal H} \otimes {\cal H})$ be the Fock space where we represent the conceptual entities. The concepts $A$, $B$ and `$A \ \textrm{or} \ B$' are represented by the unit vectors $|A\rangle, |B\rangle \in {\cal H}$ and $|A \ \textrm{or} \ B\rangle \in {\cal F}$, respectively, where
\begin{eqnarray}
|A \ \textrm{or} \ B \rangle=m e^{i\lambda}|A\rangle\otimes|B\rangle+ne^{i\nu}{1\over \sqrt{2}}(|A\rangle+|B\rangle)
\end{eqnarray}
The superposition vector ${1 \over \sqrt{2}}(|A\rangle+|B\rangle)$ describes `$A$ or $B$' as a new emergent concept, while the product vector $|A\rangle\otimes|B\rangle$ describes `$A$ or $B$' in terms of concepts $A$ and $B$. The positive numbers $m$ and $n$ are such that $m^{2}+n^{2}=1$, and they estimate the `degree of participation' of sectors 2 and 1, respectively, in the disjunction case. The decision measurement of a subject who estimates the membership of the item $x$ with respect to the concept `$A \ \textrm{or} \ B$' is represented by the orthogonal projection operator $M \oplus ( M \otimes \mathbbmss{1}+\mathbbmss{1}\otimes M - M \otimes M)$ on ${\cal F}$, where $M$ has been introduced above. We observe that
\begin{equation}
M \otimes \mathbbmss{1}+\mathbbmss{1}\otimes M - M \otimes M= \mathbbmss{1}-
(\mathbbmss{1}-M)\otimes(\mathbbmss{1}-M)
\end{equation}
that is, in the transition from conjunction to disjunction we have applied de Morgan's laws of logic in sector 2 of Fock space. The membership weight of $x$ with respect to `$A \ \textrm{or} \ B$' is given by
\begin{eqnarray} \label{OR}
\mu(A \ \textrm{or} \ B)=\langle A \ \textrm{or} \ B | M \oplus (M \otimes \mathbbmss{1}+\mathbbmss{1}\otimes M - M \otimes M)
|A \ \textrm{or} \ B \rangle \nonumber \\
m^2 \left (\mu(A)+\mu(B)-\mu(A)\mu(B) \right )+n^2 \left ( {\mu(A)+\mu(B) \over 2}+\Re\langle A|M|B\rangle \right )
\end{eqnarray}
where $\mu(A)=\langle A|M|A\rangle$ and $\mu(B)=\langle B|M|B\rangle$. The term $\Re\langle A|M|B\rangle$ is the interference term. A solution of Eq. (\ref{OR}) exists where this interference term given by
\begin{equation}
\Re\langle A|M|B\rangle=
\left\{
\begin{array}{ccc}
\sqrt{1-\mu(A)}\sqrt{1-\mu(B)}\cos\theta & & {\rm if} \ \mu(A)+\mu(B)>1 \\
\sqrt{\mu(A)}\sqrt{\mu(B)}\cos\theta & & {\rm if} \ \mu(A)+\mu(B)\le 1
\end{array}
\right.
\end{equation}
Coming to the example above, namely, the item
{\it Sunglasses} with respect to {\it Sportswear}, {\it Sports Equipment} and {\it Sportswear Or Sports Equipment}, we have that Eq. (\ref{OR}) is satisfied with $m^2=0.03$, $n^2=0.97$ and $\theta=155.00^{\circ}$.
Keeping in mind the explanation we have given in Sect. \ref{nature}, we interpret this result on conceptual disjunction as follows. Whenever a subject is asked to estimate whether a given item $x$ belongs to the vague concepts $A$, $B$, $`A \ {\rm or} \ B'$, two mechanisms act simultaneously and in superposition in the subject's thought. A `quantum logical thought', which is a probabilistic version of the classical logical reasoning, where the subject considers two copies of item $x$ and estimates whether the first copy belongs to $A$ or the second copy of $x$ belongs to $B$, and further the probabilistic version of the disjunction is applied to both estimates. And also a `quantum conceptual thought' acts, where the subject estimates whether the item $x$ belongs to the newly emergent concept `$A \ {\rm or} \ B$'. The place whether these superposed processes are structured is again Fock space. Sector 1 hosts the latter process, while sector 2 hosts the former, while the weights $m^2$ and $n^2$ measure the `degree of participation' of sectors 2 and 1, respectively, in the case of disjunction.
In the case of {\it Sunglasses}, subjects consider {\it Sunglasses} to be less strongly a member of the concept {\it Sportswear Or Sports Equipment}, than they consider it to be a member of {\it Sportswear} or of {\it Sports Equipment}. This is an effect due to a strong presence of quantum conceptual thought, the newly formed concept {\it Sportswear Or Sports Equipment} being found to be a less well fitting category for {\it Sunglasses} than the original concepts {\it Sportswear} or {\it Sports Equipment}.
And indeed, in the case of {\it Sunglasses}, considering the values of $n^2$ and $m^2$, the combination process mainly occurs in sector 1 of Fock space, which means that emergence aspects prevails over logical aspects in the reasoning process.
\section{Extending Fock space modeling to conceptual negation\label{CC-Neg}}
The first studies on the negation of natural concepts were also performed by Hampton \cite{h1997}. He tested membership weights on conceptual conjunctions of the form {\it Tools Which Are Not Weapons} in experiments on human subjects,
finding overextension
and deviations from
Boolean behaviour in the negation. More recently, we have performed a more general cognitive test inquiring into the membership weights of items with respect to conjunctions of the form {\it Fruits And Vegetables}, {\it Fruits And Not Vegetables}, {\it Not Fruits And Vegetables} and {\it Not Fruits And Not Vegetables} \cite{s2014b,asv2014b}. Our
data confirmed significant deviations from classicality
and evidenced a very stable pattern of such deviations to the classicality conditions. The data could very faithfully be represented in two-sector Fock space, thus providing support to our quantum-theoretic modeling in Sect. \ref{conjdisj}. More, they allowed us to attain new fundamental results in concept research and to sustain and corroborate our explanatory hypothesis in Sect. \ref{nature}. Hence, it is worth to briefly review our recent results starting from the conditions for classicality of conceptual data sets, i.e. representability of empirical membership weights in a
Kolmogorovian probability space.
Let $\mu(A), \mu(B), \mu(A'), \mu(B')$, $\mu(A\ {\rm and}\ B)$, $\mu(A\ {\rm and}\ B')$, $\mu(A'\ {\rm and}\ B)$, and $\mu(A'\ {\rm and}\ B')$ be the membership weights of an item $x$ with respect to the concepts $A$, $B$, their negations `not $A$', `not $B$' and the conjunctions `$A$ and $B$', `$A$ and not $B$', `not $A$ and $B$' and `not $A$ and not $B$', respectively, and suppose that all these membership weights are contained in the interval $[0,1]$ (which they will be in case they are experimentally determined as limits of relative frequencies of respective memberships). Then, they are `classical conjunction data' if and only if they satisfy the following conditions.
\begin{eqnarray} \label{condbis01}
&\mu(A)=\mu(A\ {\rm and}\ B)+\mu(A \ {\rm and}\ B') \\ \label{condbis02}
&\mu(B)=\mu(A\ {\rm and}\ B)+\mu(A' \ {\rm and}\ B) \\ \label{condbis03}
&\mu(A')=\mu(A'\ {\rm and}\ B')+\mu(A' \ {\rm and}\ B) \\ \label{condbis04}
&\mu(B')=\mu(A'\ {\rm and}\ B')+\mu(A \ {\rm and}\ B') \\ \label{condbis05}
&\mu(A\ {\rm and}\ B)+\mu(A\ {\rm and}\ B')+\mu(A'\ {\rm and}\ B)+\mu(A'\ {\rm and}\ B')=1
\end{eqnarray}
(see \cite{asv2014b} for the proof).
A large amount of data collected in \cite{asv2014b} violates very strongly and also very systematically Eqs. (\ref{condbis01})--(\ref{condbis05}), hence these data cannot be generally reproduced in a classical Kolmogorovian probability framework. It can instead be shown that almost all these data can be represented by using our quantum-theoretic modeling in two-sector Fock space, as in Sect. \ref{conjdisj}. For the sake of simplicity, let us work out separate representations for the two sectors.
Let us start from sector 1 of Fock space, which models genuine emergence. We represent the concepts $A$, $B$ and their negations `not $A$', `not $B$' by the mutually orthogonal unit vectors $|A\rangle$, $|B\rangle$ and $|A'\rangle$, $|B'\rangle$, respectively, in the individual Hilbert space ${\cal H}$. The corresponding membership weights for a given item $x$ are then given by the quantum probabilistic Born rule
\begin{eqnarray}
\mu(A)=\langle A|M|A\rangle & \quad & \mu(B)=\langle B|M|B\rangle \\
\mu(A')=\langle A'|M|A'\rangle& \quad & \mu(B')=\langle B'|M|B'\rangle
\end{eqnarray}
in sector 1. The conjunctions `$A$ and $B$', `$A$ and not $B$', `not $A$ and $B$', and `not $A$ and not $B$' are represented by the superposition vectors $\frac{1}{\sqrt{2}}(|A\rangle+|B\rangle)$, $\frac{1}{\sqrt{2}}(|A\rangle+|B'\rangle)$, $\frac{1}{\sqrt{2}}(|A'\rangle+|B\rangle)$ and $\frac{1}{\sqrt{2}}(|A'\rangle+|B'\rangle)$, respectively, in ${\cal H}$, i.e. sector 1 of Fock space, which expresses the fact `$A$ and $B$', `$A$ and not $B$', `not $A$ and $B$', and `not $A$ and not $B$' are considered as newly emergent concepts in sector 1.
Let us come to sector 2 of Fock space, which models logical reasoning.
Here we introduce a new element,
expressing an insight which we had not
yet in our earlier application of Fock space \cite{a2009a,ags2013,s2014b,s2014a}, and which we explain in detail in \cite{asv2014b}. In short it comes to `taking into account that possibly $A$ and $B$ are meaning-connected and hence their probability weights mutually
dependent'. If this is the case, we cannot represent, e.g., the conjunction `$A$ and $B$' by the tensor product vector $|A\rangle \otimes |B\rangle$ of ${\cal H} \otimes {\cal H}$, as we have supposed in Sect. \ref{conjdisj}. This would indeed entail that the membership weight for the conjunction is $\mu(A \ {\rm and} \ B)=\mu(A)\mu(B)$ in sector 2, that is, probabilistic independence between the membership estimations of $A$ and $B$. We instead, following this new insight, represent the conjunction `$A$ and $B$' by an arbitrary vector $|C\rangle\in {\cal H} \otimes {\cal H}$, in sector 2, which in general will be entangled if $A$ and $B$ are meaning dependent. If we represent the decision measurements of a subject estimating the membership of the item $x$ with respect to the concepts $A$ and $B$ by the orthogonal projection operators $M\otimes \mathbbm{1}$ and $\mathbbm{1}\otimes M$, respectively, we have
\begin{equation}
\mu(A)=\langle C|M\otimes \mathbbm{1}|C\rangle \quad
\mu(B)=\langle C| \mathbbm{1}\otimes M|C\rangle
\end{equation}
in sector 2. We have now to formalize the fact that this sector 2 has to express logical relationships between the concepts. More explicitly, the decision measurements of a subject estimating the membership of the item $x$ with respect to the negations `not $A$' and `not $B$' should be represented by the orthogonal projection operators $(\mathbbmss{1}-M)\otimes \mathbbm{1}$ and $\mathbbm{1}\otimes (\mathbbmss{1}-M)$, respectively, in sector 2, in such a way that
\begin{equation}
\mu(A')=1-\mu(A)=\langle C|(\mathbbm{1}-M)\otimes \mathbbm{1}|C\rangle \quad
\mu(B')=1-\mu(B)=\langle C| \mathbbm{1}\otimes (\mathbbm{1}-M|C\rangle)
\end{equation}
in this sector.
Interestingly enough, there is a striking connection between logic and classical probability when conjunction and negation of concepts are at stake. Namely, the logical probabilistic structure of sector 2 of Fock space sets the limits of classical probabilistic models, and vice versa. In other words, if the experimentally collected membership weights $\mu(A)$, $\mu(B)$, $\mu(A')$, $\mu(B')$, $\mu(A\ {\rm and}\ B)$, $\mu(A \ {\rm and}\ B')$, $\mu(A'\ {\rm and}\ B)$ and $\mu(A'\ {\rm and}\ B')$ can be represented in sector 2 of Fock space for a given choice of the state vector $|C\rangle$ and the decision measurement projection operator $M$, then the membership weights satisfy (\ref{condbis01})--(\ref{condbis05}), hence they are classical data. Vice versa, if $\mu(A)$, $\mu(B)$, $\mu(A')$, $\mu(B')$, $\mu(A\ {\rm and}\ B)$, $\mu(A \ {\rm and}\ B')$, $\mu(A'\ {\rm and}\ B)$ and $\mu(A'\ {\rm and}\ B')$ satisfy (\ref{condbis01})--(\ref{condbis05}), hence they are classical data, then an entangled state vector $|C\rangle$ and a decision measurement projection operator $M$ can always be found such that $\mu(A)$, $\mu(B)$, $\mu(A')$, $\mu(B')$, $\mu(A\ {\rm and}\ B)$, $\mu(A \ {\rm and}\ B')$, $\mu(A'\ {\rm and}\ B)$ and $\mu(A'\ {\rm and}\ B')$ can be represented in sector 2 of Fock space (see \cite{asv2014b} for the proof).
Let us finally come to the general representation in two-sector Fock space. We can now introduce the general form of the vector representing the state of the conjunction of the concepts $A, B$ and their respective negations.
\begin{eqnarray}
| \Psi_{AB} \rangle&=& m_{AB}e^{i \lambda_{AB}} |C\rangle + \frac{n_{AB}e^{i \nu_{AB}}}{\sqrt{2}} (|A\rangle+|B\rangle) \\
| \Psi_{AB'} \rangle&=& m_{AB'}e^{i \lambda_{AB'}} |C\rangle + \frac{n_{AB'}e^{i \nu_{AB'}}}{\sqrt{2}} (|A\rangle+|B'\rangle) \\
| \Psi_{A'B} \rangle&=& m_{A'B}e^{i \lambda_{A'B}} |C\rangle + \frac{n_{A'B}e^{i \nu_{A'B}}}{\sqrt{2}} (|A'\rangle+|B\rangle) \\
| \Psi_{A'B'} \rangle&=& m_{A'B'}e^{i \lambda_{A'B'}} |C\rangle + \frac{n_{A'B'}e^{i \nu_{A'B'}}}{\sqrt{2}} (|A'\rangle+|B'\rangle)
\end{eqnarray}
where $m^2_{XY}+n^2_{XY}=1$, $X=A,A',Y=B,B'$.
The corresponding membership weights are
\begin{eqnarray}
\mu(A\ {\rm and}\ B)&=&m_{AB}^2 \alpha_{AB}+n_{AB}^2({1 \over 2}(\mu(A)+\mu(B))+\beta_{AB}\cos\phi_{AB})
\label{FockSpaceSolutionAB} \\
\mu(A \ {\rm and} \ B')&=&m_{AB'}^2 \alpha_{AB'}+n_{AB'}^2({1 \over 2}(\mu(A)+\mu(B'))+\beta_{AB'}\cos\phi_{AB'}) \label{FockSpaceSolutionAB'} \\
\mu(A' \ {\rm and}\ B)&=&m_{A'B}^2 \alpha_{A'B}+n_{A'B}^2({1 \over 2}(\mu(A')+\mu(B))+\beta_{A'B}\cos\phi_{A'B})
\label{FockSpaceSolutionA'B} \\
\mu(A'\ {\rm and}\ B')&=&m_{A'B'}^2\alpha_{A'B'}+n_{A'B'}^2({1 \over 2}(\mu(A')+\mu(B'))+\beta_{A'B'}\cos\phi_{A'B'}) \label{FockSpaceSolutionA'B'}
\end{eqnarray}
where $0\le \alpha_{XY},\beta_{XY}\le 1$, $X=A,A',Y=B,B'$.
Let us consider a relevant example, {\it Goldfish}, with respect to ({\it Pets}, {\it Farmyard Animals}) (big overextension in all experiments, but also double overextension with respect to {\it Not Pets And Farmyard Animals}). {\it Goldfish} scored $\mu(A)=0.93$ with respect to {\it Pets}, $\mu(B)=0.17$ with respect to {\it Farmyard Animals}, $\mu(A')=0.12$ with respect to {\it Not Pets}, $\mu(B')=0.81$ with respect to {\it Not Farmyard Animals}, $\mu(A \ {\rm and} \ B)=0.43$ with respect to {\it Pets And Farmyard Animals}, $\mu(A \ {\rm and} \ B')=0.91$ with respect to {\it Pets And Not Farmyard Animals}, $\mu(A' \ {\rm and} \ B)=0.18$ with respect to {\it Not Pets And Farmyard Animals}, and $\mu(A' \ {\rm and} \ B')=0.43$ with respect to {\it Not Pets And Not Farmyard Animals}. A complete modeling in the Fock space satisfying Eqs. (\ref{FockSpaceSolutionAB}), (\ref{FockSpaceSolutionAB'}), (\ref{FockSpaceSolutionA'B}) and (\ref{FockSpaceSolutionA'B'}) is characterized by coefficients:
(i) interference angles $\phi_{AB}=78.9^{\circ}$, $\phi_{AB'}=43.15^{\circ}$, $\phi_{A'B}=54.74^{\circ}$ and $\phi_{A'B'}=77.94^{\circ}$;
(ii) coefficents $\alpha_{AB}=0.35$, $\alpha_{AB'}=0.9$, $\alpha_{A'B}=0.22$ and $\alpha_{A'B'}=0.17$;
(iii) coefficients
$\beta_{AB}=-0.24$,
$\beta_{AB'}=0.10$, $\beta_{A'B}=0.12$ and $\beta_{A'B'}=0.30$;
(iv) convex weights
$m_{AB}=0.45$,
$n_{AB}=0.89$,
$m_{AB'}=0.45$,
$n_{AB'}=0.9$,
$m_{A'B}=0.48$, $n_{A'B}=0.88$, $m_{A'B'}=0.45$, and $n_{A'B'}=0.89$.
\section{Origin and foundations of human reasoning\label{aristotle}}
In the previous sections we have provided concrete evidence that human reasoning is actually guided by two drivers, namely, emergence -- which stands for the continuous creation of new conceptual structures carrying also new meaning whenever the human mind combines pieces of existing conceptual structure containing given meaning -- and logic -- which consists of applying the probabilistic rules of logic for conjunctions, disjunctions and negations appearing in these conceptual combinations. And that these aspects of human reasoning can be formalized by using the mathematical formalism of quantum theory in Fock space. We have also made clear that emergent reasoning generally prevails over logical reasoning in concrete human decisions.
In this respect, the systematic effects, fallacies and paradoxes discovered in experimental psychology should not be considered as biases of human reasoning but,
rather, as fundamental expressions of conceptual emergence at the deepest level.
There is
further empirical evidence revealing that what really guides human subjects in a concrete decision is not logical but emergent reasoning. Consider, for example, the item {\it Olive} and its membership weights with respect to the concepts {\it Fruits}, {\it Vegetables} and their disjunction {\it Fruits Or Vegetables}, measured by Hampton \cite{h1988b}, and its membership weights with respect to the concepts {\it Fruits}, {\it Vegetables} and their conjunction {\it Fruits And Vegetables}, measured by ourselves \cite{s2014b,asv2014b}.
{\it Olive} scored $\mu(A)=0.5$ with respect to {\it Fruits}, $\mu(B)=0.1$ with respect to {\it Vegetables} and $\mu(A \ {\rm
or} \ B)=0.8$ with respect to {\it Fruits Or Vegetables}, that is, {\it Olive} was double overextended with respect to the disjunction. In addition, {\it Olive} scored $\mu(A)=0.56$ with respect to {\it Fruits}, $\mu(B)=0.63$ with respect to {\it Vegetables} and $\mu(A \ {\rm and} \ B)=0.65$ with respect to {\it Fruits
And Vegetables}, that is, {\it Olive} was double overextended also with respect to the
conjunction. This means that
situations exist where
people do not really take into account whether the connective `or', or the connective `and', is considered, but they actually estimate whether the item $x$ is a member of the new emergent concept, be it `$A$ or $B$', or `$A$ and $B$'. Indeed, the membership weight oscillates around the average $\frac{1}{2}(\mu(A)+\mu(B))$, due to the interference term. But, sector 1 emergence almost completely drives the cognitive dynamics.
Now, since the origins of logic can be traced back to the ancient Greeks, while emergent phenomena are a relatively recent discovery of modern science, one is naturally led to inquire into the deep motivations why logical reasoning has been historically formalized much before emergent reasoning. Why did logic appear first? We think that there are two main reaons for this, as follows.
(i) Emergent reasoning, as formalized in sector 1 of Fock space, needs sophistiacted technical tools, as
algebraic structures, Hilbert spaces, etc., while logical reasoning can basically be formalized by combining more intuitive structures. It is then not implausible that the first set-theoretical models by ancient Greeks could better capture the latter than the former reasoning.
(ii) Ancient Greeks formalized logical reasoning with the objective of fully understanding mathematical reasoning. Since then, mathematics develops a strong method of reasoning (the one formalized in sector 2 of Fock space), and limits its application range to the domains where such a mathematical reasoning is applicable.
Point (ii) is important,
because it helps understanding why the effects discovered by Tversky, Khaneman and other cognitive psychologists were first classified as biases or even fallacies of human reasoning. If mathematical reasoning is the most rigorous guide of classical logical reasoning, then each deviation from mathematical reasoning, e.g., wrong application of probabilities in concrete human judgements, should be considered as a real fallacy of human reasoning.
This leads us to another interesting question. Which are the domains of our world where logical reasoning is more valuable or perhaps even the only valuable type of reasoning, such as in mathematics? And, which are the domains of our world where emergent reasoning is the more valuable type of reasoning to be used? We believe that, whenever reasoning has as subject
`concrete material objects being present in real space-like configuration', logical reasoning is the one which is more appropriate. People who give dominance to emergent reasoning in these situations
defined mainly by presences of material objects in space and time will commit fallacies of reasoning, and possibly their reasoning will lead to expectations about things to happen in this material object world that do not happen following these expectations. The reason is that this material world of objects in space-time can be modelled by considering sets, subsets, movements from one set into another, hence a dynamics of classical physics nature. Of course, for real situations such a description might be very complex, but for sector 2 Fock space reasoning to be valid, it is sufficient that in principle
such a classical mechanics model is possible. To give
an easier picture of what we mean, manuals that exist as guides for the functioning of technical mechanical or electric devices are good examples of descriptions which allow such sector 2 of Fock space logical reasoning. So, the question `whether a specific type of happenings can be captured completely in a manual' is a good criterion for
sector 2 Fock space reasoning validity. And, what are the domains of our world where
sector 1 Fock space reasoning is the more valid one? We believe that, whenever the subject of our reasoning consists of complex enough conceptual situations, the human mind is in need of the emergence of new meaning and new conceptual structures at every instant. It is what we call `creativity'. And again, in case in these domains where emergent conceptual reasoning is the
most valid one people prefer to give dominance to sector 2 Fock space logical reasoning, this will not necessarily be experienced as fallacies but, rather, as `nerdy type of not understanding'. To state it metaphorically, `A poem is not a manual'.
Again to mathematics, is this not a discipline of reasoning focused completely on conceptual entities and not at all on material objects? It is indeed. But, in some way, mathematicians have delimited their subject material in function of
sector 2
logical reasoning being applicable to it, and even a strictly deterministic version of it. We believe that, first the liar paradox, and
then G\"odel's paradox, and all its equivalents, e.g., the halting problem,
are situations where the systematics of applying classical logic failed also in the mathematical realm. It is most probably not a coincidence that quite some time ago
we produced in Brussels a quantum model description for the cognitive dynamics of the liar paradox situation \cite{aertsbroekaertsmets1999}.
To conclude, we have unveiled in this paper crucial new aspects of the foundations of human reasoning with our explanatory
hypothesis on the existence of two layers in human thought guiding cognitive dynamics. However, we believe that still other aspects of the nature of human reasoning remain to be identified,
and we plan to perform new experiments to deepen this fascinating argument in future research.
|
\section{INTRODUCTION}
Coherent perfect polarization
rotation (or CPR)~\cite{CPR1} is a conservative, reversible
example of a multiport, maximally efficient, optical mode conversion process.
As such
it shares phenomenological correspondences with
the coherent perfect absorber
(antilaser or CPA)~\cite{chong10.01,wan11.01} which has been
well studied~\cite{longhi10.01, chong11.01, longhi11.01,lin11.01}.
While many optical devices such as laser wavelength locks,
field sensors, optical isolators, and modulators
are based on the non-reciprocal nature of
Faraday rotation, one way to improve all of these devices
is to process all of the incident light coherently. CPR-based design is
an intrinsically multi-(input)port approach that combines the
non-reciprocal nature of
the Faraday effect with interference to convert {\it all} of the incident light
into its orthogonal polarization. An example of the basic two-port
CPR device is shown in Fig.~\ref{CPRbasic}.
\begin{figure}[b]
\includegraphics[width=\linewidth]{CPR_schematic2.eps}
\caption{Basic schematic of two-port CPR device. Not shown are
attenuators and delay plates that balance the input field's
amplitude and phase. When the CPR resonance condition is achieved, the
reflected light $r_1$ and $r_2$ vanishes.}
\label{CPRbasic}
\end{figure}
For the apparatus shown in Fig.~\ref{CPRbasic}, one still must tune the magnetic field
to specific values to achieve complete conversion of the polarization. The required
field is significantly below that of a single port rotator, however. For one example, the complete conversion of one polarization into the orthogonal polarization using an uncoated terbium gallium garnet (TGG) slab requires only 60\% of the field-length product needed for the same rotation in just transmitted light~\cite{CPR1}. Achieving
complete polarization conversion at a lower field-length
product is technologically
useful because it is precisely the seeming
``incompressibility'' of this product and
the modest Verdet coefficients of commercially available optical materials that
pose a major obstacle to the diffusion of single port designs into integrated
optical assemblies and low cost devices. The primary motivation for
the work reported here is to quantify in typical
one-dimensional optical geometries
how CPR-based optical design significantly lowers the
field-length product. We characterize the reduction of
the threshold field-length product at which coherent perfect processes first
occur consequent to specific design choices
in optical dispersion, structural dispersion,
broken parity, and localization.
A second motivation for this study is to reveal general principles
common among coherent perfect processes. Enlarging the context for
these phenomena builds intuition useful for finding routes
to improving optical devices. We do this primarily
by comparing and contrasting CPR and CPA.
Note that in both CPR and CPA, a tunable time-odd optical process
(magneto-optical rotation for CPR versus absorption in CPA)
is combined with multiport wave interference
to achieve perfectly efficient mode conversion
(to orthogonal polarization in CPR and to electronic excitation in CPA).
Among other commonalities between CPA and CPR are their
critical dependence on the relative optical phases among
the input lightfields. As in the single port case of
critical coupling, CPA and CPR both require a particular
(hereafter ``threshold'') magnitude for the time-odd
process. In both CPR and CPA the conversion efficiency has a
resonance-like structure. Also, for a fixed wavelength
going both above or below threshold makes complete conversion impossible.
Because this ``resonance''
is not associated with a particular decay timescale, coherent perfect
resonances are intrinsically zero-width.
Understanding of coherent perfect absorption (CPA) phenomena in diverse optical systems has advanced steadily. Theory relates CPA states to self-dual spectral singularities~\cite{chong10.01,mosta} of the S-matrix. The CPA threshold's dependence on the depth of the sample is well understood theoretically, and this dependence has been modeled in complex (but still linear) media such as gold-silica composites and other plasmonic systems~\cite{agarwal,agarwal2}, and nonlinear media~\cite{longhi_nonlin}. Both ordinary and ${\cal P}{\cal T}$-symmetric systems elicit a diverse set of CPA phenomena with those most relevant to CPA thresholds including gratings~\cite{grating1,PTgrating,metasurfaceCPA,metasurfaceCPA2}, surface plasmonic polaritons~\cite{metasurfaceCPA3}, photonic crystals~\cite{PT_PhC}, near-zero-$\epsilon$ materials~\cite{eps_zero}, cavities with absorbers~\cite{cavity_abs}, controlled disorder or other spatial ordering~\cite{disorderedCPA,spatial_variation}, and very thin absorptive layers~\cite{absorb_thin}. Some of these ideas are also being explored for technological uses including all optical switching~\cite{CPAswitch,fastswitch,yet_another_CPA_switch,yep_another_one,diffractive_switch} and CPA enhancement of photoluminescence~\cite{PLenhance}.
After fixing notation and giving
a brief review of the basic phenomena,
Section III explores CPR and contrasts it with CPA
in different optical environments, highlighting the roles played by dispersion, parity, and
transport of mixed symmetry type. In Section IIIA we focus on coherent
perfect phenomena in model multilayer systems, followed in Section IIIB by breaking parity two
different ways (first softly with randomness and then explicitly with
trinary multilayers). Thematically up to that point,
one sees a direct correspondence between the layering effects on CPR and CPA states
and thresholds. Subtle differences between the two are discussed
in one archetypal example combining CPR and optical rotation in Section
IIIC. A brief conclusion highlights new directions prompted by this study.
\section{NOTATION AND PRELIMINARIES}
We use matrices to represent linear transport and
work in the basis where the local field (complex) amplitudes
for light traveling along the ${\hat z}$-axis are
${\vec v} = (E_x, H_y, E_y, -H_x)$. In terms of the individual polarization
and motional states, we use ${\vec e}_R = (E_x, H_y) = (1,1)$ for a right-moving wave and ${\vec e}_L = (-1,1)$ for a left-moving one.
Throughout this paper we restrict ourselves to
materials without linear birefringence (in contrast with ~\cite{wang1})
The $O(2)$ symmetry about the axial
direction implies for the transport
${\vec v}_{i+1} = {\cal M}_i {\vec v}_i$ that the 4$\times$4 ${\cal M}$ can be written
(in this basis) in terms of the 2$\times$2's $M$ and $C$ as
${\cal M} =
\left( \begin{array}{cc}
M & C \\
-C & M \end{array} \right)$,
where $C$ is only nonzero for transport that mixes the polarization states.
For dielectrics (also the only case we consider below), the
matrix $M$ is proportional to the familiar 2$\times$2 transfer matrix
for the individual polarizations. For example, for a unit intensity wave
incident from the left, in steady state, the field amplitudes at the
surface are
${\vec e}_{\rm in} = (1,1) + r(-1,1)$, where $r$ is the reflected amplitude.
The outgoing field amplitude to the right of the system is given via
${\vec e}_{\rm out} = t(1,1) = M{\vec e}_{\rm in}$,
where $t$ is the transmission amplitude. In this basis, for a
purely dielectric material of thickness $L$, index $n$,
\begin{equation}
M =
\left[ \begin{array}{cc}
\cos\delta & {\frac{i}{n}}\sin\delta \\
in\sin\delta & \cos\delta\\
\end{array} \right],
\label{Mdielectric_2by2}
\end{equation}
where $\delta = nk_0L$ and $k_0$ is the vacuum wavenumber.
Note that $det(M)=1$ always, but $M_{11}$ and $M_{22}$ are only
equal in systems that have overall spatial parity symmetry.
We identify the real part of the index $n$ with refraction and its positive/negative imaginary part with absorption/gain.
Analytically for a slab dielectric Faraday rotator
the $M$ and $C$ parts of the ${\cal M}$ in our field
basis are~\cite{CPR1,kato03.01}
\begin{equation}
M = {\frac{1}{2}} \left[ \begin{array}{cc}
C_1+C_2 & i(S_1/n_1+S_2/n_2) \\
i(n_1S_1+n_2S_2) & C_1 + C_2 \end{array} \right]
\label{Mdielectric}
\end{equation}
and
\begin{equation}
C = {\frac {1}{2}} \left[ \begin{array}{cc}
i(C_1-C_2) & -(S_1/n_1-S_2/n_2) \\
-(n_1S_1-n_2S_2) & i(C_1 - C_2) \end{array} \right] \, ,
\label{MnCdielectric}
\end{equation}
where $C_{1,2}$ ($S_{1,2}$) refer to the cosine (sine) of
$\delta_{1,2} = n_{1,2} k_0 L$ in which the $n_1,n_2$ are the indices
of refraction
of the left- and right- circular polarization in the slab,
$k_0$ refers to the vacuum wavevector, and $L$ is the thickness of
the slab. For a dielectric slab in an
external magnetic field pointing along the direction of propagation,
$\Delta n=n_1-n_2 \propto VB$, the product of the Verdet and the magnetic
field. Note that the resulting 4$\times$4 matrix
${\cal M}$ is quite different from one
representing optical activity (a time-even rotation process)
which has the form $M= \cos\alpha M_0$ and $C=\sin\alpha M_0$, where $\alpha$ is proportional to the density of
chiral centers in the slab and $M_0$ is the usual 2$\times$2 transfer matrix
given by Eq.~(\ref{Mdielectric_2by2}).
Because CPR is a reversible optical process we require constant local power flux
throughout in steady state. This condition thus requires the $n$'s and the $\alpha$ to be real throughout for both the time-even and time-odd rotation
processes we consider below.
For a single polarization whose linear transport is given
entirely in terms of a net 2$\times$2 transfer matrix $M$ in the basis
described above and used throughout,
the CPA state is reached when the condition
$(1,1)M(1,1)^t = M_{11} + M_{12} + M_{21} + M_{22} = 0$
is satisfied. For a general
2$\times$2 matrix,
this condition combined with the determinant indicates that CPA
implies four real conditions for four complex numbers. A remaining freedom
of optical field (amplitude and phase) then implies that
CPA requires, at minimum, tuning two dimensionless
experimental parameters, typically, the ratio $L/\lambda$ and the absorptive
index $Im(n)$.
It is also straightforward to find the condition associated with
CPR resonances using the
4$\times$4 basis. For fields incident from the left, take
${\vec v}_l = (1,1,-l, l)$, where $l$ is the amplitude of the reflected,
rotated wave. On the right, take ${\vec v}_r =
(-d, d,s,s)$; this configuration thus consists of incoming
fields of one polarization and outgoing fields of the orthogonal
polarization only, the CPR state.
In analogy with the CPA state, these boundary conditions
lead to a condition on the
size, wavelength and rotary power of the system.
For CPR resonance in uniaxial systems
with the 4$\times$4 form of ${\cal M}$ as described earlier, we require
\begin{equation}
M \left( \begin{array}{c} 1 \\ 1 \end{array} \right) + C \left( \begin{array}{c} -1 \\
1 \end{array} \right) l
=
\left( {\begin{array}{c}
-1 \\
1 \end{array}} \right) d
\label{CPRa}
\end{equation}
and
\begin{equation}
-C
\left( \begin{array}{c}
1 \\
1 \end{array} \right)
+ M \left( \begin{array}{c}
-1 \\
1 \end{array} \right) l
=
\left( \begin{array}{c}
1 \\
1 \end{array} \right) s \, .
\label{CPRb}
\end{equation}
Counting conditions (four complex) for the three complex
fields ($d,l,s$), we see
that to achieve CPR by simultaneously
solving Eqs.~(\ref{CPRa}) and (\ref{CPRb}) requires, at minimum, tuning
two experimental parameters (here, generically, the ratio ${L/\lambda}$
and the circular birefringence $\Delta n=n_1-n_2$), which we note is
analogous to the CPA case (where the parameters are $(L/\lambda)$ and the absorption coefficient).
Eliminating the fields $d,l,s$, we can write the CPR condition succinctly for
a general ${\cal M}$ as
a single complex condition $det(R)=0$, where the 2$\times$2 matrix $R$ has the
following elements:
\begin{equation}
R_{11} = (-1,1)C^{-1}M\left( \begin{array}{c}
-1 \\
1 \end{array} \right) \,
\label{CPR_simpler1}
\end{equation}
\begin{equation}
R_{12}= -(-1,1)C^{-1} \left( \begin{array}{c}
1 \\
1 \end{array} \right) \,
\end{equation}
\begin{equation}
R_{21} = (1,1)[MC^{-1}M+C]\left( \begin{array}{c}
-1 \\
1 \end{array} \right) \,
\end{equation}
\begin{equation}
R_{22} = -(1,1)MC^{-1}\left( \begin{array}{c}
1 \\
1 \end{array} \right) \, .
\label{CPR_simpler4}
\end{equation}
We now summarize
CPR phenomenology in a series of optical systems
in order to build a deeper intuition about the CPR state and
its connection to and contrast with CPA, with an eye towards its potential utility
in optical devices.
\section{ CPR IN MODEL SYSTEMS}
\subsection {CPR in layered binary systems}
Studying CPR in multilayer interference films provides a
straightforward comparison of CPR and CPA phenomena and their dependence on
dispersion, both material and structural. For simplicity, consider
first a perfectly periodic multilayer composed of $N$ alternating layers of
a material $A$ that is a dielectric
with zero Verdet and a material $B$ that has a non-zero Verdet.
We compare these systems to the CPA model system in which the bilayers
have one non-absorbing species ($A$) and the other absorbing ($B$).
In all the model systems described here, only the $B$ species rotates (for CPR)
or absorbs (for CPA).
The species
have different indices of refraction in the absence of a magnetic field (for CPR) or absorption (for CPA) that we denote $n_A$ and $n_B$, creating
an optical (reflection) bandgap. We denote these layered systems
as $(AB)^N$, but here, to eliminate any spurious effect
from explicitly broken parity, we restrict our attention to
parity symmetric layered systems
formed by adding one terminal $A$ layer, that is, $(AB)^NA$.
\begin{figure}[t]
\includegraphics[width=\linewidth]{1_typical_CPR_30lay_rev.eps}
\caption{(Color online) Typical CPR resonance in a 1-d photonic crystal. The green dashed trace is the transmission trace and the red solid trace is the total output light in the same polarization as the input polarization.
This example is 31
alternating layers (each 100 nm thick)
of index 1.55 (non Faraday) and 1.38 (Faraday),
corresponding to the first row in Table \ref{table1}~\cite{plastic}.
Clearly seen in the dashed green trace is the reflection bandgap that extends from 550-625 nm.
The CPR resonance is the pronounced reduction in the output light polarized along the input polarization for wavelengths near the short wavelength edge of the band.}
\label{foldedDFB_CPR}
\end{figure}
\begin{table
\caption{CPR thresholds ($\Delta n$ values for the circular polarization propagation eigenstates in layer $B$) at the reflection band edge for the layered binary systems described in Fig. \ref{foldedDFB_CPR}. Throughout this paper the letters after the threshold values indicate the spatial symmetry ($O$ for odd, $E$ for even) of that CPR resonance's fields.}
\label{table1}
\begin{tabular}{ c c | c c c c }
\hline
\hline
configuration &&& 31 layers && 33 layers \\
\hline
$n_A>n_B$ &&& .049(E) && .044(O) \\
$n_A<n_B$ &&& .090(E) && .080(O) \\
\hline
\hline
\end{tabular}
\end{table}
One finding of these simulations (see Fig.~\ref{foldedDFB_CPR})
is that for the lowest
thresholds, layered films with an
odd number of bilayers ({\it e.g.}, 31 layers) had threshold CPR states of
even parity and those with
an even number of bilayers ({\it e.g.}, 33 layers) had threshold CPR states
of odd parity.
A simple explanation of this observation is given in
the next subsection on consequences of parity symmetry (and parity breaking).
Note also that the
wavelength at which the lowest CPR resonance occurs is
at a band edge.
As is well known, across the reflection
bandgap there is pronounced optical dispersion
resulting in large increases
in the group velocity delay symmetrically
at the band edges, and significant reductions in the delay in the middle of
the band. (Ref.\cite{cresc12.01} is a recent relevant summary.)
The reduction in the threshold for CPR/CPA with
increases in group velocity delay at the band edge is most clearly seen by plotting
the product of
the threshold value of the rotary power of the $B$ layers times the number of layers versus
the number of layers, as in Fig. \ref{multilayer_effect} (dashed green trace)
which hews closely to a plot of the
group velocity minima versus the number of layers (solid red trace).
Changes in the group velocity delay
are, in and of themselves, not enough
to explain the pattern of CPR (and CPA) resonances in these systems; the
lowest threshold for the CPR/CPA resonances for $n_A>n_B$ occurs at the
short wavelength side of the bandgap but occurs at the long wavelength
edge of the bandgap for $n_A<n_B$.
As the time reverse of CPA, actual
lasing~\cite{Dowling}, indicates, the simplest way to explain this
difference is by apportioning the group velocity delay
across the two species of the multilayer,
and noting that only in the $B$ species is the light subject to polarization
rotation (CPR) or absorption/gain (CPA/lasing). Simulations of the local
electric field of the light traversing the multilayer indicate that, off-band,
the apportionment of the total velocity delay
should follow the ratio of the indices. Near the
long wavelength side of the band edge, however, the light's integrated electric energy
is greater in the larger index species,
whereas the reverse occurs at the short wavelength edge of the
bandgap~\cite{YSUreview}. Equating field energy to the probability that the light visits that
species, and apportioning part of the total propagation time
(and thus the overall Faraday rotation) to each species in proportion to that
probability, qualitatively explains both the wavelength and
the threshold of the CPR/CPA resonances.
\begin{figure
\includegraphics[width=\linewidth]{2_CPR_vs_n.eps}
\caption{(Color online) The correspondence between the group velocity delay and the CPR thresholds
in layered media. Varying the number of layers only, we plot
the product of the CPR threshold Verdet-field ($VB$)
product times the number of layers
versus the number of layers. We also plot the
group velocity minimum at the band edge versus the number of layers (right axis, solid red trace). }
\label{multilayer_effect}
\end{figure}
Just as in CPA, there are other CPR states that arise at
different wavelengths as one increases $\Delta n$ beyond threshold.
Note that for these simple binary multilayers, the
next-to-lowest CPR resonance typically
occurs at the opposite band edge, as expected and formerly noted~\cite{grating1} for CPA. To summarize the results from this study of CPR in multilayers
in experimental terms, a 0.76 mm thick multilayer of 160 nm layers of each
Bismuth-substituted iron garnet (called ``BIG'', index of
refraction of $\sim$ 2.3 at 633 nm, at which the Verdet is $\sim$ -7 x 10$^3$ rad/Tm ~\cite{kato03.01}) and ordinary SiO$_2$ glass would achieve CPR at 0.5 T, whereas
a slab of BIG alone of length 2.4 cm would be needed at this field, indicating in concrete terms the substantial reductions in CPR thresholds
associated with photonic bands.
In practice with real multilayer systems, non-ideality typically moves CPR/CPA
lowest threshold states from the band edge
to defect states in the band gap itself.
This is consonant with the experience in lasing where it is well
documented that layer nonuniformity and other perturbations cause
lasing to occur first through defect states typically located within
the band gap itself. The defect states still correspond to maxima of the
group velocity delay~\cite{wu09.01}. (See the discussion of parity breaking in subsection B below.)
It is illustrative to explicitly compare thresholds for
lasing/CPA and CPR in simple multilayer systems with
deliberate structural defects, such as the ``phase-slip'' (sometimes called
``folded'') distributed feedback (DFB) systems~\cite{YSUreview,yablanovich}.
Here we compare the multilayers $(AB)^N(BA)^N$ and $(BA)^N(AB)^N$, where the
time-odd process (either Faraday rotation in the case of CPR or absorption/gain
in the case of CPA/lasing) is again only in the $B$-layers.
Table \ref{table2} gives calculated threshold $\Delta n$ values for CPR for four configurations of simply-folded symmetric systems,
which agree qualitatively with the corresponding results for CPA/lasing
summarized in Fig. \ref{foldedDFBgain} (adapted from Ref.~\cite{DFBlaser}).
For example, controlling for overall gain, the folded DFB structure with the lowest lasing
threshold (as inferred from the largest gain in the figure) is that which has the gain medium
in the low index material and is folded on the low index material.
This result agrees with our simulations of the CPR threshold
as shown in Table \ref{table2} (folded on $B$, $n_A>n_B$).
\begin{table}
\caption{The CPR threshold values of $\Delta n$ for ``folded'' layered systems comprised of 52 total layers. Every entry in the table is for a CPR resonance occurring on the defect state inside the reflection band. The lowest CPR threshold occurs with even parity when rotation occurs in the lower index material and the fold is on that low index material. The CPR threshold ordering in the chart is in one-to-one agreement with that of lasing thresholds in these ``folded'' DFB systems reproduced in
Fig.~\ref{foldedDFBgain}.}
\label{table2}
\begin{tabular}{ c c | c c c c }
\hline
\hline
configuration &&& Fold on $A$ && Fold on $B$ \\
\hline
$n_A>n_B$ &&& .032(O) && .013(E) \\
$n_A<n_B$ &&& .017(E) && .028(O) \\
\hline
\hline
\end{tabular}
\end{table}
\begin{figure}[t]
\includegraphics[width=\linewidth]{fig3rev.eps}
\caption{(Color online) Transmission gain versus wavelength for the folded structures analogous to those in Table \ref{table2}, where instead of rotation in the $B$-layers a complex index of refraction is used to represent optical gain. In this case the optical band stretched from 450 nm to 520 nm, and the prominent fold defect state appears near the center of the band. (Figure adapted from Ref.~\cite{DFBlaser} with permission of The Optical Society of America.)}
\label{foldedDFBgain}
\end{figure}
For contrast we conclude this section with a case in which dispersion, but
not field placement, is important: the loaded optical
cavity as a layered optical system. Consider
a dielectric Faraday rotator inside an optical cavity composed of transversely
isotropic perfectly thin mirrors of reflectivity amplitude $r$ (so that the
reflectivity is $R = |r|^2$). The mirrors are represented by the transfer
matrix ${\cal M}_{r} = \left[ \begin{array}{cc}
M_r & 0 \\
0 & M_r \end{array} \right]$ where the 2$\times$2
matrices $M_r$ for the simple case of completely non-absorbing mirrors
are given in our 'field'
basis as
$M_r = {{1}\over{\sqrt{1-|r|^2}}}\left[ \begin{array}{cc}
1 & i|r| \\
-i|r| & 1 \end{array} \right]$.
Algebra indicates that all effects of the cavity reflectivity
modify the conditions for CPR via a single parameter,
$\gamma = {{2|r|}\over{1+|r|^2}}$.
One finds for this loaded cavity configuration
(mirror-rotator-mirror) that the CPR condition becomes ~(compare the
$r \rightarrow 0$ limit with Eq.(14) in Ref.~\cite{CPR1}):
\begin{equation}
(n_1+{{1}\over{n_1}}) S_1C_2-(n_2+{{1}\over{n_2}})S_2C_1+\gamma
\bigg({{n_1}\over{n_2}}-{{n_2}\over{n_1}}\biggr)S_1S_2 =
\nonumber
\end{equation}
\begin{equation}
\pm\biggr[(n_1-{{1}\over{n_1}})S_1-(n_2-{{1}\over{n_2}})S_2 + 2\gamma(C_2-C_1)\biggr] \, ,
\label{CPR_cavity}
\end{equation}
where, as before, $n_{1,2}= n_0\pm \Delta n/2$ and $k_0 \Delta n = 2VB$.
To compare this result with CPA, it is straightforward to show that
the lowest CPA resonance threshold for a cavity loaded with a lossy dielectric
modeled as a complex index $n$ is given by the solution of
(compare with the $r \rightarrow 0$ limit of Eq.~(7) of Ref.~\cite{chong10.01} also reproduced below in Eq.~(\ref{CPA_basic}) for completeness):
\begin{equation}
e^{i2nk_0L} =
{{(n-1)^2-(n^2+1){{2R}\over{1+R}} +i\gamma (n^2-1)}\over
{(n+1)^2-(n^2+1){{2R}\over{1+R}} +i\gamma (n^2-1)}}
\, .
\label{CPA_cavity}
\end{equation}
In Fig. \ref{CPR_cav}, we have
used Eq.~(\ref{CPR_cavity}) for $n_0 = 2.0$ and $k_0L \sim 820$
to plot the fractional
reduction in the lowest CPR resonance threshold ($\Delta n$) as
a function of the reflectivity, $R$.
The graph shows strong similarity to the inverse of the
time spent in the cavity ({\it i.e.} the fractional reduction in the group velocity), as expected, and also corresponds with the reduction in
the CPA threshold of an absorber-loaded cavity shown in the graph.
Not included in Fig.~\ref{CPR_cav},
we have also analyzed
a realistic ({\it e.g.} complex dielectric) gold mirrored cavity at
780 nm and find qualitatively the same behavior as in Fig. ~\ref{CPR_cav}
with increasing
gold layer thickness. In that study there are no 780 nm CPA states
from tuning the loss in the dielectric slab inside the cavity if the gold
layer thickness exceeds $~$35 nm (corresponding to an $R$ of about 85\% in each
mirror) because at that depth the absorption in the gold itself is
above the CPA threshold.
\begin{figure}[t]
\includegraphics[width=\linewidth]{8_thresh_cavity.eps}
\caption{(Color online) Lowest CPR (red, dot-dash trace) and CPA (blue, dashed trace) resonance thresholds in a loaded cavity decrease as one increases the finesse of the mirrors, going to zero with the inverse of the group velocity delay (green, solid). }
\label{CPR_cav}
\end{figure}
\subsection{CPR with explicitly broken parity}
In a parity symmetric absorbing structure, the fields of all CPA states must also be of definite parity, even or odd.
These two possibilities
generally occur at different absorption thresholds.
Since we have already
discussed layered optical systems, one particularly intuitive
way to understand this difference is shown in Fig. \ref{cylinders}, where one of the species ($B$) is absorbing (or rotating in the CPR case) and the other species ($A$) is not.
\begin{figure}[t]
\includegraphics[width=200pt]{cylinders.eps}
\caption{ Parity and CPA/CPR. For an odd number of bilayers, we see that
(a) the parity even fields have a
maximum on the $B$ species, whereas (b), the parity odd case, the field on the $B$ layer
is significantly smaller. For an even number of bilayers, however, the situation is reversed so that (c) parity even fields do not have their maxima on the $B$ layers, but (d) parity odd fields do.
}
\label{cylinders}
\end{figure}
For wavelengths nearly four times the layer thickness (near the band edge)
as in the example shown in Fig.~\ref{cylinders},
the parity even case has larger field overlap
on the absorber/rotator (species $B$) than the parity odd field for an
odd number of bilayers (in this case one bilayer), thus the former will have a lower CPA threshold (compare with Table~\ref{table1}).
When a rotator (or absorber) is not parity symmetric, there are still
CPR (CPA) states, but the state's fields will not be of definite parity.
To illustrate the effect of parity breaking on CPR and
its comparison with CPA (See Refs.~\cite{chong11.01,brokenP_CPA}), in this section we
consider two examples of parity broken systems:
(i) an $(AB)^NA$ multilayer, but with layer-to-layer
thickness variations, and (ii)
a trinary regular layered system of the
type $(ABC)^N$ (in both cases only $B$ is rotary (CPR) or absorptive (CPA)).
As one introduces layer thickness variations into the $(AB)^NA$ structures
discussed in the preceding subsection, formerly localized reflection band states
mix with extended states whereas some formerly extended states become localized~\cite{wu09.01}.
Initially, weak localization increases the group velocity delay and thus reduces the CPA/CPR threshold for some states near the band edge (see Fig.~\ref{localization} for one example). As the localization length shrinks further with increasing layer thickness variations, random scattering reduces the coherent band edge reflections that were responsible for the increase in the group velocity delay in the first place.
As the level of randomness is increased, the lowest resonant CPR/CPA state's wavelength at threshold moves into what was previously the reflection band.
\begin{figure}[t]
\includegraphics[width=\linewidth]{5_threshold_reduction.eps}
\caption{(Color online) Percentage reduction of CPR (closed circles, green) and CPA (open squares, red) in the thresholds for the lowest CPR/CPA state, which, as described in the text, occur near the band edge. For this example, a particular random layer thickness variation map for the 65 layer multilayer with $n_B>n_A$ (only species $B$ is Faraday (CPR case) or absorptive (CPA case)) is programmed into the simulation and increased across the horizontal axis.}
\label{localization}
\end{figure}
Note also that adding layer thickness
randomness explicitly breaks the original parity
symmetry of the system. As a consequence, at finite randomness
in the CPA case, the amplitude ratio of the input fields is no longer $\pm 1$.
The consequence of parity breaking through broken structural symmetry in CPR is different from that of CPA~\cite{paritynote}.
Solving Eqs.~(\ref{CPRa}),(\ref{CPRb}) for the amplitude ratio of the incident fields, $l$, indicates
\begin{equation}
l={ {(M_{11} + M_{12}+M_{21}+M_{22})} \over { (C_{11}+C_{21}-C_{12}-C_{22})}}
\nonumber
\end{equation}
\begin{equation}
= {{(C_{11}+C_{21}+C_{12}+C_{22})} \over {(M_{22}+M_{12}-M_{11}-M_{21})}}
\label{CPR_eq_for_l}
\end{equation}
on the CPR state.
For any optical system composed of sections without birefringence or optical activity, it was shown in Ref.~\cite{CPR1} that the 4$\times$4 ${\cal M}$ has underlying 2$\times$2 matrices $M$ and $C$ with the $M$ being time-even and of the form $\left[ \begin{array}{cc}
{\cal R} & {\cal I} \\
{\cal I} & {\cal R} \\
\end{array} \right]$
and the $C$ being time-odd and of the form
$\left[ \begin{array}{cc}
{\cal I} & {\cal R} \\
{\cal R} & {\cal I} \\
\end{array} \right]$, where ${\cal R}$ means a purely real quantity and
${\cal I}$ a purely imaginary one. This matrix structure in the formulas for $l$ gives
$|l|=1$, which in turn forces all the optical fields to have the
same intensity for a CPR state in any system.
The broken parity in the CPR state is instead manifest by the appearance of mismatched phases (not shown)
between the input fields.
\begin{table}
\caption{Thresholds for $(ABC)^N$ layered systems where the indices of refraction are chosen from the list \{1.55, 1.51, 1.38\} and the total number of layers is 45 (each layer is 100 nm thick.)}
\label{table3}
\begin{tabular}{ c c c c c c c }
\hline
\hline
TYPE && configuration && threshold && amplitude ratio \\
\hline
CPR && $n_A>n_C>n_B$ && .0465 && 1 \\
CPR && $n_A>n_B>n_C$ && .0652 && 1 \\
CPR && $n_B>n_A>n_C$ && .0752 && 1 \\
CPA && $n_A>n_C>n_B$ && .0528 && 1.44 \\
CPA && $n_A>n_B>n_C$ && .092 && .215 \\
CPA && $n_B>n_A>n_C$ && .087 && 1.52 \\
\hline
\hline
\end{tabular}
\end{table}
Explicit parity breaking via structural asymmetry is also evident in the trinary films, $(ABC)^N$, again, in which only $B$ is rotary (CPR) or absorptive (CPA). As an example, Table \ref{table3} shows a comparison of CPR and CPA lowest resonance thresholds near the band edges of the very first reflection band of trinary films.
Thus, even for perfectly ordered trinary films, it is the phase mismatch between the left and right input fields that varies universally, while the amplitude ratios only vary for the CPA case. Note also that the threshold values for the cases $n_A>n_C>n_B$ and $n_B>n_A>n_C$ are ordered the same in both CPR and CPA. In particular, for each of these cases in Table ~\ref{table3}, the CPR/CPA state forms at the appropriate band edge as discussed in the previous section.
The spectral location of the CPR/CPA state in the intermediate case $n_A>n_B>n_C$ depends on the indices' values.
Two additional facts of interest emerge from these simulations. As one might expect, the intensity ratios are more varied for the trinary films (explicit parity breaking) than for the random $(AB)^NA$ layered system (which breaks parity more softly) studied here to only 15\% layer thickness variation. Also, for the case of random $(AB)^NA$ layered systems, the variation in the phase is much larger in the CPR case than in the CPA case. Note in this regard that the CPR state forces the intensity ratio to remain unity, whereas for CPA both the amplitude ratio and the phase adjust to stay resonant in a parity broken system.
\subsection{Combined Faraday rotation and optical activity}
To highlight the time-reversal symmetries underlying CPR and CPA, we now
address the effect that the time-even part of the transport has on the
CPR/CPA threshold. In the original derivation of the CPR effect~\cite{CPR1}
in a simple slab dielectric,
increasing the index of refraction of the material reduces
the CPR threshold, as shown
graphically in Fig. \ref{fig_CPR_index} using the formula in Ref.~\cite{CPR1}.
\begin{figure}[t]
\includegraphics[width=\linewidth]{fig8redux2.eps
\caption{The CPR threshold reduction (red, dashed) is monotonic to zero as one increases the index of refraction of the slab, following the same relation as in the CPA case ($\sim 1/n$ at large $n$ as described in the text, here as a green, solid line). }
\label{fig_CPR_index}
\end{figure}
This is also the case with the CPA
threshold, which (see Eq.(7) of Ref.~\cite{chong10.01}) for an absorbing
slab dielectric of index $n = n_0+in_{abs}$ and length $L$ is:
\begin{equation}
e^{ink_0L} = \pm{{(n-1)}\over{(n+1)}}.
\label{CPA_basic}
\end{equation}
In the large $n_0$ limit, because the log is vanishing as $\sim 1/n_0^2$, the threshold $n_{abs}$ must decrease as $\sim 1/n_0$ at large $n_0$. A graph of this
reduction of CPA in a bulk absorber from Eq. ~(\ref{CPA_basic}) is included
in Fig. \ref{CPR_OFOgraph}b. We note in passing that this reduction is what one would expect for the single transit time reduction and not that associated with the etaloning as was the case for the optical cavity-assisted reduction in the thresholds.
In CPA, the index of
refraction real and imaginary parts can be considered as the time-even and time-odd contributions
to the transport. The analogous processes for the transport of the polarization
are optical activity (time-even) and Faraday rotation (time-odd).
Recall that one cannot achieve CPR with optical activity
alone, but the question we would like to address is how the presence of
optical activity in a system modifies the threshold Faraday rotation needed
for CPR.
Consider
a system with both of these processes operating. Instead of a single bulk
piece, for simplicity
we analyze a three-layer system composed of two optically active
blocks
with a Faraday rotator in between. (See discussion below Eq.~(\ref{Mdielectric_2by2}) for the matrix representation of optical activity.)
It is then straightforward to
identify the CPR state in this system,
again in terms of the equation $det(R)=0$, where
the matrix elements of the 2$\times$2 complex matrix $R$ are as in
Eqs.~(\ref{CPR_simpler1})-(\ref{CPR_simpler4}),
but where we make the substitutions for $M$ and $C$ via;
\begin{equation}
\left( \begin{array}{c} M \\ C \end{array} \right) =
\left( \begin{array}{cc}
\cos 2\alpha & -\sin 2\alpha \\
\sin 2\alpha & \cos 2\alpha \end{array} \right)
\left( \begin{array}{c} M_0MM_0 \\ M_0CM_0 \end{array} \right) \, ,
\label{CPR_OFO}
\end{equation}
where $M_0$ is given by Eq.~(\ref{Mdielectric_2by2}) for the optically active blocks (with chiral density proportional to $\alpha$) and the $M$ and $C$
on the RHS of Eq.~(\ref{CPR_OFO}) are given by Eq.~(\ref{Mdielectric}) and Eq.~(\ref{MnCdielectric}),
respectively, for the Faraday block.
Keeping the indices and length the same, but changing only the optical
activity, we can determine the location of the CPR state
(see Fig. \ref{CPR_OFOgraph}(a)).
We see that, as in the decrease of the
CPA threshold with increasing real part of the refractive index $n_0$, the Faraday rotation
needed to achieve CPR resonance decreases monotonically as one increases
the optical rotation in the adjoining slabs.
We note that this reduction continues with increasing optical activity beyond the
value at which the optical rotary part of the assembly
by itself would rotate a single input ray to its orthogonal polarization (rotation by $\pi/2$)
upon exiting in transmission. This result is true for both
positive and negative Verdet irrespective of the handedness of the optical activity; the trace in Fig. \ref{CPR_OFOgraph}(a) is
symmetric about zero optical activity. Both of these (CPR and CPA)
curves asymptote to zero threshold.
\begin{figure}[t]
\includegraphics[width=\linewidth]{fig9alt2
\caption{Lowest CPR and CPA resonance thresholds decrease monotonically to zero
as one increases the time-even part of the transport. (a) Reduction in CPR threshold for a Faraday (time odd) constituent sandwiched between two optically active slabs as a function of total optical activity of the time even part standing alone (in radians). (b) Reduction in CPA threshold as a function of the real part of the material's refractive index. }
\label{CPR_OFOgraph}
\end{figure}
This shows that increasing the time-even part of
an optical process reduces the time-odd
threshold for achieving CPR/CPA, and is expected to be useful for reducing
the size, complexity and cost of devices based on CPR or CPA, for example by
reducing the required magnetic field.
\section{Conclusions}
CPR and CPA are phenomenologically congruent in how their thresholds depend on the system's symmetry, composition and geometry. As both are coherent perfect processes, this congruence follows from the underlying commonality they share through wave interference and discrete symmetry. Furthermore, this study reveals potential design routes to decrease the size and/or magnetic field requirements for achieving CPR. For example, as detailed above, multilayering the rotating species can yield a 30-fold reduction in the naive length-field product. Similarly, even a poor optical cavity with just 60\% reflective mirrors reduces the CPR threshold length-field product by nearly 80\%. By layering with suitable optically active materials, high index materials, tertiary layered systems and layered systems with small layer thickness variations in the stack, we have shown further reduction in the naive length-field product is achievable in CPR-based devices.
\acknowledgments
The authors are grateful to the National Science Foundation for
financial support under grant number ECCS-1360725 and for
financial support from the Science and Technology Center for
Layered Polymeric Systems under grant number DMR
0423914.
|
\section{Introduction}
Local realism was a model introduced by Einstein, Podolsky, and Rosen \cite{EPR} in an attempt to reach agreement between
predictions of quantum mechanics with our classical intuition. However, in a simple argument Bell has demonstrated \cite{Bell}
that this attempt must necessarily fail. Not only has the contradiction between the assumptions of local realism and quantum
mechanics deep fundamental meaning, it also lies at the very heart of more efficient solutions to certain specific communication
tasks \cite{bellreview}.
It is then natural to confront these two contradicting theories in an experiment. For the Bell theorem, many systems have been
studied, but quantum interferometry is the most common field of implementation. However, any experiment will suffer from some
imperfections, which could be both of a technical nature (to be removed by more effort of the experimentalist) and intrinsically
irremovable. They degrade the quality of quantum correlations, making them more similar to those reproducible by local operations and classical communication. This fact is reflected in the theoretical description by mixing the desired state with a certain
noise. The simplest noise model is the so-called white noise. It has a uniform structure and degrades all correlations of the
state by a constant factor, $v$, which corresponds to Michelson's interference visibility.
Despite lack of a physical motivation in some cases, robustness of Bell inequality violation against white noise is often
considered a benchmark of the strength of nonclassicality of states. In this contribution we conduct a numerical investigation of multiqubit states that are the most robust against this imperfection, in the framework of two measurements per side. We find
that the most noise is tolerated in protocols similar to those described in Ref.~\cite{SEN}, where a group of observers simply
projects on fixed local states. More surprisingly, the most robust states are products of GHZ and pure one-qubit states.
Lately, considerable research effort has been dedicated to the study of noise robustness of nonlocal multipartite entangled sattes (see e.g., \cite{decoherence},\cite{dickenoise}). The present study can be considered as part of this research program.
\section{Noise properties of the GHZ and Dicke states}
At the beginning we analyze how correlations of two prominent families of states are resistant to white noise admixture. The
states which we consider are: the $N$-qubit GHZ state \cite{GHZ}: \begin{equation} \ket{{\rm GHZ}_N} =
\frac{1}{\sqrt{2}}(\ket{0\dots 0} + \ket{1\dots 1}) \end{equation} and the $N$-qubit Dicke states (with $e$ excitations)~\cite{Dicke}:
\begin{equation}
\ket{D^e_N} = \frac{1}{\sqrt{{N \choose e}}} \sum_{\pi} \ket{\pi(0 \dots 0\underbrace{1 \dots 1}_{e}0\dots 0)},
\label{DEN}
\end{equation}
where $\pi$ denotes a permutation of $e$ ones and $(N-e)$ zeros in the ket, and ${N \choose e}$ gives
a number of such permutations. The special case of $e=1$ corresponds to N-qubit W states. Our task is to find, for a given state $\rho$, the critical value $v_{crit}$ of the parameter $v$
in the mixture: \begin{equation} \rho(v) = v \rho + \frac{1-v}{2^N} \openone. \end{equation} If $v > v_{crit}$, there does not
exist any local realistic model describing quantum correlations of experimental events, i.e. the state violates a certain Bell
inequality. The parameter $v_{crit}$ is usually called ``critical visibility''.
\subsection{$N$-qubit GHZ state}
The GHZ states exhibit maximal violation of the Bell inequalities \cite{Mermin, WWWZB} for experiments with two alternative
measurements setting per party. A noisy $N$-qubit GHZ state, given by:
\begin{equation}
\rho_{GHZ}(v) = v_{GHZ} |GHZ_N\rangle
\langle GHZ_N| + \frac{1-v}{2^N} \openone \end{equation} leads to the critical visibility: \begin{equation}
v_{crit}^{GHZ} = \frac{1}{2^{(N-1)/2}}. \label{vGHZ}
\end{equation}
Actually, the set of inequalities in \cite{WWWZB} comprises all tight two-setting correlation-type Bell inequalities, hence the above threshold~(\ref{vGHZ}) for the visibilities defines the lowest one among all such correlations.
\subsection{$N$-qubit Dicke state}
For the case of the noisy Dicke state: \begin{equation} \rho_{D_N^e}(v) = v_{GHZ} |D_N^e\rangle \langle D_N^e| + \frac{1-v}{2^N}
\openone, \end{equation} we introduce a new Bell inequality $\langle C_N \rangle \leq 2$ for $N \geq 3$ in the following
iterative form: \begin{equation} C_N= (1-A^{(N)}_1) C_{N-1} + 2 A^{(N)}_1, \label{ineq_dicke} \end{equation} where $C_2 =
A^{(1)}_1 A^{(2)}_1+A^{(1)}_1 A^{(2)}_2+A^{(1)}_2 A^{(2)}_1-A^{(1)}_2 A^{(2)}_2$ is the CHSH expression \cite{CHSH} and
$A^{(i)}_j$ denotes a dichotomic observable measured by $i$th observer when he/she chooses $j$th measurement setting. Note that
($N-2$) observers perform only a single measurement $A^{(k)}_1~(k=3,\dots,N)$. The Bell expression (\ref{ineq_dicke}) can be
explicitly written as: \begin{eqnarray} C_N &=& C_2 + (1 - C_2) \left(\sum_{i=3}^N A^{(i)}_1 - \sum_{3 \leq i < j \leq N}
A^{(i)}_1 A^{(j)}_1 \nonumber \right.\\
&+& \sum_{3 \leq i < j < k \leq N} A^{(i)}_1 A^{(j)}_1 A^{(k)}_1 - \cdots \label{ineq_dicke1} \\
&+& \left. (-1)^{N+1} A^{(3)}_1 ... A^{(N)}_1 \right).\nonumber \end{eqnarray}
We also mention that this inequality reduces to the one appeared in \cite{BV} for $N=3$.
In order to find the quantum value of (\ref{ineq_dicke1}) let us choose measurement settings for the last ($N-2$) observers
$A^{(i)}_1 =-\sigma_z$ $(i=3,\dots, N)$. Note that due to the permutational symmetry of the Dicke states the correlations are the
same for any particular set of subsystems. The quantum value of the inequality (\ref{ineq_dicke1}) can read: \begin{eqnarray}
\avg{C_N}_{D_N^e} &=& \avg{C_2}_{D_N^e} + (1 - \avg{C_2}_{D_N^e})\nonumber \\ &\times& \sum_{k=1}^N (-1)^{2k+1} {N \choose k}
T_{\underbrace{z...z}_{k}\underbrace{0...0}_{N-k}}, \end{eqnarray} where the expectation value of the CHSH operator is equal to
$\avg{C_2}_{D_N^e} = 2\sqrt{2} \cdot \frac{2}{{N \choose e}} v$ and \begin{eqnarray} T_{\underbrace{z...z}_{k}0...0} &=&
\frac{v}{{N \choose e}} \sum_{j=0}^{\left\lceil e/2\right\rceil} \Big[{n-k\choose e-2j} {k\choose 2j}\nonumber \\ &-&{n-k\choose
e-(2j+1)}{k\choose 2j+1} \Big]. \end{eqnarray} After calculations we get: \begin{equation} \avg{C_N}_{D_N^e} = v
\left(2+\frac{2^{N} (\sqrt{2}-1)}{{N \choose e}}\right). \end{equation} Therefore, the critical visibility is equal to:
\begin{equation} v_{crit}^{_{D_N^e}} = \left(1+\frac{2^{N-1} (\sqrt{2}-1)}{{N \choose e}}\right)^{-1}. \label{vDicke}
\end{equation}
Comparing Eqs.~(\ref{vGHZ}) and (\ref{vDicke}) we can conclude that for any type of the Dicke state $D_N^e$ (for any arbitrary
number of excitations $e$), there is a critical number of particles $N^e_{crit}$ such that $v_{crit}$ for ${D_{N_{crit}}^e}$ is lower then critical visibility for the $N$-qubit GHZ state. For example: for $e=1$, $N^1_{crit}=11$; for $e=2$, $N^2_{crit}=19$; for
$e=5$, $N^3_{crit}=44$; for $e=10$, $N^4_{crit}=88$. It means that for $N>N^e_{crit}$ the Dicke states $D_N^e$ becomes more
robust against white noise admixture than the GHZ states in terms of violation of local realism. The value $N^1_{crit}=11$ for $e=1$ was the first time calculated in \cite{SEN} with some projective method. These results are in good agreement with the recent ones~\cite{dickenoise} where the most robust Dicke states subject to losses are found to correspond to only few excitations $e$.
\subsection{Numerical method}
We also use the numerical method (see e.g. \cite{SR}) based on linear programming to find critical visibilities for the
GHZ and W states up to 10 qubits in experiments with two alternative measurement settings per side. The method does not need any
knowledge at all of the forms of Bell inequalities. However, the results obtained by this method are equivalent to the analysis
of the full set of (probabilistic) Bell inequalities formulated for a given experimental situation. The results are presented in
Table~\ref{ghzw}.
\begin{table} \begin{tabular}{ccc} \hline \hline N & $v_{crit}^{GHZ}$ & $v_{crit}^{W}$ \\ \hline 3& 0.5000& 0.6442\\ 4& 0.3536& 0.5294\\ 5&
0.2500& 0.4018\\ 6& 0.1768& 0.2774\\ 7& 0.1250& 0.1736\\ 8& 0.0884& 0.1034\\ 9& 0.0625& 0.0578\\ 10& 0.0442& 0.0313 \\\hline \hline \end{tabular}
\caption{\label{ghzw} The critical visibilities for the GHZ and W states for $N\leq 10$ and two measurement settings per
observer. If $v > v_{crit}$, there does not exist any local realistic model describing quantum probabilities of experimental
events. For $N \geq 9$ the W states lead to lower critical visibility than the GHZ states.} \end{table}
Analyzing the critical values one can see that the W state is more resistant to white noise than the GHZ state already for
$N=9$. This result is stronger than the one obtained by means of the inequality (\ref{ineq_dicke}) or the projective method
\cite{SEN}.
\section{Partially product states are highly resistant to noise}
We also apply the linear programming method to analyze noise resistance of $k$-product states of $N$ qubits in the following
form: \begin{equation} |\psi^{k-prod}_N \rangle = |\underbrace{0 \dots 0}_{1 \dots k} \rangle |GHZ_{N-k} \rangle.
\label{state_prod} \end{equation} The results for two measurement settings per party are presented in Tab. \ref{product}.
Surprisingly, the obtained critical parameters $v_{crit}$ are very low and lower then the corresponding ones for the $N$-qubit
GHZ state. The surprise is due to the fact that the $N$-qubit GHZ state is genuinely $N$-partite entangled and maximizes many
entanglement conditions and measures \cite{HORODECKI-review}.
Moreover, for $2\leq N \leq 6$ we identify states that lead to the lowest critical visibility. This was possible after including
an optimization over {\em all pure states} in the numerical method. The optimal states are: for $N=2$, $\ket{GHZ_2}$; for $N=3$,
$\ket{GHZ_3}$; for $N=4$, $\ket{0}_1 \ket{GHZ_3}$; for $N=5$, $\ket{00}_{12}\ket{GHZ_3}$; and for $N=6$,
$\ket{000}_{123}\ket{GHZ_3}$. These states are the most robust states against white noise admixture.
\begin{table} \begin{tabular}{lll} \hline \hline $N$ & State & $v_{crit}$ \\ \hline 2 & $\ket{GHZ}_{AB}$ & {\bf 0.707}\\ \hline 3
& $\ket{0}_A \ket{GHZ}_{BC}$ & 0.547\\
& $\ket{GHZ}_{ABC}$ & {\bf 0.500}\\ \hline
4 & $\ket{0}_A \ket{GHZ}_{BCD}$ & {\bf 0.333}\\ & $\ket{00}_{AB} \ket{GHZ}_{CD}$ & 0.377\\ & $\ket{GHZ}_{ABCD}$ & 0.354\\ \hline
5 &$\ket{0}_A\ket{GHZ}_{BCDE}$ & 0.215 \\ &$\ket{00}_{AB}\ket{GHZ}_{CDE}$ & {\bf 0.200} \\ &$\ket{000}_{ABC}\ket{GHZ}_{DE}$ &
0.232 \\ &$\ket{GHZ}_{ABCDE}$ & 0.250 \\ \hline 6 &$\ket{0}_A\ket{GHZ}_{BCDEF}$ & 0.144\\ &$\ket{00}_{AB}\ket{GHZ}_{CDEF}$ &
0.121\\ &$\ket{000}_{ABC}\ket{GHZ}_{DEF}$ & {\bf 0.111}\\ &$\ket{0000}_{ABCD}\ket{GHZ}_{EF}$ & 0.131 \\ &$\ket{GHZ}_{ABCDEF}$ &
0.177 \\ \hline 7 &$\ket{0}_A\ket{GHZ}_{BCDEFG}$ & 0.098 \\ &$\ket{00}_{AB}\ket{GHZ}_{CDEFG}$ & 0.078 \\
&$\ket{000}_{ABC}\ket{GHZ}_{DEFG}$ & 0.064\\ &$\ket{0000}_{ABCD}\ket{GHZ}_{EFG}$ & 0.059 \\ &$\ket{00000}_{ABCDE}\ket{GHZ}_{FG}$
& 0.070 \\ &$\ket{GHZ}_{ABCDEFG}$ & 0.125 \\ \hline 8&$\ket{000}_{ABC}\ket{GHZ}_{DEFGH}$ & 0.041 \\
&$\ket{0000}_{ABCD}\ket{GHZ}_{EFGH}$ & 0.033 \\ &$\ket{00000}_{ABCDE}\ket{GHZ}_{FGH}$& 0.030 \\
&$\ket{00000}_{ABCDEF}\ket{GHZ}_{GH}$& 0.036 \\ &$\ket{GHZ}_{ABCDEFG}$& 0.088\\ \hline \hline \end{tabular}
\caption{\label{product} Comparison of the critical visibilities for the $k$-product states of the form (\ref{state_prod}) with
corresponding values for the GHZ states for experiments with two measurement settings per observer. The states that appear in bold are the most resistant to white noise for a given number of qubits.} \end{table}
\subsection{Inequalities}
We present two families of the Bell-type inequalities, which recover the results presented in Table~\ref{product} for the states:
$\ket{\psi_{(N-2)-prod}^N} = \ket{0...0}_{1...N-2}\ket{GHZ_2}$ and $\ket{\psi_{(N-3)-prod}^N} =
\ket{0...0}_{1...N-3}\ket{GHZ_3}$.
The first inequality, optimal for the state $\ket{\psi_{(N-2)-prod}^N}$, has the same form as the inequality given in Eq.
(\ref{ineq_dicke}). Taking again $A^{(i)} = -\sigma_z$ for $2<i\leq N$ we obtain $\avg{A^{(i)}}_{\psi_{(N-2)-prod}^N} =-1$. The
expectation value of $\avg{C_2}_{\psi_{(N-2)-prod}^N} = 2\sqrt{2}$. Then the critical visibility is given by: \begin{equation}
v_{crit}^{\psi_{(N-2)-prod}^N} = \frac{1}{1+(\sqrt{2}-1)2^{N-2}} \end{equation} and for $N>N_{crit} = \left\lceil \log_2(12 + 8
\sqrt{2}) \right\rceil = 5$ is lower than for the GHZ state.
The second inequality, optimal for the state $\ket{\psi_{(N-3)-prod}^N}$ has a form $\langle M_N \rangle \leq 2$ for $N \geq 4$,
where \begin{equation} C_N= (1-A^{(N)}_1) M_{N-1} + 2 A^{(N)}_1 \label{ineq_mermin} \end{equation} and $M_{3} = - A^{(1)}_1
A^{(2)}_1 A^{(3)}_1 + A^{(1)}_1 A^{(2)}_2 A^{(3)}_2 + A^{(1)}_2 A^{(2)}_2 A^{(3)}_1 + A^{(1)}_2 A^{(2)}_1 A^{(3)}_2$ is the
Mermin expression \cite{Mermin}. The observers choose again their measurements as $A^{(i)} = -\sigma_z$ for $3<i\leq N$. We
obtain $\avg{A^{(i)}}_{\psi_{(N-3)-prod}^N} =-1$ and the expectation value of $\avg{C_3}_{\psi_{(N-3)-prod}^N} = 4$. Therefore,
the critical visibility is equal to: \begin{equation} v_{crit}^{\psi_{(N-3)-prod}^N} = \frac{8}{8+2^N}, \end{equation} and is
lower than the corresponding value for the GHZ state for $N\geq 4$.
\section{A single Bell inequality violated by any pure entangled state}
Our construction is based on the inequality~(\ref{ineq_dicke1}). It is not difficult to see that this inequality is equivalent to the following one:
\begin{equation}
CH^{(1,2)}\prod_{k=3}^{N}p(A^{(k)}_1)\leq 0,
\label{Wineq}
\end{equation}
where $CH^{(i,j)} = p(A^{(i)}_2, A^{(j)}_2) + p(A^{(i)}_2, A^{(j)}_3) + p(A^{(i)}_3, A^{(j)}_2) - p(A^{(i)}_3, A^{(j)}_3) - p(A_2^{(i)}) - p(A_3^{(j)})$ is the Clauser-Horne expression \cite{CH}, $p(A_k^{(i)})$ denotes the probability of obtaining a ``1'' result by the $i$-th observer choosing observable $A_k$ and $p(A^{(i)}_l, A^{(j)}_m)$ is the probability of obtaining ``1'' by $i$th and $j$th observers choosing $A_l$ and $A_m$ observable, respectively.
From the above Bell inequality~(\ref{Wineq}), we construct the following symmetrized one (for any $N\ge 3$):
\begin{equation}
\sum_{1\le i<j\le N}{CH^{(i,j)}\prod_{\substack{{k=1}\\{k\neq i,j}}}^N p(A^{(k)}_1)}\leq 0.
\label{Wineqsym}
\end{equation}
This inequality involves three binary outcome settings ($A_1, A_2, A_3$) for each party. For the simplest case of three observables ($N=3$), it looks as follows:
\begin{equation}
CH^{(1,2)}p(A^{(3)}_1)+CH^{(1,3)}p(A^{(2)}_1)+CH^{(2,3)}p(A^{(1)}_1)\leq 0.
\label{Wineqsym3}
\end{equation}
In order to demonstrate that the single N-party Bell inequality~(\ref{Wineqsym}) is violated by any N-party pure entangled states, we make use of the result of Popescu and Rohrlich \cite{PR} (see also \cite{Cavalcanti}). They showed that for any N-party pure entangled state there exist $N-2$ local projections which leave the remaining two systems (say, systems $i$ and $j$) in a pure entangled state. Since any pair of pure entangled states violate the Clauser-Horne (CH) inequality~\cite{Gisin}, it implies that inequality~(\ref{Wineq}) is violated by this state (where the successful projections are associated with outcomes ``1''). The fact that $i$ and $j$ can be any two systems out of the N systems is captured by the symmetrized inequality~(\ref{Wineqsym}). Indeed, let the parties $i$ and $j$ output 0 for their first measurement (i.e. they carry out a degenerate measurement resulting in $p(A^{(i)}_1)=0$ and $p(A^{(j)}_1)=0$), whereas the successful projections on the rest of the parties are associated with outcome ``1''. In this case the only term which survives in (\ref{Wineqsym}) is the one of (\ref{Wineq}), the other terms giving zero contribution.
Though there exists a more economical Bell inequality in terms of number of settings (it consists of only two settings per party) which is violated by any pure entangled states~\cite{Yu}, we believe the present proof based on our new inequality~(\ref{Wineqsym}) is simpler.
\section{Conclusions}
We have analyzed quantum states, which offer the highest Bell inequality violation robustness against an admixture of the withe noise. This robustness was studied in a specific experimental scenario, with two measurement settings per side. Up to $N=8$ we find that the optimal states are products of three-qubit GHZ states and $(N-3)$ pure single qubit states. The corresponding inequalities represent a protocol, in which $(N-3)$ observers attempt to locally project the states, while the remaining three conduct a GHZ experiment. This result challenges common, but contradictory beliefs: that entanglement distributed among many parties is more strongly nonclassical, and that the robustness against white noise is a proper measure of quantumness.
\begin{acknowledgments}
This work is a part of the FNP TEAM project cofinanced by the EU Regional Development Fund.
W.L. and M.W. are supported by the Polish Ministry of Science and Higher Education Grant No. IdP2011 000361.
T.V. acknowledges financial support from a J\'anos Bolyai Grant of the Hungarian Academy of Sciences,
the Hungarian National Research Fund OTKA (K111734), and SEFRI (COST action MP1006).
\end{acknowledgments}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.